alibaba/open-code-review

Open Code Review originates from Alibaba's internal official AI code review assistant, which the author states has served tens of thousands of developers over two years. It reads Git changes, packages related files into subtasks for review by tool-equipped agents, and uses independent localization and reflection modules to correct the position and content of comments; `ocr scan` can also perform full-file reviews across entire directories. On a benchmark of 50 open-source repositories and 200 real PRs, the author claims that with the same base model, accuracy and F1 are higher than Claude Code, token consumption is approximately one-ninth, but recall is lower.

Open Code Review originates from Alibaba's internal official AI code review assistant, which the author states has served tens of thousands of developers over two years. It reads Git changes, packages related files into subtasks for review by tool-equipped agents, and uses independent localization and reflection modules to correct the position and content of comments; ocr scan can also perform full-file reviews across entire directories. On a benchmark of 50 open-source repositories and 200 real PRs, the author claims that with the same base model, accuracy and F1 are higher than Claude Code, token consumption is approximately one-ninth, but recall is lower.

Why it matters File selection, rule matching, and comment localization are handled by deterministic programs, while the model only performs reviews and context retrieval, making it less likely to miss files or misalign line numbers even with large changes.

Stars 50628 (+215 / 24h)

Reasons

  • 215 new stars in 24 hours, total stars 50628 (indicating attention only)
  • The README clearly outlines design trade-offs: prioritizing accuracy and actively sacrificing recall, with the benchmark dataset publicly available on Hugging Face AACR-Bench

Usage

  • After installing globally via npm install -g @alibaba-group/open-code-review, configure the model using ocr config provider / ocr config model, then run ocr review in the project directory; plugins are also available for Claude Code, Codex, Cursor, and Kimi Code

Risks

  • Requires providing your own model API (delegation mode uses the host agent's own model); requires Git 2.41 or higher
  • The benchmark was constructed and evaluated by the author's team

Open

Trendshift(GitHub Trending 快照)

Explore the latest AI Radar updates