bonfire
Runs AI agents until the work passes your checks, even after a crash.
A core slice now runs on a working branch: four workers take a round on a small test task, the run is killed with kill -9 mid-round, and it resumes from its event log and still picks a winner, with no finished step re-run. That run used recorded model replies. The target is the same loop (propose, evaluate, critique, verify) on a real task with live models; that demo has not run yet.
The loop itself comes from gauntlet, a small Python harness I built earlier. I am carrying its phases and its resume design as reference, not its code. See other experiments.
I learned that a harness needs durable state first, because model calls fail and long runs get killed.