release-please Silently Skipping or Under-Bumping a Release
release-please Silently Skipping or Under-Bumping a Release¶
Symptom: the release-please GitHub Actions job runs green after a Weekly:
Promote nightly to main merge, but either no release PR is opened, or the
release PR it opens proposes the wrong bump type (e.g. a patch bump when real
feat: commits were merged). The job log for the "skip" case ends with:
✔ Splitting 1 commits by path
❯ commits: 1
✔ Considering: 1 commits
✔ No user facing commits found since <last-release-sha> - skipping
This happens even when the promotion genuinely brought in feat:/fix:
commits — the commits are on main, they are not yet in any tag, but
release-please never counted them.
Root cause¶
release-please's commit-collection walk effectively drops commits once it reaches the previous release commit's SHA, and the commits it fetches from GitHub come back ordered by commit date, not strict first-parent topology. This means the walk stops the instant it reaches a commit at or before the release boundary's own commit date — regardless of whether older, structurally-unreleased commits exist elsewhere in the graph.
The failure condition: a commit's own committer date is older than the
date of whatever commit currently sits as the release boundary on main,
at the moment it finally lands via a real (non-squash) merge. Concretely,
here is how it happened on 2026-09-28:
feat/auth-rate-limit-1317merged intonightlyon 2026-09-25, carrying realfeat:/fix(security):commits dated that day.nightlysat un-promoted for several days (weekly cadence).- Meanwhile, an unrelated fix landed directly on
mainand was released asv0.42.1on the morning of 2026-09-28 — before that week's nightly promotion had run. - When the weekly promotion merged later that same day, its second-parent
commits (the real
nightlywork) were now chronologically older thanv0.42.1's own commit timestamp. release-please's date-ordered walk reached thev0.42.1boundary before it ever paged far enough back to see those older, still-unreleased commits, so it silently dropped them.
This is not about branch routing — the same stranding can happen to a
long-lived branch merged directly into main, if it sits long enough that
an interim release lands first. It is not a missing Conventional
Commits prefix either: a structurally identical prior promotion (with an
equally non-conventional merge title) correctly picked up a buried fix:
commit, because in that case no interim release had jumped ahead of it.
And it is not fixed by landing spacer commits on main afterward — a
spacer only pushes the boundary later, widening the date gap rather than
closing it (this was tried on 2026-09-28 via PR #1408 and only produced a
9-commit-late Set(1) truncation instead of a full skip — see PR #1410).
The two things that actually matter:
- How long does non-squashed work sit before it lands on
mainvia a real merge? The longer the gap, the bigger the window for an interim release to jump ahead of it. - Sequencing between releases and promotions. If a same-week
release-pleasecut is allowed to land onmainbefore that week's pending nightly promotion, anything still waiting innightlyis at risk of being dated older than the new boundary.
How to confirm you're hitting this¶
- Pull the run's log:
gh run view <run-id> --job <job-id> --log. Look at how many commits it reports considering (commits: N) versus how many real conventional commits actually exist in the range. - Verify real unreleased commits exist and check their dates against the boundary's date:
git log --format='%H|%cI|%s' <last-release-sha>..origin/main | grep -E '\|(feat|fix|perf|revert)(\(|:)'
git log -1 --format='%cI' <last-release-sha>
Any matching commit with a committer date older than the boundary's is
at risk of being silently dropped, even though
git merge-base --is-ancestor <sha> <last-release-sha> confirms it is
genuinely new (non-zero exit).
Recovery¶
Do not rely on spacer commits or repeated re-runs — the walk is deterministic on the current git graph and dates, so it will reproduce the same result. Instead, hand-correct the release-please PR directly:
- Check out the
release-please--branches--mainbranch. - Recompute the real commit set and correct bump type:
git log --format='%H|%s' <last-release-sha>..origin/main | grep -E '^\w+\|(feat|fix|perf|revert)(\(|:)'
feat: present → minor bump (this repo has bump-minor-pre-major: true);
only fix:/perf: → patch.
3. Edit .release-please-manifest.json to the corrected version and push a
fix: commit to that branch.
4. Rewrite the PR body/release notes to list the real commits (release-please
only touches the manifest here since skip-changelog: true — there is no
CHANGELOG.md to reconcile), matching the existing bullet format:
* <description> ([<short-sha>](.../commit/<sha>)), with (#N) references
in the subject converted to issue links and <scope>: rendered as
**scope:** prefix.
5. Merge the corrected PR.
To prevent recurrence structurally, see the sequencing fix tracked for
weekly-nightly-promotion.yml / release-please.yml (ensure nightly always
promotes before an interim release is allowed to cut that week).