The Agent Skills Directory

COMMAND_EXECUTION (MEDIUM): The skill performs runtime compilation and library injection for sandbox compatibility.
File: scripts/office/soffice.py
Evidence: The script writes a C source string to a temporary file, compiles it using gcc, and then sets LD_PRELOAD to load the resulting shared object into the soffice process.
Context: This shim is designed to facilitate interaction with LibreOffice in environments where AF_UNIX sockets are blocked. Per the instruction, the severity is dropped one level as it is essential for the skill's primary function.
PROMPT_INJECTION (LOW): The skill processes untrusted PPTX files, creating an indirect prompt injection surface.
Ingestion points: Data enters the agent context includes XML and text content parsed from user-provided .pptx files via scripts/office/unpack.py and scripts/thumbnail.py.
Boundary markers: Absent. The skill does not use delimiters or instructions to treat extracted content as non-executable data.
Capability inventory: Execution of subprocesses (soffice, pdftoppm, gcc), file system manipulation, and XML reconstruction.
Sanitization: The skill follows security best practices by using defusedxml to protect against XML-based vulnerabilities (XXE).

pptx