The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run and subprocess.Popen in specific reference examples (e.g., in gpu.md and web-endpoints.md) for legitimate purposes such as launching training scripts or model servers. The documentation explicitly warns against building argument lists from unsanitized user input and recommends using fixed lists or validation checks.
[EXTERNAL_DOWNLOADS]: The skill documents methods to download external resources during container image builds, such as using wget or apt_install for system packages and uv_pip_install for Python dependencies. These operations are standard for defining cloud environments and are directed at well-known registries or user-defined URLs for model weights.
[REMOTE_CODE_EXECUTION]: While the platform is designed to execute code in the cloud, the skill emphasizes security best practices. It introduces modal.Sandbox as the recommended mechanism for executing untrusted or agent-generated code, providing isolation through containerization and network egress control using CIDR allowlists.
[CREDENTIALS_UNSAFE]: Authentication is handled safely via environment variables (MODAL_TOKEN_ID, MODAL_TOKEN_SECRET) or the modal.Secret API. The instructions specifically advise against logging or exposing credentials and provide guidance on using .env files for local development in a secure manner, emphasizing that only specific keys should be loaded.

modal