fix(llm): adapt async tool results for compatible APIs

backend/services/fixed_speech.py

@@ -3,14 +3,17 @@
 from __future__ import annotations
 
 from collections.abc import Awaitable
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 from pipecat.frames.frames import OutputTransportMessageUrgentFrame, TTSSpeakFrame
 from pipecat.utils.time import time_now_iso8601
 
-from services.brains.base import BrainRuntime
+from services.pipecat.call_lifecycle import playback_marker_for
 from services.runtime_variables import DynamicVariableStore
 
+if TYPE_CHECKING:
+    from services.brains.base import BrainRuntime
+
 
 FIXED_SPEECH_CONTEXT_MARKER = "[会话事实：以下固定消息已向用户播报]"
 
@@ -79,6 +82,10 @@ class FixedSpeechOutput:
         await self._runtime.queue_frame(
             TTSSpeakFrame(content, append_to_context=False)
         )
+        playback_marker = playback_marker_for(playback_completion)
+        if playback_marker is not None:
+            await self._runtime.queue_frame(playback_marker)
+            playback_marker.completion.mark_queued()
         return playback_completion
 
     async def emit(self, message: dict[str, Any]) -> None:

backend/services/pipecat/call_lifecycle.py

@@ -3,14 +3,66 @@
 from __future__ import annotations
 
 import asyncio
-from collections import deque
 from collections.abc import Awaitable, Callable
+from dataclasses import dataclass
 
 from loguru import logger
-from pipecat.frames.frames import BotStartedSpeakingFrame, BotStoppedSpeakingFrame
+from pipecat.frames.frames import (
+    BotStartedSpeakingFrame,
+    BotStoppedSpeakingFrame,
+    DataFrame,
+    InterruptionFrame,
+)
 from pipecat.processors.frame_processor import FrameDirection, FrameProcessor
 
 
+class SpeechPlaybackCompletion:
+    """Awaitable completed by its exact marker at the transport output."""
+
+    def __init__(self, coordinator: CallEndCoordinator):
+        self._coordinator = coordinator
+        self._future: asyncio.Future[None] = (
+            asyncio.get_running_loop().create_future()
+        )
+        self._queued = False
+
+    def __await__(self):
+        return self._future.__await__()
+
+    def done(self) -> bool:
+        return self._future.done()
+
+    @property
+    def queued(self) -> bool:
+        return self._queued
+
+    def mark_queued(self) -> None:
+        self._queued = True
+
+    async def mark_played(self) -> None:
+        await self._coordinator.complete_tracked_speech(self)
+
+    def _resolve(self) -> None:
+        if not self._future.done():
+            self._future.set_result(None)
+
+
+@dataclass
+class FixedSpeechPlaybackMarkerFrame(DataFrame):
+    """Ordered frame that identifies one fixed utterance at transport output."""
+
+    completion: SpeechPlaybackCompletion
+
+
+def playback_marker_for(
+    completion: Awaitable[None] | None,
+) -> FixedSpeechPlaybackMarkerFrame | None:
+    """Build a marker only for the production call-end coordinator."""
+    if not isinstance(completion, SpeechPlaybackCompletion):
+        return None
+    return FixedSpeechPlaybackMarkerFrame(completion=completion)
+
+
 class CallEndCoordinator:
     """End immediately or after the currently armed closing speech finishes."""
 
@@ -22,17 +74,10 @@ class CallEndCoordinator:
         self._speech_stopped = asyncio.Event()
         self._speech_stopped.set()
         self._response_speech_started = False
-        self._tracked_speeches = 0
-        self._tracked_speech_completions: deque[asyncio.Future[None]] = deque()
+        self._tracked_speech_completions: set[SpeechPlaybackCompletion] = set()
         self._finish_after_tracked_speech = False
         self._finished = False
         self._reason = "completed"
-        # Only BotStoppedSpeakingFrame that follows a tracked BotStartedSpeakingFrame
-        # decrements the tracked-speech counter.  This prevents a stale or cross-talk
-        # stop frame (e.g. from a preceding LLM utterance) from consuming the counter
-        # meant for a fixed end-node message.
-        self._pending_tracked_starts = 0
-        self._current_speech_is_tracked = False
 
