osborn 0.9.104 → 0.9.106

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/claude-llm.js

@@ -990,6 +990,17 @@ class ClaudeLLMStream extends llm.LLMStream {
                     loadAllSkills(this.#opts.sessionBaseDir || this.#opts.workingDirectory || process.cwd()),
                 ].filter(Boolean).join('\n\n'),
                 canUseTool: async (toolName, input, _options) => {
+                    // Auto-approve Bash calls to the agent's OWN local API (/canvas,
+                    // /tts) — how the agent SPEAKS into a meeting. A spoken reply must
+                    // never wait on a permission dialog. Localhost-only, own endpoints
+                    // only, no shell chaining — not a general curl allowance.
+                    if (toolName === 'Bash') {
+                        const cmd = String(input?.command || '');
+                        if (/^\s*curl\b[^;&|]*https?:\/\/(localhost|127\.0\.0\.1):\d+\/(canvas|tts)\b/.test(cmd) && !/[;&|]/.test(cmd)) {
+                            console.log(`✅ Auto-approved Bash → own API (canvas/tts speak path)`);
+                            return { behavior: 'allow', updatedInput: input };
+                        }
+                    }
                     // Auto-approve writes to session workspace (but block spec.md and library/ — fast brain manages those)
                     if (toolName === 'Write' || toolName === 'Edit') {
                         const filePath = String(input?.file_path || '');

package/dist/index.js

@@ -1617,6 +1617,10 @@ 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 +1632,17 @@ 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 +2507,15 @@ 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) {
+                    console.log(`🔊➡️📽️ tts_say → canvas (addressed meeting turn): "${data.text.slice(0, 60)}"`);
+                    pushCanvas({ kind: 'say', text: data.text });
+                    markMeetingSpeaking(data.text);
+                    return;
+                }
                 console.log(`🔇 tts_say suppressed (meeting turn — response goes to /canvas, not browser): "${data.text.slice(0, 60)}"`);
                 return;
             }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.104",
+  "version": "0.9.106",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {