fix(workflow): prevent fixed speech continuation

backend/services/brains/workflow_brain.py

@@ -42,6 +42,7 @@ from services.action_runtime import (
     ActionStatus,
 )
 from services.action_stage import ActionStageRunner, ActionStageSpec, StageAction
+from services.fixed_speech import fixed_speech_context_message
 from services.knowledge import search as search_knowledge
 from services.message_policy import (
     MESSAGE_COMPLETION_POLICIES,
@@ -794,9 +795,9 @@ class WorkflowBrain(BaseBrain):
                     source="workflow-edge-transition",
                     node_id=str(edge.get("target") or "") or None,
                 )
-                context_messages.append(
-                    {"role": "assistant", "content": content}
-                )
+                context_message = fixed_speech_context_message(content)
+                if context_message is not None:
+                    context_messages.append(context_message)
         return await self._resolve_path(
             str(edge.get("target") or ""),
             leading_messages=context_messages,
@@ -828,10 +829,11 @@ class WorkflowBrain(BaseBrain):
                         if triggering_user_message
                         else {"role": "user", "content": triggering_user_text}
                     )
-                    agent_messages = [
-                        current_user_message,
-                        *context_messages,
-                    ]
+                    # Fixed speech is represented by system facts. Put those
+                    # facts before the triggering user turn so RESET contexts
+                    # still end with the real user input rather than a control
+                    # message.
+                    agent_messages = [*context_messages, current_user_message]
                 return self._agent_config(node_id, agent_messages)
             if node_type == "end":
                 await self._enter_end(node_id)
@@ -871,9 +873,9 @@ class WorkflowBrain(BaseBrain):
                         source="workflow-edge-transition",
                         node_id=target_id or None,
                     )
-                    context_messages.append(
-                        {"role": "assistant", "content": content}
-                    )
+                    context_message = fixed_speech_context_message(content)
+                    if context_message is not None:
+                        context_messages.append(context_message)
             node_id = str(edge.get("target") or "")
         raise RuntimeError("工作流连续自动跳转超过安全上限")
 
@@ -974,9 +976,9 @@ class WorkflowBrain(BaseBrain):
                 dict(message) for message in continuation.context_messages
             ]
             if result.speech:
-                context_messages.append(
-                    {"role": "assistant", "content": result.speech}
-                )
+                context_message = fixed_speech_context_message(result.speech)
+                if context_message is not None:
+                    context_messages.append(context_message)
 
             if not self._engine.has_outgoing(continuation.node_id):
                 self._state.enter(

backend/services/fixed_speech.py

@@ -12,6 +12,20 @@ from services.brains.base import BrainRuntime
 from services.runtime_variables import DynamicVariableStore
 
 
+FIXED_SPEECH_CONTEXT_MARKER = "[会话事实：以下固定消息已向用户播报]"
+
+
+def fixed_speech_context_message(content: str) -> dict[str, str] | None:
+    """Keep deterministic speech in context without inviting assistant continuation."""
+    text = content.strip()
+    if not text:
+        return None
+    return {
+        "role": "system",
+        "content": f"{FIXED_SPEECH_CONTEXT_MARKER}\n{text}",
+    }
+
+
 class FixedSpeechOutput:
     """Display and synthesize fixed speech without waiting for playback."""
 

backend/tests/test_brains.py

@@ -30,6 +30,7 @@ from services.brains.dify_llm import (
     normalize_api_base,
 )
 from services.brains.workflow_brain import ConfiguredFlowManager, WorkflowBrain
+from services.fixed_speech import FIXED_SPEECH_CONTEXT_MARKER
 from services.runtime_variables import prepare_dynamic_config
 from services.action_runtime import ActionOutcome, ActionStatus
 from services.workflow.models import (
@@ -1723,7 +1724,10 @@ class WorkflowBrainTests(unittest.IsolatedAsyncioTestCase):
                 {
                     "id": "agent1",
                     "type": "agent",
-                    "data": {"prompt": "收集基本信息"},
+                    "data": {
+                        "prompt": "收集基本信息",
+                        "entryMode": "generate",
+                    },
                 },
                 {
                     "id": "middle",
@@ -1846,6 +1850,21 @@ class WorkflowBrainTests(unittest.IsolatedAsyncioTestCase):
             if manager.current_node == "agent1":
                 break
         self.assertEqual(manager.current_node, "agent1")
+        self.assertEqual(
+            manager.configs[-1]["task_messages"],
+            [
+                {
+                    "role": "system",
+                    "content": f"{FIXED_SPEECH_CONTEXT_MARKER}\n欢迎使用。",
+                }
+            ],
+        )
+        self.assertFalse(
+            any(
+                message["role"] == "assistant"
+                for message in manager.configs[-1]["task_messages"]
+            )
+        )
 
         # The user-turn processor must return while the second Message is
         # still waiting for its transport playback boundary.
@@ -1868,8 +1887,13 @@ class WorkflowBrainTests(unittest.IsolatedAsyncioTestCase):
         self.assertEqual(
             manager.configs[-1]["task_messages"],
             [
+                {
+                    "role": "system",
+                    "content": (
+                        f"{FIXED_SPEECH_CONTEXT_MARKER}\n现在进入信息确认。"
+                    ),
+                },
                 {"role": "user", "content": "基本信息已经收集完成"},
-                {"role": "assistant", "content": "现在进入信息确认。"},
             ],
         )
 
@@ -2550,7 +2574,14 @@ class WorkflowBrainTests(unittest.IsolatedAsyncioTestCase):
             for frame in worker.frames
             if isinstance(frame, LLMMessagesAppendFrame)
             and frame.messages
-            == [{"role": "assistant", "content": "正在为你结束流程"}]
+            == [
+                {
+                    "role": "system",
+                    "content": (
+                        f"{FIXED_SPEECH_CONTEXT_MARKER}\n正在为你结束流程"
+                    ),
+                }
+            ]
         ]
         self.assertTrue(transition_context_frames)
         transition_events = [