Code review without the comment spam.

A single pass over a diff finds the obvious things and misses the rest, because one reader with one prompt has one set of concerns. So a review runs several reviewers over the same change, each looking for something different (logic, security, reliability, tests, operability) and none of them aware of what the others found.

Then two more passes. A sweep reads every reviewer's output together and looks for what they all missed, which is the largest amount of context the pipeline ever assembles. And a judging pass goes through the candidate findings and throws out everything it cannot stand behind.

The judge is the part that decides whether the feature is useful. On the first real review it ran, it kept one finding out of six. A reviewer that surfaces six things when one is real has not saved you any time.

Depth is a choice you make per review, and it is a real trade rather than a slider with no meaning behind it. Quick runs one reviewer and skips the sweep and the judge. Standard (the default) runs several reviewers plus both passes. Deep adds more reviewers, reads a large pull request in chunks, and uses the strongest model your connected subscription has.

The numbers are measured, not aspirational. Quick is usually under ten minutes. Standard typically takes about half an hour. Deep can take an hour or more on a large pull request. We publish those because promising 'a few minutes' is how a feature that is working comes to feel broken.

Tick the findings that are worth acting on and Talyn sends an agent to make those changes and push a single commit to the branch. Not a suggestion per finding for you to click through: one commit, with the work done.

There is an automatic version, and it is off by default and deliberately narrow. It will only act on a finding the judge confirmed, at or above a severity floor that defaults to blockers only, and whose quoted anchor was verified against the actual file. Each of those bounds comes from the first real review we ran, where the one finding that survived judging quoted code that was not at the line it named.

If a review turns up a blocker (blocker severity, confirmed by the judging pass, on a review of the current head, not dismissed) the merge queue parks the pull request rather than landing it. Dismiss the finding, land a fix, or push a new commit and the queue picks it up again.

Nothing about this touches your branch protection rules. It is Talyn declining to press merge, not Talyn overriding GitHub.