osborn 0.9.84 → 0.9.86

package/dist/index.js

@@ -303,6 +303,12 @@ function startApiServer(workingDir, port) {
             req.on('end', () => {
                 try {
                     const payload = JSON.parse(body);
+                    // 0.9.86: log every receipt so we can SEE what Recall streams
+                    // (event type + word count) — the handler was silent, making it
+                    // impossible to tell whether webhooks arrive or are just filtered.
+                    const evt = payload?.event ?? 'unknown';
+                    const wc = (payload?.data?.data?.words ?? []).length;
+                    console.log(`📨 Recall webhook: event=${evt} words=${wc}`);
                     const recall = getRecallClient();
                     if (recall)
                         recall.handleWebhook(payload);
@@ -4624,7 +4630,11 @@ async function main() {
                             // 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);
+                            // 0.9.85: optional castUrl — a webpage to display as the bot's
+                            // camera (live feed / seeded site). Passed from the frontend or
+                            // env OSBORN_MEETING_CAST_URL for testing.
+                            const castUrl = data.castUrl || process.env.OSBORN_MEETING_CAST_URL || undefined;
+                            const botId = await recallJoin.joinMeeting(meetingUrl, webhookBase, { castUrl });
                             const sessionId = currentLLM?.sessionId || currentResumeSessionId || 'default';
                             recallJoin.registerBot(botId, sessionId);
                             activeMeetingBotId = botId;

package/dist/recall-client.d.ts

@@ -81,6 +81,7 @@ export declare class RecallClient extends EventEmitter {
      */
     joinMeeting(meetingUrl: string, webhookBaseUrl: string, opts?: {
         botName?: string;
+        castUrl?: string;
     }): Promise<string>;
     /**
      * Fetch the bot's current transcript. Returns an array of "transcript turns"

package/dist/recall-client.js

@@ -23,6 +23,19 @@ export class RecallClient extends EventEmitter {
      */
     async joinMeeting(meetingUrl, webhookBaseUrl, opts) {
         const botName = opts?.botName ?? 'Osborn';
+        // 0.9.85: optional CAST — display a webpage as the bot's camera in the
+        // meeting (Recall output_media, the piece the v0.9.44 rewrite removed).
+        // Point castUrl at the session-engine live-feed viewer, a seeded site
+        // (e.g. voice-native's LinkedIn), or a research/output page. This is the
+        // "the client sees content" visual side of the copilot. Left off by
+        // default (silent, invisible observer) — only casts when castUrl is set.
+        const cast = opts?.castUrl && /^https?:\/\//.test(opts.castUrl)
+            ? { output_media: { camera: { kind: 'webpage', config: { url: opts.castUrl } } } }
+            : {};
+        if (opts?.castUrl && !Object.keys(cast).length)
+            console.log(`⚠️ castUrl ignored (not http(s)): ${opts.castUrl}`);
+        else if (Object.keys(cast).length)
+            console.log(`📺 Meeting cast enabled — bot camera = ${opts.castUrl}`);
         // 0.9.84: LIVE transcript via realtime_endpoints. The v0.9.44 rewrite
         // removed this and left only 30s polling of the batch download_url — but
         // that URL is a POST-PROCESSING artifact (empty until the meeting ends),
@@ -58,6 +71,7 @@ export class RecallClient extends EventEmitter {
             body: JSON.stringify({
                 meeting_url: meetingUrl,
                 bot_name: botName,
+                ...cast,
                 recording_config: {
                     transcript: {
                         provider: {
@@ -135,9 +149,19 @@ export class RecallClient extends EventEmitter {
         const bot = (await res.json());
         return bot.status_changes?.at(-1)?.code ?? 'unknown';
     }
+    // 0.9.86: track the last text we emitted per speaker so partials don't
+    // spam the LLM with every incremental word — only emit when the utterance
+    // grows meaningfully or finalizes.
+    #lastEmitted = new Map();
     handleWebhook(payload) {
-        // Only process final transcripts (transcript.data), skip partials
-        if (payload.event !== 'transcript.data')
+        // Accept BOTH finals (transcript.data) AND partials (transcript.partial_data).
+        // recallai_streaming in prioritize_low_latency mode emits partials DURING
+        // the call and finals lag — ignoring partials (the old behavior) meant no
+        // live transcript at all (confirmed 2026-07-29: 0 webhook-processed turns
+        // mid-call). We emit partials for liveness, deduped against the last text.
+        const isFinal = payload.event === 'transcript.data';
+        const isPartial = payload.event === 'transcript.partial_data';
+        if (!isFinal && !isPartial)
             return;
         const words = payload.data?.data?.words ?? [];
         const text = words.map(w => w.text).join(' ').trim();
@@ -145,7 +169,15 @@ export class RecallClient extends EventEmitter {
             return;
         const speaker = payload.data?.data?.participant?.name ?? 'Unknown';
         const botId = payload.data?.bot?.id ?? 'unknown';
-        this.emit('transcript', { botId, speaker, text });
+        const key = `${botId}:${speaker}`;
+        // Dedup: skip a partial that just repeats/prefixes what we already sent.
+        const prev = this.#lastEmitted.get(key) ?? '';
+        if (!isFinal && (text === prev || (prev && text.startsWith(prev) && text.length - prev.length < 8)))
+            return;
+        this.#lastEmitted.set(key, text);
+        if (isFinal)
+            this.#lastEmitted.delete(key); // reset for the next utterance
+        this.emit('transcript', { botId, speaker, text, partial: !isFinal });
     }
     registerBot(botId, sessionId) {
         this.#activeBots.set(botId, sessionId);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "osborn",
-  "version": "0.9.84",
+  "version": "0.9.86",
   "description": "Voice AI coding assistant - local agent that connects to Osborn frontend",
   "type": "module",
   "bin": {