Changed#
Merging the release PR is now the release. The workflow gained a decide
job: every push to main runs a seconds-long registry preflight, and the one
push whose manifests name an unpublished version — a merged release PR — goes
on to verify, publish all three packages, create the v<version> tag on the
merge commit, and publish the GitHub release from RELEASES.md. Ordinary
merges skip in seconds; a pushed tag remains the manual override;
workflow_dispatch stays dry-run only. One release at a time, enforced by a
non-cancelling concurrency group — cancelling a publish mid-flight is how a
half-released set of packages happens.
The publish can no longer fail on npm's broken trusted publishing. The
publish steps accept an environment-scoped secrets.NPM_TOKEN as the
authentication fallback: absent, OIDC is the auth exactly as before; present,
the token authenticates and the release goes out either way. Provenance
survives both paths — the attestation is signed with the job's OIDC identity,
which is independent of how the upload authenticates. The security.test.js
rule narrowed rather than vanished: only release.yml may reference the
secret, only as NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}, and npm token
material committed anywhere still fails the build. The old absolute rule
had forced four releases onto a laptop, which protects the secret by moving
the publish to the least auditable place available.
The documentation sweep is enforced, not remembered. publish.test.js
now requires the manifest version to appear as a heading in CHANGELOG.md
and by name in ROADMAP.md — RELEASES.md was already enforced — so a
release prep that skips the docs fails verify before it can merge. The
release checklist in docs/releasing.md adds the grep sweep for the stale
references no test can know about.
The documentation caught up with the registry. 1.25.0 is on npm — published
by hand from a clean clone of the tag after npm's trusted publishing rejected
the workflow's OIDC token on four real attempts against v1.11.0, with every
GitHub-side claim verified correct. The docs now say so instead of describing
the state of a week ago:
RELEASES.mdheader: all three packages at 1.25.0; 1.11.0–1.24.0 named as real releases the registry never saw; no provenance on 1.25.0, and why.docs/releasing.md: the trusted-publisher section downgraded from "reported done" to still not working, with the v1.11.0 evidence that the mismatch lives on npm's side; a Releasing by hand procedure added — the one 1.25.0 actually used — including publishing before pushing the tag, so the workflow's preflight skips the uploads and still creates the GitHub release.README.md: the three action pins advanced from the 1.10.0 commit to the 1.25.0 one, labels verified against the pinned commit bysecurity.test.js.ROADMAP.md: the Released section no longer stops at 1.9.0 — the 1.10.0→1.25.0 arc is recorded, with the registry gap stated.packages/cli/README.md,packages/mcp/README.mdand the Claude skill:profile's rows now name the surface that exists — directories and.gz,--what-if, the drill-downs, the four money gates, the three output formats.