Traditional AI programming assistants have a clear flaw: after the model completes a task, it often stops directly, lacking external verification. This leads to frequent issues such as poor code organization, missing edge cases, and security risks.

The claude-review-loop plugin developed by Hamel Husain attempts to solve this problem. It builds an automated "code review loop," letting Claude be responsible for task implementation, while the OpenAI Codex CLI provides a completely independent second opinion review.

## Two-Stage Workflow

The plugin uses a state machine mechanism to achieve seamless automation. When the user inputs the `/review-loop` command:

**Task Stage**: Claude implements the specified task completely according to the regular process.

**Review and Fix Stage**: When Claude attempts to stop, the plugin's stop hook triggers automatically. It executes the Codex CLI for a comprehensive analysis, generating a structured review report covering five dimensions: code quality, test coverage, security, documentation, etc.

Subsequently, Claude is asked to carefully read the review results, implement fixes for items it agrees with, and provide reasons if it disagrees. The entire process forms a closed loop.

## Interesting Findings from Real-World Usage

Some netizens discovered during usage that Claude sometimes shows resistance to Codex's review opinions, such as responding "This is over-engineering" or "It looks like they don't understand the overall architecture." Codex's comments, on the other hand, are often straightforward, even with a hint of passive-aggression.

This "dialogue" between models actually becomes an effective mechanism for improving code quality. As the developer said: "This is truly a perfect match."

## Technical Implementation Details

The core of the plugin lies in Claude Code's stop hook mechanism. When it detects it is in the task stage, the hook will:

1. Run Codex for an independent review

2. Write the review results to the `reviews/review-

3. Block Claude from exiting and require it to handle the review comments

4. Allow normal exit only after fixes are completed

The state file `.claude/review-loop.local.md` records the progress of the entire loop; it is recommended to add it to `.gitignore`.

## Installation and Usage

Installation requires just two two commands:

```bash

claude plugin marketplace add hamelsmu/claude-review-loop

claude plugin install review-loop@hamel-review

```

If Codex is not installed, the plugin will gracefully fall back to letting Claude perform a self-review.

This plugin is particularly suitable for engineering projects that require high-quality, maintainable code. It forcibly introduces a cross-model validation step, making AI programming no longer a one-sided output but a collaborative process with quality gatekeeping.

Project address: https://github.com/hamelsmu/claude-review-loop