     @property
     def ending(self) -> bool:
@@ -59,18 +104,32 @@ class CallEndCoordinator:
         """Wait for the next observed bot speech to finish."""
         self._armed = True
 
-    def track_speech(self) -> Awaitable[None]:
+    def track_speech(self) -> SpeechPlaybackCompletion:
         """Register fixed speech and return its transport completion signal."""
-        completion = asyncio.get_running_loop().create_future()
-        self._tracked_speech_completions.append(completion)
-        self._tracked_speeches += 1
-        self._pending_tracked_starts += 1
+        completion = SpeechPlaybackCompletion(self)
+        self._tracked_speech_completions.add(completion)
         return completion
 
+    async def complete_tracked_speech(
+        self,
+        completion: SpeechPlaybackCompletion,
+    ) -> None:
+        """Complete one fixed utterance when its marker reaches output."""
+        if completion not in self._tracked_speech_completions:
+            return
+        self._tracked_speech_completions.remove(completion)
+        completion._resolve()
+        if (
+            self._finish_after_tracked_speech
+            and not self._tracked_speech_completions
+        ):
+            logger.info("所有工作流结束语播报完毕,挂断通话")
+            await self.finish()
+
     async def arm_after_tracked_speech(self) -> None:
         """Finish after every already queued fixed utterance has played."""
         self._finish_after_tracked_speech = True
-        if self._tracked_speeches == 0:
+        if not self._tracked_speech_completions:
             await self.finish()
 
     async def finish_after_current_speech(self, *, has_text: bool) -> None:
@@ -90,30 +149,25 @@ class CallEndCoordinator:
         await self._queue_end(self._reason)
 
     async def observe(self, frame) -> None:
-        if isinstance(frame, BotStartedSpeakingFrame):
+        if isinstance(frame, InterruptionFrame):
+            # Pipecat discards queued data frames on interruption, including
+            # playback markers. Treat already queued fixed speech as stopped so
+            # an interrupted Message cannot block a later EndNode forever.
+            interrupted = tuple(
+                completion
+                for completion in self._tracked_speech_completions
+                if completion.queued
+            )
+            for completion in interrupted:
+                await self.complete_tracked_speech(completion)
+        elif isinstance(frame, BotStartedSpeakingFrame):
             self._speaking = True
             self._speech_stopped.clear()
             self._response_speech_started = True
-            if self._pending_tracked_starts > 0:
-                self._pending_tracked_starts -= 1
-                self._current_speech_is_tracked = True
         elif isinstance(frame, BotStoppedSpeakingFrame) and self._speaking:
             self._speaking = False
             self._speech_stopped.set()
-            if self._current_speech_is_tracked:
-                self._current_speech_is_tracked = False
-                if self._tracked_speeches > 0:
-                    self._tracked_speeches -= 1
-                    completion = self._tracked_speech_completions.popleft()
-                    if not completion.done():
-                        completion.set_result(None)
-            if (
-                self._finish_after_tracked_speech
-                and self._tracked_speeches == 0
-            ):
-                logger.info("所有工作流结束语播报完毕,挂断通话")
-                await self.finish()
-            elif self._armed:
+            if self._armed:
                 logger.info("结束语播报完毕,挂断通话")
                 await self.finish()
 

backend/services/pipecat/service_factory.py

@@ -117,7 +117,7 @@ def create_llm(cfg: AssistantConfig):
         raise ValueError(f"不支持的 LLM 接口类型: {cfg.llm_interface_type}")
     extra_body = cfg.llm_values.get("extraBody")
     extra = {"extra_body": extra_body} if isinstance(extra_body, dict) else {}
-    return OpenAILLMService(
+    service = OpenAILLMService(
         api_key=_require(cfg.llm_api_key, "LLM apiKey"),
         base_url=_require(cfg.llm_base_url, "LLM apiUrl"),
         settings=OpenAILLMService.Settings(
@@ -125,6 +125,15 @@ def create_llm(cfg: AssistantConfig):
             extra=extra,
         ),
     )
+    # Pipecat represents late async-tool results with the newer `developer`
+    # role. Most OpenAI-compatible providers (including DeepSeek) only accept
+    # system/assistant/user/tool. This flag uses Pipecat's built-in adapter to
+    # downgrade developer messages before sending the request. Official or
+    # otherwise compatible endpoints can opt in through the resource values.
+    service.supports_developer_role = (
+        cfg.llm_values.get("supportsDeveloperRole") is True
+    )
+    return service
 
