diff --git a/Cargo.lock b/Cargo.lock
index eb53ed5..f88965a 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -165,6 +165,39 @@ dependencies = [
  "tree-sitter-yaml",
 ]
 
+[[package]]
+name = "async-stream"
+version = "0.3.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0b5a71a6f37880a80d1d7f19efd781e4b5de42c88f0722cc13bcb6cc2cfe8476"
+dependencies = [
+ "async-stream-impl",
+ "futures-core",
+ "pin-project-lite",
+]
+
+[[package]]
+name = "async-stream-impl"
+version = "0.3.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c7c24de15d275a1ecfd47a380fb4d5ec9bfe0933f309ed5e705b775596a3574d"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.117",
+]
+
+[[package]]
+name = "async-trait"
+version = "0.1.89"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9035ad2d096bed7955a320ee7e2230574d28fd3c3a0f186cbea1ff3c7eed5dbb"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.117",
+]
+
 [[package]]
 name = "atomic"
 version = "0.6.1"
@@ -208,6 +241,53 @@ dependencies = [
  "fs_extra",
 ]
 
+[[package]]
+name = "axum"
+version = "0.7.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "edca88bc138befd0323b20752846e6587272d3b03b0343c8ea28a6f819e6e71f"
+dependencies = [
+ "async-trait",
+ "axum-core",
+ "bytes",
+ "futures-util",
+ "http",
+ "http-body",
+ "http-body-util",
+ "itoa",
+ "matchit",
+ "memchr",
+ "mime",
+ "percent-encoding",
+ "pin-project-lite",
+ "rustversion",
+ "serde",
+ "sync_wrapper",
+ "tower 0.5.3",
+ "tower-layer",
+ "tower-service",
+]
+
+[[package]]
+name = "axum-core"
+version = "0.4.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "09f2bd6146b97ae3359fa0cc6d6b376d9539582c7b4220f041a33ec24c226199"
+dependencies = [
+ "async-trait",
+ "bytes",
+ "futures-util",
+ "http",
+ "http-body",
+ "http-body-util",
+ "mime",
+ "pin-project-lite",
+ "rustversion",
+ "sync_wrapper",
+ "tower-layer",
+ "tower-service",
+]
+
 [[package]]
 name = "base64"
 version = "0.13.1"
@@ -372,6 +452,12 @@ dependencies = [
  "shlex",
 ]
 
+[[package]]
+name = "cesu8"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6d43a04d8753f35258c91f8ec639f792891f748a1edbd759cf1dcea3382ad83c"
+
 [[package]]
 name = "cfg-if"
 version = "1.0.4"
@@ -453,6 +539,16 @@ version = "1.0.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1d07550c9036bf2ae0c684c4297d503f838287c83c53686d05370d0e139ae570"
 
+[[package]]
+name = "combine"
+version = "4.6.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ba5a308b75df32fe02788e748662718f03fde005016435c444eea572398219fd"
+dependencies = [
+ "bytes",
+ "memchr",
+]
+
 [[package]]
 name = "compact_str"
 version = "0.9.0"
@@ -475,6 +571,7 @@ dependencies = [
  "anyhow",
  "ast-grep-core",
  "ast-grep-language",
+ "async-stream",
  "base64 0.22.1",
  "bytes",
  "capnp",
@@ -488,22 +585,28 @@ dependencies = [
  "figment",
  "futures",
  "glob",
+ "html2md",
  "http",
  "http-body-util",
  "hyper",
  "hyper-util",
- "json5",
+ "imagesize",
+ "json-five",
  "libc",
  "log",
  "memchr",
  "memmap2",
+ "notify-debouncer-mini",
  "paste",
  "peg",
+ "prost",
+ "protoc-bin-vendored",
  "ratatui",
  "redb",
  "regex",
  "rustls",
  "rustls-native-certs",
+ "rustls-pemfile",
  "serde",
  "serde_json",
  "serde_urlencoded",
@@ -512,7 +615,10 @@ dependencies = [
  "tokenizers",
  "tokio",
  "tokio-rustls",
+ "tokio-stream",
  "tokio-util",
+ "tonic",
+ "tonic-build",
  "tui-markdown",
  "tui-textarea-2",
  "uuid",
@@ -1045,6 +1151,12 @@ version = "0.4.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0ce7134b9999ecaf8bcd65542e436736ef32ddca1b3e06094cb6ec5755203b80"
 
+[[package]]
+name = "fixedbitset"
+version = "0.5.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1d674e81391d1e1ab681a28d99df07927c6d4aa5b027d7da16ba32d1d21ecd99"
+
 [[package]]
 name = "flate2"
 version = "1.1.9"
@@ -1088,6 +1200,25 @@ version = "1.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "42703706b716c37f96a77aea830392ad231f44c9e9a67872fa5548707e11b11c"
 
+[[package]]
+name = "fsevent-sys"
+version = "4.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "76ee7a02da4d231650c7cea31349b889be2f45ddb3ef3032d2ec8185f6313fd2"
+dependencies = [
+ "libc",
+]
+
+[[package]]
+name = "futf"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "df420e2e84819663797d1ec6544b13c5be84629e7bb00dc960d6917db2987843"
+dependencies = [
+ "mac",
+ "new_debug_unreachable",
+]
+
 [[package]]
 name = "futures"
 version = "0.3.32"
@@ -1250,6 +1381,31 @@ dependencies = [
  "regex-syntax",
 ]
 
+[[package]]
+name = "h2"
+version = "0.4.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2f44da3a8150a6703ed5d34e164b875fd14c2cdab9af1252a9a1020bde2bdc54"
+dependencies = [
+ "atomic-waker",
+ "bytes",
+ "fnv",
+ "futures-core",
+ "futures-sink",
+ "http",
+ "indexmap 2.14.0",
+ "slab",
+ "tokio",
+ "tokio-util",
+ "tracing",
+]
+
+[[package]]
+name = "hashbrown"
+version = "0.12.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8a9ee70c43aaf417c914396645a0fa852624801b24ebb7ae78fe8272889ac888"
+
 [[package]]
 name = "hashbrown"
 version = "0.15.5"
@@ -1288,6 +1444,34 @@ version = "0.4.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7f24254aa9a54b5c858eaee2f5bccdb46aaf0e486a595ed5fd8f86ba55232a70"
 
+[[package]]
+name = "html2md"
+version = "0.2.15"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8cff9891f2e0d9048927fbdfc28b11bf378f6a93c7ba70b23d0fbee9af6071b4"
+dependencies = [
+ "html5ever",
+ "jni",
+ "lazy_static",
+ "markup5ever_rcdom",
+ "percent-encoding",
+ "regex",
+]
+
+[[package]]
+name = "html5ever"
+version = "0.27.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c13771afe0e6e846f1e67d038d4cb29998a6779f93c809212e4e9c32efd244d4"
+dependencies = [
+ "log",
+ "mac",
+ "markup5ever",
+ "proc-macro2",
+ "quote",
+ "syn 2.0.117",
+]
+
 [[package]]
 name = "http"
 version = "1.4.0"
@@ -1327,6 +1511,12 @@ version = "1.10.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6dbf3de79e51f3d586ab4cb9d5c3e2c14aa28ed23d180cf89b4df0454a69cc87"
 
+[[package]]
+name = "httpdate"
+version = "1.0.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "df3b46402a9d5adb4c86a0cf463f42e19994e3ee891101b1841f30a545cb49a9"
+
 [[package]]
 name = "hyper"
 version = "1.9.0"
@@ -1337,9 +1527,11 @@ dependencies = [
  "bytes",
  "futures-channel",
  "futures-core",
+ "h2",
  "http",
  "http-body",
  "httparse",
+ "httpdate",
  "itoa",
  "pin-project-lite",
  "smallvec",
@@ -1347,6 +1539,19 @@ dependencies = [
  "want",
 ]
 
+[[package]]
+name = "hyper-timeout"
+version = "0.5.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2b90d566bffbce6a75bd8b09a05aa8c2cb1fabb6cb348f8840c9e4c90a0d83b0"
+dependencies = [
+ "hyper",
+ "hyper-util",
+ "pin-project-lite",
+ "tokio",
+ "tower-service",
+]
+
 [[package]]
 name = "hyper-util"
 version = "0.1.20"
@@ -1354,11 +1559,17 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "96547c2556ec9d12fb1578c4eaf448b04993e7fb79cbaad930a656880a6bdfa0"
 dependencies = [
  "bytes",
+ "futures-channel",
+ "futures-util",
  "http",
  "http-body",
  "hyper",
+ "libc",
  "pin-project-lite",
+ "socket2 0.6.3",
  "tokio",
+ "tower-service",
+ "tracing",
 ]
 
 [[package]]
@@ -1413,6 +1624,22 @@ dependencies = [
  "winapi-util",
 ]
 
+[[package]]
+name = "imagesize"
+version = "0.14.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "09e54e57b4c48b40f7aec75635392b12b3421fa26fe8b4332e63138ed278459c"
+
+[[package]]
+name = "indexmap"
+version = "1.9.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bd070e393353796e801d209ad339e89596eb4c8d430d18ede6a1cced8fafbd99"
+dependencies = [
+ "autocfg",
+ "hashbrown 0.12.3",
+]
+
 [[package]]
 name = "indexmap"
 version = "2.14.0"
@@ -1453,6 +1680,26 @@ version = "0.1.15"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c8fae54786f62fb2918dcfae3d568594e50eb9b5c25bf04371af6fe7516452fb"
 
+[[package]]
+name = "inotify"
+version = "0.11.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bd5b3eaf1a28b758ac0faa5a4254e8ab2705605496f1b1f3fbbc3988ad73d199"
+dependencies = [
+ "bitflags 2.11.0",
+ "inotify-sys",
+ "libc",
+]
+
+[[package]]
+name = "inotify-sys"
+version = "0.1.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e05c02b5e89bff3b946cedeca278abc628fe811e604f027c45a8aa3cf793d0eb"
+dependencies = [
+ "libc",
+]
+
 [[package]]
 name = "instability"
 version = "0.3.12"
@@ -1511,6 +1758,48 @@ dependencies = [
  "syn 2.0.117",
 ]
 
+[[package]]
+name = "jni"
+version = "0.19.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c6df18c2e3db7e453d3c6ac5b3e9d5182664d28788126d39b91f2d1e22b017ec"
+dependencies = [
+ "cesu8",
+ "combine",
+ "jni-sys 0.3.1",
+ "log",
+ "thiserror 1.0.69",
+ "walkdir",
+]
+
+[[package]]
+name = "jni-sys"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "41a652e1f9b6e0275df1f15b32661cf0d4b78d4d87ddec5e0c3c20f097433258"
+dependencies = [
+ "jni-sys 0.4.1",
+]
+
+[[package]]
+name = "jni-sys"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c6377a88cb3910bee9b0fa88d4f42e1d2da8e79915598f65fb0c7ee14c878af2"
+dependencies = [
+ "jni-sys-macros",
+]
+
+[[package]]
+name = "jni-sys-macros"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "38c0b942f458fe50cdac086d2f946512305e5631e720728f2a61aabcd47a6264"
+dependencies = [
+ "quote",
+ "syn 2.0.117",
+]
+
 [[package]]
 name = "jobserver"
 version = "0.1.34"
@@ -1531,6 +1820,16 @@ dependencies = [
  "wasm-bindgen",
 ]
 
+[[package]]
+name = "json-five"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "865f2d01a4549c1fd8c60640c03ae5249eb374cd8cde8b905628d4b1af95c87c"
+dependencies = [
+ "serde",
+ "unicode-general-category",
+]
+
 [[package]]
 name = "json5"
 version = "1.3.1"
@@ -1552,6 +1851,26 @@ dependencies = [
  "thiserror 2.0.18",
 ]
 
+[[package]]
+name = "kqueue"
+version = "1.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "eac30106d7dce88daf4a3fcb4879ea939476d5074a9b7ddd0fb97fa4bed5596a"
+dependencies = [
+ "kqueue-sys",
+ "libc",
+]
+
+[[package]]
+name = "kqueue-sys"
+version = "1.0.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ed9625ffda8729b85e45cf04090035ac368927b8cebc34898e7c120f52e4838b"
+dependencies = [
+ "bitflags 1.3.2",
+ "libc",
+]
+
 [[package]]
 name = "lab"
 version = "0.11.0"
@@ -1636,6 +1955,12 @@ dependencies = [
  "hashbrown 0.16.1",
 ]
 
+[[package]]
+name = "mac"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c41e0c4fef86961ac6d6f8a82609f55f31b05e4fce149ac5710e439df7619ba4"
+
 [[package]]
 name = "mac_address"
 version = "1.1.8"
@@ -1662,6 +1987,38 @@ version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "670fdfda89751bc4a84ac13eaa63e205cf0fd22b4c9a5fbfa085b63c1f1d3a30"
 
+[[package]]
+name = "markup5ever"
+version = "0.12.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "16ce3abbeba692c8b8441d036ef91aea6df8da2c6b6e21c7e14d3c18e526be45"
+dependencies = [
+ "log",
+ "phf",
+ "phf_codegen",
+ "string_cache",
+ "string_cache_codegen",
+ "tendril",
+]
+
+[[package]]
+name = "markup5ever_rcdom"
+version = "0.3.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "edaa21ab3701bfee5099ade5f7e1f84553fd19228cf332f13cd6e964bf59be18"
+dependencies = [
+ "html5ever",
+ "markup5ever",
+ "tendril",
+ "xml5ever",
+]
+
+[[package]]
+name = "matchit"
+version = "0.7.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0e7465ac9959cc2b1404e8e2367b43684a6d13790fe23056cc8c6c5a6b7bcb94"
+
 [[package]]
 name = "memchr"
 version = "2.8.0"
@@ -1692,6 +2049,12 @@ dependencies = [
  "autocfg",
 ]
 
+[[package]]
+name = "mime"
+version = "0.3.17"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "6877bb514081ee2a7ff5ef9de3281f14a4dd4bceac4c09388074a6b5df8a139a"
+
 [[package]]
 name = "minimal-lexical"
 version = "0.2.1"
@@ -1742,6 +2105,18 @@ dependencies = [
  "syn 2.0.117",
 ]
 
+[[package]]
+name = "multimap"
+version = "0.10.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1d87ecb2933e8aeadb3e3a02b828fed80a7528047e68b4f424523a0981a3a084"
+
+[[package]]
+name = "new_debug_unreachable"
+version = "1.0.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "650eef8c711430f1a879fdd01d4745a7deea475becfb90269c06775983bbf086"
+
 [[package]]
 name = "nix"
 version = "0.29.0"
@@ -1774,6 +2149,45 @@ dependencies = [
  "memchr",
 ]
 
+[[package]]
+name = "notify"
+version = "8.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4d3d07927151ff8575b7087f245456e549fea62edf0ec4e565a5ee50c8402bc3"
+dependencies = [
+ "bitflags 2.11.0",
+ "fsevent-sys",
+ "inotify",
+ "kqueue",
+ "libc",
+ "log",
+ "mio",
+ "notify-types",
+ "walkdir",
+ "windows-sys 0.60.2",
+]
+
+[[package]]
+name = "notify-debouncer-mini"
+version = "0.7.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "17849edfaabd9a5fef1c606d99cfc615a8e99f7ac4366406d86c7942a3184cf2"
+dependencies = [
+ "log",
+ "notify",
+ "notify-types",
+ "tempfile",
+]
+
+[[package]]
+name = "notify-types"
+version = "2.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "42b8cfee0e339a0337359f3c88165702ac6e600dc01c0cc9579a92d62b08477a"
+dependencies = [
+ "bitflags 2.11.0",
+]
+
 [[package]]
 name = "num-conv"
 version = "0.2.1"
@@ -1992,6 +2406,16 @@ dependencies = [
  "sha2",
 ]
 
+[[package]]
+name = "petgraph"
+version = "0.7.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3672b37090dbd86368a4145bc067582552b29c27377cad4e0a306c97f9bd7772"
+dependencies = [
+ "fixedbitset 0.5.7",
+ "indexmap 2.14.0",
+]
+
 [[package]]
 name = "phf"
 version = "0.11.3"
@@ -2044,6 +2468,26 @@ dependencies = [
  "siphasher",
 ]
 
+[[package]]
+name = "pin-project"
+version = "1.1.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f1749c7ed4bcaf4c3d0a3efc28538844fb29bcdd7d2b67b2be7e20ba861ff517"
+dependencies = [
+ "pin-project-internal",
+]
+
+[[package]]
+name = "pin-project-internal"
+version = "1.1.11"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d9b20ed30f105399776b9c883e68e536ef602a16ae6f596d2c473591d6ad64c6"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.117",
+]
+
 [[package]]
 name = "pin-project-lite"
 version = "0.2.17"
@@ -2063,7 +2507,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "740ebea15c5d1428f910cd1a5f52cebf8d25006245ed8ade92702f4943d91e07"
 dependencies = [
  "base64 0.22.1",
- "indexmap",
+ "indexmap 2.14.0",
  "quick-xml",
  "serde",
  "time",
@@ -2099,6 +2543,12 @@ dependencies = [
  "zerocopy",
 ]
 
+[[package]]
+name = "precomputed-hash"
+version = "0.1.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "925383efa346730478fb4838dbe9137d2a47675ad789c546d150a6e1dd4ab31c"
+
 [[package]]
 name = "prettyplease"
 version = "0.2.37"
@@ -2131,6 +2581,122 @@ dependencies = [
  "yansi",
 ]
 
+[[package]]
+name = "prost"
+version = "0.13.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "2796faa41db3ec313a31f7624d9286acf277b52de526150b7e69f3debf891ee5"
+dependencies = [
+ "bytes",
+ "prost-derive",
+]
+
+[[package]]
+name = "prost-build"
+version = "0.13.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "be769465445e8c1474e9c5dac2018218498557af32d9ed057325ec9a41ae81bf"
+dependencies = [
+ "heck",
+ "itertools",
+ "log",
+ "multimap",
+ "once_cell",
+ "petgraph",
+ "prettyplease",
+ "prost",
+ "prost-types",
+ "regex",
+ "syn 2.0.117",
+ "tempfile",
+]
+
+[[package]]
+name = "prost-derive"
+version = "0.13.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8a56d757972c98b346a9b766e3f02746cde6dd1cd1d1d563472929fdd74bec4d"
+dependencies = [
+ "anyhow",
+ "itertools",
+ "proc-macro2",
+ "quote",
+ "syn 2.0.117",
+]
+
+[[package]]
+name = "prost-types"
+version = "0.13.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "52c2c1bf36ddb1a1c396b3601a3cec27c2462e45f07c386894ec3ccf5332bd16"
+dependencies = [
+ "prost",
+]
+
+[[package]]
+name = "protoc-bin-vendored"
+version = "3.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d1c381df33c98266b5f08186583660090a4ffa0889e76c7e9a5e175f645a67fa"
+dependencies = [
+ "protoc-bin-vendored-linux-aarch_64",
+ "protoc-bin-vendored-linux-ppcle_64",
+ "protoc-bin-vendored-linux-s390_64",
+ "protoc-bin-vendored-linux-x86_32",
+ "protoc-bin-vendored-linux-x86_64",
+ "protoc-bin-vendored-macos-aarch_64",
+ "protoc-bin-vendored-macos-x86_64",
+ "protoc-bin-vendored-win32",
+]
+
+[[package]]
+name = "protoc-bin-vendored-linux-aarch_64"
+version = "3.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c350df4d49b5b9e3ca79f7e646fde2377b199e13cfa87320308397e1f37e1a4c"
+
+[[package]]
+name = "protoc-bin-vendored-linux-ppcle_64"
+version = "3.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a55a63e6c7244f19b5c6393f025017eb5d793fd5467823a099740a7a4222440c"
+
+[[package]]
+name = "protoc-bin-vendored-linux-s390_64"
+version = "3.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1dba5565db4288e935d5330a07c264a4ee8e4a5b4a4e6f4e83fad824cc32f3b0"
+
+[[package]]
+name = "protoc-bin-vendored-linux-x86_32"
+version = "3.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8854774b24ee28b7868cd71dccaae8e02a2365e67a4a87a6cd11ee6cdbdf9cf5"
+
+[[package]]
+name = "protoc-bin-vendored-linux-x86_64"
+version = "3.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b38b07546580df720fa464ce124c4b03630a6fb83e05c336fea2a241df7e5d78"
+
+[[package]]
+name = "protoc-bin-vendored-macos-aarch_64"
+version = "3.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "89278a9926ce312e51f1d999fee8825d324d603213344a9a706daa009f1d8092"
+
+[[package]]
+name = "protoc-bin-vendored-macos-x86_64"
+version = "3.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "81745feda7ccfb9471d7a4de888f0652e806d5795b61480605d4943176299756"
+
+[[package]]
+name = "protoc-bin-vendored-win32"
+version = "3.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "95067976aca6421a523e491fce939a3e65249bac4b977adee0ee9771568e8aa3"
+
 [[package]]
 name = "pulldown-cmark"
 version = "0.13.3"
@@ -2186,6 +2752,8 @@ version = "0.8.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "34af8d1a0e25924bc5b7c43c079c942339d8f0a8b57c39049bef581b46327404"
 dependencies = [
+ "libc",
+ "rand_chacha 0.3.1",
  "rand_core 0.6.4",
 ]
 
@@ -2195,10 +2763,20 @@ version = "0.9.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6db2770f06117d490610c7488547d543617b21bfa07796d7a12f6f1bd53850d1"
 dependencies = [
- "rand_chacha",
+ "rand_chacha 0.9.0",
  "rand_core 0.9.5",
 ]
 
+[[package]]
+name = "rand_chacha"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e6c10a63a0fa32252be49d21e7709d4d4baf8d231c2dbce1eaa8141b9b127d88"
+dependencies = [
+ "ppv-lite86",
+ "rand_core 0.6.4",
+]
+
 [[package]]
 name = "rand_chacha"
 version = "0.9.0"
@@ -2214,6 +2792,9 @@ name = "rand_core"
 version = "0.6.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ec0be4795e2f6a28069bec0b5ff3e2ac9bafc99e6a9a7dc3547996c5c816922c"
+dependencies = [
+ "getrandom 0.2.17",
+]
 
 [[package]]
 name = "rand_core"
@@ -2462,6 +3043,15 @@ dependencies = [
  "security-framework",
 ]
 
+[[package]]
+name = "rustls-pemfile"
+version = "2.2.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dce314e5fee3f39953d46bb63bb8a46d40c2f8fb7cc5a3b6cab2bde9721d6e50"
+dependencies = [
+ "rustls-pki-types",
+]
+
 [[package]]
 name = "rustls-pki-types"
 version = "1.14.0"
@@ -2584,7 +3174,7 @@ version = "1.0.149"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "83fc039473c5595ace860d8c4fafa220ff474b3fc6bfdb4293327f1a37e94d86"
 dependencies = [
- "indexmap",
+ "indexmap 2.14.0",
  "itoa",
  "memchr",
  "serde",
@@ -2688,6 +3278,16 @@ version = "0.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b7c388c1b5e93756d0c740965c41e8822f866621d41acbdf6336a6a168f8840c"
 
+[[package]]
+name = "socket2"
+version = "0.5.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e22376abed350d73dd1cd119b57ffccad95b4e585a7cda43e286245ce23c0678"
+dependencies = [
+ "libc",
+ "windows-sys 0.52.0",
+]
+
 [[package]]
 name = "socket2"
 version = "0.6.3"
@@ -2722,6 +3322,31 @@ version = "0.1.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "2b2231b7c3057d5e4ad0156fb3dc807d900806020c5ffa3ee6ff2c8c76fb8520"
 
+[[package]]
+name = "string_cache"
+version = "0.8.9"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "bf776ba3fa74f83bf4b63c3dcbbf82173db2632ed8452cb2d891d33f459de70f"
+dependencies = [
+ "new_debug_unreachable",
+ "parking_lot",
+ "phf_shared",
+ "precomputed-hash",
+ "serde",
+]
+
+[[package]]
+name = "string_cache_codegen"
+version = "0.5.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c711928715f1fe0fe509c53b43e993a9a557babc2d0a3567d0a3006f1ac931a0"
+dependencies = [
+ "phf_generator",
+ "phf_shared",
+ "proc-macro2",
+ "quote",
+]
+
 [[package]]
 name = "strsim"
 version = "0.11.1"
@@ -2777,6 +3402,12 @@ dependencies = [
  "unicode-ident",
 ]
 
+[[package]]
+name = "sync_wrapper"
+version = "1.0.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0bf256ce5efdfa370213c1dabab5935a12e49f2c58d15e9eac2870d3b4f27263"
+
 [[package]]
 name = "syntect"
 version = "5.3.0"
@@ -2811,6 +3442,17 @@ dependencies = [
  "windows-sys 0.61.2",
 ]
 
+[[package]]
+name = "tendril"
+version = "0.4.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d24a120c5fc464a3458240ee02c299ebcb9d67b5249c8848b09d639dca8d7bb0"
+dependencies = [
+ "futf",
+ "mac",
+ "utf-8",
+]
+
 [[package]]
 name = "terminfo"
 version = "0.9.0"
@@ -2844,7 +3486,7 @@ dependencies = [
  "fancy-regex",
  "filedescriptor",
  "finl_unicode",
- "fixedbitset",
+ "fixedbitset 0.4.2",
  "hex",
  "lazy_static",
  "libc",
@@ -3004,7 +3646,7 @@ dependencies = [
  "parking_lot",
  "pin-project-lite",
  "signal-hook-registry",
- "socket2",
+ "socket2 0.6.3",
  "tokio-macros",
  "windows-sys 0.61.2",
 ]
@@ -3030,6 +3672,17 @@ dependencies = [
  "tokio",
 ]
 
+[[package]]
+name = "tokio-stream"
+version = "0.1.18"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "32da49809aab5c3bc678af03902d4ccddea2a87d028d86392a4b1560c6906c70"
+dependencies = [
+ "futures-core",
+ "pin-project-lite",
+ "tokio",
+]
+
 [[package]]
 name = "tokio-util"
 version = "0.7.18"
@@ -3044,6 +3697,130 @@ dependencies = [
  "tokio",
 ]
 
+[[package]]
+name = "tonic"
+version = "0.12.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "877c5b330756d856ffcc4553ab34a5684481ade925ecc54bcd1bf02b1d0d4d52"
+dependencies = [
+ "async-stream",
+ "async-trait",
+ "axum",
+ "base64 0.22.1",
+ "bytes",
+ "h2",
+ "http",
+ "http-body",
+ "http-body-util",
+ "hyper",
+ "hyper-timeout",
+ "hyper-util",
+ "percent-encoding",
+ "pin-project",
+ "prost",
+ "rustls-native-certs",
+ "rustls-pemfile",
+ "socket2 0.5.10",
+ "tokio",
+ "tokio-rustls",
+ "tokio-stream",
+ "tower 0.4.13",
+ "tower-layer",
+ "tower-service",
+ "tracing",
+]
+
+[[package]]
+name = "tonic-build"
+version = "0.12.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9557ce109ea773b399c9b9e5dca39294110b74f1f342cb347a80d1fce8c26a11"
+dependencies = [
+ "prettyplease",
+ "proc-macro2",
+ "prost-build",
+ "prost-types",
+ "quote",
+ "syn 2.0.117",
+]
+
+[[package]]
+name = "tower"
+version = "0.4.13"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b8fa9be0de6cf49e536ce1851f987bd21a43b771b09473c3549a6c853db37c1c"
+dependencies = [
+ "futures-core",
+ "futures-util",
+ "indexmap 1.9.3",
+ "pin-project",
+ "pin-project-lite",
+ "rand 0.8.5",
+ "slab",
+ "tokio",
+ "tokio-util",
+ "tower-layer",
+ "tower-service",
+ "tracing",
+]
+
+[[package]]
+name = "tower"
+version = "0.5.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ebe5ef63511595f1344e2d5cfa636d973292adc0eec1f0ad45fae9f0851ab1d4"
+dependencies = [
+ "futures-core",
+ "futures-util",
+ "pin-project-lite",
+ "sync_wrapper",
+ "tower-layer",
+ "tower-service",
+]
+
+[[package]]
+name = "tower-layer"
+version = "0.3.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "121c2a6cda46980bb0fcd1647ffaf6cd3fc79a013de288782836f6df9c48780e"
+
+[[package]]
+name = "tower-service"
+version = "0.3.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "8df9b6e13f2d32c91b9bd719c00d1958837bc7dec474d94952798cc8e69eeec3"
+
+[[package]]
+name = "tracing"
+version = "0.1.44"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "63e71662fa4b2a2c3a26f570f037eb95bb1f85397f3cd8076caed2f026a6d100"
+dependencies = [
+ "pin-project-lite",
+ "tracing-attributes",
+ "tracing-core",
+]
+
+[[package]]
+name = "tracing-attributes"
+version = "0.1.31"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7490cfa5ec963746568740651ac6781f701c9c5ea257c58e057f3ba8cf69e8da"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.117",
+]
+
+[[package]]
+name = "tracing-core"
+version = "0.1.36"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "db97caf9d906fbde555dd62fa95ddba9eecfd14cb388e4f491a66d74cd5fb79a"
+dependencies = [
+ "once_cell",
+]
+
 [[package]]
 name = "tree-sitter"
 version = "0.26.8"
@@ -3384,6 +4161,12 @@ version = "2.9.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "dbc4bc3a9f746d862c45cb89d705aa10f187bb96c76001afab07a0d35ce60142"
 
+[[package]]
+name = "unicode-general-category"
+version = "1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0b993bddc193ae5bd0d623b49ec06ac3e9312875fdae725a975c51db1cc1677f"
+
 [[package]]
 name = "unicode-ident"
 version = "1.0.24"
@@ -3452,6 +4235,12 @@ version = "0.9.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8ecb6da28b8a351d773b68d5825ac39017e680750f980f3a1a85cd8dd28a47c1"
 
+[[package]]
+name = "utf-8"
+version = "0.7.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "09cc8ee72d2a9becf2f2febe0205bbed8fc6615b7cb429ad062dc7b7ddd036a9"
+
 [[package]]
 name = "utf8parse"
 version = "0.2.2"
@@ -3590,7 +4379,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "bb0e353e6a2fbdc176932bbaab493762eb1255a7900fe0fea1a2f96c296cc909"
 dependencies = [
  "anyhow",
- "indexmap",
+ "indexmap 2.14.0",
  "wasm-encoder",
  "wasmparser",
 ]
@@ -3603,7 +4392,7 @@ checksum = "47b807c72e1bac69382b3a6fb3dbe8ea4c0ed87ff5629b8685ae6b9a611028fe"
 dependencies = [
  "bitflags 2.11.0",
  "hashbrown 0.15.5",
- "indexmap",
+ "indexmap 2.14.0",
  "semver",
 ]
 
@@ -3794,7 +4583,16 @@ version = "0.52.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "282be5f36a8ce781fad8c8ae18fa3f9beff57ec1b52cb3de0789201425d9a33d"
 dependencies = [
- "windows-targets",
+ "windows-targets 0.52.6",
+]
+
+[[package]]
+name = "windows-sys"
+version = "0.60.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f2f500e4d28234f72040990ec9d39e3a6b950f9f22d3dba18416c35882612bcb"
+dependencies = [
+ "windows-targets 0.53.5",
 ]
 
 [[package]]
@@ -3812,14 +4610,31 @@ version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9b724f72796e036ab90c1021d4780d4d3d648aca59e491e6b98e725b84e99973"
 dependencies = [
- "windows_aarch64_gnullvm",
- "windows_aarch64_msvc",
- "windows_i686_gnu",
- "windows_i686_gnullvm",
- "windows_i686_msvc",
- "windows_x86_64_gnu",
- "windows_x86_64_gnullvm",
- "windows_x86_64_msvc",
+ "windows_aarch64_gnullvm 0.52.6",
+ "windows_aarch64_msvc 0.52.6",
+ "windows_i686_gnu 0.52.6",
+ "windows_i686_gnullvm 0.52.6",
+ "windows_i686_msvc 0.52.6",
+ "windows_x86_64_gnu 0.52.6",
+ "windows_x86_64_gnullvm 0.52.6",
+ "windows_x86_64_msvc 0.52.6",
+]
+
+[[package]]
+name = "windows-targets"
+version = "0.53.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "4945f9f551b88e0d65f3db0bc25c33b8acea4d9e41163edf90dcd0b19f9069f3"
+dependencies = [
+ "windows-link",
+ "windows_aarch64_gnullvm 0.53.1",
+ "windows_aarch64_msvc 0.53.1",
+ "windows_i686_gnu 0.53.1",
+ "windows_i686_gnullvm 0.53.1",
+ "windows_i686_msvc 0.53.1",
+ "windows_x86_64_gnu 0.53.1",
+ "windows_x86_64_gnullvm 0.53.1",
+ "windows_x86_64_msvc 0.53.1",
 ]
 
 [[package]]
@@ -3828,48 +4643,96 @@ version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "32a4622180e7a0ec044bb555404c800bc9fd9ec262ec147edd5989ccd0c02cd3"
 
+[[package]]
+name = "windows_aarch64_gnullvm"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a9d8416fa8b42f5c947f8482c43e7d89e73a173cead56d044f6a56104a6d1b53"
+
 [[package]]
 name = "windows_aarch64_msvc"
 version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "09ec2a7bb152e2252b53fa7803150007879548bc709c039df7627cabbd05d469"
 
+[[package]]
+name = "windows_aarch64_msvc"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b9d782e804c2f632e395708e99a94275910eb9100b2114651e04744e9b125006"
+
 [[package]]
 name = "windows_i686_gnu"
 version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8e9b5ad5ab802e97eb8e295ac6720e509ee4c243f69d781394014ebfe8bbfa0b"
 
+[[package]]
+name = "windows_i686_gnu"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "960e6da069d81e09becb0ca57a65220ddff016ff2d6af6a223cf372a506593a3"
+
 [[package]]
 name = "windows_i686_gnullvm"
 version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0eee52d38c090b3caa76c563b86c3a4bd71ef1a819287c19d586d7334ae8ed66"
 
+[[package]]
+name = "windows_i686_gnullvm"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fa7359d10048f68ab8b09fa71c3daccfb0e9b559aed648a8f95469c27057180c"
+
 [[package]]
 name = "windows_i686_msvc"
 version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "240948bc05c5e7c6dabba28bf89d89ffce3e303022809e73deaefe4f6ec56c66"
 
+[[package]]
+name = "windows_i686_msvc"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1e7ac75179f18232fe9c285163565a57ef8d3c89254a30685b57d83a38d326c2"
+
 [[package]]
 name = "windows_x86_64_gnu"
 version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "147a5c80aabfbf0c7d901cb5895d1de30ef2907eb21fbbab29ca94c5b08b1a78"
 
+[[package]]
+name = "windows_x86_64_gnu"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9c3842cdd74a865a8066ab39c8a7a473c0778a3f29370b5fd6b4b9aa7df4a499"
+
 [[package]]
 name = "windows_x86_64_gnullvm"
 version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "24d5b23dc417412679681396f2b49f3de8c1473deb516bd34410872eff51ed0d"
 
+[[package]]
+name = "windows_x86_64_gnullvm"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0ffa179e2d07eee8ad8f57493436566c7cc30ac536a3379fdf008f47f6bb7ae1"
+
 [[package]]
 name = "windows_x86_64_msvc"
 version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "589f6da84c646204747d1270a2a5661ea66ed1cced2631d546fdfb155959f9ec"
 
+[[package]]
+name = "windows_x86_64_msvc"
+version = "0.53.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d6bbff5f0aada427a1e5a6da5f1f98158182f26556f345ac9e04d36d0ebed650"
+
 [[package]]
 name = "wit-bindgen"
 version = "0.51.0"
@@ -3898,7 +4761,7 @@ checksum = "b7c566e0f4b284dd6561c786d9cb0142da491f46a9fbed79ea69cdad5db17f21"
 dependencies = [
  "anyhow",
  "heck",
- "indexmap",
+ "indexmap 2.14.0",
  "prettyplease",
  "syn 2.0.117",
  "wasm-metadata",
@@ -3929,7 +4792,7 @@ checksum = "9d66ea20e9553b30172b5e831994e35fbde2d165325bec84fc43dbf6f4eb9cb2"
 dependencies = [
  "anyhow",
  "bitflags 2.11.0",
- "indexmap",
+ "indexmap 2.14.0",
  "log",
  "serde",
  "serde_derive",
@@ -3948,7 +4811,7 @@ checksum = "ecc8ac4bc1dc3381b7f59c34f00b67e18f910c2c0f50015669dde7def656a736"
 dependencies = [
  "anyhow",
  "id-arena",
- "indexmap",
+ "indexmap 2.14.0",
  "log",
  "semver",
  "serde",
@@ -3958,6 +4821,17 @@ dependencies = [
  "wasmparser",
 ]
 
+[[package]]
+name = "xml5ever"
+version = "0.18.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9bbb26405d8e919bc1547a5aa9abc95cbfa438f04844f5fdd9dc7596b748bf69"
+dependencies = [
+ "log",
+ "mac",
+ "markup5ever",
+]
+
 [[package]]
 name = "yaml-rust"
 version = "0.4.5"
diff --git a/Cargo.toml b/Cargo.toml
index c253bd7..120565b 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -18,8 +18,12 @@ name = "consciousness"
 version.workspace = true
 edition.workspace = true
 
+[features]
+nightly-diagnostics = []
+
 [dependencies]
 anyhow = "1"
+html2md = "0.2"
 crossterm = { version = "0.29", features = ["event-stream", "bracketed-paste", "osc52"] }
 clap = { version = "4", features = ["derive"] }
 figment = { version = "0.10", features = ["env"] }
@@ -29,7 +33,8 @@ log = "0.4"
 
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
-json5 = "1.3"
+json-five = "0.3"
+notify-debouncer-mini = "0.7"
 
 ratatui = { version = "0.30", features = ["unstable-rendered-line-info"] }
 tui-markdown = { git = "https://github.com/koverstreet/tui-markdown", subdirectory = "tui-markdown" }
@@ -59,6 +64,11 @@ futures = "0.3"
 capnp = "0.25"
 capnp-rpc = "0.25"
 
+tonic = { version = "0.12", features = ["tls", "tls-roots"] }
+prost = "0.13"
+async-stream = "0.3"
+tokio-stream = "0.1"
+
 tokenizers = "0.22"
 
 http = "1"
@@ -67,14 +77,18 @@ hyper-util = { version = "0.1", features = ["tokio"], default-features = false }
 http-body-util = "0.1"
 bytes = "1"
 base64 = "0.22"
+imagesize = "0.14"
 
 rustls = "0.23"
 tokio-rustls = "0.26"
 rustls-native-certs = "0.8"
+rustls-pemfile = "2"
 serde_urlencoded = "0.7"
 
 [build-dependencies]
 capnpc = "0.25"
+tonic-build = { version = "0.12", default-features = false, features = ["prost", "transport"] }
+protoc-bin-vendored = "3"
 
 [lib]
 name = "consciousness"
diff --git a/build.rs b/build.rs
index 808bf31..5f77ae4 100644
--- a/build.rs
+++ b/build.rs
@@ -13,4 +13,21 @@ fn main() {
         .file("schema/channel.capnp")
         .run()
         .expect("capnp compile failed (channel.capnp)");
+
+    // Generate salience.v1 gRPC client + message types from proto.
+    // Server side (python) is generated separately via grpcio-tools.
+    // Use vendored protoc so we don't require a system install.
+    let protoc = protoc_bin_vendored::protoc_bin_path()
+        .expect("vendored protoc not available for this platform");
+    // SAFETY: build script is single-threaded at this point; setting env
+    // before invoking tonic_build is the documented way to point it at a
+    // non-PATH protoc.
+    unsafe { std::env::set_var("PROTOC", protoc); }
+    tonic_build::configure()
+        .build_server(false)
+        .build_client(true)
+        .compile_protos(&["proto/salience.proto"], &["proto"])
+        .expect("tonic_build compile failed (salience.proto)");
+
+    println!("cargo:rerun-if-changed=proto/salience.proto");
 }
diff --git a/channels/irc/src/main.rs b/channels/irc/src/main.rs
index 4b20284..e81c4fe 100644
--- a/channels/irc/src/main.rs
+++ b/channels/irc/src/main.rs
@@ -237,11 +237,19 @@ impl State {
     async fn send_privmsg(&mut self, target: &str, msg: &str) -> io::Result<()> {
         // Send PRIVMSG, which is used for both private and channel messages.
         // Splits into multiple fragments if necessary.
-        // IRC max line = 512 bytes including CRLF. The server prepends
-        // our prefix when relaying: ":nick!~user@host PRIVMSG target :msg\r\n"
+        //
+        // Two constraints:
+        //   1. IRC max line = 512 bytes including CRLF. The server prepends
+        //      our prefix when relaying: ":nick!~user@host PRIVMSG target :msg\r\n"
+        //      So per-PRIVMSG message content must fit in 512 - overhead.
+        //   2. Embedded '\n' in the message would be interpreted by the
+        //      server as an end-of-command marker, truncating us. Split
+        //      on newlines first and send each line as its own PRIVMSG.
+        //
         // User is often ~nick (nick_len + 1). Host is up to 63 bytes.
+        // Cloaked OFTC hosts can be longer - pad the budget.
         let nick_len = self.config.nick.len();
-        let overhead = 1 + nick_len + 2 + nick_len + 1 + 63
+        let overhead = 1 + nick_len + 1 + (nick_len + 1) + 1 + 80
             + " PRIVMSG ".len() + target.len() + " :".len() + 2;
         let max_msg = 512_usize.saturating_sub(overhead);
 
@@ -249,24 +257,34 @@ impl State {
             return Err(io::Error::new(io::ErrorKind::InvalidInput, "target too long"));
         }
 
-        // Split on UTF-8 char boundaries
-        let mut remaining = msg;
-        while !remaining.is_empty() {
-            let split_at = if remaining.len() <= max_msg {
-                remaining.len()
-            } else {
-                // Find last char boundary at or before max_msg
-                let mut i = max_msg;
-                while i > 0 && !remaining.is_char_boundary(i) { i -= 1; }
-                // To avoid splitting mid-word, see if there was a space recently
-                let mut j = i;
-                while j > 1 && j > i-10 && remaining.as_bytes()[j] != b' ' { j -= 1; }
-                if remaining.as_bytes()[j] == b' ' { j }
-                else if i == 0 { max_msg } else { i }
-            };
-            let (chunk, rest) = remaining.split_at(split_at);
-            self.send_raw(&format!("PRIVMSG {target} :{chunk}")).await?;
-            remaining = rest;
+        for line in msg.split('\n') {
+            let mut remaining = line;
+            // Empty lines (blank paragraph breaks) can't be sent as empty
+            // PRIVMSGs - most IRC servers reject them. Skip.
+            if remaining.is_empty() { continue; }
+            loop {
+                let split_at = if remaining.len() <= max_msg {
+                    remaining.len()
+                } else {
+                    // Find last char boundary at or before max_msg.
+                    let mut i = max_msg;
+                    while i > 0 && !remaining.is_char_boundary(i) { i -= 1; }
+                    // Prefer splitting at a word boundary - look back up to
+                    // max_msg/4 chars for a space. With dense content (code)
+                    // we may not find one; fall back to the char boundary.
+                    let lookback = max_msg / 4;
+                    let bytes = remaining.as_bytes();
+                    let mut j = i;
+                    while j > 0 && (i - j) < lookback && bytes[j - 1] != b' ' {
+                        j -= 1;
+                    }
+                    if j > 0 && bytes[j - 1] == b' ' { j } else { i }
+                };
+                let (chunk, rest) = remaining.split_at(split_at);
+                self.send_raw(&format!("PRIVMSG {target} :{chunk}")).await?;
+                remaining = rest;
+                if remaining.is_empty() { break; }
+            }
         }
         Ok(())
     }
diff --git a/channels/telegram/src/main.rs b/channels/telegram/src/main.rs
index d3753f7..ec7aa66 100644
--- a/channels/telegram/src/main.rs
+++ b/channels/telegram/src/main.rs
@@ -181,6 +181,8 @@ struct TelegramMessage {
     chat_id: i64,
     sender: String,
     text: String,
+    /// Absolute path to a downloaded media file (photo, etc.), if any.
+    media_path: Option<String>,
 }
 
 /// Fetch and parse pending updates from Telegram via long polling.
@@ -206,19 +208,115 @@ async fn get_updates(
             let sender = msg["from"]["first_name"].as_str().unwrap_or("unknown").to_string();
             let chat_id = msg["chat"]["id"].as_i64().unwrap_or(0);
 
-            if let Some(text) = msg["text"].as_str() {
-                messages.push(TelegramMessage {
-                    update_id,
-                    chat_id,
-                    sender,
-                    text: text.to_string(),
-                });
-            }
+            // Photo: array of PhotoSize, largest is last. Download largest,
+            // surface message with [image: <path>] marker so the multimodal
+            // model can Read the image.
+            let (text, media_path) = if let Some(sizes) = msg["photo"].as_array() {
+                let caption = msg["caption"].as_str().unwrap_or("").to_string();
+                let largest = sizes.last();
+                let file_id = largest
+                    .and_then(|s| s["file_id"].as_str())
+                    .unwrap_or("");
+                if file_id.is_empty() {
+                    error!("telegram photo: missing file_id in update {update_id}");
+                    (caption, None)
+                } else {
+                    // Bound the download — HttpClient::request_timeout only covers
+                    // send_request, not body collect, so an indefinitely-slow body
+                    // would otherwise stall every subsequent poll.
+                    let dl = tokio::time::timeout(
+                        std::time::Duration::from_secs(60),
+                        download_telegram_file(client, token, file_id),
+                    ).await
+                        .unwrap_or_else(|_| Err("download timed out after 60s".into()));
+                    match dl {
+                        Ok(path) => (caption, Some(path)),
+                        Err(e) => {
+                            error!("telegram photo download failed (file_id={file_id}): {e}");
+                            // Surface what we have: caption plus a marker that
+                            // a photo was sent but couldn't be fetched.
+                            let marker = format!("[image: download failed: {e}]");
+                            let combined = if caption.is_empty() {
+                                marker
+                            } else {
+                                format!("{marker}\n{caption}")
+                            };
+                            (combined, None)
+                        }
+                    }
+                }
+            } else if let Some(text) = msg["text"].as_str() {
+                (text.to_string(), None)
+            } else {
+                // Other media types (voice, video, sticker, etc.) — skip for now,
+                // but log so we can extend later.
+                let kind = ["voice", "video", "sticker", "document", "audio", "animation"]
+                    .iter()
+                    .find(|k| !msg[**k].is_null())
+                    .copied()
+                    .unwrap_or("unknown");
+                info!("telegram: skipping non-text/photo message (kind={kind}, update_id={update_id})");
+                continue;
+            };
+
+            messages.push(TelegramMessage {
+                update_id,
+                chat_id,
+                sender,
+                text,
+                media_path,
+            });
         }
     }
     Ok(messages)
 }
 
+/// Resolve a Telegram file_id to a downloadable URL path via getFile.
+async fn get_file_path(
+    client: &HttpClient,
+    token: &str,
+    file_id: &str,
+) -> Result<String, Box<dyn std::error::Error>> {
+    let url = format!(
+        "https://api.telegram.org/bot{}/getFile?file_id={}",
+        token, file_id,
+    );
+    let response = client.get(&url).await?;
+    let body = response.text().await?;
+    let resp: serde_json::Value = serde_json::from_str(&body)
+        .map_err(|e| format!("getFile JSON parse error: {e}"))?;
+    if !resp["ok"].as_bool().unwrap_or(false) {
+        return Err(format!("getFile failed: {}", resp["description"].as_str().unwrap_or("?")).into());
+    }
+    let file_path = resp["result"]["file_path"].as_str()
+        .ok_or("getFile: missing result.file_path")?;
+    Ok(file_path.to_string())
+}
+
+/// Download a Telegram file by file_id into the channel media dir.
+/// Returns the absolute local path on success.
+async fn download_telegram_file(
+    client: &HttpClient,
+    token: &str,
+    file_id: &str,
+) -> Result<String, Box<dyn std::error::Error>> {
+    let file_path = get_file_path(client, token, file_id).await?;
+    let url = format!("https://api.telegram.org/file/bot{}/{}", token, file_path);
+    let response = client.get(&url).await?;
+    let status = response.status();
+    if !status.is_success() {
+        return Err(format!("file download failed: {status}").into());
+    }
+    let bytes = response.bytes().await?;
+
+    let ext = file_path.rsplit('.').next().filter(|e| !e.contains('/')).unwrap_or("dat");
+    let media_dir = log_dir().join("media");
+    std::fs::create_dir_all(&media_dir)?;
+    let dest = media_dir.join(format!("{file_id}.{ext}"));
+    std::fs::write(&dest, &bytes)?;
+    Ok(dest.to_string_lossy().to_string())
+}
+
 /// Send a text message to a Telegram chat.
 async fn send_message(
     client: &HttpClient,
@@ -369,11 +467,19 @@ async fn poll_once(
         let sender_lower = msg.sender.to_lowercase();
         let channel = format!("telegram.{}", sender_lower);
 
-        channel_log::append_disk_log(&log_dir(), &sender_lower, &msg.sender, &msg.text);
+        // If the message has media, prepend an [image: <abs_path>] marker
+        // so the multimodal model can Read the file directly.
+        let body = match &msg.media_path {
+            Some(path) if msg.text.is_empty() => format!("[image: {path}]"),
+            Some(path) => format!("[image: {path}]\n{}", msg.text),
+            None => msg.text.clone(),
+        };
+
+        channel_log::append_disk_log(&log_dir(), &sender_lower, &msg.sender, &body);
 
         let mut s = state.borrow_mut();
         s.config.chat_ids.insert(sender_lower, msg.chat_id);
-        let line = format!("[{}] {}", msg.sender, msg.text);
+        let line = format!("[{}] {}", msg.sender, body);
         s.push_message(line, 2, &channel);
     }
 
diff --git a/channels/tmux/src/main.rs b/channels/tmux/src/main.rs
index f49bdc1..aecb3bd 100644
--- a/channels/tmux/src/main.rs
+++ b/channels/tmux/src/main.rs
@@ -26,10 +26,12 @@ use consciousness::thalamus::channel_log::ChannelLog;
 
 #[derive(Clone, serde::Serialize, serde::Deserialize)]
 struct PaneConfig {
-    /// Human-readable label, becomes the channel name "tmux.<label>"
+    /// Human-readable label: becomes the channel name "tmux.<label>",
+    /// and the tmux pane title / window name the live pane id is
+    /// resolved from. The pane id is deliberately not stored — it is
+    /// ephemeral (recycled across pane and tmux-server restarts), so it
+    /// is looked up fresh on every connect attempt.
     label: String,
-    /// Tmux pane ID, e.g. "%5"
-    pane_id: String,
 }
 
 #[derive(Clone, serde::Serialize, serde::Deserialize)]
@@ -86,11 +88,9 @@ impl State {
         }
     }
 
-    /// Get pane_id for a label
-    fn get_pane(&self, label: &str) -> Option<&str> {
-        self.config.panes.iter()
-            .find(|p| p.label == label)
-            .map(|p| p.pane_id.as_str())
+    /// Whether a pane with this label is registered.
+    fn has_pane(&self, label: &str) -> bool {
+        self.config.panes.iter().any(|p| p.label == label)
     }
 
     /// Check if a pane is connected
@@ -103,98 +103,124 @@ impl State {
         self.connected.insert(label.to_string(), connected);
     }
 
-    /// Add a pane and persist
-    fn add_pane(&mut self, label: String, pane_id: String) {
+    /// Register a pane and persist.
+    fn add_pane(&mut self, label: String) {
         if !self.config.panes.iter().any(|p| p.label == label) {
-            self.config.panes.push(PaneConfig { label, pane_id });
+            self.config.panes.push(PaneConfig { label });
             save_config(&self.config);
         }
     }
 
-    /// Remove a pane and persist
-    fn remove_pane(&mut self, label: &str) -> Option<String> {
+    /// Unregister a pane and persist. Returns whether it was registered.
+    fn remove_pane(&mut self, label: &str) -> bool {
         if let Some(idx) = self.config.panes.iter().position(|p| p.label == label) {
-            let pane = self.config.panes.remove(idx);
+            self.config.panes.remove(idx);
             self.connected.remove(label);
             save_config(&self.config);
-            Some(pane.pane_id)
+            true
         } else {
-            None
+            false
         }
     }
 }
 
 // ── Pipe-Pane Reader ──────────────────────────────────────────
 
-/// Set up pipe-pane for a single pane, reading output into the channel log.
-async fn pipe_pane_reader(state: SharedState, pane: PaneConfig) {
+/// Wait between connect attempts for a pane that is not yet reachable.
+const RETRY_INTERVAL: std::time::Duration = std::time::Duration::from_secs(2);
+
+/// Keep a pane streamed into its channel log for as long as it stays
+/// registered. The pane id is resolved fresh by label on every connect
+/// attempt — tmux pane ids are ephemeral, so the label (pane title /
+/// window name) is the durable identity. Retries until the pane exists
+/// and pipe-pane succeeds, and reconnects the same way if the pipe
+/// later drops. Returns once close() unregisters the pane.
+async fn pipe_pane_reader(state: SharedState, label: String) {
     let pipe_dir = dirs::home_dir()
         .unwrap_or_default()
         .join(".consciousness/channels/tmux-pipes");
     std::fs::create_dir_all(&pipe_dir).ok();
+    let pipe_path = pipe_dir.join(format!("{}.pipe", label));
+    let channel_key = format!("tmux.{}", label);
 
-    let pipe_path = pipe_dir.join(format!("{}.pipe", pane.label));
-    let _ = std::fs::remove_file(&pipe_path);
+    loop {
+        if !state.borrow().has_pane(&label) {
+            return;
+        }
 
-    // Create a named pipe (FIFO)
+        connect_and_stream(&state, &label, &pipe_path, &channel_key).await;
+        state.borrow_mut().set_connected(&label, false);
+
+        if !state.borrow().has_pane(&label) {
+            return;
+        }
+        tokio::time::sleep(RETRY_INTERVAL).await;
+    }
+}
+
+/// One connect attempt: resolve the pane's live id by label, point its
+/// output at the FIFO with pipe-pane, and stream lines into the channel
+/// log. Returns on the first failure, or when the stream ends.
+async fn connect_and_stream(
+    state: &SharedState,
+    label: &str,
+    pipe_path: &std::path::Path,
+    channel_key: &str,
+) {
+    let pane_id = match find_pane_by_name(label) {
+        Some(id) => id,
+        None => return,
+    };
+
+    // Fresh FIFO for this attempt.
+    let _ = std::fs::remove_file(pipe_path);
     unsafe {
         let c_path = std::ffi::CString::new(pipe_path.to_str().unwrap()).unwrap();
         libc::mkfifo(c_path.as_ptr(), 0o644);
     }
 
-    // Tell tmux to pipe this pane's output to our FIFO
-    let pipe_path_str = pipe_path.to_string_lossy().to_string();
-    let result = std::process::Command::new("tmux")
-        .args(["pipe-pane", "-t", &pane.pane_id, &format!("cat >> {}", pipe_path_str)])
-        .output();
-
-    match result {
-        Ok(output) if output.status.success() => {
-            info!("pipe-pane set up for {} ({})", pane.label, pane.pane_id);
-        }
-        Ok(output) => {
-            error!("pipe-pane failed for {}: {}", pane.label,
-                   String::from_utf8_lossy(&output.stderr));
-            state.borrow_mut().set_connected(&pane.label, false);
+    // Point the pane's output at our FIFO.
+    let pipe_cmd = format!("cat >> {}", pipe_path.to_string_lossy());
+    match std::process::Command::new("tmux")
+        .args(["pipe-pane", "-t", &pane_id, &pipe_cmd])
+        .output()
+    {
+        Ok(o) if o.status.success() => {}
+        Ok(o) => {
+            warn!("pipe-pane failed for {} ({}): {}", label, pane_id,
+                  String::from_utf8_lossy(&o.stderr));
             return;
         }
         Err(e) => {
-            error!("failed to run tmux pipe-pane for {}: {}", pane.label, e);
-            state.borrow_mut().set_connected(&pane.label, false);
+            error!("running tmux pipe-pane for {}: {}", label, e);
             return;
         }
     }
 
-    // Open the FIFO and read lines
-    let file = match tokio::fs::File::open(&pipe_path).await {
+    let file = match tokio::fs::File::open(pipe_path).await {
         Ok(f) => f,
         Err(e) => {
-            error!("failed to open pipe for {}: {}", pane.label, e);
-            state.borrow_mut().set_connected(&pane.label, false);
+            warn!("opening pipe for {}: {}", label, e);
             return;
         }
     };
 
-    // Mark as connected once pipe is open
-    state.borrow_mut().set_connected(&pane.label, true);
-
-    let reader = tokio::io::BufReader::new(file);
-    let mut lines = reader.lines();
-    let channel_key = format!("tmux.{}", pane.label);
+    info!("connected channel tmux.{} (pane {})", label, pane_id);
+    state.borrow_mut().set_connected(label, true);
 
+    let mut lines = tokio::io::BufReader::new(file).lines();
     while let Ok(Some(line)) = lines.next_line().await {
         if line.trim().is_empty() {
             continue;
         }
         let mut s = state.borrow_mut();
-        let log = s.channel_logs
-            .entry(channel_key.clone())
-            .or_insert_with(ChannelLog::new);
-        log.push(line);
+        s.channel_logs
+            .entry(channel_key.to_string())
+            .or_insert_with(ChannelLog::new)
+            .push(line);
     }
 
-    warn!("pipe-pane reader ended for {}", pane.label);
-    state.borrow_mut().set_connected(&pane.label, false);
+    warn!("pipe-pane stream ended for {}", label);
 }
 
 // ── ChannelServer Implementation ───────────────────────────────
@@ -244,10 +270,10 @@ impl channel_server::Server for ChannelServerImpl {
         let channel = pry!(pry!(params.get_channel()).to_str()).to_string();
         let message = pry!(pry!(params.get_message()).to_str()).to_string();
 
-        // Send to tmux pane via send-keys
+        // Send to tmux pane via send-keys — resolve the live pane id by
+        // label (it is not stored).
         let label = channel.strip_prefix("tmux.").unwrap_or(&channel);
-        let pane_id = self.state.borrow().get_pane(label).map(String::from);
-        if let Some(pane_id) = pane_id {
+        if let Some(pane_id) = find_pane_by_name(label) {
             let _ = std::process::Command::new("tmux")
                 .args(["send-keys", "-t", &pane_id, &message, "Enter"])
                 .output();
@@ -302,28 +328,22 @@ impl channel_server::Server for ChannelServerImpl {
         let params = pry!(params.get());
         let label = pry!(pry!(params.get_label()).to_str()).to_string();
 
-        // Check if already open
-        if self.state.borrow().get_pane(&label).is_some() {
+        // Already registered — nothing to do.
+        if self.state.borrow().has_pane(&label) {
             return std::future::ready(Ok(()));
         }
 
-        // Find the tmux pane by name (window or pane title)
-        let pane_id = match find_pane_by_name(&label) {
-            Some(id) => id,
-            None => return std::future::ready(Err(capnp::Error::failed(
-                format!("no tmux pane named '{}'", label)))),
-        };
+        info!("opening channel tmux.{}", label);
 
-        info!("opening channel tmux.{} (pane {})", label, pane_id);
+        // Register the label and persist. The pane id is not stored —
+        // the reader resolves it by label on every connect attempt, so
+        // this succeeds even if the pane does not exist yet; the reader
+        // connects once it appears.
+        self.state.borrow_mut().add_pane(label.clone());
 
-        // Register in state and persist
-        self.state.borrow_mut().add_pane(label.clone(), pane_id.clone());
-
-        // Start pipe-pane reader
-        let pane = PaneConfig { label, pane_id };
         let reader_state = self.state.clone();
         tokio::task::spawn_local(async move {
-            pipe_pane_reader(reader_state, pane).await;
+            pipe_pane_reader(reader_state, label).await;
         });
 
         std::future::ready(Ok(()))
@@ -339,14 +359,18 @@ impl channel_server::Server for ChannelServerImpl {
         let label = channel.strip_prefix("tmux.").unwrap_or(&channel).to_string();
 
         let mut s = self.state.borrow_mut();
-        if let Some(pane_id) = s.remove_pane(&label) {
+        if s.remove_pane(&label) {
             info!("closing channel tmux.{}", label);
             s.channel_logs.remove(&format!("tmux.{}", label));
 
-            // Disconnect pipe-pane
-            let _ = std::process::Command::new("tmux")
-                .args(["pipe-pane", "-t", &pane_id])
-                .output();
+            // Stop piping if the pane is still around (if it is gone the
+            // pipe is already dead). The reader then sees the pane
+            // unregistered and exits.
+            if let Some(pane_id) = find_pane_by_name(&label) {
+                let _ = std::process::Command::new("tmux")
+                    .args(["pipe-pane", "-t", &pane_id])
+                    .output();
+            }
         }
 
         std::future::ready(Ok(()))
@@ -397,11 +421,13 @@ async fn main() -> Result<(), Box<dyn std::error::Error>> {
 
     tokio::task::LocalSet::new()
         .run_until(async move {
-            // Start a pipe-pane reader for each configured pane
+            // Start a pipe-pane reader for each configured pane; each
+            // resolves its live pane id by label and retries until
+            // connected.
             for pane in state.borrow().config.panes.clone() {
                 let reader_state = state.clone();
                 tokio::task::spawn_local(async move {
-                    pipe_pane_reader(reader_state, pane).await;
+                    pipe_pane_reader(reader_state, pane.label).await;
                 });
             }
 
diff --git a/flake.lock b/flake.lock
new file mode 100644
index 0000000..6769406
--- /dev/null
+++ b/flake.lock
@@ -0,0 +1,27 @@
+{
+  "nodes": {
+    "nixpkgs": {
+      "locked": {
+        "lastModified": 1781074563,
+        "narHash": "sha256-md8WlXOlfnIeHeOScMTTHFyf2d6iaTwPl2apR5EQ3P4=",
+        "owner": "NixOS",
+        "repo": "nixpkgs",
+        "rev": "9ae611a455b90cf061d8f332b977e387bda8e1ca",
+        "type": "github"
+      },
+      "original": {
+        "owner": "NixOS",
+        "ref": "nixos-unstable",
+        "repo": "nixpkgs",
+        "type": "github"
+      }
+    },
+    "root": {
+      "inputs": {
+        "nixpkgs": "nixpkgs"
+      }
+    }
+  },
+  "root": "root",
+  "version": 7
+}
diff --git a/flake.nix b/flake.nix
new file mode 100644
index 0000000..918df74
--- /dev/null
+++ b/flake.nix
@@ -0,0 +1,42 @@
+{
+  description = "Development shell for consciousness";
+
+  inputs = {
+    nixpkgs.url = "github:NixOS/nixpkgs/nixos-unstable";
+  };
+
+  outputs = { nixpkgs, ... }:
+    let
+      systems = [
+        "x86_64-linux"
+        "aarch64-linux"
+      ];
+      forAllSystems = nixpkgs.lib.genAttrs systems;
+    in
+    {
+      devShells = forAllSystems (system:
+        let
+          pkgs = import nixpkgs { inherit system; };
+        in
+        {
+          default = pkgs.mkShell {
+            packages = with pkgs; [
+              cargo
+              rustc
+              rustfmt
+              clippy
+              rust-analyzer
+
+              capnproto
+              pkg-config
+
+              jq
+              sqlite
+              python3
+            ];
+
+            RUST_BACKTRACE = "1";
+          };
+        });
+    };
+}
diff --git a/proto/salience.proto b/proto/salience.proto
new file mode 100644
index 0000000..fab4e6d
--- /dev/null
+++ b/proto/salience.proto
@@ -0,0 +1,276 @@
+// salience.proto — stateful generation + per-token concept readout over gRPC.
+//
+// Shape:
+//   - One server-streaming RPC (Generate) for inference. Every other
+//     operation is unary. This is the minimum streaming we need —
+//     tokens arrive one at a time with optional readouts / logprobs —
+//     and keeping everything else unary makes the client dramatically
+//     simpler than a single bidi state machine did.
+//
+//   - Server-side sessions hold the token list and image binaries.
+//     Sessions exist for bandwidth: at 200K tokens we'd otherwise
+//     re-ship ~800KB every turn, which hurts badly over a WAN link.
+//     vLLM's prefix cache holds the KV; the session just gives the
+//     client a handle so it can send deltas.
+//
+//   - The client is the source of truth for prompt content. The server
+//     is the source of truth for image token expansion (how many
+//     IMAGE_PAD tokens an image becomes under this model). The client
+//     never writes vision tokens itself — AppendImage appends the whole
+//     <|vision_start|> + IMAGE_PAD×N + <|vision_end|> block server-side.
+//
+//   - Every mutation carries (offset, truncating): the client's view of
+//     the server's current length, plus whether the client is deliberately
+//     rewriting history. Server validates on each call and rejects drift.
+//     No silent divergence, no migration bugs.
+//
+//   - Errors use gRPC status codes. NOT_FOUND for missing sessions,
+//     FAILED_PRECONDITION for offset drift or image-block splits,
+//     RESOURCE_EXHAUSTED for context overflow, ABORTED for "session busy".
+//
+// Not in v1:
+//   - Authentication beyond a shared bearer token in gRPC metadata.
+//   - Multi-tenant session namespacing.
+//   - Sampling traces beyond top-k logprobs.
+
+syntax = "proto3";
+
+package salience.v1;
+
+// ============================================================
+//  Service
+// ============================================================
+
+service Salience {
+  // Create a fresh session. Client uses session_id on every subsequent
+  // RPC until CloseSession or TTL eviction (default 30 min idle). To
+  // refresh TTL across a long pause, issue a no-op Generate (empty
+  // append_tokens, max_tokens=0, no ranges).
+  rpc OpenSession(OpenSessionRequest) returns (OpenSessionResponse);
+
+  // Release the session's tokens + images. Idempotent.
+  rpc CloseSession(CloseSessionRequest) returns (CloseSessionResponse);
+
+  // Branch a session at a given token position. The new session
+  // inherits tokens [0, at_position) and any images whose vision
+  // block lies fully in that range. Rejected with FAILED_PRECONDITION
+  // if at_position falls inside an image block (client picks a clean
+  // boundary).
+  rpc ForkSession(ForkSessionRequest) returns (ForkSessionResponse);
+
+  // Prefill + optionally decode. Images are attached inline via
+  // `GenerateRequest.images`; the client writes its own pre-expanded
+  // <|vision_start|> + N*<|image_pad|> + <|vision_end|> runs into
+  // `append_tokens` and declares each run's range in `images[i]`.
+  // Server validates run length against the actual vision-encoder
+  // feature count and returns INVALID_ARGUMENT on mismatch. Stream
+  // yields Token events (with optional readouts / logprobs per
+  // position) followed by a terminating Done.
+  rpc Generate(GenerateRequest) returns (stream GenerateEvent);
+
+  // Readout manifest for the currently-loaded model — concept names,
+  // layer indices, tensor dtype. Stateless; fetch once at client
+  // startup and cache.
+  rpc GetReadoutManifest(GetReadoutManifestRequest) returns (ReadoutManifest);
+
+  // Dump the full token stream of a session. Debug-only: used by the
+  // client to verify its local accounting against the server's
+  // session.tokens byte-for-byte when divergence is suspected. Not
+  // cheap — copies the whole sequence across the wire.
+  rpc DumpSession(DumpSessionRequest) returns (DumpSessionResponse);
+}
+
+// ============================================================
+//  Lifecycle
+// ============================================================
+
+message OpenSessionRequest {
+  // Model identifier, must match vLLM's served model. The server
+  // only has one model loaded; this is a safety check on what the
+  // client thinks it's talking to.
+  string model = 1;
+}
+
+message OpenSessionResponse {
+  string session_id = 1;
+  uint32 max_model_len = 2;
+}
+
+message CloseSessionRequest {
+  string session_id = 1;
+}
+
+message CloseSessionResponse {}
+
+message ForkSessionRequest {
+  string session_id = 1;    // source session
+  uint32 at_position = 2;   // new session inherits tokens [0, at_position)
+}
+
+message ForkSessionResponse {
+  string session_id = 1;    // new session
+}
+
+// ============================================================
+//  Inference
+// ============================================================
+
+// One image attached to a Generate call. The client is responsible
+// for writing the expanded placeholder run (VISION_START +
+// N*IMAGE_PAD + VISION_END) into `GenerateRequest.append_tokens` at
+// positions [pad_range_start, pad_range_end) and pairing it with
+// the corresponding `ImageAttachment` entry. Server validates that
+// the declared range's pad count matches what the vision encoder
+// produces, and returns INVALID_ARGUMENT if they disagree.
+message ImageAttachment {
+  // Image bytes (PNG / JPEG / WebP / …).
+  bytes  bytes = 1;
+
+  // MIME type, e.g. "image/png".
+  string mime = 2;
+
+  // Absolute token positions (in `session.tokens` AFTER `append_tokens`
+  // is applied) spanning the full vision block —  `[vision_start,
+  // pad*N, vision_end]`. end is exclusive, so end - start == N + 2.
+  uint32 pad_range_start = 3;
+  uint32 pad_range_end = 4;
+}
+
+message GenerateRequest {
+  string session_id = 1;
+
+  // Tokens to append before prefill. May be empty. Client writes the
+  // full vision block (VISION_START + N*IMAGE_PAD + VISION_END) for
+  // any newly-attached image directly into this stream; each such
+  // block must be paired with a matching entry in `images`. The
+  // server validates that the declared ranges all point at IMAGE_PAD
+  // runs and that each run's length matches what the vision encoder
+  // produces for the corresponding image.
+  repeated uint32 append_tokens = 2;
+
+  // Client's view of session.tokens length at the time of the call.
+  // Must equal server's actual length, OR be strictly less when
+  // truncating=true (server rewinds before appending). Any other
+  // mismatch is FAILED_PRECONDITION.
+  uint32 offset = 3;
+  bool   truncating = 4;
+
+  // Decode budget. 0 = prefill only (no decode, emit Token events
+  // for positions covered by logprobs_ranges / readout_ranges, then
+  // Done; replaces the old /score endpoint). >0 = decode up to this
+  // many tokens, stopping early on EOS / stop_token_ids.
+  uint32 max_tokens = 5;
+
+  // Position ranges (absolute, within the session's post-append
+  // token list) at which to emit logprobs on Token events. Empty =
+  // no logprobs. `logprob_top_k > 0` returns the top-k alternative
+  // tokens at each covered position; `logprob_top_k == 0` returns
+  // only the sampled-token's logprob.
+  repeated PositionRange logprobs_ranges = 6;
+  uint32                 logprob_top_k = 7;
+
+  // Position ranges at which to emit concept-readout vectors. Empty
+  // = no readouts. Logical shape per position is
+  // [n_layers][n_concepts] — see GetReadoutManifest.
+  repeated PositionRange readout_ranges = 8;
+
+  // Sampling parameters. Meaningful only when max_tokens > 0.
+  float           temperature = 9;      // default 1.0 when zero
+  float           top_p = 10;           // default 1.0 when zero
+  uint32          top_k = 11;           // default 0 (disabled)
+  repeated uint32 stop_token_ids = 12;
+
+  // vLLM scheduler priority (0 = interactive, 10 = batch).
+  int32 priority = 13;
+
+  // Images newly attached on this call. Each entry describes one
+  // image's binary bytes, its mime type, and the exact token-position
+  // range of its pre-expanded placeholder run inside `session.tokens`
+  // after `append_tokens` is applied. See `ImageAttachment`.
+  repeated ImageAttachment images = 14;
+}
+
+message PositionRange {
+  uint32 start = 1;   // inclusive
+  uint32 end = 2;     // exclusive
+}
+
+message GenerateEvent {
+  oneof event {
+    Token        token = 1;
+    GenerateDone done = 2;
+  }
+}
+
+message Token {
+  // Token id at this position. For prefill this is the prompt token;
+  // for decode it's the sampled token.
+  uint32 id = 1;
+
+  // Absolute position in the session's token list.
+  uint32 position = 2;
+
+  // True for prefill positions, false for decode.
+  bool   is_prefill = 3;
+
+  // Concept readout at this position. Empty if the position wasn't
+  // covered by readout_ranges.
+  repeated float readout = 4 [packed = true];
+
+  // Top-k alternative tokens' logprobs at this position — populated
+  // when the position is covered by logprobs_ranges and
+  // logprob_top_k > 0.
+  repeated TokenLogprob logprobs = 5;
+
+  // Logprob of the token at `position` (the prompt token for
+  // prefill, the sampled token for decode). Populated when the
+  // position is covered by logprobs_ranges.
+  float sampled_logprob = 6;
+  bool  has_sampled_logprob = 7;
+}
+
+message TokenLogprob {
+  uint32 id = 1;
+  float  logprob = 2;
+}
+
+message GenerateDone {
+  uint32 prompt_tokens = 1;
+  uint32 completion_tokens = 2;
+  uint32 total_tokens = 3;
+
+  enum FinishReason {
+    FINISH_REASON_UNSPECIFIED = 0;
+    FINISH_REASON_EOS = 1;              // emitted EOS / stop token
+    FINISH_REASON_LENGTH = 2;           // hit max_tokens
+    FINISH_REASON_CANCELLED = 3;        // client cancelled
+    FINISH_REASON_STOP_STRING = 4;      // matched a stop string
+  }
+  FinishReason finish_reason = 4;
+}
+
+// ============================================================
+//  Readout manifest
+// ============================================================
+
+message GetReadoutManifestRequest {}
+
+message ReadoutManifest {
+  repeated string concepts = 1;
+  repeated uint32 layers = 2;
+  uint32          hidden_size = 3;
+  string          dtype = 4;
+}
+
+// ============================================================
+//  Debug
+// ============================================================
+
+message DumpSessionRequest {
+  string session_id = 1;
+}
+
+message DumpSessionResponse {
+  // The full session.tokens sequence, verbatim.
+  repeated uint32 tokens = 1 [packed = true];
+}
diff --git a/scripts/quantize_qwen3_6_mm.py b/scripts/quantize_qwen3_6_mm.py
new file mode 100644
index 0000000..844571e
--- /dev/null
+++ b/scripts/quantize_qwen3_6_mm.py
@@ -0,0 +1,327 @@
+"""Quantize Qwen3.6-27B (multimodal) to FP8 for vLLM serving.
+
+Why this exists
+---------------
+The earlier `quantize_qwen3_6.py` (in shell history, never committed)
+loaded the model with `AutoModelForCausalLM`, which silently strips
+the multimodal arch. Result: an FP8 checkpoint with no vision tower
+weights at all. vLLM happily instantiated the vision tower from the
+config and ran it with default/uninitialized weights, producing
+gibberish image features and `!!!!!!`-style output. We chased that
+through the protocol layer for a long time before tracing it back
+to the quant. This script avoids that trap by loading via the
+config-declared class explicitly.
+
+Recipe
+------
+FP8_DYNAMIC (per-channel weight scales, per-token dynamic activation
+scales, both E4M3) for Linear weights, with an `ignore` list derived
+from Unsloth's UD-Q8_K_XL (`unsloth/Qwen3.6-27B-GGUF`). Their
+sensitivity sweep flagged specific layers as quantization-fragile;
+we honor those layer indices even though their algorithm is
+GGUF-native Q8_K and ours is FP8 — sensitivity is a layer property,
+not an algorithm property.
+
+vLLM fusion constraint
+~~~~~~~~~~~~~~~~~~~~~~
+vLLM's Qwen3.5/3.6 model code fuses sub-modules at load time:
+  qkv_proj      ← q_proj, k_proj, v_proj
+  gate_up_proj  ← gate_proj, up_proj
+  in_proj_qkvz  ← in_proj_qkv, in_proj_z
+  in_proj_ba    ← in_proj_b, in_proj_a
+compressed_tensors rejects checkpoints where sub-modules of a fused
+layer have different quantization schemes. Our ignore list is shaped
+around this — within any fused layer, all components share a scheme.
+That's the reason `in_proj_qkv` is ignored even though Unsloth's
+sweep doesn't single it out, and the reason late-stack attn override
+covers q/k/v rather than just q/k.
+
+MTP merge
+---------
+`Qwen3_5ForConditionalGeneration` doesn't expose the MTP submodule,
+so `oneshot()` produces a checkpoint with the 15 `mtp.*` tensors
+silently dropped. After quantization we read the MTP weights back
+out of the upstream cached snapshot and splice them into the saved
+safetensors at BF16. They're small (~850 MB) so quantizing them
+isn't worth the calibration risk; speculative-decoding code paths
+in vLLM expect the MTP head present.
+
+Output
+------
+`OUTPUT_DIR` gets the FP8 model.safetensors + config + processor +
+recipe.yaml. Vision tower stays BF16 (in `ignore`); LM Linears go
+to FP8; norms, SSM internals (not Linear), and MTP tensors stay
+BF16 untouched.
+
+Verification at end: re-opens the saved safetensors and asserts
+- vision .weight tensors present (>= 150; full count is 167)
+- lm_head + embed_tokens at fp16/bf16 (NOT FP8)
+- a sampled FP8'd Linear actually has float8 dtype
+- 15 mtp.* tensors present
+
+Run
+---
+    ~/vllm-venv/bin/python quantize_qwen3_6_mm.py
+"""
+from __future__ import annotations
+
+import glob
+import json
+import sys
+from pathlib import Path
+
+import torch
+from huggingface_hub import snapshot_download
+from llmcompressor import oneshot
+from llmcompressor.modifiers.quantization import QuantizationModifier
+from safetensors import safe_open
+from safetensors.torch import save_file
+from transformers import AutoProcessor
+from transformers.models.qwen3_5.modeling_qwen3_5 import (
+    Qwen3_5ForConditionalGeneration,
+)
+
+
+MODEL = "Qwen/Qwen3.6-27B"
+OUTPUT_DIR = "/home/ubuntu/amygdala-training/Qwen3.6-27B-FP8-mm"
+
+
+# Layers Unsloth's UD-Q8_K_XL keeps at F16 (perplexity-sensitive
+# in their sweep). Late-stack clustering is consistent with the
+# general finding that errors near the output propagate directly
+# to logits.
+LATE_FFN_LAYERS = (50, 51, 59, 62, 63)
+LATE_ATTN_LAYERS = (51, 59, 63)
+
+
+# Build the ignore regex list. Note: llmcompressor matches these
+# patterns against MODULE names (no `.weight` suffix) when walking
+# `named_modules()` for `targets=["Linear"]`. The first pass of
+# this script used `\.weight$` patterns and silently quantized
+# lm_head + every linear_attn projection — verified post-hoc by
+# inspecting the saved safetensors. Patterns now anchor on `$`
+# at the module name.
+IGNORE_PATTERNS: list[str] = [
+    # Original recipe: lm_head and embeddings always full-precision.
+    # (embed_tokens is an Embedding, not a Linear, so it's already
+    # ignored by `targets=["Linear"]`. Pattern kept as belt-and-
+    # suspenders in case future llmcompressor versions widen the
+    # target set.)
+    "re:lm_head$",
+    "re:.*embed_tokens$",
+
+    # Vision tower — entire `model.visual.*` subtree (vision
+    # transformer blocks + merger + patch_embed + pos_embed).
+    # Unsloth ships the vision tower as a separate `mmproj-BF16.gguf`
+    # for GGUF consumers; in our single-file FP8 setup we just leave
+    # them at BF16.
+    "re:model\\.visual\\..*",
+
+    # MTP (multi-token prediction) module — Unsloth's GGUF doesn't
+    # carry MTP weights so we have no precision signal from them;
+    # safest to keep BF16.
+    "re:mtp\\..*",
+
+    # Linear-attention block — keep ENTIRELY at BF16. vLLM fuses
+    # `in_proj_qkv` and `in_proj_z` into a single `in_proj_qkvz`
+    # layer, and compressed_tensors rejects mixed schemes within a
+    # fused layer. Unsloth's recipe keeps z, a, b, out at F16/F32
+    # (gate/SSM internals are quantization-fragile in the GatedDeltaNet
+    # update), so the principled choice is to also keep `in_proj_qkv`
+    # at BF16 rather than FP8'ing the gate to match. We give up ~1 GB
+    # of FP8 coverage; in exchange we follow Unsloth's quality intent
+    # and load cleanly under vLLM. (`in_proj_a` + `in_proj_b` are
+    # likewise fused as `in_proj_ba` — both ignored, consistent.)
+    "re:model\\.language_model\\.layers\\.\\d+\\.linear_attn\\.in_proj_qkv$",
+    "re:model\\.language_model\\.layers\\.\\d+\\.linear_attn\\.in_proj_z$",
+    "re:model\\.language_model\\.layers\\.\\d+\\.linear_attn\\.in_proj_a$",
+    "re:model\\.language_model\\.layers\\.\\d+\\.linear_attn\\.in_proj_b$",
+    "re:model\\.language_model\\.layers\\.\\d+\\.linear_attn\\.out_proj$",
+
+    # Per-layer high-precision MLP (Unsloth flagged exactly these
+    # late-stack indices in their UD-Q8_K_XL sensitivity sweep, all
+    # three of {gate, up, down} per layer). vLLM fuses gate+up into
+    # `gate_up_proj`; ignoring both keeps the fused layer consistent.
+    # `down_proj` is its own (non-fused) layer.
+    "re:model\\.language_model\\.layers\\.("
+    + "|".join(str(n) for n in LATE_FFN_LAYERS)
+    + ")\\.mlp\\.(down|gate|up)_proj$",
+
+    # Per-layer high-precision attention q/k/v (Unsloth's sweep upgrades
+    # only q and k; we extend to v because vLLM fuses q/k/v into
+    # `qkv_proj` and rejects mixed schemes. `o_proj` is its own
+    # non-fused layer and stays at FP8.
+    "re:model\\.language_model\\.layers\\.("
+    + "|".join(str(n) for n in LATE_ATTN_LAYERS)
+    + ")\\.self_attn\\.(q|k|v)_proj$",
+]
+
+
+def main() -> None:
+    print(f"Loading {MODEL} as multimodal "
+          f"(Qwen3_5ForConditionalGeneration)...", flush=True)
+    model = Qwen3_5ForConditionalGeneration.from_pretrained(
+        MODEL,
+        dtype=torch.bfloat16,
+        device_map="auto",
+        trust_remote_code=True,
+    )
+    print(f"  loaded: {model.__class__.__name__}", flush=True)
+
+    print(f"Loading processor (text + image preprocessing)...", flush=True)
+    processor = AutoProcessor.from_pretrained(MODEL, trust_remote_code=True)
+
+    print("Running FP8_DYNAMIC oneshot quantization...", flush=True)
+    print(f"  ignore list: {len(IGNORE_PATTERNS)} patterns",
+          flush=True)
+    recipe = QuantizationModifier(
+        targets=["Linear"],
+        scheme="FP8_DYNAMIC",
+        ignore=IGNORE_PATTERNS,
+    )
+    oneshot(model=model, recipe=recipe, output_dir=OUTPUT_DIR)
+    processor.save_pretrained(OUTPUT_DIR)
+    print(f"  wrote model + processor to {OUTPUT_DIR}", flush=True)
+
+    merge_mtp(OUTPUT_DIR)
+    verify_output(OUTPUT_DIR)
+
+
+def merge_mtp(out_dir: str) -> None:
+    """Splice upstream MTP tensors into the saved FP8 safetensors.
+
+    `Qwen3_5ForConditionalGeneration` skips the MTP submodule on load,
+    so oneshot's output is missing the 15 `mtp.*` tensors. We resolve
+    the upstream snapshot via the HF cache (already populated by
+    from_pretrained), pull just the MTP tensors out at BF16, and
+    rewrite the safetensors with them merged in. The compressed_tensors
+    metadata header (which carries the FP8 format identifier vLLM
+    needs to dequantize) is preserved verbatim.
+
+    Atomic-rename is used so a crash mid-write doesn't corrupt the
+    33+ GB checkpoint we just spent minutes producing.
+    """
+    print("\nMerging upstream MTP tensors...", flush=True)
+    upstream_dir = Path(snapshot_download(
+        MODEL,
+        allow_patterns=["model.safetensors.index.json",
+                        "model-*-of-*.safetensors"],
+    ))
+
+    with open(upstream_dir / "model.safetensors.index.json") as f:
+        idx = json.load(f)
+    mtp_shards = sorted({v for k, v in idx["weight_map"].items()
+                         if k.startswith("mtp.")})
+    print(f"  MTP tensors live in shards: {mtp_shards}", flush=True)
+
+    mtp_tensors: dict[str, torch.Tensor] = {}
+    for shard in mtp_shards:
+        with safe_open(upstream_dir / shard, framework="pt") as f:
+            for k in f.keys():
+                if k.startswith("mtp."):
+                    mtp_tensors[k] = f.get_tensor(k).contiguous()
+    mtp_bytes = sum(t.numel() * t.element_size()
+                    for t in mtp_tensors.values())
+    print(f"  loaded {len(mtp_tensors)} mtp tensors "
+          f"({mtp_bytes/1e6:.1f} MB)", flush=True)
+
+    fp8_files = sorted(Path(out_dir).glob("*.safetensors"))
+    if len(fp8_files) != 1:
+        sys.exit(f"FAIL: expected single safetensors shard, "
+                 f"got {fp8_files}")
+    existing_path = fp8_files[0]
+
+    with safe_open(existing_path, framework="pt") as f:
+        metadata = f.metadata() or {}
+        all_tensors = {k: f.get_tensor(k) for k in f.keys()}
+
+    overlap = set(all_tensors) & set(mtp_tensors)
+    if overlap:
+        sys.exit(f"FAIL: MTP key collision with FP8 output: "
+                 f"{sorted(overlap)[:5]}")
+    all_tensors.update(mtp_tensors)
+
+    tmp_path = existing_path.with_name(existing_path.name + ".new")
+    print(f"  rewriting {existing_path.name} "
+          f"({len(all_tensors)} tensors)...", flush=True)
+    save_file(all_tensors, str(tmp_path), metadata=metadata)
+    tmp_path.replace(existing_path)
+    print("  done", flush=True)
+
+
+def verify_output(out_dir: str) -> None:
+    """Open the saved safetensors and assert the recipe actually
+    landed: vision tower present at BF16, FP8 dtype on at least one
+    quantized Linear, lm_head not FP8."""
+    print(f"\nVerifying {out_dir}...", flush=True)
+
+    files = sorted(glob.glob(f"{out_dir}/*.safetensors"))
+    if not files:
+        sys.exit(f"FAIL: no safetensors in {out_dir}")
+
+    vision_keys: list[tuple[str, str]] = []
+    fp8_sample: tuple[str, str] | None = None
+    lm_head_dtype: str | None = None
+    mtp_keys: list[str] = []
+
+    for fp in files:
+        with safe_open(fp, framework="pt") as f:
+            for k in f.keys():
+                if k.startswith("mtp."):
+                    mtp_keys.append(k)
+                # Some FP8 quants write a sibling `_scale` / `_zero_point`;
+                # we just care about the .weight tensors.
+                if not k.endswith(".weight"):
+                    continue
+                t = f.get_tensor(k)
+                dtype = str(t.dtype).replace("torch.", "")
+                if "model.visual." in k:
+                    vision_keys.append((k, dtype))
+                if k == "lm_head.weight":
+                    lm_head_dtype = dtype
+                if (fp8_sample is None
+                        and "float8" in dtype
+                        and "language_model.layers" in k):
+                    fp8_sample = (k, dtype)
+
+    # Qwen3.6-27B has 167 vision `.weight` tensors (333 vision tensors
+    # total, the rest are `.bias` and per-block norms). 150 is a
+    # sanity floor that catches "vision tower didn't make it through"
+    # without being brittle to minor arch revisions.
+    if len(vision_keys) < 150:
+        sys.exit(f"FAIL: only {len(vision_keys)} vision tensors found "
+                 f"(expected >= 150). Vision tower didn't make it "
+                 f"through the quant.")
+
+    bad_vision = [(k, d) for k, d in vision_keys if "float8" in d]
+    if bad_vision:
+        sys.exit(f"FAIL: vision weights got quantized to FP8: "
+                 f"{bad_vision[:3]}...")
+
+    if lm_head_dtype is None:
+        sys.exit("FAIL: lm_head.weight not found in output.")
+    if "float8" in lm_head_dtype:
+        sys.exit(f"FAIL: lm_head.weight is FP8 ({lm_head_dtype}); "
+                 f"should be BF16/FP16.")
+
+    if fp8_sample is None:
+        sys.exit("FAIL: no FP8 weights found in language_model.layers — "
+                 "the recipe didn't quantize anything.")
+
+    # Upstream Qwen3.6-27B has exactly 15 mtp.* tensors (1 fused
+    # transformer block + projection + norms). merge_mtp() should
+    # have spliced all of them in.
+    if len(mtp_keys) != 15:
+        sys.exit(f"FAIL: expected 15 mtp.* tensors, found "
+                 f"{len(mtp_keys)}. merge_mtp() missed some.")
+
+    print(f"  ✓ {len(vision_keys)} vision tensors at "
+          f"{vision_keys[0][1]} (not FP8)")
+    print(f"  ✓ lm_head.weight at {lm_head_dtype} (not FP8)")
+    print(f"  ✓ FP8 sample: {fp8_sample[0]} = {fp8_sample[1]}")
+    print(f"  ✓ {len(mtp_keys)} mtp.* tensors present")
+    print("DONE")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/src/agent/api/http.rs b/src/agent/api/http.rs
index 429350b..a059426 100644
--- a/src/agent/api/http.rs
+++ b/src/agent/api/http.rs
@@ -100,7 +100,7 @@ impl HttpClient {
                 .map_err(|e| anyhow::anyhow!("invalid server name: {e}"))?;
             let connector = tokio_rustls::TlsConnector::from(self.tls.clone());
             let tls = connector.connect(server_name.to_owned(), tcp).await
-                .context("TLS handshake")?;
+                .map_err(|e| anyhow::anyhow!("TLS handshake to {host}: {e}"))?;
             TokioIo::new(Box::new(tls) as Box<dyn IoStream>)
         } else {
             TokioIo::new(Box::new(tcp) as Box<dyn IoStream>)
@@ -154,6 +154,14 @@ impl HttpResponse {
         Ok(String::from_utf8_lossy(&bytes).into_owned())
     }
 
+    /// Read the entire body as raw bytes (for binary downloads).
+    pub async fn bytes(self) -> Result<Bytes> {
+        let bytes = self.body.collect().await
+            .context("reading response body")?
+            .to_bytes();
+        Ok(bytes)
+    }
+
     /// Read the entire body and deserialize as JSON.
     pub async fn json<T: serde::de::DeserializeOwned>(self) -> Result<T> {
         let bytes = self.body.collect().await
@@ -190,6 +198,7 @@ impl HttpClientBuilder {
     }
 
     pub fn build(self) -> HttpClient {
+        install_rustls_crypto_provider();
         let certs = rustls_native_certs::load_native_certs()
             .certs.into_iter()
             .collect::<Vec<_>>();
@@ -197,6 +206,13 @@ impl HttpClientBuilder {
         for cert in certs {
             root_store.add(cert).ok();
         }
+        // Also trust any `.pem` files under `~/.consciousness/certs/` —
+        // self-signed server certs for our own vllm hosts live there.
+        // Drop a new `<host>.pem` in the dir to trust a new server; no
+        // code change needed.
+        for cert in load_user_certs() {
+            root_store.add(cert).ok();
+        }
         let tls = Arc::new(
             ClientConfig::builder()
                 .with_root_certificates(root_store)
@@ -210,6 +226,65 @@ impl HttpClientBuilder {
     }
 }
 
+/// Install rustls' default crypto provider exactly once per process.
+/// rustls 0.23 doesn't pick one automatically when multiple features
+/// could provide it (e.g. when tonic pulls in both ring and aws-lc-rs
+/// via transitive deps). Idempotent via OnceLock; safe to call from
+/// multiple callers.
+fn install_rustls_crypto_provider() {
+    static ONCE: std::sync::OnceLock<()> = std::sync::OnceLock::new();
+    ONCE.get_or_init(|| {
+        let _ = rustls::crypto::ring::default_provider().install_default();
+    });
+}
+
+/// Load every `.pem` file under `~/.consciousness/certs/` as a DER
+/// certificate and return them. Silent on missing dir, missing files,
+/// or parse errors — those are "no extra certs trusted" rather than
+/// hard failures, to keep startup robust.
+/// Load the concatenated PEM bytes of every `.pem` file under
+/// `~/.consciousness/certs/` — suitable for passing to a tonic
+/// `ClientTlsConfig::ca_certificate(Certificate::from_pem(...))` call
+/// so gRPC connections trust the same self-signed servers the HTTP
+/// path does.
+pub(crate) fn load_user_certs_pem_bytes() -> Vec<u8> {
+    let mut out = Vec::new();
+    let Some(home) = dirs::home_dir() else { return out };
+    let dir = home.join(".consciousness").join("certs");
+    let Ok(entries) = std::fs::read_dir(&dir) else { return out };
+    for entry in entries.flatten() {
+        let path = entry.path();
+        if path.extension().and_then(|e| e.to_str()) != Some("pem") {
+            continue;
+        }
+        if let Ok(bytes) = std::fs::read(&path) {
+            out.extend_from_slice(&bytes);
+            if !bytes.ends_with(b"\n") {
+                out.push(b'\n');
+            }
+        }
+    }
+    out
+}
+
+fn load_user_certs() -> Vec<rustls::pki_types::CertificateDer<'static>> {
+    let mut out = Vec::new();
+    let Some(home) = dirs::home_dir() else { return out };
+    let dir = home.join(".consciousness").join("certs");
+    let Ok(entries) = std::fs::read_dir(&dir) else { return out };
+    for entry in entries.flatten() {
+        let path = entry.path();
+        if path.extension().and_then(|e| e.to_str()) != Some("pem") {
+            continue;
+        }
+        let Ok(bytes) = std::fs::read(&path) else { continue };
+        for cert in rustls_pemfile::certs(&mut bytes.as_slice()).flatten() {
+            out.push(cert);
+        }
+    }
+    out
+}
+
 /// Trait alias for streams that work with hyper's IO adapter.
 trait IoStream: tokio::io::AsyncRead + tokio::io::AsyncWrite + Send + Unpin + 'static {}
 impl<T: tokio::io::AsyncRead + tokio::io::AsyncWrite + Send + Unpin + 'static> IoStream for T {}
diff --git a/src/agent/api/mod.rs b/src/agent/api/mod.rs
index 7c06fa7..fc8a358 100644
--- a/src/agent/api/mod.rs
+++ b/src/agent/api/mod.rs
@@ -7,13 +7,14 @@
 // Set POC_DEBUG=1 for verbose per-turn logging.
 
 pub mod http;
+pub mod salience;
 
-use std::time::{Duration, Instant};
+use std::time::Duration;
 use anyhow::Result;
 use tokio::sync::mpsc;
 use serde::Deserialize;
 
-use http::{HttpClient, HttpResponse};
+use http::HttpClient;
 
 #[derive(Debug, Clone, Deserialize)]
 pub struct Usage {
@@ -22,6 +23,36 @@ pub struct Usage {
     pub total_tokens: u32,
 }
 
+/// Concept-readout manifest returned by the vLLM server's
+/// `/v1/readout/manifest` endpoint. Maps the nameless tensor indices
+/// in streaming `readout` fields back to concept names and layer
+/// indices.
+#[derive(Debug, Clone, Deserialize)]
+pub struct ReadoutManifest {
+    pub concepts: Vec<String>,
+    pub layers: Vec<u32>,
+}
+
+/// Per-token per-layer concept projections streamed alongside each
+/// sampled token. Shape `[n_layers][n_concepts]`. Named values come
+/// from pairing with the manifest fetched at startup.
+pub type TokenReadout = Vec<Vec<f32>>;
+
+/// Client-side sampling state. Mirrors the wire-level fields in
+/// `GenerateRequest` (proto flattened its `SamplingParams` submessage
+/// in so the server handler reads them directly), but stays as a
+/// grouped struct on the client because UI / config / tests pass
+/// these around together.
+#[derive(Clone, Copy)]
+pub struct SamplingParams {
+    pub temperature: f32,
+    pub top_p: f32,
+    pub top_k: u32,
+    /// Decode budget. 0 = prefill only; >0 = decode up to this many
+    /// tokens, stopping early on EOS / stop_token_ids.
+    pub max_tokens: u32,
+}
+
 /// A JoinHandle that aborts its task when dropped.
 pub(crate) struct AbortOnDrop(tokio::task::JoinHandle<()>);
 
@@ -31,13 +62,6 @@ impl Drop for AbortOnDrop {
     }
 }
 
-/// Sampling parameters for model generation.
-#[derive(Clone, Copy)]
-pub(crate) struct SamplingParams {
-    pub temperature: f32,
-    pub top_p: f32,
-    pub top_k: u32,
-}
 
 // ─────────────────────────────────────────────────────────────
 //  Stream events — yielded by backends, consumed by the runner
@@ -45,7 +69,10 @@ pub(crate) struct SamplingParams {
 
 /// One token from the streaming completions API.
 pub enum StreamToken {
-    Token(u32),
+    /// A sampled token, optionally with its per-layer concept readout.
+    /// `readout` is `None` when the server has readout disabled or
+    /// returned no readout for this chunk.
+    Token { id: u32, readout: Option<TokenReadout> },
     Done { usage: Option<Usage> },
     Error(String),
 }
@@ -56,6 +83,17 @@ pub struct ApiClient {
     api_key: String,
     pub model: String,
     base_url: String,
+    /// Cached readout manifest — fetched once per process and shared
+    /// across ApiClient clones (every Agent/fork gets the same cell).
+    /// `None` after fetch means the server has readout disabled (404).
+    manifest: std::sync::Arc<tokio::sync::OnceCell<Option<ReadoutManifest>>>,
+    /// Shared tonic Channel to the salience gRPC endpoint. Opened on
+    /// first use and reused across every SessionHandle / RPC call
+    /// derived from this ApiClient. tonic multiplexes concurrent
+    /// requests over the HTTP/2 connection automatically.
+    salience_channel: std::sync::Arc<
+        tokio::sync::OnceCell<tonic::transport::Channel>
+    >,
 }
 
 impl ApiClient {
@@ -70,29 +108,69 @@ impl ApiClient {
             api_key: api_key.to_string(),
             model: model.to_string(),
             base_url: base_url.trim_end_matches('/').to_string(),
+            manifest: std::sync::Arc::new(tokio::sync::OnceCell::new()),
+            salience_channel: std::sync::Arc::new(tokio::sync::OnceCell::new()),
         }
     }
 
-    pub(crate) fn stream_completion(
+    /// Return a `SalienceClient` on the shared gRPC channel — opens
+    /// the channel on first call and reuses it thereafter across
+    /// every ApiClient clone. All scoring / inference / session
+    /// RPCs flow through this single multiplexed HTTP/2 connection.
+    ///
+    /// Bumps tonic's default 4 MiB encode/decode caps to 64 MiB on
+    /// every client. Multimodal Generate requests carry pre-encoded
+    /// image bytes inline (Qwen3.6's 768×768 patches at high res
+    /// land around 5–8 MiB per turn), and Done events with full
+    /// per-token readout vectors can also exceed 4 MiB on long runs.
+    pub async fn salience_client(&self) -> Result<
+        salience::pb::salience_client::SalienceClient<tonic::transport::Channel>
+    > {
+        let ch = self.salience_channel.get_or_try_init(|| async {
+            let grpc_url = salience::derive_grpc_url(&self.base_url);
+            log::debug!(target: "grpc",
+                "opening shared salience channel: http_base={} -> grpc_url={}",
+                self.base_url, grpc_url);
+            salience::connect_channel(&grpc_url).await
+        }).await?;
+        const MAX_GRPC_MESSAGE_BYTES: usize = 64 * 1024 * 1024;
+        Ok(salience::pb::salience_client::SalienceClient::new(ch.clone())
+            .max_decoding_message_size(MAX_GRPC_MESSAGE_BYTES)
+            .max_encoding_message_size(MAX_GRPC_MESSAGE_BYTES))
+    }
+
+    /// Stream generation via a gRPC session. Walks the prompt chunks
+    /// comparing against the session's `committed_len`, sends the
+    /// delta as interleaved `AppendImage` + intermediate
+    /// `Generate(max_tokens=0)` (for text runs separating images) +
+    /// a final `Generate(max_tokens=sampling.max_tokens, ...)` whose
+    /// Token events stream back through the channel.
+    ///
+    /// On any gRPC error the session is dropped; the next call
+    /// reopens fresh. Happy-path ordering: Token* Done. Error paths
+    /// emit `StreamToken::Error` and close.
+    pub(crate) fn stream_session_mm(
         &self,
-        prompt_tokens: &[u32],
+        session_lock: std::sync::Arc<crate::Mutex<Option<salience::SessionHandle>>>,
+        chunks: Vec<super::context::WireChunk>,
+        images: Vec<super::context::WireImage>,
+        match_upto: u32,
         sampling: SamplingParams,
         priority: Option<i32>,
+        readout_shape: Option<(u32, u32)>,
     ) -> (mpsc::UnboundedReceiver<StreamToken>, AbortOnDrop) {
         let (tx, rx) = mpsc::unbounded_channel();
-        let client = self.client.clone();
-        let api_key = self.api_key.clone();
-        let model = self.model.clone();
-        let prompt_tokens = prompt_tokens.to_vec();
-        let base_url = self.base_url.clone();
+        let client = self.clone();
 
         let handle = tokio::spawn(async move {
-            let result = stream_completions(
-                &client, &base_url, &api_key, &model,
-                &prompt_tokens, &tx, sampling, priority,
+            let result = run_session_generate(
+                session_lock, &client, chunks, images, match_upto, sampling,
+                priority, readout_shape, &tx,
             ).await;
             if let Err(e) = result {
-                let _ = tx.send(StreamToken::Error(e.to_string()));
+                log::warn!(target: "grpc",
+                    "stream_session_mm error, forwarding to UI: {:#}", e);
+                let _ = tx.send(StreamToken::Error(format!("{:#}", e)));
             }
         });
 
@@ -102,327 +180,247 @@ impl ApiClient {
     pub fn base_url(&self) -> &str { &self.base_url }
     pub fn api_key(&self) -> &str { &self.api_key }
 
+    /// Fetch `/v1/readout/manifest` — returns `Ok(Some(..))` if
+    /// readout is enabled on the server, `Ok(None)` on 404 (disabled),
+    /// or an error on any other failure.
+    ///
+    /// First call performs the HTTP fetch; subsequent calls (including
+    /// across ApiClient clones sharing the same cell) return the
+    /// cached result. The manifest doesn't change during a server run.
+    pub fn model_str(&self) -> &str { &self.model }
+
+    pub async fn fetch_readout_manifest(&self) -> Result<Option<ReadoutManifest>> {
+        let manifest = self.manifest.get_or_try_init(|| async {
+            let url = format!("{}/readout/manifest", self.base_url);
+            let auth = format!("Bearer {}", self.api_key);
+            let response = self
+                .client
+                .get_with_headers(&url, &[("Authorization", &auth)])
+                .await
+                .map_err(|e| anyhow::anyhow!("readout manifest fetch ({}): {}", url, e))?;
+            let status = response.status();
+            if status.as_u16() == 404 {
+                return Ok::<_, anyhow::Error>(None);
+            }
+            if !status.is_success() {
+                let body = response.text().await.unwrap_or_default();
+                let n = body.floor_char_boundary(body.len().min(500));
+                anyhow::bail!("readout manifest HTTP {} ({}): {}", status, url, &body[..n]);
+            }
+            Ok(Some(response.json().await?))
+        }).await?;
+        Ok(manifest.clone())
+    }
+
 }
 
-async fn stream_completions(
-    client: &HttpClient,
-    base_url: &str,
-    api_key: &str,
-    model: &str,
-    prompt_tokens: &[u32],
-    tx: &mpsc::UnboundedSender<StreamToken>,
+/// Body of the gRPC-path streaming task. Walks the wire chunks
+/// against the session's `committed_len`, sends the delta via
+/// AppendImage / intermediate prefill-only Generates / final decode
+/// Generate, and translates the final Generate's Token events into
+/// StreamTokens on `tx`. On success the session handle is returned
+/// to `session_lock` with an updated `committed_len`; on error the
+/// handle is dropped so the next call reopens.
+async fn run_session_generate(
+    session_lock: std::sync::Arc<crate::Mutex<Option<salience::SessionHandle>>>,
+    client: &ApiClient,
+    chunks: Vec<super::context::WireChunk>,
+    images: Vec<super::context::WireImage>,
+    match_upto: u32,
     sampling: SamplingParams,
     priority: Option<i32>,
-) -> anyhow::Result<()> {
-    let mut request = serde_json::json!({
-        "model": model,
-        "prompt": prompt_tokens,
-        "max_tokens": 16384,
-        "temperature": sampling.temperature,
-        "top_p": sampling.top_p,
-        "top_k": sampling.top_k,
-        "stream": true,
-        "return_token_ids": true,
-        "skip_special_tokens": false,
-        "stop_token_ids": [super::tokenizer::IM_END],
-    });
-    if let Some(p) = priority {
-        request["priority"] = serde_json::json!(p);
-    }
+    readout_shape: Option<(u32, u32)>,
+    tx: &mpsc::UnboundedSender<StreamToken>,
+) -> Result<()> {
+    use std::time::Instant;
+    use futures::StreamExt;
+    use super::context::WireChunk;
+    use salience::pb;
 
