The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Node.js script vision.js via a shell command in SKILL.md. While the script performs file system operations (scanning directories like ~/.kimi-code/cache and ~/.claude/cache), this behavior is explicitly documented as a feature to find the latest cached images for analysis.
[DATA_EXPOSURE]: The script reads files from common AI client cache directories and the system temporary folder. While this constitutes broad file system access, it is restricted to identifying image files (.png, .jpg, etc.) to support the skill's primary vision analysis function.
[CREDENTIALS_SAFE]: The skill uses environment variables (CUSTOM_API_KEY) for authentication, which is a standard and safe practice for secret management.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data in the form of remote image URLs and user-supplied prompts which are then sent to an external API.
Ingestion points: User-provided URL or local file path, and a text prompt.
Boundary markers: None identified in the script logic when interpolating data into the API request.
Capability inventory: File read (via fs), network requests (via fetch).
Sanitization: Basic check for image mime types, but no sanitization of the text prompt or image content against downstream model injection.

luma-vision