merge guideline-assistant-message-error

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
@@ -124,8 +124,9 @@ Action: Provide concise, friendly, and personified natural responses.
         response.additional_kwargs["message_tag"] = "THINK"
         response.content = f"<think>{response.content}</think>"
 
-        # 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:
@@ -148,10 +149,23 @@ Action: Provide concise, friendly, and personified natural responses.
         response.additional_kwargs["message_tag"] = "THINK"
         response.content = f"<think>{response.content}</think>"
 
-        # 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,

prompt/FACT_RETRIEVAL_PROMPT.md

@@ -167,6 +167,7 @@ Remember the following:
   - For colloquial or grammatically informal expressions (common in spoken Japanese, Chinese, Korean, etc.), understand the full intended meaning and record it in a clear, semantically complete form.
   - In Japanese, spoken language often omits particles (e.g., が, を, に). When extracting facts, include the necessary particles to make the meaning unambiguous. For example: "私は林檎好きです" should be understood as "林檎が好き" (likes apples), not literally "私は林檎好き".
   - When the user expresses a preference or opinion in casual speech, record the core preference/opinion clearly. Remove the subject pronoun (私は/I) since facts are about the user by default, but keep all other semantic components intact.
+
 - **CRITICAL for People/Relationship Tracking**:
   - Write people-related facts in plain, natural language. Do NOT use structured formats like "Contact:", "referred as", or "DEFAULT when user says".
   - Good examples: "Michael Johnson is a colleague, also called Mike", "田中さんは友達", "滨田太郎は「滨田」とも呼ばれている"