adversarial-review（对抗性审查）

Adversarial Review

Use this skill to get independent review instead of same-thread self-validation.

First read

Open these before launching reviewers:

• references/reviewer-lenses.md
• references/reviewer-prompt.md
• references/cli-harnesses.md
• references/compatibility-matrix.md
• references/verdict-format.md

If the repo has any of these files and they matter to the change, read them before prompting reviewers:

• AGENTS.md
• CLAUDE.md
• GEMINI.md
• brain/principles.md

Workflow

1. Identify the review target: diff, staged changes, PR branch, implementation plan, or a tight file list.
2. Write one sentence of intent: what the author is trying to achieve.
3. Gather the smallest useful context:

- primary diff or staged changes - directly touched files - 1-3 nearby dependency files when needed - repo principles or project rules

1. Assign reviewer lenses:

- Codex reviewer -> Skeptic - Claude Code -> Architect - Gemini CLI -> Minimalist

1. Launch reviewers:

- in a Codex host, prefer a fresh Codex subagent for the Codex reviewer - otherwise use scripts/run_codex_reviewer.sh - use scripts/run_claude_reviewer.sh for Claude Code - use scripts/run_gemini_reviewer.sh for Gemini CLI

1. Write each reviewer result to its own file, for example:

- .tmp/adversarial-review/codex-skeptic.md - .tmp/adversarial-review/claude-architect.md - .tmp/adversarial-review/gemini-minimalist.md

1. Synthesize with references/verdict-format.md:

- verify every expected output file exists and is non-empty - record missing reviewers, auth failures, timeouts, capacity failures, malformed output, or caller misuse under reviewer coverage or harness failures - reject reviewer claims about flags, auth, or runtime behavior unless they match current docs or local CLI evidence - reject weak or hand-wavy claims; keep only evidence-backed findings - make a lead judgment for each finding: accept or reject with a one-line rationale

1. Return PASS, CONTESTED, or REJECT.

Defaults

• Prefer cross-family reviewer diversity first: Codex reviewer + Claude Code + Gemini CLI when all three are available.
• If only one reviewer path works, still run it and report reduced reviewer diversity.
• If Gemini is missing, unauthenticated, rate-limited, capacity-blocked, or times out, continue with Codex + Claude and report reduced reviewer diversity.
• Claude agent teams are researched but non-default. This skill is optimized for cross-family reviewer diversity, not same-family fan-out.
• Keep prompts compact. Let reviewers inspect the repo directly through their own tools.

Guardrails

• Review only. Do not edit files, commit, or open a PR as part of this skill.
• In a Codex host, do not call Codex CLI recursively when a fresh Codex subagent is available.
• Do not use unsafe approval or sandbox bypass modes for reviewer runs.
• Do not silently drop a failed reviewer.
• Do not block the whole review on Gemini availability.
• In Claude print mode, describe the task directly. Do not rely on slash commands or skills.
• In Gemini plan mode, do not ask Gemini to implement anything or exit plan mode.