osborn 0.9.58 → 0.9.61

package/dist/index.js

@@ -2116,28 +2116,48 @@ async function main() {
         const session = new voice.AgentSession({
             turnDetection: 'stt',
             preemptiveGeneration: false, // Only fire LLM on final committed transcript, not partial preemptives
+            // First-line echo defense: drop mic frames from BOTH the recognition stream
+            // and the realtime audio stream for this many ms after the agent first
+            // enters 'speaking' state. STT receives no audio during the warmup → no
+            // interim/final transcripts can fire → echo cannot trigger an interrupt.
+            // 1.4.x default is 3000; bumping to 5000 widens the safe zone at session start.
+            // One-shot per session (NOT re-armed each turn), so this protects only the
+            // first agent response. After that the in-block interruption settings handle it.
+            aecWarmupDuration: 5000,
+            // TTS stall mitigations (0.9.61). The 1.4.x SDK added a 10s default
+            // readIdleTimeout in generation.js:519 (PR livekit/agents-js#1461) — when
+            // the TTS stream goes silent for >10s, it force-closes via reader.cancel()
+            // which trips the OpenAI SDK's AbortSignal → APIUserAbortError →
+            // tts_error recoverable:false. Root cause is upstream: the OpenAI plugin
+            // BUFFERS the entire tts-1 PCM response (arrayBuffer()) before emitting a
+            // single frame. Long sentences intermittently exceed 10s end-to-end with
+            // tts-1. Raising both watchdogs to 30s gives slow OpenAI responses room
+            // to complete; raising maxUnrecoverableErrors from default 3 to 15 prevents
+            // a transient burst of stalls from killing the AgentSession outright (the
+            // counter resets on every successful speaking transition).
+            ttsReadIdleTimeout: 30_000,
+            forwardAudioIdleTimeout: 30_000,
+            connOptions: {
+                maxUnrecoverableErrors: 15,
+            },
             turnHandling: {
                 endpointing: {
                     mode: 'fixed',
                     minDelay: 500, // Wait 500ms after STT commits before generating reply
                     maxDelay: 2000, // Force end-of-turn after 2s to prevent hangs
                 },
-                // 0.9.57: bump falseInterruptionTimeout from default 2000ms → 3000ms.
-                // This is the silence-after-interrupt window the SDK waits before
-                // emitting agentFalseInterruption + resuming. Extending it gives the
-                // user a fuller breath between low-level audio activity moments to
-                // accumulate a clean silence, which helps when echo or ambient noise
-                // keeps resetting the 2s window. Other tunables in this same block
-                // (NOT changed yet — try the timeout first, escalate if needed):
-                //   - minDuration (default 500ms) — minimum sustained speech to count
-                //   - minWords (default 0) — minimum word count in interim transcript
-                //   - enabled (default true) — kept ON (auto-interrupt path active)
-                //   - resumeFalseInterruption (default true) — auto-resume kept ON
-                //   - discardAudioIfUninterruptible (default true)
+                // 1.4.x SDK fully wires these — minDuration now applies to the STT path
+                // (not just VAD), falseInterruptionTimeout actually fires the
+                // agentFalseInterruption event with auto-resume, discardAudioIfUninterruptible
+                // is checked at runtime. All inert in 1.2.1; live in 1.4.x.
                 interruption: {
-                    falseInterruptionTimeout: 3000, // 2000 → 3000 (extra second of silence before resume)
-                    minDuration: 1000, // 500 → 1000 (need 1s sustained speech to count)
-                    minWords: 3, // 0 → 3 (interim transcript needs ≥3 words)
+                    // enabled defaults true — kept default (don't set to false; cascades into
+                    // allowInterruptions:false which breaks manual interrupt() calls).
+                    minDuration: 1000, // 1.4.x: now gates STT-path; require 1s sustained speech
+                    minWords: 3, // require ≥3 words in interim transcript
+                    falseInterruptionTimeout: 2000, // emit agentFalseInterruption after 2s silence
+                    resumeFalseInterruption: true, // auto-resume TTS on false interrupt detection
+                    discardAudioIfUninterruptible: true, // drop buffered echo audio
                 },
             },
         });
@@ -3041,13 +3061,11 @@ async function main() {
                 userState = ev.newState;
                 console.log(`👤 User state: ${prev} → ${ev.newState} (agent: ${agentState})`);
                 if (ev.newState === 'speaking' && agentState === 'speaking' && sessionVoiceMode !== 'realtime') {
-                    // Reverted to the simple post-May-22 (c345c98 / 0.9.39) shape in 0.9.56.
-                    // The self-echo guard via lastRemoteSpeakerAt was defeated by the same
-                    // physics it was trying to filter — TTS bleeds into the user's mic →
-                    // LiveKit registers their participant as a remote speaker → the guard
-                    // passes → we interrupt anyway. Verified in osbornojure logs 2026-06-16
-                    // (2 of 3 interrupts that session were from this handler firing on echo).
-                    // Echo prevention moved to browser AEC on the publisher side.
+                    // 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.
                     try {
                         console.log('🎤 user_state_changed=speaking + agent speaking → interrupting TTS');
                         currentSession?.interrupt();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.58",
+  "version": "0.9.61",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {
@@ -33,14 +33,14 @@
     "@anthropic-ai/claude-agent-sdk": "^0.2.91",
     "@anthropic-ai/sdk": "^0.80.0",
     "@google/genai": "^1.0.0",
-    "@livekit/agents": "1.2.1",
-    "@livekit/agents-plugin-deepgram": "1.2.1",
-    "@livekit/agents-plugin-elevenlabs": "1.2.1",
-    "@livekit/agents-plugin-google": "1.2.1",
-    "@livekit/agents-plugin-livekit": "1.2.1",
-    "@livekit/agents-plugin-openai": "1.2.1",
-    "@livekit/agents-plugin-silero": "1.2.1",
-    "@livekit/rtc-node": "0.13.24",
+    "@livekit/agents": "1.4.6",
+    "@livekit/agents-plugin-deepgram": "1.4.6",
+    "@livekit/agents-plugin-elevenlabs": "1.4.6",
+    "@livekit/agents-plugin-google": "1.4.6",
+    "@livekit/agents-plugin-livekit": "1.4.6",
+    "@livekit/agents-plugin-openai": "1.4.6",
+    "@livekit/agents-plugin-silero": "1.4.6",
+    "@livekit/rtc-node": "0.13.29",
     "@modelcontextprotocol/sdk": "^1.29.0",
     "@openai/codex-sdk": "^0.77.0",
     "@smithery/api": "^0.48.0",