fix: normalize system messages before model calls

offline-backend-20260512/backend/packages/harness/deerflow/agents/middlewares/rag_citation_middleware.py

@@ -12,7 +12,7 @@ from typing import Any, override
 from langchain.agents import AgentState
 from langchain.agents.middleware import AgentMiddleware
 from langchain.agents.middleware.types import ModelCallResult, ModelRequest, ModelResponse
-from langchain_core.messages import SystemMessage
+from langchain_core.messages import HumanMessage
 
 from deerflow.runtime.references import (
     build_latest_reference_prompt,
@@ -105,7 +105,10 @@ class RagCitationMiddleware(AgentMiddleware[AgentState]):
         )
         patched_messages = [
             *messages,
-            SystemMessage(content=prompt),
+            HumanMessage(
+                content=prompt,
+                additional_kwargs={"hide_from_ui": True, "lc_source": "rag_citation"},
+            ),
         ]
         return request.override(messages=patched_messages)
 

offline-backend-20260512/backend/packages/harness/deerflow/agents/middlewares/writing_setup_round_cap_middleware.py

@@ -14,13 +14,13 @@ from typing import override
 from langchain.agents import AgentState
 from langchain.agents.middleware import AgentMiddleware
 from langchain.agents.middleware.types import ModelCallResult, ModelRequest, ModelResponse
-from langchain_core.messages import HumanMessage, SystemMessage
+from langchain_core.messages import HumanMessage
 
 
 class WritingSetupRoundCapMiddleware(AgentMiddleware[AgentState]):
     """达到澄清轮数上限时，强制模型调用 ``setup_writing`` 提交。
 
-    软约束（注入 SystemMessage）+ 硬约束（``tool_choice`` 锁定工具）双保险：即使个别模型
+    软约束（注入隐藏 HumanMessage）+ 硬约束（``tool_choice`` 锁定工具）双保险：即使个别模型
     无视提示，``tool_choice`` 也会逼它产出一次 ``setup_writing`` 调用。
     """
 
@@ -33,8 +33,8 @@ class WritingSetupRoundCapMiddleware(AgentMiddleware[AgentState]):
         human_count = sum(1 for m in messages if isinstance(m, HumanMessage))
         return human_count >= self._max_rounds
 
-    def _reminder(self) -> SystemMessage:
+    def _reminder(self) -> HumanMessage:
-        return SystemMessage(
+        return HumanMessage(
             content=(
                 f"【系统强制】意图澄清已达 {self._max_rounds} 轮上限。"
                 "现在必须立即调用 setup_writing 提交写作配置，禁止再向用户提任何问题。"
@@ -42,7 +42,8 @@ class WritingSetupRoundCapMiddleware(AgentMiddleware[AgentState]):
                 "word_count_target 默认 800、target_audience 默认「普通读者」、"
                 "writing_mode 默认 loose、material_source 默认 knowledge_base、"
                 "article_type 不确定就留空。"
-            )
+            ),
+            additional_kwargs={"hide_from_ui": True, "lc_source": "writing_setup_round_cap"},
         )
 
     def _patch(self, request: ModelRequest) -> ModelRequest:

offline-backend-20260512/backend/packages/harness/deerflow/models/message_ordering.py

@@ -0,0 +1,82 @@
+"""Utilities for provider-specific chat message ordering constraints."""
+
+from __future__ import annotations
+
+from typing import Any
+
+from langchain_core.messages import BaseMessage, SystemMessage
+
+
+_SYSTEM_SEPARATOR = "\n\n---\n\n"
+
+
+def normalize_chat_payload_messages(payload: dict[str, Any]) -> None:
+    """Ensure payload ``system`` messages are represented once at the beginning.
+
+    Several OpenAI-compatible gateways reject requests when a transient
+    ``system`` message appears after user/assistant/tool history. LangChain and
+    DeerFlow middleware can legitimately create those transient messages, so the
+    provider boundary folds every system payload into a single leading message
+    while preserving all non-system messages in their original order.
+    """
+    messages = payload.get("messages")
+    if not isinstance(messages, list):
+        return
+
+    normalized = _normalize_payload_message_list(messages)
+    if normalized is not messages:
+        payload["messages"] = normalized
+
+
+def normalize_langchain_messages(messages: list[BaseMessage]) -> list[BaseMessage]:
+    """Return LangChain messages with all system context merged at the front."""
+    system_messages = [message for message in messages if isinstance(message, SystemMessage)]
+    if not system_messages:
+        return messages
+
+    first_non_system_index = next((idx for idx, message in enumerate(messages) if not isinstance(message, SystemMessage)), len(messages))
+    if len(system_messages) == first_non_system_index:
+        return messages
+
+    merged = system_messages[0].model_copy(
+        update={
+            "content": _merge_contents([message.content for message in system_messages]),
+        }
+    )
+    return [merged, *(message for message in messages if not isinstance(message, SystemMessage))]
+
+
+def _normalize_payload_message_list(messages: list[Any]) -> list[Any]:
+    system_messages = [message for message in messages if isinstance(message, dict) and message.get("role") == "system"]
+    if not system_messages:
+        return messages
+
+    first_non_system_index = next(
+        (
+            idx
+            for idx, message in enumerate(messages)
+            if not (isinstance(message, dict) and message.get("role") == "system")
+        ),
+        len(messages),
+    )
+    if len(system_messages) == first_non_system_index:
+        return messages
+
+    merged_system = dict(system_messages[0])
+    merged_system["content"] = _merge_contents([message.get("content") for message in system_messages])
+    return [merged_system, *(message for message in messages if not (isinstance(message, dict) and message.get("role") == "system"))]
+
+
+def _merge_contents(contents: list[Any]) -> Any:
+    if all(isinstance(content, str) for content in contents):
+        return _SYSTEM_SEPARATOR.join(content for content in contents if content)
+
+    blocks: list[Any] = []
+    for idx, content in enumerate(contents):
+        if idx > 0:
+            blocks.append({"type": "text", "text": _SYSTEM_SEPARATOR})
+        if isinstance(content, list):
+            blocks.extend(content)
+        elif content is not None:
+            blocks.append({"type": "text", "text": str(content)})
+    return blocks

offline-backend-20260512/backend/packages/harness/deerflow/models/mindie_provider.py

@@ -6,10 +6,12 @@ import uuid
 from collections.abc import Iterator
 
 import httpx
-from langchain_core.messages import AIMessage, AIMessageChunk, HumanMessage, ToolMessage
+from langchain_core.messages import AIMessage, AIMessageChunk, HumanMessage, ToolMessage
-from langchain_core.outputs import ChatGenerationChunk, ChatResult
+from langchain_core.outputs import ChatGenerationChunk, ChatResult
-from langchain_openai import ChatOpenAI
+from langchain_openai import ChatOpenAI
-
+
+from deerflow.models.message_ordering import normalize_langchain_messages
+
 
 def _fix_messages(messages: list) -> list:
     """Sanitize incoming messages for MindIE compatibility.
@@ -55,7 +57,7 @@ def _fix_messages(messages: list) -> list:
 
         fixed.append(msg.model_copy(update={"content": text}))
 
-    return fixed
+    return normalize_langchain_messages(fixed)
 
 
 def _parse_xml_tool_call_to_dict(content: str) -> tuple[str, list[dict]]:

offline-backend-20260512/backend/packages/harness/deerflow/models/patched_deepseek.py

@@ -9,10 +9,12 @@ on all assistant messages when thinking mode is enabled.
 
 from typing import Any
 
-from langchain_core.language_models import LanguageModelInput
+from langchain_core.language_models import LanguageModelInput
-from langchain_core.messages import AIMessage
+from langchain_core.messages import AIMessage
-from langchain_deepseek import ChatDeepSeek
+from langchain_deepseek import ChatDeepSeek
-
+
+from deerflow.models.message_ordering import normalize_chat_payload_messages
+
 
 class PatchedChatDeepSeek(ChatDeepSeek):
     """ChatDeepSeek with proper reasoning_content preservation.
@@ -65,9 +67,10 @@ class PatchedChatDeepSeek(ChatDeepSeek):
             ai_messages = [m for m in original_messages if isinstance(m, AIMessage)]
             assistant_payloads = [(i, m) for i, m in enumerate(payload_messages) if m.get("role") == "assistant"]
 
-            for (idx, payload_msg), ai_msg in zip(assistant_payloads, ai_messages):
+            for (idx, payload_msg), ai_msg in zip(assistant_payloads, ai_messages):
-                reasoning_content = ai_msg.additional_kwargs.get("reasoning_content")
+                reasoning_content = ai_msg.additional_kwargs.get("reasoning_content")
-                if reasoning_content is not None:
+                if reasoning_content is not None:
-                    payload_messages[idx]["reasoning_content"] = reasoning_content
+                    payload_messages[idx]["reasoning_content"] = reasoning_content
-
+
-        return payload
+        normalize_chat_payload_messages(payload)
+        return payload

offline-backend-20260512/backend/packages/harness/deerflow/models/patched_minimax.py

