Claude and Codex together

Two different models miss different things. When you are signed in to both Claude and Codex, Limn can run both on the same commit and give you one review. In Limn's own tests, one Claude review and one Codex review, merged, found about four in five known issues; either one alone found about half.

#When it runs

• Reviews you start. In the Model menu of Run a review, Claude + Codex is listed last while both providers can review, and it is the default. Pick a single model from the menu to review with that one only.
• Automatic reviews. Settings → AI review, Auto reviews use: Claude + Codex (the default), or your default provider only. If only one is signed in when a review starts, that one reviews alone and the setting says Only one is signed in, so it reviews alone.

#How the two are merged

1. Claude and Codex start at the same time, each on its own. Neither sees the other's findings.
2. When both have finished, the tab says Merging the two reviews…. Limn groups the findings that make the same point and keeps one of each group. A short summary covers both.
3. Limn, not a model, decides which finding of a group is kept. The more severe one wins, so a blocker is never merged into a nit. A finding already on GitHub is always the one kept.
4. Each finding says who raised it: Claude, Codex, or Claude + Codex when both did. When posted, both halves go out as one review.

If the merge step itself fails, Limn still gives you one review: it keeps every finding except the clear repeats, and joins the two summaries.

#One verdict

A Claude + Codex review gives one verdict. You choose how strict it is in Settings → AI review, When Claude and Codex disagree:

Either way, if one asks for changes, Limn does not approve. A model may hold back an approval because it could not check something, such as a binary file in the diff; strict treats that as a reason not to approve, relaxed does not. Open blockers and warnings count the same whichever model found them. See Verdicts and posting to GitHub.

#When one of them fails

A Claude + Codex review always gives you a result while at least one half finishes. If one half fails, is stopped by its usage limit, or is never started (for example because a teammate's Limn is already reviewing with that model), the half that finished is the review. The tab says Only one of the two reviews finished, so this is that review. and names why the other did not finish.

Before anything is posted, the half that finished also checks the other model's earlier findings that are still open. This short check is labelled, for example, Checked Codex's findings, with a line such as "Checked Codex's 3 open findings: 1 dealt with, 2 still open."

• An earlier blocker that is still open makes the review request changes.
• An earlier warning that is still open, or one that could not be checked, stops an approval.
• The review says why it did not approve, for example "Not approved: 2 earlier Codex points are still open." or "Not approved: 1 earlier Codex point was not checked."
• The checking model never argues with the other model's finding. It may find it fixed, and then Limn may resolve its thread.

#When a model is at its usage limit

Until the limit resets, Limn uses the other model wherever it can. Replies to the locked model's findings are judged by the other model. Check threads, CI and ticket reviews and the review chat run on the other model. A Claude + Codex review you start runs the other model alone, and the menu does not offer the pair as its first choice. A model you pick by name still runs on that model.

#Usage and stopping

• Both runs count towards each provider's own usage.
• Stop stops both halves. A half that has not started yet does not start.
• An automatic Claude + Codex review asks the spend budget for each half. When the budget is used up, the first half runs alone. See AI usage and spend.
• Turning off design diagrams (Draw a design diagram after each review) saves about a third of a provider's review usage. See Design diagrams.