fix(subagents): update SubagentConfig.system_prompt to str | None and add astream regression test

Agent-Logs-Url: https://github.com/bytedance/deer-flow/sessions/2ee03a26-e19b-4106-abc5-c76a2906383b

Co-authored-by: WillemJiang <219644+WillemJiang@users.noreply.github.com>

backend/packages/harness/deerflow/subagents/config.py

@@ -26,7 +26,7 @@ class SubagentConfig:
 
     name: str
     description: str
-    system_prompt: str
+    system_prompt: str | None = None
     tools: list[str] | None = None
     disallowed_tools: list[str] | None = field(default_factory=lambda: ["task"])
     skills: list[str] | None = None

backend/tests/test_subagent_executor.py

@@ -601,6 +601,77 @@ class TestAsyncExecutionPath:
         assert result.status == SubagentStatus.COMPLETED
         assert "Task" in result.result
 
+    @pytest.mark.anyio
+    async def test_aexecute_passes_at_most_one_system_message_to_agent(
+        self,
+        classes,
+        base_config,
+        monkeypatch: pytest.MonkeyPatch,
+        tmp_path,
+    ):
+        """Regression: messages sent to agent.astream must contain at most one
+        SystemMessage and it must be the first message.
+
+        This catches any regression where system_prompt would be re-injected
+        via create_agent() (e.g. system_prompt not passed as None) and appear
+        as a second SystemMessage, which providers like vLLM and Xinference
+        reject with "System message must be at the beginning."
+        """
+        from langchain_core.messages import AIMessage, SystemMessage
+
+        SubagentExecutor = classes["SubagentExecutor"]
+        SubagentStatus = classes["SubagentStatus"]
+
+        # Set up a skill so both system_prompt AND skill content are present,
+        # maximising the chance of catching a double-SystemMessage regression.
+        skill_dir = tmp_path / "regression-skill"
+        skill_dir.mkdir()
+        (skill_dir / "SKILL.md").write_text("Skill instruction text", encoding="utf-8")
+
+        monkeypatch.setattr(
+            "deerflow.skills.storage.get_or_new_skill_storage",
+            lambda *, app_config=None: SimpleNamespace(
+                load_skills=lambda *, enabled_only: [
+                    SimpleNamespace(name="regression-skill", skill_file=skill_dir / "SKILL.md")
+                ]
+            ),
+        )
+
+        captured_states: list[dict] = []
+
+        async def capturing_astream(state, **kwargs):
+            captured_states.append(state)
+            yield {"messages": [AIMessage(content="Done", id="msg-1")]}
+
+        mock_agent = MagicMock()
+        mock_agent.astream = capturing_astream
+
+        executor = SubagentExecutor(
+            config=base_config,
+            tools=[],
+            thread_id="test-thread",
+        )
+
+        with patch.object(executor, "_create_agent", return_value=mock_agent):
+            result = await executor._aexecute("Do something")
+
+        assert result.status == SubagentStatus.COMPLETED
+        assert len(captured_states) == 1, "astream should be called exactly once"
+        initial_messages = captured_states[0]["messages"]
+
+        system_messages = [m for m in initial_messages if isinstance(m, SystemMessage)]
+        assert len(system_messages) <= 1, (
+            f"Expected at most 1 SystemMessage but got {len(system_messages)}: {system_messages}"
+        )
+        if system_messages:
+            assert initial_messages[0] is system_messages[0], (
+                "SystemMessage must be the first message in the conversation"
+            )
+            # The consolidated SystemMessage must carry both the system_prompt
+            # and all skill content — nothing should be split across two messages.
+            assert base_config.system_prompt in system_messages[0].content
+            assert "Skill instruction text" in system_messages[0].content
+
 
 # -----------------------------------------------------------------------------
 # Sync Execution Path Tests