feat：Agent上下文裁剪

app/agent/__init__.py

@@ -1,12 +1,17 @@
 import asyncio
-from typing import Dict, List, Any
+from typing import Dict, List, Any, Union
+import json
+import tiktoken
 
-from langchain.agents import AgentExecutor, create_openai_tools_agent
+from langchain.agents import AgentExecutor
 from langchain.prompts import ChatPromptTemplate, MessagesPlaceholder
 from langchain_community.callbacks import get_openai_callback
 from langchain_core.chat_history import InMemoryChatMessageHistory
-from langchain_core.messages import HumanMessage, AIMessage, ToolCall, ToolMessage, SystemMessage
+from langchain_core.messages import HumanMessage, AIMessage, ToolCall, ToolMessage, SystemMessage, trim_messages
+from langchain_core.runnables import RunnablePassthrough
 from langchain_core.runnables.history import RunnableWithMessageHistory
+from langchain.agents.format_scratchpad.openai_tools import format_to_openai_tool_messages
+from langchain.agents.output_parsers.openai_tools import OpenAIToolsAgentOutputParser
 
 from app.agent.callback import StreamingCallbackHandler
 from app.agent.memory import conversation_manager
@@ -140,15 +145,77 @@ class MoviePilotAgent:
             logger.error(f"初始化提示词失败: {e}")
             raise e
 
+    def _token_counter(self, messages: List[Union[HumanMessage, AIMessage, ToolMessage, SystemMessage]]) -> int:
+        """
+        通用的Token计数器
+        """
+        try:
+            # 尝试从模型获取编码集，如果失败则回退到 cl100k_base (大多数现代模型使用的编码)
+            try:
+                encoding = tiktoken.encoding_for_model(settings.LLM_MODEL)
+            except KeyError:
+                encoding = tiktoken.get_encoding("cl100k_base")
+
+            num_tokens = 0
+            for message in messages:
+                # 基础开销 (每个消息大约 3 个 token)
+                num_tokens += 3
+                
+                # 1. 处理文本内容 (content)
+                if isinstance(message.content, str):
+                    num_tokens += len(encoding.encode(message.content))
+                elif isinstance(message.content, list):
+                    for part in message.content:
+                        if isinstance(part, dict) and part.get("type") == "text":
+                            num_tokens += len(encoding.encode(part.get("text", "")))
+
+                # 2. 处理工具调用 (仅 AIMessage 包含 tool_calls)
+                if getattr(message, "tool_calls", None):
+                    for tool_call in message.tool_calls:
+                        # 函数名
+                        num_tokens += len(encoding.encode(tool_call.get("name", "")))
+                        # 参数 (转为 JSON 估算)
+                        args_str = json.dumps(tool_call.get("args", {}), ensure_ascii=False)
+                        num_tokens += len(encoding.encode(args_str))
+                        # 额外的结构开销 (ID 等)
+                        num_tokens += 3
+
+                # 3. 处理角色权重
+                num_tokens += 1
+
+            # 加上回复的起始 Token (大约 3 个 token)
+            num_tokens += 3
+            return num_tokens
+        except Exception as e:
+            logger.error(f"Token计数失败: {e}")
+            # 发生错误时返回一个保守的估算值
+            return len(str(messages)) // 4
+
     def _create_agent_executor(self) -> RunnableWithMessageHistory:
         """
         创建Agent执行器
         """
         try:
-            agent = create_openai_tools_agent(
-                llm=self.llm,
-                tools=self.tools,
-                prompt=self.prompt
+            # 消息裁剪器，防止上下文超出限制
+            trimmer = trim_messages(
+                max_tokens=settings.LLM_MAX_CONTEXT_TOKENS * 1000 * 0.8,
+                strategy="last",
+                token_counter=self._token_counter,
+                include_system=True,
+                allow_partial=False,
+                start_on="human",
+            )
+            # 创建Agent执行链
+            agent = (
+                RunnablePassthrough.assign(
+                    agent_scratchpad=lambda x: format_to_openai_tool_messages(
+                        x["intermediate_steps"]
+                    )
+                )
+                | self.prompt
+                | trimmer
+                | self.llm.bind_tools(self.tools)
+                | OpenAIToolsAgentOutputParser()
             )
             executor = AgentExecutor(
                 agent=agent,
@@ -190,7 +257,7 @@ class MoviePilotAgent:
 
             # 执行Agent
             logger.info(f"Agent执行推理: session_id={self.session_id}, input={message}")
-            await self._execute_agent(input_context)
+            result = await self._execute_agent(input_context)
 
             # 获取Agent回复
             agent_message = await self.callback_handler.get_message()
@@ -208,7 +275,7 @@ class MoviePilotAgent:
                     content=agent_message
                 )
             else:
-                agent_message = "很抱歉，智能体出错了，未能生成回复内容。"
+                agent_message = result.get("output") or "很抱歉，智能体出错了，未能生成回复内容。"
                 await self.send_agent_message(agent_message)
 
             return agent_message

app/core/config.py

@@ -427,6 +427,8 @@ class ConfigModel(BaseModel):
     LLM_API_KEY: Optional[str] = None
     # LLM基础URL（用于自定义API端点）
     LLM_BASE_URL: Optional[str] = "https://api.deepseek.com"
+    # LLM最大上下文Token数量（K）
+    LLM_MAX_CONTEXT_TOKENS: int = 64
     # LLM温度参数
     LLM_TEMPERATURE: float = 0.1
     # LLM最大迭代次数