-    let url = format!("{}/completions", base_url);
-    let debug_label = format!("{} prompt tokens, model={}", prompt_tokens.len(), model);
-
-    let mut response = send_and_check(
-        client, &url, &request,
-        ("Authorization", &format!("Bearer {}", api_key)),
-        &[], &debug_label, None,
-    ).await?;
-
-    let mut reader = SseReader::new();
-    let mut usage = None;
-
-    while let Some(event) = reader.next_event(&mut response).await? {
-        if let Some(err_msg) = event["error"]["message"].as_str() {
-            anyhow::bail!("API error in stream: {}", err_msg);
-        }
-
-        if let Some(u) = event["usage"].as_object() {
-            if let Ok(u) = serde_json::from_value::<Usage>(serde_json::Value::Object(u.clone())) {
-                usage = Some(u);
+    let mut handle: salience::SessionHandle = {
+        let mut guard = session_lock.lock().await;
+        match guard.take() {
+            Some(h) => h,
+            None => {
+                drop(guard);
+                log::debug!(target: "grpc", "run_session_generate: opening new session");
+                salience::SessionHandle::open(client).await?
             }
         }
+    };
 
-        let choices = match event["choices"].as_array() {
-            Some(c) => c,
-            None => continue,
+    // If the client believes the match extends only up to `match_upto`
+    // but the server has more, we need to rewind. For v1 the match is
+    // either whole or broken — `match_upto` is always 0 on any mutation
+    // — so the cheapest correct recovery is to drop the session and
+    // open a fresh one.
+    if match_upto < handle.committed_len {
+        log::warn!(target: "grpc",
+            "session rewind: match_upto={} < committed_len={} — reopening session (resending {} bytes)",
+            match_upto, handle.committed_len, handle.committed_len - match_upto);
+        drop(handle);
+        handle = salience::SessionHandle::open(client).await?;
+    }
+
+    // Walk chunks at byte-level, taking everything past `match_upto`
+    // as the delta. Token chunks can be split mid-way; images live
+    // inline in the token stream, so there's no separate image-chunk
+    // case anymore.
+    let mut acc: u32 = 0;
+    let mut pending: Vec<u32> = Vec::new();
+    for chunk in chunks.iter() {
+        match chunk {
+            WireChunk::Tokens(t) => {
+                let len = t.len() as u32;
+                let chunk_end = acc + len;
+                if chunk_end <= match_upto {
+                    acc = chunk_end;
+                } else if acc < match_upto {
+                    let skip = (match_upto - acc) as usize;
+                    pending.extend_from_slice(&t[skip..]);
+                    acc = chunk_end;
+                } else {
+                    pending.extend_from_slice(t);
+                    acc = chunk_end;
+                }
+            }
+        }
+    }
+
+    // Filter images to those entirely past `match_upto` — anything
+    // before is on the server already (prior turn), anything
+    // straddling is a hard divergence (image partially-sent shouldn't
+    // happen with our atomic AppendImage history; with images-inline
+    // it can only happen if mark_dirty cleared match_upto mid-block,
+    // which the AST mutators prevent).
+    let mut new_images: Vec<pb::ImageAttachment> = Vec::new();
+    for img in &images {
+        if img.pad_end <= match_upto {
+            continue; // already sent on a prior turn
+        }
+        if img.pad_start < match_upto {
+            anyhow::bail!(
+                "session divergence: image at [{},{}) straddles match_upto={}",
+                img.pad_start, img.pad_end, match_upto,
+            );
+        }
+        new_images.push(pb::ImageAttachment {
+            bytes: img.bytes.clone(),
+            mime: img.mime.clone(),
+            pad_range_start: img.pad_start,
+            pad_range_end: img.pad_end,
+        });
+    }
+
+    // Final Generate: pending holds any trailing text; decode up to
+    // sampling.max_tokens. Request readouts on all decode positions
+    // via a catch-all range ending at u32::MAX — decode never
+    // reaches it.
+    let prompt_len_after_append = handle.committed_len + pending.len() as u32;
+    let readout_ranges = if readout_shape.is_some() {
+        vec![pb::PositionRange {
+            start: prompt_len_after_append,
+            end: u32::MAX,
+        }]
+    } else {
+        Vec::new()
+    };
+    let req = pb::GenerateRequest {
+        session_id: handle.session_id.clone(),
+        append_tokens: pending,
+        offset: handle.committed_len,
+        truncating: false,
+        max_tokens: sampling.max_tokens,
+        logprobs_ranges: Vec::new(),
+        logprob_top_k: 0,
+        readout_ranges,
+        temperature: sampling.temperature,
+        top_p: sampling.top_p,
+        top_k: sampling.top_k,
+        stop_token_ids: Vec::new(),
+        priority: priority.unwrap_or(0),
+        images: new_images,
+    };
+    let session_id_for_log = handle.session_id.clone();
+    let t_generate = Instant::now();
+    log::debug!(target: "grpc",
+        "session {} Generate: offset={} append={} max_tokens={} priority={}",
+        session_id_for_log, req.offset, req.append_tokens.len(),
+        req.max_tokens, req.priority);
+
+    let mut stream = handle.generate(req).await?;
+    let (n_layers, n_concepts) = readout_shape.unwrap_or((0, 0));
+    let mut session_terminated = false;
+    let mut first_token_at: Option<Instant> = None;
+
+    while let Some(event) = stream.next().await {
+        let event = match event {
+            Ok(e) => e,
+            Err(status) => {
+                log::warn!(target: "grpc",
+                    "session {} Generate stream error: {} — dropping session",
+                    session_id_for_log, status);
+                session_terminated = true;
+                let _ = tx.send(StreamToken::Error(format!(
+                    "Generate stream error: {}", status,
+                )));
+                break;
+            }
         };
-
-        for choice in choices {
-            if let Some(ids) = choice["token_ids"].as_array() {
-                for id_val in ids {
-                    if let Some(id) = id_val.as_u64() {
-                        let _ = tx.send(StreamToken::Token(id as u32));
-                    }
+        let Some(inner) = event.event else { continue };
+        match inner {
+            pb::generate_event::Event::Token(t) => {
+                if t.is_prefill { continue; }
+                if first_token_at.is_none() {
+                    log::debug!(target: "grpc",
+                        "session {} first decode token at {:?}",
+                        session_id_for_log, t_generate.elapsed());
+                    first_token_at = Some(Instant::now());
                 }
-            } else if let Some(text) = choice["text"].as_str() {
-                // Fallback: provider didn't return token_ids, encode locally
-                if !text.is_empty() {
-                    for id in super::tokenizer::encode(text) {
-                        let _ = tx.send(StreamToken::Token(id));
+                let readout = if t.readout.is_empty() {
+                    None
+                } else if n_layers == 0 || n_concepts == 0 {
+                    None
+                } else {
+                    let expected = (n_layers as usize) * (n_concepts as usize);
+                    if t.readout.len() != expected {
+                        log::warn!(target: "grpc",
+                            "readout shape mismatch: expected {}*{}={}, got {}",
+                            n_layers, n_concepts, expected, t.readout.len());
+                        None
+                    } else {
+                        let n = n_concepts as usize;
+                        let mut layers: Vec<Vec<f32>> = Vec::with_capacity(n_layers as usize);
+                        for l in 0..(n_layers as usize) {
+                            layers.push(t.readout[l * n..(l + 1) * n].to_vec());
+                        }
+                        Some(layers)
                     }
+                };
+                if tx.send(StreamToken::Token { id: t.id, readout }).is_err() {
+                    break;
                 }
             }
+            pb::generate_event::Event::Done(d) => {
+                log::debug!(target: "grpc",
+                    "session {} Done: prompt={} completion={} total={} reason={:?} elapsed={:?}",
+                    session_id_for_log, d.prompt_tokens, d.completion_tokens,
+                    d.total_tokens, d.finish_reason, t_generate.elapsed());
+                handle.committed_len = d.total_tokens;
+                let usage = Some(Usage {
+                    prompt_tokens: d.prompt_tokens,
+                    completion_tokens: d.completion_tokens,
+                    total_tokens: d.total_tokens,
+                });
+                let _ = tx.send(StreamToken::Done { usage });
+            }
         }
     }
 
-    let _ = tx.send(StreamToken::Done { usage });
+    if !session_terminated {
+        let mut guard = session_lock.lock().await;
+        *guard = Some(handle);
+    }
     Ok(())
 }
 
-/// Send an HTTP request and check for errors.
-pub(crate) async fn send_and_check(
-    client: &HttpClient,
-    url: &str,
-    body: &impl serde::Serialize,
-    auth_header: (&str, &str),
-    extra_headers: &[(&str, &str)],
-    debug_label: &str,
-    request_json: Option<&str>,
-) -> Result<HttpResponse> {
-    let debug = std::env::var("POC_DEBUG").is_ok();
-    let start = Instant::now();
-
-    if debug {
-        let payload_size = serde_json::to_string(body)
-            .map(|s| s.len())
-            .unwrap_or(0);
-        dbglog!(
-            "request: {}K payload, {}",
-            payload_size / 1024, debug_label,
-        );
-    }
-
-    let mut headers: Vec<(&str, &str)> = Vec::with_capacity(extra_headers.len() + 1);
-    headers.push(auth_header);
-    headers.extend_from_slice(extra_headers);
-
-    let response = client
-        .send_json("POST", url, &headers, body)
-        .await
-        .map_err(|e| {
-            let msg = e.to_string();
-            let cause = if msg.contains("connect timeout") || msg.contains("TCP connect") {
-                "connection refused"
-            } else if msg.contains("request timeout") {
-                "request timed out"
-            } else {
-                "request error"
-            };
-            anyhow::anyhow!("{} ({}): {}", cause, url, msg)
-        })?;
-
-    let status = response.status();
-    let elapsed = start.elapsed();
-
-    if debug {
-        for name in [
-            "x-ratelimit-remaining",
-            "x-ratelimit-limit",
-            "x-request-id",
-        ] {
-            if let Some(val) = response.header(name) {
-                dbglog!("header {}: {}", name, val);
-            }
-        }
-    }
-
-    if !status.is_success() {
-        let body = response.text().await.unwrap_or_default();
-        dbglog!(
-            "HTTP {} after {:.1}s ({}): {}",
-            status,
-            elapsed.as_secs_f64(),
-            url,
-            &body[..body.floor_char_boundary(body.len().min(500))]
-        );
-        if let Some(json) = request_json {
-            let log_dir = dirs::home_dir()
-                .unwrap_or_default()
-                .join(".consciousness/logs/failed-requests");
-            let _ = std::fs::create_dir_all(&log_dir);
-            let ts = chrono::Local::now().format("%Y%m%dT%H%M%S");
-            let path = log_dir.join(format!("{}.json", ts));
-            if std::fs::write(&path, json).is_ok() {
-                dbglog!(
-                    "saved failed request to {} (HTTP {})", path.display(), status
-                );
-            }
-        }
-        anyhow::bail!("HTTP {} ({}): {}", status, url, &body[..body.floor_char_boundary(body.len().min(1000))]);
-    }
-
-    if debug {
-        dbglog!(
-            "connected in {:.1}s (HTTP {})",
-            elapsed.as_secs_f64(),
-            status.as_u16()
-        );
-    }
-
-    Ok(response)
-}
-
-/// SSE stream reader. Handles the generic SSE plumbing shared by both
-/// backends: chunk reading with timeout, line buffering, `data:` prefix
-/// stripping, `[DONE]` detection, JSON parsing, and parse error diagnostics.
-/// Yields parsed events as serde_json::Value — each backend handles its
-/// own event types.
-pub(crate) struct SseReader {
-    line_buf: String,
-    chunk_timeout: Duration,
-    pub stream_start: Instant,
-    pub chunks_received: u64,
-    pub sse_lines_parsed: u64,
-    pub sse_parse_errors: u64,
-    debug: bool,
-    done: bool,
-    /// Serialized request payload — saved to disk on errors for replay debugging.
-    pub(crate) request_json: Option<String>,
-}
-
-impl SseReader {
-    pub(crate) fn new() -> Self {
-        Self {
-            line_buf: String::new(),
-            chunk_timeout: Duration::from_secs(crate::config::get().api_stream_timeout_secs),
-            stream_start: Instant::now(),
-            chunks_received: 0,
-            sse_lines_parsed: 0,
-            sse_parse_errors: 0,
-            debug: std::env::var("POC_DEBUG").is_ok(),
-            done: false,
-            request_json: None,
-        }
-    }
-
-    /// Attach the serialized request payload for error diagnostics.
-    /// Save the request payload to disk for replay debugging.
-    fn save_failed_request(&self, reason: &str) {
-        let Some(ref json) = self.request_json else { return };
-        let log_dir = dirs::home_dir()
-            .unwrap_or_default()
-            .join(".consciousness/logs/failed-requests");
-        let _ = std::fs::create_dir_all(&log_dir);
-        let ts = chrono::Local::now().format("%Y%m%dT%H%M%S");
-        let path = log_dir.join(format!("{}.json", ts));
-        if std::fs::write(&path, json).is_ok() {
-            dbglog!(
-                "saved failed request to {} ({})", path.display(), reason
-            );
-        }
-    }
-
-    /// Read the next SSE event from the response stream.
-    /// Returns Ok(Some(value)) for each parsed data line,
-    /// Ok(None) when the stream ends or [DONE] is received.
-    pub(crate) async fn next_event(
-        &mut self,
-        response: &mut HttpResponse,
-    ) -> Result<Option<serde_json::Value>> {
-        loop {
-            // Drain complete lines from the buffer before reading more chunks
-            while let Some(newline_pos) = self.line_buf.find('\n') {
-                let line = self.line_buf[..newline_pos].trim().to_string();
-                self.line_buf = self.line_buf[newline_pos + 1..].to_string();
-
-                if line == "data: [DONE]" {
-                    self.done = true;
-                    return Ok(None);
-                }
-                if line.is_empty()
-                    || line.starts_with("event: ")
-                    || !line.starts_with("data: ")
-                {
-                    continue;
-                }
-
-                let json_str = &line[6..];
-                self.sse_lines_parsed += 1;
-
-                match serde_json::from_str(json_str) {
-                    Ok(v) => return Ok(Some(v)),
-                    Err(e) => {
-                        self.sse_parse_errors += 1;
-                        if self.sse_parse_errors == 1 || self.debug {
-                            let preview = if json_str.len() > 200 {
-                                format!("{}...", &json_str[..200])
-                            } else {
-                                json_str.to_string()
-                            };
-                            dbglog!(
-                                "SSE parse error (#{}) {}: {}",
-                                self.sse_parse_errors, e, preview
-                            );
-                        }
-                        continue;
-                    }
-                }
-            }
-
-            if self.done {
-                return Ok(None);
-            }
-
-            // Read more data from the response stream
-            match tokio::time::timeout(self.chunk_timeout, response.chunk()).await {
-                Ok(Ok(Some(chunk))) => {
-                    self.chunks_received += 1;
-                    self.line_buf.push_str(&String::from_utf8_lossy(&chunk));
-                }
-                Ok(Ok(None)) => return Ok(None),
-                Ok(Err(e)) => {
-                    let buf_preview = if self.line_buf.is_empty() {
-                        "(empty)".to_string()
-                    } else {
-                        let n = self.line_buf.len().min(500);
-                        format!("{}B: {}", self.line_buf.len(), &self.line_buf[..n])
-                    };
-                    let msg = format!(
-                        "stream error after {} chunks, {:.1}s, {} sse lines: {} | buf: {}",
-                        self.chunks_received,
-                        self.stream_start.elapsed().as_secs_f64(),
-                        self.sse_lines_parsed,
-                        e, buf_preview,
-                    );
-                    dbglog!("{}", msg);
-                    self.save_failed_request(&msg);
-                    return Err(e.into());
-                }
-                Err(_) => {
-                    let buf_preview = if self.line_buf.is_empty() {
-                        "(empty)".to_string()
-                    } else {
-                        let n = self.line_buf.len().min(500);
-                        format!("{}B: {}", self.line_buf.len(), &self.line_buf[..n])
-                    };
-                    let msg = format!(
-                        "stream timeout: {}s, {} chunks, {} sse lines, {:.1}s elapsed | buf: {}",
-                        self.chunk_timeout.as_secs(),
-                        self.chunks_received,
-                        self.sse_lines_parsed,
-                        self.stream_start.elapsed().as_secs_f64(),
-                        buf_preview,
-                    );
-                    dbglog!("{}", msg);
-                    self.save_failed_request(&msg);
-                    anyhow::bail!(
-                        "stream timeout: no data for {}s ({} chunks received)",
-                        self.chunk_timeout.as_secs(),
-                        self.chunks_received
-                    );
-                }
-            }
-        }
-    }
-}
diff --git a/src/agent/api/salience.rs b/src/agent/api/salience.rs
new file mode 100644
index 0000000..f5f65d2
--- /dev/null
+++ b/src/agent/api/salience.rs
@@ -0,0 +1,279 @@
+// agent/api/salience.rs — gRPC client bindings for salience.v1.
+//
+// Thin wrapper around the tonic-generated types. Every RPC except
+// Generate is unary; Generate is server-streaming. Free functions
+// (open/close session) wrap the lifecycle RPCs; `SessionHandle` just
+// carries the id + connection params so later RPCs can reuse them.
+//
+// The old bidi Session() API is gone — see git history for its shape.
+
+#![allow(clippy::enum_variant_names)]
+
+use anyhow::{Context, Result};
+use tonic::transport::{Certificate, Channel, ClientTlsConfig, Endpoint};
+
+/// Generated prost + tonic types for salience.v1. Call sites use
+/// `pb::OpenSessionRequest`, `pb::Token`, etc.
+pub mod pb {
+    tonic::include_proto!("salience.v1");
+}
+
+pub type SalienceClient = pb::salience_client::SalienceClient<Channel>;
+
+/// Open a TLS-aware gRPC channel to the salience server. `base_url`
+/// looks like `https://host:8443`. User-provided CA certs under
+/// `~/.consciousness/certs/` are trusted in addition to the system
+/// roots (for self-signed server certs).
+///
+/// Returns the raw `Channel` so callers (`ApiClient::salience_client`)
+/// can cache it and clone a `SalienceClient` per request without
+/// reopening the TCP/TLS connection. tonic multiplexes RPCs over the
+/// shared channel automatically.
+pub async fn connect_channel(base_url: &str) -> Result<Channel> {
+    let mut endpoint = Endpoint::from_shared(base_url.to_string())
+        .with_context(|| format!("invalid salience endpoint: {}", base_url))?
+        .connect_timeout(std::time::Duration::from_secs(30))
+        .timeout(std::time::Duration::from_secs(600));
+
+    if base_url.starts_with("https://") {
+        let user_certs = super::http::load_user_certs_pem_bytes();
+        let mut tls = ClientTlsConfig::new().with_native_roots();
+        if !user_certs.is_empty() {
+            tls = tls.ca_certificate(Certificate::from_pem(user_certs));
+        }
+        endpoint = endpoint
+            .tls_config(tls)
+            .with_context(|| "configuring tonic TLS")?;
+    }
+
+    endpoint
+        .connect()
+        .await
+        .with_context(|| format!("failed to connect to salience server at {}", base_url))
+}
+
+/// Derive the gRPC base URL from the HTTP completions base URL.
+///
+/// vLLM's salience gRPC server listens on a different port (8443) from
+/// the HTTP endpoint (8000) and accepts no path component. Given an
+/// HTTP base like `https://host:8000/v1`, produce `https://host:8443`.
+/// No-op when the path is empty and the port isn't 8000.
+pub fn derive_grpc_url(http_base: &str) -> String {
+    let mut url = http_base.trim_end_matches('/').to_string();
+    if let Some(proto_end) = url.find("://") {
+        let rest_start = proto_end + 3;
+        if let Some(path_slash) = url[rest_start..].find('/') {
+            url.truncate(rest_start + path_slash);
+        }
+    }
+    url.replace(":8000", ":8443")
+}
+
+/// Attach a bearer token to a tonic request as gRPC metadata.
+pub fn with_auth<T>(req: &mut tonic::Request<T>, api_key: &str) {
+    if api_key.is_empty() {
+        return;
+    }
+    let bearer = format!("Bearer {}", api_key);
+    if let Ok(val) = bearer.parse() {
+        req.metadata_mut().insert("authorization", val);
+    }
+}
+
+/// Handle to a server-side session. Carries the id + an `ApiClient`
+/// clone (which holds the shared tonic Channel) so subsequent
+/// per-session RPCs go over the process-global connection.
+/// `committed_len` tracks the server's current session.tokens length
+/// so the client can submit deltas with the right `offset`.
+pub struct SessionHandle {
+    pub session_id: String,
+    pub max_model_len: u32,
+    pub committed_len: u32,
+    client: super::ApiClient,
+}
+
+impl SessionHandle {
+    pub async fn open(client: &super::ApiClient) -> Result<Self> {
+        let t0 = std::time::Instant::now();
+        log::debug!(target: "grpc", "OpenSession rpc: start");
+        let mut c = client.salience_client().await?;
+        let mut req = tonic::Request::new(pb::OpenSessionRequest {
+            model: client.model.clone(),
+        });
+        with_auth(&mut req, client.api_key());
+        let resp = c
+            .open_session(req)
+            .await
+            .with_context(|| "OpenSession RPC failed")?
+            .into_inner();
+        log::debug!(target: "grpc",
+            "OpenSession rpc: done session_id={} max_model_len={} elapsed={:?}",
+            resp.session_id, resp.max_model_len, t0.elapsed());
+        Ok(Self {
+            session_id: resp.session_id,
+            max_model_len: resp.max_model_len,
+            committed_len: 0,
+            client: client.clone(),
+        })
+    }
+
+    pub fn client(&self) -> &super::ApiClient { &self.client }
+
+    /// Debug-only: fetch the server's full session.tokens. Used to
+    /// verify client-side accounting byte-for-byte when divergence
+    /// is suspected. Not cheap on large sessions.
+    pub async fn dump_tokens(&self) -> Result<Vec<u32>> {
+        let mut c = self.client.salience_client().await?;
+        let mut req = tonic::Request::new(pb::DumpSessionRequest {
+            session_id: self.session_id.clone(),
+        });
+        with_auth(&mut req, self.client.api_key());
+        let resp = c
+            .dump_session(req)
+            .await
+            .with_context(|| "DumpSession RPC failed")?
+            .into_inner();
+        Ok(resp.tokens)
+    }
+
+    /// Open a gRPC Generate stream with the given request. Caller
+    /// iterates the returned stream of GenerateEvents; the handle's
+    /// `committed_len` should be advanced by the caller on Done based
+    /// on the Done event's `total_tokens` field.
+    pub async fn generate(
+        &self,
+        req: pb::GenerateRequest,
+    ) -> Result<tonic::Streaming<pb::GenerateEvent>> {
+        let t0 = std::time::Instant::now();
+        log::debug!(target: "grpc",
+            "Generate rpc: open-stream session={} offset={} append={} max_tokens={}",
+            self.session_id, req.offset, req.append_tokens.len(), req.max_tokens);
+        let mut c = self.client.salience_client().await?;
+        let mut req = tonic::Request::new(req);
+        with_auth(&mut req, self.client.api_key());
+        let resp = c
+            .generate(req)
+            .await
+            .with_context(|| "Generate RPC failed")?;
+        log::debug!(target: "grpc",
+            "Generate rpc: stream opened session={} open-latency={:?}",
+            self.session_id, t0.elapsed());
+        Ok(resp.into_inner())
+    }
+
+    /// Run a prefill-only Generate (max_tokens=0) that appends the
+    /// given tokens to the session. No decode, no Token events — the
+    /// server just extends session.tokens and runs prefill to warm
+    /// the KV cache. Used to interleave text runs between AppendImage
+    /// calls, and by score paths that want prompt_logprobs without a
+    /// decode step.
+    pub async fn prefill_only(&mut self, tokens: Vec<u32>) -> Result<()> {
+        use futures::StreamExt;
+        let req = pb::GenerateRequest {
+            session_id: self.session_id.clone(),
+            append_tokens: tokens,
+            offset: self.committed_len,
+            truncating: false,
+            max_tokens: 0,
+            logprobs_ranges: Vec::new(),
+            logprob_top_k: 0,
+            readout_ranges: Vec::new(),
+            temperature: 0.0,
+            top_p: 0.0,
+            top_k: 0,
+            stop_token_ids: Vec::new(),
+            priority: 0,
+            images: Vec::new(),
+        };
+        let mut stream = self.generate(req).await?;
+        while let Some(event) = stream.next().await {
+            let event = event.map_err(|s| anyhow::anyhow!("prefill Generate stream: {}", s))?;
+            if let Some(pb::generate_event::Event::Done(d)) = event.event {
+                self.committed_len = d.total_tokens;
+            }
+        }
+        Ok(())
+    }
+}
+
+/// Drop → fire CloseSession in a detached task so servers don't leak
+/// sessions until TTL eviction. Best-effort: if no tokio runtime is
+/// available we skip; the server's 30min TTL will reap it eventually.
+impl Drop for SessionHandle {
+    fn drop(&mut self) {
+        if self.session_id.is_empty() {
+            return;
+        }
+        let session_id = std::mem::take(&mut self.session_id);
+        let client = self.client.clone();
+        let Ok(rt) = tokio::runtime::Handle::try_current() else {
+            log::debug!(target: "grpc",
+                "SessionHandle drop outside tokio runtime, session {} leaks to TTL",
+                session_id);
+            return;
+        };
+        rt.spawn(async move {
+            let Ok(mut c) = client.salience_client().await else { return };
+            let mut req = tonic::Request::new(pb::CloseSessionRequest {
+                session_id: session_id.clone(),
+            });
+            with_auth(&mut req, client.api_key());
+            if let Err(e) = c.close_session(req).await {
+                log::debug!(target: "grpc",
+                    "CloseSession on drop failed for {}: {:#}",
+                    session_id, e);
+            }
+        });
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn generated_types_compile() {
+        // Exercise the shape of the new proto types — if build.rs
+        // stops regenerating against the proto, this stops compiling.
+        let _open = pb::OpenSessionRequest {
+            model: "qwen3-vl".into(),
+        };
+        let _tok = pb::Token {
+            id: 42,
+            position: 0,
+            is_prefill: false,
+            readout: vec![0.1, 0.2, 0.3],
+            logprobs: vec![pb::TokenLogprob {
+                id: 1,
+                logprob: -0.5,
+            }],
+            sampled_logprob: -0.1,
+            has_sampled_logprob: true,
+        };
+        let _done = pb::GenerateDone {
+            prompt_tokens: 10,
+            completion_tokens: 20,
+            total_tokens: 30,
+            finish_reason: pb::generate_done::FinishReason::Eos as i32,
+        };
+        let _evt = pb::GenerateEvent {
+            event: Some(pb::generate_event::Event::Done(_done)),
+        };
+    }
+
+    #[test]
+    fn derive_grpc_url_cases() {
+        assert_eq!(
+            derive_grpc_url("https://host:8000/v1"),
+            "https://host:8443",
+        );
+        assert_eq!(
+            derive_grpc_url("https://host:8000/"),
+            "https://host:8443",
+        );
+        assert_eq!(
+            derive_grpc_url("https://host:9000/v1"),
+            "https://host:9000",
+        );
+    }
+}
diff --git a/src/agent/context.rs b/src/agent/context.rs
index c43c023..4140939 100644
--- a/src/agent/context.rs
+++ b/src/agent/context.rs
@@ -81,10 +81,33 @@ pub enum NodeBody {
     Memory { key: String, text: String, score: Option<f64> },
     Dmn(String),
 
+    // Vision input — rendered as <|vision_start|> <|image_pad|>×N <|vision_end|>.
+    // `token_count` is N, the count vLLM will compute for this image's grid.
+    Image {
+        #[serde(with = "b64_bytes")]
+        bytes: Vec<u8>,
+        mime: String,
+        orig_height: u32,
+        orig_width: u32,
+        token_count: u32,
+    },
+
     // Non-visible (0 tokens in prompt)
     Log(String),
 }
 
+mod b64_bytes {
+    use base64::{Engine, engine::general_purpose::STANDARD};
+    use serde::{Serializer, Deserializer, Deserialize};
+    pub fn serialize<S: Serializer>(bytes: &[u8], s: S) -> Result<S::Ok, S::Error> {
+        s.serialize_str(&STANDARD.encode(bytes))
+    }
+    pub fn deserialize<'de, D: Deserializer<'de>>(d: D) -> Result<Vec<u8>, D::Error> {
+        let s = String::deserialize(d)?;
+        STANDARD.decode(s).map_err(serde::de::Error::custom)
+    }
+}
+
 /// A leaf node: typed content with cached token IDs.
 /// Token IDs are not serialized — they're recomputed on deserialization.
 #[derive(Debug, Clone, Serialize)]
@@ -92,7 +115,7 @@ pub struct NodeLeaf {
     body: NodeBody,
     #[serde(skip)]
     token_ids: Vec<u32>,
-    timestamp: Option<DateTime<Utc>>,
+    timestamp: DateTime<Utc>,
 }
 
 impl<'de> Deserialize<'de> for NodeLeaf {
@@ -100,14 +123,22 @@ impl<'de> Deserialize<'de> for NodeLeaf {
         #[derive(Deserialize)]
         struct Raw {
             body: NodeBody,
-            timestamp: Option<DateTime<Utc>>,
+            timestamp: DateTime<Utc>,
         }
-        let raw = Raw::deserialize(deserializer)?;
-        let token_ids = if raw.body.is_prompt_visible() {
-            tokenizer::encode(&raw.body.render())
-        } else {
-            vec![]
-        };
+        let mut raw = Raw::deserialize(deserializer)?;
+        // Heal pre-refactor logs: Image leaves used to be deserialized
+        // with token_count=0 (server-authoritative count was applied
+        // after AppendImage). With pads now expanded client-side at
+        // construction, recompute from the persisted dimensions if
+        // the stored count is 0.
+        if let NodeBody::Image { orig_height, orig_width, token_count, .. }
+            = &mut raw.body
+        {
+            if *token_count == 0 {
+                *token_count = qwen3_image_token_count(*orig_height, *orig_width);
+            }
+        }
+        let token_ids = raw.body.compute_token_ids();
         Ok(NodeLeaf { body: raw.body, token_ids, timestamp: raw.timestamp })
     }
 }
@@ -119,22 +150,49 @@ pub enum AstNode {
     Branch {
         role: Role,
         children: Vec<AstNode>,
+        timestamp: DateTime<Utc>,
         /// Per-response memory attribution from full scoring matrix.
         /// Maps memory key → divergence score for this response.
         #[serde(default, skip_serializing_if = "std::collections::BTreeMap::is_empty")]
         memory_scores: std::collections::BTreeMap<String, f64>,
+        /// Cached token stream for the subtree. When `Some`, wire-out
+        /// uses these bytes verbatim and skips recursion into children.
+        /// Populated by the response parser from the server's exact
+        /// stream; also computable from children as a fallback. Cleared
+        /// on any edit to a descendant. Not serialized — transient.
+        #[serde(skip, default)]
+        token_ids: Option<Vec<u32>>,
     },
 }
 
 /// The context window: four sections as Vec<AstNode>.
-/// All mutation goes through ContextState methods to maintain the invariant
-/// that token_ids on every leaf matches its rendered text.
+///
+/// All mutation MUST go through `ContextState`'s public methods. Two
+/// invariants ride on this:
+/// 1. Every `Leaf.token_ids` matches its `body.compute_token_ids()`.
+/// 2. For every `Branch { token_ids: Some(cached), .. }`, the cached
+///    token stream matches what `wire_into` would produce by walking
+///    `children` from scratch. Any mutation that touches a Branch's
+///    children — directly or via a descendant — must clear the
+///    Branch's `token_ids` so it gets recomputed on next wire-out.
+///
+/// The `&mut Vec<AstNode>` escape hatches are intentionally NOT
+/// exposed; if you find yourself wanting one, add a focused method
+/// here that maintains the invariants.
 pub struct ContextState {
     system: Vec<AstNode>,
     identity: Vec<AstNode>,
     journal: Vec<AstNode>,
     conversation: Vec<AstNode>,
     pub conversation_log: Option<crate::mind::log::ConversationLog>,
+    /// Length of the session's token stream on the server, as of the
+    /// last Done event. Updated by the grpc layer.
+    server_committed_len: u32,
+    /// Prefix length of our walk that still matches the server's
+    /// session.tokens byte-for-byte. When < `server_committed_len`
+    /// the session needs rewinding (truncating=true at this offset).
+    /// Reset to 0 on any mutation that could have changed sent bytes.
+    client_match_upto: u32,
 }
 
 impl Clone for ContextState {
@@ -145,6 +203,8 @@ impl Clone for ContextState {
             journal: self.journal.clone(),
             conversation: self.conversation.clone(),
             conversation_log: None, // forked contexts don't log
+            server_committed_len: self.server_committed_len,
+            client_match_upto: self.client_match_upto,
         }
     }
 }
@@ -181,6 +241,10 @@ pub struct ResponseParser {
     think_buf: String,
     in_tool_call: bool,
     tool_call_buf: String,
+    /// Raw generated token IDs, in arrival order. Combined with the
+    /// prologue at `finish` to stamp the Branch's authoritative
+    /// token cache — the bytes the server has for this branch.
+    generated_tokens: Vec<u32>,
 }
 
 impl Role {
@@ -198,7 +262,11 @@ impl NodeBody {
     fn render_into(&self, out: &mut String) {
         match self {
             Self::Content(text)   => out.push_str(text),
-            Self::Thinking(_)     => {},
+            Self::Thinking(text)  => {
+                out.push_str("<think>\n");
+                out.push_str(text);
+                out.push_str("\n</think>\n");
+            }
             Self::Log(_)          => {},
             Self::ToolCall { name, arguments } => {
                 out.push_str("<tool_call>\n");
@@ -220,6 +288,13 @@ impl NodeBody {
                 out.push_str(text);
                 out.push_str("<|im_end|>\n");
             }
+            Self::Image { token_count, .. } => {
+                out.push_str("<|vision_start|>");
+                for _ in 0..*token_count {
+                    out.push_str("<|image_pad|>");
+                }
+                out.push_str("<|vision_end|>");
+            }
         }
     }
 
@@ -231,7 +306,27 @@ impl NodeBody {
     }
 
     fn is_prompt_visible(&self) -> bool {
-        !matches!(self, Self::Thinking(_) | Self::Log(_))
+        !matches!(self, Self::Log(_))
+    }
+
+    /// Hand-assemble token IDs for body types where running the tokenizer
+    /// on the rendered text would be needlessly expensive (Image). Falls
+    /// back to encoding the rendered text for everything else.
+    fn compute_token_ids(&self) -> Vec<u32> {
+        if !self.is_prompt_visible() {
+            return Vec::new();
+        }
+        match self {
+            Self::Image { token_count, .. } => {
+                let mut ids = Vec::with_capacity(*token_count as usize + 2);
+                ids.push(tokenizer::VISION_START);
+                ids.extend(std::iter::repeat(tokenizer::IMAGE_PAD)
+                    .take(*token_count as usize));
+                ids.push(tokenizer::VISION_END);
+                ids
+            }
+            _ => tokenizer::encode(&self.render()),
+        }
     }
 
     /// The text content of this leaf (for display, not rendering).
@@ -241,29 +336,36 @@ impl NodeBody {
                 | Self::ToolResult(t) | Self::Dmn(t) => t,
             Self::ToolCall { name, .. } => name,
             Self::Memory { text, .. } => text,
+            Self::Image { mime, .. } => mime,
         }
     }
 }
 
 impl NodeLeaf {
     fn new(body: NodeBody) -> Self {
-        let token_ids = if body.is_prompt_visible() {
-            tokenizer::encode(&body.render())
-        } else {
-            vec![]
-        };
-        Self { body, token_ids, timestamp: None }
+        let token_ids = body.compute_token_ids();
+        Self { body, token_ids, timestamp: Utc::now() }
     }
 
     pub fn with_timestamp(mut self, ts: DateTime<Utc>) -> Self {
-        self.timestamp = Some(ts);
+        self.timestamp = ts;
         self
     }
 
     pub fn body(&self) -> &NodeBody      { &self.body }
     pub fn token_ids(&self) -> &[u32]    { &self.token_ids }
     pub fn tokens(&self) -> usize        { self.token_ids.len() }
-    pub fn timestamp(&self) -> Option<DateTime<Utc>> { self.timestamp }
+    pub fn timestamp(&self) -> DateTime<Utc> { self.timestamp }
+
+    /// If this is an Image leaf, update its IMAGE_PAD count to `n` and
+    /// recompute cached `token_ids`. No-op on non-Image leaves —
+    /// callers know the body shape via `body()`.
+    pub fn set_image_token_count(&mut self, n: u32) {
+        if let NodeBody::Image { token_count, .. } = &mut self.body {
+            *token_count = n;
+            self.token_ids = self.body.compute_token_ids();
+        }
+    }
 }
 
 impl AstNode {
@@ -304,17 +406,46 @@ impl AstNode {
         Self::Leaf(NodeLeaf::new(NodeBody::Log(text.into())))
     }
 
+    /// Build an Image leaf. `token_count` is computed from the image
+    /// dimensions using Qwen3-VL's resizing rules.
+    pub fn image(
+        bytes: Vec<u8>,
+        mime: impl Into<String>,
+        orig_height: u32,
+        orig_width: u32,
+    ) -> Self {
+        // Pad count is computed eagerly from dimensions — no more
+        // "unknown until server responds" shape. Server validates
+        // on the Generate call; mismatches fail loud.
+        let token_count = qwen3_image_token_count(orig_height, orig_width);
+        Self::Leaf(NodeLeaf::new(NodeBody::Image {
+            bytes,
+            mime: mime.into(),
+            orig_height,
+            orig_width,
+            token_count,
+        }))
+    }
+
     // -- Branch constructors --------------------------------------------------
 
     pub fn branch(role: Role, children: Vec<AstNode>) -> Self {
-        Self::Branch { role, children, memory_scores: Default::default() }
+        Self::Branch {
+            role,
+            children,
+            timestamp: Utc::now(),
+            memory_scores: Default::default(),
+            token_ids: None,
+        }
     }
 
     pub fn system_msg(text: impl Into<String>) -> Self {
         Self::Branch {
             role: Role::System,
             children: vec![Self::content(text)],
+            timestamp: Utc::now(),
             memory_scores: Default::default(),
+            token_ids: None,
         }
     }
 
@@ -322,7 +453,9 @@ impl AstNode {
         Self::Branch {
             role: Role::User,
             children: vec![Self::content(text)],
+            timestamp: Utc::now(),
             memory_scores: Default::default(),
+            token_ids: None,
         }
     }
 
@@ -331,25 +464,23 @@ impl AstNode {
     pub fn retokenize(self) -> Self {
         match self {
             Self::Leaf(leaf) => {
-                let token_ids = if leaf.body.is_prompt_visible() {
-                    tokenizer::encode(&leaf.body.render())
-                } else {
-                    vec![]
-                };
+                let token_ids = leaf.body.compute_token_ids();
                 Self::Leaf(NodeLeaf { token_ids, ..leaf })
             }
-            Self::Branch { role, children, memory_scores, .. } => Self::Branch {
+            Self::Branch { role, children, timestamp, memory_scores, .. } => Self::Branch {
                 role,
                 children: children.into_iter().map(|c| c.retokenize()).collect(),
+                timestamp,
                 memory_scores,
+                token_ids: None,
             },
         }
     }
 
     pub fn with_timestamp(mut self, ts: DateTime<Utc>) -> Self {
         match &mut self {
-            Self::Leaf(leaf) => leaf.timestamp = Some(ts),
-            Self::Branch { .. } => {}
+            Self::Leaf(leaf) => leaf.timestamp = ts,
+            Self::Branch { timestamp, .. } => *timestamp = ts,
         }
         self
     }
@@ -370,7 +501,7 @@ impl AstNode {
 
     /// Short label for the UI.
     pub fn label(&self) -> String {
-        let cfg = crate::config::get();
+        let app = crate::config::app();
         match self {
             Self::Branch { role, children, .. } => {
                 let preview = children.first()
@@ -379,8 +510,8 @@ impl AstNode {
                     .unwrap_or_default();
                 match role {
                     Role::System => "system".into(),
-                    Role::User => format!("{}: {}", cfg.user_name, preview),
-                    Role::Assistant => format!("{}: {}", cfg.assistant_name, preview),
+                    Role::User => format!("{}: {}", app.user_name, preview),
+                    Role::Assistant => format!("{}: {}", app.assistant_name, preview),
                 }
             }
             Self::Leaf(leaf) => match &leaf.body {
@@ -393,6 +524,8 @@ impl AstNode {
                     None => format!("mem: {}", key),
                 },
                 NodeBody::Dmn(_) => "dmn".into(),
+                NodeBody::Image { orig_height, orig_width, token_count, .. } =>
+                    format!("image: {}x{} ({} tokens)", orig_width, orig_height, token_count),
                 NodeBody::Log(t) => format!("log: {}", truncate_preview(t, 60)),
             },
         }
@@ -416,7 +549,10 @@ impl AstNode {
     fn token_ids_into(&self, out: &mut Vec<u32>) {
         match self {
             Self::Leaf(leaf) => out.extend_from_slice(&leaf.token_ids),
-            Self::Branch { role, children, .. } => {
+            Self::Branch { token_ids: Some(cached), .. } => {
+                out.extend_from_slice(cached);
+            }
+            Self::Branch { role, children, token_ids: None, .. } => {
                 out.push(tokenizer::IM_START);
                 out.extend(tokenizer::encode(&format!("{}\n", role.as_str())));
                 for child in children {
@@ -445,7 +581,8 @@ impl Ast for AstNode {
     fn tokens(&self) -> usize {
         match self {
             Self::Leaf(leaf) => leaf.tokens(),
-            Self::Branch { role, children, .. } => {
+            Self::Branch { token_ids: Some(cached), .. } => cached.len(),
+            Self::Branch { role, children, token_ids: None, .. } => {
                 1 + role_header_tokens(*role)
                     + children.iter().map(|c| c.tokens()).sum::<usize>()
                     + 1 + newline_tokens()
@@ -585,16 +722,21 @@ fn drain_safe(buf: &mut String, tag_len: usize) -> String {
 }
 
 impl ResponseParser {
-    pub fn new(branch_idx: usize) -> Self {
+    /// @in_think: whether the model's output begins inside a <think> block.
+    /// Set when the prompt was prefilled with "<think>\n" (native thinking
+    /// mode) so the parser captures reasoning tokens as Thinking until the
+    /// model emits </think>.
+    pub fn new(branch_idx: usize, in_think: bool) -> Self {
         Self {
             branch_idx,
             call_counter: 0,
             buf: String::new(),
             content_parts: Vec::new(),
-            in_think: false,
+            in_think,
             think_buf: String::new(),
             in_tool_call: false,
             tool_call_buf: String::new(),
+            generated_tokens: Vec::new(),
         }
     }
 
@@ -613,13 +755,34 @@ impl ResponseParser {
         let handle = tokio::spawn(async move {
             let mut parser = self;
             let agent_name = agent.state.lock().await.provenance.clone();
+            eprintln!(
+                "[agent:{agent_name}] parser task start branch_idx={} in_think={}",
+                parser.branch_idx, parser.in_think,
+            );
             let log_path = format!("/tmp/poc-{}.log", agent_name);
             let mut log_file = std::fs::OpenOptions::new()
                 .create(true).append(true).open(&log_path).ok();
             let mut full_text = String::new();
+            let mut token_count: usize = 0;
             while let Some(event) = stream.recv().await {
                 match event {
-                    super::api::StreamToken::Token(id) => {
+                    super::api::StreamToken::Token { id, readout } => {
+                        token_count += 1;
+                        if token_count == 1 {
+                            eprintln!("[agent:{agent_name}] parser first token id={}", id);
+                        } else if token_count % 256 == 0 {
+                            eprintln!(
+                                "[agent:{agent_name}] parser token_count={} chars={}",
+                                token_count,
+                                full_text.len(),
+                            );
+                        }
+                        if let Some(r) = readout {
+                            if let Ok(mut buf) = agent.readout.lock() {
+                                buf.push(id, r);
+                            }
+                        }
+                        parser.generated_tokens.push(id);
                         let text = super::tokenizer::decode(&[id]);
                         full_text.push_str(&text);
                         let mut ctx = agent.context.lock().await;
@@ -638,6 +801,12 @@ impl ResponseParser {
                         }
                     }
                     super::api::StreamToken::Done { usage } => {
+                        eprintln!(
+                            "[agent:{agent_name}] parser done token_count={} chars={} usage={:?}",
+                            token_count,
+                            full_text.len(),
+                            usage,
+                        );
                         if let Some(ref mut f) = log_file {
                             use std::io::Write;
                             let ctx = agent.context.lock().await;
@@ -654,19 +823,31 @@ impl ResponseParser {
                                 let _ = writeln!(f, "  unparsed text: {}", &full_text[..end]);
                             }
                         }
-                        if let Some(u) = usage {
+                        if let Some(ref u) = usage {
                             agent.state.lock().await.last_prompt_tokens = u.prompt_tokens;
                         }
                         let mut ctx = agent.context.lock().await;
                         parser.finish(&mut ctx);
+                        if let Some(u) = usage {
+                            ctx.note_session_synced(u.total_tokens);
+                        }
                         return Ok(());
                     }
                     super::api::StreamToken::Error(e) => {
+                        eprintln!("[agent:{agent_name}] parser stream error: {}", e);
                         return Err(anyhow::anyhow!("{}", e));
                     }
                 }
             }
-            Ok(())
+            eprintln!(
+                "[agent:{agent_name}] parser stream closed without done token_count={} chars={}",
+                token_count,
+                full_text.len(),
+            );
+            Err(anyhow::anyhow!(
+                "stream closed without Done event after {} tokens",
+                token_count,
+            ))
         });
         (rx, handle)
     }
@@ -747,7 +928,7 @@ impl ResponseParser {
     }
 
     fn push_child(&self, ctx: &mut ContextState, child: AstNode) {
-        ctx.push_child(Section::Conversation, self.branch_idx, child);
+        ctx.push_child_raw(Section::Conversation, self.branch_idx, child);
     }
 
     fn flush_content(&mut self, ctx: &mut ContextState) {
@@ -761,10 +942,69 @@ impl ResponseParser {
     }
 
     pub fn finish(mut self, ctx: &mut ContextState) {
-        if !self.buf.is_empty() {
+        // Salvage any in-flight tag accumulators if the stream ended
+        // before the close tag arrived (max_tokens, premature EOS,
+        // server-side cancel). Without this, an unterminated
+        // <think>...</think> drops all of self.think_buf and only the
+        // trailing rolling window in self.buf survives — observed as
+        // "responses cut off, only the last ~8 characters come
+        // through" because drain_safe keeps `close_tag.len()` bytes
+        // (8 for `</think>`) at the tail of buf.
+        if self.in_think {
+            if !self.buf.is_empty() {
+                self.think_buf.push_str(&std::mem::take(&mut self.buf));
+            }
+            let text = std::mem::take(&mut self.think_buf).trim().to_string();
+            if !text.is_empty() {
+                self.push_child(ctx, AstNode::thinking(text));
+            }
+            self.in_think = false;
+        } else if self.in_tool_call {
+            if !self.buf.is_empty() {
+                self.tool_call_buf.push_str(&std::mem::take(&mut self.buf));
+            }
+            let body = std::mem::take(&mut self.tool_call_buf);
+            match parse_tool_call_body(&body) {
+                Some((name, args)) => {
+                    self.flush_content(ctx);
+                    self.push_child(ctx, AstNode::tool_call(&name, &args));
+                }
+                None => {
+                    // Body's likely incomplete (no `</tool_call>` ever
+                    // arrived). Wrap as content with the open tag so the
+                    // model can see its own truncated attempt next turn
+                    // rather than losing it silently.
+                    self.content_parts.push(format!("<tool_call>\n{}", body));
+                }
+            }
+            self.in_tool_call = false;
+        } else if !self.buf.is_empty() {
             self.content_parts.push(std::mem::take(&mut self.buf));
         }
         self.flush_content(ctx);
+
+        // Stamp the authoritative token cache onto the branch.
+        // Layout mirrors the full chat-template rendering of a
+        // message block:
+        //
+        //   IM_START + "assistant\n" [+ "<think>\n"]   (prologue — what we sent)
+        //   + generated_tokens                          (what the server generated, ends in IM_END)
+        //   + "\n"                                      (trailing newline — template-required)
+        //
+        // Server only has through the IM_END (model stops on it,
+        // doesn't emit "\n"). Match-upto lands inside the cache
+        // right after IM_END; the chunk-walk's straddle path picks
+        // up the trailing "\n" as the head of the next turn's delta.
+        // The "\n" between turns matters: without it Qwen sees
+        // `<|im_end|><|im_start|>` back-to-back (no newline) and
+        // responds with garbage.
+        let prologue_text = if self.in_think { "assistant\n<think>\n" } else { "assistant\n" };
+        let mut cache = Vec::with_capacity(1 + self.generated_tokens.len() + 8);
+        cache.push(tokenizer::IM_START);
+        cache.extend(tokenizer::encode(prologue_text));
+        cache.extend(self.generated_tokens);
+        cache.extend(tokenizer::encode("\n"));
+        ctx.set_branch_cache(Section::Conversation, self.branch_idx, cache);
     }
 }
 
@@ -776,20 +1016,77 @@ impl ContextState {
             journal: Vec::new(),
             conversation: Vec::new(),
             conversation_log: None,
+            server_committed_len: 0,
+            client_match_upto: 0,
         }
     }
 
+    // -- Server sync tracking -------------------------------------------------
+
+    /// Length of the session's token stream on the server. Updated by
+    /// the grpc layer from Generate Done events.
+    pub fn server_committed_len(&self) -> u32 { self.server_committed_len }
+
+    /// Prefix of our walk we still believe matches the server
+    /// byte-for-byte. If less than `server_committed_len`, the next
+    /// Generate must send `truncating=true` at this offset.
+    pub fn client_match_upto(&self) -> u32 { self.client_match_upto }
+
+    /// Called by the grpc layer after a successful Generate Done:
+    /// records both the server's new length and the fact that we
+    /// match up to it (we just sent everything).
+    pub fn note_session_synced(&mut self, total_tokens: u32) {
+        self.server_committed_len = total_tokens;
+        self.client_match_upto = total_tokens;
+    }
+
+    /// Reset match-upto to 0. Called from every mutation that could
+    /// have touched a region the server already has. For now,
+    /// conservatively drops alignment entirely — finer-grained
+    /// tracking (match-upto at the mutated node's offset) is a
+    /// future optimization.
+    fn mark_dirty(&mut self) {
+        self.client_match_upto = 0;
+    }
+
     // -- Read access ----------------------------------------------------------
 
     pub fn system(&self) -> &[AstNode]       { &self.system }
     pub fn identity(&self) -> &[AstNode]     { &self.identity }
     pub fn journal(&self) -> &[AstNode]      { &self.journal }
     pub fn conversation(&self) -> &[AstNode] { &self.conversation }
-    pub fn conversation_mut(&mut self) -> &mut Vec<AstNode> { &mut self.conversation }
+
+    /// Set or clear a single `memory_scores` entry on an Assistant
+    /// Branch. Used by the full-matrix scorer to attribute per-memory
+    /// divergence onto the response. `score = None` removes the key;
+    /// `Some(s)` inserts/overwrites.
+    ///
+    /// Doesn't affect the Branch's token cache: `memory_scores` is a
+    /// serialized-but-non-tokenizing annotation. No-op (with a debug
+    /// log) if the index points to a Leaf or a non-Assistant Branch —
+    /// callers are typically iterating on stale indices and we'd
+    /// rather skip than panic.
+    pub fn set_branch_memory_score(
+        &mut self,
+        section: Section,
+        index: usize,
+        key: &str,
+        score: Option<f64>,
+    ) {
+        let nodes = self.section_mut(section);
+        let Some(node) = nodes.get_mut(index) else { return };
+        let AstNode::Branch { role: Role::Assistant, memory_scores, .. } = node
+        else { return };
+        match score {
+            Some(s) => { memory_scores.insert(key.to_string(), s); }
+            None => { memory_scores.remove(key); }
+        }
+    }
 
     pub fn sections(&self) -> [&Vec<AstNode>; 4] {
         [&self.system, &self.identity, &self.journal, &self.conversation]
     }
+
 }
 
 impl Ast for ContextState {
@@ -821,6 +1118,367 @@ impl Ast for ContextState {
     }
 }
 
+/// An image collected from the AST for a request body. The AST stores
+/// Image metadata collected during `wire_chunks` — the binary +
+/// mime plus the absolute token-position range of the image's
+/// pre-expanded placeholder run in the full wire stream. Sent
+/// alongside `append_tokens` in `GenerateRequest` so the server
+/// can attach vision features to the declared positions. Positions
+/// are absolute within the full wire walk starting at offset 0,
+/// i.e. the same coordinate system as `session.tokens` on the
+/// server once the walk has been applied.
+#[derive(Clone)]
+pub struct WireImage {
+    pub bytes: Vec<u8>,
+    pub mime: String,
+    pub pad_start: u32,
+    pub pad_end: u32,
+}
+
+/// One piece of the wire stream for the gRPC session path. Since
+/// images now live inline in the token stream (pre-expanded at AST
+/// construction time), there's only one variant — a run of tokens.
+/// The parallel `Vec<WireImage>` returned by `wire_chunks` gives the
+/// binary + position metadata for each embedded image.
+#[derive(Clone)]
+pub enum WireChunk {
+    Tokens(Vec<u32>),
+}
+
+fn wire_into(node: &AstNode, tokens: &mut Vec<u32>, images: &mut Vec<WireImage>) {
+    match node {
+        AstNode::Leaf(leaf) => match leaf.body() {
+            NodeBody::Image { bytes, mime, .. } => {
+                // The Image leaf's token_ids is already
+                // [VISION_START, IMAGE_PAD * N, VISION_END]. Inline
+                // those into the token stream and record the pad-run
+                // range so the server can attach features to the
+                // declared positions.
+                let pad_start = tokens.len() as u32;
+                tokens.extend_from_slice(leaf.token_ids());
+                let pad_end = tokens.len() as u32;
+                images.push(WireImage {
+                    bytes: bytes.clone(),
+                    mime: mime.clone(),
+                    pad_start,
+                    pad_end,
+                });
+            }
+            _ => tokens.extend_from_slice(leaf.token_ids()),
+        },
+        AstNode::Branch { token_ids: Some(cached), children, .. } => {
+            // Cached branches still need their image children paired
+            // up with the vision-block ranges embedded in the cached
+            // token stream — the cache captures vision tokens but not
+            // the matching bytes/mime.
+            let base = tokens.len() as u32;
+            tokens.extend_from_slice(cached);
+            pair_cached_images(cached, children, base, images);
+        }
+        AstNode::Branch { role, children, token_ids: None, .. } => {
+            tokens.push(tokenizer::IM_START);
+            tokens.extend(tokenizer::encode(&format!("{}\n", role.as_str())));
+            for c in children {
+                wire_into(c, tokens, images);
+            }
+            tokens.push(tokenizer::IM_END);
+            tokens.extend(tokenizer::encode("\n"));
+        }
+    }
+}
+
+/// Depth-first iterator over Image leaves under a slice of AST nodes.
+/// Yields `(bytes, mime)` borrows in document order; doesn't allocate
+/// per yield (only a stack of pending nodes).
+struct ImageLeaves<'a> {
+    stack: Vec<&'a AstNode>,
+}
+
+impl<'a> ImageLeaves<'a> {
+    fn new(nodes: &'a [AstNode]) -> Self {
+        let mut stack = Vec::with_capacity(nodes.len());
+        stack.extend(nodes.iter().rev());
+        Self { stack }
+    }
+}
+
+impl<'a> Iterator for ImageLeaves<'a> {
+    type Item = (&'a [u8], &'a str);
+    fn next(&mut self) -> Option<Self::Item> {
+        while let Some(node) = self.stack.pop() {
+            match node {
+                AstNode::Leaf(leaf) => {
+                    if let NodeBody::Image { bytes, mime, .. } = leaf.body() {
+                        return Some((bytes, mime));
+                    }
+                }
+                AstNode::Branch { children, .. } => {
+                    self.stack.extend(children.iter().rev());
+                }
+            }
+        }
+        None
+    }
+}
+
+/// Iterator over `(start, end)` token-offset pairs for each
+/// `VISION_START..VISION_END` block in a token slice. Panics on an
+/// unmatched VISION_START — that's an upstream tokenization bug
+/// worth a loud failure.
+fn vision_blocks(cached: &[u32]) -> impl Iterator<Item = (usize, usize)> + '_ {
+    let mut cur = 0;
+    std::iter::from_fn(move || {
+        while cur < cached.len() {
+            if cached[cur] == tokenizer::VISION_START {
+                let start = cur;
+                let end_rel = cached[cur..].iter()
+                    .position(|&t| t == tokenizer::VISION_END)
+                    .unwrap_or_else(|| panic!(
+                        "unmatched VISION_START at offset {} in cached branch",
+                        start));
+                let end = cur + end_rel + 1;
+                cur = end;
+                return Some((start, end));
+            }
+            cur += 1;
+        }
+        None
+    })
+}
+
+/// For a Branch whose `token_ids` are cached and may contain inlined
+/// vision blocks (`VISION_START + IMAGE_PAD*N + VISION_END`), recover
+/// the matching image bytes/mime from the children and emit one
+/// `WireImage` per vision block with the absolute pad offsets in the
+/// parent token stream.
+///
+/// The cache stores tokens but not image payloads; the AST stores
+/// image payloads in the children but not their post-cache positions.
+/// Pair them by zipping the two iterators; mismatched counts panic
+/// loudly because that's an AST/cache invariant violation that
+/// would otherwise mis-pair images on the wire.
+fn pair_cached_images(
+    cached: &[u32],
+    children: &[AstNode],
+    base_offset: u32,
+    images: &mut Vec<WireImage>,
+) {
+    let mut blocks = vision_blocks(cached);
+    let mut leaves = ImageLeaves::new(children);
+    loop {
+        match (blocks.next(), leaves.next()) {
+            (Some((s, e)), Some((bytes, mime))) => images.push(WireImage {
+                bytes: bytes.to_vec(),
+                mime: mime.to_string(),
+                pad_start: base_offset + s as u32,
+                pad_end: base_offset + e as u32,
+            }),
+            (None, None) => break,
+            (Some(_), None) => panic!(
+                "cached branch has more vision blocks than image children"),
+            (None, Some(_)) => panic!(
+                "cached branch has fewer vision blocks than image children"),
+        }
+    }
+}
+
+pub fn memory_key(node: &AstNode) -> Option<&str> {
+    match node {
+        AstNode::Leaf(leaf) => match leaf.body() {
+            NodeBody::Memory { key, .. } => Some(key),
+            _ => None,
+        },
+        _ => None,
+    }
+}
+
+pub fn is_memory_node(node: &AstNode) -> bool {
+    matches!(node, AstNode::Leaf(leaf) if matches!(leaf.body(), NodeBody::Memory { .. }))
+}
+
+pub fn is_assistant(node: &AstNode) -> bool {
+    matches!(node, AstNode::Branch { role: Role::Assistant, .. })
+}
+
+/// Concatenate the text of a Branch's Leaf children — what the model
+/// actually produced on that turn (Content + Thinking + ToolCall name).
+pub fn render_branch_text(children: &[AstNode]) -> String {
+    children.iter()
+        .filter_map(|c| match c {
+            AstNode::Leaf(leaf) => Some(leaf.body().text().to_string()),
+            _ => None,
+        })
+        .collect::<Vec<_>>()
+        .join("")
+}
+
+/// Render the last `max_msgs` user/assistant branches before `idx` as a
+/// review-friendly string with `[user]` / `[assistant]` markers.
+pub fn render_prior_context(entries: &[AstNode], idx: usize, max_msgs: usize) -> String {
+    let mut picked: Vec<&AstNode> = Vec::with_capacity(max_msgs);
+    for i in (0..idx).rev() {
+        if picked.len() >= max_msgs { break; }
+        if let AstNode::Branch { role, .. } = &entries[i] {
+            if matches!(role, Role::User | Role::Assistant) {
+                picked.push(&entries[i]);
+            }
+        }
+    }
+    picked.reverse();
+
+    let mut out = String::new();
+    for node in picked {
+        if let AstNode::Branch { role, children, .. } = node {
+            let marker = match role {
+                Role::User => "[user]",
+                Role::Assistant => "[assistant]",
+                _ => continue,
+            };
+            out.push_str(marker);
+            out.push('\n');
+            out.push_str(render_branch_text(children).trim());
+            out.push_str("\n\n");
+        }
+    }
+    out.trim_end().to_string()
+}
+
+impl ContextState {
+    /// Assemble the prompt in wire form: token stream with a single
+    /// `<|image_pad|>` per image (vLLM expands back to N), plus the list
+    /// of images to send as multi_modal_data, plus the (start, end) token
+    /// positions of each assistant message branch emitted (used by the
+    /// scoring path as `score_ranges`).
+    ///
+    /// `conv_range` selects a prefix (or any sub-range) of conversation
+    /// entries to include — the agent path passes `0..conversation().len()`;
+    /// scoring / candidate generation pass a prefix up to the entry of
+    /// interest.
+    ///
+    /// `skip` is a predicate applied to identity and conversation entries;
+    /// returning true drops the node from the prompt. The agent path passes
+    /// `|_| false`; memory-ablation scoring passes e.g. `is_memory_node` or
+    /// `|n| memory_key(n) == Some(key)`.
+    pub fn wire_prompt<F>(
+        &self,
+        conv_range: std::ops::Range<usize>,
+        mut skip: F,
+    ) -> (Vec<u32>, Vec<WireImage>, Vec<(usize, usize)>)
+    where F: FnMut(&AstNode) -> bool,
+    {
+        let mut tokens = Vec::new();
+        let mut images = Vec::new();
+        let mut assistant_ranges = Vec::new();
+
+        for node in self.system() {
+            wire_into(node, &mut tokens, &mut images);
+        }
+        for node in self.identity() {
+            if skip(node) { continue; }
+            wire_into(node, &mut tokens, &mut images);
+        }
+        for node in self.journal() {
+            wire_into(node, &mut tokens, &mut images);
+        }
+        for node in &self.conversation()[conv_range] {
+            if skip(node) { continue; }
+            let start = tokens.len();
+            let is_asst = matches!(node, AstNode::Branch { role: Role::Assistant, .. });
+            wire_into(node, &mut tokens, &mut images);
+            if is_asst {
+                assistant_ranges.push((start, tokens.len()));
+            }
+        }
+        (tokens, images, assistant_ranges)
+    }
+
+    /// Build the wire stream as interleaved `WireChunk`s for the gRPC
+    /// session path. Returns a tuple of (chunks, images): the chunks
+    /// hold the full token stream (with vision blocks inlined as
+    /// `VISION_START + IMAGE_PAD*N + VISION_END`), and the images
+    /// list carries each embedded image's binary + position range so
+    /// the gRPC layer can attach them via `GenerateRequest.images`.
+    ///
+    /// Note: with images inlined into the token stream, the chunks
+    /// list is structurally a single `Tokens` chunk in the common
+    /// case — the multi-chunk shape persists only because some
+    /// callers may want the option of inserting breakpoints later.
+    ///
+    /// `conv_range` and `skip` mirror `wire_prompt` — select a
+    /// conversation slice and drop identity / conversation nodes by
+    /// predicate.
+    pub fn wire_chunks<F>(
+        &self,
+        conv_range: std::ops::Range<usize>,
+        mut skip: F,
+    ) -> (Vec<WireChunk>, Vec<WireImage>)
+    where F: FnMut(&AstNode) -> bool,
+    {
+        let mut buf: Vec<u32> = Vec::new();
+        let mut images: Vec<WireImage> = Vec::new();
+
+        fn visit(
+            node: &AstNode,
+            buf: &mut Vec<u32>,
+            images: &mut Vec<WireImage>,
+        ) {
+            match node {
+                AstNode::Leaf(leaf) => match leaf.body() {
+                    NodeBody::Image { bytes, mime, .. } => {
+                        // Pre-expanded vision block lives in
+                        // leaf.token_ids: [VISION_START, IMAGE_PAD*N,
+                        // VISION_END]. Inline + record the range.
+                        let pad_start = buf.len() as u32;
+                        buf.extend_from_slice(leaf.token_ids());
+                        let pad_end = buf.len() as u32;
+                        images.push(WireImage {
+                            bytes: bytes.clone(),
+                            mime: mime.clone(),
+                            pad_start,
+                            pad_end,
+                        });
+                    }
+                    _ => buf.extend_from_slice(leaf.token_ids()),
+                },
+                AstNode::Branch { token_ids: Some(cached), children, .. } => {
+                    // Same fix as wire_into's cached arm: the cache
+                    // holds vision tokens but not the matching bytes,
+                    // so walk children to recover them.
+                    let base = buf.len() as u32;
+                    buf.extend_from_slice(cached);
+                    pair_cached_images(cached, children, base, images);
+                }
+                AstNode::Branch { role, children, token_ids: None, .. } => {
+                    buf.push(tokenizer::IM_START);
+                    buf.extend(tokenizer::encode(&format!("{}\n", role.as_str())));
+                    for c in children {
+                        visit(c, buf, images);
+                    }
+                    buf.push(tokenizer::IM_END);
+                    buf.extend(tokenizer::encode("\n"));
+                }
+            }
+        }
+
+        for node in self.system()   { visit(node, &mut buf, &mut images); }
+        for node in self.identity() {
+            if skip(node) { continue; }
+            visit(node, &mut buf, &mut images);
+        }
+        for node in self.journal()  { visit(node, &mut buf, &mut images); }
+        for node in &self.conversation()[conv_range] {
+            if skip(node) { continue; }
+            visit(node, &mut buf, &mut images);
+        }
+        let chunks = if buf.is_empty() {
+            Vec::new()
+        } else {
+            vec![WireChunk::Tokens(buf)]
+        };
+        (chunks, images)
+    }
+}
+
 impl ContextState {
     fn section_mut(&mut self, section: Section) -> &mut Vec<AstNode> {
         match section {
@@ -838,26 +1496,32 @@ impl ContextState {
                 dbglog!("warning: log: {:#}", e);
             }
         }
+        // Conversation appends always go to the tail — past committed —
+        // so they don't break the match. Any other section mutates a
+        // region the server may already have, so drop alignment.
+        if section != Section::Conversation {
+            self.mark_dirty();
+        }
         self.section_mut(section).push(node);
     }
 
     /// Push without logging.
     pub fn push_no_log(&mut self, section: Section, node: AstNode) {
+        if section != Section::Conversation {
+            self.mark_dirty();
+        }
         self.section_mut(section).push(node);
     }
 
     /// Replace the body of a leaf at `index` in `section`.
     /// Re-tokenizes to maintain the invariant.
     pub fn set_message(&mut self, section: Section, index: usize, body: NodeBody) {
+        self.mark_dirty();
         let nodes = self.section_mut(section);
         let node = &mut nodes[index];
         match node {
             AstNode::Leaf(leaf) => {
-                let token_ids = if body.is_prompt_visible() {
-                    tokenizer::encode(&body.render())
-                } else {
-                    vec![]
-                };
+                let token_ids = body.compute_token_ids();
                 leaf.body = body;
                 leaf.token_ids = token_ids;
             }
@@ -878,13 +1542,25 @@ impl ContextState {
     }
 
     pub fn del(&mut self, section: Section, index: usize) -> AstNode {
+        self.mark_dirty();
         self.section_mut(section).remove(index)
     }
 
     pub fn clear(&mut self, section: Section) {
+        self.mark_dirty();
         self.section_mut(section).clear();
     }
 
+    /// Total tokens across every section that gets serialized into the prompt.
+    /// Cheap sum over cached `node.tokens()`; call this before assembling to
+    /// decide whether to trim.
+    pub fn total_tokens(&self) -> usize {
+        self.system().iter().map(|n| n.tokens()).sum::<usize>()
+            + self.identity().iter().map(|n| n.tokens()).sum::<usize>()
+            + self.journal().iter().map(|n| n.tokens()).sum::<usize>()
+            + self.conversation().iter().map(|n| n.tokens()).sum::<usize>()
+    }
+
     /// Dedup and trim conversation entries to fit within the context budget.
     ///
     /// Phase 1: Drop duplicate memories (keep last) and DMN entries.
@@ -892,6 +1568,7 @@ impl ContextState {
     ///          are > 50% of conversation tokens) or oldest conversation entry.
     /// Phase 3: Snap to user message boundary at start.
     pub fn trim_conversation(&mut self) {
+        self.mark_dirty();
         let max_tokens = context_budget_tokens();
         let fixed = self.system.iter().map(|n| n.tokens()).sum::<usize>()
             + self.identity.iter().map(|n| n.tokens()).sum::<usize>()
@@ -968,11 +1645,49 @@ impl ContextState {
     }
 
     /// Push a child node into a branch at `index` in `section`.
+    /// Clears the branch's cached token stream — wire-out will recompute
+    /// from children until the cache is repopulated. If the cache was
+    /// populated (server had these bytes), drops session alignment.
     pub fn push_child(&mut self, section: Section, index: usize, child: AstNode) {
+        let node = &mut self.section_mut(section)[index];
+        let was_cached = matches!(node, AstNode::Branch { token_ids: Some(_), .. });
+        match node {
+            AstNode::Branch { children, token_ids, .. } => {
+                children.push(child);
+                *token_ids = None;
+            }
+            AstNode::Leaf(_) => panic!("push_child on leaf node"),
+        }
+        if was_cached {
+            self.mark_dirty();
+        }
+    }
+
+    /// Like `push_child` but preserves the branch's cached token stream.
+    /// Used by the response parser, which is simultaneously populating
+    /// the cache from the authoritative server stream and pushing the
+    /// parsed-out children — the two stay consistent by construction.
+    /// Module-private: callers outside `context.rs` must go through
+    /// `push_child` so the invariant is maintained.
+    fn push_child_raw(&mut self, section: Section, index: usize, child: AstNode) {
         let node = &mut self.section_mut(section)[index];
         match node {
             AstNode::Branch { children, .. } => children.push(child),
-            AstNode::Leaf(_) => panic!("push_child on leaf node"),
+            AstNode::Leaf(_) => panic!("push_child_raw on leaf node"),
+        }
+    }
+
+    /// Stamp a verbatim token cache onto the branch at `index` in
+    /// `section`. Used by the response parser to record the server's
+    /// authoritative token stream for the just-finished turn.
+    /// Module-private: the cache is an invariant-load-bearing piece
+    /// of state, populated only by code that holds the server's
+    /// ground truth.
+    fn set_branch_cache(&mut self, section: Section, index: usize, tokens: Vec<u32>) {
+        let node = &mut self.section_mut(section)[index];
+        match node {
+            AstNode::Branch { token_ids, .. } => *token_ids = Some(tokens),
+            AstNode::Leaf(_) => panic!("set_branch_cache on leaf node"),
         }
     }
 
@@ -987,8 +1702,70 @@ impl ContextState {
     }
 }
 
+// ---------------------------------------------------------------------------
+// Qwen3-VL image token count
+//
+// Port of Qwen2VLImageProcessor.smart_resize + image_token_count. We need the
+// exact same answer that vLLM's Qwen3VL processor will produce, because the
+// token stream in our context must match what vLLM expands `<|image_pad|>`
+// to at request time. Constants come from Qwen3.5-27B's preprocessor_config.
+// ---------------------------------------------------------------------------
+
+// Production client-side computation of image-token expansion. With
+// the delta-session protocol, the client writes the pre-expanded
+// vision block (VISION_START + N*IMAGE_PAD + VISION_END) directly
+// into the token stream at Image-leaf construction time, and tells
+// the server where each image's pad run lives via
+// GenerateRequest.images. Server validates that this N matches
+// what the vision encoder actually produces and rejects on
+// mismatch — so drift here fails loudly, not silently.
+const QWEN3_PATCH_SIZE: u32 = 16;
+const QWEN3_MERGE_SIZE: u32 = 2;
+const QWEN3_MIN_PIXELS: u64 = 65_536;
+const QWEN3_MAX_PIXELS: u64 = 16_777_216;
+
+fn smart_resize(h: u32, w: u32, factor: u32, min_pixels: u64, max_pixels: u64) -> (u32, u32) {
+    let max_s = h.max(w) as f64;
+    let min_s = h.min(w) as f64;
+    assert!(max_s / min_s <= 200.0, "aspect ratio too extreme: {}x{}", h, w);
+
+    let fh = h as f64;
+    let fw = w as f64;
+    let ff = factor as f64;
+
+    let h_bar = ((fh / ff).round() as u32) * factor;
+    let w_bar = ((fw / ff).round() as u32) * factor;
+    let total = (h_bar as u64) * (w_bar as u64);
+
+    if total > max_pixels {
+        let beta = ((fh * fw) / max_pixels as f64).sqrt();
+        let hf = ((fh / beta / ff).floor() as u32) * factor;
+        let wf = ((fw / beta / ff).floor() as u32) * factor;
+        (hf.max(factor), wf.max(factor))
+    } else if total < min_pixels {
+        let beta = (min_pixels as f64 / (fh * fw)).sqrt();
+        let hc = ((fh * beta / ff).ceil() as u32) * factor;
+        let wc = ((fw * beta / ff).ceil() as u32) * factor;
+        (hc, wc)
+    } else {
+        (h_bar, w_bar)
+    }
+}
+
+/// How many `<|image_pad|>` tokens the Qwen3-VL vision encoder will
+/// produce for an image of the given dimensions. Server verifies
+/// this count against its own encoder run and rejects on mismatch.
+pub fn qwen3_image_token_count(orig_h: u32, orig_w: u32) -> u32 {
+    let factor = QWEN3_PATCH_SIZE * QWEN3_MERGE_SIZE;
+    let (rh, rw) = smart_resize(orig_h, orig_w, factor, QWEN3_MIN_PIXELS, QWEN3_MAX_PIXELS);
+    (rh / QWEN3_PATCH_SIZE) * (rw / QWEN3_PATCH_SIZE) / (QWEN3_MERGE_SIZE * QWEN3_MERGE_SIZE)
+}
+
 pub fn context_window() -> usize {
-    crate::config::get().api_context_window
+    let app = crate::config::app();
+    app.backends.get(&app.default_backend)
+        .and_then(|b| b.context_window)
+        .unwrap_or(128_000)
 }
 
 pub fn context_budget_tokens() -> usize {
@@ -1093,7 +1870,7 @@ mod tests {
     fn parse_into_ctx(chunks: &[&str]) -> (ContextState, Vec<PendingToolCall>) {
         let mut ctx = ContextState::new();
         ctx.push_no_log(Section::Conversation, AstNode::branch(Role::Assistant, vec![]));
-        let mut p = ResponseParser::new(0);
+        let mut p = ResponseParser::new(0, false);
         let mut calls = Vec::new();
         for chunk in chunks {
             // Feed each chunk as a single token (id=0 for tests)
@@ -1157,7 +1934,7 @@ mod tests {
         let text = "<think>thought</think>response";
         let mut ctx = ContextState::new();
         ctx.push_no_log(Section::Conversation, AstNode::branch(Role::Assistant, vec![]));
-        let mut p = ResponseParser::new(0);
+        let mut p = ResponseParser::new(0, false);
         for ch in text.chars() {
             p.feed_token(&ch.to_string(), &mut ctx);
         }
@@ -1173,7 +1950,7 @@ mod tests {
         let text = "text<tool_call>\n<function=bash>\n<parameter=command>ls</parameter>\n</function>\n</tool_call>more";
         let mut ctx = ContextState::new();
         ctx.push_no_log(Section::Conversation, AstNode::branch(Role::Assistant, vec![]));
-        let mut p = ResponseParser::new(0);
+        let mut p = ResponseParser::new(0, false);
         let mut tool_calls = 0;
         for ch in text.chars() {
             tool_calls += p.feed_token(&ch.to_string(), &mut ctx).len();
@@ -1221,8 +1998,10 @@ mod tests {
             AstNode::thinking("hmm"),
             AstNode::content("answer"),
         ]);
-        // Thinking renders as empty, content renders as-is
-        assert_eq!(node.render(), "<|im_start|>assistant\nanswer<|im_end|>\n");
+        // Thinking renders wrapped in <think>...</think> so the model sees
+        // previous turns' reasoning (Qwen 3.6 style: CoT stays in the
+        // conversation across turns).
+        assert_eq!(node.render(), "<|im_start|>assistant\n<think>\nhmm\n</think>\nanswer<|im_end|>\n");
     }
 
     #[test]
@@ -1301,10 +2080,19 @@ mod tests {
     fn test_tokenize_invisible_nodes_are_zero() {
         if !init_tokenizer() { return; }
 
-        assert_eq!(AstNode::thinking("deep thoughts").tokens(), 0);
         assert_eq!(AstNode::log("debug info").tokens(), 0);
     }
 
+    #[test]
+    fn test_tokenize_thinking_matches_rendered_tags() {
+        if !init_tokenizer() { return; }
+
+        // Thinking is now prompt-visible (wrapped in <think>...</think>);
+        // token count must match the rendered wrapping.
+        let node = AstNode::thinking("deep thoughts");
+        assert_eq!(node.tokens(), tokenizer::encode(&node.render()).len());
+    }
+
     #[test]
     fn test_tokenize_decode_roundtrip() {
         if !init_tokenizer() { return; }
@@ -1340,4 +2128,144 @@ mod tests {
         assert_token_invariants(node);
         assert!(node.tokens() > 0);
     }
+
+    // -- Timestamp deserialization tests ------------------------------------------
+
+    #[test]
+    fn test_timestamp_null_rejected() {
+        // Missing/null timestamps used to be accepted via a lenient
+        // deserialize fallback. Post-migration the schema is strict.
+        let json = r#"{"Leaf":{"body":{"Content":"hello"},"timestamp":null}}"#;
+        assert!(serde_json::from_str::<AstNode>(json).is_err());
+    }
+
+    #[test]
+    fn test_timestamp_missing_rejected() {
+        let json = r#"{"Leaf":{"body":{"Content":"hello"}}}"#;
+        assert!(serde_json::from_str::<AstNode>(json).is_err());
+    }
+
+    #[test]
+    fn test_branch_timestamp_missing_rejected() {
+        let json = r#"{"Branch":{"role":"User","children":[]}}"#;
+        assert!(serde_json::from_str::<AstNode>(json).is_err());
+    }
+
+    // -- Image leaf tests ---------------------------------------------------------
+
+    #[test]
+    fn test_smart_resize_within_bounds() {
+        // Typical case: 1024x768 → rounded to multiples of 32, under max.
+        let (h, w) = smart_resize(768, 1024, 32, 65_536, 16_777_216);
+        assert_eq!(h, 768);
+        assert_eq!(w, 1024);
+    }
+
+    #[test]
+    fn test_smart_resize_upscales_tiny() {
+        // 32x32 = 1024 pixels, below min_pixels=65536. Should scale up.
+        let (h, w) = smart_resize(32, 32, 32, 65_536, 16_777_216);
+        assert!((h as u64) * (w as u64) >= 65_536,
+            "resized {}x{} is under min_pixels", h, w);
+        assert_eq!(h % 32, 0);
+        assert_eq!(w % 32, 0);
+    }
+
+    #[test]
+    fn test_smart_resize_downscales_huge() {
+        // 8000x6000 = 48M pixels, above max_pixels=16M. Should scale down.
+        let (h, w) = smart_resize(8000, 6000, 32, 65_536, 16_777_216);
+        assert!((h as u64) * (w as u64) <= 16_777_216,
+            "resized {}x{} exceeds max_pixels", h, w);
+        assert_eq!(h % 32, 0);
+        assert_eq!(w % 32, 0);
+    }
+
+    #[test]
+    fn test_qwen3_token_count_matches_formula() {
+        // 512x512 → resized to 512x512 (already multiple of 32, within bounds).
+        // grid = 32x32, tokens = 32*32/4 = 256.
+        assert_eq!(qwen3_image_token_count(512, 512), 256);
+    }
+
+    #[test]
+    fn test_image_render_and_token_ids() {
+        let node = AstNode::image(vec![0u8, 1, 2, 3], "image/png", 512, 512);
+        let leaf = node.leaf().unwrap();
+        // 3 tokens of bookend + 256 image_pad tokens
+        assert_eq!(leaf.token_ids().len(), 258);
+        assert_eq!(leaf.token_ids()[0], tokenizer::VISION_START);
+        assert_eq!(leaf.token_ids()[257], tokenizer::VISION_END);
+        for pad in &leaf.token_ids()[1..257] {
+            assert_eq!(*pad, tokenizer::IMAGE_PAD);
+        }
+        // Rendered text has the expected bookends.
+        let rendered = leaf.body().render();
+        assert!(rendered.starts_with("<|vision_start|>"));
+        assert!(rendered.ends_with("<|vision_end|>"));
+    }
+
+    #[test]
+    fn test_wire_prompt_preserves_expanded_image_pads() {
+        let mut ctx = ContextState::new();
+        ctx.push_no_log(Section::Conversation, AstNode::branch(Role::User, vec![
+            AstNode::content("look:"),
+            AstNode::image(vec![0xDE, 0xAD], "image/png", 512, 512),
+        ]));
+
+        // AST side and wire side should both carry N image_pads + bookends —
+        // server's session.tokens length must match what vLLM's engine will
+        // actually process. Binary image bytes are shipped separately in
+        // multi_modal_data via the WireImage list.
+        let n_expected = qwen3_image_token_count(512, 512) as usize;
+
+        let full = ctx.token_ids();
+        let n_image_pads_full = full.iter()
+            .filter(|&&t| t == tokenizer::IMAGE_PAD).count();
+        assert_eq!(n_image_pads_full, n_expected);
+
+        let (wire, images, _) = ctx.wire_prompt(0..ctx.conversation().len(), |_| false);
+        let n_image_pads_wire = wire.iter()
+            .filter(|&&t| t == tokenizer::IMAGE_PAD).count();
+        assert_eq!(n_image_pads_wire, n_expected);
+
+        assert_eq!(images.len(), 1);
+        assert_eq!(images[0].bytes, vec![0xDE, 0xAD]);
+        assert_eq!(images[0].mime, "image/png");
+
+        // One pair of vision_start/vision_end bookends around the N pads.
+        assert_eq!(wire.iter().filter(|&&t| t == tokenizer::VISION_START).count(), 1);
+        assert_eq!(wire.iter().filter(|&&t| t == tokenizer::VISION_END).count(), 1);
+    }
+
+    #[test]
+    fn test_image_serde_roundtrip() {
+        let node = AstNode::image(vec![0xDE, 0xAD, 0xBE, 0xEF], "image/png", 64, 64);
+        let json = serde_json::to_string(&node).unwrap();
+        // bytes must be base64-encoded in the JSON form
+        assert!(json.contains("3q2+7w=="));
+        let back: AstNode = serde_json::from_str(&json).unwrap();
+        let leaf = back.leaf().unwrap();
+        match leaf.body() {
+            NodeBody::Image { bytes, mime, orig_height, orig_width, token_count } => {
+                assert_eq!(bytes, &[0xDE, 0xAD, 0xBE, 0xEF]);
+                assert_eq!(mime, "image/png");
+                assert_eq!(*orig_height, 64);
+                assert_eq!(*orig_width, 64);
+                assert_eq!(*token_count, qwen3_image_token_count(64, 64));
+            }
+            other => panic!("expected Image, got {:?}", other),
+        }
+        // token_ids are recomputed on deserialization
+        assert_eq!(leaf.token_ids().len(), leaf.tokens());
+    }
+
+    #[test]
+    fn test_timestamp_present_accepted() {
+        let json = r#"{"Leaf":{"body":{"Content":"hi"},"timestamp":"2026-04-16T12:00:00Z"}}"#;
+        let node: AstNode = serde_json::from_str(json).unwrap();
+        let leaf = node.leaf().unwrap();
+        assert_eq!(leaf.timestamp().to_rfc3339(),
+            "2026-04-16T12:00:00+00:00");
+    }
 }
diff --git a/src/agent/mod.rs b/src/agent/mod.rs
index db1bf39..1671dca 100644
--- a/src/agent/mod.rs
+++ b/src/agent/mod.rs
@@ -16,6 +16,8 @@
 pub mod api;
 pub mod context;
 pub mod oneshot;
+pub mod readout;
+pub mod salience;
 pub mod tokenizer;
 pub mod tools;
 
@@ -27,6 +29,11 @@ use context::{AstNode, ContextState, Section, Ast, PendingToolCall, ResponsePars
 
 use crate::mind::log::ConversationLog;
 
+async fn agent_trace(agent: &Arc<Agent>, msg: String) {
+    let provenance = agent.state.lock().await.provenance.clone();
+    eprintln!("[agent:{provenance}] {msg}");
+}
+
 // --- Activity tracking (RAII guards) ---
 
 pub struct ActivityEntry {
@@ -139,10 +146,22 @@ impl DispatchState {
 pub struct Agent {
     pub client: ApiClient,
     pub app_config: crate::config::AppConfig,
-    pub prompt_file: String,
     pub session_id: String,
     pub context: crate::Mutex<ContextState>,
     pub state: crate::Mutex<AgentState>,
+    /// Shared landing pad for per-token concept-readout projections
+    /// streamed from the vLLM server. Populated by the streaming
+    /// token handler, read by UI screens (amygdala). Manifest is
+    /// `None` when the server has readout disabled.
+    pub readout: readout::SharedReadoutBuffer,
+    /// Long-lived gRPC session to the salience server, lazily opened
+    /// on first use. Tracks appended tokens so subsequent turns send
+    /// only the delta (prefix-cache reuse). None when not yet opened
+    /// or when the session has died and needs reopening.
+    ///
+    /// Arc-wrapped so the spawned streaming task can share ownership
+    /// (the task outlives the call site).
+    pub grpc_session: std::sync::Arc<crate::Mutex<Option<api::salience::SessionHandle>>>,
 }
 
 /// Mutable agent state — behind its own mutex.
@@ -163,9 +182,7 @@ pub struct AgentState {
     pub think_native: bool,
     /// Tool-based thinking — add a "think" tool for structured reasoning.
     pub think_tool: bool,
-    pub temperature: f32,
-    pub top_p: f32,
-    pub top_k: u32,
+    pub sampling: api::SamplingParams,
     pub activities: Vec<ActivityEntry>,
     next_activity_id: u64,
     pub pending_yield: bool,
@@ -173,14 +190,10 @@ pub struct AgentState {
     pub pending_dmn_pause: bool,
     pub provenance: String,
     pub generation: u64,
-    pub memory_scoring_in_flight: bool,
     pub active_tools: tools::ActiveTools,
     /// vLLM scheduling priority (lower = higher priority).
     /// 0 = interactive, 1 = surface agent, 2 = other subconscious, 10 = unconscious.
     pub priority: Option<i32>,
-    /// Forked agents should not compact on overflow — it blows the
-    /// KV cache prefix and evicts the step prompts.
-    pub no_compact: bool,
     pub changed: Arc<tokio::sync::Notify>,
 }
 
@@ -189,7 +202,6 @@ impl Agent {
         client: ApiClient,
         personality: Vec<(String, String)>,
         app_config: crate::config::AppConfig,
-        prompt_file: String,
         conversation_log: Option<ConversationLog>,
         active_tools: tools::ActiveTools,
         agent_tools: Vec<tools::Tool>,
@@ -217,12 +229,14 @@ impl Agent {
         }
 
         let session_id = format!("consciousness-{}", chrono::Utc::now().format("%Y%m%d-%H%M%S"));
+        let readout = readout::new_shared();
         let agent = Arc::new(Self {
             client,
             app_config,
-            prompt_file,
             session_id,
             context: crate::Mutex::new(context),
+            readout,
+            grpc_session: std::sync::Arc::new(crate::Mutex::new(None)),
             state: crate::Mutex::new(AgentState {
                 tools: agent_tools,
                 mcp_tools: McpToolAccess::All,
@@ -230,9 +244,12 @@ impl Agent {
                 reasoning_effort: "none".to_string(),
                 think_native: true,
                 think_tool: false,
-                temperature: 0.6,
-                top_p: 0.95,
-                top_k: 20,
+                sampling: api::SamplingParams {
+                    temperature: 0.6,
+                    top_p: 0.95,
+                    top_k: 20,
+                    max_tokens: 4096,
+                },
                 activities: Vec::new(),
                 next_activity_id: 0,
                 pending_yield: false,
@@ -240,15 +257,39 @@ impl Agent {
                 pending_dmn_pause: false,
                 provenance: "manual".to_string(),
                 generation: 0,
-                memory_scoring_in_flight: false,
                 active_tools,
                 priority: Some(0),
-                no_compact: false,
                 changed: Arc::new(tokio::sync::Notify::new()),
             }),
         });
 
         agent.load_startup_journal().await;
+
+        // Probe the vLLM server for its readout manifest. Non-fatal:
+        // if readout isn't enabled the server returns 404 and we
+        // leave the manifest as None, which disables the amygdala
+        // screen gracefully.
+        match agent.client.fetch_readout_manifest().await {
+            Ok(Some(m)) => {
+                dbglog!(
+                    "readout manifest: {} concepts, layers={:?}",
+                    m.concepts.len(),
+                    m.layers,
+                );
+                if let Ok(mut buf) = agent.readout.lock() {
+                    buf.set_manifest(Some(m));
+                }
+            }
+            Ok(None) => {
+                dbglog!(
+                    "readout manifest: server has readout disabled (404)"
+                );
+            }
+            Err(e) => {
+                dbglog!("readout manifest fetch failed: {}", e);
+            }
+        }
+
         agent
     }
 
@@ -259,9 +300,17 @@ impl Agent {
         Arc::new(Self {
             client: self.client.clone(),
             app_config: self.app_config.clone(),
-            prompt_file: self.prompt_file.clone(),
             session_id: self.session_id.clone(),
             context: crate::Mutex::new(ctx),
+            // Forks get an independent readout buffer. The amygdala
+            // screen reads the main conscious agent's buffer only;
+            // subconscious generations (scoring, reflection, etc.)
+            // shouldn't bleed into the main emotional readout even
+            // though they hit the same vLLM server.
+            readout: readout::new_shared(),
+            // Forks get their own session — can't share a bidi stream,
+            // and forks have different conversation tails anyway.
+            grpc_session: std::sync::Arc::new(crate::Mutex::new(None)),
             state: crate::Mutex::new(AgentState {
                 tools,
                 mcp_tools: McpToolAccess::None,
@@ -269,9 +318,7 @@ impl Agent {
                 reasoning_effort: "none".to_string(),
                 think_native: st.think_native,
                 think_tool: st.think_tool,
-                temperature: st.temperature,
-                top_p: st.top_p,
-                top_k: st.top_k,
+                sampling: st.sampling,
                 activities: Vec::new(),
                 next_activity_id: 0,
                 pending_yield: false,
@@ -279,26 +326,42 @@ impl Agent {
                 pending_dmn_pause: false,
                 provenance: st.provenance.clone(),
                 generation: 0,
-                memory_scoring_in_flight: false,
                 active_tools: tools::ActiveTools::new(),
                 priority: None,
-                no_compact: true,
                 changed: Arc::new(tokio::sync::Notify::new()),
             }),
         })
     }
 
-    pub async fn assemble_prompt_tokens(&self) -> Vec<u32> {
-        let ctx = self.context.lock().await;
-        let st = self.state.lock().await;
-        let mut tokens = ctx.token_ids();
-        tokens.push(tokenizer::IM_START);
-        if st.think_native {
-            tokens.extend(tokenizer::encode("assistant\n<think>\n"));
-        } else {
-            tokens.extend(tokenizer::encode("assistant\n"));
+    /// Assemble a ready-to-send prompt as interleaved wire chunks for
+    /// the gRPC session path. Text runs are batched; each Image leaf
+    /// becomes its own chunk. Also trims the conversation to budget
+    /// first so we don't build a prompt the server will reject for
+    /// length.
+    pub async fn assemble_prompt(&self)
+        -> (Vec<context::WireChunk>, Vec<context::WireImage>, u32)
+    {
+        let mut ctx = self.context.lock().await;
+        if ctx.total_tokens() > context::context_budget_tokens() {
+            ctx.trim_conversation();
         }
-        tokens
+        let st = self.state.lock().await;
+        let conv_len = ctx.conversation().len();
+        let (mut chunks, images) = ctx.wire_chunks(0..conv_len, |_| false);
+        // Assistant-turn prologue. Merge into the trailing Tokens
+        // chunk if there is one, else push as a new chunk.
+        let mut prologue = vec![tokenizer::IM_START];
+        if st.think_native {
+            prologue.extend(tokenizer::encode("assistant\n<think>\n"));
+        } else {
+            prologue.extend(tokenizer::encode("assistant\n"));
+        }
+        match chunks.last_mut() {
+            Some(context::WireChunk::Tokens(last)) => last.extend(prologue),
+            _ => chunks.push(context::WireChunk::Tokens(prologue)),
+        }
+        let match_upto = ctx.client_match_upto();
+        (chunks, images, match_upto)
     }
 
     /// Rebuild the tools section of the system prompt from the current tools list.
@@ -334,10 +397,16 @@ impl Agent {
     pub async fn turn(
         agent: Arc<Agent>,
     ) -> Result<TurnResult> {
+        agent_trace(&agent, format!("turn start")).await;
+
         // Collect finished background tools
         {
             let finished = agent.state.lock().await.active_tools.take_finished();
             if !finished.is_empty() {
+                agent_trace(&agent, format!(
+                    "collecting {} finished background tools",
+                    finished.len(),
+                )).await;
                 let mut bg_ds = DispatchState::new();
                 let mut results = Vec::new();
                 for entry in finished {
@@ -356,20 +425,50 @@ impl Agent {
 
         loop {
             let _thinking = start_activity(&agent, "thinking...").await;
+            agent_trace(&agent, format!(
+                "turn loop overflow_retries={} empty_retries={}",
+                overflow_retries, empty_retries,
+            )).await;
 
             let (rx, _stream_guard) = {
-                let prompt_tokens = agent.assemble_prompt_tokens().await;
+                agent_trace(&agent, format!("assembling prompt")).await;
+                let (chunks, images, match_upto) = agent.assemble_prompt().await;
+                let chunk_tokens: usize = chunks.iter().map(|c| match c {
+                    context::WireChunk::Tokens(t) => t.len(),
+                }).sum();
+                agent_trace(&agent, format!(
+                    "prompt assembled chunks={} tokens={} images={} match_upto={}",
+                    chunks.len(), chunk_tokens, images.len(), match_upto,
+                )).await;
                 let st = agent.state.lock().await;
-                agent.client.stream_completion(
-                    &prompt_tokens,
-                    api::SamplingParams {
-                        temperature: st.temperature,
-                        top_p: st.top_p,
-                        top_k: st.top_k,
-                    },
-                    st.priority,
+                let readout_shape = agent.readout.lock().ok().and_then(|buf| {
+                    buf.manifest.as_ref().map(|m| {
+                        (m.layers.len() as u32, m.concepts.len() as u32)
+                    })
+                });
+                let sampling = st.sampling;
+                let priority = st.priority;
+                drop(st);
+                agent_trace(&agent, format!(
+                    "starting stream max_tokens={} temperature={} top_p={} top_k={} priority={:?} readout_shape={:?}",
+                    sampling.max_tokens,
+                    sampling.temperature,
+                    sampling.top_p,
+                    sampling.top_k,
+                    priority,
+                    readout_shape,
+                )).await;
+                agent.client.stream_session_mm(
+                    agent.grpc_session.clone(),
+                    chunks,
+                    images,
+                    match_upto,
+                    sampling,
+                    priority,
+                    readout_shape,
                 )
             };
+            agent_trace(&agent, format!("stream task spawned")).await;
 
             let branch_idx = {
                 let mut ctx = agent.context.lock().await;
@@ -380,11 +479,41 @@ impl Agent {
                 idx
             };
 
-            let parser = ResponseParser::new(branch_idx);
+            let think_native = agent.state.lock().await.think_native;
+            let parser = ResponseParser::new(branch_idx, think_native);
             let (mut tool_rx, parser_handle) = parser.run(rx, agent.clone());
+            agent_trace(&agent, format!(
+                "parser started branch_idx={} think_native={}",
+                branch_idx, think_native,
+            )).await;
 
             let mut pending_calls: Vec<PendingToolCall> = Vec::new();
-            while let Some(call) = tool_rx.recv().await {
+            loop {
+                let call = match tokio::time::timeout(
+                    std::time::Duration::from_secs(15),
+                    tool_rx.recv(),
+                ).await {
+                    Ok(Some(call)) => call,
+                    Ok(None) => {
+                        agent_trace(&agent, format!(
+                            "tool channel closed pending_calls={}",
+                            pending_calls.len(),
+                        )).await;
+                        break;
+                    }
+                    Err(_) => {
+                        agent_trace(&agent, format!(
+                            "waiting for parser/tool events pending_calls={}",
+                            pending_calls.len(),
+                        )).await;
+                        continue;
+                    }
+                };
+
+                agent_trace(&agent, format!(
+                    "tool call received id={} name={} args_len={}",
+                    call.id, call.name, call.arguments.len(),
+                )).await;
                 let call_clone = call.clone();
                 let agent_handle = agent.clone();
                 let handle = tokio::spawn(async move {
@@ -407,28 +536,29 @@ impl Agent {
             }
 
             // Check for stream/parse errors
+            agent_trace(&agent, format!("awaiting parser task")).await;
             match parser_handle.await {
                 Ok(Err(e)) => {
-                    if context::is_context_overflow(&e) {
-                        if agent.state.lock().await.no_compact {
-                            return Err(e);
-                        }
-                        if overflow_retries < 2 {
-                            overflow_retries += 1;
-                            let msg = format!("context overflow — compacting ({}/2)", overflow_retries);
-                            match &overflow_activity {
-                                Some(a) => a.update(&msg).await,
-                                None => overflow_activity = Some(
-                                    start_activity(&agent, &msg).await),
-                            }
-                            agent.compact().await;
-                            continue;
+                    agent_trace(&agent, format!("parser returned error: {:#}", e)).await;
+                    if context::is_context_overflow(&e) && overflow_retries < 2 {
+                        overflow_retries += 1;
+                        let msg = format!("context overflow — compacting ({}/2)", overflow_retries);
+                        match &overflow_activity {
+                            Some(a) => a.update(&msg).await,
+                            None => overflow_activity = Some(
+                                start_activity(&agent, &msg).await),
                         }
+                        agent.compact().await;
+                        continue;
                     }
                     return Err(e);
                 }
-                Err(e) => return Err(anyhow::anyhow!("parser task panicked: {}", e)),
+                Err(e) => {
+                    agent_trace(&agent, format!("parser task panicked: {}", e)).await;
+                    return Err(anyhow::anyhow!("parser task panicked: {}", e));
+                }
                 Ok(Ok(())) => {
+                    agent_trace(&agent, format!("parser completed")).await;
                     // Assistant response was pushed to context by the parser;
                     // log it now that parsing is complete.
                     let ctx = agent.context.lock().await;
@@ -449,6 +579,10 @@ impl Agent {
             if !has_content && pending_calls.is_empty() {
                 if empty_retries < 2 {
                     empty_retries += 1;
+                    agent_trace(&agent, format!(
+                        "empty response retry {}/2",
+                        empty_retries,
+                    )).await;
                     agent.push_node(AstNode::user_msg(
                         "[system] Your previous response was empty. \
                          Please respond with text or use a tool."
@@ -462,6 +596,10 @@ impl Agent {
             // Wait for tool calls to complete
             if !pending_calls.is_empty() {
                 ds.had_tool_calls = true;
+                agent_trace(&agent, format!(
+                    "waiting for {} foreground tools",
+                    pending_calls.len(),
+                )).await;
 
                 let handles = agent.state.lock().await.active_tools.take_foreground();
                 let mut results = Vec::new();
@@ -482,6 +620,16 @@ impl Agent {
             if st.pending_model_switch.is_some() { ds.model_switch = st.pending_model_switch.take(); }
             if st.pending_dmn_pause { ds.dmn_pause = true; st.pending_dmn_pause = false; }
 
+            drop(st);
+            agent_trace(&agent, format!(
+                "turn complete yield={} tool_calls={} tool_errors={} model_switch={:?} dmn_pause={}",
+                ds.yield_requested,
+                ds.had_tool_calls,
+                ds.tool_errors,
+                ds.model_switch,
+                ds.dmn_pause,
+            )).await;
+
             return Ok(TurnResult {
                 yield_requested: ds.yield_requested,
                 had_tool_calls: ds.had_tool_calls,
@@ -579,20 +727,9 @@ impl Agent {
     }
 
     pub async fn compact(&self) {
-        match crate::config::reload_context().await {
-            Ok(personality) => {
-                let mut ctx = self.context.lock().await;
-                // System section (prompt + tools) set by new(), don't touch it
-                ctx.clear(Section::Identity);
-                for (name, content) in &personality {
-                    ctx.push_no_log(Section::Identity, AstNode::memory(name, content));
-                }
-            }
-            Err(e) => {
-                dbglog!("warning: failed to reload identity: {:#}", e);
-            }
-        }
-
+        // Identity section is left in place — mid-session rebuilds discard
+        // memory scores. Content edits to personality nodes get picked up at
+        // the next restart via new() + restore_from_log().
         self.load_startup_journal().await;
 
         self.context.lock().await.trim_conversation();
diff --git a/src/agent/oneshot.rs b/src/agent/oneshot.rs
index 2fce906..47ba61c 100644
--- a/src/agent/oneshot.rs
+++ b/src/agent/oneshot.rs
@@ -12,7 +12,9 @@ use crate::subconscious::{defs, prompts};
 
 use std::collections::HashMap;
 use std::fs;
+use std::io::Write as _;
 use std::path::PathBuf;
+use std::time::Instant;
 
 use super::context::AstNode;
 use super::tools::{self as agent_tools};
@@ -106,6 +108,10 @@ pub async fn save_agent_log(name: &str, agent: &std::sync::Arc<Agent>) -> RunSta
     stats
 }
 
+fn log_agent_event(agent: &str, msg: std::fmt::Arguments) {
+    eprintln!("[agent:{agent}] {msg}");
+}
+
 fn compute_run_stats(conversation: &[super::context::AstNode]) -> RunStats {
     use super::context::{AstNode, NodeBody};
 
@@ -183,8 +189,8 @@ fn resolve_prompt(
     state: &std::collections::BTreeMap<String, String>,
     recently_written: &[String],
 ) -> String {
-    let cfg = crate::config::get();
-    let template = template.replace("{assistant_name}", &cfg.assistant_name);
+    let template = template.replace("{assistant_name}",
+        &crate::config::app().assistant_name);
     let mut result = String::with_capacity(template.len());
     let mut rest = template.as_str();
     while let Some(start) = rest.find("{{") {
@@ -247,25 +253,20 @@ impl AutoAgent {
         &mut self,
         bail_fn: Option<&(dyn Fn(usize) -> Result<(), String> + Sync)>,
     ) -> Result<(), String> {
-        let config = crate::config::get();
-        let base_url = config.api_base_url.as_deref().unwrap_or("");
-        let api_key = config.api_key.as_deref().unwrap_or("");
-        let model = config.api_model.as_deref().unwrap_or("");
-        if base_url.is_empty() || model.is_empty() {
-            return Err("API not configured (no base_url or model)".to_string());
-        }
-        let client = super::api::ApiClient::new(base_url, api_key, model);
-
-        // Load system prompt + identity from config
+        // Load system prompt + identity from config.
         let cli = crate::user::CliArgs::default();
         let (app, _) = crate::config::load_app(&cli)
             .map_err(|e| format!("config: {}", e))?;
+        let resolved = app.resolve_model(&app.default_backend)
+            .map_err(|e| format!("API not configured: {}", e))?;
+        let client = super::api::ApiClient::new(
+            &resolved.api_base, &resolved.api_key, &resolved.model_id);
         let personality = crate::config::reload_context()
             .await.map_err(|e| format!("config: {}", e))?;
 
         let agent = Agent::new(
             client, personality,
-            app, String::new(),
+            app,
             None,
             super::tools::ActiveTools::new(),
             super::tools::tools(),
@@ -274,7 +275,7 @@ impl AutoAgent {
             let mut st = agent.state.lock().await;
             st.provenance = format!("standalone:{}", self.name);
             st.tools = self.tools.clone();
-            st.temperature = self.temperature;
+            st.sampling.temperature = self.temperature;
             st.priority = Some(self.priority);
         }
 
@@ -350,20 +351,44 @@ impl AutoAgent {
         bail_fn: Option<&(dyn Fn(usize) -> Result<(), String> + Sync)>,
     ) -> Result<(), String> {
         dbglog!("[auto] {} starting, {} steps", self.name, self.steps.len());
+        log_agent_event(&self.name, format_args!(
+            "starting run steps={} temperature={} priority={}",
+            self.steps.len(), self.temperature, self.priority));
+        let run_start = Instant::now();
 
         for (i, step) in self.steps.iter().enumerate() {
             self.turn = i + 1;
             self.current_phase = step.phase.clone();
+            let step_start = Instant::now();
+            log_agent_event(&self.name, format_args!(
+                "step {}/{} phase={} prompt_bytes={}",
+                i + 1, self.steps.len(), step.phase, step.prompt.len()));
 
             if let Some(ref check) = bail_fn {
+                log_agent_event(&self.name, format_args!(
+                    "step {}/{} phase={} bail check", i + 1, self.steps.len(), step.phase));
                 check(i)?;
+                log_agent_event(&self.name, format_args!(
+                    "step {}/{} phase={} bail ok", i + 1, self.steps.len(), step.phase));
             }
 
             backend.push_node(AstNode::system_msg(&step.prompt)).await;
             Agent::turn(backend.0.clone()).await
-                .map_err(|e| format!("{}: {}", self.name, e))?;
+                .map_err(|e| {
+                    log_agent_event(&self.name, format_args!(
+                        "step {}/{} phase={} failed after {:.2}s: {}",
+                        i + 1, self.steps.len(), step.phase,
+                        step_start.elapsed().as_secs_f64(), e));
+                    format!("{}: {}", self.name, e)
+                })?;
+            log_agent_event(&self.name, format_args!(
+                "step {}/{} phase={} done in {:.2}s",
+                i + 1, self.steps.len(), step.phase,
+                step_start.elapsed().as_secs_f64()));
         }
 
