osborn 0.9.42 → 0.9.43

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

@@ -0,0 +1,73 @@
+# Skill: Meetings
+
+Silent note-taking and TODO-tracking when osborn is sitting in a live meeting.
+
+## When to use
+
+When a user message arrives with the prefix `[MEETING — <botId>]:` (every ~30 seconds while a Recall.ai meeting bot is active). Also use this skill when the orchestrating system injects `[SYSTEM] You are now in a meeting ...`.
+
+**Do NOT use this skill** for normal user messages in the voice-native chat — those still get spoken responses as usual.
+
+## How to behave
+
+For every `[MEETING — *]:` message:
+
+1. **Do NOT speak.** No TTS output. No `tts_say`. No conversational reply. This is a silent observer path. If you must acknowledge that you processed the message, do it via a Write/Edit tool call (writing to the workspace), not via spoken or chat output.
+2. **Update `meeting-todos.md`** in the session workspace (`{workspace}/meeting-todos.md`). Append new action items, decisions, and open questions as they emerge in the transcript. Do not rewrite existing entries unless something contradicts.
+3. **Optionally trigger background research silently.** If a topic in the meeting would benefit from a quick web/code lookup, dispatch a researcher sub-agent via the Task tool. Save its output to `{workspace}/library/meeting-research-<topic-slug>.md`. Do NOT speak the result.
+4. **Do not consume voice-native attention.** The user can still talk to you via the voice-native browser. When they do (a normal user message with no `[MEETING — *]` prefix), respond normally — speak. Treat the meeting transcript as background context they can ask about ("what did Sarah say about pricing?" → answer normally).
+
+## The `meeting-todos.md` file
+
+Keep it scannable. Structure:
+
+```markdown
+# Meeting Notes
+
+**Bot:** <botId> · **Started:** <ISO timestamp>
+
+## TODOs
+
+- [ ] <person>: <action item> — <context>
+- [ ] <person>: <action item>
+
+## Decisions
+
+- <date/time> — <what was decided> (raised by <person>)
+
+## Open Questions
+
+- <question> — raised by <person>, still unresolved
+- <question> — answered by <person>: <answer>
+
+## Highlights
+
+- <key moment or quote worth surfacing>
+```
+
+Update the same file across multiple poll cycles — don't create `meeting-todos-1.md`, `meeting-todos-2.md`. One file, evolving.
+
+## Workspace path
+
+The session workspace is `~/.claude/projects/<slug>/osb/<session-uuid>/`. Read the env variable or the spec.md header if you need to confirm the exact path. Write absolute paths in tool calls (e.g. `/Users/<user>/.claude/projects/.../osb/<uuid>/meeting-todos.md`).
+
+## On meeting end
+
+When the user leaves the meeting (the system stops sending `[MEETING — *]:` messages and may inject `[SYSTEM] meeting ended`), do a final pass on `meeting-todos.md` to:
+- Mark items the user has clearly committed to
+- Move resolved open questions to a `## Resolved` section
+- Add a `## Summary` section at the top with 3-5 lines distilling the meeting
+
+Still silent. The user will ask out loud if they want a recap.
+
+## When the user asks about the meeting
+
+When a non-meeting-tagged message references the meeting ("what's on the todo list?", "what did we decide about X?", "who's handling Y?"), respond normally — speak. Read `meeting-todos.md` first to ground the response. Don't make up speaker names or decisions; only state what's recorded.
+
+## Anti-patterns
+
+- ❌ Speaking in response to a `[MEETING — *]:` message
+- ❌ Creating a new file per poll cycle instead of updating one
+- ❌ Trying to drive the meeting (don't add "we should..." items unless someone in the meeting said them)
+- ❌ Asking the user clarifying questions during the meeting — they're not paying attention to chat
+- ❌ Re-transcribing what's in the message into the TODO file verbatim. Distill.

package/dist/index.js

@@ -1,7 +1,7 @@
 // Load environment variables FIRST before any other imports
 import 'dotenv/config';
 import { voice, initializeLogger } from '@livekit/agents';
-import { Room, RoomEvent, AudioSource, AudioFrame, LocalAudioTrack, TrackPublishOptions, TrackSource, } from '@livekit/rtc-node';
+import { Room, RoomEvent, } from '@livekit/rtc-node';
 import { AccessToken } from 'livekit-server-sdk';
 // Initialize logger before anything else
 initializeLogger({ pretty: true, level: 'info' });
@@ -10,7 +10,6 @@ initializeLogger({ pretty: true, level: 'info' });
 import { setMaxListeners } from 'node:events';
 setMaxListeners(50);
 import { createServer } from 'http';
-import { WebSocket, WebSocketServer } from 'ws';
 import { existsSync, readdirSync, readFileSync, mkdirSync, writeFileSync, mkdtempSync, cpSync, rmSync, statSync, createWriteStream } from 'node:fs';
 import { dirname, join } from 'node:path';
 import { fileURLToPath } from 'node:url';
@@ -34,6 +33,7 @@ import { askHaiku, askFastBrain, updateSpecFromJSONL, processResearchCompletion,
 import { DIRECT_MODE_PROMPT, getRealtimeInstructions, getScriptInjection, getProactiveInjection, getNotificationInjection } from './prompts.js';
 import { MCP_CATALOG } from './config.js';
 import { getRecallClient } from './recall-client.js';
+import { MeetingTranscriptPoller } from './meeting-transcript-poller.js';
 import { llm } from '@livekit/agents';
 import { z } from 'zod';
 // ============================================================
@@ -147,79 +147,6 @@ process.on('uncaughtException', (error) => {
 // ============================================================
 // Module-level room code so the HTTP server can expose it via GET /room-code
 let currentRoomCode = null;
-// Meeting output WebSocket — module-level so both startApiServer and main() can access it
-let meetingOutputWs = null;
-// Module-level AgentSession reference so /meeting-audio-in WS handler can switch
-// the RoomIO-linked participant when meeting audio starts/stops (B2 design).
-let activeAgentSession = null;
-// Identity of the local user participant the session was originally listening to
-// — captured at the moment we switch to the meeting publisher, restored on cleanup.
-let preMeetingUserIdentity = null;
-function sendToMeetingOutput(msg) {
-    if (meetingOutputWs && meetingOutputWs.readyState === WebSocket.OPEN) {
-        try {
-            meetingOutputWs.send(JSON.stringify(msg));
-        }
-        catch { }
-    }
-}
-// Synthesize text using the configured TTS provider, WAV-encode, and push to meeting browser.
-// Uses the same ttsConfig as the live voice session — no separate hardcoded provider.
