Trampoline-AI/fractal — reverse-engineered prompt

Build me a Python command line tool called Fractal that feels like a bare bones terminal AI agent for working inside a real codebase. I want the core idea to be that the agent can handle big, messy, deep context by recursively breaking work down, inspecting lots of files, and then coming back with a distilled answer or making edits in place. It should support normal interactive terminal use with multi turn conversation history, and also a headless mode so another coding agent like Cursor or Claude Code could hand Fractal a heavy analysis task and get the result back.

Please make it run user tasks against the actual workspace through an isolated sandbox, with setup checks for the sandbox tool, Python 3.11 or newer, and a model provider or local model login. A simple first run config flow and a status command would be great. Include a one line installer, sensible docs, and tests. Keep the product intentionally simple and proof of concept feeling, not overdesigned. Look up current docs online if you need to.