Hugging Face went from shipping huggingface_hub every 4 to 6 weeks to shipping every 7 days using a single GitHub Actions workflow that combines an open-weights LLM (GLM-5.2) for drafting release notes with a human-in-the-loop review. The old manual process cost a maintainer half a day of scattered focus; the new pipeline handles 12 jobs automatically and leaves only two decisions to a human: reviewing the draft changelog and posting the internal announcement.
The Old Process: A Half-Day Chore Every Month
The earlier release cadence meant fixes and features rotted on main for weeks. Each minor release required manually creating a release branch, bumping version, committing, tagging, pushing, monitoring downstream CI in transformers, datasets, diffusers, and sentence-transformers, triaging failures, writing release notes from tens of PRs, drafting Slack announcements, and opening a post-release PR. Writing good notes was the heavy part: aggregating PRs by theme, adding context, avoiding a git log dump. Nothing technically hard but a few hours of focused attention spread over days.
The New Pipeline: One Workflow, One Input
The entire automation lives in a single file: .github/workflows/release.yml. It takes exactly one input from the Actions UI: release_type (minor-prerelease, minor-release, or patch-release). From there the jobs run in order: prepare the branch and bump version, publish to PyPI (and the hf CLI as a separate package), diff commits since last tag, have the GLM-5.2 model draft structured release notes, open downstream test branches pinned to the RC for automated CI confirmation, produce a Slack announcement in team voice, archive both AI and human-edited notes to a Hugging Face Bucket, comment on every shipped PR with the version number, sync CLI docs, and post progress as a threaded Slack report. The stack uses GitHub Actions, OpenCode agent, GLM-5.2 served via HF Inference Providers, and PyPI Trusted Publishing. Every component is open, no vendor lock-in required.
Trust but Verify: Why the Model Drafts and Humans Decide
The failure mode everyone worries about is a model quietly dropping a PR or inventing one. A changelog that is almost right is worse than no changelog. So the workflow is designed with a human checkpoint after the draft. A deterministic script verifies the model's output against the actual PR list before the human sees it. The human then reviews and edits before hitting publish. That two-step guard keeps the release notes accurate without sacrificing the 90% speedup from having the model turn thirty terse PR titles into readable prose. The remaining manual steps are reviewing the draft notes and posting the internal Slack message. That is where judgment matters, and that is where a person stays in the loop.
The workflow is open source on GitHub, so any maintainer can adapt it to their own library. No closed model, no proprietary platform, no secret sauce required.
Source: Shipping huggingface_hub every week with AI, open tools, and a human in the loop
Domain: huggingface.co
Comments load interactively on the live page.