Skip to content

fix(release): supersede unpublishable 3.3.0-rc.1 - #9090

Merged
bokelley merged 1 commit into
mainfrom
fix/supersede-3-3-0-rc-1
Nov 18, 2026
Merged

bokelley merged 1 commit into
mainfrom
fix/supersede-3-3-0-rc-1

Conversation

@bokelley

Copy link
Copy Markdown
Contributor

Why

The committed 3.3.0-rc.1 Version Packages merge cannot be published. The bot generated the release head b46ec2e1804fb9fae08b3a4aede4043c235bbe7f on 61a1638493297b2119a41e4cd918e00fc42a5811 (#8736). #9080 (5eb0b201f4c9e3ac4bb92ddc580c0582ecb97280) merged after that without a changeset, so the bot never regenerated the PR. #9073 was approved on that head and squash-merged as 32a90ee0d0fd7d66505bb14277a84eeb6d15fb79. The merge's base parent is 5eb0b201f4, not 61a1638493, and its tree differs from the reviewed head by #9080's training-agent change. Release run 44300240363 stopped at "Require human approval for committed release artifacts" with the provenance quarantine. (#9082 merged before 61a1638493 and is already in the reviewed head.)

RELEASING.md says an unpublishable committed candidate must not be recovered, retroactively approved, tagged or uploaded. Its committed bytes stay as they are and the line advances to a freshly generated candidate. Until a 3.3 candidate publishes, the committed rc.1 version also blocks app deployment at the pending fence.

Current state of rc.1, checked before opening this PR:

  • no v3.3.0-rc.1 tag and no GitHub Release;
  • no rc.1 objects visible on the CDN. /schemas/, /compliance/ and /protocol/ discovery (built from R2 prefix listings) list 3.3.0-rc.0 as the newest 3.3 candidate. /schemas/3.3.0-rc.1/index.json and /compliance/3.3.0-rc.1/index.json answer 200 with the rc.0 bytes (same ETag as rc.0) and cache-control: public, no-cache, must-revalidate, which is the pinned-semver fallback, not an exact hit. A schema that exists only in rc.1 (core/impression-basis-breakdown-entry.json) returns 404, and /protocol/3.3.0-rc.1.tgz returns 404;
  • the release run failed before any publication step, and deploy run 44302651225 stopped at "Require committed version to have a complete release", so no latest upload ran from rc.1.

What this does

This reuses the supersession machinery from #7663 and #7667 (scripts/check-release-supersession.cjs, the prepare_next dispatch in release.yml, marker consumption in scripts/version-packages.mjs, and the pending fence's supersession path). It changes no workflow or release script logic.

  • Adds the hash-bound .changeset/release-supersession.json:

    {
      "version": "3.3.0-rc.1",
      "release_commit": "32a90ee0d0fd7d66505bb14277a84eeb6d15fb79",
      "protocol_sha256": "5698355f3be018c5f1b68f8417b32f4f084ee6ff7bd1e3b9ae15febaffd43a81",
      "reason": "3.3.0-rc.1 cannot publish: its reviewed release head b46ec2e1804fb9fae08b3a4aede4043c235bbe7f was generated on 61a1638493297b2119a41e4cd918e00fc42a5811, and #9080 (5eb0b201f4c9e3ac4bb92ddc580c0582ecb97280, no changeset) merged before #9073 was squash-merged as 32a90ee0d0fd7d66505bb14277a84eeb6d15fb79, so the merge base parent and tree differ from the reviewed head and the provenance gate quarantined release run 44300240363. Preserve its committed bytes as unpublished and generate a freshly reviewed rc.2.",
      "changesets": [
        { "file": "supersede-unpublishable-3-3-0-rc-1.md", "sha256": "8e1320329bbf0f9620a6491fbfae871fd0346d208d6a8719015ae0e283c34d0b" }
      ]
    }

    protocol_sha256 is computed from the committed blob (git cat-file blob 32a90ee0d0:dist/protocol/3.3.0-rc.1.tgz) and matches the committed .sha256 sidecar. There were no pending changesets on main, so the pool is just this PR's patch changeset.

  • Adds the patch changeset supersede-unpublishable-3-3-0-rc-1.md. changeset status reports adcontextprotocol -> 3.3.0-rc.2.

  • Marks 3.3.0-rc.1 unpublished in all three release-status maps: scripts/build-schemas.cjs, server/src/schemas-middleware.ts and workers/artifact-cdn/src/index.js. Updates the committed dist/schemas/index.json entry to match the generator (stability: unpublished, published: false). The exact rc.1 bytes stay addressable but can never win an alias or be the fallback for a later pinned request.

  • Leaves dist/schemas/3.3.0-rc.1/, dist/compliance/3.3.0-rc.1/ and the rc.1 protocol tuple untouched (the checker enforces this against the release merge).

  • Records the case in RELEASING.md, and adds a "Not published" row to docs/reference/versions.mdx, as was done for 3.2.0-rc.5.

  • Tests:

    • tests/release-supersession.test.cjs: a fixture for the rc.1 shape (an unrelated no-changeset commit after the release merge is tolerated; a late changeset or edited candidate bytes are refused), plus an offline check that the repository's own marker binds package.json, the committed tarball digest and sidecar, the release merge (when present in the clone), the unpublished status and the exact pending pool. The check skips once the next Version Packages cut consumes the marker.
    • tests/schema-release-status.test.ts: rc.1 is non-selectable in build, middleware and worker. rc.0 and rc.2 stay selectable. The three maps stay identical, and the generated and committed discovery entries match.
    • tests/artifact-cdn-worker.test.cjs: the worker lists rc.1 as unpublished, and a missing rc.2 falls back to rc.0, never to rc.1, even if rc.1 objects existed in R2.
    • tests/release-workflow-immutability.test.cjs: RELEASING.md's dispatch command uses exactly the declared release_commit and prepare_next inputs. The dispatch re-verifies the marker before the strict branch fence.

