Open Code Review

Open Code Review

This Codex plugin skill intentionally mirrors the canonical skill at skills/open-code-review/SKILL.md. Keep both files synchronized when updating OCR agent instructions; a symlink is avoided because plugin installs may only materialize the plugin subtree.

A skill for invoking open-code-review (ocr) — an open-source AI code review CLI that reads Git diffs and generates structured, line-level review comments.

Workflow

Step 1: Gather Business Context

Analyze the review target (commits, branch, or changes) to extract concise business context. Pass this context via --background to improve review quality.

Step 2: Run Code Review

Run the OCR command with appropriate flags. Always pass business context via --background when available:

ocr review --audience agent --background "business context here" [user-args]

Argument handling:

• Background context (RECOMMENDED): use --background "context" or -b "context" to provide business context for better review quality
• Default (no user arguments): reviews staged, unstaged, and untracked changes (workspace mode)
• Specific commit: use --commit or -c to review a single commit against its parent
• Branch comparison: use --from <ref> and --to <ref> to review diff between two refs
• Timeout: effective timeout per review group = --timeout × review rounds. Default --timeout 15 with default effort medium (2 rounds) gives 30 minutes; low/high give 15/45 minutes.
• Concurrency: default concurrency is 8 file workers; reduce with --concurrency <n> if rate limits are hit
• Preview mode: use --preview or -p to preview which files will be reviewed without running the LLM
• Output file: use --output <path> to write the full result to a file instead of stdout. If the command fails with unknown flag: --output, do not continue the review with plain stdout. Ask the user whether to upgrade (npm i -g @alibaba-group/open-code-review@latest) and wait for the answer before proceeding. After the user confirms and the upgrade succeeds, rerun with --output.
• Installation: if ocr command is not found, install it by running npm i -g @alibaba-group/open-code-review

Common invocation patterns:

Output mode:

• Always use --audience agent to suppress progress UI and emit only the final summary
• Prevent output truncation: For large reviews or restricted tool environments, pass --output /tmp/ocr_out.txt and inspect the file in full via a file reading tool instead of piping stdout through tail or head, which drops earlier review comments.

On failure: If ocr review exits non-zero (e.g. an LLM connection error), do not retry blindly — consult the Troubleshooting section below for the matching fix before re-running.

Step 3: Report

OCR output includes structured severity (critical / high / medium / low) and category (bug / security / performance / maintainability / test / style / documentation / other) on each comment. Present results grouped by severity, discarding low severity items that are likely false positives or nitpicks.

Step 4: Fix

Before applying fixes, check whether the user requested automatic fixes:

• If the user explicitly requested "review and fix" or similar, proceed with automatic fixes
• If the user only requested "review" without fix intent, ask for permission before applying any changes

When fixing issues and suggestions:

• Focus on critical, high, and medium severity items
• Apply fixes directly to the code when safe and well-defined
• For complex fixes requiring manual intervention, clearly describe what needs to be done
• Always verify fixes with the user before committing

Output Format

Each comment in OCR's output contains:

• path: File path
• content: Review comment text
• start_line / end_line: Line range (both 0 means positioning failed)
• category: Issue category (bug, security, performance, maintainability, test, style, documentation, other)
• severity: Issue severity (critical, high, medium, low)
• suggestion_code: Optional fix suggestion
• existing_code: Optional original code snippet
• thinking: Optional LLM reasoning process

Present results grouped by severity using this template:

## Code Review Results

**Files reviewed**: N
**Issues found**: X critical, Y high, Z medium

### Critical

- **`path/to/file.java:42`** [bug] — Brief description
  > Recommendation: How to fix

### High

- **`path/to/file.java:26`** [bug] — Brief description
  > Recommendation: How to fix

### Medium

- **`path/to/file.ts:88`** [performance] — Brief description
  > Recommendation: How to fix (if applicable)

If no critical, high, or medium severity issues remain after filtering, state: "Review complete — no critical, high, or medium issues found in N files."

Handling mispositioned comments:

When start_line and end_line are both 0, the comment failed to locate the exact position in the file. In such cases:

1. Read the comment content to understand the issue
2. Examine the target file mentioned in the comment
3. Identify the relevant code section based on the comment's context
4. Apply the fix or suggestion to the correct location

Custom Review Rules

If the user wants project-specific rules, OCR resolves them in this priority order:

1. --rule <path> flag (highest)
2. <repo>/.opencodereview/rule.json
3. ~/.opencodereview/rule.json
4. Built-in system defaults (lowest)

By default, the first matching user rule replaces the built-in system rule. Set merge_system_rule: true on a rule entry when the matched system rule and user rule should both be included.

