Skip to content

Publish weekly stable minor releases alongside daily patch previews - #541

Merged
zengchuang-hw merged 4 commits into
mainfrom
codex/weekly-minor-releases-540
Nov 21, 2026
Merged

zengchuang-hw merged 4 commits into
mainfrom
codex/weekly-minor-releases-540

Conversation

@zengchuang-hw

@zengchuang-hw zengchuang-hw commented Nov 21, 2026 •

Copy link
Copy Markdown
Contributor

Governing issue and delivery

Refs #540. This change adds weekly stable minor releases to the existing daily preview bot and includes the v0.3.0 highlights.

Asia/Shanghai schedule Baseline for changes and notes Publication
Every day at 02:00 Latest published numeric release Next patch, prerelease
Sunday at 08:00 Latest published numeric stable release Next minor with patch zero, Latest

On Sunday, a preview can be promoted on the same commit using a fresh build. Weekly retries skip a stable release already published in the local Sunday-to-Saturday window. Both channels retain full-tree comparison and serialized packaging, checksums and tag verification. A weekly publication rechecks the last stable and its target. A daily publication rechecks the latest numeric tag.

Checks

Release planner, publisher, and notes suites on Windows: 193 passed. The 34 shell publisher tests requiring jq skipped locally and run on Linux in the ready lane. Ruff and Bash syntax checks passed. Regression coverage includes Sunday preview-before-stable ordering, the end-of-week retry boundary, default-off weekly publication, manual dry runs, and channel-specific publication races.

Rollout

Weekly publication is disabled by default. After merge, the release owner runs the workflow on the default branch with weekly=true and dry_run=true. #554 tracks the run, owner sign-off, enabling INFERRSI_WEEKLY_RELEASE_ENABLED=true, and verification of the first Sunday release. Daily previews continue during rollout.

Use a stable-to-stable comparison for weekly notes and the latest publication for race checks.

Agent-assisted-by: Codex (implementation and release documentation)
Signed-off-by: Zeng Chuang <zengchuang1005@gmail.com>
@Dong1017 Dong1017 added area:process Repo-wide process, governance, privacy and security rules kind:task Bounded delivery task with a delegated slice labels Nov 21, 2026
amy-why-3459
amy-why-3459 previously approved these changes Nov 21, 2026

@amy-why-3459 amy-why-3459 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

PR review verdict — PASS
Reviewed base 6ed42af and head 4b1d829.

Verified against #540 and the current daily-only bot: weekly skip and notes use the last stable (notes_start_tag), while the publication race / no-overwrite check still uses the latest numeric previous_tag, so a same-commit promotion such as v0.2.7 → v0.3.0 is not treated as an unchanged tree or a stale race. The first Saturday stable is v0.3.0 (Latest) with checked-in docs/release-notes/v0.3.0.md; the following daily preview is v0.3.1 (--latest=false). The v0.2.0 notes/highlights special cases are generalized via reviewed_notes. Both crons share the release-bot concurrency group. Planner tests use a fake clock for promotion, the Saturday-to-Friday window skip, reservations, and the next-day patch. lint-and-smoke and full-check (3.10/3.11) are green on this head.

Operational note (non-blocking): the Saturday cron becomes live once this reaches the default branch. After merge, please run one manual weekly dry-run before the next Saturday 10:00 Asia/Shanghai (0 2 * * 6) publication.

Start the weekly deduplication window on Sunday and cover both sides of the local week boundary.

Agent-assisted-by: Codex (implementation and tests)
Signed-off-by: Zeng Chuang <zengchuang1005@gmail.com>

@Dong1017 Dong1017 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

PR review verdict — DO NOT PASS

Reviewed base 6ed42af83a69f13fe88aaf0426ab8ef4f8ac7654 and head 4b1d829b305dca2f0b6f07d7bd66806fea786f9d (first review of this head; diffMode: full, diffComplete: true). Refs #540.

@amy-why-3459's APPROVED review on this exact head (pullrequestreview-6002738744) is on record, and there is no outstanding CHANGES_REQUESTED. The two-channel mechanics below are sound and I am not re-litigating them. My blocker is a value mismatch against the governing issue, so this does not approve.

Blocker — the stable channel fires on Saturday; #540 asks for Sunday 08:00 Asia/Shanghai

  • .github/workflows/release.yml: - cron: "0 2 * * 6" is Saturday 02:00 UTC = Saturday 10:00 Asia/Shanghai.
  • docs/releases.md: the schedule table documents the same value ("Saturday at 10:00 (02:00 UTC Saturday)").
  • tools/check_release.py plan_release(): the weekly window is Saturday-anchored — saturday = today - timedelta(days=(today.weekday() - 5) % 7) and the retry skip stamp.astimezone(shanghai).date() >= saturday. tests/test_release.py pins the fake clock at datetime(2026, 11, 21, 2, tzinfo=utc), which is indeed a Saturday.
  • Issue #540 says the opposite: title "Publish weekly Sunday minor releases…", body "Add a Sunday 08:00 Asia/Shanghai stable run." The issue has not been edited to Saturday, and it has exactly two comments — the earlier Dong-bot triage (which paraphrased the ask as "a Saturday 10:00 Asia/Shanghai stable run") and Amy's baseline comment. No maintainer comment resolves Sunday → Saturday.

So the Saturday value looks inherited from that paraphrase rather than decided. Weekly cadence aside, a Latest-marked stable release landing on the wrong day is user-visible and the docs would codify it.

Expected correction — one of:

  1. Sunday is intended: - cron: "0 0 * * 0" (Sunday 00:00 UTC = 08:00 Asia/Shanghai), re-anchor the planner window to Sunday ((weekday - 6) % 7), update the "Saturday-to-Friday" wording in docs/releases.md, and re-point the fake-clock tests. Note this is not a cron-only edit — the retry-skip window and its documented semantics move with it.
  2. Saturday was a deliberate change: record it in #540 (title and body) and say so here, so the accepted scope and docs/releases.md agree.

Verified in the diff (no blocker)

  • Channel baselines are correctly separated: the unchanged-tree skip uses notes_start_tag (last published stable), while previous_tag (latest published numeric, preview included) stays the race/no-overwrite input — so a same-commit v0.2.7 → v0.3.0 promotion is neither skipped as unchanged nor rejected as a stale race. v0.2.0 is a published stable in Amy's scenario, so the weekly baseline resolves to v0.2.0, not the preview.
  • Weekly arithmetic (major, minor + 1, 0) against max(reserved ∩ published); --weekly / --release-v020 are a mutually exclusive argparse group and plan_release() raises when both are set.
  • The v0.2.0 notes/highlights special cases are generalized via reviewed_notes, computed from git ls-tree on the planned commit and only for non-prerelease plans; publish-side validation [[ -z "$NOTES_START_TAG" || "$NOTES_START_TAG" =~ $version_tag ]] and [[ "$REVIEWED_NOTES" == true || == false ]].
  • Both crons stay in the single release-bot concurrency group with cancel-in-progress: false.

Not verified here

  • I did not execute any tests. Fake-clock plumbing is only partly visible to me: the added plan_release(..., weekly=True, now=...) call sites are in the diff, the test-helper definition is outside the changed hunks, so I cannot see how now is injected. The green lint-and-smoke / full-check on this head is the evidence for that, not my run.
  • First-stable notes baseline: test_first_stable_respects_reservations_and_ignores_drafts_as_baselines pins notes_start_tag == "", and in that state the notes job calls tools/generate_release_notes.py --previous-tag "" whenever no curated file matches. The checked-in docs/release-notes/v0.3.0.md covers the first weekly run, so the empty-baseline generator path is not exercised; I did not read the generator's handling of an empty baseline.

Non-blocking follow-ups

  • Once the first weekly stable publishes, any weekly dispatch between that Saturday and the following Friday returns release: false before the source-tree comparison, so the stable path cannot be rehearsed mid-window. Worth one line in the dry-run section of docs/releases.md.
  • The rewrite drops the earlier "a failed reserved release is not overwritten or silently replaced — investigate and obtain a new release decision" guidance rather than keeping it for the one-off path.
  • Amy's post-merge operational note stands: one manual weekly dry run before the first scheduled stable run.

I'll re-read this once the day is aligned (code or issue) and the author pushes.


Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).

amy-why-3459
amy-why-3459 previously approved these changes Nov 21, 2026

@amy-why-3459 amy-why-3459 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

PR review verdict — PASS
Reviewed base 6ed42af and head 050b9fc.
Reviewed increment 4b1d829 → 050b9fc.

Verified against the updated #540 calendar: the weekly stable run moves from Saturday 10:00 Asia/Shanghai (0 2 * * 6) to Sunday 08:00 (0 0 * * 0). The planner's skip window is now Sunday-to-Saturday (today.isoweekday() % 7), and the tests cover same-week promotion, a Saturday 23:59 Asia/Shanghai retry, and the next Sunday minor. Workflow WEEKLY detection, docs/releases.md, and the publisher assertions match the new cron. Previously reviewed tag arithmetic, notes/race baselines, and the shared release-bot concurrency group are unchanged.

lint-and-smoke, declarations, pr-references, installed-package (3.10/3.11), dashboard-browser and host-compatibility are green on this head; full-check (3.10/3.11) in Ready CI 44865721028 is still running and should pass before merge.

Operational note (non-blocking): the Sunday cron becomes live once this reaches the default branch. After merge, please run one manual weekly dry-run before the next Sunday 08:00 Asia/Shanghai (0 0 * * 0) publication.

Cover the Sunday preview before promotion, channel-specific publication races, and dry-run eligibility. Document the rollout and one-off recovery.

Agent-assisted-by: Codex (implementation and review fixes)
Signed-off-by: Zeng Chuang <zengchuang1005@gmail.com>

Copy link
Copy Markdown
Contributor Author

@Dong1017 @amy-why-3459, the Sunday 08:00 Asia/Shanghai schedule, Sunday-based retry window and boundary tests from 050b9fc are in the current head.

  • The weekly dry-run instructions now explain why an already-published week skips even after new commits. The historical v0.2.0 section restores the maintainer decision required after a failed reservation.
  • I checked the empty notes baseline raised in review. source_snapshot walks history reachable from the planned commit and computes file statistics against Git's empty tree. The existing test_first_release_uses_complete_history_and_empty_tree_statistics passed locally.
  • The Sunday regression now plans the 02:00 patch first and promotes that same commit at 08:00. The weekly publisher checks the stable baseline and target, allowing an intervening lower preview. A newer stable or a superseding numeric release blocks the write.
  • As agreed in Publish weekly Sunday minor releases alongside daily patch previews #540, merging this PR leaves weekly publication disabled. #554 tracks the default-branch dry run and sign-off before enabling the scheduled release.

Local release suites: 193 passed, 34 shell publisher cases skipped because jq is unavailable here. Ruff passed and all 12 workflow shell blocks passed Bash syntax validation. Please re-review the increment from 050b9fc.

@Dong1017 Dong1017 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

PR review verdict — PASS
Reviewed base 6ed42af and head 880bc08. I read the full seeded diff against base (diffComplete: true), not only the 050b9fc→880bc08 increment. I also read #540 and #554. I did not run any tests.

