Michael Hood

Doc is a voice-driven recovery tool for AI coding agents. Agents like Muse Code change many files per turn and sometimes break things before anyone commits; git can't tell you which turn did it, and reverting loses the good work around it. Doc closes that gap with one loop: - Find it: "Hey Doc, which turn broke the tests?" Doc bisects recorded agent turns by testing past snapshots (never your working files) and names the bad turn, with a plain-English explanation. - Fix it: Muse Wonder plans three different fixes, three Muse Code agents try them in parallel git worktrees, Doc tests each, and Wonder ranks them. Fixes that edit the tests can't win. You pick; nothing is applied for you. - Remember it: undo just the bad turn (by voice if you like) and Doc turns the mistake into an approved lesson that's handed to the next agent before it writes a line. Safety is built in: the AI never writes your files. Every restore goes through dmc's preview-first, journaled engine and your explicit yes. It runs as a CLI and a full terminal UI with push-to-talk. Pre-existing work: Doc drives dmc, my open-source rollback tool from before the event (separate repo, pinned commit; the event's changes to it are included as patch files). Everything in this repository was built during the hackathon.