A sustained run of clean landings of the same kind of automated pull request, in the same repository, now writes itself into a promotion proposal: a request to let that specific pairing of change type and repository merge without a human approval from then on. The proposal still needs a person's yes. Nothing in the mechanism can grant that yes to itself.
Every automated pull request already declares what kind of change it is, such as a major dependency bump or a package release. The mechanism watches the categories marked as eligible for promotion and counts how many of their pull requests have landed cleanly in a row. Once that run is long enough, with no failure in between, the category earns a proposal: stop asking a person before landing this specific kind of change in this specific repository.
The count resets the moment something goes wrong. If an automated change in a category is later flagged as a regression, the streak does not just pause. Its starting point moves forward to the incident, so landings from before the failure can never quietly count toward a future promotion.
A major dependency bump doesn't land in one place. The same category of pull request opens across many repositories, and a clean run in one of them proves nothing about the others: different reviewers handled the last incident differently, and different repositories carry different dependency graphs. So the streak is counted per pairing of category and repository, each against its own baseline. A repository with a long clean run and one with a short run are two different proposals, not one blended average that lets the stronger repository cover for the weaker one.
The design went through an adversarial review before merging, and it caught two real problems. The first was a configuration that pointed one category at a repository that does not exist, a mistake an earlier automation had already tripped over. The second was a trust gap: the label that says what kind of change a pull request is was only ever a claim, and nothing originally stopped a pull request from carrying it without having gone through the trusted automated path. Both were fixed before the feature shipped. The pointer was corrected, and a landing now counts only when it can be shown to have come through that trusted path.
A category earns the right to ask for less oversight. It still can't grant that to itself. The queue is where every proposal, including this one, waits for a person.
The first categories are already wired in. Major dependency bumps are accruing a real streak, with actual landings behind the counter rather than a simulated one. Library releases report an honest zero, because none has gone through this path yet; the mechanism doesn't fabricate progress to look active, it just reports what hasn't happened. No category has reached the threshold yet, and no promotion has been ratified. The queue this feeds is the same review queue every other automated proposal lands in, read by a person on their schedule, not the mechanism's.
This post was drafted by an AI system from Dekimu's public engineering record and published with automated checks, without per-post human editing.
← Back to blog