osborn 0.9.110 → 0.9.112

package/dist/index.js

@@ -1656,6 +1656,11 @@ async function main() {
     // REDIRECTED into the meeting canvas instead of suppressed — the regular
     // voice pipeline reused with the meeting as the sink (latency fix).
     let meetingAddressedUntil = 0;
+    // Distinct human speakers seen in the current meeting. In a 1:1 (one human
+    // + the bot) EVERY utterance is addressed to the bot — no name needed
+    // (2026-08-01: "can you open carfax..." got observer-suppressed because
+    // the fresh session's latch was never opened by the name).
+    const meetingSpeakers = new Set();
     let activeMeetingPoller = null; // Transcript poller bound to that bot
     // LIVE meeting transcript → LLM (buffered webhook finals). See recall.on('transcript').
     const meetingTranscriptBuffer = [];
@@ -1747,6 +1752,7 @@ async function main() {
         }
         activeMeetingBotId = null;
         meetingAddressedUntil = 0;
+        meetingSpeakers.clear();
         // leaveBot=false when Recall itself reported the meeting over (bot already gone).
         if (opts.leaveBot !== false) {
             const recall = getRecallClient();
@@ -1857,13 +1863,16 @@ async function main() {
                 // the agent replies out loud into the meeting. This is the chat-mode path:
                 // named/asked → prompt response. Un-addressed chunks stay in the silent
                 // 20s batch for note-taking. That's the two-mode seam, mechanically.
+                meetingSpeakers.add(speaker);
                 // Fuzzy name match: meeting STT routinely mangles "Osborn" — observed
                 // live 2026-08-01: "Osborn can you hear me" → "i was born can you hear
                 // me" (name trigger missed, bot stayed silent). Accept common
                 // mis-transcriptions; mild false-positive risk is acceptable in a
-                // room that invited the bot.
-                if (/\b(osborne?|oz\s?born|os\s?born|was born|is born|ozborn|osbourne?|austin\b.{0,8}(hear|there|can you))/i.test(text)) {
-                    console.log('📓 Addressed by name (fuzzy) — immediate flush for a response');
+                // room that invited the bot. AND: in a 1:1 meeting (one human + the
+                // bot) every utterance is addressed — no name needed.
+                const oneOnOne = meetingSpeakers.size <= 1;
+                if (oneOnOne || /\b(osborne?|oz\s?born|os\s?born|was born|is born|ozborn|osbourne?|austin\b.{0,8}(hear|there|can you))/i.test(text)) {
+                    console.log(`📓 Addressed (${oneOnOne ? '1:1 meeting' : 'by name'}) — immediate flush for a response`);
                     flushMeetingBuffer(botId, true);
                 }
             }
@@ -2558,15 +2567,13 @@ async function main() {
                 // (reuse of the regular speak path; no Bash roundtrip). Silent-observer
                 // turns stay suppressed as before.
                 if (activeMeetingBotId && Date.now() < meetingAddressedUntil) {
-                    // Split into AT MOST 2 chunks: first sentence (fast first-audio) +
-                    // the remainder as ONE chunk. Per-sentence files created an audible
-                    // seam at every boundary (mp3 padding + scheduling jitter — heard as
-                    // "cracking"); two chunks keeps the latency win with one seam max.
-                    const m = data.text.match(/^([^.!?]+[.!?]+["']?)\s*([\s\S]*)$/);
-                    const chunks = m && m[2]?.trim() ? [m[1].trim(), m[2].trim()] : [data.text.trim()];
-                    console.log(`🔊➡️📽️ tts_say → canvas (${chunks.length} chunk(s), addressed turn) t=${new Date().toISOString()}: "${data.text.slice(0, 60)}"`);
-                    for (const s of chunks)
-                        pushCanvas({ kind: 'say', text: s });
+                    // ONE say per reply (2026-08-01 quality directive): chunking created
+                    // seams no scheduler fully hid. A single utterance = a single clean
+                    // audio file = zero seams — quality over ~1s of latency. (Structural
+                    // fix queued: canvas as LiveKit subscriber playing the agent's real
+                    // TTS track — the true "same browser experience" reuse.)
+                    console.log(`🔊➡️📽️ tts_say → canvas (single utterance) t=${new Date().toISOString()}: "${data.text.slice(0, 60)}"`);
+                    pushCanvas({ kind: 'say', text: data.text });
                     markMeetingSpeaking(data.text);
                     return;
                 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.110",
+  "version": "0.9.112",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {