Recovery Room

Recovery Room repairs a failed, retried import with an AI agent whose every fix is tested on a copy and checked before anyone applies it. The problem: when an import times out and is retried, the retry can write the same orders twice. Genuine purchases that arrive afterwards can look almost identical to the duplicates, so cleaning up by similarity deletes real orders too. How it works: a Meta Muse Spark 1.3 agent investigates with five tools and no write access to the source database: read the schema, list the evidence, read evidence with line numbers, run read-only SQL, and test a repair on a disposable copy. Every cancellation must cite the exact log lines that prove the order is a repeat. The runtime rejects wrong or invented citations and tells the agent why, and the agent corrects itself. Each proposed repair must pass ten independent checks on the copy. What a person receives is a guarded SQL script to review: it refuses to run if the data has changed, and the original database is never touched. When the evidence is missing, the agent stops and names the gap. Results: 16 recorded Muse Spark runs on six synthetic cases. 9 repairs passed every check, 4 stopped because evidence was missing, 2 hit errors and 1 did not finish. In 6 of the 9 repairs, the runtime rejected a first attempt and the agent corrected it. A hand-written script also solves these cases; the point is repairs you can verify. Interactive recorded demo. (replays real runs, no key needed): https://gyslainm.github.io/recovery-room/ Github repo: https://github.com/GyslainM/recovery-room Youtube Demo video: https://youtu.be/NQRHuGjj458