菜单
Comprehensive code review workflow - parallel specialized reviews → synthesis (includes cross-model Codex adversarial pass by default)
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。
基于 SOC 职业分类
The Code reporting discipline AND the reporting-system operations command. Discipline half — the "one spine" rule (every reportable unit of work lands one Weekly Work Tracker row, Source-tagged), field mapping, Bridge auto-mirror tie-in. Operations half (proposal 07) — /reporting status (registry tail, trust metrics, drift, next scheduled runs, log locations), /reporting run <pipeline>, /reporting heal. Use when finishing reportable work, when asked "how should I report this" / "log this for the report", at session end to capture unreported work, or for pipeline ops — "/reporting", "reporting status", "is reporting healthy", "run the health check", "why is a report row missing/stale".
Generate the weekly FourthOS sponsor update package for Carly (VP) and Christian (CTO). Pulls live portfolio data from the FourthOS Notion cockpit, refreshes the Notion Update Package page, renders a progressive-disclosure HTML artifact set (Tier 1 briefing dashboard, Tier 2 teaching deep-dive), stages it UNLISTED to the ai-enablement-decks GitHub Pages site, and notifies Dave to review before promoting it live. Use when asked to generate, build, preview, or promote the FourthOS weekly sponsor update, or when the CCv3-FourthOS-Weekly scheduled task runs. Triggers on "fourthos weekly", "sponsor update", "Carly update", "Christian update", "weekly portfolio update".
Run a feature through the full tri-model Game Plan pipeline — rostered roles (Claude hub, Grok builder/researcher, Codex reviewer/fixer), a workroom disk bus, and human gates. Use when the user types /game-plan [feature], says "run the game plan on X", "full crew on this", "roster this feature", or wants a multi-milestone feature built with cross-model build/review separation. Opt-in orchestrator on top of /workroom; for one-shot tasks use /grok or /codex directly, for trivial edits use neither.
Audit and curate the memory system -- identify signal vs noise, score entries, and archive low-quality learnings.
Detect drift between the continuous-claude repo and the active ~/.claude/ directory
Initialize Continuous Claude v3 for a new project with full toolset activation. Creates project CLAUDE.md (interview-driven, 8 sections), ROADMAP.md, knowledge tree, Serena code intelligence, and project registry entry. Use when opening a new project folder for the first time, starting a new project, setting up CCv3 in an existing repo, or the user says "init project", "setup project", "new project", "initialize", "start new project".
| name | review |
| description | Comprehensive code review workflow - parallel specialized reviews → synthesis (includes cross-model Codex adversarial pass by default) |
Multi-perspective code review with parallel specialists, including a cross-model adversarial pass via OpenAI Codex. Same-family reviewers share blind spots — Codex (different training family) provides cross-model triangulation. See .claude/rules/codex-adversarial.md for the full convention.
┌──────────────┐
│ critic │ ─┐
│ (code) │ │
└──────────────┘ │
│
┌──────────────┐ │
│ plan-reviewer│ ─┤
│ (plan) │ │ ┌──────────────┐
└──────────────┘ ├────▶ │ review-agent │
│ │ (synthesis) │
┌──────────────┐ │ └──────────────┘
│ plan-reviewer│ ─┤
│ (change) │ │
└──────────────┘ │
│
┌──────────────┐ │
│codex-adversary│ ─┘ ← Cross-model (OpenAI Codex)
│ (mode=code) │
└──────────────┘
Parallel Sequential
perspectives synthesis
| # | Agent | Focus | Execution |
|---|---|---|---|
| 1 | critic | Code quality, patterns, readability | Parallel |
| 1 | plan-reviewer | Architecture, plan adherence | Parallel |
| 1 | plan-reviewer | Change impact, risk assessment | Parallel |
| 1 | codex-adversary | Cross-model adversarial review (different training family) | Parallel |
| 2 | review-agent | Synthesize all reviews, final verdict | After 1 |
.claude/rules/codex-adversarial.md)# Code quality review
Task(
subagent_type="critic",
prompt="""
Review code quality: [SCOPE]
Evaluate:
- Code style and consistency
- Design patterns used
- Readability and maintainability
- Error handling
- Test coverage
Output: List of issues with severity (critical/major/minor)
""",
run_in_background=true
)
# Architecture review
Task(
subagent_type="plan-reviewer",
prompt="""
Review architecture alignment: [SCOPE]
Check:
- Follows established patterns
- Matches implementation plan (if exists)
- Consistent with system design
- No architectural violations
Output: Alignment assessment with concerns
""",
run_in_background=true
)
# Change impact review
Task(
subagent_type="plan-reviewer",
prompt="""
Review change impact: [SCOPE]
Assess:
- Risk level of changes
- Affected systems/components
- Backward compatibility
- Potential regressions
- Security implications
Output: Risk assessment with recommendations
""",
run_in_background=true
)
