Both PR-Agent findings were confirmed. A tag at the merge commit now counts only when it is exactly v<manifest version>; foreign CI or annotation tags are logged and ignored. Pending release PRs are fetched through the autorelease: pending label filter instead of paging every closed PR, so a concurrent close cannot shift a page and hide one.
Also fail loudly on a merged PR without merge_commit_sha or a manifest without a version, instead of producing vundefined.
Rebased on main after the separate audit fix (e0c69c1); the dependency overrides and lockfile changes from the earlier local commit were dropped. Verified on the rebased tree: lint, typecheck, build, the exact audit command, actionlint and a live dry run.
8.6 KiB
type, title, description, tags, status, generated, verified, sources
| type | title | description | tags | status | generated | verified | sources | ||||||||||||||||||||||||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Decision | Release engineering | Releases are produced by release-please in manifest mode, driven by conventional commits, with the message format enforced at commit time by commitizen. |
|
draft |
|
|
|
Decision
On every push to main, release-please (manifest mode) collects
conventional commits, derives the next version, and opens a release PR
that bumps package.json, updates CHANGELOG.md and moves the manifest
forward. Merging that PR creates the release. The message format that
feeds this machine is enforced locally by the commitizen pre-commit hook.
Why
- Manifest mode (config JSON + manifest JSON) over inline
release-typein the workflow: the configuration becomes part of the repository, reviewable in PRs, instead of buried in workflow YAML. - Plain tags (
v1.0.0,include-component-in-tag: false): single-package repository; component-prefixed tags would be noise. - A
versionfield inpackage.json: thenoderelease type reads it; without the field the release machinery cannot track anything. - Enforcement at commit time: a non-conventional message is not an error release-please reports — it just silently produces no release. Commitizen turns that silent failure into a blocked commit.
How a release happens, end to end
- Work lands on branches as conventional commits (
feat:,fix:,chore:, …), enforced by the local hook. - Merge to
main→ release-please opens/updates a release PR (chore(main): release …). - Merge the release PR → version bumped, changelog written, tag
vX.Y.Zcreated.
The label is load-bearing (2026-10-24)
Steps 2 and 3 are connected by a label, not by the merge alone:
- release-please creates the release PR with
autorelease: pending; - on the next run it lists merged PRs and keeps only those carrying that label; those are the releases it tags.
Gitea resolves label names in POST /issues/{index}/labels with
GetLabelIDsInRepoByNames, which is a SQL IN filter: names that do not
exist are dropped without an error. A repository with no labels
therefore accepts the call, returns 200, and attaches nothing.
That is exactly what happened here: PRs #20 and #21 were merged as release PRs with no label at all, so every run found zero merged release PRs, created zero tags, and opened the next release PR instead — while reporting success. Three release PRs, no tags.
The workflow now creates both labels (autorelease: pending,
autorelease: tagged) before the action runs. Idempotent: it reads the
label list and creates only what is missing.
The head branch must survive the merge (2026-10-27)
The label fix was necessary but not sufficient. PR #22 carried
autorelease: pending when it was merged, and the run after it still
created no tag. From its log (task 475):
❯ Found pull request #22: 'chore: release main'
✖ Bad branch name: refs/pull/22/head
⚠ Expected 1 releases, only found 0
Gitea rewrites a merged PR's head.ref to refs/pull/<n>/head once the
head branch is deleted; the branch name survives only in head.label.
The Gitea build of release-please reads head.ref
(BranchName.parse(mergedPullRequest.headBranchName), bundled at
dist/index.js:15288) and abandons the release when it is not a release
branch. GitHub keeps the branch name there, which is why the plugin's
assumption holds on GitHub and not here.
The repository deleted the head branch of every merged PR
(default_delete_branch_after_merge: true). It is false now, and a
release PR must be merged with its branch kept. The mechanism was
confirmed by comparison, not inference: while fix/release_please still
existed, PR #23 reported head.ref = fix/release_please; every PR whose
branch was gone reported refs/pull/<n>/head.