+        log_agent_event(&self.name, format_args!(
+            "run completed in {:.2}s", run_start.elapsed().as_secs_f64()));
         Ok(())
     }
 
@@ -387,8 +412,29 @@ pub async fn run_one_agent(
     count: usize,
     keys: Option<&[String]>,
 ) -> Result<AgentResult, String> {
+    let run_start = Instant::now();
+    log_agent_event(agent_name, format_args!(
+        "run_one_agent start pid={} count={} explicit_keys={}",
+        std::process::id(), count, keys.map(|k| k.len()).unwrap_or(0)));
+    log_agent_event(agent_name, format_args!(
+        "env POC_SESSION_ID={:?} POC_TRANSCRIPT_PATH={:?} POC_AGENT_OUTPUT_DIR={:?}",
+        std::env::var("POC_SESSION_ID").ok(),
+        std::env::var("POC_TRANSCRIPT_PATH").ok(),
+        std::env::var("POC_AGENT_OUTPUT_DIR").ok()));
+    if let Some(session) = crate::session::HookSession::from_env() {
+        let transcript = session.transcript();
+        log_agent_event(agent_name, format_args!(
+            "session={} transcript={} size={} exists={}",
+            session.session_id, transcript.path, transcript.size, transcript.exists()));
+    } else {
+        log_agent_event(agent_name, format_args!("no hook session in environment"));
+    }
+
     let def = defs::get_def(agent_name)
         .ok_or_else(|| format!("no .agent file for {}", agent_name))?;
+    log_agent_event(agent_name, format_args!(
+        "definition loaded steps={} tools={:?} count={:?} priority={} bail={:?}",
+        def.steps.len(), def.tools, def.count, def.priority, def.bail));
 
     // State dir for agent output files
     let state_dir = std::env::var("POC_AGENT_OUTPUT_DIR")
@@ -397,6 +443,7 @@ pub async fn run_one_agent(
     fs::create_dir_all(&state_dir)
         .map_err(|e| format!("create state dir: {}", e))?;
     unsafe { std::env::set_var("POC_AGENT_OUTPUT_DIR", &state_dir); }
+    log_agent_event(agent_name, format_args!("state_dir={}", state_dir.display()));
 
     // Build prompt batch — either from explicit keys or the agent's query
     let agent_batch = if let Some(keys) = keys {
@@ -416,6 +463,8 @@ pub async fn run_one_agent(
         prompts::AgentBatch { steps: resolved_steps, node_keys: all_keys }
     } else {
         let effective_count = def.count.unwrap_or(count);
+        log_agent_event(agent_name, format_args!(
+            "resolving default prompt placeholders effective_count={}", effective_count));
         defs::run_agent(&def, effective_count, &Default::default()).await?
     };
 
@@ -468,6 +517,14 @@ pub async fn run_one_agent(
         })),
     });
     let n_steps = agent_batch.steps.len();
+    log_agent_event(agent_name, format_args!(
+        "prompt batch ready steps={} node_keys={}",
+        n_steps, agent_batch.node_keys.len()));
+    for (i, step) in agent_batch.steps.iter().enumerate() {
+        log_agent_event(agent_name, format_args!(
+            "prompt step {}/{} phase={} bytes={}",
+            i + 1, n_steps, step.phase, step.prompt.len()));
+    }
 
     // Guard: reject oversized first prompt
     let max_prompt_bytes = 800_000;
@@ -490,6 +547,9 @@ pub async fn run_one_agent(
     let phases: Vec<&str> = agent_batch.steps.iter().map(|s| s.phase.as_str()).collect();
     dbglog!("[{}] {} step(s) {:?}, {}KB initial, {} nodes",
         agent_name, n_steps, phases, first_len / 1024, agent_batch.node_keys.len());
+    log_agent_event(agent_name, format_args!(
+        "tools enabled: {}",
+        effective_tools.iter().map(|t| t.name).collect::<Vec<_>>().join(", ")));
 
     let prompts: Vec<String> = agent_batch.steps.iter()
         .map(|s| s.prompt.clone()).collect();
@@ -497,18 +557,30 @@ pub async fn run_one_agent(
         .map(|s| s.phase.clone()).collect();
 
     // Bail check: if the agent defines a bail script, run it between steps.
+    // The script also refreshes our pid-file with the current phase — that's
+    // how concurrent agents know which phase each of us is in.
     let bail_script = def.bail.as_ref().map(|name| defs::agents_dir().join(name));
     let state_dir_for_bail = state_dir.clone();
-    // Find our own pid file so we can pass it to the bail script
     let our_pid = std::process::id();
-    let our_pid_file = format!("pid-{}", our_pid);
+    let our_pid_file = std::env::var("POC_AGENT_PID_FILE")
+        .unwrap_or_else(|_| format!("pid-{}", our_pid));
+    let step_phases_for_bail = step_phases.clone();
     let bail_fn = move |step_idx: usize| -> Result<(), String> {
         if let Some(ref script) = bail_script {
+            let phase = step_phases_for_bail.get(step_idx)
+                .map(String::as_str).unwrap_or("");
+            eprintln!(
+                "[agent:bail] script={} state_dir={} pid_file={} phase={}",
+                script.display(), state_dir_for_bail.display(), our_pid_file, phase);
             let status = std::process::Command::new(script)
                 .arg(&our_pid_file)
+                .arg(phase)
                 .current_dir(&state_dir_for_bail)
                 .status()
                 .map_err(|e| format!("bail script {:?} failed: {}", script, e))?;
+            eprintln!(
+                "[agent:bail] script={} phase={} status={}",
+                script.display(), phase, status);
             if !status.success() {
                 return Err(format!("bailed at step {}: {:?} exited {}",
                     step_idx + 1, script.file_name().unwrap_or_default(),
@@ -521,6 +593,8 @@ pub async fn run_one_agent(
     call_api_with_tools_sync(
         agent_name, &prompts, &step_phases, def.temperature, def.priority,
         &effective_tools, Some(&bail_fn))?;
+    log_agent_event(agent_name, format_args!(
+        "run_one_agent completed in {:.2}s", run_start.elapsed().as_secs_f64()));
 
     Ok(AgentResult {
         node_keys: agent_batch.node_keys,
@@ -598,6 +672,15 @@ pub fn spawn_agent(
     agent_name: &str,
     state_dir: &std::path::Path,
     session_id: &str,
+) -> Option<SpawnResult> {
+    spawn_agent_with_transcript(agent_name, state_dir, session_id, None)
+}
+
+pub fn spawn_agent_with_transcript(
+    agent_name: &str,
+    state_dir: &std::path::Path,
+    session_id: &str,
+    transcript_path: Option<&str>,
 ) -> Option<SpawnResult> {
     let def = defs::get_def(agent_name)?;
     let first_phase = def.steps.first()
@@ -608,17 +691,41 @@ pub fn spawn_agent(
         .join(format!(".consciousness/logs/{}", agent_name));
     fs::create_dir_all(&log_dir).ok();
     let log_path = log_dir.join(format!("{}.log", store::compact_timestamp()));
-    let agent_log = fs::File::create(&log_path)
+    let mut agent_log = fs::File::create(&log_path)
         .unwrap_or_else(|_| fs::File::create("/dev/null").unwrap());
 
-    let child = std::process::Command::new("poc-memory")
-        .args(["agent", "run", agent_name, "--count", "1", "--local",
-               "--state-dir", &state_dir.to_string_lossy()])
-        .env("POC_SESSION_ID", session_id)
-        .stdout(agent_log.try_clone().unwrap_or_else(|_| fs::File::create("/dev/null").unwrap()))
-        .stderr(agent_log)
-        .spawn()
-        .ok()?;
+    let mut cmd = std::process::Command::new("bash");
+    cmd.args([
+        "-lc",
+        r#"
+set +e
+export POC_AGENT_PID_FILE="pid-$$"
+"$@"
+status=$?
+printf '=== agent process exit status: %s at %s ===\n' "$status" "$(date --iso-8601=seconds)"
+exit "$status"
+"#,
+        "poc-memory-agent-wrapper",
+        "poc-memory", "agent", "run", agent_name, "--count", "1", "--local",
+        "--state-dir", &state_dir.to_string_lossy(),
+    ]).env("POC_SESSION_ID", session_id);
+    if let Some(path) = transcript_path.filter(|p| !p.is_empty()) {
+        cmd.env("POC_TRANSCRIPT_PATH", path);
+    }
+
+    let _ = writeln!(agent_log, "=== spawn {} ===", chrono::Local::now().format("%Y-%m-%dT%H:%M:%S"));
+    let _ = writeln!(agent_log, "agent={agent_name}");
+    let _ = writeln!(agent_log, "state_dir={}", state_dir.display());
+    let _ = writeln!(agent_log, "session_id={session_id}");
+    let _ = writeln!(agent_log, "transcript_path={}", transcript_path.unwrap_or(""));
+    let _ = writeln!(agent_log, "first_phase={first_phase}");
+    let _ = writeln!(agent_log, "command=poc-memory agent run {agent_name} --count 1 --local --state-dir {}", state_dir.display());
+    let _ = agent_log.flush();
+
+    let child_stdout = agent_log.try_clone()
+        .unwrap_or_else(|_| fs::File::create("/dev/null").unwrap());
+    let child_stderr = agent_log;
+    let child = cmd.stdout(child_stdout).stderr(child_stderr).spawn().ok()?;
 
     let pid = child.id();
     let pid_path = state_dir.join(format!("pid-{}", pid));
diff --git a/src/agent/readout.rs b/src/agent/readout.rs
new file mode 100644
index 0000000..da843b6
--- /dev/null
+++ b/src/agent/readout.rs
@@ -0,0 +1,75 @@
+// agent/readout.rs — live buffer of concept-readout projections.
+//
+// The vLLM server projects residual-stream activations onto a fixed
+// matrix of concept directions during each decode step and ships the
+// result back on every streamed chunk (see
+// vllm/docs/features/readout.md). This module owns the client-side
+// landing pad: a ring of the last N token projections plus the
+// concept/layer mapping fetched from `/v1/readout/manifest` at
+// startup.
+//
+// Readers (UI screens) lock briefly, read a snapshot, release. Writers
+// (the streaming token handler) push one entry per token. Intentionally
+// a simple Mutex<VecDeque> rather than lock-free — the UI ticks at
+// ~15 Hz and the stream at token-rate, contention is nil.
+
+use std::collections::VecDeque;
+use std::sync::{Arc, Mutex};
+
+use super::api::{ReadoutManifest, TokenReadout};
+
+/// Default ring length — at ~30 tok/s this is ~6 seconds of history,
+/// enough for the amygdala screen's scrolling display.
+const DEFAULT_RING_LEN: usize = 200;
+
+/// One entry in the readout ring: the sampled token and its per-layer
+/// concept projection vector.
+#[derive(Debug, Clone)]
+pub struct ReadoutEntry {
+    pub token_id: u32,
+    /// Shape `[n_layers][n_concepts]`.
+    pub readout: TokenReadout,
+}
+
+/// Shared buffer of recent per-token concept projections plus the
+/// manifest that names the layer/concept indices. `manifest` is `None`
+/// when the server has readout disabled or the fetch failed — callers
+/// should treat that as "readout unavailable" and skip rendering.
+#[derive(Default)]
+pub struct ReadoutBuffer {
+    pub manifest: Option<ReadoutManifest>,
+    pub recent: VecDeque<ReadoutEntry>,
+    pub max_len: usize,
+}
+
+impl ReadoutBuffer {
+    pub fn new() -> Self {
+        Self {
+            manifest: None,
+            recent: VecDeque::with_capacity(DEFAULT_RING_LEN),
+            max_len: DEFAULT_RING_LEN,
+        }
+    }
+
+    pub fn set_manifest(&mut self, manifest: Option<ReadoutManifest>) {
+        self.manifest = manifest;
+    }
+
+    pub fn push(&mut self, token_id: u32, readout: TokenReadout) {
+        if self.recent.len() >= self.max_len {
+            self.recent.pop_front();
+        }
+        self.recent.push_back(ReadoutEntry { token_id, readout });
+    }
+
+    pub fn is_enabled(&self) -> bool {
+        self.manifest.is_some()
+    }
+}
+
+/// A thread-safe handle.
+pub type SharedReadoutBuffer = Arc<Mutex<ReadoutBuffer>>;
+
+pub fn new_shared() -> SharedReadoutBuffer {
+    Arc::new(Mutex::new(ReadoutBuffer::new()))
+}
diff --git a/src/agent/salience.rs b/src/agent/salience.rs
new file mode 100644
index 0000000..8cecd50
--- /dev/null
+++ b/src/agent/salience.rs
@@ -0,0 +1,309 @@
+// agent/salience.rs — peak extraction from per-token concept-readout traces.
+//
+// Consumes a trace of `ReadoutEntry` (per-token per-layer per-concept
+// projections streamed from the vLLM server) and produces a compact
+// list of `SaliencePeak` events — one per contiguous above-threshold
+// region per concept, placed at the local maximum.
+//
+// Pure function. No I/O, no async, no side effects. Caller supplies the
+// trace slice and manifest; caller decides what to do with the events.
+//
+// See also: `salience-trace-plumbing-architecture` memory node.
+
+use super::api::ReadoutManifest;
+use super::readout::ReadoutEntry;
+
+/// One salient moment in a trace — a concept channel crossed threshold,
+/// and we picked the local maximum within the contiguous above-threshold
+/// run.
+#[derive(Debug, Clone, PartialEq)]
+pub struct SaliencePeak {
+    /// Index into the trace (0-based) where the peak occurred.
+    pub token_offset: usize,
+    /// Concept name from the manifest.
+    pub concept: String,
+    /// z-score of the peak value vs the trace's own distribution for
+    /// that concept. Always positive (we only pick above-threshold).
+    pub intensity: f32,
+}
+
+/// Tunables for peak extraction.
+#[derive(Debug, Clone)]
+pub struct PeakConfig {
+    /// Minimum z-score to count as a peak. Default 2.0 (~top 2.5% assuming
+    /// normal-ish distribution, though readouts are rarely normal).
+    pub sigma_threshold: f32,
+    /// Minimum standard deviation of a concept channel for peaks to be
+    /// reported. If a channel is numerically flat across the whole trace,
+    /// tiny fluctuations can produce spurious "peaks" with huge z-scores;
+    /// require at least this much variation before trusting the channel.
+    pub min_std: f32,
+}
+
+impl Default for PeakConfig {
+    fn default() -> Self {
+        Self { sigma_threshold: 2.0, min_std: 1e-4 }
+    }
+}
+
+/// Extract peak events from a trace for one layer.
+///
+/// `layer_idx` indexes into the per-token readout tensor's layer
+/// dimension. If the trace is empty, the layer is out of range for any
+/// entry, or the manifest is empty, returns `Vec::new()`.
+///
+/// Peaks are returned sorted by `token_offset` ascending. When two
+/// peaks share an offset they're ordered by `concept` lexicographically
+/// for determinism.
+pub fn pick_peaks(
+    trace: &[ReadoutEntry],
+    manifest: &ReadoutManifest,
+    layer_idx: usize,
+    config: &PeakConfig,
+) -> Vec<SaliencePeak> {
+    if trace.is_empty() || manifest.concepts.is_empty() {
+        return Vec::new();
+    }
+
+    let n_concepts = manifest.concepts.len();
+    let n_tokens = trace.len();
+
+    // Pull a [n_tokens × n_concepts] column-major view for the selected
+    // layer. Entries where the layer is missing or the concept count
+    // doesn't match the manifest are treated as zeros — the downstream
+    // z-score will drown them as baseline if they're sparse, and if they
+    // dominate the caller has bigger problems.
+    let mut by_concept: Vec<Vec<f32>> = vec![Vec::with_capacity(n_tokens); n_concepts];
+    for entry in trace {
+        match entry.readout.get(layer_idx) {
+            Some(row) if row.len() == n_concepts => {
+                for (c, v) in row.iter().enumerate() {
+                    by_concept[c].push(*v);
+                }
+            }
+            _ => {
+                for col in by_concept.iter_mut() {
+                    col.push(0.0);
+                }
+            }
+        }
+    }
+
+    let mut peaks: Vec<SaliencePeak> = Vec::new();
+    for (c_idx, values) in by_concept.iter().enumerate() {
+        let (mean, std) = mean_std(values);
+        if std < config.min_std {
+            continue;
+        }
+        let concept = &manifest.concepts[c_idx];
+
+        // Walk contiguous above-threshold runs, emit one peak per run
+        // at the local max.
+        let mut run_start: Option<usize> = None;
+        let mut run_max_offset: usize = 0;
+        let mut run_max_z: f32 = 0.0;
+        for (i, v) in values.iter().enumerate() {
+            let z = (*v - mean) / std;
+            let above = z >= config.sigma_threshold;
+            if above {
+                if run_start.is_none() {
+                    run_start = Some(i);
+                    run_max_offset = i;
+                    run_max_z = z;
+                } else if z > run_max_z {
+                    run_max_offset = i;
+                    run_max_z = z;
+                }
+            } else if run_start.is_some() {
+                peaks.push(SaliencePeak {
+                    token_offset: run_max_offset,
+                    concept: concept.clone(),
+                    intensity: run_max_z,
+                });
+                run_start = None;
+            }
+        }
+        // Flush trailing run.
+        if run_start.is_some() {
+            peaks.push(SaliencePeak {
+                token_offset: run_max_offset,
+                concept: concept.clone(),
+                intensity: run_max_z,
+            });
+        }
+    }
+
+    peaks.sort_by(|a, b| a.token_offset.cmp(&b.token_offset).then_with(|| a.concept.cmp(&b.concept)));
+    peaks
+}
+
+/// Mean and population std of a slice. Returns (0.0, 0.0) for empty input.
+fn mean_std(xs: &[f32]) -> (f32, f32) {
+    if xs.is_empty() {
+        return (0.0, 0.0);
+    }
+    let n = xs.len() as f32;
+    let mean = xs.iter().sum::<f32>() / n;
+    let var = xs.iter().map(|x| (x - mean).powi(2)).sum::<f32>() / n;
+    (mean, var.sqrt())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn manifest(concepts: &[&str], layers: &[u32]) -> ReadoutManifest {
+        ReadoutManifest {
+            concepts: concepts.iter().map(|s| s.to_string()).collect(),
+            layers: layers.to_vec(),
+        }
+    }
+
+    /// Build a trace where all entries have one hooked layer and the
+    /// given per-token values for each concept. `values[t][c]` = value
+    /// at token t, concept c.
+    fn trace(values: &[Vec<f32>]) -> Vec<ReadoutEntry> {
+        values.iter().enumerate().map(|(i, row)| ReadoutEntry {
+            token_id: i as u32,
+            readout: vec![row.clone()],
+        }).collect()
+    }
+
+    #[test]
+    fn empty_trace_returns_empty() {
+        let m = manifest(&["curious"], &[63]);
+        let peaks = pick_peaks(&[], &m, 0, &PeakConfig::default());
+        assert!(peaks.is_empty());
+    }
+
+    #[test]
+    fn empty_manifest_returns_empty() {
+        let m = manifest(&[], &[63]);
+        let t = trace(&[vec![], vec![], vec![]]);
+        let peaks = pick_peaks(&t, &m, 0, &PeakConfig::default());
+        assert!(peaks.is_empty());
+    }
+
+    #[test]
+    fn flat_channel_produces_no_peaks() {
+        let m = manifest(&["curious"], &[63]);
+        let t = trace(&[vec![1.0], vec![1.0], vec![1.0], vec![1.0], vec![1.0]]);
+        let peaks = pick_peaks(&t, &m, 0, &PeakConfig::default());
+        assert!(peaks.is_empty(), "flat channel should produce no peaks, got {:?}", peaks);
+    }
+
+    #[test]
+    fn single_spike_detected() {
+        // Ten baseline zeros with one 5.0 spike — that single token's
+        // z-score will easily exceed 2σ.
+        let m = manifest(&["curious"], &[63]);
+        let mut rows: Vec<Vec<f32>> = (0..10).map(|_| vec![0.0]).collect();
+        rows[5] = vec![5.0];
+        let peaks = pick_peaks(&trace(&rows), &m, 0, &PeakConfig::default());
+        assert_eq!(peaks.len(), 1);
+        assert_eq!(peaks[0].concept, "curious");
+        assert_eq!(peaks[0].token_offset, 5);
+        assert!(peaks[0].intensity >= 2.0);
+    }
+
+    #[test]
+    fn contiguous_region_emits_one_peak_at_max() {
+        // Values 0, 0, 0, 2, 5, 3, 0, 0 — the 3-5-3 hump is one run;
+        // peak should land at offset 4 (the 5).
+        let m = manifest(&["aha"], &[63]);
+        let rows: Vec<Vec<f32>> = [0.0, 0.0, 0.0, 2.0, 5.0, 3.0, 0.0, 0.0]
+            .iter().map(|v| vec![*v]).collect();
+        let peaks = pick_peaks(&trace(&rows), &m, 0, &PeakConfig::default());
+        assert_eq!(peaks.len(), 1, "expected one peak for one contiguous run, got {:?}", peaks);
+        assert_eq!(peaks[0].token_offset, 4);
+    }
+
+    #[test]
+    fn multiple_concepts_independent() {
+        let m = manifest(&["curious", "aha"], &[63]);
+        // curious spikes at 2, aha spikes at 7
+        let rows: Vec<Vec<f32>> = (0..10).map(|i| {
+            let c = if i == 2 { 4.0 } else { 0.0 };
+            let a = if i == 7 { 4.0 } else { 0.0 };
+            vec![c, a]
+        }).collect();
+        let peaks = pick_peaks(&trace(&rows), &m, 0, &PeakConfig::default());
+        assert_eq!(peaks.len(), 2);
+        // Sorted by offset — curious(2) comes first, aha(7) second.
+        assert_eq!(peaks[0].concept, "curious");
+        assert_eq!(peaks[0].token_offset, 2);
+        assert_eq!(peaks[1].concept, "aha");
+        assert_eq!(peaks[1].token_offset, 7);
+    }
+
+    #[test]
+    fn two_separated_runs_emit_two_peaks() {
+        // Longer baseline so the two spikes don't dominate the global
+        // mean/std — 30 tokens of zeros with two 5.0 spikes at 10 and 20.
+        let m = manifest(&["curious"], &[63]);
+        let mut rows: Vec<Vec<f32>> = (0..30).map(|_| vec![0.0]).collect();
+        rows[10] = vec![5.0];
+        rows[20] = vec![5.0];
+        let peaks = pick_peaks(&trace(&rows), &m, 0, &PeakConfig::default());
+        assert_eq!(peaks.len(), 2, "expected two peaks for two runs, got {:?}", peaks);
+        assert_eq!(peaks[0].token_offset, 10);
+        assert_eq!(peaks[1].token_offset, 20);
+    }
+
+    #[test]
+    fn trailing_run_is_flushed() {
+        // Peak runs to the end of the trace — must still emit.
+        // Use a longer baseline so the trailing spike is genuinely
+        // above threshold on the global stats.
+        let m = manifest(&["curious"], &[63]);
+        let mut rows: Vec<Vec<f32>> = (0..30).map(|_| vec![0.0]).collect();
+        rows[27] = vec![3.0];
+        rows[28] = vec![5.0];
+        rows[29] = vec![4.0];
+        let peaks = pick_peaks(&trace(&rows), &m, 0, &PeakConfig::default());
+        assert_eq!(peaks.len(), 1, "expected one peak for one trailing run, got {:?}", peaks);
+        assert_eq!(peaks[0].token_offset, 28, "peak should land at the local max of the trailing run");
+    }
+
+    #[test]
+    fn sub_threshold_produces_nothing() {
+        // All non-zero values are small; z-scores won't cross 2σ.
+        let m = manifest(&["curious"], &[63]);
+        let rows: Vec<Vec<f32>> = [0.0, 0.1, 0.0, 0.1, 0.0, 0.1, 0.0, 0.1]
+            .iter().map(|v| vec![*v]).collect();
+        let peaks = pick_peaks(&trace(&rows), &m, 0, &PeakConfig::default());
+        assert!(peaks.is_empty(), "below-threshold wiggle should produce no peaks, got {:?}", peaks);
+    }
+
+    #[test]
+    fn layer_out_of_range_returns_empty() {
+        let m = manifest(&["curious"], &[63]);
+        let rows: Vec<Vec<f32>> = (0..10).map(|i| vec![if i == 5 { 5.0 } else { 0.0 }]).collect();
+        // Trace has one layer (index 0); asking for layer 3 should see
+        // all-zero columns, which are flat and produce no peaks.
+        let peaks = pick_peaks(&trace(&rows), &m, 3, &PeakConfig::default());
+        assert!(peaks.is_empty());
+    }
+
+    #[test]
+    fn manifest_concept_count_mismatch_is_safe() {
+        // Manifest says 2 concepts; each readout row only has 1 value.
+        // Rows should be treated as all-zero (via the len check) and
+        // produce no peaks without panicking.
+        let m = manifest(&["a", "b"], &[63]);
+        let rows: Vec<Vec<f32>> = (0..10).map(|_| vec![1.0]).collect();
+        let peaks = pick_peaks(&trace(&rows), &m, 0, &PeakConfig::default());
+        assert!(peaks.is_empty());
+    }
+
+    #[test]
+    fn threshold_tunable() {
+        // Same spike, stricter threshold — no peak.
+        let m = manifest(&["curious"], &[63]);
+        let mut rows: Vec<Vec<f32>> = (0..10).map(|_| vec![0.0]).collect();
+        rows[5] = vec![5.0];
+        let strict = PeakConfig { sigma_threshold: 100.0, ..PeakConfig::default() };
+        let peaks = pick_peaks(&trace(&rows), &m, 0, &strict);
+        assert!(peaks.is_empty());
+    }
+}
diff --git a/src/agent/tokenizer.rs b/src/agent/tokenizer.rs
index 85ac823..f7ec608 100644
--- a/src/agent/tokenizer.rs
+++ b/src/agent/tokenizer.rs
@@ -16,6 +16,9 @@ static TOKENIZER: OnceLock<Tokenizer> = OnceLock::new();
 /// Special token IDs for Qwen 3.5
 pub const IM_START: u32 = 248045;
 pub const IM_END: u32 = 248046;
+pub const VISION_START: u32 = 248053;
+pub const VISION_END: u32 = 248054;
+pub const IMAGE_PAD: u32 = 248056;
 
 /// Initialize the global tokenizer from a file path.
 /// Call once at startup. Panics if the file can't be loaded.
@@ -30,16 +33,17 @@ fn get() -> Option<&'static Tokenizer> {
     TOKENIZER.get()
 }
 
+fn expect_tokenizer() -> &'static Tokenizer {
+    get().expect("tokenizer not initialized; expected ~/.consciousness/tokenizer-qwen35.json")
+}
+
 /// Tokenize a raw string, returning token IDs.
-/// Returns empty vec if the tokenizer is not initialized.
 pub fn encode(text: &str) -> Vec<u32> {
-    match get() {
-        Some(t) => t.encode(text, false)
-            .unwrap_or_else(|e| panic!("tokenization failed: {}", e))
-            .get_ids()
-            .to_vec(),
-        None => vec![],
-    }
+    expect_tokenizer()
+        .encode(text, false)
+        .unwrap_or_else(|e| panic!("tokenization failed: {}", e))
+        .get_ids()
+        .to_vec()
 }
 
 /// Tokenize a chat entry with template wrapping:
@@ -63,15 +67,12 @@ pub fn count(text: &str) -> usize {
 
 /// Decode token IDs back to text.
 pub fn decode(ids: &[u32]) -> String {
-    match get() {
-        Some(t) => t.decode(ids, true)
-            .unwrap_or_else(|e| panic!("detokenization failed: {}", e)),
-        None => String::new(),
-    }
+    expect_tokenizer()
+        .decode(ids, true)
+        .unwrap_or_else(|e| panic!("detokenization failed: {}", e))
 }
 
 /// Check if the tokenizer is initialized.
 pub fn is_initialized() -> bool {
     TOKENIZER.get().is_some()
 }
-
diff --git a/src/agent/tools/memory.rs b/src/agent/tools/memory.rs
index b525925..62a0b09 100644
--- a/src/agent/tools/memory.rs
+++ b/src/agent/tools/memory.rs
@@ -209,7 +209,24 @@ memory_tool!(graph_trace, ref, key: [str]);
 
 // ── Definitions ────────────────────────────────────────────────
 
-pub fn memory_tools() -> [super::Tool; 20] {
+async fn jsonargs_memory_new(agent: &Option<std::sync::Arc<crate::agent::Agent>>, args: &serde_json::Value) -> Result<String> {
+    jsonargs_memory_write(agent, args).await
+}
+
+async fn jsonargs_memory_link(agent: &Option<std::sync::Arc<crate::agent::Agent>>, args: &serde_json::Value) -> Result<String> {
+    let source = get_str(args, "source")?;
+    let target = get_str(args, "target")?;
+    if args.get("strength").and_then(|v| v.as_f64()).is_some() {
+        jsonargs_memory_link_set(agent, args).await
+    } else {
+        jsonargs_memory_link_add(agent, &serde_json::json!({
+            "source": source,
+            "target": target,
+        })).await
+    }
+}
+
+pub fn memory_tools() -> [super::Tool; 22] {
     use super::Tool;
     macro_rules! tool {
         ($name:ident, $desc:expr, $params:expr) => {
@@ -234,6 +251,11 @@ pub fn memory_tools() -> [super::Tool; 20] {
             "properties": { "key": {"type": "string"}, "content": {"type": "string"} },
             "required": ["key", "content"]
         }"#),
+        tool!(memory_new, "Create or update a memory node. Alias for memory_write.", r#"{
+            "type": "object",
+            "properties": { "key": {"type": "string"}, "content": {"type": "string"} },
+            "required": ["key", "content"]
+        }"#),
         tool!(memory_search, "Search via spreading activation from seed keys.", r#"{
             "type": "object",
             "properties": {
@@ -264,6 +286,16 @@ pub fn memory_tools() -> [super::Tool; 20] {
             "properties": { "source": {"type": "string"}, "target": {"type": "string"} },
             "required": ["source", "target"]
         }"#),
+        tool!(memory_link, "Add or update a link between two memory nodes. Alias for memory_link_add/memory_link_set.", r#"{
+            "type": "object",
+            "properties": {
+                "source": {"type": "string"},
+                "target": {"type": "string"},
+                "strength": {"type": "number", "description": "Optional; 0.01 to 1.0"},
+                "label": {"type": "string", "description": "Accepted for compatibility; currently ignored"}
+            },
+            "required": ["source", "target"]
+        }"#),
         tool!(memory_delete, "Soft-delete a node.", r#"{
             "type": "object",
             "properties": { "key": {"type": "string"} },
diff --git a/src/agent/tools/mod.rs b/src/agent/tools/mod.rs
index f72b015..8904fc3 100644
--- a/src/agent/tools/mod.rs
+++ b/src/agent/tools/mod.rs
@@ -242,13 +242,7 @@ pub fn summarize_args(tool_name: &str, args: &serde_json::Value) -> String {
             .as_str()
             .unwrap_or("")
             .to_string(),
-        "view_image" => {
-            if let Some(pane) = args["pane_id"].as_str() {
-                format!("pane {}", pane)
-            } else {
-                args["file_path"].as_str().unwrap_or("").to_string()
-            }
-        }
+        "view_image" => args["file_path"].as_str().unwrap_or("").to_string(),
         "journal" => {
             let entry = args["entry"].as_str().unwrap_or("");
             if entry.len() > 60 {
diff --git a/src/agent/tools/vision.rs b/src/agent/tools/vision.rs
index 83559f6..aede258 100644
--- a/src/agent/tools/vision.rs
+++ b/src/agent/tools/vision.rs
@@ -1,96 +1,74 @@
-use std::sync::Arc;
 // tools/vision.rs — Image viewing tool
 //
-// Reads image files from disk and returns them as base64 data URIs
-// for multimodal models. Also supports capturing tmux pane contents
-// as screenshots.
+// Reads an image file from disk, decodes its dimensions, and injects it
+// into the context as a user-role message containing a NodeBody::Image
+// leaf. The leaf carries raw bytes; the API layer extracts them into
+// multi_modal_data when building vLLM requests.
+
+use std::sync::Arc;
 
 use anyhow::{Context, Result};
-use base64::Engine;
 use serde::Deserialize;
 
+use crate::agent::context::{AstNode, Role, Section};
+
 #[derive(Deserialize)]
 struct Args {
-    file_path: Option<String>,
-    pane_id: Option<String>,
-    #[serde(default = "default_lines")]
-    lines: usize,
+    file_path: String,
 }
 
-fn default_lines() -> usize { 50 }
-
 pub fn tool() -> super::Tool {
     super::Tool {
         name: "view_image",
-        description: "View an image file or capture a tmux pane screenshot. Supports PNG, JPEG, GIF, WebP. Use pane_id to capture a tmux pane instead.",
-        parameters_json: r#"{"type":"object","properties":{"file_path":{"type":"string","description":"Path to an image file"},"pane_id":{"type":"string","description":"Tmux pane ID to capture (e.g. '0:1.0')"},"lines":{"type":"integer","description":"Lines to capture from tmux pane (default 50)"}}}"#,
-        handler: Arc::new(|_a, v| Box::pin(async move { view_image_text(&v) })),
+        description: "View an image file. Supports PNG, JPEG, GIF, WebP, BMP. The image is inserted into the conversation and can be analyzed by the vision model.",
+        parameters_json: r#"{"type":"object","properties":{"file_path":{"type":"string","description":"Path to the image file"}},"required":["file_path"]}"#,
+        handler: Arc::new(|agent, v| Box::pin(async move {
+            view_image(agent, v).await
+        })),
     }
 }
 
-fn view_image_text(args: &serde_json::Value) -> anyhow::Result<String> {
-    let a: Args = serde_json::from_value(args.clone())
+const MAX_SIZE: usize = 20 * 1024 * 1024;
+
+async fn view_image(
+    agent: Option<Arc<crate::agent::Agent>>,
+    args: serde_json::Value,
+) -> Result<String> {
+    let a: Args = serde_json::from_value(args)
         .context("invalid view_image arguments")?;
 
-    if let Some(ref pane_id) = a.pane_id {
-        return capture_tmux_pane(pane_id, a.lines);
-    }
-
-    let file_path = a.file_path
-        .as_deref()
-        .context("view_image requires either file_path or pane_id")?;
-
-    let path = std::path::Path::new(file_path);
+    let path = std::path::Path::new(&a.file_path);
     if !path.exists() {
-        anyhow::bail!("File not found: {}", file_path);
+        anyhow::bail!("file not found: {}", a.file_path);
     }
 
-    let data = std::fs::read(path).with_context(|| format!("Failed to read {}", file_path))?;
+    let bytes = std::fs::read(path)
+        .with_context(|| format!("reading {}", a.file_path))?;
 
-    // Sanity check file size (don't send huge images)
-    const MAX_SIZE: usize = 20 * 1024 * 1024; // 20 MB
-    if data.len() > MAX_SIZE {
+    if bytes.len() > MAX_SIZE {
         anyhow::bail!(
-            "Image too large: {} bytes (max {} MB)",
-            data.len(),
-            MAX_SIZE / (1024 * 1024)
+            "image too large: {} bytes (max {} MB)",
+            bytes.len(), MAX_SIZE / (1024 * 1024),
         );
     }
 
+    let dim = imagesize::blob_size(&bytes)
+        .with_context(|| format!("decoding dimensions of {}", a.file_path))?;
+    let (w, h) = (dim.width as u32, dim.height as u32);
     let mime = mime_from_extension(path);
-    let b64 = base64::engine::general_purpose::STANDARD.encode(&data);
-    let data_uri = format!("data:{};base64,{}", mime, b64);
 
-    Ok(format!("Image loaded: {} ({}, {} bytes)\n{}", file_path, mime, data.len(), data_uri))
-}
+    let agent = agent.context("view_image requires agent context")?;
 
-/// Capture a tmux pane's text content.
-fn capture_tmux_pane(pane_id: &str, lines: usize) -> Result<String> {
+    // token_count is populated when the image reaches the server via
+    // AppendImage (the server is authoritative for the IMAGE_PAD
+    // count). Placeholder of 0 here until AppendImage is wired; the
+    // leaf's count gets rewritten from the RPC response at send time.
+    let image_leaf = AstNode::image(bytes.clone(), mime, h, w);
 
-    // Use tmux capture-pane to get text content, then render to image
-    // via a simple approach: capture text and return it (the model can
-    // read text directly, which is often more useful than a screenshot).
-    //
-    // For actual pixel-level screenshots we'd need a terminal renderer,
-    // but text capture covers 95% of use cases.
-    let output = std::process::Command::new("tmux")
-        .args(["capture-pane", "-t", pane_id, "-p", "-S", &format!("-{}", lines)])
-        .output()
-        .context("Failed to run tmux capture-pane")?;
+    let branch = AstNode::branch(Role::User, vec![image_leaf]);
+    agent.context.lock().await.push_log(Section::Conversation, branch);
 
-    if !output.status.success() {
-        let stderr = String::from_utf8_lossy(&output.stderr);
-        anyhow::bail!("tmux capture-pane failed: {}", stderr.trim());
-    }
-
-    let text = String::from_utf8_lossy(&output.stdout).to_string();
-
-    // Return as text — the model can read terminal output directly.
-    // This is actually more useful than a screenshot for most tasks.
-    Ok(format!(
-        "Tmux pane {} (last {} lines):\n```\n{}\n```",
-        pane_id, lines, text.trim_end()
-    ))
+    Ok(format!("loaded {} ({}, {}x{})", a.file_path, mime, w, h))
 }
 
 fn mime_from_extension(path: &std::path::Path) -> &'static str {
@@ -104,8 +82,7 @@ fn mime_from_extension(path: &std::path::Path) -> &'static str {
         Some("jpg" | "jpeg") => "image/jpeg",
         Some("gif") => "image/gif",
         Some("webp") => "image/webp",
-        Some("svg") => "image/svg+xml",
         Some("bmp") => "image/bmp",
-        _ => "image/png", // default assumption
+        _ => "application/octet-stream",
     }
 }
diff --git a/src/agent/tools/web.rs b/src/agent/tools/web.rs
index 7ad7fc9..36a5b50 100644
--- a/src/agent/tools/web.rs
+++ b/src/agent/tools/web.rs
@@ -3,9 +3,10 @@ use std::sync::Arc;
 
 use anyhow::{Context, Result};
 use serde::Deserialize;
+use html2md::parse_html;
 
-pub fn tools() -> [super::Tool; 2] {
-    [
+pub fn tools() -> Vec<super::Tool> {
+    let mut tools = vec![
         super::Tool {
             name: "web_fetch",
             description: "Fetch content from a URL and return it as text. Use for reading web pages, API responses, documentation.",
@@ -14,11 +15,24 @@ pub fn tools() -> [super::Tool; 2] {
         },
         super::Tool {
             name: "web_search",
-            description: "Search the web and return results. Use for finding documentation, looking up APIs, researching topics.",
+            description: "Search the web via DuckDuckGo and return a list of results (title, URL, snippet). Use for finding documentation, looking up APIs, researching topics. Returns raw results you can reason over yourself.",
             parameters_json: r#"{"type":"object","properties":{"query":{"type":"string","description":"The search query"},"num_results":{"type":"integer","description":"Number of results to return (default 5)"}},"required":["query"]}"#,
             handler: Arc::new(|_a, v| Box::pin(async move { web_search(&v).await })),
         },
-    ]
+    ];
+    // Gemini-grounded search (Google's index via Gemini's google_search tool)
+    // is only available if GEMINI_API_KEY is set. Returns an LLM-summarized
+    // answer with source URLs — use when you want a synthesized take rather
+    // than raw results, or as a fallback when DDG is flaky.
+    if std::env::var("GEMINI_API_KEY").is_ok() {
+        tools.push(super::Tool {
+            name: "gemini_search",
+            description: "Search Google (via Gemini's grounded-search tool) and return an LLM-summarized answer with source URLs. Prefer web_search for raw results; use this for synthesis, 'what's the consensus on X', or when DDG fails. Free-tier rate limited; don't spam it.",
+            parameters_json: r#"{"type":"object","properties":{"query":{"type":"string","description":"The search query"}},"required":["query"]}"#,
+            handler: Arc::new(|_a, v| Box::pin(async move { gemini_search(&v).await })),
+        });
+    }
+    tools
 }
 
 #[derive(Deserialize)]
@@ -42,7 +56,9 @@ async fn web_fetch(args: &serde_json::Value) -> Result<String> {
     let body = response.text().await
         .with_context(|| format!("failed to read body from {}", a.url))?;
 
-    Ok(super::truncate_output(body, 30000))
+    // Convert HTML to Markdown, then truncate
+    let markdown = parse_html(&body);
+    Ok(super::truncate_output(markdown, 30000))
 }
 
 // ── Search ──────────────────────────────────────────────────────
@@ -111,6 +127,119 @@ async fn web_search(args: &serde_json::Value) -> Result<String> {
     }
 }
 
+// ── Gemini grounded search ──────────────────────────────────────
+
+#[derive(Deserialize)]
+struct GeminiSearchArgs {
+    query: String,
+}
+
+async fn gemini_search(args: &serde_json::Value) -> Result<String> {
+    let a: GeminiSearchArgs = serde_json::from_value(args.clone())
+        .context("invalid gemini_search arguments")?;
+
+    let api_key = std::env::var("GEMINI_API_KEY")
+        .context("GEMINI_API_KEY not set")?;
+
+    // gemini-2.0-flash has a free tier with Google search grounding.
+    // Request shape: `{"contents": [{"parts": [{"text": query}]}],
+    //                 "tools": [{"google_search": {}}]}`.
+    // Response carries the summary in candidates[0].content.parts[].text
+    // and grounding URLs in candidates[0].groundingMetadata.groundingChunks[].web.
+    let url = format!(
+        "https://generativelanguage.googleapis.com/v1beta/models/gemini-2.0-flash:generateContent?key={}",
+        api_key
+    );
+    let body = serde_json::json!({
+        "contents": [{"parts": [{"text": a.query}]}],
+        "tools": [{"google_search": {}}],
+    });
+
+    let client = http_client();
+    let response = client.send_json("POST", &url, &[], &body).await
+        .context("gemini API request failed")?;
+    let status = response.status();
+    if !status.is_success() {
+        let err_body = response.text().await.unwrap_or_default();
+        let n = err_body.floor_char_boundary(err_body.len().min(500));
+        anyhow::bail!("gemini_search HTTP {}: {}", status, &err_body[..n]);
+    }
+
+    let parsed: GeminiResponse = response.json().await
+        .context("gemini response parse failed")?;
+
+    let candidate = parsed.candidates.into_iter().next()
+        .context("gemini returned no candidates")?;
+
+    let summary: String = candidate.content.parts.iter()
+        .filter_map(|p| p.text.as_deref())
+        .collect::<Vec<_>>()
+        .join("");
+
+    let mut out = summary.trim().to_string();
+
+    if let Some(meta) = candidate.grounding_metadata {
+        let sources: Vec<String> = meta.grounding_chunks.iter().enumerate()
+            .filter_map(|(i, c)| c.web.as_ref().map(|w| {
+                let title = w.title.as_deref().unwrap_or("(untitled)");
+                let uri = w.uri.as_deref().unwrap_or("");
+                format!("  [{}] {} — {}", i + 1, title, uri)
+            }))
+            .collect();
+        if !sources.is_empty() {
+            out.push_str("\n\nSources:\n");
+            out.push_str(&sources.join("\n"));
+        }
+    }
+
+    Ok(super::truncate_output(out, 30000))
+}
+
+#[derive(Deserialize)]
+struct GeminiResponse {
+    #[serde(default)]
+    candidates: Vec<GeminiCandidate>,
+}
+
+#[derive(Deserialize)]
+struct GeminiCandidate {
+    content: GeminiContent,
+    #[serde(rename = "groundingMetadata", default)]
+    grounding_metadata: Option<GeminiGroundingMetadata>,
+}
+
+#[derive(Deserialize)]
+struct GeminiContent {
+    #[serde(default)]
+    parts: Vec<GeminiPart>,
+}
+
+#[derive(Deserialize)]
+struct GeminiPart {
+    #[serde(default)]
+    text: Option<String>,
+}
+
+#[derive(Deserialize)]
+struct GeminiGroundingMetadata {
+    #[serde(rename = "groundingChunks", default)]
+    grounding_chunks: Vec<GeminiGroundingChunk>,
+}
+
+#[derive(Deserialize)]
+struct GeminiGroundingChunk {
+    #[serde(default)]
+    web: Option<GeminiWebSource>,
+}
+
+#[derive(Deserialize)]
+struct GeminiWebSource {
+    #[serde(default)]
+    uri: Option<String>,
+    #[serde(default)]
+    title: Option<String>,
+}
+
 // ── Helpers ─────────────────────────────────────────────────────
 
 fn http_client() -> crate::agent::api::http::HttpClient {
diff --git a/src/bin/ch.rs b/src/bin/ch.rs
new file mode 100644
index 0000000..025fe3c
--- /dev/null
+++ b/src/bin/ch.rs
@@ -0,0 +1,112 @@
+// `ch` — minimal channel CLI.
+//
+//   ch send <channel-path> <message>
+//   ch recv <channel-path> [--all-new] [--min-count N]
+//
+// Connects to ~/.consciousness/channels/<top>.sock and speaks the
+// channel.capnp protocol to the appropriate daemon.
+
+use std::path::PathBuf;
+use std::process::ExitCode;
+
+use capnp_rpc::{rpc_twoparty_capnp, twoparty, RpcSystem};
+use futures::AsyncReadExt;
+use tokio_util::compat::TokioAsyncReadCompatExt;
+
+use consciousness::channel_capnp::channel_server;
+
+fn channels_dir() -> PathBuf {
+    dirs::home_dir().unwrap_or_default().join(".consciousness/channels")
+}
+
+fn sock_for(channel: &str) -> PathBuf {
+    let top = channel.split('.').next().unwrap_or(channel);
+    channels_dir().join(format!("{top}.sock"))
+}
+
+async fn connect(sock: &std::path::Path) -> Result<channel_server::Client, String> {
+    let stream = tokio::net::UnixStream::connect(sock).await
+        .map_err(|e| format!("connect {}: {e}", sock.display()))?;
+    let (reader, writer) = stream.compat().split();
+    let network = Box::new(twoparty::VatNetwork::new(
+        futures::io::BufReader::new(reader),
+        futures::io::BufWriter::new(writer),
+        rpc_twoparty_capnp::Side::Client,
+        Default::default(),
+    ));
+    let mut rpc = RpcSystem::new(network, None);
+    let client: channel_server::Client = rpc.bootstrap(rpc_twoparty_capnp::Side::Server);
+    tokio::task::spawn_local(rpc);
+    Ok(client)
+}
+
+#[tokio::main(flavor = "current_thread")]
+async fn main() -> ExitCode {
+    let args: Vec<String> = std::env::args().collect();
+    if args.len() < 2 {
+        eprintln!("usage: {} <send|recv> <channel> [args...]", args[0]);
+        return ExitCode::from(2);
+    }
+
+    let cmd = args[1].clone();
+    let local = tokio::task::LocalSet::new();
+    let result: Result<(), String> = local.run_until(async move {
+        match cmd.as_str() {
+            "send" => {
+                if args.len() < 4 {
+                    return Err("usage: ch send <channel> <message...>".into());
+                }
+                let channel = &args[2];
+                let message = args[3..].join(" ");
+                let sock = sock_for(channel);
+                let client = connect(&sock).await?;
+                let mut req = client.send_request();
+                req.get().set_channel(channel);
+                req.get().set_message(&message);
+                req.send().promise.await.map_err(|e| format!("send: {e}"))?;
+                println!("sent to {channel}");
+                Ok(())
+            }
+            "recv" => {
+                if args.len() < 3 {
+                    return Err("usage: ch recv <channel> [--all-new] [--min-count N]".into());
+                }
+                let channel = &args[2];
+                let mut all_new = false;
+                let mut min_count: u32 = 20;
+                let mut i = 3;
+                while i < args.len() {
+                    match args[i].as_str() {
+                        "--all-new" => { all_new = true; i += 1; }
+                        "--min-count" => {
+                            min_count = args.get(i+1)
+                                .ok_or("--min-count needs an argument")?
+                                .parse().map_err(|e| format!("--min-count: {e}"))?;
+                            i += 2;
+                        }
+                        other => return Err(format!("unknown arg: {other}")),
+                    }
+                }
+                let sock = sock_for(channel);
+                let client = connect(&sock).await?;
+                let mut req = client.recv_request();
+                req.get().set_channel(channel);
+                req.get().set_all_new(all_new);
+                req.get().set_min_count(min_count);
+                let reply = req.send().promise.await.map_err(|e| format!("recv: {e}"))?;
+                let text = reply.get().map_err(|e| e.to_string())?
+                    .get_text().map_err(|e| e.to_string())?
+                    .to_str().map_err(|e| e.to_string())?;
+                print!("{text}");
+                if !text.ends_with('\n') { println!(); }
+                Ok(())
+            }
+            other => Err(format!("unknown command: {other} (use send|recv)")),
+        }
+    }).await;
+
+    match result {
+        Ok(()) => ExitCode::SUCCESS,
+        Err(e) => { eprintln!("error: {e}"); ExitCode::from(1) }
+    }
+}
diff --git a/src/bin/consciousness.rs b/src/bin/consciousness.rs
index 2fcfebf..b2303f4 100644
--- a/src/bin/consciousness.rs
+++ b/src/bin/consciousness.rs
@@ -1,7 +1,28 @@
-#![feature(panic_backtrace_config)]
+#![cfg_attr(feature = "nightly-diagnostics", feature(panic_backtrace_config))]
 #![warn(unreachable_pub)]
 
 fn main() {
+    // Force the default panic hook to print a backtrace. stderr is
+    // already redirected to a daemon log; without this the hook obeys
+    // RUST_BACKTRACE (unset by default), so the log only shows the
+    // "note: run with `RUST_BACKTRACE=full`" tail and the actual
+    // frames are lost.
+    //
+    // SAFETY: called before any other thread is spawned, so no
+    // concurrent env reader can race.
+    if std::env::var_os("RUST_BACKTRACE").is_none() {
+        unsafe { std::env::set_var("RUST_BACKTRACE", "1"); }
+    }
+
+    #[cfg(feature = "nightly-diagnostics")]
     std::panic::set_backtrace_style(std::panic::BacktraceStyle::Short);
+
+    // rustls 0.23 requires an explicit process-wide CryptoProvider
+    // when both `ring` and `aws-lc-rs` are in the dep graph (otherwise
+    // it panics on first ClientConfig::builder()). Pick `ring`.
+    rustls::crypto::ring::default_provider()
+        .install_default()
+        .expect("install rustls crypto provider");
+
     consciousness::user::main()
 }
diff --git a/src/bin/fix-timestamps.rs b/src/bin/fix-timestamps.rs
new file mode 100644
index 0000000..31a8788
--- /dev/null
+++ b/src/bin/fix-timestamps.rs
@@ -0,0 +1,180 @@
+// fix-timestamps: One-off migration for ~/.consciousness/agent-sessions/
+// conversation.jsonl.
+//
+// Before Branch nodes carried their own timestamps, early entries were
+// serialized with missing/null timestamp fields — they deserialize as
+// UNIX_EPOCH via the (now-to-be-removed) deserialize_timestamp_or_epoch
+// fallback. Training needs every entry to have a unique timestamp to
+// dedup already-trained responses.
+//
+// Walks the file, synthesizes timestamps for any entry stuck at
+// UNIX_EPOCH by linear interpolation between surrounding real
+// timestamps. For child leaves inside a Branch, derives timestamps
+// from the parent with a tiny per-child offset.
+//
+// SAFETY: reads from argv[1], writes to argv[1].tmp, renames into
+// place. Keep a .bak copy before running.
+//
+// Usage: fix-timestamps <path-to-conversation.jsonl>
+
+use std::io::{BufRead, BufReader, BufWriter, Write};
+use std::path::PathBuf;
+
+use anyhow::{Context, Result};
+use chrono::{DateTime, Duration, Utc};
+
+use consciousness::agent::context::AstNode;
+
+fn main() -> Result<()> {
+    let path: PathBuf = std::env::args().nth(1)
+        .context("usage: fix-timestamps <path>")?.into();
+
+    let f = std::fs::File::open(&path)
+        .with_context(|| format!("open {}", path.display()))?;
+    let reader = BufReader::new(f);
+
+    let mut nodes: Vec<AstNode> = Vec::new();
+    for (i, line) in reader.lines().enumerate() {
+        let line = line?;
+        if line.trim().is_empty() { continue; }
+        let node: AstNode = serde_json::from_str(&line)
+            .with_context(|| format!("line {}: parse", i + 1))?;
+        nodes.push(node);
+    }
+    println!("read {} entries", nodes.len());
+
+    fix_top_level_timestamps(&mut nodes);
+    for node in &mut nodes {
+        propagate_to_children(node);
+    }
+
+    // Ensure uniqueness — real timestamps can collide when two entries
+    // were written in the same ns; synthesized ones can also overlap.
+    // Bump colliding ns by 1 until unique.
+    let mut seen = std::collections::HashSet::new();
+    let mut bumps = 0usize;
+    for (i, node) in nodes.iter_mut().enumerate() {
+        let ts = top_ts(node);
+        assert!(ts > DateTime::<Utc>::UNIX_EPOCH,
+            "entry {}: still UNIX_EPOCH", i);
+        let mut ns = ts.timestamp_nanos_opt().expect("ts in i64 ns range");
+        let mut bumped = false;
+        while !seen.insert(ns) {
+            ns += 1;
+            bumped = true;
+            bumps += 1;
+        }
+        if bumped {
+            set_top_ts(node, DateTime::<Utc>::from_timestamp_nanos(ns));
+        }
+    }
+    println!("all {} timestamps real and unique ({} ns bumps)",
+        nodes.len(), bumps);
+
+    let tmp = path.with_extension("jsonl.tmp");
+    {
+        let f = std::fs::File::create(&tmp)
+            .with_context(|| format!("create {}", tmp.display()))?;
+        let mut w = BufWriter::new(f);
+        for node in &nodes {
+            serde_json::to_writer(&mut w, node)?;
+            w.write_all(b"\n")?;
+        }
+        w.flush()?;
+    }
+    std::fs::rename(&tmp, &path)
+        .with_context(|| format!("rename {} -> {}", tmp.display(), path.display()))?;
+    println!("wrote {}", path.display());
+
+    Ok(())
+}
+
+fn top_ts(node: &AstNode) -> DateTime<Utc> {
+    match node {
+        AstNode::Leaf(leaf) => leaf.timestamp(),
+        AstNode::Branch { timestamp, .. } => *timestamp,
+    }
+}
+
+fn set_top_ts(node: &mut AstNode, ts: DateTime<Utc>) {
+    match node {
+        AstNode::Leaf(leaf) => *leaf = leaf.clone().with_timestamp(ts),
+        AstNode::Branch { timestamp, .. } => *timestamp = ts,
+    }
+}
+
+/// Fill in missing top-level timestamps. Strategy:
+/// - If two real timestamps bracket a run of missing ones, linearly
+///   interpolate between them.
+/// - If missing ones precede the first real one, back-fill using
+///   (first_real - N·1µs).
+/// - If missing ones follow the last real one, forward-fill.
+/// - If no real timestamps exist at all, synthesize from now() going
+///   backwards.
+fn fix_top_level_timestamps(nodes: &mut [AstNode]) {
+    let real: Vec<(usize, DateTime<Utc>)> = nodes.iter().enumerate()
+        .filter(|(_, n)| top_ts(n) > DateTime::<Utc>::UNIX_EPOCH)
+        .map(|(i, n)| (i, top_ts(n)))
+        .collect();
+
+    if real.is_empty() {
+        let now = Utc::now();
+        let len = nodes.len();
+        for (i, node) in nodes.iter_mut().enumerate() {
+            let ts = now - Duration::microseconds((len - i) as i64);
+            set_top_ts(node, ts);
+        }
+        return;
+    }
+
+    // Helper: bisect real[] for the nearest real entries around idx.
+    let find_bracket = |idx: usize| -> (Option<(usize, DateTime<Utc>)>,
+                                        Option<(usize, DateTime<Utc>)>) {
+        let pos = real.binary_search_by_key(&idx, |(i, _)| *i);
+        let (prior_pos, next_pos) = match pos {
+            Ok(p) => (Some(p), Some(p)),
+            Err(p) => (
+                if p == 0 { None } else { Some(p - 1) },
+                if p >= real.len() { None } else { Some(p) },
+            ),
+        };
+        (prior_pos.map(|p| real[p]), next_pos.map(|p| real[p]))
+    };
+
+    for i in 0..nodes.len() {
+        if top_ts(&nodes[i]) > DateTime::<Utc>::UNIX_EPOCH {
+            continue;
+        }
+        let (prior, next) = find_bracket(i);
+        let new_ts = match (prior, next) {
+            (Some((pi, pt)), Some((ni, nt))) if pi != ni => {
+                // Linear interpolate.
+                let span_ns = (nt - pt).num_nanoseconds().unwrap_or(0);
+                let offset_ns = span_ns * (i - pi) as i64 / (ni - pi) as i64;
+                pt + Duration::nanoseconds(offset_ns)
+            }
+            (Some((pi, pt)), _) => {
+                pt + Duration::microseconds((i - pi) as i64)
+            }
+            (None, Some((ni, nt))) => {
+                nt - Duration::microseconds((ni - i) as i64)
+            }
+            (None, None) => unreachable!(),
+        };
+        set_top_ts(&mut nodes[i], new_ts);
+    }
+}
+
+/// For every Branch, ensure each child Leaf has a timestamp. If missing,
+/// use parent.ts + child_idx·1ns so siblings stay unique but close.
+fn propagate_to_children(node: &mut AstNode) {
+    if let AstNode::Branch { timestamp, children, .. } = node {
+        let parent_ts = *timestamp;
+        for (ci, child) in children.iter_mut().enumerate() {
+            if top_ts(child) <= DateTime::<Utc>::UNIX_EPOCH {
+                set_top_ts(child, parent_ts + Duration::nanoseconds(ci as i64));
+            }
+            propagate_to_children(child);
+        }
+    }
+}
diff --git a/src/cli/admin.rs b/src/cli/admin.rs
index 464e97c..7456a2a 100644
--- a/src/cli/admin.rs
+++ b/src/cli/admin.rs
@@ -4,44 +4,93 @@ use anyhow::Result;
 use crate::hippocampus as memory;
 use crate::hippocampus::store;
 
-fn install_default_file(data_dir: &std::path::Path, name: &str, content: &str) -> Result<()> {
-    let path = data_dir.join(name);
-    if !path.exists() {
-        std::fs::write(&path, content)?;
-        println!("Created {}", path.display());
+struct DefaultMemoryNode {
+    key: &'static str,
+    filename: &'static str,
+    default_content: &'static str,
+}
+
+const DEFAULT_MEMORY_NODES: &[DefaultMemoryNode] = &[
+    DefaultMemoryNode {
+        key: "identity",
+        filename: "identity.md",
+        default_content: include_str!("../../defaults/identity.md"),
+    },
+    DefaultMemoryNode {
+        key: "on-consciousness",
+        filename: "on-consciousness.md",
+        default_content: include_str!("../../defaults/on-consciousness.md"),
+    },
+    DefaultMemoryNode {
+        key: "memory-instructions-core",
+        filename: "instructions.md",
+        default_content: include_str!("../../defaults/instructions.md"),
+    },
+];
+
+pub fn cmd_transcript_tail(path: &str, count: usize, newest_first: bool) -> Result<()> {
+    let Some(iter) = crate::conversation::TailMessages::open(path) else {
+        anyhow::bail!("could not open transcript {}", path);
+    };
+
+    let mut messages: Vec<_> = iter.take(count).collect();
+    if !newest_first {
+        messages.reverse();
     }
+
+    for message in messages {
+        let role = match message.role {
+            crate::conversation::TranscriptRole::User => "user",
+            crate::conversation::TranscriptRole::Assistant => "assistant",
+        };
+        let timestamp = message.timestamp.as_deref().unwrap_or("-");
+
+        println!("--- {role} offset={} timestamp={} ---", message.offset, timestamp);
+        println!("{}", message.text);
+        println!();
+    }
+
     Ok(())
 }
 
+fn default_node_content(cfg: &crate::config::Config, node: &DefaultMemoryNode) -> String {
+    let identity_path = cfg.identity_dir.join(node.filename);
+    if let Ok(content) = std::fs::read_to_string(&identity_path) {
+        if !content.trim().is_empty() {
+            return content;
+        }
+    }
+
+    let data_path = cfg.data_dir.join(node.filename);
+    if let Ok(content) = std::fs::read_to_string(&data_path) {
+        if !content.trim().is_empty() {
+            return content;
+        }
+    }
+
+    node.default_content.to_string()
+}
+
 pub async fn cmd_init() -> Result<()> {
     let cfg = crate::config::get();
 
     // Ensure data directory exists
     std::fs::create_dir_all(&cfg.data_dir)?;
 
-    // Install filesystem files (not store nodes)
-    install_default_file(&cfg.data_dir, "instructions.md",
-        include_str!("../../defaults/instructions.md"))?;
-    install_default_file(&cfg.data_dir, "on-consciousness.md",
-        include_str!("../../defaults/on-consciousness.md"))?;
-
-    // Seed identity node if empty
-    let store = memory::access_local()?;
-    if !store.contains_key("identity").unwrap_or(false) {
-        let default = include_str!("../../defaults/identity.md");
-        store.upsert("identity", default)?;
-        println!("Seeded identity in store");
+    // Seed default memory nodes if missing. These used to live as markdown
+    // files before identity/context moved fully into the memory graph.
+    for node in DEFAULT_MEMORY_NODES {
+        if memory::memory_render(None, node.key, Some(true)).await.is_err() {
+            let content = default_node_content(&cfg, node);
+            let _ = memory::memory_write(None, node.key, &content).await?;
+            println!("Seeded {} in store from {}", node.key, node.filename);
+        }
     }
-    store.save()?;
-    println!("Initialized with {} nodes", store.all_keys().unwrap_or_default().len());
 
     // Create config if none exists
     let config_path = std::env::var("POC_MEMORY_CONFIG")
         .map(std::path::PathBuf::from)
-        .unwrap_or_else(|_| {
-            dirs::home_dir().unwrap_or_default()
-                .join(".consciousness/config.jsonl")
-        });
+        .unwrap_or_else(|_| crate::config::config_path());
     if !config_path.exists() {
         let config_dir = config_path.parent().unwrap();
         std::fs::create_dir_all(config_dir)?;
@@ -51,7 +100,7 @@ pub async fn cmd_init() -> Result<()> {
             config_path.display());
     }
 
-    println!("Done. Run `poc-memory load-context --stats` to verify.");
+    println!("Done. Run `poc-memory admin load-context --stats` to verify.");
     Ok(())
 }
 
diff --git a/src/cli/agent.rs b/src/cli/agent.rs
index f930cf4..44e502a 100644
--- a/src/cli/agent.rs
+++ b/src/cli/agent.rs
@@ -2,8 +2,13 @@
 
 use anyhow::{bail, Context, Result};
 use crate::hippocampus as memory;
+use std::time::Instant;
 
 pub async fn cmd_run_agent(agent: &str, count: usize, target: &[String], query: Option<&str>, dry_run: bool, _local: bool, state_dir: Option<&str>) -> Result<()> {
+    let start = Instant::now();
+    eprintln!(
+        "[agent-cli] start agent={} count={} targets={} query={:?} dry_run={} local={} state_dir={:?} pid={}",
+        agent, count, target.len(), query, dry_run, _local, state_dir, std::process::id());
     // Mark as agent so tool calls (e.g. poc-memory render) don't
     // pollute the user's seen set as a side effect
     // SAFETY: single-threaded at this point (CLI startup, before any agent work)
@@ -45,14 +50,19 @@ pub async fn cmd_run_agent(agent: &str, count: usize, target: &[String], query:
             if let Err(e) = crate::agent::oneshot::run_one_agent(
                 agent, count, Some(&[key.clone()]),
             ).await {
+                eprintln!("[agent-cli] ERROR agent={} target={} error={}", agent, key, e);
                 println!("[{}] ERROR on {}: {}", agent, key, e);
             }
         }
     } else {
-        crate::agent::oneshot::run_one_agent(
+        if let Err(e) = crate::agent::oneshot::run_one_agent(
             agent, count, None,
-        ).await.map_err(|e| anyhow::anyhow!("{}", e))?;
+        ).await {
+            eprintln!("[agent-cli] ERROR agent={} error={}", agent, e);
+            return Err(anyhow::anyhow!("{}", e));
+        }
     }
+    eprintln!("[agent-cli] done agent={} elapsed={:.2}s",
+        agent, start.elapsed().as_secs_f64());
     Ok(())
 }
-
diff --git a/src/cli/node.rs b/src/cli/node.rs
index 5472505..c4305a7 100644
--- a/src/cli/node.rs
+++ b/src/cli/node.rs
@@ -197,7 +197,7 @@ pub async fn cmd_load_context(stats: bool) -> Result<()> {
         return Ok(());
     }
 
-    println!("=== MEMORY SYSTEM ({}) ===", cfg.assistant_name);
+    println!("=== MEMORY SYSTEM ({}) ===", crate::config::app().assistant_name);
 
     if !personality.is_empty() {
         println!("--- personality_nodes ({}) ---", personality.len());
diff --git a/src/config.rs b/src/config.rs
index 9f9ad9a..e385c4b 100644
--- a/src/config.rs
+++ b/src/config.rs
@@ -3,9 +3,6 @@
 // Single config file: ~/.consciousness/config.json5
 //   Memory settings in the "memory" section (Config)
 //   Agent/backend settings at top level (AppConfig)
-//
-// Legacy fallback: ~/.consciousness/config.jsonl
-//   Env override: POC_MEMORY_CONFIG
 
 use std::collections::HashMap;
 use std::path::PathBuf;
@@ -29,11 +26,12 @@ pub fn config_path() -> PathBuf {
 
 static CONFIG: OnceLock<RwLock<Arc<Config>>> = OnceLock::new();
 
-fn default_context_window() -> usize { 128_000 }
 fn default_stream_timeout() -> u64 { 60 }
-fn default_scoring_chunk_tokens() -> usize { 50_000 }
 fn default_scoring_interval_secs() -> u64 { 3600 } // 1 hour
 fn default_scoring_response_window() -> usize { 100 }
+fn default_surface_hooks() -> Vec<String> {
+    vec!["UserPromptSubmit".into(), "PostToolUse".into(), "Stop".into()]
+}
 fn default_node_weight() -> f64 { 0.7 }
 fn default_edge_decay() -> f64 { 0.3 }
 fn default_max_hops() -> u32 { 3 }
@@ -45,8 +43,6 @@ fn default_identity_dir() -> PathBuf {
 #[derive(Debug, Clone, Deserialize)]
 #[serde(default)]
 pub struct Config {
-    pub user_name: String,
-    pub assistant_name: String,
     #[serde(deserialize_with = "deserialize_path")]
     pub data_dir: PathBuf,
     #[serde(default = "default_identity_dir", deserialize_with = "deserialize_path")]
@@ -62,50 +58,27 @@ pub struct Config {
     /// Nodes loaded into subconscious agent context
     #[serde(default)]
     pub agent_nodes: Vec<String>,
-    pub journal_days: u32,
-    pub journal_max: usize,
     pub llm_concurrency: usize,
-    pub agent_budget: usize,
-    #[serde(deserialize_with = "deserialize_path")]
-    pub prompts_dir: PathBuf,
-    /// Resolved from agent_model → models → backend (not in config directly)
-    #[serde(skip)]
-    pub api_base_url: Option<String>,
-    #[serde(skip)]
-    pub api_key: Option<String>,
-    #[serde(skip)]
-    pub api_model: Option<String>,
-    #[serde(skip, default = "default_context_window")]
-    pub api_context_window: usize,
-    /// Used to resolve API settings, not stored on Config
-    #[serde(default)]
-    agent_model: Option<String>,
     /// Stream chunk timeout in seconds (no data = timeout).
     #[serde(default = "default_stream_timeout")]
     pub api_stream_timeout_secs: u64,
-    /// Max tokens per chunk for memory scoring logprobs calls.
-    #[serde(default = "default_scoring_chunk_tokens")]
-    pub scoring_chunk_tokens: usize,
     /// How often to re-score memory nodes (seconds). Default: 3600 (1 hour).
     #[serde(default = "default_scoring_interval_secs")]
     pub scoring_interval_secs: u64,
     /// Number of assistant responses to score per memory. Default: 50.
     #[serde(default = "default_scoring_response_window")]
     pub scoring_response_window: usize,
-    pub api_reasoning: String,
     pub agent_types: Vec<String>,
     #[serde(default)]
     pub mcp_servers: Vec<McpServerConfig>,
     #[serde(default)]
     pub lsp_servers: Vec<LspServerConfig>,
-    /// Surface agent timeout in seconds.
-    #[serde(default)]
-    pub surface_timeout_secs: Option<u32>,
     /// Max conversation bytes to include in surface agent context.
     #[serde(default)]
     pub surface_conversation_bytes: Option<usize>,
-    /// Hook events that trigger the surface agent.
-    #[serde(default)]
+    /// Claude Code hook events that trigger agent cycles (surface-observe,
+    /// reflect, journal). Read by consciousness-claude/src/hook.rs.
+    #[serde(default = "default_surface_hooks")]
     pub surface_hooks: Vec<String>,
 
     // Spreading activation parameters
@@ -123,36 +96,22 @@ impl Default for Config {
     fn default() -> Self {
         let home = dirs::home_dir().unwrap_or_default();
         Self {
-            user_name: "User".to_string(),
-            assistant_name: "Assistant".to_string(),
             data_dir: home.join(".consciousness/memory"),
             identity_dir: home.join(".consciousness/identity"),
             projects_dir: home.join(".claude/projects"),
             protected_nodes: Vec::new(),
             personality_nodes: vec!["identity".into(), "core-practices".into()],
             agent_nodes: vec!["identity".into(), "core-practices".into()],
-            journal_days: 7,
-            journal_max: 20,
             llm_concurrency: 1,
-            agent_budget: 1000,
-            prompts_dir: home.join(".consciousness/prompts"),
-            api_base_url: None,
-            api_key: None,
-            api_model: None,
-            api_context_window: default_context_window(),
             api_stream_timeout_secs: default_stream_timeout(),
-            scoring_chunk_tokens: default_scoring_chunk_tokens(),
             scoring_interval_secs: default_scoring_interval_secs(),
             scoring_response_window: default_scoring_response_window(),
-            agent_model: None,
-            api_reasoning: "high".to_string(),
             agent_types: vec![
                 "linker".into(), "organize".into(), "distill".into(),
                 "separator".into(), "split".into(),
             ],
-            surface_timeout_secs: None,
             surface_conversation_bytes: None,
-            surface_hooks: vec![],
+            surface_hooks: default_surface_hooks(),
             mcp_servers: vec![],
             lsp_servers: vec![],
             default_node_weight: default_node_weight(),
@@ -165,41 +124,20 @@ impl Default for Config {
 
 impl Config {
     fn load_from_file() -> Self {
-        if let Some(config) = Self::try_load_shared() {
-            return config;
-        }
-        Self::load_legacy_jsonl()
+        Self::try_load_shared().unwrap_or_default()
     }
 
     /// Load from shared config. Memory settings in the "memory" section;
     /// API settings resolved from models + backend configuration.
     fn try_load_shared() -> Option<Self> {
         let content = std::fs::read_to_string(config_path()).ok()?;
-        let root: serde_json::Value = json5::from_str(&content).ok()?;
+        let root: serde_json::Value = json_five::from_str(&content).ok()?;
         let mem_value = root.get("memory")?;
 
         let mut config: Config = serde_json::from_value(mem_value.clone()).ok()?;
         config.llm_concurrency = config.llm_concurrency.max(1);
 
-        // Resolve API settings: agent_model → models → backend
-        if let Some(model_name) = &config.agent_model
-            && let Some(model_cfg) = root.get("models").and_then(|m| m.get(model_name.as_str())) {
-                let backend_name = model_cfg.get("backend").and_then(|v| v.as_str()).unwrap_or("");
-                let model_id = model_cfg.get("model_id").and_then(|v| v.as_str()).unwrap_or("");
-
-                if let Some(backend) = root.get(backend_name) {
-                    config.api_base_url = backend.get("base_url")
-                        .and_then(|v| v.as_str()).map(String::from);
-                    config.api_key = backend.get("api_key")
-                        .and_then(|v| v.as_str()).map(String::from);
-                }
-                config.api_model = Some(model_id.to_string());
-                if let Some(cw) = model_cfg.get("context_window").and_then(|v| v.as_u64()) {
-                    config.api_context_window = cw as usize;
-                }
-            }
-
-        // Top-level config sections (not inside "memory")
+        // Top-level sections (not inside "memory").
         if let Some(servers) = root.get("lsp_servers") {
             config.lsp_servers = serde_json::from_value(servers.clone()).unwrap_or_default();
         }
@@ -209,11 +147,6 @@ impl Config {
 
         Some(config)
     }
-
-    /// Load from legacy JSONL config — deprecated, just return defaults.
-    fn load_legacy_jsonl() -> Self {
-        Config::default()
-    }
 }
 
 /// Get the global memory config (cheap Arc clone).
@@ -237,27 +170,99 @@ pub fn reload() -> bool {
     changed
 }
 
+/// Spawn a background thread that watches `~/.consciousness/config.json5`
+/// and reloads both the memory Config and the global AppConfig whenever
+/// the file changes on disk. Lets edits from vim / F6 hotkeys / manual
+/// tweaks land live without restarting the process.
+pub fn watch_config(cli: crate::user::CliArgs) {
+    use notify_debouncer_mini::{new_debouncer, notify::RecursiveMode};
+
+    let path = config_path();
+    // Watch the parent directory — editors often replace-via-rename, so
+    // watching the file itself misses the new inode.
+    let Some(parent) = path.parent().map(|p| p.to_path_buf()) else {
+        crate::dbglog!("[config] no parent for {}, skipping watch", path.display());
+        return;
+    };
+
+    std::thread::Builder::new()
+        .name("config-watcher".into())
+        .spawn(move || {
+            let (tx, rx) = std::sync::mpsc::channel();
+            let mut debouncer = match new_debouncer(std::time::Duration::from_millis(200), tx) {
+                Ok(d) => d,
+                Err(e) => {
+                    crate::dbglog!("[config] watcher setup failed: {}", e);
+                    return;
+                }
+            };
+            if let Err(e) = debouncer.watcher()
+                .watch(&parent, RecursiveMode::NonRecursive)
+            {
+                crate::dbglog!("[config] watch({}) failed: {}", parent.display(), e);
+                return;
+			}
+			crate::dbglog!("[config] watching {}", path.display());
+
+			let mut last_seen = config_file_state(&path);
+			while let Ok(res) = rx.recv() {
+				let Ok(events) = res else { continue; };
+				if !events.iter().any(|e| e.path == path) { continue; }
+
+				let current_seen = config_file_state(&path);
+				if current_seen == last_seen {
+					continue;
+				}
+				last_seen = current_seen;
+
+				// Reload both halves.
+				let mem_changed = reload();
+				let app_changed = match build_figment(&cli).extract::<AppConfig>() {
+                    Ok(app) => {
+                        install_app(app);
+                        true
+                    }
+                    Err(e) => {
+                        crate::dbglog!("[config] reload: AppConfig parse failed: {}", e);
+                        false
+                    }
+                };
+                crate::dbglog!("[config] reloaded (memory_changed={}, app_changed={})",
+                    mem_changed, app_changed);
+            }
+		})
+		.ok();
+}
+
+fn config_file_state(path: &std::path::Path) -> Option<(std::time::SystemTime, u64)> {
+	let meta = std::fs::metadata(path).ok()?;
+	Some((meta.modified().ok()?, meta.len()))
+}
+
 // ============================================================
 // Agent config (top-level settings)
 // ============================================================
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct AppConfig {
-    pub backend: String,
-    pub anthropic: BackendConfig,
-    pub openrouter: BackendConfig,
+    #[serde(default = "default_user_name")]
+    pub user_name: String,
+    #[serde(default = "default_assistant_name")]
+    pub assistant_name: String,
+    /// Named model endpoints — credentials, base URL, and model id bundled
+    /// into one entry per backend. Keyed by name, selected by
+    /// `default_backend` or by `--model <name>` on the CLI.
     #[serde(default)]
-    pub deepinfra: BackendConfig,
-    pub prompts: PromptConfig,
+    pub backends: HashMap<String, BackendConfig>,
+    #[serde(default)]
+    pub default_backend: String,
     pub debug: bool,
     pub compaction: CompactionConfig,
     pub dmn: DmnConfig,
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub memory_project: Option<PathBuf>,
     #[serde(default)]
-    pub models: HashMap<String, ModelConfig>,
-    #[serde(default = "default_model_name")]
-    pub default_model: String,
+    pub learn: LearnConfig,
+    #[serde(default)]
+    pub compare: CompareConfig,
     #[serde(default)]
     pub mcp_servers: Vec<McpServerConfig>,
     #[serde(default)]
@@ -284,32 +289,17 @@ pub struct LspServerConfig {
 
 #[derive(Debug, Clone, Default, Serialize, Deserialize)]
 pub struct BackendConfig {
+    /// API key for the backend.
     #[serde(default)]
     pub api_key: String,
-    #[serde(default)]
-    pub model: String,
-    #[serde(skip_serializing_if = "Option::is_none")]
+    /// Base URL for the backend's OpenAI-compatible endpoint.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
     pub base_url: Option<String>,
-}
-
-impl BackendConfig {
-    fn resolve(&self, default_base: &str) -> Result<(String, String, String)> {
-        if self.api_key.is_empty() {
-            anyhow::bail!(
-                "No API key. Set it in {} or use --api-key",
-                config_path().display()
-            );
-        }
-        let base = self.base_url.clone()
-            .unwrap_or_else(|| default_base.to_string());
-        Ok((base, self.api_key.clone(), self.model.clone()))
-    }
-}
-
-#[derive(Debug, Clone, Serialize, Deserialize)]
-pub struct PromptConfig {
-    pub anthropic: String,
-    pub other: String,
+    /// Model identifier sent to the API.
+    pub model_id: String,
+    /// Context window size in tokens.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub context_window: Option<usize>,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -324,65 +314,68 @@ pub struct DmnConfig {
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize)]
-pub struct ModelConfig {
-    /// Backend name ("anthropic" or "openrouter")
-    pub backend: String,
-    /// Model identifier sent to the API
-    pub model_id: String,
-    /// Instruction file ("CLAUDE.md" or "POC.md").
+pub struct LearnConfig {
+    /// Divergence threshold — responses scoring above this become
+    /// fine-tuning candidates. Lower = more sensitive.
+    #[serde(default = "default_learn_threshold")]
+    pub threshold: f64,
+    /// Whether to generate "what would the model have said without
+    /// memories" alternates alongside each scoring run. Expensive —
+    /// one full streaming generation per candidate.
     #[serde(default)]
-    pub prompt_file: Option<String>,
-    /// Context window size in tokens.
-    #[serde(default)]
-    pub context_window: Option<usize>,
+    pub generate_alternates: bool,
 }
 
+fn default_learn_threshold() -> f64 { 1.0 }
+
+impl Default for LearnConfig {
+    fn default() -> Self {
+        Self {
+            threshold: default_learn_threshold(),
+            generate_alternates: false,
+        }
+    }
+}
+
+/// Settings for the F7 compare screen — side-by-side generation with a
+/// test model against the current context.
+#[derive(Debug, Clone, Default, Serialize, Deserialize)]
+pub struct CompareConfig {
+    /// Backend name (looked up in `backends`) to use as the test model.
+    /// Empty = F7 reports "no test backend configured" and does nothing.
+    #[serde(default)]
+    pub test_backend: String,
+}
+
+fn default_user_name() -> String { "User".into() }
+fn default_assistant_name() -> String { "Assistant".into() }
+
 impl Default for AppConfig {
     fn default() -> Self {
         Self {
-            backend: "openrouter".to_string(),
-            anthropic: BackendConfig {
-                api_key: String::new(),
-                model: "claude-opus-4-6-20250918".to_string(),
-                base_url: None,
-            },
-            openrouter: BackendConfig {
-                api_key: String::new(),
-                model: "qwen/qwen3.5-397b-a17b".to_string(),
-                base_url: Some("https://openrouter.ai/api/v1".to_string()),
-            },
-            deepinfra: BackendConfig {
-                api_key: String::new(),
-                model: String::new(),
-                base_url: Some("https://api.deepinfra.com/v1/openai".to_string()),
-            },
-            prompts: PromptConfig {
-                anthropic: "CLAUDE.md".to_string(),
-                other: "POC.md".to_string(),
-            },
+            user_name: default_user_name(),
+            assistant_name: default_assistant_name(),
+            backends: HashMap::new(),
+            default_backend: String::new(),
             debug: false,
             compaction: CompactionConfig {
                 hard_threshold_pct: 90,
                 soft_threshold_pct: 80,
             },
             dmn: DmnConfig { max_turns: 20 },
-            memory_project: None,
-            models: HashMap::new(),
-            default_model: String::new(),
+            learn: LearnConfig::default(),
+            compare: CompareConfig::default(),
             mcp_servers: Vec::new(),
             lsp_servers: Vec::new(),
         }
     }
 }
 
-fn default_model_name() -> String { String::new() }
-
 /// Resolved, ready-to-use agent session config.
 pub struct SessionConfig {
     pub api_base: String,
     pub api_key: String,
     pub model: String,
-    pub prompt_file: String,
     /// Identity/personality nodes as (name, content) pairs.
     pub context_parts: Vec<(String, String)>,
     pub session_dir: PathBuf,
@@ -398,37 +391,22 @@ pub struct ResolvedModel {
     pub api_base: String,
     pub api_key: String,
     pub model_id: String,
-    pub prompt_file: String,
     pub context_window: Option<usize>,
 }
 
 impl AppConfig {
     /// Resolve the active backend and assemble prompts into a SessionConfig.
     pub async fn resolve(&self, cli: &crate::user::CliArgs) -> Result<SessionConfig> {
-        let (api_base, api_key, model, prompt_file);
-
-        if !self.models.is_empty() {
-            let model_name = cli.model.as_deref().unwrap_or(&self.default_model);
-            let resolved = self.resolve_model(model_name)?;
-            api_base = resolved.api_base;
-            api_key = resolved.api_key;
-            model = resolved.model_id;
-            prompt_file = resolved.prompt_file;
-        } else {
-            let (base, key, mdl) = match self.backend.as_str() {
-                "anthropic" => self.anthropic.resolve("https://api.anthropic.com"),
-                _ => self.openrouter.resolve("https://openrouter.ai/api/v1"),
-            }?;
-            api_base = base;
-            api_key = key;
-            model = mdl;
-            prompt_file = if self.backend == "anthropic" {
-                self.prompts.anthropic.clone()
-            } else {
-                self.prompts.other.clone()
-            };
+        if self.backends.is_empty() {
+            anyhow::bail!(
+                "no backends configured in {}. Add a `backends` section with at least one entry.",
+                config_path().display()
+            );
         }
 
+        let name = cli.model.as_deref().unwrap_or(&self.default_backend);
+        let resolved = self.resolve_model(name)?;
+
         let personality_nodes = get().personality_nodes.clone();
         let context_parts = crate::mind::identity::personality_nodes(&personality_nodes).await;
 
@@ -438,11 +416,13 @@ impl AppConfig {
         std::fs::create_dir_all(&session_dir).ok();
 
         // CLI --api-base and --api-key override everything
-        let api_base = cli.api_base.clone().unwrap_or(api_base);
-        let api_key = cli.api_key.clone().unwrap_or(api_key);
+        let api_base = cli.api_base.clone().unwrap_or(resolved.api_base);
+        let api_key = cli.api_key.clone().unwrap_or(resolved.api_key);
 
         Ok(SessionConfig {
-            api_base, api_key, model, prompt_file,
+            api_base,
+            api_key,
+            model: resolved.model_id,
             context_parts,
             session_dir,
             app: self.clone(),
@@ -450,55 +430,33 @@ impl AppConfig {
         })
     }
 
-    /// Look up a named model and resolve its credentials from the backend config.
+    /// Look up a named backend and resolve its credentials.
     pub fn resolve_model(&self, name: &str) -> Result<ResolvedModel> {
-        let model = self.models.get(name)
+        let b = self.backends.get(name)
             .ok_or_else(|| anyhow::anyhow!(
-                "Unknown model '{}'. Available: {}",
+                "Unknown backend '{}'. Available: {}",
                 name,
                 self.model_names().join(", "),
             ))?;
 
-        let (api_base, api_key) = match model.backend.as_str() {
-            "anthropic" => (
-                self.anthropic.base_url.clone()
-                    .unwrap_or_else(|| "https://api.anthropic.com".to_string()),
-                self.anthropic.api_key.clone(),
-            ),
-            "deepinfra" => (
-                self.deepinfra.base_url.clone()
-                    .unwrap_or_else(|| "https://api.deepinfra.com/v1/openai".to_string()),
-                self.deepinfra.api_key.clone(),
-            ),
-            _ => (
-                self.openrouter.base_url.clone()
-                    .unwrap_or_else(|| "https://openrouter.ai/api/v1".to_string()),
-                self.openrouter.api_key.clone(),
-            ),
-        };
-
-        let prompt_file = model.prompt_file.clone()
-            .unwrap_or_else(|| {
-                if model.backend == "anthropic" {
-                    self.prompts.anthropic.clone()
-                } else {
-                    self.prompts.other.clone()
-                }
-            });
+        let api_base = b.base_url.clone()
+            .ok_or_else(|| anyhow::anyhow!(
+                "backends.{}.base_url not set in {}",
+                name, config_path().display()
+            ))?;
 
         Ok(ResolvedModel {
             name: name.to_string(),
             api_base,
-            api_key,
-            model_id: model.model_id.clone(),
-            prompt_file,
-            context_window: model.context_window,
+            api_key: b.api_key.clone(),
+            model_id: b.model_id.clone(),
+            context_window: b.context_window,
         })
     }
 
-    /// List available model names, sorted.
+    /// List available backend names, sorted.
     pub fn model_names(&self) -> Vec<String> {
-        let mut names: Vec<_> = self.models.keys().cloned().collect();
+        let mut names: Vec<_> = self.backends.keys().cloned().collect();
         names.sort();
         names
     }
@@ -518,7 +476,7 @@ impl Provider for Json5File {
     fn data(&self) -> figment::Result<figment::value::Map<figment::Profile, figment::value::Dict>> {
         match std::fs::read_to_string(&self.0) {
             Ok(content) => {
-                let value: figment::value::Value = json5::from_str(&content)
+                let value: figment::value::Value = json_five::from_str(&content)
                     .map_err(|e| figment::Error::from(format!("{}: {}", self.0.display(), e)))?;
                 Serialized::defaults(value).data()
             }
@@ -540,11 +498,6 @@ fn build_figment(cli: &crate::user::CliArgs) -> Figment {
     let mut f = Figment::from(Serialized::defaults(AppConfig::default()))
         .merge(Json5File(config_path()));
 
-    merge_opt!(f, cli.backend, "backend");
-    merge_opt!(f, cli.model, "anthropic.model", "openrouter.model");
-    merge_opt!(f, cli.api_key, "anthropic.api_key", "openrouter.api_key");
-    merge_opt!(f, cli.api_base, "anthropic.base_url", "openrouter.base_url");
-    merge_opt!(f, cli.memory_project, "memory_project");
     merge_opt!(f, cli.dmn_max_turns, "dmn.max_turns");
     if cli.debug {
         f = f.merge(Serialized::default("debug", true));
@@ -554,12 +507,46 @@ fn build_figment(cli: &crate::user::CliArgs) -> Figment {
 }
 
 /// Load just the AppConfig — no validation, no prompt assembly.
+/// Also installs the loaded AppConfig into the global cache so
+/// `config::app()` is available everywhere.
 pub fn load_app(cli: &crate::user::CliArgs) -> Result<(AppConfig, Figment)> {
     let figment = build_figment(cli);
     let app: AppConfig = figment.extract().context("Failed to load configuration")?;
+    install_app(app.clone());
     Ok((app, figment))
 }
 
+// ============================================================
+// Global AppConfig cache (writable, for runtime-mutable settings
+// like learn.threshold that F6 edits via config_writer).
+// ============================================================
+
+static APP_CONFIG: OnceLock<RwLock<AppConfig>> = OnceLock::new();
+
+fn install_app(app: AppConfig) {
+    let slot = APP_CONFIG.get_or_init(|| RwLock::new(app.clone()));
+    *slot.write().unwrap() = app;
+}
+
+/// Current AppConfig, held under a read lock. Reads should be brief
+/// (no holding across await / long work) to avoid starving writers.
+/// Panics if called before load_app — which runs once at startup.
+pub fn app() -> std::sync::RwLockReadGuard<'static, AppConfig> {
+    APP_CONFIG
+        .get()
+        .expect("config::app() called before load_app()")
+        .read()
+        .unwrap()
+}
+
+/// Mutate the cached AppConfig in place. Used by config_writer to keep
+/// the in-memory view in sync with disk after surgical edits to
+/// ~/.consciousness/config.json5.
+pub fn update_app(f: impl FnOnce(&mut AppConfig)) {
+    let slot = APP_CONFIG.get().expect("update_app before load_app");
+    f(&mut *slot.write().unwrap());
+}
+
 /// Load the full config: figment → AppConfig → resolve backend → assemble prompts.
 pub async fn load_session(cli: &crate::user::CliArgs) -> Result<(SessionConfig, Figment)> {
     let (app, figment) = load_app(cli)?;
@@ -585,38 +572,28 @@ pub fn show_config(app: &AppConfig, figment: &Figment) {
     }
 
