agentsclimarketplace

Cross model review

Skill i-stack/ai-coding-kit/skills-engineering/cross-model-review

Sync your MCP servers, Agent Skills, and AI rules across Cursor, Claude Code, Cline & more in one click.

Install
npx -y skills add i-stack/ai-coding-kit --skill cross-model-review

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

跨模型对抗审查已锁定的 PLAN.md——自动发现可用 CLI(codex/gemini/claude),推荐两个不同 provider 的组合并让用户选择,reviewer 只读运行输出 VERDICT:APPROVED|REVISE,主 agent 仲裁并把理由写进 PLAN-REVIEW-LOG.md,MAX_ROUNDS 不收敛输出 deadlock。基于 chaseai-yt/grill-me-codex(MIT)的 Act 2 思路。

SKILL.md

3.3 KB, as published. Nobody here has run it

Cross Model Review

强制入口

命中本 skill 时,必须先完整阅读 references/cross_model_review.md 并按其中条款执行。

  • 不得以 preamble、Cursor 规则摘要或其它二次摘要代替该文件全文。
  • 本 skill 是 plan-grill 的 Act 2;plan-grill 锁定 PLAN.md 后接力本 skill。

五条核心规则

  • [CMR-001] 自动发现 reviewer:直接探测 codex/gemini/claude 三个 CLI 的可用性、版本、non-interactive 与只读模式支持(可用 command -v <cli> + <cli> --version,本仓库可选辅助脚本为 skills-engineering/scripts/detect-review-clis.sh)。可用 provider < 2 时停止并提示安装,不伪造 cross-model。
  • [CMR-002] 推荐组合 + 用户选择:从可用 CLI 中推荐两个不同 provider 的组合(如 codex + gemini),让用户确认。不静默替用户选死。
  • [CMR-003] reviewer 只读:每个 reviewer 必须以只读模式运行——codex 用 -s read-only,gemini 用 --approval-mode plan,claude 用 --permission-mode plan。reviewer 不写代码,只输出 VERDICT: APPROVEDVERDICT: REVISE + 具体修改建议。
  • [CMR-004] 主 agent 仲裁:主 agent(Claude/Codex,视宿主而定)是最终仲裁者。每轮必须收集所有已选 reviewer 的 verdict;reviewer 原始输出、中间输出和交付日志必须保存在当前项目根目录下(推荐 .plan-reviews/<date>-<slug>/raw/),不得用 /tmp 作为 reviewer 输出缓冲。只有全部 APPROVED 才能收敛,任一 REVISE 都必须仲裁并进入修订/下一轮。采纳有证据的批评,拒绝不成立的批评并写明理由,记录进 PLAN-REVIEW-LOG.md
  • [CMR-005] MAX_ROUNDS + deadlock:到 MAX_ROUNDS(默认 5)仍不收敛时,输出 deadlock——列出每个未决点 + 主 agent 的反立场,交给用户裁决。禁止假装 approved。

细则见 references/cross_model_review.md。登录限流场景的完整运行样例见 examples/regression-login-rate-limit.md

何时加载

  • 默认触发:用户说 cross-model-review / cross review / "对抗审查" / "让两个模型审计划" / "model debate" / "stress-test PLAN.md" / "review PLAN.md" / "让 Gemini/Codex/Claude 审一下计划"。
  • 接力 plan-grill:plan-grill 锁定 PLAN.md 后,用户说"让另一个模型审查"则加载本 skill。
  • 跳过:没有 PLAN.md(先跑 plan-grill);trivial 改动;用户明确"直接实施"。

与相邻 skill 的分工

Skill分工
plan-grill(PG-001~004)盘问锁定 PLAN.md(Act 1)
cross-model-review(本 skill)跨模型对抗审查 PLAN.md(Act 2)
problem-analysis(PA-001~003)问题审查,先于 plan-grill
epistemic-integrity(GR-011~013)主 agent 仲裁时的真值接地纪律

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.