Skip to content

Releases: markdown-office/mdoffice

Release list

v0.1.1

Choose a tag to compare

@github-actions github-actions released this 26 Oct 13:52
94a8802

The first release that was published. The v0.1.0 tag was cut and the release
was attempted, and it never happened: the pipeline stopped on the macOS build for
Intel, and a release waits for every binary, so nothing was published on any
platform. What this release adds is the fix for that and the corrections that
went with it.

The features are in the notes for 0.1.0, which is what the v0.1.0 tag holds
and what a consumer of that tag would have got.

Changed

  • The install section says the package is not on Packagist. It is not, and the
    README's first instruction was composer require, which does not resolve. What
    works today is a clone or a release; the Composer install is kept for the day the
    package is published, with the plugin a consumer has to allow. The 0.1.0 notes
    below said the package was on Packagist, which was never true, and no longer say
    so.

  • The path repositories are not stripped by Packagist. Composer reads
    repositories from the root package only, so an installed copy never sees the two
    path entries whatever Packagist does with them. The README gave a different
    reason, and the reason it gave was not the reason.

  • markdown-presentation reads a deck back, and the README described it as
    writing one only.

Fixed

  • The release pipeline signs the interpreter when upstream published it
    unsigned.
    Upstream publishes an ad-hoc signature with the arm64 build and
    none at all with the Intel one, and the pipeline stopped on the Intel one —
    so nothing was published on any platform, because the release waits for every
    binary. The signature is put on here rather than refused for want of it: the
    same kind of signature the arm64 build already carries, and the one the
    kernel asks for.

  • The archive is stamped with the commit's date. The step that read it
    named a directory that is not there, and a git error inside a command
    substitution is swallowed by the echo around it — the step succeeded, wrote
    nothing, and the archive fell back to the modification times of the sources.
    The date is now read from the repository the step runs in, and the build stops
    rather than writing an empty one.