The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The analyze_image tool allows the agent to fetch images from remote URLs provided in the image_url parameter.
Evidence: The handler passes the image_url to the VisionService which performs the network fetch for analysis.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data (images and text prompts) via a vision LLM, which is a potential surface for indirect prompt injection attacks.
Ingestion points: The skill ingests data from remote URLs (image_url), local files (image_path), and base64 strings (image_base64) provided at runtime in handler.ts.
Boundary markers: No explicit delimiters or boundary markers are applied to the image content or prompt in the handler; isolation depends on the underlying VisionService and model provider.
Capability inventory: The handler can read configuration files (fs.readFileSync), query a database for generated images (prisma), and make network requests to a vision endpoint.
Sanitization: The skill includes input validation for image_path in handler.ts, rejecting paths that exceed 400 characters or contain newline characters to mitigate path manipulation or injection attempts.

image-analysis