🎯Basically, Cloudflare built an AI system to help engineers review code faster and better.
What Happened
Cloudflare has developed a novel AI code review system utilizing OpenCode, aiming to improve the efficiency of their engineering teams. Traditional code review processes often lead to bottlenecks, with long wait times for reviews. To address this, Cloudflare experimented with various AI tools but found them lacking in flexibility and customization. Thus, they created a CI-native orchestration system that leverages multiple specialized AI agents to perform code reviews more effectively.
The Architecture: Plugins All the Way to the Moon
The architecture of the new system is designed to be modular and adaptable. By employing a composable plugin architecture, Cloudflare ensures that the system can easily integrate with different version control systems and AI providers without hardcoding dependencies. Each plugin is responsible for specific tasks, such as fetching data from GitLab or checking compliance with internal standards. This isolation allows for seamless updates and maintenance, ensuring longevity and adaptability in a fast-evolving tech landscape.
How We Use OpenCode Under the Hood
OpenCode was chosen for its open-source nature and existing internal use at Cloudflare. The system operates in two layers: the Coordinator Process and the Review Plugin. The coordinator manages the review process, while the review plugin handles the execution of sub-reviewers, which are specialized AI agents that evaluate the code. This setup allows for concurrent processing and efficient handling of large merge requests.
What’s JSONL, and What Do We Use It For?
To manage logging effectively, Cloudflare employs JSONL (JSON Lines) format for structured logging. This format allows each line of log data to be a valid JSON object, facilitating easier parsing and reducing memory overhead. It ensures that even if a process exits early, the logs remain accessible and usable, which is crucial for debugging and monitoring.
The Streaming Pipeline
The output from the coordinator is processed in real-time, allowing for immediate feedback on code quality. This approach not only speeds up the review process but also enhances the accuracy of the reviews, as multiple AI agents contribute their findings, which are then consolidated into a single structured comment for the engineer.
Conclusion
Cloudflare's innovative approach to AI code reviews showcases the potential for AI to transform engineering practices. By leveraging OpenCode and a modular architecture, they have created a system that not only improves code quality but also enhances the efficiency of their engineering teams. As AI continues to evolve, such implementations may become standard practice in software development, paving the way for faster and safer code deployment.
🔒 Pro insight: This orchestration model highlights the growing trend of integrating AI into CI/CD pipelines, enhancing both speed and accuracy in code reviews.
_NicoElNino_Alamy.png?width=1280&auto=webp&quality=80&disable=upscale)
