Preserve full user transcript across multiple inferences in one turn

When a stop-strategy chain splits inference-triggered from
finalization (e.g. `LLMTurnCompletionUserTurnStopStrategy` gating a
deferred detector), more than one inference can fire inside a single
user turn — each adds the new transcription segment to the context.
Previously each inference overwrote `_pending_user_turn_aggregation`,
so the eventual `on_user_turn_stopped` event surfaced only the
segment from the last inference, dropping anything the user said
before it.

Concatenate each segment into `_full_user_turn_aggregation` instead
of overwriting, and combine that running buffer with any post-final-
inference segment when emitting the public event.

src/pipecat/processors/aggregators/llm_response_universal.py

@@ -588,10 +588,14 @@ class LLMUserAggregator(LLMContextAggregator):
 
         self._user_is_muted = False
         self._user_turn_start_timestamp = ""
-        # Aggregation captured at inference-trigger time and surfaced again
+        # Full transcript across the user turn. Each
-        # in `on_user_turn_stopped`. Set to None when no inference-triggered
+        # `_on_user_turn_inference_triggered` push captures only the
-        # event has fired since the last finalization.
+        # new segment since the previous push (push_aggregation resets
-        self._pending_user_turn_aggregation: str | None = None
+        # `_aggregation` after writing to context); we accumulate those
+        # segments here so the eventual `on_user_turn_stopped` event
+        # surfaces the full turn transcript even when several
+        # inferences fire before finalization.
+        self._full_user_turn_aggregation: str | None = None
 
         self._user_turn_controller = UserTurnController(
             user_turn_strategies=user_turn_strategies,
@@ -862,7 +866,7 @@ class LLMUserAggregator(LLMContextAggregator):
         logger.debug(f"{self}: User started speaking (strategy: {strategy})")
 
         self._user_turn_start_timestamp = time_now_iso8601()
-        self._pending_user_turn_aggregation = None
+        self._full_user_turn_aggregation = None
 
         if params.enable_user_speaking_frames:
             await self.broadcast_frame(UserStartedSpeakingFrame)
@@ -881,12 +885,20 @@ class LLMUserAggregator(LLMContextAggregator):
     ):
         logger.debug(f"{self}: User turn inference triggered (strategy: {strategy})")
 
-        # Push aggregation now: this writes the user message to the context
+        # Push aggregation now: this writes the user message segment to
-        # and pushes LLMContextFrame, which is what kicks LLM inference.
+        # the context and emits LLMContextFrame, which kicks LLM
-        # `on_user_turn_stopped` later fires when the turn is semantically
+        # inference. Concatenate the segment into
-        # final and surfaces the aggregated message via the public event.
+        # `_full_user_turn_aggregation` so multiple inferences in the
-        aggregation = await self.push_aggregation()
+        # same turn don't lose earlier segments from the eventual
-        self._pending_user_turn_aggregation = aggregation
+        # `on_user_turn_stopped` event.
+        segment = await self.push_aggregation()
+        if segment:
+            if self._full_user_turn_aggregation:
+                self._full_user_turn_aggregation = (
+                    f"{self._full_user_turn_aggregation} {segment}".strip()
+                )
+            else:
+                self._full_user_turn_aggregation = segment
 
         await self._call_event_handler("on_user_turn_inference_triggered", strategy)
 
@@ -924,27 +936,25 @@ class LLMUserAggregator(LLMContextAggregator):
     ):
         """Maybe emit user turn stopped event.
 
-        The aggregation has typically already been pushed at
+        Earlier inference triggers in the same turn have already pushed
-        inference-trigger time and is cached in
+        their segments to the context and accumulated them into
-        ``self._pending_user_turn_aggregation``. Any aggregation that has
+        ``self._full_user_turn_aggregation``. Any aggregation that
-        accumulated since the last inference-trigger (e.g. transcriptions
+        arrived after the last inference trigger is flushed here so
-        that arrived between inference trigger and finalization) is flushed
+        end-of-turn content is never lost from the public event.
-        here so end-of-turn content is never lost.
 
