OpenCodeReview

About

Running a general-purpose coding agent over every pull request is expensive and imprecise, which is why Alibaba built OpenCodeReview: two years of internal use across tens of thousands of developers shaped a design it calls Deterministic Engineering x Agent Hybrid. Deterministic stages select files, bundle them intelligently (i18n properties grouped, each bundle an isolated sub-agent), match template rules, and position comments precisely; a scenario-tuned LLM agent with a distilled toolset reads full files, searches the codebase, and makes the judgment calls. The ocr CLI reads Git diffs or scans whole trees, and a Delegation Mode lets an existing agent (Claude Code, Codex, Cursor, OpenCode) run the review with its own model — no separate API key. Distribution covers npm, install scripts, release binaries, CI integrations (GitHub Actions, GitLab, Gerrit), a VS Code extension, MCP server, and per-harness plugins. Alibaba publishes AACR-Bench alongside, claiming higher precision at roughly one-ninth the token cost of general-purpose agents.