RaceLab

RaceLab - AI proposes the repair. Execution checks the fix. Try the live MVP: https://racelab.tangvu.dev/ No installation, sign-in or visitor API key required. Background jobs can pass normal tests yet become permanently stuck when a worker crashes after claiming work. An AI-generated patch may look convincing while introducing another failure, such as stealing a live worker's job. RaceLab makes these repairs testable. Developers paste or import a supported job-claim method, and Muse Spark 1.3 investigates the failure and proposes replacement source. An independent runner executes the candidate against ten real SQLite/process conditions. Developers can inspect original versus repaired code, process and SQL traces, and actual versus expected results. The demo reproduces a failure in a pinned TypeScript method from the public TEND repository: terminate the worker after its claim commits, restart another process on the same database, and observe the stranded job. A Muse Spark repair passes all ten declared recovery conditions. The MVP supports new model investigations, browser-saved projects, private run recovery for up to 24 hours, and regression export/import. Reopening a regression executes its checks again instead of trusting saved PASS labels. I built RaceLab's workbench, execution adapters and repair workflow during the hackathon, reusing transport/accounting infrastructure from my earlier event prototype. TEND is an existing source input. Verification covers the declared methods and finite conditions, not arbitrary repositories or whole-application correctness.