osborn 0.9.52 → 0.9.54

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, } from '@livekit/rtc-node';
+import { Room, RoomEvent, RemoteParticipant, } from '@livekit/rtc-node';
 import { AccessToken } from 'livekit-server-sdk';
 // Initialize logger before anything else
 initializeLogger({ pretty: true, level: 'info' });
@@ -1205,6 +1205,16 @@ 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
@@ -2052,6 +2062,19 @@ async function main() {
                     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.
+                interruption: {
+                    minDuration: 750, // 500 → 750: require 750ms of sustained audio activity
+                    minWords: 2, // 0 → 2: require ≥2 transcript words (filters single-word echo blips)
+                    // SDK defaults kept: enabled=true, resumeFalseInterruption=true,
+                    // falseInterruptionTimeout=2000ms (that 2s timer is what resumed your audio
+                    // exactly where it stopped — confirmed working as designed).
+                },
             },
         });
         return { session, agent };
@@ -2589,6 +2612,20 @@ 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
@@ -2954,8 +2991,25 @@ 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;
+                    }
                     try {
-                        console.log('🎤 user_state_changed=speaking + agent speaking → interrupting TTS');
+                        console.log('🎤 user_state_changed=speaking + agent speaking + remote-speaker confirmed → interrupting TTS');
                         currentSession?.interrupt();
                     }
                     catch (err) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.52",
+  "version": "0.9.54",
   "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": "1.2.1",
+    "@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",
     "@modelcontextprotocol/sdk": "^1.29.0",
     "@openai/codex-sdk": "^0.77.0",
     "@smithery/api": "^0.48.0",