conda

The conda channel for clice: toolchain packages we build ourselves and do not publish to conda-forge, starting with xclang.

[workspace]
channels = ["conda-forge", "https://conda.clice.io"]

Needs pixi >= 0.40, conda >= 24.5 or mamba >= 2.0.

How it works

The channel is static files, no server:

• Every version of every package is one GitHub release, tagged <name>-<version>, holding that version's files for all platforms.
• <subdir>/repodata.json lives on the gh-pages branch, served at https://conda.clice.io. Its info.base_url (CEP-15) is https://dl.clice.io/<subdir>/, a Cloudflare Worker that redirects each file to the release holding it.

Lock files record the dl.clice.io URL, never the GitHub one, so the files can move to other storage later without breaking them.

A published file name never changes content. Publishing the same name with different bytes is refused; bump the build number instead.

Publishing

A repository publishes its packages with this repository's action and the organization's UPLOAD_CONDA token:

- uses: clice-io/conda@main
  with:
    packages: dist/*.conda
    token: ${{ secrets.UPLOAD_CONDA }}

Any repository of the organization can publish any number of packages, for any subdir; every run gets its own pull request, so publishers never wait on or overwrite each other. The action uploads each file to its release and commits its repodata record to index/<subdir>/<file>.json through a pull request that merges itself once check has passed: every record must name a file its release holds with those bytes, and no published record may change. On main, pages rebuilds <subdir>/repodata.json on gh-pages from index/.

To take a package down, delete its files from the release (or the release itself), then remove their records in a pull request; check refuses a removal while the file is still there.

action.yml          the action
scripts/publish.ts  upload and write the records
scripts/check.ts    the check pull requests wait for
scripts/pages.ts    index/ -> <subdir>/repodata.json
index/              one record per published file