About this video

• Video Title: Automatic code reviews with OpenAI Codex
• Channel: OpenAI
• Speakers: Roma, Maya
• Duration: 00:08:20

Overview

This video demonstrates how OpenAI Codex can be integrated into team workflows for automatic code reviews. It highlights Codex's ability to function as a coding teammate by working with existing tools and plugging into team workflows, specifically focusing on the code review process. The video explains how Codex is trained to identify bugs, its comparison to static analysis, how it can be customized, and its application both in the cloud and through the Codex CLI.

Key takeaways

• Codex as a Coding Teammate: Codex is designed to work with existing tools and integrate into team workflows, with code review being a key area of focus.
• Enhanced Bug Detection: Codex models are trained specifically to find bugs and investigate issues, going beyond traditional static analysis by having access to the entire repository and the ability to run tests.
• GitHub Integration: Codex code reviews can be automatically triggered for pull requests submitted to a repository, with options to provide specific instructions to the agent.
• Customization and Control: Users can customize code reviews by providing specific instructions through comments or by leveraging the AGENTS.md file for more detailed guidelines, including what to focus on or ignore.
• CLI and Cloud Availability: Code reviews can be performed both in the cloud through GitHub integration and locally using the Codex CLI, allowing for reviews before code is even submitted.
• Iterative Improvement: The system supports an iterative workflow where Codex can not only identify issues but also be prompted to fix them, facilitating continuous improvement.

Codex's code review differs from traditional static analysis in that the model has access to the entire repository, not just the diff. It can track dependencies, understand the broader codebase, and even write and run Python code to test hypotheses about potential issues. This allows it to investigate issues more deeply and catch bugs that might not be apparent from a simple diff analysis.

Users can customize or provide specific instructions for Codex code reviews in several ways:

• Commenting on Pull Requests: Users can comment @codex review this PR and add extra information to help the agent understand the PR or to instruct it to focus on specific areas.
• Using AGENTS.md: The AGENTS.md file provides an open format for coding agents, including Codex, to follow specific instructions. This can include custom code review guidelines, requirements on what to pay special attention to, or types of problems to ignore.
• Custom Response Styles: Users can provide custom instructions about the style in which Codex responds, such as requesting it to always point out bugs while also offering encouragement.