Shipping huggingface_hub every week with AI, open tools, and a human in the loop

Summary

huggingface_hub, the Python client underlying libraries such as transformers, datasets, diffusers, and sentence-transformers, moved from releases every four to six weeks to a weekly cadence. A single manually triggered GitHub Actions workflow now computes versions, manages branches, tags and publishes both huggingface_hub and the hf CLI, creates downstream RC test branches, and handles post-release tasks. An open-weights model drafts structured release notes and Slack announcements using PR metadata and documentation diffs, while OpenCode, HF Inference Providers, PyPI Trusted Publishing, and pinned, checksum-verified tooling provide the stack. Deterministic manifest validation checks missing or extra PR references, then a human reviews and publishes the drafts. The reported result is weekly shipping at about $0.25 per full release, with more consistent notes, earlier downstream breakage detection, and shorter contributor feedback loops.

Takeaways

1. 1

   The workflow validates AI-generated release notes against a deterministic manifest of PRs from squash-merge commits, identifying missing or extra references and re-prompting the agent to correct discrepancies.

2. 2

   Documentation diffs from changed Markdown files under docs/ are supplied to the model so release-note details such as CLI examples are grounded in the actual PR documentation.

3. 3

   The reported cadence increased to weekly releases, with full release-note and announcement generation costing about $0.25; downstream RC branches also surface integration breakages earlier.