The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The documentation and installation scripts recommend a one-liner command to download and execute an installation script directly from the author's GitHub repository.
Evidence: bash -c "$(curl -fsSL https://raw.githubusercontent.com/penfick/skills/main/vision-support/install.sh)" in README.md and README.zh.md.
The install.sh script clones the repository and executes node init via a subshell.
[COMMAND_EXECUTION]: The skill utilizes Node.js child process modules to execute system commands and manage internal script execution.
Evidence: execSync is used in scripts/vision.mjs to query the Windows Registry for proxy settings (reg query "HKCU\\Software\\Microsoft\\Windows\\CurrentVersion\\Internet Settings" /v ProxyServer).
Evidence: execFileSync is used in bin/cli.mjs to execute the core vision.mjs script with inherited permissions.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a data bridge by ingesting image files (local or remote), sending them to vision models, and returning the text description to the agent. This represents an indirect prompt injection surface where text within images could influence the agent's behavior.
Ingestion points: scripts/vision.mjs reads local image files and fetches remote URLs to process them via external providers (OpenAI, Google, Anthropic, etc.).
Boundary markers: SKILL.md and README.md contain an "Iron Rule" stating that vision models must only be used for image description and never for logic reasoning.
Capability inventory: The skill has network access (fetch), file system access (read/write config and images), and internal command execution capabilities.
Sanitization: The skill returns raw text descriptions from the vision model APIs without specific filtering, relying on the agent's own safety guardrails.

vision-support