The Agent Skills Directory

[SAFE]: The skill implements legitimate functionality for image brightness, contrast, and saturation adjustments using verified libraries like Pillow and NumPy.
[SAFE]: External communication is conducted through the official OpenRouter endpoint (a well-known service) using standard SDKs, and credentials are correctly managed through environment variables.
[SAFE]: The implementation includes robust input validation, including range checks for color multipliers (0.5-2.0) and non-destructive file operations that prevent overwriting original user data.
[INDIRECT_PROMPT_INJECTION]: The skill processes user prompts and image files. The primary implementation in photo_color.py effectively mitigates injection risks by using local keyword matching for style selection rather than passing raw user input to the LLM. The alternative implementation described in references/implementation.md is noted as a potential injection surface but is not currently active in the main execution path.
Ingestion points: prompt argument in the main and enhance_photo functions.
Boundary markers: Not required as user input is isolated from the LLM prompt in the script.
Capability inventory: Local file system read/write, network API access via HTTPS.
Sanitization: The prompt is normalized to lowercase for local keyword matching; all numerical outputs are validated against safe ranges.

photo-color