Given a repo + a specific commit, it iteratively searches for a plausible high-level spec prompt that could have produced that change. It runs agent loops, scores candidates for technical similarity and "realism" (does this look like a prompt a human would actually write), and outputs the best spec.
Inspiration: I saw Mitchell Hashimoto mention experimenting with agents to reproduce manual code edits, and around the same time GitHub released the Copilot SDK.
Repo: github.com/igolaizola/retrospec