# 03 — Operate
Cutting a release from the trunk, end to end. The binding supplies the concrete commands; the sequence is the same everywhere. The command form of this chapter is the release runbook, `rk guide release`.
## The sequence
1. Land the work through squash-merged pull requests, each with its release intent captured in the squash title on a change the release bot attributes to the package, and the local check suite green. Which changes the bot counts is [the binding's](../bindings/README.md) answer.
2. Read the release request. The bot keeps one pull request open against `master`. Every merged pull request that changes a file the release packs refreshes it, so the proposed version and changelog describe the tip used by its latest refresh.
3. Decide whether this release ships now. No landed job arms or merges the request, so a release waits for the operator's order and a hold takes no action. A changelog that reads wrong is answered first where it is generated, in the squash titles and bodies the landed gates already judge, and in which changes the bot counts at all: an entry that never appeared belonged to a commit the bot attributed to no package. A hand correction on the request's branch is the last step before the merge, in a worktree where the project's mode requires one, because the bot refreshes the request as work lands — by force-push while only its own commits sit on the branch, or by closing it and opening a fresh one — and a correction made early is dropped while a correction made last rides into the merge.
4. Order the merge with `rk release merge --apply`, once its check is green and it carries the trunk's tip. The verb dispatches the landed merge job, which runs as the release App, proves that nothing moved since the operator read it, and merges with the head pinned; a release line's request is merged by a person. That merge is the release: automation tags `v<version>` on the push that lands the bump and runs the binding's publish path — a registry publish over OIDC, a dedicated artifact build on the tag, or a tarball attached to the release page, as the binding declares.
5. Wait for the artifact build before verifying anything. Where the binding runs a dedicated artifact builder, it creates the release page in its final job, after every platform builds, so for some minutes after the merge there is no release to look at and a check fails on timing rather than on the release. A binding whose publish stage carries the artifacts has nothing separate to wait for.
6. Verify what the binding ships. Universally: the tag and `master` resolve to the same commit. Then per binding: a registry serves the new version, a release page carries its artifacts and is not empty, an installed binary reports the new version, and the provenance verifies — the run that built each release artifact signed it, proven with the pair's own verifier — each check applying only where the binding declares that surface. A pair that honestly declares no provenance has nothing to verify there, and a check that failed on it would be wrong; an invariant with no verification step is a comment, which is why the release is not done until this one passes.
A request merges only when it carries the trunk's tip. A release request that falls behind waits: its contents are computed, so merging them unchanged could ship a version computed against a trunk that moved. The bot's own refresh rebuilds the request on the new tip and recomputes the version and changelog; once its checks pass, it can merge at the recomputed version. An ordinary GitHub request the forge reports as `BEHIND`, whose own description of that state is "The head ref is out of date", needs `gh pr update-branch`, then one more run of every workflow that triggers on the request before it can merge, which is the whole cost strictness adds and is paid once per stale attempt rather than once per trunk commit. The refusal reads differently in the two places an operator meets it, and the second is the one that misleads: the CLI says "the head branch is not up to date with the base branch", while the merge endpoint returns 405 with "Repository rule violations found" and then counts the required checks as expected rather than naming the staleness, because a check that passed on the stale head does not count toward a merge the forge has not tested. A request whose checks are all green can therefore report every check as missing, and the branch state is what tells the operator why. A further trunk move can require another update. The trunk moving alone dispatches no request workflow. [The drawn model](./06-release-from-trunk.md#why-a-stale-release-waits) owns the race, the repair, and the daily path.
## The two steps that get missed
Step 3 is where a release is stopped, and it is the only point a changelog correction reaches the release it describes: compare the entry against the commit range since the previous tag, correct, and merge with nothing landing in between. A stop after the merge is a withdrawal rather than an abandon, and [recovery](./04-recovery.md) owns it.
Step 5 is why a check run straight after the merge reports the release as missing under a dedicated artifact builder. The tag lands within about a minute; the release page arrives only when the slowest platform build finishes.
## When an older line needs the fix
This sequence releases the trunk's tip; a patch-only release for users on an older version is a different path. [Branch for release](./07-branch-for-release.md) owns it, and [recovery](./04-recovery.md) carries the entry point for a line whose branch does not exist yet.