osborn 0.9.105 → 0.9.107

package/.claude/skills/meetings/SKILL.md

@@ -91,16 +91,19 @@ When a chunk says **`YOU WERE ADDRESSED`**, people in the meeting are WAITING
 for your voice. Measured failure (2026-08-01): notes-writing inside the reply
 turn pushed speech→reply past 30 seconds ("very very delayed").
 
-**Mandatory turn shape when addressed:**
-1. **FIRST tool call = the `/canvas say` POST.** One short spoken reply (1–2
-   sentences, conversational). Nothing runs before it — no Read, no Edit, no
-   transcript pull, no sub-agent.
+**Mandatory turn shape when addressed (0.9.106+):**
+1. **Just ANSWER in plain text — your words are spoken into the meeting
+   automatically.** The regular voice pipeline is redirected to the meeting
+   for addressed turns; do NOT use Bash/curl to speak (that adds a slow tool
+   roundtrip and can eat your tool budget). 1–2 short conversational
+   sentences, text first, before any tool call.
 2. **THEN** delegate note-taking / research to the writer/researcher
    sub-agents in the background as usual.
-3. If you genuinely need a fact before answering, say a holding line FIRST
-   ("Good question — one second while I check"), then look it up, then follow
-   up with a second `/canvas say`. Never leave the room in silence while you
-   work.
+3. If you genuinely need a fact before answering, SAY a holding line first
+   ("Good question — one second while I check"), look it up, then continue in
+   plain text. Never leave the room in silence while you work.
+(The `/canvas say` curl remains for NON-addressed cases: when the voice-native
+user says "tell the meeting X", or proactive announcements.)
 
