generate-openenv-env

作成者: huggingface

Generate OpenEnv environments from a concrete use case (for example, "generate an env for the library textarena"). Use when asked to design or implement a new…

npx skills add https://github.com/huggingface/openenv --skill generate-openenv-env

/generate-openenv-env

Build a production-ready OpenEnv environment from a use-case prompt.

Execute Workflow

When invoked, execute this workflow end-to-end.

1. Parse the use case and name the environment

Derive a repo path in the form envs/<name>_env/.

  • Normalize to snake_case.
  • Keep names short and domain-specific.
  • Example: "generate an env for the library textarena" -> envs/textarena_env/.

2. Research the target library/API before coding

Gather the minimum interface facts needed to implement reset, step, and state serialization.

  • Search local docs/examples first.
  • Search upstream docs/repo for the target library when local context is insufficient.
  • Extract only implementation-critical details:
    • installation/dependency requirements
    • environment creation API
    • action format
    • observation format
    • reward and done semantics
    • special setup (model files, downloads, auth, etc.)

3. Mine matching OpenEnv examples

Select 2-3 existing environments as implementation templates.

  • Always read references/openenv-tutorial-01-environments.md (Part 10) and references/openenv-docs-environment-builder.md.
  • Prefer envs/textarena_env for external-library wrappers with richer state.
  • Add one simpler baseline (for example envs/snake_env or envs/echo_env) to keep the implementation minimal.
  • Follow patterns, do not copy blindly.
  • Exclude generated or vendored files when mining examples (.venv/, build/, site-packages/, __pycache__/).

For a compact checklist and mapping, read references/env-generation-checklist.md.

4. Ask focused implementation questions

Ask only the questions that materially affect architecture. Use the question bank in references/env-generation-checklist.md.

Cover at least:

  • action space contract
  • observation fields needed by agents
  • reward design and terminal conditions
  • episode/session configuration knobs
  • deployment target and dependency constraints

If answers are unavailable, proceed with explicit assumptions and document them.

5. Choose the environment archetype

Choose one archetype before scaffolding:

  • Typed step/reset environment (default): use EnvClient + typed Action/Observation[/State] models.
  • MCP tool environment: use MCPEnvironment + MCPToolClient and MCP action/observation types.
  • Specialized client flow (rare): only when the standard clients cannot express required behavior (for example local+remote hybrid clients).

6. Scaffold the environment

Use the CLI to scaffold:

PYTHONPATH=src uv run openenv init <name>_env --output-dir envs

This generates all files with correct placeholders replaced, including pyproject.toml, Dockerfile, and uv.lock.

If the CLI is unavailable (import errors, missing dependencies), create the structure manually matching:

envs/<name>_env/
├── __init__.py
├── client.py
├── models.py
├── openenv.yaml
├── pyproject.toml
└── server/
    ├── __init__.py
    ├── app.py
    ├── <name>_environment.py
    └── Dockerfile

Use assets/openenv_env_template/ as a reference for file contents when scaffolding manually.

7. Implement with OpenEnv contracts

Implement these files in order:

  1. models.py
  2. server/<name>_environment.py
  3. server/app.py
  4. client.py
  5. openenv.yaml
  6. README.md

Use these standards:

  • Use typed models (Action/Observation/State).
  • Use create_app(<factory_or_class>, ActionType, ObservationType, env_name=...) in server/app.py. Pass a class or factory callable, not an instantiated environment.
  • Dual-import pattern (required in server/app.py and server/<name>_environment.py): Use try: from ..models import X / except ImportError: from models import X. Relative imports work in-repo (PYTHONPATH=src:envs); bare imports work in Docker (PYTHONPATH=/app/env). The same pattern applies to intra-server imports (e.g., from .foo import Bar vs from server.foo import Bar).
  • client.py uses EnvClient[ActionType, ObservationType, State] (three type parameters).
  • Keep server logic in server/, keep client parsing in client.py.
  • Expose config through environment variables when behavior is likely to vary.
  • Keep reward logic inside the environment.
  • Prefer reset/step signatures compatible with Environment:
    • reset(seed=None, episode_id=None, **kwargs)
    • step(action, timeout_s=None, **kwargs)
  • Set SUPPORTS_CONCURRENT_SESSIONS=True only when isolation is real. Set max_concurrent_envs in create_app accordingly (1 when False, >1 when True).
  • For MCP/tool-call UIs that send stringified JSON arguments, add action validators/parsers in server/app.py.
  • Export public client/models symbols in __init__.py.
  • Keep openenv.yaml aligned with current scaffold format (spec_version: 1, name, type, runtime, app, port).
  • Avoid training/evaluation code paths in this skill.

8. Validate before handoff

Run the narrowest useful checks:

# Verify in-repo imports work (catches missing dual-import pattern)
PYTHONPATH=src:envs uv run python -c "from envs.<name>_env.server.<name>_environment import <ClassName>Environment"

# Build and validate
cd envs/<name>_env
openenv build
openenv validate --verbose
PYTHONPATH=src:envs uv run pytest envs/<name>_env -q

If tests do not exist, run a smoke check:

PYTHONPATH=src:envs uv run uvicorn envs.<name>_env.server.app:app --port 8000
curl http://localhost:8000/health
openenv validate --url http://localhost:8000

9. Deliver with assumptions and gaps

Report:

  • files created/updated
  • chosen archetype (typed vs MCP vs specialized)
  • assumptions made due to missing answers
  • validation commands executed and outcomes
  • remaining risks or follow-up questions

Guardrails

  • Do not route into model training/evaluation workflows.
  • Do not invent library APIs; confirm against source docs.
  • Do not skip reading at least one existing OpenEnv env before implementation.
  • Do not copy outdated manifest patterns from older envs (name/version/action/observation-only manifests).
  • Do not copy build artifacts or virtualenv files from example envs.
  • Do not set max_concurrent_envs > 1 unless the environment explicitly supports concurrent sessions.

huggingfaceのその他のスキル

cpu-kernels
huggingface
C++ CPUカーネルをSIMDイントリンシクス(AVX2/AVX512)を用いて記述、最適化、ベンチマークするためのガイダンスを提供します。Hugging Faceカーネルエコシステム向けです。含まれるもの…
official
generate-openenv-env
huggingface
具体的なユースケース(例:「ライブラリtextarena用のenvを生成する」)からOpenEnv環境を生成します。新しい環境の設計や実装を求められた際に使用します。
official
hf-mcp
huggingface
Hugging Face HubをMCPサーバーツール経由で利用します。モデル、データセット、Spaces、論文を検索できます。リポジトリの詳細を取得し、ドキュメントを取得し、計算ジョブを実行し、Gradioを使用します…
official
trl-training
huggingface
TRL(Transformers Reinforcement Learning)を使用してトランスフォーマー言語モデルをトレーニングおよびファインチューニングします。SFT、DPO、GRPO、KTO、RLOO、および報酬モデルのトレーニングをサポートしています…
official
deploy-hf
huggingface
OpenEnv環境をHugging Face Spacesにデプロイします。デプロイ、Hugging Faceへのプッシュ、またはスペースの更新を求められたときに使用します。
official
hf-space-recovery
huggingface
Diagnose and recover failing or stuck Hugging Face Space deployments for OpenEnv environments. Use when deploying envs from `envs/` to the Hub (`openenv`…
official
pre-submit-pr
huggingface
プルリクエストを送信する前に変更を検証します。lint、テスト、アライメントレビュー、RFC分析を含む包括的なチェックを実行します。作成前に使用…
official
example-skill
huggingface
アクションスモークテスト用のフィクスチャスキルの例
official