fix(agent_activity): enhance metrics reporting for user turn completion

Author: hudson-worden

livekit-agents/livekit/agents/voice/agent_activity.py

@@ -1877,6 +1877,28 @@ async def _user_turn_completed_task(
             transcript_confidence=info.transcript_confidence,
         )
 
+
+        metrics_report: llm.MetricsReport = {}
+        if self.stt:
+            metrics_report["stt_metadata"] = {
+                "model_name": self.stt.model,
+                "model_provider": self.stt.provider,
+            }
+        if info.started_speaking_at is not None:
+            metrics_report["started_speaking_at"] = info.started_speaking_at
+
+        if info.stopped_speaking_at is not None:
+            metrics_report["stopped_speaking_at"] = info.stopped_speaking_at
+
+        if info.transcription_delay is not None:
+            metrics_report["transcription_delay"] = info.transcription_delay
+
+        if info.end_of_turn_delay is not None:
+            metrics_report["end_of_turn_delay"] = info.end_of_turn_delay
+
+        if user_message is not None:
+            user_message.metrics = metrics_report
+
         if isinstance(self.llm, llm.RealtimeModel):
             if self.llm.capabilities.turn_detection:
                 return
@@ -1936,6 +1958,7 @@ async def _user_turn_completed_task(
             return
 
         on_user_turn_completed_delay = time.perf_counter() - start_time
+        metrics_report["on_user_turn_completed_delay"] = on_user_turn_completed_delay
 
         if isinstance(self.llm, llm.RealtimeModel):
             # ignore stt transcription for realtime model
@@ -1953,29 +1976,6 @@ async def _user_turn_completed_task(
                 self._session._conversation_item_added(user_message)
             return
 
-        metrics_report: llm.MetricsReport = {}
-        if self.stt:
-            metrics_report["stt_metadata"] = {
-                "model_name": self.stt.model,
-                "model_provider": self.stt.provider,
-            }
-        if info.started_speaking_at is not None:
-            metrics_report["started_speaking_at"] = info.started_speaking_at
-
-        if info.stopped_speaking_at is not None:
-            metrics_report["stopped_speaking_at"] = info.stopped_speaking_at
-
-        if info.transcription_delay is not None:
-            metrics_report["transcription_delay"] = info.transcription_delay
-
-        if info.end_of_turn_delay is not None:
-            metrics_report["end_of_turn_delay"] = info.end_of_turn_delay
-
-        metrics_report["on_user_turn_completed_delay"] = on_user_turn_completed_delay
-
-        if user_message is not None:
-            user_message.metrics = metrics_report
-
         speech_handle: SpeechHandle | None = None
         if preemptive := self._preemptive_generation:
             # make sure the on_user_turn_completed didn't change some request parameters