fix: relay confirmation interrupts past muted input

backend/services/pipecat/pipeline.py

@@ -8,7 +8,6 @@
 
 import asyncio
 import base64
-from collections.abc import Awaitable, Callable
 from io import BytesIO
 from typing import Any
 
@@ -107,21 +106,6 @@ ON_DEMAND_KNOWLEDGE_SYSTEM_HINT = (
 )
 
 
-async def _wait_for_output_stop(
-    wait_until_stopped: Callable[[], Awaitable[None]],
-    *,
-    timeout_seconds: float = 3.0,
-) -> None:
-    """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:
     """Keep tool metadata useful without letting it dominate the model context."""
     compact = " ".join(value.split())
@@ -437,6 +421,11 @@ async def run_pipeline(
         should_ignore_update=lambda: call_end.ending,
     )
     client_tools = ClientToolBroker()
+    # Confirmation nodes mute the user aggregator, which intentionally drops
+    # InterruptionFrame. Relay confirmation interruptions from the next
+    # processor instead so LLM, assistant aggregation, TTS, and output all see
+    # the frame before the workflow continues.
+    client_tools.set_interrupt_handler(user_turn_router.broadcast_interruption)
     vision_capture = VisionCaptureProcessor()
     knowledge_retrieval = KnowledgeRetrievalProcessor(
         automatic_knowledge_id,
@@ -751,11 +740,6 @@ async def run_pipeline(
         current_enable_interrupt = enable_interrupt
         current_turn_config = normalized
 
-    async def wait_for_output_stopped() -> None:
-        """Keep workflow continuation behind the interrupted output."""
-        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)。"""
         messages = context.get_messages()
@@ -788,7 +772,6 @@ async def run_pipeline(
             set_vision_scope=lambda scope: workflow_vision_scope.update(scope),
             vision_function=workflow_vision_function,
             set_input_enabled=lambda enabled: input_state.__setitem__("enabled", enabled),
-            wait_for_output_stopped=wait_for_output_stopped,
             apply_turn_config=apply_workflow_turn_config,
             flow_global_functions=flow_global_functions,
         ),
@@ -877,6 +860,7 @@ async def run_pipeline(
         vision_enabled=vision_enabled,
         vision_state=vision_state,
         submit_user_input=submit_user_input,
+        client_tools=client_tools,
     )
     runner = WorkerRunner(handle_sigint=False)
     run_status = "completed"

backend/services/pipecat/pipeline_events.py

@@ -16,6 +16,7 @@ from pipecat.runner.utils import (
     maybe_capture_participant_camera,
 )
 from pipecat.utils.time import time_now_iso8601
+from services.client_tools import ClientToolBroker
 from services.pipecat.processors import UserInput
 
 
@@ -32,6 +33,7 @@ def bind_cascade_pipeline_events(
     vision_enabled: bool,
     vision_state: dict[str, str | None],
     submit_user_input: Callable[[UserInput], Awaitable[None]] | None = None,
+    client_tools: ClientToolBroker | None = None,
 ) -> None:
     """Connect processors to transport events without owning pipeline assembly."""
 
@@ -159,6 +161,8 @@ def bind_cascade_pipeline_events(
 
     @assistant_aggregator.event_handler("on_interruption_processed")
     async def on_interruption_processed(_aggregator):
+        if client_tools is not None:
+            client_tools.on_interruption_processed()
         if not pending_user_inputs:
             return
         await finish_user_input(pending_user_inputs.pop(0))