diagnose

Raise the reproduction rate instead of chasing a clean repro: loop the trigger 100 times, parallelise, add stress, narrow timing windows, or inject sleeps. A 50% flake is debuggable; a 1% flake is not, so keep increasing the rate before hypothesising.

Try in order: a failing test at the right seam, a CLI invocation diffed against a snapshot, a curl script against a dev server, a headless browser script, a replayed captured trace, a throwaway harness, a fuzz loop, or a git bisect harness. Make it fast, sharp, and deterministic.

Stop and say so explicitly, listing what you tried. Ask for environment access, a captured artifact such as a HAR file or log dump, or permission to add temporary production instrumentation. As a last resort, use the HITL bash template so a human triggers the bug while output is captured.

Establish a baseline measurement first using a timing harness, profiler, or query plan, then bisect against it. Measure first and fix second; never change code before you can quantify the regression.

Tagging every debug log with a unique prefix like [DEBUG-a4f2] makes cleanup a single grep at the end of the session. Untagged logs tend to survive into production; tagged logs are reliably removed.

Skip the test-first step when no correct seam exists, meaning no test can exercise the real bug pattern at the actual call site. A shallow test gives false confidence; instead document the missing seam and flag it for post-mortem follow-up.