Input
- A bug report, failing command, test, error, performance complaint, repro steps, or affected area.
- Use explicit input first; otherwise infer from context, recent edits, selected files, or branch.
- Safest default: find an agent-runnable repro before patching.
Workflow
- Build a feedback loop. Reproduce the issue with a runnable command, test, script, request, or browser flow.
- Confirm the symptom. Verify the repro matches the user's reported behavior, not a nearby failure.
- Minimize the repro. Reduce inputs, state, or steps until the failing behavior is clear.
- Form hypotheses. List 2-4 falsifiable causes and test one at a time.
- Instrument narrowly. Add targeted logs, assertions, debugger probes, or measurements only where they distinguish hypotheses.
- Fix the root cause. Keep the change scoped to the proven cause.
- Add regression coverage. Add or update a test when there is a suitable behavior-facing seam.
- Clean up. Remove temporary probes, scripts, and logs.
Output
- Repro used
- Root cause
- Fix applied
- Regression coverage added or skipped
- Validation run
Guardrails
- Do not patch before reproducing unless reproduction is impossible.
- If no repro is possible, state what was tried and what artifact or access is needed.
- Do not leave temporary debug logs, scripts, or instrumentation behind.
- For performance issues, measure before and after.