Open Code Review, an AI code review tool, allows users to enhance the review capabilities of existing AI by setting various rules, and has already detected 1 million code defects within the Alibaba Group.
Software development using AI agents is rapidly gaining popularity, and there is a growing trend to entrust code reviews to AI rather than humans. However, code reviews using AI tend to be prone to oversights and inconsistencies in quality. To solve these problems, Alibaba, one of China's leading technology companies, has developed a code review agent system called ' Open Code Review ,' which is already being used by tens of thousands of developers within the Alibaba Group and has reportedly detected one million code defects.
Open Code Review — Agent Native Code Review
Alibaba has identified the following three issues when entrusting code reviews to AI agents such as Claude Code.
Lack of comprehensiveness: When reviewing large-scale changes involving multiple files, the agent may focus on only some files, missing changes in other files.
Positional misalignment: When reporting a problem, line numbers or file references are reported incorrectly.
Unstable quality: Even slight changes to the prompt can drastically alter the review quality.
Alibaba points out that the reason for problems with existing agents is that 'logic based on language models lacks strict constraints on the review process.' Open Code Review incorporates engineering logic-based mechanisms rather than language model-based ones for file selection and rule matching, enabling it to perform deterministic reviews despite being an AI agent system.
Open Code Review allows you to perform reviews using any AI model, and it can reduce token usage to one-fifth compared to existing agents. It is already being used by more than 20,000 Alibaba Group employees and has successfully detected more than 1 million defects.
The table below shows the review performance benchmark results for Open Code Review and Claude Code. Open Code Review achieved the highest review quality when using Claude Opus 4.6, and even when using the same GLM-4.7, Open Code Review performed better than Claude Code.
The source code for Open Code Review is available at the following link.
GitHub - alibaba/open-code-review: Battle-tested at Alibaba's scale. Hybrid architecture code review tool: deterministic pipelines + LLM Agent, precise line-level comments, built-in fine-tuned ruleset (NPE, thread-safety, XSS, SQL injection), OpenAI & Anthropic compatible. · GitHub
https://github.com/alibaba/open-code-review
Related Posts:
in AI, Posted by log1o_hf