fix(#2009): emit truncation on VAD interrupt

Author: ruskaruma

src/agents/realtime/openai_realtime.py

@@ -621,10 +621,30 @@ async def _handle_ws_event(self, event: dict[str, Any]):
             last_audio = self._audio_state_tracker.get_last_audio_item()
             if last_audio is not None:
                 item_id, content_index = last_audio
+                playback_state = self._get_playback_state()
+                playback_item_id = playback_state.get("current_item_id")
+                playback_content_index = playback_state.get("current_item_content_index") or 0
+                playback_elapsed_ms = playback_state.get("elapsed_ms")
                 await self._emit_event(
                     RealtimeModelAudioInterruptedEvent(item_id=item_id, content_index=content_index)
                 )
 
+                elapsed_override = getattr(parsed, "audio_end_ms", None)
+                if elapsed_override is None or elapsed_override <= 0:
+                    effective_elapsed_ms = playback_elapsed_ms
+                else:
+                    effective_elapsed_ms = float(elapsed_override)
+
+                if playback_item_id and effective_elapsed_ms is not None:
+                    truncated_ms = max(int(round(effective_elapsed_ms)), 0)
+                    await self._send_raw_message(
+                        _ConversionHelper.convert_interrupt(
+                            playback_item_id,
+                            playback_content_index,
+                            truncated_ms,
+                        )
+                    )
+
                 # Reset trackers so subsequent playback state queries don't
                 # reference audio that has been interrupted client‑side.
                 self._audio_state_tracker.on_interrupted()
@@ -643,9 +663,6 @@ async def _handle_ws_event(self, event: dict[str, Any]):
                 )
                 if not automatic_response_cancellation_enabled:
                     await self._cancel_response()
-            # Avoid sending conversation.item.truncate here. When the session's
-            # turn_detection.interrupt_response is enabled (GA default), the server emits
-            # conversation.item.truncated after the VAD start and takes care of history updates.
         elif parsed.type == "response.created":
             self._ongoing_response = True
             await self._emit_event(RealtimeModelTurnStartedEvent())

tests/realtime/test_openai_realtime.py

@@ -443,7 +443,7 @@ async def test_transcription_related_and_timeouts_and_speech_started(self, model
 
         # Prepare tracker state to simulate ongoing audio
         model._audio_state_tracker.set_audio_format("pcm16")
-        model._audio_state_tracker.on_audio_delta("i1", 0, b"aaaa")
+        model._audio_state_tracker.on_audio_delta("i1", 0, b"a" * 48)
         model._ongoing_response = True
 
         # Patch sending to avoid websocket dependency
@@ -464,6 +464,17 @@ async def test_transcription_related_and_timeouts_and_speech_started(self, model
             }
         )
 
+        truncate_events = [
+            call.args[0]
+            for call in model._send_raw_message.await_args_list
+            if getattr(call.args[0], "type", None) == "conversation.item.truncate"
+        ]
+        assert truncate_events
+        truncate_event = truncate_events[0]
+        assert truncate_event.item_id == "i1"
+        assert truncate_event.content_index == 0
+        assert truncate_event.audio_end_ms == 1
+
         # Output transcript delta
         await model._handle_ws_event(
             {