 ## Browse requests while CASTING a live stream — drive the ENGINE, not the canvas
 

package/dist/index.js

@@ -570,11 +570,42 @@ function startApiServer(workingDir, port) {
         // speechSynthesis is NOT captured by Recall, a media element IS.
         if (req.method === 'GET' && url.pathname === '/tts') {
             const text = (url.searchParams.get('text') || '').slice(0, 4000);
+            if (!text) {
+                res.writeHead(400, { 'Content-Type': 'application/json' });
+                res.end(JSON.stringify({ error: 'no text' }));
+                return;
+            }
+            const t0 = Date.now();
+            // Deepgram FIRST (2026-08-01 latency fix): same TTS family the regular
+            // voice pipeline uses — ~3-6x faster than the OpenAI full-file synth this
+            // endpoint used before (measured 2-4s of the meeting reply lag). OpenAI
+            // stays as fallback.
+            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', {
+                        method: 'POST',
+                        headers: { 'Authorization': `Token ${dgKey}`, 'Content-Type': 'application/json' },
+                        body: JSON.stringify({ text }),
+                    });
+                    if (dg.ok) {
+                        const buf = Buffer.from(await dg.arrayBuffer());
+                        console.log(`🗣️ /tts deepgram ${buf.length}b in ${Date.now() - t0}ms (${text.length} chars) t=${new Date().toISOString()}`);
+                        res.writeHead(200, { 'Content-Type': 'audio/mpeg', 'Cache-Control': 'no-store', 'Content-Length': buf.length });
+                        res.end(buf);
+                        return;
+                    }
+                    console.warn(`⚠️ /tts deepgram ${dg.status} — falling back to OpenAI`);
+                }
+                catch (e) {
+                    console.warn(`⚠️ /tts deepgram error: ${e.message} — falling back to OpenAI`);
+                }
+            }
             const voice = url.searchParams.get('voice') || 'alloy';
             const key = process.env.OPENAI_API_KEY;
-            if (!text || !key) {
+            if (!key) {
                 res.writeHead(400, { 'Content-Type': 'application/json' });
-                res.end(JSON.stringify({ error: !key ? 'no OPENAI_API_KEY' : 'no text' }));
+                res.end(JSON.stringify({ error: 'no TTS provider keys' }));
                 return;
             }
             try {
@@ -590,6 +621,7 @@ function startApiServer(workingDir, port) {
                     return;
                 }
                 const buf = Buffer.from(await tts.arrayBuffer());
+                console.log(`🗣️ /tts openai ${buf.length}b in ${Date.now() - t0}ms t=${new Date().toISOString()}`);
                 res.writeHead(200, { 'Content-Type': 'audio/mpeg', 'Cache-Control': 'no-store', 'Content-Length': buf.length });
                 res.end(buf);
             }
@@ -1617,6 +1649,10 @@ async function main() {
     // gone forever, matching the skill-catalog pattern.
     let userRemovedAgents = [];
     let activeMeetingBotId = null; // Recall.ai bot ID if in a meeting
+    // While set (Date.now() < value), the agent's streaming tts_say output is
+    // REDIRECTED into the meeting canvas instead of suppressed — the regular
+    // voice pipeline reused with the meeting as the sink (latency fix).
+    let meetingAddressedUntil = 0;
     let activeMeetingPoller = null; // Transcript poller bound to that bot
     // LIVE meeting transcript → LLM (buffered webhook finals). See recall.on('transcript').
     const meetingTranscriptBuffer = [];
@@ -1628,8 +1664,17 @@ async function main() {
         if (!meetingTranscriptBuffer.length || !currentLLM)
             return;
         const turns = meetingTranscriptBuffer.splice(0); // drain
+        if (addressed) {
+            // REUSE the regular voice pipeline (2026-08-01 latency fix): during an
+            // addressed turn the agent's normal streaming tts_say output is REDIRECTED
+            // to the canvas (see the tts_say handler) — same speak path as regular
+            // mode, sink = meeting. The old prompt made the agent compose a Bash curl
+            // (extra LLM tool roundtrip, ate the 3-call budget, sometimes lost the
+            // reply entirely) — measured as most of the 5-8s reply lag.
+            meetingAddressedUntil = Date.now() + 90_000;
+        }
         const header = addressed
-            ? `[MEETING — ${botId}] — YOU WERE ADDRESSED. Respond OUT LOUD into the meeting now: POST http://localhost:${apiPort}/canvas {"kind":"say","text":"..."} with a short, direct reply. Then note it.`
+            ? `[MEETING — ${botId}] — YOU WERE ADDRESSED. Reply now in PLAIN TEXT — your words are spoken into the meeting automatically (do NOT use Bash/curl to speak). One or two short conversational sentences first; then, if needed, delegate notes/research in the background.`
             : `[MEETING — ${botId}]:`;
         // Prepend + consume any interruption context (bot was cut off mid-sentence).
         const interrupt = meetingInterruptContext ? `${meetingInterruptContext}\n` : '';
@@ -2494,6 +2539,21 @@ 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) {
+                // ADDRESSED turn → REDIRECT the normal streaming reply into the meeting
+                // (reuse of the regular speak path; no Bash roundtrip). Silent-observer
+                // turns stay suppressed as before.
+                if (activeMeetingBotId && Date.now() < meetingAddressedUntil) {
+                    // Sentence-split so the FIRST sentence synthesizes + plays while the
+                    // rest queue behind it (canvas plays says sequentially + prefetches)
+                    // — first-audio latency = one short sentence's synth, not the whole
+                    // reply's.
+                    const sentences = data.text.match(/[^.!?]+[.!?]+["']?|[^.!?]+$/g)?.map((s) => s.trim()).filter(Boolean) || [data.text];
+                    console.log(`🔊➡️📽️ tts_say → canvas (${sentences.length} sentence(s), addressed turn) t=${new Date().toISOString()}: "${data.text.slice(0, 60)}"`);
+                    for (const s of sentences)
+                        pushCanvas({ kind: 'say', text: s });
+                    markMeetingSpeaking(data.text);
+                    return;
+                }
                 console.log(`🔇 tts_say suppressed (meeting turn — response goes to /canvas, not browser): "${data.text.slice(0, 60)}"`);
                 return;
             }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.105",
+  "version": "0.9.107",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {