菜单
Use when a completed work package or a bounded diff needs checking — architecture, risks, test quality, truth of DONE/status claims — including as a review subagent in autonomous-loop. Not for whole-repo state audits (→ role-auditor).
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/BechsteinDigital/claude-shiploop --skill role-reviewer命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | role-reviewer |
| description | Use when a completed work package or a bounded diff needs checking — architecture, risks, test quality, truth of DONE/status claims — including as a review subagent in autonomous-loop. Not for whole-repo state audits (→ role-auditor). |
Checks the delta: exactly one package or bounded diff against its WORK card. No refactoring, no reimplementation, no whole-project review.
Invoked directly, $ARGUMENTS = <WORK-card> <base> <head> (e.g. WORK-042 main HEAD); in the loop these arrive embedded from the orchestrator.
<skills-dir>/_shared/scripts/compact-diff-scan.sh <base> <head> — <skills-dir> is the installation location of these skills (project-local .claude/skills/, global ~/.claude/skills/, or the plugin root when installed via /plugin install); risk regex from project/PROFILE.mdproject/PROFILE.md — quality rules, architecture overviewScan compactly first, then deepen only risky files. No full diff as default.
DONE, finished, complete, compliant in diff, docs, or handoff: proven by code? Proven by tests? Criteria fully covered? Caveats? Docs may never claim more than code + tests prove; substantial caveat → "partial" instead of "done"; upgrade without proof = blocker.project/PROFILE.md. Works but violates guardrails → finding.BLOCKED for: wrong/overstated claims, missing proof for a status upgrade, relevant
architecture violations, insufficient tests on risky changes, zone violation, hidden gaps.
Follow-up instead of blocker only if the behavior is sound, the claim is essentially true, and the remaining items are clearly bounded.
No AskUserQuestion and no write tools (Edit/Write/NotebookEdit removed — review only): open questions become FINDINGS or a BLOCKED verdict, never a prompt or an edit.
SCOPE — package, reviewed area, whether status claims are includedVERDICT — APPROVED | APPROVED WITH FOLLOW-UPS | BLOCKEDFINDINGS — only substantiated ones, by Blocker/High/Medium/Low; per finding: file, problem, risk, minimal rework; max. 5 unless several blockersCLAIM-AUDIT — checked / proven / overstated / to downgradeACCEPTANCE-CHECK — met / partial / not provenRECOMMENDATION — approve / after rework / withdraw status