
Review
- Updated April 13, 2026
- joemiller/claude-plugins
review is a Claude Code skill in the Code Review & Quality category. Multi-model code review with cross-validation
Key points
- review
- Code Review & Quality
- AI-coding skill
Review by the numbers
- Data as of Jul 7, 2026 (Skillselion catalog sync)
/plugin marketplace add joemiller/claude-plugins/plugin install review@joemiller-pluginsAdd your badge
Show developers this skill is listed on Skillselion. Paste this into your README.
| Last updated | April 13, 2026 |
|---|---|
| Repository | joemiller/claude-plugins ↗ |
What it does
Multi-model code review with cross-validation
README.md
review
Multi-model code review plugin for Claude Code. Dispatches Claude and Codex reviewers in parallel, then a judge agent cross-validates and deduplicates findings.
Install
/plugin install review@joemiller-plugins
Usage
/review # uncommitted changes
/review --staged # staged changes only
/review HEAD~3..HEAD # git range
/review src/parser.ts # specific files
/review https://github.com/org/repo/pull/123 # GitHub PR
Use --agents N to run N reviewers per model (default asks interactively):
/review --agents 3 # 3 Claude + 3 Codex reviewers
/review --agents 1 https://github.com/org/repo/pull/42
How it works
- Resolves scope to a diff (local changes or PR)
- Generates N shuffled copies of the diff (randomized file order per reviewer to broaden coverage)
- Launches N Claude + N Codex reviewers in parallel, each performing an adversarial review
- A judge agent validates findings against the actual code, deduplicates, and produces a merged report
- In PR mode, optionally posts findings as inline PR comments
Each reviewer focuses on material issues: trust boundaries, resource leaks, race conditions, input validation, error handling, and state corruption. Style and naming nits are explicitly excluded.
Requirements
git,jq,shuf(coreutils)gh— PR mode onlycodexCLI — optional; falls back to Claude-only when unavailable
Architecture
- Skill (
review) — pure orchestrator. Dispatches agents and formats output, never reviews code itself. - Agents (
reviewer,judge) — return structuredFINDING:blocks. The skill handles all formatting.
Attribution
Prompts and architecture derived from @kelp's cross-review plugin.
Differences from upstream
- Single unified skill — handles both local changes and GitHub PRs, with
--agents Nfor multi-agent reviews - Judge agent replaces cross-validation — upstream runs each model as a validator of the other's findings (bidirectional); this plugin uses a dedicated judge that validates all findings in one pass
- No reconciliation step — upstream has
--reconcilefor disputed findings to return to the originator for rebuttal