How to use the Open Code Review Delegate agent skill

What the skill makes the agent do

Normally, a code review tool such as open-code-review calls a separate LLM service to do the reasoning. This skill switches that around. The OCR CLI handles deterministic work: deciding which changed files are reviewable and which review rules apply to each file. The host agent, the AI assistant you are already using, reads the diffs and writes the review comments with its own model.

The agent works through a fixed sequence: preview the changes, fetch rules, read each diff, review each file, format comments, and report. It then checks coverage, so every file is listed as either reviewed or skipped with a reason, and findings are grouped by severity (critical and high, medium, low).

Install the skill

Download the skill ZIP from this page and unzip it into one of two locations. For a personal install available in all your projects, use ~/.claude/skills/. For a single project, use the .claude/skills/ folder inside that project.

If you use the Claude apps instead, upload the ZIP under Settings → Capabilities → Skills.

The skill needs the ocr CLI on your machine. The source describes two ways to get it: install it with npm, or download a binary from the GitHub release. The npm package name is @alibaba-group/open-code-review.

• Personal install: unzip into ~/.claude/skills/
• Project install: unzip into the project's .claude/skills/ folder
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

Requirements before you start

You need three things: the ocr CLI installed, git available in the repository you want reviewed, and a host AI agent that can read files and run shell commands. The skill does not need a configured LLM endpoint, because the review is done by your agent rather than by OCR.

• ocr CLI installed via npm (@alibaba-group/open-code-review) or a GitHub release binary
• git available in the repository
• A host AI agent capable of reading files and running shell commands

Trigger the review with natural requests

You do not need to type commands yourself. Ask the agent in plain language, for example to review your current branch against main, or to review one commit. The agent chooses the matching ocr delegate preview options. The source's common invocations map to these scenarios: uncommitted workspace changes, a branch comparison, and a single commit.

If you have requirement context, such as a ticket description, you can include it. The source says --background passes business context into the preview output for reference during review, and --background-file reads it from a Markdown file.

• Review my uncommitted changes before I commit.
• Review the feature branch against main and flag anything risky.
• Review commit abc123 and list every file as reviewed or skipped.
• Review this branch and fix only the critical and high issues.

Read the report

Each finding is a comment with a path, the review text, optional line numbers in the new file, and optional category and severity. The summary includes total_files, reviewed_files, skipped_files, and coverage_rate. Any skipped file must come with a reason.

Findings are grouped by severity. Critical and high items are always reported. Medium items are reported with context. Low items appear only when they are clearly valuable. The source says likely false positives are discarded silently.

• Critical/High: bugs, security issues, data loss risks
• Medium: performance, error handling gaps, maintainability
• Low: style nits and minor suggestions, only if clearly valuable

Safety notes: scripts and network

The skill runs scripts, because it calls the ocr CLI and git. The source lists the skill as using no network of its own. Diffs are read by the host agent and sent to that agent's model, so check that your agent's model provider is approved to see the source code you are reviewing. If the repository is sensitive, do not use this mode.

The diff commands disable external diff programs, text conversion, and color, so the output is plain patches. The skill also tells the agent to write large diffs to a file outside the repository and delete that file after reading it.

• Runs scripts: yes (ocr and git commands)
• Uses network: no, according to the source
• Diffs are sent to the host agent's model provider

Tips for better reviews

Run the review inside the repository you want checked, because ocr delegate uses the current directory. Use --repo /path to point elsewhere. Keep large changes manageable: the agent reviews them in batches grouped by shared rules.

Remember the limit of this skill. It reads diffs and can optionally apply fixes, but it does not run your tests to confirm correctness.

• Run from the repo root, or pass --repo /path
• Pass --exclude with comma-separated patterns to skip files
• Use --background-file for long requirement documents; the source sets a 1 MiB raw file limit and an 8000-character sanitized limit