# Cross-model adversarial review (skip if --no-codex flag passed)
# See .claude/rules/codex-adversarial.md for cost/quota guidance.
Task(
subagent_type="codex-adversary",
prompt="""
## Mode
code
## Scope
base ref: [BASE_REF, e.g. main or HEAD~1]
## Focus
[optional — e.g. "auth boundary", "data loss paths", or omit to let Codex pick]
## Codebase
$CLAUDE_PROJECT_DIR
""",
run_in_background=true
)
# Wait for all four parallel reviews
[Check TaskOutput for all four]
Task(
subagent_type="review-agent",
prompt="""
Synthesize reviews for: [SCOPE]
Reviews:
- critic: [code quality findings]
- plan-reviewer: [architecture findings]
- plan-reviewer: [change impact findings]
- codex-adversary: [cross-model findings — prefix each with [Codex] in output]
Treat codex-adversary as a DISTINCT cross-model input — NOT pooled with critic.
Findings BOTH critic and codex-adversary flag are high-confidence.
Findings only codex-adversary flags are the cross-model lift.
THIN-APPROVE RULE (dogfood-proven 2026-07-11 — this exact check caught the run's
only HIGH finding): if codex-adversary returns approve with ZERO findings AND its
run showed instability (fork-storms, timeouts, near-instant return, no visible
engagement with the diff), do NOT bank the approve — evidence quality inherits
from process health. Add a supplemental critic pass over the same scope before
the verdict. Asymmetry: distrust thin approvals; respect surviving objections
(instability suppresses detection, it doesn't fabricate findings).
Create final review:
- Overall verdict (APPROVE / REQUEST_CHANGES / NEEDS_DISCUSSION)
- Prioritized action items (group: both-flagged, claude-only, [Codex]-only)
- Blocking vs non-blocking issues
- Summary for PR description
"""
)
After synthesis, append one row to .claude/logs/codex-lift.jsonl:
mkdir -p "$CLAUDE_PROJECT_DIR/.claude/logs"
jq -nc \
--arg ts "$(date -u +%Y-%m-%dT%H:%M:%SZ)" \
--arg skill "review" \
--arg scope "[SCOPE]" \
--argjson critic_count <N> \
--argjson codex_count <N> \
--argjson both_count <N> \
'{ts:$ts, skill:$skill, scope:$scope, critic_only:($critic_count - $both_count), codex_only:($codex_count - $both_count), both:$both_count, via:"direct"}' \
>> "$CLAUDE_PROJECT_DIR/.claude/logs/codex-lift.jsonl"
This is the actual signal of value — track cross-model lift over time.
User: /review
→ All five agents (critic + 2x plan-reviewer + codex-adversary + review-agent), comprehensive review
User: /review --quick
→ critic only, fast feedback
User: /review --no-codex
→ Skip codex-adversary (saves a Codex turn). Use for trivial diffs, doc-only changes,
or when ChatGPT subscription quota is constrained. See .claude/rules/codex-adversarial.md.
User: /review --security
→ Add aegis (security agent) to parallel phase
User: /review PR #123
→ Fetch PR diff, review changes
User: /review the authentication changes
Claude: Starting /review workflow...
Phase 1: Running parallel reviews...
┌────────────────────────────────────────────┐
│ critic: Reviewing code quality... │
│ plan-reviewer: Checking architecture... │
│ plan-reviewer: Assessing change impact... │
│ codex-adversary: Cross-model review... │
└────────────────────────────────────────────┘
critic: Found 2 issues
- [minor] Inconsistent error messages in auth.ts
- [major] Missing input validation in login()
plan-reviewer: ✅ Matches authentication plan
plan-reviewer: Medium risk
- Affects: login, signup, password reset
- Breaking change: session token format
codex-adversary: verdict=needs-attention, 3 findings
- [Codex] auth.ts:42 — session token rotation not idempotent under concurrent login
- [Codex] login.ts:88 — missing input validation in login() (agrees with critic)
- [Codex] session.ts:14 — token format change has no migration path for in-flight sessions
Phase 2: Synthesizing...
┌─────────────────────────────────────────────┐
│ Review Summary │
├─────────────────────────────────────────────┤
│ Verdict: REQUEST_CHANGES │
│ │
│ Cross-model agreement (high confidence): │
│ 1. Add input validation to login() │
│ (both critic + [Codex]) │
│ │
│ [Codex]-only (cross-model lift): │
│ 2. Session token rotation race condition │
│ 3. No migration path for in-flight sessions │
│ │
│ Non-blocking: │
│ 4. Standardize error messages (critic) │
│ │
│ Telemetry: critic_only=1, codex_only=2, │
│ both=1 (logged to codex-lift.jsonl) │
└─────────────────────────────────────────────┘