The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions and provided deploy.py script involve direct execution of shell commands, including docker, vastai, ssh, rsync, and wget to manage infrastructure and data.
[DYNAMIC_EXECUTION]: The deploy.py script uses the subprocess.run module to dynamically build and execute CLI commands based on user-provided arguments such as GPU type, image name, and custom startup commands.
[EXTERNAL_DOWNLOADS]: The skill provides a command to download datasets from an external URL (https://storage.example.com/dataset.tar.gz) directly to a remote instance using wget. While the specific URL is a placeholder, it establishes a workflow for fetching external remote content.
[REMOTE_CODE_EXECUTION]: The skill facilitates the execution of commands on remote GPU instances through SSH and the Vast.ai onstart-cmd feature, allowing for automated setup and health monitoring. This includes a heredoc pattern that executes a shell script directly on the target host.
[INDIRECT_PROMPT_INJECTION]: The deploy.py script provides a vulnerability surface by accepting a --onstart argument that is passed without sanitization to the remote instance's startup command. This creates a surface where instructions from processed data or user queries could be executed as shell commands on the remote instance.
Ingestion points: Command-line arguments (--image, --onstart) processed by deploy.py in SKILL.md.
Boundary markers: None identified in the provided script or instructions.
Capability inventory: subprocess.run (executing local vastai CLI), ssh (remote command execution), and docker (container management) as seen in SKILL.md.
Sanitization: No input validation or escaping is performed on the arguments before they are passed to the subprocess call.

vastai-deploy-integration