The change profile is in preview and is turned on per workspace. Ask us to
enable it for yours.
What you see
Every full review’s summary gets one line:Change: bug fix · high risk (concurrency, stored data) · reaches a few other modules
Re-reviews of new commits that only look at what changed keep the line from the review they build on.
What it changes
- How deep the review goes. A high-risk change gets a thorough review. A low-risk one, such as a change to tests or docs only, can get a lighter review.
- Whether a second reviewer runs. In workspaces with adversarial reviews in
automode, the profile answers the gate’s risk, contract and security checks. See Model-assisted gate. - What the reviewer checks first. In workspaces where it’s turned on, the reviewer is told how the change was classified and gets a short checklist for up to three risk areas. For example, “every new or changed entry point checks who the caller is” for security, or “retries are idempotent” for concurrency. The reviewer is told to follow the code wherever it says otherwise.
Guard rails
The classification makes a review more careful only when it is fairly sure (three to one). It makes a review lighter only when it is very sure (nine to one), and never past what Otto’s own signals require:- A review is never lightened past a design or spec document, a migration that ships with code, a path your
otto.ymlreviews at Deep or Max, an open blocking finding from an earlier review, a change touching authentication, permissions or secrets, or a removed definition that other files still use. - When the classifier is unsure about a part, Otto’s own signals decide that part. When it is slow or unavailable, they decide all of it, and the line says it came from Otto’s rules.
otto.yml settings still apply. Review depth picks the review tier, and the adversarial gate settings still turn checks off, scope them to paths and set their thresholds. See Model-assisted gate for how thresholds combine with the profile.
What the classifier sees
A small classification model reads only the pull request’s metadata:- the changed file paths, their sizes and line counts, and the risk areas their paths suggest;
- the names of changed definitions and how many other files use them;
- counts of security-relevant patterns on the changed lines, such as new endpoints or removed guards;
- the title and description, with code blocks removed.
