osborn 0.9.55 → 0.9.60

package/dist/index.js

@@ -1,7 +1,7 @@
 // Load environment variables FIRST before any other imports
 import 'dotenv/config';
 import { voice, initializeLogger } from '@livekit/agents';
-import { Room, RoomEvent, RemoteParticipant, } from '@livekit/rtc-node';
+import { Room, RoomEvent, } from '@livekit/rtc-node';
 import { AccessToken } from 'livekit-server-sdk';
 // Initialize logger before anything else
 initializeLogger({ pretty: true, level: 'info' });
@@ -1275,16 +1275,6 @@ async function main() {
     // Session-level always-allow list: paths the user has approved for this session without prompting
     let sessionAlwaysAllowPaths = new Set();
     let userState = 'listening'; // Track user speech state for queue safety
-    // Self-echo guard for the TTS interrupt below. Updated by the
-    // ActiveSpeakersChanged listener registered near the other room.on(...) handlers.
-    // user_state_changed carries NO speaker identity (verified against the SDK type
-    // — UserStateChangedEvent has only oldState/newState/createdAt), so a separate
-    // remote-speaker timestamp is the only way to distinguish "real user spoke" from
-    // "agent's own TTS echoed through the mic". Independent producer: rtc-node
-    // emits activeSpeakersChanged from server WebRTC audio-level reports
-    // (room.js:213), with NO reference to AgentSession or STT — so there's no
-    // dependency loop with user_state_changed's STT-driven producer.
-    let lastRemoteSpeakerAt = 0;
     let currentVoiceMode = voiceMode; // Track active voice mode for data handlers
     let currentProvider = realtimeConfig.provider; // Track active realtime provider
     // Authenticated Supabase userId from participant metadata. Used to scope
@@ -2126,35 +2116,32 @@ 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,
             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
                 },
-                // Echo-driven false-interrupt protection at the SDK level (1.2.x has these knobs,
-                // we just never set them — defaults are minDuration:500ms / minWords:0 which let
-                // through every short echo blip). Both knobs gate the SDK's internal
-                // interruptByAudioActivity() (agent_activity.js — runs on Deepgram interim
-                // transcripts AND speechDuration updates), which is the path that was firing
-                // even after our user_state_changed handler skipped the trigger.
+                // 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: {
-                    // SDK auto-interrupt fully DISABLED (0.9.55). Even with minDuration:750
-                    // and minWords:2 in 0.9.54, the SDK's onInterimTranscript path bypasses
-                    // duration gating (it fires on first interim text) and minWords gates
-                    // against accumulated transcript wordcount — so once a real user utters
-                    // ≥2 words, every subsequent echo passes. Worse: double-fires within
-                    // 200ms corrupt SegmentSynchronizerImpl state (pushAudio called after
-                    // close → markPlaybackFinished before input done → playback hangs).
-                    // With enabled:false the SDK won't fire interruptByAudioActivity at all;
-                    // our user_state_changed handler at index.ts:3162 with the self-echo
-                    // guard (lastRemoteSpeakerAt + ActiveSpeakersChanged) becomes the SOLE
-                    // interrupt path. We control timing, deduplication, and identity.
-                    enabled: false,
-                    // The values below have no effect with enabled:false but kept for
-                    // documentation in case enabled is flipped back on for testing.
-                    minDuration: 750,
-                    minWords: 2,
+                    // 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
                 },
             },
         });
@@ -2693,20 +2680,6 @@ async function main() {
         // rather than hold it indefinitely. Cancelled in ParticipantConnected.
         armAloneTimer();
     });
-    // Self-echo guard producer. Server WebRTC audio-level reports drive this
-    // (rtc-node room.js:213, ~50-100ms latency from mic onset — faster than
-    // Deepgram STT classification, so by the time user_state_changed fires
-    // lastRemoteSpeakerAt is already current). Filter speakers to RemoteParticipant
-    // — LocalParticipant is the agent itself and including it would defeat the
-    // whole point (the echo we're guarding against IS the agent's local audio).
-    // This is the speaker-identity filter the removed ActiveSpeakersChanged
-    // handler had (May 21 / c345c98) — minus the interrupt() call, since the
-    // user_state_changed handler now owns interrupt firing.
-    room.on(RoomEvent.ActiveSpeakersChanged, (speakers) => {
-        if (speakers.some((s) => s instanceof RemoteParticipant)) {
-            lastRemoteSpeakerAt = Date.now();
-        }
-    });
     // NOTE: previously this section also had a RoomEvent.ActiveSpeakersChanged
     // handler that interrupted TTS on any sustained audio activity (~50ms after
     // mic onset). That fired too eagerly — coughs, paper rustles, the agent's
@@ -3072,34 +3045,14 @@ async function main() {
                 userState = ev.newState;
                 console.log(`👤 User state: ${prev} → ${ev.newState} (agent: ${agentState})`);
                 if (ev.newState === 'speaking' && agentState === 'speaking' && sessionVoiceMode !== 'realtime') {
-                    const now = Date.now();
-                    // Self-echo guard. Reject this trigger entirely if no remote
-                    // participant has been heard speaking in the last 500ms — at that
-                    // point user_state=speaking is almost certainly TTS bleeding through
-                    // the mic (Deepgram correctly identifies it as "speech", we add the
-                    // identity filter the high-level event lacks). 500ms is wider than
-                    // the ~50-300ms gap between ActiveSpeakersChanged and user_state_changed
-                    // firing, so a real user is comfortably inside the window.
-                    //
-                    // The 1s leading-edge debounce that used to live here was removed in
-                    // 0.9.54 — the SDK-side `turnHandling.interruption.minDuration:750` +
-                    // `minWords:2` now do the heavy lifting on echo filtering, and stacking
-                    // an extra cooldown on top risked masking the SDK's own resume timing.
-                    if (now - lastRemoteSpeakerAt > 500) {
-                        console.log('🔇 Skipping interrupt — no recent remote-speaker activity (self-echo guard)');
-                        return;
-                    }
+                    // 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 {
-                        // force:true bypasses the SpeechHandle's allowInterruptions check
-                        // (speech_handle.js:93-99). Required because turnHandling.interruption.enabled=false
-                        // sets allowInterruptions=false on every SpeechHandle (agent_activity.js:329-331),
-                        // which is what blocks the SDK's auto-interrupt path — but without
-                        // force:true, this manual call from our handler would also throw
-                        // "This generation handle does not allow interruptions". Combined,
-                        // they let US interrupt (with self-echo guard already verified above)
-                        // while keeping the SDK's auto-trigger off.
-                        console.log('🎤 user_state_changed=speaking + agent speaking + remote-speaker confirmed → interrupting TTS (force)');
-                        currentSession?.interrupt({ force: true });
+                        console.log('🎤 user_state_changed=speaking + agent speaking → interrupting TTS');
+                        currentSession?.interrupt();
                     }
                     catch (err) {
                         console.warn('⚠️ user-state interrupt failed:', err instanceof Error ? err.message : err);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.55",
+  "version": "0.9.60",
   "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",