TraceFix turns a production traceback into a failing regression test that reproduces the crash, the commit that caused it, and a fix that keeps the whole suite green. IBM Bob writes the code; TraceFix checks every step with pytest before accepting it.
Three stages need no AI at all. The two that do are wrapped in hard, machine-checked gates, so a wrong answer from Bob gets rejected and sent back, never merged.
Every frame of the traceback mapped to a tracked file, with code context.
git blame around the crash lines, ranked, to find the suspect commit.
Accepted only if it fails with the same exception in the same function as production.
Accepted only if the repro passes and the full existing suite stays green.
Two commits on an incident branch plus an HTML/Markdown report ready to paste into a PR.
Real output from the coupons incident, with two Bob agents racing:
Getting from a stack trace to a reproduction is usually the slowest part of fixing a production bug. TraceFix does it in seconds, with code context and the suspect commit already attached.
Every incident leaves behind a regression test that reproduces the real failure, so the bug can't come back quietly.
Nothing Bob writes is accepted on its word. Wrong exception, wrong function, a touched file it shouldn't touch, or a red suite: the change is rejected and the evidence goes back to Bob.
Recorded from real TraceFix runs against the shopcart demo service. Replays play back the actual event timings.