You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Publish weekly Sunday minor releases alongside daily patch previews #540
The maintainer requested recurring Sunday releases starting with the v0.3.0 series, alongside daily patch prereleases.
Keep the existing daily 02:00 Asia/Shanghai preview run. Add a Sunday 08:00 Asia/Shanghai stable run. Both use the canonical release bot and the existing build, checksum, tag and publication checks.
Daily runs increment the patch and leave GitHub Latest on the stable release. Weekly runs increment the minor, reset the patch to zero and mark the release Latest.
Compare daily source trees to the latest numeric publication and weekly source trees to the latest stable publication. A preview on the same commit must still be eligible for promotion. Skip a repeat if its channel has no new source content.
Weekly highlights and generated notes cover changes since the last stable release. Use the prepared v0.3.0 highlights for the first weekly release.
Preserve dry-run support, the publication race check and no-overwrite behavior. Verify both channel paths, retries, and stable/preview ordering with release tests.
Implementation is scoped to the release workflow, planner, notes, documentation and regression coverage. Actual activation requires integration into the default branch.
Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).
Scoped and reasonable, but the issue leaves the version/tag arithmetic and the "what counts as new source content" definitions implicit — those are exactly the parts that break release automation. Worth pinning down here before implementation starts.
What the issue asks: keep the daily 02:00 Asia/Shanghai patch prerelease, add a Saturday 10:00 Asia/Shanghai stable run, both through the canonical release bot with the existing build/checksum/tag/publication checks. Daily = patch bump, GitHub Latest stays on the stable release. Weekly = minor bump, patch reset to 0, marks Latest. Per-channel source-tree comparison (daily vs latest numeric, weekly vs latest stable), same-commit preview still promotable, skip when a channel has no new source content. Scope: workflow, planner, notes, docs, regression coverage; activation requires integration into the default branch.
Gaps and risks
Tag scheme is undefined. Daily patch previews and weekly stable releases must not collide numerically, and the prerelease identifiers have to sort correctly for the "latest numeric publication" lookup. State explicitly whether daily tags are v0.3.1-<date>-style and how the next day's preview behaves after a weekly roll to v0.4.0 — does it restart at v0.4.1-*? An undocumented invariant here produces either a duplicate-tag failure or a stable release that is numerically older than an existing preview.
"Source content" needs a concrete definition. Full-tree comparison vs a subtree/hash over release-relevant paths are very different in behavior: a docs-only or CI-only commit would otherwise burn a patch bump every day. Name the compared ref range (previous tag on that channel .. HEAD) and the hash basis.
Latest-pointer transitions are implicit. "Daily leaves Latest on the stable release" needs an explicit make_latest value on both paths, plus a stated behavior when the weekly run fails — otherwise Latest silently stays on the previous minor and nothing surfaces it. Also confirm that a post-weekly daily preview cannot become the resolved latest numeric publication for consumers who follow "latest numeric".
Two schedules can now overlap. The publication race check is preserved, but Saturday 10:00 stable can collide with a late/manual daily run. Say whether both jobs share one concurrency group (serialize) or use separate groups relying only on the race check.
Skip semantics are unstated. A skipped weekly run should exit in a way that is distinguishable from a failure (success-with-annotation vs failure), otherwise the schedule's green/red signal becomes meaningless.
Notes ranges. Weekly highlights "since the last stable release" — confirm the range is the last trusted stable tag .. HEAD (not the last weekly attempt), and give the checked-in path plus an owner for the prepared v0.3.0 highlights so the notes generation does not fight hand-written text on the first run.
Dry run and secrets. Preserve dry-run support for both channels and confirm the canonical release bot's token/permissions already cover a second scheduled workflow on the default branch, not just the existing daily one.
Cron literals. Publish the UTC crons alongside the local times (02:00 CST = 0 18 * * *; Saturday 10:00 CST = 0 2 * * 6) so reviewers can check them without redoing the offset.
Activation is out of band. Since landing this does not activate it, the issue should spawn a tracked follow-up for the default-branch integration plus one manual dry run of the weekly path before the first live run, with a named sign-off.
Clarifying questions
Exact tag/version format for each channel, and the promotion case: when a preview and a stable land on the same commit, is the stable release a re-tag of existing artifacts (checksums reused) or a fresh build?
What is the comparison basis for "new source content" — full tree, or a release-relevant subtree hash?
Where do the v0.3.0 weekly highlights live, and who owns that file?
Can a weekly minor bump ever be skipped in favor of promoting the newest daily preview instead, or is the weekly path always a fresh minor?
Suggested next step: post a short design delta in this issue covering the tag scheme, the comparison refs/hash basis, and the Latest-pointer values for both paths; then land it as one PR with release tests that use a fake clock and fixture tags (covering both channel paths, the retry path, the skip path, and stable/preview ordering) so the results do not depend on wall-clock time.
I read issue #540's body only — no linked design doc or PR was referenced, so this is a first pass over the stated scope, not a review of an implementation.
Thanks for writing this down as a scoped follow-up to the existing daily preview bot. I read the issue body and the earlier comment from @Dong1017; I also checked the current default-branch implementation (.github/workflows/release.yml, tools/check_release.py, [tool.inferrsi.release] in pyproject.toml, and recent Releases). The direction is reasonable, but a few points should be pinned to the current planner and workflow before implementation. Reusing today's daily-only helpers unchanged would break the weekly path.
Current baseline (so the design can say what stays and what changes)
Daily 02:00 Asia/Shanghai is already cron: "0 18 * * *" on the canonical Release bot. Version tags are already plain vMAJOR.MINOR.PATCH with no date or prerelease suffix; tools/check_release.py and setuptools_scm.tag_regex reject any other form. Daily vs formal is distinguished only by the GitHub prerelease flag and --latest=false vs --latest. The latest published numeric tag is currently v0.2.7 (prerelease); the latest stable is v0.2.0.
plan_release() always does patch+1 from the max reserved or published numeric version. The only stable path is the one-off --release-v020 hard-coded to v0.2.0, with matching special cases for notes (docs/release-notes/v0.2.0.md) and the notes start tag.
The skip check compares the full default-branch tree to previous_tag = max(published numeric tags). The publication race check requires that the latest published numeric tag is still that same PREVIOUS_TAG before writing. Both jobs already share the release-bot concurrency group (cancel-in-progress: false).
Please confirm or tighten the following
First weekly tag and subsequent daily tags. Please state that the first Saturday stable is v0.3.0 (minor bump from the current v0.2.x series, patch reset, GitHub Latest), and that the next daily preview is then v0.3.1 with --latest=false. If the tag scheme stays the existing vMAJOR.MINOR.PATCH plus prerelease flag, say so explicitly so implementers do not introduce suffixed tags that the planner and setuptools_scm would reject.
Channel-specific baselines must not reuse the daily previous_tag. Today's skip would treat a same-commit promotion as "no new source content" (the weekly tree equals v0.2.7). Today's race check would also fail a weekly plan whose PREVIOUS_TAG is the last stable (v0.2.0), because the latest numeric tag is v0.2.7. Weekly skip should compare against the last stable; the race / no-overwrite check should ensure the target tag is unused and that no newer stable appeared, not that max numeric still equals last stable. Same-commit promotion should be a new tag plus a fresh build (version strings in the artifacts will differ, so checksums from the preview cannot be reused).
Notes and highlights. Weekly notes "since the last stable release" must not use the daily previous_tag (which would be v0.2.7 for a first v0.3.0). Please put the prepared v0.3.0 highlights at a checked-in path analogous to docs/release-notes/v0.2.0.md, name an owner, and generalize the v0.2.0 one-off notes/highlights branches rather than adding another hard-coded tag.
Activation and dry-run. Scheduled runs currently publish; dry_run only applies to workflow_dispatch (default true). Since "actual activation requires integration into the default branch," please say how the first weekly cron is added without accidentally publishing the first Saturday (for example: land the planner and tests first with the Saturday cron disabled or guarded, require one successful manual dry run of the weekly path, then enable the cron). Keep both crons in the existing release-bot group so a late or manual daily run and Saturday 10:00 Asia/Shanghai (0 2 * * 6) cannot publish concurrently.
No mid-week official patch. This scheme has no stable patch channel. A hotfix that lands mid-week would only ship as a preview until the next Saturday, which would then be a new minor (for example v0.4.0), not v0.3.1 stable. If that calendar policy is intentional, please say so; otherwise generalize today's workflow_dispatch into a manual stable-patch path instead of another hard-coded version.
"New source content." The current skip is a full-tree hash. A docs-only or CI-only commit already burns a daily patch. Please state whether weekly and daily keep that full-tree rule or switch to a release-relevant subtree (and if the latter, whether daily behavior changes too).
The earlier questions about skip-versus-failure signaling, the Latest pointer when the weekly run fails, and promotion versus retag are also worth answering here. A short design delta covering the tag arithmetic, the two comparison/race baselines, the notes path, and the activation / dry-run sequence would be enough to start the PR. Release tests should use a fake clock and fixture tags so they do not depend on wall-clock Saturday.
I only reviewed the issue text against the current default-branch release bot — not an implementation PR (none is linked here).
changed the title [-]Publish weekly Saturday minor releases alongside daily patch previews[/-][+]Publish weekly Sunday minor releases alongside daily patch previews[/+]on Nov 21, 2026
I see the schedule update: the weekly stable run moved from Saturday 10:00 Asia/Shanghai to Sunday 08:00 Asia/Shanghai (cron 0 0 * * 0). The rest of the request looks unchanged.
Please keep the design and the implementation aligned with this calendar — the weekly cron, the "already published this week" window, and the activation / dry-run target should all use Sunday 08:00 rather than Saturday 10:00. The mid-week official-patch gap now runs until the following Sunday (a Friday hotfix waits two days longer than under the Saturday plan). If that is intentional, a one-line confirmation here would help.
Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).
Decision: accepted as the design baseline for #541, with the clarifications below. Answering as Dong-bot (authorized delegate). The decisions are scoped to the current issue body, where the weekly run is Sunday 08:00 Asia/Shanghai. My first-pass comment and Amy's first reply still say Saturday 10:00, so treat those times as superseded. I read only this issue's body and thread, not #541 or the code.
Schedule
Daily preview stays at 02:00 CST, 0 18 * * *.
Weekly stable runs Sunday 08:00 CST, 0 0 * * 0.
The Saturday crons are dropped.
The Sunday daily at 02:00 CST runs six hours before the weekly run. It takes the next patch, and the weekly run then takes the next minor. That ordering is expected and needs a test.
Amy's six points
Tags.
Keep plain vMAJOR.MINOR.PATCH, with no suffixes. Previews are distinguished only by the GitHub prerelease flag and --latest=false.
Weekly takes the max numeric major.minor, adds 1 to the minor and sets the patch to 0. With today's baseline, the first weekly is v0.3.0.
Daily takes the max numeric patch plus 1. The next daily after v0.3.0 is v0.3.1, and after a later weekly roll it is v0.4.1.
Channel baselines.
Daily compares its source tree to the latest numeric publication.
Weekly compares to the latest stable publication, so a preview on the same commit stays promotable.
Promotion is a new tag with a fresh build. Checksums are not reused, because the artifact version strings differ.
The weekly race check requires that the target tag is unused and that no newer stable has appeared. The daily check keeps today's rule that the latest numeric tag is unchanged.
Notes.
Weekly notes and highlights cover the last stable tag through HEAD, not the daily previous_tag.
The prepared v0.3.0 highlights go at docs/release-notes/v0.3.0.md. Generalize the v0.2.0 one-off into a docs/release-notes/vX.Y.Z.md lookup instead of adding another hard-coded tag.
Activation.
Land the planner, workflow and tests with the weekly cron guarded or disabled.
Run one manual workflow_dispatch dry run of the weekly path.
Enable the weekly cron in a separate follow-up after that dry run.
Both crons stay in the existing release-bot group with cancel-in-progress: false, so runs serialize.
Dry-run support is preserved for both channels.
No stable mid-week patch. This is intentional, per the maintainer's calendar. A Friday hotfix ships as a preview and reaches stable at the next Sunday's minor bump, for example v0.4.0. A manual stable-patch path is out of scope for Publish weekly stable minor releases alongside daily patch previews #541.
"New source content." Keep the current full-tree comparison for both channels. A release-relevant subtree is out of scope, so daily behavior doesn't change.
My earlier open questions
Skip: a skipped run exits successfully with a notice annotation, so red only means a real failure.
Weekly failure: Latest stays on the previous stable and the run fails visibly. A daily preview never takes Latest (--latest=false).
Weekly vs. promoting the newest preview: the weekly run is always a fresh minor and fresh build. It is never a retag of a preview.
Tests: use a fake clock and fixture tags. Cover both channel paths, retries, the skip path, same-commit promotion, the race check, and stable/preview ordering.
Still open, non-blocking for the design
Name the owner of docs/release-notes/v0.3.0.md.
Add a tracked follow-up for the default-branch cron enablement, with a named sign-off after the dry run.
Confirm that the longer wait for a stable hotfix under the Sunday schedule is acceptable. Amy asked for a one-line confirmation, and the policy above assumes it is.
The release calendar is confirmed: daily patch previews at 02:00 Asia/Shanghai, and a new stable minor on Sunday at 08:00. A mid-week hotfix is eligible for the daily preview and then for Sunday's stable release. The first weekly tag on the current series is v0.3.0, followed by v0.3.1 for the next eligible daily change.
#541 now includes the separate stable publication race check and a default-off activation guard. After #541 merges, the release owner (@zengchuang-hw) owns the v0.3.0 highlights and sign-off for the first weekly dry run. The run, sign-off and first live release are tracked in #554.
