feat(workflow): add edge-tool routing and realtime runtime

backend/services/pipecat/pipeline.py

@@ -16,6 +16,7 @@ from models import AssistantConfig
 from openai import AsyncOpenAI
 from PIL import Image
 from services.brains import Brain, BrainRuntime, build_brain
+from services.brains.base import RealtimeBrainRuntime
 from services.conversation_history import ConversationRecorder
 from services.pipecat.call_lifecycle import (
     CallEndCoordinator,
@@ -73,6 +74,7 @@ from services.pipecat.processors import (
     KnowledgeRetrievalProcessor,
     PassthroughLLMAssistantAggregator,
     RealtimeDynamicVariableProcessor,
+    RealtimeInputAudioGateProcessor,
     RealtimeUserInputProcessor,
     SessionUpdateProcessor,
     UserInput,
@@ -890,7 +892,32 @@ async def run_realtime_pipeline(
         instructions=brain.system_prompt(cfg),
     )
     input_sample_rate, output_sample_rate = realtime_audio_sample_rates(cfg)
-    user_input = RealtimeUserInputProcessor()
+    worker_holder: dict[str, PipelineWorker] = {}
+    input_state = {"enabled": True}
+
+    async def queue_call_end(reason: str) -> None:
+        worker = worker_holder.get("worker")
+        if worker is None:
+            return
+        logger.info(f"结束 Realtime 通话: reason={reason}")
+        await worker.queue_frame(
+            OutputTransportMessageUrgentFrame(
+                message={"type": "call-ended", "reason": reason}
+            )
+        )
+        await worker.queue_frame(EndFrame())
+
+    call_end = CallEndCoordinator(queue_call_end)
+    client_tools = ClientToolBroker()
+    client_tools.set_interrupt_handler(realtime.interrupt)
+    user_input = RealtimeUserInputProcessor(
+        should_ignore_input=lambda: (
+            call_end.ending or not input_state["enabled"]
+        )
+    )
+    input_gate = RealtimeInputAudioGateProcessor(
+        lambda: not call_end.ending and input_state["enabled"]
+    )
     dynamic_variables = RealtimeDynamicVariableProcessor(brain, cfg, realtime)
 
     async def refresh_realtime_instructions() -> None:
@@ -910,14 +937,22 @@ async def run_realtime_pipeline(
         channel=channel,
         runtime_mode=cfg.runtimeMode,
         session_id=cfg.conversation_id or None,
+        extra=(
+            WorkflowEngine(cfg.graph).session_metadata()
+            if cfg.type == "workflow"
+            else None
+        ),
     )
     pipeline = Pipeline(
         [
             transport.input(),
+            client_tools,
             session_update,
             user_input,
+            input_gate,
             realtime,
             dynamic_variables,
+            EndCallAfterSpeechProcessor(call_end),
             ConversationHistoryProcessor(recorder),
             transport.output(),
         ]
@@ -931,11 +966,28 @@ async def run_realtime_pipeline(
         ),
         enable_rtvi=False,
     )
+    worker_holder["worker"] = worker
+
+    def set_input_enabled(enabled: bool) -> None:
+        input_state["enabled"] = enabled
+
+    await brain.setup_realtime(
+        cfg,
+        RealtimeBrainRuntime(
+            realtime=realtime,
+            queue_frame=worker.queue_frame,
+            call_end=call_end,
+            session_id=cfg.conversation_id or "",
+            client_tools=client_tools,
+            set_input_enabled=set_input_enabled,
+        ),
+    )
 
     bind_realtime_pipeline_events(
         transport=transport,
         worker=worker,
         realtime=realtime,
+        brain=brain,
         text_input=user_input,
         greeting=greeting,
     )