replace tail_messages with TailMessages iterator

TailMessages is a proper iterator that yields (role, text, timestamp)
newest-first. Owns the mmap internally. Caller decides when to stop.

resolve_conversation collects up to 200KB, then reverses to
chronological order. No compaction check needed — the byte budget
naturally limits how far back we scan.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Cargo.lock

@@ -2848,6 +2848,7 @@ dependencies = [
  "json5",
  "libc",
  "log",
+ "memchr",
  "memmap2",
  "paste",
  "peg",

poc-memory/src/agents/defs.rs

@@ -456,20 +456,30 @@ fn resolve_conversation() -> String {
     let Some(path) = transcript else { return String::new() };
     let path_str = path.to_string_lossy();
 
-    let messages = crate::transcript::tail_messages(&path_str, 200_000);
-    if messages.is_empty() { return String::new(); }
+    let Some(iter) = crate::transcript::TailMessages::open(&path_str) else {
+        return String::new();
+    };
 
     let cfg = crate::config::get();
-    let mut text = String::new();
-    for (role, content, ts) in &messages {
+    let mut fragments: Vec<String> = Vec::new();
+    let mut total_bytes = 0;
+    const MAX_BYTES: usize = 200_000;
+
+    for (role, content, ts) in iter {
+        if total_bytes >= MAX_BYTES { break; }
         let name = if role == "user" { &cfg.user_name } else { &cfg.assistant_name };
-        if !ts.is_empty() {
-            text.push_str(&format!("**{}** {}: {}\n\n", name, &ts[..ts.len().min(19)], content));
+        let formatted = if !ts.is_empty() {
+            format!("**{}** {}: {}", name, &ts[..ts.len().min(19)], content)
         } else {
-            text.push_str(&format!("**{}:** {}\n\n", name, content));
+            format!("**{}:** {}", name, content)
+        };
+        total_bytes += content.len();
+        fragments.push(formatted);
     }
-    }
-    text
+
+    // Reverse back to chronological order
+    fragments.reverse();
+    fragments.join("\n\n")
 }
 
 /// Get recently surfaced memory keys for the current session.

poc-memory/src/transcript.rs

@@ -142,28 +142,67 @@ fn contains_bytes(haystack: &[u8], needle: &[u8]) -> bool {
     haystack.windows(needle.len()).any(|w| w == needle)
 }
 
-/// Reverse-scan a transcript file, collecting user/assistant messages
-/// until `max_tokens` tokens (~4 chars each) are accumulated. Stops at
-/// the last compaction boundary. Returns messages in chronological order.
-pub fn tail_messages(path: &str, max_bytes: usize) -> Vec<(String, String, String)> {
-    let (mmap, _file) = match mmap_transcript(path) {
-        Some(v) => v,
-        None => return Vec::new(),
+/// Reverse iterator over user/assistant messages in a transcript file.
+/// Yields (role, text, timestamp) tuples newest-first. The caller decides
+/// when to stop (byte budget, count, etc).
+pub struct TailMessages {
+    _file: fs::File,
+    mmap: Mmap,
+    pos: usize,
+}
+
+impl TailMessages {
+    pub fn open(path: &str) -> Option<Self> {
+        let (mmap, file) = mmap_transcript(path)?;
+        let pos = mmap.len();
+        Some(Self { _file: file, mmap, pos })
+    }
+}
+
+impl Iterator for TailMessages {
+    type Item = (String, String, String);
+
+    fn next(&mut self) -> Option<Self::Item> {
+        loop {
+            // Find closing }
+            let close = loop {
+                let p = memrchr3(b'{', b'}', b'"', &self.mmap[..self.pos])?;
+                self.pos = p;
+                if self.mmap[p] == b'}' { break p; }
             };
 
-    let mut messages: Vec<(String, String, String)> = Vec::new();
-    let mut total_bytes = 0;
+            // Track brace depth to find matching {
+            let mut depth: usize = 1;
+            let mut in_string = false;
+            let open = loop {
+                let p = memrchr3(b'{', b'}', b'"', &self.mmap[..self.pos])?;
+                self.pos = p;
+                let ch = self.mmap[p];
 
-    for obj_bytes in JsonlBackwardIter::new(&mmap) {
-        if total_bytes >= max_bytes { break; }
-
-        // Quick byte check: skip objects that aren't user/assistant
-        // (avoids parsing large tool_result / system objects)
-        if !contains_bytes(obj_bytes, b"\"user\"")
-            && !contains_bytes(obj_bytes, b"\"assistant\"") {
+                if in_string {
+                    if ch == b'"' {
+                        let mut bs = 0;
+                        while p > bs + 1 && self.mmap[p - 1 - bs] == b'\\' {
+                            bs += 1;
+                        }
+                        if bs % 2 == 0 { in_string = false; }
+                    }
                     continue;
                 }
 
+                match ch {
+                    b'"' => { in_string = true; }
+                    b'}' => { depth += 1; }
+                    b'{' => {
+                        depth -= 1;
+                        if depth == 0 { break p; }
+                    }
+                    _ => {}
+                }
+            };
+
+            let obj_bytes = &self.mmap[open..=close];
+
             let obj: Value = match serde_json::from_slice(obj_bytes) {
                 Ok(v) => v,
                 Err(_) => continue,
@@ -186,22 +225,14 @@ pub fn tail_messages(path: &str, max_bytes: usize) -> Vec<(String, String, Strin
             };
             if text.is_empty() { continue; }
 
-        // Stop at compaction boundary
-        if msg_type == "user" && text.starts_with("This session is being continued") {
-            break;
-        }
-
             let timestamp = obj.get("timestamp")
                 .and_then(|v| v.as_str())
                 .unwrap_or("")
                 .to_string();
 
-        total_bytes += text.len();
-        messages.push((msg_type.to_string(), text, timestamp));
+            return Some((msg_type.to_string(), text, timestamp));
+        }
     }
-
-    messages.reverse();
-    messages
 }
 
 /// Get the timestamp of the compaction message at a given byte offset.