diagnosing-bugs

Non-deterministic bugs require raising the reproduction rate rather than finding a clean repro. Loop the trigger 100 times, parallelize, add stress, narrow timing windows, or inject sleeps until the failure rate is high enough to debug against.

Build a feedback loop that goes red on the specific symptom: a failing test, a curl script against a dev server, a CLI invocation with fixture input, or a headless browser script. Then minimize it by removing inputs and steps one at a time until every remaining element is load-bearing.

Stop and say so explicitly rather than guessing. Ask the user for environment access, a redacted captured artifact such as a HAR file or log dump, or permission to add temporary production instrumentation.

For performance regressions, establish a baseline measurement first using a timing harness, profiler, or query plan, then bisect between known states. Measure first and fix second; scattered logging is usually the wrong tool for perf issues.

Write the regression test before the fix, but only at a correct seam where the test exercises the real bug pattern as it occurs at the call site. If no correct seam exists, document that as an architectural finding instead of writing a shallow test.