osborn 0.9.114 → 0.9.116

package/dist/index.js

@@ -242,7 +242,7 @@ async function synthMp3(text) {
     const dgKey = process.env.DEEPGRAM_API_KEY;
     if (dgKey) {
         try {
-            const dg = await fetch('https://api.deepgram.com/v1/speak?model=aura-2-thalia-en&encoding=mp3&bit_rate=48000', {
+            const dg = await fetch('https://api.deepgram.com/v1/speak?model=aura-2-asteria-en&encoding=mp3&bit_rate=48000', {
                 method: 'POST',
                 headers: { 'Authorization': `Token ${dgKey}`, 'Content-Type': 'application/json' },
                 body: JSON.stringify({ text: text.slice(0, 4000) }),
@@ -659,7 +659,7 @@ function startApiServer(workingDir, port) {
                     // WAV/linear16, not mp3: mp3 files carry encoder padding (leading/
                     // trailing silence + boundary click) — back-to-back sentence clips
                     // were heard as "cracking". WAV is gapless-safe and cheaper to decode.
-                    const dg = await fetch('https://api.deepgram.com/v1/speak?model=aura-2-thalia-en&encoding=linear16&sample_rate=48000&container=wav', {
+                    const dg = await fetch('https://api.deepgram.com/v1/speak?model=aura-2-asteria-en&encoding=linear16&sample_rate=48000&container=wav', {
                         method: 'POST',
                         headers: { 'Authorization': `Token ${dgKey}`, 'Content-Type': 'application/json' },
                         body: JSON.stringify({ text }),
@@ -1984,10 +1984,15 @@ async function main() {
                 }
             }
             else {
-                // speech_off — a natural silence boundary. Flush the buffered turns now
-                // (chunk on conversational pauses, not an arbitrary clock), unless empty.
-                if (!isBot && meetingTranscriptBuffer.length)
-                    flushMeetingBuffer(botId, false);
+                // speech_off — a natural silence boundary = the speaker finished their
+                // utterance. When the conversation latch is open (1:1 or interactive
+                // mode), that boundary IS the bot's turn — flush ADDRESSED so it
+                // replies at natural pauses, VAD-driven, no name required
+                // (user directive 2026-08-01).
+                if (!isBot && meetingTranscriptBuffer.length) {
+                    const latchOpen = Date.now() < meetingAddressedUntil || meetingSpeakers.size <= 1;
+                    flushMeetingBuffer(botId, latchOpen);
+                }
             }
         });
     }
@@ -2653,11 +2658,11 @@ async function main() {
             // re-captures in the same room → feedback). Set by PipelineDirectLLM.chat()
             // when the turn is a [MEETING —] chunk. Normal user turns are unaffected.
             if (directLLM.suppressMeetingTTS) {
-                // BROWSER-PARITY PATH: canvas is in the LiveKit room → the normal
-                // session.say audio reaches the meeting through it. Fall through to
-                // the regular pipeline for addressed turns (identical audio to the
-                // browser experience); observer turns still stay silent.
-                if (meetingCanvasInRoom && activeMeetingBotId && Date.now() < meetingAddressedUntil) {
+                // MEETING MODE CONDITION (user directive 2026-08-01): says route to
+                // the NATIVE Recall sink by default — no room/Chrome relay (that path
+                // was moderately-to-extremely choppy). The LiveKit parity relay stays
+                // available behind OSBORN_MEETING_AUDIO=parity for future testing.
+                if (process.env.OSBORN_MEETING_AUDIO === 'parity' && meetingCanvasInRoom && activeMeetingBotId && Date.now() < meetingAddressedUntil) {
                     console.log(`🔊🎼 meeting reply via NATIVE session.say (canvas relays): "${data.text.slice(0, 50)}"`);
                     markMeetingSpeaking(data.text);
                     // no return — normal say proceeds below
@@ -5205,17 +5210,16 @@ async function main() {
                             // the bot casts the meeting canvas as its camera+mic (below), so
                             // it can show visuals and speak into the meeting on demand.
                             await sendToFrontend({ type: 'meeting_joining', message: 'Osborn is joining your meeting...' });
-                            // Cast target = the bot's camera+mic webpage (Recall output_media).
-                            // DEFAULT is the meeting canvas (frontend /meeting-canvas), pointed
-                            // at THIS agent's public URL so it can subscribe to /canvas-stream
-                            // and become the bot's face (visuals) + voice (TTS it speaks). An
-                            // explicit castUrl from the frontend or OSBORN_MEETING_CAST_URL
-                            // overrides it (e.g. to cast a live feed or a seeded site instead).
-                            const frontendUrl = (process.env.OSBORN_FRONTEND_URL || 'https://www.voice-native.com').replace(/\/$/, '');
-                            const canvasUrl = /^https:\/\//.test(webhookBase)
-                                ? `${frontendUrl}/meeting-canvas?agent=${encodeURIComponent(webhookBase)}`
-                                : undefined;
-                            const castUrl = data.castUrl || process.env.OSBORN_MEETING_CAST_URL || canvasUrl;
+                            // AUDIO-FIRST DEFAULT (2026-08-01, proven live): the webpage-camera
+                            // output_media routes the bot's AUDIO through the canvas page's
+                            // Web Audio — which sits SUSPENDED in Recall's headless Chrome (no
+                            // user gesture) → the bot was inaudible all night. Default now = NO
+                            // webpage camera, so the bot's voice is the direct Recall
+                            // output_audio track (speakIntoMeeting → recall.outputAudio, ~1.8s
+                            // synth, verified audible + clear). The browser CAST is opt-in via
+                            // explicit data.castUrl / OSBORN_MEETING_CAST_URL until we split
+                            // video(webpage) from audio(output_audio) to reclaim it cleanly.
+                            const castUrl = data.castUrl || process.env.OSBORN_MEETING_CAST_URL || undefined;
                             const botId = await recallJoin.joinMeeting(meetingUrl, webhookBase, { castUrl });
                             const sessionId = currentLLM?.sessionId || currentResumeSessionId || 'default';
                             recallJoin.registerBot(botId, sessionId);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.114",
+  "version": "0.9.116",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {