diagnosing-bugs

Build a feedback loop that goes red on the exact symptom: replay a captured trace, write a curl script against a dev server, or create a throwaway harness exercising the bug code path. If no loop is possible, request a redacted HAR file, log dump, or permission for temporary production instrumentation.

Raise the reproduction rate instead of chasing a clean repro: loop the trigger 100 times, parallelize, add stress, narrow timing windows, and inject sleeps. A 50%-flake bug is debuggable; keep raising the rate until the loop reliably goes red.

Establish a baseline measurement first using a timing harness, performance.now(), a profiler, or a query plan, then bisect between known states. Logs are usually wrong for perf work; measure first and fix second.

Writing the test first proves it actually catches the bug by watching it fail, then pass after the fix. Only do this at a correct seam that exercises the real bug pattern; if no correct seam exists, that architectural gap is itself the finding.

Use it as a last resort when a human must physically interact with the UI. The hitl-loop.template.sh script structures the session with step prompts and captured answers, so the agent still receives parseable output while the human performs clicks or sign-ins.