fix(voice): clear stale paused-speech state across generation steps (ports livekit/agents#5594)

.changeset/clear-paused-speech-leak.md

@@ -0,0 +1,11 @@
+---
+'@livekit/agents': patch
+---
+
+fix(voice): clear stale paused-speech state across generation steps
+
+Ports livekit/agents#5594. Resets `pausedSpeech`, the false-interruption
+timer, and the paused audio output at the scheduling-loop boundary in
+`AgentActivity` after each generation step finishes, so paused state
+captured during an earlier silent step (e.g. a silent tool call) does not
+leak into the next step on the same `SpeechHandle` (e.g. the tool reply).

agents/src/voice/agent_activity.ts

@@ -1498,6 +1498,22 @@ export class AgentActivity implements RecognitionHooks {
         this._currentSpeech = speechHandle;
         speechHandle._authorizeGeneration();
         await speechHandle.waitIfNotInterrupted([speechHandle._waitForGeneration()]);
+
+        // Ref: python livekit-agents/livekit/agents/voice/agent_activity.py - 1365-1373 lines
+        // Clear stale paused-speech state captured by an earlier generation step
+        // (e.g. a silent tool-call step) so it doesn't leak into the next step
+        // running on the same SpeechHandle (e.g. the tool reply).
+        if (this.pausedSpeech && this.pausedSpeech.handle === this._currentSpeech) {
+          this.pausedSpeech = undefined;
+          if (this.falseInterruptionTimer) {
+            clearTimeout(this.falseInterruptionTimer);
+            this.falseInterruptionTimer = undefined;
+          }
+          const audioOutput = this.agentSession.output.audio;
+          if (audioOutput && audioOutput.canPause) {
+            audioOutput.resume();
+          }
+        }
         this._currentSpeech = undefined;
       }