A bundle used to cost a second pull request whose entire content was a number.
The workflow published only when a merge to `main` left `module.json` at a
version with no release yet, so a merge that did not touch that line released
nothing. Measured rather than argued: v0.3.0 (2026-08-12) is the only release
this repo has ever cut, while the nine Teams phases and the cutover landed on
`main` in the week since.
Adopt the engine `link` and `installer` already run - feat!/BREAKING CHANGE ->
major, feat -> minor, fix|perf -> patch, nothing releasable -> no release. The
number that ships is the tag, and CI writes it into the `module.json` inside the
bundle, with an assertion that the rewrite happened: a bundle carrying the wrong
version would install under a number that is not the one it was released as.
`module.json`'s version is kept as a floor rather than deleted - a version above
the newest tag still releases at that version - so the declared model survives as
the special case it always was, and is still how a `coreApi` bump overrules the
subjects. `workflow_dispatch` covers what the rules cannot reach: leave `version`
blank to bump the newest tag by `bump`, or name an exact version. Where the log
and the input disagree the larger bump wins, because a button pressed on a log
full of `feat:` would otherwise publish its `patch` default over a minor's worth
of work.
Two things carried over from `link` at the same time: a tag pushed without a
release behind it is now recovered rather than making that state permanent, and
the changelog moved into the plan step (so the assemble step clears `$OUT`, not
`dist/`, which now holds it).
On this repo's `main` the engine computes v0.4.0. The workflow still never
writes to a branch.
Co-Authored-By: Claude <noreply@anthropic.com>
wtclaude
8ec21086b5