[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-co11ap5e-agent":3},{"id":4,"slug":5,"name":6,"summary":7,"stack":8,"background":13,"responsibility":14,"workflow":15,"architecture":22,"highlights":23,"retrospective":28,"repositoryUrl":29,"liveUrl":30,"featured":31},5,"co11ap5e-agent","co11ap5e_agent（轻量 ReAct Agent 框架）","不依赖 LangChain 的轻量 ReAct Agent 框架，用原生 OpenAI SDK 从零构建，兼容 DeepSeek \u002F DashScope \u002F OpenAI。",[9,10,11,12],"Python","OpenAI SDK","Pydantic","ReAct","从零构建一个不依赖重型框架的 Agent 运行内核：只保留 openai \u002F pydantic \u002F httpx，用原生 SDK 实现 ReAct 主循环、工具调用与记忆管理，目标是让每一层都能逐行解释。","独立完成框架设计、实现与测试。\n\n- 实现 Thought → Action → Observation 的 ReAct 核心循环\n- 用 @tool 装饰器从函数签名自动生成 OpenAI Function Calling JSON Schema\n- 构建短期（摘要压缩）、工作（任务 scratchpad）、长期（JSON 持久化）三层记忆体系\n- 实现容错 JSON 解析与 Plan-and-Execute 规划器\n- 处理工具超时、连续失败与最大迭代保护",[16,17,18,19,20,21],"接收用户任务","规划器判断是否需要拆解子任务","进入 ReAct 循环：思考 → 调工具 → 观察结果","工具执行与参数容错解析","结果写入短期记忆并维护工作记忆","达到目标后输出最终回答","仅依赖 openai \u002F pydantic \u002F httpx，核心为自研 ReAct 循环；记忆分短期 \u002F 工作 \u002F 长期三层；工具系统基于函数签名自动生成 Schema；提供 astream_chat 流式输出。",[24,25,26,27],"零框架依赖，核心循环全自研","三层记忆体系按任务生命周期管理","容错解析 LLM 输出的不规范 JSON","支持 DeepSeek \u002F DashScope \u002F OpenAI 多后端","Agent 框架的价值不在于炫技，而在于把「思考-行动-观察」的主循环、记忆边界和错误恢复设计清楚。自研让每个行为都可解释，也方便面试时讲清设计取舍。","https:\u002F\u002Fgithub.com\u002FS1rryNut\u002Fco11ap5e_agent",null,true]