cross-review-policy
Actionable review routing policy — which agent reviews which agent's work, default three-agent adversarial review
Best use case
cross-review-policy is best used when you need a repeatable AI agent workflow instead of a one-off prompt.
Actionable review routing policy — which agent reviews which agent's work, default three-agent adversarial review
Teams using cross-review-policy should expect a more consistent output, faster repeated execution, less prompt rewriting.
When to use this skill
- You want a reusable workflow that can be run more than once with consistent structure.
When not to use this skill
- You only need a quick one-off answer and do not need a reusable workflow.
- You cannot install or maintain the underlying files, dependencies, or repository context.
Installation
Claude Code / Cursor / Codex
Manual Installation
- Download SKILL.md from GitHub
- Place it in
.claude/skills/cross-review-policy/SKILL.mdinside your project - Restart your AI agent — it will auto-discover the skill
How cross-review-policy Compares
| Feature / Agent | cross-review-policy | Standard Approach |
|---|---|---|
| Platform Support | Not specified | Limited / Varies |
| Context Awareness | High | Baseline |
| Installation Complexity | Unknown | N/A |
Frequently Asked Questions
What does this skill do?
Actionable review routing policy — which agent reviews which agent's work, default three-agent adversarial review
Where can I find the source code?
You can find the source code on GitHub using the link provided at the top of the page.
SKILL.md Source
# Cross-Review Policy Actionable enforcement of the AI Review Routing Policy (`docs/standards/AI_REVIEW_ROUTING_POLICY.md`). ## Provider Roles | Provider | Role | Scope | |----------|------|-------| | **Codex** | 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. Codex 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. Codex 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: `.Codex/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
Related Skills
plan-review-prompt-refresh-after-plan-edits
Refresh reviewer prompt files from the latest on-disk plan before every adversarial re-review. Prevents Codex/Gemini from critiquing stale plan text after local edits.
tdd-verification-and-adversarial-review
Verify pre-written TDD tests pass, conduct adversarial code review on committed diffs, and route findings to existing issues
multi-role-agent-contract-review-pipeline
Execute a 4-role agent team (Planner/Architect/Reviewer/Integrator) pipeline for self-reviewing knowledge artifacts before delivery
gsd-adversarial-review-pattern
Catch hidden test failures by running adversarial review on sparse-data edge cases before final push
boundary-policy-classification-by-role
Classify artifacts as durable vs transient by their functional role rather than directory path, using multi-layer architectural validation
adversarial-code-review-tdd
Systematic adversarial review pattern to identify breaking assumptions in already-passing test suites
adversarial-code-review-for-committed-diffs
Systematic process for reviewing already-committed code changes to catch type inconsistencies, edge cases, and docstring gaps
adversarial-code-review-and-fix
Systematic pattern for catching design flaws in already-passing code through adversarial review, then fixing them with TDD confirmation.
learned-git-worktree-hook-path-and-real-hook-shape-review
Catch hook-installation and governance bugs that only appear in linked git worktrees or against the real generated hook shape, not simplified test fixtures.
ten-agent-pre-plan-review-wave
Launch and verify a 10-agent planning-only wave that moves open GitHub issues into status:plan-review using one isolated worktree per issue, wave-specific continuation cron, and post-run artifact-reconciliation checks.
static-site-build-artifact-plan-review
Plan-review pattern for static-site fixes where the deployed artifact is generated from source files (e.g. sitemap/robots/static assets). Prevents review churn by separating durable regression checks from one-time migration verification and by validating built output, not just source files.
preserved-plan-refile-with-attested-review-wave
Reopen a previously closed GitHub issue with a preserved local plan, rewrite it into a conservative draft, and drive iterative attested adversarial review waves until it is truly approval-ready.