The Agent Skills Directory

[DYNAMIC_EXECUTION]: The module scripts/office/soffice.py contains an embedded C source code string. At runtime, the skill writes this code to a temporary directory, compiles it using gcc, and then uses the LD_PRELOAD environment variable to inject the generated shared library (.so) into the LibreOffice (soffice) process. This shims low-level socket functions (socket, listen, accept, close, read) to allow communication in environments where standard UNIX domain sockets are restricted.
[COMMAND_EXECUTION]: The skill frequently invokes external system binaries and compilers.
It calls gcc to compile its own injection shim.
It executes soffice (LibreOffice) to perform document conversions and execute macros.
It utilizes git for diffing text content during validation.
It uses standard utilities like zip, unzip, and pandoc for document processing.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process Word documents from untrusted sources, which exposes a potential attack surface.
Ingestion points: Document XML content is parsed and processed in multiple scripts, such as scripts/merge_runs.py and scripts/office/validators/docx.py.
Boundary markers: While the skill performs XSD schema validation, it does not use explicit boundary delimiters when extracting text for the agent to read.
Capability inventory: The skill has broad capabilities including file system access (read/write/delete), archive manipulation, and arbitrary shell command execution via subprocesses.
Sanitization: The skill effectively uses defusedxml to prevent XML External Entity (XXE) attacks and performs rigorous schema checks to ensure document integrity.

docx