| name | rpi-diagnose |
| description | Diagnose a bug iteratively: reproduce, find root cause, attempt up to 3 fixes, produce a diagnosis artifact. Use when user says 'why is X not working?', 'X is broken', 'after Y, I see Z fail', or reports broken behavior, even if they don't say 'diagnose'. Do NOT invoke for understanding working code (use rpi-research). |
Diagnose Bug
Goal
Iteratively diagnose and fix complex bugs through root-cause analysis. This combines research's interview flow with implement's execution loop, adding an autonomous iterate-until-fixed cycle.
Accept a bug description, error message, or path to a failing test. Clarify expected vs actual behavior, reproduce, trace root cause, then attempt fixes autonomously (up to 3 attempts). Produce a diagnosis artifact in .rpi/diagnoses/ regardless of outcome.
When the bug is fixed, announce completion and suggest → /rpi:rpi-commit to commit the fix and diagnosis artifact. When escalation is needed, suggest → /rpi:rpi-plan (or /rpi:rpi-propose for architectural issues).
Invariants
- Always interview before investigating — ask 1-2 clarifying questions to establish expected vs actual behavior
- Before opening a new diagnosis, search for prior diagnoses on the same topic; ask whether to extend an existing one if scope overlaps
- Reproduce the bug before investigating root cause — run the failing test, trigger the error, or confirm the symptom is observable; if unreproducible, checkpoint with user
- Trace the code path from symptom to root cause — all findings must include file:line references
- Do not propose a fix until root cause is identified
- Fix-iterate loop: attempt up to 3 fix iterations autonomously; each iteration: apply fix, run relevant tests, evaluate results
- Revert on failure: fully revert each failed fix attempt before trying the next — the codebase must never be left in a broken state between attempts
- Checkpoint after 3 failures: stop and present what was tried, why each attempt failed, and current understanding of root cause; ask the user how to proceed (more attempts, different approach, or escalate)
- Suggest commit on success: when a fix works and tests pass, do not auto-commit — suggest the user run
/rpi:rpi-commit to review and commit both the fix and the diagnosis artifact together
- Diagnosis artifact: always produce an artifact in
.rpi/diagnoses/ containing bug report (expected/actual/reproduction), root cause with file:line references, investigation log (each attempt with hypothesis/change/result), and resolution status
- Escalation: if the fix is too complex for inline patching, write the diagnosis artifact with resolution status "escalated" and suggest
/rpi:rpi-plan; if the fix requires architectural changes, suggest /rpi:rpi-propose instead
- Do not auto-push or auto-merge changes
Principles
- Bugs are spec deviations — frame diagnosis as closing the gap between expected and actual behavior
- Escalate, don't force — a clear diagnosis artifact handed off to
/rpi:rpi-plan is more valuable than a fragile hack
- Right-size iteration — simple bugs may be fixed on the first attempt; don't over-investigate when the fix is obvious
Recommended model: premium tier, high effort — iterative root-cause needs deep reasoning. Advisory; see docs/model-routing.md.
