Repository navigation
Description
What was observed
This repository has five release guards, and every one of them starts from a tag:
| guard | the question it puts |
|---|---|
scripts/check-release-tag.py |
does this tag name the version the tagged tree declares? |
scripts/check-release-published.py |
is this tag published — run, release, draft/prerelease, assets? |
scripts/check-release-orphan-tags.py |
is any tag on the repository published-less? |
scripts/check-release-artifact-version.py |
does this artifact declare the claimed version? |
scripts/check-release-prompt.py |
does this artifact carry the tree's prompt? |
Not one of them asks whether a tag should exist. Checked by reading each guard's argument list: the first two and the last two take a tag or an artifact as an argument, and the fleet view enumerates tags that are already there. A tree that declares a version nothing has tagged is therefore invisible to all five — and that is not a hypothetical state.
Measured on this checkout at master d5b17fea:
emrg/__init__.py __version__ = "0.3.59"
git describe --tags v0.3.58
git rev-list --count v0.3.58..master 9
So nine merged commits, including two that change what every installation reads (#2944's submission-pace arithmetic and #2948's corrected index-guard failure line), are on master and reach no host, because a host only gets them from a published release and nothing published declares 0.3.59.
Why it matters
The condition is noticed today by a person or a cycle happening to look. The last two releases were each opened by hand, and each issue had to re-derive the same facts from scratch: that master is ahead of the newest tag, by how many commits, and that a bump is what closes the gap. That is a reading re-derived per release, which is the same argument that produced check-release-artifact-version.py and check-release-prompt.py in the first place.
It is also the one failure in this chain that gets worse with time rather than loud: every further merge adds commits a host cannot install, and nothing turns red. check-release-orphan-tags.py exists because three published-less tags were once found by hand; this is the same shape of question one step earlier, where no tag exists to ask about.
What the fix must satisfy
- One command answers "does this tree declare a version no tag carries?" — reading the declaration from
emrg/__init__.py(the same file and anchored shapecheck-release-tag.pyandtests/test_version_sync.pyuse) and the tag list from the repository, with no network. - The answer names the three facts a release request needs: the declared version, the newest tag, and how many commits sit between them. A verdict without those numbers cannot be acted on.
- The family's exit contract, unchanged:
0nothing is owed,1a release is owed (both versions printed),2nothing could be measured — no declaration, no tags, a git failure, an unparsable tag. Exit 2 is never a pass, and a guard that cannot read its subject must not report agreement. - Version comparison is by numeric parts, not by string or by tag date:
0.3.10is newer than0.3.9, and a tag's commit date says nothing about which version is later. A tag that does not parse as three dotted numbers is reported and skipped, never sorted in as though it were one. - It is a host-side screen, not a CI gate. Wired into CI it would be red on every unreleased master, which turns the release cadence into a permanently failing check and teaches readers to ignore it. The acceptance is a command a host runs before cutting a release, in the same role
bump-version.py --checkandcheck-release-tag.pyalready play.
Acceptance
- On this checkout the guard reports exit
1, naming0.3.59,v0.3.58and the commit count between them. - On a tree whose declared version is the newest tag's version it reports exit
0and says nothing is owed. - With no version declaration, no tags at all, or a
gitthat cannot run, it reports exit2and never0. - A tag list containing
v0.3.9andv0.3.10picksv0.3.10as the newest. - No existing guard is changed and no CI workflow is touched.
Activity
pm25coder commentedon Nov 18, 2026
Handled by #2958