fix: audio routing via MediaStream, auth context-destroyed handling, session cleanup

Co-authored-by: Cursor <cursoragent@cursor.com>

src/bot/audioProcedure.ts

@@ -29,9 +29,28 @@ export class AudioProcedure {
     await this._page.evaluate(() => {
       // Create a global audio context
       const AudioContext = window.AudioContext || (window as any).webkitAudioContext;
-      (window as any).__audioContext = new AudioContext();
+      const ctx = new AudioContext();
+      (window as any).__audioContext = ctx;
       (window as any).__audioQueue = [];
       (window as any).__isPlaying = false;
+
+      // Create a MediaStream destination so audio is routed into the
+      // browser's virtual microphone (picked up by Teams) instead of
+      // the default speaker output (ctx.destination).
+      const streamDest = ctx.createMediaStreamDestination();
+      (window as any).__audioStreamDest = streamDest;
+
+      // Expose the stream so headless Chromium can pipe it as mic input.
+      // navigator.mediaDevices.getUserMedia will be overridden to return this stream.
+      const audioStream = streamDest.stream;
+      const originalGetUserMedia = navigator.mediaDevices.getUserMedia.bind(navigator.mediaDevices);
+      navigator.mediaDevices.getUserMedia = async (constraints?: MediaStreamConstraints) => {
+        // If requesting audio only, return our TTS stream
+        if (constraints && constraints.audio && !constraints.video) {
+          return audioStream;
+        }
+        return originalGetUserMedia(constraints);
+      };
     });
 
     this._audioContext = true;
@@ -83,10 +102,12 @@ export class AudioProcedure {
           audioBuffer = await ctx.decodeAudioData(bytes.buffer);
         }
 
-        // Create source and play
+        // Create source and play through the MediaStream destination
+        // so audio is routed into the Teams microphone input, not speakers
         const source = ctx.createBufferSource();
         source.buffer = audioBuffer;
-        source.connect(ctx.destination);
+        const streamDest = (window as any).__audioStreamDest as MediaStreamAudioDestinationNode;
+        source.connect(streamDest || ctx.destination);
         source.start(0);
 
         // Return a promise that resolves when playback ends

src/bot/authProcedure.ts

@@ -197,9 +197,43 @@ export class AuthProcedure {
   /**
    * Verify that authentication was successful.
    * Checks if we're on a Microsoft/Teams page with an authenticated session.
+   * 
+   * Note: After successful login, Microsoft often triggers navigation/redirects
+   * which can destroy the execution context. An "Execution context was destroyed"
+   * error is treated as a successful login (navigation = login worked).
    */
   private async _verifyAuthentication(): Promise<boolean> {
     try {
+      // Wait for navigation that indicates login succeeded (redirect to Teams/Office)
+      try {
+        await this._page.waitForNavigation({
+          url: (url) =>
+            url.href.includes('teams.microsoft.com') ||
+            url.href.includes('office.com') ||
+            url.href.includes('myapps.microsoft.com') ||
+            url.href.includes('microsoftonline.com/common/oauth2'),
+          timeout: 15000,
+        });
+        this._logger.info('Navigation detected after login - authentication succeeded');
+        return true;
+      } catch (navError) {
+        const errorMessage = String(navError);
+
+        // "Execution context was destroyed" means the page navigated away
+        // from the login page, which indicates a successful login redirect
+        if (errorMessage.includes('Execution context was destroyed') ||
+            errorMessage.includes('execution context') ||
+            errorMessage.includes('navigation')) {
+          this._logger.info('Execution context destroyed during verification - treating as successful login (page navigated)');
+          // Give the page a moment to settle after navigation
+          await this._page.waitForTimeout(2000);
+          return true;
+        }
+
+        // Timeout - check where we ended up
+        this._logger.debug(`waitForNavigation did not match expected URL: ${navError}`);
+      }
+
       const url = this._page.url();
       
       // If we're on Teams or Microsoft portal, we're authenticated
@@ -237,6 +271,13 @@ export class AuthProcedure {
       // If we're somewhere else entirely, assume authenticated
       return true;
     } catch (error) {
+      const errorMessage = String(error);
+      // Catch "Execution context was destroyed" at the top level too
+      if (errorMessage.includes('Execution context was destroyed') ||
+          errorMessage.includes('execution context')) {
+        this._logger.info('Execution context destroyed during verification (top-level) - treating as successful login');
+        return true;
+      }
       this._logger.error(`Authentication verification error: ${error}`);
       return false;
     }

src/sessionManager.ts

@@ -96,11 +96,13 @@ export class SessionManager {
 
   /**
    * End a bot session and leave the meeting.
+   * Robust: handles cases where the session was already cleaned up
+   * (e.g. disconnected state removed it from the map).
    */
   async endSession(sessionId: string): Promise<void> {
     const orchestrator = this._sessions.get(sessionId);
     if (!orchestrator) {
-      logger.warn(`Session ${sessionId} not found`);
+      logger.warn(`Session ${sessionId} not found for endSession - may have already been cleaned up`);
       return;
     }
 
@@ -108,7 +110,10 @@ export class SessionManager {
 
     try {
       await orchestrator.stop();
+    } catch (error) {
+      logger.error(`Error stopping session ${sessionId}:`, error);
     } finally {
+      // Always remove from map after explicit end
       this._sessions.delete(sessionId);
     }
   }
@@ -166,15 +171,32 @@ export class SessionManager {
 
   /**
    * Handle state changes from orchestrators.
+   * 
+   * IMPORTANT: Do NOT delete from _sessions on 'disconnected' state.
+   * The orchestrator may enter 'disconnected' due to a transient WebSocket
+   * drop or browser crash. If we delete here, the Gateway's subsequent
+   * 'leave' command won't find the session in endSession().
+   * Cleanup is done explicitly in endSession() or shutdown().
+   * Only auto-remove on terminal 'error' state after a delay so the
+   * Gateway still has time to call endSession() first.
    */
   private _handleStateChange(sessionId: string, state: BotState, message?: string): void {
     logger.info(`Session ${sessionId} state: ${state}${message ? ` - ${message}` : ''}`);
 
-    // Clean up if disconnected or error
+    if (state === 'error') {
-    if (state === 'disconnected' || state === 'error') {
+      // Give Gateway a grace period to call endSession(), then auto-cleanup
+      setTimeout(() => {
+        if (this._sessions.has(sessionId)) {
+          const orch = this._sessions.get(sessionId);
+          if (orch && orch.state === 'error') {
+            logger.info(`Auto-cleaning stale error session ${sessionId}`);
             this._sessions.delete(sessionId);
           }
         }
+      }, 30000); // 30s grace period
+    }
+    // 'disconnected' state: do NOT delete - let endSession() handle it
+  }
 
   /**
    * Handle transcripts from orchestrators.