Triage Hugging Face discussions on your models and datasets

Every weekday at 9am, sweep the discussions and pull requests on every model and dataset my organization publishes on Hugging Face, and triage whatever came in since the last run.

Start by enumerating our repos. Use Hugging Face List Models and List Datasets, both filtered with the author set to my organization name. These list endpoints are cursor paginated, so follow the pagination all the way through instead of stopping at the first page, otherwise repos get missed once we publish more than one page worth.

For each repo, call List Discussions. Repo ids are namespace/name, and repo scoped paths split them into separate namespace and repo segments, so handle that when building each call. This single endpoint returns both discussions and pull requests, so keep track of which is which. Skip automated bot pull requests such as model card updates, config bumps, and similar housekeeping opened by bot accounts, because those are high volume and would drown out the real threads.

For anything opened or updated since the last run, call Get Discussion to read the full thread along with its events. Monday's run should cover the whole weekend, not just Sunday.

Classify each thread as one of these: a bug report, a feature or model request, a usage question, a licensing or access question, or spam.

For genuine bug reports, create a Linear issue with Create Issue. Each issue should include the repo name, a link back to the discussion, the reporter's environment details if they gave any, and a short summary of what they hit. Do not open a second issue for a discussion that already produced one on an earlier run.

Then post a single digest to our ML team channel with Slack Bot Send a Message. One message, not one per repo. Group the threads by category, lead with the unanswered usage questions since those are the ones that cost us goodwill, and flag anything that has been sitting without a reply for more than 48 hours. If nothing new came in, post one short line saying so rather than an empty digest.

Finally, for stale usage questions only, meaning the ones with no reply for over 48 hours, leave a short acknowledgment on the thread with Comment on Discussion so the community knows we saw it. Never auto reply to bug reports or to licensing and access questions, because those need a human.