source: hugging face blog: shipping huggingface_hub every week with ai, open tools, and a human in the loop
level: technical
huggingface_hub, the python client at the base of the hugging face ecosystem, used to release every four to six weeks. the manual process involved writing release notes, triaging downstream tests, and drafting announcements, taking half a day of focused work. the team split the work into mechanical steps, like version bumping and tagging, and brain work, like writing notes. they automated the mechanical parts with a github actions workflow and used an open-weights model to draft the brain work, keeping a human in the loop for final judgment.
the pipeline is triggered manually from the actions ui with a release type. it computes the next version, bumps it, publishes to pypi using trusted publishing, and drafts release notes by diffing commits and pulling pr metadata. a deterministic script checks the model's draft against a manifest of pr numbers to catch missing or invented entries. if discrepancies exist, the agent is re-prompted to fix them. the model also gets documentation diffs from each pr to ground its summaries in real examples. a human reviewer then edits the draft before the final release, turning hours of writing into a fifteen-minute editing session.
the entire stack uses open tools: github actions, opencode agent runtime, an open-weights model served via inference providers, and pypi trusted publishing. security is tightened with oidc tokens and checksum verification. the cost is about $0.25 per release. the new weekly cadence catches integration issues earlier through downstream test branches, shortens contributor feedback loops with automatic comments on shipped prs, and produces better release notes. the workflow is public and designed to be adapted by other maintainers, with the trust-but-verify loop being the key reusable component.
why it matters: this shows how to safely use ai for software release notes by combining model drafts with deterministic verification, a pattern any maintainer can adopt to save time and improve consistency.
source: hugging face blog: shipping huggingface_hub every week with ai, open tools, and a human in the loop