Review Scope Guard

Overview

A scope-aware triage skill that sits between a review tool and a user-facing summary. It takes a list of review findings, a Definition of Done (DoD), and optional explicit project context, classifies each finding into one of four action categories, and maintains a rejected-findings ledger so the same complaint is never re-litigated across cycles. The skill never applies fixes itself — it only decides which findings are worth escalating and which should be suppressed.

This skill exists because adversarial review tools (including codex's adversarial-review) are calibrated for "correctness gaps from a theoretical ideal", not "impact on the stated scope". Without a scope filter the implementer chases edge cases and semantic implementations that were never in scope, then reverts them. A 19-cycle curl-import session that reverted ~50% of its Phase 2-3 additions is the empirical baseline this skill is designed to prevent.

Language

All user-facing output is rendered in the user's language (the language the user has been using in the conversation, or as configured in the Claude Code system-level language setting). This section is the authoritative translation contract — any per-language sample reference (e.g. references/output-samples.ja.md) is illustrative only and MUST NOT contradict these rules.

Translate into the user's language:

• Section headings and column labels (target-language equivalents of Category / Rationale / Action)
• Free-text fields Claude authors: rationale body, recommended_action values, stop-signal evidence prose, next-action hints, degraded-mode warnings
• AskUserQuestion question, header, and option label / description fields (e.g. during DoD interview)

Keep verbatim (do NOT translate), regardless of user language: