The Agent Skills Directory

[COMMAND_EXECUTION]: Runtime compilation and process injection.\n
The script scripts/office/soffice.py dynamically writes C source code to a temporary file, compiles it using gcc, and uses the LD_PRELOAD environment variable to inject the resulting shared library into the LibreOffice (soffice) process.\n
This mechanism is used to bypass environment restrictions on Unix domain sockets in sandboxed environments but employs patterns typical of advanced process manipulation.\n- [COMMAND_EXECUTION]: Execution of system tools via subprocess.\n
Multiple scripts (scripts/thumbnail.py, scripts/office/soffice.py, scripts/office/validators/redlining.py) invoke external binaries including soffice, pdftoppm, and git diff using subprocess.run().\n- [PROMPT_INJECTION]: Indirect prompt injection surface.\n
Ingestion points: Untrusted content is extracted from .pptx files in scripts/thumbnail.py and scripts/office/unpack.py.\n
Boundary markers: The skill instructions do not explicitly mandate the use of delimiters or isolation warnings for extracted content.\n
Capability inventory: The agent has access to subprocess.run (via soffice.py and thumbnail.py) and broad file system modification permissions.\n
Sanitization: Content is parsed using defusedxml, but text content is not sanitized or filtered against embedded AI instructions, allowing for potential hijacking of the agent's logic.

pptx