huggingface/tau — reverse-engineered prompt

Build me a small, readable Python terminal coding agent called Tau. I want it to feel like a real tool I can run inside any project folder, type requests like explain this repo, add tests, or fix this stack trace, and have it read files, edit code, run shell commands, and stream what it’s doing.

Please make it work both as an interactive terminal app and a simple print mode for one off prompts. It should keep durable session history on disk in JSONL so I can resume or branch past work. I also want basic slash commands for things like logging into a model provider, picking a model, managing sessions, exporting, and compacting context.

The important part is keeping the core clean and teachable. The agent brain should be reusable as a small library that emits events, and the terminal UI should just consume those events. Support common providers like OpenAI, Anthropic, Hugging Face, OpenRouter, and custom OpenAI compatible endpoints.

Include tests, docs, and a quickstart. You can check current docs online if needed.