Deploy Hf Agent Skill

Deploy to Hugging Face Spaces

Deploy an OpenEnv environment to Hugging Face Spaces using the OpenEnv CLI.

When to Use This Skill

• User asks to "deploy to Hugging Face"
• User says "push to Hugging Face Spaces"
• User wants to update an existing space
• After implementing new features that need to be tested in production

Prerequisites

Before deploying, ensure:

• The environment has an openenv.yaml file
• The environment has a server/Dockerfile
• You have Hugging Face credentials configured (automatic via huggingface-cli)

Instructions

1. Identify the Environment

Determine which environment to deploy:

• If user specifies: use that environment (e.g., "carla_env", "browser_env")
• If in environment directory: use current directory
• Otherwise: ask the user

2. Determine the Repository ID

Repository ID format: username/space-name

• If user provides full ID: use it (e.g., "sergiopaniego/carla-env-real-updated")
• If user provides only space name: construct ID with their username
• Check openenv.yaml for default repo-id
• Otherwise: ask the user

3. Pre-Deployment Setup

IMPORTANT: Always run from the project root directory.

Before deploying, ensure OpenEnv is installed:

cd /path/to/OpenEnv  # Navigate to project root if needed
uv pip install -e .

If this fails with "does not appear to be a Python project", you're not in the project root.

4. Run the Deployment Command

Execute the deployment:

PYTHONPATH=src uv run python -m openenv.cli push <environment-dir> --repo-id <username/space-name>

Parameters:

• <environment-dir>: Path to environment (e.g., envs/carla_env)
• --repo-id: Hugging Face Spaces repository ID (e.g., sergiopaniego/carla-env-real-updated)

Optional flags:

• --private: Deploy as a private space
• --no-interface: Disable the web interface (deploy API-only)
• --base-image <image>: Override the base Docker image
• --hardware <hw> / -H <hw>: Request Hugging Face Space hardware (e.g. t4-medium, a10g-small, cpu-basic)

5. Verify Deployment

After successful deployment:

• Note the Space URL returned by the command
• Wait for build to complete:
• CPU environments: ~5 minutes
• GPU environments (CARLA): ~30-60 minutes
• Check the space status at the URL
• Test with a simple health check once build completes:

   curl https://<username>-<space-name>.hf.space/health

Example Usage

Deploy carla_env to existing space

PYTHONPATH=src uv run python -m openenv.cli push envs/carla_env --repo-id sergiopaniego/carla-env-real-updated

Deploy echo_env as private space

PYTHONPATH=src uv run python -m openenv.cli push envs/echo_env --repo-id username/my-echo-env --private

Deploy with GPU hardware

PYTHONPATH=src uv run python -m openenv.cli push envs/carla_env --repo-id username/carla-env --hardware t4-medium

Deploy with custom base image

PYTHONPATH=src uv run python -m openenv.cli push envs/browser_env --repo-id username/browser-env --base-image nvidia/cuda:11.8.0-runtime-ubuntu22.04

Output Format

Report deployment status:

## Hugging Face Deployment

### Environment
- Environment: <env-name>
- Directory: <path>
- Dockerfile: <path-to-dockerfile>

### Deployment
- Repository ID: <username/space-name>
- Space URL: <https://huggingface.co/spaces/username/space-name>
- Status: ✓ Deployed successfully

### Next Steps
1. Wait for space to build (5 min for CPU, 30-60 min for GPU/CARLA)
2. Visit space URL to check build status
3. Test environment once build completes

### Testing Commands

Health check

curl https://<username>-<space-name>.hf.space/health

Reset environment

curl -X POST https://<username>-<space-name>.hf.space/reset

Step action

curl -X POST https://<username>-<space-name>.hf.space/step \

-H "Content-Type: application/json" \

-d '{"action_type": "observe"}'

Troubleshooting

Error: "ModuleNotFoundError: No module named 'openenv'"

Solution: Install OpenEnv first (must be run from project root):

cd /path/to/OpenEnv  # Navigate to project root
uv pip install -e .

Error: "does not appear to be a Python project"

Cause: You're not in the project root directory.

Solution: Navigate to the OpenEnv project root where pyproject.toml exists:

cd /Users/sergiopaniegoblanco/Documents/Projects/OpenEnv  # Adjust path
uv pip install -e .

Error: "Directory does not exist"

Solution: Ensure you're passing the correct environment directory path:

# Correct
PYTHONPATH=src uv run python -m openenv.cli push envs/carla_env --repo-id ...

# Incorrect
PYTHONPATH=src uv run python -m openenv.cli push carla_env --repo-id ...

Error: "Authentication required"

Solution: Login to Hugging Face CLI first:

huggingface-cli login

Space build fails

Solutions:

• Check Dockerfile syntax and dependencies
• Verify hardware requirements (GPU spaces need --hardware setting on HF)
• Check space logs on Hugging Face for detailed errors
• Ensure openenv.yaml is valid

Common Environments

| Environment | Path | Typical Repo ID | Hardware |

|-------------|------|-----------------|----------|

| carla_env (standalone) | envs/carla_env | username/carla-env-real | GPU (T4/A10G) |

| carla_env (mock) | envs/carla_env | username/carla-env-mock | CPU |

| echo_env | envs/echo_env | username/echo-env | CPU |

| browser_env | envs/browser_env | username/browser-env | CPU |

| tbench2_env | envs/tbench2_env | username/tbench2-env | CPU |

Notes

• Deployment requires Hugging Face authentication (automatic if huggingface-cli is logged in)
• By default, spaces are public (use --private for private spaces)
• By default, web interface is enabled (use --no-interface for API-only)
• GPU spaces can request hardware via --hardware (e.g. --hardware t4-medium)
• Build times vary: CPU (~5 min), GPU with CARLA (~30-60 min)
• The CLI automatically moves Dockerfile to repository root for Hugging Face compatibility

Related Documentation

• DEPLOYMENT_GUIDE.md - Detailed deployment modes
• README.md - OpenEnv overview
• Hugging Face Spaces Docs: https://huggingface.co/docs/hub/spaces