A pending label outlives the branch and blocks every later release (2026-10-31)
Flipping default_delete_branch_after_merge to false only protects
branches deleted after the change. PR #22's branch was already gone, so
its head.ref stayed refs/pull/22/head and its autorelease: pending
label stayed attached. Run 143 (task 475, 2026-10-27, the run immediately
after #22 merged) shows the full deadlock:
❯ Found pull request #22: 'chore: release main'
✖ Bad branch name: refs/pull/22/head
...
✔ Building candidate release pull request for path: .
❯ Found pull request #22: 'chore: release main'
⚠ There are untagged, merged release PRs outstanding - aborting
::set-output:: prs_created=false
::set-output:: releases_created=false
The first block fails the tagging phase. The second finds the same
still-pending PR and aborts the propose phase, so no new release PR is
even opened. Every later run (145, 148, 151, 153, 162, 164, 167) has the
same shape. The hand-created v1.4.0 tag restores the baseline (found: v1.4.0) but does not clear the label, so it does not break the
deadlock. This is the current reason the pipeline is still healthy-looking
and tagless.
The workflow now reconciles that state before the action runs
(scripts/reconcile-release-please.mjs):
- merged +
autorelease: pending+ a tag already at the merge commit -> replace the label withautorelease: tagged; - merged +
autorelease: pending+ head branch gone + no tag -> create the tag from the version recorded in the manifest at the merge commit (the exact tag release-please would have made), then swap the label; - merged +
autorelease: pending+ head branch still present -> leave it; release-please tags it in the same run.
The step is idempotent and runs on every push to main and every manual
dispatch, so a deleted branch cannot silently poison the release pipeline
again. #20 and #21 are not labelled pending, so they do not block; their
tags remain the one-time manual recovery in the table below.
Two details keep the reconciliation from suppressing a real release:
- pending PRs are fetched through the
autorelease: pendinglabel filter, not by paging every closed PR, so concurrent PR closes cannot shift a page and hide one; - a tag at the merge commit counts only when its name is exactly
v<manifest version>. A CI or annotation tag on the same commit is logged and ignored, and the release tag is still created.
Recovering the versions that never got a tag
The label alone cannot recover them — #20, #21 and #22 all lost their head branch, so release-please skips them for the reason above. Creating those tags is a one-time manual step:
| Version | Merge commit |
|---|---|
v1.1.0 |
67cf232 (#12) |
v1.2.0 |
4e8aa28 (#20) |
v1.3.0 |
08eb042 (#21) |
v1.4.0 |
0c937e6 (#22) |
v1.0.0 is the bootstrap changelog entry from the configuration commit
c89747e (2026-10-18); no release PR produced it.
The manifest moved to 1.4.0 without any tag existing, so until a tag is present release-please has no baseline: it reported "No latest release found for path: ., component: , but a previous version (1.4.0) was specified in the manifest" and would compute the next release from the whole history.
Not verified
- That the release phase tags correctly once the head branch survives a merge. The next release cycle proves it.
- That the reconcile step's tag-creation fallback works with a live
token. Only the tag-already-exists path was exercised, in dry-run,
against the live PR #22. The fallback is written from the same Gitea
API shapes (
POST /tags,PUT /issues/{index}/labels) the rest of the workflow uses. - That
RELEASE_PLEASE_TOKEN_CV_SITE_NUXTcarries thewrite:issuescope the same wayPR_GITEA_TOKENneeded it; the previous failure was an immediate API rejection consistent with a scope problem.
Revisit when
- A second package appears in the repository (then
include-component-in-tagandseparate-pull-requestsmatter). - A release must trigger a deploy — the workflow gains a step after
release_createdis set. - The label step breaks: if Gitea ever starts rejecting unknown label
names, the
Ensure release-please labels existstep will fail loudly instead of silently, which is the point.