Claude Code identifies 55 real bugs in Ceph via two-week fault-injection loop

Details

Claude Code ran an autonomous fault-injection and fix loop on the Ceph distributed storage system for two weeks, identifying 55 real defects.
32 bugs were found in code already on main, with the oldest dating back to 2020, while 16 additional bugs were discovered and fixed in the fault-injection tool itself.
The process cost approximately $840 in API fees for 87 hours of runtime, though human review remained the primary bottleneck for commit quality.
The author's rough estimate is that human time per bug decreased from about two days to a couple of hours, with over 90% of the automated fixes deemed sound.
The workflow relied on unit test failures to validate bugs, background commands for periodic check-ins, and subagents for diagnosis and review.
Operational issues included a detached wake-up command causing a six-hour delay in detecting a failure and disk space exhaustion due to incorrect process termination logic.

Key point