osborn 0.9.64 → 0.9.67

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

@@ -2151,8 +2151,8 @@ async function main() {
             // Bumping both watchdogs to 30s gives the forwarder room to ride out
             // normal between-message pauses without timing out. Independent of the
             // interruption block above, which is doing its job (0 interrupts fired).
-            ttsReadIdleTimeout: 30_000, // default 10000 → 30000
-            forwardAudioIdleTimeout: 30_000, // default 10000 → 30000
+            ttsReadIdleTimeout: 40_000, // default 10000 → 30000
+            forwardAudioIdleTimeout: 40_000, // default 10000 → 30000
             // connOptions: {
             //   maxUnrecoverableErrors: 15,               // default 3 (left at default)
             // },
@@ -2167,9 +2167,9 @@ async function main() {
                 // a full 3s window to keep talking before deciding it was false and
                 // resuming. Other two knobs left at SDK defaults.
                 interruption: {
-                    minDuration: 2000, // default 500  — require 2s sustained speech
-                    minWords: 3, // default 0    — require ≥3 transcript words
-                    falseInterruptionTimeout: 3000, // default 2000 — wait 3s before auto-resume
+                    minDuration: 2500, // default 500  — require 2s sustained speech
+                    minWords: 4, // default 0    — require ≥3 transcript words
+                    falseInterruptionTimeout: 4000, // default 2000 — wait 3s before auto-resume
                     // resumeFalseInterruption: true,      // default true  (unchanged)
                     // discardAudioIfUninterruptible: true,// default true  (unchanged)
                 },
@@ -3075,17 +3075,47 @@ 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

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": "0.9.64",
+  "version": "0.9.67",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {