The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes product descriptions, factual details, and image URLs from external sources, creating a potential vector for malicious instructions to influence agent behavior.
Ingestion points: The catalog-client.mjs script reads input from operation.json (prepared by the agent from user/external input) and fetches content from external HTTPS image URLs.
Boundary markers: The skill strictly enforces a 'Preview and approve' workflow, requiring the agent to present a JSON plan to the user for explicit approval before any mutations or uploads occur.
Capability inventory: The script can read and write local files (staging images and plans) and perform network requests to the store API and external image hosts.
Sanitization: The catalog-client.mjs script performs basic sanitization by verifying image magic numbers (JPEG, PNG, WebP) and enforcing a 5MB file size limit.
[EXTERNAL_DOWNLOADS]: The catalog-client.mjs script downloads data from arbitrary HTTPS URLs provided in the catalog operation to retrieve product images.
[COMMAND_EXECUTION]: The skill instructs the agent to execute a local Node.js script (catalog-client.mjs) to manage the catalog. This is the primary and intended method of operation for the skill.

zehri-catalog