From 513dda8bbb7640d7e8c974fc99e810bb803149e0 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E6=9C=B1=E6=BD=AE?= Date: Thu, 9 Apr 2026 15:02:28 +0800 Subject: [PATCH] =?UTF-8?q?=F0=9F=90=9B=20fix:=20=E4=BF=AE=E5=A4=8D=20Guid?= =?UTF-8?q?elineMiddleware=20=E5=AF=BC=E8=87=B4=20assistant=20message=20pr?= =?UTF-8?q?efill=20=E6=8A=A5=E9=94=99?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit enable_thinking 开启时，thinking 中间件将 AIMessage 追加到 messages 末尾， 导致不支持 assistant prefill 的模型返回 400 错误。 修复方式：在 AIMessage 后追加多语言 HumanMessage，确保消息以 user 结尾。 Co-Authored-By: Claude Opus 4.6 (1M context) --- agent/guideline_middleware.py | 26 ++++++++++++++++++++------ 1 file changed, 20 insertions(+), 6 deletions(-) diff --git a/agent/guideline_middleware.py b/agent/guideline_middleware.py index 9ddfb17..a9b2911 100644 --- a/agent/guideline_middleware.py +++ b/agent/guideline_middleware.py @@ -6,7 +6,7 @@ from utils.fastapi_utils import (extract_block_from_system_prompt, format_messag from langchain.chat_models import BaseChatModel from langgraph.runtime import Runtime -from langchain_core.messages import SystemMessage +from langchain_core.messages import SystemMessage, HumanMessage from typing import Any, Callable from langchain_core.callbacks import BaseCallbackHandler from langchain_core.outputs import LLMResult @@ -124,10 +124,11 @@ Action: Provide concise, friendly, and personified natural responses. response.additional_kwargs["message_tag"] = "THINK" response.content = f"{response.content}" - # 将响应添加到原始消息列表 - state['messages'] = state['messages'] + [response] + # 将响应添加到原始消息列表，并追加 HumanMessage 确保消息以 user 结尾 + # 某些模型不支持 assistant message prefill，要求最后一条消息必须是 user + state['messages'] = state['messages'] + [response, HumanMessage(content=self._get_follow_up_prompt())] return state - + async def abefore_agent(self, state: AgentState, runtime: Runtime) -> dict[str, Any] | None: if not self.guidelines: return None @@ -148,10 +149,23 @@ Action: Provide concise, friendly, and personified natural responses. response.additional_kwargs["message_tag"] = "THINK" response.content = f"{response.content}" - # 将响应添加到原始消息列表 - state['messages'] = state['messages'] + [response] + # 将响应添加到原始消息列表，并追加 HumanMessage 确保消息以 user 结尾 + # 某些模型不支持 assistant message prefill，要求最后一条消息必须是 user + state['messages'] = state['messages'] + [response, HumanMessage(content=self._get_follow_up_prompt())] return state + def _get_follow_up_prompt(self) -> str: + """根据语言返回引导主 agent 回复的提示""" + prompts = { + "ja": "以上の分析に基づいて、ユーザーに返信してください。", + "jp": "以上の分析に基づいて、ユーザーに返信してください。", + "zh": "请根据以上分析，回复用户。", + "zh-TW": "請根據以上分析，回覆用戶。", + "ko": "위 분석을 바탕으로 사용자에게 답변해 주세요.", + "en": "Based on the above analysis, please respond to the user.", + } + return prompts.get(self.language, prompts["en"]) + def wrap_model_call( self, request: ModelRequest,