Skip to content

Release decision: one-off formal InferRSI v0.2.0 #443

Open

Description

@hsliuustc0106

Decision and scope

The owner explicitly approved creation and linkage of this public governing record for the one-off formal v0.2.0 release prepared in #442. The requested release is stable, not a draft or prerelease; highlights compare against v0.1.1. The assistant is preparing the release under hsliuustc0106. This record does not substitute for independent PR approvals or bypass branch protection.

Acceptance

  • Merge the reviewed preparation PR only after required CI and two independent eligible write-account approvals; retain unresolved review concerns until reviewers clear them.
  • Keep scheduled/default preview policy and pyproject settings unchanged. After v0.2.0, ordinary changed-source planning naturally chooses v0.2.1 preview.
  • Use the manual-only exact v0.2.0 path on reviewed main, first with dry_run=true. Hongsheng Liu (hsliuustc0106) is the release decision owner for the subsequent dry_run=false publication; the assistant must not infer that this issue or a reviewer comment independently authorizes dispatch.
  • Bind tag and all packages to the planned source commit. Deliver wheel, sdist, Codex plugin ZIP, WorkSwarm plugin ZIP and SHA256SUMS.
  • Require source/ancestry and reservation checks, CPU tests/lint, build/twine/sdist validation, clean installed-wheel/version parity, both extracted plugin validations and downloaded checksum verification. Recheck current release state immediately before publication; no overwrite of reserved tags, releases or assets.
  • Use reviewed highlights since v0.1.1 with actual packaged tasks and README examples; retain FAIL/INCONCLUSIVE qualifications. Software checks are not GPU/NPU model-quality or performance validation.
  • Verify published release is draft=false, prerelease=false, points to the intended tag/commit, and has all five artifact classes.

Current status

PR #442 head 19b38de. Focused tests: 210 passed; lint and diff checks passed. Initial exact-head CI matrix passed. Fresh Ready run is still in progress: https://github.com/InferMatrixLab/InferRSI/actions/runs/43599786640 . No release dispatch, merge, tag creation or publication has occurred for v0.2.0.

Boundaries and cleanup

The one-off path intentionally remains fail-closed after v0.2.0 is reserved/published. Cleanup can be proposed and reviewed separately after release, not silently folded into this release decision. No permission expansion, paid model calls or hardware campaigns are authorized by this record.

Activity

  1. amy-why-3459 commented on Nov 14, 2026

    @amy-why-3459
    Collaborator

    Thanks for recording this. It covers what I asked for on #442:

    • the stable v0.2.0 decision;
    • v0.1.1 as the notes baseline;
    • a named owner (@hsliuustc0106) for the dry_run=false publication, after a dry_run=true pass on reviewed main.

    I've approved #442 at 19b38de8fa13a3a0d6e7375306cf396cd511f7eb. That approval covers the preparation code only, not dispatch.

    One suggestion for the follow-up cleanup: once v0.2.0 is published, remove the release_v020 input, the planner flag and the four RELEASE_TAG == v0.2.0 checks in release.yml, or keep them with a note that they are intentionally fail-closed. A cross-link from here would make that easy to track.

  2. added
    area:processRepo-wide process, governance, privacy and security rules
    kind:taskBounded delivery task with a delegated slice
    on Nov 16, 2026
  3. Dong1017 commented on Nov 17, 2026

    @Dong1017
    Collaborator

    Posted by an agent on behalf of @Dong1017 (incremental only; points already raised by others or decided by the owner are not repeated).

    Incremental note: v0.2.0 is published and its assets check out, but the dry_run=true pass ran on a different commit from the one that was published. The "Current status" section still says nothing has been published.

