fix(voice): emit turn_ended when workflow yields no audio

src/agents/voice/result.py

@@ -220,6 +220,12 @@ async def _turn_done(self):
                 )
             )
             self._text_buffer = ""
+        elif self._started_processing_turn:
+            # Turn was started (turn_started emitted) but produced no synthesizable text.
+            # Emit turn_ended directly so consumers see balanced lifecycle events.
+            local_queue = asyncio.Queue()
+            await local_queue.put(VoiceStreamEventLifecycle(event="turn_ended"))
+            self._ordered_tasks.append(local_queue)
         self._done_processing = True
         if self._dispatcher_task is None:
             self._dispatcher_task = asyncio.create_task(self._dispatch_audio())

tests/voice/test_pipeline.py

@@ -314,3 +314,21 @@ async def test_voicepipeline_multi_turn_on_start_exception_does_not_abort() -> N
 
     assert events[-1] == "session_ended"
     assert "error" not in events
+
+
+@pytest.mark.asyncio
+async def test_voicepipeline_empty_workflow_yield_emits_turn_ended() -> None:
+    # Workflow yields only an empty string (e.g. an LLM keepalive delta). turn_started must
+    # be balanced by turn_ended even though no audio is synthesized.
+    fake_stt = FakeSTT(["first"])
+    workflow = FakeWorkflow([[""]])
+    fake_tts = FakeTTS()
+    config = VoicePipelineConfig(tts_settings=TTSModelSettings(buffer_size=1))
+    pipeline = VoicePipeline(
+        workflow=workflow, stt_model=fake_stt, tts_model=fake_tts, config=config
+    )
+    audio_input = AudioInput(buffer=np.zeros(2, dtype=np.int16))
+    result = await pipeline.run(audio_input)
+    events, audio_chunks = await asyncio.wait_for(extract_events(result), timeout=2.0)
+    assert events == ["turn_started", "turn_ended", "session_ended"]
+    assert audio_chunks == []