 
 def create_tts(cfg: AssistantConfig):

backend/services/pipecat/transports.py

@@ -14,15 +14,67 @@ from pipecat.transports.base_transport import TransportParams
 
 # WebRTC
 from pipecat.transports.smallwebrtc.connection import SmallWebRTCConnection
-from pipecat.transports.smallwebrtc.transport import SmallWebRTCTransport
+from pipecat.transports.smallwebrtc.transport import (
+    SmallWebRTCOutputTransport,
+    SmallWebRTCTransport,
+)
 
 # 裸 WS 音频流
 from pipecat.transports.websocket.fastapi import (
+    FastAPIWebsocketOutputTransport,
     FastAPIWebsocketTransport,
     FastAPIWebsocketParams,
 )
 from pipecat.serializers.protobuf import ProtobufFrameSerializer
 
+from services.pipecat.call_lifecycle import FixedSpeechPlaybackMarkerFrame
+
+
+class _PlaybackMarkerOutputMixin:
+    """Resolve fixed-speech markers after preceding audio has been sent."""
+
+    async def write_transport_frame(self, frame):
+        if isinstance(frame, FixedSpeechPlaybackMarkerFrame):
+            await frame.completion.mark_played()
+            return
+        await super().write_transport_frame(frame)
+
+
+class _WebRTCOutputTransport(
+    _PlaybackMarkerOutputMixin,
+    SmallWebRTCOutputTransport,
+):
+    pass
+
+
+class _WebRTCTransport(SmallWebRTCTransport):
+    def output(self) -> SmallWebRTCOutputTransport:
+        if not self._output:
+            self._output = _WebRTCOutputTransport(
+                self._client,
+                self._params,
+                name=self._input_name,
+            )
+        return self._output
+
+
+class _WebsocketOutputTransport(
+    _PlaybackMarkerOutputMixin,
+    FastAPIWebsocketOutputTransport,
+):
+    pass
+
+
+class _WebsocketTransport(FastAPIWebsocketTransport):
+    def __init__(self, websocket: WebSocket, params: FastAPIWebsocketParams):
+        super().__init__(websocket=websocket, params=params)
+        self._output = _WebsocketOutputTransport(
+            self,
+            self._client,
+            self._params,
+            name=self._output_name,
+        )
+
 
 def _base_params(*, video_in_enabled: bool = False) -> dict:
     """两种 transport 共享的音频参数。"""
@@ -41,7 +93,7 @@ def build_webrtc_transport(
     *,
     video_in_enabled: bool = False,
 ) -> SmallWebRTCTransport:
-    return SmallWebRTCTransport(
+    return _WebRTCTransport(
         webrtc_connection=connection,
         params=TransportParams(**_base_params(video_in_enabled=video_in_enabled)),
     )
@@ -51,7 +103,7 @@ def build_ws_transport(websocket: WebSocket) -> FastAPIWebsocketTransport:
     """裸 WS 输出。序列化用 protobuf(自定义客户端用同款解码);
     若对接电话商,把 serializer 换成对应的 TwilioFrameSerializer 等即可。
     """
-    return FastAPIWebsocketTransport(
+    return _WebsocketTransport(
         websocket=websocket,
         params=FastAPIWebsocketParams(
             serializer=ProtobufFrameSerializer(),