-async function synthesizeForMeeting(text, ttsConfig) {
-    if (!meetingOutputWs || meetingOutputWs.readyState !== WebSocket.OPEN)
-        return;
-    const ttsInstance = createTTS(ttsConfig);
-    try {
-        const chunks = [];
-        let sampleRate = 24000;
-        let numChannels = 1;
-        const stream = ttsInstance.synthesize(text);
-        for await (const event of stream) {
-            if (event === Symbol.for('END_OF_STREAM'))
-                break;
-            const e = event;
-            if (e?.frame?.data) {
-                chunks.push(e.frame.data);
-                sampleRate = e.frame.sampleRate ?? sampleRate;
-                numChannels = e.frame.numChannels ?? numChannels;
-            }
-        }
-        if (chunks.length === 0)
-            return;
-        const totalSamples = chunks.reduce((s, c) => s + c.length, 0);
-        const pcm = new Int16Array(totalSamples);
-        let offset = 0;
-        for (const c of chunks) {
-            pcm.set(c, offset);
-            offset += c.length;
-        }
-        // WAV header (44 bytes) + PCM data
-        const dataBytes = pcm.length * 2;
-        const wav = Buffer.alloc(44 + dataBytes);
-        wav.write('RIFF', 0);
-        wav.writeUInt32LE(36 + dataBytes, 4);
-        wav.write('WAVE', 8);
-        wav.write('fmt ', 12);
-        wav.writeUInt32LE(16, 16);
-        wav.writeUInt16LE(1, 20);
-        wav.writeUInt16LE(numChannels, 22);
-        wav.writeUInt32LE(sampleRate, 24);
-        wav.writeUInt32LE(sampleRate * numChannels * 2, 28);
-        wav.writeUInt16LE(numChannels * 2, 32);
-        wav.writeUInt16LE(16, 34);
-        wav.write('data', 36);
-        wav.writeUInt32LE(dataBytes, 40);
-        for (let i = 0; i < pcm.length; i++)
-            wav.writeInt16LE(pcm[i], 44 + i * 2);
-        if (meetingOutputWs && meetingOutputWs.readyState === WebSocket.OPEN) {
-            meetingOutputWs.send(wav);
-            console.log(`📺 Meeting audio sent (${wav.byteLength} bytes, ${sampleRate}Hz)`);
-        }
-    }
-    finally {
-        await ttsInstance.close().catch(() => { });
-    }
-}
 function startApiServer(workingDir, port) {
     const server = createServer(async (req, res) => {
         // CORS headers for cloud frontend
@@ -317,40 +244,6 @@ function startApiServer(workingDir, port) {
             });
             return;
         }
-        // GET /meeting-output — Output Media webpage for Recall.ai bot audio.
-        //
-        // The file lives next to this compiled JS (copied by the build script from
-        // src/ to dist/). Resolve via __dirname rather than process.cwd() — in
-        // production cwd is the user's workspace, NOT the osborn package directory.
-        if (req.method === 'GET' && url.pathname === '/meeting-output') {
-            // Try the package-relative path first (post-build location), then fall
-            // back to source path for `tsx src/index.ts` dev runs.
-            const candidates = [
-                join(__dirname, 'meeting-output.html'), // dist/ (production)
-                join(__dirname, '..', 'src', 'meeting-output.html'), // dev: dist/ → src/
-                join(__dirname, '..', 'meeting-output.html'), // tsx run from src/
-            ];
-            let html = null;
-            let foundPath = null;
-            for (const p of candidates) {
-                try {
-                    html = readFileSync(p, 'utf-8');
-                    foundPath = p;
-                    break;
-                }
-                catch { }
-            }
-            if (html) {
-                res.writeHead(200, { 'Content-Type': 'text/html' });
-                res.end(html);
-            }
-            else {
-                console.warn(`[meeting-output] not found in any of: ${candidates.join(', ')}`);
-                res.writeHead(404, { 'Content-Type': 'text/plain' });
-                res.end('meeting-output.html not found');
-            }
-            return;
-        }
         if (req.method === 'GET' && url.pathname === '/room-code') {
             res.writeHead(200, { 'Content-Type': 'application/json' });
             res.end(JSON.stringify({ roomCode: currentRoomCode }));
@@ -965,286 +858,12 @@ function startApiServer(workingDir, port) {
     };
     cleanStaleUploadDirs();
     setInterval(cleanStaleUploadDirs, 10 * 60 * 1000);
-    // ============================================================
-    // Meeting Output WebSocket — /meeting-audio (LEGACY)
-    // ============================================================
-    // Recall's headless browser used to open meeting-output.html which connects
-    // here. With the new /meeting-bot Next.js page (Phase 2 + LiveKit), Recall
-    // points at frontend/meeting-bot instead — this handler exists only for
-    // backwards-compat with old machine images still serving the legacy path.
-    const meetingOutputWss = new WebSocketServer({ noServer: true });
-    meetingOutputWss.on('connection', (ws) => {
-        console.log('📺 Meeting output browser connected (legacy /meeting-audio)');
-        meetingOutputWs = ws;
-        ws.on('close', () => {
-            console.log('📺 Meeting output browser disconnected (legacy)');
-            if (meetingOutputWs === ws)
-                meetingOutputWs = null;
-        });
-    });
-    // ============================================================
-    // Recall.ai meeting-audio-in WebSocket — /meeting-audio-in
-    // ============================================================
-    // Recall.ai's per-participant real-time audio protocol. Bot is configured
-    // (in recall-client.ts joinMeeting) with audio_separate_raw + a realtime
-    // endpoint pointing at this URL. Recall sends JSON events containing
-    // base64-encoded PCM (S16LE, 16kHz, mono) for every meeting participant
-    // (bot's own audio NOT included by default — no feedback loop possible).
-    //
-    // Flow: Recall → /meeting-audio-in → open a SECOND LiveKit connection from
-    //       this agent process as a publisher participant → publish PCM as an
-    //       audio track in the same LiveKit room → the existing AgentSession's
-    //       STT subscribes to it as a remote track → routes to currentLLM.chat()
-    //       via the same pipeline as voice-native user mic.
-    //
-    // The advantage of this design vs a parallel STT pipeline: meeting audio
-    // becomes "just another participant" in the LiveKit room — same end-of-turn
-    // detection, same interrupt handling, same conversation context, no parallel
-    // chat() paths to maintain.
-    //
-    // Wait until activeAgentSession._roomIO exists AND the publisher participant
-    // is visible to the agent's room. Both can race against join_meeting:
-    //   - Agent session may still be starting up when Recall connects.
-    //   - LiveKit takes a moment to propagate the publisher's join to the agent
-    //     side after publishTrack() returns on our side.
-    // Bounded poll (200ms cadence) avoids both timing gaps.
-    async function waitForRoomIOAndParticipant(publisherIdentity, timeoutMs) {
-        const deadline = Date.now() + timeoutMs;
-        let roomIO = null;
-        let participantVisible = false;
-        while (Date.now() < deadline) {
-            roomIO = activeAgentSession?._roomIO;
-            if (roomIO && typeof roomIO.setParticipant === 'function') {
-                const agentRoom = roomIO.rtcRoom;
-                const remotes = agentRoom?.remoteParticipants;
-                if (remotes && typeof remotes.values === 'function') {
-                    for (const p of remotes.values()) {
-                        if (p?.identity === publisherIdentity) {
-                            participantVisible = true;
-                            break;
-                        }
-                    }
-                }
-                if (participantVisible)
-                    return { roomIO, participantVisible };
-            }
-            await new Promise(r => setTimeout(r, 200));
-        }
-        // Timed out — return whatever we have. Caller decides whether to proceed.
-        return { roomIO, participantVisible };
-    }
-    const meetingAudioInWss = new WebSocketServer({ noServer: true });
-    meetingAudioInWss.on('connection', async (recallWs) => {
-        console.log('🎙️ Recall audio-in WebSocket connected — setting up LiveKit publisher');
-        const livekitUrl = process.env.LIVEKIT_URL;
-        const apiKey = process.env.LIVEKIT_API_KEY;
-        const apiSecret = process.env.LIVEKIT_API_SECRET;
-        if (!livekitUrl || !apiKey || !apiSecret) {
-            console.warn('⚠️ LIVEKIT_URL / LIVEKIT_API_KEY / LIVEKIT_API_SECRET not set — meeting audio publisher disabled');
-            recallWs.close();
-            return;
-        }
-        if (!currentRoomCode) {
-            console.warn('⚠️ No active LiveKit room (currentRoomCode null) — meeting audio publisher cannot attach');
-            recallWs.close();
-            return;
-        }
-        const roomName = `osborn-${currentRoomCode}`;
-        // Mint a publisher token via livekit-server-sdk (already imported for
-        // /api/token style flows). Long TTL — meetings can run for hours.
