ngosangns Knowledge Base

OpenCodeReview

#technology#coding-agents#code-review#ai-tools

Updated: 2026-09-29

Source

OpenCodeReview (ocr) is Alibaba's open-source AI code review CLI, written in Go under Apache-2.0. It originated as Alibaba Group's internal AI review assistant — two years serving tens of thousands of developers and flagging millions of defects — before being released as open source. It reads Git diffs, sends changed files to a configurable LLM through an agent with tool-use capabilities, and produces structured review comments with line-level precision. Beyond diffs, ocr scan reviews whole files for auditing unfamiliar codebases with no meaningful diff.

Architecture: Deterministic Engineering × Agent Hybrid

The core design splits responsibility between hard engineering constraints and the LLM agent.

Deterministic engineering — steps that must not go wrong:

  • Precise file selection — decides exactly which files get reviewed and which are filtered, so no important change is missed.
  • Smart file bundling — groups related files (e.g., message_en.properties + message_zh.properties) into one review unit; each bundle runs as a sub-agent with isolated context, keeping quality stable on large changesets and enabling concurrency.
  • Fine-grained rule matching — template-engine-based rules matched to each file's characteristics (built-in ruleset covers NPE, thread-safety, XSS, SQL injection across multiple languages), keeping model attention focused.
  • External positioning and reflection modules — independent comment-positioning and comment-reflection modules that systematically improve location accuracy and content accuracy of AI feedback.

Agent — dynamic decisions only:

  • Scenario-tuned prompts — templates optimized for review, reducing token cost.
  • Scenario-tuned toolset — distilled from large-scale production tool-call traces; the agent can read full files, search the codebase, and inspect other changed files for context.

Why It Exists

Motivation: general-purpose agents with review skills (e.g., Claude Code) suffer on real changesets — incomplete coverage (agents selectively skip files on large diffs), position drift (reported line numbers don't match actual code), and unstable quality (natural-language skills fluctuate with prompt variations). Pure language-driven review lacks hard constraints.

On AACR-Bench — 50 open-source repos, 200 real PRs, 10 languages, 1,505 ground-truth issues annotated by 80+ senior engineers — OCR achieves significantly higher Precision and F1 than general-purpose agents on the same underlying model, at ~1/9 the tokens and faster wall-clock time. Recall is deliberately lower: precision over noise.

Usage

Requires Git ≥ 2.41. Install:

npm install -g @alibaba-group/open-code-review   # `ocr` command
# also: install script, GitHub Release binary, build from source

Configure an LLM provider (OpenAI- and Anthropic-compatible, plus Bedrock and custom providers) via interactive ocr config provider + ocr config model, or use Delegation Mode instead.

ocr review                                   # staged + unstaged + untracked workspace changes
ocr review --from main --to feature-branch   # branch range, merge-base mode
ocr review --commit abc123                   # single commit
ocr review --from main --to feature-branch --resume <session-id>  # resume interrupted run
ocr scan --path internal/agent               # full-file scan, no diff needed
ocr review --format json --output result.json  # structured output for host agents
ocr delegate preview / ocr delegate rule <files>  # delegation mode, no LLM config needed

Execution Modes and Integrations

  • Default (OCR-managed) — OCR runs the review with its configured LLM.
  • Delegation Mode — a host coding agent (Claude Code, Codex, Cursor, Kimi Code, OpenCode, QCA, any skill-compatible agent) performs the review with its own LLM; OCR handles file selection and rule resolution only, so no separate API key is needed.
  • Extensions/plugins — Claude Code plugin with slash commands, Codex/Cursor/Kimi skills, OpenCode native tools, VS Code and IntelliJ IDEA extensions, MCP server for external tools, portable agent skill.
  • CI/CD — GitHub Action (action.yml, ocr-review.yml self-reviewing workflow), GitLab CI, GitFlic CI, Gerrit.
  • Session Viewer — browser UI to replay review sessions and mark comments fixed/ignored. OpenTelemetry telemetry for observability.

Project Facts

  • ~42k GitHub stars; OpenSSF Best Practices Gold badge; source files enforced English-only in CI; 90% coverage threshold (make coverage); SPDX headers on every source file.
  • Contributor rules require disclosing AI usage in issues/PRs and forbid attributing commits to AI — notable for an AI-tooling project.
  • Repo dogfoods itself: contributors run ocr review before committing, and CI runs an OCR review workflow.
  • Docs site: https://open-codereview.ai/docs — quickstart, CLI reference, review rules, delegate, CI/CD.

See also: AI Coding Productivity Tools · Coding Agents · Agents Overview