Cursor
Review what Cursor changed.
Cursor edits in your editor, so you saw some of it happen. That is the trap: partial familiarity feels like review, and the files you were not looking at are the ones nothing has checked.
You watched some of it
Accepting edits as they appear is a form of review, and a good one for the edits you actually read. The problem is what it does to your model of the change as a whole: a session that produced forty edits and showed you eleven leaves you feeling like the change has been seen. It has not. Twenty-nine edits went in with no reader, and they are not marked.
What to check first
Sort the changed files by whether you remember them.
This is uncomfortable and it is the highest-value five minutes available to you. The files you cannot recall seeing are the unreviewed change; everything else already had a reader. Most people find the list is longer than they expected.
Check what the accepted edits reach.
Reading an edit tells you the edit is fine. It does not tell you what else now behaves differently — callers, routes and jobs whose own source did not change and so appear nowhere in the diff.
Re-read anything accepted quickly in a run of accepts.
Acceptance rate goes up as a session goes on. The edits late in a long session got the least attention and are disproportionately where a signature quietly changed.
Questions
I already reviewed the edits as they happened. What does this add?
Two things you cannot get from reading edits one at a time: the shape of the change as a whole — which components moved, what is new, what depends on what now — and its reach, meaning the code affected by the change whose own source was never edited and so never appeared in front of you.
Does it work on a branch, or only a pull request?
Both. A pull request by number, or any two refs. With no base it uses the merge-base with your default branch and says so in the report.
Does it matter whether I used tab-completion, chat, or agent mode?
Not to the analysis. Cutplane reads the code, so the mode that produced it changes nothing about what the parser finds. It changes what you should be suspicious of, which is what the list above is for.
The method behind all of this is in how to review a pull request your AI agent wrote, and a real report on a public pull request is the output on something you can go and check.