OpenEnv: Skill for Claude Code

Release Skill

This skill orchestrates the repo-embedded deployment tooling and documents the openenv-namespace canonicalization process.

Prerequisites

• hf CLI authenticated (run hf auth login or set HF_TOKEN).
• Local build dependencies installed so scripts/prepare_hf_deployment.sh can stage Docker contexts.
• Access to scripts/manage_hf_collection.py for collection updates and discovery.

Primary Flow

1. Stage envs with scripts/prepare_hf_deployment.sh. Default arguments deploy every deployable env from envs/. Pass --env <name> to target a subset. The script:
   • Resolves the requested OpenEnv ref for staged dependency rewrites. If 0.2.2 is only a release-candidate label and no v0.2.2 tag exists yet, the script should fall back to main for env dependency rewrites while keeping the Hub suffix at -0.2.2.
   • Rewrites loose openenv-core[core]>=... specs and direct Dockerfile installs to git+https://github.com/huggingface/OpenEnv.git@<resolved-ref> so the sweep does not silently install 0.2.1 from PyPI.
   • Builds a staging tree with src/, envs/<env>/, and a rewritten Dockerfile that sets BASE_IMAGE to ghcr.io/huggingface/openenv-base:latest unless a hash is supplied.
   • Generates a README with Hub metadata, enforces openenv/openenv-<version> tags, and adds the HUB_TAG used in collection sync.
   • Uses hf repo create/hf upload plus visibility flags to push Docker spaces.
2. Suffix naming and privacy. Deploy to private spaces named <env>-0.2.2 (set SPACE_SUFFIX=-0.2.2 or rely on the version-derived default). Use --private to keep the collection private for now. The repo deploy script should update only the versioned collection during this phase, not the global tagged collection.
3. Runtime verification. For private Spaces, verify through authenticated hf.space domains, not anonymous browser URLs. Use scripts/verify_private_spaces.py --hf-namespace openenv --suffix -0.2.2 or inspect hf spaces info <space> plus the runtime domain from runtime.raw.domains. Prefer hf spaces ls --expand=runtime to confirm RUNNING/SLEEPING. For stuck or error states, consult scripts/prepare_hf_deployment.sh logic or fall back to the hf-space-recovery skill.
   • Do not treat /health alone as sufficient. This sweep found false positives where /health was 200 but runtime use still failed: chat_env-0.2.2 returned 500 on /reset, and tbench2_env-0.2.2 also returned 500 on /reset.
   • For HTTP envs, an authenticated POST /reset is the minimum usability gate. If the action schema is simple, follow with a schema-correct POST /step probe as well; snake_env, finrl_env, and sumo_rl_env all surfaced step-time failures that a health-only check would miss.
   • Some envs are slow enough that POST /reset may time out on a 20-second probe (git_env, unity_env). Treat those as inconclusive until manually retried with a longer timeout or an env-specific probe.
   • Certain canonical spaces such as openenv/echo_env and the TextArena variants expose multiple READY runtime domains. The verifier now probes every READY domain and reports success once any domain returns usable data, so the canonical status is based on the first passing domain rather than the first listed domain.