The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill documents and implements mechanisms for remote script execution from vendor-controlled sources.
SKILL.md and cli/selfupdate.py reference an installation and update process that fetches scripts from vast.ai and pipes them to bash.
cli/commands/offers.py contains examples of downloading launch scripts from the official vast-ai GitHub repository.
This behavior is consistent with a platform management CLI and utilizes the vendor's own infrastructure.
[EXTERNAL_DOWNLOADS]: The CLI fetches configuration, manifests, and SSL certificates from vast.ai domains.
Fetches update manifests from https://vast.ai/cli/manifest.json.
Downloads the Vast.ai root certificate from https://console.vast.ai/static/jvastai_root.cer to establish secure connections to workers.
[INDIRECT_PROMPT_INJECTION]: The skill provides functions that ingest untrusted data which could influence agent behavior.
Ingestion points: cli/commands/misc.py (retrieving container logs via vastai logs) and cli/commands/instances.py (retrieving instance labels or status messages).
Boundary markers: Absent; data from remote instances is interpolated directly into the agent's context.
Capability inventory: Extensive; the agent can execute shell commands, manage billing, and transfer files.
Sanitization: No specific filtering is performed on data retrieved from instance logs or metadata.
[DYNAMIC_EXECUTION]: Several components use dynamic loading for extensibility.
serverless/remote/serialization.py uses importlib.import_module to deserialize objects for remote function calls.
serverless/remote/serve_deployment.py extracts deployment tarballs and executes setup scripts from a config.json inside the worker container.
[COMMAND_EXECUTION]: The skill wraps numerous CLI utilities to perform its tasks.
Invokes ssh, scp, rsync, and nvidia-smi as subprocesses to manage remote connectivity and hardware diagnostics.
Uses the uv package manager for self-updates.

vastai