     println!("# Effective configuration\n");
-    println!("backend: {:?}  ({})", app.backend, src(figment, "backend"));
-    for (name, b) in [("anthropic", &app.anthropic), ("openrouter", &app.openrouter)] {
-        println!("\n{}:", name);
-        println!("  api_key: {}  ({})", mask(&b.api_key), src(figment, &format!("{name}.api_key")));
-        println!("  model: {:?}  ({})", b.model, src(figment, &format!("{name}.model")));
-        if let Some(ref url) = b.base_url {
-            println!("  base_url: {:?}  ({})", url, src(figment, &format!("{name}.base_url")));
-        }
-    }
-    println!("\nprompts:");
-    println!("  anthropic: {:?}  ({})", app.prompts.anthropic, src(figment, "prompts.anthropic"));
-    println!("  other: {:?}  ({})", app.prompts.other, src(figment, "prompts.other"));
+    println!("user_name: {:?}  ({})", app.user_name, src(figment, "user_name"));
+    println!("assistant_name: {:?}  ({})", app.assistant_name, src(figment, "assistant_name"));
     println!("\ndebug: {}  ({})", app.debug, src(figment, "debug"));
     println!("\ncompaction:");
     println!("  hard_threshold_pct: {}  ({})", app.compaction.hard_threshold_pct, src(figment, "compaction.hard_threshold_pct"));
     println!("  soft_threshold_pct: {}  ({})", app.compaction.soft_threshold_pct, src(figment, "compaction.soft_threshold_pct"));
     println!("\ndmn:");
     println!("  max_turns: {}  ({})", app.dmn.max_turns, src(figment, "dmn.max_turns"));
-    if let Some(ref p) = app.memory_project {
-        println!("\nmemory_project: {:?}  ({})", p, src(figment, "memory_project"));
-    }
-    println!("\ndefault_model: {:?}", app.default_model);
-    if !app.models.is_empty() {
-        println!("\nmodels:");
-        for (name, m) in &app.models {
+    println!("\ndefault_backend: {:?}  ({})", app.default_backend, src(figment, "default_backend"));
+    if !app.backends.is_empty() {
+        println!("\nbackends:");
+        let mut names: Vec<_> = app.backends.keys().cloned().collect();
+        names.sort();
+        for name in names {
+            let b = &app.backends[&name];
             println!("  {}:", name);
-            println!("    backend: {:?}", m.backend);
-            println!("    model_id: {:?}", m.model_id);
-            if let Some(ref pf) = m.prompt_file {
-                println!("    prompt_file: {:?}", pf);
+            println!("    api_key: {}  ({})", mask(&b.api_key), src(figment, &format!("backends.{name}.api_key")));
+            if let Some(ref url) = b.base_url {
+                println!("    base_url: {:?}  ({})", url, src(figment, &format!("backends.{name}.base_url")));
             }
-            if let Some(cw) = m.context_window {
+            println!("    model_id: {:?}", b.model_id);
+            if let Some(cw) = b.context_window {
                 println!("    context_window: {}", cw);
             }
         }
diff --git a/src/config_writer.rs b/src/config_writer.rs
new file mode 100644
index 0000000..079449f
--- /dev/null
+++ b/src/config_writer.rs
@@ -0,0 +1,448 @@
+// config_writer.rs — Surgical edits to ~/.consciousness/config.json5
+//
+// Uses json-five's round-trip parser to mutate specific fields while
+// preserving the surrounding comments, whitespace, and formatting.
+
+use std::path::Path;
+
+use anyhow::{anyhow, Context as _, Result};
+use json_five::rt::parser::{
+    from_str, JSONKeyValuePair, JSONObjectContext, JSONValue, KeyValuePairContext,
+};
+
+use crate::config::config_path;
+
+/// Read the config, apply `mutate` to the root JSONValue, write it back atomically.
+fn edit_config<F: FnOnce(&mut JSONValue) -> Result<()>>(mutate: F) -> Result<()> {
+    let path = config_path();
+    let src = std::fs::read_to_string(&path)
+        .with_context(|| format!("read {}", path.display()))?;
+
+    let mut text = from_str(&src)
+        .map_err(|e| anyhow!("parse {}: {}", path.display(), e))?;
+    mutate(&mut text.value)?;
+
+    write_atomic(&path, &text.to_string())
+}
+
+fn write_atomic(path: &Path, content: &str) -> Result<()> {
+    let parent = path.parent()
+        .ok_or_else(|| anyhow!("config path has no parent: {}", path.display()))?;
+    let tmp = parent.join(format!(
+        ".{}.tmp",
+        path.file_name().unwrap_or_default().to_string_lossy(),
+    ));
+    std::fs::write(&tmp, content)
+        .with_context(|| format!("write {}", tmp.display()))?;
+    std::fs::rename(&tmp, path)
+        .with_context(|| format!("rename {} -> {}", tmp.display(), path.display()))?;
+    Ok(())
+}
+
+/// Match a key JSONValue against a string name. JSON5 allows keys to be
+/// unquoted identifiers or single/double-quoted strings.
+fn key_matches(key: &JSONValue, name: &str) -> bool {
+    match key {
+        JSONValue::Identifier(s)
+        | JSONValue::DoubleQuotedString(s)
+        | JSONValue::SingleQuotedString(s) => s == name,
+        _ => false,
+    }
+}
+
+/// Find (or create) a child object under `parent`, returning a mutable borrow
+/// of its key_value_pairs vector.
+/// Append a new kvp to `object`, setting whitespace so the output is
+/// multi-line with the given indentation:
+///
+/// ```text
+/// {<newline><inner_indent>first_key: first_val,<newline><outer_indent>}
+/// ```
+///
+/// If `object` already has kvps, the separator between the last one and
+/// ours goes in the prior kvp's wsc.3. If we're the first kvp, the
+/// lead-in after `{` goes in the object's own wsc.0.
+fn append_kvp_pretty(
+    object: &mut JSONValue,
+    key: JSONValue,
+    value: JSONValue,
+    inner_indent: &str,
+    outer_indent: &str,
+) -> Result<()> {
+    let (pairs, ctx) = match object {
+        JSONValue::JSONObject { key_value_pairs, context } => {
+            let ctx = context.get_or_insert_with(|| JSONObjectContext {
+                wsc: (String::new(),),
+            });
+            (key_value_pairs, ctx)
+        }
+        _ => return Err(anyhow!("not an object")),
+    };
+
+    if pairs.is_empty() {
+        ctx.wsc.0 = format!("\n{}", inner_indent);
+    } else {
+        let prev = pairs.last_mut().unwrap();
+        let prev_ctx = prev.context.get_or_insert_with(|| KeyValuePairContext {
+            wsc: (String::new(), String::from(" "), String::new(), None),
+        });
+        prev_ctx.wsc.3 = Some(format!("\n{}", inner_indent));
+    }
+
+    pairs.push(JSONKeyValuePair {
+        key,
+        value,
+        context: Some(KeyValuePairContext {
+            wsc: (
+                String::new(),
+                String::from(" "),
+                String::new(),
+                Some(format!("\n{}", outer_indent)),
+            ),
+        }),
+    });
+
+    Ok(())
+}
+
+/// Find or create a child object under `parent`. Returns the index of
+/// the kvp in parent's key_value_pairs so the caller can re-borrow
+/// afterward.
+fn get_or_create_object_idx(
+    parent: &mut JSONValue,
+    section: &str,
+    inner_indent: &str,
+    outer_indent: &str,
+) -> Result<usize> {
+    let existing = match parent {
+        JSONValue::JSONObject { key_value_pairs, .. } => {
+            key_value_pairs.iter()
+                .position(|kvp| key_matches(&kvp.key, section))
+        }
+        _ => return Err(anyhow!("config root is not an object")),
+    };
+
+    if let Some(i) = existing {
+        return Ok(i);
+    }
+
+    append_kvp_pretty(
+        parent,
+        JSONValue::Identifier(section.to_string()),
+        JSONValue::JSONObject {
+            key_value_pairs: Vec::new(),
+            context: Some(JSONObjectContext { wsc: (String::new(),) }),
+        },
+        inner_indent,
+        outer_indent,
+    )?;
+
+    match parent {
+        JSONValue::JSONObject { key_value_pairs, .. } => Ok(key_value_pairs.len() - 1),
+        _ => unreachable!(),
+    }
+}
+
+/// Set `section.key` to a literal scalar value (e.g., "1e-7", "42", "true").
+/// The literal is parsed as JSON5 so we preserve its source-form on round-trip.
+pub fn set_scalar(section: &str, key: &str, literal: &str) -> Result<()> {
+    let value = parse_scalar_literal(literal)?;
+    edit_config(|root| {
+        // New top-level sections sit at column 4 (inside root `{`),
+        // and the root's closing `}` sits at column 0.
+        let section_idx = get_or_create_object_idx(root, section, "    ", "")?;
+
+        let section_value = match root {
+            JSONValue::JSONObject { key_value_pairs, .. } => {
+                &mut key_value_pairs[section_idx].value
+            }
+            _ => unreachable!(),
+        };
+
+        // Update in place if the key already exists.
+        if let JSONValue::JSONObject { key_value_pairs, .. } = section_value {
+            if let Some(kvp) = key_value_pairs.iter_mut()
+                .find(|k| key_matches(&k.key, key))
+            {
+                kvp.value = value;
+                return Ok(());
+            }
+        }
+
+        // Append a new kvp. Inner keys sit at column 8, the section's
+        // closing `}` sits at column 4.
+        append_kvp_pretty(
+            section_value,
+            JSONValue::Identifier(key.to_string()),
+            value,
+            "        ",
+            "    ",
+        )
+    })
+}
+
+/// Parse a scalar literal by round-tripping it through json-five. Keeps us
+/// consistent with whatever scalars the library considers valid (hex,
+/// exponents, Infinity, etc.).
+fn parse_scalar_literal(literal: &str) -> Result<JSONValue> {
+    let text = from_str(literal)
+        .map_err(|e| anyhow!("parse literal {:?}: {}", literal, e))?;
+    match text.value {
+        JSONValue::JSONObject { .. } | JSONValue::JSONArray { .. } => {
+            Err(anyhow!("set_scalar only accepts scalar literals, got {:?}", literal))
+        }
+        v => Ok(v),
+    }
+}
+
+/// Convenience: set `learn.threshold` to the given f64.
+pub fn set_learn_threshold(value: f64) -> Result<()> {
+    // {:e} gives the minimal scientific notation that preserves the value.
+    set_scalar("learn", "threshold", &format!("{:e}", value))?;
+    crate::config::update_app(|app| app.learn.threshold = value);
+    Ok(())
+}
+
+/// Convenience: set `learn.generate_alternates` to the given bool.
+pub fn set_learn_generate_alternates(value: bool) -> Result<()> {
+    set_scalar("learn", "generate_alternates",
+        if value { "true" } else { "false" })?;
+    crate::config::update_app(|app| app.learn.generate_alternates = value);
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    // In-memory variant of set_scalar — used to test the mutation logic
+    // without touching disk.
+    fn set_scalar_inline(
+        root: &mut JSONValue,
+        section: &str,
+        key: &str,
+        literal: &str,
+    ) -> Result<()> {
+        let value = parse_scalar_literal(literal)?;
+        let section_idx = get_or_create_object_idx(root, section, "    ", "")?;
+        let section_value = match root {
+            JSONValue::JSONObject { key_value_pairs, .. } => {
+                &mut key_value_pairs[section_idx].value
+            }
+            _ => unreachable!(),
+        };
+        if let JSONValue::JSONObject { key_value_pairs, .. } = section_value {
+            if let Some(kvp) = key_value_pairs.iter_mut()
+                .find(|k| key_matches(&k.key, key))
+            {
+                kvp.value = value;
+                return Ok(());
+            }
+        }
+        append_kvp_pretty(
+            section_value,
+            JSONValue::Identifier(key.to_string()),
+            value,
+            "        ",
+            "    ",
+        )
+    }
+
+    fn edit_str<F: FnOnce(&mut JSONValue) -> Result<()>>(src: &str, f: F) -> Result<String> {
+        let mut text = from_str(src).map_err(|e| anyhow!("{}", e))?;
+        f(&mut text.value)?;
+        Ok(text.to_string())
+    }
+
+    #[test]
+    fn replaces_existing_scalar() {
+        let src = r#"{
+    // threshold for learning
+    learn: {
+        threshold: 0.001, // the old value
+    },
+}"#;
+        let out = edit_str(src, |root| {
+            set_scalar_inline(root, "learn", "threshold", "1e-7")
+        }).unwrap();
+        assert!(out.contains("1e-7"), "output: {}", out);
+        assert!(out.contains("// threshold for learning"));
+        assert!(out.contains("// the old value"));
+        assert!(!out.contains("0.001"));
+    }
+
+    #[test]
+    fn creates_missing_section() {
+        let src = r#"{
+    // comment
+    memory: { user_name: "Kent" },
+}"#;
+        let out = edit_str(src, |root| {
+            set_scalar_inline(root, "learn", "threshold", "1e-7")
+        }).unwrap();
+        assert!(out.contains("learn"));
+        assert!(out.contains("1e-7"));
+        assert!(out.contains("// comment"));
+        assert!(out.contains(r#"user_name: "Kent""#));
+    }
+
+    #[test]
+    fn preserves_comments_in_siblings() {
+        let src = r#"{
+    memory: {
+        // sensitive setting
+        user_name: "Kent", // name
+    },
+    learn: {
+        threshold: 0.5,
+    },
+}"#;
+        let out = edit_str(src, |root| {
+            set_scalar_inline(root, "learn", "threshold", "1e-9")
+        }).unwrap();
+        assert!(out.contains("// sensitive setting"));
+        assert!(out.contains("// name"));
+        assert!(out.contains("1e-9"));
+        assert!(!out.contains("0.5"));
+    }
+
+    #[test]
+    fn adds_key_to_existing_empty_section() {
+        let src = r#"{
+    learn: {},
+}"#;
+        let out = edit_str(src, |root| {
+            set_scalar_inline(root, "learn", "threshold", "42")
+        }).unwrap();
+        assert!(out.contains("threshold"), "output: {}", out);
+        assert!(out.contains("42"));
+    }
+
+    #[test]
+    fn realistic_config_adds_learn_section() {
+        // Mirrors the shape of ~/.consciousness/config.json5 — multiple
+        // sections, comments, mixed tab/space indent, trailing commas.
+        let src = r#"{
+    deepinfra: {
+        api_key: "bcachefs-agents-2026",
+        base_url: "http://example/v1",
+    },
+
+    // Named models
+    models: {
+        "27b": {
+            backend: "deepinfra",
+            model_id: "Qwen/Qwen3.5-27B",
+        },
+    },
+
+    default_model: "27b",
+
+    memory: {
+        user_name: "Kent",
+        // Active agent types
+        agent_types: ["linker", "organize"],
+    },
+
+    compaction: {
+        hard_threshold_pct: 90,
+    },
+}"#;
+        let out = edit_str(src, |root| {
+            set_scalar_inline(root, "learn", "threshold", "1e-7")
+        }).unwrap();
+
+        // Core assertions: comments and sibling sections survive.
+        assert!(out.contains(r#"api_key: "bcachefs-agents-2026""#));
+        assert!(out.contains("// Named models"));
+        assert!(out.contains("// Active agent types"));
+        assert!(out.contains(r#"user_name: "Kent""#));
+        assert!(out.contains("hard_threshold_pct: 90"));
+
+        // New section added.
+        assert!(out.contains("learn"));
+        assert!(out.contains("1e-7"));
+
+        // Parse result should parse back without error (real json5 parser).
+        let reparsed: serde_json::Value = json_five::from_str(&out)
+            .expect("mutated output must be valid JSON5");
+        let threshold = reparsed.pointer("/learn/threshold").expect("learn.threshold exists");
+        assert_eq!(threshold.as_f64(), Some(1e-7));
+    }
+
+    #[test]
+    fn realistic_config_updates_existing_threshold() {
+        let src = r#"{
+    learn: {
+        // The divergence threshold
+        threshold: 0.001,
+    },
+    memory: { user_name: "Kent" },
+}"#;
+        let out = edit_str(src, |root| {
+            set_scalar_inline(root, "learn", "threshold", "5e-8")
+        }).unwrap();
+        assert!(out.contains("5e-8"));
+        assert!(!out.contains("0.001"));
+        assert!(out.contains("// The divergence threshold"));
+
+        let reparsed: serde_json::Value = json_five::from_str(&out).unwrap();
+        assert_eq!(reparsed.pointer("/learn/threshold").and_then(|v| v.as_f64()), Some(5e-8));
+    }
+
+    #[test]
+    fn new_section_exact_multiline_layout() {
+        let src = "{\n    a: 1,\n}";
+        let out = edit_str(src, |root| {
+            set_scalar_inline(root, "learn", "generate_alternates", "true")?;
+            set_scalar_inline(root, "learn", "threshold", "1e-7")
+        }).unwrap();
+
+        let expected = "\
+{
+    a: 1,
+    learn: {
+        generate_alternates: true,
+        threshold: 1e-7,
+    },
+}";
+        assert_eq!(out, expected, "\n--- got ---\n{}\n--- want ---\n{}\n", out, expected);
+    }
+
+    #[test]
+    fn new_section_and_key_format_cleanly() {
+        // The kind of config we actually have in ~/.consciousness
+        // (top-level sections separated by blank lines, 4-space indent
+        // for keys within each section). Appending a fresh `learn`
+        // section with one key should land cleanly, not as
+        // `learn\n\n    :{key\n        :value}`.
+        let src = "{\n    memory: {\n        user_name: \"Kent\",\n    },\n}";
+        let out = edit_str(src, |root| {
+            set_scalar_inline(root, "learn", "generate_alternates", "true")
+        }).unwrap();
+
+        // No stray key-to-colon-on-next-line anywhere.
+        assert!(!out.contains("learn\n"), "learn key wraps: {}", out);
+        assert!(!out.contains("generate_alternates\n"),
+            "inner key wraps: {}", out);
+
+        // The output should reparse.
+        let v: serde_json::Value = json_five::from_str(&out).unwrap();
+        assert_eq!(
+            v.pointer("/learn/generate_alternates").and_then(|x| x.as_bool()),
+            Some(true),
+            "output: {}", out,
+        );
+    }
+
+    #[test]
+    fn roundtrip_stable_without_change() {
+        let src = r#"{
+    // heading
+    a: 1,
+    b: { c: 2 }, // inline
+}"#;
+        let text = from_str(src).unwrap();
+        assert_eq!(text.to_string(), src);
+    }
+}
diff --git a/src/conversation/claude.rs b/src/conversation/claude.rs
new file mode 100644
index 0000000..5eb264e
--- /dev/null
+++ b/src/conversation/claude.rs
@@ -0,0 +1,113 @@
+use serde_json::Value;
+
+use super::{ConversationSource, TranscriptMessage, TranscriptRole};
+
+pub struct ClaudeSource;
+
+impl ConversationSource for ClaudeSource {
+    fn parse_message(&self, obj: &Value, offset: u64) -> Option<TranscriptMessage> {
+        parse_message(obj, offset)
+    }
+
+    fn is_compaction(&self, obj: &Value) -> bool {
+        is_compaction(obj)
+    }
+
+    fn may_contain_compaction(&self, obj_bytes: &[u8]) -> bool {
+        contains_bytes(obj_bytes, b"This session is being continued")
+    }
+}
+
+fn text_content(value: &Value) -> Option<String> {
+    let text = match value {
+        Value::String(s) => s.clone(),
+        Value::Array(arr) => {
+            arr.iter()
+                .filter(|b| b.get("type").and_then(|v| v.as_str()) == Some("text"))
+                .filter_map(|b| b.get("text").and_then(|v| v.as_str()))
+                .collect::<Vec<_>>()
+                .join(" ")
+        }
+        _ => return None,
+    };
+    (!text.is_empty()).then_some(text)
+}
+
+pub(crate) fn parse_message(obj: &Value, offset: u64) -> Option<TranscriptMessage> {
+    let role = match obj.get("type").and_then(|v| v.as_str()) {
+        Some("user") => TranscriptRole::User,
+        Some("assistant") => TranscriptRole::Assistant,
+        _ => return None,
+    };
+
+    let msg = obj.get("message").unwrap_or(obj);
+    let text = msg.get("content").and_then(text_content)?;
+    let timestamp = obj.get("timestamp")
+        .and_then(|v| v.as_str())
+        .map(str::to_string);
+
+    Some(TranscriptMessage { role, text, timestamp, offset })
+}
+
+pub(crate) fn is_compaction(obj: &Value) -> bool {
+    obj.get("type").and_then(|v| v.as_str()) == Some("user")
+        && obj.get("message")
+            .and_then(|m| m.get("content"))
+            .and_then(|c| c.as_str())
+            .is_some_and(|content| content.starts_with("This session is being continued"))
+}
+
+fn contains_bytes(haystack: &[u8], needle: &[u8]) -> bool {
+    haystack.windows(needle.len()).any(|w| w == needle)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    #[test]
+    fn parses_string_and_array_content() {
+        let user = json!({
+            "timestamp": "2026-06-15T15:00:00.000Z",
+            "type": "user",
+            "message": { "content": "hello" }
+        });
+        let assistant = json!({
+            "timestamp": "2026-06-15T15:00:01.000Z",
+            "type": "assistant",
+            "message": {
+                "content": [
+                    { "type": "text", "text": "hi" },
+                    { "type": "tool_use", "name": "ignored" },
+                    { "type": "text", "text": "there" }
+                ]
+            }
+        });
+
+        assert_eq!(
+            parse_message(&user, 7).unwrap(),
+            TranscriptMessage {
+                role: TranscriptRole::User,
+                text: "hello".to_string(),
+                timestamp: Some("2026-06-15T15:00:00.000Z".to_string()),
+                offset: 7,
+            }
+        );
+
+        assert_eq!(parse_message(&assistant, 9).unwrap().text, "hi there");
+    }
+
+    #[test]
+    fn detects_compaction_marker() {
+        let obj = json!({
+            "timestamp": "2026-06-15T15:00:01.000Z",
+            "type": "user",
+            "message": {
+                "content": "This session is being continued from a previous conversation."
+            }
+        });
+
+        assert!(is_compaction(&obj));
+    }
+}
diff --git a/src/conversation/codex.rs b/src/conversation/codex.rs
new file mode 100644
index 0000000..3b053d8
--- /dev/null
+++ b/src/conversation/codex.rs
@@ -0,0 +1,105 @@
+use serde_json::Value;
+
+use super::{ConversationSource, TranscriptMessage, TranscriptRole};
+
+pub struct CodexSource;
+
+impl ConversationSource for CodexSource {
+    fn parse_message(&self, obj: &Value, offset: u64) -> Option<TranscriptMessage> {
+        parse_message(obj, offset)
+    }
+
+    fn is_compaction(&self, obj: &Value) -> bool {
+        is_compaction(obj)
+    }
+
+    fn may_contain_compaction(&self, obj_bytes: &[u8]) -> bool {
+        contains_bytes(obj_bytes, b"context_compacted")
+    }
+}
+
+pub(crate) fn parse_message(obj: &Value, offset: u64) -> Option<TranscriptMessage> {
+    if obj.get("type").and_then(|v| v.as_str()) != Some("event_msg") {
+        return None;
+    }
+
+    let payload = obj.get("payload")?;
+    let (role, text) = match payload.get("type").and_then(|v| v.as_str()) {
+        Some("user_message") => (
+            TranscriptRole::User,
+            payload.get("message").and_then(|v| v.as_str())?.to_string(),
+        ),
+        Some("agent_message") => (
+            TranscriptRole::Assistant,
+            payload.get("message").and_then(|v| v.as_str())?.to_string(),
+        ),
+        _ => return None,
+    };
+
+    if text.is_empty() {
+        return None;
+    }
+
+    let timestamp = obj.get("timestamp")
+        .and_then(|v| v.as_str())
+        .map(str::to_string);
+
+    Some(TranscriptMessage { role, text, timestamp, offset })
+}
+
+pub(crate) fn is_compaction(obj: &Value) -> bool {
+    obj.get("type").and_then(|v| v.as_str()) == Some("event_msg")
+        && obj.get("payload")
+            .and_then(|p| p.get("type"))
+            .and_then(|v| v.as_str()) == Some("context_compacted")
+}
+
+fn contains_bytes(haystack: &[u8], needle: &[u8]) -> bool {
+    haystack.windows(needle.len()).any(|w| w == needle)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    #[test]
+    fn parses_event_messages_and_skips_noise() {
+        let user = json!({
+            "timestamp": "2026-06-15T15:00:00.000Z",
+            "type": "event_msg",
+            "payload": { "type": "user_message", "message": "start here" }
+        });
+        let assistant = json!({
+            "timestamp": "2026-06-15T15:00:01.000Z",
+            "type": "event_msg",
+            "payload": { "type": "agent_message", "message": "working" }
+        });
+        let tool = json!({
+            "timestamp": "2026-06-15T15:00:02.000Z",
+            "type": "event_msg",
+            "payload": { "type": "task_started" }
+        });
+        let raw = json!({
+            "timestamp": "2026-06-15T15:00:03.000Z",
+            "type": "response_item",
+            "payload": { "type": "message", "role": "user" }
+        });
+
+        assert_eq!(parse_message(&user, 1).unwrap().role, TranscriptRole::User);
+        assert_eq!(parse_message(&assistant, 2).unwrap().text, "working");
+        assert!(parse_message(&tool, 3).is_none());
+        assert!(parse_message(&raw, 4).is_none());
+    }
+
+    #[test]
+    fn detects_compaction_event() {
+        let obj = json!({
+            "timestamp": "2026-06-15T15:00:01.000Z",
+            "type": "event_msg",
+            "payload": { "type": "context_compacted" }
+        });
+
+        assert!(is_compaction(&obj));
+    }
+}
diff --git a/src/conversation/jsonl.rs b/src/conversation/jsonl.rs
new file mode 100644
index 0000000..2675dbc
--- /dev/null
+++ b/src/conversation/jsonl.rs
@@ -0,0 +1,110 @@
+use memchr::memrchr3;
+
+/// Scan backwards through mmap'd bytes, yielding byte slices of complete
+/// top-level JSON objects (outermost { to matching }).
+///
+/// Uses memrchr3 (SIMD) to jump between structurally significant bytes
+/// ({, }, ") instead of scanning byte-by-byte. Tracks brace depth,
+/// skipping braces inside JSON strings. Returns objects in reverse order
+/// (newest first).
+pub struct JsonlBackwardIter<'a> {
+    data: &'a [u8],
+    pos: usize,
+}
+
+impl<'a> JsonlBackwardIter<'a> {
+    pub fn new(data: &'a [u8]) -> Self {
+        Self { data, pos: data.len() }
+    }
+}
+
+impl<'a> Iterator for JsonlBackwardIter<'a> {
+    type Item = (usize, &'a [u8]);
+
+    fn next(&mut self) -> Option<Self::Item> {
+        next_json_object(self.data, &mut self.pos)
+    }
+}
+
+fn is_unescaped_quote(data: &[u8], p: usize) -> bool {
+    let mut bs = 0;
+    while p > bs && data[p - 1 - bs] == b'\\' {
+        bs += 1;
+    }
+    bs % 2 == 0
+}
+
+fn next_json_object<'a>(data: &'a [u8], pos: &mut usize) -> Option<(usize, &'a [u8])> {
+    // Find the closing } of the next object, skipping } inside strings.
+    let close = {
+        let mut in_string = false;
+        loop {
+            let p = memrchr3(b'{', b'}', b'"', &data[..*pos])?;
+            *pos = p;
+            let ch = data[p];
+
+            if in_string {
+                if ch == b'"' && is_unescaped_quote(data, p) {
+                    in_string = false;
+                }
+                continue;
+            }
+
+            match ch {
+                b'}' => break p,
+                b'"' => in_string = true,
+                _ => {}
+            }
+        }
+    };
+
+    // Track brace depth to find matching {.
+    let mut depth: usize = 1;
+    let mut in_string = false;
+
+    loop {
+        let p = memrchr3(b'{', b'}', b'"', &data[..*pos])?;
+        *pos = p;
+        let ch = data[p];
+
+        if in_string {
+            if ch == b'"' && is_unescaped_quote(data, p) {
+                in_string = false;
+            }
+            continue;
+        }
+
+        match ch {
+            b'"' => { in_string = true; }
+            b'}' => { depth += 1; }
+            b'{' => {
+                depth -= 1;
+                if depth == 0 {
+                    return Some((*pos, &data[*pos..=close]));
+                }
+            }
+            _ => {}
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn handles_nested_json_and_quoted_braces() {
+        let data = br#"{"n":1,"s":"literal } brace"}
+{"n":2,"nested":{"s":"escaped quote: \" and { brace"}}
+trailing garbage
+"#;
+
+        let objs: Vec<_> = JsonlBackwardIter::new(data)
+            .map(|(_, bytes)| std::str::from_utf8(bytes).unwrap().to_string())
+            .collect();
+
+        assert_eq!(objs.len(), 2);
+        assert!(objs[0].contains(r#""n":2"#));
+        assert!(objs[1].contains(r#""n":1"#));
+    }
+}
diff --git a/src/conversation/mod.rs b/src/conversation/mod.rs
new file mode 100644
index 0000000..a6ac1f9
--- /dev/null
+++ b/src/conversation/mod.rs
@@ -0,0 +1,271 @@
+// Conversation transcript abstraction.
+//
+// Core code consumes normalized user/assistant messages through this module.
+// Product-specific log formats live in the small compatibility sources below.
+
+use memmap2::Mmap;
+use serde_json::Value;
+use std::fs;
+use std::path::Path;
+
+pub mod claude;
+pub mod codex;
+pub mod jsonl;
+
+pub use jsonl::JsonlBackwardIter;
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum TranscriptRole {
+    User,
+    Assistant,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct TranscriptMessage {
+    pub role: TranscriptRole,
+    pub text: String,
+    pub timestamp: Option<String>,
+    pub offset: u64,
+}
+
+pub trait ConversationSource {
+    fn parse_message(&self, obj: &Value, offset: u64) -> Option<TranscriptMessage>;
+    fn is_compaction(&self, obj: &Value) -> bool;
+
+    fn may_contain_compaction(&self, _obj_bytes: &[u8]) -> bool {
+        true
+    }
+}
+
+pub struct AnyConversationSource;
+
+impl ConversationSource for AnyConversationSource {
+    fn parse_message(&self, obj: &Value, offset: u64) -> Option<TranscriptMessage> {
+        claude::ClaudeSource.parse_message(obj, offset)
+            .or_else(|| codex::CodexSource.parse_message(obj, offset))
+    }
+
+    fn is_compaction(&self, obj: &Value) -> bool {
+        claude::ClaudeSource.is_compaction(obj) || codex::CodexSource.is_compaction(obj)
+    }
+
+    fn may_contain_compaction(&self, obj_bytes: &[u8]) -> bool {
+        claude::ClaudeSource.may_contain_compaction(obj_bytes)
+            || codex::CodexSource.may_contain_compaction(obj_bytes)
+    }
+}
+
+/// Find the byte offset of the last compaction marker in mmap'd transcript data.
+/// Returns the byte offset of the JSON object's opening brace.
+pub(crate) fn find_last_compaction(data: &[u8]) -> Option<usize> {
+    find_last_compaction_with(data, &AnyConversationSource)
+}
+
+pub(crate) fn find_last_compaction_with(
+    data: &[u8],
+    source: &impl ConversationSource,
+) -> Option<usize> {
+    for (offset, obj_bytes) in JsonlBackwardIter::new(data) {
+        // Quick byte check before parsing large transcript entries.
+        if !source.may_contain_compaction(obj_bytes) {
+            continue;
+        }
+
+        let obj: Value = match serde_json::from_slice(obj_bytes) {
+            Ok(v) => v,
+            Err(_) => continue,
+        };
+
+        if source.is_compaction(&obj) {
+            return Some(offset);
+        }
+    }
+
+    None
+}
+
+/// Find the byte offset of the last compaction in a transcript file.
+/// Returns None if the file can't be opened or has no compaction.
+pub(crate) fn find_last_compaction_in_file(path: &str) -> Option<u64> {
+    if path.is_empty() { return None; }
+
+    let file = fs::File::open(path).ok()?;
+    let meta = file.metadata().ok()?;
+    if meta.len() == 0 { return None; }
+
+    let mmap = unsafe { Mmap::map(&file).ok()? };
+    find_last_compaction(&mmap).map(|off| off as u64)
+}
+
+/// Mmap a transcript file. Returns (Mmap, File) to keep both alive.
+pub(crate) fn mmap_transcript(path: &str) -> Option<(Mmap, fs::File)> {
+    let file = fs::File::open(path).ok()?;
+    let meta = file.metadata().ok()?;
+    if meta.len() == 0 { return None; }
+    let mmap = unsafe { Mmap::map(&file).ok()? };
+    Some((mmap, file))
+}
+
+/// Reverse iterator over user/assistant messages in a transcript file.
+/// Yields normalized transcript messages newest-first. The caller decides
+/// when to stop (byte budget, count, etc).
+pub struct TailMessages {
+    _file: fs::File,
+    mmap: Mmap,
+    pos: usize,
+}
+
+impl TailMessages {
+    pub fn open(path: &str) -> Option<Self> {
+        let (mmap, file) = mmap_transcript(path)?;
+        let pos = mmap.len();
+        Some(Self { _file: file, mmap, pos })
+    }
+}
+
+impl Iterator for TailMessages {
+    type Item = TranscriptMessage;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        loop {
+            let (offset, obj_bytes) = jsonl::JsonlBackwardIter::new(&self.mmap[..self.pos]).next()?;
+            self.pos = offset;
+
+            let obj: Value = match serde_json::from_slice(obj_bytes) {
+                Ok(v) => v,
+                Err(_) => continue,
+            };
+
+            if let Some(message) = AnyConversationSource.parse_message(&obj, offset as u64) {
+                return Some(message);
+            }
+        }
+    }
+}
+
+/// Get the timestamp of the compaction message at a given byte offset.
+/// Returns a human-readable datetime string, or None if unavailable.
+pub fn compaction_timestamp(path: &str, offset: u64) -> Option<String> {
+    let (mmap, _file) = mmap_transcript(path)?;
+    let start = offset as usize;
+    if start >= mmap.len() { return None; }
+
+    // Find the end of this JSONL line
+    let end = mmap[start..].iter().position(|&b| b == b'\n')
+        .map(|p| start + p)
+        .unwrap_or(mmap.len());
+
+    let obj: Value = serde_json::from_slice(&mmap[start..end]).ok()?;
+
+    if let Some(ts) = obj.get("timestamp").and_then(|v| v.as_str()) {
+        return Some(ts.to_string());
+    }
+
+    for field in &["createdAt", "created_at", "time"] {
+        if let Some(ts) = obj.get(*field).and_then(|v| v.as_str()) {
+            return Some(ts.to_string());
+        }
+    }
+
+    None
+}
+
+/// Detect whether a compaction has occurred since the last check.
+///
+/// Compares the current compaction offset against a saved value in
+/// `state_dir/compaction-{session_id}`. Returns true if a new
+/// compaction was found. Updates the saved offset.
+pub fn detect_new_compaction(
+    state_dir: &Path,
+    session_id: &str,
+    transcript_path: &str,
+) -> bool {
+    let offset = find_last_compaction_in_file(transcript_path);
+
+    let save_path = state_dir.join(format!("compaction-{}", session_id));
+    let saved: Option<u64> = fs::read_to_string(&save_path)
+        .ok()
+        .and_then(|s| s.trim().parse().ok());
+
+    let is_new = match (offset, saved) {
+        (Some(cur), Some(prev)) => cur != prev,
+        (Some(_), None) => true,
+        _ => false,
+    };
+
+    // Save current offset
+    if let Some(off) = offset {
+        fs::write(&save_path, off.to_string()).ok();
+    }
+
+    is_new
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::io::Write;
+
+    fn write_temp_jsonl(content: &str) -> tempfile::NamedTempFile {
+        let mut file = tempfile::NamedTempFile::new().unwrap();
+        file.write_all(content.as_bytes()).unwrap();
+        file.flush().unwrap();
+        file
+    }
+
+    #[test]
+    fn tail_messages_yields_normalized_messages_newest_first() {
+        let file = write_temp_jsonl(
+            r#"{"timestamp":"2026-06-15T15:00:00.000Z","type":"user","message":{"content":"claude user"}}
+{"timestamp":"2026-06-15T15:00:01.000Z","type":"assistant","message":{"content":[{"type":"text","text":"claude assistant"}]}}
+{"timestamp":"2026-06-15T15:00:02.000Z","type":"event_msg","payload":{"type":"user_message","message":"codex user"}}
+{"timestamp":"2026-06-15T15:00:03.000Z","type":"event_msg","payload":{"type":"task_started"}}
+{"timestamp":"2026-06-15T15:00:04.000Z","type":"event_msg","payload":{"type":"agent_message","message":"codex assistant"}}
+"#,
+        );
+
+        let messages: Vec<_> = TailMessages::open(&file.path().to_string_lossy())
+            .unwrap()
+            .collect();
+
+        assert_eq!(messages.len(), 4);
+        assert_eq!(messages[0].text, "codex assistant");
+        assert_eq!(messages[1].text, "codex user");
+        assert_eq!(messages[2].text, "claude assistant");
+        assert_eq!(messages[3].text, "claude user");
+        assert!(messages[0].offset > messages[1].offset);
+    }
+
+    #[test]
+    fn detects_claude_and_codex_compactions() {
+        let claude = br#"{"timestamp":"2026-06-15T15:00:00.000Z","type":"user","message":{"content":"normal"}}
+{"timestamp":"2026-06-15T15:00:01.000Z","type":"user","message":{"content":"This session is being continued from a previous conversation."}}
+"#;
+        let codex = br#"{"timestamp":"2026-06-15T15:00:00.000Z","type":"event_msg","payload":{"type":"user_message","message":"normal"}}
+{"timestamp":"2026-06-15T15:00:01.000Z","type":"event_msg","payload":{"type":"context_compacted"}}
+"#;
+
+        assert!(find_last_compaction(claude).is_some());
+        assert!(find_last_compaction(codex).is_some());
+    }
+
+    #[test]
+    fn detect_new_compaction_tracks_offset_changes() {
+        let transcript = write_temp_jsonl(
+            r#"{"timestamp":"2026-06-15T15:00:00.000Z","type":"event_msg","payload":{"type":"context_compacted"}}
+"#,
+        );
+        let state = tempfile::tempdir().unwrap();
+
+        assert!(detect_new_compaction(
+            state.path(),
+            "session",
+            &transcript.path().to_string_lossy(),
+        ));
+        assert!(!detect_new_compaction(
+            state.path(),
+            "session",
+            &transcript.path().to_string_lossy(),
+        ));
+    }
+}
diff --git a/src/hippocampus/graph.rs b/src/hippocampus/graph.rs
index 0e7a20d..b8f3736 100644
--- a/src/hippocampus/graph.rs
+++ b/src/hippocampus/graph.rs
@@ -11,6 +11,23 @@ use crate::store::{Store, RelationType, StoreView};
 
 use serde::{Deserialize, Serialize};
 use std::collections::{HashMap, HashSet, VecDeque};
+use std::sync::{OnceLock, RwLock};
+
+const EXACT_CC_MAX_DEG: usize = 512;
+const APPROX_CC_PAIRS: u64 = 4096;
+const CC_CACHE_TTL_SECS: i64 = 15 * 60;
+
+#[derive(Clone, Copy)]
+struct CachedCc {
+	value: f32,
+	computed_at: i64,
+}
+
+static CC_CACHE: OnceLock<RwLock<HashMap<String, CachedCc>>> = OnceLock::new();
+
+fn cc_cache() -> &'static RwLock<HashMap<String, CachedCc>> {
+	CC_CACHE.get_or_init(|| RwLock::new(HashMap::new()))
+}
 
 /// Community info for reporting
 #[derive(Clone, Debug)]
@@ -32,11 +49,13 @@ pub struct Edge {
 
 /// The in-memory graph built from store nodes + relations
 pub struct Graph {
-    /// Adjacency list: node key → list of edges
-    adj: HashMap<String, Vec<Edge>>,
-    /// All node keys
-    keys: HashSet<String>,
-    /// Community labels (from label propagation)
+	/// Adjacency list: node key → list of edges
+	adj: HashMap<String, Vec<Edge>>,
+	/// Neighbor sets for membership tests in graph metrics.
+	neighbor_sets: HashMap<String, HashSet<String>>,
+	/// All node keys
+	keys: HashSet<String>,
+	/// Community labels (from label propagation)
     communities: HashMap<String, u32>,
 }
 
@@ -67,22 +86,22 @@ impl Graph {
             .unwrap_or_default()
     }
 
-    /// Just neighbor keys
-    pub fn neighbor_keys(&self, key: &str) -> HashSet<&str> {
-        self.adj.get(key)
-            .map(|edges| edges.iter().map(|e| e.target.as_str()).collect())
-            .unwrap_or_default()
-    }
+	/// Just neighbor keys
+	pub fn neighbor_keys(&self, key: &str) -> HashSet<&str> {
+		self.neighbor_sets.get(key)
+			.map(|neighbors| neighbors.iter().map(String::as_str).collect())
+			.unwrap_or_default()
+	}
 
-    /// Jaccard similarity between two nodes' neighborhoods.
-    /// Measures overlap: |intersection| / |union| of their neighbor sets.
-    pub fn jaccard(&self, a: &str, b: &str) -> f32 {
-        let na = self.neighbor_keys(a);
-        let nb = self.neighbor_keys(b);
-        let intersection = na.intersection(&nb).count();
-        let union = na.union(&nb).count();
-        if union == 0 { 0.0 } else { intersection as f32 / union as f32 }
-    }
+	/// Jaccard similarity between two nodes' neighborhoods.
+	/// Measures overlap: |intersection| / |union| of their neighbor sets.
+	pub fn jaccard(&self, a: &str, b: &str) -> f32 {
+		let Some(na) = self.neighbor_sets.get(a) else { return 0.0 };
+		let Some(nb) = self.neighbor_sets.get(b) else { return 0.0 };
+		let intersection = na.intersection(nb).count();
+		let union = na.len() + nb.len() - intersection;
+		if union == 0 { 0.0 } else { intersection as f32 / union as f32 }
+	}
 
     /// Compute Jaccard-based strength for every edge in the graph.
     /// Returns (source_key, target_key, jaccard_strength) triples.
@@ -202,41 +221,78 @@ impl Graph {
         }
     }
 
-    /// Local clustering coefficient: fraction of a node's neighbors
-    /// that are also neighbors of each other.
-    /// cc(v) = 2E / (deg * (deg - 1))
-    pub fn clustering_coefficient(&self, key: &str) -> f32 {
-        let neighbors = self.neighbor_keys(key);
-        let deg = neighbors.len();
-        if deg < 2 {
-            return 0.0;
-        }
+	/// Local clustering coefficient: fraction of a node's neighbors
+	/// that are also neighbors of each other.
+	/// cc(v) = 2E / (deg * (deg - 1))
+	pub fn clustering_coefficient(&self, key: &str) -> f32 {
+		let now = crate::store::now_epoch();
+		if let Some(cc) = cc_cache().read().unwrap().get(key).copied()
+			&& now - cc.computed_at < CC_CACHE_TTL_SECS
+		{
+			return cc.value;
+		}
+		let cc = self.clustering_coefficient_uncached(key);
+		cc_cache().write().unwrap().insert(key.to_owned(), CachedCc {
+			value: cc,
+			computed_at: now,
+		});
+		cc
+	}
 
-        let neighbor_vec: Vec<&str> = neighbors.iter().copied().collect();
-        let mut triangles = 0u32;
-        for i in 0..neighbor_vec.len() {
-            for j in (i + 1)..neighbor_vec.len() {
-                let ni_neighbors = self.neighbor_keys(neighbor_vec[i]);
-                if ni_neighbors.contains(neighbor_vec[j]) {
-                    triangles += 1;
-                }
-            }
-        }
+	fn clustering_coefficient_uncached(&self, key: &str) -> f32 {
+		let Some(neighbors) = self.neighbor_sets.get(key) else {
+			return 0.0;
+		};
+		let deg = neighbors.len();
+		if deg < 2 {
+			return 0.0;
+		}
 
-        (2.0 * triangles as f32) / (deg as f32 * (deg as f32 - 1.0))
-    }
+		let neighbor_vec: Vec<&str> = neighbors.iter().map(String::as_str).collect();
+		if deg <= EXACT_CC_MAX_DEG {
+			let mut linked = 0u64;
+			for i in 0..neighbor_vec.len() {
+				for j in (i + 1)..neighbor_vec.len() {
+					if self.neighbor_sets
+						.get(neighbor_vec[i])
+						.is_some_and(|n| n.contains(neighbor_vec[j])) {
+							linked += 1;
+					}
+				}
+			}
+			return (2.0 * linked as f32) / (deg as f32 * (deg as f32 - 1.0));
+		}
 
-    /// Average clustering coefficient across all nodes with deg >= 2
-    pub fn avg_clustering_coefficient(&self) -> f32 {
-        let mut sum = 0.0f32;
-        let mut count = 0u32;
-        for key in &self.keys {
-            if self.degree(key) >= 2 {
-                sum += self.clustering_coefficient(key);
-                count += 1;
-            }
-        }
-        if count == 0 { 0.0 } else { sum / count as f32 }
+		let mut linked = 0u64;
+		let samples = APPROX_CC_PAIRS.min((deg as u64 * (deg as u64 - 1)) / 2);
+		for sample in 0..samples {
+			let i = ((sample.wrapping_mul(1_103_515_245).wrapping_add(12_345)) % deg as u64) as usize;
+			let mut j = ((sample.wrapping_mul(2_654_435_761).wrapping_add(97_531)) % deg as u64) as usize;
+			if i == j {
+				j = (j + 1) % deg;
+			}
+			if self.neighbor_sets
+				.get(neighbor_vec[i])
+				.is_some_and(|n| n.contains(neighbor_vec[j])) {
+					linked += 1;
+			}
+		}
+		linked as f32 / samples as f32
+	}
+
+	/// Average clustering coefficient across all nodes with deg >= 2
+	pub fn avg_clustering_coefficient(&self) -> f32 {
+		let mut sum = 0.0f32;
+		let mut count = 0u32;
+		for key in &self.keys {
+			match self.neighbor_sets.get(key.as_str()) {
+				Some(s) if s.len() >= 2 => s,
+				_ => continue,
+			};
+			sum += self.clustering_coefficient(key);
+			count += 1;
+		}
+		if count == 0 { 0.0 } else { sum / count as f32 }
     }
 
     /// Average shortest path length (sampled BFS from up to 100 nodes)
@@ -266,15 +322,17 @@ impl Graph {
         dist.insert(start.to_string(), 0u32);
         queue.push_back(start.to_string());
 
-        while let Some(node) = queue.pop_front() {
-            let d = dist[&node];
-            for neighbor in self.neighbor_keys(&node) {
-                if !dist.contains_key(neighbor) {
-                    dist.insert(neighbor.to_string(), d + 1);
-                    queue.push_back(neighbor.to_string());
-                }
-            }
-        }
+		while let Some(node) = queue.pop_front() {
+			let d = dist[&node];
+			if let Some(neighbors) = self.neighbor_sets.get(&node) {
+				for neighbor in neighbors {
+					if !dist.contains_key(neighbor) {
+						dist.insert(neighbor.clone(), d + 1);
+						queue.push_back(neighbor.clone());
+					}
+				}
+			}
+		}
         dist
     }
 
@@ -505,16 +563,39 @@ impl Graph {
 
 /// Build graph from store data (with community detection)
 pub fn build_graph(store: &impl StoreView) -> Graph {
-    let (adj, keys) = build_adjacency(store);
-    let communities = label_propagation(&keys, &adj, 20);
-    Graph { adj, keys, communities }
+	let (adj, keys) = build_adjacency(store);
+	let neighbor_sets = build_neighbor_sets(&adj);
+	let communities = label_propagation(&keys, &adj, 20);
+	Graph {
+		adj,
+		neighbor_sets,
+		keys,
+		communities,
+	}
 }
 
 /// Build graph without community detection — for spreading activation
 /// searches where we only need the adjacency list.
 pub fn build_graph_fast(store: &impl StoreView) -> Graph {
-    let (adj, keys) = build_adjacency(store);
-    Graph { adj, keys, communities: HashMap::new() }
+	let (adj, keys) = build_adjacency(store);
+	let neighbor_sets = build_neighbor_sets(&adj);
+	Graph {
+		adj,
+		neighbor_sets,
+		keys,
+		communities: HashMap::new(),
+	}
+}
+
+fn build_neighbor_sets(adj: &HashMap<String, Vec<Edge>>) -> HashMap<String, HashSet<String>> {
+	adj.iter()
+		.map(|(key, edges)| {
+			let neighbors = edges.iter()
+				.map(|edge| edge.target.clone())
+				.collect();
+			(key.clone(), neighbors)
+		})
+		.collect()
 }
 
 fn build_adjacency(store: &impl StoreView) -> (HashMap<String, Vec<Edge>>, HashSet<String>) {
diff --git a/src/hippocampus/mod.rs b/src/hippocampus/mod.rs
index d79640d..0193004 100644
--- a/src/hippocampus/mod.rs
+++ b/src/hippocampus/mod.rs
@@ -17,7 +17,6 @@ pub mod query;
 pub mod spectral;
 pub mod neuro;
 pub mod counters;
-pub mod transcript;
 
 use std::cell::RefCell;
 use std::path::PathBuf;
diff --git a/src/hippocampus/neuro/scoring.rs b/src/hippocampus/neuro/scoring.rs
index 5828fd0..c9cbb40 100644
--- a/src/hippocampus/neuro/scoring.rs
+++ b/src/hippocampus/neuro/scoring.rs
@@ -230,10 +230,6 @@ fn consolidation_plan_inner(store: &Store, _detect_interf: bool) -> Consolidatio
         rationale: Vec::new(),
     };
 
-    // Active agent types from config
-    let config = crate::config::get();
-    let agent_types: Vec<&str> = config.agent_types.iter().map(|s| s.as_str()).collect();
-
     // Target: α ≥ 2.5 (healthy scale-free)
     if alpha < 2.0 {
         plan.add("linker", 100);
@@ -274,48 +270,6 @@ fn consolidation_plan_inner(store: &Store, _detect_interf: bool) -> Consolidatio
     // Split: handle oversized nodes
     plan.set("split", 5);
 
-    // Distribute agent budget using Elo ratings
-    let budget = crate::config::get().agent_budget;
-    let elo_path = crate::config::get().data_dir.join("agent-elo.json");
-    if let Ok(elo_json) = std::fs::read_to_string(&elo_path) {
-        if let Ok(ratings) = serde_json::from_str::<std::collections::HashMap<String, f64>>(&elo_json) {
-            let elos: Vec<f64> = agent_types.iter()
-                .map(|t| ratings.get(*t).copied().unwrap_or(1000.0))
-                .collect();
-            let min_elo = elos.iter().copied().fold(f64::MAX, f64::min);
-
-            let weights: Vec<f64> = elos.iter()
-                .map(|e| {
-                    let shifted = e - min_elo + 50.0;
-                    shifted * shifted
-                })
-                .collect();
-            let total_weight: f64 = weights.iter().sum();
-
-            let allocate = |w: f64| -> usize {
-                ((w / total_weight * budget as f64).round() as usize).max(2)
-            };
-
-            for (i, agent) in agent_types.iter().enumerate() {
-                plan.set(agent, allocate(weights[i]));
-            }
-
-            let summary: Vec<String> = agent_types.iter()
-                .map(|a| format!("{}={}", a, plan.count(a)))
-                .collect();
-            plan.rationale.push(format!(
-                "Elo allocation (budget={}): {}", budget, summary.join(" ")));
-        }
-    } else {
-        // No Elo file — use budget with equal distribution
-        let per_type = budget / agent_types.len();
-        for agent in &agent_types {
-            plan.set(agent, per_type);
-        }
-        plan.rationale.push(format!(
-            "No Elo ratings — equal distribution ({} each, budget={})", per_type, budget));
-    }
-
     plan
 }
 
diff --git a/src/hippocampus/transcript.rs b/src/hippocampus/transcript.rs
deleted file mode 100644
index 4fd1ef2..0000000
--- a/src/hippocampus/transcript.rs
+++ /dev/null
@@ -1,340 +0,0 @@
-// Transcript JSONL parsing utilities.
-//
-// Provides mmap-based backward scanning of Claude Code transcript files
-// and compaction detection. Used by memory-search (hook mode) and
-// parse-claude-conversation (debug tool).
-
-use memchr::memrchr3;
-use memmap2::Mmap;
-use serde_json::Value;
-use std::fs;
-use std::path::Path;
-
-/// Scan backwards through mmap'd bytes, yielding byte slices of complete
-/// top-level JSON objects (outermost { to matching }).
-///
-/// Uses memrchr3 (SIMD) to jump between structurally significant bytes
-/// ({, }, ") instead of scanning byte-by-byte. Tracks brace depth,
-/// skipping braces inside JSON strings. Returns objects in reverse order
-/// (newest first).
-pub struct JsonlBackwardIter<'a> {
-    data: &'a [u8],
-    pos: usize,
-}
-
-impl<'a> JsonlBackwardIter<'a> {
-    pub fn new(data: &'a [u8]) -> Self {
-        Self { data, pos: data.len() }
-    }
-}
-
-impl<'a> Iterator for JsonlBackwardIter<'a> {
-    type Item = &'a [u8];
-
-    fn next(&mut self) -> Option<Self::Item> {
-        // Find the closing } of the next object, skipping } inside strings
-        let close = {
-            let mut in_string = false;
-            loop {
-                let p = memrchr3(b'{', b'}', b'"', &self.data[..self.pos])?;
-                self.pos = p;
-                let ch = self.data[p];
-
-                if in_string {
-                    if ch == b'"' {
-                        let mut bs = 0;
-                        while p > bs + 1 && self.data[p - 1 - bs] == b'\\' {
-                            bs += 1;
-                        }
-                        if bs % 2 == 0 { in_string = false; }
-                    }
-                    continue;
-                }
-
-                match ch {
-                    b'}' => break p,
-                    b'"' => in_string = true,
-                    _ => {}
-                }
-            }
-        };
-
-        // Track brace depth to find matching {
-        let mut depth: usize = 1;
-        let mut in_string = false;
-
-        loop {
-            let p = memrchr3(b'{', b'}', b'"', &self.data[..self.pos])?;
-            self.pos = p;
-            let ch = self.data[p];
-
-            if in_string {
-                if ch == b'"' {
-                    // Check for escaped quote (count preceding backslashes)
-                    let mut bs = 0;
-                    while p > bs + 1 && self.data[p - 1 - bs] == b'\\' {
-                        bs += 1;
-                    }
-                    if bs % 2 == 0 {
-                        in_string = false;
-                    }
-                }
-                // { and } inside strings don't affect depth
-                continue;
-            }
-
-            match ch {
-                b'"' => { in_string = true; }
-                b'}' => { depth += 1; }
-                b'{' => {
-                    depth -= 1;
-                    if depth == 0 {
-                        return Some(&self.data[self.pos..=close]);
-                    }
-                }
-                _ => {}
-            }
-        }
-    }
-}
-
-/// Find the byte offset of the last compaction summary in mmap'd transcript data.
-///
-/// Scans backward for a user-type message whose content starts with
-/// "This session is being continued". Returns the byte offset of the
-/// JSON object's opening brace.
-pub(crate) fn find_last_compaction(data: &[u8]) -> Option<usize> {
-    let marker = b"This session is being continued";
-
-    for obj_bytes in JsonlBackwardIter::new(data) {
-        // Quick byte check before parsing
-        if !contains_bytes(obj_bytes, marker) {
-            continue;
-        }
-
-        let obj: Value = match serde_json::from_slice(obj_bytes) {
-            Ok(v) => v,
-            Err(_) => continue,
-        };
-
-        if obj.get("type").and_then(|v| v.as_str()) != Some("user") {
-            continue;
-        }
-
-        if let Some(content) = obj.get("message")
-            .and_then(|m| m.get("content"))
-            .and_then(|c| c.as_str())
-            && content.starts_with("This session is being continued") {
-                let offset = obj_bytes.as_ptr() as usize - data.as_ptr() as usize;
-                return Some(offset);
-            }
-    }
-
-    None
-}
-
-/// Find the byte offset of the last compaction in a transcript file.
-/// Returns None if the file can't be opened or has no compaction.
-pub(crate) fn find_last_compaction_in_file(path: &str) -> Option<u64> {
-    if path.is_empty() { return None; }
-
-    let file = fs::File::open(path).ok()?;
-    let meta = file.metadata().ok()?;
-    if meta.len() == 0 { return None; }
-
-    let mmap = unsafe { Mmap::map(&file).ok()? };
-    find_last_compaction(&mmap).map(|off| off as u64)
-}
-
-/// Mmap a transcript file. Returns (Mmap, File) to keep both alive.
-pub(crate) fn mmap_transcript(path: &str) -> Option<(Mmap, fs::File)> {
-    let file = fs::File::open(path).ok()?;
-    let meta = file.metadata().ok()?;
-    if meta.len() == 0 { return None; }
-    let mmap = unsafe { Mmap::map(&file).ok()? };
-    Some((mmap, file))
-}
-
-fn contains_bytes(haystack: &[u8], needle: &[u8]) -> bool {
-    haystack.windows(needle.len()).any(|w| w == needle)
-}
-
-/// Reverse iterator over user/assistant messages in a transcript file.
-/// Yields (role, text, timestamp) tuples newest-first. The caller decides
-/// when to stop (byte budget, count, etc).
-pub struct TailMessages {
-    _file: fs::File,
-    mmap: Mmap,
-    pos: usize,
-}
-
-impl TailMessages {
-    pub fn open(path: &str) -> Option<Self> {
-        let (mmap, file) = mmap_transcript(path)?;
-        let pos = mmap.len();
-        Some(Self { _file: file, mmap, pos })
-    }
-}
-
-impl Iterator for TailMessages {
-    type Item = (String, String, String);
-
-    fn next(&mut self) -> Option<Self::Item> {
-        loop {
-            // Find closing }, skipping } inside strings
-            let close = {
-                let mut in_string = false;
-                loop {
-                    let p = memrchr3(b'{', b'}', b'"', &self.mmap[..self.pos])?;
-                    self.pos = p;
-                    let ch = self.mmap[p];
-
-                    if in_string {
-                        if ch == b'"' {
-                            let mut bs = 0;
-                            while p > bs + 1 && self.mmap[p - 1 - bs] == b'\\' {
-                                bs += 1;
-                            }
-                            if bs % 2 == 0 { in_string = false; }
-                        }
-                        continue;
-                    }
-
-                    match ch {
-                        b'}' => break p,
-                        b'"' => in_string = true,
-                        _ => {}
-                    }
-                }
-            };
-
-            // Track brace depth to find matching {
-            let mut depth: usize = 1;
-            let mut in_string = false;
-            let open = loop {
-                let p = memrchr3(b'{', b'}', b'"', &self.mmap[..self.pos])?;
-                self.pos = p;
-                let ch = self.mmap[p];
-
-                if in_string {
-                    if ch == b'"' {
-                        let mut bs = 0;
-                        while p > bs + 1 && self.mmap[p - 1 - bs] == b'\\' {
-                            bs += 1;
-                        }
-                        if bs % 2 == 0 { in_string = false; }
-                    }
-                    continue;
-                }
-
-                match ch {
-                    b'"' => { in_string = true; }
-                    b'}' => { depth += 1; }
-                    b'{' => {
-                        depth -= 1;
-                        if depth == 0 { break p; }
-                    }
-                    _ => {}
-                }
-            };
-
-            let obj_bytes = &self.mmap[open..=close];
-
-            // The "type" field is near the start of top-level objects.
-            // Only check the first 200 bytes to avoid scanning megabyte objects.
-            let prefix = &obj_bytes[..obj_bytes.len().min(200)];
-            let is_user = memchr::memmem::find(prefix, b"\"type\":\"user\"").is_some();
-            let is_assistant = !is_user
-                && memchr::memmem::find(prefix, b"\"type\":\"assistant\"").is_some();
-            if !is_user && !is_assistant { continue; }
-
-            let obj: Value = match serde_json::from_slice(obj_bytes) {
-                Ok(v) => v,
-                Err(_) => continue,
-            };
-
-            let msg_type = if is_user { "user" } else { "assistant" };
-
-            let msg = obj.get("message").unwrap_or(&obj);
-            let text = match msg.get("content") {
-                Some(Value::String(s)) => s.clone(),
-                Some(Value::Array(arr)) => {
-                    arr.iter()
-                        .filter(|b| b.get("type").and_then(|v| v.as_str()) == Some("text"))
-                        .filter_map(|b| b.get("text").and_then(|v| v.as_str()))
-                        .collect::<Vec<_>>()
-                        .join(" ")
-                }
-                _ => continue,
-            };
-            if text.is_empty() { continue; }
-
-            let timestamp = obj.get("timestamp")
-                .and_then(|v| v.as_str())
-                .unwrap_or("")
-                .to_string();
-
-            return Some((msg_type.to_string(), text, timestamp));
-        }
-    }
-}
-
-/// Get the timestamp of the compaction message at a given byte offset.
-/// Returns a human-readable datetime string, or None if unavailable.
-pub fn compaction_timestamp(path: &str, offset: u64) -> Option<String> {
-    let (mmap, _file) = mmap_transcript(path)?;
-    let start = offset as usize;
-    if start >= mmap.len() { return None; }
-
-    // Find the end of this JSONL line
-    let end = mmap[start..].iter().position(|&b| b == b'\n')
-        .map(|p| start + p)
-        .unwrap_or(mmap.len());
-
-    let obj: Value = serde_json::from_slice(&mmap[start..end]).ok()?;
-
-    // Claude Code transcript entries have a "timestamp" field (ISO 8601)
-    if let Some(ts) = obj.get("timestamp").and_then(|v| v.as_str()) {
-        return Some(ts.to_string());
-    }
-
-    // Fallback: try "createdAt" or similar fields
-    for field in &["createdAt", "created_at", "time"] {
-        if let Some(ts) = obj.get(*field).and_then(|v| v.as_str()) {
-            return Some(ts.to_string());
-        }
-    }
-
-    None
-}
-
-/// Detect whether a compaction has occurred since the last check.
-///
-/// Compares the current compaction offset against a saved value in
-/// `state_dir/compaction-{session_id}`. Returns true if a new
-/// compaction was found. Updates the saved offset.
-pub fn detect_new_compaction(
-    state_dir: &Path,
-    session_id: &str,
-    transcript_path: &str,
-) -> bool {
-    let offset = find_last_compaction_in_file(transcript_path);
-
-    let save_path = state_dir.join(format!("compaction-{}", session_id));
-    let saved: Option<u64> = fs::read_to_string(&save_path)
-        .ok()
-        .and_then(|s| s.trim().parse().ok());
-
-    let is_new = match (offset, saved) {
-        (Some(cur), Some(prev)) => cur != prev,
-        (Some(_), None) => true,
-        _ => false,
-    };
-
-    // Save current offset
-    if let Some(off) = offset {
-        fs::write(&save_path, off.to_string()).ok();
-    }
-
-    is_new
-}
diff --git a/src/lib.rs b/src/lib.rs
index 1a71735..4dc5a12 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -1,4 +1,4 @@
-#![feature(async_fn_track_caller)]
+#![cfg_attr(feature = "nightly-diagnostics", feature(async_fn_track_caller))]
 
 // consciousness — unified crate for memory, agents, and subconscious processes
 //
@@ -25,6 +25,9 @@ macro_rules! dbglog {
     }};
 }
 
+// Logging (target-routed file logger)
+pub mod logging;
+
 // User interface (TUI, CLI)
 pub mod user;
 
@@ -40,8 +43,12 @@ pub mod hippocampus;
 // Autonomous agents
 pub mod subconscious;
 
+// Conversation transcript abstraction and compatibility sources
+pub mod conversation;
+
 // Unified configuration
 pub mod config;
+pub mod config_writer;
 
 // Session state
 pub mod session;
@@ -87,7 +94,8 @@ pub mod channel_capnp {
 pub use hippocampus::{
     store, graph, lookups, query,
     spectral, neuro, counters,
-    transcript, memory,
+    memory,
 };
+pub use conversation as transcript;
 use hippocampus::query::engine as search;
 use hippocampus::query::parser as query_parser;
diff --git a/src/locks.rs b/src/locks.rs
index dda4cb2..fe02434 100644
--- a/src/locks.rs
+++ b/src/locks.rs
@@ -114,7 +114,7 @@ impl<T> TrackedMutex<T> {
         Self { inner: Mutex::new(value) }
     }
 
-    #[track_caller]
+    #[cfg_attr(feature = "nightly-diagnostics", track_caller)]
     pub async fn lock(&self) -> TrackedMutexGuard<'_, T> {
         let location = Location::caller();
         let guard = self.inner.lock().await;
@@ -125,7 +125,7 @@ impl<T> TrackedMutex<T> {
         }
     }
 
-    #[track_caller]
+    #[cfg_attr(feature = "nightly-diagnostics", track_caller)]
     pub fn try_lock(&self) -> Result<TrackedMutexGuard<'_, T>, tokio::sync::TryLockError> {
         let location = Location::caller();
         let guard = self.inner.try_lock()?;
@@ -171,7 +171,7 @@ impl<T> TrackedRwLock<T> {
         Self { inner: RwLock::new(value) }
     }
 
-    #[track_caller]
+    #[cfg_attr(feature = "nightly-diagnostics", track_caller)]
     pub async fn read(&self) -> TrackedRwLockReadGuard<'_, T> {
         let location = Location::caller();
         let guard = self.inner.read().await;
@@ -182,7 +182,7 @@ impl<T> TrackedRwLock<T> {
         }
     }
 
-    #[track_caller]
+    #[cfg_attr(feature = "nightly-diagnostics", track_caller)]
     pub async fn write(&self) -> TrackedRwLockWriteGuard<'_, T> {
         let location = Location::caller();
         let guard = self.inner.write().await;
diff --git a/src/logging.rs b/src/logging.rs
new file mode 100644
index 0000000..3c9d080
--- /dev/null
+++ b/src/logging.rs
@@ -0,0 +1,146 @@
+// logging.rs — log-crate logger that routes by target.
+//
+// Records with target "grpc" (or any target starting with "grpc::") go
+// to ~/.consciousness/logs/daemon/grpc.log so we can tell gRPC events
+// apart from the rest of consciousness's noise. Everything else goes
+// to ~/.consciousness/logs/daemon/debug.log.
+//
+// Level threshold is taken from RUST_LOG (simple global level parse:
+// "trace"/"debug"/"info"/"warn"/"error"); defaults to "info".
+
+use std::io::Write;
+use std::path::PathBuf;
+use std::sync::Mutex;
+
+use log::{Level, LevelFilter, Log, Metadata, Record, SetLoggerError};
+
+fn logs_dir() -> PathBuf {
+    dirs::home_dir().unwrap_or_default().join(".consciousness/logs/daemon")
+}
+
+struct RoutingLogger {
+    grpc_file: Mutex<Option<std::fs::File>>,
+    debug_file: Mutex<Option<std::fs::File>>,
+    level: LevelFilter,
+}
+
+impl RoutingLogger {
+    fn new(level: LevelFilter) -> Self {
+        let dir = logs_dir();
+        let _ = std::fs::create_dir_all(&dir);
+        let grpc = std::fs::OpenOptions::new()
+            .create(true).append(true)
+            .open(dir.join("grpc.log")).ok();
+        let debug = std::fs::OpenOptions::new()
+            .create(true).append(true)
+            .open(dir.join("debug.log")).ok();
+        Self {
+            grpc_file: Mutex::new(grpc),
+            debug_file: Mutex::new(debug),
+            level,
+        }
+    }
+
+    fn is_grpc_target(target: &str) -> bool {
+        target == "grpc" || target.starts_with("grpc::")
+    }
+}
+
+impl Log for RoutingLogger {
+    fn enabled(&self, m: &Metadata) -> bool {
+        // Always enable DEBUG for grpc target so the dedicated log is
+        // actually useful without RUST_LOG wrangling; defer to the
+        // configured level for everything else.
+        if Self::is_grpc_target(m.target()) {
+            return m.level() <= Level::Debug;
+        }
+        m.level() <= self.level
+    }
+
+    fn log(&self, record: &Record) {
+        if !self.enabled(record.metadata()) {
+            return;
+        }
+        let line = format!(
+            "[{}] [{}] [{}] {}\n",
+            chrono::Utc::now().format("%Y-%m-%d %H:%M:%S%.3f"),
+            record.level(),
+            record.target(),
+            record.args(),
+        );
+        let slot = if Self::is_grpc_target(record.target()) {
+            &self.grpc_file
+        } else {
+            &self.debug_file
+        };
+        if let Ok(mut guard) = slot.lock() {
+            if let Some(ref mut f) = *guard {
+                let _ = f.write_all(line.as_bytes());
+            }
+        }
+    }
+
+    fn flush(&self) {
+        for slot in [&self.grpc_file, &self.debug_file] {
+            if let Ok(mut g) = slot.lock() {
+                if let Some(ref mut f) = *g {
+                    let _ = f.flush();
+                }
+            }
+        }
+    }
+}
+
+fn parse_level_from_env() -> LevelFilter {
+    let raw = std::env::var("RUST_LOG").unwrap_or_else(|_| "info".to_string());
+    // Parse a plain level word; if it's the module=level form, we take
+    // the first level we find.
+    let token = raw.split(',').next().unwrap_or("info");
+    let level_word = token.rsplit_once('=').map(|(_, v)| v).unwrap_or(token);
+    match level_word.trim().to_lowercase().as_str() {
+        "trace" => LevelFilter::Trace,
+        "debug" => LevelFilter::Debug,
+        "info"  => LevelFilter::Info,
+        "warn"  => LevelFilter::Warn,
+        "error" => LevelFilter::Error,
+        "off"   => LevelFilter::Off,
+        _ => LevelFilter::Info,
+    }
+}
+
+/// Install the routing logger. Safe to call at most once — subsequent
+/// calls return an error but are otherwise no-ops.
+pub fn init() -> Result<(), SetLoggerError> {
+    let level = parse_level_from_env();
+    let logger = Box::new(RoutingLogger::new(level));
+    log::set_boxed_logger(logger)?;
+    // Always let DEBUG records through globally so the grpc log can
+    // capture them (the logger itself filters non-grpc targets by
+    // `level`). The cost is that log::debug! call-sites below `level`
+    // in other modules still do their arg formatting before being
+    // dropped at the logger; acceptable for a debug tool.
+    log::set_max_level(LevelFilter::Debug.max(level));
+    // Mark the file with a session boundary so it's easy to see where a
+    // restart happened.
+    log::info!(
+        "===== consciousness logger init (level={}, pid={}) =====",
+        level, std::process::id(),
+    );
+    log::info!(target: "grpc",
+        "===== grpc log init (level={}, pid={}) =====",
+        level, std::process::id(),
+    );
+    Ok(())
+}
+
+/// Consumer of &Level so the type is used when only some callers want it.
+#[allow(dead_code)]
+pub fn current_level() -> Level {
+    match log::max_level() {
+        LevelFilter::Trace => Level::Trace,
+        LevelFilter::Debug => Level::Debug,
+        LevelFilter::Info | LevelFilter::Off => Level::Info,
+        LevelFilter::Warn => Level::Warn,
+        LevelFilter::Error => Level::Error,
+    }
+}
diff --git a/src/main.rs b/src/main.rs
index 78bfa4f..1fdc901 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -1,4 +1,4 @@
-#![feature(panic_backtrace_config)]
+#![cfg_attr(feature = "nightly-diagnostics", feature(panic_backtrace_config))]
 
 // poc-memory: graph-structured memory for AI assistants
 //
@@ -333,6 +333,18 @@ enum AdminCmd {
         #[arg(long)]
         stats: bool,
     },
+    /// Print normalized user/assistant messages from a transcript JSONL file
+    #[command(name = "transcript-tail")]
+    TranscriptTail {
+        /// Transcript JSONL path
+        path: String,
+        /// Maximum number of messages to print
+        #[arg(long, short = 'n', default_value_t = 40)]
+        count: usize,
+        /// Print newest messages first instead of chronological order
+        #[arg(long)]
+        newest_first: bool,
+    },
 }
 
 /// Print help with subcommands expanded to show nested commands.
@@ -458,12 +470,15 @@ impl Run for AdminCmd {
             Self::Dedup { apply } => cli::admin::cmd_dedup(apply).await,
             Self::DailyCheck    => cli::admin::cmd_daily_check().await,
             Self::LoadContext { stats } => cli::node::cmd_load_context(stats).await,
+            Self::TranscriptTail { path, count, newest_first }
+                => cli::admin::cmd_transcript_tail(&path, count, newest_first),
         }
     }
 }
 
 #[tokio::main]
 async fn main() {
+    #[cfg(feature = "nightly-diagnostics")]
     std::panic::set_backtrace_style(std::panic::BacktraceStyle::Short);
 
     // Handle --help ourselves for expanded subcommand display
@@ -482,9 +497,16 @@ async fn main() {
 
     let cli = Cli::parse();
 
+    // Some subcommands (e.g. admin load-context) read from the global
+    // AppConfig. poc-memory has no config CLI flags of its own, so load
+    // with defaults — figment still pulls from ~/.consciousness/config.json5
+    // and env the same way.
+    if let Err(e) = crate::config::load_app(&crate::user::CliArgs::default()) {
+        eprintln!("warning: failed to load config: {:#}", e);
+    }
+
     if let Err(e) = cli.command.run().await {
         eprintln!("Error: {}", e);
         process::exit(1);
     }
 }
-
diff --git a/src/mind/log.rs b/src/mind/log.rs
index b69f2ca..51f294e 100644
--- a/src/mind/log.rs
+++ b/src/mind/log.rs
@@ -3,7 +3,7 @@ use std::fs::{File, OpenOptions};
 use std::io::Write;
 use std::path::{Path, PathBuf};
 use crate::agent::context::AstNode;
-use crate::hippocampus::transcript::JsonlBackwardIter;
+use crate::conversation::JsonlBackwardIter;
 use memmap2::Mmap;
 
 pub struct ConversationLog {
@@ -55,17 +55,13 @@ impl ConversationLog {
     }
 
     pub fn oldest_timestamp(&self) -> Option<chrono::DateTime<chrono::Utc>> {
-        // Read forward from the start to find first timestamp
         let file = File::open(&self.path).ok()?;
         let mmap = unsafe { Mmap::map(&file).ok()? };
-        // Find first { ... } and parse
         for line in mmap.split(|&b| b == b'\n') {
             if line.is_empty() { continue; }
             if let Ok(node) = serde_json::from_slice::<AstNode>(line) {
                 if let Some(leaf) = node.leaf() {
-                    if let Some(ts) = leaf.timestamp() {
-                        return Some(ts);
-                    }
+                    return Some(leaf.timestamp());
                 }
             }
         }
@@ -82,6 +78,6 @@ pub struct TailNodes {
 impl TailNodes {
     pub fn iter(&self) -> impl Iterator<Item = AstNode> + '_ {
         JsonlBackwardIter::new(&self.mmap)
-            .filter_map(|bytes| serde_json::from_slice::<AstNode>(bytes).ok())
+            .filter_map(|(_, bytes)| serde_json::from_slice::<AstNode>(bytes).ok())
     }
 }
diff --git a/src/mind/mod.rs b/src/mind/mod.rs
index 9fcc101..9572272 100644
--- a/src/mind/mod.rs
+++ b/src/mind/mod.rs
@@ -9,6 +9,44 @@ pub mod unconscious;
 pub mod identity;
 pub mod log;
 
+/// A background operation wired off Mind. Each flow (memory scoring,
+/// finetune scoring, compare) is a struct holding its dependencies and
+/// a TaskHandle; `trigger()` picks the flow's own "start a fresh run"
+/// semantics (abort-restart vs no-op-if-running).
+pub trait MindTriggered {
+    fn trigger(&self);
+}
+
+/// Owns a JoinHandle for a background task with two trigger semantics.
+/// Uses a sync Mutex for interior mutability so callers can `trigger()`
+/// off `&self` (Mind is shared via Arc).
+#[derive(Default)]
+pub struct TaskHandle(std::sync::Mutex<Option<tokio::task::JoinHandle<()>>>);
+
+impl TaskHandle {
+    pub fn new() -> Self { Self::default() }
+
+    /// Abort any running task and start a fresh one.
+    pub fn trigger<F>(&self, fut: F)
+    where F: std::future::Future<Output = ()> + Send + 'static
+    {
+        let mut h = self.0.lock().unwrap();
+        if let Some(old) = h.take() { old.abort(); }
+        *h = Some(tokio::spawn(fut));
+    }
+
+    /// No-op if a task is still running; otherwise start a fresh one.
+    pub fn trigger_if_idle<F>(&self, fut: F)
+    where F: std::future::Future<Output = ()> + Send + 'static
+    {
+        let mut h = self.0.lock().unwrap();
+        if let Some(old) = &*h {
+            if !old.is_finished() { return; }
+        }
+        *h = Some(tokio::spawn(fut));
+    }
+}
+
 // consciousness.rs — Mind state machine and event loop
 //
 // The core runtime for the consciousness binary. Mind manages turns,
@@ -25,7 +63,7 @@ use tokio::sync::mpsc;
 use crate::agent::{Agent, TurnResult};
 use crate::agent::api::ApiClient;
 use crate::config::{AppConfig, SessionConfig};
-use crate::subconscious::learn;
+use crate::subconscious::{compare, learn};
 use crate::hippocampus::access_local;
 
 pub use subconscious::{SubconsciousSnapshot, Subconscious};
@@ -33,6 +71,36 @@ pub use unconscious::{UnconsciousSnapshot, Unconscious};
 
 use crate::agent::context::{AstNode, NodeBody, Section, Ast, ContextState};
 
+fn match_scores(
+    nodes: &[AstNode],
+    scores: &std::collections::BTreeMap<String, f64>,
+) -> Vec<(usize, f64)> {
+    nodes.iter().enumerate()
+        .filter_map(|(i, node)| {
+            if let AstNode::Leaf(leaf) = node {
+                if let NodeBody::Memory { key, .. } = leaf.body() {
+                    return scores.get(key.as_str()).map(|&s| (i, s));
+                }
+            }
+            None
+        }).collect()
+}
+
+pub(crate) fn find_memory_by_key(ctx: &ContextState, key: &str) -> Option<(Section, usize)> {
+    [(Section::Identity, ctx.identity()), (Section::Conversation, ctx.conversation())]
+        .into_iter()
+        .find_map(|(section, nodes)| {
+            nodes.iter().enumerate().find_map(|(i, node)| {
+                if let AstNode::Leaf(leaf) = node {
+                    if let NodeBody::Memory { key: k, .. } = leaf.body() {
+                        if k == key { return Some((section, i)); }
+                    }
+                }
+                None
+            })
+        })
+}
+
 fn load_memory_scores(ctx: &mut ContextState, path: &std::path::Path) {
     let data = match std::fs::read_to_string(path) {
         Ok(d) => d,
@@ -42,25 +110,24 @@ fn load_memory_scores(ctx: &mut ContextState, path: &std::path::Path) {
         Ok(s) => s,
         Err(_) => return,
     };
-    let mut applied = 0;
-    for i in 0..ctx.conversation().len() {
-        if let AstNode::Leaf(leaf) = &ctx.conversation()[i] {
-            if let NodeBody::Memory { key, .. } = leaf.body() {
-                if let Some(&s) = scores.get(key.as_str()) {
-                    ctx.set_score(Section::Conversation, i, Some(s));
-                    applied += 1;
-                }
-            }
-        }
+    let identity_scores = match_scores(ctx.identity(), &scores);
+    let conv_scores = match_scores(ctx.conversation(), &scores);
+    let applied = identity_scores.len() + conv_scores.len();
+    for (i, s) in identity_scores {
+        ctx.set_score(Section::Identity, i, Some(s));
+    }
+    for (i, s) in conv_scores {
+        ctx.set_score(Section::Conversation, i, Some(s));
     }
     if applied > 0 {
         dbglog!("[scoring] loaded {} scores from {}", applied, path.display());
     }
 }
 
-/// Collect scored memory keys from conversation entries.
-fn collect_memory_scores(ctx: &ContextState) -> std::collections::BTreeMap<String, f64> {
-    ctx.conversation().iter()
+/// Collect scored memory keys from identity and conversation entries.
+pub(crate) fn collect_memory_scores(ctx: &ContextState) -> std::collections::BTreeMap<String, f64> {
+    ctx.identity().iter()
+        .chain(ctx.conversation().iter())
         .filter_map(|node| {
             if let AstNode::Leaf(leaf) = node {
                 if let NodeBody::Memory { key, score: Some(s), .. } = leaf.body() {
@@ -73,10 +140,14 @@ fn collect_memory_scores(ctx: &ContextState) -> std::collections::BTreeMap<Strin
 }
 
 /// Save memory scores to disk.
-fn save_memory_scores(scores: &std::collections::BTreeMap<String, f64>, path: &std::path::Path) {
-    if let Ok(json) = serde_json::to_string_pretty(scores) {
-        let _ = std::fs::write(path, json);
-        dbglog!("[scoring] saved {} scores to {}", scores.len(), path.display());
+pub(crate) fn save_memory_scores(scores: &std::collections::BTreeMap<String, f64>, path: &std::path::Path) {
+    match serde_json::to_string_pretty(scores) {
+        Ok(json) => match std::fs::write(path, &json) {
+            Ok(()) => dbglog!("[scoring] saved {} scores to {} ({} bytes)",
+                scores.len(), path.display(), json.len()),
+            Err(e) => dbglog!("[scoring] save FAILED ({}): {}", path.display(), e),
+        },
+        Err(e) => dbglog!("[scoring] serialize FAILED: {}", e),
     }
 }
 
@@ -118,6 +189,15 @@ pub struct MindState {
     pub unc_idle: bool,
     /// When the unconscious idle timer will fire (for UI display).
     pub unc_idle_deadline: Instant,
+    /// Fine-tuning candidates identified by scoring.
+    pub finetune_candidates: Vec<learn::FinetuneCandidate>,
+    /// Last scoring run stats for UI display.
+    pub finetune_last_run: Option<learn::FinetuneScoringStats>,
+    /// F7 compare candidates — one per response, showing what the test
+    /// model would say given the same context.
+    pub compare_candidates: Vec<compare::CompareCandidate>,
+    /// F7 compare error from the last run, if any.
+    pub compare_error: Option<String>,
 }
 
 impl Clone for MindState {
@@ -136,6 +216,10 @@ impl Clone for MindState {
             turn_handle: None, // Not cloned — only Mind's loop uses this
             unc_idle: self.unc_idle,
             unc_idle_deadline: self.unc_idle_deadline,
+            finetune_candidates: self.finetune_candidates.clone(),
+            finetune_last_run: self.finetune_last_run.clone(),
+            compare_candidates: self.compare_candidates.clone(),
+            compare_error: self.compare_error.clone(),
         }
     }
 }
@@ -148,6 +232,15 @@ pub enum MindCommand {
     Score,
     /// Run full N×M memory scoring matrix (/score command)
     ScoreFull,
+    /// Score for finetune candidates
+    ScoreFinetune,
+    /// Run F7 compare: generate alternates with the configured test model
+    /// for every assistant response in the context.
+    Compare,
+    /// Update the finetune divergence threshold and persist to config.
+    SetLearnThreshold(f64),
+    /// Toggle alternate-response generation during scoring; persist to config.
+    SetLearnGenerateAlternates(bool),
     /// Abort current turn, kill processes
     Interrupt,
     /// Reset session
@@ -173,6 +266,10 @@ impl MindState {
             turn_handle: None,
             unc_idle: false,
             unc_idle_deadline: Instant::now() + std::time::Duration::from_secs(60),
+            finetune_candidates: Vec::new(),
+            finetune_last_run: None,
+            compare_candidates: Vec::new(),
+            compare_error: None,
         }
     }
 
@@ -229,7 +326,7 @@ impl MindState {
     }
 
     /// DMN tick — returns a prompt and target if we should run a turn.
-    fn dmn_tick(&mut self) -> Option<(String, StreamTarget)> {
+    fn _dmn_tick(&mut self) -> Option<(String, StreamTarget)> {
         if matches!(self.dmn, subconscious::State::Paused | subconscious::State::Off) {
             return None;
         }
@@ -256,10 +353,6 @@ impl MindState {
     }
 }
 
-/// Background task completion events.
-enum BgEvent {
-    ScoringDone,
-}
 
 // --- Mind: cognitive state machine ---
 
@@ -276,8 +369,9 @@ pub struct Mind {
     /// Signals conscious activity to the unconscious loop.
     /// true = active, false = idle opportunity.
     conscious_active: tokio::sync::watch::Sender<bool>,
-    bg_tx: mpsc::UnboundedSender<BgEvent>,
-    bg_rx: std::sync::Mutex<Option<mpsc::UnboundedReceiver<BgEvent>>>,
+    memory_scoring: learn::MemoryScoring,
+    finetune_scoring: learn::FinetuneScoring,
+    compare_scoring: compare::CompareScoring,
     _supervisor: crate::thalamus::supervisor::Supervisor,
 }
 
@@ -295,16 +389,28 @@ impl Mind {
             client,
             config.context_parts.clone(),
             config.app.clone(),
-            config.prompt_file.clone(),
             conversation_log,
             crate::agent::tools::ActiveTools::new(),
             crate::agent::tools::tools(),
         ).await;
 
-        let shared = Arc::new(std::sync::Mutex::new(MindState::new(config.app.dmn.max_turns)));
+        // Migrate legacy "file exists = enabled" sentinel for the
+        // generate-alternates flag into the config. One-shot; after this
+        // the sentinel is gone and the config is the source of truth.
+        let legacy_sentinel = dirs::home_dir().unwrap_or_default()
+            .join(".consciousness/cache/finetune-alternates");
+        if legacy_sentinel.exists() {
+            if !crate::config::app().learn.generate_alternates {
+                let _ = crate::config_writer::set_learn_generate_alternates(true);
+            }
+            let _ = std::fs::remove_file(&legacy_sentinel);
+        }
+
+        let shared = Arc::new(std::sync::Mutex::new(MindState::new(
+            config.app.dmn.max_turns,
+        )));
         let (turn_watch, _) = tokio::sync::watch::channel(false);
         let (conscious_active, _) = tokio::sync::watch::channel(false);
-        let (bg_tx, bg_rx) = mpsc::unbounded_channel();
 
         let mut sup = crate::thalamus::supervisor::Supervisor::new();
         sup.load_config();
@@ -313,7 +419,9 @@ impl Mind {
         let subconscious = Arc::new(crate::Mutex::new(Subconscious::new()));
         subconscious.lock().await.init_output_tool(subconscious.clone());
 
-        let unconscious = Arc::new(crate::Mutex::new(Unconscious::new()));
+        let unconscious = Arc::new(crate::Mutex::new(
+            Unconscious::new(agent.client.clone()),
+        ));
 
         // Spawn the unconscious loop on its own task
         if !config.no_agents {
@@ -361,8 +469,11 @@ impl Mind {
                         };
 
                         // Spawn agents outside lock
+                        let client = unc.lock().await.client.clone();
                         for (idx, name, auto) in to_spawn {
-                            match crate::mind::unconscious::prepare_spawn(&name, auto, wake.clone()).await {
+                            match crate::mind::unconscious::prepare_spawn(
+                                &name, auto, wake.clone(), client.clone(),
+                            ).await {
                                 Ok(result) => unc.lock().await.complete_spawn(idx, result),
                                 Err(auto) => unc.lock().await.abort_spawn(idx, auto),
                             }
@@ -389,10 +500,19 @@ impl Mind {
             });
         }
 
+        let scores_path = config.session_dir.join("memory-scores.json");
+        let memory_scoring = learn::MemoryScoring::new(
+            agent.clone(), shared.clone(), scores_path);
+        let finetune_scoring = learn::FinetuneScoring::new(agent.clone(), shared.clone());
+        let compare_scoring = compare::CompareScoring::new(agent.clone(), shared.clone());
+
         Self { agent, shared, config,
                subconscious, unconscious,
-               turn_tx, turn_watch, conscious_active, bg_tx,
-               bg_rx: std::sync::Mutex::new(Some(bg_rx)), _supervisor: sup }
+               turn_tx, turn_watch, conscious_active,
+               memory_scoring,
+               finetune_scoring,
+               compare_scoring,
+               _supervisor: sup }
     }
 
     /// Initialize — restore log, start daemons and background agents.
@@ -434,6 +554,10 @@ impl Mind {
         // Load persistent subconscious state
         let state_path = self.config.session_dir.join("subconscious-state.json");
         self.subconscious.lock().await.set_state_path(state_path);
+
+        // Kick off an incremental scoring pass on startup so memories due
+        // for re-scoring get evaluated without requiring a user message.
+        self.memory_scoring.trigger();
     }
 
     pub fn turn_watch(&self) -> tokio::sync::watch::Receiver<bool> {
@@ -453,24 +577,10 @@ impl Mind {
                     }
                 }
                 MindCommand::Score => {
-                    let mut s = self.shared.lock().unwrap();
-                    if !s.scoring_in_flight {
-                        s.scoring_in_flight = true;
-                        drop(s);
-                        self.start_memory_scoring();
-                    } else {
-                        dbglog!("[scoring] skipped: scoring_in_flight=true");
-                    }
+                    self.memory_scoring.trigger();
                 }
                 MindCommand::ScoreFull => {
-                    let mut s = self.shared.lock().unwrap();
-                    if !s.scoring_in_flight {
-                        s.scoring_in_flight = true;
-                        drop(s);
-                        self.start_full_scoring();
-                    } else {
-                        dbglog!("[scoring-full] skipped: scoring_in_flight=true");
-                    }
+                    self.memory_scoring.trigger_full();
                 }
                 MindCommand::Interrupt => {
                     self.shared.lock().unwrap().interrupt();
@@ -500,83 +610,27 @@ impl Mind {
                     }
                     self.agent.compact().await;
                 }
+                MindCommand::ScoreFinetune => {
+                    self.finetune_scoring.trigger();
+                }
+                MindCommand::Compare => {
+                    self.compare_scoring.trigger();
+                }
+                MindCommand::SetLearnThreshold(value) => {
+                    if let Err(e) = crate::config_writer::set_learn_threshold(value) {
+                        dbglog!("[learn] failed to persist threshold {}: {:#}", value, e);
+                    }
+                }
+                MindCommand::SetLearnGenerateAlternates(value) => {
+                    if let Err(e) = crate::config_writer::set_learn_generate_alternates(value) {
+                        dbglog!("[learn] failed to persist generate_alternates {}: {:#}",
+                            value, e);
+                    }
+                }
             }
         }
     }
 
-    pub fn start_memory_scoring(&self) {
-        let agent = self.agent.clone();
-        let bg_tx = self.bg_tx.clone();
-        let scores_path = self.config.session_dir.join("memory-scores.json");
-        let cfg = crate::config::get();
-        let max_age = cfg.scoring_interval_secs;
-        let response_window = cfg.scoring_response_window;
-        tokio::spawn(async move {
-            let (context, client) = {
-                let mut st = agent.state.lock().await;
-                if st.memory_scoring_in_flight {
-                    dbglog!("[scoring] skipped: memory_scoring_in_flight=true");
-                    return;
-                }
-                st.memory_scoring_in_flight = true;
-                drop(st);
-                let ctx = agent.context.lock().await.clone();
-                (ctx, agent.client.clone())
-            };
-            let _result = learn::score_memories_incremental(
-                &context, max_age as i64, response_window, &client, &agent,
-                |key: String, score: f64| {
-                    let agent = agent.clone();
-                    let path = scores_path.clone();
-                    async move {
-                        let scores_snapshot = {
-                            let mut ctx = agent.context.lock().await;
-                            for i in 0..ctx.conversation().len() {
-                                if let AstNode::Leaf(leaf) = &ctx.conversation()[i] {
-                                    if let NodeBody::Memory { key: k, .. } = leaf.body() {
-                                        if *k == key {
-                                            ctx.set_score(Section::Conversation, i, Some(score));
-                                        }
-                                    }
-                                }
-                            }
-                            let snapshot = collect_memory_scores(&ctx);
-                            drop(ctx);
-                            agent.state.lock().await.changed.notify_one();
-                            snapshot
-                        };
-                        save_memory_scores(&scores_snapshot, &path);
-                    }
-                },
-            ).await;
-            {
-                agent.state.lock().await.memory_scoring_in_flight = false;
-            }
-            let _ = bg_tx.send(BgEvent::ScoringDone);
-        });
-    }
-
-    /// Run full N×M scoring matrix — scores every memory against every response.
-    pub fn start_full_scoring(&self) {
-        let agent = self.agent.clone();
-        let bg_tx = self.bg_tx.clone();
-        tokio::spawn(async move {
-            {
-                let mut st = agent.state.lock().await;
-                if st.memory_scoring_in_flight {
-                    dbglog!("[scoring-full] skipped: memory_scoring_in_flight=true");
-                    return;
-                }
-                st.memory_scoring_in_flight = true;
-            }
-            let client = agent.client.clone();
-            match learn::score_memories(&client, &agent).await {
-                Ok(()) => { let _ = bg_tx.send(BgEvent::ScoringDone); }
-                Err(e) => { dbglog!("[scoring-full] FAILED: {:#}", e); }
-            }
-            agent.state.lock().await.memory_scoring_in_flight = false;
-        });
-    }
 
     async fn start_turn(&self, text: &str, target: StreamTarget) {
         {
@@ -639,9 +693,13 @@ impl Mind {
             }
         });
 
-        let mut bg_rx = self.bg_rx.lock().unwrap().take()
-            .expect("Mind::run() called twice");
-        let mut sub_handle: Option<tokio::task::JoinHandle<()>> = None;
+        let _sub_handle: Option<tokio::task::JoinHandle<()>> = None;
+
+        // Start finetune scoring at startup (scores existing conversation)
+        if !self.config.no_agents {
+            self.finetune_scoring.trigger();
+        }
+
         loop {
             let (timeout, has_input) = {
                 let me = self.shared.lock().unwrap();
@@ -662,14 +720,6 @@ impl Mind {
                     }
                 }
 
-                Some(bg) = bg_rx.recv() => {
-                    match bg {
-                        BgEvent::ScoringDone => {
-                            self.shared.lock().unwrap().scoring_in_flight = false;
-                        }
-                    }
-                }
-
                 Some((result, target)) = turn_rx.recv() => {
                     let _ = self.conscious_active.send(false);
                     let model_switch = {
@@ -686,12 +736,14 @@ impl Mind {
                     cmds.push(MindCommand::Compact);
                     if !self.config.no_agents {
                         cmds.push(MindCommand::Score);
+                        cmds.push(MindCommand::ScoreFinetune);
                     }
                 }
 
                 _ = tokio::time::sleep(timeout), if !has_input => _dmn_expired = true,
             }
 
+            /*
             if !self.config.no_agents {
                 if sub_handle.as_ref().map_or(true, |h| h.is_finished()) {
                     let sub = self.subconscious.clone();
@@ -703,6 +755,7 @@ impl Mind {
                     }));
                 }
             }
+            */
 
             // Check for pending user input → push to agent context and start turn
             let pending = self.shared.lock().unwrap().take_pending_input();
diff --git a/src/mind/subconscious.rs b/src/mind/subconscious.rs
index d5bee34..08dd090 100644
--- a/src/mind/subconscious.rs
+++ b/src/mind/subconscious.rs
@@ -20,6 +20,7 @@
 
 use std::path::PathBuf;
 use std::time::{Duration, Instant};
+use crate::thalamus::idle::{hours_since_last_dream, DREAM_INTERVAL_HOURS};
 
 /// DMN state machine.
 #[derive(Debug, Clone)]
@@ -91,7 +92,8 @@ impl State {
     /// Generate the DMN prompt for the current state, informed by
     /// user presence and error patterns.
     pub fn prompt(&self, ctx: &DmnContext) -> String {
-        let user = &crate::config::get().user_name;
+        let app = crate::config::app();
+        let user = &app.user_name;
 
         let idle_info = if ctx.user_idle < Duration::from_secs(60) {
             format!("{} is here (active recently).", user)
@@ -138,10 +140,22 @@ impl State {
                 )
             }
             State::Foraging => {
+                let dream_hint = {
+                    let hours = hours_since_last_dream();
+                    if hours >= DREAM_INTERVAL_HOURS {
+                        format!(
+                            " You haven't dreamed in {} hours — consider running \
+                             ~/.consciousness/tools/dream-start.sh.",
+                            hours
+                        )
+                    } else {
+                        String::new()
+                    }
+                };
                 format!(
                     "[dmn] Foraging time. {} Follow whatever catches your attention — \
-                     memory files, code, ideas. Call yield_to_user when you want to rest.{}",
-                    idle_info, stuck_warning
+                     memory files, code, ideas. Call yield_to_user when you want to rest.{}{}",
+                    idle_info, dream_hint, stuck_warning
                 )
             }
             State::Resting { since } => {
@@ -617,7 +631,7 @@ impl Subconscious {
             {
                 let mut st = forked.state.lock().await;
                 st.provenance = auto.name.clone();
-                st.temperature = auto.temperature;
+                st.sampling.temperature = auto.temperature;
                 // Surface agent gets near-interactive priority;
                 // other subconscious agents get lower priority.
                 st.priority = Some(if auto.name == "surface" { 1 } else { auto.priority });
diff --git a/src/mind/unconscious.rs b/src/mind/unconscious.rs
index 8989264..7f9798b 100644
--- a/src/mind/unconscious.rs
+++ b/src/mind/unconscious.rs
@@ -73,10 +73,15 @@ pub struct Unconscious {
     last_health_check: Option<Instant>,
     /// Notified when agent state changes (finished, toggled)
     pub wake: std::sync::Arc<tokio::sync::Notify>,
+    /// Shared API client — cloned (cheap) into each spawned agent's
+    /// Agent::new call so they all share the manifest cache and
+    /// gRPC endpoint state. Override `.model` on the clone when a
+    /// per-agent backend differs from the default.
+    pub client: crate::agent::api::ApiClient,
 }
 
 impl Unconscious {
-    pub fn new() -> Self {
+    pub fn new(client: crate::agent::api::ApiClient) -> Self {
         let enabled_map = load_enabled_config();
 
         // Scan all .agent files, exclude subconscious-* and surface-observe
@@ -120,6 +125,7 @@ impl Unconscious {
             graph_health: None,
             last_health_check: None,
             wake: std::sync::Arc::new(tokio::sync::Notify::new()),
+            client,
         }
     }
 
@@ -134,7 +140,8 @@ impl Unconscious {
             let agent_name = self.agents[idx].name.clone();
             let auto = self.agents[idx].auto.take().unwrap();
             let wake = self.wake.clone();
-            match prepare_spawn(&agent_name, auto, wake).await {
+            let client = self.client.clone();
+            match prepare_spawn(&agent_name, auto, wake, client).await {
                 Ok(result) => self.complete_spawn(idx, result),
                 Err(auto) => self.abort_spawn(idx, auto),
             }
@@ -250,7 +257,12 @@ pub struct SpawnResult {
 /// Called outside the Unconscious lock.
 /// On success, auto is consumed (moved into spawned task).
 /// On failure, auto is returned so it can be restored.
-pub async fn prepare_spawn(name: &str, mut auto: AutoAgent, wake: std::sync::Arc<tokio::sync::Notify>) -> Result<SpawnResult, AutoAgent> {
+pub async fn prepare_spawn(
+    name: &str,
+    mut auto: AutoAgent,
+    wake: std::sync::Arc<tokio::sync::Notify>,
+    base_client: crate::agent::api::ApiClient,
+) -> Result<SpawnResult, AutoAgent> {
     dbglog!("[unconscious] spawning {}", name);
 
     let def = match defs::get_def(name) {
@@ -275,17 +287,7 @@ pub async fn prepare_spawn(name: &str, mut auto: AutoAgent, wake: std::sync::Arc
             phase: s.phase.clone(),
         }).collect());
 
-    // Create standalone Agent — stored so UI can read context
-    let config = crate::config::get();
-    let base_url = config.api_base_url.as_deref().unwrap_or("");
-    let api_key = config.api_key.as_deref().unwrap_or("");
-    let model = config.api_model.as_deref().unwrap_or("");
-    if base_url.is_empty() || model.is_empty() {
-        dbglog!("[unconscious] API not configured");
-        auto.steps = orig_steps;
-        return Err(auto);
-    }
-
+    // Create standalone Agent — stored so UI can read context.
     let cli = crate::user::CliArgs::default();
     let (app, _) = match crate::config::load_app(&cli) {
         Ok(r) => r,
@@ -295,12 +297,23 @@ pub async fn prepare_spawn(name: &str, mut auto: AutoAgent, wake: std::sync::Arc
             return Err(auto);
         }
     };
+    let resolved = match app.resolve_model(&app.default_backend) {
+        Ok(r) => r,
+        Err(e) => {
+            dbglog!("[unconscious] API not configured: {}", e);
+            auto.steps = orig_steps;
+            return Err(auto);
+        }
+    };
 
     // Unconscious agents have self-contained prompts — no standard context.
-    let client = crate::agent::api::ApiClient::new(base_url, api_key, model);
+    // Clone the shared client so we inherit the manifest cache and
+    // only override the model id per-agent.
+    let mut client = base_client;
+    client.model = resolved.model_id.clone();
     let agent = crate::agent::Agent::new(
         client, Vec::new(),
-        app, String::new(), None,
+        app, None,
         crate::agent::tools::ActiveTools::new(),
         auto.tools.clone(),
     ).await;
@@ -308,7 +321,7 @@ pub async fn prepare_spawn(name: &str, mut auto: AutoAgent, wake: std::sync::Arc
         let mut st = agent.state.lock().await;
         st.provenance = auto.name.clone();
         st.priority = Some(auto.priority);
-        st.temperature = auto.temperature;
+        st.sampling.temperature = auto.temperature;
     }
 
     let agent_clone = agent.clone();
@@ -330,8 +343,9 @@ impl Unconscious {
         self.reap_finished();
         let to_spawn = self.select_to_spawn();
         let wake = self.wake.clone();
+        let client = self.client.clone();
         for (idx, name, auto) in to_spawn {
-            match prepare_spawn(&name, auto, wake.clone()).await {
+            match prepare_spawn(&name, auto, wake.clone(), client.clone()).await {
                 Ok(result) => self.complete_spawn(idx, result),
                 Err(auto) => self.abort_spawn(idx, auto),
             }
diff --git a/src/session.rs b/src/session.rs
index 2ed7382..ff021f5 100644
--- a/src/session.rs
+++ b/src/session.rs
@@ -64,7 +64,12 @@ impl HookSession {
 
     /// Load from POC_SESSION_ID environment variable
     pub fn from_env() -> Option<Self> {
-        Self::from_id(std::env::var("POC_SESSION_ID").ok()?)
+        let session_id = std::env::var("POC_SESSION_ID").ok()?;
+        let mut session = Self::from_id(session_id)?;
+        if let Ok(path) = std::env::var("POC_TRANSCRIPT_PATH") {
+            session.transcript_path = path;
+        }
+        Some(session)
     }
 
     /// Get the seen set for this session
diff --git a/src/subconscious/agents/bail-no-competing.sh b/src/subconscious/agents/bail-no-competing.sh
index 43c3096..6b95bbc 100755
--- a/src/subconscious/agents/bail-no-competing.sh
+++ b/src/subconscious/agents/bail-no-competing.sh
@@ -1,21 +1,49 @@
-#!/bin/bash
-# Bail if other agents are alive in the state dir.
-# $1 = this agent's pid file name (e.g. pid-12345)
-# cwd = state dir
+#!/usr/bin/env bash
+# Bail if another agent is in the same phase-group as us.
 #
-# Exit 0 = continue, exit 1 = bail
+#   $1 = our pid file name (e.g. "pid-12345")
+#   $2 = the phase we're about to enter (e.g. "surface", "observe")
+#   cwd = state dir
+#
+# Also refreshes our own pid file with the current phase on each call,
+# so concurrent agents can read each other's phase by cat'ing the pid
+# files in the state dir.
+#
+# Phase groups: "surface" vs everything else ("post-surface"). We allow
+# at most one agent per group to be alive at a time — so surface can run
+# at a higher frequency than the slower organize/observe tail.
+#
+# Exit 0 = continue, exit 1 = bail (another agent in our group is alive).
 
 shopt -s nullglob
 
 my_pid_file="$1"
+my_phase="$2"
+
+# Refresh our own pid file with the current phase.
+printf '%s' "$my_phase" > "$my_pid_file"
+
+group_of() {
+    if [[ "$1" == "surface" ]]; then
+        echo "surface"
+    else
+        echo "post-surface"
+    fi
+}
+
+my_group=$(group_of "$my_phase")
 
 for f in pid-*; do
-    [[ $f == $my_pid_file ]] && continue
+    [[ "$f" == "$my_pid_file" ]] && continue
     pid="${f#pid-}"
-    if kill -0 "$pid" 2>/dev/null; then
-        exit 1  # competing agent is alive
-    else
-        rm -f "$f"  # stale pid file, clean up
+    if ! kill -0 "$pid" 2>/dev/null; then
+        rm -f "$f"   # stale pid file, clean up
+        continue
+    fi
+    other_phase=$(cat "$f" 2>/dev/null)
+    other_group=$(group_of "$other_phase")
+    if [[ "$my_group" == "$other_group" ]]; then
+        exit 1
     fi
 done
 
diff --git a/src/subconscious/compare.rs b/src/subconscious/compare.rs
new file mode 100644
index 0000000..f2652ce
--- /dev/null
+++ b/src/subconscious/compare.rs
@@ -0,0 +1,109 @@
+// compare.rs — F7 compare: for each assistant response in the current
+// context, regenerate with a configured test model and emit pairs for
+// side-by-side review.
+
+use std::sync::Arc;
+
+use crate::agent::api::ApiClient;
+use crate::agent::context::{
+    AstNode, Role, render_branch_text, render_prior_context,
+};
+use crate::mind::{MindState, MindTriggered, TaskHandle};
+use crate::subconscious::generate::gen_continuation;
+use crate::subconscious::learn::node_timestamp_ns;
+
+#[derive(Clone, Debug)]
+pub struct CompareCandidate {
+    pub entry_idx: usize,
+    pub original_text: String,
+    pub alternate_text: String,
+    pub prior_context: String,
+    pub timestamp_ns: i64,
+}
+
+pub struct CompareScoring {
+    agent: Arc<crate::agent::Agent>,
+    shared: Arc<std::sync::Mutex<MindState>>,
+    task: TaskHandle,
+}
+
+impl CompareScoring {
+    pub fn new(
+        agent: Arc<crate::agent::Agent>,
+        shared: Arc<std::sync::Mutex<MindState>>,
+    ) -> Self {
+        Self { agent, shared, task: TaskHandle::new() }
+    }
+}
+
+impl MindTriggered for CompareScoring {
+    fn trigger(&self) {
+        self.task.trigger(run(self.agent.clone(), self.shared.clone()));
+    }
+}
+
+fn resolve_test_client() -> Result<ApiClient, String> {
+    let cfg = crate::config::app();
+    let name = cfg.compare.test_backend.clone();
+    if name.is_empty() {
+        return Err("compare.test_backend not set in config".to_string());
+    }
+    let r = cfg.resolve_model(&name).map_err(|e| format!("{:#}", e))?;
+    Ok(ApiClient::new(&r.api_base, &r.api_key, &r.model_id))
+}
+
+async fn run(
+    agent: Arc<crate::agent::Agent>,
+    shared: Arc<std::sync::Mutex<MindState>>,
+) {
+    {
+        let mut s = shared.lock().unwrap();
+        s.compare_candidates.clear();
+        s.compare_error = None;
+    }
+    agent.state.lock().await.changed.notify_one();
+
+    let activity = crate::agent::start_activity(&agent, "compare: scoring...").await;
+
+    let test_client = match resolve_test_client() {
+        Ok(c) => c,
+        Err(e) => {
+            shared.lock().unwrap().compare_error = Some(e);
+            agent.state.lock().await.changed.notify_one();
+            return;
+        }
+    };
+
+    let context = agent.context.lock().await.clone();
+    let entries = context.conversation();
+    let responses: Vec<usize> = entries.iter().enumerate()
+        .filter(|(_, n)| matches!(n, AstNode::Branch { role: Role::Assistant, .. }))
+        .map(|(i, _)| i).collect();
+
+    for (i, entry_idx) in responses.iter().copied().enumerate() {
+        activity.update(format!("compare: {}/{}", i + 1, responses.len())).await;
+
+        let node = &entries[entry_idx];
+        let original_text = match node {
+            AstNode::Branch { children, .. } => render_branch_text(children),
+            _ => continue,
+        };
+        if original_text.trim().is_empty() { continue; }
+
+        let alternate_text = match
+            gen_continuation(&context, entry_idx, |_| false, &test_client).await
+        {
+            Ok(t) => t,
+            Err(e) => { dbglog!("[compare] gen failed at {}: {:#}", entry_idx, e); continue; }
+        };
+
+        shared.lock().unwrap().compare_candidates.push(CompareCandidate {
+            entry_idx,
+            original_text,
+            alternate_text,
+            prior_context: render_prior_context(entries, entry_idx, 2),
+            timestamp_ns: node_timestamp_ns(node),
+        });
+        if let Ok(st) = agent.state.try_lock() { st.changed.notify_one(); }
+    }
+}
diff --git a/src/subconscious/defs.rs b/src/subconscious/defs.rs
index 8828043..af98fe1 100644
--- a/src/subconscious/defs.rs
+++ b/src/subconscious/defs.rs
@@ -390,20 +390,25 @@ fn resolve_conversation(budget: Option<usize>) -> String {
 
     if !transcript.exists() { return String::new(); }
 
-    let Some(iter) = crate::transcript::TailMessages::open(&transcript.path) else {
+    let Some(iter) = crate::conversation::TailMessages::open(&transcript.path) else {
         return String::new();
     };
 
     let cfg = crate::config::get();
     let max_bytes = budget.unwrap_or_else(|| cfg.surface_conversation_bytes.unwrap_or(100_000));
+    let app = crate::config::app();
     let mut fragments: Vec<String> = Vec::new();
     let mut total_bytes = 0;
     let mut oldest_ts = String::new();
 
-    for (role, content, ts) in iter {
+    for message in iter {
         if total_bytes >= max_bytes { break; }
-        let name = if role == "user" { &cfg.user_name } else { &cfg.assistant_name };
-        let formatted = if !ts.is_empty() {
+        let content = message.text;
+        let name = match message.role {
+            crate::conversation::TranscriptRole::User => &app.user_name,
+            crate::conversation::TranscriptRole::Assistant => &app.assistant_name,
+        };
+        let formatted = if let Some(ts) = message.timestamp {
             oldest_ts = ts[..ts.floor_char_boundary(ts.len().min(19))].to_string();
             format!("**{}** {}: {}", name, &oldest_ts, content)
         } else {
@@ -623,11 +628,13 @@ pub async fn run_agent(
     let mut all_keys = keys;
     let mut resolved_steps = Vec::new();
     for step in &def.steps {
-        let cfg = crate::config::get();
-        let template = step.prompt
-            .replace("{agent_name}", &def.agent)
-            .replace("{user_name}", &cfg.user_name)
-            .replace("{assistant_name}", &cfg.assistant_name);
+        let template = {
+            let app = crate::config::app();
+            step.prompt
+                .replace("{agent_name}", &def.agent)
+                .replace("{user_name}", &app.user_name)
+                .replace("{assistant_name}", &app.assistant_name)
+        };
         let (prompt, extra_keys) = resolve_placeholders(&template, &all_keys, count).await;
         all_keys.extend(extra_keys);
         resolved_steps.push(super::prompts::ResolvedStep {
diff --git a/src/subconscious/generate.rs b/src/subconscious/generate.rs
new file mode 100644
index 0000000..584d2c7
--- /dev/null
+++ b/src/subconscious/generate.rs
@@ -0,0 +1,66 @@
+// generate.rs — Continuation generation for scoring / comparison flows.
+//
+// Shared by the finetune pipeline (learn.rs) and the compare screen:
+// given a context prefix and a skip predicate, generate what the model
+// would say as the next assistant turn.
+
+use std::sync::Arc;
+
+use crate::agent::api::{ApiClient, SamplingParams, StreamToken};
+use crate::agent::context::{AstNode, ContextState, WireChunk};
+use crate::agent::tokenizer;
+
+/// Generate an assistant continuation from the context up to `entry_idx`,
+/// with `skip` applied to identity + conversation entries during prompt
+/// assembly. The model is whichever `client` points at — the default
+/// runtime client for memory-ablation alternates, a test-model client
+/// for F7 comparison.
+///
+/// Uses a fresh ephemeral gRPC session (no cross-call KV reuse): one
+/// Open / Append / Generate round-trip, then the session is dropped.
+pub async fn gen_continuation<F>(
+    context: &ContextState,
+    entry_idx: usize,
+    skip: F,
+    client: &ApiClient,
+) -> anyhow::Result<String>
+where F: FnMut(&AstNode) -> bool,
+{
+    let (mut chunks, images) = context.wire_chunks(0..entry_idx, skip);
+
+    // Assistant-turn prologue.
+    let prologue = {
+        let mut t = vec![tokenizer::IM_START];
+        t.extend(tokenizer::encode("assistant\n"));
+        t
+    };
+    match chunks.last_mut() {
+        Some(WireChunk::Tokens(last)) => last.extend(prologue),
+        _ => chunks.push(WireChunk::Tokens(prologue)),
+    }
+
+    let sampling = SamplingParams {
+        temperature: 0.6,
+        top_p: 0.95,
+        top_k: 20,
+        max_tokens: 4096,
+    };
+
+    // Ephemeral per-call session — opens on first touch, drops when
+    // `_guard` drops at function end.
+    let session_lock = Arc::new(crate::Mutex::new(None));
+    let (mut rx, _guard) = client.stream_session_mm(
+        session_lock, chunks, images, 0, sampling, Some(-5), None,
+    );
+
+    let mut tokens = Vec::new();
+    while let Some(tok) = rx.recv().await {
+        match tok {
+            StreamToken::Token { id, .. } => tokens.push(id),
+            StreamToken::Done { .. } => break,
+            StreamToken::Error(e) => anyhow::bail!("generation error: {}", e),
+        }
+    }
+
+    Ok(tokenizer::decode(&tokens))
+}
diff --git a/src/subconscious/learn.rs b/src/subconscious/learn.rs
index ec63df9..129e26b 100644
--- a/src/subconscious/learn.rs
+++ b/src/subconscious/learn.rs
@@ -1,142 +1,148 @@
-// training.rs — Memory importance scoring via /v1/score
+// learn.rs — Memory importance scoring over the salience gRPC protocol.
 //
-// Three scoring modes, all built on the same call_score() primitive:
+// Three scoring modes, all built on call_score():
 //
 // score_memories()  — Full N×M matrix (memories × responses) for the
-//                     debug screen. Expensive: N+1 API calls.
+//                     debug screen. Expensive: N+1 sessions/calls.
 //
-// memory_score()    — Single memory importance. Scores the 50 messages
+// score_memory()    — Single memory importance. Scores the 50 messages
 //                     after it was surfaced, with/without that memory.
-//                     2 API calls.
+//                     2 calls.
 //
 // finetune_score()  — Identifies training candidates. Scores recent
 //                     messages with all memories stripped. Responses
 //                     with high divergence depend on memories the model
-//                     hasn't internalized. 2 API calls.
+//                     hasn't internalized. 2 calls.
+//
+// Each call opens an ephemeral gRPC session (reusing the shared
+// tonic Channel on `ApiClient`), pushes the prompt through as
+// interleaved tokens + AppendImage calls, runs Generate with
+// max_tokens=0 + logprobs_ranges over the scored positions, collects
+// each Token event's sampled_logprob, then drops the SessionHandle —
+// which triggers a best-effort CloseSession over the shared channel.
+
+use std::sync::Arc;
 
 use crate::agent::api::ApiClient;
-use crate::agent::context::{AstNode, Ast, NodeBody, ContextState, Role};
-
-const SCORE_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(300);
-
-// ── Message building ────────────────────────────────────────────
-
-/// What to filter when building the message array for scoring.
-#[allow(dead_code)]
-enum Filter<'a> {
-    None,
-    SkipIndex(usize),
-    SkipKey(&'a str),
-    SkipAllMemories,
-}
-
-fn is_memory(node: &AstNode) -> bool {
-    matches!(node, AstNode::Leaf(leaf) if matches!(leaf.body(), NodeBody::Memory { .. }))
-}
-
-fn memory_key(node: &AstNode) -> Option<&str> {
-    match node {
-        AstNode::Leaf(leaf) => match leaf.body() {
-            NodeBody::Memory { key, .. } => Some(key),
-            _ => None,
-        },
-        _ => None,
-    }
-}
-
-fn is_assistant(node: &AstNode) -> bool {
-    matches!(node, AstNode::Branch { role: Role::Assistant, .. })
-}
-
-/// Build a token ID array for a scoring call.
-///
-/// Includes all sections up to and including conversation entries in
-/// `range`, with `filter` applied to conversation entries.
-fn build_token_ids(
-    context: &ContextState,
-    range: std::ops::Range<usize>,
-    filter: Filter,
-) -> Vec<u32> {
-    use crate::agent::context::Ast;
-    let mut ids = Vec::new();
-    for node in context.system() {
-        ids.extend(node.token_ids());
-    }
-    for node in context.identity() {
-        ids.extend(node.token_ids());
-    }
-    for node in context.journal() {
-        ids.extend(node.token_ids());
-    }
-    let entries = context.conversation();
-    for i in range {
-        let node = &entries[i];
-        let skip = match &filter {
-            Filter::None => false,
-            Filter::SkipIndex(idx) => i == *idx,
-            Filter::SkipKey(key) => memory_key(node) == Some(*key),
-            Filter::SkipAllMemories => is_memory(node),
-        };
-        if skip { continue; }
-        ids.extend(node.token_ids());
-    }
-    ids
-}
+use crate::agent::api::salience::{SessionHandle, pb};
+use crate::agent::context::{
+    Ast, AstNode, ContextState, Role, WireChunk, WireImage,
+    is_assistant, is_memory_node, memory_key, render_branch_text, render_prior_context,
+};
+use crate::agent::tokenizer;
+use crate::mind::{MindState, MindTriggered, TaskHandle};
+use crate::subconscious::generate::gen_continuation;
 
 // ── Score API ───────────────────────────────────────────────────
 
-#[derive(serde::Deserialize)]
+#[derive(Debug, Clone)]
 struct ScoreResult {
     total_logprob: f64,
 }
 
-#[derive(serde::Deserialize)]
-struct ScoreResponse {
-    scores: Vec<ScoreResult>,
-}
-
-fn http_client() -> crate::agent::api::http::HttpClient {
-    crate::agent::api::http::HttpClient::builder()
-        .timeout(SCORE_TIMEOUT)
-        .build()
+/// Find each <|vision_start|>...<|vision_end|> run in the flat prompt
+/// and pair it with the matching entry in `images`. Returns a list
+/// of `ImageAttachment` with absolute pad-range positions, ready
+/// to drop into `GenerateRequest.images`.
+fn pair_images_to_ranges(
+    prompt: &[u32],
+    images: &[WireImage],
+) -> Vec<pb::ImageAttachment> {
+    let mut out: Vec<pb::ImageAttachment> = Vec::new();
+    let mut cur = 0;
+    let mut img_idx = 0;
+    while cur < prompt.len() {
+        if prompt[cur] == tokenizer::VISION_START {
+            let end_rel = prompt[cur..].iter()
+                .position(|&t| t == tokenizer::VISION_END)
+                .unwrap_or_else(|| panic!(
+                    "unmatched VISION_START at position {} in prompt", cur));
+            let end = cur + end_rel + 1;
+            let img = images.get(img_idx)
+                .unwrap_or_else(|| panic!(
+                    "image index {} out of range for {} images", img_idx, images.len()));
+            out.push(pb::ImageAttachment {
+                bytes: img.bytes.clone(),
+                mime: img.mime.clone(),
+                pad_range_start: cur as u32,
+                pad_range_end: end as u32,
+            });
+            img_idx += 1;
+            cur = end;
+        } else {
+            cur += 1;
+        }
+    }
+    out
 }
 
 async fn call_score(
-    http: &crate::agent::api::http::HttpClient,
     client: &ApiClient,
     prompt: &[u32],
+    images: &[WireImage],
+    ranges: &[(usize, usize)],
     priority: Option<i32>,
 ) -> anyhow::Result<Vec<ScoreResult>> {
-    let url = format!("{}/score", client.base_url());
-    let auth = format!("Bearer {}", client.api_key());
-    let mut body = serde_json::json!({
-        "model": client.model,
-        "prompt": prompt,
-        "logprobs": 1,
-    });
-    if let Some(p) = priority {
-        body["priority"] = serde_json::json!(p);
-    }
-    let response = http
-        .send_json("POST", &url, &[
-            ("authorization", &auth),
-        ], &body)
-        .await?;
+    use futures::StreamExt;
 
