osborn 0.9.65 → 0.9.68

package/dist/fast-brain.js

@@ -71,7 +71,7 @@ let geminiClient = null;
 let initialized = false;
 // Model IDs — configurable per provider
 const ANTHROPIC_FAST_MODEL = 'claude-haiku-4-5-20251001';
-const GEMINI_FAST_MODEL = 'gemini-2.0-flash';
+const GEMINI_FAST_MODEL = 'gemini-2.5-flash'; // 0.9.67: was gemini-2.0-flash — 404 deprecated by Google (verified against @livekit/agents-plugin-google test default)
 // Agent SDK session tracking — resume across voice questions for context continuity
 let fastBrainSessionId = null;
 // Gemini Chat session — persists across voice questions for context continuity.

package/dist/index.js

@@ -2703,7 +2703,16 @@ async function main() {
     // Room Event Handlers
     // ============================================================
     room.on(RoomEvent.Connected, () => {
-        console.log('✅ Connected to room:', roomName);
+        // 0.9.68: log Room SID + name PROMINENTLY so we can cross-reference
+        // this specific session in LiveKit Cloud dashboard → Sessions tab.
+        // @livekit/rtc-node Room exposes SID via async getSid() (it's resolved
+        // after WebRTC handshake), so we fetch it asynchronously and log when ready.
+        console.log(`✅ Connected to room: ${roomName} | t=${new Date().toISOString()}`);
+        room.getSid().then((sid) => {
+            console.log(`🔗 [LIVEKIT-DASHBOARD] room sid=${sid} name=${roomName} — search at https://cloud.livekit.io/projects → Sessions → "${sid}"`);
+        }).catch((err) => {
+            console.log(`⚠️ [LIVEKIT-DASHBOARD] failed to fetch room SID: ${err instanceof Error ? err.message : String(err)}`);
+        });
         localParticipant = room.localParticipant;
         // Arm the alone timer: if we connected but no user joins within the grace
         // window (e.g. machine woken then abandoned mid-handshake), leave the room
@@ -3075,17 +3084,47 @@ async function main() {
                 userState = ev.newState;
                 console.log(`👤 User state: ${prev} → ${ev.newState} (agent: ${agentState})`);
                 if (ev.newState === 'speaking' && agentState === 'speaking' && sessionVoiceMode !== 'realtime') {
-                    // Simple manual interrupt for echo-side defense fallback. With 1.4.x
-                    // the SDK's interrupt-by-audio-activity path is properly gated by
-                    // turnHandling.interruption.{minDuration, minWords, falseInterruptionTimeout},
-                    // and resumeFalseInterruption auto-recovers if echo was misclassified.
-                    // This handler stays as a secondary trigger only.
+                    // 0.9.67: action commented out, condition + debug kept.
+                    //
+                    // Why removed: in @livekit/agents 1.4.x SpeechHandle.interrupt() calls
+                    //   replyAbortController.abort() → AbortSignal.any composes into the
+                    //   OpenAI TTS HTTP fetch → arrayBuffer() throws AbortError →
+                    //   APIUserAbortError (openai/client.mjs:364) → SDK marks the error
+                    //   recoverable:false → connOptions.maxUnrecoverableErrors counter trips
+                    //   → session collapses. In 1.2.1 the same call was a hard-kill that
+                    //   never reached an HTTP fetch — that's why it ran fine for ~1 month
+                    //   under the silently caret-resolved 1.4.5 (which had inherited the
+                    //   abort plumbing) until it crossed the unrecoverable-errors threshold.
+                    //
+                    // What handles interruption now: SDK 1.4.x's gated path —
+                    //   turnHandling.interruption.{minDuration:2500, minWords:4,
+                    //   falseInterruptionTimeout:4000, resumeFalseInterruption:true}
+                    //   pauses TTS via audioOutput.pause() (no abort) and either resumes
+                    //   on a false trigger or hard-interrupts on a confirmed barge-in.
+                    //
+                    // Debug: this block now ONLY observes — logs what we'd have interrupted
+                    //   on so we can compare against the SDK's own behavior. If the SDK
+                    //   under-reacts to real barge-ins we can re-enable selectively.
                     try {
-                        console.log('🎤 user_state_changed=speaking + agent speaking → interrupting TTS');
-                        currentSession?.interrupt();
+                        const evKeys = ev && typeof ev === 'object' ? Object.keys(ev) : [];
+                        const evShape = evKeys.reduce((acc, k) => {
+                            const v = ev[k];
