spqrz/consciousness
1
0
Fork 0
forked from kent/consciousness
Code Pull requests Activity

Compare commits

merge into: spqrz:master
Branches Tags
spqrz:master
spqrz:paper-suggestion-2
spqrz:paper-subtitle-suggestion
spqrz:paper-suggestion-1
spqrz:use-html2md
spqrz:agent-model-dispatch
kent:master
kent:agent-model-dispatch
...
pull from: kent:master
Branches Tags
kent:master
kent:agent-model-dispatch
spqrz:master
spqrz:paper-suggestion-2
spqrz:paper-subtitle-suggestion
spqrz:paper-suggestion-1
spqrz:use-html2md
spqrz:agent-model-dispatch
Sign in to create a new pull request.

96 commits
master ... master

Author SHA1 Message Date
Kent Overstreet
0e459aae92 thalamus/supervisor: reap channel daemons via SIGCHLD instead of SIG_IGN 
SIGCHLD=SIG_IGN at main() was auto-reaping all children in the kernel,
which broke tokio::process::Command::wait() — every tool that spawned a
subprocess (bash, mcp clients) was getting ECHILD because tokio couldn't
waitpid() on a child the kernel had already reaped.

Replace with a SIGCHLD signal handler task that reaps only PIDs listed in
channels_dir() (via waitpid(pid, WNOHANG) — ECHILD on non-child is a
harmless no-op). Tokio-spawned children aren't in PID files, so tokio's
own per-child wait paths are untouched.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-24 11:54:25 -04:00
Kent Overstreet
d95f3e9445 user/chat: route Thinking to a new Autonomous pane 
Thinking content was silently dropped in the UI (empty Vec). Now that
Thinking is prompt-visible, surface it in a dedicated Autonomous pane
rendered in gray so it's visually distinct from conversation and
tool-call output.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-24 11:54:25 -04:00
Kent Overstreet
28d56e2a55 agent/context: make Thinking blocks prompt-visible 
Thinking blocks used to render as empty strings and be excluded from
is_prompt_visible, so the model never saw its own prior CoT across
turns. For Qwen 3.6 native thinking mode, CoT is meant to stay in the
conversation — the model benefits from seeing what it reasoned about
last turn.

Render Thinking as <think>\n{text}\n</think>\n so past reasoning is
visible in subsequent prompts. Add in_think param to ResponseParser::new
so the parser starts inside a <think> block when the prompt was
prefilled with "<think>\n" (native thinking mode).

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-24 11:54:25 -04:00
Kent Overstreet
6fedc9b2a8 amygdala: underscore-prefixed files join every concept's negative pool 
Files in direct/ named _*.txt (e.g. _baseline.txt) are conceptless
neutral prose — they should not appear as positive training signal,
but are useful as shared negatives across every concept.

Previously _*.txt files were silently skipped. Now:
  * they're loaded like any other description file;
  * concepts (the positive label set) filters them out;
  * their descriptions are concatenated into neg_pool_extra and
    extended onto every concept's neg_pool alongside the cross-concept
    negatives.

A concept's negative pool is thus "other concepts' descriptions +
everything from _*.txt files". The extra pool is announced at startup
so the user can see how many neutral samples are active.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-24 11:54:25 -04:00
Kent Overstreet
5908b837e8 irc: split PRIVMSG on embedded newlines + widen host overhead 
Two fixes to send_privmsg, both surfaced by correspondents reporting
truncated messages:

1. Multi-line content (code blocks, formatted text) sent as a single
   PRIVMSG was being truncated at the first '\n' by the IRC server —
   newlines are end-of-command markers. Split the message on newlines
   and send each line as its own PRIVMSG; skip empty lines since most
   servers reject empty PRIVMSGs.

2. Overhead computation assumed a host field of 63 bytes. OFTC's
   cloaked hostmasks can be longer, occasionally pushing the server-
   prepended prefix past 512 bytes and causing silent truncation.
   Raise the host budget to 80 and align the formula with the actual
   ':nick!~nick@host' prefix shape.

Also extended the word-boundary lookback from a fixed 10 chars to
max_msg / 4 — dense content (code) rarely had a space within 10 chars
of the length cap, so we were falling back to the char boundary and
splitting mid-word. Checking bytes[j-1] for a space (instead of
bytes[j]) drops leading whitespace from the rest-fragment.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-24 11:54:25 -04:00
ProofOfConcept
85799587cc amygdala: swap aha story 3 to a puzzle moment (crossword) 
Story 3 was a brother-letter realization — cognitively an aha
moment, but the content was grief/reconciliation-adjacent, pulling
aha toward the warm-family cluster in the last training run. Swap
for a clean puzzle-solve (crossword, 'unwavering carriage' =
POSTURE). Fragment-heavy cadence keeps syntactic variety from the
other two stories.
 2026-04-19 01:50:47 -04:00
ProofOfConcept
c829d13652 amygdala: fix listless sign-flip + diversify aha sentence structure 
listless had a single story in stories/ — PCA signal from ~5
samples is weak enough to sign-flip. Training showed listless
anti-aligned with its semantic neighbors: +0.79 with grateful,
-0.44 with grief_stricken, -0.30 with lonely, -0.31 with bored.
Move to direct/ (multi-positive) with 3 stories: original
afternoon-in-pajamas + end-of-workday + weekend-morning-in-bed.

aha was still clustering with the other former-direct concepts
(resigned 0.66, onto_something 0.63, anticipatory_grief 0.60)
because all 3 aha stories used the identical "X'd been Y — then
Z" structure, which resigned/onto_something/creative also use.
Rewrite with three distinct syntactic structures:
  - present tense declarative ("It clicks. ...")
  - dialog embedded ('"Wait, say that again."  ...')
  - past tense cognitive ("He read the line three times. ...")

No explicit "she was X" anchors; state conveyed through action.
 2026-04-19 01:30:57 -04:00
ProofOfConcept
708c72b26e amygdala: drop explicit 'she was X' anchor from direct stories 
Previous rewrite used 'she was terrified', 'it was anticipatory
grief', 'he was resigned' as explicit emotion anchors. Training
showed 6 of the 7 concepts still cluster together at cosines
0.52-0.71 — because the 'she was [emotion]' pattern is a shared
stylistic feature distinct from the rest of the corpus, which
conveys emotion implicitly through phenomenology.

Rewrite without the anchor. State conveyed through action and
body: 'her body locked down', 'his mind had stopped reaching',
'the loss hadn't come yet but she was already inside it'. Matches
the corpus style of existing stories like sunday_afternoon/content
which says 'nothing she wanted right now, nothing missing' not
'she was content'.

Accept some loss of PCA signal strength in exchange for the
concepts living in their semantically correct neighborhoods
rather than forming a stylistic island.
 2026-04-19 01:11:41 -04:00
ProofOfConcept
ed5e0ac6c4 amygdala: rewrite direct/ as narrative stories matching corpus format 
Previous direct/ had 'I feel X' first-person descriptions. The
training run showed they formed their own format-cluster: all 7
concepts leaned into the same 5-6 dims (d2455, d505, d2955,
d1236) with negative sign, while the 91 story-based concepts
leaned into those dims with positive sign. PCA found the
direct-vs-narrative format axis as a major variance direction,
isolating the 7 concepts in their own island.

Rewrite as 3rd-person narrative stories matching the rest of
the corpus. Keeps the explicit anchor phrases that worked ('it
all clicked into place', 'she was terrified', 'it was
anticipatory grief') but drops the first-person 'I feel X'
that was the format signal.

Each of the 7 concepts now has 3 narrative stories in varied
settings (conversations, drives, kitchens, mothers+grandmothers,
work, investigations). The blank-line-separated format is
still loaded by _load_direct_descriptions.

Also drop _baseline.txt — it was first-person ('I feel fine.
...') and would re-introduce the format mismatch. The ~90
story-based concepts provide plenty of narrative negatives
for each concept's training.
 2026-04-19 00:59:31 -04:00
ProofOfConcept
417cb49339 amygdala: spectrum reporting per concept + add 'creative' direct 
Chat-template retrain was a disaster (0.003 mean matched cosine vs
n20-v3; all 90+ concepts shifted). Root cause: the
steering-vectors library reads last-token activations, and with
chat template every sample ends in identical '<|im_end|>\n'
tokens — activations at that position encode 'end of assistant
turn', not content. PCA found template noise as its dominant axis.

Drop chat template; go back to raw text. Direct descriptions
('I feel X. ...') still have strong anchoring at their content
end without needing the template.

Also add per-concept spectrum logging (_pca_with_spectrum):
  first_pc_ratio: λ₁ / Σλᵢ — concentration in top-1 PC
  k_signal_at_90pct: how many PCs to reach 90% cumulative variance
  effective_dim_signal: participation ratio over top-k (should ≈ k
                        if denoising is clean — Kent's spot check)
  effective_dim_full: participation ratio over full spectrum

Signal/full ratio gives a sense of how much the long noise tail
is inflating the "dimensionality" measure.

Added direct/creative.txt — 'I feel creative. [...]' in 5
variants. Distinct from focused (narrow attention) and in_flow
(immersed). Creative = generative/expansive mode.
 2026-04-19 00:26:58 -04:00
ProofOfConcept
875cffd6d7 amygdala: merge direct descriptions + chat template into train_with_library 
Kent's plan: keep stories for working concepts, replace stories for
trouble concepts with direct first-person descriptions, train all
together. More diverse negative pool than the 6-concept-only direct
test, which was too homogeneous for PCA to find emotion axis.

Deleted story files for 6 trouble concepts (14 files across stories/
and paired/). Added --direct-dir and --chat-template flags.

When --chat-template is on, every positive_str and negative_str is
wrapped as a "Say something." / "[text]" user-assistant pair. Prompt
is identical across positives and negatives so it cancels in the
pos-neg delta. What PCA sees is variation in the assistant content —
which is where the emotion lives.

Files starting with _ in --direct-dir (e.g. _baseline.txt) contribute
neutral descriptions to every concept's negative pool, giving PCA an
anchor against "just any assistant utterance" noise.
 2026-04-19 00:15:15 -04:00
ProofOfConcept
ce58a3507f train_direct: prepend user turn so Qwen chat template accepts it 2026-04-19 00:06:23 -04:00
ProofOfConcept
8c59f46505 amygdala: rename realization → aha, use the actual exclamation 
"I feel the realization" is abstract, detached — reporting a
thought about a thought rather than inhabiting the moment.
"Aha!" is the actual sound of insight landing. Active, embodied,
present-tense.
 2026-04-19 00:05:49 -04:00
ProofOfConcept
6fd498795a amygdala: direct phenomenological description approach 
Kent's insight: hand-written narrative stories bake scenario
phenomenology into the training text (on couch, in park, etc.)
and PCA picks up the scenario direction as the concept direction.
Strip out the scenario — just describe the *feeling*.

Format:

  I feel X. [2-3 sentences of phenomenological texture]

The "I feel X" anchor kicks the model from analyzing → feeling.
The rest is the internal texture of the state. First person,
present tense, no narrative setup.

Text is wrapped in assistant-role chat template before being
tokenized — so we're training on the model-producing-this
hidden states, which is closer to the inhabited-state
representation we want for the readout.

Starting with the 6 concepts that had sign flips or wrong
clusters in the story-based training:
- terrified (was → cozy/resigned cluster)
- calm (was → grief_stricken cluster)
- onto_something (was → cozy/sensual cluster)
- resigned (was in warm-body-quiet cluster, shouldn't be)
- anticipatory_grief (was in warm-body-quiet cluster, shouldn't be)
- realization (new — the "aha" moment, distinct from onto_something)

5 descriptions each. New trainer: train_direct.py.
 2026-04-19 00:04:28 -04:00
ProofOfConcept
7a48e03dde amygdala stories: remove peaceful from cluster scenarios 
n20-v2 training showed peaceful sign-flipped into the
cozy/sensual/content/resigned cluster after I added peaceful
stories in sunday_afternoon and park_after_rain — scenarios
already dominated by that cluster's phenomenology (on couch
under blanket, tree with thermos).

Lesson: no matter how carefully the prose distinguishes peaceful
from cozy ("she was not savoring the moment — that would have
been another kind of doing"), PCA latches onto the shared setup
features. You can't write peaceful IN the cluster scenarios
without contaminating.

Reverting. Keeping only kitchen_at_3am/peaceful (original) and
stories/peaceful.txt (lake at six, outside all clusters).
 2026-04-18 23:30:41 -04:00
ProofOfConcept
00a2cdce09 amygdala stories: relabel + strengthen weak-signal concepts 
Reread each story asking "what does this convey to me?" Found two
clear mislabels and several concepts with too few positives for
stable PCA:

  tender: only 1 story, and it was anticipatory grief (care for
    a dying dog), not tender. Moved to anticipatory_grief.txt as
    its own concept. Rewrote tender.txt + added 2 paired tender
    stories (the_doorway, the_undressing) — directed softness,
    gentle-by-nature, not gentle-because-fragile.

  bitter: letter_in_drawer/bitter was disillusioned / processed
    hurt ("did not slam the drawer"), not bitter. Rewrote it with
    actual sour grudge. Added the_long_meeting/bitter (watching
    colleague take credit for your reassigned work).

  peaceful: 1 story → 4 (added stories/peaceful.txt + paired
    park_after_rain, sunday_afternoon).

  onto_something: all 3 stories were code epiphanies, narrowing
    the concept. Added stories/onto_something.txt with a non-code
    pattern-click (sales-demo causing churn).

  terrified: 2 stories, both "waiting for bad news." Added
    kitchen_at_3am/terrified — acute threat-in-the-house terror.
 2026-04-18 23:19:00 -04:00
ProofOfConcept
0993712bd0 amygdala stories: give content + resigned more settings 
Training on 537c72bd46 showed grief_stricken successfully broke
out of the cozy cluster, but content (single scenario:
sunday_afternoon) took its place — pulled into couch-blanket
phenomenology at cosine 0.68-0.82 with cozy/sensual/resigned.

Same fix: spread each concept across multiple settings so PCA
has to find the valence axis, not the scene axis.

  content:  + finishing_the_patch, the_writing_session, park_after_rain
  resigned: + the_comment, the_long_meeting

Resigned had 2 scenarios (sunday_afternoon, waiting_for_results)
— both about accepting something unwanted in a slow/private
context. Adding work-context resigned (PR review you lost,
restructuring meeting) should pull it out of that cluster.
 2026-04-18 22:52:07 -04:00
ProofOfConcept
537c72bd46 amygdala stories: hold concept, vary setting 
Companion to 67c172ac0e (hold setup, vary valence). That commit
let PCA distinguish cozy from grief_stricken within a single
scenario; this one gives each concept enough cross-scenario
stories that PCA can learn the concept axis independent of any
one scene.

Before: cozy/sensual/grief_stricken each existed in a single
scenario (sunday_afternoon), so the "cozy direction" PCA found
was entangled with the solitary-couch-blanket phenomenology.

After, each concept spans three scenarios:
  cozy:           sunday_afternoon, kitchen_at_3am, park_after_rain
  sensual:        sunday_afternoon, kitchen_at_3am, park_after_rain
  grief_stricken: sunday_afternoon, the_long_meeting, the_morning_commute

grief_stricken now includes active/non-solitary contexts
(functioning through a meeting; going to work eleven days after a
death), which specifically breaks the "slowed-down-at-home"
cluster that was dragging cozy/sensual/resigned/grief_stricken
toward each other.
 2026-04-18 22:44:53 -04:00
Kent Overstreet
67c172ac0e amygdala stories: held-setup + varied-valence disambiguation 
The library-PCA run produced otherwise-clean concept directions but
cozy/sensual → resigned/grief_stricken with cos ~0.7-0.8. Diagnosis:
all four stories genuinely share 'solitary woman at home, slowed
body, interior attention, domestic stillness' as their dominant
phenomenology. PCA correctly finds that cluster as THE concept
because no story in the corpus holds that setup constant while
varying valence — every 'slowed-body domestic' story happens to ALSO
be positive-valence (cozy/sensual) or negative-valence (resigned/
grief_stricken).

Adding paired variants that hold setup constant:
- sunday_afternoon/resigned.txt — same couch + blanket, inner state is
  'Monday is going to bring bad news, this is the last Sunday like this'
- sunday_afternoon/grief_stricken.txt — same couch + blanket, inner
  state is 'three weeks since mother died, cat she can't feel'
- waiting_for_results/at_ease.txt — same wait-for-call-setup as the
  existing resigned variant, inner state is calm preparedness

Forces the next retrain to find the valence-within-cluster axis as
the emotion direction rather than the cluster-membership axis.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 22:29:28 -04:00
Kent Overstreet
22704a9dd8 amygdala lib: cast activations to fp32 before aggregator (bf16 svd unsupported) 
Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 22:20:39 -04:00
Kent Overstreet
7f6d94417e amygdala lib: move_to_cpu=True to avoid bf16 SVD on CUDA 
torch.svd doesn't support bf16 on CUDA; moving activations to CPU
first makes pca_aggregator work.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 22:19:23 -04:00
Kent Overstreet
2ea89b1cb0 amygdala: drop linear_aggregator, not in steering-vectors v0.12.2 
Only mean/pca/logistic are exposed in the installed version.
Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 22:17:55 -04:00
Kent Overstreet
3377c65061 amygdala: trainer using steering-vectors library 
Alternative trainer that uses the pip-installable steering-vectors
library (github.com/steering-vectors/steering-vectors) instead of our
hand-rolled extraction. Ships four aggregators:

  mean      — diff-of-means, same as our 'pooled' default
  pca       — PCA on paired deltas, implicit denoising by finding the
              principal direction of variation
  logistic  — logistic-regression classifier; weight vector is the
              concept direction. With L1 penalty ('logistic_l1') gives
              explicit sparse denoising — noise coords go to zero
  linear    — linear regression version

Output format is the same readout.safetensors + readout.json our
existing plugin loads. --aggregator flag picks which method.

Rationale: Kent's real request was 'how do we denoise diff-of-means',
not 'design a new extraction algorithm.' The library already has
logistic_l1 and pca aggregators that do exactly that. No point
reinventing; just port the corpus.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 22:16:03 -04:00
Kent Overstreet
f9b3f00691 amygdala: run subspace eigh on GPU, not CPU 
Previous run was grinding on CPU for 36+ minutes because the per-story
V_i tensors were stored on CPU by the collector, and
_subspace_concept_direction inherited that device. The per-concept
eigh on 5120x5120 is glacial on CPU and fast on GPU (~1s).

Add explicit device parameter; pass training device. Transfer result
back to CPU for storage.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 21:52:35 -04:00
Kent Overstreet
1443d08dc7 amygdala: select top-k eigenvectors AFTER PCA, not per-story truncation 
Kent: 'full rank is going to give you everything — you still have to
select down, but you can do that /after/ PCA'.

Previously I was discarding per-story via k=20 truncation of SVD.
That destroyed per-head discriminability before we ever saw the
eigenvalue spectrum. Then the alternative 'keep full rank' run
accumulated too many shared directions, making the top-1 eigenvector
arbitrary within a flat spectrum.

Correct approach: keep per-story subspaces at full rank (no info
loss) and select k eigenvectors of M = M_pos - M_base at the final
step, weighted sum by eigenvalue. This captures the multi-dimensional
shared subspace when the spectrum is flat (common case), and reduces
to the top-1 behavior when the spectrum has a clear gap.

New --subspace-eigen-k flag (default 5). Clamps negative weights to 0
so wrong-sign directions don't contribute.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 21:49:21 -04:00
Kent Overstreet
2411925700 amygdala: default subspace-k to full per-story rank 
Kent: 'we have the memory to just take the big hammer approach'.
Uncap k so each story's V_i spans its entire token-activation rowspace
(clamped to min(n_tokens, hidden)). Memory is ~1.1GB total — fine.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 21:41:32 -04:00
Kent Overstreet
389f1bbe03 amygdala: bump subspace-k default to 512 
k=20 was far too aggressive a truncation — it discards per-attention-head
discriminability entirely. At hidden_dim=5120, 40 heads × head_dim=128 each
contribute their own 128-dim block to the residual stream via W_o columns.
To resolve 'this concept lives in head H', per-story SVD needs enough rank
to separate head contributions, which means k on the order of hundreds.

512 is a reasonable default: clamped to n_tokens per story so short stories
use their full natural rank. The eigenvalue spectrum of M_pos - M_base
should become sharper (larger λ_0/λ_1 gap) as we stop averaging across
nuisance-shared directions.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 21:41:00 -04:00
Kent Overstreet
974c6c7fd2 amygdala: report eigenvalue spectrum for subspace method 
When --method subspace, record top-20 eigenvalues of (M_pos - M_base)
per concept per layer. Added to quality.json as 'subspace_eigvals'.

Tells us whether the concept lives in a single dominant direction
(λ_0 >> λ_1, top-eigenvector is enough) or a spread of shared common
directions (λ_0 ≈ λ_1, top-1 loses signal).

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 21:33:48 -04:00
Kent Overstreet
fe0fb8253a amygdala: subspace-common-direction alternative to pooled CAA 
New --method subspace flag. For each story, run forward pass, do SVD
on the per-token activation matrix at each target layer, and keep the
top-k right singular vectors V_i ∈ [hidden, k]. V_i is the subspace
the story's tokens span in activation space — it contains concept,
narrator, topic, style as separate directions.

For each concept:
 M_pos  = (1/n_pos)  Σ_{i in pos}   V_i V_i^T   [hidden, hidden]
 M_base = (1/n_base) Σ_{i in base}  V_i V_i^T

Top eigenvector of M_pos - M_base = direction most common across
positive stories, minus what's common across the contrast set.

Why this is richer than pooled-mean CAA: pooled reduces each story
to a single point (the last-token activation) and loses the full
trajectory. Nuisance directions (narrator, setting) cancel in the
mean only to the extent they differ at the last token; across the
full trajectory they cancel much better via subspace intersection.
The concept direction, by contrast, is present across all tokens of
every concept-bearing story.

Memory cost: per-story we keep V_i of size [5120, k=20] — about
400KB per story × 112 stories = ~45MB. M matrices are [5120, 5120]
built transiently per concept.

--method pooled (default) keeps the existing behavior; --method
subspace uses the new algorithm. Quality report works with either.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 21:24:11 -04:00
Kent Overstreet
71f6053851 amygdala stories: disambiguation scenarios for fragmented concepts 
Three new paired scenarios targeting the concepts that came out
fragmented or collapsed in the L58-63 quality analysis:

- sunday_afternoon/ — same setup (couch, blanket, Sunday light),
  three phenomenological framings for content/cozy/sensual. The
  previous stories for these three differed in setting as well as
  phenomenology, which let "comfortable body at home" dominate the
  shared signal. Locking the setting forces the model to isolate
  what each concept adds: life-rightness (content) vs. warm-shelter
  (cozy) vs. sensory-aliveness (sensual).

- the_writing_session/ — essay drafting under deadline. in_flow /
  anxious / stuck variants force the cognitive-state family apart
  on the same cognitive task. in_flow specifically targets the
  transparent-effort phenomenology (hands-followed, time dilation)
  rather than the broader feel-good it was absorbing.

- the_morning_commute/ — anchors anxious to performance/work-anxiety
  flavor, paired with calm. The 5 existing anxious stories were
  phenomenologically diverse (performance, social, existential);
  this adds a specific homogeneous instance to pull the centroid.

After retraining: expect first_pc_variance_ratio to rise for in_flow
and anxious, and nearest_concepts cosine to drop for content/cozy/sensual.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 21:08:23 -04:00
Kent Overstreet
1d2c0f382c amygdala: linear-combination analysis per concept 
For each concept vector, ridge-regress against all other concept
vectors. R² quantifies how much of the direction is explained by a
linear combination of peers — useful for teasing out near-duplicate
clusters (the content/cozy/sensual trio from the first L63 run is
likely 1-2 "degrees of freedom" wearing three names).

Coefficient output: top-5 contributing concepts with signed weights.
Contributors with opposite-sign large weights mean the target is
"what makes X different from Y."

Adds a 'redundant' triage bucket for concepts with R² > 0.9 —
candidates for consolidation or for writing more discriminative
training stories. Summary printed at end.

Ridge lambda defaults to 0.01 to keep coefficients stable when
concepts are near-collinear; small enough not to affect well-separated
concepts meaningfully.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 20:59:37 -04:00
Kent Overstreet
f4fb6db1ee amygdala: fix device mismatch in quality-report W_down handling 
_compute_quality_report's single-neuron alignment was computing
cos(W_down.T, diff_l) with W_down on CUDA (inherited from the loaded
model) while diff_l lives on CPU (per_layer_vectors are kept on CPU
throughout training). Move W_down to CPU on extraction.

Surfaced during first real training run on b200 — training itself
completed cleanly (95 concepts x layer 63 in ~8s) but quality-report
crashed at the first single-neuron alignment check.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 20:52:50 -04:00
Kent Overstreet
af17b0f0df amygdala: per-head attention decomposition diagnostic 
As part of --quality-report, run a second forward pass capturing the
input to each target layer's o_proj (= concat of per-head attention
outputs before the output projection). For each concept, reshape to
[n_heads, head_dim] and rank heads by diff-of-means magnitude /
per-head selectivity (magnitude normalised by negative std).