backend/services/workflow/realtime.py

@@ -23,6 +23,7 @@ from services.message_stage import (
     MessageStageRunner,
     MessageStageSpec,
 )
+from services.pipecat.call_lifecycle import playback_marker_for
 from services.pipecat.realtime_tools import RealtimeTool, RealtimeToolResult
 from services.runtime_variables import DynamicVariableError, DynamicVariableStore
 from services.system_tools import state_update_properties, system_tool_kind
@@ -99,6 +100,18 @@ class RealtimeWorkflowOutput(WorkflowOutput):
             content,
             suppress_transcript=True,
         )
+        playback_marker = playback_marker_for(playback_completion)
+        if playback_marker is not None:
+            async def finish_playback_tracking() -> None:
+                # The provider boundary is emitted after its final audio frame.
+                # Queueing the marker then places it behind that audio at output.
+                if provider_completion is not None:
+                    await provider_completion
+                await self._runtime.queue_frame(playback_marker)
+                playback_marker.completion.mark_queued()
+                await playback_completion
+
+            return asyncio.create_task(finish_playback_tracking())
         # Message playback policy and deterministic continuation must use the
         # transport boundary.  Provider response.done only means generation
         # has finished; audio may still be buffered at the output transport.
@@ -983,12 +996,20 @@ class WorkflowRealtimeController:
             return
         self._runtime.call_end.begin("workflow_completed")
         if message:
-            self._runtime.call_end.arm_after_speech()
             completion = await self._output.speak(
                 message,
                 source="workflow-end-speech",
                 node_id=node_id,
             )
+            arm_tracked = getattr(
+                self._runtime.call_end,
+                "arm_after_tracked_speech",
+                None,
+            )
+            if callable(arm_tracked):
+                await arm_tracked()
+            else:
+                self._runtime.call_end.arm_after_speech()
             if completion:
                 await completion
         else:

backend/tests/test_call_lifecycle.py

@@ -3,8 +3,15 @@ from __future__ import annotations
 import asyncio
 import unittest
 
-from pipecat.frames.frames import BotStartedSpeakingFrame, BotStoppedSpeakingFrame
-from services.pipecat.call_lifecycle import CallEndCoordinator
+from pipecat.frames.frames import (
+    BotStartedSpeakingFrame,
+    BotStoppedSpeakingFrame,
+    InterruptionFrame,
+)
+from services.pipecat.call_lifecycle import (
+    CallEndCoordinator,
+    playback_marker_for,
+)
 
 
 class CallEndCoordinatorTest(unittest.IsolatedAsyncioTestCase):
@@ -65,17 +72,68 @@ class CallEndCoordinatorTest(unittest.IsolatedAsyncioTestCase):
         self.coordinator.begin("workflow_completed")
         await self.coordinator.arm_after_tracked_speech()
 
-        await self.coordinator.observe(BotStartedSpeakingFrame())
-        await self.coordinator.observe(BotStoppedSpeakingFrame())
+        first_marker = playback_marker_for(first_completion)
+        second_marker = playback_marker_for(second_completion)
+        self.assertIsNotNone(first_marker)
+        self.assertIsNotNone(second_marker)
+
+        await first_marker.completion.mark_played()
         self.assertTrue(first_completion.done())
         self.assertFalse(second_completion.done())
         self.assertEqual(self.reasons, [])
 
-        await self.coordinator.observe(BotStartedSpeakingFrame())
-        await self.coordinator.observe(BotStoppedSpeakingFrame())
+        await second_marker.completion.mark_played()
         self.assertTrue(second_completion.done())
         self.assertEqual(self.reasons, ["workflow_completed"])
 
