Run reviews

After creating your profiles, choose the reviewer that matches your input:

Review a project

oar run ./profiles/code-reviewer \
  --task review-repository \
  --input ./my-project \
  --output ./code-review.json

The reviewer looks for concrete issues in correctness, security, complexity, testing, and usability. Its instructions keep recommendations appropriate to the project's purpose and guard against overengineering and scope creep.

--input accepts a directory; it does not require Git metadata. OAR uploads a copy and makes it the agent's working directory. It does not select a PR diff or synchronize changes back to your machine. Tests run only if the agent can execute them with the tools, dependencies, and permissions available in the sandbox; the review should state what it could verify.

Review a document

oar run ./profiles/technical-writing-reviewer \
  --task review-document \
  --input ./design-notes.txt \
  --prompt-var context="Audience: engineers evaluating the proposed design." \
  --output ./writing-review.json

The reviewer assesses accuracy, clarity, completeness, organization, audience fit, and usefulness. Use readable text such as Markdown or .txt. OAR uploads the file as provided; it does not convert PDFs or Word documents to text.

Set focus and context

Both reviewers accept two optional prompt variables:

Add either or both to a run:

--prompt-var focus="src/auth and tests/auth" \
--prompt-var context="A small internal tool; keep recommendations proportionate."

Without these options, the reviewer considers the whole input. Focus guides attention; it does not restrict file access or reduce the upload. To limit what you share, supply a smaller input directory.

To include supporting files, use --upload and tell the reviewer where to find them in the sandbox. Upload paths explains how destinations work.

Understand the result

Both reviewers write JSON with this structure:

review.json
├── verdict           pass | needs_changes | inconclusive
├── summary           Main conclusion
├── criterion_scores  A score and explanation for each criterion
├── overall_score     Rounded average of the criterion scores
├── findings          Specific issues and recommended fixes
├── strengths         What works well
└── limitations       What the reviewer could not establish

Each criterion is scored from 0 to 100, where 100 is best:

Code review uses five criteria: correctness, robustness and security, maintainability and complexity, tests and verification, and usability and integration. Writing review uses six: accuracy, clarity, completeness, structure, audience fit, and actionability. The profile's skills/ directory contains the full rubric; schemas/review.json defines the JSON fields.

The packaged reviewers require a score of at least 90 and no findings for pass. A reported finding means needs_changes. Missing context can lead to inconclusive, with the uncertainty recorded in limitations.

Read the evidence and limitations alongside the score. OAR checks that the result matches the schema; it does not establish that the review is correct. The rubric tells the agent how to calculate the average; OAR's schema validation does not recalculate it. CI behavior explains why a valid review with findings still returns exit code 0.