Rule file format:

{
  "rules": [
    {
      "path": "**/*.java",
      "rule": "All new methods must validate required parameters for null",
      "merge_system_rule": true
    },
    {
      "path": "**/*mapper*.xml",
      "rule": "Check SQL for injection risks and missing closing tags"
    }
  ]
}

To preview which rule applies to a file before reviewing:

ocr rules check src/main/java/com/example/Foo.java

Advanced Review Options

Beyond the common flags above, ocr review exposes a few groups of controls. Run ocr review --help for the complete list.

Scoping

• --exclude '<patterns>' — comma-separated gitignore-style patterns (for example --exclude '**/generated/*,**/testdata/*'), merged with rule.json excludes.
• --background-file <path> — read review context from a Markdown file. Takes precedence over --background.

Output

• --format text|json|sarif — text (default) for humans; json for machine-readable findings; sarif for code-scanning integrations such as GitHub Code Scanning.

Model

• --provider <name> / --model <name> — override the configured provider/model for this run only (for example, to recheck a diff with a different model; the user names the model, ocr llm providers lists the built-ins).

Budget

• --max-tokens <n> — per-group prompt ceiling; defaults to the configured value or the template default (200000).
• --max-tokens-budget <n> — cap total input + output tokens for the run. Once exceeded, dispatch stops, partial results are still published, and skipped files are reported as failed(budget).
• --no-filter — keep all review comments and skip the LLM post-filtering call.

Gotchas

• LLM must be configured first — ocr review will fail loudly if no LLM is reachable. See the Troubleshooting section below if this happens.
• Working directory matters — ocr review operates on the Git repo at the current directory. Use --repo /path/to/repo to run from elsewhere.
• Untracked files are reviewed in workspace mode — running bare ocr review includes staged, unstaged, and untracked changes. Stage selectively if you want narrower scope.
• Large diffs may hit token limits — MAX_TOKENS sets the prompt budget (200000 in the review template; ocr scan uses 58888); conversation context is compressed to stay within this prompt budget. Model output is capped separately by MAX_COMPLETION_TOKENS (16384). A file whose diff alone exceeds ~80% of MAX_TOKENS is skipped before the LLM is called.
• Plan phase triggers on either of two thresholds — a group runs an extra risk-analysis phase before main review when its largest changed file reaches PLAN_MODE_LINE_THRESHOLD (default 50) or it holds 2+ files whose combined changed lines reach PLAN_MODE_GROUP_LINE_THRESHOLD (default 100). This adds latency but improves quality.
• Don't pass --audience human — it streams progress UI that pollutes output. Always use --audience agent.
• Comment language follows config — the language config controls review comment language, defaults to English, and accepts any language name (for example English or 中文).
• Avoid output truncation — Large review runs produce verbose output. Never pipe command output to tail or head as it drops review comments from earlier sections. Use --output <path> and read it in full; on older CLIs, follow the Output file guidance above.
• Resume an interrupted review — a failed or interrupted range/commit review can be continued with ocr review --resume <id> using the same --from/--to or --commit target (the id is printed as retry with: --resume <id> on failure, or find it with ocr session list). Workspace resume is not supported.

Validation

After the review completes, verify success by checking:

1. The command exited with code 0
2. Comments were generated (or "No comments generated" message appears)
3. Warnings (if any) are displayed in stderr

If errors occurred, check the stderr warnings for details about which files failed and why.

Troubleshooting

ocr: command not found

Install the CLI:

npm install -g @alibaba-group/open-code-review

unknown flag: --output

The CLI is older than v1.10.0. Do not continue the review with plain stdout. Ask the user whether to upgrade (npm i -g @alibaba-group/open-code-review@latest) and wait for the answer before proceeding. After the user confirms and the upgrade succeeds, rerun with --output.

ocr review fails with LLM connection error

Prompt the user to configure an LLM provider.

Interactive setup (recommended):

ocr config provider

Manual setup (alternative):

ocr config set llm.url https://api.anthropic.com/v1/messages
ocr config set llm.auth_token <api-key>
ocr config set llm.model claude-opus-4-6
ocr config set llm.use_anthropic true

Verify connectivity with ocr llm test. Stop here and ask the user to provide credentials — never invent or hardcode API keys.

References

• Full docs: https://github.com/alibaba/open-code-review
• NPM package: https://www.npmjs.com/package/@alibaba-group/open-code-review
• Issue tracker: https://github.com/alibaba/open-code-review/issues