diagnosing-bugs

For non-deterministic bugs, raise the reproduction rate instead of seeking a clean repro: loop the trigger 100 times, parallelize, add stress, and narrow timing windows. A 50% flake rate is debuggable; keep increasing the rate until the failure is reliable enough to instrument.

Build a feedback loop that goes red on the exact symptom: a failing test, curl script against a dev server, CLI invocation with a fixture, or a headless browser script. The loop must be deterministic, fast, and assert the user's specific failure, not just the absence of crashes.

Stop and say so explicitly rather than hypothesizing without a signal. 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. Logs are usually the wrong tool for perf work; measure first and fix second.

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 architectural gap instead of writing a shallow test that gives false confidence.