The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data (image descriptions and filenames) which are interpolated into a prompt file (prompt.md) and subsequently executed by the Codex agent. This represents a potential injection surface where malicious input could attempt to influence agent behavior or command execution.
Ingestion points: User-provided text for image descriptions, filenames, and styles in prompt.md.
Capability inventory: The skill possesses file system write access, network access (via danger-full-access sandbox), and execution of shell commands and Python scripts.
Boundary markers: The skill uses some template structure but lacks explicit instructions for the agent to treat user-provided content as non-executable data or to sanitize specific fields like filenames.
Sanitization: No explicit sanitization or validation of the user-provided strings is documented before they are used in automated workflows.
[DYNAMIC_EXECUTION]: The skill generates and executes shell commands and Python code at runtime to handle image acceptance and compression.
Evidence: The skill uses python3 heredocs to drive loops and wait for files, and it constructs sips, magick, and cwebp commands using interpolated variables for paths and dimensions.
Evidence: Python scripts using the PIL (Pillow) library are dynamically generated to calculate image alpha channel bounding boxes.
[COMMAND_EXECUTION]: The skill relies on executing external binaries for its core functionality.
Evidence: Use of macOS built-in sips for image manipulation.
Evidence: Use of magick (ImageMagick) for complex image processing, including transparency extraction and resizing.
Evidence: Use of cwebp for WebP conversion.

imagegen