coral: Skill for Claude Code

Review Gate

Run project review agents by tier taxonomy, consolidate findings, issue a verdict.

<Review_Protocol> 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. 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), and add coral:architect to the list of agents with tier 1 by default.
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]."

**Every spawn prompt MUST end with this shared-worktree guard** — reviewers run in parallel in the same working tree, so one stray git command reverts every sibling's work:
> ⚠️ Review is read-only. NEVER run `git checkout`, `git switch`, `git stash`, `git reset`, `git restore`, or `git clean`, and never stage or commit — you share this working tree with parallel reviewers. To inspect another revision, use `git diff <ref>`, `git show <ref>:<path>`, or `git log <ref>` — never check it out.

**Default**: spawn ALL INVOKE agents in parallel, wait for all.

**With `--gated`** (cost-saving short-circuit — use when API budget is tight):
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 |

| 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} |