> ## Documentation Index
> Fetch the complete documentation index at: https://docs.novita.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Hugging Face OpenEnv

[OpenEnv](https://github.com/huggingface/OpenEnv) is an end-to-end framework for creating, deploying, and using isolated execution environments for agentic reinforcement learning (RL) training. It uses simple, Gymnasium-style APIs so training code and agents can interact with many kinds of environments through a consistent interface.

Novita Sandbox is now integrated with OpenEnv as a provider for running OpenEnv server environments in isolated cloud sandboxes.

## Run a Tbench2Env Server

The following example uses `NovitaSandboxProvider` to build and start the Tbench2 OpenEnv server in a Novita Sandbox. It then connects to the server with `Tbench2Env`, resets a task, executes commands, and releases the sandbox when the run finishes.

The example is based on [`examples/novita_tbench2_simple.py`](https://github.com/huggingface/OpenEnv/blob/main/examples/novita_tbench2_simple.py) in the OpenEnv repository.

Set `NOVITA_API_KEY` and run the example from the root of a local OpenEnv checkout:

```bash CLI icon="terminal" theme={"system"}
export NOVITA_API_KEY="<YOUR_NOVITA_API_KEY>"
PYTHONPATH=src:envs uv run python examples/novita_tbench2_simple.py
```

```python Python icon="python" theme={"system"}
import asyncio
import os

from openenv.core.containers.runtime.novita_provider import NovitaSandboxProvider
from tbench2_env import Tbench2Action, Tbench2Env


async def main() -> int:
    tasks_dir = os.environ.get("TB2_TASKS_DIR")
    if not tasks_dir:
        print("TB2_TASKS_DIR not set. TB2 repo will be downloaded.")

    task_id = os.environ.get("TB2_TASK_ID", "headless-terminal")

    image = NovitaSandboxProvider.image_from_dockerfile(
        "envs/tbench2_env/server/Dockerfile",
    )
    provider = NovitaSandboxProvider()
    # The first start builds a Novita template; later starts reuse the cache.
    base_url = provider.start_container(image=image)
    provider.wait_for_ready(base_url, timeout_s=300)

    try:
        async with Tbench2Env(base_url=base_url, provider=provider) as env:
            result = await env.reset(task_id=task_id)
            print("Instruction head:")
            print(result.observation.instruction[:200])

            result = await env.step(
                Tbench2Action(action_type="exec", command="ls -la")
            )
            print("Command output:")
            print(result.observation.output)

            result = await env.step(
                Tbench2Action(
                    action_type="exec", command="curl -v http://127.0.0.1:8000"
                )
            )
            print("Command output:")
            print(result.observation.output)
    finally:
        provider.stop_container()

    return 0


if __name__ == "__main__":
    raise SystemExit(asyncio.run(main()))
```

For a complete end-to-end evaluation, see [`examples/novita_tbench2_e2e_eval.py`](https://github.com/huggingface/OpenEnv/blob/main/examples/novita_tbench2_e2e_eval.py):

```bash CLI icon="terminal" theme={"system"}
export NOVITA_API_KEY="<YOUR_NOVITA_API_KEY>"
export API_KEY="<YOUR_MODEL_API_KEY>"
export API_BASE_URL="<YOUR_MODEL_BASE_URL>"

PYTHONPATH=src:envs uv run python examples/novita_tbench2_e2e_eval.py \
  --task regex-log \
  --model "<YOUR_MODEL_NAME>" \
  --max-steps 20 \
  --verbose
```