         Args:
             strategy: The strategy that triggered the turn stop.
             on_session_end: If True, only emit if there's unemitted content
                 (avoids duplicate events when session ends).
         """
-        aggregation = await self.push_aggregation()
+        segment = await self.push_aggregation()
-        previous_aggregation = self._pending_user_turn_aggregation
+        full_aggregation = self._full_user_turn_aggregation
-        self._pending_user_turn_aggregation = None
+        self._full_user_turn_aggregation = None
 
-        content = None
+        if segment and full_aggregation:
-        if aggregation and previous_aggregation:
+            content = f"{full_aggregation} {segment}".strip()
-            content = f"{previous_aggregation} {aggregation}".strip()
         else:
-            content = previous_aggregation or aggregation
+            content = full_aggregation or segment
 
         if not on_session_end or content:
             message = UserTurnStoppedMessage(

tests/test_context_aggregators_universal.py

@@ -697,6 +697,75 @@ class TestLLMUserAggregator(unittest.IsolatedAsyncioTestCase):
 
         self.assertEqual(events, ["inference_triggered", "stopped"])
 
+    async def test_multiple_inferences_in_one_turn_preserve_aggregation(self):
+        """Two inference triggers before finalization should preserve the full user transcript.
+
+        When the LLM marks the first inference incomplete (○ / ◐) and the
+        user keeps speaking, the deferred upstream strategy fires a
+        second inference. Both the public ``on_user_turn_stopped`` event
+        and the conversation context should reflect the full user
+        utterance, not just the segment from the last inference.
+        """
+        from pipecat.frames.frames import UserTurnCompletedFrame
+        from pipecat.turns.user_stop import LLMTurnCompletionUserTurnStopStrategy, deferred
+
+        gating = LLMTurnCompletionUserTurnStopStrategy()
+        upstream = deferred(
+            SpeechTimeoutUserTurnStopStrategy(user_speech_timeout=TRANSCRIPTION_TIMEOUT)
+        )
+        context = LLMContext()
+        user_aggregator = LLMUserAggregator(
+            context,
+            params=LLMUserAggregatorParams(
+                user_turn_strategies=UserTurnStrategies(stop=[upstream, gating]),
+            ),
+        )
+
+        inference_count = 0
+        stop_message = None
+
+        @user_aggregator.event_handler("on_user_turn_inference_triggered")
+        async def on_inference_triggered(aggregator, strategy):
+            nonlocal inference_count
+            inference_count += 1
+
+        @user_aggregator.event_handler("on_user_turn_stopped")
+        async def on_stopped(aggregator, strategy, message):
+            nonlocal stop_message
+            stop_message = message
+
+        pipeline = Pipeline([user_aggregator])
+
+        frames_to_send = [
+            VADUserStartedSpeakingFrame(),
+            TranscriptionFrame(text="I'm thinking", user_id="", timestamp="now"),
+            SleepFrame(),
+            VADUserStoppedSpeakingFrame(),
+            SleepFrame(sleep=TRANSCRIPTION_TIMEOUT + 0.1),
+            # First inference fired here. Imagine the LLM returned ○;
+            # the turn is not yet finalized, so the user keeps talking.
+            VADUserStartedSpeakingFrame(),
+            TranscriptionFrame(text="about pizza", user_id="", timestamp="now"),
+            SleepFrame(),
+            VADUserStoppedSpeakingFrame(),
+            SleepFrame(sleep=TRANSCRIPTION_TIMEOUT + 0.1),
+            # Second inference fired here. Now the LLM returns ✓ and the
+            # turn finalizes via UserTurnCompletedFrame.
+            UserTurnCompletedFrame(),
+            SleepFrame(),
+        ]
+        await run_test(pipeline, frames_to_send=frames_to_send)
+
+        self.assertEqual(inference_count, 2)
+        self.assertIsNotNone(stop_message)
+        # The public event should report the full transcript, even
+        # though each inference push only writes its own segment to
+        # the context.
+        self.assertEqual(stop_message.content, "I'm thinking about pizza")
+
+        user_messages = [m for m in context.get_messages() if m.get("role") == "user"]
+        self.assertEqual([m["content"] for m in user_messages], ["I'm thinking", "about pizza"])
+
 
 class TestLLMAssistantAggregator(unittest.IsolatedAsyncioTestCase):
     async def test_empty(self):