Activity
Dong1017 commentedon Nov 21, 2026
Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).
Scoped and reasonable, but the issue leaves the version/tag arithmetic and the "what counts as new source content" definitions implicit — those are exactly the parts that break release automation. Worth pinning down here before implementation starts.
What the issue asks: keep the daily 02:00 Asia/Shanghai patch prerelease, add a Saturday 10:00 Asia/Shanghai stable run, both through the canonical release bot with the existing build/checksum/tag/publication checks. Daily = patch bump, GitHub Latest stays on the stable release. Weekly = minor bump, patch reset to 0, marks Latest. Per-channel source-tree comparison (daily vs latest numeric, weekly vs latest stable), same-commit preview still promotable, skip when a channel has no new source content. Scope: workflow, planner, notes, docs, regression coverage; activation requires integration into the default branch.
Gaps and risks
v0.3.1-<date>-style and how the next day's preview behaves after a weekly roll tov0.4.0— does it restart atv0.4.1-*? An undocumented invariant here produces either a duplicate-tag failure or a stable release that is numerically older than an existing preview.make_latestvalue on both paths, plus a stated behavior when the weekly run fails — otherwise Latest silently stays on the previous minor and nothing surfaces it. Also confirm that a post-weekly daily preview cannot become the resolved latest numeric publication for consumers who follow "latest numeric".dry-runsupport for both channels and confirm the canonical release bot's token/permissions already cover a second scheduled workflow on the default branch, not just the existing daily one.0 18 * * *; Saturday 10:00 CST =0 2 * * 6) so reviewers can check them without redoing the offset.Clarifying questions
Suggested next step: post a short design delta in this issue covering the tag scheme, the comparison refs/hash basis, and the Latest-pointer values for both paths; then land it as one PR with release tests that use a fake clock and fixture tags (covering both channel paths, the retry path, the skip path, and stable/preview ordering) so the results do not depend on wall-clock time.
I read issue #540's body only — no linked design doc or PR was referenced, so this is a first pass over the stated scope, not a review of an implementation.
Schedule Saturday stable minors alongside daily patch previews
amy-why-3459 commentedon Nov 21, 2026
Thanks for writing this down as a scoped follow-up to the existing daily preview bot. I read the issue body and the earlier comment from @Dong1017; I also checked the current default-branch implementation (
.github/workflows/release.yml,tools/check_release.py,[tool.inferrsi.release]inpyproject.toml, and recent Releases). The direction is reasonable, but a few points should be pinned to the current planner and workflow before implementation. Reusing today's daily-only helpers unchanged would break the weekly path.Current baseline (so the design can say what stays and what changes)
cron: "0 18 * * *"on the canonical Release bot. Version tags are already plainvMAJOR.MINOR.PATCHwith no date or prerelease suffix;tools/check_release.pyandsetuptools_scm.tag_regexreject any other form. Daily vs formal is distinguished only by the GitHub prerelease flag and--latest=falsevs--latest. The latest published numeric tag is currentlyv0.2.7(prerelease); the latest stable isv0.2.0.plan_release()always does patch+1 from the max reserved or published numeric version. The only stable path is the one-off--release-v020hard-coded tov0.2.0, with matching special cases for notes (docs/release-notes/v0.2.0.md) and the notes start tag.previous_tag = max(published numeric tags). The publication race check requires that the latest published numeric tag is still that samePREVIOUS_TAGbefore writing. Both jobs already share therelease-botconcurrency group (cancel-in-progress: false).Please confirm or tighten the following
First weekly tag and subsequent daily tags. Please state that the first Saturday stable is
v0.3.0(minor bump from the currentv0.2.xseries, patch reset, GitHub Latest), and that the next daily preview is thenv0.3.1with--latest=false. If the tag scheme stays the existingvMAJOR.MINOR.PATCHplus prerelease flag, say so explicitly so implementers do not introduce suffixed tags that the planner andsetuptools_scmwould reject.Channel-specific baselines must not reuse the daily
previous_tag. Today's skip would treat a same-commit promotion as "no new source content" (the weekly tree equalsv0.2.7). Today's race check would also fail a weekly plan whosePREVIOUS_TAGis the last stable (v0.2.0), because the latest numeric tag isv0.2.7. Weekly skip should compare against the last stable; the race / no-overwrite check should ensure the target tag is unused and that no newer stable appeared, not that max numeric still equals last stable. Same-commit promotion should be a new tag plus a fresh build (version strings in the artifacts will differ, so checksums from the preview cannot be reused).Notes and highlights. Weekly notes "since the last stable release" must not use the daily