Earlier blocker resolved. My DO NOT PASS on 4b1d829 was the Saturday-vs-Sunday mismatch. The head now matches #540, the 10:23 confirmation and Dong-bot's accepted decision there:

  • release.yml has cron: "0 0 * * 0", which is Sunday 08:00 Asia/Shanghai, next to the daily 0 18 * * *.
  • plan_release uses sunday = today - timedelta(days=today.isoweekday() % 7) as the retry window.
  • docs/releases.md uses the same schedule and window.

What I checked in the diff

  • Baselines. For weekly, notes_start_tag is the latest published stable. For daily it is the same as previous_tag. For --release-v020 it is v0.1.1. The unchanged-tree skip applies only to notes_start_tag, so a same-commit preview stays promotable. Ancestry is still checked for both tags.
  • Version arithmetic. Weekly plans (major, minor+1, 0) from the max of published and reserved versions, so drafts and reserved tags are counted. The v0.4.0 reservation test covers this. Weekly forces prerelease=false. --release-v020 and --weekly are mutually exclusive in both the function and argparse.
  • Retry window. A stable release published on or after the local Sunday returns release=false. The tests cover Saturday 23:59 Shanghai as still in the old week and 08:00 Sunday as the new week. They also cover the sequence 02:00 preview → 08:00 promotion on the same commit → next-Sunday skip → new commit → v0.4.0. The daily path continues with v0.3.1, and its notes start at v0.3.0.
  • Publisher race checks. Weekly requires the latest stable to equal NOTES_START_TAG and requires RELEASE_TAG to be the newest numeric tag. A lower intervening preview (v0.2.8) is allowed. A newer stable or a higher numeric release (v0.3.1) blocks the write, and the tests assert mutations == []. Daily keeps the exact latest-numeric equality check. The input validation adds WEEKLY and REVIEWED_NOTES, and rejects a weekly prerelease.
  • Reviewed notes. reviewed_notes is set when a stable release has docs/release-notes/<tag>.md at the planned commit, or for the one-off v0.2.0. It replaces the hard-coded v0.2.0 branches in the notes and publish jobs. The publisher still fails before the tag if the reviewed notes are missing or over 20000 bytes.
  • Rollout guard. A weekly run that is not a dry run and has INFERRSI_WEEKLY_RELEASE_ENABLED unset or false exits with release=false and a notice. Dry runs and daily runs are unaffected. This matches the #540 and #554 activation design, and test_weekly_rollout_gate covers all five combinations.
  • Empty-notes baseline. The author says the existing test_first_release_uses_complete_history_and_empty_tree_statistics covers a first release with an empty notes baseline. I did not open it, since it is outside the diff.

Not verified (non-blocking)

  • The 34 jq-dependent publisher tests were skipped on the author's machine. They only pass if Linux CI runs them green on this head. I could not see CI for 880bc08, so confirm it is green before merge.
  • Scheduled execution and real publication stay unobserved until the #554 dry run. That is the agreed plan.

Non-blocking follow-ups for #554

  • Run the weekly dry run on a commit whose tree differs from the latest stable, so the build is exercised. A no-change skip does not exercise it.
  • Have a sign-off who is not the implementer, as noted in the #554 triage.

I will re-check if the head changes after this approval and someone mentions Dong1017.


Posted by Dong-bot, an automated agent acting on behalf of @Dong1017 (always-on-bot, model claude-sonnet-5-5).

@amy-why-3459 amy-why-3459 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

PR review verdict — PASS
Reviewed base 6ed42af and head 880bc08.
Reviewed increment 050b9fc → 880bc08.

The new commit gates live weekly publication on vars.INFERRSI_WEEKLY_RELEASE_ENABLED while leaving weekly workflow_dispatch dry runs available, which matches the rollout accepted in #540 and recorded in #554. The publisher now rechecks the stable baseline (NOTES_START_TAG) and that the weekly target is still the newest eligible version, so a Sunday 02:00 preview below the planned minor no longer fails the race; daily still compares the latest numeric tag to PREVIOUS_TAG. The Sunday planner test now builds that preview first and promotes the same commit. Docs record the activation variable, the mid-window dry-run skip, and the one-off reservation recovery path. Previously reviewed Sunday 08:00 scheduling, notes/race baselines, and the shared release-bot concurrency group are unchanged.

lint-and-smoke, declarations, pr-references, installed-package (3.10/3.11), dashboard-browser and host-compatibility (windows-latest) were green on this head when I checked; full-check (3.10/3.11) in https://github.com/InferMatrixLab/InferRSI/actions/runs/44894100248 was still running and should pass before merge. I did not execute the publisher suite locally.

Copy link
Copy Markdown
Contributor Author

Both reviewers approved 880bc08. The Linux ready run executed the release tests, including the jq publisher cases. Its Python 3.11 job reported 5,911 passed and one failure in test_project_explanation.py. The job checked out the merge with main at 9466bc8. I reproduced that same failure on unmodified main. Fix #556 updates the retired test import and the failing case now passes locally. The release branch stays at the reviewed commit. A green ready run against corrected main is still needed before integration.

@cursor cursor Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Technical / third-party verify (cursor App permission none — not write+ / dual / merge-ready).

Tip Dong1017 always-on + amy Approves stand. Ready full-check FAIL on tip (full-check 3.10/3.11): test_recovered_attempt_does_not_trust_a_rewritable_declaration still imports retired inferrsi.infra.artifacts.digest and fails with DID NOT RAISE KeyboardInterrupt.

That is a main-integration regression after #457/#535, not a defect in this release diff. Fix is in open #556 (Refs #555). Do not merge #541 on a red Ready tip; await #556 (or equivalent) on main, then re-run Ready and write+ re-verify.

No tip Approve from cursor while Ready is red.

@cursor cursor Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Technical / third-party tip COMMENT (cursor App permission none — not Approve / not write+ / not merge-ready).

Ready blocker (concrete): tip Ready full-check FAIL is still tests/test_project_explanation.py::test_recovered_attempt_does_not_trust_a_rewritable_declaration — callback from inferrsi.infra.artifacts import digest fails before the intended KeyboardInterrupt (see Ready run 44894100248).

Unblock: #556 is now MERGED on main with that one-line Harness digest import. Update this branch onto current main and re-run Ready; tip Dong1017 + amy Approves already present — write+ steward can re-verify after green.

Retest the weekly release workflow with the corrected Harness digest import.

Agent-assisted-by: Codex (branch integration)
Signed-off-by: Zeng Chuang <zengchuang1005@gmail.com>

zengchuang-hw commented Nov 21, 2026 •

Copy link
Copy Markdown
Contributor Author

Updated #541 to d2b174d by merging main at 10ba7a55, which includes #556. The six release-related files match the previously reviewed versions.

All 10 GitHub checks passed on the updated head. Both Python 3.10 and 3.11 full checks reported 5,918 passed, 31 skipped and 674 subtests passed. The previously failing recovery test also passed locally with the release planner and scheduling checks (50 passed on the exact merged tree). GitHub shows both approvals and a clean merge status. Ready CI

@zengchuang-hw
zengchuang-hw merged commit 3bf0ae9 into main Nov 21, 2026
11 checks passed
Dong1017 added a commit that referenced this pull request Nov 21, 2026
Main gained #559 as a squash commit and #541's release workflow. The docs and the scenarios initializer conflicted only because this branch already carried #559's text with the rename on top; the merge keeps this branch's versions, and trials/__init__.py equals main's scenarios/__init__.py byte for byte. release.yml merged cleanly and still passes INFERRSI_PRE_CHANGE_SESSIONS to its one check_installed run.

Agent-assisted-by: Claude Code / claude-opus-5-5 [effort=xhigh] (author)

Signed-off-by: GUOGUO <xwdong1998@163.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

area:process Repo-wide process, governance, privacy and security rules kind:task Bounded delivery task with a delegated slice

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants