The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script scripts/office/soffice.py generates a C source file at runtime from a hardcoded string, compiles it into a shared object using gcc, and then injects it into the soffice process via the LD_PRELOAD environment variable. This technique is used to bypass system restrictions on AF_UNIX sockets in sandboxed environments but constitutes a high-risk execution pattern.\n- [COMMAND_EXECUTION]: Instructions in SKILL.md direct the AI agent to autonomously execute sudo apt install commands to install system dependencies such as pandoc, libreoffice, and poppler-utils. This promotes the use of root privileges without human intervention.\n- [COMMAND_EXECUTION]: Several skill components, including scripts/accept_changes.py, scripts/office/soffice.py, and scripts/office/validators/redlining.py, use subprocess.run to execute external binary tools like soffice, gcc, and git.\n- [EXTERNAL_DOWNLOADS]: The skill requires the installation of external software and packages via npm install docx and pip install defusedxml lxml to function.\n- [PROMPT_INJECTION]: The skill processes untrusted user-supplied .docx files by unpacking and parsing their internal XML content. While the skill employs mitigations like defusedxml and provides warnings in its documentation, the exposure of these data ingestion points to powerful system tools like soffice and git creates a surface for indirect prompt injection.

docx