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/cwijayasundara/claude_harness_eng_v4 --skill review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。
Discover and map an existing codebase before planning or changing it.
Change the behavior of existing code — story-driven by default, or --issue N for a GitHub bug fix. Test-first, full verification, code review.
Use when a planned change touches persisted data shape — ORM models, migration files, schema definitions, serialized formats, or message contracts — in /change, /refactor, or /implement on an existing codebase. Routes schema changes through expand-contract and proves reversibility before any deploy.
基于 SOC 职业分类
正在显示 SKILL.md
| name | review |
| description | Run evaluator and security reviewer concurrently for comprehensive quality gate. |
| argument-hint | [story-id] |
| context | fork |
On-demand, pre-merge entry point to the harness's one quality gate: it spawns the same evaluator + security-reviewer agents that /evaluate (Layer 4) and /auto (Gate 7) run in-pipeline, with identical verdict semantics. This skill owns only the on-demand orchestration; the gate's definitions live in /evaluate.
Ultracode tip: Multi-dimension review with adversarial verification is a natural fan-out, so
/effort ultracodepays off on this plain skill form.
/review # reviews the current group in context
/review E3-S1 # reviews a specific story and its group
Use the Agent tool to spawn both agents in a single call — concurrent execution is the point of this skill:
specs/reviews/evaluator-report.md and updates features.json.specs/reviews/security-review.md and the canonical specs/reviews/security-verdict.json.Both agents get the same changed-file set and group context. Every changed file in the group is in scope for both — never pass a subset to avoid findings.
Severity levels (BLOCK/WARN/INFO), the BLOCK self-healing loop (generator fix → full re-run, max 3 cycles, then escalate), and the security verdict format are defined once in /evaluate (.claude/skills/evaluate/SKILL.md) — follow them exactly from there. Do not merge or mark a group complete while any BLOCK finding remains open, and always re-run the full review after fixes.
specs/reviews/evaluator-report.md — PASS/FAIL with per-check detailspecs/reviews/security-review.md — BLOCK/WARN/INFO findings with file referencesspecs/reviews/security-verdict.json — the canonical machine-readable security verdictAll three must exist before the review is complete; a missing output is itself a BLOCK finding.
/evaluate)There are two entry points to the same gate, not two gates:
/evaluate Layer 4 / /auto Gate 7 — the authoritative in-pipeline owner./review — the on-demand pre-merge entry point (this skill).In /auto, Gate 7 already covers it — a separate /review is only needed for manual gating before a merge.