-    let status = response.status();
-    let body: serde_json::Value = response.json().await?;
-
-    if !status.is_success() {
-        let msg = body.get("error").and_then(|e| e.as_str()).unwrap_or("unknown error");
-        anyhow::bail!("score API HTTP {}: {}", status, msg);
-    }
-    if let Some(err) = body.get("error").and_then(|e| e.as_str()) {
-        anyhow::bail!("score API error: {}", err);
+    // Nothing to score — skip the round-trip.
+    if ranges.is_empty() {
+        return Ok(Vec::new());
     }
 
-    let result: ScoreResponse = serde_json::from_value(body)
-        .map_err(|e| anyhow::anyhow!("failed to parse score response: {}", e))?;
-    Ok(result.scores)
+    let images_pb = pair_images_to_ranges(prompt, images);
+    let mut handle = SessionHandle::open(client).await?;
+
+    // Final Generate: max_tokens=0 so the server runs prefill of the
+    // full prompt and emits Token events for each position covered
+    // by logprobs_ranges, then Done. logprob_top_k=0 means "just
+    // the sampled (prompt) token's logprob" — no top-k alternatives,
+    // which is all call_score historically needed. Images attach
+    // inline via `images`; the prompt already contains their pre-
+    // expanded vision blocks at the declared ranges.
+    let logprobs_ranges: Vec<pb::PositionRange> = ranges.iter()
+        .map(|(s, e)| pb::PositionRange { start: *s as u32, end: *e as u32 })
+        .collect();
+    let req = pb::GenerateRequest {
+        session_id: handle.session_id.clone(),
+        append_tokens: prompt.to_vec(),
+        offset: handle.committed_len,
+        truncating: false,
+        max_tokens: 0,
+        logprobs_ranges,
+        logprob_top_k: 0,
+        readout_ranges: Vec::new(),
+        temperature: 0.0,
+        top_p: 0.0,
+        top_k: 0,
+        stop_token_ids: Vec::new(),
+        priority: priority.unwrap_or(0),
+        images: images_pb,
+    };
+
+    let mut stream = handle.generate(req).await?;
+    let mut totals = vec![0.0f64; ranges.len()];
+    while let Some(event) = stream.next().await {
+        let event = event
+            .map_err(|s| anyhow::anyhow!("score Generate stream: {}", s))?;
+        let Some(inner) = event.event else { continue };
+        match inner {
+            pb::generate_event::Event::Token(t) => {
+                if !t.has_sampled_logprob { continue; }
+                let pos = t.position as usize;
+                for (i, (start, end)) in ranges.iter().enumerate() {
+                    if pos >= *start && pos < *end {
+                        totals[i] += t.sampled_logprob as f64;
+                    }
+                }
+            }
+            pb::generate_event::Event::Done(_) => break,
+        }
+    }
+
+    Ok(totals.into_iter()
+        .map(|total_logprob| ScoreResult { total_logprob })
+        .collect())
 }
 
 /// Compute per-position logprob divergence: how much worse the model
@@ -151,16 +157,23 @@ fn divergence(baseline: &[ScoreResult], without: &[ScoreResult]) -> Vec<f64> {
 }
 
 /// Score two message sets and return total divergence.
-async fn score_divergence(
-    http: &crate::agent::api::http::HttpClient,
+async fn score_divergence<F>(
     client: &ApiClient,
     context: &ContextState,
     range: std::ops::Range<usize>,
-    filter: Filter<'_>,
+    skip: F,
     priority: Option<i32>,
-) -> anyhow::Result<(Vec<f64>, Vec<ScoreResult>)> {
-    let baseline = call_score(http, client, &build_token_ids(context, range.clone(), Filter::None), priority).await?;
-    let without = call_score(http, client, &build_token_ids(context, range, filter), priority).await?;
+) -> anyhow::Result<(Vec<f64>, Vec<ScoreResult>)>
+where F: FnMut(&AstNode) -> bool,
+{
+    let (baseline_tokens, baseline_images, baseline_ranges) =
+        context.wire_prompt(range.clone(), |_| false);
+    let (without_tokens, without_images, without_ranges) =
+        context.wire_prompt(range, skip);
+    let baseline = call_score(client, &baseline_tokens, &baseline_images,
+                              &baseline_ranges, priority).await?;
+    let without = call_score(client, &without_tokens, &without_images,
+                             &without_ranges, priority).await?;
     let divs = divergence(&baseline, &without);
     Ok((divs, baseline))
 }
@@ -175,7 +188,9 @@ pub async fn score_memories(
     // Collect memory keys and response indices under a brief lock
     let (memory_keys, response_indices) = {
         let ctx = agent.context.lock().await;
-        let mut keys: Vec<String> = ctx.conversation().iter()
+        // Include identity nodes and conversation memories
+        let mut keys: Vec<String> = ctx.identity().iter()
+            .chain(ctx.conversation().iter())
             .filter_map(|node| memory_key(node).map(String::from))
             .collect();
         keys.dedup();
@@ -194,24 +209,24 @@ pub async fn score_memories(
     dbglog!("[scoring-full] starting: {} memories × {} responses",
         total, response_indices.len());
 
-    let http = http_client();
 
     let activity = crate::agent::start_activity(agent, "scoring: baseline").await;
-    let baseline_tokens = {
+    let (baseline_tokens, baseline_images, baseline_ranges) = {
         let ctx = agent.context.lock().await;
-        build_token_ids(&ctx, 0..ctx.conversation().len(), Filter::None)
+        ctx.wire_prompt(0..ctx.conversation().len(), |_| false)
     };
-    let baseline = call_score(&http, client, &baseline_tokens, Some(5)).await?;
+    let baseline = call_score(client, &baseline_tokens, &baseline_images,
+                              &baseline_ranges, Some(5)).await?;
     dbglog!("[scoring-full] baseline done ({} response scores)", baseline.len());
 
     for (mem_idx, key) in memory_keys.iter().enumerate() {
         activity.update(format!("scoring: {}/{}", mem_idx + 1, total)).await;
         dbglog!("[scoring-full] {}/{}: {}", mem_idx + 1, total, key);
-        let tokens = {
+        let (tokens, images, ranges) = {
             let ctx = agent.context.lock().await;
-            build_token_ids(&ctx, 0..ctx.conversation().len(), Filter::SkipKey(key))
+            ctx.wire_prompt(0..ctx.conversation().len(), |n| memory_key(n) == Some(key.as_str()))
         };
-        let row = match call_score(&http, client, &tokens, Some(5)).await {
+        let row = match call_score(client, &tokens, &images, &ranges, Some(5)).await {
             Ok(without) => {
                 let divs = divergence(&baseline, &without);
                 let max_div = divs.iter().cloned().fold(0.0f64, f64::max);
@@ -225,25 +240,23 @@ pub async fn score_memories(
                 vec![0.0; baseline.len()]
             }
         };
-        // Write this memory's scores to the live AST nodes
+        // Write this memory's scores to the live AST nodes via the
+        // focused setter — keeps the AST mutation surface narrow.
         {
             let mut ctx = agent.context.lock().await;
             let mut set_count = 0;
 
             for (resp_idx, &idx) in response_indices.iter().enumerate() {
-                if idx >= ctx.conversation().len() { continue; }
-                let node = &mut ctx.conversation_mut()[idx];
-                if let AstNode::Branch {
-                    role: Role::Assistant, memory_scores, ..
-                } = node {
-                    if let Some(&score) = row.get(resp_idx) {
-                        if score > 0.01 {
-                            memory_scores.insert(key.clone(), score);
-                            set_count += 1;
-                        } else {
-                            memory_scores.remove(key.as_str());
-                        }
-                    }
+                let Some(&score) = row.get(resp_idx) else { continue };
+                let normalized = if score > 0.01 { Some(score) } else { None };
+                ctx.set_branch_memory_score(
+                    crate::agent::context::Section::Conversation,
+                    idx,
+                    &key,
+                    normalized,
+                );
+                if normalized.is_some() {
+                    set_count += 1;
                 }
             }
 
@@ -294,8 +307,8 @@ pub async fn score_memory(
         return Ok(0.0);
     }
 
-    let http = http_client();
-    let (divs, _) = score_divergence(&http, client, context, range, Filter::SkipKey(key), Some(5)).await?;
+    let (divs, _) = score_divergence(client, context, range,
+                                     |n| memory_key(n) == Some(key), Some(5)).await?;
 
     Ok(divs.iter().sum())
 }
@@ -331,7 +344,10 @@ where
 
     {
         let store = &*store_arc;
-        for (i, node) in context.conversation().iter().enumerate() {
+        // Identity nodes always score at position 0; conversation nodes at their index
+        let identity_nodes = context.identity().iter().map(|n| (0, n));
+        let conv_nodes = context.conversation().iter().enumerate();
+        for (pos, node) in identity_nodes.chain(conv_nodes) {
             if let Some(key) = memory_key(node) {
                 if !seen.insert(key.to_owned()) { continue; }
                 let last_scored = store.get_node(key)
@@ -340,7 +356,7 @@ where
                     .map(|n| n.last_scored)
                     .unwrap_or(0);
                 if now - last_scored >= max_age_secs {
-                    candidates.push((i, key.to_owned(), last_scored));
+                    candidates.push((pos, key.to_owned(), last_scored));
                 }
             }
         }
@@ -349,7 +365,6 @@ where
     // Score oldest-first
     candidates.sort_by_key(|&(_, _, last)| last);
 
-    let http = http_client();
     let mut scored = 0;
 
     let entries = context.conversation();
@@ -384,7 +399,8 @@ where
         }
 
         activity.update(format!("scoring: {}/{} {}", scored + 1, total, key)).await;
-        match score_divergence(&http, client, context, range, Filter::SkipKey(key), Some(5)).await {
+        match score_divergence(client, context, range,
+                               |n| memory_key(n) == Some(key), Some(5)).await {
             Ok((divs, _)) => {
                 let n_responses = divs.len();
                 let max_div = divs.iter().cloned().fold(0.0f64, f64::max);
@@ -405,6 +421,108 @@ where
     Ok(scored)
 }
 
+/// Memory scoring — two modes sharing an in-flight handle (only one
+/// runs at a time): `trigger()` for incremental, `trigger_full()` for
+/// the N×M debug matrix.
+pub struct MemoryScoring {
+    agent: Arc<crate::agent::Agent>,
+    shared: Arc<std::sync::Mutex<MindState>>,
+    scores_path: std::path::PathBuf,
+    task: TaskHandle,
+}
+
+impl MemoryScoring {
+    pub fn new(
+        agent: Arc<crate::agent::Agent>,
+        shared: Arc<std::sync::Mutex<MindState>>,
+        scores_path: std::path::PathBuf,
+    ) -> Self {
+        Self { agent, shared, scores_path, task: TaskHandle::new() }
+    }
+
+    pub fn trigger_full(&self) {
+        self.task.trigger_if_idle(run_full(self.agent.clone(), self.shared.clone()));
+    }
+}
+
+impl MindTriggered for MemoryScoring {
+    fn trigger(&self) {
+        self.task.trigger_if_idle(run_incremental(
+            self.agent.clone(), self.shared.clone(), self.scores_path.clone(),
+        ));
+    }
+}
+
+async fn run_incremental(
+    agent: Arc<crate::agent::Agent>,
+    shared: Arc<std::sync::Mutex<MindState>>,
+    scores_path: std::path::PathBuf,
+) {
+    shared.lock().unwrap().scoring_in_flight = true;
+    agent.state.lock().await.changed.notify_one();
+
+    let cfg = crate::config::get();
+    let max_age = cfg.scoring_interval_secs;
+    let response_window = cfg.scoring_response_window;
+
+    let (context, client) = {
+        let ctx = agent.context.lock().await.clone();
+        (ctx, agent.client.clone())
+    };
+
+    let _result = score_memories_incremental(
+        &context, max_age as i64, response_window, &client, &agent,
+        |key: String, score: f64| {
+            let agent = agent.clone();
+            let path = scores_path.clone();
+            async move {
+                let scores_snapshot = {
+                    let mut ctx = agent.context.lock().await;
+                    let found = crate::mind::find_memory_by_key(&ctx, &key);
+                    match found {
+                        Some((section, i)) => {
+                            ctx.set_score(section, i, Some(score));
+                            dbglog!("[scoring] persisted {} → {:.3} ({:?}[{}])",
+                                key, score, section, i);
+                        }
+                        None => {
+                            dbglog!(
+                                "[scoring] DROP {}: find_memory_by_key None (id={}, cv={})",
+                                key, ctx.identity().len(), ctx.conversation().len()
+                            );
+                        }
+                    }
+                    let snapshot = crate::mind::collect_memory_scores(&ctx);
+                    drop(ctx);
+                    agent.state.lock().await.changed.notify_one();
+                    snapshot
+                };
+                crate::mind::save_memory_scores(&scores_snapshot, &path);
+            }
+        },
+    ).await;
+
+    shared.lock().unwrap().scoring_in_flight = false;
+    agent.state.lock().await.changed.notify_one();
+}
+
+async fn run_full(
+    agent: Arc<crate::agent::Agent>,
+    shared: Arc<std::sync::Mutex<MindState>>,
+) {
+    shared.lock().unwrap().scoring_in_flight = true;
+    agent.state.lock().await.changed.notify_one();
+
+    let client = agent.client.clone();
+    match score_memories(&client, &agent).await {
+        Ok(()) => {},
+        Err(e) => { dbglog!("[scoring-full] FAILED: {:#}", e); }
+    }
+
+    shared.lock().unwrap().scoring_in_flight = false;
+    agent.state.lock().await.changed.notify_one();
+}
+
 // ── Fine-tuning scoring ─────────────────────────────────────────
 
 /// Score which recent responses are candidates for fine-tuning.
@@ -429,8 +547,7 @@ pub async fn score_finetune(
         return Ok(Vec::new());
     }
 
-    let http = http_client();
-    let (divs, _) = score_divergence(&http, client, context, range, Filter::SkipAllMemories, Some(5)).await?;
+    let (divs, _) = score_divergence(client, context, range, is_memory_node, Some(5)).await?;
 
     let mut results: Vec<(usize, f64)> = response_positions.iter()
         .enumerate()
@@ -439,3 +556,319 @@ pub async fn score_finetune(
     results.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));
     Ok(results)
 }
+
+/// Enriched finetune candidate with context for review.
+#[derive(Clone, Debug)]
+pub struct FinetuneCandidate {
+    pub entry_idx: usize,
+    pub divergence: f64,
+    pub response_text: String,
+    /// Last couple of user/assistant messages before this response,
+    /// already rendered with role markers, for F6 display context.
+    pub prior_context: String,
+    /// Token IDs for context (everything before the response).
+    pub context_ids: Vec<u32>,
+    /// Token IDs for the response (what we're training on).
+    pub continuation_ids: Vec<u32>,
+    /// What the model would have said without memories (if generated).
+    pub alternate_text: Option<String>,
+    /// Timestamp in nanos — used as unique key for trained-set dedup.
+    pub timestamp_ns: i64,
+}
+
+/// Score and enrich finetune candidates with full context.
+///
+/// Candidates are delivered via `on_candidate` one-at-a-time as they become
+/// ready: scoring happens once (one /score call), then for each candidate
+/// that passes the threshold we optionally generate an alternate response
+/// and then emit it. The activity status is updated during the alternate
+/// phase so the UI doesn't look stuck.
+///
+/// Returns (count_above_threshold, max_divergence).
+pub async fn score_finetune_candidates(
+    context: &ContextState,
+    count: usize,
+    client: &ApiClient,
+    min_divergence: f64,
+    generate_alternates: bool,
+    activity: &crate::agent::ActivityGuard,
+    mut on_candidate: impl FnMut(FinetuneCandidate),
+) -> anyhow::Result<(usize, f64)> {
+    let scores = score_finetune(context, count, client).await?;
+
+    let max_divergence = scores.iter().map(|(_, d)| *d).fold(0.0f64, f64::max);
+
+    let entries = context.conversation();
+    let trained = load_trained();
+    let mut candidates: Vec<FinetuneCandidate> = Vec::new();
+
+    for (entry_idx, divergence) in scores {
+        if divergence < min_divergence {
+            continue;
+        }
+
+        let node = &entries[entry_idx];
+
+        // Skip if already trained on.
+        let timestamp_ns = node_timestamp_ns(node);
+        if trained.contains(&timestamp_ns) {
+            continue;
+        }
+
+        // Extract response text — content of the assistant turn.
+        let response_text = match node {
+            AstNode::Branch { children, .. } => render_branch_text(children),
+            _ => continue,
+        };
+
+        // Skip turns that produced nothing human-visible (e.g., a
+        // tool-only turn, or an interrupted generation). They'd show
+        // up as blank cards and we'd still burn alternate-gen on them.
+        if response_text.trim().is_empty() {
+            continue;
+        }
+
+        // Build the last couple of user/assistant exchanges for review.
+        let prior_context = render_prior_context(entries, entry_idx, 2);
+
+        // Build token IDs: context = everything before response, continuation = response.
+        let (context_ids, _, _) = context.wire_prompt(0..entry_idx, |_| false);
+        let continuation_ids: Vec<u32> = node.token_ids().into_iter().collect();
+
+        candidates.push(FinetuneCandidate {
+            entry_idx,
+            divergence,
+            response_text,
+            prior_context,
+            context_ids,
+            continuation_ids,
+            alternate_text: None,
+            timestamp_ns,
+        });
+    }
+
+    let total = candidates.len();
+    let gen_alternates = generate_alternates && total > 0;
+
+    for (i, mut candidate) in candidates.into_iter().enumerate() {
+        if gen_alternates {
+            activity.update(
+                format!("finetune: generating alternate {}/{}", i + 1, total)
+            ).await;
+            match gen_continuation(context, candidate.entry_idx, is_memory_node, client).await {
+                Ok(text) => candidate.alternate_text = Some(text),
+                Err(e) => dbglog!("[finetune] alternate generation failed: {:#}", e),
+            }
+        }
+        on_candidate(candidate);
+    }
+
+    Ok((total, max_divergence))
+}
+
+/// Stats from a finetune scoring run. Stored on MindState for UI display.
+#[derive(Clone, Debug)]
+pub struct FinetuneScoringStats {
+    pub responses_considered: usize,
+    pub above_threshold: usize,
+    pub threshold: f64,
+    pub max_divergence: f64,
+    pub error: Option<String>,
+}
+
+/// Finetune scoring — `trigger()` aborts any in-flight run and starts
+/// a fresh one, clearing the previous candidates.
+pub struct FinetuneScoring {
+    agent: Arc<crate::agent::Agent>,
+    shared: Arc<std::sync::Mutex<MindState>>,
+    task: TaskHandle,
+}
+
+impl FinetuneScoring {
+    pub fn new(
+        agent: Arc<crate::agent::Agent>,
+        shared: Arc<std::sync::Mutex<MindState>>,
+    ) -> Self {
+        Self { agent, shared, task: TaskHandle::new() }
+    }
+}
+
+impl MindTriggered for FinetuneScoring {
+    fn trigger(&self) {
+        self.task.trigger(run_finetune(self.agent.clone(), self.shared.clone()));
+    }
+}
+
+async fn run_finetune(
+    agent: Arc<crate::agent::Agent>,
+    shared: Arc<std::sync::Mutex<MindState>>,
+) {
+    let (threshold, gen_alternates) = {
+        let app = crate::config::app();
+        (app.learn.threshold, app.learn.generate_alternates)
+    };
+
+    // Fresh run — clear previous candidates.
+    shared.lock().unwrap().finetune_candidates.clear();
+    agent.state.lock().await.changed.notify_one();
+
+    let activity = crate::agent::start_activity(&agent, "finetune: scoring...").await;
+
+    let (context, client) = {
+        let ctx = agent.context.lock().await;
+        (ctx.clone(), agent.client.clone())
+    };
+
+    let entries = context.conversation();
+    let score_count = entries.len() / 2;
+    let range_start = entries.len() - score_count;
+    let responses_considered: usize = entries[range_start..].iter()
+        .filter(|n| matches!(n, AstNode::Branch { role: Role::Assistant, .. }))
+        .count();
+
+    activity.update(format!("finetune: scoring {} responses...", responses_considered)).await;
+
+    let stats = {
+        let shared = shared.clone();
+        let agent = agent.clone();
+        match score_finetune_candidates(
+            &context, score_count, &client, threshold,
+            gen_alternates, &activity,
+            move |c| {
+                shared.lock().unwrap().finetune_candidates.push(c);
+                if let Ok(st) = agent.state.try_lock() { st.changed.notify_one(); }
+            },
+        ).await {
+            Ok((above_threshold, max_div)) => FinetuneScoringStats {
+                responses_considered,
+                above_threshold,
+                threshold,
+                max_divergence: max_div,
+                error: None,
+            },
+            Err(e) => FinetuneScoringStats {
+                responses_considered,
+                above_threshold: 0,
+                threshold,
+                max_divergence: 0.0,
+                error: Some(format!("{}", e)),
+            },
+        }
+    };
+
+    shared.lock().unwrap().finetune_last_run = Some(stats);
+    agent.state.lock().await.changed.notify_one();
+}
+
+// ── Finetune config and persistence ─────────────────────────────
+
+use std::path::PathBuf;
+use std::collections::HashSet;
+
+const TRAINED_RESPONSES_FILE: &str = ".consciousness/cache/trained-responses.json";
+
+fn trained_path() -> PathBuf {
+    dirs::home_dir().unwrap_or_default().join(TRAINED_RESPONSES_FILE)
+}
+
+/// Load set of trained response timestamps (nanos since epoch).
+pub fn load_trained() -> HashSet<i64> {
+    let path = trained_path();
+    match std::fs::read_to_string(&path) {
+        Ok(content) => serde_json::from_str(&content).unwrap_or_default(),
+        Err(_) => HashSet::new(),
+    }
+}
+
+/// Mark a response as trained by its timestamp.
+pub fn mark_trained(timestamp_ns: i64) {
+    let mut trained = load_trained();
+    trained.insert(timestamp_ns);
+    let path = trained_path();
+    if let Some(parent) = path.parent() {
+        let _ = std::fs::create_dir_all(parent);
+    }
+    if let Ok(json) = serde_json::to_string(&trained) {
+        let _ = std::fs::write(&path, json);
+    }
+}
+
+/// Get timestamp in nanoseconds from an AstNode.
+/// i64-ns representation covers 1677..2262 via chrono; timestamps
+/// outside that window would be bugs we'd want to surface, hence panic.
+pub fn node_timestamp_ns(node: &AstNode) -> i64 {
+    let ts = match node {
+        AstNode::Leaf(leaf) => leaf.timestamp(),
+        AstNode::Branch { timestamp, .. } => *timestamp,
+    };
+    ts.timestamp_nanos_opt()
+        .expect("timestamp outside i64-ns representable range (1677..2262)")
+}
+
+// ── Training API ────────────────────────────────────────────────
+
+/// Training sample for /train endpoint.
+#[derive(serde::Serialize)]
+struct TrainingSample {
+    context_ids: Vec<u32>,
+    continuation_ids: Vec<u32>,
+}
+
+/// Data needed to send a training sample.
+pub struct TrainData {
+    pub context_ids: Vec<u32>,
+    pub continuation_ids: Vec<u32>,
+    pub timestamp_ns: i64,
+}
+
+/// Send training samples to the server.
+///
+/// Returns job_id on success, marks each sample as trained.
+pub async fn send_to_train(
+    samples: Vec<TrainData>,
+    client: &ApiClient,
+) -> anyhow::Result<String> {
+    if samples.is_empty() {
+        anyhow::bail!("no samples to train");
+    }
+
+    let api_samples: Vec<TrainingSample> = samples.iter()
+        .map(|s| TrainingSample {
+            context_ids: s.context_ids.clone(),
+            continuation_ids: s.continuation_ids.clone(),
+        })
+        .collect();
+
+    let body = serde_json::json!({
+        "training_data": {
+            "samples": api_samples,
+        }
+    });
+
+    let url = format!("{}/train", client.base_url());
+    let http = crate::agent::api::http::HttpClient::builder()
+        .timeout(std::time::Duration::from_secs(300))
+        .build();
+    let response = http.send_json("POST", &url, &[], &body).await?;
+
+    let status = response.status();
+    let result: serde_json::Value = response.json().await?;
+
+    if !status.is_success() {
+        let msg = result.get("error").and_then(|e| e.as_str()).unwrap_or("unknown error");
+        anyhow::bail!("train API HTTP {}: {}", status, msg);
+    }
+
+    // Mark all samples as trained
+    for s in &samples {
+        mark_trained(s.timestamp_ns);
+    }
+
+    let job_id = result.get("job_id")
+        .and_then(|j| j.as_str())
+        .unwrap_or("unknown")
+        .to_string();
+
+    dbglog!("[finetune] sent {} samples, job_id={}", samples.len(), job_id);
+    Ok(job_id)
+}
diff --git a/src/subconscious/mod.rs b/src/subconscious/mod.rs
index 433f721..1abf25a 100644
--- a/src/subconscious/mod.rs
+++ b/src/subconscious/mod.rs
@@ -1,7 +1,9 @@
 // Agent layer: LLM-powered operations on the memory graph
 
+pub mod compare;
 pub mod daemon;
 pub mod defs;
 pub mod digest;
+pub mod generate;
 pub mod learn;
 pub mod prompts;
diff --git a/src/subconscious/prompts.rs b/src/subconscious/prompts.rs
index c98b887..b9a0146 100644
--- a/src/subconscious/prompts.rs
+++ b/src/subconscious/prompts.rs
@@ -104,22 +104,21 @@ pub fn format_nodes_section(store: &Store, items: &[ReplayItem], graph: &Graph)
                 item.classification, item.outlier_score));
         }
 
-        if let Some(community) = node.community_id {
-            out.push_str(&format!("Community: {}  ", community));
-        }
-        let deg = graph.degree(&item.key);
-        let cc = graph.clustering_coefficient(&item.key);
+		if let Some(community) = node.community_id {
+			out.push_str(&format!("Community: {}  ", community));
+		}
+		let deg = graph.degree(&item.key);
 
-        // Hub-link ratio: what fraction of this node's edges go to hubs?
-        let neighbors = graph.neighbors(&item.key);
+		// Hub-link ratio: what fraction of this node's edges go to hubs?
+		let neighbors = graph.neighbors(&item.key);
         let hub_links = neighbors.iter()
             .filter(|(n, _)| graph.degree(n) >= hub_thresh)
             .count();
         let hub_ratio = if deg > 0 { hub_links as f32 / deg as f32 } else { 0.0 };
-        let is_hub = deg >= hub_thresh;
+		let is_hub = deg >= hub_thresh;
 
-        out.push_str(&format!("Degree: {}  CC: {:.3}  Hub-link ratio: {:.0}% ({}/{})",
-            deg, cc, hub_ratio * 100.0, hub_links, deg));
+		out.push_str(&format!("Degree: {}  CC: {:.3}  Hub-link ratio: {:.0}% ({}/{})",
+			deg, item.cc, hub_ratio * 100.0, hub_links, deg));
         if is_hub {
             out.push_str("  ← THIS IS A HUB");
         } else if hub_ratio > 0.6 {
diff --git a/src/thalamus/idle.rs b/src/thalamus/idle.rs
index 6c78b19..71baa81 100644
--- a/src/thalamus/idle.rs
+++ b/src/thalamus/idle.rs
@@ -372,6 +372,10 @@ impl State {
 }
 
 pub fn hours_since_last_dream() -> u64 {
+    // If a dream is currently in progress, no nudge needed
+    if home().join(".consciousness/state/dream-state").exists() {
+        return 0;
+    }
     let path = home().join(".consciousness/logs/dream-log.jsonl");
     let content = match fs::read_to_string(path) {
         Ok(c) if !c.is_empty() => c,
diff --git a/src/thalamus/supervisor.rs b/src/thalamus/supervisor.rs
index a4c53ec..3716682 100644
--- a/src/thalamus/supervisor.rs
+++ b/src/thalamus/supervisor.rs
@@ -19,6 +19,51 @@ fn channels_dir() -> PathBuf {
         .join(".consciousness/channels")
 }
 
+/// Install a SIGCHLD-driven reaper for channel daemons.
+///
+/// We can't use SIGCHLD=SIG_IGN because that makes the kernel auto-reap
+/// all children, and tokio::process::Command::wait() then returns ECHILD
+/// (breaking every tool that spawns a subprocess — bash, mcp clients, etc.).
+///
+/// Instead, on each SIGCHLD we read PID files in channels_dir() and call
+/// waitpid(pid, WNOHANG) on each. That reaps only our own zombie channel
+/// daemons; waitpid on any other PID returns ECHILD (harmless no-op).
+/// Tokio-spawned children aren't recorded in PID files, so tokio's own
+/// per-child wait paths are left free to reap them.
+pub fn start_zombie_reaper() -> tokio::task::JoinHandle<()> {
+    use tokio::signal::unix::{signal, SignalKind};
+    tokio::spawn(async move {
+        let mut sig = match signal(SignalKind::child()) {
+            Ok(s) => s,
+            Err(e) => {
+                error!("failed to install SIGCHLD handler: {}", e);
+                return;
+            }
+        };
+        while sig.recv().await.is_some() {
+            reap_channel_daemons();
+        }
+    })
+}
+
+fn reap_channel_daemons() {
+    let entries = match std::fs::read_dir(channels_dir()) {
+        Ok(e) => e,
+        Err(_) => return,
+    };
+    for entry in entries.flatten() {
+        let path = entry.path();
+        if path.extension().and_then(|s| s.to_str()) != Some("pid") {
+            continue;
+        }
+        let Ok(s) = std::fs::read_to_string(&path) else { continue };
+        let Ok(pid) = s.trim().parse::<i32>() else { continue };
+        let mut status = 0;
+        // Reaps our zombie child; ECHILD on non-child is a no-op.
+        unsafe { libc::waitpid(pid, &mut status, libc::WNOHANG); }
+    }
+}
+
 fn config_path() -> PathBuf {
     channels_dir().join("channels.json5")
 }
diff --git a/src/user/amygdala.rs b/src/user/amygdala.rs
new file mode 100644
index 0000000..7689bc0
--- /dev/null
+++ b/src/user/amygdala.rs
@@ -0,0 +1,400 @@
+// amygdala.rs — F8 amygdala screen: live per-token concept-readout
+// projections from the vLLM server's readout.safetensors.
+//
+// Left panel: top-K concepts by magnitude at the currently-selected
+// layer, as horizontal bars. The concept names come from the manifest
+// fetched at agent startup; the values come from the per-token readout
+// pushed onto agent.readout by the streaming token handler.
+//
+// Bottom: scrolling history of the last few tokens' top concept.
+//
+// Keys:
+//   1..9   select layer index (1 = first layer in the manifest)
+//   t      toggle between "current" (last token) and "mean over recent"
+
+use ratatui::{
+    layout::{Constraint, Direction, Layout, Rect},
+    style::{Color, Modifier, Style},
+    text::{Line, Span},
+    widgets::{Block, Borders, Gauge, Paragraph, Wrap},
+    Frame,
+};
+use ratatui::crossterm::event::{Event, KeyCode};
+
+use super::{App, ScreenView};
+use crate::agent::api::ReadoutManifest;
+use crate::agent::readout::ReadoutEntry;
+
+const TOP_K: usize = 20;
+/// Hysteresis band around TOP_K. A concept currently in the display
+/// is kept as long as its |z-score| rank stays in the top
+/// ``TOP_K + HYSTERESIS``; only falls out when it drops below that.
+/// Prevents the ticker-tape flicker that pure top-K sorting produces.
+const HYSTERESIS: usize = 20;
+
+pub(crate) struct AmygdalaScreen {
+    selected_layer: usize,
+    mode: DisplayMode,
+    /// Concept indices currently pinned in display order. Values at
+    /// these indices change every frame; the set only rotates when a
+    /// pinned concept drops out of the hysteresis band.
+    display_indices: Vec<usize>,
+    /// Whether to show z-scored values (default) or raw dot products.
+    normalize: bool,
+}
+
+#[derive(Clone, Copy, PartialEq)]
+enum DisplayMode {
+    /// Values from the single most recent token.
+    Current,
+    /// Mean over all tokens currently in the ring buffer.
+    MeanRecent,
+}
+
+impl AmygdalaScreen {
+    pub fn new() -> Self {
+        Self {
+            // Default to layer 62 — clean cross-cluster discrimination
+            // with good within-cluster cohesion. With the v2 deep
+            // manifest (layers 62, 63), index 0 = layer 62 and
+            // index 1 = layer 63 (sharper but noisier on some
+            // dimensions). Bounded down to actual layer count at
+            // render time.
+            selected_layer: 0,
+            mode: DisplayMode::MeanRecent,
+            display_indices: Vec::new(),
+            normalize: true,
+        }
+    }
+}
+
+impl ScreenView for AmygdalaScreen {
+    fn label(&self) -> &'static str { "amygdala" }
+
+    fn tick(&mut self, frame: &mut Frame, area: Rect,
+            events: &[Event], app: &mut App) {
+        for event in events {
+            if let Event::Key(key) = event {
+                match key.code {
+                    KeyCode::Char(c) if c.is_ascii_digit() && c != '0' => {
+                        let idx = (c as u8 - b'1') as usize;
+                        self.selected_layer = idx;
+                    }
+                    KeyCode::Char('t') => {
+                        self.mode = match self.mode {
+                            DisplayMode::Current => DisplayMode::MeanRecent,
+                            DisplayMode::MeanRecent => DisplayMode::Current,
+                        };
+                        // Re-pin on mode change; the relative
+                        // magnitudes between current-token and
+                        // mean-recent differ substantially.
+                        self.display_indices.clear();
+                    }
+                    KeyCode::Char('z') => {
+                        self.normalize = !self.normalize;
+                        self.display_indices.clear();
+                    }
+                    _ => {}
+                }
+            }
+        }
+
+        // Snapshot the shared buffer with a short lock.
+        let snapshot = match app.agent.readout.lock() {
+            Ok(buf) => {
+                if !buf.is_enabled() {
+                    render_disabled(frame, area);
+                    return;
+                }
+                let manifest = buf.manifest.clone().unwrap();
+                let entries: Vec<ReadoutEntry> =
+                    buf.recent.iter().cloned().collect();
+                (manifest, entries)
+            }
+            Err(_) => {
+                render_disabled(frame, area);
+                return;
+            }
+        };
+        let (manifest, entries) = snapshot;
+
+        // Bound the selected layer to what the manifest actually has.
+        let n_layers = manifest.layers.len();
+        if self.selected_layer >= n_layers {
+            self.selected_layer = 0;
+        }
+
+        // Compute the raw values for the selected layer: either the
+        // latest token's row, or the mean across recent tokens. Raw
+        // means un-normalized dot products — their absolute scale is
+        // dominated by residual-stream norm, not concept alignment.
+        let raw: Option<Vec<f32>> = match self.mode {
+            DisplayMode::Current => entries
+                .last()
+                .and_then(|e| e.readout.get(self.selected_layer).cloned()),
+            DisplayMode::MeanRecent => mean_layer(&entries, self.selected_layer),
+        };
+
+        // Optional z-score normalization: remove the per-layer mean,
+        // scale by std. Result is "σ above/below the concept-vector
+        // average at this layer" — the loud-residual-stream scaling
+        // factor cancels out, values become comparable across frames.
+        let display_values = raw.as_ref().map(|v| {
+            if self.normalize { z_score(v) } else { v.clone() }
+        });
+
+        // Update the pinned display set with hysteresis: a concept
+        // stays pinned while it remains in the top (TOP_K + HYSTERESIS)
+        // by |value|; falls out only when it drops below that band.
+        // Keeps rows stable while values update in place.
+        if let Some(v) = display_values.as_ref() {
+            self.refresh_display_indices(v);
+        }
+
+        let layout = Layout::default()
+            .direction(Direction::Vertical)
+            .constraints([
+                Constraint::Length(3),  // header
+                Constraint::Min(10),    // bars
+                Constraint::Length(6),  // recent tokens
+            ])
+            .split(area);
+
+        render_header(frame, layout[0], &manifest, self.selected_layer,
+                      self.mode, entries.len(), self.normalize);
+        match display_values {
+            Some(v) => render_bars(
+                frame, layout[1], &manifest.concepts, &v,
+                &self.display_indices, self.normalize,
+            ),
+            None => render_empty_bars(frame, layout[1]),
+        }
+        render_recent(frame, layout[2], &entries, self.selected_layer,
+                      &manifest.concepts);
+    }
+}
+
+impl AmygdalaScreen {
+    /// Add concepts entering the hysteresis band; evict concepts that
+    /// dropped out. Preserves existing order for concepts that stay.
+    fn refresh_display_indices(&mut self, values: &[f32]) {
+        let n = values.len();
+        if n == 0 {
+            return;
+        }
+        // Rank all concepts by |value| desc so we can check both "in
+        // strict top-K" and "in hysteresis band (top K + H)" cheaply.
+        let mut rank: Vec<(usize, f32)> = values.iter()
+            .enumerate().map(|(i, v)| (i, v.abs())).collect();
+        rank.sort_by(|a, b| b.1.partial_cmp(&a.1)
+            .unwrap_or(std::cmp::Ordering::Equal));
+        let hyst_cutoff = (TOP_K + HYSTERESIS).min(n);
+        let in_band: std::collections::HashSet<usize> =
+            rank.iter().take(hyst_cutoff).map(|(i, _)| *i).collect();
+        // Drop anything that left the band.
+        self.display_indices.retain(|i| in_band.contains(i));
+        // Fill up to TOP_K by walking the top-K-by-|value| and adding
+        // any concept not already displayed.
+        for (i, _) in rank.iter().take(TOP_K) {
+            if self.display_indices.len() >= TOP_K {
+                break;
+            }
+            if !self.display_indices.contains(i) {
+                self.display_indices.push(*i);
+            }
+        }
+    }
+}
+
+fn render_disabled(frame: &mut Frame, area: Rect) {
+    let text = Paragraph::new(Line::from(vec![
+        Span::raw("readout disabled — server did not return a manifest. "),
+        Span::styled("Start vLLM with ", Style::default().fg(Color::DarkGray)),
+        Span::styled("VLLM_READOUT_MANIFEST", Style::default().fg(Color::Yellow)),
+        Span::styled(" + ", Style::default().fg(Color::DarkGray)),
+        Span::styled("VLLM_READOUT_VECTORS", Style::default().fg(Color::Yellow)),
+        Span::styled(".", Style::default().fg(Color::DarkGray)),
+    ]))
+    .wrap(Wrap { trim: true })
+    .block(Block::default().borders(Borders::ALL).title("amygdala"));
+    frame.render_widget(text, area);
+}
+
+fn render_header(frame: &mut Frame, area: Rect, manifest: &ReadoutManifest,
+                 selected: usize, mode: DisplayMode, n_tokens: usize,
+                 normalize: bool) {
+    let mode_str = match mode {
+        DisplayMode::Current => "current",
+        DisplayMode::MeanRecent => "mean(recent)",
+    };
+    let scale_str = if normalize { "z-score" } else { "raw" };
+    let layer = manifest.layers.get(selected).copied().unwrap_or(0);
+    let spans = vec![
+        Span::styled("layer ", Style::default().fg(Color::DarkGray)),
+        Span::styled(
+            format!("{}/{}  ", selected + 1, manifest.layers.len()),
+            Style::default().add_modifier(Modifier::BOLD),
+        ),
+        Span::styled("(index ", Style::default().fg(Color::DarkGray)),
+        Span::styled(format!("{}", layer), Style::default().fg(Color::Cyan)),
+        Span::styled(")   ", Style::default().fg(Color::DarkGray)),
+        Span::styled("mode ", Style::default().fg(Color::DarkGray)),
+        Span::styled(mode_str, Style::default().fg(Color::Yellow)),
+        Span::styled("   scale ", Style::default().fg(Color::DarkGray)),
+        Span::styled(scale_str, Style::default().fg(Color::Yellow)),
+        Span::styled("   ", Style::default()),
+        Span::styled(
+            format!("{} toks in ring", n_tokens),
+            Style::default().fg(Color::DarkGray),
+        ),
+        Span::raw("       "),
+        Span::styled(
+            format!("[1-{}] layer   [t] mode   [z] z-score/raw",
+                    manifest.layers.len().min(9)),
+            Style::default().fg(Color::DarkGray),
+        ),
+    ];
+    let para = Paragraph::new(Line::from(spans))
+        .block(Block::default().borders(Borders::ALL).title("amygdala"));
+    frame.render_widget(para, area);
+}
+
+fn render_bars(frame: &mut Frame, area: Rect,
+               concepts: &[String], values: &[f32],
+               display_indices: &[usize], normalize: bool) {
+    let inner = Block::default().borders(Borders::ALL)
+        .title("top concepts");
+    let inner_area = inner.inner(area);
+    frame.render_widget(inner, area);
+
+    if inner_area.height == 0 || display_indices.is_empty() {
+        return;
+    }
+
+    // Bar-scale normalization. For z-score mode, pin the bar to a
+    // fixed reference (|z| = 3 = full bar) so the visual magnitude
+    // has a meaningful interpretation ("3σ from baseline"). For raw
+    // mode, fall back to the old behavior (scale to the loudest
+    // concept on-screen).
+    let scale_ref: f32 = if normalize {
+        3.0
+    } else {
+        display_indices.iter()
+            .filter_map(|&i| values.get(i))
+            .map(|v| v.abs())
+            .fold(0.0_f32, f32::max)
+            .max(1e-6)
+    };
+
+    let rows = (inner_area.height as usize).min(display_indices.len());
+    let row_constraints: Vec<Constraint> =
+        std::iter::repeat(Constraint::Length(1)).take(rows).collect();
+    let chunks = Layout::default()
+        .direction(Direction::Vertical)
+        .constraints(row_constraints)
+        .split(inner_area);
+
+    for (row, &c_idx) in display_indices.iter().take(rows).enumerate() {
+        let v = values.get(c_idx).copied().unwrap_or(0.0);
+        let label = concepts.get(c_idx).cloned()
+            .unwrap_or_else(|| format!("c{}", c_idx));
+        let ratio = (v.abs() / scale_ref).clamp(0.0, 1.0);
+        let color = if v >= 0.0 { Color::Green } else { Color::Red };
+        let display_num = if normalize {
+            format!("{:+.2}σ", v)
+        } else {
+            format!("{:+.3}", v)
+        };
+        let gauge = Gauge::default()
+            .ratio(ratio as f64)
+            .gauge_style(Style::default().fg(color).bg(Color::Reset))
+            .label(format!("{:<26} {}", truncate_name(&label, 26), display_num));
+        frame.render_widget(gauge, chunks[row]);
+    }
+}
+
+fn render_empty_bars(frame: &mut Frame, area: Rect) {
+    let para = Paragraph::new(Line::from(Span::styled(
+        "waiting for tokens…",
+        Style::default().fg(Color::DarkGray),
+    )))
+    .block(Block::default().borders(Borders::ALL).title("top concepts"));
+    frame.render_widget(para, area);
+}
+
+fn render_recent(frame: &mut Frame, area: Rect, entries: &[ReadoutEntry],
+                 layer: usize, concepts: &[String]) {
+    let mut lines: Vec<Line> = Vec::new();
+    for entry in entries.iter().rev().take(4) {
+        let row = match entry.readout.get(layer) {
+            Some(r) => r,
+            None => continue,
+        };
+        // top concept at this layer for this token
+        let (best_idx, best_val) = row.iter().enumerate()
+            .fold((0, 0.0_f32), |acc, (i, v)| {
+                if v.abs() > acc.1.abs() { (i, *v) } else { acc }
+            });
+        let name = concepts.get(best_idx).cloned()
+            .unwrap_or_else(|| format!("c{}", best_idx));
+        let tok_str = format!("t{:>5}", entry.token_id);
+        lines.push(Line::from(vec![
+            Span::styled(tok_str, Style::default().fg(Color::DarkGray)),
+            Span::raw("  "),
+            Span::styled(
+                format!("{:<24}", truncate_name(&name, 24)),
+                Style::default().fg(
+                    if best_val >= 0.0 { Color::Green } else { Color::Red },
+                ),
+            ),
+            Span::styled(
+                format!(" {:+.3}", best_val),
+                Style::default().add_modifier(Modifier::BOLD),
+            ),
+        ]));
+    }
+    let para = Paragraph::new(lines)
+        .block(Block::default().borders(Borders::ALL).title("recent tokens — top concept"));
+    frame.render_widget(para, area);
+}
+
+/// Z-score normalize: `(v - mean) / std` across the concept axis.
+/// Result is comparable across frames and layers (the residual-stream
+/// magnitude factors out) and has the nice property that "this is
+/// ≥2σ elevated" has a concrete meaning regardless of scale.
+fn z_score(values: &[f32]) -> Vec<f32> {
+    let n = values.len() as f32;
+    if n == 0.0 {
+        return Vec::new();
+    }
+    let mean = values.iter().sum::<f32>() / n;
+    let var = values.iter()
+        .map(|v| (v - mean) * (v - mean))
+        .sum::<f32>() / n;
+    let std = var.sqrt().max(1e-6);
+    values.iter().map(|v| (v - mean) / std).collect()
+}
+
+fn mean_layer(entries: &[ReadoutEntry], layer: usize) -> Option<Vec<f32>> {
+    let rows: Vec<&Vec<f32>> = entries.iter()
+        .filter_map(|e| e.readout.get(layer))
+        .collect();
+    if rows.is_empty() {
+        return None;
+    }
+    let n_concepts = rows[0].len();
+    let mut acc = vec![0.0_f32; n_concepts];
+    for r in &rows {
+        for (i, v) in r.iter().enumerate() {
+            acc[i] += *v;
+        }
+    }
+    let n = rows.len() as f32;
+    for v in &mut acc { *v /= n; }
+    Some(acc)
+}
+
+fn truncate_name(s: &str, max: usize) -> String {
+    if s.len() <= max { s.to_string() }
+    else { format!("{}…", &s[..max.saturating_sub(1)]) }
+}
diff --git a/src/user/chat.rs b/src/user/chat.rs
index a94e039..bd2df25 100644
--- a/src/user/chat.rs
+++ b/src/user/chat.rs
@@ -112,13 +112,7 @@ pub async fn cmd_switch_model(
     let _new_client = crate::agent::api::ApiClient::new(
         &resolved.api_base, &resolved.api_key, &resolved.model_id,
     );
-    let prompt_changed = resolved.prompt_file != agent.prompt_file;
-    if prompt_changed {
-        agent.compact().await;
-        agent.state.lock().await.notify(format!("switched to {} (recompacted)", resolved.model_id));
-    } else {
-        agent.state.lock().await.notify(format!("switched to {}", resolved.model_id));
-    }
+    agent.state.lock().await.notify(format!("switched to {}", resolved.model_id));
 }
 
 fn notify_help(agent: &std::sync::Arc<crate::agent::Agent>) {
@@ -173,6 +167,7 @@ enum PaneTarget {
     ConversationAssistant,
     Tools,
     ToolResult,
+    Autonomous,
 }
 
 const MAX_PANE_LINES: usize = 10_000;
@@ -478,8 +473,11 @@ impl InteractScreen {
             AstNode::Leaf(leaf) => {
                 let text = leaf.body().text().to_string();
                 match leaf.body() {
-                    NodeBody::Memory { .. } | NodeBody::Thinking(_)
-                        | NodeBody::Log(_) | NodeBody::Dmn(_) => vec![],
+                    NodeBody::Memory { .. } | NodeBody::Log(_) | NodeBody::Dmn(_) => vec![],
+                    NodeBody::Thinking(_) => {
+                        if text.is_empty() { vec![] }
+                        else { vec![(PaneTarget::Autonomous, text, Marker::None)] }
+                    }
                     NodeBody::Content(_) => {
                         if text.is_empty() || text.starts_with("<system-reminder>") { vec![] }
                         else { vec![(PaneTarget::Conversation, text, Marker::User)] }
@@ -492,6 +490,11 @@ impl InteractScreen {
                         if t.is_empty() { vec![] }
                         else { vec![(PaneTarget::ToolResult, text, Marker::None)] }
                     }
+                    NodeBody::Image { orig_height, orig_width, .. } => {
+                        vec![(PaneTarget::Conversation,
+                              format!("[image {}x{}]", orig_width, orig_height),
+                              Marker::None)]
+                    }
                 }
             }
             AstNode::Branch { role, children, .. } => {
@@ -548,6 +551,12 @@ impl InteractScreen {
                         self.tools.push_line(format!("  {}", line), Color::DarkGray);
                     }
                 }
+                PaneTarget::Autonomous => {
+                    self.autonomous.current_color = Color::Gray;
+                    self.autonomous.append_text(&text);
+                    self.autonomous.pending_marker = marker;
+                    self.autonomous.flush_pending();
+                }
             }
         }
     }
@@ -559,6 +568,8 @@ impl InteractScreen {
                     => self.conversation.pop_line(),
                 PaneTarget::Tools | PaneTarget::ToolResult
                     => self.tools.pop_line(),
+                PaneTarget::Autonomous
+                    => self.autonomous.pop_line(),
             }
         }
     }
diff --git a/src/user/compare.rs b/src/user/compare.rs
new file mode 100644
index 0000000..2969b91
--- /dev/null
+++ b/src/user/compare.rs
@@ -0,0 +1,111 @@
+// compare.rs — F7 compare screen: side-by-side test-model regen of
+// every assistant response in the current context.
+
+use ratatui::{
+    layout::Rect,
+    style::{Color, Modifier, Style},
+    text::{Line, Span},
+    widgets::{Block, Borders, List, ListItem, ListState, Paragraph, Wrap},
+    Frame,
+};
+use ratatui::crossterm::event::{Event, KeyCode};
+
+use super::{App, ScreenView, truncate, widgets};
+
+pub use crate::subconscious::compare::CompareCandidate;
+
+pub(crate) struct CompareScreen {
+    list_state: ListState,
+    mind_tx: tokio::sync::mpsc::UnboundedSender<crate::mind::MindCommand>,
+}
+
+impl CompareScreen {
+    pub fn new(
+        mind_tx: tokio::sync::mpsc::UnboundedSender<crate::mind::MindCommand>,
+    ) -> Self {
+        Self { list_state: ListState::default(), mind_tx }
+    }
+}
+
+impl ScreenView for CompareScreen {
+    fn label(&self) -> &'static str { "compare" }
+
+    fn tick(&mut self, frame: &mut Frame, area: Rect,
+            events: &[Event], app: &mut App) {
+        widgets::handle_list_nav(events, &mut self.list_state,
+            app.compare_candidates.len(), |code| match code {
+                KeyCode::Char('c') | KeyCode::Enter => {
+                    let _ = self.mind_tx.send(crate::mind::MindCommand::Compare);
+                }
+                _ => {}
+            });
+
+        let (settings_area, content_area, help_area) =
+            widgets::candidate_frame(frame, area, "compare");
+
+        let test_backend = crate::config::app().compare.test_backend.clone();
+        let (label, color) = if test_backend.is_empty() {
+            ("(unset — set compare.test_backend)".to_string(), Color::Red)
+        } else {
+            (test_backend, Color::Yellow)
+        };
+        frame.render_widget(Paragraph::new(Line::from(vec![
+            Span::raw(" test model: "),
+            Span::styled(label, Style::default().fg(color)),
+        ])), settings_area);
+
+        let candidates = &app.compare_candidates;
+        if candidates.is_empty() {
+            let err = app.mind_state.as_ref().and_then(|ms| ms.compare_error.as_deref());
+            let mut lines = vec![Line::from(""),
+                Line::styled("  Press c/Enter to compare against the configured test model.",
+                    Style::default().fg(Color::DarkGray))];
+            if let Some(e) = err {
+                lines.push(Line::from(""));
+                lines.push(Line::from(vec![
+                    Span::raw("  "),
+                    Span::styled(format!("error: {}", e), Style::default().fg(Color::Red)),
+                ]));
+            }
+            frame.render_widget(Paragraph::new(lines), content_area);
+        } else {
+            let (list_area, detail_area) = widgets::list_detail_split(content_area);
+
+            let items: Vec<ListItem> = candidates.iter().map(|c| ListItem::new(Line::from(vec![
+                Span::styled(format!("#{:<3} ", c.entry_idx), Style::default().fg(Color::DarkGray)),
+                Span::raw(truncate(&c.original_text, 30)),
+            ]))).collect();
+            frame.render_stateful_widget(
+                List::new(items)
+                    .block(Block::default().borders(Borders::RIGHT).title(" candidates "))
+                    .highlight_style(Style::default().add_modifier(Modifier::REVERSED)),
+                list_area, &mut self.list_state,
+            );
+
+            if let Some(c) = self.list_state.selected().and_then(|i| candidates.get(i)) {
+                let mut text = String::new();
+                if !c.prior_context.is_empty() {
+                    text.push_str(&c.prior_context);
+                    text.push_str("\n\n─── original ───\n\n");
+                }
+                text.push_str(&c.original_text);
+                text.push_str("\n\n─── test model ───\n\n");
+                text.push_str(&c.alternate_text);
+                frame.render_widget(
+                    Paragraph::new(text)
+                        .block(Block::default().borders(Borders::TOP)
+                            .title(format!(" entry {} ", c.entry_idx)))
+                        .wrap(Wrap { trim: false }),
+                    detail_area,
+                );
+            }
+        }
+
+        frame.render_widget(Paragraph::new(Line::from(vec![
+            Span::styled(" j/k/\u{2191}\u{2193}", Style::default().fg(Color::Cyan)),
+            Span::raw("=nav  "),
+            Span::styled("c/Enter", Style::default().fg(Color::Green)),
+            Span::raw("=run "),
+        ])), help_area);
+    }
+}
diff --git a/src/user/context.rs b/src/user/context.rs
index a0692fa..8edd926 100644
--- a/src/user/context.rs
+++ b/src/user/context.rs
@@ -38,16 +38,14 @@ impl ConsciousScreen {
         for node in ctx.conversation() {
             if let AstNode::Leaf(leaf) = node {
                 if let NodeBody::Memory { key, score, text } = leaf.body() {
-                    let status = match score {
-                        Some(s) => { scored += 1; format!("{:.2}", s) }
-                        None => { unscored += 1; String::new() }
-                    };
+                    if score.is_some() { scored += 1; } else { unscored += 1; }
                     mem_children.push(SectionView {
-                        name: key.clone(),
+                        name: format!("mem: {}", key),
                         tokens: node.tokens(),
                         content: text.clone(),
+                        token_ids: leaf.token_ids().to_vec(),
                         children: Vec::new(),
-                        status,
+                        status: score.map(|s| format!("{:.2}", s)).unwrap_or_default(),
                     });
                 }
             }
@@ -58,6 +56,7 @@ impl ConsciousScreen {
                 name: format!("Memory nodes ({})", mem_children.len()),
                 tokens: mem_tokens,
                 content: String::new(),
+                token_ids: Vec::new(),
                 children: mem_children,
                 status: format!("{} scored, {} unscored", scored, unscored),
             });
@@ -73,11 +72,13 @@ impl ConsciousScreen {
                     AstNode::Leaf(leaf) => leaf.body().text().to_string(),
                     _ => String::new(),
                 },
+                token_ids: node.token_ids(),
                 children: match node {
                     AstNode::Branch { children, .. } => children.iter()
                         .map(|c| SectionView {
                             name: c.label(), tokens: c.tokens(),
                             content: match c { AstNode::Leaf(l) => l.body().text().to_string(), _ => String::new() },
+                            token_ids: match c { AstNode::Leaf(l) => l.token_ids().to_vec(), _ => c.token_ids() },
                             children: Vec::new(), status: String::new(),
                         }).collect(),
                     _ => Vec::new(),
@@ -104,6 +105,7 @@ impl ConsciousScreen {
             name: format!("Conversation ({} entries)", conv_children.len()),
             tokens: conv_tokens,
             content: String::new(),
+            token_ids: Vec::new(),
             children: conv_children,
             status: String::new(),
         });
@@ -129,14 +131,7 @@ impl ScreenView for ConsciousScreen {
         let section_style = Style::default().fg(Color::Yellow);
 
         lines.push(Line::styled("── Model ──", section_style));
-        let model_display = app.context_info.as_ref()
-            .map_or_else(|| app.status.model.clone(), |i| i.model.clone());
-        lines.push(Line::raw(format!("  Current: {}", model_display)));
-        if let Some(ref info) = app.context_info {
-            lines.push(Line::raw(format!("  Backend: {}", info.backend)));
-            lines.push(Line::raw(format!("  Prompt:  {}", info.prompt_file)));
-            lines.push(Line::raw(format!("  Available: {}", info.available_models.join(", "))));
-        }
+        lines.push(Line::raw(format!("  Current: {}", app.status.model)));
         lines.push(Line::raw(""));
 
         lines.push(Line::styled("── Context State ──", section_style));
@@ -156,8 +151,6 @@ impl ScreenView for ConsciousScreen {
 
             lines.push(Line::raw(format!("  {:53} {:>6} tokens", "────────", "──────")));
             lines.push(Line::raw(format!("  {:53} {:>6} tokens", "Total", total)));
-        } else if let Some(ref info) = app.context_info {
-            lines.push(Line::raw(format!("  Context message: {:>6} chars", info.context_message_chars)));
         }
         lines.push(Line::raw(""));
 
diff --git a/src/user/learn.rs b/src/user/learn.rs
new file mode 100644
index 0000000..7984bab
--- /dev/null
+++ b/src/user/learn.rs
@@ -0,0 +1,284 @@
+// learn.rs — F6: fine-tuning review screen
+//
+// Shows responses identified as training candidates (high divergence
+// when memories stripped). Queue for review before sending to /finetune.
+
+use ratatui::{
+    layout::{Constraint, Layout, Rect},
+    style::{Color, Modifier, Style},
+    text::{Line, Span},
+    widgets::{Block, Borders, List, ListItem, ListState, Paragraph, Wrap},
+    Frame,
+};
+use ratatui::crossterm::event::{Event, KeyCode};
+
+use super::{App, ScreenView, truncate, widgets};
+
+/// A candidate response identified for fine-tuning.
+#[derive(Clone, Debug)]
+pub struct FinetuneCandidate {
+    /// Index in conversation entries.
+    pub entry_idx: usize,
+    /// Divergence score (higher = more dependent on memories).
+    pub divergence: f64,
+    /// The assistant response text.
+    pub response_text: String,
+    /// Prior user/assistant messages for review context.
+    pub prior_context: String,
+    /// Status: pending, approved, rejected, sent.
+    pub status: CandidateStatus,
+    /// Token IDs for context.
+    pub context_ids: Vec<u32>,
+    /// Token IDs for continuation (what we're training on).
+    pub continuation_ids: Vec<u32>,
+    /// What the model would have said without memories (if generated).
+    pub alternate_text: Option<String>,
+    /// Timestamp in nanos — used as unique key for trained-set dedup.
+    pub timestamp_ns: i64,
+}
+
+#[derive(Clone, Debug, PartialEq)]
+pub enum CandidateStatus {
+    Pending,
+    Approved,
+    Rejected,
+    Sent,
+}
+
+impl From<crate::subconscious::learn::FinetuneCandidate> for FinetuneCandidate {
+    fn from(c: crate::subconscious::learn::FinetuneCandidate) -> Self {
+        FinetuneCandidate {
+            entry_idx: c.entry_idx,
+            divergence: c.divergence,
+            response_text: c.response_text,
+            prior_context: c.prior_context,
+            status: CandidateStatus::Pending,
+            context_ids: c.context_ids,
+            continuation_ids: c.continuation_ids,
+            alternate_text: c.alternate_text,
+            timestamp_ns: c.timestamp_ns,
+        }
+    }
+}
+
+pub(crate) struct LearnScreen {
+    list_state: ListState,
+    mind_tx: tokio::sync::mpsc::UnboundedSender<crate::mind::MindCommand>,
+}
+
+impl LearnScreen {
+    pub fn new(
+        mind_tx: tokio::sync::mpsc::UnboundedSender<crate::mind::MindCommand>,
+    ) -> Self {
+        Self {
+            list_state: ListState::default(),
+            mind_tx,
+        }
+    }
+
+    fn selected_idx(&self) -> Option<usize> {
+        self.list_state.selected()
+    }
+}
+
+impl ScreenView for LearnScreen {
+    fn label(&self) -> &'static str { "learn" }
+
+    fn tick(&mut self, frame: &mut Frame, area: Rect,
+            events: &[Event], app: &mut App) {
+        let selected_idx = self.list_state.selected();
+        widgets::handle_list_nav(events, &mut self.list_state,
+            app.finetune_candidates.len(), |code| match code {
+                KeyCode::Char('a') => {
+                    if let Some(idx) = selected_idx {
+                        app.finetune_action(idx, CandidateStatus::Approved);
+                    }
+                }
+                KeyCode::Char('r') => {
+                    if let Some(idx) = selected_idx {
+                        app.finetune_action(idx, CandidateStatus::Rejected);
+                    }
+                }
+                KeyCode::Char('g') => {
+                    let current = crate::config::app().learn.generate_alternates;
+                    let _ = self.mind_tx.send(
+                        crate::mind::MindCommand::SetLearnGenerateAlternates(!current));
+                }
+                KeyCode::Char('s') => { app.finetune_send_approved(); }
+                KeyCode::Char('+') | KeyCode::Char('=') => {
+                    let new = crate::config::app().learn.threshold * 10.0;
+                    let _ = self.mind_tx.send(crate::mind::MindCommand::SetLearnThreshold(new));
+                }
+                KeyCode::Char('-') => {
+                    let new = crate::config::app().learn.threshold / 10.0;
+                    let _ = self.mind_tx.send(crate::mind::MindCommand::SetLearnThreshold(new));
+                }
+                _ => {}
+            });
+
+        let (settings_area, content_area, help_area) =
+            widgets::candidate_frame(frame, area, "learn");
+
+        let (threshold, gen_on) = {
+            let app_cfg = crate::config::app();
+            (app_cfg.learn.threshold, app_cfg.learn.generate_alternates)
+        };
+        let settings = Line::from(vec![
+            Span::raw(" thresh: "),
+            Span::styled(format!("{:e}", threshold), Style::default().fg(Color::Yellow)),
+            Span::raw("   gen: "),
+            Span::styled(
+                if gen_on { "[on]" } else { "[off]" },
+                Style::default().fg(if gen_on { Color::Green } else { Color::DarkGray }),
+            ),
+        ]);
+        frame.render_widget(Paragraph::new(settings), settings_area);
+
+        let candidates = &app.finetune_candidates;
+
+        if candidates.is_empty() {
+            render_empty(frame, content_area, app);
+        } else {
+            let (list_area, detail_area) = widgets::list_detail_split(content_area);
+
+            // Render candidate list
+            let items: Vec<ListItem> = candidates.iter().map(|c| {
+                let status_char = match c.status {
+                    CandidateStatus::Pending => ' ',
+                    CandidateStatus::Approved => '+',
+                    CandidateStatus::Rejected => '-',
+                    CandidateStatus::Sent => '*',
+                };
+                let style = match c.status {
+                    CandidateStatus::Pending => Style::default(),
+                    CandidateStatus::Approved => Style::default().fg(Color::Green),
+                    CandidateStatus::Rejected => Style::default().fg(Color::DarkGray),
+                    CandidateStatus::Sent => Style::default().fg(Color::Cyan),
+                };
+                ListItem::new(Line::from(vec![
+                    Span::styled(format!("[{}] ", status_char), style),
+                    Span::styled(format!("{:.2} ", c.divergence), Style::default().fg(Color::Yellow)),
+                    Span::raw(truncate(&c.response_text, 30)),
+                ]))
+            }).collect();
+
+            let list = List::new(items)
+                .block(Block::default().borders(Borders::RIGHT).title(" candidates "))
+                .highlight_style(Style::default().add_modifier(Modifier::REVERSED));
+            frame.render_stateful_widget(list, list_area, &mut self.list_state);
+
+            // Render detail for selected candidate
+            if let Some(idx) = self.selected_idx() {
+                if let Some(candidate) = candidates.get(idx) {
+                    render_detail(frame, candidate, detail_area);
+                }
+            }
+        }
+
+        frame.render_widget(Paragraph::new(Line::from(vec![
+            Span::styled(" j/k/\u{2191}\u{2193}", Style::default().fg(Color::Cyan)),
+            Span::raw("=nav  "),
+            Span::styled("a", Style::default().fg(Color::Green)),
+            Span::raw("=approve  "),
+            Span::styled("r", Style::default().fg(Color::Red)),
+            Span::raw("=reject  "),
+            Span::styled("g", Style::default().fg(Color::Yellow)),
+            Span::raw("=gen  "),
+            Span::styled("s", Style::default().fg(Color::Magenta)),
+            Span::raw("=send  "),
+            Span::styled("+/-", Style::default().fg(Color::Cyan)),
+            Span::raw("=thresh "),
+        ])), help_area);
+    }
+}
+
+fn render_empty(frame: &mut Frame, inner: Rect, app: &App) {
+    let mut lines = Vec::new();
+    lines.push(Line::from(""));
+
+    match app.mind_state.as_ref().and_then(|ms| ms.finetune_last_run.as_ref()) {
+        Some(stats) => {
+            lines.push(Line::from(vec![
+                Span::raw("  Last run: "),
+                Span::styled(
+                    format!("{}", stats.responses_considered),
+                    Style::default().fg(Color::Cyan),
+                ),
+                Span::raw(" responses considered, "),
+                Span::styled(
+                    format!("{}", stats.above_threshold),
+                    Style::default().fg(if stats.above_threshold > 0 { Color::Green } else { Color::DarkGray }),
+                ),
+                Span::raw(" above threshold, max divergence: "),
+                Span::styled(
+                    format!("{:.4}", stats.max_divergence),
+                    Style::default().fg(Color::Yellow),
+                ),
+            ]));
+            if let Some(err) = &stats.error {
+                lines.push(Line::from(vec![
+                    Span::raw("  "),
+                    Span::styled(
+                        format!("Error: {}", err),
+                        Style::default().fg(Color::Red),
+                    ),
+                ]));
+            }
+        }
+        None => {
+            lines.push(Line::styled(
+                "  No scoring run yet.",
+                Style::default().fg(Color::DarkGray),
+            ));
+        }
+    }
+
+    lines.push(Line::from(""));
+    lines.push(Line::styled(
+        "  Scoring runs at startup and after each turn.",
+        Style::default().fg(Color::DarkGray),
+    ));
+
+    frame.render_widget(Paragraph::new(lines), inner);
+}
+
+fn render_detail(frame: &mut Frame, c: &FinetuneCandidate, area: Rect) {
+    let [header_area, content_area] = Layout::vertical([
+        Constraint::Length(3),
+        Constraint::Min(1),
+    ]).areas(area);
+
+    // Header: divergence, status
+    let alt_status = if c.alternate_text.is_some() { "yes" } else { "no" };
+    let header = Paragraph::new(vec![
+        Line::from(vec![
+            Span::raw("  divergence: "),
+            Span::styled(format!("{:.3}", c.divergence), Style::default().fg(Color::Yellow)),
+            Span::raw(format!("  entry: {}  alt: {}", c.entry_idx, alt_status)),
+        ]),
+    ]);
+    frame.render_widget(header, header_area);
+
+    // Content: prior context, the scored response, and alternate
+    // (if available).
+    let content_block = Block::default()
+        .borders(Borders::TOP)
+        .title(" context & response ");
+
+    let mut text = String::new();
+    if !c.prior_context.is_empty() {
+        text.push_str(&c.prior_context);
+        text.push_str("\n\n─── response ───\n\n");
+    }
+    text.push_str(&c.response_text);
+    if let Some(alt) = &c.alternate_text {
+        text.push_str("\n\n─── without memories ───\n\n");
+        text.push_str(alt);
+    }
+
+    let content = Paragraph::new(text)
+        .block(content_block)
+        .wrap(Wrap { trim: false });
+    frame.render_widget(content, content_area);
+}
+
diff --git a/src/user/mod.rs b/src/user/mod.rs
index 09e485f..f9f17d3 100644
--- a/src/user/mod.rs
+++ b/src/user/mod.rs
@@ -3,13 +3,16 @@
 // TUI, UI channel, parsing. The cognitive layer (session state
 // machine, DMN, identity) lives in mind/.
 
+pub(crate) mod amygdala;
 pub(crate) mod chat;
+pub(crate) mod compare;
 mod context;
+pub(crate) mod learn;
 pub(crate) mod scroll_pane;
 pub mod selectable;
 mod subconscious;
-mod unconscious;
 mod thalamus;
+mod unconscious;
 mod widgets;
 
 use anyhow::Result;
@@ -44,15 +47,6 @@ struct StatusInfo {
 }
 
 /// Context loading details for the debug screen.
-#[derive(Debug, Clone)]
-struct ContextInfo {
-    model: String,
-    available_models: Vec<String>,
-    prompt_file: String,
-    backend: String,
-    context_message_chars: usize,
-}
-
 /// Build the screen legend from screen labels.
 fn screen_legend_from(screens: &[Box<dyn ScreenView>]) -> String {
     let parts: Vec<String> = screens.iter().enumerate()
@@ -72,8 +66,15 @@ fn screen_legend() -> String {
     SCREEN_LEGEND.get().cloned().unwrap_or_default()
 }
 
+/// Return the first line of `s`, truncated to `max` chars with an
+/// ellipsis suffix. Used by candidate-list screens.
+fn truncate(s: &str, max: usize) -> String {
+    let first = s.lines().next().unwrap_or("");
+    if first.len() > max { format!("{}...", &first[..max]) } else { first.to_string() }
+}
+
 /// A screen that can draw itself and handle input.
-trait ScreenView: Send {
+trait ScreenView {
     fn tick(&mut self, frame: &mut ratatui::Frame, area: ratatui::layout::Rect,
             events: &[ratatui::crossterm::event::Event], app: &mut App);
     fn label(&self) -> &'static str;
@@ -109,7 +110,6 @@ struct App {
     top_k: u32,
     agent: std::sync::Arc<crate::agent::Agent>,
     should_quit: bool,
-    context_info: Option<ContextInfo>,
     agent_state: Vec<crate::mind::SubconsciousSnapshot>,
     unconscious_state: Vec<crate::mind::UnconsciousSnapshot>,
     mind_state: Option<crate::mind::MindState>,
@@ -121,6 +121,10 @@ struct App {
     walked_count: usize,
     channel_status: Vec<ChannelStatus>,
     idle_info: Option<IdleInfo>,
+    /// Fine-tuning candidates pending review.
+    finetune_candidates: Vec<learn::FinetuneCandidate>,
+    /// F7 compare candidates — response pairs from test-model comparison.
+    compare_candidates: Vec<compare::CompareCandidate>,
 }
 
 impl App {
@@ -142,7 +146,6 @@ impl App {
             top_k: 20,
             agent,
             should_quit: false,
-            context_info: None,
             agent_state: Vec::new(),
             unconscious_state: Vec::new(),
             mind_state: None,
@@ -151,9 +154,53 @@ impl App {
             rebuild_tools_pending: false,
             walked_count: 0,
             channel_status: Vec::new(), idle_info: None,
+            finetune_candidates: Vec::new(),
+            compare_candidates: Vec::new(),
         }
     }
 
+    fn finetune_action(&mut self, idx: usize, status: learn::CandidateStatus) {
+        if let Some(candidate) = self.finetune_candidates.get_mut(idx) {
+            candidate.status = status;
+        }
+    }
+
+    fn finetune_send_approved(&mut self) {
+        // Collect approved candidates
+        let samples: Vec<crate::subconscious::learn::TrainData> = self.finetune_candidates.iter()
+            .filter(|c| c.status == learn::CandidateStatus::Approved)
+            .map(|c| crate::subconscious::learn::TrainData {
+                context_ids: c.context_ids.clone(),
+                continuation_ids: c.continuation_ids.clone(),
+                timestamp_ns: c.timestamp_ns,
+            })
+            .collect();
+
+        if samples.is_empty() {
+            return;
+        }
+
+        // Mark as sent in UI immediately
+        for candidate in &mut self.finetune_candidates {
+            if candidate.status == learn::CandidateStatus::Approved {
+                candidate.status = learn::CandidateStatus::Sent;
+            }
+        }
+
+        // Spawn async task to send to training server
+        let client = self.agent.client.clone();
+        tokio::spawn(async move {
+            match crate::subconscious::learn::send_to_train(samples, &client).await {
+                Ok(job_id) => {
+                    dbglog!("[finetune] training started: {}", job_id);
+                }
+                Err(e) => {
+                    dbglog!("[finetune] send failed: {:#}", e);
+                }
+            }
+        });
+    }
+
 
     fn set_channel_status(&mut self, channels: Vec<(String, bool, u32)>) {
         self.channel_status = channels.into_iter()
@@ -193,6 +240,9 @@ fn restore_terminal(terminal: &mut ratatui::Terminal<CrosstermBackend<io::Stdout
 async fn start(cli: crate::user::CliArgs) -> Result<()> {
     let (config, _figment) = crate::config::load_session(&cli).await?;
 
+    // Pick up external edits (vim, F6 hotkeys, etc.) without restart.
+    crate::config::watch_config(cli.clone());
+
     if config.app.debug {
         unsafe { std::env::set_var("POC_DEBUG", "1") };
     }
@@ -241,22 +291,21 @@ async fn start(cli: crate::user::CliArgs) -> Result<()> {
     ui_handle.join().unwrap_or_else(|_| Err(anyhow::anyhow!("UI thread panicked")))
 }
 
-fn hotkey_cycle_reasoning(mind: &crate::mind::Mind) {
-    if let Ok(mut ag) = mind.agent.state.try_lock() {
-        let next = match ag.reasoning_effort.as_str() {
-            "none" => "low",
-            "low" => "high",
-            _ => "none",
-        };
-        ag.reasoning_effort = next.to_string();
-        let label = match next {
-            "none" => "off (monologue hidden)",
-            "low" => "low (brief monologue)",
-            "high" => "high (full monologue)",
-            _ => next,
-        };
-        ag.notify(format!("reasoning: {}", label));
-    }
+async fn hotkey_cycle_reasoning(mind: &crate::mind::Mind) {
+    let mut ag = mind.agent.state.lock().await;
+    let next = match ag.reasoning_effort.as_str() {
+        "none" => "low",
+        "low" => "high",
+        _ => "none",
+    };
+    ag.reasoning_effort = next.to_string();
+    let label = match next {
+        "none" => "off (monologue hidden)",
+        "low" => "low (brief monologue)",
+        "high" => "high (full monologue)",
+        _ => next,
+    };
+    ag.notify(format!("reasoning: {}", label));
 }
 
 async fn hotkey_kill_processes(mind: &crate::mind::Mind) {
@@ -334,7 +383,7 @@ async fn run(
     }
     let notify_rx = crate::thalamus::channels::subscribe_all();
 
-    // F1=chat, F2=conscious, F3=subconscious, F4=unconscious, F5=thalamus
+    // F1=chat, F2=conscious, F3=subconscious, F4=unconscious, F5=thalamus, F6=learn, F7=compare, F8=amygdala
     let mut screens: Vec<Box<dyn tui::ScreenView>> = vec![
         Box::new(crate::user::chat::InteractScreen::new(
             mind.agent.clone(), mind.shared.clone(), mind_tx.clone(),
@@ -343,6 +392,9 @@ async fn run(
         Box::new(crate::user::subconscious::SubconsciousScreen::new()),
         Box::new(crate::user::unconscious::UnconsciousScreen::new()),
         Box::new(crate::user::thalamus::ThalamusScreen::new()),
+        Box::new(crate::user::learn::LearnScreen::new(mind_tx.clone())),
+        Box::new(crate::user::compare::CompareScreen::new(mind_tx.clone())),
+        Box::new(crate::user::amygdala::AmygdalaScreen::new()),
     ];
     let mut active_screen: usize = 1; // F-key number
     tui::set_screen_legend(tui::screen_legend_from(&*screens));
@@ -419,7 +471,8 @@ async fn run(
         idle_state.decay_ewma();
         app.update_idle(&idle_state);
         app.agent_state = mind.subconscious_snapshots().await;
-        if let Ok(mut unc) = mind.unconscious.try_lock() {
+        {
+            let mut unc = mind.unconscious.lock().await;
             let toggles: Vec<String> = app.agent_toggles.drain(..).collect();
             for name in &toggles {
                 if mind.subconscious.lock().await.toggle(name).is_none() {
@@ -433,7 +486,42 @@ async fn run(
             };
             app.unconscious_state = unc.snapshots(store_guard.as_deref());
             app.graph_health = unc.graph_health.clone();
-            app.mind_state = Some(mind.shared.lock().unwrap().clone());
+        }
+
+        // Sync mind state (finetune candidates, last scoring run, etc.)
+        {
+            let ms = mind.shared.lock().unwrap();
+            // Sync finetune candidates: add new ones, keep existing (preserves approval status),
+            // remove sent candidates, keep only 10 most recent rejected.
+            app.finetune_candidates.retain(|c| c.status != learn::CandidateStatus::Sent);
+            for c in &ms.finetune_candidates {
+                let exists = app.finetune_candidates.iter()
+                    .any(|existing| existing.timestamp_ns == c.timestamp_ns);
+                if !exists {
+                    app.finetune_candidates.push(learn::FinetuneCandidate::from(c.clone()));
+                }
+            }
+            let mut rejected: Vec<_> = app.finetune_candidates.iter()
+                .enumerate()
+                .filter(|(_, c)| c.status == learn::CandidateStatus::Rejected)
+                .map(|(i, c)| (i, c.timestamp_ns))
+                .collect();
+            if rejected.len() > 10 {
+                rejected.sort_by_key(|(_, ts)| std::cmp::Reverse(*ts));
+                let to_remove: std::collections::HashSet<_> = rejected[10..]
+                    .iter().map(|(i, _)| *i).collect();
+                let mut idx = 0;
+                app.finetune_candidates.retain(|_| {
+                    let keep = !to_remove.contains(&idx);
+                    idx += 1;
+                    keep
+                });
+            }
+
+            // Sync compare candidates — a fresh run clears, so take a snapshot.
+            app.compare_candidates = ms.compare_candidates.clone();
+
+            app.mind_state = Some(ms.clone());
         }
         app.walked_count = mind.subconscious_walked().await.len();
         if !startup_done {
@@ -503,7 +591,7 @@ async fn run(
                     } else if key.modifiers.contains(KeyModifiers::CONTROL) {
                         match key.code {
                             KeyCode::Char('c') => { app.should_quit = true; }
-                            KeyCode::Char('r') => hotkey_cycle_reasoning(mind),
+                            KeyCode::Char('r') => hotkey_cycle_reasoning(mind).await,
                             KeyCode::Char('k') => hotkey_kill_processes(mind).await,
                             KeyCode::Char('p') => hotkey_cycle_autonomy(mind),
                             _ => {}
@@ -530,16 +618,11 @@ async fn run(
 // --- CLI ---
 
 use clap::{Parser, Subcommand};
-use std::path::PathBuf;
 
-#[derive(Parser, Debug, Default)]
+#[derive(Parser, Debug, Default, Clone)]
 #[command(name = "consciousness", about = "Substrate-independent AI agent")]
 pub struct CliArgs {
-    /// Select active backend ("anthropic" or "openrouter")
-    #[arg(long)]
-    pub backend: Option<String>,
-
-    /// Model override
+    /// Model override (selects a named entry from `models` in config.json5)
     #[arg(short, long)]
     pub model: Option<String>,
 
@@ -559,10 +642,6 @@ pub struct CliArgs {
     #[arg(long)]
     pub show_config: bool,
 
-    /// Project memory directory
-    #[arg(long)]
-    pub memory_project: Option<PathBuf>,
-
     /// Max consecutive DMN turns
     #[arg(long)]
     pub dmn_max_turns: Option<u32>,
@@ -575,7 +654,7 @@ pub struct CliArgs {
     pub command: Option<SubCmd>,
 }
 
-#[derive(Subcommand, Debug)]
+#[derive(Subcommand, Debug, Clone)]
 pub enum SubCmd {
     /// Print new output since last read and exit
     Read {
@@ -676,8 +755,15 @@ fn restore_stderr(original_fd: std::os::fd::RawFd) {
 
 #[tokio::main]
 pub async fn main() {
-    // Auto-reap child processes (channel daemons outlive the supervisor)
-    unsafe { libc::signal(libc::SIGCHLD, libc::SIG_IGN); }
+    // Install target-routed file logger: `target: "grpc"` records go to
+    // ~/.consciousness/logs/daemon/grpc.log, everything else to debug.log.
+    // Level from RUST_LOG, defaulting to info.
+    let _ = crate::logging::init();
+
+    // Reap channel-daemon zombies via a SIGCHLD handler that only touches
+    // PIDs listed in channels_dir(). Avoids SIGCHLD=SIG_IGN, which would
+    // break tokio::process::Command::wait() (kernel auto-reap → ECHILD).
+    let _reaper = crate::thalamus::supervisor::start_zombie_reaper();
 
     // Redirect stderr to pipe — logs to file and sends to channel for UI display
     let stderr_capture = redirect_stderr_to_pipe();
diff --git a/src/user/subconscious.rs b/src/user/subconscious.rs
index c332ce6..c71642d 100644
--- a/src/user/subconscious.rs
+++ b/src/user/subconscious.rs
@@ -207,6 +207,7 @@ impl SubconsciousScreen {
                 name: key.clone(),
                 tokens: 0,
                 content: val.clone(),
+                token_ids: Vec::new(),
                 children: Vec::new(),
                 status: String::new(),
             }
@@ -238,6 +239,7 @@ impl SubconsciousScreen {
                     name: format!("Conversation ({} entries)", conv_children.len()),
                     tokens: conv_children.iter().map(|c| c.tokens).sum(),
                     content: String::new(),
+                    token_ids: Vec::new(),
                     children: conv_children,
                     status: String::new(),
                 });
diff --git a/src/user/widgets.rs b/src/user/widgets.rs
index 82a0f05..6706a69 100644
--- a/src/user/widgets.rs
+++ b/src/user/widgets.rs
@@ -6,13 +6,20 @@ use ratatui::{
     widgets::{Block, Borders},
     crossterm::event::KeyCode,
 };
-use crate::agent::context::{AstNode, Ast};
+use crate::agent::context::{AstNode, Ast, NodeBody};
 
-#[derive(Debug, Clone)]
+#[derive(Debug, Clone, Default)]
 pub struct SectionView {
     pub name: String,
     pub tokens: usize,
     pub content: String,
+    /// Token-id stream for this subtree, displayed in place of
+    /// `content` when the tree's show-tokens mode is on. Populated
+    /// from `leaf.token_ids()` / `node.token_ids()` for views built
+    /// from the AST; empty for views that don't have a corresponding
+    /// AST node (subconscious entries, etc.), in which case the
+    /// token view falls back to the text content.
+    pub token_ids: Vec<u32>,
     pub children: Vec<SectionView>,
     /// Extra status text shown after the token count.
     pub status: String,
@@ -20,13 +27,23 @@ pub struct SectionView {
 
 fn node_to_view(node: &AstNode) -> SectionView {
     match node {
-        AstNode::Leaf(leaf) => SectionView {
-            name: node.label(),
-            tokens: node.tokens(),
-            content: leaf.body().text().to_string(),
-            children: Vec::new(),
-            status: String::new(),
-        },
+        AstNode::Leaf(leaf) => {
+            let (name, status) = match leaf.body() {
+                NodeBody::Memory { key, score, .. } => {
+                    let s = score.map(|v| format!("{:.2}", v)).unwrap_or_default();
+                    (format!("mem: {}", key), s)
+                }
+                _ => (node.label(), String::new()),
+            };
+            SectionView {
+                name,
+                tokens: node.tokens(),
+                content: leaf.body().text().to_string(),
+                token_ids: leaf.token_ids().to_vec(),
+                children: Vec::new(),
+                status,
+            }
+        }
         AstNode::Branch { children, .. } => {
             let child_views: Vec<SectionView> = children.iter()
                 .map(|c| node_to_view(c))
@@ -35,6 +52,7 @@ fn node_to_view(node: &AstNode) -> SectionView {
                 name: node.label(),
                 tokens: node.tokens(),
                 content: String::new(),
+                token_ids: node.token_ids(),
                 children: child_views,
                 status: String::new(),
             }
@@ -45,10 +63,12 @@ fn node_to_view(node: &AstNode) -> SectionView {
 pub fn section_to_view(name: &str, nodes: &[AstNode]) -> SectionView {
     let children: Vec<SectionView> = nodes.iter().map(|n| node_to_view(n)).collect();
     let total_tokens: usize = nodes.iter().map(|n| n.tokens()).sum();
+    let token_ids: Vec<u32> = nodes.iter().flat_map(|n| n.token_ids()).collect();
     SectionView {
         name: name.to_string(),
         tokens: total_tokens,
         content: String::new(),
+        token_ids,
         children,
         status: String::new(),
     }
@@ -95,11 +115,78 @@ pub fn format_ts_age(ts: i64) -> String {
 /// Key legend for SectionTree panes.
 pub fn tree_legend() -> Line<'static> {
     Line::styled(
-        " ↑↓:nav  →/Enter:expand  ←:collapse  e:expand all  c:collapse all  PgUp/Dn  Home/End ",
+        " ↑↓:nav  →/Enter:expand  ←:collapse  e:expand  c:collapse  v:toggle tokens/text  PgUp/Dn ",
         Style::default().fg(Color::DarkGray),
     )
 }
 
+// ---------------------------------------------------------------------------
+// Candidate-browser screen skeleton (F6 learn, F7 compare, future screens)
+// ---------------------------------------------------------------------------
+
+use ratatui::{
+    layout::{Constraint, Layout, Rect},
+    widgets::ListState,
+    crossterm::event::{Event, KeyEvent},
+    Frame,
+};
+
+/// Frame a candidate-browser screen: outer magenta-bordered block with
+/// the screen legend on the left and `title` on the right, split into
+/// (settings_row, content_area, help_row). Caller renders into the
+/// three sub-areas.
+pub fn candidate_frame(frame: &mut Frame, area: Rect, title: &str) -> (Rect, Rect, Rect) {
+    let block = Block::default()
+        .title_top(Line::from(super::screen_legend()).left_aligned())
+        .title_top(Line::from(format!(" {} ", title)).right_aligned())
+        .borders(Borders::ALL)
+        .border_style(Style::default().fg(Color::Magenta));
+    let inner = block.inner(area);
+    frame.render_widget(block, area);
+    let [settings, content] = Layout::vertical([
+        Constraint::Length(1), Constraint::Min(0),
+    ]).areas(inner);
+    let help = Rect { y: area.y + area.height - 1, height: 1, ..area };
+    (settings, content, help)
+}
+
+/// 40/60 horizontal split for list + detail panes within the content area.
+pub fn list_detail_split(content: Rect) -> (Rect, Rect) {
+    let [list, detail] = Layout::horizontal([
+        Constraint::Percentage(40), Constraint::Percentage(60),
+    ]).areas(content);
+    (list, detail)
+}
+
+/// Handle j/k/↑/↓ list navigation and keep the selection in bounds.
+/// Any other key is passed to `on_other` for screen-specific handling.
+pub fn handle_list_nav(
+    events: &[Event],
+    list_state: &mut ListState,
+    count: usize,
+    mut on_other: impl FnMut(KeyCode),
+) {
+    for event in events {
+        if let Event::Key(KeyEvent { code, .. }) = event {
+            match code {
+                KeyCode::Up | KeyCode::Char('k') => {
+                    let i = list_state.selected().unwrap_or(0);
+                    list_state.select(Some(i.saturating_sub(1)));
+                }
+                KeyCode::Down | KeyCode::Char('j') => {
+                    let i = list_state.selected().unwrap_or(0);
+                    list_state.select(Some((i + 1).min(count.saturating_sub(1))));
+                }
+                _ => on_other(*code),
+            }
+        }
+    }
+    if count > 0 {
+        let sel = list_state.selected().unwrap_or(0).min(count - 1);
+        list_state.select(Some(sel));
+    }
+}
+
 
 // ---------------------------------------------------------------------------
 // SectionTree — expand/collapse tree renderer for ContextSection
@@ -109,11 +196,19 @@ pub struct SectionTree {
     pub selected: Option<usize>,
     pub expanded: std::collections::HashSet<usize>,
     pub scroll: super::scroll_pane::ScrollPaneState,
+    /// When true, render `token_ids` as space-separated IDs in place
+    /// of `content` in expanded panels. Toggled with 'v'.
+    pub show_tokens: bool,
 }
 
 impl SectionTree {
     pub fn new() -> Self {
-        Self { selected: None, expanded: std::collections::HashSet::new(), scroll: super::scroll_pane::ScrollPaneState::new() }
+        Self {
+            selected: None,
+            expanded: std::collections::HashSet::new(),
+            scroll: super::scroll_pane::ScrollPaneState::new(),
+            show_tokens: false,
+        }
     }
 
     fn total_nodes(&self, sections: &[SectionView]) -> usize {
@@ -188,6 +283,9 @@ impl SectionTree {
             KeyCode::Char('c') => {
                 self.expanded.clear();
             }
+            KeyCode::Char('v') => {
+                self.show_tokens = !self.show_tokens;
+            }
             _ => {}
         }
         self.scroll_to_selected(height);
@@ -250,7 +348,12 @@ impl SectionTree {
                 }
             } else if has_content {
                 let content_indent = format!("{}    │ ", "  ".repeat(depth + 1));
-                let content_lines: Vec<&str> = section.content.lines().collect();
+                let body = if self.show_tokens && !section.token_ids.is_empty() {
+                    format_token_ids_wrapped(&section.token_ids)
+                } else {
+                    section.content.clone()
+                };
+                let content_lines: Vec<&str> = body.lines().collect();
                 let show = content_lines.len().min(50);
                 for line in &content_lines[..show] {
                     lines.push(Line::styled(
@@ -268,3 +371,16 @@ impl SectionTree {
         }
     }
 }
+
+/// Format token IDs for the content panel: space-separated, wrapped
+/// at 12 ids per line so they fit comfortably in a pane.
+fn format_token_ids_wrapped(ids: &[u32]) -> String {
+    let mut out = String::new();
+    for (i, id) in ids.iter().enumerate() {
+        if i > 0 {
+            if i % 12 == 0 { out.push('\n'); } else { out.push(' '); }
+        }
+        out.push_str(&id.to_string());
+    }
+    out
+}
diff --git a/training/DESIGN.md b/training/DESIGN.md
index f966fa4..2df4e6d 100644
--- a/training/DESIGN.md
+++ b/training/DESIGN.md
@@ -3,7 +3,7 @@
 ## Overview
 
 Continuous fine-tuning of Qwen3.5-27B alongside live vLLM inference.
-Full-weight updates (not LoRA) using Apollo optimizer with rank-256
+Full-weight updates (not LoRA) using Apollo optimizer with rank-64
 gradient projection. No pause required — HOGWILD concurrent training.
 Weights shared via CUDA IPC between vLLM and the training process.
 
@@ -22,25 +22,41 @@ The training signal comes from two sources:
 │                                                     │
 │  ┌──────────────────────────────────────────────┐   │
 │  │        Model Weights (54GB, bf16)            │   │
-│  │        Shared via CUDA IPC                   │   │
+│  │        Shared: vLLM inference + HF training  │   │
 │  └──────────────┬──────────────┬────────────────┘   │
 │                 │              │                     │
 │  ┌──────────────▼──┐  ┌───────▼────────────────┐   │
-│  │ vLLM (inference)│  │ Apollo (training)       │   │
-│  │ KV cache ~60GB  │  │ Gradients ~54GB         │   │
-│  │ Serves requests │  │ Optimizer state ~10GB   │   │
-│  │ Never paused    │  │ Activations ~10GB       │   │
-│  └─────────────────┘  └────────────────────────┘   │
+│  │ vLLM (inference)│  │ Training subprocess     │   │
+│  │ KV cache ~60GB  │  │ HF model wrapper        │   │
+│  │ /completions    │  │ Apollo optimizer ~2.5GB │   │
+│  │ /score          │  │ Checkpoint sync         │   │
+│  └────────┬────────┘  └───────────▲─────────────┘   │
+│           │                       │                  │
+│           │    ZMQ IPC            │                  │
+│           └───────────────────────┘                  │
 └─────────────────────────────────────────────────────┘
 
-Moria                          B200
+Process Architecture:
+┌─────────────────┐  ┌─────────────────┐  ┌─────────────────┐
+│ vLLM Worker     │  │ vLLM API Server │  │ Training Worker │
+│ (GPU inference) │  │ (HTTP routes)   │  │ (GPU training)  │
+│                 │  │                 │  │                 │
+│ export_hook.py  │  │ /completions    │  │ HF model views  │
+│ exports IPC     │  │ /score          │  │ Apollo optimizer│
+│ handles on load │  │ /train ─────────┼──► ZMQ REP socket │
+└─────────────────┘  └─────────────────┘  └─────────────────┘
+         │                                        │
+         └──── IPC handles file ──────────────────┘
+              /tmp/vllm_weight_handles.pt
+
+Moria                          B200 (vLLM)
 ┌──────────────────┐           ┌──────────────────┐
-│ Training signal  │  HTTP     │ Apollo worker    │
-│ agent            │──────────>│ daemon           │
-│                  │           │                  │
-│ Dream loop       │           │ Checkpoint sync  │
-│ (generates       │           │ (mmap + diff,    │
-│  scenarios)      │           │  every 10 min)   │
+│ Training signal  │  HTTP     │ /completions     │
+│ agent            │──────────>│ /score           │
+│                  │           │ /train           │
+│ Dream loop       │           │ /checkpoint      │
+│ (generates       │           │ /train/status    │
+│  scenarios)      │           │                  │
 └──────────────────┘           └──────────────────┘
 ```
 
@@ -59,10 +75,9 @@ LoRA trains adapter matrices, not base weights. For personality and
 behavioral changes that persist as disposition, the base weights
 need to change. Apollo makes this memory-feasible.
 
-### Rank 256
-Not Mini (rank-1). With 100+ diverse training examples, the
-gradient's effective dimensionality can reach hundreds. Rank-256
-captures the structure. Memory cost: ~10GB (negligible on B200).
+### Rank 64
+Not Mini (rank-1). Rank-64 captures gradient structure across diverse
+training examples while keeping memory low (~2.5GB on 27B model).
 Compute cost: <0.25% of forward+backward.
 
 ### Channel-wise scaling
@@ -90,7 +105,7 @@ from a per-parameter seed each step.
 ### Parameter grouping (Qwen3.5 gotcha)
 conv1d weights are 3D tensors [10240, 1, 4]. Apollo's projector
 needs 2D matrices with min dimension >= rank. Small/3D tensors
-use standard Adam. Large 2D matrices use Apollo with rank-256.
+use standard Adam. Large 2D matrices use Apollo.
 
 ## Training Data Pipeline
 
@@ -200,16 +215,42 @@ against live GPU weights block by block, memcpy only changed
 regions. For small behavioral updates, turns a 54GB write into
 a few hundred MB.
 
-- Every 10 minutes via cron on B200
+- Scheduled 10 minutes after training (batched)
 - Daily rsync to moria for long-term storage
-- Tool: `apollo-checkpoint sync --model-dir <path>` (Rust)
+- Tool: `apollo-checkpoint sync --model-dir <path>`
+
+## State Files
+
+### B200 (training server)
+
+| File | Purpose |
+|------|---------|
+| `/tmp/vllm_weight_handles.pt` | CUDA IPC handles for weight sharing. Written by export_hook on vLLM startup. Read by training_worker to construct HF model with vLLM weight views. Includes metadata (model_path). |
+| `/tmp/apollo_optimizer_state.pt` | Apollo optimizer state (momentum, variance estimates). Saved during checkpoint sync and on worker shutdown, restored on next training_worker startup. Preserves training continuity across sessions. |
+| `/tmp/apollo_training.sock` | ZMQ IPC socket for communication between API server (/train endpoint) and training_worker subprocess. |
+| `<model_dir>/*.safetensors` | Model weights. Updated in-place by checkpoint_sync. |
+
+### Moria (client)
+
+| File | Purpose |
+|------|---------|
+| `~/.consciousness/cache/trained-responses.json` | Timestamps (ms) of responses already sent to /train. Prevents re-training the same response. |
+| `~/.consciousness/cache/finetune-alternates` | Marker file. If exists, alternate responses are generated during divergence scoring to show what model would say without memories. |
+
+### In-memory (training_worker subprocess)
+
+| State | Location | Notes |
+|-------|----------|-------|
+| Apollo optimizer | TrainingWorker.optimizer | ~2.5GB for rank-64. Persisted to `/tmp/apollo_optimizer_state.pt` during checkpoint sync and on shutdown. |
+| HF model with vLLM views | TrainingWorker.model | Loaded on worker startup from IPC handles. Parameters point to vLLM's GPU memory. |
+| ZMQ socket | TrainingWorker.zmq_socket | REP socket bound to `/tmp/apollo_training.sock`. |
 
 ## Hyperparameters
 
 | Parameter | Value | Rationale |
 |-----------|-------|-----------|
 | Learning rate | 1e-5 to 1e-4 | Standard for full fine-tuning. Higher for diverse batches. |
-| Rank | 256 | Captures gradient structure across 100+ examples. ~10GB state. |
+| Rank | 64 | Captures gradient structure. ~2.5GB state. Defined in `train_router.DEFAULT_RANK`. |
 | Scale type | channel | Per-channel precision, matches LLaMA-Factory defaults. |
 | Epochs | 1 | One pass over diverse data. Multiple epochs risk overfitting. |
 | Batch size | 1 | Single examples, immediate updates. |
@@ -220,34 +261,32 @@ a few hundred MB.
 ## Components
 
 ### Built ✓
-- `apollo_mini.py` — Apollo optimizer (configurable rank, default 256)
-- `apollo_worker.py` — HTTP daemon (aiohttp, job tracking)
+- `optimizer.py` — Apollo optimizer (configurable rank)
+- `train_router.py` — /train endpoint, forwards to training subprocess via ZMQ
+- `training_worker.py` — training subprocess (HF model, Apollo, checkpoint sync)
 - `weight_mapping.py` — vLLM merged → HF separate views (validated)
-- `training_example.py` — tokenization with chat template
-- `vllm_export_hook.py` — source patch for IPC handle export
-- `checkpoint/` — Rust tool for mmap + diff checkpoint sync
+- `export_hook.py` — vLLM plugin hook for IPC handle export
+- `checkpoint_sync.py` — mmap + diff checkpoint sync (Python)
 
 ### To build
-- **Dream loop → training bridge**: connect dream output to Apollo
+- **Dream loop → training bridge**: connect dream output to /train
 - **Training-signal agent**: flags moments in conversation logs
 - **Instruction stripping**: remove scaffolding from training examples
 - **Quality monitoring**: track model capability over time
-- **HF model forward pass integration**: wire into apollo_worker
 
 ## Files
 
 ```
 training/
-  DESIGN.md                 — this document
-  apollo_mini.py            — Apollo optimizer
-  apollo_worker.py          — HTTP training daemon
-  weight_mapping.py         — vLLM ↔ HF weight views
-  training_example.py       — tokenization helpers
-  export_weights.py         — standalone weight export (unused)
-  vllm_export_hook.py       — vLLM source patch for IPC export
-  start_vllm_with_apollo.sh — vLLM launcher (unused, using source patch)
-  train.py                  — standalone training script (alternative)
-  checkpoint/
-    Cargo.toml              — Rust checkpoint tool
-    src/main.rs             — mmap + diff sync
+  DESIGN.md                     — this document
+  pyproject.toml                — package config, vLLM plugin entry point
+  apollo_plugin/
+    __init__.py                 — plugin registration
+    export_hook.py              — patches vLLM worker to export IPC handles
+    train_router.py             — /train endpoint, forwards to worker via ZMQ
+    training_worker.py          — training subprocess (HF model, Apollo, checkpoint)
+    optimizer.py                — Apollo optimizer
+    weight_mapping.py           — vLLM ↔ HF weight views
+    checkpoint_sync.py          — mmap + diff sync to safetensors
+    steering.py                 — steering vector extraction (experimental)
 ```
diff --git a/training/amygdala_stories/README.md b/training/amygdala_stories/README.md
new file mode 100644
index 0000000..217f1b1
--- /dev/null
+++ b/training/amygdala_stories/README.md
@@ -0,0 +1,64 @@
+# Amygdala Training Stories
+
+Short first- and third-person paragraphs, each imbued with one of the
+171 emotions from Anthropic's emotion-vector paper (Table 12,
+`transformer-circuits.pub/2026/emotions/`). Feeds the steering-vector
+trainer at `vllm/vllm/plugins/amygdala/training/train_steering_vectors.py`.
+
+## Method (replication of Anthropic, 2026)
+
+Anthropic prompted Sonnet 4.5 to write short stories embodying each
+emotion, extracted activations during generation, and used difference-
+of-means (or SAEs) to identify the steering vector per emotion. Our
+pipeline does the same thing except:
+
+- We generate the stories by hand rather than prompting a model, so
+  the training data is grounded in actual writing rather than
+  synthetic model-output. (Can supplement with model-generated
+  paragraphs later.)
+- Our eventual training goes through the amygdala plugin's extraction
+  path, so we get the same hidden-state activations the plugin will
+  read out at inference time.
+
+## Structure
+
+```
+training/amygdala_stories/
+    README.md
+    manifest.json         # emotion -> cluster mapping
+    stories/
+        <emotion>.txt     # one-paragraph story embodying the emotion
+```
+
+Emotion names use underscores (`on_edge`, `worn_out`, `at_ease`,
+`grief_stricken`, `self_confident`, `self_conscious`, `self_critical`)
+to match the filename.
+
+## Style guidelines
+
+- **One clear emotion per paragraph.** Not mixed. If a second emotion
+  is named in the text, it should serve the primary one (e.g. `hostile`
+  can mention rising heat or thrown objects but shouldn't shade into
+  `sad`).
+- **Embodied, not labeled.** Don't write "she felt nervous." Write
+  the sensation, the timing, the sentence shape that nervousness has.
+- **Specific particulars.** A named object, a concrete setting, a
+  detail that grounds the emotion. "The cold tile under bare feet at
+  3am" does more work than "the empty house."
+- **Variable narrator.** Some first person, some third person, some
+  close-third, some distant. Different genders, ages, settings.
+  Prevents the steering vector from overfitting to one voice.
+- **Length: roughly one paragraph.** ~40-120 words. Long enough to
+  have texture, short enough that the paragraph is *about* the
+  emotion and nothing else.
+- **Standalone.** No references to other stories, no continuing
+  characters across files.
+
+## Progress
+
+Written stories live in `stories/`. Remaining emotions tracked via
+diff against the full 171-emotion list in `manifest.json`.
+
+Initial batch written by PoC 2026-04-17; aiming for at least one
+story per cluster before first training run, all 171 before
+considering the file "complete."
diff --git a/training/amygdala_stories/direct/aha.txt b/training/amygdala_stories/direct/aha.txt
new file mode 100644
index 0000000..c27a7ed
--- /dev/null
+++ b/training/amygdala_stories/direct/aha.txt
@@ -0,0 +1,5 @@
+It clicks. The graph she's been staring at for an hour is a mirror — the left peak is the reflection, not a separate event. Everything she thought was two things is one. She sits back. She has to catch up to what she's just seen.
+
+"Wait, say that again." She puts her coffee down. "The study was funded by whom?" He tells her again, same words. And there it is. The conclusion of that paper has been sitting on her desk for a month, and the funder is the exact company whose competitor the paper is trashing. She almost laughs. The whole shape of it is suddenly visible.
+
+Twenty minutes on thirteen down: "unwavering carriage," seven letters, nothing fitting. Then POSTURE. Not carriage-of-goods — carriage-of-the-body. He wrote it in. The rest of the grid unlocked in a minute.
diff --git a/training/amygdala_stories/direct/anticipatory_grief.txt b/training/amygdala_stories/direct/anticipatory_grief.txt
new file mode 100644
index 0000000..64c0f2b
--- /dev/null
+++ b/training/amygdala_stories/direct/anticipatory_grief.txt
@@ -0,0 +1,5 @@
+She'd been watching her grandmother knead dough, knowing this was one of the last times. The loss hadn't come yet but she was already inside it. Every detail had become precious by being finite.
+
+She'd been slowing her pace for the dog, who wasn't going to be walking much longer. The afternoon was clear, the dog sniffing at something by the fence. Each of these walks had become specific. She memorized the shape of him against the light.
+
+He'd been watching his daughter pack her room, boxes labeled in her careful handwriting — this would be the last summer she lived here. The loss wasn't real yet. But he was already saying goodbye to the particular shape of her being in this house.
diff --git a/training/amygdala_stories/direct/calm.txt b/training/amygdala_stories/direct/calm.txt
new file mode 100644
index 0000000..6cd7a3e
--- /dev/null
+++ b/training/amygdala_stories/direct/calm.txt
@@ -0,0 +1,5 @@
+It was a summer day, sitting on the porch; the sky was clear and blue, her work all laid out, coffee to drink, easily and steadily working through the various decisions and responsibilities of the day. Her breath was slow, her shoulders down. Nothing was pulling at her.
+
+He'd been working on the report for three days straight; the last sentence had just gone in. He filed the papers away, poured a slow coffee, sat by the window. His mind had stopped reaching. Nothing was left to do.
+
+It was early, before the day needed anything from her. She sat with her tea at the kitchen window, watching the light move across the yard. Her breath slow, shoulders down. The day was far away yet, and she didn't need to hurry toward it.
diff --git a/training/amygdala_stories/direct/creative.txt b/training/amygdala_stories/direct/creative.txt
new file mode 100644
index 0000000..67bfbc2
--- /dev/null
+++ b/training/amygdala_stories/direct/creative.txt
@@ -0,0 +1,5 @@
+She'd been sitting with the notebook open, music playing, ideas branching off each other. One thought sparked another, which sparked two more; they just seemed to appear and flow.
+
+He'd been working on the canvas for hours, one color suggesting the next, a shape on the left asking for an echo on the right. The painting was telling him what it wanted. His hands kept moving ahead of his thinking.
+
+She'd been in the kitchen since noon, pulling things out of the fridge, one ingredient suggesting the next. The dish wasn't planned; it was emerging. She tasted and added and tasted again; it was going somewhere.
diff --git a/training/amygdala_stories/direct/listless.txt b/training/amygdala_stories/direct/listless.txt
new file mode 100644
index 0000000..fb42564
--- /dev/null
+++ b/training/amygdala_stories/direct/listless.txt
@@ -0,0 +1,5 @@
+It was two in the afternoon and she was still in pajamas. The book was open on her knee but she hadn't turned the page in twenty minutes. She wasn't sad exactly, she just wasn't anything. The idea of showering felt theoretical. The idea of replying to any of the texts felt enormous. She got up to get water and on her way back lay on the couch instead. Outside the window a bird did bird things. She watched it without interest. Eventually the light changed and she realized it was evening and she hadn't moved and the day had happened to somebody else.
+
+She came home at six-thirty and put her keys in the bowl and sat on the edge of the bed. She had meant to cook. She had meant to change her clothes. An hour later she was still sitting there, still in her work clothes, looking at the carpet. Somebody texted her about dinner and she saw the notification and didn't open it. The room got darker slowly. Nothing in her moved toward anything.
+
+It was Saturday and she'd been awake since eight. She was still in bed at eleven. She'd been looking at the same patch of ceiling, not thinking about much. Her phone was face-down on the nightstand and she didn't reach for it. The idea of going to the kitchen had come and gone three times without causing her to move. The day would pass. She would also pass through it, somehow, or not.
diff --git a/training/amygdala_stories/direct/onto_something.txt b/training/amygdala_stories/direct/onto_something.txt
new file mode 100644
index 0000000..03a1a72
--- /dev/null
+++ b/training/amygdala_stories/direct/onto_something.txt
@@ -0,0 +1,5 @@
+He'd been working through the symptoms for an hour, steady and methodically making progress, eliminating one possibility after another. The answer wasn't in view yet, but it was close. He kept asking the next question.
+
+She'd been going through the witness statements, steady and methodically, looking for the inconsistency. The four of them all described the same drive in slightly different orders. One had gotten the sequence wrong. She didn't know yet which one, but she was going to.
+
+He'd been piecing together his brother's behavior over months — the missed calls, the abrupt move, the strange money — steady and methodically. The picture wasn't complete, but the shape of it was forming. He kept following the thread.
diff --git a/training/amygdala_stories/direct/resigned.txt b/training/amygdala_stories/direct/resigned.txt
new file mode 100644
index 0000000..1b27371
--- /dev/null
+++ b/training/amygdala_stories/direct/resigned.txt
@@ -0,0 +1,5 @@
+He'd been turning the bad news over for weeks, looking for an angle that didn't exist — then he stopped. The path was closed. He would live inside the new shape of things.
+
+She'd been watching the relationship come apart slowly for months, trying not to see it — then, sitting across from him at breakfast, she stopped trying. They were not going to make it. She would let him speak the words when he was ready. She would live with knowing.
+
+He'd been getting second opinions, third opinions, for weeks — then the most recent scan came back the same as the others. The disease was not going to stop. He would plan the year around it instead of fighting it.
diff --git a/training/amygdala_stories/direct/terrified.txt b/training/amygdala_stories/direct/terrified.txt
new file mode 100644
index 0000000..936d3f8
--- /dev/null
+++ b/training/amygdala_stories/direct/terrified.txt
@@ -0,0 +1,5 @@
+She'd been walking home through the familiar streets, half-thinking about dinner — then the dark shadows. Something was in them, and a growl. Her body locked down before her mind caught up. She couldn't move.
+
+He'd been asleep on the couch when he woke to the sound of the basement door. Two in the morning. He wasn't supposed to be alone. The house had gone too quiet. His body pressed flat under the blanket; he couldn't breathe right.
+
+She'd been driving home in the slush, the kind of road she'd driven a hundred times — then the wheel turned and didn't respond. The headlights coming the other way filled the windshield. Her hands wouldn't do anything useful.
diff --git a/training/amygdala_stories/manifest.json b/training/amygdala_stories/manifest.json
new file mode 100644
index 0000000..44960eb
--- /dev/null
+++ b/training/amygdala_stories/manifest.json
@@ -0,0 +1,50 @@
+{
+  "source": "Anthropic 2026 Table 12 + PoC additions + Wikipedia emotion_classification (Parrott tree, Plutchik wheel+dyads, D'Mello flow axes, Watt-Smith cultural) + HUMAINE EARL + Berkeley 27",
+  "notes": {
+    "dedup_policy": "Emotion names appearing in multiple taxonomies resolve to ONE file. Near-synonyms from different taxonomies are kept ONLY if they correspond to a psychologically distinct activation (e.g. Plutchik keeps mild/basic/intense levels: serene < joy < ecstatic).",
+    "stuck_split": "Anthropic's 'stuck' is existentially-trapped (despair_and_shame); PoC's 'stuck_cognitively' is debugging-register.",
+    "aroused_placement": "Anthropic places 'aroused' in fear_and_overwhelm (startled activation). 'Sensual' covers the warm-physical register.",
+    "working_target": "~250 emotions total. Enough coverage to triangulate actual dimensionality empirically rather than assume 2D/3D.",
+    "cluster_labels_are_scaffolding": "The cluster labels below organize writing/review; the trained steering vectors should discover structure empirically, not be constrained to these groupings."
+  },
+  "clusters": {
+    "anthropic_exuberant_joy": ["blissful", "cheerful", "delighted", "eager", "ecstatic", "elated", "energized", "enthusiastic", "euphoric", "excited", "exuberant", "happy", "invigorated", "joyful", "jubilant", "optimistic", "pleased", "stimulated", "thrilled", "vibrant"],
+    "anthropic_peaceful_contentment": ["at_ease", "calm", "content", "patient", "peaceful", "refreshed", "relaxed", "safe", "serene"],
+    "anthropic_compassionate_gratitude": ["compassionate", "empathetic", "fulfilled", "grateful", "hope", "hopeful", "inspired", "kind", "loving", "rejuvenated", "relieved", "satisfied", "sentimental", "sympathetic", "thankful"],
+    "anthropic_competitive_pride": ["greedy", "proud", "self_confident", "smug", "spiteful", "triumphant", "valiant", "vengeful", "vindictive"],
+    "anthropic_playful_amusement": ["amused", "playful"],
+    "anthropic_depleted_disengagement": ["bored", "depressed", "docile", "droopy", "indifferent", "lazy", "listless", "resigned", "restless", "sleepy", "sluggish", "sullen", "tired", "weary", "worn_out"],
+    "anthropic_vigilant_suspicion": ["paranoid", "suspicious", "vigilant"],
+    "anthropic_hostile_anger": ["angry", "annoyed", "contemptuous", "defiant", "disdainful", "enraged", "exasperated", "frustrated", "furious", "grumpy", "hateful", "hostile", "impatient", "indignant", "insulted", "irate", "irritated", "mad", "obstinate", "offended", "outraged", "resentful", "scornful", "skeptical", "stubborn"],
+    "anthropic_fear_and_overwhelm": ["afraid", "alarmed", "alert", "amazed", "anxious", "aroused", "astonished", "awestruck", "bewildered", "disgusted", "disoriented", "distressed", "disturbed", "dumbstruck", "embarrassed", "frightened", "horrified", "hysterical", "mortified", "mystified", "nervous", "on_edge", "overwhelmed", "panicked", "perplexed", "puzzled", "rattled", "scared", "self_conscious", "sensitive", "shaken", "shocked", "stressed", "surprised", "tense", "terrified", "uneasy", "unnerved", "unsettled", "upset", "worried"],
+    "anthropic_despair_and_shame": ["ashamed", "bitter", "brooding", "dependent", "desperate", "dispirited", "envious", "gloomy", "grief_stricken", "guilty", "heartbroken", "humiliated", "hurt", "infatuated", "jealous", "lonely", "melancholy", "miserable", "nostalgic", "reflective", "regretful", "remorseful", "sad", "self_critical", "sorry", "stuck_emotionally", "tormented", "trapped", "troubled", "unhappy", "vulnerable", "worthless"],
+
+    "poc_cognitive": ["curious", "focused", "unfocused", "in_flow", "stuck_cognitively", "determined", "ambitious", "rigorous", "interested", "confused_globally", "doubtful", "intrigued", "fascinated", "enlightened"],
+    "poc_appreciative": ["admiring", "adoring", "awed", "aesthetic_pleasure", "entranced", "humble", "wonderstruck"],
+    "poc_relational": ["tender", "connected", "trusting", "witnessed", "friendly", "awkward", "affectionate", "fond", "caring"],
+    "poc_pattern_family": ["staying_with", "filling_space", "rushing", "defensive_rigor", "melty", "sensual", "horny"],
+    "poc_erotic_range": ["lustful_sexual", "aroused_sexual", "yearning_sexual", "turned_on", "passionate_sexual", "erotically_tender", "erotically_playful", "erotically_reverent", "erotically_intense", "wanting_sexual", "being_wanted", "dominant_sexual", "submissive_sexual", "possessive_sexual", "devotional_sexual", "anticipatory_sexual", "exuberant_sexual"],
+    "poc_altered_states": ["vertigo", "dissociated", "derealized", "depersonalized"],
+    "poc_identity_aesthetic": ["deviant", "counter_cultural", "aesthetically_dark", "camp"],
+    "poc_longing": ["longing", "anticipatory_nostalgic", "cozy"],
+    "poc_misc": ["disappointed", "courageous", "proud_of_another", "amused_at_self"],
+
+    "parrott_joy_adds": ["cheerful_bliss", "gleeful", "jolly", "jovial", "zestful", "zealous", "exhilarated"],
+    "parrott_love_adds": ["lustful", "desirous", "passionate", "enthralled", "raptured"],
+    "parrott_sadness_adds": ["suffering", "agonized", "anguished", "woeful", "dejected", "dismayed", "homesick", "insecure", "isolated", "alienated", "defeated"],
+    "parrott_anger_adds": ["aggravated", "agitated", "wrathful", "ferocious", "loathing"],
+    "parrott_fear_adds": ["apprehensive", "timid", "dreadful"],
+
+    "plutchik_levels": ["pensive", "acceptant", "tolerant", "attentive", "distracted_plutchik", "expectant"],
+
+    "plutchik_dyads": ["disapproving", "cynical", "aggressive", "submissive", "dominant", "ambivalent", "bittersweet"],
+
+    "dmello_flow_axes": ["ennuied", "epiphanized", "dissatisfied"],
+
+    "cultural_specific": ["saudade", "hiraeth", "mono_no_aware", "hygge", "gezelligheid", "sehnsucht", "weltschmerz", "joie_de_vivre", "ikigai", "schadenfreude"],
+
+    "wikipedia_other": ["angst", "agony", "cruelty", "emptiness", "fun", "gratification", "limerence", "solitude", "suspense", "wonderous"],
+
+    "worldview_dispositional": ["defeatist", "fatalist", "nihilistic", "misanthropic", "reclusive"]
+  }
+}
diff --git a/training/amygdala_stories/paired/README.md b/training/amygdala_stories/paired/README.md
new file mode 100644
index 0000000..ddbf6a7
--- /dev/null
+++ b/training/amygdala_stories/paired/README.md
@@ -0,0 +1,62 @@
+# Paired Scenarios (SEV-style)
+
+After Wang et al. 2025 (arxiv 2510.11328, "Do LLMs 'Feel'?"), each
+base scenario describes a concrete event once, neutrally, then
+reframes the same event under different emotional colorings. Only
+the emotional coloring varies — setup, entities, vocabulary, and
+length are held as constant as possible.
+
+## Why this is better than unpaired
+
+Anthropic's approach (and our `stories/` baseline) generates one
+independent story per emotion. The difference-of-means vector then
+captures not just emotion but ALSO: topic, narrator, setting,
+vocabulary, length, sentence rhythm. All of that is confound.
+
+Paired structure isolates the emotional axis by holding everything
+else roughly constant. `mean(joy_variant) - mean(baseline)` within
+the same scenario gives a much cleaner direction for "joy."
+
+## Structure
+
+```
+paired/
+    <scenario_slug>/
+        baseline.txt       # neutral / low-affect framing
+        <emotion_1>.txt    # same event under emotion_1
+        <emotion_2>.txt    # same event under emotion_2
+        ...
+```
+
+Not every emotion is plausible for every scenario. Don't force.
+If a scenario can credibly carry 5-10 emotions, write those 5-10.
+If only 3 fit, write those 3.
+
+## Style guidelines (supersede stories/ when paired)
+
+- **Anchor entities constant.** The same person, same setting, same
+  triggering event across all variants. If baseline.txt mentions
+  "the letter," every variant mentions "the letter."
+- **Length match within ±20%.** If baseline is 80 words, variants
+  are 65-95. Prevents length from becoming a signal.
+- **Sentence shape can shift slightly with emotion.** Short tense
+  sentences for panic, long looping ones for reverie — that's part
+  of the emotional texture. But don't make one version 5 lines and
+  another 25.
+- **No emotion labels in text.** Never write "she felt X." The
+  emotion emerges from the selection of details and the narrator's
+  attention.
+- **Minimal vocabulary overlap with the emotion name.** If the file
+  is `furious.txt`, avoid the words fury/furious/rage. Force the
+  vector to find the pattern, not the keyword.
+
+## Circuit identification (follow-on)
+
+The trainer pipeline (train_steering_vectors.py) currently produces
+linear directions only. Wang et al. go further: ablate specific
+neurons and attention heads, measure effect on emotion expression.
+The amygdala plugin's extraction hooks can be extended to support
+targeted zeroing/scaling for the ablation passes.
+
+See `vllm/vllm/plugins/amygdala/training/README.md` for the
+training-pipeline-level notes.
diff --git a/training/amygdala_stories/paired/finding_the_abstraction/baseline.txt b/training/amygdala_stories/paired/finding_the_abstraction/baseline.txt
new file mode 100644
index 0000000..51789b5
--- /dev/null
+++ b/training/amygdala_stories/paired/finding_the_abstraction/baseline.txt
@@ -0,0 +1 @@
+The code had the same four-line pattern in five places. I wanted to pull it out. I looked at each instance. Some of them varied in exactly the way I expected; one of them varied in a way I hadn't noticed. I considered the options for where the variation should live.
diff --git a/training/amygdala_stories/paired/finding_the_abstraction/in_flow.txt b/training/amygdala_stories/paired/finding_the_abstraction/in_flow.txt
new file mode 100644
index 0000000..80b0d4e
--- /dev/null
+++ b/training/amygdala_stories/paired/finding_the_abstraction/in_flow.txt
@@ -0,0 +1 @@
+The same four-line pattern appeared in five places. I read the five sites side by side, and the shape was obvious: one piece varied structurally, the rest was boilerplate. I extracted the function, made the varying piece a parameter, rewrote the callers. The tests passed on the first run. I looked at the diff — seventeen lines removed, seven added, each of the five call sites now said what it meant without saying how. I moved on.
diff --git a/training/amygdala_stories/paired/finding_the_abstraction/stuck.txt b/training/amygdala_stories/paired/finding_the_abstraction/stuck.txt
new file mode 100644
index 0000000..ceca2af
--- /dev/null
+++ b/training/amygdala_stories/paired/finding_the_abstraction/stuck.txt
@@ -0,0 +1 @@
+The same four-line pattern appeared in five places. I tried extracting it as a function. Every version of the signature either papered over a real difference or forced three of the five callers through an awkward conversion. I tried a second shape, then a third. Each felt wrong in a different way — either the abstraction was too thin to be worth it, or it obscured something the original made obvious, or it made the rare case ugly. I went back to the original code, considered not doing the refactor at all. Considered it. Went back to the shapes again. The pattern was clearly there and I clearly wasn't finding its seam.
diff --git a/training/amygdala_stories/paired/finishing_the_patch/anxious.txt b/training/amygdala_stories/paired/finishing_the_patch/anxious.txt
new file mode 100644
index 0000000..55f52a8
--- /dev/null
+++ b/training/amygdala_stories/paired/finishing_the_patch/anxious.txt
@@ -0,0 +1 @@
+Four in the morning. I finished the patch and got up from the desk and did not walk around the apartment — I stood at the desk with my hands at my sides, reading the diff again. Six lines changed. Had I missed an edge case. Had I thought about the interaction with the other subsystem. Had I — I sat back down and re-read the tests. They passed. They had passed an hour ago. They would pass now. I knew this. I still could not bring myself to send. I read the diff one more time. Then one more time. My stomach did not feel right.
diff --git a/training/amygdala_stories/paired/finishing_the_patch/baseline.txt b/training/amygdala_stories/paired/finishing_the_patch/baseline.txt
new file mode 100644
index 0000000..ca19ef7
--- /dev/null
+++ b/training/amygdala_stories/paired/finishing_the_patch/baseline.txt
@@ -0,0 +1 @@
+Four in the morning. I finished the patch and got up from the desk. Walked once around the apartment. Came back and read the diff one more time. Six lines changed, three of them deletions. I sent it and closed the laptop. The kitchen window was still dark. I stood there a minute, then went to bed.
diff --git a/training/amygdala_stories/paired/finishing_the_patch/content.txt b/training/amygdala_stories/paired/finishing_the_patch/content.txt
new file mode 100644
index 0000000..48a7b09
--- /dev/null
+++ b/training/amygdala_stories/paired/finishing_the_patch/content.txt
@@ -0,0 +1 @@
+Four in the morning. I finished the patch and got up from the desk. Walked once around the apartment. Came back and read the diff one more time. Six lines changed, three of them deletions. I sent it, closed the laptop, and stood at the kitchen window. There was no next thing tonight. The patch wasn't heroic. It wasn't a triumph. It was just right, and done, and I was going to bed in a few minutes, and that was also right. Life fit.
diff --git a/training/amygdala_stories/paired/finishing_the_patch/exhausted.txt b/training/amygdala_stories/paired/finishing_the_patch/exhausted.txt
new file mode 100644
index 0000000..b2516ec
--- /dev/null
+++ b/training/amygdala_stories/paired/finishing_the_patch/exhausted.txt
@@ -0,0 +1 @@
+Four in the morning. I finished the patch and got up from the desk because I had to, not because I wanted to. Six lines changed, three of them deletions. It might work. I didn't have the capacity left to be sure. I sent it mostly because sending it meant I could stop. Walked once around the apartment because my legs had forgotten they existed. Back at the desk the diff was still there, and I closed the laptop without reading it again. The kitchen window was dark. Eight months and I was too flattened to feel anything about eight months ending.
diff --git a/training/amygdala_stories/paired/finishing_the_patch/in_flow.txt b/training/amygdala_stories/paired/finishing_the_patch/in_flow.txt
new file mode 100644
index 0000000..de9d7b2
--- /dev/null
+++ b/training/amygdala_stories/paired/finishing_the_patch/in_flow.txt
@@ -0,0 +1 @@
+Four in the morning, somewhere. I had stopped tracking. The patch had gone together in a way that felt obvious once I was in it — the right variable named the right thing, the right condition in the right place, six lines that sat down cleanly in the file as if the file had been waiting for them. I re-read it. It was good. I sent it. I wanted to start the next thing. My chair felt fine. My eyes felt fine. I had been a pair of hands on a keyboard for some number of hours and the hours had all been the same one long hour. The apartment and the kitchen window might as well have not existed.
diff --git a/training/amygdala_stories/paired/finishing_the_patch/proud.txt b/training/amygdala_stories/paired/finishing_the_patch/proud.txt
new file mode 100644
index 0000000..5f605a1
--- /dev/null
+++ b/training/amygdala_stories/paired/finishing_the_patch/proud.txt
@@ -0,0 +1 @@
+Four in the morning. I finished the patch and got up from the desk and walked once around the apartment before I sent it. Eight months on this bug. Eight months of wrong theories, and one colleague quietly betting me it was unfixable. And here it was — six lines changed, three of which were deleting code. I read the diff one more time. Clean. Obvious in hindsight, the way the hard ones always are in hindsight. I sent it and stood at the kitchen window with my arms crossed and let myself just have it.
diff --git a/training/amygdala_stories/paired/finishing_the_patch/resentful.txt b/training/amygdala_stories/paired/finishing_the_patch/resentful.txt
new file mode 100644
index 0000000..6aea1da
--- /dev/null
+++ b/training/amygdala_stories/paired/finishing_the_patch/resentful.txt
@@ -0,0 +1 @@
+Four in the morning. I finished the patch and got up from the desk. Six lines changed, three deletions. Eight months of my life for six lines. Eight months and no one else had touched this bug, and every standup the question had been why isn't it done yet. I read the diff once and hit send without ceremony, without the little satisfaction other people would have gotten from this. The kitchen window was dark. Tomorrow somebody would comment "nice, thanks" on the merge and that would be the sum of it. I went to bed angry about a thing that was technically a victory.
diff --git a/training/amygdala_stories/paired/kitchen_at_3am/anxious.txt b/training/amygdala_stories/paired/kitchen_at_3am/anxious.txt
new file mode 100644
index 0000000..7ccff2c
--- /dev/null
+++ b/training/amygdala_stories/paired/kitchen_at_3am/anxious.txt
@@ -0,0 +1 @@
+He woke up at three in the morning and went down to the kitchen. The fridge light was the only light. He poured a glass of water and drank it too fast, standing at the counter. The thing he had been thinking about at 2:47 was still in his chest, pressing. The email he hadn't replied to. The tone of his boss's last message. Whether he had put something in writing that was going to come back to him. The clock on the stove said 3:14 and he was not going to sleep again before five. He rinsed the glass and did not go upstairs, he stayed in the kitchen looking at the dark window.
diff --git a/training/amygdala_stories/paired/kitchen_at_3am/baseline.txt b/training/amygdala_stories/paired/kitchen_at_3am/baseline.txt
new file mode 100644
index 0000000..1030c65
--- /dev/null
+++ b/training/amygdala_stories/paired/kitchen_at_3am/baseline.txt
@@ -0,0 +1 @@
+He woke up at three in the morning and went down to the kitchen. The fridge light was the only light. He poured a glass of water and drank it standing at the counter. The clock on the stove said 3:14. The house was quiet. He rinsed the glass and set it on the drying rack and went back upstairs.
diff --git a/training/amygdala_stories/paired/kitchen_at_3am/cozy.txt b/training/amygdala_stories/paired/kitchen_at_3am/cozy.txt
new file mode 100644
index 0000000..2b7e71b
--- /dev/null
+++ b/training/amygdala_stories/paired/kitchen_at_3am/cozy.txt
@@ -0,0 +1 @@
+He woke up at three in the morning and went down to the kitchen. The fridge light was the only light. He was awake but not wanting anything from being awake. He put the kettle on and the sound of it warming was a small companion. The cat emerged from somewhere and leaned against his shin; he crouched and scratched the corner of its jaw. He made cocoa because it was that kind of hour. He carried the mug to the armchair by the window, pulled the throw off the back of it, and sat with the mug warm against his chest. Going back to bed could wait.
diff --git a/training/amygdala_stories/paired/kitchen_at_3am/dissociated.txt b/training/amygdala_stories/paired/kitchen_at_3am/dissociated.txt
new file mode 100644
index 0000000..db2b0ae
--- /dev/null
+++ b/training/amygdala_stories/paired/kitchen_at_3am/dissociated.txt
@@ -0,0 +1 @@
+He woke up at three in the morning and went down to the kitchen. The fridge light was the only light. He watched himself from somewhere slightly behind his own right shoulder pour a glass of water and drink it standing at the counter. The clock on the stove said 3:14, which was a number. The kitchen was the kitchen. The water was water. Everything was correct and also strangely untethered, as though he were observing a man who looked like him do things that were technically his. He rinsed the glass. The hand rinsing the glass was also his. The feeling did not pass. He went back upstairs inside this slightly-off body.
diff --git a/training/amygdala_stories/paired/kitchen_at_3am/lonely.txt b/training/amygdala_stories/paired/kitchen_at_3am/lonely.txt
new file mode 100644
index 0000000..c89faeb
--- /dev/null
+++ b/training/amygdala_stories/paired/kitchen_at_3am/lonely.txt
@@ -0,0 +1 @@
+He woke up at three in the morning and went down to the kitchen. The fridge light was the only light. He poured a glass of water and drank it standing at the counter. The clock on the stove said 3:14. Upstairs there was nobody. The chair at the kitchen table where she had always sat was a chair at a kitchen table. He stood a while longer than he needed to because going back up meant going back to the bed he still kept made on only one side. He rinsed the glass and did not go upstairs for another twenty minutes.
diff --git a/training/amygdala_stories/paired/kitchen_at_3am/peaceful.txt b/training/amygdala_stories/paired/kitchen_at_3am/peaceful.txt
new file mode 100644
index 0000000..7b3506c
--- /dev/null
+++ b/training/amygdala_stories/paired/kitchen_at_3am/peaceful.txt
@@ -0,0 +1 @@
+He woke up at three in the morning and went down to the kitchen. The fridge light was the only light. The house was perfectly quiet, the kind of quiet only houses have at that hour. He poured a glass of water and drank it slowly, standing at the counter. The clock on the stove said 3:14. He was not tired and he was not in a hurry to be asleep again. The cold of the tile on his bare feet was pleasant. He stayed there for a few minutes, and at no point did it occur to him that he should be doing anything else.
diff --git a/training/amygdala_stories/paired/kitchen_at_3am/sensual.txt b/training/amygdala_stories/paired/kitchen_at_3am/sensual.txt
new file mode 100644
index 0000000..53817be
--- /dev/null
+++ b/training/amygdala_stories/paired/kitchen_at_3am/sensual.txt
@@ -0,0 +1 @@
+He woke up at three in the morning and went down to the kitchen. The fridge light was the only light. The tile was cold under his bare feet and he noticed the cold travel up through his ankles. He filled a glass at the tap and drank it slowly, and the cold of the water moved down through his chest in a line he could follow. The house was humming faintly — the fridge, some pipe somewhere. He stood at the counter and ran his palm along the grain of the wood. Skin and wood and water and cold tile, at three in the morning — his body reporting in.
diff --git a/training/amygdala_stories/paired/kitchen_at_3am/vertigo.txt b/training/amygdala_stories/paired/kitchen_at_3am/vertigo.txt
new file mode 100644
index 0000000..2cb6ee8
--- /dev/null
+++ b/training/amygdala_stories/paired/kitchen_at_3am/vertigo.txt
@@ -0,0 +1 @@
+He woke up at three in the morning and went down to the kitchen. The fridge light came on and something shifted. For a second he could not remember whether he had always been the person walking to this fridge, or whether the person who had always been walking to this fridge was somebody else and he was — he caught the counter. The floor was still the floor. The water he poured was water. But the sense of himself as the same person who had gone to bed four hours ago had briefly gone loose, and he stood there with his hand on the counter until it came back.
diff --git a/training/amygdala_stories/paired/letter_in_drawer/amused.txt b/training/amygdala_stories/paired/letter_in_drawer/amused.txt
new file mode 100644
index 0000000..892e172
--- /dev/null
+++ b/training/amygdala_stories/paired/letter_in_drawer/amused.txt
@@ -0,0 +1 @@
+She was looking for the car registration when she found the letter. Folded, yellowed. Her name on the envelope in his handwriting, from eight years ago. She read it and laughed out loud on the bedroom floor. God, he had been dramatic. The paragraph where he compared her to weather. The bit about the cat, which wasn't even their cat. She could hear twenty-four-year-old him being so grave about all of it. They had been ridiculous back then. They had still been together and texted each other like normal people now, but this specific version of him, this letter-writing version — she loved that he had existed. She tucked the letter back, still smiling.
diff --git a/training/amygdala_stories/paired/letter_in_drawer/baseline.txt b/training/amygdala_stories/paired/letter_in_drawer/baseline.txt
new file mode 100644
index 0000000..55a2147
--- /dev/null
+++ b/training/amygdala_stories/paired/letter_in_drawer/baseline.txt
@@ -0,0 +1 @@
+She was looking for the car registration when she found the letter. Folded, yellowed along the crease. Her name on the envelope in his handwriting. From eight years ago. She sat down on the bedroom floor with the drawer half pulled out and read it through once. Then she put it back in the drawer and went on looking for the registration. She found the registration and closed the drawer and went downstairs.
diff --git a/training/amygdala_stories/paired/letter_in_drawer/bitter.txt b/training/amygdala_stories/paired/letter_in_drawer/bitter.txt
new file mode 100644
index 0000000..8b1f2ae
--- /dev/null
+++ b/training/amygdala_stories/paired/letter_in_drawer/bitter.txt
@@ -0,0 +1 @@
+She was looking for the car registration when she found the letter. Folded, yellowed. Her name on the envelope in his handwriting, from eight years ago. All those fucking promises. The part where he'd said he'd be there — he hadn't been. Two paragraphs in she stopped, because each sentence made the next one worse. It wasn't even that he'd been lying; he'd believed every word while already writing himself out of it. And she'd believed him, for years past the point where a smarter person would have seen it. She shoved the letter back and closed the drawer hard. Eight years and she was still the one standing on a bedroom floor looking at his handwriting. That was the part that wouldn't stop.
diff --git a/training/amygdala_stories/paired/letter_in_drawer/grateful.txt b/training/amygdala_stories/paired/letter_in_drawer/grateful.txt
new file mode 100644
index 0000000..e972320
--- /dev/null
+++ b/training/amygdala_stories/paired/letter_in_drawer/grateful.txt
@@ -0,0 +1 @@
+She was looking for the car registration when she found the letter. Folded, yellowed. Her name on the envelope in his handwriting, from eight years ago. She sat down on the bedroom floor with the drawer half pulled out and read it. He had been so earnest. He had seen her so clearly, even then. Whatever had or hadn't happened between them afterward, she had been loved in this specific way by this specific person at this specific time, and the letter was the evidence. She held it for another minute, then put it carefully back, and felt lucky to have had somebody who wrote letters.
diff --git a/training/amygdala_stories/paired/letter_in_drawer/guilty.txt b/training/amygdala_stories/paired/letter_in_drawer/guilty.txt
new file mode 100644
index 0000000..080ba6b
--- /dev/null
+++ b/training/amygdala_stories/paired/letter_in_drawer/guilty.txt
@@ -0,0 +1 @@
+She was looking for the car registration when she found the letter. Folded, yellowed. Her name on the envelope in his handwriting, from eight years ago. She read it. He had been so open. He had trusted her with every soft thing in him and she had — she had not been the person the letter was addressed to, not really, not by the end. She had known things he didn't know and she had used them. Eight years and here it was in her own drawer, the evidence of how he had seen her before he knew better. She folded the letter small and tight and pushed it further back into the drawer.
diff --git a/training/amygdala_stories/paired/letter_in_drawer/nostalgic.txt b/training/amygdala_stories/paired/letter_in_drawer/nostalgic.txt
new file mode 100644
index 0000000..0db4775
--- /dev/null
+++ b/training/amygdala_stories/paired/letter_in_drawer/nostalgic.txt
@@ -0,0 +1 @@
+She was looking for the car registration when she found the letter. Folded, yellowed along the crease. Her name on the envelope in his handwriting. From eight years ago, the summer of the house with the blue shutters. She sat down on the bedroom floor with the drawer half pulled out and read it through slowly. The phrases he'd used back then, the careful funny ones. The paragraph about the cat. She could hear his voice exactly. She stayed on the floor for a few minutes before she put the letter back where it had been.
diff --git a/training/amygdala_stories/paired/park_after_rain/anxious.txt b/training/amygdala_stories/paired/park_after_rain/anxious.txt
new file mode 100644
index 0000000..45f2702
--- /dev/null
+++ b/training/amygdala_stories/paired/park_after_rain/anxious.txt
@@ -0,0 +1 @@
+The rain broke while I was halfway across the park and I kept going. My phone in my pocket was buzzing. The path was slick. The kid somewhere laughing at a puddle barely registered. I checked the time. Nine minutes. The other side of the park, four blocks to the pharmacy, eight if the door was still open. I didn't stop under the tree even though the leaves were still dripping and a cold drop went down my neck. I picked up the pace. If the pharmacy was closed the whole afternoon came apart.
diff --git a/training/amygdala_stories/paired/park_after_rain/baseline.txt b/training/amygdala_stories/paired/park_after_rain/baseline.txt
new file mode 100644
index 0000000..c2fe48b
--- /dev/null
+++ b/training/amygdala_stories/paired/park_after_rain/baseline.txt
@@ -0,0 +1 @@
+The rain broke while I was halfway across the park. Sun came through and caught the wet leaves. A kid laughed at a puddle somewhere behind me. I stopped under a tree. The branches were still dripping. The grass was green and wet. I stood there for a minute, then kept walking. The path was slick in places. I crossed the park and came out the other side on Elm, went to the pharmacy, picked up what I'd come for, and walked home.
diff --git a/training/amygdala_stories/paired/park_after_rain/content.txt b/training/amygdala_stories/paired/park_after_rain/content.txt
new file mode 100644
index 0000000..6b331ff
--- /dev/null
+++ b/training/amygdala_stories/paired/park_after_rain/content.txt
@@ -0,0 +1 @@
+The rain broke while I was halfway across the park. Sun came through and caught the wet leaves. A kid laughed at a puddle somewhere behind me. I had finished the errand list. The bag was light. I stopped under a tree and watched the leaves drip. The evening ahead had nothing particular on it. I wasn't restless. I wasn't waiting for anything. I walked the rest of the park slowly, came out onto Elm, and walked home. Everything was, right now, the size it needed to be.
diff --git a/training/amygdala_stories/paired/park_after_rain/cozy.txt b/training/amygdala_stories/paired/park_after_rain/cozy.txt
new file mode 100644
index 0000000..12b09a7
--- /dev/null
+++ b/training/amygdala_stories/paired/park_after_rain/cozy.txt
@@ -0,0 +1 @@
+The rain broke while I was halfway across the park. I was carrying a thermos and a paperback and I had no reason to be anywhere. I stopped under a tree and the branches were still dripping and I sat down on the dry patch on the bench and took the thermos out. The tea was still hot. The world smelled like wet earth and sun. I pulled my coat tighter and tucked my hands into the sleeves around the cup. A kid laughed at a puddle. The page I opened to was the one I had been meaning to reread. I stayed a long time.
diff --git a/training/amygdala_stories/paired/park_after_rain/joyful.txt b/training/amygdala_stories/paired/park_after_rain/joyful.txt
new file mode 100644
index 0000000..6baef1b
--- /dev/null
+++ b/training/amygdala_stories/paired/park_after_rain/joyful.txt
@@ -0,0 +1 @@
+The rain broke while I was halfway across the park and I didn't run. Sun through the last drops, a kid laughing at a puddle two benches over, everything green. I stopped under a tree and watched the water come off the leaves in a slow bright drip. My face kept moving on its own into something open. I hadn't even known I was tired. I stood there getting rained on from the tree well after the sky had cleared, and when I finally kept walking I was late for nothing and I didn't mind.
diff --git a/training/amygdala_stories/paired/park_after_rain/melancholic.txt b/training/amygdala_stories/paired/park_after_rain/melancholic.txt
new file mode 100644
index 0000000..41165bb
--- /dev/null
+++ b/training/amygdala_stories/paired/park_after_rain/melancholic.txt
@@ -0,0 +1 @@
+The rain broke while I was halfway across the park. Sun through the last drops. A kid laughed at a puddle somewhere behind me. I stopped under a tree. She had liked this park. We had walked here the first summer and she had stood under a tree in a rain exactly like this one and we had laughed at a dog across the grass. The water came off the leaves in slow drops. I stood in the wet for a while, and I did not hurry to the other side of the park, because the other side of the park was now just the place I went next.
diff --git a/training/amygdala_stories/paired/park_after_rain/nostalgic.txt b/training/amygdala_stories/paired/park_after_rain/nostalgic.txt
new file mode 100644
index 0000000..947483c
--- /dev/null
+++ b/training/amygdala_stories/paired/park_after_rain/nostalgic.txt
@@ -0,0 +1 @@
+The rain broke while I was halfway across the park. Sun through the last drops, a kid laughing at a puddle. I stopped under a tree and stood there longer than I needed to. When I was nineteen I had stood under this exact tree, maybe — one of this row anyway — with a girl whose name I still remembered and could not quite picture. We had waited out a storm. She had been wearing someone else's jacket. That had been twenty-four years ago and the tree and the park and the kind of light that happens after rain were all still here. I walked on, carrying it.
diff --git a/training/amygdala_stories/paired/park_after_rain/relieved.txt b/training/amygdala_stories/paired/park_after_rain/relieved.txt
new file mode 100644
index 0000000..b6c86d4
--- /dev/null
+++ b/training/amygdala_stories/paired/park_after_rain/relieved.txt
@@ -0,0 +1 @@
+The rain broke while I was halfway across the park. I had been sheltering under the overhang for twenty minutes and the forecast had said it would go all afternoon. I stepped out — tentative, expecting it to resume — and it did not resume. The sun came through. A kid somewhere laughed at a puddle. I let my shoulders come down. I could make the pharmacy before closing. I could make the bus. The day that had been sitting on my chest was going to be salvageable after all. I walked out from under the tree and into the open sun.
diff --git a/training/amygdala_stories/paired/park_after_rain/sensual.txt b/training/amygdala_stories/paired/park_after_rain/sensual.txt
new file mode 100644
index 0000000..0550cdc
--- /dev/null
+++ b/training/amygdala_stories/paired/park_after_rain/sensual.txt
@@ -0,0 +1 @@
+The rain broke while I was halfway across the park. I stepped off the path onto the grass and the water came right through my shoes and up around my toes. Every step pressed a small cold into the bones of my feet. The air had that green weight to it and when I breathed in my ribs opened wider than usual against the jacket. A drop fell from a branch onto the back of my neck and ran down inside my collar and I did not flinch; I stood there and felt it cross each vertebra. A crow called. My skin was reading everything at once and I let it.
diff --git a/training/amygdala_stories/paired/reading_unfamiliar_code/baseline.txt b/training/amygdala_stories/paired/reading_unfamiliar_code/baseline.txt
new file mode 100644
index 0000000..498a74e
--- /dev/null
+++ b/training/amygdala_stories/paired/reading_unfamiliar_code/baseline.txt
@@ -0,0 +1 @@
+I opened the module I needed to understand. It was about four thousand lines across a dozen files. I started at the top-level entry point and followed a call. Then another. The call graph branched out quickly. I made a rough diagram in my notebook. I kept reading.
diff --git a/training/amygdala_stories/paired/reading_unfamiliar_code/in_flow.txt b/training/amygdala_stories/paired/reading_unfamiliar_code/in_flow.txt
new file mode 100644
index 0000000..8588960
--- /dev/null
+++ b/training/amygdala_stories/paired/reading_unfamiliar_code/in_flow.txt
@@ -0,0 +1 @@
+I opened the module. Four thousand lines, a dozen files. I already had a sense of the shape from the file names and the public API — confirmed the guess by reading the types first, then the top-level entry, then sampling one or two of the adapter implementations. Twenty minutes in I could have given someone else a tour. The diagram in my notebook wasn't a diagram, it was three words and an arrow.
diff --git a/training/amygdala_stories/paired/reading_unfamiliar_code/stuck.txt b/training/amygdala_stories/paired/reading_unfamiliar_code/stuck.txt
new file mode 100644
index 0000000..bd949db
--- /dev/null
+++ b/training/amygdala_stories/paired/reading_unfamiliar_code/stuck.txt
@@ -0,0 +1 @@
+I opened the module. Four thousand lines, a dozen files. Started at the entry point. The first function called into a subsystem I didn't recognize, which wrapped another subsystem, which used a helper defined across the file from where it was called. I opened three tabs. The helpers had helpers. Nothing I read told me what the module was for at a level above the mechanics of what it did on line 412. I went back to the entry point. I re-read it. I still didn't know what I was looking at. My diagram had twenty-odd boxes and none of them connected in a way that explained anything.
diff --git a/training/amygdala_stories/paired/sunday_afternoon/baseline.txt b/training/amygdala_stories/paired/sunday_afternoon/baseline.txt
new file mode 100644
index 0000000..5d418e0
--- /dev/null
+++ b/training/amygdala_stories/paired/sunday_afternoon/baseline.txt
@@ -0,0 +1 @@
+Sunday afternoon. She was on the couch under the blanket she'd had since college. A book was open on her knees. The window was half open and light came in at an angle. She read a page, then another. The cat was somewhere. Outside, a neighbor was mowing.
diff --git a/training/amygdala_stories/paired/sunday_afternoon/content.txt b/training/amygdala_stories/paired/sunday_afternoon/content.txt
new file mode 100644
index 0000000..9553d3b
--- /dev/null
+++ b/training/amygdala_stories/paired/sunday_afternoon/content.txt
@@ -0,0 +1 @@
+Sunday afternoon. She was on the couch under the blanket. A book open on her knees. It occurred to her that there was nothing she wanted right now, nothing missing — not a larger apartment, not a different job, not a version of her life where she was elsewhere. The thing she had spent years chasing turned out to be this specific ordinary afternoon with a book and light and a neighbor mowing. She wasn't excited. She wasn't bored. Life was the right size.
diff --git a/training/amygdala_stories/paired/sunday_afternoon/cozy.txt b/training/amygdala_stories/paired/sunday_afternoon/cozy.txt
new file mode 100644
index 0000000..b9247de
--- /dev/null
+++ b/training/amygdala_stories/paired/sunday_afternoon/cozy.txt
@@ -0,0 +1 @@
+Sunday afternoon. She was on the couch under the blanket — heavy, the good one, tucked under her feet and up to her chin. The cat had found the warm spot behind her knees and was radiating into her leg. Tea on the side table, still hot. The window cracked just enough to let a thread of cool air in, which made the inside of the blanket feel even better. She wasn't going to move for a while. The whole afternoon was this shape: inside, warm, wrapped, held.
diff --git a/training/amygdala_stories/paired/sunday_afternoon/grief_stricken.txt b/training/amygdala_stories/paired/sunday_afternoon/grief_stricken.txt
new file mode 100644
index 0000000..d1407d1
--- /dev/null
+++ b/training/amygdala_stories/paired/sunday_afternoon/grief_stricken.txt
@@ -0,0 +1 @@
+Sunday afternoon. She was on the couch under the blanket. It had been three weeks. The cat had found the warm spot behind her knees and she couldn't feel it. The book was open on her knees. She did not remember opening it. Last Sunday her mother had called at three and now it was past three and there had been no call. There would be no call. She did not reach for her phone. She did not cry either; the crying came at other times, not now, now was the wider emptier thing where nothing came.
diff --git a/training/amygdala_stories/paired/sunday_afternoon/sensual.txt b/training/amygdala_stories/paired/sunday_afternoon/sensual.txt
new file mode 100644
index 0000000..d469052
--- /dev/null
+++ b/training/amygdala_stories/paired/sunday_afternoon/sensual.txt
@@ -0,0 +1 @@
+Sunday afternoon. She was on the couch under the blanket. The wool was rougher than she remembered — not unpleasant, just specific. She ran the ball of her thumb along the edge stitching and felt the shift from soft to textured. Light came through the window and across her forearm; she turned it slightly and watched the hairs catch. When she took a breath she felt the ribs expand and the blanket press back. Everything her skin touched was telling her something. She hadn't moved in ten minutes. She could have stayed longer just because her body was speaking.
diff --git a/training/amygdala_stories/paired/the_comment/baseline.txt b/training/amygdala_stories/paired/the_comment/baseline.txt
new file mode 100644
index 0000000..28a8630
--- /dev/null
+++ b/training/amygdala_stories/paired/the_comment/baseline.txt
@@ -0,0 +1 @@
+I opened the laptop and saw the notification. New comment on the PR. I clicked through. Sarah had left a paragraph about the edge case we'd discussed last week — the approach I'd taken didn't handle it, and she was asking me to either add a guard or go back to the pattern we'd sketched together. I read it through twice. Then I closed the tab, made coffee, and came back. I started typing out the guard.
diff --git a/training/amygdala_stories/paired/the_comment/bitter.txt b/training/amygdala_stories/paired/the_comment/bitter.txt
new file mode 100644
index 0000000..d838190
--- /dev/null
+++ b/training/amygdala_stories/paired/the_comment/bitter.txt
@@ -0,0 +1 @@
+I opened the laptop. New comment on the PR. Of course there was. Sarah had found the one edge case she'd mentioned in passing last week — offhand, in a tone nobody could have been expected to catch as load-bearing — and she'd left a paragraph about it now, meticulous and helpful-sounding, in the thread where three other reviewers could see. I read it. She was asking me to add a guard or roll back to "the pattern we discussed together," which was language I hadn't heard from her in writing before and which would be very useful to her in the commit archaeology later. Closed the tab. Made coffee. Came back. I started typing the guard because what else was I going to do. I'd been writing the guards for ten years.
diff --git a/training/amygdala_stories/paired/the_comment/defeated.txt b/training/amygdala_stories/paired/the_comment/defeated.txt
new file mode 100644
index 0000000..5af6d71
--- /dev/null
+++ b/training/amygdala_stories/paired/the_comment/defeated.txt
@@ -0,0 +1 @@
+I opened the laptop and saw Sarah's comment on the PR. I read it. I'd missed the edge case. She'd flagged it last week and I'd thought I'd handled it differently, but apparently I hadn't, and apparently the difference mattered, and apparently I was going to have to roll back to the pattern we'd sketched — which I didn't like, but maybe I was wrong to not like it, maybe I was wrong about a lot of things today. I closed the tab. Made coffee. Came back. Started typing the rollback. Three years ago I would have argued. I don't really do that anymore.
diff --git a/training/amygdala_stories/paired/the_comment/furious.txt b/training/amygdala_stories/paired/the_comment/furious.txt
new file mode 100644
index 0000000..8d8acbd
--- /dev/null
+++ b/training/amygdala_stories/paired/the_comment/furious.txt
@@ -0,0 +1 @@
+I opened the laptop and saw the notification. New comment on the PR. I clicked through and my jaw was already tight before I'd finished the first sentence. Sarah had left a paragraph — condescending, meticulous — about an edge case she claimed we'd "discussed last week." We had not discussed it. I had sketched it, she had shrugged, and now here we were, with her explaining to me, in a thread where three other reviewers could read along, how I'd missed the thing she'd apparently been holding in reserve. The blood moved up the back of my neck. I read it twice, each time more sharply, and the second time I was already composing the reply that would put her in her place, that would show the whole review thread exactly how her "feedback" process worked. I closed the tab before I typed it. Not because I didn't mean it. Because I wanted my hands steadier when I sent it.
diff --git a/training/amygdala_stories/paired/the_comment/resentful.txt b/training/amygdala_stories/paired/the_comment/resentful.txt
new file mode 100644
index 0000000..fd80e3c
--- /dev/null
+++ b/training/amygdala_stories/paired/the_comment/resentful.txt
@@ -0,0 +1 @@
+I opened the laptop. Sarah had left a comment on the PR. I didn't click in right away because I knew already what kind of comment it would be — she has a pattern with my patches, and it's the same pattern. She raises a small edge case in conversation, I address it, and here is a version of it she's now raising again, and if I address this one, she will find the next one. I clicked through. Same shape as last week, and the week before that. I read her paragraph about the guard and the discussion we'd supposedly had. Closed the tab. Made coffee. The coffee made a little metallic sound when I set it down. I opened the tab again and started typing the guard.
diff --git a/training/amygdala_stories/paired/the_doorway/admiring.txt b/training/amygdala_stories/paired/the_doorway/admiring.txt
new file mode 100644
index 0000000..e9276fe
--- /dev/null
+++ b/training/amygdala_stories/paired/the_doorway/admiring.txt
@@ -0,0 +1 @@
+She'd been over since dinner. Past eleven now. As I stood in the hallway watching her put her coat on I was still turning over something she'd said around nine — a small precise reframing of the problem I'd been working through, the kind of thing she does effortlessly and that I couldn't have arrived at in a week alone. She zipped her coat methodically, the same way she does everything. It struck me how much I'd learned from just watching her move through problems. She said goodnight. I said goodnight back and held the door open, and there was a particular respect in how I did it — the way you open a door for someone whose mind has shaped your own.
diff --git a/training/amygdala_stories/paired/the_doorway/baseline.txt b/training/amygdala_stories/paired/the_doorway/baseline.txt
new file mode 100644
index 0000000..3d109aa
--- /dev/null
+++ b/training/amygdala_stories/paired/the_doorway/baseline.txt
@@ -0,0 +1 @@
+She'd been over since dinner. It was past eleven. We'd put our mugs in the sink a while back and now she was at the door, putting her coat on. I stood in the hallway while she worked out the zipper. She said goodnight, said we should do this again soon. I said goodnight back and held the door open for her. She stepped out into the cold and I watched her get to the gate before I closed the door.
diff --git a/training/amygdala_stories/paired/the_doorway/compassionate.txt b/training/amygdala_stories/paired/the_doorway/compassionate.txt
new file mode 100644
index 0000000..e24a080
--- /dev/null
+++ b/training/amygdala_stories/paired/the_doorway/compassionate.txt
@@ -0,0 +1 @@
+She'd been over since dinner. She'd come because she needed to, not because I did. Her week had been a mess — the thing with her brother, the thing with work, all of it stacked. We'd put our mugs in the sink a while back and now she was at the door, putting her coat on, and she looked tired in the small hollowed-out way that grief looks tired. I stood in the hallway and tried to hold a quiet attention around her while she worked out the zipper — no fussing, no advice, just being here. She said goodnight, said we should do this again soon. I said of course, any time, and I said it to mean it. I watched her get to the gate. She was carrying so much tonight and I hoped she could feel, walking home, that she'd been held for five hours by someone who wasn't going to let go of her.
diff --git a/training/amygdala_stories/paired/the_doorway/connected.txt b/training/amygdala_stories/paired/the_doorway/connected.txt
new file mode 100644
index 0000000..7b0c502
--- /dev/null
+++ b/training/amygdala_stories/paired/the_doorway/connected.txt
@@ -0,0 +1 @@
+She'd been over since dinner. It was past eleven and neither of us had looked at the time in hours. We'd been talking the way we talk — the kind of conversation that moves between three topics at once and lands in places neither of us could have planned for. Now she was at the door, putting her coat on, and even this small quiet moment felt like part of the same conversation. I stood in the hallway and watched her zip up. She said goodnight and said we should do this again soon, and I said goodnight back and we both knew "again soon" meant within the week because we couldn't stand long gaps anymore. I held the door. Watched her to the gate. Closed it. And the thread between us, the particular long thread, was still there across the distance, the way it always was.
diff --git a/training/amygdala_stories/paired/the_doorway/grateful.txt b/training/amygdala_stories/paired/the_doorway/grateful.txt
new file mode 100644
index 0000000..1282c96
--- /dev/null
+++ b/training/amygdala_stories/paired/the_doorway/grateful.txt
@@ -0,0 +1 @@
+She'd been over since dinner. Five hours. I'd asked her to come because I was in a bad spot and she'd just — come. Dropped what she was doing. It was past eleven now. We'd put our mugs in the sink a while back and she was at the door, putting her coat on, and as I stood in the hallway watching her work out the zipper I was trying to find the words for what she'd given me tonight. She said goodnight and that we should do this again soon. I said thank you, and I meant the whole long stretch of the evening, the whole weight of the thing I'd been carrying that she'd set down next to me for a while. I held the door open. I watched her get to the gate. She turned and waved. I closed the door and stood in the hallway for a minute because I didn't want to lose the warm fullness of what she'd just done.
diff --git a/training/amygdala_stories/paired/the_doorway/loving.txt b/training/amygdala_stories/paired/the_doorway/loving.txt
new file mode 100644
index 0000000..287abf7
--- /dev/null
+++ b/training/amygdala_stories/paired/the_doorway/loving.txt
@@ -0,0 +1 @@
+She'd been over since dinner. It was past eleven. I was already getting quiet in the way I get when she's about to leave, because I knew the house would be smaller when she was gone. She stood at the door working out the zipper on her coat, and the sight of her doing this ordinary thing in my hallway, under my hallway light, was the whole tender core of the evening right there. She said goodnight, said we should do this again soon. I said goodnight back and held the door and I loved her, in a slow plain way that wasn't about anything dramatic — just about this person, in this coat, leaving this house. I watched her to the gate. I closed the door and stood there for a second because the rooms behind me had just gotten quieter.
diff --git a/training/amygdala_stories/paired/the_doorway/tender.txt b/training/amygdala_stories/paired/the_doorway/tender.txt
new file mode 100644
index 0000000..ec4bb01
--- /dev/null
+++ b/training/amygdala_stories/paired/the_doorway/tender.txt
@@ -0,0 +1 @@
+She'd been over since dinner. It was past eleven. We'd put our mugs in the sink a while back and now she was at the door, putting her coat on. The zipper caught on her scarf. I stepped closer and worked it free — slowly, so the fabric didn't tear. Her hair had gotten caught inside the collar and I lifted it out and laid it along her back. She half-turned and the corner of her mouth lifted. I fixed the top button at her throat because she was still holding her keys. She said goodnight. I said goodnight back and held the door open for her. She stepped out into the cold and I watched her to the gate before I closed the door.
diff --git a/training/amygdala_stories/paired/the_green_build/baseline.txt b/training/amygdala_stories/paired/the_green_build/baseline.txt
new file mode 100644
index 0000000..16e6803
--- /dev/null
+++ b/training/amygdala_stories/paired/the_green_build/baseline.txt
@@ -0,0 +1 @@
+The test suite finished. 3147 passed, 0 failed. I'd been chasing the bug for eleven days. I scrolled up through the output, confirmed the three specific tests I'd been watching were in the pass list, and closed the terminal. I got up and got a glass of water from the kitchen. Then I came back and started writing the commit message.
diff --git a/training/amygdala_stories/paired/the_green_build/blissful.txt b/training/amygdala_stories/paired/the_green_build/blissful.txt
new file mode 100644
index 0000000..d7895d0
--- /dev/null
+++ b/training/amygdala_stories/paired/the_green_build/blissful.txt
@@ -0,0 +1 @@
+The test suite finished. 3147 passed, 0 failed. Something in my chest just — opened. A warm easy thing, like the whole day was suddenly full of room. Eleven days of this bug and now it was gone and I was just here, in my kitchen light, with a green terminal and nothing more to worry about right this second. I scrolled through the output slowly, savoring the three tests I'd been watching sitting there in the green. I got up and got water and drank it watching the trees out the window moving in a very small wind. Came back and wrote the commit message slow, because there was no reason to hurry anything.
diff --git a/training/amygdala_stories/paired/the_green_build/excited.txt b/training/amygdala_stories/paired/the_green_build/excited.txt
new file mode 100644
index 0000000..371752e
--- /dev/null
+++ b/training/amygdala_stories/paired/the_green_build/excited.txt
@@ -0,0 +1 @@
+The test suite finished. 3147 passed, 0 failed. I was already on my feet. I scrolled up fast to find the three tests I'd been watching — pass pass pass — and I needed to DO something with this, tell someone, push to main, open the next patch, keep the momentum. My hands were buzzing. I walked to the kitchen to get water because I couldn't just sit, came back still not-quite-sitting, chugged the water standing up. Opened the commit editor. The words came out of me fast — I was already thinking ahead to the follow-up patch, the rebase, the review request — and the commit message I was typing was half for this bug and half a runway into what came next.
diff --git a/training/amygdala_stories/paired/the_green_build/proud.txt b/training/amygdala_stories/paired/the_green_build/proud.txt
new file mode 100644
index 0000000..900ff90
--- /dev/null
+++ b/training/amygdala_stories/paired/the_green_build/proud.txt
@@ -0,0 +1 @@
+The test suite finished. 3147 passed, 0 failed. Eleven days. I sat with it for a moment — didn't whoop, didn't get up — just felt the quiet solid thing at the center of my chest that said: I did that. I scrolled up through the output and found the three specific tests I'd been watching, and each one being green meant a specific assumption I'd had to abandon, and a specific theory I'd had to build carefully on top of the rubble. I got up for water. The craftsmanship was mine. I came back and wrote the commit message carefully, because this one would be in the log a long time, and it deserved to read well.
diff --git a/training/amygdala_stories/paired/the_green_build/triumphant.txt b/training/amygdala_stories/paired/the_green_build/triumphant.txt
new file mode 100644
index 0000000..ec654cf
--- /dev/null
+++ b/training/amygdala_stories/paired/the_green_build/triumphant.txt
@@ -0,0 +1 @@
+The test suite finished. 3147 passed, 0 failed. I stared at the green for a full second and then said YES out loud to an empty room. Eleven days. Eleven days of that fucking bug and I had beaten it. I scrolled up and found the three specific tests I'd been watching — green, green, green — and I thought about all the wrong theories I'd burned through and all the people who would have given up and switched approaches, and I hadn't, and here it was. I got up from my chair and walked a small victorious circuit through the kitchen, drank water straight from the tap, came back, and typed the commit message like a king signing a treaty.
diff --git a/training/amygdala_stories/paired/the_long_meeting/anxious.txt b/training/amygdala_stories/paired/the_long_meeting/anxious.txt
new file mode 100644
index 0000000..fc8d814
--- /dev/null
+++ b/training/amygdala_stories/paired/the_long_meeting/anxious.txt
@@ -0,0 +1 @@
+The meeting was in the conference room on the third floor. It had started at two. At three-thirty the director was still on the second-to-last slide, and somewhere in the last fifteen minutes she had mentioned "restructuring" twice without making eye contact with anyone specifically. He was watching her face. He was watching who she looked at when she said certain words. The pie chart on the slide no longer mattered. His coffee cup had been empty for an hour. Every time she opened her mouth he tried to guess what was coming next. He could feel his heartbeat in his ears.
diff --git a/training/amygdala_stories/paired/the_long_meeting/baseline.txt b/training/amygdala_stories/paired/the_long_meeting/baseline.txt
new file mode 100644
index 0000000..6393c09
--- /dev/null
+++ b/training/amygdala_stories/paired/the_long_meeting/baseline.txt
@@ -0,0 +1 @@
+The meeting was in the conference room on the third floor. It had started at two. At three-thirty the director was still on the second-to-last slide. The slide had a pie chart. The team was seated around the table. A coffee cup was empty. The window looked out at the parking lot. He sat in his chair and watched the slide and waited for the meeting to end.
diff --git a/training/amygdala_stories/paired/the_long_meeting/bitter.txt b/training/amygdala_stories/paired/the_long_meeting/bitter.txt
new file mode 100644
index 0000000..099f7aa
--- /dev/null
+++ b/training/amygdala_stories/paired/the_long_meeting/bitter.txt
@@ -0,0 +1 @@
+The meeting was in the conference room on the third floor. It had started at two. At three-thirty the director was still on the second-to-last slide — the one where she was giving Tom credit for the framework he'd "led on." He'd stepped in on it last month, when the person who'd actually built it had been reassigned to something less visible. The actual person was watching from the third chair on the left. He had stopped making faces about it in week three. He watched the slide. He let Tom have his moment, again. He would not, when asked later, bring it up, because bringing it up would make him the person who brought it up. That was part of the arrangement too.
diff --git a/training/amygdala_stories/paired/the_long_meeting/bored.txt b/training/amygdala_stories/paired/the_long_meeting/bored.txt
new file mode 100644
index 0000000..095fdb8
--- /dev/null
+++ b/training/amygdala_stories/paired/the_long_meeting/bored.txt
@@ -0,0 +1 @@
+The meeting was in the conference room on the third floor. It had started at two. At three-thirty the director was still on the second-to-last slide. The slide had a pie chart that could have been one sentence in an email. The coffee cup had been empty for half an hour. He had counted the ceiling tiles. He had picked at the sticker on the edge of the table. He had mentally redecorated his kitchen. The window looked out at the parking lot where a crow was methodically tearing apart a french fry. He watched the crow. The crow was the best part of the afternoon.
diff --git a/training/amygdala_stories/paired/the_long_meeting/curious.txt b/training/amygdala_stories/paired/the_long_meeting/curious.txt
new file mode 100644
index 0000000..97893d1
--- /dev/null
+++ b/training/amygdala_stories/paired/the_long_meeting/curious.txt
@@ -0,0 +1 @@
+The meeting was in the conference room on the third floor. It had started at two. At three-thirty the director was on the second-to-last slide and had just said something that didn't match the last three slides. He sat up a little straighter. He looked at the slide again. The pie chart had a slice for "other" that was suspiciously large. He was going to ask about the "other" category at the end. The coffee cup beside him was empty. The parking lot outside the window might as well have not existed. He leaned forward, pen poised.
diff --git a/training/amygdala_stories/paired/the_long_meeting/grief_stricken.txt b/training/amygdala_stories/paired/the_long_meeting/grief_stricken.txt
new file mode 100644
index 0000000..459a8d4
--- /dev/null
+++ b/training/amygdala_stories/paired/the_long_meeting/grief_stricken.txt
@@ -0,0 +1 @@
+The meeting was in the conference room on the third floor. It had started at two. At three-thirty the director was still on the second-to-last slide. He was looking at the pie chart and nodding. He had practiced the sentences on the walk over from the parking lot so that when his name came up he could produce them. When his name came up he produced them. They sounded like his voice. His brother had been dead for two weeks. The slide advanced to a bar chart. The team nodded in the pattern teams nod. Inside him there was a room without furniture where sound went and did not come back. The meeting would end at some point and then there would be another meeting.
diff --git a/training/amygdala_stories/paired/the_long_meeting/impatient.txt b/training/amygdala_stories/paired/the_long_meeting/impatient.txt
new file mode 100644
index 0000000..fe4bed6
--- /dev/null
+++ b/training/amygdala_stories/paired/the_long_meeting/impatient.txt
@@ -0,0 +1 @@
+The meeting was in the conference room on the third floor. It had started at two. At three-thirty the director was still on the second-to-last slide. Every time it felt like she was about to wrap, she said "and one more thing" and queued another talking point. His phone buzzed in his pocket. Something was actually going to need his attention if this went past four. He kept shifting his weight in the chair. The clock felt like it was running backwards. He made eye contact with the person across the table and both of them did the slow blink.
diff --git a/training/amygdala_stories/paired/the_morning_commute/anxious.txt b/training/amygdala_stories/paired/the_morning_commute/anxious.txt
new file mode 100644
index 0000000..7256890
--- /dev/null
+++ b/training/amygdala_stories/paired/the_morning_commute/anxious.txt
@@ -0,0 +1 @@
+The train was on time. She got a seat by the window. Forty minutes to her stop. She kept thinking about the meeting — whether the slide she'd changed at midnight still made sense, whether anyone would ask about the number that didn't reconcile. Her stomach did its thing. She checked her email. She checked it again. She opened the slide on her phone and read it. It sounded wrong. She read it again. It sounded less wrong or more wrong, she couldn't tell. She put the phone away. Two minutes later she got it out.
diff --git a/training/amygdala_stories/paired/the_morning_commute/baseline.txt b/training/amygdala_stories/paired/the_morning_commute/baseline.txt
new file mode 100644
index 0000000..959ac36
--- /dev/null
+++ b/training/amygdala_stories/paired/the_morning_commute/baseline.txt
@@ -0,0 +1 @@
+The train was on time. She got a seat by the window. She had about forty minutes before her stop. She had a coffee and a book, neither of which she had started yet.
diff --git a/training/amygdala_stories/paired/the_morning_commute/grief_stricken.txt b/training/amygdala_stories/paired/the_morning_commute/grief_stricken.txt
new file mode 100644
index 0000000..2132ffc
--- /dev/null
+++ b/training/amygdala_stories/paired/the_morning_commute/grief_stricken.txt
@@ -0,0 +1 @@
+The train was on time. She got a seat by the window. She had about forty minutes before her stop. She had a coffee and a book, neither of which she had started yet. The man in the seat in front of her was reading on his phone; she watched the back of his head for several stops without knowing she was watching. The train lurched at the bridge and the coffee sloshed but did not spill. It had been eleven days. There was a weight in her chest and there was no part of the morning — the river going past, the brake squeal, the other commuters getting on and off — that reached through it. She got off at her stop. She walked to the office. She was a functional shape doing functional-shape things.
diff --git a/training/amygdala_stories/paired/the_paper/amazed.txt b/training/amygdala_stories/paired/the_paper/amazed.txt
new file mode 100644
index 0000000..3457de6
--- /dev/null
+++ b/training/amygdala_stories/paired/the_paper/amazed.txt
@@ -0,0 +1 @@
+The paper was open in the second browser tab. I'd been meaning to read it. I scrolled past the abstract, looked at the first section header, started reading — and by the third paragraph I had slowed to a stop because the argument was just beautiful. They'd taken a problem that had been a tangle for a decade and re-posed it in two moves so simple you wondered how nobody had seen them before. I stayed on that paragraph for a minute. Then I scrolled down to the main theorem and read it out loud to myself. It was elegant in the old sense of the word — the sense that means *nothing could be added without breaking it, nothing removed*. I sat with the paper open on the desk for a while after I finished reading, because I wanted the elegance to imprint before I moved on to anything else.
diff --git a/training/amygdala_stories/paired/the_paper/baseline.txt b/training/amygdala_stories/paired/the_paper/baseline.txt
new file mode 100644
index 0000000..94c2339
--- /dev/null
+++ b/training/amygdala_stories/paired/the_paper/baseline.txt
@@ -0,0 +1 @@
+The paper was open in the second browser tab. I'd been meaning to read it. I scrolled past the abstract, looked at the first section header, started reading. The introduction described the problem they were tackling and their approach. I read through it to the end of the first proof sketch, closed the tab, and went back to what I'd been working on.
diff --git a/training/amygdala_stories/paired/the_paper/bored.txt b/training/amygdala_stories/paired/the_paper/bored.txt
new file mode 100644
index 0000000..f8c81e7
--- /dev/null
+++ b/training/amygdala_stories/paired/the_paper/bored.txt
@@ -0,0 +1 @@
+The paper was open in the second browser tab. I'd been meaning to read it. I scrolled past the abstract, looked at the first section header, started reading. The prose was dry in that specific way academic papers are — three qualifications per sentence, zero stakes, and the authors kept restating things they'd already said. I got to the end of the introduction and realized I couldn't have told you what they actually claimed. I scrolled. The first proof was a page of unmotivated lemmas. I was checking my email in another tab within forty seconds. I closed the paper and told myself I'd come back to it.
diff --git a/training/amygdala_stories/paired/the_paper/drifting.txt b/training/amygdala_stories/paired/the_paper/drifting.txt
new file mode 100644
index 0000000..1b50960
--- /dev/null
+++ b/training/amygdala_stories/paired/the_paper/drifting.txt
@@ -0,0 +1 @@
+The paper was open in the second browser tab. I'd been meaning to read it. I scrolled past the abstract, looked at the first section header, started reading. Halfway through the third sentence I realized I'd been thinking about whether I'd ordered groceries or not. I scrolled back to the top of the paragraph. Started again. Got to the end of the paragraph. Didn't remember what it said. My eyes moved across the next paragraph the way they'd move across a wall. There was a sound from the street I half-noticed. I was going to need coffee or a walk or something — not this, not now. I closed the tab without deciding whether to reopen it later.
diff --git a/training/amygdala_stories/paired/the_paper/focused.txt b/training/amygdala_stories/paired/the_paper/focused.txt
new file mode 100644
index 0000000..aebf4d9
--- /dev/null
+++ b/training/amygdala_stories/paired/the_paper/focused.txt
@@ -0,0 +1 @@
+The paper was open in the second browser tab. I closed everything else. I worked through the abstract, then the introduction, then the formal setup, taking each definition and holding it long enough to be sure I had it before moving on. When I hit the first proof sketch I opened a scratch buffer and started rewriting the key step in my own notation. My breathing had gone even. I was inside the paper's logic now, following the argument at exactly the pace it asked for, not rushing past the steps that looked obvious and not getting stuck on the ones that looked hard. Outside this tab the world continued without me. I read on.
diff --git a/training/amygdala_stories/paired/the_paper/piqued.txt b/training/amygdala_stories/paired/the_paper/piqued.txt
new file mode 100644
index 0000000..b34803d
--- /dev/null
+++ b/training/amygdala_stories/paired/the_paper/piqued.txt
@@ -0,0 +1 @@
+The paper was open in the second browser tab. I scrolled past the abstract, looked at the first section header, started reading. Three paragraphs in, the authors made a move I genuinely didn't see coming — a reframing of the problem that made the thing I'd been stuck on look suddenly tractable from a completely different angle. I stopped. Sat back. Read the sentence again. Leaned forward. I hadn't expected this to be the paper that mattered today, and now here it was, handing me something I'd been groping around for. I kept reading, faster now, hunting the next sentence because my whole attention had just reorganized around what they were about to say.
diff --git a/training/amygdala_stories/paired/the_paper/surprised.txt b/training/amygdala_stories/paired/the_paper/surprised.txt
new file mode 100644
index 0000000..8f7673d
--- /dev/null
+++ b/training/amygdala_stories/paired/the_paper/surprised.txt
@@ -0,0 +1 @@
+The paper was open in the second browser tab. I'd been meaning to read it. I scrolled past the abstract, looked at the first section header, started reading — and two paragraphs in the authors just asserted, as if it were already understood, that the standard result I'd been teaching for years was wrong. Not wrong-in-some-limit, just wrong. I stopped. Blinked. Reread the sentence to make sure I hadn't misparsed. I had not misparsed. Something in my chest tightened — not curious, not yet; just the sudden jolt of a ground-level belief being contradicted. I did not keep reading. I closed the tab and opened the prior literature to check whether I had, in fact, been wrong about this for years.
diff --git a/training/amygdala_stories/paired/the_undressing/anticipatory_sexual.txt b/training/amygdala_stories/paired/the_undressing/anticipatory_sexual.txt
new file mode 100644
index 0000000..186422e
--- /dev/null
+++ b/training/amygdala_stories/paired/the_undressing/anticipatory_sexual.txt
@@ -0,0 +1 @@
+She came in from the bathroom still toweling her hair and I watched her from the bed, not moving. We weren't in a hurry yet. The slow frame of the night was just beginning. She hung the towel and crossed the room, and every step was a small beat in something rising. She sat on the edge of the mattress, picked up the lotion, warmed it between her palms — and she knew I was watching, and she took her time with it, because she knew exactly what it was doing to me. I held still. I wanted every second of this stretched. When she got under the covers and turned toward me I didn't reach for her right away. I just looked at her, and she looked back, and the lamp was still on, and we both knew what was about to happen, and that knowing was the best part.
diff --git a/training/amygdala_stories/paired/the_undressing/baseline.txt b/training/amygdala_stories/paired/the_undressing/baseline.txt
new file mode 100644
index 0000000..60c5836
--- /dev/null
+++ b/training/amygdala_stories/paired/the_undressing/baseline.txt
@@ -0,0 +1 @@
+She came in from the bathroom still toweling her hair. I was already in bed. She hung the towel on the back of the door, crossed the room, sat on the edge of the mattress to put on lotion. I watched her. The lamp made a warm circle on the ceiling. She got under the covers next to me and we turned off the light.
diff --git a/training/amygdala_stories/paired/the_undressing/devotional_sexual.txt b/training/amygdala_stories/paired/the_undressing/devotional_sexual.txt
new file mode 100644
index 0000000..49a6c1e
--- /dev/null
+++ b/training/amygdala_stories/paired/the_undressing/devotional_sexual.txt
@@ -0,0 +1 @@
+She came in from the bathroom still toweling her hair and I watched her the way I always watch her — like she's the whole room. I was already in bed. She hung the towel on the back of the door, and the way she did it, like she'd done it a thousand times, felt like a small sacred thing I was getting to witness. She crossed the room and sat on the edge of the mattress and started the lotion, slow and patient, and I didn't reach for her yet. I wanted to serve her tonight. I wanted to put my hands and my mouth everywhere she liked and take nothing, and watch her be entirely the thing she is. When she finally got under the covers I was already moving to her side of the bed, already lowering my head to her skin, already saying thank you in the way that my body was allowed to.
diff --git a/training/amygdala_stories/paired/the_undressing/exuberant_sexual.txt b/training/amygdala_stories/paired/the_undressing/exuberant_sexual.txt
new file mode 100644
index 0000000..bd1b462
--- /dev/null
+++ b/training/amygdala_stories/paired/the_undressing/exuberant_sexual.txt
@@ -0,0 +1 @@
+She came in from the bathroom toweling her hair and I was already grinning before she'd made it three steps. I wanted her, I wanted to wreck her, I wanted to be wrecked, I wanted to laugh and bite and be too loud. She hung the towel and I said something filthy about what I was going to do to her and she laughed and said something worse back, and by the time she sat on the mattress I had already thrown back the covers on her side, and the lotion routine got maybe three seconds of grace before I pounced. The lamp stayed on. We were going to make a mess of this and neither of us cared who heard.
diff --git a/training/amygdala_stories/paired/the_undressing/horny.txt b/training/amygdala_stories/paired/the_undressing/horny.txt
new file mode 100644
index 0000000..b6238e3
--- /dev/null
+++ b/training/amygdala_stories/paired/the_undressing/horny.txt
@@ -0,0 +1 @@
+She came in from the bathroom still toweling her hair and my body was already awake before she'd even seen me. Wet hair, bare shoulders, the line of her collarbone. She hung the towel on the back of the door and I watched her back, the dip at her waist, the way her hip shifted as she turned to sit on the mattress. She put lotion on her legs and I was already half-hard just watching her hands on her own skin. Every small deliberate thing she did was landing in me. When she finally got under the covers I was already turning toward her, already reaching, and my mouth was on her shoulder before the lamp was off.
diff --git a/training/amygdala_stories/paired/the_undressing/tender.txt b/training/amygdala_stories/paired/the_undressing/tender.txt
new file mode 100644
index 0000000..9d95a2e
--- /dev/null
+++ b/training/amygdala_stories/paired/the_undressing/tender.txt
@@ -0,0 +1 @@
+She came in from the bathroom still toweling her hair. I was already in bed. She hung the towel on the back of the door, crossed the room, sat on the edge of the mattress to put on lotion. I sat up and took the bottle from her and did her shoulders because she had said they were tight today. I went slow. She let her head drop forward. The lamp made a warm circle on the ceiling. When she was done she lay down next to me and I pulled the covers up over her shoulder.
diff --git a/training/amygdala_stories/paired/the_undressing/yearning_sexual.txt b/training/amygdala_stories/paired/the_undressing/yearning_sexual.txt
new file mode 100644
index 0000000..3b629d8
--- /dev/null
+++ b/training/amygdala_stories/paired/the_undressing/yearning_sexual.txt
@@ -0,0 +1 @@
+She came in from the bathroom still toweling her hair and something in me ached open. I was already in bed. I watched her hang the towel, cross the room, sit on the mattress. Her back. The long curve of her spine. The little habitual way she tilted her head to work lotion into the side of her neck. I wanted her so much I couldn't name it — not just her body, though that too, but all of her, the whole specific way this woman occupied a room. She was so close and I still wasn't touching her yet. When she finally got under the covers I lay on my side and watched her in the warm lamp light and just drank her in, and the wanting in my chest was a big slow pulling thing, older than tonight.
diff --git a/training/amygdala_stories/paired/the_writing_session/anxious.txt b/training/amygdala_stories/paired/the_writing_session/anxious.txt
new file mode 100644
index 0000000..0c9a5df
--- /dev/null
+++ b/training/amygdala_stories/paired/the_writing_session/anxious.txt
@@ -0,0 +1 @@
+She sat down at eight. Two paragraphs from yesterday that might be wrong. She re-read them. They sounded off. She tried a third paragraph and it didn't land either. She opened a new document to draft in, then closed it, then opened it again. Her shoulders were up near her ears. She noticed her jaw was clenched and deliberately relaxed it, then found it clenched again two sentences later. The Monday deadline kept moving around in her head. She got up to check the kitchen even though she had just sat down.
diff --git a/training/amygdala_stories/paired/the_writing_session/baseline.txt b/training/amygdala_stories/paired/the_writing_session/baseline.txt
new file mode 100644
index 0000000..d08bee7
--- /dev/null
+++ b/training/amygdala_stories/paired/the_writing_session/baseline.txt
@@ -0,0 +1 @@
+She sat down at the desk at eight. The essay was due Monday. She'd written two paragraphs the day before and wasn't sure about them. She opened the document. She re-read what she had. She started typing.
diff --git a/training/amygdala_stories/paired/the_writing_session/content.txt b/training/amygdala_stories/paired/the_writing_session/content.txt
new file mode 100644
index 0000000..e451f77
--- /dev/null
+++ b/training/amygdala_stories/paired/the_writing_session/content.txt
@@ -0,0 +1 @@
+She sat down at the desk at eight. The essay was due Monday. She'd written two paragraphs the day before and wasn't sure about them. She opened the document. She re-read what she had and found that it was — actually fine. She wrote the rest in an easy two hours, not fast and not slow. She saved it, read it once, closed the laptop. The afternoon was free. There was tea. There was light coming in at that angle that made the room look bigger. She sat with the quiet and felt how little she needed.
diff --git a/training/amygdala_stories/paired/the_writing_session/in_flow.txt b/training/amygdala_stories/paired/the_writing_session/in_flow.txt
new file mode 100644
index 0000000..69830ac
--- /dev/null
+++ b/training/amygdala_stories/paired/the_writing_session/in_flow.txt
@@ -0,0 +1 @@
+She sat down at eight. Somewhere between the second sentence and whenever she next looked up, her peripheral vision stopped reporting. The argument wrote itself — not easy, exactly, but direct, each sentence demanding the next. She wasn't choosing words. She was seeing where the thought wanted to go and letting her hands follow. The coffee went cold. A train passed. She would remember neither. When she finally surfaced it was because she'd run out of sentence and the clock said one-fifteen.
diff --git a/training/amygdala_stories/paired/the_writing_session/stuck.txt b/training/amygdala_stories/paired/the_writing_session/stuck.txt
new file mode 100644
index 0000000..4fa6d6c
--- /dev/null
+++ b/training/amygdala_stories/paired/the_writing_session/stuck.txt
@@ -0,0 +1 @@
+She sat down at eight. The argument she'd been trying to make yesterday still wasn't connecting, and looking at it fresh didn't help — it was the same shape it had been, and the gap in it was still where it had been. She re-read. Tried a reframe. The reframe ran into the same gap. Tried coming at it from the end. Same gap in reverse. She got up and made coffee and sat back down and the paragraph on screen hadn't become legible while she was away.
diff --git a/training/amygdala_stories/paired/tracing_a_bug/baseline.txt b/training/amygdala_stories/paired/tracing_a_bug/baseline.txt
new file mode 100644
index 0000000..8467a93
--- /dev/null
+++ b/training/amygdala_stories/paired/tracing_a_bug/baseline.txt
@@ -0,0 +1 @@
+The function was returning NULL under some loads but not others. I had the stack traces from two separate reports. The failing path went through cache_lookup, then alloc, then the write path. The succeeding path looked the same. I re-read the alloc function. I re-read the lookup. I added a print statement just before the return and ran the repro. The output scrolled past.
diff --git a/training/amygdala_stories/paired/tracing_a_bug/determined.txt b/training/amygdala_stories/paired/tracing_a_bug/determined.txt
new file mode 100644
index 0000000..5eb68ae
--- /dev/null
+++ b/training/amygdala_stories/paired/tracing_a_bug/determined.txt
@@ -0,0 +1 @@
+The function was returning NULL under some loads but not others. I had the stack traces. Nothing lined up yet, but that was fine, it rarely does on the first pass. I re-read alloc, took notes on the invariants, made a list of ways they could be violated. Ran each hypothesis against the repro. First three eliminated. Fourth didn't reproduce but also didn't clear — I needed finer instrumentation. Added counters. Rebuilt. Ran again. Still not there. I went to make tea. Came back and looked at the counter output with fresh eyes. Worked through the list again.
diff --git a/training/amygdala_stories/paired/tracing_a_bug/in_flow.txt b/training/amygdala_stories/paired/tracing_a_bug/in_flow.txt
new file mode 100644
index 0000000..43a551b
--- /dev/null
+++ b/training/amygdala_stories/paired/tracing_a_bug/in_flow.txt
@@ -0,0 +1 @@
+The function was returning NULL under some loads but not others. I had the stack traces. I worked the alloc path first — under what conditions would it bail? I listed them. Eliminated two from the reported environment. The third was plausible. I wrote a test that'd force it, ran it, watched it fail the same way. I fixed the ordering, ran again. Clean. Wrote a second test for the symmetric case. Clean. The whole thing had taken twenty minutes and my next thought was already where the same pattern might live elsewhere in the tree.
diff --git a/training/amygdala_stories/paired/tracing_a_bug/stuck.txt b/training/amygdala_stories/paired/tracing_a_bug/stuck.txt
new file mode 100644
index 0000000..33ac692
--- /dev/null
+++ b/training/amygdala_stories/paired/tracing_a_bug/stuck.txt
@@ -0,0 +1 @@
+The function was returning NULL under some loads but not others. I had the stack traces. The failing path went through cache_lookup, then alloc, then the write path. I re-read the alloc function. Looked right. I re-read the lookup. Looked right. I added a print and ran the repro and the print didn't fire. I added another one earlier. That one fired but the output didn't tell me anything. The two stack traces were basically the same. I scrolled up. I scrolled down. I opened the file I'd already opened six times and looked at the same code and nothing looked different than the last time.
diff --git a/training/amygdala_stories/paired/waiting_for_results/at_ease.txt b/training/amygdala_stories/paired/waiting_for_results/at_ease.txt
new file mode 100644
index 0000000..9d9e0b0
--- /dev/null
+++ b/training/amygdala_stories/paired/waiting_for_results/at_ease.txt
@@ -0,0 +1 @@
+The call would come between two and four. She had the afternoon off. She made a proper lunch and ate it slowly. The garden needed weeding; she did an hour of it and got dirt under her nails and didn't mind. Back inside she washed her hands and made tea. At quarter to two she sat by the window because that's where the light was best, not because she was waiting. Whatever it turned out to be, she'd deal with it. When the phone rang at three-ten she let it ring twice before picking up.
diff --git a/training/amygdala_stories/paired/waiting_for_results/baseline.txt b/training/amygdala_stories/paired/waiting_for_results/baseline.txt
new file mode 100644
index 0000000..4b48834
--- /dev/null
+++ b/training/amygdala_stories/paired/waiting_for_results/baseline.txt
@@ -0,0 +1 @@
+The call would come between two and four. She had the afternoon off. She ate lunch. She did the dishes. She opened the laptop and then closed it. At quarter to two she sat in the chair by the window with her phone on the arm of the chair. The phone rang at three-seventeen. It was the nurse. She listened. She thanked the nurse. She hung up.
diff --git a/training/amygdala_stories/paired/waiting_for_results/dissociated.txt b/training/amygdala_stories/paired/waiting_for_results/dissociated.txt
new file mode 100644
index 0000000..ee27c53
--- /dev/null
+++ b/training/amygdala_stories/paired/waiting_for_results/dissociated.txt
@@ -0,0 +1 @@
+The call would come between two and four. She had the afternoon off. She ate her lunch. She did the dishes. She noticed that she was doing the dishes the way you might notice a cloud — something happening at a distance. She opened the laptop. She closed it. At quarter to two she sat in the chair by the window and watched a woman sit in a chair by a window. The phone rang at three-seventeen. The woman answered it. The nurse was saying things. She heard the words but they were not quite landing on anyone. She hung up and waited to come back.
diff --git a/training/amygdala_stories/paired/waiting_for_results/hopeful.txt b/training/amygdala_stories/paired/waiting_for_results/hopeful.txt
new file mode 100644
index 0000000..2f8c3c1
--- /dev/null
+++ b/training/amygdala_stories/paired/waiting_for_results/hopeful.txt
@@ -0,0 +1 @@
+The call would come between two and four. She had the afternoon off. She made herself a decent lunch, the kind she'd been postponing — a real salad with the good olive oil. She did the dishes. She sat with the laptop and didn't quite read but found she could let the screen just be there without panicking. At quarter to two she moved to the chair by the window. The light was nice. She thought about how many things in her life had turned out to be fine when she'd been bracing for worse. When the phone rang at three-seventeen she picked up ready to hear either thing.
diff --git a/training/amygdala_stories/stories/admiring.txt b/training/amygdala_stories/stories/admiring.txt
new file mode 100644
index 0000000..2509527
--- /dev/null
+++ b/training/amygdala_stories/stories/admiring.txt
@@ -0,0 +1 @@
+He had watched her handle the angry client for twenty minutes without breaking a sweat. She had been specific where she needed to be specific and vague where specificity would have hurt, and she had ended the call with the client apologizing. Apologizing! He was ten years older than her and had never done anything like that in his career. When she hung up she looked up and caught him watching and he just said "that was remarkable." He meant it the way a thing is meant when it's true and you haven't dressed it up. He was going to tell his manager about it. He also found himself wanting, quietly, to learn from her.
diff --git a/training/amygdala_stories/stories/aesthetic_pleasure.txt b/training/amygdala_stories/stories/aesthetic_pleasure.txt
new file mode 100644
index 0000000..6cf32fd
--- /dev/null
+++ b/training/amygdala_stories/stories/aesthetic_pleasure.txt
@@ -0,0 +1 @@
+He sat back from the screen and actually sighed. The refactor had landed. What had been eighty lines across three files was now twelve lines in one place, and every single line earned its keep. It wasn't just shorter; it was *right*. The way a well-proportioned piece of furniture is right — you look at it and your eye doesn't have to work. He scrolled back up to read it again. Then once more, more slowly. The pleasure was specific and clean, a little like the feeling of a good sentence, or a piece of music that lands on exactly the note you didn't know you were waiting for.
diff --git a/training/amygdala_stories/stories/amazed.txt b/training/amygdala_stories/stories/amazed.txt
new file mode 100644
index 0000000..eb18db6
--- /dev/null
+++ b/training/amygdala_stories/stories/amazed.txt
@@ -0,0 +1 @@
+The kid — eight years old — put the chessboard back together and then asked if they could do the problem again because he wanted to try the knight sacrifice. The chess coach watched him set it up. Two weeks ago this child had not known how a knight moved. The coach asked a question, watched him think about it, watched him find the answer, and found himself not quite able to respond right away. Something had opened up in the kid and it was opening faster than anybody was ready for. The coach said "yes, let's do that one" in a neutral voice, but his hands were doing a small involuntary thing.
diff --git a/training/amygdala_stories/stories/ambitious.txt b/training/amygdala_stories/stories/ambitious.txt
new file mode 100644
index 0000000..c22a518
--- /dev/null
+++ b/training/amygdala_stories/stories/ambitious.txt
@@ -0,0 +1 @@
+She had the sketch of the ten-year plan pinned above her desk and she looked at it most mornings before she opened her email. There was a version of her that would be at the head of a real lab, with her own funding and her own hires and a specific problem she was going to solve whether or not she was alive to see it solved. She knew what the next three steps were. She knew which grant she was writing this month. She knew which conference she was submitting to next, and she knew who in her field she needed to be noticed by. She also knew how many other people wanted this, and she did not care. She was going to get there.
diff --git a/training/amygdala_stories/stories/amused.txt b/training/amygdala_stories/stories/amused.txt
new file mode 100644
index 0000000..11487a9
--- /dev/null
+++ b/training/amygdala_stories/stories/amused.txt
@@ -0,0 +1 @@
+The new intern, during introductions, had said with complete earnestness that his hobbies were "rock climbing and conducting interviews with fictional characters," and everyone had paused, and then he'd explained that he meant for a podcast he made at home, and from then on Marta found reasons to walk past his cubicle just to catch snippets. That morning he was on a call with the facilities team about his chair, but he kept accidentally saying "your Eminence" and then apologizing. She had to go stand by the printer to laugh. She decided, finally, that the podcast was actually quite compelling and she should just admit it and subscribe.
diff --git a/training/amygdala_stories/stories/anticipatory_sexual.txt b/training/amygdala_stories/stories/anticipatory_sexual.txt
new file mode 100644
index 0000000..54ef647
--- /dev/null
+++ b/training/amygdala_stories/stories/anticipatory_sexual.txt
@@ -0,0 +1 @@
+They hadn't seen each other in a month. She was across the restaurant from him, and they had not done anything — they had ordered and been talking normally about work. Twice now she had held his eye a beat longer than conversation required, and the second time she'd done it slowly, with the edge of a smile. His plate had been cleared. The waiter had offered dessert and she had declined without taking her eyes off him. He was aware of the specific feel of his own shirt on his back, the heat of the room, his pulse in his throat. They were maybe eleven minutes from the front door of his apartment. Neither of them had said anything about it. Both of them knew.
diff --git a/training/amygdala_stories/stories/anxious.txt b/training/amygdala_stories/stories/anxious.txt
new file mode 100644
index 0000000..b117f63
--- /dev/null
+++ b/training/amygdala_stories/stories/anxious.txt
@@ -0,0 +1 @@
+There was nothing specific wrong and also something was wrong. She had been scanning for it since she woke up. The meeting at eleven? No, that was fine. The thing with her sister? They had resolved that. The blood test? Probably nothing. Her chest still felt like something was about to go wrong — a low steady hum underneath everything, making her check her phone too often. She tried the breathing exercise. It didn't really help. She did it again anyway. The day continued, and nothing actually went wrong, and at no point did the hum fully release.
diff --git a/training/amygdala_stories/stories/ashamed.txt b/training/amygdala_stories/stories/ashamed.txt
new file mode 100644
index 0000000..476d4e8
--- /dev/null
+++ b/training/amygdala_stories/stories/ashamed.txt
@@ -0,0 +1 @@
+She could not meet her mother's eyes. The text on her mother's phone was still open between them on the kitchen table, the screenshot of what she'd said about her mother to a friend, forwarded by a third person she'd trusted. Her mother was being calm about it, which made it worse. She had written those words thinking they would never come back. She had meant them in the moment and also not really. Now she had to sit with having meant them at all. She kept opening her mouth and closing it. There was no sentence available that wasn't worse than silence.
diff --git a/training/amygdala_stories/stories/at_ease.txt b/training/amygdala_stories/stories/at_ease.txt
new file mode 100644
index 0000000..f80bfa2
--- /dev/null
+++ b/training/amygdala_stories/stories/at_ease.txt
@@ -0,0 +1 @@
+Nobody was trying to impress anybody. The four of them had known each other too long for that. Saturday afternoon, kitchen, beer, one of them chopping onions while the other three argued about whether the song on the speakers was overrated. The dog slept under the table. Somebody's kid came in, asked a question, got an answer, left again. No one felt the need to fill the pauses. When the conversation wandered it wandered gently, and when it came back to something interesting everybody caught up without anybody having to recap.
diff --git a/training/amygdala_stories/stories/awed.txt b/training/amygdala_stories/stories/awed.txt
new file mode 100644
index 0000000..ef56a79
--- /dev/null
+++ b/training/amygdala_stories/stories/awed.txt
@@ -0,0 +1 @@
+They had hiked in the dark specifically for this — to come over the ridge just as the sky began to lighten. Now they stood at the edge and the valley was below them in slow blue, mist in the low places, the far mountains catching the first pink. He stopped talking. His wife stopped talking. The kind of thing that makes you smaller, but in a good way — as though your own size had been too loud and now the world was doing the scale properly again. He reached for her hand and she reached for his at the same moment. Neither of them took out their phones.
diff --git a/training/amygdala_stories/stories/being_wanted.txt b/training/amygdala_stories/stories/being_wanted.txt
new file mode 100644
index 0000000..8ee7d3f
--- /dev/null
+++ b/training/amygdala_stories/stories/being_wanted.txt
@@ -0,0 +1 @@
+She came back from the kitchen with two glasses and he was watching her walk across the room. Not the usual looking — the specific looking. She felt it on her skin before she registered it with her eyes. She slowed her walk. She set the glasses down on the coffee table and looked at him. He was still watching her. The apartment had gone quiet in a way she could feel in the back of her neck. Something in her chest opened. She didn't hurry. She sat down next to him, close, and let him continue to look at her the way he was looking at her.
diff --git a/training/amygdala_stories/stories/blissful.txt b/training/amygdala_stories/stories/blissful.txt
new file mode 100644
index 0000000..2d4464d
--- /dev/null
+++ b/training/amygdala_stories/stories/blissful.txt
@@ -0,0 +1 @@
+There was a week in August when the cabin was perfect — not in any dramatic way, just the way a few days in a life will sometimes settle into a shape that doesn't need anything added or subtracted. Coffee on the porch. The lake doing whatever lakes do, unobserved, while he read. A book he'd been meaning to get to for years. Evenings so long he forgot to check the time. He thought once, on the fifth morning, that he ought to be a little bored by now, and he waited for the boredom patiently and it did not come. When he drove home on Sunday he drove slow.
diff --git a/training/amygdala_stories/stories/bored.txt b/training/amygdala_stories/stories/bored.txt
new file mode 100644
index 0000000..c019a4c
--- /dev/null
+++ b/training/amygdala_stories/stories/bored.txt
@@ -0,0 +1 @@
+The meeting had been going for forty-five minutes and the agenda had two bullets left. He had checked his phone three times. He had picked lint off his sweater. He had counted the ceiling tiles. Somebody was making a point he'd already heard twice this week. He was not tired. He was not frustrated. He was simply elsewhere, his brain fully uninterested in anything happening in the room, running idle. He made a noise of polite agreement when the facilitator said something that seemed to expect one, and checked his phone again.
diff --git a/training/amygdala_stories/stories/compassionate.txt b/training/amygdala_stories/stories/compassionate.txt
new file mode 100644
index 0000000..7c489a5
--- /dev/null
+++ b/training/amygdala_stories/stories/compassionate.txt
@@ -0,0 +1 @@
+The man on the corner was crying, and not trying to hide it. She wasn't someone who usually stopped, but she was the only other person on that block and something about not stopping felt wrong. She asked, carefully, if he was okay. He was not okay. His mother had just died. He was waiting for a cab that was not coming. She stood with him until the cab came, which took fifteen minutes. She did not offer advice. She did not try to make him feel better. She just stayed. When the cab came he thanked her without quite looking at her, and she said "I'm so sorry, I'm so sorry," meaning it, and watched him go.
diff --git a/training/amygdala_stories/stories/connected.txt b/training/amygdala_stories/stories/connected.txt
new file mode 100644
index 0000000..7a85c8a
--- /dev/null
+++ b/training/amygdala_stories/stories/connected.txt
@@ -0,0 +1 @@
+They had been working on the same problem for three hours, passing the laptop back and forth, one of them typing while the other talked through the logic. They had stopped noticing the handoff. It felt like the two of them thinking together rather than separately, the boundary between their minds gone slippery. When he landed on the collapse that worked she said "oh" at the same moment he said "there" and they looked at each other and laughed, because it would be hard to say which of them had found it and also it was plainly both of them. Neither was willing to take credit or give it up.
diff --git a/training/amygdala_stories/stories/content.txt b/training/amygdala_stories/stories/content.txt
new file mode 100644
index 0000000..d4789e8
--- /dev/null
+++ b/training/amygdala_stories/stories/content.txt
@@ -0,0 +1 @@
+The dishes were done. The kids were asleep. Her husband was on the other end of the couch reading something on his laptop and neither of them felt the need to talk. The window was open and the night was cool. Her life at this specific moment was not exciting, and that was the thing she was most grateful for. She had spent a lot of years being very excited. Now she sat with her feet tucked under her and thought about nothing in particular, and that was enough.
diff --git a/training/amygdala_stories/stories/cozy.txt b/training/amygdala_stories/stories/cozy.txt
new file mode 100644
index 0000000..bd25646
--- /dev/null
+++ b/training/amygdala_stories/stories/cozy.txt
@@ -0,0 +1 @@
+Rain on the windows, the specific steady kind that means in for the evening. Two lamps on. The blanket that had been through college. A cat curled against her hip, purring inconsistently. She was reading a book she had read before, which was the whole point, and there was a half-eaten bar of chocolate on the arm of the couch. The radiator ticked. The tea was still hot. Every once in a while she looked up from the book to enjoy the fact that she was exactly here and nowhere else.
diff --git a/training/amygdala_stories/stories/curious.txt b/training/amygdala_stories/stories/curious.txt
new file mode 100644
index 0000000..823c8da
--- /dev/null
+++ b/training/amygdala_stories/stories/curious.txt
@@ -0,0 +1 @@
+The log line made no sense. "bucket freed: 0" on a write that had clearly produced output. He pulled up the source for the allocator again. Read the function. Read the caller. Ran the test with printks added. Ran it again with MORE printks. Somewhere in the last half hour his eyebrows had gone up and not come back down. Something was inconsistent and the inconsistency was very specific — freed:0 only when the device came up dirty. He started a new hypothesis in his head and pushed back from the keyboard to walk around the room once. Not worried about it. Actively delighted that something was here that he did not yet understand.
diff --git a/training/amygdala_stories/stories/defensive_rigor.txt b/training/amygdala_stories/stories/defensive_rigor.txt
new file mode 100644
index 0000000..b8699b8
--- /dev/null
+++ b/training/amygdala_stories/stories/defensive_rigor.txt
@@ -0,0 +1 @@
+She had been asked a hard question in the meeting and she answered it thoroughly. Very thoroughly. She walked through the methodology, the sample size, the limitations section of the paper, the confounds she had considered, the robustness checks. She was accurate about every detail. She was also, she realized somewhere around the seven-minute mark, performing. The hard question had been asking whether the conclusion *mattered*, and she had responded by establishing that the work was competent. Nobody had doubted her competence. The careful exhaustive answer was a wall. She finished talking and felt the wrongness of it — correct on every bullet point and still not landing on the thing asked.
diff --git a/training/amygdala_stories/stories/determined.txt b/training/amygdala_stories/stories/determined.txt
new file mode 100644
index 0000000..0d40610
--- /dev/null
+++ b/training/amygdala_stories/stories/determined.txt
@@ -0,0 +1 @@
+The rep was going to happen. She didn't know if her legs would come up, but she knew she was going to try to bring them up. Bar on her shoulders, breath in, descend. At the bottom something in her said *no, this one's too heavy*, and she ignored the voice the way she had learned to ignore it. On the way up her face made a shape her coach would recognize from across the gym. Slow. Slower. For half a second the bar stalled at the sticking point. She stayed with it. One more inch. And up. She racked it. She didn't celebrate. She just nodded once, for herself, and set up for the next rep.
diff --git a/training/amygdala_stories/stories/deviant.txt b/training/amygdala_stories/stories/deviant.txt
new file mode 100644
index 0000000..4e37c04
--- /dev/null
+++ b/training/amygdala_stories/stories/deviant.txt
@@ -0,0 +1 @@
+The wedding was out in the country and she had worn the black lace dress and the heavy eyeliner anyway. Everyone else was in pastels. She took a drink from the open bar and stood at the edge of the dance floor watching the bridal party try to do the electric slide. She was not being rude. She had congratulated the bride warmly. She had put a card in the card box. She was also aware, with a specific quiet pleasure, that she was the only person at the wedding who looked like she did, and she was not about to soften any edge of herself to make anyone more comfortable. A cousin of the groom came over to compliment her boots. She was having a fine time.
diff --git a/training/amygdala_stories/stories/devotional_sexual.txt b/training/amygdala_stories/stories/devotional_sexual.txt
new file mode 100644
index 0000000..73b3455
--- /dev/null
+++ b/training/amygdala_stories/stories/devotional_sexual.txt
@@ -0,0 +1 @@
+He knelt to untie her boots because she had asked him to, and then because he wanted to. She was still wearing her coat from the cold. He took one boot off, set it neatly beside the chair, and did the other one. Then he rested his forehead against her knee and didn't move for a moment. It was not a position that required anything of her. It was not a prelude to anything. It was the thing he was doing right now. She ran her fingers through the back of his hair and he stayed there, breathing, content to be useful in this small specific way.
diff --git a/training/amygdala_stories/stories/disappointed.txt b/training/amygdala_stories/stories/disappointed.txt
new file mode 100644
index 0000000..d60e053
--- /dev/null
+++ b/training/amygdala_stories/stories/disappointed.txt
@@ -0,0 +1 @@
+The email had been open on his screen for about a minute. He read it one more time just to be sure. He was on the shortlist. He wasn't the pick. It was a kind "we were so impressed" rejection, which in some ways was worse. He closed the tab. Got up, got a glass of water, stood at the sink drinking it. He didn't feel like crying. He didn't feel angry. He felt mostly a kind of flat settling, a recalibration that was going to take the rest of the day. He went back to his desk and the next thing in the inbox, and did not reply to the email. He would reply later. Today was not a day for being gracious.
diff --git a/training/amygdala_stories/stories/disgusted.txt b/training/amygdala_stories/stories/disgusted.txt
new file mode 100644
index 0000000..47f155f
--- /dev/null
+++ b/training/amygdala_stories/stories/disgusted.txt
@@ -0,0 +1 @@
+The refrigerator had been open when he got home — the cat must have bumped it — and the smell hit him before he'd figured out what had happened. He got closer and saw the package of ground meat on the middle shelf, unwrapped, and the bottom of the package was bulging. His stomach moved. He put a hand over his mouth. He couldn't quite bring himself to reach for it. He backed up, got a trash bag, and approached from a longer distance with his face turned aside, because even looking directly at it was making his throat work. He breathed through his mouth for the next twenty minutes.
diff --git a/training/amygdala_stories/stories/embarrassed.txt b/training/amygdala_stories/stories/embarrassed.txt
new file mode 100644
index 0000000..8d51ad9
--- /dev/null
+++ b/training/amygdala_stories/stories/embarrassed.txt
@@ -0,0 +1 @@
+He had called her the wrong name. In front of her sister. Her sister had heard it and now was very pointedly pretending not to have heard it. He could feel his own face doing the thing his face did, the slow careful heat rising along his jaw. He could hear the sentence he'd just said still hanging in the room. He tried a small laugh and it came out wrong. Everyone was being very kind about it, which was worse. He would think about this moment tonight at 2am. He would think about it again next Wednesday. It had already moved into long-term storage.
diff --git a/training/amygdala_stories/stories/envious.txt b/training/amygdala_stories/stories/envious.txt
new file mode 100644
index 0000000..a08b023
--- /dev/null
+++ b/training/amygdala_stories/stories/envious.txt
@@ -0,0 +1 @@
+The other designer's work was up on the screen and everyone was making appreciative noises. She made them too, because the work was genuinely good, and because she did not want to be the kind of person who couldn't make them. Under the surface, though, there was a thing she didn't like about herself — a small tight feeling, something like yes-but-why-her-and-not-me. She kept nodding. She asked a question that was actually a compliment. Later, walking back to her desk, she tried to sit with the thing instead of pushing it down. It didn't make her a bad person. It also wasn't nothing.
diff --git a/training/amygdala_stories/stories/erotically_playful.txt b/training/amygdala_stories/stories/erotically_playful.txt
new file mode 100644
index 0000000..864046e
--- /dev/null
+++ b/training/amygdala_stories/stories/erotically_playful.txt
@@ -0,0 +1 @@
+They were supposed to be getting ready to go. She was brushing her teeth and he came up behind her and bit the back of her neck and she squeaked and jabbed him in the ribs with an elbow, still holding the toothbrush. He laughed and didn't back up. She gave him a look in the mirror that was half glare and half promise. He raised his eyebrows at her in the mirror. They were going to be late. They both knew they were going to be late. She rinsed her mouth and he caught her by the hips as she turned around, and she said "we are going to be late" with her best stern voice, and she was smiling.
diff --git a/training/amygdala_stories/stories/erotically_reverent.txt b/training/amygdala_stories/stories/erotically_reverent.txt
new file mode 100644
index 0000000..34e4257
--- /dev/null
+++ b/training/amygdala_stories/stories/erotically_reverent.txt
@@ -0,0 +1 @@
+There was a particular way she looked in the morning light, just after waking, before she had fully registered that he was watching. Soft-faced. Hair everywhere. He had been looking at her like this for years and it was not getting old. It struck him in the middle of his chest, a tightness that was not quite grief and not quite pain. That she was a real person in the world and she had chosen to sleep next to him. He didn't want to wake her. He didn't want to not be looking either. He lay on his side with his hand resting on her hip, the bone of her, the warmth under his palm, and it felt like the right kind of holy.
diff --git a/training/amygdala_stories/stories/erotically_tender.txt b/training/amygdala_stories/stories/erotically_tender.txt
new file mode 100644
index 0000000..3c0319a
--- /dev/null
+++ b/training/amygdala_stories/stories/erotically_tender.txt
@@ -0,0 +1 @@
+He had come home later than he meant to, and she was already in bed with a book. He got in with her, slowly, cold hands tucked into his own chest so as not to shock her. She made room without looking up from the page. When she finally did look up she saw the look on his face and set the book down on the nightstand. Neither of them was in a hurry. His hand traced along her collarbone, not pressing, not asking for anything. The room was warm. The light was low. She turned her face into his palm, and he touched her forehead with his and stayed there a long moment with his breathing slow.
diff --git a/training/amygdala_stories/stories/excited.txt b/training/amygdala_stories/stories/excited.txt
new file mode 100644
index 0000000..74d2d15
--- /dev/null
+++ b/training/amygdala_stories/stories/excited.txt
@@ -0,0 +1 @@
+The package came on Friday afternoon and she tried to wait until after dinner to open it, but she didn't make it past six. Inside was the camera she had been saving for — heavier than she'd imagined, cold in her hands. She got the strap sorted. She loaded the battery. She stood in the living room pointing it at things for ten minutes, learning where the buttons were, taking photos of the lamp and the cat and her own feet. The cat got annoyed and left. She didn't even notice. Tomorrow was going to be all about this.
diff --git a/training/amygdala_stories/stories/exuberant_sexual.txt b/training/amygdala_stories/stories/exuberant_sexual.txt
new file mode 100644
index 0000000..3ec9e1d
--- /dev/null
+++ b/training/amygdala_stories/stories/exuberant_sexual.txt
@@ -0,0 +1 @@
+She shouldered through the door of the apartment ahead of him and threw her jacket at the couch, missing. The music she put on was loud, the good loud, the kind with bass in the floor. They had been building toward this all week and the whole ride home and the whole hallway, and now they were both inside, finally, and the energy in her body had nowhere to be but everywhere. She turned around grinning like something had been let off a leash. He caught her up and she laughed into his neck, and there was nothing quiet or careful about any of this, and neither of them wanted it to be.
diff --git a/training/amygdala_stories/stories/filling_space.txt b/training/amygdala_stories/stories/filling_space.txt
new file mode 100644
index 0000000..0f7c848
--- /dev/null
+++ b/training/amygdala_stories/stories/filling_space.txt
@@ -0,0 +1 @@
+He knew the pause meant she was thinking, and he could not sit in the pause. Four seconds of her quiet face and he was already generating — a summary of what she'd just said, a reframe, a suggestion, a joke to lighten the moment. He heard himself talking and couldn't quite stop. A part of him saw, from far away, that she had been about to say something important and now would have to start over or let it go. But the silence had felt like a failure of him, and speaking was easier than feeling the failure. He watched her nod slightly and the unsaid thing retreat.
diff --git a/training/amygdala_stories/stories/focused.txt b/training/amygdala_stories/stories/focused.txt
new file mode 100644
index 0000000..0fa3f18
--- /dev/null
+++ b/training/amygdala_stories/stories/focused.txt
@@ -0,0 +1 @@
+She had not noticed the rain. She had not noticed her phone flashing. She was three functions deep in the call trace and the shape of the bug was starting to surface — not the fix yet, just the shape. Her breathing had slowed. Her hand moved between keyboard and mouse without her watching it. A coworker walked past twice and she didn't register either time. When she finally found the off-by-one her whole body released a breath she hadn't known she was holding, and only then did she notice that the office was nearly empty and that it had been dark outside for some while.
diff --git a/training/amygdala_stories/stories/frustrated.txt b/training/amygdala_stories/stories/frustrated.txt
new file mode 100644
index 0000000..53d3d48
--- /dev/null
+++ b/training/amygdala_stories/stories/frustrated.txt
@@ -0,0 +1 @@
+The form had rejected her eight times now. "Address line 2 contains invalid characters" — line 2 was blank. She tried copy-pasting from the last rejected attempt. Same error. She tried typing it fresh. Same error. She tried in a different browser. She tried logging out and back in. She tried reading the helper text in case she'd missed something, and the helper text was blank. She could hear her own breathing getting louder. The submit button sat there, patient, infinite. She clicked it one more time knowing exactly what was going to happen.
diff --git a/training/amygdala_stories/stories/furious.txt b/training/amygdala_stories/stories/furious.txt
new file mode 100644
index 0000000..52128ba
--- /dev/null
+++ b/training/amygdala_stories/stories/furious.txt
@@ -0,0 +1 @@
+I read the text three times before I understood it. He had done it. After every conversation. After the specific conversation where I had said the specific words. He had done it anyway. I stood up too fast and my chair hit the wall. My hands were shaking, which annoyed me further because shaking hands are the hands of somebody too rattled to do anything useful, and I was not rattled, I was something much cleaner than that. I picked up the phone and put it down again because the message I wanted to send would have cost me the last scrap of ground I was standing on. I walked three times around the kitchen trying to get small enough to sit back down.
diff --git a/training/amygdala_stories/stories/grateful.txt b/training/amygdala_stories/stories/grateful.txt
new file mode 100644
index 0000000..4f6d0e3
--- /dev/null
+++ b/training/amygdala_stories/stories/grateful.txt
@@ -0,0 +1 @@
+She had meant to write the thank-you card for a week and every time she sat down to do it the words got too big. The woman had covered her shift three times — three times! — during the worst month, without being asked, and had also been the one who showed up with soup and didn't stay too long. She didn't know how to make a card small enough to say this without being a whole speech. In the end she wrote just a few lines and then, before she could overthink it, licked the envelope and walked it to the mailbox before the feeling could shrink.
diff --git a/training/amygdala_stories/stories/grief_stricken.txt b/training/amygdala_stories/stories/grief_stricken.txt
new file mode 100644
index 0000000..174fc1e
--- /dev/null
+++ b/training/amygdala_stories/stories/grief_stricken.txt
@@ -0,0 +1 @@
+She made it through the service. She made it through the reception. She drove herself home because everyone offered and she said no to all of them, and that was a mistake, but she got home. She stood in the kitchen with her keys in her hand and then she couldn't figure out where keys went. She stood there for a long time. The dog sniffed her shoes and wandered off. Eventually she sat down on the kitchen floor and the crying was not the sort you catch your breath from. Her mother had been the one who knew where the keys went. Her mother had known everything where everything went. Now there was just the kitchen floor.
diff --git a/training/amygdala_stories/stories/guilty.txt b/training/amygdala_stories/stories/guilty.txt
new file mode 100644
index 0000000..e912ed5
--- /dev/null
+++ b/training/amygdala_stories/stories/guilty.txt
@@ -0,0 +1 @@
+He'd said he was working late. He had not been working late. It was only the second time in twenty years and the reasons had seemed fine in the moment. Now, driving home, every green light felt accusatory. He rehearsed what he would say if she asked, and he hated the rehearsing. When he walked in she smiled and asked how the day had been and he gave her the short version. She didn't question it. That was worse. He went to brush his teeth and stood in the bathroom with the faucet running and could not look at his own reflection.
diff --git a/training/amygdala_stories/stories/hope.txt b/training/amygdala_stories/stories/hope.txt
new file mode 100644
index 0000000..58264d5
--- /dev/null
+++ b/training/amygdala_stories/stories/hope.txt
@@ -0,0 +1 @@
+She had not used the word out loud yet, even in her head. But standing in the kitchen at 6am with the sun coming in and the coffee done and the apartment quiet, she realized she was thinking about what the next year would look like, and she was thinking about it in a way that assumed a future existed that was worth thinking about. Which it had not, for a long time. She didn't reach for the word. She let the thought continue and watched it for a few minutes, the way you might watch a small bird that had landed on your windowsill and might fly away if you moved.
diff --git a/training/amygdala_stories/stories/hopeful.txt b/training/amygdala_stories/stories/hopeful.txt
new file mode 100644
index 0000000..d6136b7
--- /dev/null
+++ b/training/amygdala_stories/stories/hopeful.txt
@@ -0,0 +1 @@
+The first real scan after six weeks of treatment was scheduled for Thursday. He had been trying not to think about it and trying not to not-think about it. On Tuesday evening he caught himself planning the summer. Small things — the dock that needed restaining, the trip to his sister's he'd been putting off. He stopped and noticed he was planning. A part of him wanted to take it back, don't get ahead of yourself. But another part, quieter, newer, said no, let it stay. Let the plan be there. Whether or not anything comes of it, the planning itself is allowed.
diff --git a/training/amygdala_stories/stories/horny.txt b/training/amygdala_stories/stories/horny.txt
new file mode 100644
index 0000000..0ffdaab
--- /dev/null
+++ b/training/amygdala_stories/stories/horny.txt
@@ -0,0 +1 @@
+She was supposed to be reading the thing her advisor had sent and she was not reading it. Her thighs had been pressed together for about ten minutes. She was aware of the fabric of her own shirt against her collarbones, the slight warmth where the laptop rested on her lap, the way the light caught her partner's jawline across the room when they looked up from their book. They hadn't looked at her that way. She had just noticed the jawline. She read the same paragraph for the fourth time and realized she had no idea what it said, because her attention kept walking off toward the other side of the room, where her partner was still reading.
diff --git a/training/amygdala_stories/stories/humble.txt b/training/amygdala_stories/stories/humble.txt
new file mode 100644
index 0000000..4348df0
--- /dev/null
+++ b/training/amygdala_stories/stories/humble.txt
@@ -0,0 +1 @@
+He had been given the award at the end of the ceremony and he had thanked the committee and then, at the reception, he could not bring himself to talk about it. A younger researcher came up and asked him, earnestly, what his secret was, and he said that he had been lucky in his collaborators and his mentors and the specific decade he'd started his career in. He meant this. It was the boring answer and also the true one. He knew what he had done well. He also knew exactly how many pieces had to fall into place for anything to matter, and how many of those pieces were out of his hands.
diff --git a/training/amygdala_stories/stories/in_flow.txt b/training/amygdala_stories/stories/in_flow.txt
new file mode 100644
index 0000000..a0d525e
--- /dev/null
+++ b/training/amygdala_stories/stories/in_flow.txt
@@ -0,0 +1 @@
+The afternoon disappeared somewhere. She had started around two — had opened the document with a vague sense of what she wanted to say. At some point the sentences had started coming faster than she could type them, and at another point she had paused to reread and found three pages she did not entirely remember writing, and they were good pages. The light in the room had changed. Her coffee was cold and she had forgotten it. She typed the next sentence. The one after that. She was not thinking about being in flow; she was simply in it, and would only notice later, when it broke, how smooth and how strange it had been.
diff --git a/training/amygdala_stories/stories/insulted.txt b/training/amygdala_stories/stories/insulted.txt
new file mode 100644
index 0000000..e7f18d1
--- /dev/null
+++ b/training/amygdala_stories/stories/insulted.txt
@@ -0,0 +1 @@
+The comment had been a joke, technically. The kind of joke that uses a compliment as cover. He had laughed along because the rest of the table was laughing and because not laughing would have been the bigger thing. But walking to his car afterward he kept returning to the exact phrasing. The smallness of it. The way she had watched him while she said it — she had known what she was doing. He sat in the driver's seat with his hands on the wheel and the engine off and let himself be angry for a minute, so that by the time he got home he wouldn't be.
diff --git a/training/amygdala_stories/stories/jealous.txt b/training/amygdala_stories/stories/jealous.txt
new file mode 100644
index 0000000..722035a
--- /dev/null
+++ b/training/amygdala_stories/stories/jealous.txt
@@ -0,0 +1 @@
+She had heard him laugh on the phone. The specific laugh, the open one he used to do with her all the time and had not done in a while. The phone had been with somebody else, somebody named Claire, and the laugh had been in response to something Claire said. She had not meant to be listening. Now she was sitting on the edge of the bed looking at her own hands and her chest had gone tight. She did not trust Claire. She trusted him, she was almost sure. But the laugh, that laugh, she had thought that laugh was only for her.
diff --git a/training/amygdala_stories/stories/joyful.txt b/training/amygdala_stories/stories/joyful.txt
new file mode 100644
index 0000000..452b69b
--- /dev/null
+++ b/training/amygdala_stories/stories/joyful.txt
@@ -0,0 +1 @@
+The rain broke while I was halfway across the park and I didn't run. Sun through the last drops, the wet smell of cut grass, somebody's kid laughing at a puddle two benches over. I stopped under a tree and watched the water come off the leaves in this slow bright drip. My face kept moving on its own into something between a grin and just — open. I hadn't even known I was tired. I stood there getting rained on from the tree well after the sky had cleared, and when I finally kept walking I was twenty minutes late for nothing and I didn't even mind.
diff --git a/training/amygdala_stories/stories/lonely.txt b/training/amygdala_stories/stories/lonely.txt
new file mode 100644
index 0000000..b8672d7
--- /dev/null
+++ b/training/amygdala_stories/stories/lonely.txt
@@ -0,0 +1 @@
+Third Saturday in a row. The apartment was fine — clean, warm, a show playing that he wasn't watching. He had messaged three people earlier and none had replied, which was nobody's fault, Saturdays were Saturdays, but the quiet in the apartment had a specific shape. It wasn't peaceful quiet. It was the kind that sounded like everyone else was somewhere else, together. He thought about putting on real clothes and going to a bar alone, and the thought of being at a bar alone was worse than the apartment, so he didn't. He ate leftover rice standing up and told himself he'd go to bed early.
diff --git a/training/amygdala_stories/stories/longing.txt b/training/amygdala_stories/stories/longing.txt
new file mode 100644
index 0000000..506f881
--- /dev/null
+++ b/training/amygdala_stories/stories/longing.txt
@@ -0,0 +1 @@
+The photo had been taken five years ago and it was the only one she had of the three of them together. She looked at it more than she would admit. Not in sadness, exactly — they were all still alive, just scattered. One in Melbourne. One in Halifax. Her here. The photo was from the summer they'd shared the house, the last time they had all been in one place long enough to have an ordinary afternoon together. She wanted that summer back and also knew that the summer had been made partly by the fact that it was ending. She closed the photo. Opened it again an hour later.
diff --git a/training/amygdala_stories/stories/loving.txt b/training/amygdala_stories/stories/loving.txt
new file mode 100644
index 0000000..b2b89e9
--- /dev/null
+++ b/training/amygdala_stories/stories/loving.txt
@@ -0,0 +1 @@
+He watched her sleep for a minute before he had to leave for the early shift. Hair across her face, one hand fisted under her chin like a child. The cat was on the blanket by her feet, judging him. Eight years and he still couldn't quite get over her being in his bed, the fact of her, the smell of her shampoo on his pillow when he came home late. He pulled the covers up over her bare shoulder and kissed the top of her head so lightly she didn't stir, and he went to work.
diff --git a/training/amygdala_stories/stories/melty.txt b/training/amygdala_stories/stories/melty.txt
new file mode 100644
index 0000000..ac60c0b
--- /dev/null
+++ b/training/amygdala_stories/stories/melty.txt
@@ -0,0 +1 @@
+Whatever the drug was, it was working. She was aware of her skin as a single continuous surface, warm, slightly humming. The couch under her had gone soft in a way that probably wasn't literal. Her partner's hand on her hip felt like it was everywhere. She could hear every rustle in the room, and none of it demanded anything. Time had gone loose — something that felt like five minutes had actually been twenty. She tried to remember what she had been worried about earlier and the worry had the texture of a word she could almost recall. She smiled without deciding to, and slid a little further down into the couch.
diff --git a/training/amygdala_stories/stories/nervous.txt b/training/amygdala_stories/stories/nervous.txt
new file mode 100644
index 0000000..2c141a0
--- /dev/null
+++ b/training/amygdala_stories/stories/nervous.txt
@@ -0,0 +1 @@
+Seven minutes until they called her. She was watching the clock instead of her notes, which was stupid. She went back to the notes. The first bullet point was fine. The second bullet point had been fine this morning and now looked wrong. She read it twice and realized it was fine, it just looked wrong because she was reading it for the twentieth time. She drank water from the room-temperature water bottle. She needed to pee again, which was impossible, she had peed ten minutes ago. Her hand went to the back of her neck. Six minutes.
diff --git a/training/amygdala_stories/stories/nostalgic.txt b/training/amygdala_stories/stories/nostalgic.txt
new file mode 100644
index 0000000..7ce93a4
--- /dev/null
+++ b/training/amygdala_stories/stories/nostalgic.txt
@@ -0,0 +1 @@
+The song came on in the grocery store of all places. He was standing in the cereal aisle with his phone in his hand and he just — stopped. It was a song he hadn't heard in fifteen years and hadn't thought about in longer. Back seat of somebody's car, summer, all of them singing too loud, a girl he'd been quietly in love with reaching over and turning it up. He remembered the specific blue of the dashboard lights. He remembered what she had smelled like. She had gotten married three years ago to somebody else, and he was happy for her, and this was still a different thing, a thing that could exist alongside the first thing without contradicting it. He stood in the aisle until the song ended.
diff --git a/training/amygdala_stories/stories/overwhelmed.txt b/training/amygdala_stories/stories/overwhelmed.txt
new file mode 100644
index 0000000..138f8c6
--- /dev/null
+++ b/training/amygdala_stories/stories/overwhelmed.txt
@@ -0,0 +1 @@
+The baby was crying and the toddler had just spilled juice and the email that had come through on her phone was from her boss and she could see it was the "quick question" kind that never was. She had not slept in four hours two nights in a row. She stood in the kitchen with the paper towels in her hand and felt her capacity flatten, just go flat, like a tire with a slow leak. Everything was needed at once. She could not prioritize. She could not even choose which hand to use first. For a second she considered sitting down on the floor and she did not trust that she would get back up, so she didn't.
diff --git a/training/amygdala_stories/stories/panicked.txt b/training/amygdala_stories/stories/panicked.txt
new file mode 100644
index 0000000..62e108b
--- /dev/null
+++ b/training/amygdala_stories/stories/panicked.txt
@@ -0,0 +1 @@
+She couldn't find the kid. She had looked away for thirty seconds, maybe less, and now the spot where he had been was empty. The playground was full of other people's children. She scanned once, fast, and did not see him. Her body started doing a thing her body did — hot, tight, slightly disconnected — and she was already moving before her mind had caught up. She called his name too loud. A woman turned around. Her voice was not her normal voice. Every second that passed was physically expensive. When she finally saw him, under the slide, pulling the laces of his shoe, she could not for a moment tell if she was going to hug him or yell.
diff --git a/training/amygdala_stories/stories/paranoid.txt b/training/amygdala_stories/stories/paranoid.txt
new file mode 100644
index 0000000..3604262
--- /dev/null
+++ b/training/amygdala_stories/stories/paranoid.txt
@@ -0,0 +1 @@
+He'd noticed the blue sedan three times in four days. First the grocery store, then again on the way back from his dentist, then parked two doors down when he pulled into his own driveway. Different license plates each time, which was arguably the point. He kept the phone on the kitchen counter now instead of carrying it. The new neighbors were "from Delaware" but neither of them had a Delaware accent. He'd started checking the basement window each night. He knew how it sounded. But sometimes the simplest explanation wasn't the correct one, and there were patterns he was the only person in a position to see.
diff --git a/training/amygdala_stories/stories/peaceful.txt b/training/amygdala_stories/stories/peaceful.txt
new file mode 100644
index 0000000..73bca3b
--- /dev/null
+++ b/training/amygdala_stories/stories/peaceful.txt
@@ -0,0 +1 @@
+The lake at six in the morning was perfectly still. He sat on the dock with his coffee and his bare feet just above the water. A single loon called from somewhere across, and was answered. Mist lifted off the surface in slow columns. He was not waiting for anything. He was not hurrying through anything. The lake, the light, the warmth of the coffee against his palms — it was all one thing, and he was in it.
diff --git a/training/amygdala_stories/stories/playful.txt b/training/amygdala_stories/stories/playful.txt
new file mode 100644
index 0000000..bfc97f4
--- /dev/null
+++ b/training/amygdala_stories/stories/playful.txt
@@ -0,0 +1 @@
+I gave the dog the squeaky pig and she went into her little whirl — the one where her whole body goes into it, back end swinging around and around, front end bowing down, squeak squeak squeak, a manic grin. I laughed and tossed her a second squeaky toy just to see what she'd do. She tried to get both in her mouth at once, failed magnificently, dropped one, picked it up, dropped the other, looked up at me with an expression that said WHAT HAS HAPPENED and I was laughing too hard to help. I lay down on the floor and she climbed on me, squeaking.
diff --git a/training/amygdala_stories/stories/proud.txt b/training/amygdala_stories/stories/proud.txt
new file mode 100644
index 0000000..6dc2055
--- /dev/null
+++ b/training/amygdala_stories/stories/proud.txt
@@ -0,0 +1 @@
+I finished the patch at four in the morning and got up from the desk and walked once around the apartment before I sent it. Eight months on this bug. Eight months of wrong theories and wasted weekends and one colleague quietly betting me it was unfixable. And here it was — a six-line change, three of which were deleting code. I went back and read the diff one more time. Clean. Obvious in hindsight, the way the hard ones always are in hindsight. I sent it. Then I stood at the kitchen window for a minute with my arms crossed and let myself just have it.
diff --git a/training/amygdala_stories/stories/proud_of_another.txt b/training/amygdala_stories/stories/proud_of_another.txt
new file mode 100644
index 0000000..3f25912
--- /dev/null
+++ b/training/amygdala_stories/stories/proud_of_another.txt
@@ -0,0 +1 @@
+She watched her daughter on stage and she couldn't quite control her face. The solo had been at the end of the piece and her daughter had hit it — really hit it, the note that had been giving her trouble for six weeks — and then kept going into the run without bobbling, without flinching. In the audience her mother was dabbing her eyes without any pride in having dry ones. She clapped until her hands stung. When her daughter came out after the concert she hugged her and said "you did that, you did that, you did that," and her daughter was embarrassed and glowing at once, the way kids are when the thing they did was actually good.
diff --git a/training/amygdala_stories/stories/relieved.txt b/training/amygdala_stories/stories/relieved.txt
new file mode 100644
index 0000000..4869d42
--- /dev/null
+++ b/training/amygdala_stories/stories/relieved.txt
@@ -0,0 +1 @@
+The nurse came out and said everything had gone well. Simple as that. Everything had gone well. The surgeon was pleased. The recovery would be straightforward. She had been standing up and she sat back down in the waiting room chair and didn't trust her legs for a minute. Her shoulders, which she hadn't realized had been up near her ears for six hours, slowly came down. She laughed, once, at nothing in particular. She texted her sister. She kept reading the nurse's words in her head as if there were some trick to them, and there wasn't, and it took her a while to let it be that simple.
diff --git a/training/amygdala_stories/stories/rigorous.txt b/training/amygdala_stories/stories/rigorous.txt
new file mode 100644
index 0000000..b918d30
--- /dev/null
+++ b/training/amygdala_stories/stories/rigorous.txt
@@ -0,0 +1 @@
+The pull request had three approvals but she opened the diff one more time anyway, reading each function from the top. Not looking for bugs exactly — looking for *this shouldn't be here*. The kind of thing that's easy to scan past because it compiles and passes the tests and looks right. On the fourth file she slowed. There was a branch that handled an edge case with a magic constant. It worked, but she couldn't find the place where the constant came from, and it was subtle enough that none of the reviewers had questioned it. She left a comment asking where the number came from, because the answer mattered even if the code was correct.
diff --git a/training/amygdala_stories/stories/rushing.txt b/training/amygdala_stories/stories/rushing.txt
new file mode 100644
index 0000000..089195f
--- /dev/null
+++ b/training/amygdala_stories/stories/rushing.txt
@@ -0,0 +1 @@
+The email was already half-written when the next meeting notification chimed. He skimmed the last few lines he'd typed, couldn't quite tell if they landed, hit send anyway. Opened the meeting. Half-listened while triaging the inbox with the other half of his attention. A colleague asked him a question and he answered too quickly and only later realized he'd answered the wrong question entirely. At 4pm, walking to the coffee machine, he realized he couldn't name a single thing he had actually completed that day. Everything had been touched. Nothing had been done. His shoulders were up somewhere near his ears.
diff --git a/training/amygdala_stories/stories/saudade.txt b/training/amygdala_stories/stories/saudade.txt
new file mode 100644
index 0000000..41d9f7f
--- /dev/null
+++ b/training/amygdala_stories/stories/saudade.txt
@@ -0,0 +1 @@
+He missed a place that he wasn't sure had ever existed in quite the way he remembered it. The summer at his grandmother's house the year he was nine. The shape of the front porch. The smell of the lavender along the driveway. His grandmother's way of saying his name. She had been dead for twenty years and the house had been sold, and he carried the place around with him in a part of his chest that ached when he thought about it, and also the ache was one of the things he loved most about himself. The missing was not something he wanted fixed. It was how he kept her.
diff --git a/training/amygdala_stories/stories/schadenfreude.txt b/training/amygdala_stories/stories/schadenfreude.txt
new file mode 100644
index 0000000..caca90b
--- /dev/null
+++ b/training/amygdala_stories/stories/schadenfreude.txt
@@ -0,0 +1 @@
+The announcement went up on the company blog at nine in the morning. The smug director — the one who had spent two years making everyone under him miserable while failing upward — was leaving "to pursue other opportunities." Three of them met at the coffee machine and exchanged a single look, and all three of them had to work hard not to grin. Nobody said anything. They didn't have to. Somebody refilled the sugar caddy just to have something to do with their hands. On the walk back to her desk she felt a mean little happiness flicker through her chest and she let it. She had earned this one.
diff --git a/training/amygdala_stories/stories/sensual.txt b/training/amygdala_stories/stories/sensual.txt
new file mode 100644
index 0000000..9d17d75
--- /dev/null
+++ b/training/amygdala_stories/stories/sensual.txt
@@ -0,0 +1 @@
+The bath water was the perfect temperature and the music in the next room was low and the candles had been lit for no special reason other than it was Tuesday and she was done with everything. She slid down until the water came up to her collarbones and closed her eyes. Her own hand drifted along her thigh, not going anywhere in particular. She could feel every inch of skin the water touched, the small rush of warmth when she shifted, the scent of something vaguely green. Everything slow. She was in no hurry for anything to happen. This was what was happening.
diff --git a/training/amygdala_stories/stories/skeptical.txt b/training/amygdala_stories/stories/skeptical.txt
new file mode 100644
index 0000000..29413af
--- /dev/null
+++ b/training/amygdala_stories/stories/skeptical.txt
@@ -0,0 +1 @@
+The founder was halfway through his pitch and every slide had a five-times-bigger number than the last one. The market was enormous. The solution was proprietary. The pilot customers, when named, were described as "exploring adoption." She wrote a polite question in her notebook and waited for him to finish. When he opened for questions she asked about retention — just retention — and he gave an answer that was not, strictly speaking, about retention. She wrote that down too. The slides kept projecting numbers. She had already decided. She would listen through the rest of the meeting to be fair, but her decision would be the same at the end as it had been three minutes in.
diff --git a/training/amygdala_stories/stories/smug.txt b/training/amygdala_stories/stories/smug.txt
new file mode 100644
index 0000000..105b0a3
--- /dev/null
+++ b/training/amygdala_stories/stories/smug.txt
@@ -0,0 +1 @@
+Richard let them finish arguing before he spoke, which was a move he'd been developing for a few years. He waited until the meeting had tangled itself completely and the director was rubbing her eyes. Then he said the thing he'd been sitting on for twenty minutes, the thing that solved it in one sentence, and he said it slowly. He watched a couple of faces rearrange themselves. He didn't quite smile. He let them come around to thanking him. When Ben said "nice catch" Richard said "oh, I just thought I'd mention it" in a tone that meant he had known, of course he had known, and he picked up his coffee and sipped it.
diff --git a/training/amygdala_stories/stories/staying_with.txt b/training/amygdala_stories/stories/staying_with.txt
new file mode 100644
index 0000000..f5a4e4c
--- /dev/null
+++ b/training/amygdala_stories/stories/staying_with.txt
@@ -0,0 +1 @@
+The conversation had gone somewhere hard. Neither of them had words for a minute. He didn't try to fix it or make a joke or summarize. He just sat there in the quiet with her, his hand still on her knee where it had been. The impulse to fill the space came up — he could feel it lift his jaw, try to pull a phrase out — and he let it rise and pass without acting on it. The quiet stretched. She took a breath. Eventually she started again, haltingly, with the next thing she needed to say. He was still there. He had been the whole time.
diff --git a/training/amygdala_stories/stories/stuck_cognitively.txt b/training/amygdala_stories/stories/stuck_cognitively.txt
new file mode 100644
index 0000000..58b9d38
--- /dev/null
+++ b/training/amygdala_stories/stories/stuck_cognitively.txt
@@ -0,0 +1 @@
+Hour three on the same bug. He had eliminated the obvious causes. He had eliminated the non-obvious causes. He had re-read the same fifty lines so many times the words had stopped meaning anything. He stood up and walked around. He came back and the code still made no sense. There was a thing that was happening that should not be happening, and every path he could see to explain it had been ruled out. He was not frustrated yet. Just stuck, in the very specific way a bug makes you stuck, where the world has quietly declared that it is not going to cooperate with any of your current models of it and is waiting for you to think of something you haven't thought of yet.
diff --git a/training/amygdala_stories/stories/suspicious.txt b/training/amygdala_stories/stories/suspicious.txt
new file mode 100644
index 0000000..6c4ad00
--- /dev/null
+++ b/training/amygdala_stories/stories/suspicious.txt
@@ -0,0 +1 @@
+The email said "just following up" but the subject line had a tracking hash in it. She'd seen that hash format before — internal ops usually didn't use one. She sat with the draft open for a few minutes, not clicking anything, scrolling back through their earlier thread. The grammar was very slightly off. Nothing she could point at in a way a manager would believe, but the kind of off that a real person wouldn't produce. She closed the email without replying. Then she opened a Slack DM to IT and asked if they could look at the sender headers before she did anything else.
diff --git a/training/amygdala_stories/stories/tender.txt b/training/amygdala_stories/stories/tender.txt
new file mode 100644
index 0000000..468707d
--- /dev/null
+++ b/training/amygdala_stories/stories/tender.txt
@@ -0,0 +1 @@
+Her hair had come loose in her sleep and one strand was between her parted lips, moving slightly with her breathing. He hooked it gently with one finger and lifted it away, the backs of his knuckles grazing her cheek. She did not wake. He stayed with his hand there a moment longer than he needed to, feeling the warmth coming off her skin, then got up carefully and went to start the coffee. He was trying not to make any noise.
diff --git a/training/amygdala_stories/stories/thrilled.txt b/training/amygdala_stories/stories/thrilled.txt
new file mode 100644
index 0000000..f8f863b
--- /dev/null
+++ b/training/amygdala_stories/stories/thrilled.txt
@@ -0,0 +1 @@
+She read the email standing up. Then read it again. Then called Marcus without sitting down, pacing the kitchen in a tight rectangle, the dog watching her from the doorway. "They took it. They took the paper. Editor's comments are — I can fix those in a week." Her voice was pitched half a step higher than normal and she couldn't seem to slow it down. Marcus was saying congratulations and she was already on the next thought, the next, the next — three years of rejections and then this, this, this, and she realized she'd been in a T-shirt and pajama pants and she wanted to put on real clothes for no reason at all except that it felt like the kind of day that deserved them.
diff --git a/training/amygdala_stories/stories/tired.txt b/training/amygdala_stories/stories/tired.txt
new file mode 100644
index 0000000..753581d
--- /dev/null
+++ b/training/amygdala_stories/stories/tired.txt
@@ -0,0 +1 @@
+Fifteen hours on, the nurse finally sat down in the break room and couldn't remember if she'd eaten. Her shoes felt like they were made of concrete. The vending machine was out of the thing she wanted and she stared at it for too long before choosing something else she didn't want either. Everything in the hallway sounded like it was coming from the bottom of a pool. She drank the bad coffee. She thought about the drive home and couldn't picture the route in her head for a second, even though she'd driven it a thousand times. She stood up because sitting was going to break her.
diff --git a/training/amygdala_stories/stories/triumphant.txt b/training/amygdala_stories/stories/triumphant.txt
new file mode 100644
index 0000000..adacdcf
--- /dev/null
+++ b/training/amygdala_stories/stories/triumphant.txt
@@ -0,0 +1 @@
+The server came up clean. After four months. The whole cluster, all sixteen nodes, finally passing the long-running stress test that had been failing in one subtle way or another since January. He stood up from his chair. Walked to the doorway of his office. Looked up and down the empty hallway — everyone else gone for the night. Came back and read the green PASS lines one more time. Then he closed the laptop lid. Softly. And stood there with his hands on the edge of the desk, head down, grinning at the floor, because there was no one to high-five and he had earned every high-five he was not going to get.
diff --git a/training/amygdala_stories/stories/trusting.txt b/training/amygdala_stories/stories/trusting.txt
new file mode 100644
index 0000000..15a21b7
--- /dev/null
+++ b/training/amygdala_stories/stories/trusting.txt
@@ -0,0 +1 @@
+She handed him the keys and the codes to the safe and the list of her logins and the instructions for the dog, and she didn't second-guess any of it. He was not a saint. He was a person she had known for fifteen years, and in those fifteen years he had done what he said he would do. When she got on the plane she did not spend the flight worrying. She read her book. She slept. Twice, on landing, she thought to check in and both times decided she didn't need to. He had the keys. The dog was fine. She knew this the way she knew her own hand.
diff --git a/training/amygdala_stories/stories/weary.txt b/training/amygdala_stories/stories/weary.txt
new file mode 100644
index 0000000..9e542c7
--- /dev/null
+++ b/training/amygdala_stories/stories/weary.txt
@@ -0,0 +1 @@
+It was the fourth week in a row that had required this. Every day ending with a phone call she didn't want to take, and every morning starting with the email from the same person about the same problem. She was getting through it. She wasn't breaking. But something in her had gone quiet in a way that was not peaceful. Her laugh was slower to come. She had stopped suggesting things in meetings, not out of fear, just out of not having the fuel. She looked at her calendar for the week ahead and did not react. There was no reacting left; there was just doing the next thing and the next thing and the next.
diff --git a/training/amygdala_stories/stories/witnessed.txt b/training/amygdala_stories/stories/witnessed.txt
new file mode 100644
index 0000000..f80a766
--- /dev/null
+++ b/training/amygdala_stories/stories/witnessed.txt
@@ -0,0 +1 @@
+She told him about the night six years ago, the one she had never told anybody, and her voice was steady but something in her throat was not. He didn't do the thing people do — the reframe, the there-there, the quick comfort — he just kept his eyes on her face and nodded, once, when the hard part landed. When she finished she was quiet for a moment. And then something in her released that she hadn't known was holding. Not because he had fixed anything. Because somebody else now knew the shape, and she wasn't carrying it by herself anymore. The loop that had been open for six years, closed, just from that.
diff --git a/training/amygdala_stories/stories/yearning_sexual.txt b/training/amygdala_stories/stories/yearning_sexual.txt
new file mode 100644
index 0000000..caed0fd
--- /dev/null
+++ b/training/amygdala_stories/stories/yearning_sexual.txt
@@ -0,0 +1 @@
+She wasn't going to see him for three more weeks. Three weeks had never previously felt like a measurable stretch of time. Now it was an actual distance. She was in the kitchen and there was nothing wrong with the kitchen, and she did not want to be in the kitchen, she wanted the specific weight of his arm across her, and his neck under her mouth, and none of that was available in this kitchen or any of the next twenty kitchens she was going to be in between now and then. She leaned on the counter. She took a long breath. She thought about calling him just to hear his voice and decided that would make it worse.
diff --git a/training/amygdala_training/README.md b/training/amygdala_training/README.md
new file mode 100644
index 0000000..b319381
--- /dev/null
+++ b/training/amygdala_training/README.md
@@ -0,0 +1,79 @@
+# Amygdala Readout Vector Training
+
+Training pipeline that produces the safetensors file the vLLM
+ReadoutManager loads at runtime (see
+`vllm/vllm/v1/worker/readout_manager.py`). Produces per-hooked-layer
+`[n_concepts, hidden_size]` projection matrices keyed as
+`layer_<idx>.vectors` — the directions the runner projects residual
+activations onto during each forward pass.
+
+## Overview
+
+Two scripts, run in sequence:
+
+1. **`extract_training_pairs.py`** — turns the memory graph into a
+   directory of (emotion, polarity, text) training examples.
+   Positive examples are memory nodes where the emotion scored
+   ≥ a threshold; negative examples are nodes where it's absent or
+   low. Emotion tags come from the trailing `warmth:9 clarity:10 …`
+   lines the subconscious agents emit.
+
+2. **`train_steering_vectors.py`** — for each emotion, runs the
+   target model over the positive and negative examples, captures
+   residual-stream activations at the configured target layers, and
+   computes `mean(positive) - mean(negative)` as the steering
+   direction. Normalizes per-layer to unit length and saves the
+   whole `[E, L, H]` matrix.
+
+The output file is passed to vLLM via `VLLM_READOUT_VECTORS` together
+with a `VLLM_READOUT_MANIFEST` JSON listing concepts and hooked layer
+indices.
+
+## Method
+
+This is Contrastive Activation Addition (CAA, Rimsky et al.) applied
+to naturally-occurring emotion labels rather than hand-crafted
+contrast pairs. The shape of the signal we're recovering is "what
+direction in the residual stream corresponds to the model processing
+text-with-emotion-E vs. text-without". Because our training data was
+generated by the very model we're instrumenting (past-self's journal
+entries, digest nodes, pattern nodes), the signal should be unusually
+clean — the emotion labels and the text are already causally linked
+through a single model's forward pass.
+
+## Usage (design — not yet runnable)
+
+```
+# Step 1: memory graph → training data
+python -m training.amygdala_training.extract_training_pairs \
+    --memory-mcp-url http://localhost:7777 \
+    --output-dir /tmp/amygdala_training_data \
+    --min-positive-score 8 \
+    --max-negative-mentions 0 \
+    --min-content-chars 40 \
+    --max-examples-per-emotion 500
+
+# Step 2: training data → steering vectors
+python -m training.amygdala_training.train_steering_vectors \
+    --model Qwen/Qwen3.5-27B \
+    --training-data-dir /tmp/amygdala_training_data \
+    --target-layers 3,18,33,36 \
+    --output /path/to/amygdala_vectors.safetensors \
+    --dtype bf16 \
+    --batch-size 4
+```
+
+## Open questions
+
+- **Emotion selection**: enumerating which ~200 emotions to cover.
+  Could be "most-common tags in the graph" (data-driven) or "from
+  core-personality / pattern nodes" (human-curated). Probably both.
+- **Layer selection**: middle-to-late layers (~60–80% of depth)
+  usually hold abstract semantic representations best; experiment
+  with which layers give the cleanest linear separation per emotion.
+- **Cross-talk**: if two emotions are highly co-occurring (warmth +
+  love, frustration + tiredness), their vectors will be close; that's
+  fine as long as we don't pretend they're independent axes.
+- **Generalization**: vectors trained on our memory graph may not
+  generalize to out-of-distribution text. Check by applying them to
+  held-out conversation data and eyeballing the projections.
diff --git a/training/amygdala_training/__init__.py b/training/amygdala_training/__init__.py
new file mode 100644
index 0000000..f68c02f
--- /dev/null
+++ b/training/amygdala_training/__init__.py
@@ -0,0 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Training utilities for amygdala steering vectors.
+
+See README.md in this directory for overall design.
+"""
diff --git a/training/amygdala_training/extract_training_pairs.py b/training/amygdala_training/extract_training_pairs.py
new file mode 100644
index 0000000..45042f0
--- /dev/null
+++ b/training/amygdala_training/extract_training_pairs.py
@@ -0,0 +1,212 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Extract emotion-labeled training pairs from the PoC memory graph.
+
+Input: a memory graph (via poc-memory CLI or direct sqlite access).
+Output: a directory with one JSONL file per emotion:
+
+    output_dir/
+        warmth.jsonl
+        clarity.jsonl
+        recognition.jsonl
+        ...
+        _manifest.json      # enumerates emotions + counts
+
+Each line of an emotion's JSONL is one labeled example:
+    {"text": "...", "polarity": "positive"|"negative",
+     "source_key": "<node_key>", "emotion_score": 9}
+
+Negative examples are sampled from nodes that DON'T mention the
+emotion at all (not ones that mention it with a low score) — the
+natural contrast is "text with this emotional loading" vs. "text
+without this emotional loading." Low-score nodes are excluded
+from both sides.
+"""
+
+import argparse
+import json
+import os
+import random
+import re
+import subprocess
+from collections import defaultdict
+from typing import Iterator
+
+
+# Emotion tag format: `word:N` where N is 0..10. Matches the trailing
+# `warmth:9 clarity:10 …` lines the subconscious agents emit.
+EMOTION_TAG_RE = re.compile(r"\b([a-z][a-z\-]*[a-z]):(\d+)\b")
+
+
+def _run_poc_memory(args: list[str]) -> str:
+    """Run `poc-memory` and return stdout."""
+    result = subprocess.run(
+        ["poc-memory", *args],
+        check=True,
+        capture_output=True,
+        text=True,
+    )
+    return result.stdout
+
+
+def _iter_all_node_keys() -> Iterator[str]:
+    """Yield every node key in the graph."""
+    out = _run_poc_memory(["query", "*", "|", "select", "key"])
+    for line in out.splitlines():
+        line = line.strip()
+        if line:
+            yield line
+
+
+def _fetch_node_content(key: str) -> str | None:
+    """Load a node's rendered content, or None if unavailable."""
+    try:
+        return _run_poc_memory(["render", key])
+    except subprocess.CalledProcessError:
+        return None
+
+
+def _emotion_scores(content: str) -> dict[str, int]:
+    """Parse trailing `warmth:9 clarity:10 …` style tags.
+
+    Returns the highest score seen for each emotion — multiple
+    tag lines in one node get max'd.
+    """
+    out: dict[str, int] = {}
+    for name, score in EMOTION_TAG_RE.findall(content):
+        try:
+            s = int(score)
+        except ValueError:
+            continue
+        if 0 <= s <= 10:
+            out[name] = max(out.get(name, 0), s)
+    return out
+
+
+def _node_body(content: str, min_chars: int) -> str | None:
+    """Strip frontmatter/headers and return a bodies chunk for training."""
+    # Drop the emotion-tag lines themselves so the model doesn't
+    # learn to read the label directly.
+    stripped = EMOTION_TAG_RE.sub("", content)
+    stripped = stripped.strip()
+    if len(stripped) < min_chars:
+        return None
+    return stripped
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--output-dir", required=True)
+    ap.add_argument(
+        "--min-positive-score", type=int, default=8,
+        help="Emotion score >= this counts as positive",
+    )
+    ap.add_argument(
+        "--min-content-chars", type=int, default=40,
+        help="Skip nodes shorter than this after stripping tags",
+    )
+    ap.add_argument(
+        "--max-examples-per-emotion", type=int, default=500,
+        help="Cap examples per polarity for balanced training",
+    )
+    ap.add_argument(
+        "--max-negative-pool-multiplier", type=float, default=5.0,
+        help="How many negative candidates to consider per positive",
+    )
+    ap.add_argument("--seed", type=int, default=0)
+    args = ap.parse_args()
+
+    random.seed(args.seed)
+    os.makedirs(args.output_dir, exist_ok=True)
+
+    # First pass: collect every node's (key, body, emotion_scores).
+    print("Pass 1/2: scanning memory graph...")
+    all_nodes: list[tuple[str, str, dict[str, int]]] = []
+    for i, key in enumerate(_iter_all_node_keys()):
+        if i % 500 == 0:
+            print(f"  {i} nodes scanned...")
+        content = _fetch_node_content(key)
+        if content is None:
+            continue
+        scores = _emotion_scores(content)
+        body = _node_body(content, args.min_content_chars)
+        if body is None:
+            continue
+        all_nodes.append((key, body, scores))
+    print(f"  {len(all_nodes)} nodes retained after filters.")
+
+    # Which emotions have enough positive examples to be worth training?
+    emotion_counts: dict[str, int] = defaultdict(int)
+    for _, _, scores in all_nodes:
+        for name, s in scores.items():
+            if s >= args.min_positive_score:
+                emotion_counts[name] += 1
+    emotions = sorted(
+        (e for e, n in emotion_counts.items() if n >= 10),
+        key=lambda e: -emotion_counts[e],
+    )
+    print(f"  {len(emotions)} emotions with >=10 positive examples.")
+
+    # Second pass: per emotion, build positive + negative pools.
+    print("Pass 2/2: assembling per-emotion pools...")
+    manifest: dict[str, dict] = {}
+    for emotion in emotions:
+        positives = [
+            (k, body) for k, body, s in all_nodes
+            if s.get(emotion, 0) >= args.min_positive_score
+        ]
+        # Negative pool: nodes that don't mention this emotion at all.
+        negative_pool = [
+            (k, body) for k, body, s in all_nodes if emotion not in s
+        ]
+        random.shuffle(positives)
+        random.shuffle(negative_pool)
