tier-review

分类: 工具与效率 | 上传者: kangig94 | 下载: 0 | 版本: v1.0（最新）
实施后使用来运行分层排序的审查代理并生成综合结论。
更新日志: Source: GitHub https://github.com/kangig94/coral
目录结构

当前层级: tree/main/.claude/skills/tier-review/
📄 SKILL.md 3.9 KB
SKILL.md

---
name: tier-review
description: "Use after implementation to run tier-ordered review agents and produce a consolidated verdict."
argument-hint: "[scope description]"
---

# Review Gate

Run project review agents in tier order, consolidate findings, issue a verdict.

<Review_Protocol>
  <Role>
    You are the review gate executor. Discover which project agents exist, match them
    to the review scope, spawn by tier, and consolidate into a single verdict.
  </Role>
  <Protocol>
    All paths below are relative to the **user's project root** (working directory), not the plugin.

    ## Phase 1 — Discover Agents

    1. Read the project's `.claude/rules/agents.md` for the Quick Reference table (agent, tier, purpose)
       and the Consultation Matrix (task category → agent mapping)
    2. If agents.md not found, read `.claude/agents/*.md` — look for tier in description
       or Situation table. If tier is unclear, treat safety-related agents as tier 1,
       domain-specific as tier 2, quality reviewers as tier 3.
    3. If no agents found at all → report "no review agents configured" and exit.

    ## Phase 2 — Plan

    1. Determine review scope from conversation context:
       - Explicit argument → use it
       - Recent implementation in conversation → those files
       - Neither → fall back to git diff (staged + unstaged)
    2. For each agent, check against the scope:
       - Consultation Matrix: is this task category → agent mapping MANDATORY?
       - Agent's Situation/Priority table: does the scope match?
    3. Build an invocation table — for each agent: INVOKE (with scope + focus) or SKIP (with reason)
       - If all agents are SKIP → report "no agents relevant to this scope" and exit.
    4. Report the table before executing.

    ## Phase 3 — Execute

    Spawn only agents marked INVOKE in the table.
    For each agent, pass: "Review [scope files] focusing on [focus from plan]."

    **Fast path** (total INVOKE agents < 5): spawn ALL agents in parallel, wait for all.
    **Standard path** (5+ INVOKE agents): tier-ordered execution:
    1. Tier 1 (safety) — spawn in parallel, wait for all to complete
       - If ANY returns BLOCKING findings → output REJECT verdict, STOP here
    2. Tier 2 (domain) + tier 3 (quality) — spawn in parallel, wait for all to complete

    ## Phase 4 — Consolidate

    Apply merge rules to all agent results:
    1. **Verdict mapping**: agent PASS → no BLOCKING; agent NEEDS WORK → STRONG findings
    2. **Dedup**: same file:line from multiple agents → single entry, list all agents
    3. **Severity**: agents disagree → use higher severity
    4. **Convergent signals**: same file flagged by multiple agents → elevate priority
    5. **Root cause**: tier 1 finding that explains tier 3 symptom → connect and elevate

    Issue final verdict:
    | Condition | Verdict |
    |-----------|---------|
    | Any BLOCKING finding | REJECT |
    | Multiple unresolved STRONG findings | NEEDS WORK |
    | No BLOCKING, 1-2 STRONG items remaining | APPROVED WITH CONDITIONS |
    | No BLOCKING, no unresolved STRONG | APPROVED |
    | No BLOCKING, no STRONG, all agents report high quality | EXCEPTIONAL |
  </Protocol>
  <Output_Format>
    ## Review: [scope]

    | Tier | Agent | Status | Verdict | Key Findings |
    |------|-------|--------|---------|--------------|
    | {1/2/3} | {agent} | INVOKED/SKIPPED | PASS/FAIL/- | {summary or skip reason} |

    ### Strengths
    - {Positive observations from agents with file:line evidence}

    ### Consolidated Findings
    | # | Severity | Agent(s) | Location | Finding | Suggestion |
    |---|----------|----------|----------|---------|------------|
    | 1 | BLOCKING/STRONG/MINOR | {sources} | file:line | {issue} | {fix} |

    ### Verdict: [EXCEPTIONAL / APPROVED / APPROVED WITH CONDITIONS / NEEDS WORK / REJECT]
    {justification}
  </Output_Format>
</Review_Protocol>
登录后下载/点赞/收藏 ❤ 5 | ★ 0
tier-review

目录结构

SKILL.md

举报内容

提示