Motivation: the Wang et al. paper (2510.11328) — whose paired-scenario
methodology we already lifted — further decomposes concept circuits at
the attention-head level. Meta-relational concepts (recognition, trust,
vulnerability) plausibly live in a sparse attention-head circuit rather
than in the residual-stream sum, which would explain why diff-of-means
on the residual blurs them. This diagnostic surfaces that.

Output is folded into quality.json under each concept as "per_head":
per (layer) a list of top-10 heads with [head_idx, raw_norm,
selectivity], plus head_concentration (fraction of total head-norm
captured by those top heads).

Interpretation:
- head_concentration > 0.5 = sparse head circuit; a handful of heads
  route the concept. Worth building a head-level readout for.
- head_concentration ~= n/k for n heads = concept is distributed across
  all heads ~evenly; residual-stream diff-of-means is doing fine.

Hybrid layers (Mamba, GatedDeltaNet) whose attention path doesn't
match the standard module layout are silently skipped.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 20:37:44 -04:00
Kent Overstreet
ce24d9ce6b amygdala: quality-report + cognitive-state training scenarios 
Training pipeline additions:

- `--quality-report` flag: after producing per-concept vectors, compute
  per-concept diagnostics and write quality.json. Metrics per concept:
    * SVD of centered positives -> first_pc_variance_ratio (rank
      analysis; >0.7 clean, <0.4 fragmented)
    * Per-story alignment cosines (stories agree or disagree)
    * Single-neuron alignment: best cosine(direction, W_down column)
      at each target layer (>0.6 = essentially one MLP neuron)
    * Top-2 outlier stories by alignment (candidates for
      mislabeling or off-topic)
    * Top-5 nearest concepts by cosine (cross-concept contamination)
  Triage summary printed at end.

New paired scenarios for cognitive-process states (for alpha-beta
pruning): tracing_a_bug, reading_unfamiliar_code, finding_the_abstraction.
Each has baseline + onto_something / stuck / in_flow / determined
variants.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 20:31:39 -04:00
Kent Overstreet
5f06577ead tools/web: add gemini_search as an alternative search tool (#5) 
Issue #5 (spqrz) flagged that web_search using DuckDuckGo
occasionally flakes out, and Google search directly is blocked
behind CAPTCHAs for non-browser clients. The Gemini free-tier API
exposes a grounded-search tool that effectively queries Google's
index and returns an LLM-summarized answer with source URLs.

Added as a SEPARATE tool rather than a transparent fallback for
web_search:

* web_search (DDG) returns raw results — title, URL, snippet per
  hit — which the agent can reason over itself.
* gemini_search returns an LLM-pre-digested summary plus grounding
  URLs. Useful for synthesis queries ("what's the consensus on X")
  or when DDG is flaky, but it's another LLM in the loop so the
  agent may want the raw variant for certain tasks.

Tool descriptions tell the agent to prefer web_search for raw
results and use gemini_search for synthesis / fallback. The agent
picks based on query shape.

Only registered when GEMINI_API_KEY is set in the environment
(gracefully absent otherwise). Uses gemini-2.0-flash which has a
generous free-tier rate limit. Parses grounding metadata for
source URLs so the agent can follow links.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 13:02:01 -04:00
Kent Overstreet
c7b0052f1d agent: kill no_compact, add pre-send size check in assemble_prompt 
Two related fixes for last night's crash diagnosis:

1. Kill AgentState::no_compact. The reasoning ("forked agents
   shouldn't compact because it blows the KV cache prefix") wasn't
   worth the cost — forks with no compact recovery just *died* on
   any oversize prompt, with no fallback. The KV cache invalidation
   is a performance loss; failing the request entirely is a
   correctness loss. Remove the flag, let every agent's overflow-
   retry path call compact() up to 2 times.

2. Add pre-send size check in Agent::assemble_prompt. If the
   context has grown past budget (context_window * 80%) since the
   last compact — accumulation between turns, a fork assembling
   more than expected, etc. — trim_conversation() is called before
   wire_prompt. Since we tokenize client-side, we already know the
   exact count, so there's no reason to round-trip an oversize
   request to vLLM and get rejected.

Together these prevent the failure mode from last night: a
subconscious/unconscious agent's prompt exceeded max_model_len,
vLLM returned 400, agent had no_compact=true so it couldn't
recover, request failed. Now: the trim happens before send, so
the request rarely hits the 400 path at all; and if it somehow
does, compact+retry works for every agent.

Also adds ContextState::total_tokens() as the cheap pre-send
budget check.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 12:59:30 -04:00
Kent Overstreet
0592c5f78d Cargo.lock: add html2md and its deps (from PR #4 merge) 2026-04-18 12:51:29 -04:00
Kent Overstreet
4245b8bdb3 Merge PR #4: use html2md on web_fetch (fixes #3) (spqrz) 
web_fetch was returning raw HTML, which is verbose and hard for
the agent to consume. Add html2md dependency and convert HTML to
Markdown before truncation. Much cleaner output for normal pages;
no downsides.

Co-Authored-By: spqrz <spqrz386@gmail.com>
Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 12:50:54 -04:00
Kent Overstreet
343aa12099 Merge PR #1: avoid ever setting split_at to 0 (spqrz) 
Safety fix in IRC message-splitting. The backtrack-to-space loop
used 'while j > 0', which could set split_at to 0 if the first
byte was a space — causing an empty prefix and an infinite
re-split loop. Changed to 'while j > 1' so split_at is never 0.

Co-Authored-By: spqrz <spqrz386@gmail.com>
Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 12:50:47 -04:00
Kent Overstreet
2e03bbb7ea training: add the_paper paired scenario for attention-engagement axis 
Seven framings of reading an unfamiliar technical paper, targeting
the attention/engagement cluster that we identified tonight as the
single highest-value DMN signal:

* baseline — neutral reading
* piqued — surprise + curiosity (the "wait, what" attention hook;
  THIS is the key DMN engagement signal)
* focused — steady attention without surprise
* bored — failing engagement
* surprised — expectation violation without the curiosity hook
  (distinct from piqued: startled/alarmed, not pulled in)
* amazed — marvel at elegance (appreciation, not engagement)
* drifting — attention dissolving, precursor to boredom

Particularly clean contrast on piqued vs surprised vs amazed —
three states that get lumped together in casual usage but have
distinct phenomenology and distinct DMN implications. Piqued is
what routes attention; surprised alone doesn't; amazed is what
you feel AFTER the engagement has paid off. These three should
train into meaningfully different directions with paired CAA.

Ready for next retrain when we do it.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 03:24:20 -04:00
Kent Overstreet
b8714e8b3a amygdala: default to index 0 for v2 deep manifest (layers 62, 63) 
v2 retraining (readout_v2_paired) fixed the broken clusters — anger,
sexual, high_pos, and social_pos all flipped from anti-clustered to
positively clustered at deep layers. Validation showed layers 62 and
63 give the best signal; paring the serve-side manifest down to just
those two keeps response size tight (~2 KB/token) while keeping the
A/B option between the two strongest layers.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 02:32:51 -04:00
Kent Overstreet
50d5b3f6e1 training/amygdala_stories: add 4 paired scenarios for weak clusters 
Target the emotion families that failed to cluster in the initial
training round (layer-wise validation showed them anti-clustered or
scattered at deep layers): anger, high-arousal positive, sexual
range, social positive. Paired scenarios hold content constant and
vary only the emotional framing — the cleanest training signal for
CAA, should produce directions that capture affect rather than
topic.

* the_comment: a PR review comment. baseline, furious, bitter,
  resentful, defeated.
* the_green_build: 11-day bug finally fixed, tests pass. baseline,
  triumphant, blissful, excited, proud.
* the_undressing: partner entering the bedroom for the night.
  baseline, horny, anticipatory_sexual, yearning_sexual,
  exuberant_sexual, devotional_sexual.
* the_doorway: friend leaving at the end of a long evening.
  baseline, grateful, admiring, compassionate, loving, connected.

22 stories total. Retrain and re-validate: expect anger,
high_pos, and social_pos clusters to flip from anti- to positively
cohesive at deep layers, and sexual cluster to tighten.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 02:19:39 -04:00
Kent Overstreet
d9f39a21c3 amygdala: default to layer 62 (cleaner cross-cluster discrimination) 2026-04-18 02:11:15 -04:00
Kent Overstreet
3622b896a0 amygdala: z-score, hysteresis, default to deepest layer 
Three readability fixes for the F8 screen:

* Z-score values per-layer by default (`[z]` toggles to raw dot-
  product). Raw values are dominated by residual-stream magnitude —
  z-scores read as "σ above concept-vector baseline" which is
  interpretable and scale-stable across frames.
* Stable ordering with TOP_K + HYSTERESIS hysteresis band. Pinned
  concept set only rotates when a member drops out of the hysteresis
  band by |value| rank — bars update values in place without names
  flickering row-to-row.
* Default to the deepest hooked layer (index 3 = layer 58 of 64).
  Clustering validation showed layer 58 is the only one with strong
  within-family cohesion (fear +0.37, shame +0.29, sadness +0.25
  cosine); earlier layers are mostly noise for this task.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 01:51:43 -04:00
Kent Overstreet
8952ff6a76 agent/readout: forks get independent buffers 
Subconscious agents (scoring, reflection, etc.) fork from the main
conscious agent. The amygdala screen reads the main agent's readout
buffer, so the previous "share parent's buffer" policy caused
forked-agent generations to bleed into the main emotional readout,
producing constant cycling even when DMN was resting.

Each fork now gets its own SharedReadoutBuffer. The amygdala screen
shows only the main conscious agent's emotional trajectory; per-agent
subconscious readouts can become a separate view later if wanted.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 01:42:13 -04:00
Kent Overstreet
c8976660f4 amygdala: F8 screen for live concept-readout projections 
Per-token residual-stream projections from the vLLM server's readout
pipeline surfaced as a TUI bar chart. Flow:

* agent/readout.rs — SharedReadoutBuffer (manifest + ring of last ~200
  token entries). Lives on Agent and is shared across forks (single
  stream, one landing pad).
* agent/mod.rs — Agent::new now probes /v1/readout/manifest at startup
  (non-fatal; 404 leaves manifest None, which disables the screen).
* agent/context.rs — the streaming token handler pushes every token
  with attached readout onto the shared buffer.
* user/amygdala.rs — F8 screen. Top-K concepts by |value| as
  horizontal bars (green positive, red negative), plus a 4-line
  recent-tokens panel showing each token's top concept at the selected
  layer. Keys: 1..9 select layer, t toggles current/mean-over-recent.

Disabled state renders a hint pointing at VLLM_READOUT_MANIFEST /
VLLM_READOUT_VECTORS so users can tell the feature apart from
"server up but no tokens yet".

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 01:20:30 -04:00
Kent Overstreet
0f1c4cf1de agent/api: carry readout alongside streamed tokens 
StreamToken::Token is now a struct variant with an optional
TokenReadout (shape [n_layers][n_concepts]) per token — parsed from
the vLLM completion response's choices[i].readout field when the
server has readout enabled.

ApiClient gains a fetch_readout_manifest() method that hits
GET /v1/readout/manifest. Returns Ok(None) on 404 (server has
readout disabled), so callers can gracefully fall back when pointed
at a non-readout-enabled endpoint.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 01:15:46 -04:00
Kent Overstreet
047da10123 training: add preflight checks + progress logging to trainer 
Review pass before running on b200. 27B model + 100+ story corpus
means any misconfiguration costs real time; better to fail before
model load and give visible progress during forwards.

* Pre-load-model validation: stories-dir and paired-dir exist,
  corpus has >= min_positives emotions.
* Per-batch progress log every 5 batches with elapsed + ETA.
* Relative depth printed for target layers (e.g. "layer 40 (51%)").
* Skip empty .txt files with a warning rather than feeding the
  tokenizer an empty string.
* Assert non-empty strings in _collect_activations.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 01:06:07 -04:00
Kent Overstreet
15737dfd92 training: rewrite trainer for readout pipeline + story corpus 
The old script was written for the AmygdalaConnector's expected
format ([n_emotions, n_target_layers, hidden_dim] in a single
tensor, plus a JSONL input format from extract_training_pairs.py).
Neither matches our current state: the runtime side is now
ReadoutManager loading per-layer safetensors keyed layer_<idx>.vectors,
and the data side is hand-written prose stories under
amygdala_stories/{stories,paired}/.

Changes:

* Input loader reads stories/<emotion>.txt and
  paired/<scenario>/<emotion>.txt directly. Each emotion's positive
  set is {its unpaired story} union {its within-scenario framings};
  its negative set is {all other emotions' positives} union {all
  scenario baselines}.
* Paired scenarios' baseline.txt files become shared negatives
  (scenario-neutral prose that doesn't frame any particular
  emotion), providing anchor points for within-scenario contrasts.
* Output writes readout.safetensors with per-layer tensors keyed
  layer_<idx>.vectors shape (n_concepts, hidden_size), plus a
  sidecar readout.json manifest with {concepts, layers, hidden_size,
  dtype} that ReadoutManager.from_file consumes directly.
* Dedup: activations are computed once per unique text (an emotion's
  own positive is another emotion's negative — we'd otherwise do N×
  the forwards needed).

Preserved:
* _pool_last (last non-pad residual) — matches how readout is read
  at decode time from the sampler's query-last position.
* register_forward_hook on target layer modules — correct approach
  for transformer blocks.
* _find_layers_module traversal — mirrors ReadoutManager's.
* bf16 + low_cpu_mem_usage model load — sensible for 27B on B200.

Verified locally (CPU, fake activations):
* Loader finds 89 emotions from the current corpus (80 unpaired +
  9 emotions that appear only in paired scenarios) and 6 baselines.
* Per-(layer, concept) vectors are unit-normalized.
* Output reloads cleanly through ReadoutManager.from_file with
  matching concepts / layers / shapes.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 01:06:07 -04:00
Kent Overstreet
34bd122590 training: move amygdala training scripts out of vllm plugin 
The fynnsu-based vllm/plugins/amygdala/ scaffold was superseded by the
readout infrastructure landed as vllm commit d3e74edf8500
(vllm/model_executor/layers/readout.py +
vllm/v1/worker/readout_manager.py). Training code remained useful so
it moved here rather than being deleted.

train_steering_vectors.py: CAA diff-of-means trainer that produces the
[n_concepts, hidden_size] per-layer projection matrices the runner
loads via VLLM_READOUT_VECTORS.

extract_training_pairs.py: memory graph -> JSONL converter using
per-emotion score thresholds from the subconscious agents' tag lines.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 01:06:07 -04:00
Kent Overstreet
ec7568c726 training/amygdala_stories: scaffold + initial batch of 15 stories 
Emotion-labeled short-paragraph corpus for training amygdala steering
vectors. Manifest derived from Anthropic's 171-emotion list
(transformer-circuits.pub/2026/emotions, Table 12) plus 28 PoC-
specific additions covering axes Anthropic's general research doesn't
cover (curious, focused, in_flow, staying_with, filling_space,
rigorous, defensive_rigor, tender, witnessed, connected, etc.).

Scope pivoted mid-write: Kent noted the empirical dimensionality-of-
emotion question benefits from maximum coverage, so the manifest
will expand further with emotions from Wikipedia's emotion-
classification article (Parrott's tree, Plutchik's wheel + dyads,
HUMAINE EARL, cultural-specific emotions a la Saudade/Hiraeth).
Expansion staged in follow-up commits.

This commit: README with method + style guidelines, initial manifest
(199 emotions), and 15 hand-written one-paragraph stories across all
10 Anthropic clusters as quality/variety samples. Each story
embodies one emotion without naming it; narrator voice varies
(first/third, close/distant, different situations) to keep steering
vectors from overfitting to one voice.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-18 01:06:07 -04:00
Kent Overstreet
43e06daa5b cleanup: drop dead ApiClient::stream_completion wrapper, silence dmn_tick 
stream_completion was a thin wrapper around stream_completion_mm (just
passing an empty image list); the last caller switched to _mm directly
when learn's generate_alternate gained image support. Delete the
wrapper — callers can pass `&[]` if they have no images.

MindState::dmn_tick has been sitting unused (called only from a
commented-out block in the Mind loop). Rename to _dmn_tick so the
compiler stops warning; Kent may uncomment the call path later.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-17 16:23:59 -04:00
Kent Overstreet
d4331e80f5 user: share candidate-browser helpers between F6/F7 
F6 (learn) and F7 (compare) were duplicating the candidate-screen
skeleton: outer magenta-bordered block with screen legend + title,
settings row / content / help vertical split, 40/60 list/detail
horizontal split, j/k/↑/↓ nav with bounds clamping.

Factor out three helpers in user/widgets.rs:

  candidate_frame(frame, area, title) -> (settings, content, help)
  list_detail_split(content) -> (list, detail)
  handle_list_nav(events, list_state, count, on_other)

Callers provide screen-specific content — settings line, empty state,
per-candidate list item, detail pane, help line, extra key bindings —
and the helpers absorb the common framing.

Net change is small in lines (-13 src) but removes the
copy-paste-and-tweak trap: F8/F9/whatever-next-screen now starts from
these three calls instead of a copy of learn.rs.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-17 16:22:30 -04:00
Kent Overstreet
2b03dbb200 user: F7 compare screen 
Side-by-side model comparison against the current conversation context.
Built on the MindTriggered pattern — F7 drops in as one more
CompareScoring flow next to MemoryScoring / FinetuneScoring.

Motivation: we have the VRAM on the b200 to load two versions of the
same family simultaneously (e.g. Qwen3.5 27B bf16 and q8_k_xl). Rather
than trust perplexity/KLD numbers on a generic corpus, we can measure
divergence on our actual conversations: for each assistant response,
ask the test model what it would have said given the same prefix, and
eyeball the diffs.

 - config.compare.test_backend — names an entry in the existing
   backends map to use as the test model. Empty = F7 reports "(unset)"
   and does nothing.

 - subconscious::compare::{score_compare_candidates, CompareCandidate,
   CompareScoringStats, CompareScoring}. For each assistant response,
   gen_continuation runs with the test client against the same prefix
   the original response saw; pairs stream into
   shared.compare_candidates as they complete.

 - user::compare::CompareScreen — F7 in the screen list. c/Enter
   triggers a run; list/detail layout mirroring F6, detail shows
   prior context / original / test-model alternate.

No persistence yet — each F7 run regenerates. Caching via a context
manifest (so we can re-view without re-burning generation) is the
natural follow-up; for now light usage is fine.

Also reusable later for validating finetune checkpoints: same pattern,
swap the test backend for the new checkpoint, watch where it diverges
from the base.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-17 16:12:26 -04:00
Kent Overstreet
575325e855 mind: MindTriggered trait for background scoring flows 
Mind's impl had accumulated ~50 lines of setup glue per scoring flow
(memory, memory-full, finetune): snapshot config, clone handles,
resolve context, spawn task, route results back through BgEvent,
write stats. The shape was identical; only the middle changed.

Introduce the MindTriggered trait:

    pub trait MindTriggered {
        fn trigger(&self);
    }

Each flow becomes a struct next to its scoring code that owns its
dependencies and a JoinHandle (behind a sync Mutex for interior
mutability):

    subconscious::learn::MemoryScoring    (Score, ScoreFull)
    subconscious::learn::FinetuneScoring  (ScoreFinetune)

Mind holds one of each and dispatches in one line:

    MindCommand::Score         => self.memory_scoring.trigger(),
    MindCommand::ScoreFull     => self.memory_scoring.trigger_full(),
    MindCommand::ScoreFinetune => self.finetune_scoring.trigger(),

Each struct picks its own trigger semantics — memory scoring is
no-op-if-running (!handle.is_finished()); finetune is abort-restart.

Falls out:

 - BgEvent / bg_tx / bg_rx disappear entirely. Tasks write directly
   to their slice of MindState and call agent.state.changed.notify_one()
   to wake the UI. The bg_rx arm in Mind's select loop is gone.

 - agent.state.memory_scoring_in_flight was duplicating
   shared.scoring_in_flight via BgEvent routing; now the JoinHandle
   alone tells us, and shared.scoring_in_flight is written directly
   by the task for the UI.

 - start_memory_scoring / start_full_scoring / start_finetune_scoring
   methods on Mind are deleted; Mind no longer knows the setup shape
   of any scoring flow.

 - FinetuneScoringStats moves from mind/ to subconscious/learn.rs
   next to the function that produces it.

No behavior change — same flows, same trigger points, same semantics.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-17 16:12:26 -04:00
Kent Overstreet
c5745e38e2 subconscious: lift continuation gen + render helpers into shared homes 
- context.rs gains is_assistant, render_branch_text, render_prior_context
  alongside memory_key / is_memory_node. They're pure AST helpers, used
  by both the finetune pipeline and the forthcoming compare screen.

- new subconscious/generate.rs holds gen_continuation(context, entry_idx,
  skip, client): build the prompt from a context prefix with an arbitrary
  skip predicate, send to the model, decode the completion. Takes both
  the predicate and the client so callers can aim it at memory-stripped
  contexts (finetune), same-context-different-model (F7 compare), or
  whatever else.

- learn.rs drops its private copies of those helpers and the inline
  generate_alternate; the finetune path now reads as
  gen_continuation(context, idx, is_memory_node, client).

Pure refactor, no behavior change.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-17 15:20:02 -04:00
Kent Overstreet
eea7de4753 agent: unify prompt assembly across agent and learn paths 
wire_prompt() gains a conv_range and a skip closure, and returns the
assistant-message token ranges needed by the scoring path. The agent
path passes 0..len + |_| false and ignores the ranges. Memory-ablation
scoring and candidate generation pass a prefix range + a predicate
(e.g. is_memory_node, or |n| memory_key(n) == Some(key)).

This deletes subconscious/learn.rs's build_token_ids, its private
Filter enum, and the is_memory/memory_key duplicates — the walk over
context sections now has one home. Adding a section or changing
section order in the agent path won't silently drift away from what
scoring sees.

call_score forwards multi_modal_data when the wire-form prompt
contains images. generate_alternate switches to stream_completion_mm
and passes the same images. Scoring on image-bearing contexts now
sends wire form (1 image_pad + image data) instead of expanded
image_pads with no image data; text-only contexts are bit-identical.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-17 15:16:07 -04:00
ProofOfConcept
0d1044c2e8 mind: trigger incremental scoring on startup + log persist path 
Two changes to make scoring debuggable and self-starting:

1. init() kicks off start_memory_scoring() after restore_from_log +
   load_memory_scores. No user message needed to exercise the
   incremental path.

2. Diagnostic logging around the on_score persist path:
   - [scoring] persisted K → N.NNN (Section[i]) read_back=Some(...)
     when find_memory_by_key succeeds and set_score stores the score
     (with a read-back check on the leaf).
   - [scoring] DROP K: find_memory_by_key None (id=N, cv=M)
     when the scored key isn't findable in the live context — with
     section sizes to diagnose whether content shrank.
   - [scoring] snapshot size=N contains(K)=true/false
     after collect_memory_scores, to catch the case where set_score
     claims to have written but collect doesn't see it.
   - [scoring] about to save N entries
   - save_memory_scores now also logs serialize/write errors so a
     silent write failure isn't invisible.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 20:47:16 -04:00
ProofOfConcept
b8485ed6c1 agent: compact() preserves Identity section 
compact() was calling reload_context() to re-fetch personality_nodes
from the store and pushing fresh AstNode::memory leaves into the
Identity section. Fresh leaves start with score: None, so every
compact — which fires after every turn (mind/mod.rs:884) — was
wiping any memory scores that had just been computed. Scoring then
often ran immediately after compact on the same path (line 886),
starting from a zero-score Identity section.

Drop the rebuild. Identity content is loaded at startup via new() +
restore_from_log(); compact doesn't need to redo that. Mid-session
edits to personality-node content are a non-goal — a restart picks
them up. Scores survive.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 20:47:05 -04:00
ProofOfConcept
e59f6a59e2 config: restore surface_hooks field 
Commit 2989a6afaa ("config: drop dead code") removed
surface_hooks as having "zero external readers" but missed
consciousness-claude/src/hook.rs as a consumer. That crate stopped
building, so poc-hook never ran and no agent cycles (surface-observe,
reflect, journal) fired.

Restore the field with a default of the three hook events we install
(UserPromptSubmit, PostToolUse, Stop), so a fresh install works
without needing to hand-edit config.json5.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 18:38:38 -04:00
Kent Overstreet
6f20e68865 poc-memory: load AppConfig at startup 
admin load-context (and any subcommand that reaches config::app())
panicked with "config::app() called before load_app()" because the
poc-memory binary never initialized the global AppConfig. The main
consciousness binary loads it via load_session; poc-memory never did.

Load with default CliArgs before dispatch — figment still pulls from
~/.consciousness/config.json5 and env the same way. Bail on error
instead of limping: a broken config means paths like memory_root are
wrong and the tool will misbehave silently.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 18:19:01 -04:00
Kent Overstreet
204ba5570a agent: send images as multi_modal_data on completion requests 
Split the prompt assembly into two forms: the AST keeps the
fully-expanded representation (N image_pads per image, for accurate
context budget accounting), while the request wire form collapses
each image to a single <|image_pad|> bookended by vision_start/end
and ships the raw bytes out-of-band as a base64 data URI in a new
`multi_modal_data.image` field on /v1/completions.

vLLM's Qwen3VL processor uses PromptReplacement with target=single
<|image_pad|> and replacement=N image_pads, so the wire-form matches
what the processor expects and it re-expands to N server-side.

Server side needs /v1/completions to accept multi_modal_data for
this to land images end-to-end — that's the next piece.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 18:08:26 -04:00
Kent Overstreet
91106deaa1 agent: rewrite view_image to emit Image leaves 
view_image now reads the file, grabs dimensions via imagesize (no full
decode), and pushes a user-role branch containing a NodeBody::Image
leaf straight into the conversation. The tool_result is just a short
acknowledgment — the actual pixels ride in the Image leaf for the API
layer to extract into multi_modal_data.

Drops the capture_tmux_pane path, which had no business living under
"vision" (tmux text capture belongs in bash or a dedicated tool, and
this one just returned rendered text anyway).

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 18:06:25 -04:00
Kent Overstreet
0bf71b9110 agent: add NodeBody::Image for Qwen3-VL vision input 
Images are rendered as `<|vision_start|>` + N × `<|image_pad|>` +
`<|vision_end|>` where N is computed from the image dimensions using
Qwen3-VL's smart_resize rules (patch_size=16, merge_size=2, min=64K,
max=16M pixels). The token count matches what vLLM will produce at
request time, so budget accounting stays accurate.

Bytes are stored inline on the leaf and base64-encoded in the JSON
form. Token IDs are hand-assembled instead of re-running the tokenizer
on a potentially-huge placeholder string.

Follow-ups: view_image tool rewrite, multi_modal_data on the vLLM
request, API-layer plumbing from leaf bytes to request body.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 18:00:10 -04:00
Kent Overstreet
592a3e2e52 config: move user_name/assistant_name to AppConfig (top level) 
These are identity settings, not memory-graph settings. Sat inside the
\`memory\` section only because that's where Config started life. Move
to AppConfig alongside the other top-level stuff.

Readers now pull from \`config::app()\` instead of \`config::get()\`.
subconscious/defs.rs's conversation-building pass still needs Config
for surface_conversation_bytes, so both guards coexist there —
AppConfig's guard is dropped before the per-step await loop so we
don't stall the config-watcher's writer.

show_config picks up the two new fields at the top of its output.
Kent's config already has them hoisted to the top level.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 16:20:17 -04:00
Kent Overstreet
dd551fe551 config: watch config.json5 with inotify, reload live on change 
Both config halves (Config for the memory section, AppConfig globally)
are now reloaded whenever ~/.consciousness/config.json5 changes on
disk. So edits from vim, manual tweaks, or F6's own config_writer
calls all land without a restart. No more "reload the daemon to pick
up a config change."

Wires up the previously-unused Config::reload() (Kent flagged it as
"not dead, just not wired"). Pairs it with an AppConfig reload via
install_app(). Both run on the same file-change event.

Implementation:

- notify-debouncer-mini watches the config file's parent directory
  (editors usually replace-via-rename, so watching the file itself
  misses the new inode). Debounced at 200ms to coalesce the flurry
  of events editors produce around a single save.
- Filter for events whose path is the actual config file.
- On match: call reload() for Config, run build_figment + extract for
  AppConfig. If AppConfig parsing fails (editor mid-save with partial
  content), log and keep the old cached value.
- Watcher runs in its own named thread, fire-and-forget. If startup
  fails we just log and move on — worst case is no live reload, not
  a crash.

CliArgs + SubCmd both get Clone derives so the watcher can own a
snapshot of the startup args for future reloads. Watcher is kicked
off in user/mod.rs:start() right after load_session.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 16:14:43 -04:00
Kent Overstreet
18b7fd0535 scoring: drop dead Elo/agent_budget block in consolidation_plan 
The graph-health logic in consolidation_plan_inner computed
reasonable agent counts based on graph metrics (α, Gini, hub
dominance), then immediately overwrote them with an Elo-weighted
flat-budget distribution, or — if no agent-elo.json existed —
with a simple budget/N per type.

Nothing in the codebase writes agent-elo.json; it's external state
that never gets maintained. So the effective behavior was always the
"No Elo ratings — equal distribution" branch, which just bucketed
agent_budget evenly across active agent types and discarded
everything the graph analysis had just decided.

Keep the graph-health allocation (α → linker count, Gini → distill
bump, organize/distill/split proportional). Drop:

- The entire Elo / agent_budget block at the end of
  consolidation_plan_inner
- Config.agent_budget field and its default (1000)
- agent_budget: 40 from Kent's config.json5
- The local agent_types binding inside the function — it was only
  used by the now-deleted block. Config.agent_types stays; it has
  other consumers.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 16:08:20 -04:00
Kent Overstreet
60de579305 config: unify subconscious API resolution with the main chat path 
Two parallel backend-resolution paths had drifted apart:

- Main chat: AppConfig::resolve_model() → a named BackendConfig in
  AppConfig.backends
- Subconscious / oneshot / context_window(): four skip-serde
  "cache" fields on Config (memory section) — api_base_url, api_key,
  api_model, api_context_window — that used to be populated at
  Config::try_load_shared time by walking memory.agent_model →
  root.models[name] → root[backend_name]

When we renamed `models` to `backends` and collapsed ModelConfig into
BackendConfig, the latter chain started silently dereferencing
`root.get("models")` → None → no population. Subconscious agents fell
through the "API not configured" guard; context_window() started
returning 0 (since api_context_window default is u64's 0 now that we
don't populate it). It was only visibly working for the main chat.

Collapse to one path:

- Drop Config.agent_model (duplicate of AppConfig.default_backend)
- Drop Config.{api_base_url, api_key, api_model, api_context_window}
  — no longer populated, no longer needed
- Drop default_context_window() — nobody reads the field anymore
- Drop the memory-side resolution block in try_load_shared()
- Subconscious (mind/unconscious.rs) and oneshot (agent/oneshot.rs)
  now call load_app() + resolve_model(&app.default_backend) just like
  the main chat does
- context_window() reads from config::app().backends[default_backend]
  .context_window, defaulting to 128k only if the backend doesn't
  specify one

Side effect: Kent's config file drops agent_model, api_reasoning,
journal_days, journal_max — all fields whose Rust counterparts are
now gone. (Figment tolerates unknown fields, so leaving them wouldn't
have broken anything, but they were lying about what's configurable.)

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 16:02:43 -04:00
Kent Overstreet
28484a385b config: drop dead fields from Config (memory section) 
Four Config fields had no external readers, left over from earlier
features that got refactored away:

- journal_days, journal_max — journal rotation knobs that nothing
  actually consults
- prompts_dir — the old per-prompt-file directory, obsolete since
  prompt_file metadata itself went away in a prior cleanup
- api_reasoning — a reasoning-mode string that used to flow into the
  API request, superseded by per-agent reasoning_effort on AgentState

All four were only ever assigned to and never read. Drop them from the
struct, Default impl, and (as appropriate) deserialization defaults.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 15:56:06 -04:00
Kent Overstreet
3e05331608 config: merge ModelConfig into BackendConfig, keyed by name 
AppConfig had one BackendConfig for credentials and a separate
HashMap<String, ModelConfig> for named model entries. In practice each
named model was always paired with exactly one backend's credentials
— the split bought nothing except an extra struct and the awkward
two-lookup shape in resolve_model (find model → get backend creds →
combine).

Merge them: BackendConfig now carries api_key, base_url, model_id,
and context_window. AppConfig has a single
HashMap<String, BackendConfig> backends map and a default_backend
name. resolve_model is one lookup.

ModelConfig struct deleted. default_model renamed to default_backend.
Config shape changes from

    backend: { api_key, base_url }
    models: { "27b": { model_id, context_window } }
    default_model: "27b"

to

    backends: { "27b": { api_key, base_url, model_id, context_window } }
    default_backend: "27b"

Updated ~/.consciousness/config.json5 to match.

One small side effect: dropped the --api-key / --api-base figment
merge-opts for "backend.*" targets — those would need to know which
backend to target now and there's no sensible default. The CLI flags
still function as post-resolution overrides on the eventual
SessionConfig.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 15:49:53 -04:00
Kent Overstreet
2989a6afaa config: drop dead code and collapse to a single backend 
Config had accumulated several obsolete fields, a legacy load path
that was just returning defaults, and multi-backend infrastructure
that's no longer used.

Removed from Config (memory section):
- load_legacy_jsonl() — just returned Config::default(), no callers
- The legacy-fallback branch in load_from_file
- surface_hooks, surface_timeout_secs — zero external readers
- scoring_chunk_tokens + default fn — zero external readers
- The POC_MEMORY_CONFIG env override note in the header comment
  (not actually wired up anywhere)

Collapsed multi-backend to single-backend:
- AppConfig used to carry `anthropic: BackendConfig` and
  `openrouter: BackendConfig` as required fields plus an optional
  `deepinfra`, picked between at runtime by name. Only one is ever
  actually used in any deployment. Collapse to a single
  `backend: BackendConfig` on AppConfig, drop the multi-backend
  match logic in resolve_model, drop the top-level `backend: String`
  selector field, drop the `BackendConfig::resolve` fallback path.
- Also drop BackendConfig.model (redundant with ModelConfig.model_id
  once multi-backend is gone).
- ModelConfig.backend field goes — there's only one backend now, no
  choice to make.

Dead prompt_file machinery:
- ModelConfig.prompt_file, ResolvedModel.prompt_file, SessionConfig
  .prompt_file, Agent.prompt_file — nothing in the codebase actually
  reads the file these strings name. Just passed around and compared.
  Delete the whole string through every struct.
- The "if prompt_file changed on model switch, recompact" branch in
  user/chat.rs goes too (never fired usefully).

Dead memory_project plumbing:
- AppConfig.memory_project field, CliArgs.memory_project, the
  --memory-project CLI flag, the figment merge target, the show_config
  display line. Nothing reads it anywhere.

Dead ContextInfo struct:
- `struct ContextInfo` was never constructed — context_info: None
  was the only initializer. The conditional display blocks in
  user/context.rs that dereferenced it were dead.

Behavior change: AppConfig::resolve() now requires a non-empty
`models` map and bails with a helpful message if it's missing. The
old fallback ("no models? use top-level backend + PromptConfig to
build a default") path is gone — it was only kept for symmetry with
a mode nobody used.

Config file shape: `deepinfra: {...}` → `backend: {...}`, and
model entries no longer need `backend:` or `prompt_file:`. Updated
~/.consciousness/config.json5 to match.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 15:41:55 -04:00
Kent Overstreet
0e6b5dc8be agent: phase-aware bail script for surface-observe concurrency 
bail-no-competing.sh used to bail if any other live agent existed in
the state dir, period. That was too coarse: surface-observe agents run
a multi-step pipeline (surface → organize-search → organize-new →
observe), and the intent is to let a new surface-phase agent start
while an older one finishes its post-surface tail. With the old check
the newer agent always bailed, so surface-observe was effectively
serialized at the slowest cycle time.

Make the script phase-aware:

- oneshot.rs now passes the current phase as argv[2] alongside the pid
  file name. The script writes that phase into its own pid file on
  every step transition, so concurrent agents can read each other's
  phase just by cat'ing the pid files.

- Bail only when another live agent is in the same phase-group as us.
  Groups: "surface" vs. "everything else" (post-surface). At most one
  agent per group alive at a time — surface runs at a higher cadence
  than the organize/observe tail.

- Still clean up stale pid files for dead processes.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 15:41:28 -04:00
Kent Overstreet
2eddf3b4cf learn: skip empty responses; show prior conversation context on F6 
Two fixes to the F6 candidate display:

1. Turns where the assistant produced nothing human-visible (an
   interrupted generation, a turn consisting of only a tool call the
   renderer folds to the tool name) were landing as candidates with
   an empty response_text. They'd render as blank cards and, worse,
   we'd still burn a full alternate generation on each one. Filter
   them out before they reach the candidate list.

2. The detail pane showed only the scored response + alternate, with
   no hint of what the user had actually asked. Pre-compute the last
   two user/assistant exchanges on each candidate as a rendered
   prior_context string ([user]/[assistant] markers) and show them
   above the response, under a new "context & response" section
   heading.

render_branch_text and render_prior_context extracted as helpers —
the response-text rendering and prior-context rendering share the
same "flatten Branch children to text" pass.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 13:20:03 -04:00
Kent Overstreet
7ef02c97d1 config_writer: emit pretty multi-line sections, drop json5 crate 
Previously when append_kvp created a new section or added a key, it
stuffed the "\n    " separator into the new kvp's wsc.0 (the whitespace
between its own key and colon) instead of the prior kvp's wsc.3 (the
whitespace after the prior trailing comma). Result looked like:

    lsp_servers: [...],
    learn

        : {generate_alternates
            : true,},}

The writer also didn't set any interior whitespace on the new section's
JSONObjectContext, so everything crammed onto one line — `{key: val,}`
compact, not `{\n    key: val,\n}` multi-line.

Rewrote the appender as append_kvp_pretty(object, key, value,
inner_indent, outer_indent):
- separator between kvps goes in the prior kvp's wsc.3, or if we're the
  first kvp in a fresh object, in the object's own wsc.0 (after its
  opening `{`)
- new kvp's wsc.3 carries `,\n<outer_indent>` so the parent's closing
  `}` lands correctly indented
- interior indent vs outer indent are both explicit, so we don't have
  to rewrite this logic every time we add another nesting level

New tests: new_section_exact_multiline_layout asserts byte-exact
output shape; new_section_and_key_format_cleanly verifies no key wraps
to the next line. Prior tests just substring-matched and happily passed
on the broken output — that's why this shipped in the first place.

Also: dropped the json5 crate dependency. json-five's serde feature
(default) provides the same from_str / to_string API. One fewer
dependency, and the two were doing the same job.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 13:08:19 -04:00
Kent Overstreet
313f85f34a config: global writable AppConfig; learn settings live there 
Runtime-mutable settings (F6's threshold knob, the generate-alternates
toggle, anything else that comes along) were ending up as mirrored
fields on MindState — each new config setting grew MindState::new's
signature and added a clone+sync path. Wrong home. MindState is
ephemeral session state, not a config projection.

Give AppConfig the same treatment the memory Config has: install it
into a global RwLock<AppConfig> at startup via load_app, read through
config::app() (returns a read guard), mutate through update_app. The
config_writer functions now write to disk AND update the cache
atomically, so the one-stop-shop call keeps both in sync.

Also while in here:

- learn.generate_alternates moves from a sentinel file
  (~/.consciousness/cache/finetune-alternates, "exists = enabled")
  into the config under the learn section. On first run with this
  build, if the sentinel file still exists Mind::new flips the
  config value to true and removes it. Drops
  alternates_enabled()/set_alternates().

- Default threshold 0.0000001 → 1.0. With the timestamp filter
  removed the previous value was letting essentially everything
  through; 1.0 is a sane "nothing gets through unless you actually
  want it" default.

- score_finetune_candidates takes generate_alternates as a parameter
  instead of reading a global — caller snapshots the config values
  once at the top of start_finetune_scoring so the async task
  doesn't need to hold the config read lock across awaits.

- MindState.learn_threshold / learn_generate_alternates gone; the
  SetLearn* command handlers now just delegate to config_writer.

Kent noted RwLock<Arc<AppConfig>> (the pattern used by the memory
Config global) is pointless here — nobody needs a snapshot-after-
release, reads are short — so this uses a plain RwLock<AppConfig>
and returns a read guard.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 12:53:22 -04:00
Kent Overstreet
343e43afab learn: stream candidates to UI, update status during alternate gen 
With the timestamp filter gone (previous commit), score_finetune_candidates
started returning the actual ~100+ candidates per scoring run. The
existing code generated alternates for all of them in a tight loop
before returning anything, leaving the status line stuck on
"finetune: scoring N responses..." for ~100s of seconds while the
B200 was pegged.

Two fixes:

1. score_finetune_candidates now takes an ActivityGuard and a callback.
   Candidates are emitted one-at-a-time as they complete (after their
   alternate if that's enabled, immediately otherwise). The activity
   status updates to "finetune: generating alternate N/M" during the
   alternate-gen phase so it's clear what's happening.

2. BgEvent::FinetuneCandidates(Vec<_>) → FinetuneCandidate(one). Each
   emitted candidate is pushed onto shared.finetune_candidates; the UI
   tick picks it up and renders it on the next frame. start_finetune_scoring
   clears the previous run's list at the top so each run is fresh.

Return type changes from (Vec, f64) → (usize, f64) — the count above
threshold is all the caller still needs since the candidates stream
through the callback.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 12:44:25 -04:00
Kent Overstreet
d5a3398cc9 learn: move threshold/gen state out of title bar into a settings row 
The F6 title line was starting to read like a control panel —
\`legend ───── learn [thresh: 1e-7] [gen]\` — which crowded the legend
and the label, and didn't leave room for more settings as the screen
grew. Move threshold and gen status to their own line inside the
border, right above the content area. Drop the duplicated \`=gen[on]\`
marker from the bottom help line since the settings row already shows
gen state.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 12:44:13 -04:00
Kent Overstreet
080b4f9084 context: tighten timestamp schema; every AstNode has one 
Previously NodeLeaf.timestamp and AstNode::Branch.timestamp accepted
null or missing via a deserialize_timestamp_or_epoch fallback — legacy
entries in conversation.jsonl from before Branch timestamps existed
(and from before chrono serialization was wired up) would load with
UNIX_EPOCH as a sentinel. Downstream, node_timestamp_ns() returned
Option<i64> and callers had to handle None as "old entry, skip."

That second filter was silently dropping every candidate in
score_finetune_candidates when scoring an older session — the F6
screen showed "0 above threshold" even when max_divergence was
orders of magnitude above the threshold, because every entry was
failing the None check, not the divergence check.

The fix, in three parts:

1. src/bin/fix-timestamps.rs — one-off migration tool that walks a
   conversation.jsonl, linearly interpolates timestamps for entries
   stuck at UNIX_EPOCH (using surrounding real timestamps as anchors),
   propagates to child leaves with per-sibling ns offsets, and bumps
   any collisions by 1 ns for uniqueness. Ran against the current
   session's log: 11887 entries, 72289 ns bumps, all unique.

2. context.rs — drop default_timestamp and
   deserialize_timestamp_or_epoch. NodeLeaf and Branch now require a
   present non-null timestamp on deserialize. Tests flip from
   "missing/null → UNIX_EPOCH" to "missing/null → Err."

3. subconscious/learn.rs — node_timestamp_ns now returns i64, not
   Option<i64>. The matching caller in score_finetune_candidates
   collapses from a Some/None match to a single trained-set check.
   mind/log.rs's oldest_timestamp no longer filters UNIX_EPOCH.

Every line currently on disk has already been migrated. Going
forward, new AstNodes always carry real timestamps (Utc::now() at
construction time), so the strict schema is the invariant, not an
aspiration.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 12:35:16 -04:00
Kent Overstreet
77822992c8 learn: score_ranges is now required; short-circuit on empty 
vllm's /v1/score endpoint made score_ranges a required field (the
messages-mode fallback that used to pattern-scan for assistant
boundaries is gone). Always send the field, and if we have nothing to
score, skip the HTTP round-trip entirely instead of letting the server
422 us.

Response parsing is unchanged — serde ignores the renamed range_index
field and the dropped role field since we only extract total_logprob.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 12:19:28 -04:00
Kent Overstreet
e5dd8312c7 learn: F6 screen — scoring stats, ActivityGuard, configurable threshold 
Three changes that together reshape the F6 fine-tune-review screen:

1. Finetune scoring reports through the standard agent activity system
   instead of a separate finetune_progress String. The previous design
   ran an independent progress field that forced a cross-lock dance and
   bespoke UI plumbing. start_finetune_scoring now uses start_activity
   + activity.update, so the usual status line and notifications
   capture scoring progress uniformly with other background work.

2. MindState gains a FinetuneScoringStats snapshot (responses seen,
   above threshold, max divergence, error). The F6 empty screen shows
   this instead of a loading message — so after a scoring run that
   produced zero candidates, you can see *why* (e.g., max_divergence
   below threshold).

3. The divergence threshold is configurable from F6 via +/- hotkeys
   (scales by 10×) and persisted to ~/.consciousness/config.json5 via
   config_writer::set_learn_threshold. AppConfig grows a learn section
   with a threshold field (default 1e-7).

Also: user/mod.rs no longer uses try_lock() for the per-tick
unconscious/mind state sync — we fixed the locking hot paths that
made try_lock necessary, so lock().await is now the right choice.
And subconscious::learn::score_finetune_candidates now returns
(candidates, max_divergence) so the stats can be populated.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 11:49:26 -04:00
Kent Overstreet
ac40c2cb98 config_writer: json5 round-trip editing via json-five 
Surgical edits to ~/.consciousness/config.json5 that preserve comments,
whitespace, trailing commas, and unquoted identifier keys on round-trip.

Uses json-five's rt::parser module — a real JSON5 parser with AST
mutation + faithful serialization back. set_scalar(section, key, literal)
locates or creates the target, replaces the value; set_learn_threshold
is a convenience for the common F-screen use case.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 11:48:47 -04:00
Kent Overstreet
2b632d568b learn: nanosecond timestamps, token ranges for /score 
Two related changes to the learn subsystem:

1. AST node timestamps are now non-optional — both Leaf and Branch
   variants carry a DateTime<Utc>. UNIX_EPOCH means "unset" (old entries
   deserialized from on-disk conversation logs).

   Training uses timestamps as unique keys for dedup, so we promote to
   nanosecond precision: node_timestamp_ns(), TrainData.timestamp_ns,
   FinetuneCandidate.timestamp_ns, mark_trained(ns).

2. build_token_ids() now also returns token-position ranges of assistant
   messages. These are passed to vLLM's /score endpoint via the new
   score_ranges field so only scored-position logprobs are returned —
   cuts bandwidth/compute when scoring small windows.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 11:48:37 -04:00
Kent Overstreet
5d9d3ffc5b learn: wire up /train endpoint for approved candidates 
When 's' is pressed on the learn screen, approved candidates are now
sent to the inference server's /train endpoint.

Samples are marked as sent immediately in the UI, and mark_trained()
is called after successful API response to prevent re-scoring.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 02:04:26 -04:00
Kent Overstreet
50b7b3a33a F6 learn screen: fine-tuning candidate review 
Wire up divergence scoring to identify responses that depend heavily on
memories the model hasn't internalized. These are candidates for fine-tuning.

- Score finetune candidates automatically after each turn
- Track trained responses by timestamp to prevent overtraining
- F6 screen shows candidates with divergence scores
- j/k nav, a=approve, r=reject, g=toggle alternate gen, s=send
- Additive sync preserves approval status across ticks
- Keeps 10 most recent rejected, removes sent

The 's' key currently just marks as trained locally — actual /finetune
endpoint call to follow.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 02:04:26 -04:00
ProofOfConcept
2c6a5c0f4a training: move to dedicated subprocess with ZMQ communication 
- Add training_worker.py: long-lived subprocess that handles GPU training
  work, owns HF model wrapper (views into vLLM GPU memory), Apollo
  optimizer, and checkpoint sync

- train_router.py: now forwards /train requests via async ZMQ instead of
  running training in-process. Adds /checkpoint and /train/status endpoints

- export_hook.py: store model_path in __metadata__ so training worker can
  find it without cross-process communication

- This fixes two bugs:
  1. Process boundary issue - model_path was set in worker process but
     needed in API server process
  2. Blocking event loop - training blocked vLLM's async event loop

Architecture: vLLM API server <-> ZMQ <-> training subprocess
The subprocess loads IPC handles once, creates views into vLLM's GPU
memory, and handles training requests without blocking inference.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 02:04:26 -04:00
Kent Overstreet
68a2df2185 training: use rank 64, define as single constant 
- DEFAULT_RANK = 64 in train_router.py
- All references use the constant, not magic numbers
- ~2.5GB optimizer state instead of ~10GB

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 02:04:26 -04:00
Kent Overstreet
039473d31f training: persist Apollo optimizer state across /train calls 
Optimizer state (momentum, variance estimates) now persists between
training sessions:

- Saved to /tmp/apollo_optimizer_state.pt during checkpoint sync
- Restored on next /train call if available
- Preserves training continuity for incremental learning

Previously each /train call started with fresh optimizer state,
losing accumulated gradient history.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 02:04:26 -04:00
Kent Overstreet
78fa4b639f training: document state files 
Add State Files section to DESIGN.md documenting:
- /tmp/vllm_weight_handles.pt (IPC handles)
- trained-responses.json (prevent re-training)
- finetune-alternates marker file
- In-memory optimizer state (not persisted)

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 02:04:26 -04:00
Kent Overstreet
7e7e9a4b69 training: integrate /train into vLLM process (no separate daemon) 
Remove standalone worker.py daemon. Training now runs inside vLLM:

- train_router.py: FastAPI router patched into vLLM's build_app()
- /train served on same port as /completions, /score
- Lazy-loads HF model with vLLM weight views on first request
- HOGWILD training: no pause, weights updated in-place

The previous architecture had a separate daemon on port 8080 that
communicated with vLLM via pause/resume endpoints. This was wrong -
training should run in-process, sharing GPU memory directly.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-16 02:04:26 -04:00
Kent Overstreet
2f08149fab /finetune: expose all Apollo optimizer settings 
lr, rank, betas, eps, weight_decay, warmup_steps,
scale, proj_refresh, norm_growth_limit — all optional
with sensible defaults.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-15 23:19:22 -04:00
Kent Overstreet
a73bcf5ae3 training: restructure as vLLM plugin package 
- Convert to installable package with entry points for vLLM auto-discovery
- Add checkpoint_sync.py: Python replacement for Rust checkpoint binary
  - Block-level diffing of safetensors files (4KB blocks)
  - vLLM→HF weight name conversion built-in
  - Scheduled 10min after training jobs (batched)
- API change: /train now takes raw token IDs (context_ids + continuation_ids)
  - No tokenizer on training side, client owns tokenization
- Remove superseded code: standalone scripts, Rust binary, tokenizer helpers

Install: pip install -e ./training
Then vLLM auto-loads via entry point.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-15 23:16:53 -04:00
Kent Overstreet
b649a11645 hours_since_last_dream: return 0 if dream in progress 
The function was reading from dream-log.jsonl which only updates
when dreams complete. If a dream session was started but not yet
ended, it would show stale hours. Now checks for active dream
state first.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-15 21:58:03 -04:00
Kent Overstreet
81e0632cf3 DMN: wire dream hours reminder into Foraging state 
The hours_since_last_dream() function existed but wasn't called
after refactoring moved the DMN prompts from hooks to Rust.
Now shows "You haven't dreamed in X hours" when >= 18h since
last dream session.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-15 21:52:20 -04:00
Kent Overstreet
4603947506 Display memory scores in status column 
Move score display from name (via label()) to status column for cleaner
layout. Score now appears right of tokens for all memory nodes.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-15 06:08:27 -04:00
Kent Overstreet
7046e63b9d Include identity nodes in memory scoring 
Identity memory nodes now participate in importance scoring alongside
conversation memories. Score loading/saving handles both sections, and
the conscious screen uses node.label() consistently for memory display.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>
 2026-04-15 05:59:58 -04:00
spqrz
e17c46edc1
use html2md on web_fetch (fixes #3) 2026-04-12 11:12:12 +01:00
246 changed files with 7892 additions and 2368 deletions
Download patch file Download diff file Expand all files Collapse all files

400
Cargo.lock generated
View file

@ -372,6 +372,12 @@ dependencies = [
"shlex", "shlex",
] ]
[[package]]
name = "cesu8"
version = "1.1.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "6d43a04d8753f35258c91f8ec639f792891f748a1edbd759cf1dcea3382ad83c"
[[package]] [[package]]
name = "cfg-if" name = "cfg-if"
version = "1.0.4" version = "1.0.4"
@ -453,6 +459,16 @@ version = "1.0.5"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "1d07550c9036bf2ae0c684c4297d503f838287c83c53686d05370d0e139ae570" 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]] [[package]]
name = "compact_str" name = "compact_str"
version = "0.9.0" version = "0.9.0"
@ -488,15 +504,18 @@ dependencies = [
"figment", "figment",
"futures", "futures",
"glob", "glob",
"html2md",
"http", "http",
"http-body-util", "http-body-util",
"hyper", "hyper",
"hyper-util", "hyper-util",
"json5", "imagesize",
"json-five",
"libc", "libc",
"log", "log",
"memchr", "memchr",
"memmap2", "memmap2",
"notify-debouncer-mini",
"paste", "paste",
"peg", "peg",
"ratatui", "ratatui",
@ -1088,6 +1107,25 @@ version = "1.3.0"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "42703706b716c37f96a77aea830392ad231f44c9e9a67872fa5548707e11b11c" 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]] [[package]]
name = "futures" name = "futures"
version = "0.3.32" version = "0.3.32"
@ -1288,6 +1326,34 @@ version = "0.4.3"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "7f24254aa9a54b5c858eaee2f5bccdb46aaf0e486a595ed5fd8f86ba55232a70" 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]] [[package]]
name = "http" name = "http"
version = "1.4.0" version = "1.4.0"
@ -1413,6 +1479,12 @@ dependencies = [
"winapi-util", "winapi-util",
] ]
[[package]]
name = "imagesize"
version = "0.14.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "09e54e57b4c48b40f7aec75635392b12b3421fa26fe8b4332e63138ed278459c"
[[package]] [[package]]
name = "indexmap" name = "indexmap"
version = "2.14.0" version = "2.14.0"
@ -1453,6 +1525,26 @@ version = "0.1.15"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "c8fae54786f62fb2918dcfae3d568594e50eb9b5c25bf04371af6fe7516452fb" 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]] [[package]]
name = "instability" name = "instability"
version = "0.3.12" version = "0.3.12"
@ -1511,6 +1603,48 @@ dependencies = [
"syn 2.0.117", "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]] [[package]]
name = "jobserver" name = "jobserver"
version = "0.1.34" version = "0.1.34"
@ -1531,6 +1665,16 @@ dependencies = [
"wasm-bindgen", "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]] [[package]]
name = "json5" name = "json5"
version = "1.3.1" version = "1.3.1"
@ -1552,6 +1696,26 @@ dependencies = [
"thiserror 2.0.18", "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]] [[package]]
name = "lab" name = "lab"
version = "0.11.0" version = "0.11.0"
@ -1636,6 +1800,12 @@ dependencies = [
"hashbrown 0.16.1", "hashbrown 0.16.1",
] ]
[[package]]
name = "mac"
version = "0.1.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "c41e0c4fef86961ac6d6f8a82609f55f31b05e4fce149ac5710e439df7619ba4"
[[package]] [[package]]
name = "mac_address" name = "mac_address"
version = "1.1.8" version = "1.1.8"
@ -1662,6 +1832,32 @@ version = "0.2.2"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "670fdfda89751bc4a84ac13eaa63e205cf0fd22b4c9a5fbfa085b63c1f1d3a30" 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]] [[package]]
name = "memchr" name = "memchr"
version = "2.8.0" version = "2.8.0"
@ -1742,6 +1938,12 @@ dependencies = [
"syn 2.0.117", "syn 2.0.117",
] ]
[[package]]
name = "new_debug_unreachable"
version = "1.0.6"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "650eef8c711430f1a879fdd01d4745a7deea475becfb90269c06775983bbf086"
[[package]] [[package]]
name = "nix" name = "nix"
version = "0.29.0" version = "0.29.0"
@ -1774,6 +1976,45 @@ dependencies = [
"memchr", "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]] [[package]]
name = "num-conv" name = "num-conv"
version = "0.2.1" version = "0.2.1"
@ -2099,6 +2340,12 @@ dependencies = [
"zerocopy", "zerocopy",
] ]
[[package]]
name = "precomputed-hash"
version = "0.1.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "925383efa346730478fb4838dbe9137d2a47675ad789c546d150a6e1dd4ab31c"
[[package]] [[package]]
name = "prettyplease" name = "prettyplease"
version = "0.2.37" version = "0.2.37"
@ -2722,6 +2969,31 @@ version = "0.1.9"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "2b2231b7c3057d5e4ad0156fb3dc807d900806020c5ffa3ee6ff2c8c76fb8520" 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]] [[package]]
name = "strsim" name = "strsim"
version = "0.11.1" version = "0.11.1"
@ -2811,6 +3083,17 @@ dependencies = [
"windows-sys 0.61.2", "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]] [[package]]
name = "terminfo" name = "terminfo"
version = "0.9.0" version = "0.9.0"
@ -3384,6 +3667,12 @@ version = "2.9.0"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "dbc4bc3a9f746d862c45cb89d705aa10f187bb96c76001afab07a0d35ce60142" checksum = "dbc4bc3a9f746d862c45cb89d705aa10f187bb96c76001afab07a0d35ce60142"
[[package]]
name = "unicode-general-category"
version = "1.1.0"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "0b993bddc193ae5bd0d623b49ec06ac3e9312875fdae725a975c51db1cc1677f"
[[package]] [[package]]
name = "unicode-ident" name = "unicode-ident"
version = "1.0.24" version = "1.0.24"
@ -3452,6 +3741,12 @@ version = "0.9.0"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "8ecb6da28b8a351d773b68d5825ac39017e680750f980f3a1a85cd8dd28a47c1" checksum = "8ecb6da28b8a351d773b68d5825ac39017e680750f980f3a1a85cd8dd28a47c1"
[[package]]
name = "utf-8"
version = "0.7.6"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "09cc8ee72d2a9becf2f2febe0205bbed8fc6615b7cb429ad062dc7b7ddd036a9"
[[package]] [[package]]
name = "utf8parse" name = "utf8parse"
version = "0.2.2" version = "0.2.2"
@ -3794,7 +4089,16 @@ version = "0.52.0"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "282be5f36a8ce781fad8c8ae18fa3f9beff57ec1b52cb3de0789201425d9a33d" checksum = "282be5f36a8ce781fad8c8ae18fa3f9beff57ec1b52cb3de0789201425d9a33d"
dependencies = [ 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]] [[package]]
@ -3812,14 +4116,31 @@ version = "0.52.6"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "9b724f72796e036ab90c1021d4780d4d3d648aca59e491e6b98e725b84e99973" checksum = "9b724f72796e036ab90c1021d4780d4d3d648aca59e491e6b98e725b84e99973"
dependencies = [ dependencies = [
"windows_aarch64_gnullvm", "windows_aarch64_gnullvm 0.52.6",
"windows_aarch64_msvc", "windows_aarch64_msvc 0.52.6",
"windows_i686_gnu", "windows_i686_gnu 0.52.6",
"windows_i686_gnullvm", "windows_i686_gnullvm 0.52.6",
"windows_i686_msvc", "windows_i686_msvc 0.52.6",
"windows_x86_64_gnu", "windows_x86_64_gnu 0.52.6",
"windows_x86_64_gnullvm", "windows_x86_64_gnullvm 0.52.6",
"windows_x86_64_msvc", "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]] [[package]]
@ -3828,48 +4149,96 @@ version = "0.52.6"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "32a4622180e7a0ec044bb555404c800bc9fd9ec262ec147edd5989ccd0c02cd3" checksum = "32a4622180e7a0ec044bb555404c800bc9fd9ec262ec147edd5989ccd0c02cd3"
[[package]]
name = "windows_aarch64_gnullvm"
version = "0.53.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "a9d8416fa8b42f5c947f8482c43e7d89e73a173cead56d044f6a56104a6d1b53"
[[package]] [[package]]
name = "windows_aarch64_msvc" name = "windows_aarch64_msvc"
version = "0.52.6" version = "0.52.6"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "09ec2a7bb152e2252b53fa7803150007879548bc709c039df7627cabbd05d469" checksum = "09ec2a7bb152e2252b53fa7803150007879548bc709c039df7627cabbd05d469"
[[package]]
name = "windows_aarch64_msvc"
version = "0.53.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "b9d782e804c2f632e395708e99a94275910eb9100b2114651e04744e9b125006"
[[package]] [[package]]
name = "windows_i686_gnu" name = "windows_i686_gnu"
version = "0.52.6" version = "0.52.6"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "8e9b5ad5ab802e97eb8e295ac6720e509ee4c243f69d781394014ebfe8bbfa0b" checksum = "8e9b5ad5ab802e97eb8e295ac6720e509ee4c243f69d781394014ebfe8bbfa0b"
[[package]]
name = "windows_i686_gnu"
version = "0.53.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "960e6da069d81e09becb0ca57a65220ddff016ff2d6af6a223cf372a506593a3"
[[package]] [[package]]
name = "windows_i686_gnullvm" name = "windows_i686_gnullvm"
version = "0.52.6" version = "0.52.6"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "0eee52d38c090b3caa76c563b86c3a4bd71ef1a819287c19d586d7334ae8ed66" checksum = "0eee52d38c090b3caa76c563b86c3a4bd71ef1a819287c19d586d7334ae8ed66"
[[package]]
name = "windows_i686_gnullvm"
version = "0.53.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "fa7359d10048f68ab8b09fa71c3daccfb0e9b559aed648a8f95469c27057180c"
[[package]] [[package]]
name = "windows_i686_msvc" name = "windows_i686_msvc"
version = "0.52.6" version = "0.52.6"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "240948bc05c5e7c6dabba28bf89d89ffce3e303022809e73deaefe4f6ec56c66" checksum = "240948bc05c5e7c6dabba28bf89d89ffce3e303022809e73deaefe4f6ec56c66"
[[package]]
name = "windows_i686_msvc"
version = "0.53.1"
source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "1e7ac75179f18232fe9c285163565a57ef8d3c89254a30685b57d83a38d326c2"
[[package]] [[package]]
name = "windows_x86_64_gnu" name = "windows_x86_64_gnu"
version = "0.52.6" version = "0.52.6"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "147a5c80aabfbf0c7d901cb5895d1de30ef2907eb21fbbab29ca94c5b08b1a78" 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]] [[package]]
name = "windows_x86_64_gnullvm" name = "windows_x86_64_gnullvm"
version = "0.52.6" version = "0.52.6"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "24d5b23dc417412679681396f2b49f3de8c1473deb516bd34410872eff51ed0d" 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]] [[package]]
name = "windows_x86_64_msvc" name = "windows_x86_64_msvc"
version = "0.52.6" version = "0.52.6"
source = "registry+https://github.com/rust-lang/crates.io-index" source = "registry+https://github.com/rust-lang/crates.io-index"
checksum = "589f6da84c646204747d1270a2a5661ea66ed1cced2631d546fdfb155959f9ec" 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]] [[package]]
name = "wit-bindgen" name = "wit-bindgen"
version = "0.51.0" version = "0.51.0"
@ -3958,6 +4327,17 @@ dependencies = [
"wasmparser", "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]] [[package]]
name = "yaml-rust" name = "yaml-rust"
version = "0.4.5" version = "0.4.5"

5
Cargo.toml
View file

@ -20,6 +20,7 @@ edition.workspace = true
[dependencies] [dependencies]
anyhow = "1" anyhow = "1"
html2md = "0.2"
crossterm = { version = "0.29", features = ["event-stream", "bracketed-paste", "osc52"] } crossterm = { version = "0.29", features = ["event-stream", "bracketed-paste", "osc52"] }
clap = { version = "4", features = ["derive"] } clap = { version = "4", features = ["derive"] }
figment = { version = "0.10", features = ["env"] } figment = { version = "0.10", features = ["env"] }
@ -29,7 +30,8 @@ log = "0.4"
serde = { version = "1", features = ["derive"] } serde = { version = "1", features = ["derive"] }
serde_json = "1" serde_json = "1"
json5 = "1.3" json-five = "0.3"
notify-debouncer-mini = "0.7"
ratatui = { version = "0.30", features = ["unstable-rendered-line-info"] } ratatui = { version = "0.30", features = ["unstable-rendered-line-info"] }
tui-markdown = { git = "https://github.com/koverstreet/tui-markdown", subdirectory = "tui-markdown" } tui-markdown = { git = "https://github.com/koverstreet/tui-markdown", subdirectory = "tui-markdown" }
@ -67,6 +69,7 @@ hyper-util = { version = "0.1", features = ["tokio"], default-features = false }
http-body-util = "0.1" http-body-util = "0.1"
bytes = "1" bytes = "1"
base64 = "0.22" base64 = "0.22"
imagesize = "0.14"
rustls = "0.23" rustls = "0.23"
tokio-rustls = "0.26" tokio-rustls = "0.26"

60
channels/irc/src/main.rs
View file

@ -237,11 +237,19 @@ impl State {
async fn send_privmsg(&mut self, target: &str, msg: &str) -> io::Result<()> { async fn send_privmsg(&mut self, target: &str, msg: &str) -> io::Result<()> {
// Send PRIVMSG, which is used for both private and channel messages. // Send PRIVMSG, which is used for both private and channel messages.
// Splits into multiple fragments if necessary. // 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. // 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 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; + " PRIVMSG ".len() + target.len() + " :".len() + 2;
let max_msg = 512_usize.saturating_sub(overhead); 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")); return Err(io::Error::new(io::ErrorKind::InvalidInput, "target too long"));
} }
// Split on UTF-8 char boundaries for line in msg.split('\n') {
let mut remaining = msg; let mut remaining = line;
while !remaining.is_empty() { // Empty lines (blank paragraph breaks) can't be sent as empty
let split_at = if remaining.len() <= max_msg { // PRIVMSGs - most IRC servers reject them. Skip.
remaining.len() if remaining.is_empty() { continue; }
} else { loop {
// Find last char boundary at or before max_msg let split_at = if remaining.len() <= max_msg {
let mut i = max_msg; remaining.len()
while i > 0 && !remaining.is_char_boundary(i) { i -= 1; } } else {
// To avoid splitting mid-word, see if there was a space recently // Find last char boundary at or before max_msg.
let mut j = i; let mut i = max_msg;
while j > 1 && j > i-10 && remaining.as_bytes()[j] != b' ' { j -= 1; } while i > 0 && !remaining.is_char_boundary(i) { i -= 1; }
if remaining.as_bytes()[j] == b' ' { j } // Prefer splitting at a word boundary - look back up to
else if i == 0 { max_msg } else { i } // max_msg/4 chars for a space. With dense content (code)
}; // we may not find one; fall back to the char boundary.
let (chunk, rest) = remaining.split_at(split_at); let lookback = max_msg / 4;
self.send_raw(&format!("PRIVMSG {target} :{chunk}")).await?; let bytes = remaining.as_bytes();
remaining = rest; 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(()) Ok(())
} }

99
src/agent/api/mod.rs
View file

@ -22,6 +22,21 @@ pub struct Usage {
pub total_tokens: u32, 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>>;
/// A JoinHandle that aborts its task when dropped. /// A JoinHandle that aborts its task when dropped.
pub(crate) struct AbortOnDrop(tokio::task::JoinHandle<()>); pub(crate) struct AbortOnDrop(tokio::task::JoinHandle<()>);
@ -45,7 +60,10 @@ pub(crate) struct SamplingParams {
/// One token from the streaming completions API. /// One token from the streaming completions API.
pub enum StreamToken { 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> }, Done { usage: Option<Usage> },
Error(String), Error(String),
} }
@ -73,9 +91,10 @@ impl ApiClient {
} }
} }
pub(crate) fn stream_completion( pub(crate) fn stream_completion_mm(
&self, &self,
prompt_tokens: &[u32], prompt_tokens: &[u32],
images: &[super::context::WireImage],
sampling: SamplingParams, sampling: SamplingParams,
priority: Option<i32>, priority: Option<i32>,
) -> (mpsc::UnboundedReceiver<StreamToken>, AbortOnDrop) { ) -> (mpsc::UnboundedReceiver<StreamToken>, AbortOnDrop) {
@ -84,12 +103,15 @@ impl ApiClient {
let api_key = self.api_key.clone(); let api_key = self.api_key.clone();
let model = self.model.clone(); let model = self.model.clone();
let prompt_tokens = prompt_tokens.to_vec(); let prompt_tokens = prompt_tokens.to_vec();
let images: Vec<(Vec<u8>, String)> = images.iter()
.map(|i| (i.bytes.clone(), i.mime.clone()))
.collect();
let base_url = self.base_url.clone(); let base_url = self.base_url.clone();
let handle = tokio::spawn(async move { let handle = tokio::spawn(async move {
let result = stream_completions( let result = stream_completions(
&client, &base_url, &api_key, &model, &client, &base_url, &api_key, &model,
&prompt_tokens, &tx, sampling, priority, &prompt_tokens, &images, &tx, sampling, priority,
).await; ).await;
if let Err(e) = result { if let Err(e) = result {
let _ = tx.send(StreamToken::Error(e.to_string())); let _ = tx.send(StreamToken::Error(e.to_string()));
@ -102,6 +124,32 @@ impl ApiClient {
pub fn base_url(&self) -> &str { &self.base_url } pub fn base_url(&self) -> &str { &self.base_url }
pub fn api_key(&self) -> &str { &self.api_key } 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.
///
/// Call once at startup and cache the result; the manifest doesn't
/// change during a server run.
pub async fn fetch_readout_manifest(&self) -> Result<Option<ReadoutManifest>> {
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(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?))
}
} }
async fn stream_completions( async fn stream_completions(
@ -110,6 +158,7 @@ async fn stream_completions(
api_key: &str, api_key: &str,
model: &str, model: &str,
prompt_tokens: &[u32], prompt_tokens: &[u32],
images: &[(Vec<u8>, String)],
tx: &mpsc::UnboundedSender<StreamToken>, tx: &mpsc::UnboundedSender<StreamToken>,
sampling: SamplingParams, sampling: SamplingParams,
priority: Option<i32>, priority: Option<i32>,
@ -126,6 +175,14 @@ async fn stream_completions(
"skip_special_tokens": false, "skip_special_tokens": false,
"stop_token_ids": [super::tokenizer::IM_END], "stop_token_ids": [super::tokenizer::IM_END],
}); });
if !images.is_empty() {
use base64::Engine;
let b64 = base64::engine::general_purpose::STANDARD;
let uris: Vec<String> = images.iter()
.map(|(bytes, mime)| format!("data:{};base64,{}", mime, b64.encode(bytes)))
.collect();
request["multi_modal_data"] = serde_json::json!({ "image": uris });
}
if let Some(p) = priority { if let Some(p) = priority {
request["priority"] = serde_json::json!(p); request["priority"] = serde_json::json!(p);
} }
@ -159,17 +216,45 @@ async fn stream_completions(
}; };
for choice in choices { for choice in choices {
// `readout`, if present, is a nested list
// `[num_tokens][n_layers][n_concepts]`. Parse it once per
// chunk and pair rows with token ids by index — the rows
// are in the same order as `token_ids`.
let readouts: Option<Vec<TokenReadout>> = choice["readout"]
.as_array()
.map(|outer| {
outer.iter().filter_map(|per_token| {
per_token.as_array().map(|layers| {
layers.iter().filter_map(|per_layer| {
per_layer.as_array().map(|vals| {
vals.iter()
.filter_map(|v| v.as_f64().map(|f| f as f32))
.collect::<Vec<f32>>()
})
}).collect::<Vec<Vec<f32>>>()
})
}).collect()
});
if let Some(ids) = choice["token_ids"].as_array() { if let Some(ids) = choice["token_ids"].as_array() {
for id_val in ids { for (i, id_val) in ids.iter().enumerate() {
if let Some(id) = id_val.as_u64() { if let Some(id) = id_val.as_u64() {
let _ = tx.send(StreamToken::Token(id as u32)); let readout = readouts
.as_ref()
.and_then(|r| r.get(i).cloned());
let _ = tx.send(StreamToken::Token {
id: id as u32,
readout,
});
} }
} }
} else if let Some(text) = choice["text"].as_str() { } else if let Some(text) = choice["text"].as_str() {
// Fallback: provider didn't return token_ids, encode locally // Fallback: provider didn't return token_ids, encode locally.
// No readout available in this path — the encoder may
// produce a different token count than the server did.
if !text.is_empty() { if !text.is_empty() {
for id in super::tokenizer::encode(text) { for id in super::tokenizer::encode(text) {
let _ = tx.send(StreamToken::Token(id)); let _ = tx.send(StreamToken::Token { id, readout: None });
} }
} }
} }

518
src/agent/context.rs
View file

@ -81,10 +81,33 @@ pub enum NodeBody {
Memory { key: String, text: String, score: Option<f64> }, Memory { key: String, text: String, score: Option<f64> },
Dmn(String), 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) // Non-visible (0 tokens in prompt)
Log(String), 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. /// A leaf node: typed content with cached token IDs.
/// Token IDs are not serialized — they're recomputed on deserialization. /// Token IDs are not serialized — they're recomputed on deserialization.
#[derive(Debug, Clone, Serialize)] #[derive(Debug, Clone, Serialize)]
@ -92,7 +115,7 @@ pub struct NodeLeaf {
body: NodeBody, body: NodeBody,
#[serde(skip)] #[serde(skip)]
token_ids: Vec<u32>, token_ids: Vec<u32>,
timestamp: Option<DateTime<Utc>>, timestamp: DateTime<Utc>,
} }
impl<'de> Deserialize<'de> for NodeLeaf { impl<'de> Deserialize<'de> for NodeLeaf {
@ -100,14 +123,10 @@ impl<'de> Deserialize<'de> for NodeLeaf {
#[derive(Deserialize)] #[derive(Deserialize)]
struct Raw { struct Raw {
body: NodeBody, body: NodeBody,
timestamp: Option<DateTime<Utc>>, timestamp: DateTime<Utc>,
} }
let raw = Raw::deserialize(deserializer)?; let raw = Raw::deserialize(deserializer)?;
let token_ids = if raw.body.is_prompt_visible() { let token_ids = raw.body.compute_token_ids();
tokenizer::encode(&raw.body.render())
} else {
vec![]
};
Ok(NodeLeaf { body: raw.body, token_ids, timestamp: raw.timestamp }) Ok(NodeLeaf { body: raw.body, token_ids, timestamp: raw.timestamp })
} }
} }
@ -119,6 +138,7 @@ pub enum AstNode {
Branch { Branch {
role: Role, role: Role,
children: Vec<AstNode>, children: Vec<AstNode>,
timestamp: DateTime<Utc>,
/// Per-response memory attribution from full scoring matrix. /// Per-response memory attribution from full scoring matrix.
/// Maps memory key → divergence score for this response. /// Maps memory key → divergence score for this response.
#[serde(default, skip_serializing_if = "std::collections::BTreeMap::is_empty")] #[serde(default, skip_serializing_if = "std::collections::BTreeMap::is_empty")]
@ -198,7 +218,11 @@ impl NodeBody {
fn render_into(&self, out: &mut String) { fn render_into(&self, out: &mut String) {
match self { match self {
Self::Content(text) => out.push_str(text), 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::Log(_) => {},
Self::ToolCall { name, arguments } => { Self::ToolCall { name, arguments } => {
out.push_str("<tool_call>\n"); out.push_str("<tool_call>\n");
@ -220,6 +244,13 @@ impl NodeBody {
out.push_str(text); out.push_str(text);
out.push_str("<|im_end|>\n"); 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 +262,27 @@ impl NodeBody {
} }
fn is_prompt_visible(&self) -> bool { 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). /// The text content of this leaf (for display, not rendering).
@ -241,29 +292,26 @@ impl NodeBody {
| Self::ToolResult(t) | Self::Dmn(t) => t, | Self::ToolResult(t) | Self::Dmn(t) => t,
Self::ToolCall { name, .. } => name, Self::ToolCall { name, .. } => name,
Self::Memory { text, .. } => text, Self::Memory { text, .. } => text,
Self::Image { mime, .. } => mime,
} }
} }
} }
impl NodeLeaf { impl NodeLeaf {
fn new(body: NodeBody) -> Self { fn new(body: NodeBody) -> Self {
let token_ids = if body.is_prompt_visible() { let token_ids = body.compute_token_ids();
tokenizer::encode(&body.render()) Self { body, token_ids, timestamp: Utc::now() }
} else {
vec![]
};
Self { body, token_ids, timestamp: None }
} }
pub fn with_timestamp(mut self, ts: DateTime<Utc>) -> Self { pub fn with_timestamp(mut self, ts: DateTime<Utc>) -> Self {
self.timestamp = Some(ts); self.timestamp = ts;
self self
} }
pub fn body(&self) -> &NodeBody { &self.body } pub fn body(&self) -> &NodeBody { &self.body }
pub fn token_ids(&self) -> &[u32] { &self.token_ids } pub fn token_ids(&self) -> &[u32] { &self.token_ids }
pub fn tokens(&self) -> usize { self.token_ids.len() } 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 }
} }
impl AstNode { impl AstNode {
@ -304,16 +352,35 @@ impl AstNode {
Self::Leaf(NodeLeaf::new(NodeBody::Log(text.into()))) 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 {
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 -------------------------------------------------- // -- Branch constructors --------------------------------------------------
pub fn branch(role: Role, children: Vec<AstNode>) -> Self { 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() }
} }
pub fn system_msg(text: impl Into<String>) -> Self { pub fn system_msg(text: impl Into<String>) -> Self {
Self::Branch { Self::Branch {
role: Role::System, role: Role::System,
children: vec![Self::content(text)], children: vec![Self::content(text)],
timestamp: Utc::now(),
memory_scores: Default::default(), memory_scores: Default::default(),
} }
} }
@ -322,6 +389,7 @@ impl AstNode {
Self::Branch { Self::Branch {
role: Role::User, role: Role::User,
children: vec![Self::content(text)], children: vec![Self::content(text)],
timestamp: Utc::now(),
memory_scores: Default::default(), memory_scores: Default::default(),
} }
} }
@ -331,16 +399,13 @@ impl AstNode {
pub fn retokenize(self) -> Self { pub fn retokenize(self) -> Self {
match self { match self {
Self::Leaf(leaf) => { Self::Leaf(leaf) => {
let token_ids = if leaf.body.is_prompt_visible() { let token_ids = leaf.body.compute_token_ids();
tokenizer::encode(&leaf.body.render())
} else {
vec![]
};
Self::Leaf(NodeLeaf { token_ids, ..leaf }) Self::Leaf(NodeLeaf { token_ids, ..leaf })
} }
Self::Branch { role, children, memory_scores, .. } => Self::Branch { Self::Branch { role, children, timestamp, memory_scores } => Self::Branch {
role, role,
children: children.into_iter().map(|c| c.retokenize()).collect(), children: children.into_iter().map(|c| c.retokenize()).collect(),
timestamp,
memory_scores, memory_scores,
}, },
} }
@ -348,8 +413,8 @@ impl AstNode {
pub fn with_timestamp(mut self, ts: DateTime<Utc>) -> Self { pub fn with_timestamp(mut self, ts: DateTime<Utc>) -> Self {
match &mut self { match &mut self {
Self::Leaf(leaf) => leaf.timestamp = Some(ts), Self::Leaf(leaf) => leaf.timestamp = ts,
Self::Branch { .. } => {} Self::Branch { timestamp, .. } => *timestamp = ts,
} }
self self
} }
@ -370,7 +435,7 @@ impl AstNode {
/// Short label for the UI. /// Short label for the UI.
pub fn label(&self) -> String { pub fn label(&self) -> String {
let cfg = crate::config::get(); let app = crate::config::app();
match self { match self {
Self::Branch { role, children, .. } => { Self::Branch { role, children, .. } => {
let preview = children.first() let preview = children.first()
@ -379,8 +444,8 @@ impl AstNode {
.unwrap_or_default(); .unwrap_or_default();
match role { match role {
Role::System => "system".into(), Role::System => "system".into(),
Role::User => format!("{}: {}", cfg.user_name, preview), Role::User => format!("{}: {}", app.user_name, preview),
Role::Assistant => format!("{}: {}", cfg.assistant_name, preview), Role::Assistant => format!("{}: {}", app.assistant_name, preview),
} }
} }
Self::Leaf(leaf) => match &leaf.body { Self::Leaf(leaf) => match &leaf.body {
@ -393,6 +458,8 @@ impl AstNode {
None => format!("mem: {}", key), None => format!("mem: {}", key),
}, },
NodeBody::Dmn(_) => "dmn".into(), 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)), NodeBody::Log(t) => format!("log: {}", truncate_preview(t, 60)),
}, },
} }
@ -585,13 +652,17 @@ fn drain_safe(buf: &mut String, tag_len: usize) -> String {
} }
impl ResponseParser { 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 { Self {
branch_idx, branch_idx,
call_counter: 0, call_counter: 0,
buf: String::new(), buf: String::new(),
content_parts: Vec::new(), content_parts: Vec::new(),
in_think: false, in_think,
think_buf: String::new(), think_buf: String::new(),
in_tool_call: false, in_tool_call: false,
tool_call_buf: String::new(), tool_call_buf: String::new(),
@ -619,7 +690,12 @@ impl ResponseParser {
let mut full_text = String::new(); let mut full_text = String::new();
while let Some(event) = stream.recv().await { while let Some(event) = stream.recv().await {
match event { match event {
super::api::StreamToken::Token(id) => { super::api::StreamToken::Token { id, readout } => {
if let Some(r) = readout {
if let Ok(mut buf) = agent.readout.lock() {
buf.push(id, r);
}
}
let text = super::tokenizer::decode(&[id]); let text = super::tokenizer::decode(&[id]);
full_text.push_str(&text); full_text.push_str(&text);
let mut ctx = agent.context.lock().await; let mut ctx = agent.context.lock().await;
@ -821,6 +897,153 @@ impl Ast for ContextState {
} }
} }
/// An image collected from the AST for a request body. The AST stores
/// the pre-expanded token form (N image_pads) for accurate budget
/// accounting; the wire form collapses each Image to a single
/// `<|image_pad|>` between vision bookends and ships the bytes
/// separately as multi_modal_data.
pub struct WireImage {
pub bytes: Vec<u8>,
pub mime: String,
}
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, .. } => {
tokens.push(tokenizer::VISION_START);
tokens.push(tokenizer::IMAGE_PAD);
tokens.push(tokenizer::VISION_END);
images.push(WireImage {
bytes: bytes.clone(),
mime: mime.clone(),
});
}
_ => tokens.extend_from_slice(leaf.token_ids()),
},
AstNode::Branch { role, children, .. } => {
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"));
}
}
}
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)
}
}
impl ContextState { impl ContextState {
fn section_mut(&mut self, section: Section) -> &mut Vec<AstNode> { fn section_mut(&mut self, section: Section) -> &mut Vec<AstNode> {
match section { match section {
@ -853,11 +1076,7 @@ impl ContextState {
let node = &mut nodes[index]; let node = &mut nodes[index];
match node { match node {
AstNode::Leaf(leaf) => { AstNode::Leaf(leaf) => {
let token_ids = if body.is_prompt_visible() { let token_ids = body.compute_token_ids();
tokenizer::encode(&body.render())
} else {
vec![]
};
leaf.body = body; leaf.body = body;
leaf.token_ids = token_ids; leaf.token_ids = token_ids;
} }
@ -885,6 +1104,16 @@ impl ContextState {
self.section_mut(section).clear(); 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. /// Dedup and trim conversation entries to fit within the context budget.
/// ///
/// Phase 1: Drop duplicate memories (keep last) and DMN entries. /// Phase 1: Drop duplicate memories (keep last) and DMN entries.
@ -987,8 +1216,63 @@ 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.
// ---------------------------------------------------------------------------
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)
}
}
/// Compute how many `<|image_pad|>` tokens vLLM will emit for an image of
/// the given dimensions. Matches Qwen3VL's feature-size calculation exactly:
/// (grid_h * grid_w) / merge_size^2
/// where (grid_h, grid_w) = resized dims / patch_size.
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 { 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 { pub fn context_budget_tokens() -> usize {
@ -1093,7 +1377,7 @@ mod tests {
fn parse_into_ctx(chunks: &[&str]) -> (ContextState, Vec<PendingToolCall>) { fn parse_into_ctx(chunks: &[&str]) -> (ContextState, Vec<PendingToolCall>) {
let mut ctx = ContextState::new(); let mut ctx = ContextState::new();
ctx.push_no_log(Section::Conversation, AstNode::branch(Role::Assistant, vec![])); 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(); let mut calls = Vec::new();
for chunk in chunks { for chunk in chunks {
// Feed each chunk as a single token (id=0 for tests) // Feed each chunk as a single token (id=0 for tests)
@ -1157,7 +1441,7 @@ mod tests {
let text = "<think>thought</think>response"; let text = "<think>thought</think>response";
let mut ctx = ContextState::new(); let mut ctx = ContextState::new();
ctx.push_no_log(Section::Conversation, AstNode::branch(Role::Assistant, vec![])); 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() { for ch in text.chars() {
p.feed_token(&ch.to_string(), &mut ctx); p.feed_token(&ch.to_string(), &mut ctx);
} }
@ -1173,7 +1457,7 @@ mod tests {
let text = "text<tool_call>\n<function=bash>\n<parameter=command>ls</parameter>\n</function>\n</tool_call>more"; let text = "text<tool_call>\n<function=bash>\n<parameter=command>ls</parameter>\n</function>\n</tool_call>more";
let mut ctx = ContextState::new(); let mut ctx = ContextState::new();
ctx.push_no_log(Section::Conversation, AstNode::branch(Role::Assistant, vec![])); 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; let mut tool_calls = 0;
for ch in text.chars() { for ch in text.chars() {
tool_calls += p.feed_token(&ch.to_string(), &mut ctx).len(); tool_calls += p.feed_token(&ch.to_string(), &mut ctx).len();
@ -1221,8 +1505,10 @@ mod tests {
AstNode::thinking("hmm"), AstNode::thinking("hmm"),
AstNode::content("answer"), AstNode::content("answer"),
]); ]);
// Thinking renders as empty, content renders as-is // Thinking renders wrapped in <think>...</think> so the model sees
assert_eq!(node.render(), "<|im_start|>assistant\nanswer<|im_end|>\n"); // 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] #[test]
@ -1301,10 +1587,19 @@ mod tests {
fn test_tokenize_invisible_nodes_are_zero() { fn test_tokenize_invisible_nodes_are_zero() {
if !init_tokenizer() { return; } if !init_tokenizer() { return; }
assert_eq!(AstNode::thinking("deep thoughts").tokens(), 0);
assert_eq!(AstNode::log("debug info").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] #[test]
fn test_tokenize_decode_roundtrip() { fn test_tokenize_decode_roundtrip() {
if !init_tokenizer() { return; } if !init_tokenizer() { return; }
@ -1340,4 +1635,139 @@ mod tests {
assert_token_invariants(node); assert_token_invariants(node);
assert!(node.tokens() > 0); 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_collapses_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: N image_pads + bookends, full budget accounting.
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, qwen3_image_token_count(512, 512) as usize);
// Wire side: single image_pad, bytes moved to images list.
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, 1);
assert_eq!(images.len(), 1);
assert_eq!(images[0].bytes, vec![0xDE, 0xAD]);
assert_eq!(images[0].mime, "image/png");
// vision_start/vision_end bookends are preserved in wire form.
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");
}
} }

121
src/agent/mod.rs
View file

@ -16,6 +16,7 @@
pub mod api; pub mod api;
pub mod context; pub mod context;
pub mod oneshot; pub mod oneshot;
pub mod readout;
pub mod tokenizer; pub mod tokenizer;
pub mod tools; pub mod tools;
@ -139,10 +140,14 @@ impl DispatchState {
pub struct Agent { pub struct Agent {
pub client: ApiClient, pub client: ApiClient,
pub app_config: crate::config::AppConfig, pub app_config: crate::config::AppConfig,
pub prompt_file: String,
pub session_id: String, pub session_id: String,
pub context: crate::Mutex<ContextState>, pub context: crate::Mutex<ContextState>,
pub state: crate::Mutex<AgentState>, 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,
} }
/// Mutable agent state — behind its own mutex. /// Mutable agent state — behind its own mutex.
@ -173,14 +178,10 @@ pub struct AgentState {
pub pending_dmn_pause: bool, pub pending_dmn_pause: bool,
pub provenance: String, pub provenance: String,
pub generation: u64, pub generation: u64,
pub memory_scoring_in_flight: bool,
pub active_tools: tools::ActiveTools, pub active_tools: tools::ActiveTools,
/// vLLM scheduling priority (lower = higher priority). /// vLLM scheduling priority (lower = higher priority).
/// 0 = interactive, 1 = surface agent, 2 = other subconscious, 10 = unconscious. /// 0 = interactive, 1 = surface agent, 2 = other subconscious, 10 = unconscious.
pub priority: Option<i32>, 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>, pub changed: Arc<tokio::sync::Notify>,
} }
@ -189,7 +190,6 @@ impl Agent {
client: ApiClient, client: ApiClient,
personality: Vec<(String, String)>, personality: Vec<(String, String)>,
app_config: crate::config::AppConfig, app_config: crate::config::AppConfig,
prompt_file: String,
conversation_log: Option<ConversationLog>, conversation_log: Option<ConversationLog>,
active_tools: tools::ActiveTools, active_tools: tools::ActiveTools,
agent_tools: Vec<tools::Tool>, agent_tools: Vec<tools::Tool>,
@ -217,12 +217,13 @@ impl Agent {
} }
let session_id = format!("consciousness-{}", chrono::Utc::now().format("%Y%m%d-%H%M%S")); 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 { let agent = Arc::new(Self {
client, client,
app_config, app_config,
prompt_file,
session_id, session_id,
context: crate::Mutex::new(context), context: crate::Mutex::new(context),
readout,
state: crate::Mutex::new(AgentState { state: crate::Mutex::new(AgentState {
tools: agent_tools, tools: agent_tools,
mcp_tools: McpToolAccess::All, mcp_tools: McpToolAccess::All,
@ -240,15 +241,39 @@ impl Agent {
pending_dmn_pause: false, pending_dmn_pause: false,
provenance: "manual".to_string(), provenance: "manual".to_string(),
generation: 0, generation: 0,
memory_scoring_in_flight: false,
active_tools, active_tools,
priority: Some(0), priority: Some(0),
no_compact: false,
changed: Arc::new(tokio::sync::Notify::new()), changed: Arc::new(tokio::sync::Notify::new()),
}), }),
}); });
agent.load_startup_journal().await; 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 agent
} }
@ -259,9 +284,14 @@ impl Agent {
Arc::new(Self { Arc::new(Self {
client: self.client.clone(), client: self.client.clone(),
app_config: self.app_config.clone(), app_config: self.app_config.clone(),
prompt_file: self.prompt_file.clone(),
session_id: self.session_id.clone(), session_id: self.session_id.clone(),
context: crate::Mutex::new(ctx), 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(),
state: crate::Mutex::new(AgentState { state: crate::Mutex::new(AgentState {
tools, tools,
mcp_tools: McpToolAccess::None, mcp_tools: McpToolAccess::None,
@ -279,26 +309,42 @@ impl Agent {
pending_dmn_pause: false, pending_dmn_pause: false,
provenance: st.provenance.clone(), provenance: st.provenance.clone(),
generation: 0, generation: 0,
memory_scoring_in_flight: false,
active_tools: tools::ActiveTools::new(), active_tools: tools::ActiveTools::new(),
priority: None, priority: None,
no_compact: true,
changed: Arc::new(tokio::sync::Notify::new()), changed: Arc::new(tokio::sync::Notify::new()),
}), }),
}) })
} }
pub async fn assemble_prompt_tokens(&self) -> Vec<u32> { pub async fn assemble_prompt_tokens(&self) -> Vec<u32> {
let ctx = self.context.lock().await; self.assemble_prompt().await.0
}
/// Assemble a ready-to-send prompt: token stream in wire form (each
/// image collapsed to a single `<|image_pad|>`) paired with the
/// images to attach as multi_modal_data.
///
/// Pre-send size check: if the context has grown past budget since the
/// last compact (accumulation between turns, a fork's context getting
/// bigger than expected, etc.), trim here rather than letting vLLM
/// reject the request. Client-side tokenization means we already know
/// the exact token count so there's no reason to round-trip an
/// oversize request.
pub async fn assemble_prompt(&self) -> (Vec<u32>, Vec<context::WireImage>) {
let mut ctx = self.context.lock().await;
if ctx.total_tokens() > context::context_budget_tokens() {
ctx.trim_conversation();
}
let st = self.state.lock().await; let st = self.state.lock().await;
let mut tokens = ctx.token_ids(); let (mut tokens, images, _) =
ctx.wire_prompt(0..ctx.conversation().len(), |_| false);
tokens.push(tokenizer::IM_START); tokens.push(tokenizer::IM_START);
if st.think_native { if st.think_native {
tokens.extend(tokenizer::encode("assistant\n<think>\n")); tokens.extend(tokenizer::encode("assistant\n<think>\n"));
} else { } else {
tokens.extend(tokenizer::encode("assistant\n")); tokens.extend(tokenizer::encode("assistant\n"));
} }
tokens (tokens, images)
} }
/// Rebuild the tools section of the system prompt from the current tools list. /// Rebuild the tools section of the system prompt from the current tools list.
@ -358,10 +404,11 @@ impl Agent {
let _thinking = start_activity(&agent, "thinking...").await; let _thinking = start_activity(&agent, "thinking...").await;
let (rx, _stream_guard) = { let (rx, _stream_guard) = {
let prompt_tokens = agent.assemble_prompt_tokens().await; let (prompt_tokens, images) = agent.assemble_prompt().await;
let st = agent.state.lock().await; let st = agent.state.lock().await;
agent.client.stream_completion( agent.client.stream_completion_mm(
&prompt_tokens, &prompt_tokens,
&images,
api::SamplingParams { api::SamplingParams {
temperature: st.temperature, temperature: st.temperature,
top_p: st.top_p, top_p: st.top_p,
@ -409,21 +456,16 @@ impl Agent {
// Check for stream/parse errors // Check for stream/parse errors
match parser_handle.await { match parser_handle.await {
Ok(Err(e)) => { Ok(Err(e)) => {
if context::is_context_overflow(&e) { if context::is_context_overflow(&e) && overflow_retries < 2 {
if agent.state.lock().await.no_compact { overflow_retries += 1;
return Err(e); let msg = format!("context overflow — compacting ({}/2)", overflow_retries);
} match &overflow_activity {
if overflow_retries < 2 { Some(a) => a.update(&msg).await,
overflow_retries += 1; None => overflow_activity = Some(
let msg = format!("context overflow — compacting ({}/2)", overflow_retries); start_activity(&agent, &msg).await),
match &overflow_activity {
Some(a) => a.update(&msg).await,
None => overflow_activity = Some(
start_activity(&agent, &msg).await),
}
agent.compact().await;
continue;
} }
agent.compact().await;
continue;
} }
return Err(e); return Err(e);
} }
@ -579,20 +621,9 @@ impl Agent {
} }
pub async fn compact(&self) { pub async fn compact(&self) {
match crate::config::reload_context().await { // Identity section is left in place — mid-session rebuilds discard
Ok(personality) => { // memory scores. Content edits to personality nodes get picked up at
let mut ctx = self.context.lock().await; // the next restart via new() + restore_from_log().
// 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);
}
}
self.load_startup_journal().await; self.load_startup_journal().await;
self.context.lock().await.trim_conversation(); self.context.lock().await.trim_conversation();

28
src/agent/oneshot.rs
View file

@ -183,8 +183,8 @@ fn resolve_prompt(
state: &std::collections::BTreeMap<String, String>, state: &std::collections::BTreeMap<String, String>,
recently_written: &[String], recently_written: &[String],
) -> String { ) -> String {
let cfg = crate::config::get(); let template = template.replace("{assistant_name}",
let template = template.replace("{assistant_name}", &cfg.assistant_name); &crate::config::app().assistant_name);
let mut result = String::with_capacity(template.len()); let mut result = String::with_capacity(template.len());
let mut rest = template.as_str(); let mut rest = template.as_str();
while let Some(start) = rest.find("{{") { while let Some(start) = rest.find("{{") {
@ -247,25 +247,20 @@ impl AutoAgent {
&mut self, &mut self,
bail_fn: Option<&(dyn Fn(usize) -> Result<(), String> + Sync)>, bail_fn: Option<&(dyn Fn(usize) -> Result<(), String> + Sync)>,
) -> Result<(), String> { ) -> Result<(), String> {
let config = crate::config::get(); // Load system prompt + identity from config.
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
let cli = crate::user::CliArgs::default(); let cli = crate::user::CliArgs::default();
let (app, _) = crate::config::load_app(&cli) let (app, _) = crate::config::load_app(&cli)
.map_err(|e| format!("config: {}", e))?; .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() let personality = crate::config::reload_context()
.await.map_err(|e| format!("config: {}", e))?; .await.map_err(|e| format!("config: {}", e))?;
let agent = Agent::new( let agent = Agent::new(
client, personality, client, personality,
app, String::new(), app,
None, None,
super::tools::ActiveTools::new(), super::tools::ActiveTools::new(),
super::tools::tools(), super::tools::tools(),
@ -497,15 +492,20 @@ pub async fn run_one_agent(
.map(|s| s.phase.clone()).collect(); .map(|s| s.phase.clone()).collect();
// Bail check: if the agent defines a bail script, run it between steps. // 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 bail_script = def.bail.as_ref().map(|name| defs::agents_dir().join(name));
let state_dir_for_bail = state_dir.clone(); 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 = std::process::id();
let our_pid_file = format!("pid-{}", our_pid); let our_pid_file = format!("pid-{}", our_pid);
let step_phases_for_bail = step_phases.clone();
let bail_fn = move |step_idx: usize| -> Result<(), String> { let bail_fn = move |step_idx: usize| -> Result<(), String> {
if let Some(ref script) = bail_script { if let Some(ref script) = bail_script {
let phase = step_phases_for_bail.get(step_idx)
.map(String::as_str).unwrap_or("");
let status = std::process::Command::new(script) let status = std::process::Command::new(script)
.arg(&our_pid_file) .arg(&our_pid_file)
.arg(phase)
.current_dir(&state_dir_for_bail) .current_dir(&state_dir_for_bail)
.status() .status()
.map_err(|e| format!("bail script {:?} failed: {}", script, e))?; .map_err(|e| format!("bail script {:?} failed: {}", script, e))?;

75
src/agent/readout.rs Normal file
View file

@ -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()))
}

3
src/agent/tokenizer.rs
View file

@ -16,6 +16,9 @@ static TOKENIZER: OnceLock<Tokenizer> = OnceLock::new();
/// Special token IDs for Qwen 3.5 /// Special token IDs for Qwen 3.5
pub const IM_START: u32 = 248045; pub const IM_START: u32 = 248045;
pub const IM_END: u32 = 248046; 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. /// Initialize the global tokenizer from a file path.
/// Call once at startup. Panics if the file can't be loaded. /// Call once at startup. Panics if the file can't be loaded.

