osborn 0.9.106 → 0.9.108

package/dist/index.js

@@ -570,11 +570,45 @@ 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 {
+                    // 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=24000&container=wav', {
+                        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 wav ${buf.length}b in ${Date.now() - t0}ms (${text.length} chars) t=${new Date().toISOString()}`);
+                        res.writeHead(200, { 'Content-Type': 'audio/wav', '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 +624,7 @@ 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);
             }
@@ -1641,8 +1676,12 @@ async function main() {
             // reply entirely) — measured as most of the 5-8s reply lag.
             meetingAddressedUntil = Date.now() + 90_000;
         }
+        // PROMPT PARITY (user directive 2026-08-01): addressed turns carry ONLY a
+        // minimal tag — no behavioral re-instruction. The agent replies exactly as
+        // it would to a regular voice turn; the tts_say→canvas redirect handles
+        // where the words go. Same agent, same behavior, both fronts.
         const header = addressed
-            ? `[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}] (addressed — reply is spoken into the meeting):`
             : `[MEETING — ${botId}]:`;
         // Prepend + consume any interruption context (bot was cut off mid-sentence).
         const interrupt = meetingInterruptContext ? `${meetingInterruptContext}\n` : '';
@@ -2511,8 +2550,15 @@ async function main() {
                 // (reuse of the regular speak path; no Bash roundtrip). Silent-observer
                 // turns stay suppressed as before.
                 if (activeMeetingBotId && Date.now() < meetingAddressedUntil) {
-                    console.log(`🔊➡️📽️ tts_say → canvas (addressed meeting turn): "${data.text.slice(0, 60)}"`);
-                    pushCanvas({ kind: 'say', text: data.text });
+                    // Split into AT MOST 2 chunks: first sentence (fast first-audio) +
+                    // the remainder as ONE chunk. Per-sentence files created an audible
+                    // seam at every boundary (mp3 padding + scheduling jitter — heard as
+                    // "cracking"); two chunks keeps the latency win with one seam max.
+                    const m = data.text.match(/^([^.!?]+[.!?]+["']?)\s*([\s\S]*)$/);
+                    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 });
                     markMeetingSpeaking(data.text);
                     return;
                 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.106",
+  "version": "0.9.108",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {