You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Credentials guidance: which credential each submission step needs, and how to supply it #4
Why: The PI raised it on 2026-11-06 during plan A2. Since #3, a Hub token is a person's, not the org's: CI holds none, and whoever pushes or promotes supplies their own. Neither people nor agents have one place that says which credential each step needs, where it comes from, and how it reaches the container.
Context:
What each step needs:
scripts/push and scripts/promote: HF_TOKEN with write access to the org's Hub datasets.
scripts/check.py, scripts/record, and CI's check: no credential.
GHCR pulls of toolshed images: none, since the images are public.
Private repo:brineylab/datasets is currently PRIVATE, so a cluster can't clone it anonymously. The cook copies a clone there (rsync) or needs GitHub access. DESIGN §3 says the repository is public.
How the token is supplied today:
Laptop: HF_TOKEN=$(< ~/.cache/huggingface/token) with podman -e HF_TOKEN.
Cluster: export HF_TOKEN=..., which pyxis passes through to the container.
flyloft: its stored secret.
Never: print the token, commit it to the repo, or set HF_HUB_OFFLINE, which blocks downloads and the push.
Files:CONTRIBUTING.md (cooking and promoting), docs/DESIGN.md §3 Guards and Promoting, AGENTS.md's Never a credential rule, and flyloft's submit_dataset flow.
Done when: CONTRIBUTING has a Credentials section covering each step and each kind of submission (laptop, srun with pyxis, flyloft campaign): which credential, where it comes from, how it reaches the container, and what never to do. DESIGN and AGENTS point to it.
Activity