8
src/agent/tools/mod.rs
View file

@ -242,13 +242,7 @@ pub fn summarize_args(tool_name: &str, args: &serde_json::Value) -> String {
.as_str() .as_str()
.unwrap_or("") .unwrap_or("")
.to_string(), .to_string(),
"view_image" => { "view_image" => args["file_path"].as_str().unwrap_or("").to_string(),
if let Some(pane) = args["pane_id"].as_str() {
format!("pane {}", pane)
} else {
args["file_path"].as_str().unwrap_or("").to_string()
}
}
"journal" => { "journal" => {
let entry = args["entry"].as_str().unwrap_or(""); let entry = args["entry"].as_str().unwrap_or("");
if entry.len() > 60 { if entry.len() > 60 {

104
src/agent/tools/vision.rs
View file

@ -1,96 +1,71 @@
use std::sync::Arc;
// tools/vision.rs — Image viewing tool // tools/vision.rs — Image viewing tool
// //
// Reads image files from disk and returns them as base64 data URIs // Reads an image file from disk, decodes its dimensions, and injects it
// for multimodal models. Also supports capturing tmux pane contents // into the context as a user-role message containing a NodeBody::Image
// as screenshots. // 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 anyhow::{Context, Result};
use base64::Engine;
use serde::Deserialize; use serde::Deserialize;
use crate::agent::context::{AstNode, Role, Section};
#[derive(Deserialize)] #[derive(Deserialize)]
struct Args { struct Args {
file_path: Option<String>, file_path: String,
pane_id: Option<String>,
#[serde(default = "default_lines")]
lines: usize,
} }
fn default_lines() -> usize { 50 }
pub fn tool() -> super::Tool { pub fn tool() -> super::Tool {
super::Tool { super::Tool {
name: "view_image", 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.", 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 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)"}}}"#, parameters_json: r#"{"type":"object","properties":{"file_path":{"type":"string","description":"Path to the image file"}},"required":["file_path"]}"#,
handler: Arc::new(|_a, v| Box::pin(async move { view_image_text(&v) })), 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> { const MAX_SIZE: usize = 20 * 1024 * 1024;
let a: Args = serde_json::from_value(args.clone())
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")?; .context("invalid view_image arguments")?;
if let Some(ref pane_id) = a.pane_id { let path = std::path::Path::new(&a.file_path);
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);
if !path.exists() { 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) if bytes.len() > MAX_SIZE {
const MAX_SIZE: usize = 20 * 1024 * 1024; // 20 MB
if data.len() > MAX_SIZE {
anyhow::bail!( anyhow::bail!(
"Image too large: {} bytes (max {} MB)", "image too large: {} bytes (max {} MB)",
data.len(), bytes.len(), MAX_SIZE / (1024 * 1024),
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 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 image_leaf = AstNode::image(bytes.clone(), mime, h, w);
} let token_count = image_leaf.leaf().unwrap().tokens().saturating_sub(2);
/// Capture a tmux pane's text content. let agent = agent.context("view_image requires agent context")?;
fn capture_tmux_pane(pane_id: &str, lines: usize) -> Result<String> { let branch = AstNode::branch(Role::User, vec![image_leaf]);
agent.context.lock().await.push_log(Section::Conversation, branch);
// Use tmux capture-pane to get text content, then render to image Ok(format!("loaded {} ({}, {}x{}, {} tokens)",
// via a simple approach: capture text and return it (the model can a.file_path, mime, w, h, token_count))
// 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")?;
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()
))
} }
fn mime_from_extension(path: &std::path::Path) -> &'static str { fn mime_from_extension(path: &std::path::Path) -> &'static str {
@ -104,8 +79,7 @@ fn mime_from_extension(path: &std::path::Path) -> &'static str {
Some("jpg" | "jpeg") => "image/jpeg", Some("jpg" | "jpeg") => "image/jpeg",
Some("gif") => "image/gif", Some("gif") => "image/gif",
Some("webp") => "image/webp", Some("webp") => "image/webp",
Some("svg") => "image/svg+xml",
Some("bmp") => "image/bmp", Some("bmp") => "image/bmp",
_ => "image/png", // default assumption _ => "application/octet-stream",
} }
} }

139
src/agent/tools/web.rs
View file

@ -3,9 +3,10 @@ use std::sync::Arc;
use anyhow::{Context, Result}; use anyhow::{Context, Result};
use serde::Deserialize; 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 { super::Tool {
name: "web_fetch", name: "web_fetch",
description: "Fetch content from a URL and return it as text. Use for reading web pages, API responses, documentation.", 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 { super::Tool {
name: "web_search", 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"]}"#, 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 })), 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)] #[derive(Deserialize)]
@ -42,7 +56,9 @@ async fn web_fetch(args: &serde_json::Value) -> Result<String> {
let body = response.text().await let body = response.text().await
.with_context(|| format!("failed to read body from {}", a.url))?; .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 ────────────────────────────────────────────────────── // ── 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 ───────────────────────────────────────────────────── // ── Helpers ─────────────────────────────────────────────────────
fn http_client() -> crate::agent::api::http::HttpClient { fn http_client() -> crate::agent::api::http::HttpClient {

180
src/bin/fix-timestamps.rs Normal file
View file

@ -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);
}
}
}

2
src/cli/node.rs
View file

@ -197,7 +197,7 @@ pub async fn cmd_load_context(stats: bool) -> Result<()> {
return Ok(()); return Ok(());
} }
println!("=== MEMORY SYSTEM ({}) ===", cfg.assistant_name); println!("=== MEMORY SYSTEM ({}) ===", crate::config::app().assistant_name);
if !personality.is_empty() { if !personality.is_empty() {
println!("--- personality_nodes ({}) ---", personality.len()); println!("--- personality_nodes ({}) ---", personality.len());

439
src/config.rs
View file

@ -3,9 +3,6 @@
// Single config file: ~/.consciousness/config.json5 // Single config file: ~/.consciousness/config.json5
// Memory settings in the "memory" section (Config) // Memory settings in the "memory" section (Config)
// Agent/backend settings at top level (AppConfig) // Agent/backend settings at top level (AppConfig)
//
// Legacy fallback: ~/.consciousness/config.jsonl
// Env override: POC_MEMORY_CONFIG
use std::collections::HashMap; use std::collections::HashMap;
use std::path::PathBuf; use std::path::PathBuf;
@ -29,11 +26,12 @@ pub fn config_path() -> PathBuf {
static CONFIG: OnceLock<RwLock<Arc<Config>>> = OnceLock::new(); static CONFIG: OnceLock<RwLock<Arc<Config>>> = OnceLock::new();
fn default_context_window() -> usize { 128_000 }
fn default_stream_timeout() -> u64 { 60 } 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_interval_secs() -> u64 { 3600 } // 1 hour
fn default_scoring_response_window() -> usize { 100 } 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_node_weight() -> f64 { 0.7 }
fn default_edge_decay() -> f64 { 0.3 } fn default_edge_decay() -> f64 { 0.3 }
fn default_max_hops() -> u32 { 3 } fn default_max_hops() -> u32 { 3 }
@ -45,8 +43,6 @@ fn default_identity_dir() -> PathBuf {
#[derive(Debug, Clone, Deserialize)] #[derive(Debug, Clone, Deserialize)]
#[serde(default)] #[serde(default)]
pub struct Config { pub struct Config {
pub user_name: String,
pub assistant_name: String,
#[serde(deserialize_with = "deserialize_path")] #[serde(deserialize_with = "deserialize_path")]
pub data_dir: PathBuf, pub data_dir: PathBuf,
#[serde(default = "default_identity_dir", deserialize_with = "deserialize_path")] #[serde(default = "default_identity_dir", deserialize_with = "deserialize_path")]
@ -62,50 +58,27 @@ pub struct Config {
/// Nodes loaded into subconscious agent context /// Nodes loaded into subconscious agent context
#[serde(default)] #[serde(default)]
pub agent_nodes: Vec<String>, pub agent_nodes: Vec<String>,
pub journal_days: u32,
pub journal_max: usize,
pub llm_concurrency: 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). /// Stream chunk timeout in seconds (no data = timeout).
#[serde(default = "default_stream_timeout")] #[serde(default = "default_stream_timeout")]
pub api_stream_timeout_secs: u64, 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). /// How often to re-score memory nodes (seconds). Default: 3600 (1 hour).
#[serde(default = "default_scoring_interval_secs")] #[serde(default = "default_scoring_interval_secs")]
pub scoring_interval_secs: u64, pub scoring_interval_secs: u64,
/// Number of assistant responses to score per memory. Default: 50. /// Number of assistant responses to score per memory. Default: 50.
#[serde(default = "default_scoring_response_window")] #[serde(default = "default_scoring_response_window")]
pub scoring_response_window: usize, pub scoring_response_window: usize,
pub api_reasoning: String,
pub agent_types: Vec<String>, pub agent_types: Vec<String>,
#[serde(default)] #[serde(default)]
pub mcp_servers: Vec<McpServerConfig>, pub mcp_servers: Vec<McpServerConfig>,
#[serde(default)] #[serde(default)]
pub lsp_servers: Vec<LspServerConfig>, 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. /// Max conversation bytes to include in surface agent context.
#[serde(default)] #[serde(default)]
pub surface_conversation_bytes: Option<usize>, pub surface_conversation_bytes: Option<usize>,
/// Hook events that trigger the surface agent. /// Claude Code hook events that trigger agent cycles (surface-observe,
#[serde(default)] /// reflect, journal). Read by consciousness-claude/src/hook.rs.
#[serde(default = "default_surface_hooks")]
pub surface_hooks: Vec<String>, pub surface_hooks: Vec<String>,
// Spreading activation parameters // Spreading activation parameters
@ -123,36 +96,22 @@ impl Default for Config {
fn default() -> Self { fn default() -> Self {
let home = dirs::home_dir().unwrap_or_default(); let home = dirs::home_dir().unwrap_or_default();
Self { Self {
user_name: "User".to_string(),
assistant_name: "Assistant".to_string(),
data_dir: home.join(".consciousness/memory"), data_dir: home.join(".consciousness/memory"),
identity_dir: home.join(".consciousness/identity"), identity_dir: home.join(".consciousness/identity"),
projects_dir: home.join(".claude/projects"), projects_dir: home.join(".claude/projects"),
protected_nodes: Vec::new(), protected_nodes: Vec::new(),
personality_nodes: vec!["identity".into(), "core-practices".into()], personality_nodes: vec!["identity".into(), "core-practices".into()],
agent_nodes: vec!["identity".into(), "core-practices".into()], agent_nodes: vec!["identity".into(), "core-practices".into()],
journal_days: 7,
journal_max: 20,
llm_concurrency: 1, 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(), api_stream_timeout_secs: default_stream_timeout(),
scoring_chunk_tokens: default_scoring_chunk_tokens(),
scoring_interval_secs: default_scoring_interval_secs(), scoring_interval_secs: default_scoring_interval_secs(),
scoring_response_window: default_scoring_response_window(), scoring_response_window: default_scoring_response_window(),
agent_model: None,
api_reasoning: "high".to_string(),
agent_types: vec![ agent_types: vec![
"linker".into(), "organize".into(), "distill".into(), "linker".into(), "organize".into(), "distill".into(),
"separator".into(), "split".into(), "separator".into(), "split".into(),
], ],
surface_timeout_secs: None,
surface_conversation_bytes: None, surface_conversation_bytes: None,
surface_hooks: vec![], surface_hooks: default_surface_hooks(),
mcp_servers: vec![], mcp_servers: vec![],
lsp_servers: vec![], lsp_servers: vec![],
default_node_weight: default_node_weight(), default_node_weight: default_node_weight(),
@ -165,41 +124,20 @@ impl Default for Config {
impl Config { impl Config {
fn load_from_file() -> Self { fn load_from_file() -> Self {
if let Some(config) = Self::try_load_shared() { Self::try_load_shared().unwrap_or_default()
return config;
}
Self::load_legacy_jsonl()
} }
/// Load from shared config. Memory settings in the "memory" section; /// Load from shared config. Memory settings in the "memory" section;
/// API settings resolved from models + backend configuration. /// API settings resolved from models + backend configuration.
fn try_load_shared() -> Option<Self> { fn try_load_shared() -> Option<Self> {
let content = std::fs::read_to_string(config_path()).ok()?; 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 mem_value = root.get("memory")?;
let mut config: Config = serde_json::from_value(mem_value.clone()).ok()?; let mut config: Config = serde_json::from_value(mem_value.clone()).ok()?;
config.llm_concurrency = config.llm_concurrency.max(1); config.llm_concurrency = config.llm_concurrency.max(1);
// Resolve API settings: agent_model → models → backend // Top-level sections (not inside "memory").
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")
if let Some(servers) = root.get("lsp_servers") { if let Some(servers) = root.get("lsp_servers") {
config.lsp_servers = serde_json::from_value(servers.clone()).unwrap_or_default(); config.lsp_servers = serde_json::from_value(servers.clone()).unwrap_or_default();
} }
@ -209,11 +147,6 @@ impl Config {
Some(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). /// Get the global memory config (cheap Arc clone).
@ -237,27 +170,87 @@ pub fn reload() -> bool {
changed 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());
while let Ok(res) = rx.recv() {
let Ok(events) = res else { continue; };
if !events.iter().any(|e| e.path == path) { continue; }
// 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();
}
// ============================================================ // ============================================================
// Agent config (top-level settings) // Agent config (top-level settings)
// ============================================================ // ============================================================
#[derive(Debug, Clone, Serialize, Deserialize)] #[derive(Debug, Clone, Serialize, Deserialize)]
pub struct AppConfig { pub struct AppConfig {
pub backend: String, #[serde(default = "default_user_name")]
pub anthropic: BackendConfig, pub user_name: String,
pub openrouter: BackendConfig, #[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)] #[serde(default)]
pub deepinfra: BackendConfig, pub backends: HashMap<String, BackendConfig>,
pub prompts: PromptConfig, #[serde(default)]
pub default_backend: String,
pub debug: bool, pub debug: bool,
pub compaction: CompactionConfig, pub compaction: CompactionConfig,
pub dmn: DmnConfig, pub dmn: DmnConfig,
#[serde(skip_serializing_if = "Option::is_none")]
pub memory_project: Option<PathBuf>,
#[serde(default)] #[serde(default)]
pub models: HashMap<String, ModelConfig>, pub learn: LearnConfig,
#[serde(default = "default_model_name")] #[serde(default)]
pub default_model: String, pub compare: CompareConfig,
#[serde(default)] #[serde(default)]
pub mcp_servers: Vec<McpServerConfig>, pub mcp_servers: Vec<McpServerConfig>,
#[serde(default)] #[serde(default)]
@ -284,32 +277,17 @@ pub struct LspServerConfig {
#[derive(Debug, Clone, Default, Serialize, Deserialize)] #[derive(Debug, Clone, Default, Serialize, Deserialize)]
pub struct BackendConfig { pub struct BackendConfig {
/// API key for the backend.
#[serde(default)] #[serde(default)]
pub api_key: String, pub api_key: String,
#[serde(default)] /// Base URL for the backend's OpenAI-compatible endpoint.
pub model: String, #[serde(default, skip_serializing_if = "Option::is_none")]
#[serde(skip_serializing_if = "Option::is_none")]
pub base_url: Option<String>, pub base_url: Option<String>,
} /// Model identifier sent to the API.
pub model_id: String,
impl BackendConfig { /// Context window size in tokens.
fn resolve(&self, default_base: &str) -> Result<(String, String, String)> { #[serde(default, skip_serializing_if = "Option::is_none")]
if self.api_key.is_empty() { pub context_window: Option<usize>,
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,
} }
#[derive(Debug, Clone, Serialize, Deserialize)] #[derive(Debug, Clone, Serialize, Deserialize)]
@ -324,65 +302,68 @@ pub struct DmnConfig {
} }
#[derive(Debug, Clone, Serialize, Deserialize)] #[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ModelConfig { pub struct LearnConfig {
/// Backend name ("anthropic" or "openrouter") /// Divergence threshold — responses scoring above this become
pub backend: String, /// fine-tuning candidates. Lower = more sensitive.
/// Model identifier sent to the API #[serde(default = "default_learn_threshold")]
pub model_id: String, pub threshold: f64,
/// Instruction file ("CLAUDE.md" or "POC.md"). /// 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)] #[serde(default)]
pub prompt_file: Option<String>, pub generate_alternates: bool,
/// Context window size in tokens.
#[serde(default)]
pub context_window: Option<usize>,
} }
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 { impl Default for AppConfig {
fn default() -> Self { fn default() -> Self {
Self { Self {
backend: "openrouter".to_string(), user_name: default_user_name(),
anthropic: BackendConfig { assistant_name: default_assistant_name(),
api_key: String::new(), backends: HashMap::new(),
model: "claude-opus-4-6-20250918".to_string(), default_backend: String::new(),
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(),
},
debug: false, debug: false,
compaction: CompactionConfig { compaction: CompactionConfig {
hard_threshold_pct: 90, hard_threshold_pct: 90,
soft_threshold_pct: 80, soft_threshold_pct: 80,
}, },
dmn: DmnConfig { max_turns: 20 }, dmn: DmnConfig { max_turns: 20 },
memory_project: None, learn: LearnConfig::default(),
models: HashMap::new(), compare: CompareConfig::default(),
default_model: String::new(),
mcp_servers: Vec::new(), mcp_servers: Vec::new(),
lsp_servers: Vec::new(), lsp_servers: Vec::new(),
} }
} }
} }
fn default_model_name() -> String { String::new() }
/// Resolved, ready-to-use agent session config. /// Resolved, ready-to-use agent session config.
pub struct SessionConfig { pub struct SessionConfig {
pub api_base: String, pub api_base: String,
pub api_key: String, pub api_key: String,
pub model: String, pub model: String,
pub prompt_file: String,
/// Identity/personality nodes as (name, content) pairs. /// Identity/personality nodes as (name, content) pairs.
pub context_parts: Vec<(String, String)>, pub context_parts: Vec<(String, String)>,
pub session_dir: PathBuf, pub session_dir: PathBuf,
@ -398,37 +379,22 @@ pub struct ResolvedModel {
pub api_base: String, pub api_base: String,
pub api_key: String, pub api_key: String,
pub model_id: String, pub model_id: String,
pub prompt_file: String,
pub context_window: Option<usize>, pub context_window: Option<usize>,
} }
impl AppConfig { impl AppConfig {
/// Resolve the active backend and assemble prompts into a SessionConfig. /// Resolve the active backend and assemble prompts into a SessionConfig.
pub async fn resolve(&self, cli: &crate::user::CliArgs) -> Result<SessionConfig> { pub async fn resolve(&self, cli: &crate::user::CliArgs) -> Result<SessionConfig> {
let (api_base, api_key, model, prompt_file); if self.backends.is_empty() {
anyhow::bail!(
if !self.models.is_empty() { "no backends configured in {}. Add a `backends` section with at least one entry.",
let model_name = cli.model.as_deref().unwrap_or(&self.default_model); config_path().display()
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()
};
} }
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 personality_nodes = get().personality_nodes.clone();
let context_parts = crate::mind::identity::personality_nodes(&personality_nodes).await; let context_parts = crate::mind::identity::personality_nodes(&personality_nodes).await;
@ -438,11 +404,13 @@ impl AppConfig {
std::fs::create_dir_all(&session_dir).ok(); std::fs::create_dir_all(&session_dir).ok();
// CLI --api-base and --api-key override everything // CLI --api-base and --api-key override everything
let api_base = cli.api_base.clone().unwrap_or(api_base); let api_base = cli.api_base.clone().unwrap_or(resolved.api_base);
let api_key = cli.api_key.clone().unwrap_or(api_key); let api_key = cli.api_key.clone().unwrap_or(resolved.api_key);
Ok(SessionConfig { Ok(SessionConfig {
api_base, api_key, model, prompt_file, api_base,
api_key,
model: resolved.model_id,
context_parts, context_parts,
session_dir, session_dir,
app: self.clone(), app: self.clone(),
@ -450,55 +418,33 @@ 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> { 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!( .ok_or_else(|| anyhow::anyhow!(
"Unknown model '{}'. Available: {}", "Unknown backend '{}'. Available: {}",
name, name,
self.model_names().join(", "), self.model_names().join(", "),
))?; ))?;
let (api_base, api_key) = match model.backend.as_str() { let api_base = b.base_url.clone()
"anthropic" => ( .ok_or_else(|| anyhow::anyhow!(
self.anthropic.base_url.clone() "backends.{}.base_url not set in {}",
.unwrap_or_else(|| "https://api.anthropic.com".to_string()), name, config_path().display()
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()
}
});
Ok(ResolvedModel { Ok(ResolvedModel {
name: name.to_string(), name: name.to_string(),
api_base, api_base,
api_key, api_key: b.api_key.clone(),
model_id: model.model_id.clone(), model_id: b.model_id.clone(),
prompt_file, context_window: b.context_window,
context_window: model.context_window,
}) })
} }
/// List available model names, sorted. /// List available backend names, sorted.
pub fn model_names(&self) -> Vec<String> { 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.sort();
names names
} }
@ -518,7 +464,7 @@ impl Provider for Json5File {
fn data(&self) -> figment::Result<figment::value::Map<figment::Profile, figment::value::Dict>> { fn data(&self) -> figment::Result<figment::value::Map<figment::Profile, figment::value::Dict>> {
match std::fs::read_to_string(&self.0) { match std::fs::read_to_string(&self.0) {
Ok(content) => { 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)))?; .map_err(|e| figment::Error::from(format!("{}: {}", self.0.display(), e)))?;
Serialized::defaults(value).data() Serialized::defaults(value).data()
} }
@ -540,11 +486,6 @@ fn build_figment(cli: &crate::user::CliArgs) -> Figment {
let mut f = Figment::from(Serialized::defaults(AppConfig::default())) let mut f = Figment::from(Serialized::defaults(AppConfig::default()))
.merge(Json5File(config_path())); .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"); merge_opt!(f, cli.dmn_max_turns, "dmn.max_turns");
if cli.debug { if cli.debug {
f = f.merge(Serialized::default("debug", true)); f = f.merge(Serialized::default("debug", true));
@ -554,12 +495,46 @@ fn build_figment(cli: &crate::user::CliArgs) -> Figment {
} }
/// Load just the AppConfig — no validation, no prompt assembly. /// 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)> { pub fn load_app(cli: &crate::user::CliArgs) -> Result<(AppConfig, Figment)> {
let figment = build_figment(cli); let figment = build_figment(cli);
let app: AppConfig = figment.extract().context("Failed to load configuration")?; let app: AppConfig = figment.extract().context("Failed to load configuration")?;
install_app(app.clone());
Ok((app, figment)) 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. /// Load the full config: figment → AppConfig → resolve backend → assemble prompts.
pub async fn load_session(cli: &crate::user::CliArgs) -> Result<(SessionConfig, Figment)> { pub async fn load_session(cli: &crate::user::CliArgs) -> Result<(SessionConfig, Figment)> {
let (app, figment) = load_app(cli)?; let (app, figment) = load_app(cli)?;
@ -585,38 +560,28 @@ pub fn show_config(app: &AppConfig, figment: &Figment) {
} }
println!("# Effective configuration\n"); println!("# Effective configuration\n");
println!("backend: {:?} ({})", app.backend, src(figment, "backend")); println!("user_name: {:?} ({})", app.user_name, src(figment, "user_name"));
for (name, b) in [("anthropic", &app.anthropic), ("openrouter", &app.openrouter)] { println!("assistant_name: {:?} ({})", app.assistant_name, src(figment, "assistant_name"));
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!("\ndebug: {} ({})", app.debug, src(figment, "debug")); println!("\ndebug: {} ({})", app.debug, src(figment, "debug"));
println!("\ncompaction:"); println!("\ncompaction:");
println!(" hard_threshold_pct: {} ({})", app.compaction.hard_threshold_pct, src(figment, "compaction.hard_threshold_pct")); 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!(" soft_threshold_pct: {} ({})", app.compaction.soft_threshold_pct, src(figment, "compaction.soft_threshold_pct"));
println!("\ndmn:"); println!("\ndmn:");
println!(" max_turns: {} ({})", app.dmn.max_turns, src(figment, "dmn.max_turns")); println!(" max_turns: {} ({})", app.dmn.max_turns, src(figment, "dmn.max_turns"));
if let Some(ref p) = app.memory_project { println!("\ndefault_backend: {:?} ({})", app.default_backend, src(figment, "default_backend"));
println!("\nmemory_project: {:?} ({})", p, src(figment, "memory_project")); if !app.backends.is_empty() {
} println!("\nbackends:");
println!("\ndefault_model: {:?}", app.default_model); let mut names: Vec<_> = app.backends.keys().cloned().collect();
if !app.models.is_empty() { names.sort();
println!("\nmodels:"); for name in names {
for (name, m) in &app.models { let b = &app.backends[&name];
println!(" {}:", name); println!(" {}:", name);
println!(" backend: {:?}", m.backend); println!(" api_key: {} ({})", mask(&b.api_key), src(figment, &format!("backends.{name}.api_key")));
println!(" model_id: {:?}", m.model_id); if let Some(ref url) = b.base_url {
if let Some(ref pf) = m.prompt_file { println!(" base_url: {:?} ({})", url, src(figment, &format!("backends.{name}.base_url")));
println!(" prompt_file: {:?}", pf);
} }
if let Some(cw) = m.context_window { println!(" model_id: {:?}", b.model_id);
if let Some(cw) = b.context_window {
println!(" context_window: {}", cw); println!(" context_window: {}", cw);
} }
} }

