Skip to content

PR Reviews

The review flow is separate from task dispatches, but it follows the same basic idea: the local WebUI orchestrates the work and the remote runner does the GitHub-facing part.

Make sure you already saved:

  • a shell prelude in Settings
  • a Main GitHub user in Runner setup

Without those, the review flow stays disabled.

In the Reviews page:

  1. click Request a review
  2. paste a full GitHub pull request URL
  3. optionally add one-off instructions
  4. choose Codex or Claude for this review

The agent will post the review directly on GitHub and the WebUI will keep the local run history.

When the pull request changes, ask for a re-review instead of creating a separate review thread. track keeps the previous review context and records each run so you can see how the review evolved.

If you saved a default review prompt in Settings, it is appended to every manual review request before the one-off instructions you type for that specific run.

You can think of it as an AGENTS.md for reviews specifically: one shared prompt that applies to every review request unless you override it with extra instructions for that run.

That is a good place for standing guidance such as:

  • prioritize bugs and regressions
  • call out risky behavior changes
  • prefer high-signal findings over style commentary