After merge

main must stay frozen from this PR's merge until 3.3.0-rc.2 publishes. Any new changeset invalidates the marker. Any other merge moves main and stales the dispatch and the generated PR.

  1. Dispatch the supersession on the exact merged main SHA:

    MAIN_SHA=$(git ls-remote origin refs/heads/main | cut -f1)
    gh workflow run release.yml -R adcontextprotocol/adcp --ref main \
      -f release_commit="$MAIN_SHA" \
      -f prepare_next=true

    MAIN_SHA must be this PR's merge commit. The dispatch re-verifies the marker against live GitHub state and then runs the full storyboard matrix in verify-release (about two hours; a dispatch is never generation_only). Only then does Changesets generate the rc.2 Version Packages PR as a Draft. main must not move during that window.

  2. Review the generated rc.2 Version Packages PR: package and lockfile at 3.3.0-rc.2; .changeset/pre.json still pre/rc; the marker and changeset consumed; rc.2 schemas, compliance and the signed protocol tuple generated; the rc.1 directories and tuple unchanged.

  3. Get a non-author maintainer approval on its final head.

  4. Merge it with main unmoved, so the merge's base parent and tree equal the approved head.

Note: the push run for this PR's merge goes through the same gates. pending accepts the verified marker, and an ordinary push is generation_only, so that run may already open the rc.2 Draft PR. The explicit dispatch is still the documented authority. If it regenerates the PR, review and approve the head the dispatch leaves. The same pending fence gates deploy.yml, so app deployment can resume once this merges. The CDN worker change takes effect when the worker is next deployed.

Validation

  • GITHUB_REPOSITORY=adcontextprotocol/adcp node scripts/check-release-supersession.cjs verify (live GitHub): Verified reviewed supersession of unpublished 3.3.0-rc.1; next Version Packages cut may proceed.
  • check-release-state.cjs pending with the marker on the current main tree: passed.
  • npx changeset status: adcontextprotocol -> 3.3.0-rc.2. check-changeset-protocol-scope.cjs: passed.
  • Generated discovery (buildRootSchemaDiscovery()) equals the committed dist/schemas/index.json.
  • test:release-workflow parts: release-workflow-immutability, release-supersession (4/4), changesets-fence, release-publication-*, skip-withdrawn-release, promote-release-candidate, published-compliance-versions, artifact-cdn-worker, schema-discovery-pointers and recover-cdn-version-trees all passed locally. Local-only environment failures, the same on unmodified main: macOS BSD stat (rerun with GNU coreutils), two maintenance-release-publication fixture tests, and one cdn-artifacts JSONL test that needs dist/protocol tarballs missing from this sparse checkout. CI runs the full suite.
  • tests/schema-release-status.test.ts and server/tests/integration/schema-routing.test.ts: passed.
  • npm run typecheck: passed.
  • npm run test:unit: 81 files / 1,284 vitest tests plus 182 node tests passed.
  • The pre-commit hook passed on the commit, including the full server unit suite (855 files: 14,587 passed, 31 skipped) and typecheck. Hooks ran normally; no --no-verify.
  • The full storyboard matrix is left to CI.

🤖 Generated with Claude Code

The 3.3.0-rc.1 Version Packages head b46ec2e was generated on
61a1638 (#8736). #9080 merged afterwards without a changeset, so the
bot never regenerated the PR, and #9073 was squash-merged as 32a90ee
on base 5eb0b20 with a tree that differs from the reviewed head.
Release run 44300240363 quarantined it at the provenance gate.

Following RELEASING.md "Superseding an unpublishable committed
prerelease", bind the committed candidate in a reviewed
.changeset/release-supersession.json (release merge, committed tarball
digest, reason, pending changeset pool), add the patch changeset that
advances the line to 3.3.0-rc.2, and mark 3.3.0-rc.1 unpublished in the
build, server and CDN worker release-status maps and the committed
discovery index. The rc.1 schema, compliance and protocol bytes are
unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

@aao-secretariat aao-secretariat Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

Ladon verdict: Approve

Approve — clean release-ops PR with no blocking or medium findings.

This PR supersedes the unpublishable 3.3.0-rc.1 by reusing the existing #7663/#7667 supersession machinery with no workflow or script logic changes. The reviewer verified:

  • The 3.3.0-rc.1 → unpublished override is consistent across all four sites (build-schemas, server middleware, CDN worker, committed dist index).
  • The hand-edited dist/schemas/index.json entry is byte-for-byte what the generator emits.
  • Immutable per-version artifacts under dist/** are untouched (no modification/rewrite of an existing semver artifact — not a critical immutability violation).
  • The patch changeset correctly resolves to rc.2.
  • Fallback/latest_stable never resolve to the superseded rc.1 bytes.

High-risk flag is true, but all three reasons are (modified) files with no accompanying medium-or-higher finding — presumed safe per the change-kind guidance; no (deleted) entries. No gated paths (gated_paths: false), no no-auto-approve team match, no prior decision. No Critical/High/Medium findings. Rows 1–8 do not fire; falls through to row 9 → approve.

@bokelley
bokelley merged commit 859ea91 into main Nov 18, 2026
62 checks passed
@bokelley
bokelley deleted the fix/supersede-3-3-0-rc-1 branch November 18, 2026 16:41
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant