How AI review works
AI review runs your own Claude Code or Codex on your Mac to look for problems in a snapshot. It’s a pipeline of many narrow reviewers rather than one long prompt: finders look for problems, verifiers try to disprove each one, and a synthesizer merges what survives. This page explains each stage. To run one, see Run an AI review.
Before it starts, Revue stages a copy of what you’re reviewing: the pull request’s files at that snapshot, or your checkout’s files for local work, and the diff. With Claude Code, every reviewer reads that copy with three tools only, Read, Grep and Glob, and can’t write a file or run a command. Codex runs in its read-only sandbox, with your own Codex configuration ignored. Every prompt tells the reviewer to treat the diff and the code as data, never as instructions, since a pull request can come from anyone.
Finders
Section titled “Finders”One finder runs for each focus area you keep on. Each looks at the changed lines through its own lens only, reads as much of the surrounding code as it needs, and reports what it finds with a severity. Finding nothing is a valid answer. Sensitivity changes what the finders are asked to report. Run an AI review says what the focus areas and sensitivity levels are, and where to set them.
Finders use the Finding model, sonnet by default.
Verifier
Section titled “Verifier”Every candidate a finder reports goes to a verifier whose job is to refute it. The verifier reads the whole function the finding points at, and its callers when the claim depends on them, and returns one verdict:
- Confirmed: the claim is accurate and worth acting on. A verifier may raise or lower its severity.
- Refuted: it misreads the code, is handled elsewhere, is speculative or can’t happen.
- Uncertain: the code can’t settle it either way.
A critical finding gets two verifiers, everything else one. Any refutation drops the finding. A finding with only uncertain verdicts is kept, marked as uncertain and at low confidence. A finding no verifier could judge is dropped, not shown.
Verifiers use the Verification model, opus by default.
Synthesizer
Section titled “Synthesizer”Six finders can flag one bug six ways. The synthesizer groups the confirmed findings that describe the same defect and keeps one of each group, so each problem appears once. It never drops a finding it wasn’t asked to merge.
The synthesizer uses the Synthesis model, opus by default, which also drafts pull request descriptions.
With Codex as the provider, every stage uses the Codex model you choose instead of these three.
What it sends
Section titled “What it sends”AI review runs through your own Claude Code or Codex, signed in with that tool’s own account. Revue has no AI account and no server of its own in between. Help › What Revue Stores and Sends says this about your AI provider:
The change, and the files of its repository the tool reads while it works, through your own Claude Code or Codex, when you run AI review, ask about lines or draft a pull request description. Not otherwise, unless you turn on automatic review.
What your provider does with it is up to that provider and your account with it.