diagnosing-bugs

For non-deterministic bugs, raise the reproduction rate instead of chasing a clean repro: loop the trigger 100 times, parallelize, add stress, narrow timing windows, or inject sleeps. A 50%-flake bug is debuggable; a 1% flake is not.

Build one agent-runnable command that drives the actual bug code path and asserts the user's exact symptom, such as a failing test, curl script, CLI fixture diff, or Playwright browser script. Confirm it goes red, then minimize inputs until every remaining element is load-bearing.

Yes, a headless browser script with Playwright or Puppeteer is one of the recommended feedback loops. It drives the UI and asserts on DOM, console, or network behavior, giving an automated red/green signal for front-end bugs.

Writing the regression test first proves the test actually catches the bug by watching it fail before the fix and pass after. It only applies when a correct seam exists where the test exercises the real bug pattern as it occurs at the call site.

Stop and say so explicitly rather than hypothesizing without a loop. 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, avoid log-based probing and instead establish a baseline measurement with a timing harness, performance.now(), a profiler, or a query plan, then bisect. Measure first, fix second.