Run a review

You start a review from the pull request's AI Review tab. You choose who reviews and whether the result stays in Limn or goes to GitHub when it is done. This page walks through one review from start to post.

#Before you start

• AI must be on for this machine and you must be signed in to Claude, Codex or both. If something is missing, the AI Review tab shows one line in place of the Run button that says what to do. See Settings → AI.
• The first time, Limn downloads the AI runtimes in the background. Signing in is the only step for you.

#Start a review

1. Open the pull request and choose the AI Review tab.
2. In Run a review, open the Model menu and pick who reviews. The menu lists each signed-in provider's models, most thorough first, for example Claude Opus 5.5 (most thorough). With both Claude and Codex signed in, Claude + Codex is last in the list and is the default.
3. Press Review only to keep the result in Limn. To post it when it finishes, open the caret beside the button and choose Review and post. The button remembers your last choice.

#While it runs

A status line says what the reviewer is doing (for example Fetching the diff, Cloning the worktree, Reviewing, Saving findings) and how long it has taken. Under it, a short log shows what the reviewer is reading. When it finishes, the tokens it used and its cost are behind the ⓘ beside the review's summary. Limn charges nothing.

Press Stop to end the run. You can leave the tab or open another pull request while it runs. A Claude + Codex review says which half is running, then Merging the two reviews… at the end.

#Read the findings

The finished review opens with its verdict, a fact line (the commit it read, for example On latest commit · abc1234, the model and the effort) and the summary. The findings follow, most serious first. Each finding names its file and line, says what is wrong and why, and may carry a suggested change.

On each finding you can:

• Reword it. Your wording is what posts; the reviewer's stays beside it. After a reword the button reads Edit wording.
• Ignore it, so it is left out of the review you post and out of AI Fix.
• Copy it as Markdown, or use Copy all for every finding you kept.
• Send to chat to ask about it in the Review chat. See Ask about a review.
• Show it in the Changes tab, or open the line on GitHub.
• Post it alone with Post as comment (on its line) or Post as PR comment (when it has no line). Once posted, the finding shows Posted with a link, and it cannot be posted twice.

On a pull request that was reviewed before, the earlier findings are in the same Findings list: still open ones under Open, the rest under Resolved or addressed, each with one line on how it was dealt with. See Follow-up reviews.

#Post your review

Post to GitHub sends one review: your summary as its top comment, with the verdict you pick, and every finding you have not ignored or already posted, each on its line.

1. Edit the summary box. After a Review only run it is filled with the reviewer's summary. It takes Markdown, and @ mentions a teammate. Save draft keeps it without posting.
2. Pick the Verdict: Comment, Request changes or Approve. The reviewer's verdict is chosen for you; you can change it.
3. Press Post to GitHub, then Confirm post.

#Run it again

After new commits, the header says how far behind the review is (for example 2 newer commits) and the Run button becomes the main action again, labelled Re-review only or Re-review and post. A re-review follows up the earlier findings rather than starting over.

If you run a review on a commit that already has one, Limn says This commit is already reviewed. and asks you to press Run anyway or Cancel.

Earlier reviews

Reviews on this PR lists every review: yours, automatic ones and your teammates' posted reviews, with the verdict, date and commit of each. Pick one to read it. Only the latest review can be edited and posted; an earlier one is shown as An earlier review · read-only.

#Effort

Settings → AI has one setting, Effort for reviews and AI Fix, for both providers: Default, Low, Medium, High, Extra high and Max. Higher effort reads more and finds more, and uses more of your subscription or costs more.

• It reaches every review (AI Review, the CI review, the story check and Check threads), AI Fix and Implement story.
• Default runs Claude at medium and Codex at high on its most thorough model.
• A quick review of a small change can run one level lower, but never below what Default would use unless you pick a lower level.
• Each review shows the effort it used beside its model, for example "Claude Opus 5.5 · high".