debugkit

Provide the expected behavior, actual behavior, and where you observed it. The agent reproduces the failure, shrinks it to a minimal case, tests falsifiable hypotheses, and proves the cause by toggling it on and off before reporting.

Force determinism first by pinning seeds, serializing concurrency, and freezing clocks. If the bug stays non-deterministic, the proof gate uses a statistical form: declare a run count N upfront, then compare failure rates with and without the suspected cause.

It never applies fixes. It may edit files temporarily as probes, but every probe is recorded as a patch in a ledger and reverse-applied afterward, so your uncommitted work is preserved. The output is a described fix, not an applied one.

Yes, because commit bisects run in a throwaway git worktree created with git worktree add, never in your working tree. The scratch worktree is removed before the run ends, leaving your branch and uncommitted edits untouched.

The run ends in an instrumentation plan rather than a guessed diagnosis. You receive ranked hypotheses, each paired with the specific experiment or log line that would discriminate it, and you rerun the diagnosis once that evidence arrives.

Skip it for code reviews of unexecuted diffs, feature requests for behavior never built, test planning before a bug exists, and general optimization without a regression. Those have no observable symptom to reproduce and toggle.