+    async def test_previous_speech_stop_does_not_complete_fixed_speech(self):
+        await self.coordinator.observe(BotStartedSpeakingFrame())
+        completion = self.coordinator.track_speech()
+        marker = playback_marker_for(completion)
+        self.coordinator.begin("workflow_completed")
+        await self.coordinator.arm_after_tracked_speech()
+
+        await self.coordinator.observe(BotStoppedSpeakingFrame())
+
+        self.assertFalse(completion.done())
+        self.assertEqual(self.reasons, [])
+        await marker.completion.mark_played()
+        self.assertEqual(self.reasons, ["workflow_completed"])
+
+    async def test_delayed_previous_speech_boundary_cannot_claim_marker(self):
+        completion = self.coordinator.track_speech()
+        marker = playback_marker_for(completion)
+        self.coordinator.begin("workflow_completed")
+        await self.coordinator.arm_after_tracked_speech()
+
+        await self.coordinator.observe(BotStartedSpeakingFrame())
+        await self.coordinator.observe(BotStoppedSpeakingFrame())
+
+        self.assertFalse(completion.done())
+        self.assertEqual(self.reasons, [])
+        await marker.completion.mark_played()
+        self.assertEqual(self.reasons, ["workflow_completed"])
+
+    async def test_interruption_completes_marker_already_in_output_queue(self):
+        completion = self.coordinator.track_speech()
+        marker = playback_marker_for(completion)
+        marker.completion.mark_queued()
+
+        await self.coordinator.observe(InterruptionFrame())
+
+        self.assertTrue(completion.done())
+
+    async def test_interruption_does_not_complete_marker_not_yet_queued(self):
+        completion = self.coordinator.track_speech()
+        marker = playback_marker_for(completion)
+
+        await self.coordinator.observe(InterruptionFrame())
+
+        self.assertFalse(completion.done())
+        marker.completion.mark_queued()
+        await marker.completion.mark_played()
+        self.assertTrue(completion.done())
+
 
 if __name__ == "__main__":
     unittest.main()

backend/tests/test_fixed_speech_playback.py

@@ -0,0 +1,146 @@
+from __future__ import annotations
+
+import asyncio
+import unittest
+from types import SimpleNamespace
+
+from models import AssistantConfig
+from pipecat.frames.frames import (
+    BotStartedSpeakingFrame,
+    BotStoppedSpeakingFrame,
+    TTSSpeakFrame,
+)
+import services.brains  # Initialize the brain registry before realtime imports.
+from services.fixed_speech import FixedSpeechOutput
+from services.pipecat.call_lifecycle import (
+    CallEndCoordinator,
+    FixedSpeechPlaybackMarkerFrame,
+)
+from services.pipecat.transports import build_ws_transport
+from services.runtime_variables import DynamicVariableStore
+from services.workflow.realtime import WorkflowRealtimeController
+from services.workflow_engine import WorkflowEngine
+
+
+class FixedSpeechPlaybackTest(unittest.IsolatedAsyncioTestCase):
+    async def test_pipeline_speech_queues_marker_immediately_after_tts(self):
+        queued = []
+
+        async def queue_end(_reason: str) -> None:
+            pass
+
+        async def queue_frame(frame) -> None:
+            queued.append(frame)
+
+        call_end = CallEndCoordinator(queue_end)
+        output = FixedSpeechOutput(
+            DynamicVariableStore({}),
+            SimpleNamespace(call_end=call_end, queue_frame=queue_frame),
+        )
+
+        completion = await output.speak(
+            "固定结束语",
+            source="test",
+            record_history=False,
+        )
+
+        self.assertIsInstance(queued[0], TTSSpeakFrame)
+        self.assertIsInstance(queued[1], FixedSpeechPlaybackMarkerFrame)
+        self.assertIs(queued[1].completion, completion)
+
+    async def test_websocket_output_resolves_marker(self):
+        async def queue_end(_reason: str) -> None:
+            pass
+
+        call_end = CallEndCoordinator(queue_end)
+        completion = call_end.track_speech()
+        marker = FixedSpeechPlaybackMarkerFrame(completion=completion)
+        websocket = SimpleNamespace(headers={})
+        output = build_ws_transport(websocket).output()
+
+        await output.write_transport_frame(marker)
+
+        self.assertTrue(completion.done())
+
+    async def test_realtime_end_ignores_unrelated_speech_boundaries(self):
+        graph = {
+            "specVersion": 3,
+            "settings": {},
+            "nodes": [
+                {"id": "start", "type": "start", "data": {}},
+                {
+                    "id": "end",
+                    "type": "end",
+                    "data": {"message": "感谢来电，再见。", "scope": "session"},
+                },
+            ],
+            "edges": [],
+        }
+        reasons = []
+        queued = []
+
+        async def queue_end(reason: str) -> None:
+            reasons.append(reason)
+
+        async def queue_frame(frame) -> None:
+            queued.append(frame)
+
+        class FakeRealtime:
+            def __init__(self):
+                self.provider_completion = None
+
+            async def update_session(self, _instructions, _tools):
+                pass
+
+            async def speak_fixed(self, _text, *, suppress_transcript=True):
+                self.provider_completion = (
+                    asyncio.get_running_loop().create_future()
+                )
+                return self.provider_completion
+
+        call_end = CallEndCoordinator(queue_end)
+        realtime = FakeRealtime()
+        controller = WorkflowRealtimeController(
+            cfg=AssistantConfig(type="workflow", graph=graph),
+            engine=WorkflowEngine(graph),
+            store=DynamicVariableStore({}),
+            runtime=SimpleNamespace(
+                realtime=realtime,
+                queue_frame=queue_frame,
+                call_end=call_end,
+                session_id="test-session",
+                client_tools=None,
+                set_input_enabled=lambda _enabled: None,
+                capture_image=None,
+            ),
+        )
+
+        end_task = asyncio.create_task(controller._enter_end("end"))
+        while realtime.provider_completion is None:
+            await asyncio.sleep(0)
+
+        await call_end.observe(BotStartedSpeakingFrame())
+        await call_end.observe(BotStoppedSpeakingFrame())
+        self.assertEqual(reasons, [])
+        self.assertFalse(end_task.done())
+
+        realtime.provider_completion.set_result(None)
+        marker = None
+        while marker is None:
+            await asyncio.sleep(0)
+            marker = next(
+                (
+                    frame
+                    for frame in queued
+                    if isinstance(frame, FixedSpeechPlaybackMarkerFrame)
+                ),
+                None,
+            )
+        await marker.completion.mark_played()
+        await end_task
+
+        self.assertEqual(reasons, ["workflow_completed"])
+
+
+if __name__ == "__main__":
+    unittest.main()