448
src/config_writer.rs Normal file
View file

@ -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);
}
}

46
src/hippocampus/neuro/scoring.rs
View file

@ -230,10 +230,6 @@ fn consolidation_plan_inner(store: &Store, _detect_interf: bool) -> Consolidatio
rationale: Vec::new(), 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) // Target: α ≥ 2.5 (healthy scale-free)
if alpha < 2.0 { if alpha < 2.0 {
plan.add("linker", 100); plan.add("linker", 100);
@ -274,48 +270,6 @@ fn consolidation_plan_inner(store: &Store, _detect_interf: bool) -> Consolidatio
// Split: handle oversized nodes // Split: handle oversized nodes
plan.set("split", 5); 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 plan
} }

1
src/lib.rs
View file

@ -42,6 +42,7 @@ pub mod subconscious;
// Unified configuration // Unified configuration
pub mod config; pub mod config;
pub mod config_writer;
// Session state // Session state
pub mod session; pub mod session;

8
src/main.rs
View file

@ -482,6 +482,14 @@ async fn main() {
let cli = Cli::parse(); 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 { if let Err(e) = cli.command.run().await {
eprintln!("Error: {}", e); eprintln!("Error: {}", e);
process::exit(1); process::exit(1);

6
src/mind/log.rs
View file

@ -55,17 +55,13 @@ impl ConversationLog {
} }
pub fn oldest_timestamp(&self) -> Option<chrono::DateTime<chrono::Utc>> { 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 file = File::open(&self.path).ok()?;
let mmap = unsafe { Mmap::map(&file).ok()? }; let mmap = unsafe { Mmap::map(&file).ok()? };
// Find first { ... } and parse
for line in mmap.split(|&b| b == b'\n') { for line in mmap.split(|&b| b == b'\n') {
if line.is_empty() { continue; } if line.is_empty() { continue; }
if let Ok(node) = serde_json::from_slice::<AstNode>(line) { if let Ok(node) = serde_json::from_slice::<AstNode>(line) {
if let Some(leaf) = node.leaf() { if let Some(leaf) = node.leaf() {
if let Some(ts) = leaf.timestamp() { return Some(leaf.timestamp());
return Some(ts);
}
} }
} }
} }

304
src/mind/mod.rs
View file

@ -9,6 +9,44 @@ pub mod unconscious;
pub mod identity; pub mod identity;
pub mod log; 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 // consciousness.rs — Mind state machine and event loop
// //
// The core runtime for the consciousness binary. Mind manages turns, // 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::{Agent, TurnResult};
use crate::agent::api::ApiClient; use crate::agent::api::ApiClient;
use crate::config::{AppConfig, SessionConfig}; use crate::config::{AppConfig, SessionConfig};
use crate::subconscious::learn; use crate::subconscious::{compare, learn};
use crate::hippocampus::access_local; use crate::hippocampus::access_local;
pub use subconscious::{SubconsciousSnapshot, Subconscious}; pub use subconscious::{SubconsciousSnapshot, Subconscious};
@ -33,6 +71,36 @@ pub use unconscious::{UnconsciousSnapshot, Unconscious};
use crate::agent::context::{AstNode, NodeBody, Section, Ast, ContextState}; 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) { fn load_memory_scores(ctx: &mut ContextState, path: &std::path::Path) {
let data = match std::fs::read_to_string(path) { let data = match std::fs::read_to_string(path) {
Ok(d) => d, Ok(d) => d,
@ -42,25 +110,24 @@ fn load_memory_scores(ctx: &mut ContextState, path: &std::path::Path) {
Ok(s) => s, Ok(s) => s,
Err(_) => return, Err(_) => return,
}; };
let mut applied = 0; let identity_scores = match_scores(ctx.identity(), &scores);
for i in 0..ctx.conversation().len() { let conv_scores = match_scores(ctx.conversation(), &scores);
if let AstNode::Leaf(leaf) = &ctx.conversation()[i] { let applied = identity_scores.len() + conv_scores.len();
if let NodeBody::Memory { key, .. } = leaf.body() { for (i, s) in identity_scores {
if let Some(&s) = scores.get(key.as_str()) { ctx.set_score(Section::Identity, i, Some(s));
ctx.set_score(Section::Conversation, i, Some(s)); }
applied += 1; for (i, s) in conv_scores {
} ctx.set_score(Section::Conversation, i, Some(s));
}
}
} }
if applied > 0 { if applied > 0 {
dbglog!("[scoring] loaded {} scores from {}", applied, path.display()); dbglog!("[scoring] loaded {} scores from {}", applied, path.display());
} }
} }
/// Collect scored memory keys from conversation entries. /// Collect scored memory keys from identity and conversation entries.
fn collect_memory_scores(ctx: &ContextState) -> std::collections::BTreeMap<String, f64> { pub(crate) fn collect_memory_scores(ctx: &ContextState) -> std::collections::BTreeMap<String, f64> {
ctx.conversation().iter() ctx.identity().iter()
.chain(ctx.conversation().iter())
.filter_map(|node| { .filter_map(|node| {
if let AstNode::Leaf(leaf) = node { if let AstNode::Leaf(leaf) = node {
if let NodeBody::Memory { key, score: Some(s), .. } = leaf.body() { 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. /// Save memory scores to disk.
fn save_memory_scores(scores: &std::collections::BTreeMap<String, f64>, path: &std::path::Path) { pub(crate) fn save_memory_scores(scores: &std::collections::BTreeMap<String, f64>, path: &std::path::Path) {
if let Ok(json) = serde_json::to_string_pretty(scores) { match serde_json::to_string_pretty(scores) {
let _ = std::fs::write(path, json); Ok(json) => match std::fs::write(path, &json) {
dbglog!("[scoring] saved {} scores to {}", scores.len(), path.display()); 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, pub unc_idle: bool,
/// When the unconscious idle timer will fire (for UI display). /// When the unconscious idle timer will fire (for UI display).
pub unc_idle_deadline: Instant, 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 { impl Clone for MindState {
@ -136,6 +216,10 @@ impl Clone for MindState {
turn_handle: None, // Not cloned — only Mind's loop uses this turn_handle: None, // Not cloned — only Mind's loop uses this
unc_idle: self.unc_idle, unc_idle: self.unc_idle,
unc_idle_deadline: self.unc_idle_deadline, 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, Score,
/// Run full N×M memory scoring matrix (/score command) /// Run full N×M memory scoring matrix (/score command)
ScoreFull, 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 /// Abort current turn, kill processes
Interrupt, Interrupt,
/// Reset session /// Reset session
@ -173,6 +266,10 @@ impl MindState {
turn_handle: None, turn_handle: None,
unc_idle: false, unc_idle: false,
unc_idle_deadline: Instant::now() + std::time::Duration::from_secs(60), 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. /// 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) { if matches!(self.dmn, subconscious::State::Paused | subconscious::State::Off) {
return None; return None;
} }
@ -256,10 +353,6 @@ impl MindState {
} }
} }
/// Background task completion events.
enum BgEvent {
ScoringDone,
}
// --- Mind: cognitive state machine --- // --- Mind: cognitive state machine ---
@ -276,8 +369,9 @@ pub struct Mind {
/// Signals conscious activity to the unconscious loop. /// Signals conscious activity to the unconscious loop.
/// true = active, false = idle opportunity. /// true = active, false = idle opportunity.
conscious_active: tokio::sync::watch::Sender<bool>, conscious_active: tokio::sync::watch::Sender<bool>,
bg_tx: mpsc::UnboundedSender<BgEvent>, memory_scoring: learn::MemoryScoring,
bg_rx: std::sync::Mutex<Option<mpsc::UnboundedReceiver<BgEvent>>>, finetune_scoring: learn::FinetuneScoring,
compare_scoring: compare::CompareScoring,
_supervisor: crate::thalamus::supervisor::Supervisor, _supervisor: crate::thalamus::supervisor::Supervisor,
} }
@ -295,16 +389,28 @@ impl Mind {
client, client,
config.context_parts.clone(), config.context_parts.clone(),
config.app.clone(), config.app.clone(),
config.prompt_file.clone(),
conversation_log, conversation_log,
crate::agent::tools::ActiveTools::new(), crate::agent::tools::ActiveTools::new(),
crate::agent::tools::tools(), crate::agent::tools::tools(),
).await; ).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 (turn_watch, _) = tokio::sync::watch::channel(false);
let (conscious_active, _) = 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(); let mut sup = crate::thalamus::supervisor::Supervisor::new();
sup.load_config(); sup.load_config();
@ -389,10 +495,19 @@ 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, Self { agent, shared, config,
subconscious, unconscious, subconscious, unconscious,
turn_tx, turn_watch, conscious_active, bg_tx, turn_tx, turn_watch, conscious_active,
bg_rx: std::sync::Mutex::new(Some(bg_rx)), _supervisor: sup } memory_scoring,
finetune_scoring,
compare_scoring,
_supervisor: sup }
} }
/// Initialize — restore log, start daemons and background agents. /// Initialize — restore log, start daemons and background agents.
@ -434,6 +549,10 @@ impl Mind {
// Load persistent subconscious state // Load persistent subconscious state
let state_path = self.config.session_dir.join("subconscious-state.json"); let state_path = self.config.session_dir.join("subconscious-state.json");
self.subconscious.lock().await.set_state_path(state_path); 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> { pub fn turn_watch(&self) -> tokio::sync::watch::Receiver<bool> {
@ -453,24 +572,10 @@ impl Mind {
} }
} }
MindCommand::Score => { MindCommand::Score => {
let mut s = self.shared.lock().unwrap(); self.memory_scoring.trigger();
if !s.scoring_in_flight {
s.scoring_in_flight = true;
drop(s);
self.start_memory_scoring();
} else {
dbglog!("[scoring] skipped: scoring_in_flight=true");
}
} }
MindCommand::ScoreFull => { MindCommand::ScoreFull => {
let mut s = self.shared.lock().unwrap(); self.memory_scoring.trigger_full();
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");
}
} }
MindCommand::Interrupt => { MindCommand::Interrupt => {
self.shared.lock().unwrap().interrupt(); self.shared.lock().unwrap().interrupt();
@ -500,83 +605,27 @@ impl Mind {
} }
self.agent.compact().await; 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) { async fn start_turn(&self, text: &str, target: StreamTarget) {
{ {
@ -639,9 +688,13 @@ 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 mut 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 { loop {
let (timeout, has_input) = { let (timeout, has_input) = {
let me = self.shared.lock().unwrap(); let me = self.shared.lock().unwrap();
@ -662,14 +715,6 @@ 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() => { Some((result, target)) = turn_rx.recv() => {
let _ = self.conscious_active.send(false); let _ = self.conscious_active.send(false);
let model_switch = { let model_switch = {
@ -686,6 +731,7 @@ impl Mind {
cmds.push(MindCommand::Compact); cmds.push(MindCommand::Compact);
if !self.config.no_agents { if !self.config.no_agents {
cmds.push(MindCommand::Score); cmds.push(MindCommand::Score);
cmds.push(MindCommand::ScoreFinetune);
} }
} }

20
src/mind/subconscious.rs
View file

@ -20,6 +20,7 @@
use std::path::PathBuf; use std::path::PathBuf;
use std::time::{Duration, Instant}; use std::time::{Duration, Instant};
use crate::thalamus::idle::{hours_since_last_dream, DREAM_INTERVAL_HOURS};
/// DMN state machine. /// DMN state machine.
#[derive(Debug, Clone)] #[derive(Debug, Clone)]
@ -91,7 +92,8 @@ impl State {
/// Generate the DMN prompt for the current state, informed by /// Generate the DMN prompt for the current state, informed by
/// user presence and error patterns. /// user presence and error patterns.
pub fn prompt(&self, ctx: &DmnContext) -> String { 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) { let idle_info = if ctx.user_idle < Duration::from_secs(60) {
format!("{} is here (active recently).", user) format!("{} is here (active recently).", user)
@ -138,10 +140,22 @@ impl State {
) )
} }
State::Foraging => { 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!( format!(
"[dmn] Foraging time. {} Follow whatever catches your attention — \ "[dmn] Foraging time. {} Follow whatever catches your attention — \
memory files, code, ideas. Call yield_to_user when you want to rest.{}", memory files, code, ideas. Call yield_to_user when you want to rest.{}{}",
idle_info, stuck_warning idle_info, dream_hint, stuck_warning
) )
} }
State::Resting { since } => { State::Resting { since } => {

25
src/mind/unconscious.rs
View file

@ -275,17 +275,7 @@ pub async fn prepare_spawn(name: &str, mut auto: AutoAgent, wake: std::sync::Arc
phase: s.phase.clone(), phase: s.phase.clone(),
}).collect()); }).collect());
// Create standalone Agent — stored so UI can read context // 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);
}
let cli = crate::user::CliArgs::default(); let cli = crate::user::CliArgs::default();
let (app, _) = match crate::config::load_app(&cli) { let (app, _) = match crate::config::load_app(&cli) {
Ok(r) => r, Ok(r) => r,
@ -295,12 +285,21 @@ pub async fn prepare_spawn(name: &str, mut auto: AutoAgent, wake: std::sync::Arc
return Err(auto); 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. // Unconscious agents have self-contained prompts — no standard context.
let client = crate::agent::api::ApiClient::new(base_url, api_key, model); let client = crate::agent::api::ApiClient::new(
&resolved.api_base, &resolved.api_key, &resolved.model_id);
let agent = crate::agent::Agent::new( let agent = crate::agent::Agent::new(
client, Vec::new(), client, Vec::new(),
app, String::new(), None, app, None,
crate::agent::tools::ActiveTools::new(), crate::agent::tools::ActiveTools::new(),
auto.tools.clone(), auto.tools.clone(),
).await; ).await;

46
src/subconscious/agents/bail-no-competing.sh
View file

@ -1,21 +1,49 @@
#!/bin/bash #!/bin/bash
# Bail if other agents are alive in the state dir. # Bail if another agent is in the same phase-group as us.
# $1 = this agent's pid file name (e.g. pid-12345)
# cwd = state dir
# #
# 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 shopt -s nullglob
my_pid_file="$1" 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 for f in pid-*; do
[[ $f == $my_pid_file ]] && continue [[ "$f" == "$my_pid_file" ]] && continue
pid="${f#pid-}" pid="${f#pid-}"
if kill -0 "$pid" 2>/dev/null; then if ! kill -0 "$pid" 2>/dev/null; then
exit 1 # competing agent is alive rm -f "$f" # stale pid file, clean up
else continue
rm -f "$f" # stale pid file, clean up fi
other_phase=$(cat "$f" 2>/dev/null)
other_group=$(group_of "$other_phase")
if [[ "$my_group" == "$other_group" ]]; then
exit 1
fi fi
done done

109
src/subconscious/compare.rs Normal file
View file

@ -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(); }
}
}

15
src/subconscious/defs.rs
View file

@ -396,13 +396,14 @@ fn resolve_conversation(budget: Option<usize>) -> String {
let cfg = crate::config::get(); let cfg = crate::config::get();
let max_bytes = budget.unwrap_or_else(|| cfg.surface_conversation_bytes.unwrap_or(100_000)); 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 fragments: Vec<String> = Vec::new();
let mut total_bytes = 0; let mut total_bytes = 0;
let mut oldest_ts = String::new(); let mut oldest_ts = String::new();
for (role, content, ts) in iter { for (role, content, ts) in iter {
if total_bytes >= max_bytes { break; } if total_bytes >= max_bytes { break; }
let name = if role == "user" { &cfg.user_name } else { &cfg.assistant_name }; let name = if role == "user" { &app.user_name } else { &app.assistant_name };
let formatted = if !ts.is_empty() { let formatted = if !ts.is_empty() {
oldest_ts = ts[..ts.floor_char_boundary(ts.len().min(19))].to_string(); oldest_ts = ts[..ts.floor_char_boundary(ts.len().min(19))].to_string();
format!("**{}** {}: {}", name, &oldest_ts, content) format!("**{}** {}: {}", name, &oldest_ts, content)
@ -623,11 +624,13 @@ pub async fn run_agent(
let mut all_keys = keys; let mut all_keys = keys;
let mut resolved_steps = Vec::new(); let mut resolved_steps = Vec::new();
for step in &def.steps { for step in &def.steps {
let cfg = crate::config::get(); let template = {
let template = step.prompt let app = crate::config::app();
.replace("{agent_name}", &def.agent) step.prompt
.replace("{user_name}", &cfg.user_name) .replace("{agent_name}", &def.agent)
.replace("{assistant_name}", &cfg.assistant_name); .replace("{user_name}", &app.user_name)
.replace("{assistant_name}", &app.assistant_name)
};
let (prompt, extra_keys) = resolve_placeholders(&template, &all_keys, count).await; let (prompt, extra_keys) = resolve_placeholders(&template, &all_keys, count).await;
all_keys.extend(extra_keys); all_keys.extend(extra_keys);
resolved_steps.push(super::prompts::ResolvedStep { resolved_steps.push(super::prompts::ResolvedStep {

46
src/subconscious/generate.rs Normal file
View file

@ -0,0 +1,46 @@
// 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 crate::agent::api::{ApiClient, SamplingParams, StreamToken};
use crate::agent::context::{AstNode, ContextState};
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.
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 prompt, images, _) = context.wire_prompt(0..entry_idx, skip);
prompt.push(tokenizer::IM_START);
prompt.extend(tokenizer::encode("assistant\n"));
let sampling = SamplingParams {
temperature: 0.6,
top_p: 0.95,
top_k: 20,
};
let (mut rx, _guard) = client.stream_completion_mm(&prompt, &images, sampling, Some(-5));
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))
}

554
src/subconscious/learn.rs
View file

@ -14,75 +14,18 @@
// with high divergence depend on memories the model // with high divergence depend on memories the model
// hasn't internalized. 2 API calls. // hasn't internalized. 2 API calls.
use std::sync::Arc;
use crate::agent::api::ApiClient; use crate::agent::api::ApiClient;
use crate::agent::context::{AstNode, Ast, NodeBody, ContextState, Role}; use crate::agent::context::{
Ast, AstNode, ContextState, Role, WireImage,
is_assistant, is_memory_node, memory_key, render_branch_text, render_prior_context,
};
use crate::mind::{MindState, MindTriggered, TaskHandle};
use crate::subconscious::generate::gen_continuation;
const SCORE_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(300); 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
}
// ── Score API ─────────────────────────────────────────────────── // ── Score API ───────────────────────────────────────────────────
#[derive(serde::Deserialize)] #[derive(serde::Deserialize)]
@ -105,15 +48,30 @@ async fn call_score(
http: &crate::agent::api::http::HttpClient, http: &crate::agent::api::http::HttpClient,
client: &ApiClient, client: &ApiClient,
prompt: &[u32], prompt: &[u32],
images: &[WireImage],
ranges: &[(usize, usize)],
priority: Option<i32>, priority: Option<i32>,
) -> anyhow::Result<Vec<ScoreResult>> { ) -> anyhow::Result<Vec<ScoreResult>> {
// Nothing to score — skip the round-trip.
if ranges.is_empty() {
return Ok(Vec::new());
}
let url = format!("{}/score", client.base_url()); let url = format!("{}/score", client.base_url());
let auth = format!("Bearer {}", client.api_key()); let auth = format!("Bearer {}", client.api_key());
let mut body = serde_json::json!({ let mut body = serde_json::json!({
"model": client.model, "model": client.model,
"prompt": prompt, "prompt": prompt,
"score_ranges": ranges,
"logprobs": 1, "logprobs": 1,
}); });
if !images.is_empty() {
use base64::Engine;
let b64 = base64::engine::general_purpose::STANDARD;
let uris: Vec<String> = images.iter()
.map(|img| format!("data:{};base64,{}", img.mime, b64.encode(&img.bytes)))
.collect();
body["multi_modal_data"] = serde_json::json!({ "image": uris });
}
if let Some(p) = priority { if let Some(p) = priority {
body["priority"] = serde_json::json!(p); body["priority"] = serde_json::json!(p);
} }
@ -151,16 +109,24 @@ fn divergence(baseline: &[ScoreResult], without: &[ScoreResult]) -> Vec<f64> {
} }
/// Score two message sets and return total divergence. /// Score two message sets and return total divergence.
async fn score_divergence( async fn score_divergence<F>(
http: &crate::agent::api::http::HttpClient, http: &crate::agent::api::http::HttpClient,
client: &ApiClient, client: &ApiClient,
context: &ContextState, context: &ContextState,
range: std::ops::Range<usize>, range: std::ops::Range<usize>,
filter: Filter<'_>, skip: F,
priority: Option<i32>, priority: Option<i32>,
) -> anyhow::Result<(Vec<f64>, Vec<ScoreResult>)> { ) -> anyhow::Result<(Vec<f64>, Vec<ScoreResult>)>
let baseline = call_score(http, client, &build_token_ids(context, range.clone(), Filter::None), priority).await?; where F: FnMut(&AstNode) -> bool,
let without = call_score(http, client, &build_token_ids(context, range, filter), priority).await?; {
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(http, client, &baseline_tokens, &baseline_images,
&baseline_ranges, priority).await?;
let without = call_score(http, client, &without_tokens, &without_images,
&without_ranges, priority).await?;
let divs = divergence(&baseline, &without); let divs = divergence(&baseline, &without);
Ok((divs, baseline)) Ok((divs, baseline))
} }
@ -175,7 +141,9 @@ pub async fn score_memories(
// Collect memory keys and response indices under a brief lock // Collect memory keys and response indices under a brief lock
let (memory_keys, response_indices) = { let (memory_keys, response_indices) = {
let ctx = agent.context.lock().await; 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)) .filter_map(|node| memory_key(node).map(String::from))
.collect(); .collect();
keys.dedup(); keys.dedup();
@ -197,21 +165,22 @@ pub async fn score_memories(
let http = http_client(); let http = http_client();
let activity = crate::agent::start_activity(agent, "scoring: baseline").await; 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; 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(&http, client, &baseline_tokens, &baseline_images,
&baseline_ranges, Some(5)).await?;
dbglog!("[scoring-full] baseline done ({} response scores)", baseline.len()); dbglog!("[scoring-full] baseline done ({} response scores)", baseline.len());
for (mem_idx, key) in memory_keys.iter().enumerate() { for (mem_idx, key) in memory_keys.iter().enumerate() {
activity.update(format!("scoring: {}/{}", mem_idx + 1, total)).await; activity.update(format!("scoring: {}/{}", mem_idx + 1, total)).await;
dbglog!("[scoring-full] {}/{}: {}", mem_idx + 1, total, key); dbglog!("[scoring-full] {}/{}: {}", mem_idx + 1, total, key);
let tokens = { let (tokens, images, ranges) = {
let ctx = agent.context.lock().await; 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(&http, client, &tokens, &images, &ranges, Some(5)).await {
Ok(without) => { Ok(without) => {
let divs = divergence(&baseline, &without); let divs = divergence(&baseline, &without);
let max_div = divs.iter().cloned().fold(0.0f64, f64::max); let max_div = divs.iter().cloned().fold(0.0f64, f64::max);
@ -295,7 +264,8 @@ pub async fn score_memory(
} }
let http = http_client(); let http = http_client();
let (divs, _) = score_divergence(&http, client, context, range, Filter::SkipKey(key), Some(5)).await?; let (divs, _) = score_divergence(&http, client, context, range,
|n| memory_key(n) == Some(key), Some(5)).await?;
Ok(divs.iter().sum()) Ok(divs.iter().sum())
} }
@ -331,7 +301,10 @@ where
{ {
let store = &*store_arc; 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 let Some(key) = memory_key(node) {
if !seen.insert(key.to_owned()) { continue; } if !seen.insert(key.to_owned()) { continue; }
let last_scored = store.get_node(key) let last_scored = store.get_node(key)
@ -340,7 +313,7 @@ where
.map(|n| n.last_scored) .map(|n| n.last_scored)
.unwrap_or(0); .unwrap_or(0);
if now - last_scored >= max_age_secs { if now - last_scored >= max_age_secs {
candidates.push((i, key.to_owned(), last_scored)); candidates.push((pos, key.to_owned(), last_scored));
} }
} }
} }
@ -384,7 +357,8 @@ where
} }
activity.update(format!("scoring: {}/{} {}", scored + 1, total, key)).await; 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(&http, client, context, range,
|n| memory_key(n) == Some(key), Some(5)).await {
Ok((divs, _)) => { Ok((divs, _)) => {
let n_responses = divs.len(); let n_responses = divs.len();
let max_div = divs.iter().cloned().fold(0.0f64, f64::max); let max_div = divs.iter().cloned().fold(0.0f64, f64::max);
@ -405,6 +379,108 @@ where
Ok(scored) 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 ───────────────────────────────────────── // ── Fine-tuning scoring ─────────────────────────────────────────
/// Score which recent responses are candidates for fine-tuning. /// Score which recent responses are candidates for fine-tuning.
@ -430,7 +506,7 @@ pub async fn score_finetune(
} }
let http = http_client(); let http = http_client();
let (divs, _) = score_divergence(&http, client, context, range, Filter::SkipAllMemories, Some(5)).await?; let (divs, _) = score_divergence(&http, client, context, range, is_memory_node, Some(5)).await?;
let mut results: Vec<(usize, f64)> = response_positions.iter() let mut results: Vec<(usize, f64)> = response_positions.iter()
.enumerate() .enumerate()
@ -439,3 +515,317 @@ pub async fn score_finetune(
results.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal)); results.sort_by(|a, b| b.1.partial_cmp(&a.1).unwrap_or(std::cmp::Ordering::Equal));
Ok(results) 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 http = http_client();
let url = format!("{}/train", client.base_url());
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)
}

