fix(realtime): complete workflow message transitions

backend/services/workflow/realtime.py

@@ -35,6 +35,11 @@ from services.workflow_engine import WorkflowEngine
 
 
 MAX_AUTOMATIC_HOPS = 50
+AGENT_ENTRY_TRIGGER = (
+    "[工作流内部事件，不是用户消息] 当前 Agent 节点已激活。"
+    "请严格依据当前 instructions 立即生成进入该节点时的回复，"
+    "不要提及或复述本事件。"
+)
 ToolHandler = Callable[[dict[str, Any]], Awaitable[RealtimeToolResult]]
 
 
@@ -80,10 +85,18 @@ class RealtimeWorkflowOutput(WorkflowOutput):
                 **({"nodeId": node_id} if node_id else {}),
             }
         )
-        return await self._realtime.speak_fixed(
+        track_speech = getattr(self._runtime.call_end, "track_speech", None)
+        playback_completion: Awaitable[None] | None = None
+        if callable(track_speech):
+            playback_completion = track_speech()
+        provider_completion = await self._realtime.speak_fixed(
             content,
             suppress_transcript=True,
         )
+        # 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.
+        return playback_completion or provider_completion
 
 
 class WorkflowRealtimeController:
@@ -138,7 +151,7 @@ class WorkflowRealtimeController:
             allow_visible_actions=True,
         )
         if activation.continue_response:
-            await self._runtime.realtime.request_response()
+            await self._request_agent_entry_response()
 
     async def on_client_ready(self) -> None:
         await self._output.mark_client_ready()
@@ -160,7 +173,7 @@ class WorkflowRealtimeController:
         if activation and activation.after_output:
             await activation.after_output()
         elif activation and activation.continue_response:
-            await self._runtime.realtime.request_response()
+            await self._request_agent_entry_response()
         return SessionVariableUpdate(
             changed=changed,
             dynamic_variables=self._store.public_values(),
@@ -321,7 +334,13 @@ class WorkflowRealtimeController:
         if activation.after_output:
             await activation.after_output()
         elif activation.continue_response:
-            await self._runtime.realtime.request_response()
+            await self._request_agent_entry_response()
+
+    async def _request_agent_entry_response(self) -> None:
+        """Give provider inference a temporary turn for Agent entry speech."""
+        await self._runtime.realtime.request_response(
+            trigger_text=AGENT_ENTRY_TRIGGER,
+        )
 
     async def _follow_edge(
         self,