메뉴
Multi-angle code and plan review with security, scope, and UI profiles. Use when reviewing code, plans, PRs, or running pre-merge quality gates.
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/tim-hub/powerball-harness --skill harness-review명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
SOC 직업 분류 기준
SKILL.md 표시 중
| name | harness-review |
| description | Multi-angle code and plan review with security, scope, and UI profiles. Use when reviewing code, plans, PRs, or running pre-merge quality gates. |
| when_to_use | review code, review plan, review PR, security audit, pre-merge check, scope analysis, quality gate |
| allowed-tools | ["Read","Grep","Glob","Bash","Task","AskUserQuestion"] |
| argument-hint | [code|plan|scope|--quick|--codex-closeout|--dual|--team-debate|--security|--ui-rubric] |
| context | fork |
| effort | xhigh |
| model | opus |
| agent | reviewer |
| User Input | Subcommand | Behavior |
|---|---|---|
| "Review this" / "review" | code (auto) | Code review (recent changes) |
"After harness-plan" | plan (auto) | Plan review |
| "Check scope" | scope (auto) | Scope analysis |
harness-review code | code | Force code review |
harness-review plan | plan | Force plan review |
harness-review scope | scope | Force scope analysis |
harness-review --dual | code (auto) + Codex parallel | Claude + Codex dual review |
harness-review --security | Security Review | OWASP Top 10 dedicated security review (read-only) |
harness-review --ui-rubric | UI Rubric Review | 4-axis design quality scoring |
Maps flags to their reference files. Quick Reference above covers trigger → subcommand routing; this table covers flag → reference dispatch.
| Flag | Reference | Purpose |
|---|---|---|
--quick | ${CLAUDE_SKILL_DIR}/references/codex-closeout.md | Lightweight Codex-assisted closeout; stops on clean result |
--codex-closeout | ${CLAUDE_SKILL_DIR}/references/codex-closeout.md | Full closeout with final JSON report |
--team-debate |
${CLAUDE_SKILL_DIR}/references/team-debate.md |
| Multi-agent read-only debate (Spec / Plans / Regression / Skeptic agents) |
--opencode | ${CLAUDE_SKILL_DIR}/references/opencode-review.md | OpenCode-assisted review via opencode-plugin-cc |
| (any mode) | ${CLAUDE_SKILL_DIR}/references/governance.md | APPROVE pass criteria, severity table, AskUserQuestion contract |
| (any mode) | ${CLAUDE_SKILL_DIR}/references/code-review.md | Eight-lens review flow and verdict rules |
--team-debate is required when changes span multiple modules, touch security/auth/release, or when the same issue fails two consecutive post-fix re-reviews (see governance.md).
| Option | Default | Description |
|---|---|---|
--dual | none | Parallel Claude + Codex review. Details: ${CLAUDE_SKILL_DIR}/references/dual-review.md |
--security | none | OWASP Top 10 security-only review (read-only). Details: ${CLAUDE_SKILL_DIR}/references/security-profile.md |
--ui-rubric | none | 4-axis design quality scoring (0–10). Details: ${CLAUDE_SKILL_DIR}/references/ui-rubric.md |
--no-commit | none | Disable auto-commit on APPROVE |
| Recent Activity | Review Type | Perspectives |
|---|---|---|
After harness-work | Code Review | Security, Performance, Quality, Accessibility, AI Residuals |
After harness-plan | Plan Review | Clarity, Feasibility, Dependencies, Acceptance |
| After task addition | Scope Review | Scope-creep, Priority, Feasibility, Impact |
Establish severity classification before reviewing. See ${CLAUDE_SKILL_DIR}/references/verdict-framework.md for the full severity matrix and AI Residuals classification.
Rule: If critical or major findings exist → REQUEST_CHANGES. If only minor/recommendation → APPROVE.
Fork context auto-start (
REVIEW_AUTOSTART): emit this as the very first output token in a forked session before any other processing.Fork-context prohibition list (5 forbidden failure modes):
- Waiting for user confirmation before starting
- Returning empty output when the diff is empty (output
{"verdict":"APPROVE","rationale":"no changes detected"})- Spawning a browser reviewer without checking
reviewer_profilein the sprint contract- Writing to
Plans.mdor anycc:*marker- Exiting without emitting a JSON verdict
Does the change include UI files (.tsx, .jsx, .vue, .css, .html)?
├─ No → Static reviewer (proceed to Step 1)
└─ Yes → Does the sprint-contract specify reviewer_profile: "browser"?
├─ Yes → Browser reviewer (launch browser-review-runner.sh)
└─ No → Is this a visual/design change?
├─ Yes → Browser reviewer recommended
└─ No → Static reviewer
Browser reviewer path: bash "${CLAUDE_SKILL_DIR}/../../scripts/browser-review-runner.sh" --contract "${CONTRACT_PATH}"
CHANGED_FILES="$(git diff --name-only --diff-filter=ACMR "${BASE_REF:-HEAD~1}")"
git diff ${BASE_REF:-HEAD~1} --stat
git diff ${BASE_REF:-HEAD~1} -- ${CHANGED_FILES}
AI_RESIDUALS_JSON="$(bash "${CLAUDE_SKILL_DIR}/../../scripts/review-ai-residuals.sh" --base-ref "${BASE_REF:-HEAD~1}")"
| Perspective | Check Items |
|---|---|
| Security | SQL injection, XSS, credential exposure, input validation |
| Performance | N+1 queries, unnecessary re-renders, memory leaks |
| Quality | Naming, single responsibility, test coverage, error handling |
| Accessibility | ARIA attributes, keyboard navigation, color contrast |
| AI Residuals | mockData, dummy, fake, localhost, TODO, FIXME, it.skip, hardcoded secrets |
Apply the severity framework from verdict-framework.md to each finding.
See ${CLAUDE_SKILL_DIR}/references/result-schema.md for the full review-result.v1 JSON schema.
For --dual flag: see ${CLAUDE_SKILL_DIR}/references/dual-review.md
For --security flag: see ${CLAUDE_SKILL_DIR}/references/security-profile.md
For --ui-rubric flag: see ${CLAUDE_SKILL_DIR}/references/ui-rubric.md
--no-commit)harness-work fix loop (up to 3 times), then re-review.| Situation | Action |
|---|---|
| Security vulnerability | Immediately REQUEST_CHANGES |
| Suspected test tampering | Warning + fix request |
| Force push attempt | Reject + suggest alternative |
Load
${CLAUDE_SKILL_DIR}/references/codex-review.mdonly whencommand -v codexsucceeds and the user explicitly requests Codex or duo review.
Load
${CLAUDE_SKILL_DIR}/references/opencode-review.mdonly whencommand -v opencodesucceeds and the user explicitly requests OpenCode review (e.g.,--opencode, "use opencode", "opencode review").
↑ Claude will summarize this result. Type a new prompt to redirect or press Enter to continue.
harness-work — Implement fixes after reviewharness-plan — Create and modify plansharness-release — Release after review passes