Purpose
Find the smallest evidence-supported explanation for a failure without speculative edits or unrelated cleanup.
When it fires
- A test, build, deployment, API, UI flow, or runtime behavior fails.
- The failure is intermittent and needs boundary isolation.
- The user asks for diagnosis, root cause, or incident analysis.
Inputs
- Expected and observed behavior.
- Reproduction steps, logs, timestamps, environment, and relevant revision.
- Repository instructions and known constraints.
Workflow
- Capture the exact failure and determine whether it reproduces.
- Reduce it to the smallest safe reproduction while preserving the failing condition.
- Trace inputs, state transitions, and outputs across the suspected boundary.
- List competing hypotheses and the observation that would falsify each one.
- Run the cheapest discriminating checks first.
- State the root cause only when evidence rules out plausible alternatives.
- If a fix is authorized, apply the smallest causal change and run focused plus regression checks.
- Record residual uncertainty, monitoring, and rollback needs.
Output contract
Return:
symptomand reproducibility;evidencewith commands, locations, and relevant revisions;hypotheses_testedand falsification results;root_causeornot_yet_proven;impact_boundary;recommended_fixwithout implementation when diagnosis-only;verificationandresidual_risk.
Tools & MCP
Prefer repository search, focused tests, logs, and read-only platform inspection. Redact credentials and personal data. Do not alter production state to make a diagnosis easier.
Quality bar
- Observation, inference, and decision are clearly separated.
- “Could be” is not presented as root cause.
- The reproduction and verification are copyable.
- The proposed fix addresses the causal boundary rather than masking the symptom.
- Unrelated working-tree changes remain untouched.
Example
Input: “The preview works, but the production API returns 415.”
Good output: evidence showing the exact content-type mismatch at the request boundary, alternatives ruled out, a minimal header-validation fix, and focused plus production-preview checks.
Built on SIP — Starlight Intelligence Protocol Vertical: starlight-agent-skills · portable capability layer