hf-space-recovery - SKILL.md Agent Skill

name: hf-space-recovery description: Diagnose and recover failing or stuck Hugging Face Space deployments for OpenEnv environments. Use when deploying envs from `envs/` to the Hub (`openenv` namespace with version suffixes), when Spaces are in `BUILDING`/`APP_STARTING`/`RUNTIME_ERROR`, or when release collections need to be reconciled after targeted redeploys.

HF Space Recovery

Use this skill to recover OpenEnv Hub deployments quickly with minimal blast radius.

Execute This Workflow

1) Confirm release tuple

Use a single release tuple across all commands:

Namespace: openenv
Version: vX.Y.Z
Space suffix: -vX-Y-Z

Default to a version suffix and treat unsuffixed Spaces as legacy.

2) Snapshot runtime status

Collect all versioned spaces and isolate non-running ones:

hf spaces ls --author openenv --limit 500 --expand=runtime \
  | jq -r '.[] | select(.id|test("-v[0-9]+-[0-9]+-[0-9]+$")) \
    | [.id, .runtime.stage, (.runtime.raw.errorMessage // "")] | @tsv' \
  | sort

Treat RUNNING and SLEEPING as healthy. Triage everything else.

3) Classify and extract signal

RUNTIME_ERROR: read traceback from .runtime.raw.errorMessage.
BUILD_ERROR: read build error text from runtime info, then patch Dockerfile/deps.
APP_STARTING longer than 10 minutes: inspect event stream and metrics before changing code.

hf spaces info openenv/<space-id> --expand=runtime
curl -sS -m 10 https://huggingface.co/api/spaces/openenv/<space-id>/events | sed -n '1,140p'
curl -sS -m 10 -i https://huggingface.co/api/spaces/openenv/<space-id>/metrics | sed -n '1,120p'

Read references/troubleshooting.md for symptom-to-fix mappings.

4) Apply minimal fix and targeted redeploy

Prefer targeted redeploys over full-fleet pushes:

scripts/prepare_hf_deployment.sh \
  --hf-namespace openenv \
  --env <env_name> \
  --skip-collection

Use openenv CLI as a supplement, not a replacement, for release triage:

Validate env layout quickly (uv run openenv validate ...) when applicable.
Keep release deploys on scripts/prepare_hf_deployment.sh to preserve suffix/pinning behavior.

5) Unstick runtime when code is already good

If Space remains in APP_STARTING with no actionable error:

uv run --with huggingface_hub python - <<'PY'
from huggingface_hub import HfApi
api = HfApi()
api.restart_space("openenv/<space-id>", factory_reboot=True)
PY

If still stuck, force recreation as last resort:

hf repo delete openenv/<space-id> --repo-type space
scripts/prepare_hf_deployment.sh --hf-namespace openenv --env <env_name> --skip-collection

6) Verify and close

Verify both runtime stage and health endpoint:

hf spaces info openenv/<space-id> --expand=runtime
curl -sS -m 10 https://<space-subdomain>.hf.space/health

Then verify fleet-wide:

hf spaces ls --author openenv --limit 500 --expand=runtime \
  | jq -r '.[] | select(.id|test("-v[0-9]+-[0-9]+-[0-9]+$")) \
    | select(.runtime.stage!="RUNNING" and .runtime.stage!="SLEEPING") \
    | [.id, .runtime.stage] | @tsv' | sort

7) Reconcile collection

When targeted deploys are done, update collection membership for the same version:

python3 scripts/manage_hf_collection.py \
  --version vX.Y.Z \
  --collection-namespace openenv \
  --space-id openenv/<space-id>

Add one --space-id per redeployed space.