@@ -20,11 +20,13 @@ from langchain_core.language_models import LanguageModelInput
 from langchain_core.messages import AIMessage, AIMessageChunk
 from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResult
 from langchain_openai import ChatOpenAI
-from langchain_openai.chat_models.base import (
+from langchain_openai.chat_models.base import (
-    _convert_delta_to_message_chunk,
+    _convert_delta_to_message_chunk,
-    _create_usage_metadata,
+    _create_usage_metadata,
-)
+)
-
+
+from deerflow.models.message_ordering import normalize_chat_payload_messages
+
 _THINK_TAG_RE = re.compile(r"<think>\s*(.*?)\s*</think>", re.DOTALL)
 
 
@@ -112,9 +114,10 @@ class PatchedChatMiniMax(ChatOpenAI):
                 **extra_body,
                 "reasoning_split": True,
             }
-        else:
+        else:
-            payload["extra_body"] = {"reasoning_split": True}
+            payload["extra_body"] = {"reasoning_split": True}
-        return payload
+        normalize_chat_payload_messages(payload)
+        return payload
 
     def _convert_chunk_to_generation_chunk(
         self,

offline-backend-20260512/backend/packages/harness/deerflow/models/patched_openai.py

@@ -29,9 +29,11 @@ from typing import Any
 
 from langchain_core.language_models import LanguageModelInput
 from langchain_core.messages import AIMessage, AIMessageChunk
-from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResult
+from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResult
-from langchain_openai import ChatOpenAI
+from langchain_openai import ChatOpenAI
-
+
+from deerflow.models.message_ordering import normalize_chat_payload_messages
+
 
 class PatchedChatOpenAI(ChatOpenAI):
     """ChatOpenAI with ``thought_signature`` preservation for Gemini thinking via OpenAI gateway.
@@ -87,15 +89,16 @@ class PatchedChatOpenAI(ChatOpenAI):
                 if payload_msg.get("role") == "assistant" and isinstance(orig_msg, AIMessage):
                     _restore_tool_call_signatures(payload_msg, orig_msg)
                     _restore_reasoning_content(payload_msg, orig_msg)
-        else:
+        else:
-            # Fallback: match assistant-role entries positionally against AIMessages.
+            # Fallback: match assistant-role entries positionally against AIMessages.
-            ai_messages = [m for m in original_messages if isinstance(m, AIMessage)]
+            ai_messages = [m for m in original_messages if isinstance(m, AIMessage)]
-            assistant_payloads = [(i, m) for i, m in enumerate(payload_messages) if m.get("role") == "assistant"]
+            assistant_payloads = [(i, m) for i, m in enumerate(payload_messages) if m.get("role") == "assistant"]
-            for (_, payload_msg), ai_msg in zip(assistant_payloads, ai_messages):
+            for (_, payload_msg), ai_msg in zip(assistant_payloads, ai_messages):
-                _restore_tool_call_signatures(payload_msg, ai_msg)
+                _restore_tool_call_signatures(payload_msg, ai_msg)
-                _restore_reasoning_content(payload_msg, ai_msg)
+                _restore_reasoning_content(payload_msg, ai_msg)
-
+
-        return payload
+        normalize_chat_payload_messages(payload)
+        return payload
 
     def _convert_chunk_to_generation_chunk(
         self,

offline-backend-20260512/backend/packages/harness/deerflow/models/vllm_provider.py

@@ -32,9 +32,11 @@ from langchain_core.messages import (
 )
 from langchain_core.messages.tool import tool_call_chunk
 from langchain_core.outputs import ChatGeneration, ChatGenerationChunk, ChatResult
-from langchain_openai import ChatOpenAI
+from langchain_openai import ChatOpenAI
-from langchain_openai.chat_models.base import _create_usage_metadata
+from langchain_openai.chat_models.base import _create_usage_metadata
-
+
+from deerflow.models.message_ordering import normalize_chat_payload_messages
+
 
 def _normalize_vllm_chat_template_kwargs(payload: dict[str, Any]) -> None:
     """Map DeerFlow's legacy ``thinking`` toggle to vLLM/Qwen's ``enable_thinking``.
@@ -182,13 +184,14 @@ class VllmChatModel(ChatOpenAI):
             for payload_msg, orig_msg in zip(payload_messages, original_messages):
                 if payload_msg.get("role") == "assistant" and isinstance(orig_msg, AIMessage):
                     _restore_reasoning_field(payload_msg, orig_msg)
-        else:
+        else:
-            ai_messages = [message for message in original_messages if isinstance(message, AIMessage)]
+            ai_messages = [message for message in original_messages if isinstance(message, AIMessage)]
-            assistant_payloads = [message for message in payload_messages if message.get("role") == "assistant"]
+            assistant_payloads = [message for message in payload_messages if message.get("role") == "assistant"]
-            for payload_msg, ai_msg in zip(assistant_payloads, ai_messages):
+            for payload_msg, ai_msg in zip(assistant_payloads, ai_messages):
-                _restore_reasoning_field(payload_msg, ai_msg)
+                _restore_reasoning_field(payload_msg, ai_msg)
-
+
-        return payload
+        normalize_chat_payload_messages(payload)
+        return payload
 
     def _create_chat_result(self, response: dict | openai.BaseModel, generation_info: dict | None = None) -> ChatResult:
         """Preserve vLLM reasoning on non-streaming responses."""

