The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local bash script (generate_image.sh) to interface with the Frevana API. This is the primary execution mechanism and is limited to the skill's own distributed scripts.
[EXTERNAL_DOWNLOADS]: The skill supports downloading reference images from remote URLs via the --image-url flag using curl. The implementation includes several security controls: a max-filesize limit of 50MB, connection and execution timeouts, and a subsequent validation step that uses Python to verify the file's magic bytes to ensure it is a valid PNG, JPEG, or WebP image before further processing.
[DYNAMIC_EXECUTION]: The bash script employs inline Python 3 snippets (using heredocs) to perform tasks that are difficult to handle natively in shell, such as retrieving file sizes, verifying image magic bytes, and performing robust JSON/header parsing. These snippets are static, embedded in the script, and do not evaluate or execute untrusted external data as code.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection as it processes untrusted data from user-provided prompts and remote image content.
Ingestion points: CLI arguments (--prompt, --image-url, --image-dir) and remote HTTP(S) responses.
Boundary markers: None identified in the prompt interpolation to the backend.
Capability inventory: File system read access (reference images), network access (curl to Frevana API and remote URLs).
Sanitization: The skill performs file format validation but relies on the backend API for prompt sanitization. The risk is considered low as the output is structured JSON intended for image rendering.

gemini-3-pro-image