| name | investigate |
| description | Multi-source root-cause analysis for a specific bug or failure. Depth-first on one symptom. Use when something is broken and you need the actual cause, not a workaround. |
/investigate — root-cause analysis
Depth-first on one symptom. The investigator's job is to find the actual cause, not patch the symptom.
Process
- Reproduce the symptom (run the failing test, hit the failing endpoint, replay the failing input). If you can't reproduce, that's the first finding.
- Inventory all code paths that touch the symptom.
- Bisect via git, logs, or feature flags to localize the regression.
- Hypothesize a cause; design a minimal test that would prove or disprove the hypothesis; run it.
- Iterate until the hypothesis holds and the proof is mechanical.
- Output: cause statement + evidence path + suggested fix shape (do NOT auto-apply; pass to /goal or /opussonnet for execution).
When to use
- A specific test is failing or flaky.
- A specific endpoint or page is broken.
- A specific user-reported bug needs a definitive cause.
Distinct from
/audit — breadth-first across many holes. Investigate is depth-first on one./debug-hard task class — the router escalates debug-hard to Opus automatically; /investigate is the user-invokable skill.
