The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documentation describes and provides examples for executing shell commands and subprocesses.
references/web-endpoints.md and references/gpu.md include examples using subprocess.run and subprocess.Popen for training launchers and custom web servers (e.g., launching vllm or accelerate).
references/api_reference.md identifies the .exec() method in modal.Sandbox for running commands within isolated containers.
The documentation explicitly warns users to use fixed, hardcoded argument lists and avoid constructing commands from unsanitized user input to prevent command injection.
[REMOTE_CODE_EXECUTION]: The platform is designed to build and execute code remotely in serverless containers.
SKILL.md and references/images.md detail methods for building container images using modal.Image.run_commands() and modal.Image.run_function(), which execute code during the image build process.
The skill enables remote execution of Python functions via the @app.function() decorator and .remote() calls.
[EXTERNAL_DOWNLOADS]: The skill utilizes several mechanisms for fetching external resources.
SKILL.md instructs the agent to install the modal SDK using uv pip install.
Multiple examples in references/examples.md and references/images.md show the installation of common Python and system packages (e.g., torch, transformers, ffmpeg) via uv_pip_install and apt_install.
The citations section in SKILL.md requires fetching scientific metadata from arxiv.org, a well-known repository for research papers.
[INDIRECT_PROMPT_INJECTION]: The skill provides surface area for indirect prompt injection through data ingestion tools, which it mitigates with security guidance.
Ingestion points: Data enters the agent context through web scraping utilities (httpx.get in references/examples.md), persistent storage (modal.Volume in references/volumes.md), and web endpoints (@modal.fastapi_endpoint).
Boundary markers: The documentation recommends using modal.Sandbox for running untrusted or dynamically generated code to isolate it from the host environment.
Capability inventory: The skill provides full access to file system writes, network operations, and command execution across its various storage and compute modules.
Sanitization: Instructions emphasize validating input against allowlists and using environment variables or files instead of command arguments for user-supplied data.
[DYNAMIC_EXECUTION]: The skill facilitates the creation and execution of dynamic code environments.
SKILL.md describes the modal.Sandbox class, which programmatically creates isolated containers to run user-generated or model-generated code.
references/images.md explains run_function(), which allows executing Python code during the container build process, and run_commands() for shell execution during builds.

modal