The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes system commands using the Python subprocess module across several scripts. Most notably, scripts/office/soffice.py writes a C source file to disk and compiles it using gcc. The resulting shared library is then utilized with the LD_PRELOAD environment variable to shim socket-related system calls, facilitating the operation of LibreOffice in restricted or sandboxed environments. Additionally, scripts/office/validators/redlining.py executes git diff for version comparison, and scripts/thumbnail.py invokes soffice and pdftoppm for document conversion and visual processing.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface due to its core functionality of ingesting external slide data.
Ingestion points: Untrusted content is read from user-provided .pptx files using the markitdown module and custom parsing in scripts/thumbnail.py.
Boundary markers: The extracted presentation text is provided to the AI agent without delimiters or instructions to ignore embedded commands.
Capability inventory: The agent possesses capabilities to execute shell-based tools such as soffice, git, pdftoppm, and gcc.
Sanitization: There is no evidence of content sanitization for extracted text before it enters the agent's context.
[EXTERNAL_DOWNLOADS]: The documentation provides instructions for installing external dependencies from official repositories. Python packages include markitdown[pptx] and Pillow. Node.js dependencies include pptxgenjs, react-icons, react, react-dom, and sharp.

pptx