The Agent Skills Directory

[COMMAND_EXECUTION]: Runtime compilation of system components.
The script scripts/office/soffice.py contains logic to dynamically generate a C source file (lo_socket_shim.c) in the system's temporary directory.
It subsequently invokes the gcc compiler to build this source into a shared library (.so) at runtime.
[COMMAND_EXECUTION]: Process and library injection.
The skill uses the LD_PRELOAD environment variable to inject the dynamically compiled shared library into the soffice (LibreOffice) subprocess.
The injected library intercepts and shims low-level socket system calls (socket, listen, accept, close) to bypass environment-specific restrictions on AF_UNIX sockets.
[COMMAND_EXECUTION]: Execution of external system utilities.
Multiple scripts within the skill utilize subprocess.run to execute external tools including soffice, pdftoppm, git, and the gcc compiler.
[REMOTE_CODE_EXECUTION]: Potential for indirect prompt injection via data ingestion.
The skill ingests and processes untrusted external data from .pptx files using scripts like thumbnail.py and unpack.py.
While the skill employs the defusedxml library to mitigate some XML-related risks, it lacks explicit boundary markers or sanitization for all content extracted from these files before further processing by the agent context.

pptx