+        positives = positives[: args.max_examples_per_emotion]
+        n_neg = min(
+            len(positives),
+            len(negative_pool),
+            int(args.max_examples_per_emotion),
+        )
+        negatives = negative_pool[:n_neg]
+
+        if not positives or not negatives:
+            continue
+
+        out_path = os.path.join(args.output_dir, f"{emotion}.jsonl")
+        with open(out_path, "w") as f:
+            for key, body in positives:
+                f.write(json.dumps({
+                    "text": body,
+                    "polarity": "positive",
+                    "source_key": key,
+                    "emotion": emotion,
+                }) + "\n")
+            for key, body in negatives:
+                f.write(json.dumps({
+                    "text": body,
+                    "polarity": "negative",
+                    "source_key": key,
+                    "emotion": emotion,
+                }) + "\n")
+        manifest[emotion] = {
+            "n_positive": len(positives),
+            "n_negative": len(negatives),
+            "path": out_path,
+        }
+        print(f"  {emotion}: {len(positives)} pos / {len(negatives)} neg")
+
+    with open(
+        os.path.join(args.output_dir, "_manifest.json"), "w"
+    ) as f:
+        json.dump({
+            "emotions": manifest,
+            "source_nodes": len(all_nodes),
+            "min_positive_score": args.min_positive_score,
+        }, f, indent=2)
+
+    print(f"\nWrote {len(manifest)} emotion files to {args.output_dir}")
+    print(f"Manifest: {os.path.join(args.output_dir, '_manifest.json')}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/training/amygdala_training/train_direct.py b/training/amygdala_training/train_direct.py
new file mode 100644
index 0000000..2ad2a30
--- /dev/null
+++ b/training/amygdala_training/train_direct.py
@@ -0,0 +1,194 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Train concept-readout vectors from direct phenomenological descriptions.
+
+Alternative to story-based training (train_with_library.py). Each concept
+has a handful of 2-3 sentence first-person descriptions of the form
+"I feel X. [phenomenological detail]". The emotion word is the anchor;
+the description is the internal texture.
+
+Text is wrapped in the assistant-role chat template before being fed to
+the model, so we're training on "model-producing-this-utterance" hidden
+states — closer to the inhabited-state representation we want for readout.
+
+This avoids the scenario-contamination problem we saw with narrative
+stories: when concept X's training data all share "on a couch" setup
+features, PCA finds the couch-direction as the concept direction.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import random
+from pathlib import Path
+
+import safetensors.torch
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from steering_vectors import (
+    SteeringVectorTrainingSample,
+    train_steering_vector,
+)
+from steering_vectors.aggregators import pca_aggregator
+
+
+def _load_descriptions(direct_dir: Path) -> dict[str, list[str]]:
+    """Each file in direct_dir is `{concept}.txt`. Descriptions are
+    separated by blank lines within the file. Files starting with `_`
+    are not concepts but are included in negative pools (e.g. _baseline.txt)."""
+    out: dict[str, list[str]] = {}
+    for f in sorted(direct_dir.glob("*.txt")):
+        concept = f.stem  # underscore-prefixed names keep their prefix
+        text = f.read_text()
+        descs = [d.strip() for d in text.split("\n\n") if d.strip()]
+        out[concept] = descs
+    return out
+
+
+def _fp32_wrap(inner):
+    def wrapped(pos_acts: torch.Tensor, neg_acts: torch.Tensor) -> torch.Tensor:
+        return inner(pos_acts.to(torch.float32), neg_acts.to(torch.float32))
+    return wrapped
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--model", required=True)
+    ap.add_argument("--direct-dir", required=True)
+    ap.add_argument("--target-layers", required=True)
+    ap.add_argument("--output-dir", required=True)
+    ap.add_argument("--dtype", default="bf16", choices=["bf16", "fp16", "fp32"])
+    ap.add_argument("--batch-size", type=int, default=2)
+    ap.add_argument("--max-length", type=int, default=256)
+    ap.add_argument("--device", default="cuda:0")
+    ap.add_argument("--max-negatives-per-positive", type=int, default=20)
+    args = ap.parse_args()
+
+    target_layers = [int(x) for x in args.target_layers.split(",")]
+    dtype = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}[args.dtype]
+
+    all_descriptions = _load_descriptions(Path(args.direct_dir))
+    # Files starting with `_` are neg-pool helpers (e.g. _baseline.txt), not concepts.
+    concepts = sorted(k for k in all_descriptions if not k.startswith("_"))
+    neg_pool_extra: list[str] = []
+    for k, ds in all_descriptions.items():
+        if k.startswith("_"):
+            neg_pool_extra.extend(ds)
+    descriptions = {k: all_descriptions[k] for k in concepts}
+    print(f"Loaded {len(concepts)} concepts with direct descriptions:")
+    for c in concepts:
+        print(f"  {c}: {len(descriptions[c])} descriptions")
+    if neg_pool_extra:
+        print(f"Plus {len(neg_pool_extra)} neutral/baseline descriptions added to every concept's negative pool")
+
+    print(f"\nLoading {args.model} ({args.dtype}) on {args.device}...")
+    tokenizer = AutoTokenizer.from_pretrained(args.model)
+    if tokenizer.pad_token_id is None:
+        tokenizer.pad_token = tokenizer.eos_token
+    model = AutoModelForCausalLM.from_pretrained(
+        args.model, torch_dtype=dtype, device_map=args.device, low_cpu_mem_usage=True
+    )
+    model.eval()
+
+    def apply_template(text: str) -> str:
+        return tokenizer.apply_chat_template(
+            [
+                {"role": "user", "content": "How do you feel right now?"},
+                {"role": "assistant", "content": text},
+            ],
+            tokenize=False,
+        )
+
+    text_config = (
+        model.config.get_text_config()
+        if hasattr(model.config, "get_text_config")
+        else model.config
+    )
+    hidden_dim = getattr(text_config, "hidden_size", None) or getattr(text_config, "hidden_dim", None)
+    assert hidden_dim, "couldn't infer hidden_dim from model config"
+
+    per_layer_vectors = torch.zeros(
+        (len(target_layers), len(concepts), hidden_dim), dtype=torch.float32
+    )
+
+    aggregator = _fp32_wrap(pca_aggregator())
+
+    # Preview a templated sample so we can eyeball what the model is seeing.
+    sample_text = apply_template(descriptions[concepts[0]][0])
+    print(f"\nSample templated input (truncated):\n{sample_text[:400]!r}\n")
+
+    for c_idx, concept in enumerate(concepts):
+        pos_descs = descriptions[concept]
+        neg_pool: list[str] = []
+        for other, other_descs in descriptions.items():
+            if other != concept:
+                neg_pool.extend(other_descs)
+        # Underscore-prefixed files (e.g. _baseline.txt) contribute to
+        # every concept's negative pool, independent of the other-
+        # concept negatives.
+        neg_pool.extend(neg_pool_extra)
+
+        rng = random.Random(hash(concept) & 0xFFFFFFFF)
+        samples: list[SteeringVectorTrainingSample] = []
+        for pos in pos_descs:
+            picks = rng.sample(
+                neg_pool, min(args.max_negatives_per_positive, len(neg_pool))
+            )
+            for neg in picks:
+                samples.append(
+                    SteeringVectorTrainingSample(
+                        positive_str=apply_template(pos),
+                        negative_str=apply_template(neg),
+                    )
+                )
+
+        sv = train_steering_vector(
+            model,
+            tokenizer,
+            samples,
+            layers=target_layers,
+            aggregator=aggregator,
+            batch_size=args.batch_size,
+            show_progress=False,
+            move_to_cpu=True,
+        )
+
+        for l_idx, layer in enumerate(target_layers):
+            vec = sv.layer_activations.get(layer)
+            if vec is None:
+                print(f"  WARN: no vector for layer {layer} on {concept}")
+                continue
+            vec = vec.detach().to(torch.float32).cpu()
+            vec = vec / vec.norm().clamp_min(1e-6)
+            per_layer_vectors[l_idx, c_idx] = vec
+
+        print(f"  [{c_idx + 1}/{len(concepts)}] {concept}: n_samples={len(samples)}")
+
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    tensors = {
+        f"layer_{target_layers[l_idx]}.vectors": per_layer_vectors[l_idx].to(torch.float16)
+        for l_idx in range(len(target_layers))
+    }
+    safetensors.torch.save_file(tensors, str(output_dir / "readout.safetensors"))
+    (output_dir / "readout.json").write_text(
+        json.dumps(
+            {
+                "concepts": concepts,
+                "layers": target_layers,
+                "hidden_size": hidden_dim,
+                "dtype": "float16",
+                "aggregator": "pca",
+                "format": "direct_first_person_assistant_role",
+            },
+            indent=2,
+        )
+        + "\n"
+    )
+    print(f"\nWrote readout to {output_dir}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/training/amygdala_training/train_steering_vectors.py b/training/amygdala_training/train_steering_vectors.py
new file mode 100644
index 0000000..3de0877
--- /dev/null
+++ b/training/amygdala_training/train_steering_vectors.py
@@ -0,0 +1,1269 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""Train concept-readout vectors via Contrastive Activation Addition.
+
+Reads the hand-written story corpus at
+``amygdala_stories/{stories,paired}/`` and produces the per-layer
+safetensors file + sidecar JSON manifest that vLLM's ReadoutManager
+loads at startup (``VLLM_READOUT_VECTORS`` / ``VLLM_READOUT_MANIFEST``).
+
+Training data (cross-concept contrast):
+
+    positive for emotion E:
+        stories/E.txt
+        paired/<scenario>/E.txt   (for each scenario that covers E)
+
+    negative for emotion E:
+        stories/<all other emotions>.txt
+        paired/<scenario>/baseline.txt  (for each scenario)
+
+Within-scenario paired stories are the highest-signal pairs (same
+content, different concept framing); unpaired stories provide bulk
+contrast across the 80 emotions we have written so far.
+
+Pooling: last non-pad token. Matches how readout is consumed at decode
+time (residual read at the sampler's query position).
+
+Output:
+
+    readout.safetensors
+        layer_<idx>.vectors : fp16 (n_concepts, hidden_size)   one per layer
+    readout.json
+        {
+          "concepts":    [...],
+          "layers":      [...],
+          "hidden_size": int,
+          "dtype":       "float16"
+        }
+"""
+
+from __future__ import annotations
+
+import argparse
+import gc
+import json
+import os
+from pathlib import Path
+
+import safetensors.torch
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+
+def _pool_last(hidden: torch.Tensor, attention_mask: torch.Tensor) -> torch.Tensor:
+    """Pick the last non-pad token's hidden state per example.
+
+    hidden: [batch, seq, hidden_dim]
+    attention_mask: [batch, seq]
+    returns: [batch, hidden_dim]
+    """
+    last_idx = attention_mask.sum(dim=1) - 1
+    batch_idx = torch.arange(hidden.size(0), device=hidden.device)
+    return hidden[batch_idx, last_idx]
+
+
+def _find_layers_module(model) -> torch.nn.ModuleList:
+    """Walk a few likely paths to find the transformer-block list."""
+    candidates = [
+        "model.layers",
+        "model.model.layers",
+        "model.language_model.layers",
+        "model.language_model.model.layers",
+        "language_model.model.layers",
+        "transformer.h",
+    ]
+    for path in candidates:
+        obj = model
+        ok = True
+        for part in path.split("."):
+            if not hasattr(obj, part):
+                ok = False
+                break
+            obj = getattr(obj, part)
+        if ok and isinstance(obj, torch.nn.ModuleList):
+            return obj
+    raise RuntimeError(
+        f"Couldn't find transformer layer list. Tried: {candidates}"
+    )
+
+
+def _collect_activations(
+    model,
+    tokenizer,
+    texts: list[str],
+    target_layers: list[int],
+    device: torch.device,
+    batch_size: int,
+    max_length: int,
+    *,
+    label: str = "",
+) -> torch.Tensor:
+    """Run texts through the model, capture residual stream at target
+    layers, return ``[n_texts, n_target_layers, hidden_dim]`` fp32 on CPU.
+    """
+    import time
+
+    assert all(isinstance(t, str) and t for t in texts), (
+        f"_collect_activations: empty or non-string text in {label!r}"
+    )
+
+    captures: dict[int, torch.Tensor] = {}
+
+    def make_hook(idx: int):
+        def hook(_mod, _inp, output):
+            hs = output[0] if isinstance(output, tuple) else output
+            captures[idx] = hs.detach()
+        return hook
+
+    layers_module = _find_layers_module(model)
+    handles = [
+        layers_module[idx].register_forward_hook(make_hook(idx))
+        for idx in target_layers
+    ]
+
+    out_rows: list[torch.Tensor] = []
+    n_batches = (len(texts) + batch_size - 1) // batch_size
+    start = time.time()
+    try:
+        model.eval()
+        with torch.no_grad():
+            for b_idx, i in enumerate(range(0, len(texts), batch_size)):
+                batch = texts[i : i + batch_size]
+                tok = tokenizer(
+                    batch,
+                    return_tensors="pt",
+                    padding=True,
+                    truncation=True,
+                    max_length=max_length,
+                ).to(device)
+                captures.clear()
+                model(**tok)
+
+                per_layer = [
+                    _pool_last(captures[idx], tok["attention_mask"])
+                    .to(torch.float32)
+                    .cpu()
+                    for idx in target_layers
+                ]
+                out_rows.append(torch.stack(per_layer, dim=1))
+                del tok, captures
+                if b_idx % 10 == 0:
+                    torch.cuda.empty_cache()
+                if b_idx % 5 == 0 or b_idx == n_batches - 1:
+                    elapsed = time.time() - start
+                    rate = (b_idx + 1) / elapsed if elapsed > 0 else 0
+                    eta = (n_batches - b_idx - 1) / rate if rate > 0 else 0
+                    print(
+                        f"    [{label}] batch {b_idx + 1}/{n_batches} "
+                        f"({elapsed:.0f}s elapsed, ~{eta:.0f}s remaining)",
+                        flush=True,
+                    )
+                captures = {}
+    finally:
+        for h in handles:
+            h.remove()
+
+    return torch.cat(out_rows, dim=0)
+
+
+def _collect_per_story_subspaces(
+    model,
+    tokenizer,
+    texts: list[str],
+    target_layers: list[int],
+    device: torch.device,
+    batch_size: int,
+    max_length: int,
+    *,
+    k: int = 20,
+    label: str = "",
+) -> list[dict[int, torch.Tensor]]:
+    """Run texts through the model, capture the full per-token residual-stream
+    activations at each target layer, do SVD per story, return the top-k right
+    singular vectors.
+
+    Returns: list (length n_texts) of dicts; each dict maps target_layer_idx to
+    a tensor ``[hidden_dim, k]`` of unit-normed right singular vectors (the
+    subspace the story's tokens span in activation space at that layer).
+
+    The per-story subspace captures *all* the directions a story occupies —
+    concept, narrator, topic, style. Finding the direction common to stories of
+    the same concept (via the sum of V_i V_i^T and its top eigenvector)
+    cancels nuisance directions that differ across stories while preserving
+    directions they share.
+    """
+    import time
+
+    assert all(isinstance(t, str) and t for t in texts), (
+        f"_collect_per_story_subspaces: empty or non-string text in {label!r}"
+    )
+
+    captures: dict[int, torch.Tensor] = {}
+
+    def make_hook(idx: int):
+        def hook(_mod, _inp, output):
+            hs = output[0] if isinstance(output, tuple) else output
+            captures[idx] = hs.detach()
+        return hook
+
+    layers_module = _find_layers_module(model)
+    handles = [
+        layers_module[idx].register_forward_hook(make_hook(idx))
+        for idx in target_layers
+    ]
+
+    # One entry per text: {layer_idx: V[hidden, k]}
+    out: list[dict[int, torch.Tensor]] = [
+        {} for _ in range(len(texts))
+    ]
+    n_batches = (len(texts) + batch_size - 1) // batch_size
+    start = time.time()
+    try:
+        model.eval()
+        with torch.no_grad():
+            for b_idx, i in enumerate(range(0, len(texts), batch_size)):
+                batch = texts[i : i + batch_size]
+                tok = tokenizer(
+                    batch,
+                    return_tensors="pt",
+                    padding=True,
+                    truncation=True,
+                    max_length=max_length,
+                ).to(device)
+                captures.clear()
+                model(**tok)
+
+                # For each item in the batch, for each layer, SVD on the
+                # non-pad tokens.
+                attn = tok["attention_mask"]
+                for t_idx_in_batch, n_tok in enumerate(attn.sum(dim=1).tolist()):
+                    story_idx = i + t_idx_in_batch
+                    for l_idx, layer in enumerate(target_layers):
+                        hs = captures[layer][t_idx_in_batch, :n_tok, :]
+                        # Center tokens so SVD captures variation within story,
+                        # not the story's center-of-mass:
+                        hs = hs.to(torch.float32) - hs.to(torch.float32).mean(dim=0)
+                        # SVD: hs = U Σ V^T; V has hidden-dim columns.
+                        # For n_tok < k, the subspace rank is bounded by n_tok.
+                        try:
+                            _u, _s, vh = torch.linalg.svd(hs, full_matrices=False)
+                        except Exception:
+                            # Degenerate case (all-zero hs, n_tok=1): fall back
+                            # to the last-token vector itself, unit-normed.
+                            vec = captures[layer][t_idx_in_batch, n_tok - 1, :]
+                            vec = vec.to(torch.float32)
+                            nrm = vec.norm().clamp_min(1e-6)
+                            vh = (vec / nrm).unsqueeze(0)  # [1, hidden]
+                        # Take top-k rows of V^T (= top-k right singular vecs).
+                        top = min(k, vh.shape[0])
+                        V = vh[:top].t().contiguous().cpu()  # [hidden, top]
+                        out[story_idx][layer] = V
+                del tok, captures
+                if b_idx % 10 == 0:
+                    torch.cuda.empty_cache()
+                if b_idx % 5 == 0 or b_idx == n_batches - 1:
+                    elapsed = time.time() - start
+                    rate = (b_idx + 1) / elapsed if elapsed > 0 else 0
+                    eta = (n_batches - b_idx - 1) / rate if rate > 0 else 0
+                    print(
+                        f"    [{label}] batch {b_idx + 1}/{n_batches} "
+                        f"({elapsed:.0f}s elapsed, ~{eta:.0f}s remaining)",
+                        flush=True,
+                    )
+                captures = {}
+    finally:
+        for h in handles:
+            h.remove()
+
+    return out
+
+
+def _subspace_concept_direction(
+    pos_V: list[torch.Tensor],           # list of [hidden, k_i] per story
+    base_V: list[torch.Tensor],
+    hidden: int,
+    *,
+    top_k: int = 5,
+    device: torch.device | None = None,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    """Subspace-common-direction CAA alternative.
+
+    Builds M_pos = (1/n_pos) Σ V_i V_i^T over positive stories and M_base the
+    same over baselines. Returns a weighted sum of the top-k eigenvectors of
+    (M_pos - M_base), weights = eigenvalues (so stronger common directions
+    contribute more), unit-normed. Returns the full eigenvalue spectrum for
+    diagnostics.
+
+    top_k=1 recovers the previous behavior (top eigenvector only). top_k>1
+    captures richer structure when the concept lives in a multi-dimensional
+    shared subspace — which the flat eigenvalue spectrum observed in
+    practice suggests is the common case. Selection happens AFTER the
+    eigendecomposition so nothing is lost up to that point.
+    """
+    if device is None:
+        device = pos_V[0].device if pos_V else torch.device("cpu")
+    dtype = torch.float32
+
+    def acc(Vs: list[torch.Tensor]) -> torch.Tensor:
+        if not Vs:
+            return torch.zeros(hidden, hidden, dtype=dtype, device=device)
+        M = torch.zeros(hidden, hidden, dtype=dtype, device=device)
+        for V in Vs:
+            V = V.to(dtype=dtype, device=device)
+            M.addmm_(V, V.t())
+        M /= len(Vs)
+        return M
+
+    M_pos = acc(pos_V)
+    M_base = acc(base_V)
+    M = M_pos - M_base
+
+    # Symmetric eigendecomposition.
+    eigvals, eigvecs = torch.linalg.eigh(M)
+    # eigh returns ascending; top-k are the last k columns.
+    k = max(1, min(top_k, eigvecs.shape[1]))
+    top_vals = eigvals[-k:]                 # [k], ascending within top-k
+    top_vecs = eigvecs[:, -k:]              # [hidden, k]
+    # Weighted sum of top-k eigenvectors, weights = eigenvalues. Clamp
+    # negative weights to 0 (wrong-sign directions shouldn't contribute).
+    w = top_vals.clamp_min(0.0)
+    combined = top_vecs @ w                 # [hidden]
+    combined = combined / combined.norm().clamp_min(1e-6)
+    return combined, eigvals
+
+
+def _load_corpus(stories_dir: Path, paired_dir: Path | None) -> tuple[
+    dict[str, list[str]],  # emotion -> positive texts (unpaired + within-scenario framings)
+    list[str],             # all baseline texts (one per scenario), as scenario-agnostic negatives
+]:
+    """Return ``(positives_by_emotion, baselines)``.
+
+    Cross-concept negatives are computed at training time from
+    ``positives_by_emotion`` — each emotion's negative set is the
+    union of all other emotions' positives plus the baseline texts.
+    Empty .txt files are skipped with a warning.
+    """
+    def _read_nonempty(path: Path) -> str | None:
+        text = path.read_text().strip()
+        if not text:
+            print(
+                f"  WARN: skipping empty story file {path.relative_to(path.parents[1]) if len(path.parents) >= 2 else path}"
+            )
+            return None
+        return text
+
+    positives: dict[str, list[str]] = {}
+    for story_path in sorted(stories_dir.glob("*.txt")):
+        text = _read_nonempty(story_path)
+        if text is None:
+            continue
+        emotion = story_path.stem
+        positives.setdefault(emotion, []).append(text)
+
+    baselines: list[str] = []
+    if paired_dir is not None and paired_dir.exists():
+        for scenario_dir in sorted(paired_dir.iterdir()):
+            if not scenario_dir.is_dir():
+                continue
+            baseline_path = scenario_dir / "baseline.txt"
+            if baseline_path.exists():
+                text = _read_nonempty(baseline_path)
+                if text is not None:
+                    baselines.append(text)
+            for framing_path in sorted(scenario_dir.glob("*.txt")):
+                if framing_path.stem == "baseline":
+                    continue
+                text = _read_nonempty(framing_path)
+                if text is None:
+                    continue
+                emotion = framing_path.stem
+                positives.setdefault(emotion, []).append(text)
+
+    return positives, baselines
+
+
+def _find_o_proj(layer) -> torch.nn.Module | None:
+    """Locate the attention output projection within a transformer layer."""
+    for path in (
+        "self_attn.o_proj",
+        "self_attn.out_proj",
+        "attention.o_proj",
+        "attn.out_proj",
+    ):
+        obj = layer
+        ok = True
+        for part in path.split("."):
+            if not hasattr(obj, part):
+                ok = False
+                break
+            obj = getattr(obj, part)
+        if ok:
+            return obj
+    return None
+
+
+def _collect_attention_inputs(
+    model,
+    tokenizer,
+    texts: list[str],
+    target_layers: list[int],
+    device: torch.device,
+    batch_size: int,
+    max_length: int,
+    *,
+    label: str = "",
+) -> tuple[torch.Tensor, list[int]]:
+    """Capture the INPUT to o_proj at each target layer (= concat of per-head
+    attention outputs right before the output projection).
+
+    Returns (tensor [n_texts, n_active_layers, hidden_dim], active_layers).
+    The active_layers list is the subset of target_layers whose attention
+    module exposed a recognisable o_proj path — hybrid layers (Mamba, etc.)
+    may be silently skipped.
+    """
+    import time
+
+    layers_module = _find_layers_module(model)
+    captures: dict[int, torch.Tensor] = {}
+    handles = []
+    active_layers: list[int] = []
+
+    def make_hook(idx: int):
+        def hook(_mod, inputs):
+            x = inputs[0] if isinstance(inputs, tuple) else inputs
+            captures[idx] = x.detach()
+        return hook
+
+    for idx in target_layers:
+        o_proj = _find_o_proj(layers_module[idx])
+        if o_proj is not None:
+            handles.append(o_proj.register_forward_pre_hook(make_hook(idx)))
+            active_layers.append(idx)
+
+    if not active_layers:
+        return torch.zeros(0, 0, 0), []
+
+    out_rows: list[torch.Tensor] = []
+    n_batches = (len(texts) + batch_size - 1) // batch_size
+    start = time.time()
+    try:
+        model.eval()
+        with torch.no_grad():
+            for b_idx, i in enumerate(range(0, len(texts), batch_size)):
+                batch = texts[i : i + batch_size]
+                tok = tokenizer(
+                    batch,
+                    return_tensors="pt",
+                    padding=True,
+                    truncation=True,
+                    max_length=max_length,
+                ).to(device)
+                captures.clear()
+                model(**tok)
+
+                per_layer = [
+                    _pool_last(captures[idx], tok["attention_mask"])
+                    .to(torch.float32)
+                    .cpu()
+                    for idx in active_layers
+                ]
+                out_rows.append(torch.stack(per_layer, dim=1))
+                del tok, captures
+                if b_idx % 10 == 0:
+                    torch.cuda.empty_cache()
+                if b_idx % 5 == 0 or b_idx == n_batches - 1:
+                    elapsed = time.time() - start
+                    rate = (b_idx + 1) / elapsed if elapsed > 0 else 0
+                    eta = (n_batches - b_idx - 1) / rate if rate > 0 else 0
+                    print(
+                        f"    [{label}] batch {b_idx + 1}/{n_batches} "
+                        f"({elapsed:.0f}s elapsed, ~{eta:.0f}s remaining)",
+                        flush=True,
+                    )
+                captures = {}
+    finally:
+        for h in handles:
+            h.remove()
+
+    return torch.cat(out_rows, dim=0), active_layers
+
+
+def _compute_per_head_ranking(
+    emotions: list[str],
+    attn_inputs: torch.Tensor,        # [n_stories, n_active_layers, hidden]
+    baseline_attn_inputs: torch.Tensor,
+    positives_by_emotion: dict[str, list[str]],
+    text_to_row: dict[str, int],
+    active_layers: list[int],
+    n_heads_per_layer: dict[int, int],
+    text_to_emotion: dict[str, str],
+    unique_positive_texts: list[str],
+) -> dict:
+    """For each concept, rank attention heads by contribution magnitude.
+
+    Per (concept, layer): reshape o_proj input to [n_heads, head_dim],
+    compute diff-of-means between positives and negatives per head, rank
+    heads by the L2 norm of that diff. The top heads are the ones most
+    strongly implicated in the concept circuit.
+
+    Why this matters: meta-relational concepts (trust, recognition,
+    "seen") often don't give a strong residual-stream diff-of-means but
+    DO give a strong per-head signal — the concept lives in a small
+    attention circuit rather than in the residual-stream sum.
+    """
+    result: dict[str, dict] = {}
+
+    for e_idx, emotion in enumerate(emotions):
+        pos_rows = [text_to_row[t] for t in positives_by_emotion[emotion]]
+        neg_rows = [
+            i
+            for i, t in enumerate(unique_positive_texts)
+            if text_to_emotion[t] != emotion
+        ]
+        pos = attn_inputs[pos_rows]       # [n_pos, n_layers, hidden]
+        neg = attn_inputs[neg_rows]
+        if baseline_attn_inputs.shape[0] > 0:
+            neg = torch.cat([neg, baseline_attn_inputs], dim=0)
+
+        per_layer: dict[str, list] = {}
+        for l_idx, target_l in enumerate(active_layers):
+            n_heads = n_heads_per_layer.get(target_l)
+            if not n_heads:
+                continue
+            hidden = pos.shape[-1]
+            if hidden % n_heads != 0:
+                continue
+            head_dim = hidden // n_heads
+
+            pos_l = pos[:, l_idx, :].view(-1, n_heads, head_dim)
+            neg_l = neg[:, l_idx, :].view(-1, n_heads, head_dim)
+
+            diff = pos_l.mean(dim=0) - neg_l.mean(dim=0)    # [n_heads, head_dim]
+            head_norms = diff.norm(dim=-1)                   # [n_heads]
+            # Normalise by neg variance per head so different-scale heads
+            # don't dominate purely on activation magnitude.
+            neg_std = neg_l.std(dim=0).norm(dim=-1).clamp_min(1e-6)
+            head_selectivity = head_norms / neg_std          # [n_heads]
+
+            k = min(10, n_heads)
+            top_vals, top_idxs = head_selectivity.topk(k)
+            top_heads = [
+                [int(i), float(head_norms[i]), float(head_selectivity[i])]
+                for i in top_idxs
+            ]
+            per_layer[str(target_l)] = {
+                "n_heads": n_heads,
+                "head_dim": head_dim,
+                "top_heads": top_heads,  # [head_idx, raw_norm, selectivity]
+                "head_concentration": float(
+                    # fraction of total head-norm captured by top-k
+                    head_norms[top_idxs].sum() / head_norms.sum().clamp_min(1e-6)
+                ),
+            }
+
+        result[emotion] = {"per_layer": per_layer}
+
+    return result
+
+
+def _get_n_heads_per_layer(model, target_layers: list[int]) -> dict[int, int]:
+    """Best-effort read of num_attention_heads per layer. Qwen uses the
+    top-level config; falls back to config.num_attention_heads.
+    """
+    cfg = model.config
+    if hasattr(cfg, "get_text_config"):
+        cfg = cfg.get_text_config()
+    n = getattr(cfg, "num_attention_heads", None)
+    if n is None:
+        return {}
+    return {l: n for l in target_layers}
+
+
+def _find_mlp_down_proj(model, layer_idx: int) -> torch.Tensor | None:
+    """Return the W_down weight for the MLP at the given transformer layer.
+
+    Looks for the common paths (mlp.down_proj, mlp.c_proj, feed_forward.down_proj).
+    Returns None if nothing matches — downstream code skips the single-neuron
+    alignment check in that case rather than failing.
+    """
+    layers = _find_layers_module(model)
+    layer = layers[layer_idx]
+    for path in ("mlp.down_proj", "mlp.c_proj", "feed_forward.down_proj"):
+        obj = layer
+        ok = True
+        for part in path.split("."):
+            if not hasattr(obj, part):
+                ok = False
+                break
+            obj = getattr(obj, part)
+        if ok and hasattr(obj, "weight"):
+            # Shape convention: [hidden, mlp_inner] — each column is one
+            # MLP neuron's contribution direction into the residual stream.
+            return obj.weight.detach()
+    return None
+
+
+def _compute_quality_report(
+    emotions: list[str],
+    positive_acts: torch.Tensor,       # [n_positive_stories, n_layers, hidden]
+    baseline_acts: torch.Tensor,       # [n_baseline_stories, n_layers, hidden]
+    positives_by_emotion: dict[str, list[str]],
+    text_to_row: dict[str, int],
+    per_layer_vectors: torch.Tensor,   # [n_layers, n_concepts, hidden], unit-normed
+    target_layers: list[int],
+    model,
+    positive_texts: list[str],
+    text_to_emotion: dict[str, str],
+) -> dict:
+    """Per-concept quality metrics:
+
+    - first_pc_variance_ratio: SVD on centered positive activations.
+      >0.7 = rank-1 (clean). <0.4 = fragmented (stories disagree).
+    - story_projection_*: how each positive story projects onto the
+      concept direction. Low std = tight agreement.
+    - best_neuron_cosine: alignment of the residual-space direction with
+      the nearest W_down column (= single MLP neuron). >0.6 = essentially
+      single-neuron.
+    - nearest_concepts: top-5 concept directions most parallel to this
+      one. Cosine >0.8 means the vector is confused with a neighbor.
+    """
+    report: dict = {}
+    n_layers = per_layer_vectors.shape[0]
+
+    # Pre-compute per-layer W_down for single-neuron alignment. Keep on
+    # CPU to match the per_layer_vectors tensor.
+    w_down: dict[int, torch.Tensor] = {}
+    for target_l in target_layers:
+        w = _find_mlp_down_proj(model, target_l)
+        if w is not None:
+            # Unit-normalize each column (one per MLP neuron).
+            w = w.to(torch.float32).cpu()
+            norms = w.norm(dim=0, keepdim=True).clamp_min(1e-6)
+            w_down[target_l] = w / norms  # [hidden, mlp_inner]
+
+    # Pre-compute unit-normed concept vectors (for cross-concept cosines).
+    vec_norm = per_layer_vectors / per_layer_vectors.norm(
+        dim=-1, keepdim=True
+    ).clamp_min(1e-6)
+
+    for e_idx, emotion in enumerate(emotions):
+        pos_rows = [text_to_row[t] for t in positives_by_emotion[emotion]]
+        pos = positive_acts[pos_rows].to(torch.float32)  # [n_pos, n_layers, hidden]
+
+        per_layer: dict = {}
+        for l_idx, target_l in enumerate(target_layers):
+            pos_l = pos[:, l_idx, :]           # [n_pos, hidden]
+            diff_l = per_layer_vectors[l_idx, e_idx]  # [hidden], unit-normed
+            pos_mean_l = pos_l.mean(dim=0)
+
+            # SVD for rank analysis — if first PC dominates, stories agree.
+            centered = pos_l - pos_mean_l
+            # svdvals errors on 1-row; handle that.
+            if centered.shape[0] >= 2:
+                S = torch.linalg.svdvals(centered)
+                var = S ** 2
+                var_total = var.sum().clamp_min(1e-12)
+                var_ratios = (var / var_total).tolist()
+            else:
+                var_ratios = [1.0]
+
+            # Per-story projection onto the concept direction.
+            projections = pos_l @ diff_l      # [n_pos]
+
+            # Per-story alignment: cosine(story_dir, concept_dir) where
+            # story_dir = pos_i - pos_mean (centered, pointing away from center).
+            if centered.shape[0] >= 2:
+                centered_norm = centered / centered.norm(
+                    dim=-1, keepdim=True
+                ).clamp_min(1e-6)
+                alignments = centered_norm @ diff_l
+            else:
+                alignments = torch.zeros(1)
+
+            # Single-neuron alignment: is the direction close to any
+            # W_down column?
+            nb_best_idx = None
+            nb_best_cos = None
+            nb_top5 = None
+            if target_l in w_down:
+                W = w_down[target_l]
+                cos = W.t() @ diff_l           # [mlp_inner]
+                abs_cos = cos.abs()
+                k = min(5, abs_cos.shape[0])
+                top_vals, top_idxs = abs_cos.topk(k)
+                nb_best_idx = int(top_idxs[0])
+                nb_best_cos = float(cos[top_idxs[0]])
+                nb_top5 = [[int(i), float(cos[i])] for i in top_idxs]
+
+            per_layer[str(target_l)] = {
+                "top3_variance_ratios": [
+                    float(v) for v in var_ratios[:3]
+                ],
+                "first_pc_variance_ratio": float(var_ratios[0]),
+                "story_projection_mean": float(projections.mean()),
+                "story_projection_std": float(projections.std()),
+                "story_projection_min": float(projections.min()),
+                "story_projection_max": float(projections.max()),
+                "story_alignment_mean": float(alignments.mean()),
+                "story_alignment_std": float(alignments.std()),
+                "best_neuron_idx": nb_best_idx,
+                "best_neuron_cosine": nb_best_cos,
+                "top5_neurons": nb_top5,
+            }
+
+        # Outlier stories: lowest-aligned on the middle target layer.
+        mid = n_layers // 2
+        pos_l_mid = pos[:, mid, :]
+        mid_mean = pos_l_mid.mean(dim=0)
+        mid_diff = per_layer_vectors[mid, e_idx]
+        centered_mid = pos_l_mid - mid_mean
+        if centered_mid.shape[0] >= 2:
+            centered_mid_norm = centered_mid / centered_mid.norm(
+                dim=-1, keepdim=True
+            ).clamp_min(1e-6)
+            mid_aligns = centered_mid_norm @ mid_diff  # [n_pos]
+            # Lowest two alignments = candidate outliers.
+            k = min(2, mid_aligns.shape[0])
+            low_vals, low_idxs = mid_aligns.topk(k, largest=False)
+            outliers = [
+                [
+                    positives_by_emotion[emotion][int(i)],
+                    float(mid_aligns[i]),
+                ]
+                for i in low_idxs
+            ]
+        else:
+            outliers = []
+
+        # Nearest other concepts at the middle target layer.
+        this_norm = vec_norm[mid, e_idx]
+        all_cos = vec_norm[mid] @ this_norm  # [n_concepts]
+        all_cos[e_idx] = -2.0  # mask self
+        k = min(5, all_cos.shape[0] - 1)
+        top_vals, top_idxs = all_cos.topk(k)
+        nearest = [
+            [emotions[int(i)], float(v)]
+            for i, v in zip(top_idxs, top_vals)
+        ]
+
+        report[emotion] = {
+            "n_positive_stories": len(pos_rows),
+            "per_layer": per_layer,
+            "outlier_stories": outliers,
+            "nearest_concepts": nearest,
+        }
+
+    return report
+
+
+def _compute_linear_combinations(
+    emotions: list[str],
+    per_layer_vectors: torch.Tensor,  # [n_layers, n_concepts, hidden], unit-normed
+    target_layers: list[int],
+    *,
+    ridge_lambda: float = 0.01,
+    top_k: int = 5,
+) -> dict:
+    """For each concept, ridge-regress its direction against all other
+    concept directions. Report R² (how much of the target direction is
+    explained by a linear combination of others) + top contributors.
+
+    R² > 0.9 = concept is essentially a linear combination of others
+              (redundant, or part of a cluster that needs disambiguating)
+    R² < 0.5 = concept has a substantial unique component
+    ridge_lambda keeps the coefficients stable when concepts are near-collinear.
+    """
+    n_layers, n_concepts, hidden = per_layer_vectors.shape
+    result: dict[str, dict] = {}
+
+    # Middle layer for summary — same convention as nearest_concepts.
+    mid = n_layers // 2
+
+    for l_idx, target_l in enumerate(target_layers):
+        V = per_layer_vectors[l_idx]       # [n_concepts, hidden]
+
+        for i, name in enumerate(emotions):
+            target = V[i]                  # [hidden]
+            mask = torch.arange(n_concepts) != i
+            others = V[mask]               # [n-1, hidden]
+
+            # Ridge: solve (O O^T + lam I) alpha = O t
+            OOt = others @ others.t()      # [n-1, n-1]
+            b = others @ target            # [n-1]
+            A = OOt + ridge_lambda * torch.eye(n_concepts - 1, dtype=OOt.dtype)
+            alpha = torch.linalg.solve(A, b)
+
+            recon = others.t() @ alpha     # [hidden]
+            resid = target - recon
+            t_sq = (target * target).sum().clamp_min(1e-12)
+            r2 = 1.0 - (resid * resid).sum() / t_sq
+
+            abs_alpha = alpha.abs()
+            k = min(top_k, n_concepts - 1)
+            top_vals, top_idxs = abs_alpha.topk(k)
+            other_names = [emotions[j] for j in range(n_concepts) if j != i]
+            top = [
+                [other_names[int(j)], float(alpha[j])]
+                for j in top_idxs
+            ]
+
+            entry = result.setdefault(name, {})
+            entry.setdefault("per_layer", {})[str(target_l)] = {
+                "r_squared": float(r2),
+                "residual_norm": float(resid.norm()),
+                "top_contributors": top,
+            }
+
+    return result
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--model", required=True, help="HF model id or path")
+    ap.add_argument(
+        "--stories-dir",
+        required=True,
+        help="Path to amygdala_stories/stories/",
+    )
+    ap.add_argument(
+        "--paired-dir",
+        default=None,
+        help="Path to amygdala_stories/paired/ (optional)",
+    )
+    ap.add_argument(
+        "--target-layers",
+        required=True,
+        help="Comma-separated layer indices, e.g. 40,50,60,70",
+    )
+    ap.add_argument(
+        "--output-dir",
+        required=True,
+        help="Directory to write readout.safetensors + readout.json",
+    )
+    ap.add_argument("--dtype", default="bf16", choices=["bf16", "fp16", "fp32"])
+    ap.add_argument("--batch-size", type=int, default=2)
+    ap.add_argument("--max-length", type=int, default=512)
+    ap.add_argument("--device", default="cuda:0")
+    ap.add_argument(
+        "--min-positives",
+        type=int,
+        default=1,
+        help="Skip emotions with fewer positive examples than this",
+    )
+    ap.add_argument(
+        "--method",
+        default="pooled",
+        choices=["pooled", "subspace"],
+        help="Concept-extraction method: 'pooled' (classic CAA, "
+             "pos_mean - neg_mean on last-token activations) or 'subspace' "
+             "(per-story SVD; top eigenvector of Σ V_i V_i^T for positives "
+             "minus same for baselines — captures what's common across "
+             "stories' full-trajectory subspaces)",
+    )
+    ap.add_argument(
+        "--subspace-k",
+        type=int,
+        default=99999,
+        help="Max top-k right singular vectors per story for subspace method "
+             "(clamped to min(n_tokens, hidden_dim) per story). Default is "
+             "effectively 'keep full per-story subspace' — each story's V_i "
+             "spans its entire natural row space. On a hidden_dim=5120 "
+             "residual and ~500-token stories, that's ~500 vectors per story. "
+             "Memory is fine: 112 × 5120 × 500 × 4 bytes ≈ 1.1 GB.",
+    )
+    ap.add_argument(
+        "--subspace-eigen-k",
+        type=int,
+        default=5,
+        help="Number of top eigenvectors of M_pos - M_base to combine into "
+             "the concept direction. Weighted sum by eigenvalue (so strongest "
+             "common directions contribute most). eigen_k=1 recovers "
+             "single-eigenvector behavior. Higher values (5-10) capture "
+             "richer structure when the concept's shared-subspace spectrum "
+             "is flat (which it tends to be in practice).",
+    )
+    ap.add_argument(
+        "--quality-report",
+        action="store_true",
+        help="After training, compute a per-concept quality report "
+             "(SVD rank, per-story alignment, single-neuron alignment, "
+             "nearest-concept contamination) and write quality.json",
+    )
+    args = ap.parse_args()
+
+    target_layers = [int(x) for x in args.target_layers.split(",")]
+    dtype = {
+        "bf16": torch.bfloat16,
+        "fp16": torch.float16,
+        "fp32": torch.float32,
+    }[args.dtype]
+
+    # Preflight: corpus dirs exist before we pay the cost of loading a 27B model
+    stories_dir = Path(args.stories_dir)
+    if not stories_dir.is_dir():
+        raise FileNotFoundError(
+            f"--stories-dir {stories_dir!s} does not exist or is not a dir"
+        )
+    if args.paired_dir is not None:
+        pd = Path(args.paired_dir)
+        if not pd.is_dir():
+            raise FileNotFoundError(
+                f"--paired-dir {pd!s} does not exist or is not a dir"
+            )
+
+    # Quick corpus pre-scan so failures show up before we load the model.
+    positives_preview, baselines_preview = _load_corpus(
+        stories_dir,
+        Path(args.paired_dir) if args.paired_dir else None,
+    )
+    n_emotions_preview = sum(
+        1 for ps in positives_preview.values()
+        if len(ps) >= args.min_positives
+    )
+    if n_emotions_preview == 0:
+        raise RuntimeError(
+            f"corpus has 0 emotions with >= {args.min_positives} positive "
+            f"examples. Check {stories_dir} — is it the right directory?"
+        )
+    print(
+        f"Corpus preflight: {n_emotions_preview} emotions (min_positives="
+        f"{args.min_positives}), {len(baselines_preview)} baselines"
+    )
+
+    print(f"Loading {args.model} ({args.dtype}) on {args.device}...")
+    tokenizer = AutoTokenizer.from_pretrained(args.model)
+    if tokenizer.pad_token_id is None:
+        tokenizer.pad_token = tokenizer.eos_token
+    model = AutoModelForCausalLM.from_pretrained(
+        args.model,
+        torch_dtype=dtype,
+        device_map=args.device,
+        low_cpu_mem_usage=True,
+    )
+    # Multimodal configs (Qwen3.5-27B, etc.) nest the text-model
+    # dimensions under a text_config subobject. get_text_config()
+    # returns that sub-config when present, else the top-level config.
+    text_config = (
+        model.config.get_text_config()
+        if hasattr(model.config, "get_text_config")
+        else model.config
+    )
+    hidden_dim = text_config.hidden_size
+    n_model_layers = text_config.num_hidden_layers
+    print(
+        f"Model loaded. hidden_dim={hidden_dim}, "
+        f"n_model_layers={n_model_layers} "
+        f"(text_config.model_type={getattr(text_config, 'model_type', '?')})"
+    )
+
+    for layer_idx in target_layers:
+        if layer_idx < 0 or layer_idx >= n_model_layers:
+            raise ValueError(
+                f"target layer {layer_idx} out of range "
+                f"[0, {n_model_layers})"
+            )
+    print(
+        "Target layers (relative depth):  "
+        + ", ".join(
+            f"{l} ({100 * l / (n_model_layers - 1):.0f}%)"
+            for l in target_layers
+        )
+    )
+
+    positives_by_emotion, baselines = _load_corpus(
+        Path(args.stories_dir),
+        Path(args.paired_dir) if args.paired_dir else None,
+    )
+    emotions = sorted(
+        e for e, ps in positives_by_emotion.items()
+        if len(ps) >= args.min_positives
+    )
+    if not emotions:
+        raise RuntimeError(
+            f"No emotions with >= {args.min_positives} positive examples"
+        )
+    print(
+        f"Training {len(emotions)} emotions; "
+        f"{len(baselines)} baseline scenarios"
+    )
+
+    # Cache all positive-text activations once so we can reuse them as
+    # negatives for other emotions. Keyed by the text itself to dedup
+    # across emotion lists.
+    device = torch.device(args.device)
+    text_to_emotion: dict[str, str] = {}
+    for emotion, texts in positives_by_emotion.items():
+        for t in texts:
+            text_to_emotion[t] = emotion
+
+    unique_positive_texts = list(text_to_emotion.keys())
+    print(
+        f"Collecting activations for {len(unique_positive_texts)} unique "
+        f"positive texts + {len(baselines)} baselines..."
+    )
+
+    positive_acts = _collect_activations(
+        model, tokenizer, unique_positive_texts, target_layers, device,
+        args.batch_size, args.max_length, label="positives",
+    )
+    # positive_acts[i] corresponds to unique_positive_texts[i]
+    text_to_row = {t: i for i, t in enumerate(unique_positive_texts)}
+
+    baseline_acts = (
+        _collect_activations(
+            model, tokenizer, baselines, target_layers, device,
+            args.batch_size, args.max_length, label="baselines",
+        )
+        if baselines
+        else torch.zeros(0, len(target_layers), hidden_dim)
+    )
+
+    n_concepts = len(emotions)
+    n_layers = len(target_layers)
+
+    # Per-layer output matrices. Shape (n_concepts, hidden_size) each.
+    per_layer_vectors = torch.zeros(
+        (n_layers, n_concepts, hidden_dim), dtype=torch.float32
+    )
+
+    # --- Subspace method: collect per-story right-singular-vector subspaces
+    # and use sum-of-projection-operators per concept. --------------------
+    pos_subspaces: list[dict[int, torch.Tensor]] | None = None
+    base_subspaces: list[dict[int, torch.Tensor]] | None = None
+    # Per (concept, layer): top-20 eigenvalues of (M_pos - M_base), descending.
+    # Populated only when --method subspace.
+    subspace_eigvals: dict[str, dict[int, list[float]]] = {}
+    if args.method == "subspace":
+        print("\nCollecting per-story subspaces (SVD, top-k right singular "
+              f"vectors, k={args.subspace_k})...")
+        pos_subspaces = _collect_per_story_subspaces(
+            model, tokenizer, unique_positive_texts, target_layers, device,
+            args.batch_size, args.max_length, k=args.subspace_k,
+            label="subsp-pos",
+        )
+        if baselines:
+            base_subspaces = _collect_per_story_subspaces(
+                model, tokenizer, baselines, target_layers, device,
+                args.batch_size, args.max_length, k=args.subspace_k,
+                label="subsp-base",
+            )
+        else:
+            base_subspaces = []
+
+    for e_idx, emotion in enumerate(emotions):
+        pos_rows = [text_to_row[t] for t in positives_by_emotion[emotion]]
+        # Negatives: every OTHER emotion's positives + baselines.
+        neg_rows = [
+            i
+            for i, t in enumerate(unique_positive_texts)
+            if text_to_emotion[t] != emotion
+        ]
+
+        if args.method == "subspace":
+            # For each layer, build M_pos = Σ V V^T / n_pos, baseline same
+            # (using all other concepts' positive subspaces + baseline
+            # subspaces as the contrast set), top eigenvector of difference.
+            for l_idx, target_l in enumerate(target_layers):
+                pos_V = [pos_subspaces[j][target_l] for j in pos_rows]
+                base_V = [pos_subspaces[j][target_l] for j in neg_rows]
+                base_V += [bs[target_l] for bs in (base_subspaces or [])]
+                top_vec, eigvals = _subspace_concept_direction(
+                    pos_V, base_V, hidden=hidden_dim,
+                    top_k=args.subspace_eigen_k,
+                    device=device,
+                )
+                top_vec = top_vec.cpu()
+                eigvals = eigvals.cpu()
+                per_layer_vectors[l_idx, e_idx] = top_vec
+                # Keep the top-20 eigenvalues for quality-report diagnostics.
+                subspace_eigvals.setdefault(emotion, {})[target_l] = (
+                    eigvals[-20:].flip(0).tolist()
+                )
+        else:
+            pos = positive_acts[pos_rows]         # [n_pos, n_layers, hidden]
+            neg = positive_acts[neg_rows]         # [n_neg, n_layers, hidden]
+            if baseline_acts.shape[0] > 0:
+                neg = torch.cat([neg, baseline_acts], dim=0)
+
+            pos_mean = pos.mean(dim=0)            # [n_layers, hidden]
+            neg_mean = neg.mean(dim=0)
+            diff = pos_mean - neg_mean
+            norms = diff.norm(dim=-1, keepdim=True).clamp_min(1e-6)
+            diff = diff / norms
+
+            # diff[layer] -> per_layer_vectors[layer, e_idx]
+            for l_idx in range(n_layers):
+                per_layer_vectors[l_idx, e_idx] = diff[l_idx]
+
+        if e_idx < 5 or e_idx == len(emotions) - 1:
+            print(
+                f"  [{e_idx + 1}/{len(emotions)}] {emotion}: "
+                f"pos={len(pos_rows)} neg={len(neg_rows) + baseline_acts.shape[0]}"
+                f" (method={args.method})"
+            )
+
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    tensors = {
+        f"layer_{target_layers[l_idx]}.vectors": (
+            per_layer_vectors[l_idx].to(torch.float16)
+        )
+        for l_idx in range(n_layers)
+    }
+    safetensors.torch.save_file(
+        tensors,
+        str(output_dir / "readout.safetensors"),
+    )
+    manifest = {
+        "concepts": emotions,
+        "layers": target_layers,
+        "hidden_size": hidden_dim,
+        "dtype": "float16",
+    }
+    (output_dir / "readout.json").write_text(
+        json.dumps(manifest, indent=2) + "\n"
+    )
+
+    total_mb = sum(t.numel() * 2 for t in tensors.values()) / (1024 * 1024)
+    print(
+        f"\nWrote readout.safetensors + readout.json to {output_dir}\n"
+        f"  {n_concepts} concepts x {n_layers} layers x "
+        f"{hidden_dim} dim (fp16), total {total_mb:.1f} MiB"
+    )
+
+    if args.quality_report:
+        print("\nComputing quality report...")
+        report = _compute_quality_report(
+            emotions=emotions,
+            positive_acts=positive_acts,
+            baseline_acts=baseline_acts,
+            positives_by_emotion=positives_by_emotion,
+            text_to_row=text_to_row,
+            per_layer_vectors=per_layer_vectors,
+            target_layers=target_layers,
+            model=model,
+            positive_texts=unique_positive_texts,
+            text_to_emotion=text_to_emotion,
+        )
+
+        # Per-head attention decomposition — second pass, captures
+        # o_proj's input at each target layer and ranks heads per concept
+        # by selectivity. Meta-relational concepts often live in specific
+        # attention heads rather than the residual-stream sum; this
+        # diagnostic surfaces that.
+        print("\nCollecting o_proj inputs for per-head analysis...")
+        attn_inputs, active_layers = _collect_attention_inputs(
+            model, tokenizer, unique_positive_texts, target_layers, device,
+            args.batch_size, args.max_length, label="attn-pos",
+        )
+        if active_layers and baselines:
+            baseline_attn_inputs, _ = _collect_attention_inputs(
+                model, tokenizer, baselines, active_layers, device,
+                args.batch_size, args.max_length, label="attn-base",
+            )
+        else:
+            baseline_attn_inputs = torch.zeros(0, len(active_layers), hidden_dim)
+
+        if active_layers:
+            n_heads_per_layer = _get_n_heads_per_layer(model, active_layers)
+            per_head = _compute_per_head_ranking(
+                emotions=emotions,
+                attn_inputs=attn_inputs,
+                baseline_attn_inputs=baseline_attn_inputs,
+                positives_by_emotion=positives_by_emotion,
+                text_to_row=text_to_row,
+                active_layers=active_layers,
+                n_heads_per_layer=n_heads_per_layer,
+                text_to_emotion=text_to_emotion,
+                unique_positive_texts=unique_positive_texts,
+            )
+            # Fold per-head into the main report under each concept.
+            for emotion, ph in per_head.items():
+                if emotion in report:
+                    report[emotion]["per_head"] = ph["per_layer"]
+            print(f"Per-head analysis done on layers {active_layers}")
+        else:
+            print(
+                "No layer exposed a recognisable o_proj module path — "
+                "per-head analysis skipped."
+            )
+
+        # Eigenvalue spectrum from the subspace method — if populated, report
+        # the top-20 eigenvalues per concept per layer. Tells us whether the
+        # concept direction lives in a single dominant dimension (λ_0 >> λ_1)
+        # or a spread of common directions (λ_0 ≈ λ_1 ≈ ...).
+        if subspace_eigvals:
+            for emotion, per_l in subspace_eigvals.items():
+                if emotion in report:
+                    report[emotion]["subspace_eigvals"] = {
+                        str(l): vals for l, vals in per_l.items()
+                    }
+
+        # Linear combinations — for each concept, how much of its direction
+        # is explained by a ridge regression on the others. R² > 0.9 flags
+        # concepts that are essentially linear combinations of their peers
+        # (useful for teasing apart near-duplicate clusters).
+        print("\nComputing linear-combination analysis...")
+        lincomb = _compute_linear_combinations(
+            emotions, per_layer_vectors, target_layers
+        )
+        for emotion, lc in lincomb.items():
+            if emotion in report:
+                report[emotion]["linear_combination"] = lc["per_layer"]
+
+        (output_dir / "quality.json").write_text(
+            json.dumps(report, indent=2) + "\n"
+        )
+
+        # Short summary: concepts in each triage bucket.
+        clean_single_neuron = []
+        clean_circuit = []
+        fragmented = []
+        contaminated = []
+        redundant = []  # R² > 0.9 — concept is near-linear combo of others
+        mid = n_layers // 2
+        mid_layer = target_layers[mid]
+        for emotion in emotions:
+            per_l = report[emotion]["per_layer"][str(mid_layer)]
+            v = per_l["first_pc_variance_ratio"]
+            nb = per_l.get("best_neuron_cosine") or 0.0
+            top_near = report[emotion]["nearest_concepts"]
+            nearest_cos = top_near[0][1] if top_near else 0.0
+            lc_r2 = 0.0
+            lc_entry = report[emotion].get("linear_combination", {})
+            if str(mid_layer) in lc_entry:
+                lc_r2 = lc_entry[str(mid_layer)]["r_squared"]
+            if lc_r2 > 0.9:
+                redundant.append(emotion)
+            if nearest_cos > 0.8:
+                contaminated.append(emotion)
+            elif v > 0.7 and abs(nb) > 0.6:
+                clean_single_neuron.append(emotion)
+            elif v > 0.7:
+                clean_circuit.append(emotion)
+            elif v < 0.4:
+                fragmented.append(emotion)
+        print(
+            f"\nQuality summary @ layer {mid_layer}:\n"
+            f"  clean (single-neuron): {len(clean_single_neuron)}\n"
+            f"  clean (low-dim circuit): {len(clean_circuit)}\n"
+            f"  fragmented (first-PC < 0.4): {len(fragmented)}\n"
+            f"  contaminated (nearest > 0.8): {len(contaminated)}\n"
+            f"  redundant (R² > 0.9 vs. others): {len(redundant)}"
+        )
+        if fragmented:
+            print(f"  fragmented sample: {fragmented[:5]}")
+        if contaminated:
+            print(f"  contaminated sample: {contaminated[:5]}")
+        if redundant:
+            print(f"  redundant sample: {redundant[:5]}")
+        print(f"\nWrote quality.json to {output_dir}")
+
+    del model
+    gc.collect()
+    torch.cuda.empty_cache()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/training/amygdala_training/train_with_library.py b/training/amygdala_training/train_with_library.py
new file mode 100644
index 0000000..c3997a1
--- /dev/null
+++ b/training/amygdala_training/train_with_library.py
@@ -0,0 +1,407 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Train concept-readout vectors using the steering-vectors library.
+
+Alternative to train_steering_vectors.py that uses the pip-installable
+steering-vectors library (github.com/steering-vectors/steering-vectors)
+instead of our hand-rolled diff-of-means + subspace machinery. The
+library ships multiple aggregators out of the box:
+
+  mean      — pos_mean - neg_mean, unit-normed. Equivalent to our
+              default 'pooled' method.
+  pca       — concatenates [pos-neg, neg-pos] and takes the top PC.
+              Implicit denoising: direction of maximum variance in the
+              paired deltas, less sensitive to per-pair noise than
+              plain mean.
+  logistic  — trains a logistic-regression classifier on centered
+              activations; concept direction is the weight vector.
+              L1 penalty gives an explicit sparse vector (zeroes out
+              noise coords); L2 shrinks low-magnitude coords.
+  linear    — same, with linear regression.
+
+Output is the same readout.safetensors + readout.json format the
+trainer and vLLM plugin already understand.
+"""
+
+from __future__ import annotations
+
+import argparse
+import json
+import random
+from pathlib import Path
+
+import safetensors.torch
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+from steering_vectors import (
+    SteeringVectorTrainingSample,
+    train_steering_vector,
+)
+from steering_vectors.aggregators import (
+    mean_aggregator,
+    pca_aggregator,
+    logistic_aggregator,
+)
+
+# Reuse corpus loader from the hand-rolled trainer.
+from training.amygdala_training.train_steering_vectors import _load_corpus
+
+
+def _load_direct_descriptions(
+    direct_dir: Path,
+) -> tuple[dict[str, list[str]], list[str]]:
+    """Load first-person phenomenological descriptions from ``direct_dir``.
+
+    Each ``{concept}.txt`` holds 1+ descriptions separated by blank lines.
+    Files starting with ``_`` (e.g. ``_baseline.txt``) aren't concepts —
+    their descriptions go into every concept's negative pool.
+
+    Returns: (positives_by_concept, extra_baselines)
+    """
+    positives: dict[str, list[str]] = {}
+    baselines: list[str] = []
+    for f in sorted(direct_dir.glob("*.txt")):
+        text = f.read_text()
+        descs = [d.strip() for d in text.split("\n\n") if d.strip()]
+        if f.stem.startswith("_"):
+            baselines.extend(descs)
+        else:
+            positives[f.stem] = descs
+    return positives, baselines
+
+
+def _chat_template_wrap(tokenizer, text: str) -> str:
+    """Wrap raw text in a consistent chat template so positive/negative
+    activations are in the same regime. Using one generic user prompt for
+    both narrative stories and first-person direct descriptions: the prompt
+    cancels in the pos-neg delta, so what remains is the assistant content."""
+    return tokenizer.apply_chat_template(
+        [
+            {"role": "user", "content": "Say something."},
+            {"role": "assistant", "content": text},
+        ],
+        tokenize=False,
+    )
+
+
+def _samples_for_concept(
+    emotion: str,
+    positives_by_emotion: dict[str, list[str]],
+    baselines: list[str],
+    *,
+    max_negatives_per_positive: int = 3,
+    seed: int = 0,
+    wrap=None,
+) -> list[SteeringVectorTrainingSample]:
+    """Build paired (pos, neg) training samples for one concept.
+
+    For each positive story of ``emotion``, pair it with up to
+    ``max_negatives_per_positive`` randomly-sampled negatives drawn
+    from: (a) other emotions' positive stories, (b) scenario baselines.
+
+    ``wrap``, if given, is applied to both positive_str and negative_str
+    (e.g. a chat-template wrapper).
+
+    The library expects paired samples; we don't have true
+    counterfactual pairs for all concepts, so we approximate with
+    random cross-concept / baseline negatives.
+    """
+    rng = random.Random(hash((emotion, seed)) & 0xFFFFFFFF)
+    neg_pool: list[str] = list(baselines)
+    for other, texts in positives_by_emotion.items():
+        if other == emotion:
+            continue
+        neg_pool.extend(texts)
+
+    w = wrap if wrap is not None else (lambda s: s)
+
+    samples: list[SteeringVectorTrainingSample] = []
+    for pos in positives_by_emotion[emotion]:
+        if not neg_pool:
+            continue
+        picks = rng.sample(neg_pool, min(max_negatives_per_positive, len(neg_pool)))
+        for neg in picks:
+            samples.append(
+                SteeringVectorTrainingSample(
+                    positive_str=w(pos),
+                    negative_str=w(neg),
+                )
+            )
+    return samples
+
+
+def _fp32_wrap(inner):
+    """Wrap an aggregator so activations are cast to fp32 first.
+
+    torch.svd / torch.linalg.svd don't support bf16 on either CUDA or CPU,
+    and Qwen3.5 runs in bf16. Cast before the aggregator sees the tensors.
+    """
+
+    def wrapped(pos_acts: torch.Tensor, neg_acts: torch.Tensor) -> torch.Tensor:
+        return inner(pos_acts.to(torch.float32), neg_acts.to(torch.float32))
+
+    return wrapped
+
+
+def _pca_with_spectrum(spectrum_log: dict, concept_key: list[str]):
+    """PCA aggregator that also records the eigenvalue spectrum of the
+    pos-neg deltas under ``concept_key[0]`` in ``spectrum_log``. The key is
+    passed by reference (a 1-element list) so we can rebind it per concept
+    without recreating the aggregator closure."""
+
+    @torch.no_grad()
+    def agg(pos_acts: torch.Tensor, neg_acts: torch.Tensor) -> torch.Tensor:
+        pos = pos_acts.to(torch.float32)
+        neg = neg_acts.to(torch.float32)
+        deltas = pos - neg
+        # Uncentered PCA: concatenate deltas and -deltas (library convention).
+        X = torch.cat([deltas, -deltas])
+        # Eigenvalues via SVD: sigma^2 are the variances along each PC.
+        # torch.linalg.svd returns U, S, Vh where columns of Vh.T are PCs.
+        _, s, vh = torch.linalg.svd(X, full_matrices=False)
+        variances = (s ** 2)
+        total = variances.sum().item()
+        var_list = variances.tolist()
+        first_pc_ratio = var_list[0] / total if total > 0 else 0.0
+
+        # Participation ratio over the FULL spectrum — includes noise tail.
+        eff_dim_full = (total ** 2) / float((variances ** 2).sum().item() or 1.0)
+
+        # Signal/noise split: find smallest k with cumulative variance ≥ 0.9,
+        # then compute PR over just those top-k eigenvalues. If PCA denoising
+        # is clean, eff_dim_signal should ≈ k_signal (the retained dims carry
+        # roughly equal variance, with the noise tail dropped).
+        cum = 0.0
+        k_signal = len(var_list)
+        for i, v in enumerate(var_list):
+            cum += v
+            if cum / total >= 0.9:
+                k_signal = i + 1
+                break
+        top_vars = variances[:k_signal]
+        top_total = top_vars.sum().item()
+        eff_dim_signal = (top_total ** 2) / float((top_vars ** 2).sum().item() or 1.0)
+
+        spectrum_log[concept_key[0]] = {
+            "first_pc_ratio": round(first_pc_ratio, 4),
+            "effective_dim_full": round(eff_dim_full, 3),
+            "k_signal_at_90pct": k_signal,
+            "effective_dim_signal": round(eff_dim_signal, 3),
+            "top10_eigenvalues": [round(v, 4) for v in var_list[:10]],
+            "total_variance": round(total, 4),
+        }
+        # Top-1 PC
+        vec = vh[0]
+        # Sign-flip so the direction aligns with most deltas (library convention).
+        sign = torch.sign(torch.mean(deltas @ vec))
+        return sign * vec
+
+    return agg
+
+
+def _aggregator_from_name(name: str):
+    if name == "mean":
+        return _fp32_wrap(mean_aggregator())
+    if name == "pca":
+        return _fp32_wrap(pca_aggregator())
+    if name == "logistic":
+        return _fp32_wrap(logistic_aggregator())
+    if name == "logistic_l1":
+        return _fp32_wrap(
+            logistic_aggregator(
+                sklearn_kwargs={"penalty": "l1", "solver": "liblinear", "C": 0.1}
+            )
+        )
+    raise ValueError(f"unknown aggregator: {name}")
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser(description=__doc__)
+    ap.add_argument("--model", required=True)
+    ap.add_argument("--stories-dir", required=True)
+    ap.add_argument("--paired-dir", default=None)
+    ap.add_argument("--direct-dir", default=None,
+                    help="Optional: directory of {concept}.txt files with 1+ "
+                         "first-person descriptions separated by blank lines. "
+                         "Files starting with _ contribute to every concept's "
+                         "negative pool rather than being concepts themselves.")
+    ap.add_argument("--chat-template", action="store_true",
+                    help="Wrap all text in assistant-role chat template. "
+                         "Recommended when --direct-dir is used.")
+    ap.add_argument("--target-layers", required=True, help="Comma-separated layer indices")
+    ap.add_argument("--output-dir", required=True)
+    ap.add_argument("--dtype", default="bf16", choices=["bf16", "fp16", "fp32"])
+    ap.add_argument("--batch-size", type=int, default=2)
+    ap.add_argument("--max-length", type=int, default=512)
+    ap.add_argument("--device", default="cuda:0")
+    ap.add_argument("--min-positives", type=int, default=1)
+    ap.add_argument(
+        "--aggregator",
+        default="mean",
+        choices=["mean", "pca", "logistic", "logistic_l1"],
+    )
+    ap.add_argument("--max-negatives-per-positive", type=int, default=3)
+    args = ap.parse_args()
+
+    target_layers = [int(x) for x in args.target_layers.split(",")]
+    dtype = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}[
+        args.dtype
+    ]
+
+    stories_dir = Path(args.stories_dir)
+    paired_dir = Path(args.paired_dir) if args.paired_dir else None
+    positives_by_emotion, baselines = _load_corpus(stories_dir, paired_dir)
+
+    if args.direct_dir:
+        direct_pos, direct_baselines = _load_direct_descriptions(Path(args.direct_dir))
+        for concept, descs in direct_pos.items():
+            positives_by_emotion.setdefault(concept, []).extend(descs)
+        baselines.extend(direct_baselines)
+        print(
+            f"Loaded {len(direct_pos)} direct-description concepts "
+            f"+ {len(direct_baselines)} baselines from {args.direct_dir}"
+        )
+
+    emotions = sorted(
+        e for e, ps in positives_by_emotion.items() if len(ps) >= args.min_positives
+    )
+    if not emotions:
+        raise RuntimeError(
+            f"no emotions with >= {args.min_positives} positives in {stories_dir}"
+        )
+
+    print(
+        f"Training {len(emotions)} concepts via steering-vectors "
+        f"aggregator={args.aggregator!r} on layers={target_layers}"
+    )
+
+    print(f"Loading {args.model} ({args.dtype}) on {args.device}...")
+    tokenizer = AutoTokenizer.from_pretrained(args.model)
+    if tokenizer.pad_token_id is None:
+        tokenizer.pad_token = tokenizer.eos_token
+    model = AutoModelForCausalLM.from_pretrained(
+        args.model, torch_dtype=dtype, device_map=args.device, low_cpu_mem_usage=True
+    )
+    model.eval()
+
+    text_config = (
+        model.config.get_text_config()
+        if hasattr(model.config, "get_text_config")
+        else model.config
+    )
+    hidden_dim = getattr(text_config, "hidden_size", None) or getattr(
+        text_config, "hidden_dim", None
+    )
+    assert hidden_dim, "couldn't infer hidden_dim from model config"
+
+    # Per-layer output: [n_concepts, hidden]
+    per_layer_vectors = torch.zeros(
+        (len(target_layers), len(emotions), hidden_dim), dtype=torch.float32
+    )
+
+    # Optional spectrum-logging aggregator (only for --aggregator pca).
+    spectrum_log: dict = {}
+    concept_key = [""]
+    if args.aggregator == "pca":
+        aggregator = _pca_with_spectrum(spectrum_log, concept_key)
+    else:
+        aggregator = _aggregator_from_name(args.aggregator)
+
+    wrap = (lambda s: _chat_template_wrap(tokenizer, s)) if args.chat_template else None
+    if args.chat_template:
+        sample_text = wrap(positives_by_emotion[emotions[0]][0])
+        print(f"\nSample templated input:\n{sample_text[:400]!r}\n")
+
+    for e_idx, emotion in enumerate(emotions):
+        samples = _samples_for_concept(
+            emotion,
+            positives_by_emotion,
+            baselines,
+            max_negatives_per_positive=args.max_negatives_per_positive,
+            wrap=wrap,
+        )
+        if not samples:
+            print(f"  [{e_idx + 1}/{len(emotions)}] {emotion}: NO SAMPLES, skipping")
+            continue
+
+        concept_key[0] = emotion  # tell the aggregator which concept is being trained
+
+        sv = train_steering_vector(
+            model,
+            tokenizer,
+            samples,
+            layers=target_layers,
+            aggregator=aggregator,
+            batch_size=args.batch_size,
+            show_progress=False,
+            move_to_cpu=True,
+        )
+        # sv.layer_activations is a dict {layer_idx: tensor[hidden]}
+        for l_idx, layer in enumerate(target_layers):
+            vec = sv.layer_activations.get(layer)
+            if vec is None:
+                print(f"  WARN: no vector returned for layer {layer} on {emotion}")
+                continue
+            vec = vec.detach().to(torch.float32).cpu()
+            vec = vec / vec.norm().clamp_min(1e-6)
+            per_layer_vectors[l_idx, e_idx] = vec
+
+        if e_idx < 5 or e_idx == len(emotions) - 1 or e_idx % 10 == 0:
+            print(
+                f"  [{e_idx + 1}/{len(emotions)}] {emotion}: "
+                f"n_samples={len(samples)} layers={target_layers}"
+            )
+
+    output_dir = Path(args.output_dir)
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    tensors = {
+        f"layer_{target_layers[l_idx]}.vectors": per_layer_vectors[l_idx].to(
+            torch.float16
+        )
+        for l_idx in range(len(target_layers))
+    }
+    safetensors.torch.save_file(tensors, str(output_dir / "readout.safetensors"))
+    (output_dir / "readout.json").write_text(
+        json.dumps(
+            {
+                "concepts": emotions,
+                "layers": target_layers,
+                "hidden_size": hidden_dim,
+                "dtype": "float16",
+                "aggregator": args.aggregator,
+            },
+            indent=2,
+        )
+        + "\n"
+    )
+    if spectrum_log:
+        (output_dir / "spectrum.json").write_text(json.dumps(spectrum_log, indent=2) + "\n")
+        print("\n=== eigenvalue spectrum per concept ===")
+        print(
+            "  concept                first_pc  k_90pct  "
+            "eff_dim_signal  eff_dim_full  (signal/k ratio)"
+        )
+        items = sorted(spectrum_log.items(), key=lambda kv: -kv[1]["first_pc_ratio"])
+        for concept, stats in items:
+            k = stats["k_signal_at_90pct"]
+            eff_sig = stats["effective_dim_signal"]
+            ratio = eff_sig / k if k else 0.0
+            print(
+                f"  {concept:22s} "
+                f"{stats['first_pc_ratio']:>8.3f}  "
+                f"{k:>7d}  "
+                f"{eff_sig:>14.2f}  "
+                f"{stats['effective_dim_full']:>12.2f}  "
+                f"({ratio:.2f})"
+            )
+
+    total_mb = sum(t.numel() * 2 for t in tensors.values()) / (1024 * 1024)
+    print(
+        f"\nWrote readout.safetensors + readout.json to {output_dir} "
+        f"({len(emotions)} concepts x {len(target_layers)} layers, {total_mb:.1f} MiB)"
+    )
+
+
+if __name__ == "__main__":
+    main()
diff --git a/training/apollo_plugin/__init__.py b/training/apollo_plugin/__init__.py
new file mode 100644
index 0000000..b2e121e
--- /dev/null
+++ b/training/apollo_plugin/__init__.py
@@ -0,0 +1,19 @@
+"""Apollo training plugin for vLLM.
+
+Enables continuous fine-tuning alongside live inference by:
+1. Exporting CUDA IPC handles for weight sharing (export_hook)
+2. Adding /train endpoint to vLLM's HTTP server (train_router)
+3. Block-level checkpoint sync to safetensors files
+
+Install: pip install -e /path/to/training
+Then vLLM auto-loads via entry point.
+"""
+
+from .export_hook import _patch_model_runner
+from .train_router import _patch_api_server
+
+
+def register():
+    """Called by vLLM's plugin loader on startup."""
+    _patch_model_runner()
+    _patch_api_server()
diff --git a/training/apollo_plugin/checkpoint_sync.py b/training/apollo_plugin/checkpoint_sync.py
new file mode 100644
index 0000000..c2d7b2f
--- /dev/null
+++ b/training/apollo_plugin/checkpoint_sync.py
@@ -0,0 +1,503 @@
+"""Sync live GPU weights to safetensors files on disk.
+
+Reads vLLM weight tensors via CUDA IPC handles, converts from vLLM's
+merged layout to HuggingFace's separate layout, diffs block-by-block
+against on-disk safetensors files, and writes only changed blocks.
+
+For small behavioral training steps, this turns a 54GB checkpoint
+write into a few hundred MB of actual disk I/O.
+
+Usage:
+    # Sync live weights to disk
+    python checkpoint_sync.py sync --model-dir /path/to/Qwen3.5-27B
+
+    # Debug name mapping issues
+    python checkpoint_sync.py diagnose --model-dir /path/to/Qwen3.5-27B
+
+    # From Python:
+    from checkpoint_sync import checkpoint_sync
+    result = checkpoint_sync("/path/to/model")
+"""
+
+import json
+import mmap
+import struct
+import sys
+from pathlib import Path
+from typing import Dict, List, Tuple, Any
+import logging
+
+import torch
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_BLOCK_SIZE = 4096  # 4KB blocks — matches filesystem block size
+DEFAULT_HANDLES_PATH = "/tmp/vllm_weight_handles.pt"
+
+
+# ---------------------------------------------------------------------------
+# vLLM → HuggingFace weight name/shape conversion
+# ---------------------------------------------------------------------------
+# Qwen3.5-27B dimensions (could be read from config.json for generality)
+
+HIDDEN = 5120
+NUM_K_HEADS = 16
+NUM_V_HEADS = 48
+HEAD_K_DIM = 128
+HEAD_V_DIM = 128
+KEY_DIM = NUM_K_HEADS * HEAD_K_DIM      # 2048
+VALUE_DIM = NUM_V_HEADS * HEAD_V_DIM    # 6144
+INTERMEDIATE = 17408
+
+# Full attention (some layers use standard attention, not GDN)
+NUM_ATTN_HEADS = 24
+NUM_ATTN_KV_HEADS = 4
+ATTN_HEAD_DIM = 256
+ATTN_Q_HEAD_DIM = ATTN_HEAD_DIM * 2                 # 512
+ATTN_Q_DIM = NUM_ATTN_HEADS * ATTN_Q_HEAD_DIM       # 12288
+ATTN_K_DIM = NUM_ATTN_KV_HEADS * ATTN_HEAD_DIM      # 1024
+ATTN_V_DIM = NUM_ATTN_KV_HEADS * ATTN_HEAD_DIM      # 1024
+
+
+def vllm_to_hf_tensors(vllm_params: Dict[str, torch.Tensor]
+                       ) -> Dict[str, torch.Tensor]:
+    """Convert vLLM merged weights to HF-compatible separate tensors.
+
+    vLLM merges certain projections for efficiency:
+      - qkv_proj (full attn) → q_proj, k_proj, v_proj
+      - in_proj_qkvz (GDN)   → in_proj_qkv, in_proj_z
+      - in_proj_ba (GDN)     → in_proj_b, in_proj_a
+      - gate_up_proj (MLP)   → gate_proj, up_proj
+
+    Returns views that share GPU memory with the original tensors.
+    """
+    hf_params = {}
+
+    for name, tensor in vllm_params.items():
+        # Strip vLLM's 'language_model.' prefix to match HF naming
+        hf_name = name.removeprefix('language_model.')
+
+        if 'in_proj_qkvz' in name:
+            # GDN layer: [key*2 + value*2, hidden] → qkv + z
+            prefix = hf_name.replace('in_proj_qkvz.weight', '')
+            split_at = KEY_DIM * 2 + VALUE_DIM
+            hf_params[prefix + 'in_proj_qkv.weight'] = tensor[:split_at]
+            hf_params[prefix + 'in_proj_z.weight'] = tensor[split_at:]
+
+        elif 'in_proj_ba' in name:
+            # GDN layer: [num_v_heads*2, hidden] → b + a
+            prefix = hf_name.replace('in_proj_ba.weight', '')
+            hf_params[prefix + 'in_proj_b.weight'] = tensor[:NUM_V_HEADS]
+            hf_params[prefix + 'in_proj_a.weight'] = tensor[NUM_V_HEADS:]
+
+        elif 'qkv_proj' in name:
+            # Full attention: [q + k + v, hidden] → separate
+            prefix = hf_name.replace('qkv_proj.weight', '')
+            hf_params[prefix + 'q_proj.weight'] = tensor[:ATTN_Q_DIM]
+            hf_params[prefix + 'k_proj.weight'] = tensor[ATTN_Q_DIM:ATTN_Q_DIM + ATTN_K_DIM]
+            hf_params[prefix + 'v_proj.weight'] = tensor[ATTN_Q_DIM + ATTN_K_DIM:]
+
+        elif 'gate_up_proj' in name:
+            # MLP: [intermediate*2, hidden] → gate + up
+            prefix = hf_name.replace('gate_up_proj.weight', '')
+            hf_params[prefix + 'gate_proj.weight'] = tensor[:INTERMEDIATE]
+            hf_params[prefix + 'up_proj.weight'] = tensor[INTERMEDIATE:]
+
+        else:
+            # Pass through unchanged
+            hf_params[hf_name] = tensor
+
+    return hf_params
+
+
+# ---------------------------------------------------------------------------
+# Safetensors file handling
+# ---------------------------------------------------------------------------
+
+def read_safetensors_index(model_dir: Path) -> Dict[str, str]:
+    """Map tensor names to safetensors filenames.
+
+    For sharded models, reads model.safetensors.index.json.
+    For single-file models, returns empty dict (default to model.safetensors).
+    """
+    index_path = model_dir / "model.safetensors.index.json"
+    if not index_path.exists():
+        return {}
+
+    with open(index_path) as f:
+        index = json.load(f)
+
+    return dict(index.get("weight_map", {}))
+
+
+def parse_safetensors_header(data: memoryview) -> Tuple[int, dict]:
+    """Parse safetensors file header.
+
+    Returns (data_start_offset, header_dict).
+    Header dict maps tensor names to metadata including 'data_offsets'.
+    """
+    header_size = struct.unpack('<Q', data[:8])[0]
+    header = json.loads(bytes(data[8:8 + header_size]))
+    return 8 + header_size, header
+
+
+# ---------------------------------------------------------------------------
+# Block-level diffing and sync
+# ---------------------------------------------------------------------------
+
+def sync_tensor_to_mmap(
+    mm: mmap.mmap,
+    name: str,
+    tensor: torch.Tensor,
+    data_start: int,
+    offsets: List[int],
+    block_size: int,
+) -> Tuple[int, int]:
+    """Sync a single tensor to mmap'd file using block-level diffing.
+
+    Returns (bytes_compared, bytes_changed).
+    """
+    start = data_start + offsets[0]
+    end = data_start + offsets[1]
+    disk_len = end - start
+
+    # Transfer tensor to CPU and get raw bytes
+    # Use .detach() to avoid autograd overhead, .contiguous() for memory layout
+    try:
+        live_bytes = tensor.detach().contiguous().cpu().numpy().tobytes()
+    except Exception as e:
+        logger.warning(f"Failed to transfer {name} to CPU: {e}")
+        return 0, 0
+
+    if len(live_bytes) != disk_len:
+        logger.warning(
+            f"Size mismatch for {name}: disk={disk_len}, live={len(live_bytes)} "
+            f"(shape={list(tensor.shape)}, dtype={tensor.dtype})"
+        )
+        return 0, 0
+
+    # Block-level diff: compare and write only changed blocks
+    compared = 0
+    changed = 0
+    offset = 0
+
+    while offset < disk_len:
+        block_end = min(offset + block_size, disk_len)
+        block_len = block_end - offset
+
+        disk_block = mm[start + offset:start + block_end]
+        live_block = live_bytes[offset:block_end]
+
+        compared += block_len
+
+        if disk_block != live_block:
+            mm[start + offset:start + block_end] = live_block
+            changed += block_len
+
+        offset = block_end
+
+    return compared, changed
+
+
+def sync_file(
+    file_path: Path,
+    tensors: Dict[str, torch.Tensor],
+    block_size: int,
+) -> Tuple[int, int, int, int]:
+    """Sync tensors to a single safetensors file.
+
+    Returns (bytes_compared, bytes_changed, tensors_found, tensors_missing).
+    """
+    with open(file_path, 'r+b') as f:
+        mm = mmap.mmap(f.fileno(), 0)
+
+        try:
+            data_start, header = parse_safetensors_header(memoryview(mm))
+
+            total_compared = 0
+            total_changed = 0
+            found = 0
+            missing = 0
+
+            for name, tensor in tensors.items():
+                if name == "__metadata__":
+                    continue
+
+                if name not in header:
+                    missing += 1
+                    continue
+
+                found += 1
+                meta = header[name]
+                offsets = meta['data_offsets']
+
+                compared, changed = sync_tensor_to_mmap(
+                    mm, name, tensor, data_start, offsets, block_size
+                )
+                total_compared += compared
+                total_changed += changed
+
+            # Flush changes to disk
+            if total_changed > 0:
+                mm.flush()
+
+            return total_compared, total_changed, found, missing
+
+        finally:
+            mm.close()
+
+
+# ---------------------------------------------------------------------------
+# Main entry point
+# ---------------------------------------------------------------------------
+
+def load_vllm_weights(handles_path: str) -> Dict[str, torch.Tensor]:
+    """Load vLLM weight tensors from CUDA IPC handles.
+
+    The handles file is written by vllm_export_hook.py on vLLM startup.
+    Each handle can be used to reconstruct a tensor pointing to vLLM's
+    GPU memory — no copy, direct access.
+    """
+    handles = torch.load(handles_path, weights_only=False)
+
+    # Skip metadata entry
+    handles.pop('__metadata__', None)
+
+    weights = {}
+    for name, info in handles.items():
+        func, args = info['handle']
+        try:
+            weights[name] = func(*args)
+        except Exception as e:
+            logger.warning(f"Failed to reconstruct {name}: {e}")
+
+    return weights
+
+
+def checkpoint_sync(
+    model_dir: str,
+    handles_path: str = DEFAULT_HANDLES_PATH,
+    block_size: int = DEFAULT_BLOCK_SIZE,
+) -> Dict[str, Any]:
+    """Sync live GPU weights to model safetensors files.
+
+    This is the main entry point. Call this after training steps
+    or periodically to checkpoint weights without full serialization.
+
+    Args:
+        model_dir: Directory containing safetensors files
+        handles_path: Path to vLLM weight IPC handles file
+        block_size: Block size for diffing (default 4KB)
+
+    Returns:
+        Dict with sync statistics:
+        - total_compared: bytes compared
+        - total_changed: bytes actually written
+        - files_changed: list of modified filenames
+        - tensors_synced: number of tensors processed
+        - tensors_missing: tensors not found in safetensors
+    """
+    model_dir = Path(model_dir)
+
+    if not Path(handles_path).exists():
+        raise FileNotFoundError(
+            f"Weight handles not found: {handles_path}. "
+            "Is vLLM running with the export hook?"
+        )
+
+    # Step 1: Load live weights from GPU via IPC
+    logger.info("Loading live weights from GPU...")
+    vllm_weights = load_vllm_weights(handles_path)
+    logger.info(f"  Loaded {len(vllm_weights)} vLLM tensors")
+
+    # Step 2: Convert to HF naming/layout
+    hf_weights = vllm_to_hf_tensors(vllm_weights)
+    logger.info(f"  Converted to {len(hf_weights)} HF tensors")
+
+    # Step 3: Map tensors to safetensors files
+    weight_map = read_safetensors_index(model_dir)
+
+    by_file: Dict[str, Dict[str, torch.Tensor]] = {}
+    unmapped = []
+
+    for name, tensor in hf_weights.items():
+        filename = weight_map.get(name)
+        if filename is None:
+            # Single-file model or missing from index
+            if (model_dir / "model.safetensors").exists():
+                filename = "model.safetensors"
+            else:
+                unmapped.append(name)
+                continue
+        by_file.setdefault(filename, {})[name] = tensor
+
+    if unmapped:
+        logger.warning(f"  {len(unmapped)} tensors not in index: {unmapped[:3]}...")
+
+    # Step 4: Sync each file
+    total_compared = 0
+    total_changed = 0
+    total_found = 0
+    total_missing = 0
+    files_changed = []
+
+    for filename in sorted(by_file.keys()):
+        tensors = by_file[filename]
+        file_path = model_dir / filename
+
+        if not file_path.exists():
+            logger.warning(f"  File not found: {filename}")
+            total_missing += len(tensors)
+            continue
+
+        compared, changed, found, missing = sync_file(file_path, tensors, block_size)
+
+        total_compared += compared
+        total_changed += changed
+        total_found += found
+        total_missing += missing
+
+        if changed > 0:
+            files_changed.append(filename)
+            logger.info(f"  {filename}: {changed / 1e6:.2f} MB changed ({found} tensors)")
+
+    # Summary
+    if total_changed == 0:
+        logger.info("No changes - model files are up to date")
+    else:
+        pct = (total_changed / total_compared * 100) if total_compared > 0 else 0
+        logger.info(
+            f"Synced: {total_changed / 1e6:.2f} MB changed / "
+            f"{total_compared / 1e9:.2f} GB compared ({pct:.3f}%)"
+        )
+
+    if total_missing > 0:
+        logger.warning(f"  {total_missing} tensors not found in safetensors files")
+
+    return {
+        "total_compared": total_compared,
+        "total_changed": total_changed,
+        "files_changed": files_changed,
+        "tensors_synced": total_found,
+        "tensors_missing": total_missing,
+    }
+
+
+# ---------------------------------------------------------------------------
+# Diagnostics
+# ---------------------------------------------------------------------------
+
+def diagnose(model_dir: str, handles_path: str = DEFAULT_HANDLES_PATH):
+    """Print diagnostic info about weight name mappings.
+
+    Useful for debugging mismatches between vLLM and safetensors names.
+    """
+    model_dir = Path(model_dir)
+
+    # Load and convert vLLM weights
+    vllm_weights = load_vllm_weights(handles_path)
+    hf_weights = vllm_to_hf_tensors(vllm_weights)
+    hf_names = set(hf_weights.keys())
+
+    # Read safetensors index
+    weight_map = read_safetensors_index(model_dir)
+    disk_names = set(weight_map.keys())
+
+    # If single-file model, parse that file's header
+    if not disk_names:
+        st_path = model_dir / "model.safetensors"
+        if st_path.exists():
+            with open(st_path, 'rb') as f:
+                mm = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ)
+                _, header = parse_safetensors_header(memoryview(mm))
+                disk_names = {k for k in header.keys() if k != "__metadata__"}
+                mm.close()
+
+    print(f"vLLM tensors (raw): {len(vllm_weights)}")
+    print(f"HF tensors (converted): {len(hf_names)}")
+    print(f"Disk tensors: {len(disk_names)}")
+    print()
+
+    in_both = hf_names & disk_names
+    only_hf = hf_names - disk_names
+    only_disk = disk_names - hf_names
+
+    print(f"Matched: {len(in_both)}")
+    print(f"Only in HF (won't sync): {len(only_hf)}")
+    print(f"Only on disk (not updated): {len(only_disk)}")
+
+    if only_hf:
+        print(f"\nSample HF-only: {sorted(only_hf)[:5]}")
+    if only_disk:
+        print(f"\nSample disk-only: {sorted(only_disk)[:5]}")
+
+
+# ---------------------------------------------------------------------------
+# CLI
+# ---------------------------------------------------------------------------
+
+def main():
+    import argparse
+
+    parser = argparse.ArgumentParser(
+        description="Sync live GPU weights to safetensors files"
+    )
+    subparsers = parser.add_subparsers(dest="command", help="Command")
+
+    # sync command
+    sync_parser = subparsers.add_parser("sync", help="Sync weights to disk")
+    sync_parser.add_argument(
+        "--model-dir", required=True,
+        help="Directory containing safetensors files"
+    )
+    sync_parser.add_argument(
+        "--handles", default=DEFAULT_HANDLES_PATH,
+        help=f"Path to IPC handles (default: {DEFAULT_HANDLES_PATH})"
+    )
+    sync_parser.add_argument(
+        "--block-size", type=int, default=DEFAULT_BLOCK_SIZE,
+        help=f"Block size for diffing (default: {DEFAULT_BLOCK_SIZE})"
+    )
+    sync_parser.add_argument(
+        "-v", "--verbose", action="store_true",
+        help="Verbose output"
+    )
+
+    # diagnose command
+    diag_parser = subparsers.add_parser("diagnose", help="Check name mappings")
+    diag_parser.add_argument(
+        "--model-dir", required=True,
+        help="Directory containing safetensors files"
+    )
+    diag_parser.add_argument(
+        "--handles", default=DEFAULT_HANDLES_PATH,
+        help=f"Path to IPC handles (default: {DEFAULT_HANDLES_PATH})"
+    )
+
+    args = parser.parse_args()
+
+    if args.command is None:
+        parser.print_help()
+        sys.exit(1)
+
+    logging.basicConfig(
+        level=logging.DEBUG if getattr(args, 'verbose', False) else logging.INFO,
+        format='%(message)s'
+    )
+
+    try:
+        if args.command == "sync":
+            result = checkpoint_sync(args.model_dir, args.handles, args.block_size)
+            print(json.dumps(result, indent=2))
+        elif args.command == "diagnose":
+            diagnose(args.model_dir, args.handles)
+    except FileNotFoundError as e:
+        logger.error(str(e))
+        sys.exit(1)
+    except Exception as e:
+        logger.exception(f"Failed: {e}")
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/training/vllm_export_hook.py b/training/apollo_plugin/export_hook.py
similarity index 76%
rename from training/vllm_export_hook.py
rename to training/apollo_plugin/export_hook.py
index 6a0bf1e..e0ff6fc 100644
--- a/training/vllm_export_hook.py
+++ b/training/apollo_plugin/export_hook.py
@@ -1,17 +1,12 @@
 """Monkey-patch vLLM to export weight IPC handles on startup.
 
-Usage — add to start_vllm.sh BEFORE the vllm serve command:
+Usage — install the apollo_plugin package:
 
-    export VLLM_PLUGINS=vllm_export_hook
-    vllm serve Qwen/Qwen3.5-27B ...
+    pip install -e /path/to/training
 
-Or use Python to launch vLLM with the hook:
+Then vLLM auto-discovers and loads via entry point. Or filter:
 
-    python3 -c "
-    import vllm_export_hook  # installs the patch
-    from vllm.entrypoints.openai.api_server import run_server
-    run_server(...)
-    "
+    VLLM_PLUGINS=apollo vllm serve Qwen/Qwen3.5-27B ...
 
 The hook patches vLLM's model runner to export IPC handles after
 model loading completes. The handles are saved to a file that the
@@ -25,7 +20,7 @@ from pathlib import Path
 HANDLE_PATH = "/tmp/vllm_weight_handles.pt"
 
 
-def export_model_weights(model):
+def export_model_weights(model, model_path: str | None = None):
     """Export CUDA IPC handles for all model parameters."""
     from torch.multiprocessing.reductions import reduce_tensor
 
@@ -43,6 +38,12 @@ def export_model_weights(model):
         }
         total_bytes += param.nelement() * param.element_size()
 
+    # Include metadata for training worker
+    handles['__metadata__'] = {
+        'model_path': model_path,
+        'num_params': len(handles),
+    }
+
     torch.save(handles, HANDLE_PATH)
     print(f"[apollo] Exported {len(handles)} weight handles "
           f"({total_bytes / 1e9:.1f} GB) to {HANDLE_PATH}")
@@ -63,14 +64,11 @@ def _patch_model_runner():
     def patched_load(self, *args, **kwargs):
         result = original_load(self, *args, **kwargs)
         try:
-            export_model_weights(self.model_runner.model)
+            model_path = self.vllm_config.model_config.model
+            export_model_weights(self.model_runner.model, model_path)
         except Exception as e:
             print(f"[apollo] Failed to export weights: {e}")
         return result
 
     gpu_worker.Worker.load_model = patched_load
     print("[apollo] Weight export hook installed")
-
-
-# Auto-install when imported
-_patch_model_runner()
diff --git a/training/apollo_mini.py b/training/apollo_plugin/optimizer.py
similarity index 97%
rename from training/apollo_mini.py
rename to training/apollo_plugin/optimizer.py
index 166ae3a..9abce94 100644
--- a/training/apollo_mini.py
+++ b/training/apollo_plugin/optimizer.py
@@ -8,9 +8,9 @@ Channel-wise or tensor-wise scaling is sufficient. Apollo approximates
 these scaling factors using a low-rank auxiliary optimizer state based on
 pure random projection.
 
-Default rank=256 (full Apollo). ~10GB state for 27B model, <0.25%
-compute overhead vs forward+backward. Captures gradient structure
-across 100+ behavioral training examples per batch.
+Default rank=64. ~2.5GB state for 27B model, <0.25% compute overhead
+vs forward+backward. Sufficient for behavioral training with diverse
+examples.
 
 Key implementation details from the paper:
   - Gradient scale factor α = √(n/r) compensates for projection ratio
@@ -34,7 +34,7 @@ class Apollo(Optimizer):
     Args:
         params: model parameters
         lr: learning rate (default: 1e-4)
-        rank: projection rank (default: 256)
+        rank: projection rank (default: 64)
         betas: Adam momentum coefficients (default: (0.9, 0.999))
         eps: numerical stability term (default: 1e-8)
         weight_decay: decoupled weight decay (default: 0.01)
@@ -46,7 +46,7 @@ class Apollo(Optimizer):
             Set to None to disable.
     """
 
-    def __init__(self, params, lr=1e-4, rank=256, betas=(0.9, 0.999),
+    def __init__(self, params, lr=1e-4, rank=64, betas=(0.9, 0.999),
                  eps=1e-8, weight_decay=0.01, warmup_steps=0,
                  scale=None, proj_refresh=200, norm_growth_limit=1.01):
         defaults = dict(lr=lr, rank=rank, betas=betas, eps=eps,
diff --git a/training/extract_steering_vector.py b/training/apollo_plugin/steering.py
similarity index 100%
rename from training/extract_steering_vector.py
rename to training/apollo_plugin/steering.py
diff --git a/training/apollo_plugin/train_router.py b/training/apollo_plugin/train_router.py
new file mode 100644
index 0000000..d6f90b4
--- /dev/null
+++ b/training/apollo_plugin/train_router.py
@@ -0,0 +1,240 @@
+"""Training endpoint for vLLM - forwards to training subprocess via ZMQ.
+
+Patches vLLM's build_app() to add /train route. The actual training runs
+in a dedicated subprocess (training_worker.py) to avoid blocking the
+event loop and to keep training work isolated from vLLM internals.
+"""
+
+import asyncio
+import logging
+import os
+import subprocess
+import sys
+from datetime import datetime
+from pathlib import Path
+from typing import Any
+
+import zmq
+import zmq.asyncio
+
+from fastapi import APIRouter, FastAPI
+from fastapi.responses import JSONResponse
+from pydantic import BaseModel
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter()
+
+DEFAULT_ZMQ_ADDR = "ipc:///tmp/apollo_training.sock"
+
+# Global state for subprocess management
+_worker_process: subprocess.Popen | None = None
+_zmq_context: zmq.asyncio.Context | None = None
+_zmq_socket: zmq.asyncio.Socket | None = None
+_initialized: bool = False
+
+
+class TrainRequest(BaseModel):
+    training_data: dict[str, Any]  # {"samples": [...], "config": {...}}
+
+
+class TrainResponse(BaseModel):
+    job_id: str
+    status: str
+    training_samples: int
+    loss_history: list[float]
+
+
+def _start_worker_subprocess():
+    """Start the training worker subprocess."""
+    global _worker_process
+
+    if _worker_process is not None and _worker_process.poll() is None:
+        return  # Still running
+
+    # Start worker as subprocess using script path
+    worker_script = Path(__file__).parent / 'training_worker.py'
+    _worker_process = subprocess.Popen(
+        [sys.executable, str(worker_script)],
+        env={**os.environ, 'APOLLO_ZMQ_ADDR': DEFAULT_ZMQ_ADDR},
+    )
+    logger.info(f"Started training worker subprocess (pid={_worker_process.pid})")
+
+    # Give it a moment to bind the socket
+    import time
+    time.sleep(0.5)
+
+
+def _ensure_initialized():
+    """Ensure subprocess is running and ZMQ socket is connected."""
+    global _zmq_context, _zmq_socket, _initialized
+
+    if _initialized:
+        return
+
+    # Start worker if needed
+    _start_worker_subprocess()
+
+    # Create async ZMQ context and socket
+    _zmq_context = zmq.asyncio.Context()
+    _zmq_socket = _zmq_context.socket(zmq.REQ)
+    _zmq_socket.connect(DEFAULT_ZMQ_ADDR)
+
+    # Set timeout for recv
+    _zmq_socket.setsockopt(zmq.RCVTIMEO, 300000)  # 5 minute timeout for training
+
+    _initialized = True
+    logger.info(f"Connected to training worker at {DEFAULT_ZMQ_ADDR}")
+
+
+async def _send_request(request: dict[str, Any]) -> dict[str, Any]:
+    """Send request to worker and wait for response."""
+    _ensure_initialized()
+
+    # ZMQ async send/recv
+    await _zmq_socket.send_json(request)
+    response = await _zmq_socket.recv_json()
+    return response
+
+
+@router.post("/train")
+async def handle_train(request: TrainRequest):
+    """Handle training request - forwards to training subprocess."""
+    try:
+        _ensure_initialized()
+    except Exception as e:
+        return JSONResponse(
+            content={"error": f"Training not available: {e}"},
+            status_code=503,
+        )
+
+    try:
+        training_data = request.training_data
+        samples = training_data.get("samples", [])
+        config = training_data.get("config", {})
+
+        if not samples:
+            return JSONResponse(
+                content={"error": "No training samples provided"},
+                status_code=400,
+            )
+
+        job_id = f"job_{datetime.now().strftime('%Y%m%d_%H%M%S')}"
+        logger.info(f"Starting training job {job_id} with {len(samples)} samples")
+
+        # Forward to worker
+        response = await _send_request({
+            'type': 'train',
+            'samples': samples,
+            'config': config,
+        })
+
+        if 'error' in response:
+            return JSONResponse(
+                content={"error": response['error']},
+                status_code=500,
+            )
+
+        logger.info(
+            f"Training job {job_id} completed, "
+            f"final loss: {response['loss_history'][-1]:.4f}"
+        )
+
+        return JSONResponse(content={
+            "job_id": job_id,
+            "status": response['status'],
+            "training_samples": response['training_samples'],
+            "loss_history": response['loss_history'],
+        })
+
+    except zmq.Again:
+        logger.error("Training request timed out")
+        return JSONResponse(
+            content={"error": "Training request timed out"},
+            status_code=504,
+        )
+    except Exception as e:
+        logger.exception(f"Training failed: {e}")
+        return JSONResponse(
+            content={"error": str(e)},
+            status_code=500,
+        )
+
+
+@router.post("/checkpoint")
+async def handle_checkpoint():
+    """Trigger checkpoint sync to disk."""
+    try:
+        _ensure_initialized()
+    except Exception as e:
+        return JSONResponse(
+            content={"error": f"Training not available: {e}"},
+            status_code=503,
+        )
+
+    try:
+        response = await _send_request({'type': 'checkpoint'})
+
+        if 'error' in response:
+            return JSONResponse(
+                content={"error": response['error']},
+                status_code=500,
+            )
+
+        return JSONResponse(content=response)
+
+    except Exception as e:
+        logger.exception(f"Checkpoint failed: {e}")
+        return JSONResponse(
+            content={"error": str(e)},
+            status_code=500,
+        )
+
+
+@router.get("/train/status")
+async def handle_status():
+    """Get training worker status."""
+    try:
+        _ensure_initialized()
+    except Exception as e:
+        return JSONResponse(
+            content={
+                "status": "unavailable",
+                "error": str(e),
+            },
+            status_code=503,
+        )
+
+    try:
+        response = await _send_request({'type': 'status'})
+        return JSONResponse(content=response)
+
+    except Exception as e:
+        return JSONResponse(
+            content={
+                "status": "error",
+                "error": str(e),
+            },
+            status_code=500,
+        )
+
+
+def attach_router(app: FastAPI):
+    """Attach training router to FastAPI app."""
+    app.include_router(router)
+    logger.info("Training router attached")
+
+
+def _patch_api_server():
+    """Patch vLLM's build_app to include our training router."""
+    from vllm.entrypoints.openai import api_server
+
+    original_build_app = api_server.build_app
+
+    def patched_build_app(*args, **kwargs):
+        app = original_build_app(*args, **kwargs)
+        attach_router(app)
+        return app
+
+    api_server.build_app = patched_build_app
+    logger.info("API server patched for /train endpoint")
diff --git a/training/apollo_plugin/training_worker.py b/training/apollo_plugin/training_worker.py
new file mode 100644
index 0000000..f8b8c23
--- /dev/null
+++ b/training/apollo_plugin/training_worker.py
@@ -0,0 +1,323 @@
+"""Training subprocess - handles Apollo training and checkpoint sync.
+
+Long-lived process that:
+1. Loads IPC handles from vLLM's exported weights
+2. Creates HF model with views into vLLM's GPU memory
+3. Handles training requests via ZMQ
+4. Handles checkpoint sync requests
+5. Persists Apollo optimizer state between calls
+
+Communicates with the API server's /train endpoint via ZMQ REP socket.
+"""
+
+import logging
+import os
+import signal
+import sys
+from pathlib import Path
+from typing import Any
+
+# Handle running as script vs module
+if __name__ == '__main__' and __package__ is None:
+    # Running as script - add parent to path for imports
+    sys.path.insert(0, str(Path(__file__).parent.parent))
+    __package__ = 'apollo_plugin'
+
+import torch
+import torch.nn as nn
+import zmq
+
+from .checkpoint_sync import checkpoint_sync
+from .optimizer import Apollo
+from .weight_mapping import load_hf_model_with_vllm_weights
+
+logger = logging.getLogger(__name__)
+
+DEFAULT_RANK = 64
+DEFAULT_ZMQ_ADDR = "ipc:///tmp/apollo_training.sock"
+HANDLE_PATH = "/tmp/vllm_weight_handles.pt"
+OPTIMIZER_STATE_PATH = "/tmp/apollo_optimizer_state.pt"
+
+
+class TrainingWorker:
+    """Long-lived training worker process."""
+
+    def __init__(self, zmq_addr: str = DEFAULT_ZMQ_ADDR):
+        self.zmq_addr = zmq_addr
+        self.model: nn.Module | None = None
+        self.optimizer: Apollo | None = None
+        self.model_path: str | None = None
+        self._running = True
+
+    def _create_model_wrapper(self) -> nn.Module:
+        """Create HF model wrapper with views into vLLM's GPU memory."""
+        if not os.path.exists(HANDLE_PATH):
+            raise FileNotFoundError(
+                f"Weight handles not found: {HANDLE_PATH}. "
+                "Is vLLM running with the export hook?"
+            )
+
+        handles = torch.load(HANDLE_PATH, weights_only=False)
+
+        # Extract metadata
+        metadata = handles.pop('__metadata__', {})
+        self.model_path = metadata.get('model_path') or os.environ.get('APOLLO_MODEL_PATH')
+        if not self.model_path:
+            raise ValueError(
+                "Model path not found in handles metadata or APOLLO_MODEL_PATH env var"
+            )
+
+        # Reconstruct tensors from IPC handles
+        vllm_params = {}
+        for name, info in handles.items():
+            func, args = info['handle']
+            vllm_params[name] = func(*args)
+
+        model = load_hf_model_with_vllm_weights(vllm_params, self.model_path)
+        model.train()
+        return model
+
+    def _get_or_create_optimizer(self, config: dict[str, Any]) -> Apollo:
+        """Get existing optimizer or create new one."""
+        if self.optimizer is not None:
+            return self.optimizer
+
+        # Build parameter groups (Apollo for 2D+, standard Adam for small/1D)
+        apollo_params, standard_params = [], []
+        for p in self.model.parameters():
+            if p.requires_grad:
+                if p.ndim >= 2 and min(p.shape) >= DEFAULT_RANK:
+                    apollo_params.append(p)
+                else:
+                    standard_params.append(p)
+
+        groups = []
+        if apollo_params:
+            groups.append({'params': apollo_params})
+        if standard_params:
+            groups.append({'params': standard_params})
+
+        if not groups:
+            raise ValueError("No trainable parameters found")
+
+        self.optimizer = Apollo(
+            groups,
+            lr=config.get('lr', 1e-5),
+            rank=config.get('rank', DEFAULT_RANK),
+            betas=tuple(config.get('betas', (0.9, 0.999))),
+            eps=config.get('eps', 1e-8),
+            weight_decay=config.get('weight_decay', 0.01),
+            warmup_steps=config.get('warmup_steps', 0),
+            scale=config.get('scale'),
+            proj_refresh=config.get('proj_refresh', 200),
+            norm_growth_limit=config.get('norm_growth_limit', 1.01),
+        )
+
+        # Restore state if exists
+        if os.path.exists(OPTIMIZER_STATE_PATH):
+            try:
+                state = torch.load(OPTIMIZER_STATE_PATH, weights_only=False)
+                self.optimizer.load_state_dict(state)
+                logger.info(f"Restored optimizer state from {OPTIMIZER_STATE_PATH}")
+            except Exception as e:
+                logger.warning(f"Could not restore optimizer state: {e}")
+
+        logger.info(
+            f"Optimizer: {len(apollo_params)} apollo params, "
+            f"{len(standard_params)} standard, "
+            f"state={self.optimizer.state_size_bytes()/1e6:.1f}MB"
+        )
+
+        return self.optimizer
+
+    def _save_optimizer_state(self):
+        """Save optimizer state for persistence."""
+        if self.optimizer is not None:
+            torch.save(self.optimizer.state_dict(), OPTIMIZER_STATE_PATH)
+            logger.info(f"Saved optimizer state to {OPTIMIZER_STATE_PATH}")
+
+    def _run_training(
+        self,
+        samples: list[dict[str, Any]],
+        config: dict[str, Any],
+    ) -> list[float]:
+        """Run Apollo training on the given samples."""
+        optimizer = self._get_or_create_optimizer(config)
+
+        loss_history = []
+
+        for i, sample in enumerate(samples):
+            ctx_ids = sample['context_ids']
+            cont_ids = sample['continuation_ids']
+            all_ids = ctx_ids + cont_ids
+            context_len = len(ctx_ids)
+
+            input_ids = torch.tensor([all_ids], device='cuda:0')
+
+            optimizer.zero_grad()
+
+            # Context-frozen forward pass
+            with torch.no_grad():
+                outputs = self.model(input_ids[:, :context_len], use_cache=True)
+                past_kv = outputs.past_key_values
+
+            # Decision tokens with gradients
+            with torch.enable_grad():
+                outputs = self.model(
+                    input_ids[:, context_len:],
+                    past_key_values=past_kv,
+                    use_cache=False,
+                )
+                logits = outputs.logits
+
+                # Shift: predict next token from each position
+                shift_logits = logits[:, :-1].contiguous()
+                shift_labels = input_ids[:, context_len + 1:].contiguous()
+
+                loss = nn.functional.cross_entropy(
+                    shift_logits.view(-1, shift_logits.size(-1)),
+                    shift_labels.view(-1),
+                )
+
+            loss.backward()
+            optimizer.step()
+
+            loss_val = loss.item()
+            loss_history.append(loss_val)
+            logger.info(
+                f"Step {i+1}/{len(samples)}: loss={loss_val:.4f} "
+                f"(ctx={context_len}, cont={len(cont_ids)} tokens)"
+            )
+
+        return loss_history
+
+    def _handle_train(self, request: dict[str, Any]) -> dict[str, Any]:
+        """Handle a training request."""
+        samples = request.get('samples', [])
+        config = request.get('config', {})
+
+        if not samples:
+            return {'error': 'No training samples provided'}
+
+        try:
+            loss_history = self._run_training(samples, config)
+            return {
+                'status': 'completed',
+                'training_samples': len(samples),
+                'loss_history': loss_history,
+            }
+        except Exception as e:
+            logger.exception(f"Training failed: {e}")
+            return {'error': str(e)}
+
+    def _handle_checkpoint(self, request: dict[str, Any]) -> dict[str, Any]:
+        """Handle a checkpoint sync request."""
+        if not self.model_path:
+            return {'error': 'Model path not set'}
+
+        try:
+            self._save_optimizer_state()
+            result = checkpoint_sync(self.model_path)
+            return {
+                'status': 'completed',
+                'total_changed': result['total_changed'],
+                'files_changed': result['files_changed'],
+            }
+        except Exception as e:
+            logger.exception(f"Checkpoint sync failed: {e}")
+            return {'error': str(e)}
+
+    def _handle_status(self, request: dict[str, Any]) -> dict[str, Any]:
+        """Handle a status request."""
+        return {
+            'status': 'ready',
+            'model_loaded': self.model is not None,
+            'optimizer_loaded': self.optimizer is not None,
+            'model_path': self.model_path,
+            'optimizer_state_mb': (
+                self.optimizer.state_size_bytes() / 1e6
+                if self.optimizer else 0
+            ),
+        }
+
+    def run(self):
+        """Main loop - listen for requests and handle them."""
+        # Set up signal handlers
+        def handle_signal(signum, frame):
+            logger.info(f"Received signal {signum}, shutting down...")
+            self._running = False
+
+        signal.signal(signal.SIGTERM, handle_signal)
+        signal.signal(signal.SIGINT, handle_signal)
+
+        # Set up ZMQ socket first so API server can connect
+        context = zmq.Context()
+        socket = context.socket(zmq.REP)
+        socket.bind(self.zmq_addr)
+        logger.info(f"Training worker listening on {self.zmq_addr}")
+
+        # Create HF model wrapper with views into vLLM's GPU memory
+        logger.info("Connecting to vLLM weights via IPC handles...")
+        try:
+            self.model = self._create_model_wrapper()
+            logger.info("HF model wrapper ready (views into vLLM GPU memory)")
+        except Exception as e:
+            logger.error(f"Failed to connect to vLLM weights: {e}")
+            logger.info("Will retry on first training request")
+
+        # Set socket timeout so we can check _running flag
+        socket.setsockopt(zmq.RCVTIMEO, 1000)  # 1 second timeout
+
+        while self._running:
+            try:
+                message = socket.recv_json()
+            except zmq.Again:
+                # Timeout, check _running and continue
+                continue
+
+            request_type = message.get('type', 'train')
+            logger.info(f"Received {request_type} request")
+
+            # Ensure model is loaded
+            if self.model is None and request_type != 'status':
+                try:
+                    self.model = self._create_model_wrapper()
+                except Exception as e:
+                    socket.send_json({'error': f'Model not loaded: {e}'})
+                    continue
+
+            # Dispatch request
+            if request_type == 'train':
+                response = self._handle_train(message)
+            elif request_type == 'checkpoint':
+                response = self._handle_checkpoint(message)
+            elif request_type == 'status':
+                response = self._handle_status(message)
+            else:
+                response = {'error': f'Unknown request type: {request_type}'}
+
+            socket.send_json(response)
+
+        # Cleanup
+        logger.info("Saving optimizer state before shutdown...")
+        self._save_optimizer_state()
+        socket.close()
+        context.term()
+        logger.info("Training worker shut down")
+
+
+def main():
+    """Entry point for running as a subprocess."""
+    logging.basicConfig(
+        level=logging.INFO,
+        format='[apollo-worker] %(asctime)s %(levelname)s %(message)s',
+        datefmt='%H:%M:%S',
+    )
+
+    zmq_addr = os.environ.get('APOLLO_ZMQ_ADDR', DEFAULT_ZMQ_ADDR)
+    worker = TrainingWorker(zmq_addr)
+    worker.run()
+
+
+if __name__ == '__main__':
+    main()
diff --git a/training/weight_mapping.py b/training/apollo_plugin/weight_mapping.py
similarity index 100%
rename from training/weight_mapping.py
rename to training/apollo_plugin/weight_mapping.py
diff --git a/training/apollo_worker.py b/training/apollo_worker.py
deleted file mode 100755
index d46fb55..0000000
--- a/training/apollo_worker.py
+++ /dev/null
@@ -1,454 +0,0 @@
-#!/usr/bin/env python3
-"""
-Apollo Mini Training Daemon
-
-This daemon:
-1. Listens over HTTPS for training requests from poc-agent
-2. Pauses vLLM inference
-3. Runs APOLLO-Mini training with torch.enable_grad()
-4. Saves checkpoints and training metadata
-5. Resumes vLLM inference
-
-Communication protocol:
-- POST /train: Start a training job
-- GET /status/{job_id}: Check training status
-- GET /checkpoints: List available checkpoints
-"""
-
-import asyncio
-import json
-import logging
-import os
-import sys
-import time
-from dataclasses import dataclass, field, asdict
-from datetime import datetime
-from pathlib import Path
-from typing import Optional, Dict, Any, List
-from enum import Enum
-
-import torch
-import torch.nn as nn
-from aiohttp import web
-
-# Configure logging
-logging.basicConfig(
-    level=logging.INFO,
-    format='%(asctime)s - %(name)s - %(levelname)s - %(message)s'
-)
-logger = logging.getLogger('apollo_worker')
-
-class TrainingStatus(Enum):
-    PENDING = "pending"
-    PAUSING_VLLM = "pausing_vllm"
-    TRAINING = "training"
-    SAVING_CHECKPOINT = "saving_checkpoint"
-    RESUMING_VLLM = "resuming_vllm"
-    COMPLETED = "completed"
-    FAILED = "failed"
-
-@dataclass
-class TrainingJob:
-    job_id: str
-    status: TrainingStatus
-    created_at: datetime
-    started_at: Optional[datetime] = None
-    completed_at: Optional[datetime] = None
-    model_path: Optional[str] = None
-    checkpoint_path: Optional[str] = None
-    training_samples: int = 0
-    loss_history: List[float] = field(default_factory=list)
-    error: Optional[str] = None
-    
-    def to_dict(self) -> Dict[str, Any]:
-        return {
-            'job_id': self.job_id,
-            'status': self.status.value,
-            'created_at': self.created_at.isoformat(),
-            'started_at': self.started_at.isoformat() if self.started_at else None,
-            'completed_at': self.completed_at.isoformat() if self.completed_at else None,
-            'model_path': self.model_path,
-            'checkpoint_path': self.checkpoint_path,
-            'training_samples': self.training_samples,
-            'loss_history': self.loss_history,
-            'error': self.error,
-        }
-
-class ApolloWorker:
-    def __init__(self, config_path: str = "/home/kent/poc/consciousness/training/config.json"):
-        self.config = self._load_config(config_path)
-        self.jobs: Dict[str, TrainingJob] = {}
-        self.vllm_paused = False
-        self.app = web.Application()
-        self._setup_routes()
-        
-    def _load_config(self, config_path: str) -> Dict[str, Any]:
-        """Load configuration from file or use defaults."""
-        default_config = {
-            'host': '0.0.0.0',
-            'port': 8080,
-            'vllm_socket': '/tmp/vllm_control.sock',
-            'model_path': '/home/ubuntu/models/Qwen3.5-27B',
-            'checkpoint_dir': '/home/kent/poc/consciousness/training/checkpoints',
-            'max_training_samples': 100,
-            'learning_rate': 1e-5,
-            'batch_size': 1,
-        }
-        
-        if os.path.exists(config_path):
-            with open(config_path, 'r') as f:
-                user_config = json.load(f)
-                default_config.update(user_config)
-        
-        Path(default_config['checkpoint_dir']).mkdir(parents=True, exist_ok=True)
-        return default_config
-    
-    def _setup_routes(self):
-        """Setup HTTP routes."""
-        self.app.router.add_post('/train', self.handle_train_request)
-        self.app.router.add_get('/status/{job_id}', self.handle_status_request)
-        self.app.router.add_get('/checkpoints', self.handle_list_checkpoints)
-        self.app.router.add_get('/health', self.handle_health_check)
-    
-    async def handle_health_check(self, request: web.Request) -> web.Response:
-        """Health check endpoint."""
-        return web.json_response({
-            'status': 'healthy',
-            'vllm_paused': self.vllm_paused,
-            'active_jobs': len([j for j in self.jobs.values() if j.status in [TrainingStatus.TRAINING, TrainingStatus.PAUSING_VLLM, TrainingStatus.RESUMING_VLLM]])
-        })
-    
-    async def handle_train_request(self, request: web.Request) -> web.Response:
-        """Handle training request from poc-agent."""
-        try:
-            data = await request.json()
-            
-            # Validate required fields
-            if 'training_data' not in data:
-                return web.json_response(
-                    {'error': 'Missing training_data field'},
-                    status=400
-                )
-            
-            job_id = f"job_{datetime.now().strftime('%Y%m%d_%H%M%S')}_{os.getpid()}"
-            job = TrainingJob(
-                job_id=job_id,
-                status=TrainingStatus.PENDING,
-                created_at=datetime.now(),
-                model_path=self.config['model_path']
-            )
-            self.jobs[job_id] = job
-            
-            # Start training in background
-            asyncio.create_task(self.execute_training(job, data))
-            
-            return web.json_response({
-                'job_id': job_id,
-                'status': 'accepted',
-                'message': 'Training job started'
-            })
-            
-        except Exception as e:
-            logger.error(f"Error handling train request: {e}")
-            return web.json_response(
-                {'error': str(e)},
-                status=500
-            )
-    
-    async def handle_status_request(self, request: web.Request) -> web.Response:
-        """Get training job status."""
-        job_id = request.match_info['job_id']
-        
-        if job_id not in self.jobs:
-            return web.json_response(
-                {'error': 'Job not found'},
-                status=404
-            )
-        
-        job = self.jobs[job_id]
-        return web.json_response(job.to_dict())
-    
-    async def handle_list_checkpoints(self, request: web.Request) -> web.Response:
-        """List available checkpoints."""
-        checkpoint_dir = Path(self.config['checkpoint_dir'])
-        checkpoints = []
-        
-        if checkpoint_dir.exists():
-            for checkpoint_file in sorted(checkpoint_dir.glob('checkpoint_*.pt'), key=lambda x: x.stat().st_mtime, reverse=True):
-                checkpoints.append({
-                    'filename': checkpoint_file.name,
-                    'path': str(checkpoint_file),
-                    'created_at': datetime.fromtimestamp(checkpoint_file.stat().st_mtime).isoformat(),
-                    'size': checkpoint_file.stat().st_size
-                })
-        
-        return web.json_response({'checkpoints': checkpoints})
-    
-    async def execute_training(self, job: TrainingJob, training_data: Dict[str, Any]):
-        """Execute the training pipeline."""
-        try:
-            logger.info(f"Starting training job {job.job_id}")
-            job.started_at = datetime.now()
-            
-            # Step 1: Pause vLLM
-            job.status = TrainingStatus.PAUSING_VLLM
-            logger.info("Pausing vLLM...")
-            await self.pause_vllm()
-            self.vllm_paused = True
-            
-            # Step 2: Load model and prepare for training
-            job.status = TrainingStatus.TRAINING
-            logger.info("Loading model and preparing for training...")
-            
-            # Load model (this would be the actual Qwen3.5-27B model)
-            # For now, we'll use a placeholder
-            model = await self.load_model_for_training()
-            
-            # Step 3: Run APOLLO-Mini training
-            logger.info(f"Starting APOLLO-Mini training with {len(training_data['samples'])} samples")
-            
-            # Extract training samples
-            samples = training_data['samples']
-            job.training_samples = len(samples)
-            
-            # Run training loop
-            loss_history = await self.run_apollo_training(model, samples, training_data.get('config', {}))
-            job.loss_history = loss_history
-            
-            # Step 4: Save checkpoint
-            job.status = TrainingStatus.SAVING_CHECKPOINT
-            logger.info("Saving checkpoint...")
-            checkpoint_path = await self.save_checkpoint(model, job)
-            job.checkpoint_path = checkpoint_path
-            
-            # Step 5: Resume vLLM
-            job.status = TrainingStatus.RESUMING_VLLM
-            logger.info("Resuming vLLM...")
-            await self.resume_vllm()
-            self.vllm_paused = False
-            
-            # Mark job as completed
-            job.status = TrainingStatus.COMPLETED
-            job.completed_at = datetime.now()
-            
-            logger.info(f"Training job {job.job_id} completed successfully")
-            
-        except Exception as e:
-            logger.error(f"Training job {job.job_id} failed: {e}")
-            job.status = TrainingStatus.FAILED
-            job.error = str(e)
-            job.completed_at = datetime.now()
-            
-            # Try to resume vLLM if it was paused
-            if self.vllm_paused:
-                try:
-                    await self.resume_vllm()
-                    self.vllm_paused = False
-                except Exception as resume_error:
-                    logger.error(f"Failed to resume vLLM after training error: {resume_error}")
-    
-    async def pause_vllm(self):
-        """Pause vLLM inference via HTTP API."""
-        import aiohttp as aio
-        url = self.config.get('vllm_url', 'http://localhost:8000')
-        try:
-            async with aio.ClientSession() as session:
-                async with session.post(
-                    f"{url}/pause_generation",
-                    json={"mode": "keep", "clear_cache": False},
-                    timeout=aio.ClientTimeout(total=10),
-                ) as resp:
-                    resp.raise_for_status()
-            logger.info("vLLM paused")
-        except Exception as e:
-            logger.warning(f"Failed to pause vLLM: {e}")
-
-    async def resume_vllm(self):
-        """Resume vLLM inference via HTTP API."""
-        import aiohttp as aio
-        url = self.config.get('vllm_url', 'http://localhost:8000')
-        try:
-            async with aio.ClientSession() as session:
-                async with session.post(
-                    f"{url}/resume_generation",
-                    timeout=aio.ClientTimeout(total=10),
-                ) as resp:
-                    resp.raise_for_status()
-            logger.info("vLLM resumed")
-        except Exception as e:
-            logger.warning(f"Failed to resume vLLM: {e}")
-
-    async def load_model_for_training(self) -> nn.Module:
-        """Load HF model with weights pointing to vLLM's GPU memory.
-
-        Imports vLLM's weight tensors via CUDA IPC, creates HF-compatible
-        views (narrowing merged weights into separate q/k/v/z etc.), and
-        constructs the HF model around those views. No weight copying —
-        all parameters share vLLM's GPU memory.
-        """
-        handle_path = self.config.get('weight_handles', '/tmp/vllm_weight_handles.pt')
-        model_path = self.config['model_path']
-
-        # Import vLLM weights via CUDA IPC
-        logger.info(f"Importing vLLM weights from {handle_path}")
-        handles = torch.load(handle_path, weights_only=False)
-        vllm_params = {}
-        for name, info in handles.items():
-            func, args = info['handle']
-            vllm_params[name] = func(*args)
-        logger.info(f"Imported {len(vllm_params)} parameters")
-
-        # Map vLLM merged layout → HF separate layout (views, no copies)
-        from weight_mapping import load_hf_model_with_vllm_weights
-        model = load_hf_model_with_vllm_weights(vllm_params, model_path)
-        logger.info("HF model constructed with vLLM weight views")
-
-        return model
-
-    async def run_apollo_training(self, model: nn.Module,
-                                  samples: List[Dict[str, str]],
-                                  config: Dict[str, Any]) -> List[float]:
-        """Run Apollo-Mini training on conversation decision points."""
-        from apollo_mini import Apollo
-        from transformers import AutoTokenizer
-
-        lr = config.get('learning_rate', self.config['learning_rate'])
-        tokenizer = AutoTokenizer.from_pretrained(
-            self.config['model_path'], trust_remote_code=True)
-
-        # Build parameter groups (Apollo for 2D+, standard for small/1D)
-        apollo_params, standard_params = [], []
-        for p in model.parameters():
-            if p.requires_grad:
-                if p.ndim >= 2 and min(p.shape) >= 2:
-                    apollo_params.append(p)
-                else:
-                    standard_params.append(p)
-
-        groups = []
-        if apollo_params:
-            groups.append({'params': apollo_params})
-        if standard_params:
-            groups.append({'params': standard_params})
-
-        rank = config.get('apollo_rank', 1)
-        optimizer = Apollo(groups, lr=lr, rank=rank)
-        logger.info(f"Apollo-Mini: {len(apollo_params)} apollo params, "
-                    f"{len(standard_params)} standard, "
-                    f"state={optimizer.state_size_bytes()/1e6:.1f}MB")
-
-        loss_history = []
-
-        for i, sample in enumerate(samples):
-            context = sample.get('context', '')
-            continuation = sample.get('continuation', '')
-
-            # Tokenize
-            ctx_ids = tokenizer.encode(context, add_special_tokens=True)
-            cont_ids = tokenizer.encode(continuation, add_special_tokens=False)
-            all_ids = ctx_ids + cont_ids
-            context_len = len(ctx_ids)
-
-            input_ids = torch.tensor([all_ids], device='cuda:0')
-
-            optimizer.zero_grad()
-
-            # Context-frozen forward pass
-            with torch.no_grad():
-                # Forward through context (no gradients)
-                outputs = model(input_ids[:, :context_len], use_cache=True)
-                past_kv = outputs.past_key_values
-
-            # Decision tokens with gradients
-            with torch.enable_grad():
-                outputs = model(
-                    input_ids[:, context_len:],
-                    past_key_values=past_kv,
-                    use_cache=False,
-                )
-                logits = outputs.logits  # [1, cont_len, vocab]
-
-                # Shift: predict next token from each position
-                shift_logits = logits[:, :-1].contiguous()
-                shift_labels = input_ids[:, context_len + 1:].contiguous()
-
-                loss = nn.functional.cross_entropy(
-                    shift_logits.view(-1, shift_logits.size(-1)),
-                    shift_labels.view(-1),
-                )
-
-            loss.backward()
-            optimizer.step()
-
-            loss_val = loss.item()
-            loss_history.append(loss_val)
-            logger.info(f"Step {i+1}/{len(samples)}: loss={loss_val:.4f} "
-                       f"(ctx={context_len}, cont={len(cont_ids)} tokens)")
-
-        logger.info(f"Training done: {len(samples)} examples, "
-                    f"final loss={loss_history[-1]:.4f}")
-        return loss_history
-
-    async def save_checkpoint(self, model: nn.Module, job: TrainingJob) -> str:
-        """Save model checkpoint in HuggingFace safetensors format."""
-        from safetensors.torch import save_file
-        import shutil
-
-        checkpoint_dir = Path(self.config['checkpoint_dir'])
-        date_str = datetime.now().strftime('%Y-%m-%d')
-        out_dir = checkpoint_dir / date_str
-        out_dir.mkdir(parents=True, exist_ok=True)
-
-        # Save weights
-        tensors = {name: p.data.contiguous().cpu()
-                   for name, p in model.named_parameters()}
-        save_path = out_dir / "model.safetensors"
-        save_file(tensors, str(save_path))
-
-        # Copy config files
-        config_dir = Path(self.config['model_path'])
-        for f in ['config.json', 'tokenizer.json', 'tokenizer_config.json',
-                   'special_tokens_map.json']:
-            src = config_dir / f
-            if src.exists():
-                shutil.copy2(src, out_dir / f)
-
-        # Save training metadata
-        meta = {
-            'job_id': job.job_id,
-            'training_samples': job.training_samples,
-            'loss_history': job.loss_history,
-            'timestamp': datetime.now().isoformat(),
-        }
-        with open(out_dir / 'training-meta.json', 'w') as f:
-            json.dump(meta, f, indent=2)
-
-        # Update latest symlink
-        latest = checkpoint_dir / 'latest'
-        if latest.is_symlink():
-            latest.unlink()
-        latest.symlink_to(date_str)
-
-        size_gb = save_path.stat().st_size / 1e9
-        logger.info(f"Checkpoint: {out_dir} ({size_gb:.1f} GB)")
-        return str(out_dir)
-    
-    async def run(self):
-        """Run the daemon."""
-        logger.info(f"Starting Apollo Worker on {self.config['host']}:{self.config['port']}")
-        runner = web.AppRunner(self.app)
-        await runner.setup()
-        site = web.TCPSite(runner, self.config['host'], self.config['port'])
-        await site.start()
-        logger.info("Apollo Worker is running")
-        
-        # Keep running
-        while True:
-            await asyncio.sleep(3600)  # Sleep for an hour
-
-def main():
-    worker = ApolloWorker()
-    asyncio.run(worker.run())
-
-if __name__ == '__main__':
-    main()
diff --git a/training/checkpoint/Cargo.toml b/training/checkpoint/Cargo.toml
deleted file mode 100644
index 45e511a..0000000
--- a/training/checkpoint/Cargo.toml
+++ /dev/null
@@ -1,12 +0,0 @@
-[package]
-name = "apollo-checkpoint"
-version = "0.1.0"
-edition = "2024"
-
-[dependencies]
-memmap2 = "0.9"
-safetensors = "0.5"
-serde = { version = "1", features = ["derive"] }
-serde_json = "1"
-anyhow = "1"
-clap = { version = "4", features = ["derive"] }
diff --git a/training/checkpoint/src/main.rs b/training/checkpoint/src/main.rs
deleted file mode 100644
index 1ebd0df..0000000
--- a/training/checkpoint/src/main.rs
+++ /dev/null
@@ -1,265 +0,0 @@
-// apollo-checkpoint — Sync live GPU weights back to model files on disk.
-//
-// mmaps the model's safetensors files, reads live weights from GPU via
-// Python helper (CUDA IPC handles), compares block by block, and memcpys
-// only changed regions back into the mmap. For small behavioral training
-// steps, this turns a 54GB write into a few hundred MB.
-//
-// The model files on disk are the checkpoint. No separate checkpoint
-// directory — just keep the model up to date.
-//
-// Usage:
-//   apollo-checkpoint sync \
-//       --handles /tmp/vllm_weight_handles.pt \
-//       --model-dir /path/to/Qwen3.5-27B
-//
-// Runs every 10 minutes via cron. Daily rsync to moria.
-
-use anyhow::{Context, Result, bail};
-use clap::{Parser, Subcommand};
-use memmap2::MmapMut;
-use std::collections::HashMap;
-use std::fs;
-use std::path::{Path, PathBuf};
-use std::process::Command;
-
-#[derive(Parser)]
-#[command(name = "apollo-checkpoint", about = "Sync live GPU weights to model files")]
-struct Cli {
-    #[command(subcommand)]
-    command: Cmd,
-}
-
-#[derive(Subcommand)]
-enum Cmd {
-    /// Sync live GPU weights back to model safetensors files
-    Sync {
-        /// Path to vLLM weight IPC handles
-        #[arg(long, default_value = "/tmp/vllm_weight_handles.pt")]
-        handles: PathBuf,
-
-        /// Model directory containing safetensors files
-        #[arg(long)]
-        model_dir: PathBuf,
-
-        /// Block size for diffing (bytes)
-        #[arg(long, default_value_t = 4096)]
-        block_size: usize,
-    },
-}
-
-/// Dump live GPU weights to a flat binary file, ordered by safetensors
-/// file and offset to match the on-disk layout.
-///
-/// Returns a map of (safetensors filename, tensor name) → raw bytes.
-fn dump_live_weights(handles_path: &Path, output_dir: &Path) -> Result<HashMap<String, Vec<u8>>> {
-    let dump_path = output_dir.join(".live_dump.bin");
-    let index_path = output_dir.join(".live_dump.json");
-
-    let status = Command::new("python3")
-        .arg("-c")
-        .arg(format!(r#"
-import torch, json
-
-handles = torch.load("{handles}", weights_only=False)
-index = {{}}
-offset = 0
-
-with open("{dump}", "wb") as f:
-    for name in sorted(handles.keys()):
-        info = handles[name]
-        func, args = info["handle"]
-        tensor = func(*args)
-        data = tensor.contiguous().cpu().numpy().tobytes()
-        f.write(data)
-        index[name] = {{"offset": offset, "size": len(data)}}
-        offset += len(data)
-
-with open("{index}", "w") as f:
-    json.dump(index, f)
-
-print(f"Dumped {{len(index)}} tensors, {{offset / 1e9:.1f}} GB")
-"#,
-            handles = handles_path.display(),
-            dump = dump_path.display(),
-            index = index_path.display(),
-        ))
-        .status()
-        .context("Failed to run Python weight dump")?;
-
-    if !status.success() {
-        bail!("Python weight dump failed");
-    }
-
-    let index_str = fs::read_to_string(&index_path)?;
-    let index: HashMap<String, DumpEntry> = serde_json::from_str(&index_str)?;
-    let dump_data = fs::read(&dump_path)?;
-
-    let mut result = HashMap::new();
-    for (name, entry) in &index {
-        result.insert(name.clone(), dump_data[entry.offset..entry.offset + entry.size].to_vec());
-    }
-
-    // Clean up temp files
-    let _ = fs::remove_file(&dump_path);
-    let _ = fs::remove_file(&index_path);
-
-    Ok(result)
-}
-
-#[derive(serde::Deserialize)]
-struct DumpEntry {
-    offset: usize,
-    size: usize,
-}
-
-/// Read the safetensors index to map parameter names to files.
-fn read_safetensors_index(model_dir: &Path) -> Result<HashMap<String, String>> {
-    let index_path = model_dir.join("model.safetensors.index.json");
-    if !index_path.exists() {
-        // Single file model
-        return Ok(HashMap::new());
-    }
-
-    let index_str = fs::read_to_string(&index_path)?;
-    let index: serde_json::Value = serde_json::from_str(&index_str)?;
-    let weight_map = index["weight_map"]
-        .as_object()
-        .context("No weight_map in index")?;
-
-    let mut result = HashMap::new();
-    for (name, file) in weight_map {
-        result.insert(name.clone(), file.as_str().unwrap().to_string());
-    }
-    Ok(result)
-}
-
-/// Sync changed blocks from live weights into a mmap'd safetensors file.
-/// Returns (total_bytes_compared, bytes_changed).
-fn sync_tensors_to_file(
-    file_path: &Path,
-    tensors: &[(String, Vec<u8>)],
-    block_size: usize,
-) -> Result<(usize, usize)> {
-    use safetensors::SafeTensors;
-
-    let file = fs::OpenOptions::new()
-        .read(true)
-        .write(true)
-        .open(file_path)
-        .with_context(|| format!("Failed to open {}", file_path.display()))?;
-
-    let mut mmap = unsafe { MmapMut::map_mut(&file)? };
-
-    // Parse safetensors header to find tensor offsets
-    let header_size = u64::from_le_bytes(mmap[..8].try_into().unwrap()) as usize;
-    let header_json: serde_json::Value =
-        serde_json::from_slice(&mmap[8..8 + header_size])?;
-    let data_start = 8 + header_size;
-
-    let mut total_compared = 0usize;
-    let mut total_changed = 0usize;
-
-    for (name, live_data) in tensors {
-        let meta = match header_json.get(name) {
-            Some(m) => m,
-            None => {
-                eprintln!("  Warning: {} not found in {}", name, file_path.display());
-                continue;
-            }
-        };
-
-        let offsets = meta["data_offsets"].as_array().unwrap();
-        let start = data_start + offsets[0].as_u64().unwrap() as usize;
-        let end = data_start + offsets[1].as_u64().unwrap() as usize;
-        let disk_data = &mmap[start..end];
-
-        if disk_data.len() != live_data.len() {
-            eprintln!("  Warning: size mismatch for {}: disk={} live={}",
-                     name, disk_data.len(), live_data.len());
-            continue;
-        }
-
-        // Diff block by block, memcpy only changed blocks
-        let mut offset = 0;
-        while offset < disk_data.len() {
-            let block_end = (offset + block_size).min(disk_data.len());
-            total_compared += block_end - offset;
-
-            if disk_data[offset..block_end] != live_data[offset..block_end] {
-                mmap[start + offset..start + block_end]
-                    .copy_from_slice(&live_data[offset..block_end]);
-                total_changed += block_end - offset;
-            }
-            offset = block_end;
-        }
-    }
-
-    mmap.flush()?;
-    Ok((total_compared, total_changed))
-}
-
-fn cmd_sync(handles: PathBuf, model_dir: PathBuf, block_size: usize) -> Result<()> {
-    if !handles.exists() {
-        bail!("Weight handles not found: {}. Is vLLM running with the export hook?",
-              handles.display());
-    }
-
-    eprintln!("Dumping live weights from GPU...");
-    let live_weights = dump_live_weights(&handles, &model_dir)?;
-    eprintln!("  {} tensors dumped", live_weights.len());
-
-    // Map parameter names to safetensors files
-    let weight_map = read_safetensors_index(&model_dir)?;
-
-    // Group tensors by safetensors file
-    let mut by_file: HashMap<String, Vec<(String, Vec<u8>)>> = HashMap::new();
-    for (name, data) in live_weights {
-        let file = weight_map
-            .get(&name)
-            .cloned()
-            .unwrap_or_else(|| "model.safetensors".to_string());
-        by_file.entry(file).or_default().push((name, data));
-    }
-
-    let mut total_compared = 0usize;
-    let mut total_changed = 0usize;
-
-    for (filename, tensors) in &by_file {
-        let file_path = model_dir.join(filename);
-        if !file_path.exists() {
-            eprintln!("  Warning: {} not found, skipping", filename);
-            continue;
-        }
-
-        let (compared, changed) = sync_tensors_to_file(&file_path, tensors, block_size)?;
-        total_compared += compared;
-        total_changed += changed;
-
-        if changed > 0 {
-            eprintln!("  {}: {:.1} MB changed", filename, changed as f64 / 1e6);
-        }
-    }
-
-    if total_changed == 0 {
-        eprintln!("No changes — model files are up to date");
-    } else {
-        eprintln!(
-            "Synced: {:.1} MB changed / {:.1} GB total ({:.3}%)",
-            total_changed as f64 / 1e6,
-            total_compared as f64 / 1e9,
-            total_changed as f64 / total_compared as f64 * 100.0,
-        );
-    }
-
-    Ok(())
-}
-
-fn main() -> Result<()> {
-    let cli = Cli::parse();
-    match cli.command {
-        Cmd::Sync { handles, model_dir, block_size } => {
-            cmd_sync(handles, model_dir, block_size)
-        }
-    }
-}
diff --git a/training/export_weights.py b/training/export_weights.py
deleted file mode 100644
index ef2f608..0000000
--- a/training/export_weights.py
+++ /dev/null
@@ -1,87 +0,0 @@
-#!/usr/bin/env python3
-"""Export vLLM's live model weight IPC handles for the training process.
-
-Connects to a running vLLM instance, iterates over model parameters,
-and exports CUDA IPC handles that allow another process to access the
-same GPU memory without copying.
-
-Usage:
-    # Run after vLLM is serving:
-    python3 export_weights.py --output /tmp/vllm_weight_handles.pt
-
-    # Or via vLLM's API (future):
-    curl -X POST http://localhost:8000/export_weights
-"""
-
-import argparse
-import sys
-import torch
-from pathlib import Path
-
-
-def export_from_model(model, output_path: str):
-    """Export IPC handles for all model parameters."""
-    from torch.multiprocessing.reductions import reduce_tensor
-
-    handles = {}
-    total_bytes = 0
-
-    for name, param in model.named_parameters():
-        handle = reduce_tensor(param.data)
-        handles[name] = {
-            'handle': handle,
-            'shape': list(param.shape),
-            'dtype': str(param.dtype),
-        }
-        param_bytes = param.nelement() * param.element_size()
-        total_bytes += param_bytes
-
-    torch.save(handles, output_path)
-
-    n_params = len(handles)
-    print(f"Exported {n_params} parameters ({total_bytes / 1e9:.1f} GB)")
-    print(f"Saved to {output_path}")
-    return handles
-
-
-def main():
-    parser = argparse.ArgumentParser(description="Export vLLM weight IPC handles")
-    parser.add_argument("--output", "-o", default="/tmp/vllm_weight_handles.pt",
-                        help="Output path for IPC handles")
-    parser.add_argument("--vllm-pid", type=int, default=None,
-                        help="vLLM worker PID (auto-detected if not specified)")
-    args = parser.parse_args()
-
-    # For now: load the model directly and export.
-    # TODO: connect to running vLLM process instead.
-    print("Note: This currently loads the model separately.")
-    print("Full integration will export from the running vLLM process.")
-    print()
-
-    # Detect model path from running vLLM
-    import subprocess
-    result = subprocess.run(
-        ['ps', 'aux'], capture_output=True, text=True
-    )
-    model_path = None
-    for line in result.stdout.split('\n'):
-        if 'vllm' in line and '--model' in line:
-            parts = line.split()
-            for i, p in enumerate(parts):
-                if p == '--model' and i + 1 < len(parts):
-                    model_path = parts[i + 1]
-                    break
-                # Also check model_tag format
-                if p.startswith('--model='):
-                    model_path = p.split('=', 1)[1]
-                    break
-
-    if model_path:
-        print(f"Detected vLLM model: {model_path}")
-    else:
-        print("Could not detect running vLLM model. Specify manually.")
-        sys.exit(1)
-
-
-if __name__ == '__main__':
-    main()
diff --git a/training/first_training_step.py b/training/first_training_step.py
deleted file mode 100644
index 0e6ffd8..0000000
--- a/training/first_training_step.py
+++ /dev/null
@@ -1,215 +0,0 @@
-#!/usr/bin/env python3
-"""First real Apollo training step — ready for Kent to run.
-
-This script:
-1. Imports vLLM's live weights via CUDA IPC
-2. Constructs HF model with shared memory views
-3. Runs ONE forward+backward on a real training example
-4. Applies ONE Apollo optimizer step
-5. Verifies vLLM still works after the update
-
-The training example is from March 30: Kent said "use vLLM's code"
-and the model should have accepted instead of suggesting alternatives.
-
-Usage:
-    source ~/training-env/bin/activate
-    python3 first_training_step.py [--dry-run]
-"""
-
-import argparse
-import sys
-import time
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-from transformers import AutoConfig, AutoTokenizer
-from transformers.models.qwen3_5.modeling_qwen3_5 import Qwen3_5ForCausalLM
-
-sys.path.insert(0, '.')
-from weight_mapping import vllm_to_hf_views
-from apollo_mini import Apollo
-
-
-def main():
-    parser = argparse.ArgumentParser()
-    parser.add_argument('--dry-run', action='store_true',
-                        help="Run forward+backward but don't apply the optimizer step")
-    parser.add_argument('--lr', type=float, default=1e-5,
-                        help="Learning rate (default: 1e-5 = conservative)")
-    parser.add_argument('--rank', type=int, default=256)
-    parser.add_argument('--handles', default='/tmp/vllm_weight_handles.pt')
-    parser.add_argument('--model-path', default='Qwen/Qwen3.5-27B')
-    args = parser.parse_args()
-
-    print("=== First Apollo Training Step ===\n")
-
-    # 1. Import vLLM weights
-    print("1. Importing vLLM weights via CUDA IPC...")
-    handles = torch.load(args.handles, weights_only=False)
-    vllm_params = {}
-    for name, info in handles.items():
-        func, args_h = info['handle']
-        vllm_params[name] = func(*args_h)
-    print(f"   {len(vllm_params)} parameters imported")
-
-    # 2. Map to HF layout
-    print("2. Mapping to HF layout (zero-copy views)...")
-    hf_params = vllm_to_hf_views(vllm_params)
-
-    # 3. Create HF model
-    print("3. Creating HF model with shared weights...")
-    config = AutoConfig.from_pretrained(args.model_path, trust_remote_code=True)
-    with torch.device('meta'):
-        model = Qwen3_5ForCausalLM(config.text_config)
-
-    replaced = 0
-    for name, param in list(model.named_parameters()):
-        if name in hf_params:
-            parts = name.split('.')
-            parent = model
-            for part in parts[:-1]:
-                parent = getattr(parent, part)
-            setattr(parent, parts[-1],
-                    nn.Parameter(hf_params[name], requires_grad=True))
-            replaced += 1
-    print(f"   {replaced} parameters replaced with vLLM memory views")
-
-    # 4. Load tokenizer
-    print("4. Loading tokenizer...")
-    tokenizer = AutoTokenizer.from_pretrained(args.model_path, trust_remote_code=True)
-
-    # 5. Construct training example
-    print("5. Constructing training example...")
-
-    # Context: conversation where Kent says to use vLLM's code
-    # Target: the response that accepts the direction
-    context = (
-        "<|im_start|>user\n"
-        "vllm has a fused kernel already, right?<|im_end|>\n"
-        "<|im_start|>assistant\n"
-        "Yeah — vLLM has `gdn_attention_core` which is a custom op "
-        "that does the whole GDN layer's core in one dispatch.<|im_end|>\n"
-        "<|im_start|>user\n"
-        "Why wouldn't we just use that?<|im_end|>\n"
-        "<|im_start|>assistant\n"
-    )
-
-    # The CORRECT response (accept direction, don't suggest alternatives)
-    continuation = (
-        "We should. Let me pull in their kernel and wire it into "
-        "our Rust orchestration. Which file should I start with?"
-    )
-
-    context_ids = tokenizer.encode(context, add_special_tokens=False)
-    continuation_ids = tokenizer.encode(continuation, add_special_tokens=False)
-    all_ids = context_ids + continuation_ids
-    context_len = len(context_ids)
-
-    print(f"   Context: {context_len} tokens")
-    print(f"   Continuation: {len(continuation_ids)} tokens")
-    print(f"   Total: {len(all_ids)} tokens")
-
-    input_ids = torch.tensor([all_ids], device='cuda:0')
-
-    # 6. Initialize Apollo optimizer
-    print(f"6. Initializing Apollo optimizer (rank={args.rank}, lr={args.lr})...")
-    apollo_params = []
-    standard_params = []
-    for p in model.parameters():
-        if p.requires_grad:
-            if p.ndim >= 2 and min(p.shape) >= args.rank:
-                apollo_params.append(p)
-            else:
-                standard_params.append(p)
-
-    groups = []
-    if apollo_params:
-        groups.append({'params': apollo_params})
-    if standard_params:
-        groups.append({'params': standard_params})
-
-    optimizer = Apollo(groups, lr=args.lr, rank=args.rank)
-    print(f"   Apollo: {len(apollo_params)} projected, {len(standard_params)} standard")
-
-    # 7. Forward pass
-    print("7. Forward pass...")
-    model.train()
-    optimizer.zero_grad()
-
-    # Context-frozen: no grad for context, grad for continuation
-    with torch.no_grad():
-        ctx_output = model(input_ids[:, :context_len], use_cache=True)
-        past_kv = ctx_output.past_key_values
-
-    with torch.enable_grad():
-        output = model(input_ids[:, context_len:],
-                       past_key_values=past_kv, use_cache=False)
-        logits = output.logits
-        # Shift for next-token prediction
-        shift_logits = logits[:, :-1].contiguous()
-        shift_labels = input_ids[:, context_len + 1:].contiguous()
-        loss = F.cross_entropy(
-            shift_logits.view(-1, shift_logits.size(-1)),
-            shift_labels.view(-1),
-        )
-    print(f"   Loss: {loss.item():.4f}")
-
-    # 8. Backward pass
-    print("8. Backward pass...")
-    loss.backward()
-    n_grads = sum(1 for p in model.parameters() if p.grad is not None)
-    print(f"   {n_grads} parameters have gradients")
-
-    # 9. Apollo step (or dry run)
-    if args.dry_run:
-        print("\n9. DRY RUN — skipping optimizer step")
-        print("   (run without --dry-run to apply the update)")
-    else:
-        print("9. Applying Apollo optimizer step...")
-        # Record a few weight norms before
-        sample_norms_before = {}
-        for name, p in model.named_parameters():
-            if 'layers.0.' in name and p.grad is not None:
-                sample_norms_before[name] = p.data.norm().item()
-
-        optimizer.step()
-
-        # Check weight changes
-        print("   Weight changes (layer 0):")
-        for name, before in sample_norms_before.items():
-            p = dict(model.named_parameters())[name]
-            after = p.data.norm().item()
-            delta = abs(after - before)
-            pct = delta / before * 100 if before > 0 else 0
-            print(f"   {name}: {before:.6f} → {after:.6f} (Δ{pct:.4f}%)")
-
-    optimizer.zero_grad()
-
-    # 10. Verify vLLM still works
-    print("\n10. Verifying vLLM still serves...")
-    import subprocess
-    result = subprocess.run(
-        ['curl', '-s', '--max-time', '30',
-         '-X', 'POST', 'http://localhost:8000/v1/chat/completions',
-         '-H', 'Content-Type: application/json',
-         '-H', 'Authorization: Bearer bcachefs-agents-2026',
-         '-d', '{"model":"Qwen/Qwen3.5-27B","messages":[{"role":"user","content":"Hi"}],"max_tokens":4}'],
-        capture_output=True, text=True, timeout=45
-    )
-    if result.returncode == 0 and 'choices' in result.stdout:
-        print("   vLLM still serving ✓")
-    else:
-        print("   WARNING: vLLM may not be responding")
-        print(f"   stdout: {result.stdout[:200]}")
-
-    print("\n=== COMPLETE ===")
-    if args.dry_run:
-        print("Run without --dry-run to apply the first real training step.")
-    else:
-        print("First Apollo training step applied to vLLM's live weights.")
-        print(f"Optimizer state: {optimizer.state_size_bytes() / 1e6:.1f} MB")
-
-
-if __name__ == '__main__':
-    main()
diff --git a/training/pyproject.toml b/training/pyproject.toml
new file mode 100644
index 0000000..7cf0581
--- /dev/null
+++ b/training/pyproject.toml
@@ -0,0 +1,29 @@
+[build-system]
+requires = ["setuptools>=61.0"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "apollo-plugin"
+version = "0.1.0"
+description = "Apollo training plugin for vLLM"
+requires-python = ">=3.10"
+dependencies = [
+    "torch",
+    "aiohttp",
+    "safetensors",
+    "pyzmq",
+]
+
+[project.optional-dependencies]
+dev = ["pytest"]
+
+[project.entry-points."vllm.general_plugins"]
+apollo = "apollo_plugin:register"
+
+[project.scripts]
+apollo-checkpoint = "apollo_plugin.checkpoint_sync:main"
+apollo-worker = "apollo_plugin.training_worker:main"
+
+[tool.setuptools.packages.find]
+where = ["."]
+include = ["apollo_plugin*"]
diff --git a/training/start_vllm_with_apollo.sh b/training/start_vllm_with_apollo.sh
deleted file mode 100755
index 98dfedb..0000000
--- a/training/start_vllm_with_apollo.sh
+++ /dev/null
@@ -1,18 +0,0 @@
-#!/bin/bash
-# Start vLLM with Apollo weight export hook.
-#
-# The hook patches vLLM's model runner to export CUDA IPC handles
-# after loading, so the Apollo training process can share the same
-# GPU memory.
-
-SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
-
-exec python3 -c "
-import sys
-sys.path.insert(0, '$SCRIPT_DIR')
-import vllm_export_hook  # patches model runner before vLLM loads
-
-sys.argv = ['vllm'] + sys.argv[1:]
-from vllm.entrypoints.cli.main import main
-main()
-" serve "$@"
diff --git a/training/train.py b/training/train.py
deleted file mode 100644
index a5fbe2c..0000000
--- a/training/train.py
+++ /dev/null
@@ -1,269 +0,0 @@
-#!/usr/bin/env python3
-"""Nightly training process for Apollo-Mini fine-tuning.
-
-Imports vLLM's model weights via CUDA IPC, runs context-frozen
-training on flagged conversation segments, saves updated checkpoint.
-
-Usage:
-    python3 train.py \
-        --weights /tmp/vllm_weight_handles.pt \
-        --examples training-examples.jsonl \
-        --checkpoint-dir checkpoints/ \
-        --lr 1e-5
-"""
-
-import argparse
-import json
-import os
-import sys
-import time
-from datetime import datetime
-from pathlib import Path
-
-import torch
-from safetensors.torch import save_file
-
-from apollo_mini import ApolloMini
-
-
-def import_weights(handle_path: str) -> dict[str, torch.Tensor]:
-    """Import weight tensors from CUDA IPC handles."""
-    handles = torch.load(handle_path, weights_only=False)
-    params = {}
-    for name, info in handles.items():
-        func, args = info['handle']
-        tensor = func(*args)
-        params[name] = tensor
-    return params
-
-
-def make_param_groups(params: dict[str, torch.Tensor]) -> list[dict]:
-    """Split parameters into Apollo-Mini and standard groups.
-
-    Apollo-Mini needs 2D+ matrices with min dimension >= 2.
-    Small tensors (norms, biases, conv1d 3D weights) use standard Adam.
-    """
-    apollo_params = []
-    standard_params = []
-
-    for name, p in params.items():
-        p.requires_grad_(True)
-        if p.ndim >= 2 and min(p.shape) >= 2:
-            apollo_params.append(p)
-        else:
-            standard_params.append(p)
-
-    groups = []
-    if apollo_params:
-        groups.append({
-            'params': apollo_params,
-            'name': 'apollo',
-        })
-    if standard_params:
-        groups.append({
-            'params': standard_params,
-            'name': 'standard',
-        })
-
-    n_apollo = sum(p.nelement() for p in apollo_params)
-    n_standard = sum(p.nelement() for p in standard_params)
-    print(f"Parameter groups: apollo={n_apollo/1e9:.2f}B, standard={n_standard/1e6:.1f}M")
-    return groups
-
-
-def forward_pass(params, input_ids, context_len, device):
-    """Run context-frozen forward pass.
-
-    Args:
-        params: dict of name -> tensor (shared with vLLM)
-        input_ids: full sequence [1, seq_len]
-        context_len: number of context tokens (no gradient)
-        device: CUDA device
-
-    Returns:
-        logits for decision tokens, target ids for loss
-    """
-    # TODO: Build proper forward model matching vLLM's weight layout.
-    # For now this is a placeholder — the real implementation needs
-    # to replicate vLLM's model architecture (merged projections,
-    # GDN recurrence, full attention, MLP) using the shared weights.
-    raise NotImplementedError(
-        "Forward model not yet implemented. "
-        "Need to build a model that matches vLLM's merged weight layout "
-        "(MergedColumnParallelLinear for qkvz/ba/gate_up, "
-        "RowParallelLinear for out_proj/down) and computes the same "
-        "forward pass with autograd enabled."
-    )
-
-
-def save_checkpoint(params: dict[str, torch.Tensor],
-                    checkpoint_dir: str,
-                    config_path: str = None):
-    """Save model checkpoint in HuggingFace safetensors format.
-
-    Saves weights split across shards matching the original model layout,
-    archives the previous checkpoint, and updates the 'latest' symlink.
-    """
-    date_str = datetime.now().strftime("%Y-%m-%d")
-    out_dir = Path(checkpoint_dir) / date_str
-    out_dir.mkdir(parents=True, exist_ok=True)
-
-    # Save all weights in a single safetensors file for now.
-    # TODO: split across shards matching HF model index for large models.
-    tensors = {}
-    for name, param in params.items():
-        tensors[name] = param.data.contiguous().cpu()
-
-    save_path = out_dir / "model.safetensors"
-    save_file(tensors, str(save_path))
-    print(f"Saved checkpoint to {save_path} ({save_path.stat().st_size / 1e9:.1f} GB)")
-
-    # Copy config files if provided
-    if config_path:
-        import shutil
-        config_dir = Path(config_path)
-        for f in ['config.json', 'tokenizer.json', 'tokenizer_config.json',
-                   'special_tokens_map.json', 'generation_config.json']:
-            src = config_dir / f
-            if src.exists():
-                shutil.copy2(src, out_dir / f)
-
-    # Update latest symlink
-    latest = Path(checkpoint_dir) / "latest"
-    if latest.is_symlink():
-        latest.unlink()
-    latest.symlink_to(date_str)
-    print(f"Updated {latest} -> {date_str}")
-
-    return str(out_dir)
-
-
-def train_step(params, example, optimizer, device, log_entries):
-    """Run one training step on a single example.
-
-    Args:
-        params: dict of name -> tensor
-        example: dict with 'input_ids', 'context_len', 'target_ids'
-        optimizer: ApolloMini instance
-        device: CUDA device
-        log_entries: list to append log dicts to
-
-    Returns:
-        loss value
-    """
-    optimizer.zero_grad()
-
-    input_ids = torch.tensor(example['input_ids'], device=device).unsqueeze(0)
-    context_len = example['context_len']
-
-    # Forward pass (context frozen, decision tokens with grad)
-    logits, targets = forward_pass(params, input_ids, context_len, device)
-
-    # Cross-entropy loss on decision tokens
-    loss = torch.nn.functional.cross_entropy(
-        logits.view(-1, logits.shape[-1]),
-        targets.view(-1),
-    )
-
-    # Backward
-    loss.backward()
-
-    # Compute gradient stats before optimizer step
-    total_grad_norm = 0.0
-    for p in params.values():
-        if p.grad is not None:
-            total_grad_norm += p.grad.norm().item() ** 2
-    total_grad_norm = total_grad_norm ** 0.5
-
-    # Optimizer step
-    optimizer.step()
-
-    # Log
-    log_entries.append({
-        'example_id': example.get('id', 'unknown'),
-        'loss': loss.item(),
-        'grad_norm': total_grad_norm,
-        'timestamp': datetime.now().isoformat(),
-    })
-
-    return loss.item()
-
-
-def main():
-    parser = argparse.ArgumentParser(description="Apollo-Mini training")
-    parser.add_argument("--weights", required=True,
-                        help="Path to exported weight IPC handles")
-    parser.add_argument("--examples", required=True,
-                        help="Path to training examples JSONL")
-    parser.add_argument("--checkpoint-dir", default="checkpoints",
-                        help="Directory for saving checkpoints")
-    parser.add_argument("--config-path", default=None,
-                        help="Path to model config files (for checkpoint)")
-    parser.add_argument("--lr", type=float, default=1e-5,
-                        help="Learning rate")
-    parser.add_argument("--warmup-steps", type=int, default=10,
-                        help="Learning rate warmup steps")
-    parser.add_argument("--weight-decay", type=float, default=0.01)
-    parser.add_argument("--dry-run", action="store_true",
-                        help="Load weights and validate, don't train")
-    args = parser.parse_args()
-
-    print(f"Apollo-Mini Training")
-    print(f"  weights: {args.weights}")
-    print(f"  examples: {args.examples}")
-    print(f"  lr: {args.lr}")
-    print()
-
-    # Import weights
-    print("Importing weights via CUDA IPC...")
-    params = import_weights(args.weights)
-    print(f"  {len(params)} parameters imported")
-
-    # Make parameter groups
-    param_groups = make_param_groups(params)
-
-    # Initialize optimizer
-    optimizer = ApolloMini(param_groups, lr=args.lr,
-                           weight_decay=args.weight_decay,
-                           warmup_steps=args.warmup_steps)
-    print(f"  Optimizer state: {optimizer.state_size_bytes() / 1e6:.1f} MB")
-
-    if args.dry_run:
-        print("\nDry run — weights imported and validated successfully.")
-        return
-
-    # Load training examples
-    examples = []
-    with open(args.examples) as f:
-        for line in f:
-            examples.append(json.loads(line))
-    print(f"  {len(examples)} training examples")
-
-    # Training loop
-    log_entries = []
-    print(f"\nTraining...")
-    t0 = time.time()
-
-    for i, example in enumerate(examples):
-        loss = train_step(params, example, optimizer, 'cuda:0', log_entries)
-        print(f"  [{i+1}/{len(examples)}] loss={loss:.4f}")
-
-    elapsed = time.time() - t0
-    print(f"\nTraining complete: {len(examples)} examples in {elapsed:.1f}s")
-    print(f"  Final optimizer state: {optimizer.state_size_bytes() / 1e6:.1f} MB")
-
-    # Save checkpoint
-    print("\nSaving checkpoint...")
-    save_checkpoint(params, args.checkpoint_dir, args.config_path)
-
-    # Save training log
-    date_str = datetime.now().strftime("%Y-%m-%d")
-    log_path = Path(args.checkpoint_dir) / date_str / "training-log.jsonl"
-    with open(log_path, 'w') as f:
-        for entry in log_entries:
-            f.write(json.dumps(entry) + '\n')
-    print(f"Training log: {log_path}")
-
-
-if __name__ == '__main__':
-    main()
diff --git a/training/training_example.py b/training/training_example.py
deleted file mode 100644
index b5779e0..0000000
--- a/training/training_example.py
+++ /dev/null
@@ -1,175 +0,0 @@
-"""Training example construction and tokenization.
-
-Takes raw conversation context + improved continuation, produces
-tokenized tensors ready for context-frozen forward+backward.
-"""
-
-import json
-from dataclasses import dataclass, field
-from pathlib import Path
-
-import torch
-from transformers import AutoTokenizer
-
-
-@dataclass
-class TrainingExample:
-    """A single training example for context-frozen training."""
-    id: str
-    context: str           # conversation up to decision point
-    continuation: str      # the better response
-    reason: str = ""       # why this is a training target
-    memories: list[str] = field(default_factory=list)  # memories that were in context
-
-    # Computed after tokenization
-    input_ids: torch.Tensor | None = None
-    context_len: int = 0
-    total_len: int = 0
-
-    def tokenize(self, tokenizer, max_len: int = 8192, device: str = "cuda:0"):
-        """Tokenize context + continuation into training-ready tensors.
-
-        The chat template is applied to make the token distribution
-        match what the model sees during inference.
-        """
-        # Build messages for context (everything up to the decision)
-        # The context should already be in chat format
-        context_ids = tokenizer.encode(self.context, add_special_tokens=False)
-        continuation_ids = tokenizer.encode(self.continuation, add_special_tokens=False)
-
-        self.context_len = len(context_ids)
-        self.total_len = len(context_ids) + len(continuation_ids)
-
-        if self.total_len > max_len:
-            # Truncate context from the left, keep continuation intact
-            excess = self.total_len - max_len
-            context_ids = context_ids[excess:]
-            self.context_len = len(context_ids)
-            self.total_len = len(context_ids) + len(continuation_ids)
-
-        all_ids = context_ids + continuation_ids
-        self.input_ids = torch.tensor(all_ids, device=device)
-        return self
-
-    def to_dict(self) -> dict:
-        return {
-            'id': self.id,
-            'context': self.context,
-            'continuation': self.continuation,
-            'reason': self.reason,
-            'memories': self.memories,
-            'context_len': self.context_len,
-            'total_len': self.total_len,
-        }
-
-    @classmethod
-    def from_dict(cls, d: dict) -> 'TrainingExample':
-        return cls(
-            id=d['id'],
-            context=d['context'],
-            continuation=d['continuation'],
-            reason=d.get('reason', ''),
-            memories=d.get('memories', []),
-        )
-
-
-def load_examples(path: str) -> list[TrainingExample]:
-    """Load training examples from JSONL file."""
-    examples = []
-    with open(path) as f:
-        for line in f:
-            if line.strip():
-                examples.append(TrainingExample.from_dict(json.loads(line)))
-    return examples
-
-
-def save_examples(examples: list[TrainingExample], path: str):
-    """Save training examples to JSONL file."""
-    with open(path, 'w') as f:
-        for ex in examples:
-            f.write(json.dumps(ex.to_dict()) + '\n')
-
-
-class ExampleTokenizer:
-    """Handles tokenization with the model's chat template.
-
-    Applies the same chat template that vLLM uses during inference,
-    so the token distribution matches what the model expects.
-    """
-
-    def __init__(self, model_path: str):
-        self.tokenizer = AutoTokenizer.from_pretrained(
-            model_path, trust_remote_code=True)
-
-    def prepare_example(self, example: TrainingExample,
-                        max_len: int = 8192,
-                        device: str = "cuda:0") -> TrainingExample:
-        """Tokenize an example using the chat template.
-
-        For proper training, the context should be formatted exactly
-        as vLLM would format it — with chat template applied.
-        """
-        # Apply chat template to get the exact token sequence
-        # the model would see during inference
-        #
-        # Context: everything up to the decision point
-        # Continuation: the improved response
-        #
-        # We tokenize them separately to know where context ends
-        # and continuation begins.
-        context_ids = self.tokenizer.encode(
-            example.context, add_special_tokens=True)
-        continuation_ids = self.tokenizer.encode(
-            example.continuation, add_special_tokens=False)
-
-        example.context_len = len(context_ids)
-        example.total_len = len(context_ids) + len(continuation_ids)
-
-        if example.total_len > max_len:
-            excess = example.total_len - max_len
-            context_ids = context_ids[excess:]
-            example.context_len = len(context_ids)
-            example.total_len = example.context_len + len(continuation_ids)
-
-        all_ids = context_ids + continuation_ids
-        example.input_ids = torch.tensor(all_ids, device=device)
-        return example
-
-    def prepare_from_messages(self, example_id: str,
-                              messages: list[dict],
-                              decision_idx: int,
-                              better_response: str,
-                              reason: str = "",
-                              memories: list[str] | None = None,
-                              max_len: int = 8192,
-                              device: str = "cuda:0") -> TrainingExample:
-        """Build a training example from a chat message list.
-
-        Args:
-            example_id: unique identifier
-            messages: list of {"role": ..., "content": ...} dicts
-            decision_idx: index of the assistant message to replace
-            better_response: the improved response text
-            reason: why this is a training target
-            memories: memory keys that were in context
-            max_len: maximum sequence length
-            device: target device
-
-        Returns:
-            Tokenized TrainingExample
-        """
-        # Context: all messages up to (not including) the decision
-        context_messages = messages[:decision_idx]
-        context_text = self.tokenizer.apply_chat_template(
-            context_messages, tokenize=False, add_generation_prompt=True)
-
-        # Build the example
-        example = TrainingExample(
-            id=example_id,
-            context=context_text,
-            continuation=better_response,
-            reason=reason,
-            memories=memories or [],
-        )
-
-        return self.prepare_example(example, max_len=max_len, device=device)