osborn 0.9.137 → 0.9.139

package/dist/pipeline-direct-llm.js

@@ -25,7 +25,7 @@ function buildSessionTail(sessionId, workingDir) {
         return '';
     if (!sessionId)
         return '';
-    const maxLines = parseInt(process.env.OSBORN_SESSION_TAIL_COUNT || '500', 10);
+    const maxLines = parseInt(process.env.OSBORN_SESSION_TAIL_COUNT || '1500', 10);
     try {
         const indexPath = getIndexPath(sessionId, workingDir);
         if (!indexPath)
@@ -63,14 +63,30 @@ export class PipelineDirectLLM extends llm.LLM {
     #turnAbort = null;
     #indexWatcher = null;
     #indexBuilding = false;
+    // Set to true when PostCompact fires — cleared after injecting the tail once
+    // on the next real user turn (seam-only injection, never per-turn).
+    #postCompactionTailPending = false;
     // True while the in-flight turn is a [MEETING —] chunk — the tts_say gate in
     // index.ts reads this to skip browser session.say() (meeting audio goes via
     // /canvas, not the laptop speakers → no same-room feedback). Set per chat().
     suppressMeetingTTS = false;
     constructor(opts) {
         super();
-        this.#claudeLLM = new ClaudeLLM(opts);
-        this.#opts = opts;
+        // Chain the onCompactionEvent callback so we can set #postCompactionTailPending
+        // when compaction completes, without breaking the existing UI-notification path.
+        const upstreamCompactionEvent = opts.onCompactionEvent;
+        const chainedOpts = {
+            ...opts,
+            onCompactionEvent: (event) => {
+                if (event.type === 'compaction_complete') {
+                    this.#postCompactionTailPending = true;
+                    console.log('[SEAM] compaction complete → will inject session_tail on next turn');
+                }
+                upstreamCompactionEvent?.(event);
+            },
+        };
+        this.#claudeLLM = new ClaudeLLM(chainedOpts);
+        this.#opts = chainedOpts;
     }
     /** Stop the index watcher (call on disconnect/session switch) */
     stopIndexWatcher() {
@@ -133,16 +149,25 @@ export class PipelineDirectLLM extends llm.LLM {
             }
         }
         console.log(`📥 [pipeline] chat() call #${callN} (${userText.length} chars): "${userText}"`);
-        // ── SESSION TAIL injection ────────────────────────────────────────────────
-        // Prepend a compact recent-conversation block so the orchestrator never
-        // loses thread across compaction. Runs synchronously (readFileSync) but the
-        // index file is ≤1MB and kept hot in the OS page cache — negligible latency.
-        // Skipped for meeting chunks (they are not real user turns).
-        const sessionTailBlock = (!userText.startsWith('[MEETING'))
-            ? buildSessionTail(this.#claudeLLM.sessionId, this.#opts.workingDirectory || process.cwd())
-            : '';
-        if (sessionTailBlock) {
-            console.log(`🧵 [pipeline] Session tail: ${sessionTailBlock.split('\n').length - 2} lines ready`);
+        // ── SESSION TAIL injection (seam-only) ───────────────────────────────────
+        // Inject the tail ONCE on the first real user turn after a compaction
+        // completes. Never injected on every turn (avoids catastrophic accumulation
+        // in the SDK persistent subprocess history). Meeting chunks are not real
+        // user turns and never get the tail.
+        let sessionTailBlock = '';
+        if (this.#postCompactionTailPending && !userText.startsWith('[MEETING') && userText.trim()) {
+            sessionTailBlock = buildSessionTail(this.#claudeLLM.sessionId, this.#opts.workingDirectory || process.cwd());
+            if (sessionTailBlock) {
+                const tailLines = sessionTailBlock.split('\n').length - 2;
+                const tailChars = sessionTailBlock.length;
+                console.log(`[SEAM] injecting post-compaction session_tail: ${tailLines} lines, ~${tailChars} chars`);
+            }
+            else {
+                console.log('[SEAM] post-compaction turn but session_tail empty/disabled');
+            }
+            // Always clear the flag after the first real turn, whether or not the
+            // index file had content (avoids re-trying on an empty index).
+            this.#postCompactionTailPending = false;
         }
         // Always check the pending playback context — it can carry two independent
         // signals: (a) an actual interruption (spokenText + recentMessages) when the
@@ -227,7 +252,7 @@ export class PipelineDirectLLM extends llm.LLM {
                 // when text is non-empty.
                 enrichedMessage = userText;
             }
-            // Modify the last user message in chatCtx — prepend session tail if available
+            // Modify the last user message in chatCtx
             const finalMessage = sessionTailBlock
                 ? sessionTailBlock + '\n\n' + enrichedMessage
                 : enrichedMessage;
@@ -240,7 +265,7 @@ export class PipelineDirectLLM extends llm.LLM {
             }
         }
         else if (sessionTailBlock && userText.trim()) {
-            // No interrupt context — inject session tail directly into the user message
+            // No interrupt context — inject post-compaction session tail into the user message
             for (let i = chatCtx.items.length - 1; i >= 0; i--) {
                 const item = chatCtx.items[i];
                 if (item.type === 'message' && item.role === 'user') {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.137",
+  "version": "0.9.139",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {