# Cross Review Policy

> Actionable review routing policy — which agent reviews which agent's work, default three-agent adversarial review

- Skill: `vamseeachanta/cross-review-policy-2` (Agent Skill)
- Install (CLI): `npx skillmds@latest add vamseeachanta/cross-review-policy-2`
- Raw SKILL.md: https://api.skillmd.com/api/skills/vamseeachanta/cross-review-policy-2/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: vamseeachanta (https://skillmd.com/u/vamseeachanta)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/vamseeachanta/cross-review-policy-2

---


# Cross-Review Policy

Actionable enforcement of the AI Review Routing Policy (`docs/standards/AI_REVIEW_ROUTING_POLICY.md`).

## Provider Roles

| Provider | Role | Scope |
|----------|------|-------|
| **Claude Code** | Default orchestrator | Task framing, planning, routing, repo-facing workflow |
| **Codex** | Default coding worker & adversarial reviewer | Bounded implementation, test writing, refactors, diff review |
| **Gemini** | Default adversarial reviewer | Architecture review, large-context research, plan & code review |

## Default: Three-Agent Adversarial Review

All plan-stage and code/artifact-stage work gets reviewed by ALL three agents unless the user explicitly scopes down.

### Plan review
1. Claude frames the plan
2. Codex reviews for implementation feasibility
3. Gemini reviews for architecture and scope

### Code review
1. Implementation agent produces the diff
2. Two other agents review independently
3. Claude synthesizes the combined review result

## Reviewer Stance (mandatory)

Routing is not enough. A reviewer who agrees charitably produces false positives that cost more to fix later than the cold review saved. Every review prompt (plan, code, artifact) MUST:

1. Frame the reviewer as adversarial: "Assume the work has defects until proven otherwise."
2. Forbid praise and restatement — findings only.
3. Bias toward non-approval: APPROVE requires affirmative verification of each correctness-critical claim.
4. Require evidence: each finding cites a file path, plan section, or quoted claim.
5. Treat the work's own cited sources as claims to verify, not facts to trust.
6. Treat empty reviews as failures: if nothing is found, the prompt must make the reviewer list what was checked.

A review that returns APPROVE without a check-list is suspect — rerun with a stronger prompt.

Rationale: user feedback 2026-04-17 on #2323 — "Make all the reviews adversarial in nature. Helps maximize productivity."

## Allowed Reductions

A narrower review is allowed ONLY when:

| Condition | Allowed adjustment |
|-----------|--------------------|
| User explicitly requests faster/lighter pass | Reduce to two-agent, document reason |
| Provider unavailable / quota exhausted | Continue with remaining, record missing reviewer |
| Purely clerical change (typo, comment) | Waive one reviewer with explicit note |

## Enforcement

- The pre-push review gate (`scripts/enforcement/require-review-on-push.sh`) checks for review markers
- Cross-review hook: `.claude/hooks/cross-review-gate.sh`
- Compliance tracked by: `scripts/enforcement/compliance-dashboard.sh`

## When to invoke this skill
- Before completing any implementation task
- Before pushing code to main
- When deciding review routing for a new task