claude-pr-review workflow
/pr-review sends a pull request through five independent reviewers, cross-checks findings and builds one report.
- 114 ★ on the repository
- Last commit
Extended access
Why this level
- A run starts background processes, `gh` and external CLIs on a local clone.
- Codex sections temporarily rewrite `~/.codex/config.toml`.
Install
- Works in
Manual install
Read the repository's `INSTALL.md`, ask which configuration I pick and run only that configuration's steps. The `cp` commands overwrite same-named files in `~/.claude/`, so show first what would collide.
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- A run starts background processes, `gh` and external CLIs on a local clone.
- Codex sections temporarily rewrite `~/.codex/config.toml`.
This is third-party code. Review the repository files before installing.
What it does
The `/pr-review` command for Claude Code runs up to five reviews of one pull request from different angles: Claude Code subagents with full repository access, a neutral `codex review` on the diff alone, an adversarial Codex pass and two Gemini passes. The principle comes from security auditing: the agent that finds an issue never verifies it. Findings are cross-checked between axes, and disputed ones are kept in the report with evidence from both sides. Before a run the command shows a matrix of axes and you choose which to enable. Every changed file gets an outcome: a finding, no issues or a deliberate skip. A second round reconciles earlier findings and marks them fixed, still open or stale. The report is written in Traditional Chinese (zh-TW) with paste-ready inline comments; change the language in a fork.
Who it is for. Teams that review large pull requests and can spend time and tokens on several independent reviews.
Good fit when
- A change touches payments or authorization and one reviewer is not enough.
- A frontend and backend pair of pull requests breaks on the contract between them.
- The author fixed comments and you need to see which findings are closed.
- A hotfix heads to staging and inherited defects must not land on its author.
Not a fit when
- The pull request is small and one quick agent pass is enough.
- There is no local clone of the repository for a temporary `git worktree`.
- The report is needed in English without editing the command.
Example request
Run /pr-review for this pull request with the Claude Code and Codex neutral axes enabled and show the report with findings.Limitations
The repository was extracted from the author's personal setup, so expect to adjust it. Claude Code with subagents, `gh` and a local clone of the reviewed repository are required. Codex axes need the OpenAI Codex CLI and its plugin, Gemini axes need `agy`. Codex sections modify `~/.codex/config.toml` during a run and restore it afterwards, so read steps 3 and 7 of the command before the first launch. It was tested on macOS. A run on a mid-size pull request takes tens of minutes and millions of Codex tokens.
How to disable. Delete the copied `pr-review.md`, agent files, scripts and `model-routing.env` from `~/.claude/`. If you installed the `pr-review-c4-dispatch-gate.py` hook, remove its settings entry too.
FAQ
How do I install claude-pr-review?
Copy the command and agents into `~/.claude/`: `mkdir -p ~/.claude/commands && cp commands/pr-review.md ~/.claude/commands/`, then `cp agents/*.md` into `~/.claude/agents/`. The full order is in `INSTALL.md`.
Can I run it without Codex and Gemini?
Yes. The minimal configuration uses Claude Code subagents only, and a missing axis appears in the report as a gap rather than a failure.
What does the coverage check mean?
After review every changed file must carry a status: a finding, no issues or a deliberate skip. A missing status counts as a violation.
Which axes are enabled by default?
Before a run the command shows an axis matrix and waits for your choice: a recommendation is not consent and silence selects nothing.
Can I review two pull requests at once?
Yes. Pass several URLs: the command prepares each in its own worktree, picks axes once and emits a group report. It was built for frontend and backend pairs.
Discussion
No questions or reviews yet.
Related
WorkflowsOpenQodex reviews your change before a push: scanners with no model first, then a separate Claude Code or Codex process triages findings on your login
Reviews the git diff and the conversation with the agent: catches bugs and gaps between your request and the agent's result. CLI, skill, GitHub Action
A 0 to 10 score for CLAUDE.md and AGENTS.md: Reporails finds vague rules, repeats and broken imports
Open-source AI agent for pull request review: descriptions, feedback and improvement suggestions on GitHub, GitLab and more
Hooks for Claude Code and Codex CLI keep a ledger of agent actions and refuse a "done" claim until a check has passed.
A Claude Code plugin that verifies a change against its plan: boots your stack, runs acceptance criteria, reports evidence.
README in short
The repository ships a command, six reviewer agents, severity references and report scripts. The author explains the methodology in a zh-TW article and states plainly that it is tested on macOS and only partly covered by tests.