Home / Everyday moves

Reviewing code and driving git in natural language

/code-review launches a local review with effort levels from low to max, and ultra is not one of them: it is a distinct facility, a multi-agent review in the cloud, billed separately. Git is then driven in plain language, with the same permission safety net as commands typed by hand.

A local code review is launched with /code-review and five effort levels, low, medium, high, xhigh, max. Ultra is not part of that scale: it escalates to a multi-agent review run in the cloud, billed separately from the five local levels included in normal plan usage.

Five local levels, one confidence report

The low and medium levels only report findings the review is most confident about, while high, xhigh, and max widen the coverage at the cost of including less certain findings. Match the effort to the risk of the change: too low lets a real bug through, too high drowns a small fix under uncertain remarks. The level is typed straight after the command, with no flag in front of it, everything that follows /code-review without the word ultra being read as the target of the review rather than as an option.

/code-review medium
# rapporte les constats les plus sûrs sur le diff en cours
# /code-review max élargirait la couverture,
# au prix d'inclure des constats moins certains

Ultra runs in the cloud

The ultra command triggers a multi-agent review in the cloud, distinct from the five local levels and billed independently: a Pro or Max account is granted three free runs allocated once, none on Team or Enterprise, beyond that the cost typically ranges from five to twenty five dollars in usage credits depending on the size of the change analysed, with confirmation requested before every interactive launch. A third facility exists as well, reserved for Team and Enterprise plans: a GitHub app service that automatically comments on every pull request, billed separately at an average of fifteen to twenty five dollars per review.

Git is driven without memorising commands

Asking in plain language for the repository status, the commit message, or the creation of a pull request actually triggers the underlying commands, gh pr create or glab mr create depending on the platform, without the user needing to know their exact syntax. The safety net stays the same as with these commands typed by hand: forcing a push or deleting a branch remains subject to the permission mode chosen for the session, exactly like any other git command. The review effort chosen above applies before this git step, just as a regression test precedes the closing of a bug.

What changed in October 2026

Code Review, the automatic review of pull requests, left preview on 15 October 2026 for Team and Enterprise organizations: an administrator picks Quick, Standard (the default) or Deep for each repository, and reviews run on their own on new pull requests. On the local command side, /code-review at medium effort also reports cleanup and CLAUDE.md conventions findings on models without tuned review settings, including Opus 5.5 and Sonnet 5.5, and --max-findings sets how many findings are reported. Finally, claude ultrareview now requires workspace trust: it asks on a terminal, otherwise it refuses, and a script passes --trust-workspace.