This guide introduces Flow code reviews. You will learn:

  • What a code review checks, and what it reports
  • The two places a code review runs
  • Which agents run it

A Flow code review reads your changes through lenses. A lens is one focus, such as access control, query cost or migration safety, written by the team behind AdonisJS. Each lens runs in its own sub-agent of your coding agent, over the changed files it covers. See Lenses.

Each finding has a severity, the lenses that found it, the problem, its impact, the fix and the lines of code it is about. The severity is critical, major, minor or nit, from what the problem can cause and how likely it is.

A review reports the most severe findings first. You can accept a finding, so that the next reviews stop reporting it. See Accepting a finding.

A code review runs on your computer, or on each pull request on GitHub. Both read the same lenses and the same .flow/review.yml.

  • On your computer. npx @adonisplus/flow review starts a session of your coding agent that reviews your branch. It needs no setup, and it can write the findings as an HTML report. See Local reviews.
  • On GitHub. Each pull request gets a review in your GitHub Actions. The findings are comments on the changed lines, and the "Flow review" check fails while a critical or major finding is open. It needs the GitHub App, a workflow and a model credential. See GitHub setup.

A local review runs on the agent of flow.lock: Claude Code, Codex, Cursor or OpenCode. A review on GitHub runs on Claude Code or Codex, with your own credential. You pay the vendor of the agent for the tokens. Flow never receives the credential.

Local reviews

Review your branch on your computer, and write an HTML report.

GitHub setup

Turn on code reviews for the pull requests of a repository.