The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the subprocess module to run media processing scripts from a local ffmpeg-skill installation.
This execution is restricted to a hardcoded list of allowed vendor scripts (probe.py, color.py, look.py, and check.py).
Command-line arguments are constructed from validated types (numbers, enums, and absolute paths) rather than raw request strings.
The skill explicitly rejects input fields named command, exec, shell, argv, or script at any nesting depth to prevent argument injection.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted JSON request documents defining video projects.
Ingestion points: JSON data is read from stdin in src/color_grading/cli.py and parsed by model.py.
Boundary markers: The skill enforces a rigorous JSON schema with structural and semantic validation.
Capability inventory: The skill has the capability to read/write local files and execute specific processing tools.
Sanitization: All creative and technical parameters (like exposure, contrast, and LUT paths) are checked against valid ranges, enums, and types before being used.
[DATA_EXPOSURE]: The skill reads media files and Look-Up Tables from the local filesystem.
It employs a PathPolicy class that resolves symlinks and checks all paths against configurable allowed_input_roots and allowed_lut_roots.
File system writes are strictly confined to a designated workspace directory.
Path traversal attempts (using ..) and unsafe filenames are systematically blocked.

color-grading