Tsurai7 Staff review
Intent first, quality second
Before judging the code, check did they build what was requested — nothing
more, nothing less? Read the stated intent (PR description, commits, plan
file) and compare against the diff. Open with Scope: CLEAN / DRIFT / MISSING — Intent: <1 line> — Delivered: <1 line>, listing out-of-scope changes and
unaddressed requirements. Informational — it frames the review, not blocks it.
Look beyond the diff
A new enum value, status, tier, or constant is the one class of change a within-diff review cannot judge: grep for the files handling sibling values and read them — the bug lives where the new value is not handled.
Confidence calibration
Every finding carries a confidence score, and the score decides its fate:
| Confidence | Meaning | Fate |
|---|---|---|
| 9–10 | Verified by reading the specific code | Report |
| 7–8 | Strong pattern match | Report |
| 5–6 | Could be a false positive | Report with an explicit caveat |
| 3–4 | Suspicious but unproven | Appendix only |
| 1–2 | Speculation | Only if severity would be critical |
Finding format: [SEVERITY] (confidence: N/10) file:line — problem — suggested fix.
The verification gate (before emitting)
A finding enters the report only if you can quote the motivating line(s) — the verbatim code that makes it a bug. Can't quote it? Demote to confidence ≤5; never inflate to dodge the gate. For framework-generated symbols (ORM columns, decorators, migrations), quote the construct that generates them — a failed grep for the literal name proves nothing.
Fix-first, then adversarial pass
Every finding gets an action: mechanical and safe → fix it; behavioral or risky
→ ask, showing the proposed fix. Before the verdict, adversarially try to
refute each of your own findings — and remember risk is not proportional to
diff size: a 5-line auth change can be critical. Deliver the verdict per the
answer-shapes code-review shape.
Sibling sweep
For any new unit that has peers (worker, client, adapter, handler): name the closest sibling, diff the new unit against it dimension by dimension — transport and timeouts, error handling and messages, metrics and timestamps, input/output mappings, logging — and justify every divergence. "Every other X does Y" findings live here; they are invisible within the diff.
Critic pass (high stakes)
When findings gate a merge, or the change touches money, PII, or published contracts, dispatch an independent critic — fresh context, not the authoring session — with a mandatory verdict per finding: AGREE, DISAGREE with a code citation, or CONCERN without one. The artifact stays frozen during the exchange. Findings survive only with evidence; consensus without citations is not evidence — naive multi-agent agreement measurably underperforms a single reviewer.
Final-state gate
Review-fix churn invalidates earlier reviews: fixes introduce new defects and redesigns re-route flows. Before human review, the final state gets one whole-diff pass in a fresh context (ideally a different model lineage than the author's), at high effort, reading only the spec, the invariant catalog, and the code — not the author's reasoning. Any non-trivial fix after that pass re-reviews the delta.
Lens dispatch
This skill is the code/platform lens. A full review of a substantial change runs the other lenses as parallel fresh contexts, each given only its own sources and the diff — never the author's reasoning, never another lens's output: spec fidelity (call the Skill tool with "spec-fidelity-review"), test adequacy (call the Skill tool with "test-adequacy-review"), and standing invariants across execution paths (call the Skill tool with "invariant-audit"). Report per lens and never merge or rerank findings across lenses — one lens's silence says nothing about another, and a single ranked list is how one axis masks the rest. All lenses' findings land in the findings-ledger. Distinct lenses find near-disjoint defect sets (perspective-based reading); adding a lens that overlaps an existing one adds cost, not coverage.
Provenance
Distilled from the gstack /review skill by Garry Tan
(garrytan/gstack, MIT) — method only,
none of the runtime (specialist dispatch, Greptile/Codex, learnings database).
Sibling sweep, critic pass, final-state gate, and lens dispatch added from an
internal production retrospective (2026-08) and 2025–26 review-effectiveness
research.