The Agent Skills Directory

[COMMAND_EXECUTION]: Runtime compilation and process injection.
File: scripts/office/soffice.py
Evidence: The script contains hardcoded C source code in the _SHIM_SOURCE variable. At runtime, it writes this source to a temporary file, compiles it using gcc -shared -fPIC, and then injects the resulting shared object into the soffice process via the LD_PRELOAD environment variable. This shims standard library functions (socket, listen, accept, etc.) to handle restricted socket environments. This is a highly privileged execution pattern.
[COMMAND_EXECUTION]: Arbitrary execution of external binaries.
Files: scripts/accept_changes.py, scripts/office/soffice.py, scripts/office/validators/redlining.py
Evidence: The skill makes extensive use of subprocess.run to invoke external system tools including soffice, gcc, git, and pandoc. Arguments to these commands include file paths and configuration flags derived from task input.
[PROMPT_INJECTION]: Indirect prompt injection surface.
Files: scripts/office/unpack.py, scripts/office/validators/redlining.py
Ingestion points: Document content enters the agent context through the unpacking and redlining validation processes.
Boundary markers: Absent. No delimiters or instructions are provided to the agent to disregard embedded instructions within document XML.
Capability inventory: The skill has read/write file system access and can execute multiple external processes.
Sanitization: Uses defusedxml to mitigate XXE attacks during parsing, but does not sanitize extracted natural language content before processing.
[EXTERNAL_DOWNLOADS]: Recommendation for global package installation from public registries.
Evidence: SKILL.md instructs the user to run npm install -g docx to enable document generation capabilities.

docx