Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseCode / SDK · ActionableGitHub Trending AI→ original

Alibaba Open-Sources AI Code Reviewer That Tens of Thousands of Developers Used for Two Years

Alibaba has released the code for Open Code Review, a tool that checks others' changes and writes comments on specific lines. It ran inside the company for two years, with tens of thousands of developers using it, and it found millions of defects. Ordinary code, not a neural network, picks the files and groups the changes, so large changes are not checked selectively. If your AI reviewer skips files and mixes up lines, this scheme is worth adopting.

AI-processed from GitHub Trending AI; edited by Hamidun News
Alibaba Open-Sources AI Code Reviewer That Tens of Thousands of Developers Used for Two Years
Source: GitHub Trending AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceGitHub Trending AI ↗
Editor-in-Chief
Installnpx skills add alibaba/open-code-review
Key metric #2Higher precision but lower recall than Claude Code: fewer false alarms
Key metric #3Works with Claude Code, Codex, Cursor, Kimi Code, and in CI

Code review is when a colleague looks over your changes before they are merged into the main branch and leaves comments on specific lines. Alibaba has released a tool that does this with a neural network: Open Code Review, a command-line program that grew out of the company's internal AI assistant. According to the authors, Alibaba's developers used it for two years, tens of thousands of them, and the tool found millions of defects in code.

After that testing, the team turned it into an open-source project, and it is now trending on GitHub, with about 206 stars in a day. The main difference from typical AI reviewers is that here the neural network does not control the entire process itself.

Why universal agents fall short

Many developers have tried handing a diff (a list of changes) to an agent such as Claude Code with a review skill. On large sets of changes, it starts to cut corners: it checks some files and silently skips the rest. A comment about "line 42" may actually refer to line 57. And a review that depends on the wording of a prompt is hard to debug: quality swings noticeably with small edits to the phrasing.

The authors' diagnosis: if a process is described entirely in words, it has no hard guarantees. Picture a reviewer who is told "look over everything carefully" but is not given a list of files. They will look at whatever seems important.

What the code does and what the model does

Open Code Review divides the work. Everything that must be reliable is handled by ordinary deterministic code:

  • It decides which files to review and which to filter out, so that an important change is not lost;
  • It groups related files into packages, and a separate subagent with an isolated context reviews each package. This keeps the review stable on huge changes and makes it easy to run in parallel;
  • It selects review rules for each specific file, so the model is not distracted by irrelevant material;
  • Separate modules anchor comments to the correct lines and re-verify the reviewer's own findings.

What remains for the agent is what cannot be done without a model: deciding which context to pull in and analyzing the code. It can open a file in full, search the repository, and look at other changed files. The authors derived the agent's toolset from analyzing real production call traces, and they optimized the prompts for review quality and token savings. There is also a mode without a diff: the command ocr scan reviews whole files or directories, for example when auditing an unfamiliar repository.

Quality, limitations, and setup

The authors claim that, on the same underlying model, their tool delivers higher precision than universal agents, uses fewer tokens, and finishes reviews faster. They tested on their own benchmark, AACR-Bench: real tasks from popular open-source repositories, with labels double-checked by more than 80 senior engineers. The benchmark is published on Hugging Face, so the comparison can be reproduced on your own code.

There is an honest caveat: recall, meaning the share of real defects found, is lower for this tool than for universal agents. This is a deliberate trade-off in favor of precision. Fewer false positives means less time spent sifting through noise, but some problems may slip through.

It can be installed with npx skills add alibaba/open-code-review or npm install -g @alibaba-group/open-code-review. There are two modes. In native mode, you need your own model endpoint, and the full pipeline runs. In delegation mode, your agent (Claude Code, Codex, Cursor, or Kimi Code) performs the review on its own model, while OCR handles file selection, rules, and report structure. No separate key is needed in this case, but an agent subscription does not replace an API key for native mode. There are integrations with GitHub Actions, GitLab CI, and Gerrit, as well as OpenTelemetry for observability.

For teams building their own review agents, the takeaway is architectural: leave file selection, grouping, and verification of comment positions to ordinary code, and reserve the language model for analysis.

Story

Key Questions Answered

Why universal agents fall short?

Many developers have tried handing a diff (a list of changes) to an agent such as Claude Code with a review skill. On large sets of changes, it starts to cut corners: it checks some files and silently skips the rest. A comment about "line 42" may actually refer to line 57. And a review that depends on the wording of a prompt is hard to debug: quality swings noticeably with small edits to the phrasing. The authors' diagnosis: if a process is described entirely in words, it has no hard guarantees.

What the code does and what the model does?

Open Code Review divides the work. Everything that must be reliable is handled by ordinary deterministic code: It decides which files to review and which to filter out, so that an important change is not lost; It groups related files into packages, and a separate subagent with an isolated context reviews each package.

What are the key details on Quality, limitations, and setup?

The authors claim that, on the same underlying model, their tool delivers higher precision than universal agents, uses fewer tokens, and finishes reviews faster. They tested on their own benchmark, AACR-Bench: real tasks from popular open-source repositories, with labels double-checked by more than 80 senior engineers. The benchmark is published on Hugging Face, so the comparison can be reproduced on your own code.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Hands-On AI Engineering · Hamidun Academy

Build & Ship Production AI Agents Without Boilerplate Code

Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.

Start 72h Free Trial →Vibe-Coding Intensive
What do you think?
Loading comments…