backend/tests/test_service_factory.py

@@ -4,13 +4,61 @@ import unittest
 from unittest.mock import patch
 
 from models import AssistantConfig
+from pipecat.processors.aggregators.llm_context import LLMContext
 from services.pipecat.service_factory import (
     HTTP_TTS_STOP_FRAME_TIMEOUT_S,
     WEBSOCKET_TTS_STOP_FRAME_TIMEOUT_S,
+    create_llm,
     create_tts,
 )
 
 
+class LLMServiceFactoryTest(unittest.TestCase):
+    def test_openai_compatible_llm_converts_async_tool_developer_result(self):
+        service = create_llm(
+            AssistantConfig(
+                llm_interface_type="openai-llm",
+                model="deepseek-chat",
+                llm_api_key="test-key",
+                llm_base_url="https://llm.example.test/v1",
+            )
+        )
+        context = LLMContext(
+            messages=[
+                {"role": "system", "content": "你是助手"},
+                {
+                    "role": "developer",
+                    "content": '{"type":"async_tool","status":"finished"}',
+                },
+                {"role": "user", "content": "ok"},
+            ]
+        )
+
+        params = service.get_llm_adapter().get_llm_invocation_params(
+            context,
+            convert_developer_to_user=not service.supports_developer_role,
+        )
+
+        self.assertEqual(
+            [message["role"] for message in params["messages"]],
+            ["system", "user", "user"],
+        )
+        self.assertNotIn("developer", str(params["messages"]))
+
+    def test_provider_can_explicitly_enable_developer_role(self):
+        service = create_llm(
+            AssistantConfig(
+                llm_interface_type="openai-llm",
+                model="gpt-compatible",
+                llm_api_key="test-key",
+                llm_base_url="https://llm.example.test/v1",
+                llm_values={"supportsDeveloperRole": True},
+            )
+        )
+
+        self.assertTrue(service.supports_developer_role)
+
+
 class TTSServiceFactoryTest(unittest.TestCase):
     def test_http_tts_keeps_wider_audio_chunk_timeout(self):
         config = AssistantConfig(