diagnosing-bugs

For non-deterministic bugs, raise the reproduction rate instead of seeking a clean repro: loop the trigger 100 times, parallelise, add stress, and narrow timing windows. A 50%-flake bug is debuggable; keep increasing the rate until the feedback loop reliably goes red.

Build a tight feedback loop first: a failing test, curl script, CLI invocation, or headless browser script that asserts the user's exact symptom. The loop must be deterministic, fast, and runnable unattended before any hypothesis is tested.

Stop and say so explicitly rather than guessing. List what you tried, then ask the user for environment access, a captured artifact like a HAR file or log dump, or permission to add temporary production instrumentation.

For performance regressions, skip log-based probing and measure first: establish a baseline with a timing harness, profiler, or query plan, then bisect between known-good and known-bad states. Measure first, fix second.

Write the regression test before the fix, but only if a correct seam exists where the test exercises the real bug pattern as it occurs at the call site. If no correct seam exists, document that architectural gap as a finding itself.

Single-hypothesis generation anchors on the first plausible idea and wastes effort. Generate 3-5 ranked, falsifiable hypotheses with explicit predictions, show them to the user for re-ranking, then test one variable at a time.