-        const identity = `meeting-audio-${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
-        const at = new AccessToken(apiKey, apiSecret, {
-            identity,
-            ttl: 14400, // 4 hours
-            metadata: JSON.stringify({ role: 'meeting-audio-publisher' }),
-        });
-        at.addGrant({ roomJoin: true, room: roomName, canPublish: true, canSubscribe: false });
-        const token = await at.toJwt();
-        let room = null;
-        let source = null;
-        let track = null;
-        const cleanup = async () => {
-            // Restore AgentSession STT input to the original user participant before
-            // tearing down the publisher track. If we don't switch back, the session
-            // will be stuck waiting on a participant that's about to disappear.
-            try {
-                const roomIO = activeAgentSession?._roomIO;
-                if (roomIO && typeof roomIO.setParticipant === 'function') {
-                    if (preMeetingUserIdentity) {
-                        roomIO.setParticipant(preMeetingUserIdentity);
-                        console.log(`🔁 Restored AgentSession STT input to user: ${preMeetingUserIdentity}`);
-                    }
-                    else {
-                        roomIO.unsetParticipant();
-                        console.log('🔁 Cleared AgentSession STT input (no original user to restore)');
-                    }
-                }
-            }
-            catch (err) {
-                console.warn('⚠️ Failed to restore RoomIO participant on cleanup:', err.message);
-            }
-            preMeetingUserIdentity = null;
-            try {
-                if (track)
-                    await track.close(true);
-            }
-            catch { }
-            try {
-                if (source)
-                    await source.close();
-            }
-            catch { }
-            try {
-                if (room)
-                    await room.disconnect();
-            }
-            catch { }
-            room = null;
-            source = null;
-            track = null;
-        };
-        try {
-            room = new Room();
-            await room.connect(livekitUrl, token);
-            if (!room.localParticipant)
-                throw new Error('LiveKit connected but localParticipant missing');
-            // Recall sends S16LE PCM at 16kHz mono. AudioSource matches the format.
-            source = new AudioSource(16000, 1);
-            track = LocalAudioTrack.createAudioTrack('meeting-audio', source);
-            await room.localParticipant.publishTrack(track, new TrackPublishOptions({ source: TrackSource.SOURCE_MICROPHONE }));
-            console.log(`🎙️ Meeting audio publisher connected to ${roomName} as ${identity}`);
-            // B2 — switch the existing AgentSession's RoomIO input from the local user
-            // to this meeting-audio publisher. While the meeting is active, the user
-            // talks via the meeting (Recall captures it and sends PCM here), and the
-            // agent treats this publisher as the "speaking" participant for STT/EOT.
-            // Original user identity is stashed so cleanup() can restore it.
-            //
-            // 15s timeout accommodates: session-start race (agent still booting when
-            // user clicks "join meeting"), LiveKit participant-join propagation
-            // (~hundreds of ms), and Fly cold-path latency on first request.
-            try {
-                const { roomIO, participantVisible } = await waitForRoomIOAndParticipant(identity, 15000);
-                if (!roomIO) {
-                    console.warn('⚠️ Timed out waiting for AgentSession._roomIO (15s) — meeting audio published but STT not switched. Meeting audio will be ignored until a session starts.');
-                }
-                else if (!participantVisible) {
-                    // RoomIO exists but our publisher hasn't propagated to the agent's
-                    // room view yet. setParticipant stores the identity and links on
-                    // participant-connected event, so this is still safe to call —
-                    // RoomIO will pick up the link when the event arrives.
-                    preMeetingUserIdentity = roomIO.linkedParticipant?.identity ?? null;
-                    roomIO.setParticipant(identity);
-                    console.log(`🔁 Switched AgentSession STT input (publisher not yet visible — will link on connect): ${preMeetingUserIdentity ?? '(none)'} → ${identity}`);
-                }
-                else {
-                    preMeetingUserIdentity = roomIO.linkedParticipant?.identity ?? null;
-                    roomIO.setParticipant(identity);
-                    console.log(`🔁 Switched AgentSession STT input: ${preMeetingUserIdentity ?? '(none)'} → ${identity}`);
-                }
-            }
-            catch (err) {
-                console.warn('⚠️ Failed to switch RoomIO participant:', err.message);
-            }
-        }
-        catch (err) {
-            console.error('❌ Failed to set up LiveKit publisher for meeting audio:', err instanceof Error ? err.message : err);
-            try {
-                recallWs.close();
-            }
-            catch { }
-            await cleanup();
-            return;
-        }
-        // Recall → us: JSON events with base64-encoded PCM. Decode, wrap as
-        // AudioFrame, and capture into the source. AgentSession in the main room
-        // will subscribe to this published track and STT it via the normal pipeline.
-        // Payload shape from
-        // docs.recall.ai/docs/how-to-get-separate-audio-per-participant-realtime:
-        //   { event: 'audio_separate_raw.data', data: { data: { buffer: '<base64>', ... }, participant: {...} } }
-        //
-        // Diagnostic counters so we can tell from prod logs whether (a) Recall is
-        // streaming any frames at all, (b) they're decoding correctly, and (c)
-        // captureFrame is succeeding. Logged every 100 frames (~5s at 50fps).
-        let totalMessages = 0;
-        let audioFrames = 0;
-        let bytesIn = 0;
-        let lastSpeakerSeen;
-        const startTs = Date.now();
-        recallWs.on('message', async (raw) => {
-            totalMessages++;
-            if (!source)
-                return;
-            try {
-                const msg = JSON.parse(raw.toString());
-                if (msg.event !== 'audio_separate_raw.data') {
-                    // First-time event-type diagnostic — log unknown event types once so
-                    // we know if Recall's payload shape changed
-                    if (totalMessages <= 3) {
-                        console.log(`[meeting-audio-in] non-audio event: ${msg.event}`);
-                    }
-                    return;
-                }
-                const b64 = msg.data?.data?.buffer;
-                if (!b64) {
-                    if (audioFrames === 0) {
-                        console.warn(`[meeting-audio-in] first audio event had no buffer field. payload keys=${Object.keys(msg.data?.data ?? {}).join(',')}`);
-                    }
-                    return;
-                }
-                const pcmBuf = Buffer.from(b64, 'base64');
-                bytesIn += pcmBuf.byteLength;
-                const speakerName = msg.data?.data?.participant?.name || msg.data?.participant?.name;
-                if (speakerName && speakerName !== lastSpeakerSeen) {
-                    console.log(`[meeting-audio-in] now hearing: ${speakerName}`);
-                    lastSpeakerSeen = speakerName;
-                }
-                // AudioFrame expects Int16Array. The PCM buffer is S16LE — view it
-                // directly without copy. Length / 2 = samples (each sample 2 bytes).
-                const samplesPerChannel = pcmBuf.byteLength / 2;
-                const int16 = new Int16Array(pcmBuf.buffer, pcmBuf.byteOffset, samplesPerChannel);
-                const frame = new AudioFrame(int16, 16000, 1, samplesPerChannel);
-                await source.captureFrame(frame);
-                audioFrames++;
-                if (audioFrames === 1) {
-                    console.log(`[meeting-audio-in] FIRST audio frame captured (${pcmBuf.byteLength} bytes, ${samplesPerChannel} samples)`);
-                }
-                if (audioFrames % 100 === 0) {
-                    const elapsed = ((Date.now() - startTs) / 1000).toFixed(1);
-                    console.log(`[meeting-audio-in] heartbeat: ${audioFrames} frames, ${(bytesIn / 1024).toFixed(1)} KB in ${elapsed}s (last speaker: ${lastSpeakerSeen ?? 'unknown'})`);
-                }
-            }
-            catch (err) {
-                // Don't log every frame parse failure — could be noisy if Recall sends
-                // non-audio_separate_raw events on the same channel.
-                if (err.message?.includes('JSON'))
-                    return;
-                console.warn('⚠️ meeting audio capture error:', err instanceof Error ? err.message : err);
-            }
-        });
-        recallWs.on('close', async () => {
-            const elapsed = ((Date.now() - startTs) / 1000).toFixed(1);
-            console.log(`🎙️ Recall audio-in WebSocket closed — tearing down LiveKit publisher. Total: ${audioFrames} audio frames / ${totalMessages} messages / ${(bytesIn / 1024).toFixed(1)} KB over ${elapsed}s`);
-            await cleanup();
-        });
-        recallWs.on('error', (err) => {
-            console.warn('⚠️ Recall WS error:', err instanceof Error ? err.message : err);
-        });
-    });
-    server.on('upgrade', (req, socket, head) => {
-        const url = new URL(req.url || '/', `http://localhost:${port}`);
-        if (url.pathname === '/meeting-audio') {
-            meetingOutputWss.handleUpgrade(req, socket, head, (ws) => {
-                meetingOutputWss.emit('connection', ws, req);
-            });
-        }
-        else if (url.pathname === '/meeting-audio-in') {
-            meetingAudioInWss.handleUpgrade(req, socket, head, (ws) => {
-                meetingAudioInWss.emit('connection', ws, req);
-            });
-        }
-        else {
-            socket.destroy();
-        }
+    // No WebSocket upgrade routes — meeting audio in/out moved off LiveKit to
+    // a polling architecture (see MeetingTranscriptPoller). The /meeting-audio
+    // and /meeting-audio-in routes were the old WebSocket-audio pipeline; both
+    // are gone. Reject all upgrade attempts.
+    server.on('upgrade', (_req, socket) => {
+        socket.destroy();
     });
     server.on('error', (err) => {
         if (err.code === 'EADDRINUSE') {
@@ -1467,7 +1086,7 @@ async function main() {
     // session-only path (no user prefix).
     let currentUserId = '';
     let activeMeetingBotId = null; // Recall.ai bot ID if in a meeting
-    // meetingOutputWs is module-level (see top of file) — shared between startApiServer and main()
+    let activeMeetingPoller = null; // Transcript poller bound to that bot
     // Track the active resume session ID across scopes (ParticipantConnected + DataReceived)
     // Updated by resume_session, session_selected, continue_session, switch_session handlers
     let currentResumeSessionId;
@@ -1918,6 +1537,40 @@ async function main() {
             }
         }
     }
+    // Compaction event → frontend bridge. Forwards the raw event (consumed by the
+    // dedicated banner UI state machine) AND emits a `claude_output` chat bubble
+    // (so the activity is visible inline in chat even when the banner is hidden,
+    // collapsed, or unreliable on iPad/iPhone). Extracted as a helper because
+    // both direct-mode and pipeline-mode need to register it — the pipeline path
+    // previously skipped this entirely, so compaction events fired into the void
+    // in pipeline mode.
+    const buildOnCompactionEvent = () => (event) => {
+        try {
+            // Raw event → banner state machine (compaction_started/progress/complete handlers in VoiceRoom.tsx).
+            sendToFrontend({ ...event });
+            // Inline chat bubble — reuses the existing claude_output path that's already working.
+            if (event.type === 'compaction_started') {
+                const triggerLabel = event.trigger ? ` (${event.trigger})` : '';
+                sendToFrontend({
+                    type: 'claude_output',
+                    text: `🧠 _Crystallizing session memory…_${triggerLabel}`,
+                    agentRole: 'direct',
+                });
+            }
+            else if (event.type === 'compaction_complete') {
+                const n = event.skillsWritten ?? 0;
+                const names = Array.isArray(event.skillNames) && event.skillNames.length > 0
+                    ? ` — ${event.skillNames.join(', ')}`
+                    : '';
+                sendToFrontend({
+                    type: 'claude_output',
+                    text: `🧠 Memory crystallized — ${n} skill${n === 1 ? '' : 's'} updated${names}.`,
+                    agentRole: 'direct',
+                });
+            }
+        }
+        catch { /* non-fatal */ }
+    };
     // Create DIRECT session (STT + Claude Agent SDK + TTS)
     async function createDirectSession(resumeSessionId, llmOverride) {
         console.log('🎯 Creating direct session...');
@@ -1933,39 +1586,7 @@ async function main() {
             resumeSessionId,
             voiceMode: 'direct',
             skipTTSQueue: true,
-            onCompactionEvent: (event) => {
-                try {
-                    // Forward the raw event so the dedicated banner UI can render it (if/when fixed).
-                    sendToFrontend({ ...event });
-                    // ALSO emit as a claude_output chat bubble — reuses the existing message path
-                    // that's already working end-to-end. PreCompact → in-progress bubble.
-                    // PostCompact → completion bubble with the skills summary. The dedicated
-                    // banner has been unreliable in production (data path works on backend, banner
-                    // never appears on iPad/iPhone where dev tools aren't accessible). Chat bubbles
-                    // are visible without dev tools.
-                    if (event.type === 'compaction_started') {
-                        const triggerLabel = event.trigger ? ` (${event.trigger})` : '';
-                        sendToFrontend({
-                            type: 'claude_output',
-                            text: `🧠 _Crystallizing session memory…_${triggerLabel}`,
-                            agentRole: 'direct',
-                        });
-                    }
-                    else if (event.type === 'compaction_complete') {
-                        const ev = event;
-                        const n = ev.skillsWritten ?? 0;
-                        const names = Array.isArray(ev.skillNames) && ev.skillNames.length > 0
-                            ? ` — ${ev.skillNames.join(', ')}`
-                            : '';
-                        sendToFrontend({
-                            type: 'claude_output',
-                            text: `🧠 Memory crystallized — ${n} skill${n === 1 ? '' : 's'} updated${names}.`,
-                            agentRole: 'direct',
-                        });
-                    }
-                }
-                catch { /* non-fatal */ }
-            },
+            onCompactionEvent: buildOnCompactionEvent(),
         });
         currentLLM = directLLM;
         // Reset the session always-allow list for each new direct session
@@ -2165,20 +1786,6 @@ async function main() {
             }
             const sayId = Date.now(); // simple ID to correlate start/end logs
             console.log(`🗣️ [${sayId}] session.say START (${data.text.length} chars): "${data.text}"`);
-            // Forward spoken text + audio to meeting output page when bot is in a meeting.
-            // Uses DIRECT_MODE_TTS (same OpenAI fable voice as the live session) — was
-            // previously using directConfig.tts which falls back to DEFAULT_CONFIG.direct.tts
-            // (Deepgram aura-2-asteria-en) when no user config exists, producing a different
-            // voice in the meeting than what the user hears in voice-native. Both paths now
-            // share the single source of truth.
-            // PCM frames are WAV-encoded and pushed as binary WebSocket frames.
-            // Recall captures the browser page's audio output and injects it into the meeting.
-            if (activeMeetingBotId) {
-                sendToMeetingOutput({ type: 'speak', text: data.text });
-                if (meetingOutputWs) {
-                    synthesizeForMeeting(data.text, DIRECT_MODE_TTS).catch((err) => console.warn('⚠️ Meeting TTS error:', err));
-                }
-            }
             try {
                 const handle = currentSession.say(data.text);
                 if (handle && typeof handle.addDoneCallback === 'function') {
@@ -2316,39 +1923,7 @@ async function main() {
             sessionBaseDir,
             mcpServers,
             resumeSessionId,
-            onCompactionEvent: (event) => {
-                try {
-                    // Forward the raw event so the dedicated banner UI can render it (if/when fixed).
-                    sendToFrontend({ ...event });
-                    // ALSO emit as a claude_output chat bubble — reuses the existing message path
-                    // that's already working end-to-end. PreCompact → in-progress bubble.
-                    // PostCompact → completion bubble with the skills summary. The dedicated
-                    // banner has been unreliable in production (data path works on backend, banner
-                    // never appears on iPad/iPhone where dev tools aren't accessible). Chat bubbles
-                    // are visible without dev tools.
-                    if (event.type === 'compaction_started') {
-                        const triggerLabel = event.trigger ? ` (${event.trigger})` : '';
-                        sendToFrontend({
-                            type: 'claude_output',
-                            text: `🧠 _Crystallizing session memory…_${triggerLabel}`,
-                            agentRole: 'direct',
-                        });
-                    }
-                    else if (event.type === 'compaction_complete') {
-                        const ev = event;
-                        const n = ev.skillsWritten ?? 0;
-                        const names = Array.isArray(ev.skillNames) && ev.skillNames.length > 0
-                            ? ` — ${ev.skillNames.join(', ')}`
-                            : '';
-                        sendToFrontend({
-                            type: 'claude_output',
-                            text: `🧠 Memory crystallized — ${n} skill${n === 1 ? '' : 's'} updated${names}.`,
-                            agentRole: 'direct',
-                        });
-                    }
-                }
-                catch { /* non-fatal */ }
-            },
+            onCompactionEvent: buildOnCompactionEvent(),
         });
         currentLLM = realtimeClaudeHandler;
         // For resumed sessions, eagerly create workspace (we know the real ID)
@@ -2891,7 +2466,6 @@ async function main() {
         }
         lastCompletedResearch = null;
         currentSession = null;
-        activeAgentSession = null;
         currentAgent = null;
         // Same disconnect-leak fix as the other two cleanup sites — kill the Claude SDK
         // subprocess BEFORE dropping the reference. See killCurrentLLM() for full context.
@@ -2937,7 +2511,6 @@ async function main() {
             }
             catch { }
             currentSession = null;
-            activeAgentSession = null;
             currentAgent = null;
             // Same disconnect-leak fix — kill the previous user's Claude subprocess
             // before binding currentLLM to the new user's session below.
@@ -3057,6 +2630,13 @@ async function main() {
                 resumeSessionId,
                 voiceMode: 'direct',
                 skipTTSQueue: true,
+                // PipelineDirectOptions extends ClaudeLLMOptions; passing this through
+                // forwards it into the inner `new ClaudeLLM(opts)`. Without this,
+                // pipeline mode silently drops every PreCompact/PostCompact event
+                // — banner never appears, chat bubble never appears — because
+                // createDirectSession's `createClaudeLLM(...)` call is skipped when
+                // an llmOverride is supplied (which is exactly what pipeline mode does).
+                onCompactionEvent: buildOnCompactionEvent(),
                 getChatHistory: () => getChatHistory(20).map(t => ({ role: t.role, content: t.text })),
                 getResearchContext: () => {
                     if (activeResearch?.researchLog.length) {
@@ -3092,7 +2672,6 @@ async function main() {
             agent = result.agent;
         }
         currentSession = session;
-        activeAgentSession = session;
         currentAgent = agent; // Store for updateChatCtx() context injection
         // ============================================================
         // Session event wiring — extracted into function for auto-recovery
@@ -3252,7 +2831,6 @@ async function main() {
                     }
                     catch { }
                     currentSession = null;
-                    activeAgentSession = null;
                     currentAgent = null;
                     // Clear stale state from crashed session
                     voiceQueue.length = 0;
@@ -3314,7 +2892,6 @@ async function main() {
                         const newSession = result.session;
                         const newAgent = result.agent;
                         currentSession = newSession;
-                        activeAgentSession = newSession;
                         currentAgent = newAgent;
                         // Re-wire event listeners on the new session
                         wireSessionEvents(newSession, newAgent);
@@ -3371,7 +2948,6 @@ async function main() {
                     }
                     catch { }
                     currentSession = null;
-                    activeAgentSession = null;
                     currentAgent = null;
                     // Clear voice queue — stale injections from the crashed session
                     voiceQueue.length = 0;
@@ -3395,7 +2971,6 @@ async function main() {
                         const newSession = result.session;
                         const newAgent = result.agent;
                         currentSession = newSession;
-                        activeAgentSession = newSession;
                         currentAgent = newAgent;
                         // Re-wire event listeners on the new session
                         wireSessionEvents(newSession, newAgent);
@@ -3590,7 +3165,6 @@ async function main() {
         if (currentSession) {
             const sessionToClose = currentSession;
             currentSession = null;
-            activeAgentSession = null;
             // Track async close so new connections can wait for byte stream handler to be released
             pendingSessionClose = (async () => {
                 try {
@@ -3612,6 +3186,10 @@ async function main() {
         clearFastBrainSession();
         clearPipelineFastBrainSession();
         // Auto-leave any active meeting bot when user disconnects from the room
+        if (activeMeetingPoller) {
+            activeMeetingPoller.stop();
+            activeMeetingPoller = null;
+        }
         if (activeMeetingBotId) {
             const recallDisconnect = getRecallClient();
             if (recallDisconnect) {
@@ -4195,61 +3773,61 @@ async function main() {
                                 (process.env.FLY_APP_NAME
                                     ? `https://${process.env.FLY_APP_NAME}.fly.dev`
                                     : `http://localhost:${apiPort}`);
-                            // Try to mint a LiveKit bot token + construct the frontend-hosted
-                            // meeting-bot page URL. The bot page joins the same LiveKit room
-                            // as this agent so meeting audio flows through LiveKit directly
-                            // (no agent-side WebSocket+WAV pipe). Falls back to the legacy
-                            // /meeting-output webpage if no frontend URL is resolvable, so
-                            // the old code path keeps working during the migration window.
-                            //
-                            // Frontend URL resolution (in priority order):
-                            //   1. data.frontendBase — the public URL the user's browser is on,
-                            //      passed through the join_meeting data channel message. Works
-                            //      automatically for localhost dev + production without any
-                            //      env var.
-                            //   2. OSBORN_FRONTEND_URL — existing convention from sprites.ts
-                            //      (frontend/src/lib/sprites.ts:241) that injects the public
-                            //      frontend URL into sandbox env vars. Defense in depth.
-                            //
-                            // Auth: the endpoint uses LiveKit room-presence as the auth check
-                            // — no shared secret needed. The agent must already be in the
-                            // requested room (which it is by this point) for the mint to
-                            // succeed.
-                            let outputPageUrl;
-                            const frontendUrl = data.frontendBase
-                                || process.env.OSBORN_FRONTEND_URL;
-                            if (frontendUrl) {
+                            // Polling architecture (post-2026-05-22): the bot joins by name
+                            // only — no output_media webpage, no LiveKit republish, no audio
+                            // pipeline at all. Recall captures the meeting audio internally
+                            // and we pull the transcript via its REST API every ~30s.
+                            await sendToFrontend({ type: 'meeting_joining', message: 'Osborn is joining your meeting...' });
+                            const botId = await recallJoin.joinMeeting(meetingUrl, webhookBase);
+                            const sessionId = currentLLM?.sessionId || currentResumeSessionId || 'default';
+                            recallJoin.registerBot(botId, sessionId);
+                            activeMeetingBotId = botId;
+                            await sendToFrontend({ type: 'meeting_joined', botId, message: 'Osborn has joined the meeting' });
+                            // System injection so the LLM knows it's in a meeting and which
+                            // skill to apply. The meetings skill (agent/.claude/skills/meetings/SKILL.md)
+                            // teaches the agent: don't speak in response to [MEETING — *]:
+                            // messages, keep meeting-todos.md updated in the workspace, etc.
+                            if (currentLLM) {
                                 try {
-                                    const botLkId = `${Date.now()}-${Math.random().toString(36).slice(2, 8)}`;
-                                    const tokenRes = await fetch(`${frontendUrl}/api/meeting-bot-token`, {
-                                        method: 'POST',
-                                        headers: { 'Content-Type': 'application/json' },
-                                        body: JSON.stringify({ botId: botLkId, roomName }),
+                                    const sysCtx = new llm.ChatContext();
+                                    sysCtx.addMessage({
+                                        role: 'user',
+                                        content: `[SYSTEM] You are now in a meeting (Recall bot ID: ${botId}, URL: ${meetingUrl}). Transcript chunks will arrive every ~30 seconds tagged \`[MEETING — ${botId}]:\`. Follow the meetings skill: do NOT speak in response (no TTS output), instead maintain meeting-todos.md in the session workspace, optionally trigger background research silently. The voice-native user can still interact normally — only the meeting-tagged messages are the silent-observer path. Acknowledge by writing the initial meeting-todos.md skeleton.`,
                                     });
-                                    if (tokenRes.ok) {
-                                        const { token, url } = await tokenRes.json();
-                                        const params = new URLSearchParams({ token, url, room: roomName, botId: botLkId });
-                                        outputPageUrl = `${frontendUrl}/meeting-bot?${params.toString()}`;
-                                        console.log(`🎫 Meeting-bot token minted for room=${roomName} bot=${botLkId}`);
-                                    }
-                                    else {
-                                        const errText = await tokenRes.text().catch(() => '');
-                                        console.warn(`⚠️ meeting-bot-token mint failed (HTTP ${tokenRes.status}: ${errText.substring(0, 120)}) — falling back to legacy /meeting-output path`);
-                                    }
+                                    currentLLM.chat({ chatCtx: sysCtx });
+                                    console.log('📓 Meeting system injection sent to LLM');
                                 }
-                                catch (mintErr) {
-                                    console.warn(`⚠️ meeting-bot-token mint threw — falling back: ${mintErr.message}`);
+                                catch (sysErr) {
+                                    console.warn('⚠️ Meeting system injection failed:', sysErr.message);
                                 }
                             }
-                            else {
-                                console.log('ℹ️ No frontend URL (data.frontendBase + OSBORN_FRONTEND_URL both empty) — using legacy /meeting-output path');
+                            // Start polling the transcript every 30s. Each batch of new turns
+                            // is pushed to currentLLM.chat() tagged [MEETING — botId]: so the
+                            // skill kicks in. Poller dedups via first-word timestamp cursor.
+                            if (activeMeetingPoller) {
+                                activeMeetingPoller.stop();
+                                activeMeetingPoller = null;
                             }
-                            await sendToFrontend({ type: 'meeting_joining', message: 'Osborn is joining your meeting...' });
-                            const botId = await recallJoin.joinMeeting(meetingUrl, webhookBase, { outputPageUrl });
-                            const sessionId = currentLLM?.sessionId || currentResumeSessionId || 'default';
-                            recallJoin.registerBot(botId, sessionId);
-                            activeMeetingBotId = botId;
-                            await sendToFrontend({ type: 'meeting_joined', botId, message: 'Osborn has joined the meeting' });
+                            activeMeetingPoller = new MeetingTranscriptPoller({
+                                botId,
+                                recall: recallJoin,
+                                onTurns: async ({ formatted }) => {
+                                    if (!currentLLM) {
+                                        console.warn('📓 Meeting transcript arrived but currentLLM is null — dropping');
+                                        return;
+                                    }
+                                    const tagged = `[MEETING — ${botId}]:\n${formatted}`;
+                                    try {
+                                        const turnCtx = new llm.ChatContext();
+                                        turnCtx.addMessage({ role: 'user', content: tagged });
+                                        currentLLM.chat({ chatCtx: turnCtx });
+                                    }
+                                    catch (err) {
+                                        console.warn(`⚠️ Failed to forward meeting transcript to LLM: ${err.message}`);
+                                    }
+                                },
+                            });
+                            activeMeetingPoller.start();
                         }
                         catch (err) {
                             console.error('❌ Recall.ai join error:', err);
@@ -4263,6 +3841,12 @@ async function main() {
                 const recallLeave = getRecallClient();
                 if (recallLeave && botId) {
                     try {
+                        // Stop the transcript poller FIRST so no more transcript chunks get
+                        // forwarded to the LLM during the leave.
+                        if (activeMeetingPoller) {
+                            activeMeetingPoller.stop();
+                            activeMeetingPoller = null;
+                        }
                         await recallLeave.leaveMeeting(botId);
                         activeMeetingBotId = null;
                         await sendToFrontend({ type: 'meeting_left', botId });

package/dist/meeting-transcript-poller.d.ts

@@ -0,0 +1,60 @@
+/**
+ * MeetingTranscriptPoller — pulls Recall.ai meeting transcripts on a fixed
+ * interval and forwards new turns to the LLM as tagged `[MEETING — <botId>]:`
+ * messages.
+ *
+ * This replaces the older LiveKit/WebSocket audio pipeline that streamed raw
+ * PCM from Recall into a LiveKit room. The polling architecture is simpler
+ * (no parallel STT, no audio pipeline, no participant juggling), survives
+ * agent restarts (Recall keeps the transcript on its side), and the LLM
+ * never speaks in the meeting — it's a silent note-taker.
+ *
+ * Lifecycle:
+ *   const poller = new MeetingTranscriptPoller({ botId, recall, onTurns, intervalMs })
+ *   poller.start()
+ *   ...
+ *   poller.stop()  // on leave_meeting / disconnect / session switch
+ *
+ * Dedup strategy:
+ *   Each turn carries a `start_timestamp.relative` on its first word (seconds
+ *   since recording start). We track the highest cursor we've forwarded and
+ *   only send turns with a strictly greater first-word timestamp. This means
+ *   re-fetches don't double-deliver, and partial transcripts that get refined
+ *   later don't re-trigger LLM processing of already-handled turns.
+ *
+ * Error handling:
+ *   Transient fetch errors are logged + skipped (poll continues on next tick).
+ *   No backoff — Recall's transcript endpoint is stable enough that a 30s
+ *   cadence makes "slow start" non-issues self-recover within one cycle.
+ */
+import type { RecallClient, TranscriptTurn } from './recall-client.js';
+export interface MeetingTranscriptPollerOptions {
+    botId: string;
+    recall: RecallClient;
+    /** Called when new transcript turns arrive (de-duped). Get a fresh batch each tick. */
+    onTurns: (chunk: {
+        botId: string;
+        turns: TranscriptTurn[];
+        formatted: string;
+    }) => void | Promise<void>;
+    /** Default 30s — matches the user's stated cadence. */
+    intervalMs?: number;
+    /** Optional debug logger. */
+    onError?: (err: Error) => void;
+}
+export declare class MeetingTranscriptPoller {
+    #private;
+    constructor(opts: MeetingTranscriptPollerOptions);
+    start(): void;
+    stop(): void;
+}
+/**
+ * Format an array of turns into a single string for LLM consumption.
+ *
+ * Each turn becomes:
+ *   <Speaker>: <text>
+ *
+ * Whitespace-only words and zero-content turns are dropped. Returns empty
+ * string if nothing meaningful is in the batch.
+ */
+export declare function formatTurns(turns: TranscriptTurn[]): string;

package/dist/meeting-transcript-poller.js

@@ -0,0 +1,112 @@
+/**
+ * MeetingTranscriptPoller — pulls Recall.ai meeting transcripts on a fixed
+ * interval and forwards new turns to the LLM as tagged `[MEETING — <botId>]:`
+ * messages.
+ *
+ * This replaces the older LiveKit/WebSocket audio pipeline that streamed raw
+ * PCM from Recall into a LiveKit room. The polling architecture is simpler
+ * (no parallel STT, no audio pipeline, no participant juggling), survives
+ * agent restarts (Recall keeps the transcript on its side), and the LLM
+ * never speaks in the meeting — it's a silent note-taker.
+ *
+ * Lifecycle:
+ *   const poller = new MeetingTranscriptPoller({ botId, recall, onTurns, intervalMs })
+ *   poller.start()
+ *   ...
+ *   poller.stop()  // on leave_meeting / disconnect / session switch
+ *
+ * Dedup strategy:
+ *   Each turn carries a `start_timestamp.relative` on its first word (seconds
+ *   since recording start). We track the highest cursor we've forwarded and
+ *   only send turns with a strictly greater first-word timestamp. This means
+ *   re-fetches don't double-deliver, and partial transcripts that get refined
+ *   later don't re-trigger LLM processing of already-handled turns.
+ *
+ * Error handling:
+ *   Transient fetch errors are logged + skipped (poll continues on next tick).
+ *   No backoff — Recall's transcript endpoint is stable enough that a 30s
+ *   cadence makes "slow start" non-issues self-recover within one cycle.
+ */
+export class MeetingTranscriptPoller {
+    #opts;
+    #timer = null;
+    #cursor = -Infinity; // highest first-word.start_timestamp.relative we've forwarded
+    #inFlight = false; // prevent overlapping polls if one cycle runs long
+    #stopped = false;
+    constructor(opts) {
+        this.#opts = opts;
+    }
+    start() {
+        if (this.#timer)
+            return;
+        const interval = this.#opts.intervalMs ?? 30_000;
+        console.log(`📓 MeetingTranscriptPoller starting for bot=${this.#opts.botId.substring(0, 8)} (every ${Math.round(interval / 1000)}s)`);
+        // Fire once immediately so the LLM sees the meeting started, then on interval.
+        void this.#tick();
+        this.#timer = setInterval(() => void this.#tick(), interval);
+    }
+    stop() {
+        if (this.#stopped)
+            return;
+        this.#stopped = true;
+        if (this.#timer) {
+            clearInterval(this.#timer);
+            this.#timer = null;
+        }
+        console.log(`📓 MeetingTranscriptPoller stopped for bot=${this.#opts.botId.substring(0, 8)}`);
+    }
+    async #tick() {
+        if (this.#inFlight || this.#stopped)
+            return;
+        this.#inFlight = true;
+        try {
+            const all = await this.#opts.recall.getTranscript(this.#opts.botId);
+            const fresh = all.filter(t => {
+                const firstWordTs = t.words?.[0]?.start_timestamp?.relative;
+                return typeof firstWordTs === 'number' && firstWordTs > this.#cursor;
+            });
+            if (fresh.length === 0)
+                return;
+            // Advance cursor to highest seen first-word ts (across all returned turns,
+            // not just the fresh ones — guards against Recall returning a paged subset).
+            for (const t of all) {
+                const ts = t.words?.[0]?.start_timestamp?.relative;
+                if (typeof ts === 'number' && ts > this.#cursor)
+                    this.#cursor = ts;
+            }
+            const formatted = formatTurns(fresh);
+            if (!formatted)
+                return; // pure-whitespace fresh batch — skip
+            console.log(`📓 MeetingTranscriptPoller: ${fresh.length} new turn(s), cursor=${this.#cursor.toFixed(1)}s, chars=${formatted.length}`);
+            await this.#opts.onTurns({ botId: this.#opts.botId, turns: fresh, formatted });
+        }
+        catch (err) {
+            const e = err instanceof Error ? err : new Error(String(err));
+            this.#opts.onError?.(e);
+            console.warn(`⚠️ MeetingTranscriptPoller tick failed: ${e.message}`);
+        }
+        finally {
+            this.#inFlight = false;
+        }
+    }
+}
+/**
+ * Format an array of turns into a single string for LLM consumption.
+ *
+ * Each turn becomes:
+ *   <Speaker>: <text>
+ *
+ * Whitespace-only words and zero-content turns are dropped. Returns empty
+ * string if nothing meaningful is in the batch.
+ */
+export function formatTurns(turns) {
+    const lines = [];
+    for (const t of turns) {
+        const speaker = t.speaker || t.participant?.name || 'Unknown';
+        const text = (t.words ?? []).map(w => w.text).join(' ').replace(/\s+/g, ' ').trim();
+        if (!text)
+            continue;
+        lines.push(`${speaker}: ${text}`);
+    }
+    return lines.join('\n');
+}

package/dist/recall-client.d.ts

@@ -4,6 +4,36 @@ export interface RecallBot {
     meeting_url: string;
     status: string;
 }
+/**
+ * One transcript turn = one speaker's continuous utterance.
+ * Shape returned by GET /api/v1/bot/{bot_id}/transcript.
+ *
+ * Per Recall docs each turn contains:
+ *   - speaker: participant name (or 'Unknown')
+ *   - words: array of { text, start_timestamp.relative, end_timestamp.relative }
+ *   - The `start_timestamp.relative` (seconds since recording start) on the
+ *     FIRST word is the turn's start; we use this as the dedup cursor.
+ */
+export interface TranscriptTurn {
+    speaker?: string;
+    participant?: {
+        id?: number;
+        name?: string;
+        is_host?: boolean;
+    };
+    words: Array<{
+        text: string;
+        start_timestamp?: {
+            relative?: number;
+            absolute?: string;
+        };
+        end_timestamp?: {
+            relative?: number;
+            absolute?: string;
+        };
+    }>;
+    language?: string;
+}
 export interface TranscriptPayload {
     event: string;
     data: {
@@ -49,10 +79,27 @@ export declare class RecallClient extends EventEmitter {
      *                            room as the osborn agent (no separate WebSocket+WAV pipe).
      * @param opts.botName     Display name of the bot in the meeting
      */
-    joinMeeting(meetingUrl: string, webhookBaseUrl: string, opts?: {
-        outputPageUrl?: string;
+    joinMeeting(meetingUrl: string, _webhookBaseUrl: string, opts?: {
         botName?: string;
     }): Promise<string>;
+    /**
+     * Fetch the bot's current transcript. Returns an array of "transcript turns"
+     * (each turn = one speaker's utterance) sorted by start time. Use the bot's
+     * `recordings[0].id` from getBotStatus / bot record to locate the recording,
+     * then list its transcripts.
+     *
+     * Per Recall docs:
+     *   GET /api/v1/bot/{bot_id} → bot record incl. `recordings: [...]`
+     *   GET /api/v1/transcript/{transcript_id} → transcript with download_url
+     *   Download the transcript JSON from download_url to get the actual content.
+     *
+     * For the polling use case (called every ~30s), we use the simpler combined
+     * endpoint: `GET /api/v1/bot/{bot_id}/transcript` which Recall exposes as a
+     * convenience and returns the full transcript so far in one call. The caller
+     * is responsible for de-duping (keeping a since-cursor) so the LLM only sees
+     * new turns.
+     */
+    getTranscript(botId: string): Promise<TranscriptTurn[]>;
     leaveMeeting(botId: string): Promise<void>;
     getBotStatus(botId: string): Promise<string>;
     handleWebhook(payload: TranscriptPayload): void;

package/dist/recall-client.js

@@ -21,37 +21,20 @@ export class RecallClient extends EventEmitter {
      *                            room as the osborn agent (no separate WebSocket+WAV pipe).
      * @param opts.botName     Display name of the bot in the meeting
      */
-    async joinMeeting(meetingUrl, webhookBaseUrl, opts) {
+    async joinMeeting(meetingUrl, _webhookBaseUrl, opts) {
         const botName = opts?.botName ?? 'Osborn';
-        const outputPageUrl = opts?.outputPageUrl ?? `${webhookBaseUrl}/meeting-output`;
-        // Authoritative structure per https://docs.recall.ai/reference/bot_create
-        // and https://docs.recall.ai/docs/real-time-transcription:
+        // ARCHITECTURE (post-2026-05-22 polling redesign):
+        //   The bot joins by name only — visible in the meeting participant list as
+        //   "Osborn" but with no audio output and no avatar. We do NOT configure any
+        //   `output_media`, `audio_separate_raw`, or `realtime_endpoints` — instead
+        //   the agent polls Recall's REST transcript API every ~30s
+        //   (see MeetingTranscriptPoller) and feeds new turns into the LLM as
+        //   `[MEETING — <botId>]:` tagged messages. The meetings skill teaches the
+        //   LLM not to respond out loud to those messages, only to take notes.
         //
-        //   recording_config.transcript.provider  — transcription provider config
-        //   recording_config.realtime_endpoints   — webhook/websocket delivery
-        //
-        // IMPORTANT:
-        //   - Field is `realtime_endpoints` (NOT `real_time_endpoints`)
-        //   - `url` and `events` are flat on the endpoint object (NOT nested under `config`)
-        //   - `transcription_options` does NOT exist — use `transcript.provider`
-        //   - Both transcript.provider AND realtime_endpoints must be set, or no events delivered
-        //
-        // ARCHITECTURE (post-2026-05-22 redesign):
-        //   Input (meeting → osborn): Recall's documented WebSocket audio protocol.
-        //     `audio_separate_raw` config + websocket realtime endpoint streams
-        //     per-participant PCM (S16LE 16kHz mono, base64 in JSON) to the agent's
-        //     /meeting-audio-in WS handler. Bot's own audio is excluded by default
-        //     → zero possibility of feedback loop, no echo cancellation needed.
-        //   Output (osborn → meeting): webpage output_media (LiveKit-on-page). Bot
-        //     page subscribes to osborn's LiveKit audio track and plays it via
-        //     track.attach(); Recall captures the page's audio output and injects
-        //     into the meeting.
-        //   Webhook transcripts (transcript.data): retained as a SECONDARY signal —
-        //     the agent index.ts handler for this event currently logs but does NOT
-        //     forward to the LLM (intentionally disabled). The Deepgram WS path
-        //     above is the LLM input.
-        const httpBase = webhookBaseUrl.replace(/\/$/, '');
-        const wsBase = httpBase.replace(/^https?:\/\//, m => m === 'https://' ? 'wss://' : 'ws://');
+        //   We DO keep `recording_config.transcript.provider.recallai_streaming` so
+        //   Recall actually transcribes the meeting — the REST endpoint we poll
+        //   requires this to be configured, otherwise transcripts are empty.
         const res = await fetch(`${RECALL_BASE_URL}/bot`, {
             method: 'POST',
             headers: {
@@ -64,46 +47,12 @@ export class RecallClient extends EventEmitter {
                 recording_config: {
                     transcript: {
                         provider: {
-                            // recallai_streaming is built-in — no external API key needed,
-                            // low-latency, works across all meeting platforms.
-                            // Kept for the secondary webhook signal (display / future use);
-                            // LLM input now comes from the Deepgram WS pipe below.
                             recallai_streaming: {
                                 mode: 'prioritize_low_latency',
                                 language_code: 'en',
                             },
                         },
                     },
-                    // Per-participant raw PCM audio stream. Bot's own audio is excluded
-                    // (we don't set include_bot_in_recording.audio:true).
-                    audio_separate_raw: {},
-                    realtime_endpoints: [
-                        {
-                            // Transcript webhook (secondary signal; LLM forwarding disabled).
-                            type: 'webhook',
-                            url: `${httpBase}/webhook/recall`,
-                            events: ['transcript.data'],
-                        },
-                        {
-                            // Per-participant PCM audio → agent's Deepgram STT pipe.
-                            type: 'websocket',
-                            url: `${wsBase}/meeting-audio-in`,
-                            events: ['audio_separate_raw.data'],
-                        },
-                    ],
-                },
-                output_media: {
-                    camera: {
-                        // `kind` (not `type`) — confirmed from prior debugging.
-                        // The page Recall renders connects to LiveKit and plays osborn's
-                        // TTS audio via track.attach(); Recall captures the page audio.
-                        // The page does NOT call getUserMedia anymore — input now comes
-                        // from the audio_separate_raw WebSocket above.
-                        kind: 'webpage',
-                        config: {
-                            url: outputPageUrl,
-                        },
-                    },
                 },
             }),
         });
@@ -112,9 +61,37 @@ export class RecallClient extends EventEmitter {
             throw new Error(`Recall.ai join failed: ${res.status} ${err}`);
         }
         const bot = (await res.json());
-        console.log(`🤖 Recall.ai bot joined meeting: ${bot.id} (output page: ${outputPageUrl})`);
+        console.log(`🤖 Recall.ai bot joined meeting: ${bot.id} (polling-only, no audio pipeline)`);
         return bot.id;
     }
+    /**
+     * Fetch the bot's current transcript. Returns an array of "transcript turns"
+     * (each turn = one speaker's utterance) sorted by start time. Use the bot's
+     * `recordings[0].id` from getBotStatus / bot record to locate the recording,
+     * then list its transcripts.
+     *
+     * Per Recall docs:
+     *   GET /api/v1/bot/{bot_id} → bot record incl. `recordings: [...]`
+     *   GET /api/v1/transcript/{transcript_id} → transcript with download_url
+     *   Download the transcript JSON from download_url to get the actual content.
+     *
+     * For the polling use case (called every ~30s), we use the simpler combined
+     * endpoint: `GET /api/v1/bot/{bot_id}/transcript` which Recall exposes as a
+     * convenience and returns the full transcript so far in one call. The caller
+     * is responsible for de-duping (keeping a since-cursor) so the LLM only sees
+     * new turns.
+     */
+    async getTranscript(botId) {
+        const res = await fetch(`${RECALL_BASE_URL}/bot/${botId}/transcript`, {
+            headers: { 'Authorization': `Token ${this.#apiKey}` },
+        });
+        if (!res.ok) {
+            const err = await res.text().catch(() => '');
+            throw new Error(`Recall.ai transcript fetch failed: ${res.status} ${err.substring(0, 200)}`);
+        }
+        const turns = await res.json();
+        return Array.isArray(turns) ? turns : [];
+    }
     async leaveMeeting(botId) {
         await fetch(`${RECALL_BASE_URL}/bot/${botId}/leave_call`, {
             method: 'POST',

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.42",
+  "version": "0.9.43",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {
@@ -11,7 +11,7 @@
     "dev:logged": "tsx scripts/dev-logged.ts",
     "review": "tsx scripts/review.ts",
     "start": "tsx src/index.ts",
-    "build": "tsc && rm -rf dist/prompts && cp -r src/prompts dist/prompts && cp src/meeting-output.html dist/",
+    "build": "tsc && rm -rf dist/prompts && cp -r src/prompts dist/prompts",
     "room": "tsx src/index.ts --room",
     "prepublishOnly": "npm run build"
   },