Alibaba open-sources its internal code-review assistant as Open Code Review: higher precision and F1 than a general agent on the same model, at about one-ninth the tokens
Synopsis
Alibaba has open-sourced Open Code Review, its internal AI code-review assistant used for two years, and a user only needs to configure a model endpoint to start. It reads Git diffs and sends changed files to a configurable LLM through an agent with tool use, which can read full files, search the codebase and consult other changed files to produce structured, line-level review comments. On a benchmark of 50 open-source repositories, 200 real pull requests and 10 programming languages, it reaches clearly higher precision and F1 than a general-purpose agent with the same underlying model while consuming about one-ninth the tokens, but its recall is lower, a deliberate trade-off favoring precision over noise.
Interpretation
Code review now has an open-source command-line tool that starts from a configured model endpoint, handing the steps that must not go wrong to deterministic engineering logic and leaving dynamic judgment to the agent. Using general-purpose agents such as Claude Code with Skills for code review previously led to incomplete coverage on larger changesets, reported issues drifting from the actual code location, and quality that fluctuated with small prompt variations. The tool originated as Alibaba Group's internal official AI code-review assistant, which over the past two years served tens of thousands of developers and identified millions of code defects before being incubated as an open-source project.
Review is split into fixed steps for file selection, file bundling, rule matching and positioning and reflection, each guaranteed by engineering logic rather than the language model. File bundling groups related files into one review unit, for example message_en.properties with message_zh.properties, and each bundle runs as a sub-agent with isolated context, keeping the process stable on very large changesets and supporting concurrent review. Rule matching uses a template engine to dispatch review rules by file characteristics, while independent comment-positioning and comment-reflection modules improve the location accuracy and content accuracy of AI feedback.
The agent handles only dynamic decisions and dynamic context retrieval, with prompts and a toolset tuned specifically for code review. The toolset was distilled from tool-call traces in large-scale production data, including call frequency distributions, per-tool repetition rates and the impact of new tools on the overall call chain. Beyond diff review, ocr scan reviews entire files for auditing unfamiliar codebases or directories that have no meaningful diff.
Perspective
The demonstrated scope is code review: reading Git diffs to produce line-level comments, plus ocr scan for reviewing entire files in unfamiliar codebases; the benchmark consists of 50 popular open-source repositories, 200 real pull requests and 10 programming languages, cross-validated by more than 80 senior engineers into 1,505 annotated issues.
Recall is lower than that of general-purpose agents, a deliberate trade-off favoring precision over noise; the material does not report how that trade-off behaves across different changeset sizes.