Refactor workflow routing and greeting management in Brain classes

- Update WorkflowBrain to handle greeting playback more effectively, ensuring that the initial greeting completes before transitioning to the first node.
- Introduce new methods for managing greeting states and conditions, enhancing the interaction flow for user turns.
- Refactor WorkflowLLMRouter to improve routing logic and ensure proper handling of conditional paths.
- Enhance tests to verify the correct behavior of greeting management and routing under various scenarios, including waiting for audio playback to finish.
- Update frontend components to reflect changes in edge handling and improve user experience in workflow configurations.

backend/services/pipecat/pipeline_events.py

@@ -3,6 +3,8 @@
 from loguru import logger
 
 from pipecat.frames.frames import (
+    BotStartedSpeakingFrame,
+    BotStoppedSpeakingFrame,
     EndFrame,
     LLMMessagesAppendFrame,
     OutputTransportMessageUrgentFrame,
@@ -33,6 +35,28 @@ def bind_cascade_pipeline_events(
     pending_text_inputs: list[str] = []
     greeting_transcript_sent = False
     greeting_timestamp = ""
+    greeting_playback_pending = False
+    greeting_playback_started = False
+
+    # FlowManager already observes downstream frames for its own actions. Add
+    # to that filter instead of replacing it, then use the real transport
+    # playback boundary to release Workflow startup.
+    worker.add_reached_downstream_filter(
+        (BotStartedSpeakingFrame, BotStoppedSpeakingFrame)
+    )
+
+    @worker.event_handler("on_frame_reached_downstream")
+    async def on_frame_reached_downstream(_worker, frame):
+        nonlocal greeting_playback_pending, greeting_playback_started
+        if not greeting_playback_pending:
+            return
+        if isinstance(frame, BotStartedSpeakingFrame):
+            greeting_playback_started = True
+            return
+        if isinstance(frame, BotStoppedSpeakingFrame) and greeting_playback_started:
+            greeting_playback_pending = False
+            greeting_playback_started = False
+            await brain.on_greeting_finished()
 
     async def queue_transcript(role: str, content: str, timestamp: str) -> None:
         if not content:
@@ -132,7 +156,7 @@ def bind_cascade_pipeline_events(
 
     @transport.event_handler("on_client_connected")
     async def on_client_connected(_transport, _client):
-        nonlocal greeting_timestamp
+        nonlocal greeting_timestamp, greeting_playback_pending
         if vision_enabled:
             try:
                 vision_state["client_id"] = get_transport_client_id(
@@ -145,16 +169,24 @@ def bind_cascade_pipeline_events(
                 )
             except Exception as exc:  # noqa: BLE001 - media availability is optional
                 logger.warning(f"视觉理解摄像头捕获初始化失败: {exc}")
-        if greeting:
+        has_greeting = bool(greeting.strip())
+        if has_greeting:
             # Preserve the actual playback order. The transcript is delivered
             # later on client-ready, but the preview sorts by this timestamp.
             greeting_timestamp = greeting_timestamp or time_now_iso8601()
             if brain.spec.owns_context:
                 brain.prepare_greeting_context(greeting, context)
+            greeting_playback_pending = True
+
+        # Initialize the Workflow before the greeting is queued so a very
+        # short TTS response cannot finish before the brain arms its startup
+        # gate. Other brain types simply ignore greeting_pending.
+        await brain.on_connected(greeting_pending=has_greeting)
+
+        if has_greeting:
             await worker.queue_frame(
                 TTSSpeakFrame(greeting, append_to_context=False)
             )
-        await brain.on_connected()
 
     @transport.event_handler("on_client_disconnected")
     async def on_client_disconnected(_transport, _client):