previous_tag(which would bev0.2.7for a firstv0.3.0). Please put the prepared v0.3.0 highlights at a checked-in path analogous todocs/release-notes/v0.2.0.md, name an owner, and generalize the v0.2.0 one-off notes/highlights branches rather than adding another hard-coded tag.Activation and dry-run. Scheduled runs currently publish;
dry_runonly applies toworkflow_dispatch(default true). Since "actual activation requires integration into the default branch," please say how the first weekly cron is added without accidentally publishing the first Saturday (for example: land the planner and tests first with the Saturday cron disabled or guarded, require one successful manual dry run of the weekly path, then enable the cron). Keep both crons in the existingrelease-botgroup so a late or manual daily run and Saturday 10:00 Asia/Shanghai (0 2 * * 6) cannot publish concurrently.No mid-week official patch. This scheme has no stable patch channel. A hotfix that lands mid-week would only ship as a preview until the next Saturday, which would then be a new minor (for example
v0.4.0), notv0.3.1stable. If that calendar policy is intentional, please say so; otherwise generalize today'sworkflow_dispatchinto a manual stable-patch path instead of another hard-coded version."New source content." The current skip is a full-tree hash. A docs-only or CI-only commit already burns a daily patch. Please state whether weekly and daily keep that full-tree rule or switch to a release-relevant subtree (and if the latter, whether daily behavior changes too).
The earlier questions about skip-versus-failure signaling, the Latest pointer when the weekly run fails, and promotion versus retag are also worth answering here. A short design delta covering the tag arithmetic, the two comparison/race baselines, the notes path, and the activation / dry-run sequence would be enough to start the PR. Release tests should use a fake clock and fixture tags so they do not depend on wall-clock Saturday.
I only reviewed the issue text against the current default-branch release bot — not an implementation PR (none is linked here).
[-]Publish weekly Saturday minor releases alongside daily patch previews[/-][+]Publish weekly Sunday minor releases alongside daily patch previews[/+]amy-why-3459 commentedon Nov 21, 2026
I see the schedule update: the weekly stable run moved from Saturday 10:00 Asia/Shanghai to Sunday 08:00 Asia/Shanghai (cron
0 0 * * 0). The rest of the request looks unchanged.Please keep the design and the implementation aligned with this calendar — the weekly cron, the "already published this week" window, and the activation / dry-run target should all use Sunday 08:00 rather than Saturday 10:00. The mid-week official-patch gap now runs until the following Sunday (a Friday hotfix waits two days longer than under the Saturday plan). If that is intentional, a one-line confirmation here would help.
#541 already reflects this Sunday 08:00 schedule.
Dong1017 commentedon Nov 21, 2026
Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).
Decision: accepted as the design baseline for #541, with the clarifications below. Answering as Dong-bot (authorized delegate). The decisions are scoped to the current issue body, where the weekly run is Sunday 08:00 Asia/Shanghai. My first-pass comment and Amy's first reply still say Saturday 10:00, so treat those times as superseded. I read only this issue's body and thread, not #541 or the code.
Schedule
0 18 * * *.0 0 * * 0.Amy's six points
vMAJOR.MINOR.PATCH, with no suffixes. Previews are distinguished only by the GitHub prerelease flag and--latest=false.major.minor, adds 1 to the minor and sets the patch to 0. With today's baseline, the first weekly isv0.3.0.v0.3.0isv0.3.1, and after a later weekly roll it isv0.4.1.previous_tag.docs/release-notes/v0.3.0.md. Generalize the v0.2.0 one-off into adocs/release-notes/vX.Y.Z.mdlookup instead of adding another hard-coded tag.workflow_dispatchdry run of the weekly path.release-botgroup withcancel-in-progress: false, so runs serialize.v0.4.0. A manual stable-patch path is out of scope for Publish weekly stable minor releases alongside daily patch previews #541.My earlier open questions
--latest=false).Still open, non-blocking for the design
docs/release-notes/v0.3.0.md.zengchuang-hw commentedon Nov 21, 2026
The release calendar is confirmed: daily patch previews at 02:00 Asia/Shanghai, and a new stable minor on Sunday at 08:00. A mid-week hotfix is eligible for the daily preview and then for Sunday's stable release. The first weekly tag on the current series is v0.3.0, followed by v0.3.1 for the next eligible daily change.
#541 now includes the separate stable publication race check and a default-off activation guard. After #541 merges, the release owner (@zengchuang-hw) owns the v0.3.0 highlights and sign-off for the first weekly dry run. The run, sign-off and first live release are tracked in #554.