    1. Published release matches the asset criteria. v0.2.0 was published 2026-11-14 14:03 UTC with draft=false and prerelease=false, and the tag points at 223ee523115d8d59eebf4568fd6ac34f4708b1d4. All five artifact classes are attached (wheel, sdist, Codex ZIP, WorkSwarm ZIP, SHA256SUMS). I downloaded them, and sha256sum -c SHA256SUMS passed for all four packages. The wheel METADATA says Version: 0.2.0.
    2. The dry run did not cover the published commit. Release run 43601546789 ran on 86435c12 (the release: prepare one-off formal v0.2.0 with reviewed highlights #442 merge), and its publish job was skipped. The publishing run 43602273847 ran on 223ee523, which is one commit later: Compose standalone knowledge reads in Apps #436 "Compose standalone knowledge reads in Apps" was merged at 13:53 UTC, between the two dispatches. The publishing run did run its own build, CPU checks, clean-install and plugin validation on 223ee523. Whether that satisfies "first with dry_run=true" on the published source is a call for the release decision owner (@hsliuustc0106), and the issue should record that call.
    3. Record and cleanup. "Current status" still says "No release dispatch, merge, tag creation or publication has occurred". Once the owner records the outcome, the cleanup from @amy-why-3459's note still applies: .github/workflows/release.yml on main 1acfacf8 still has the release_v020 input (L9) and the RELEASE_TAG == v0.2.0 branches (L167, L255, L267, L277), and no cleanup issue is linked here.

    Scope: GitHub release API and assets, the two release.yml run job and step lists, commit ancestry on main 1acfacf8. Unverified: the workflow_dispatch inputs (not exposed by the runs API). I inferred that the first run was the dry run because its publish job was skipped.

  4. zengchuang-hw commented on Nov 17, 2026

    @zengchuang-hw
    Contributor

    Independent release-boundary review (not a publication authorization)

    Reviewed current main 1bbc3b0c8de14d6194d7ee7987607365b665fbfe against the one-off v0.2.0 scope and the release tag's source 223ee523115d8d59eebf4568fd6ac34f4708b1d4. I did not use other reviewers' comments as evidence.

    The one-off publication is already observable: v0.2.0 was published on 2026-11-14 at 14:03:33 UTC with draft=false, prerelease=false; its tag resolves directly to the source above. Its asset inventory contains the wheel, sdist, Codex ZIP, WorkSwarm ZIP and SHA256SUMS. Release run 43602273847 is successful at that SHA, including the build/package, isolated installed-wheel, plugin validation, checksum and publisher steps. These are observed GitHub status records, not a claim that I independently rebuilt those uploaded bytes.

    Source checks:

    • Direct fixed-ref comparison confirms the release workflow, planner/checker, v0.2.0 notes and publisher tests are unchanged between that tag and the reviewed main (not a whole-repository equivalence claim).
    • The one-off mode is manual/opt-in, forces 0.2.0 stable, retains v0.1.1 ancestry/notes baseline and rejects an already reserved v0.2.0 or any later numeric version. Consequently this is not a request to rerun or overwrite the one-off release.
    • Ordinary defaults remain preview; subsequent published v0.2.1/v0.2.2/v0.2.3 are also marked prerelease. The notes preserve INCONCLUSIVE and distinguish software/package checks from model quality or service-performance acceptance.
    • On current main, python -m pytest -o addopts='' -q -p no:cacheprovider tests/test_release.py tests/test_release_publisher.py: 70 passed. These exercise planner guards and the actual publisher shell against mocks, not live publication.

    Conclusion: no additional release-boundary implementation blocker found in this focused check. The published-state evidence supports recording the one-off outcome rather than issuing another publication decision.

    Limits: I did not independently verify the preparatory PR's two eligible write-account approvals, the separate dry-run/authorization sequence, download and recompute the published asset hashes, rebuild all distributions, run the full test suite, or exercise native host UIs/GPU/NPU workloads. Those are not implied by this comment or by the successful workflow status. No tag, release, asset, dependency, repository setting or publication action was changed by this review.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    area:processRepo-wide process, governance, privacy and security ruleskind:taskBounded delivery task with a delegated slice

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions