farifran/aegis-harness — reverse-engineered prompt

Build me a small shell based AI harness called Aegis that runs language model tasks in a controlled and repeatable way. I want it to feel like a safety wrapper around an AI coding assistant, where the runtime is in charge, not the model.

It should support separate modes like discovery, forensics, validation, adversarial review, repair, and optimize. The read only modes should only inspect evidence the runtime gives them. The repair and optimize modes can make changes, but only through clearly allowed surfaces.

Please include capability handlers for reading files, listing the tree, searching symbols, checking git status and diffs, reading runtime handover notes, and reading a simple architecture graph. The runtime should create temporary execution areas, inject capability evidence, call an OpenAI compatible chat endpoint, require machine readable JSON output, validate it, save any unresolved attention into an explicit handover file, and clean up after itself.

Keep it simple, mostly bash, with clear setup notes and examples for running the full harness or one mode.