The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: In scripts/office/soffice.py, the skill implements a mechanism to bypass socket restrictions by writing C source code to a temporary file, compiling it into a shared object using gcc, and then using the LD_PRELOAD environment variable to inject this library into the soffice process.
[COMMAND_EXECUTION]: The skill extensively uses subprocess.run to execute various system commands, including gcc for compilation, soffice for document conversion, pandoc for text extraction, and pdftoppm for image generation. It also writes and executes LibreOffice macros at runtime in scripts/accept_changes.py.
[EXTERNAL_DOWNLOADS]: The skill depends on external software packages including pandoc, docx (Node.js package via npm), LibreOffice, and Poppler. These tools are retrieved from trusted repositories and organizations.
[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection.
Ingestion points: Content enters the agent context from external .docx files processed by the skill.
Boundary markers: No delimiters or specific ignore-embedded-instruction warnings are used for document content interpolation.
Capability inventory: Significant capabilities including system command execution, file system writes, and code injection are available across the skill's scripts.
Sanitization: XML parsing is secured via the defusedxml library to prevent XXE attacks, but no logical content-level sanitization is performed on the parsed document text.

docx