Root-cause debug

Reproduce, isolate, test, then verify.

Steps

Record expected and observed behavior, exact reproduction, environment, and error evidence. If it is intermittent, gather evidence instead of choosing a fix by intuition.
Inspect recent changes and compare a working case. Trace inputs and outputs across component boundaries, logging only non-secret data needed for the hypothesis.
State one candidate cause and the observation that would disprove it. Make the smallest safe test that separates it from alternatives; change one variable at a time.
Once supported, add a regression case and make the narrow fix within the user's approved scope. Avoid bundling unrelated refactors. If tools are absent, provide a testable plan rather than claiming execution.
Run the reproduction and relevant regression checks. Report commands, actual outcomes, and what remains untested. Repeated failed hypotheses call for revisiting assumptions, not stacking patches.