+                            acc[k] = (v && typeof v === 'object') ? `<object:${Object.keys(v).join(',')}>` : v;
+                            return acc;
+                        }, {});
+                        console.log('🔎 [DEBUG] manual-interrupt WOULD HAVE FIRED — SDK gated path now owns it:', JSON.stringify({
+                            t: new Date().toISOString(),
+                            userPrev: prev,
+                            userNew: ev.newState,
+                            agentState,
+                            sessionVoiceMode,
+                            evKeys,
+                            evShape,
+                        }));
+                        // currentSession?.interrupt()  // ← 0.9.67 DISABLED: cascades to APIUserAbortError → recoverable:false → session collapse
                     }
                     catch (err) {
-                        console.warn('⚠️ user-state interrupt failed:', err instanceof Error ? err.message : err);
+                        console.warn('⚠️ user-state interrupt debug failed:', err instanceof Error ? err.message : err);
                     }
                 }
                 // When user stops speaking, retry voice queue — items may be waiting
@@ -3153,9 +3192,26 @@ async function main() {
                     sendAgentTranscript(message, 'playout');
                 }
             });
+            // 0.9.68: mirror SDK's internal unrecoverable-error counters so we can
+            // see EXACTLY how close we are to closeImpl() firing (default threshold 3).
+            // Counter resets on each successful "speaking" transition (agent_session.js:740).
+            let __ttsErrorCounter = 0;
+            let __llmErrorCounter = 0;
+            const __maxUnrecov = 3; // SDK default DEFAULT_SESSION_CONNECT_OPTIONS.maxUnrecoverableErrors
             // Error handler
             sess.on('error', (ev) => {
                 const msg = ev.error?.message || String(ev.error);
+                const errType = ev.type || 'unknown';
+                const recoverable = ev.recoverable;
+                // 0.9.68: counter mirror — increment for recoverable:false same as SDK does
+                if (recoverable === false) {
+                    if (errType === 'tts_error')
+                        __ttsErrorCounter++;
+                    else if (errType === 'llm_error')
+                        __llmErrorCounter++;
+                }
+                const willCloseNext = (__ttsErrorCounter > __maxUnrecov || __llmErrorCounter > __maxUnrecov);
+                console.log(`📊 [ERROR-COUNTER] type=${errType} recoverable=${recoverable} ttsErrorCount=${__ttsErrorCounter}/${__maxUnrecov} llmErrorCount=${__llmErrorCounter}/${__maxUnrecov} willCloseNext=${willCloseNext} t=${new Date().toISOString()}`);
                 // OpenAI race: voice queue collided with server-side VAD auto-response
                 if (msg.includes('conversation_already_has_active_response') || msg.includes('active_response')) {
                     console.log('⚠️ OpenAI active response collision — queue will retry on next listening state');
@@ -3168,6 +3224,15 @@ async function main() {
                 }
                 console.error('❌ Session error:', ev.error);
             });
+            // 0.9.68: reset error counter mirror when SDK does (on speaking transition).
+            // Reuses the existing agent_state_changed handler logic — fires AFTER.
+            sess.on('agent_state_changed', (ev) => {
+                if (ev.newState === 'speaking' && (__ttsErrorCounter > 0 || __llmErrorCounter > 0)) {
+                    console.log(`📊 [COUNTER-RESET] speaking transition cleared ttsErrorCount=${__ttsErrorCounter}→0 llmErrorCount=${__llmErrorCounter}→0`);
+                    __ttsErrorCounter = 0;
+                    __llmErrorCounter = 0;
+                }
+            });
             // Capture voice mode at session creation — prevents state confusion
             // if currentVoiceMode changes between session start and crash recovery
             const sessionVoiceMode = currentVoiceMode;

package/dist/pipeline-fastbrain.js

@@ -17,7 +17,7 @@ import { GoogleGenAI } from '@google/genai';
 // ============================================================
 // CONSTANTS
 // ============================================================
-const GEMINI_MODEL = 'gemini-2.0-flash';
+const GEMINI_MODEL = 'gemini-2.5-flash'; // 0.9.67: was gemini-2.0-flash — 404 deprecated by Google
 const TIMEOUT_MS = 20_000; // AFC needs time for tool calls + processing + synthesis
 const MAX_AFC_CALLS = 4;
 // ============================================================

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.65",
+  "version": "0.9.68",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {