yzy1223/jarvis

J.A.R.V.I.S. is a Python based multi-agent coding assistant, documented mainly in Chinese, built around the idea that it should get smarter every time you use it. The README points out that tools like Cursor, GitHub Copilot, and Devin start every conversation from scratch and never carry lessons forward. This project tries to fix that by remembering what worked in past sessions and reusing it. The system is split into three layers. The first layer is an orchestrator that figures out what you want, recommends a technology stack, and breaks the request into smaller sub-tasks. The second layer is a group of worker agents, covering frontend, backend, database, and testing, that actually write code and run commands. The third layer checks the work and learns from it: an Inspector confirms the agents followed the rules, a Cortex module turns successful sessions into searchable memories stored in a vector database, and a Skill Forge promotes an experience into a reusable named skill once it has been used three times, so the fourth similar request skips straight to execution. A Reflection Engine also studies failures across five categories and updates the system's understanding of the project to avoid repeating mistakes. Safety is handled through six layers of protection: file operations are restricted to the project folder, a blocklist blocks sensitive files like .env, a whitelist limits which shell commands can run, a matching blacklist blocks dangerous ones, and all agent output is checked before being accepted. To run it, you need Python 3.10 or newer and an OpenAI compatible API key, with support listed for OpenAI, DeepSeek, SiliconFlow, Qwen, or a local Ollama server. Setup is a pip install plus an .env file with your API key, then a single command to start. It is released under the Apache 2.0 license. The full README is longer than what was shown.