Skip to content

Run an AI review

AI review reads a snapshot and points out problems before you or your reviewers spend time on them. It runs on your Mac with the Claude Code or Codex you already use. Use it on a pull request you’re reviewing, or on your own branch before you push.

  • Claude Code or Codex, installed on this Mac and signed in. Revue uses that tool’s own account, so there’s no Revue account to set up.
  • The provider chosen in Settings › AI review. Claude Code is the default. If the tool isn’t on your PATH, set Claude executable or Codex executable there.
  • Enable AI left on in the same section.

AI review doesn’t need the Claude Code plugin. Only sending comments to a session needs it.

Settings › AI review with Claude Code chosen as the provider and found installed, Enable AI on, Balanced sensitivity and all six focus areas selectedSettings › AI review with Claude Code chosen as the provider and found installed, Enable AI on, Balanced sensitivity and all six focus areas selected

Open the pull request or branch, choose the snapshot you want reviewed, then choose Review › Run AI Review, or AI review in the context bar’s actions menu. While it runs, the same item reads Stop AI review. When it’s done it shows how many findings there are, and if it failed it offers Retry AI review.

A review reads the snapshot you have open. On a range, it reviews what changed between the two snapshots.

Each finding is a card under the lines it’s about, marked Revue AI, with a severity, what’s wrong and, often, a suggested change.

A Revue AI finding card with its title, an explanation, a suggested change and the Apply fix, Ask, Dismiss and Send to Claude buttonsA Revue AI finding card with its title, an explanation, a suggested change and the Apply fix, Ask, Dismiss and Send to Claude buttons
ActionKey
Next findingF8
Previous finding⇧F8

To list findings under each file’s header instead of between the lines, turn off Show findings inline in Settings › AI review.

Action Key What it does
Apply fix F On a branch checked out on this Mac, writes the suggested change into your working file. It commits nothing, and Undo puts the file back.
Suggest on GitHub On a pull request, adds the suggested change to your pending review as a suggestion.
Ask A Opens Ask AI on the finding, to ask why or how.
Dismiss D Hides the finding. It folds in place with Undo for a moment first. Dismissing an AI review finding stays on this Mac.
Send to Claude On a local branch with a Claude Code or Codex session, hands the finding to the agent.

Review › Add Finding as Comment turns a finding into a comment of your own, and Review › Show Dismissed Findings brings dismissed ones back into view.

Sensitivity sets how much a review reports. Quiet asks for consequential, high-confidence bugs only. Thorough also looks at edge cases and less certain defects. Balanced sits between them.

Focus areas choose which specialist reviewers run, one for each area. Turn off the ones you don’t want, such as Tests on a branch that has none. Keep at least one on.

SettingWhereOptionsDefault
SensitivitySettings › AI reviewQuiet, Balanced, ThoroughBalanced
Focus areasSettings › AI reviewCorrectness, Error handling, Security, Performance, Tests, ReuseCorrectness, Error handling, Security, Performance, Tests, Reuse

How AI review works explains what each specialist and stage does.

With this on, Revue starts a review on its own each time a new recorded snapshot arrives on the layer you have open. It’s off by default, because each review uses your provider’s account.

SettingWhereOptionsDefault
Review new snapshots automaticallySettings › AI reviewOn or offOff

A review runs many short turns of your own Claude Code or Codex, and any usage or charges are on that tool’s account. Revue never pays for or meters it. These settings bound a review:

SettingWhereOptionsDefault
Per-turn timeoutSettings › AI review15 s to 1800 s, in steps of 15180 s
Concurrent review turnsSettings › AI review1 to 16, in steps of 18
Finding modelSettings › AI reviewsonnet
Verification modelSettings › AI reviewopus
Synthesis modelSettings › AI reviewopus
Codex modelSettings › AI reviewNone

A turn that runs past the timeout fails with AI-002, and an account with no credit left fails with AI-003.