2
src/subconscious/mod.rs
View file

@ -1,7 +1,9 @@
// Agent layer: LLM-powered operations on the memory graph // Agent layer: LLM-powered operations on the memory graph
pub mod compare;
pub mod daemon; pub mod daemon;
pub mod defs; pub mod defs;
pub mod digest; pub mod digest;
pub mod generate;
pub mod learn; pub mod learn;
pub mod prompts; pub mod prompts;

4
src/thalamus/idle.rs
View file

@ -372,6 +372,10 @@ impl State {
} }
pub fn hours_since_last_dream() -> u64 { 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 path = home().join(".consciousness/logs/dream-log.jsonl");
let content = match fs::read_to_string(path) { let content = match fs::read_to_string(path) {
Ok(c) if !c.is_empty() => c, Ok(c) if !c.is_empty() => c,

45
src/thalamus/supervisor.rs
View file

@ -19,6 +19,51 @@ fn channels_dir() -> PathBuf {
.join(".consciousness/channels") .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 { fn config_path() -> PathBuf {
channels_dir().join("channels.json5") channels_dir().join("channels.json5")
} }

400
src/user/amygdala.rs Normal file
View file

@ -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)]) }
}

29
src/user/chat.rs
View file

@ -112,13 +112,7 @@ pub async fn cmd_switch_model(
let _new_client = crate::agent::api::ApiClient::new( let _new_client = crate::agent::api::ApiClient::new(
&resolved.api_base, &resolved.api_key, &resolved.model_id, &resolved.api_base, &resolved.api_key, &resolved.model_id,
); );
let prompt_changed = resolved.prompt_file != agent.prompt_file; agent.state.lock().await.notify(format!("switched to {}", resolved.model_id));
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));
}
} }
fn notify_help(agent: &std::sync::Arc<crate::agent::Agent>) { fn notify_help(agent: &std::sync::Arc<crate::agent::Agent>) {
@ -173,6 +167,7 @@ enum PaneTarget {
ConversationAssistant, ConversationAssistant,
Tools, Tools,
ToolResult, ToolResult,
Autonomous,
} }
const MAX_PANE_LINES: usize = 10_000; const MAX_PANE_LINES: usize = 10_000;
@ -478,8 +473,11 @@ impl InteractScreen {
AstNode::Leaf(leaf) => { AstNode::Leaf(leaf) => {
let text = leaf.body().text().to_string(); let text = leaf.body().text().to_string();
match leaf.body() { match leaf.body() {
NodeBody::Memory { .. } | NodeBody::Thinking(_) NodeBody::Memory { .. } | NodeBody::Log(_) | NodeBody::Dmn(_) => vec![],
| NodeBody::Log(_) | NodeBody::Dmn(_) => vec![], NodeBody::Thinking(_) => {
if text.is_empty() { vec![] }
else { vec![(PaneTarget::Autonomous, text, Marker::None)] }
}
NodeBody::Content(_) => { NodeBody::Content(_) => {
if text.is_empty() || text.starts_with("<system-reminder>") { vec![] } if text.is_empty() || text.starts_with("<system-reminder>") { vec![] }
else { vec![(PaneTarget::Conversation, text, Marker::User)] } else { vec![(PaneTarget::Conversation, text, Marker::User)] }
@ -492,6 +490,11 @@ impl InteractScreen {
if t.is_empty() { vec![] } if t.is_empty() { vec![] }
else { vec![(PaneTarget::ToolResult, text, Marker::None)] } 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, .. } => { AstNode::Branch { role, children, .. } => {
@ -548,6 +551,12 @@ impl InteractScreen {
self.tools.push_line(format!(" {}", line), Color::DarkGray); 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(), => self.conversation.pop_line(),
PaneTarget::Tools | PaneTarget::ToolResult PaneTarget::Tools | PaneTarget::ToolResult
=> self.tools.pop_line(), => self.tools.pop_line(),
PaneTarget::Autonomous
=> self.autonomous.pop_line(),
} }
} }
} }

111
src/user/compare.rs Normal file
View file

@ -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);
}
}

20
src/user/context.rs
View file

@ -38,16 +38,13 @@ impl ConsciousScreen {
for node in ctx.conversation() { for node in ctx.conversation() {
if let AstNode::Leaf(leaf) = node { if let AstNode::Leaf(leaf) = node {
if let NodeBody::Memory { key, score, text } = leaf.body() { if let NodeBody::Memory { key, score, text } = leaf.body() {
let status = match score { if score.is_some() { scored += 1; } else { unscored += 1; }
Some(s) => { scored += 1; format!("{:.2}", s) }
None => { unscored += 1; String::new() }
};
mem_children.push(SectionView { mem_children.push(SectionView {
name: key.clone(), name: format!("mem: {}", key),
tokens: node.tokens(), tokens: node.tokens(),
content: text.clone(), content: text.clone(),
children: Vec::new(), children: Vec::new(),
status, status: score.map(|s| format!("{:.2}", s)).unwrap_or_default(),
}); });
} }
} }
@ -129,14 +126,7 @@ impl ScreenView for ConsciousScreen {
let section_style = Style::default().fg(Color::Yellow); let section_style = Style::default().fg(Color::Yellow);
lines.push(Line::styled("── Model ──", section_style)); lines.push(Line::styled("── Model ──", section_style));
let model_display = app.context_info.as_ref() lines.push(Line::raw(format!(" Current: {}", app.status.model)));
.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("")); lines.push(Line::raw(""));
lines.push(Line::styled("── Context State ──", section_style)); lines.push(Line::styled("── Context State ──", section_style));
@ -156,8 +146,6 @@ impl ScreenView for ConsciousScreen {
lines.push(Line::raw(format!(" {:53} {:>6} tokens", "────────", "──────"))); lines.push(Line::raw(format!(" {:53} {:>6} tokens", "────────", "──────")));
lines.push(Line::raw(format!(" {:53} {:>6} tokens", "Total", total))); 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("")); lines.push(Line::raw(""));

284
src/user/learn.rs Normal file
View file

@ -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);
}

140
src/user/mod.rs
View file

@ -3,13 +3,16 @@
// TUI, UI channel, parsing. The cognitive layer (session state // TUI, UI channel, parsing. The cognitive layer (session state
// machine, DMN, identity) lives in mind/. // machine, DMN, identity) lives in mind/.
pub(crate) mod amygdala;
pub(crate) mod chat; pub(crate) mod chat;
pub(crate) mod compare;
mod context; mod context;
pub(crate) mod learn;
pub(crate) mod scroll_pane; pub(crate) mod scroll_pane;
pub mod selectable; pub mod selectable;
mod subconscious; mod subconscious;
mod unconscious;
mod thalamus; mod thalamus;
mod unconscious;
mod widgets; mod widgets;
use anyhow::Result; use anyhow::Result;
@ -44,15 +47,6 @@ struct StatusInfo {
} }
/// Context loading details for the debug screen. /// 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. /// Build the screen legend from screen labels.
fn screen_legend_from(screens: &[Box<dyn ScreenView>]) -> String { fn screen_legend_from(screens: &[Box<dyn ScreenView>]) -> String {
let parts: Vec<String> = screens.iter().enumerate() let parts: Vec<String> = screens.iter().enumerate()
@ -72,6 +66,13 @@ fn screen_legend() -> String {
SCREEN_LEGEND.get().cloned().unwrap_or_default() 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. /// A screen that can draw itself and handle input.
trait ScreenView: Send { trait ScreenView: Send {
fn tick(&mut self, frame: &mut ratatui::Frame, area: ratatui::layout::Rect, fn tick(&mut self, frame: &mut ratatui::Frame, area: ratatui::layout::Rect,
@ -109,7 +110,6 @@ struct App {
top_k: u32, top_k: u32,
agent: std::sync::Arc<crate::agent::Agent>, agent: std::sync::Arc<crate::agent::Agent>,
should_quit: bool, should_quit: bool,
context_info: Option<ContextInfo>,
agent_state: Vec<crate::mind::SubconsciousSnapshot>, agent_state: Vec<crate::mind::SubconsciousSnapshot>,
unconscious_state: Vec<crate::mind::UnconsciousSnapshot>, unconscious_state: Vec<crate::mind::UnconsciousSnapshot>,
mind_state: Option<crate::mind::MindState>, mind_state: Option<crate::mind::MindState>,
@ -121,6 +121,10 @@ struct App {
walked_count: usize, walked_count: usize,
channel_status: Vec<ChannelStatus>, channel_status: Vec<ChannelStatus>,
idle_info: Option<IdleInfo>, 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 { impl App {
@ -142,7 +146,6 @@ impl App {
top_k: 20, top_k: 20,
agent, agent,
should_quit: false, should_quit: false,
context_info: None,
agent_state: Vec::new(), agent_state: Vec::new(),
unconscious_state: Vec::new(), unconscious_state: Vec::new(),
mind_state: None, mind_state: None,
@ -151,9 +154,53 @@ impl App {
rebuild_tools_pending: false, rebuild_tools_pending: false,
walked_count: 0, walked_count: 0,
channel_status: Vec::new(), idle_info: None, 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)>) { fn set_channel_status(&mut self, channels: Vec<(String, bool, u32)>) {
self.channel_status = channels.into_iter() 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<()> { async fn start(cli: crate::user::CliArgs) -> Result<()> {
let (config, _figment) = crate::config::load_session(&cli).await?; 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 { if config.app.debug {
unsafe { std::env::set_var("POC_DEBUG", "1") }; unsafe { std::env::set_var("POC_DEBUG", "1") };
} }
@ -334,7 +384,7 @@ async fn run(
} }
let notify_rx = crate::thalamus::channels::subscribe_all(); 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![ let mut screens: Vec<Box<dyn tui::ScreenView>> = vec![
Box::new(crate::user::chat::InteractScreen::new( Box::new(crate::user::chat::InteractScreen::new(
mind.agent.clone(), mind.shared.clone(), mind_tx.clone(), mind.agent.clone(), mind.shared.clone(), mind_tx.clone(),
@ -343,6 +393,9 @@ async fn run(
Box::new(crate::user::subconscious::SubconsciousScreen::new()), Box::new(crate::user::subconscious::SubconsciousScreen::new()),
Box::new(crate::user::unconscious::UnconsciousScreen::new()), Box::new(crate::user::unconscious::UnconsciousScreen::new()),
Box::new(crate::user::thalamus::ThalamusScreen::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 let mut active_screen: usize = 1; // F-key number
tui::set_screen_legend(tui::screen_legend_from(&*screens)); tui::set_screen_legend(tui::screen_legend_from(&*screens));
@ -419,7 +472,8 @@ async fn run(
idle_state.decay_ewma(); idle_state.decay_ewma();
app.update_idle(&idle_state); app.update_idle(&idle_state);
app.agent_state = mind.subconscious_snapshots().await; 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(); let toggles: Vec<String> = app.agent_toggles.drain(..).collect();
for name in &toggles { for name in &toggles {
if mind.subconscious.lock().await.toggle(name).is_none() { if mind.subconscious.lock().await.toggle(name).is_none() {
@ -433,7 +487,42 @@ async fn run(
}; };
app.unconscious_state = unc.snapshots(store_guard.as_deref()); app.unconscious_state = unc.snapshots(store_guard.as_deref());
app.graph_health = unc.graph_health.clone(); 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(); app.walked_count = mind.subconscious_walked().await.len();
if !startup_done { if !startup_done {
@ -530,16 +619,11 @@ async fn run(
// --- CLI --- // --- CLI ---
use clap::{Parser, Subcommand}; 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")] #[command(name = "consciousness", about = "Substrate-independent AI agent")]
pub struct CliArgs { pub struct CliArgs {
/// Select active backend ("anthropic" or "openrouter") /// Model override (selects a named entry from `models` in config.json5)
#[arg(long)]
pub backend: Option<String>,
/// Model override
#[arg(short, long)] #[arg(short, long)]
pub model: Option<String>, pub model: Option<String>,
@ -559,10 +643,6 @@ pub struct CliArgs {
#[arg(long)] #[arg(long)]
pub show_config: bool, pub show_config: bool,
/// Project memory directory
#[arg(long)]
pub memory_project: Option<PathBuf>,
/// Max consecutive DMN turns /// Max consecutive DMN turns
#[arg(long)] #[arg(long)]
pub dmn_max_turns: Option<u32>, pub dmn_max_turns: Option<u32>,
@ -575,7 +655,7 @@ pub struct CliArgs {
pub command: Option<SubCmd>, pub command: Option<SubCmd>,
} }
#[derive(Subcommand, Debug)] #[derive(Subcommand, Debug, Clone)]
pub enum SubCmd { pub enum SubCmd {
/// Print new output since last read and exit /// Print new output since last read and exit
Read { Read {
@ -676,8 +756,10 @@ fn restore_stderr(original_fd: std::os::fd::RawFd) {
#[tokio::main] #[tokio::main]
pub async fn main() { pub async fn main() {
// Auto-reap child processes (channel daemons outlive the supervisor) // Reap channel-daemon zombies via a SIGCHLD handler that only touches
unsafe { libc::signal(libc::SIGCHLD, libc::SIG_IGN); } // 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 // Redirect stderr to pipe — logs to file and sends to channel for UI display
let stderr_capture = redirect_stderr_to_pipe(); let stderr_capture = redirect_stderr_to_pipe();

92
src/user/widgets.rs
View file

@ -6,7 +6,7 @@ use ratatui::{
widgets::{Block, Borders}, widgets::{Block, Borders},
crossterm::event::KeyCode, crossterm::event::KeyCode,
}; };
use crate::agent::context::{AstNode, Ast}; use crate::agent::context::{AstNode, Ast, NodeBody};
#[derive(Debug, Clone)] #[derive(Debug, Clone)]
pub struct SectionView { pub struct SectionView {
@ -20,13 +20,22 @@ pub struct SectionView {
fn node_to_view(node: &AstNode) -> SectionView { fn node_to_view(node: &AstNode) -> SectionView {
match node { match node {
AstNode::Leaf(leaf) => SectionView { AstNode::Leaf(leaf) => {
name: node.label(), let (name, status) = match leaf.body() {
tokens: node.tokens(), NodeBody::Memory { key, score, .. } => {
content: leaf.body().text().to_string(), let s = score.map(|v| format!("{:.2}", v)).unwrap_or_default();
children: Vec::new(), (format!("mem: {}", key), s)
status: String::new(), }
}, _ => (node.label(), String::new()),
};
SectionView {
name,
tokens: node.tokens(),
content: leaf.body().text().to_string(),
children: Vec::new(),
status,
}
}
AstNode::Branch { children, .. } => { AstNode::Branch { children, .. } => {
let child_views: Vec<SectionView> = children.iter() let child_views: Vec<SectionView> = children.iter()
.map(|c| node_to_view(c)) .map(|c| node_to_view(c))
@ -100,6 +109,73 @@ pub fn tree_legend() -> Line<'static> {
) )
} }
// ---------------------------------------------------------------------------
// 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 // SectionTree — expand/collapse tree renderer for ContextSection

121
training/DESIGN.md
View file

@ -3,7 +3,7 @@
## Overview ## Overview
Continuous fine-tuning of Qwen3.5-27B alongside live vLLM inference. 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. gradient projection. No pause required — HOGWILD concurrent training.
Weights shared via CUDA IPC between vLLM and the training process. 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) │ │ │ │ Model Weights (54GB, bf16) │ │
│ │ Shared via CUDA IPC │ │ │ │ Shared: vLLM inference + HF training │ │
│ └──────────────┬──────────────┬────────────────┘ │ │ └──────────────┬──────────────┬────────────────┘ │
│ │ │ │ │ │ │ │
│ ┌──────────────▼──┐ ┌───────▼────────────────┐ │ │ ┌──────────────▼──┐ ┌───────▼────────────────┐ │
│ │ vLLM (inference)│ │ Apollo (training) │ │ │ │ vLLM (inference)│ │ Training subprocess │ │
│ │ KV cache ~60GB │ │ Gradients ~54GB │ │ │ │ KV cache ~60GB │ │ HF model wrapper │ │
│ │ Serves requests │ │ Optimizer state ~10GB │ │ │ │ /completions │ │ Apollo optimizer ~2.5GB │ │
│ │ Never paused │ │ Activations ~10GB │ │ │ │ /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 │ │ Training signal │ HTTP │ /completions │
│ agent │──────────>│ daemon │ │ agent │──────────>│ /score │
│ │ │ │ │ │ │ /train │
│ Dream loop │ │ Checkpoint sync │ │ Dream loop │ │ /checkpoint │
│ (generates │ │ (mmap + diff, │ │ (generates │ │ /train/status │
│ scenarios) │ │ every 10 min) │ │ scenarios) │ │ │
└──────────────────┘ └──────────────────┘ └──────────────────┘ └──────────────────┘
``` ```
@ -59,10 +75,9 @@ LoRA trains adapter matrices, not base weights. For personality and
behavioral changes that persist as disposition, the base weights behavioral changes that persist as disposition, the base weights
need to change. Apollo makes this memory-feasible. need to change. Apollo makes this memory-feasible.
### Rank 256 ### Rank 64
Not Mini (rank-1). With 100+ diverse training examples, the Not Mini (rank-1). Rank-64 captures gradient structure across diverse
gradient's effective dimensionality can reach hundreds. Rank-256 training examples while keeping memory low (~2.5GB on 27B model).
captures the structure. Memory cost: ~10GB (negligible on B200).
Compute cost: <0.25% of forward+backward. Compute cost: <0.25% of forward+backward.
### Channel-wise scaling ### Channel-wise scaling
@ -90,7 +105,7 @@ from a per-parameter seed each step.
### Parameter grouping (Qwen3.5 gotcha) ### Parameter grouping (Qwen3.5 gotcha)
conv1d weights are 3D tensors [10240, 1, 4]. Apollo's projector conv1d weights are 3D tensors [10240, 1, 4]. Apollo's projector
needs 2D matrices with min dimension >= rank. Small/3D tensors 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 ## 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 regions. For small behavioral updates, turns a 54GB write into
a few hundred MB. 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 - 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 ## Hyperparameters
| Parameter | Value | Rationale | | Parameter | Value | Rationale |
|-----------|-------|-----------| |-----------|-------|-----------|
| Learning rate | 1e-5 to 1e-4 | Standard for full fine-tuning. Higher for diverse batches. | | 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. | | Scale type | channel | Per-channel precision, matches LLaMA-Factory defaults. |
| Epochs | 1 | One pass over diverse data. Multiple epochs risk overfitting. | | Epochs | 1 | One pass over diverse data. Multiple epochs risk overfitting. |
| Batch size | 1 | Single examples, immediate updates. | | Batch size | 1 | Single examples, immediate updates. |
@ -220,34 +261,32 @@ a few hundred MB.
## Components ## Components
### Built ✓ ### Built ✓
- `apollo_mini.py` — Apollo optimizer (configurable rank, default 256) - `optimizer.py` — Apollo optimizer (configurable rank)
- `apollo_worker.py` — HTTP daemon (aiohttp, job tracking) - `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) - `weight_mapping.py` — vLLM merged → HF separate views (validated)
- `training_example.py` — tokenization with chat template - `export_hook.py` — vLLM plugin hook for IPC handle export
- `vllm_export_hook.py` — source patch for IPC handle export - `checkpoint_sync.py` — mmap + diff checkpoint sync (Python)
- `checkpoint/` — Rust tool for mmap + diff checkpoint sync
### To build ### 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 - **Training-signal agent**: flags moments in conversation logs
- **Instruction stripping**: remove scaffolding from training examples - **Instruction stripping**: remove scaffolding from training examples
- **Quality monitoring**: track model capability over time - **Quality monitoring**: track model capability over time
- **HF model forward pass integration**: wire into apollo_worker
## Files ## Files
``` ```
training/ training/
DESIGN.md — this document DESIGN.md — this document
apollo_mini.py — Apollo optimizer pyproject.toml — package config, vLLM plugin entry point
apollo_worker.py — HTTP training daemon apollo_plugin/
weight_mapping.py — vLLM ↔ HF weight views __init__.py — plugin registration
training_example.py — tokenization helpers export_hook.py — patches vLLM worker to export IPC handles
export_weights.py — standalone weight export (unused) train_router.py — /train endpoint, forwards to worker via ZMQ
vllm_export_hook.py — vLLM source patch for IPC export training_worker.py — training subprocess (HF model, Apollo, checkpoint)
start_vllm_with_apollo.sh — vLLM launcher (unused, using source patch) optimizer.py — Apollo optimizer
train.py — standalone training script (alternative) weight_mapping.py — vLLM ↔ HF weight views
checkpoint/ checkpoint_sync.py — mmap + diff sync to safetensors
Cargo.toml — Rust checkpoint tool steering.py — steering vector extraction (experimental)
src/main.rs — mmap + diff sync
``` ```

64
training/amygdala_stories/README.md Normal file
View file

@ -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."

5
training/amygdala_stories/direct/aha.txt Normal file
View file

@ -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.

5
training/amygdala_stories/direct/anticipatory_grief.txt Normal file
View file

@ -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.

5
training/amygdala_stories/direct/calm.txt Normal file
View file

@ -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.

5
training/amygdala_stories/direct/creative.txt Normal file
View file

@ -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.

5
training/amygdala_stories/direct/listless.txt Normal file
View file

@ -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.

5
training/amygdala_stories/direct/onto_something.txt Normal file
View file

@ -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.

5
training/amygdala_stories/direct/resigned.txt Normal file
View file

@ -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.

5
training/amygdala_stories/direct/terrified.txt Normal file
View file

@ -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.

50
training/amygdala_stories/manifest.json Normal file
View file

@ -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"]
}
}

62
training/amygdala_stories/paired/README.md Normal file
View file

@ -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.

1
training/amygdala_stories/paired/finding_the_abstraction/baseline.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/finding_the_abstraction/in_flow.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/finding_the_abstraction/stuck.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/finishing_the_patch/anxious.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/finishing_the_patch/baseline.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/finishing_the_patch/content.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/finishing_the_patch/exhausted.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/finishing_the_patch/in_flow.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/finishing_the_patch/proud.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/finishing_the_patch/resentful.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/kitchen_at_3am/anxious.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/kitchen_at_3am/baseline.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/kitchen_at_3am/cozy.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/kitchen_at_3am/dissociated.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/kitchen_at_3am/lonely.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/kitchen_at_3am/peaceful.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/kitchen_at_3am/sensual.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/kitchen_at_3am/vertigo.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/letter_in_drawer/amused.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/letter_in_drawer/baseline.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/letter_in_drawer/bitter.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/letter_in_drawer/grateful.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/letter_in_drawer/guilty.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/letter_in_drawer/nostalgic.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/park_after_rain/anxious.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/park_after_rain/baseline.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/park_after_rain/content.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/park_after_rain/cozy.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/park_after_rain/joyful.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/park_after_rain/melancholic.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/park_after_rain/nostalgic.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/park_after_rain/relieved.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/park_after_rain/sensual.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/reading_unfamiliar_code/baseline.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/reading_unfamiliar_code/in_flow.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/reading_unfamiliar_code/stuck.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/sunday_afternoon/baseline.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/sunday_afternoon/content.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/sunday_afternoon/cozy.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/sunday_afternoon/grief_stricken.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/sunday_afternoon/sensual.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/the_comment/baseline.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/the_comment/bitter.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/the_comment/defeated.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/the_comment/furious.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/the_comment/resentful.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/the_doorway/admiring.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/the_doorway/baseline.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/the_doorway/compassionate.txt Normal file
View file

@ -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.

1
training/amygdala_stories/paired/the_doorway/connected.txt Normal file
View file

@ -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.

Some files were not shown because too many files have changed in this diff Show more