References

Skill: Cross Model Review

This is an English mirror of the authoritative Chinese SKILL.md. In case of discrepancies, the Chinese source takes precedence.

name: cross-model-review description: Adversarial cross-model review of locked PLAN.md — auto-discover available CLIs (codex/gemini/claude), recommend a combination of two different providers and let user choose, reviewer runs read-only outputting VERDICT:APPROVED|REVISE, main agent arbitrates and records reasoning in PLAN-REVIEW-LOG.md, MAX_ROUNDS without convergence outputs deadlock. Based on chaseai-yt/grill-me-codex (MIT) Act 2 approach. locale: zh-CN supported_locales: [zh-CN, en-US]

Mandatory Entry

When this skill is triggered, you must first read in full references/cross_model_review.md and execute according to its terms.

• Do not substitute the full text with preamble, Cursor rule summaries, or other secondary summaries.
• This skill is Act 2 of plan-grill; after plan-grill locks PLAN.md, this skill takes over.

Five Core Rules

• [CMR-001] Auto-discover reviewers: Directly probe availability, versions, non-interactive and read-only mode support for three CLIs: codex/gemini/claude (using command -v <cli> + <cli> --version; optional auxiliary script in this repository: skills-engineering/scripts/detect-review-clis.sh). When available providers < 2, stop and prompt to install; do not fabricate cross-model.
• [CMR-002] Recommended combination + user selection: From available CLIs, recommend a combination of two different providers (e.g., codex + gemini) and let user confirm. Do not silently choose for the user.
• [CMR-003] Reviewer read-only: Each reviewer must run in read-only mode — codex uses -s read-only, gemini uses --approval-mode plan, claude uses --permission-mode plan. Reviewer does not write code, only outputs VERDICT: APPROVED or VERDICT: REVISE + specific modification suggestions.
• [CMR-004] Main agent arbitration: The main agent (Claude/Codex, depending on host) is the final arbitrator. Each round must collect verdicts from all selected reviewers; reviewer raw output, intermediate output, and delivery logs must be saved under the current project root directory (recommended .plan-reviews/<date>-<slug>/raw/); must not use /tmp as reviewer output buffer. Only all APPROVED can converge; any REVISE must be arbitrated and proceed to revision/next round. Adopt criticisms with evidence, reject criticisms that don't hold with reasoning written, recorded in PLAN-REVIEW-LOG.md.
• [CMR-005] MAX_ROUNDS + deadlock: When MAX_ROUNDS (default 5) still not converged, output deadlock — list each unresolved point + main agent's counter-position, hand to user for adjudication. Do not pretend approved.

Details in references/cross_model_review.md. Full running example for login rate limiting scenario in examples/regression-login-rate-limit.md.

When to Load

• Default trigger: User says cross-model-review / cross review / "adversarial review" / "let two models review the plan" / "model debate" / "stress-test PLAN.md" / "review PLAN.md" / "let Gemini/Codex/Claude review the plan".
• Relay from plan-grill: After plan-grill locks PLAN.md, user says "let another model review" then load this skill.
• Skip: No PLAN.md (run plan-grill first); trivial changes; user explicitly says "implement directly".

Division of Labor with Adjacent Skills