offline-backend-20260512/backend/packages/harness/tests/test_message_ordering.py

@@ -0,0 +1,84 @@
+import pytest
+from langchain_core.messages import AIMessage, HumanMessage, SystemMessage
+
+from deerflow.agents.middlewares.writing_setup_round_cap_middleware import WritingSetupRoundCapMiddleware
+from deerflow.models.message_ordering import normalize_chat_payload_messages, normalize_langchain_messages
+from deerflow.models.patched_openai import PatchedChatOpenAI
+
+
+def _multi_turn_payload_with_late_systems(turns: int):
+    messages = [{"role": "system", "content": "base system"}]
+    for idx in range(turns):
+        messages.append({"role": "user", "content": f"user {idx}"})
+        messages.append({"role": "assistant", "content": f"assistant {idx}"})
+        if idx >= 9 and (idx + 1) % 10 == 0:
+            messages.append({"role": "system", "content": f"late summary after {idx + 1} turns"})
+    messages.append({"role": "system", "content": "tail forced instruction"})
+    return {"messages": messages}
+
+
+@pytest.mark.parametrize("turns", [12, 25, 50])
+def test_normalize_chat_payload_messages_merges_late_systems_after_many_turns(turns):
+    payload = _multi_turn_payload_with_late_systems(turns)
+
+    normalize_chat_payload_messages(payload)
+
+    roles = [message["role"] for message in payload["messages"]]
+    assert roles[0] == "system"
+    assert roles.count("system") == 1
+    assert roles[1:5] == ["user", "assistant", "user", "assistant"]
+    assert "base system" in payload["messages"][0]["content"]
+    assert "late summary after 10 turns" in payload["messages"][0]["content"]
+    assert "tail forced instruction" in payload["messages"][0]["content"]
+
+
+def test_normalize_langchain_messages_moves_late_systems_before_history():
+    messages = [SystemMessage(content="base system")]
+    for idx in range(12):
+        messages.append(HumanMessage(content=f"user {idx}"))
+        messages.append(AIMessage(content=f"assistant {idx}"))
+        if idx == 10:
+            messages.append(SystemMessage(content="late compaction summary"))
+
+    normalized = normalize_langchain_messages(messages)
+
+    assert isinstance(normalized[0], SystemMessage)
+    assert sum(isinstance(message, SystemMessage) for message in normalized) == 1
+    assert normalized[0].content == "base system\n\n---\n\nlate compaction summary"
+    assert isinstance(normalized[1], HumanMessage)
+    assert normalized[1].content == "user 0"
+
+
+@pytest.mark.parametrize("turns", [12, 25, 50])
+def test_patched_openai_payload_has_only_leading_system_after_many_turns(turns):
+    model = PatchedChatOpenAI(
+        model="test-model",
+        api_key="test-key",
+        base_url="http://127.0.0.1:9/v1",
+    )
+    messages = [SystemMessage(content="base system")]
+    for idx in range(turns):
+        messages.append(HumanMessage(content=f"user {idx}"))
+        messages.append(AIMessage(content=f"assistant {idx}"))
+        if idx >= 9 and (idx + 1) % 10 == 0:
+            messages.append(SystemMessage(content=f"late system after {idx + 1} turns"))
+    messages.append(SystemMessage(content="tail system"))
+
+    payload = model._get_request_payload(messages)
+
+    roles = [message["role"] for message in payload["messages"]]
+    assert roles[0] == "system"
+    assert roles.count("system") == 1
+    assert "base system" in payload["messages"][0]["content"]
+    assert "late system after 10 turns" in payload["messages"][0]["content"]
+    assert "tail system" in payload["messages"][0]["content"]
+
+
+def test_writing_round_cap_reminder_is_not_late_system_message():
+    middleware = WritingSetupRoundCapMiddleware(max_rounds=3)
+
+    reminder = middleware._reminder()
+
+    assert isinstance(reminder, HumanMessage)
+    assert not isinstance(reminder, SystemMessage)
+    assert reminder.additional_kwargs["hide_from_ui"] is True