Try to fix confirm and interrupt

backend/services/pipecat/pipeline.py

@@ -107,19 +107,19 @@ ON_DEMAND_KNOWLEDGE_SYSTEM_HINT = (
 )
 
 
-async def _wait_for_interrupted_output(
-    worker: PipelineWorker,
+async def _wait_for_output_stop(
+    wait_until_stopped: Callable[[], Awaitable[None]],
     *,
-    wait_until_stopped: Callable[[], Awaitable[None]] | None = None,
+    timeout_seconds: float = 3.0,
 ) -> None:
-    """Wait until an in-band interruption crosses pipeline and playback."""
-    if not await worker.flush_pipeline(timeout=2.0):
-        raise RuntimeError("输出中断帧未能及时穿过媒体管线")
-    if wait_until_stopped is not None:
-        try:
-            await asyncio.wait_for(wait_until_stopped(), timeout=2.0)
-        except TimeoutError as exc:
-            raise RuntimeError("等待客户端语音停止超时") from exc
+    """Wait for the transport's real stop event after an interruption."""
+    try:
+        await asyncio.wait_for(
+            wait_until_stopped(),
+            timeout=timeout_seconds,
+        )
+    except TimeoutError as exc:
+        raise RuntimeError("等待客户端语音停止超时") from exc
 
 
 def _compact_knowledge_metadata(value: str, max_length: int) -> str:
@@ -753,14 +753,8 @@ async def run_pipeline(
 
     async def wait_for_output_stopped() -> None:
         """Keep workflow continuation behind the interrupted output."""
-        wait_until_stopped = (
-            call_end.wait_until_silent if call_end.speaking else None
-        )
-        await _wait_for_interrupted_output(
-            worker,
-            wait_until_stopped=wait_until_stopped,
-        )
-
+        if call_end.speaking:
+            await _wait_for_output_stop(call_end.wait_until_silent)
 
     def set_system_prompt(text: str) -> None:
         """替换上下文里的系统提示(节点切换时整体替换,而非追加)。"""

backend/services/pipecat/service_factory.py

@@ -20,10 +20,12 @@ from services.pipecat.xfyun_super_tts import (
 )
 from services.pipecat.xfyun_tts import DEFAULT_XFYUN_TTS_URL, XfyunTTSService
 
-# TTS「说完」判定的空闲时长:默认 3.0s 过长(导致工作流结束节点说完后还要等约 3s
-# 才挂断,也拖慢日常轮次的交还)。设 1.0s 既能让结束语文字/音频送达,又更跟手。
-# 流式 TTS 句间音频间隔通常远小于 1s,不会把一段多句回复误判为结束。
-TTS_STOP_FRAME_TIMEOUT_S = 1.0
+# HTTP TTS may pause between response chunks. Keep Pipecat's wider default so
+# a long utterance is not marked stopped and routed onward before late chunks
+# arrive. WebSocket TTS has an explicit completion event and can use the short
+# idle fallback without delaying the end of a call.
+HTTP_TTS_STOP_FRAME_TIMEOUT_S = 3.0
+WEBSOCKET_TTS_STOP_FRAME_TIMEOUT_S = 1.0
 
 
 def config_with_resource(
@@ -158,7 +160,7 @@ def create_tts(cfg: AssistantConfig):
             volume=int(cfg.tts_values.get("volume") or 50),
             pitch=int(cfg.tts_values.get("pitch") or 50),
             push_stop_frames=True,
-            stop_frame_timeout_s=TTS_STOP_FRAME_TIMEOUT_S,
+            stop_frame_timeout_s=WEBSOCKET_TTS_STOP_FRAME_TIMEOUT_S,
         )
     if cfg.tts_interface_type not in {"openai-tts", "dashscope-tts"}:
         raise ValueError(f"不支持的 TTS 接口类型: {cfg.tts_interface_type}")
@@ -169,7 +171,7 @@ def create_tts(cfg: AssistantConfig):
     return OpenAITTSService(
         api_key=_require(cfg.tts_api_key, "TTS apiKey"),
         base_url=_require(cfg.tts_base_url, "TTS apiUrl"),
-        stop_frame_timeout_s=TTS_STOP_FRAME_TIMEOUT_S,
+        stop_frame_timeout_s=HTTP_TTS_STOP_FRAME_TIMEOUT_S,
         settings=OpenAITTSService.Settings(
             model=_require(cfg.tts_model, "TTS modelId"),
             voice=voice,