The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/office/soffice.py implements a sophisticated workaround for socket restrictions by writing C source code to a temporary file, compiling it into a shared library using gcc, and then injecting it into the LibreOffice process using the LD_PRELOAD environment variable.
[COMMAND_EXECUTION]: Multiple scripts execute system commands via the subprocess module to perform document conversion and validation:
scripts/office/soffice.py executes gcc to compile shims and soffice for document conversion.
scripts/accept_changes.py executes soffice to run specialized macros.
scripts/office/validators/redlining.py executes git diff to analyze document changes.
[EXTERNAL_DOWNLOADS]: The skill instructions in SKILL.md suggest using npm install docx if the preinstalled dependency is unavailable.
[PROMPT_INJECTION]: The skill is designed to process external Word documents, which represents an indirect prompt injection surface. The analysis found the following evidence chain:
Ingestion points: Untrusted document content is ingested via unzip and XML parsing in scripts/merge_runs.py and scripts/office/validators/docx.py.
Boundary markers: Present. SKILL.md explicitly labels docx files from external parties as "untrusted" and mandates the removal of symlink entries.
Capability inventory: The skill has significant capabilities including arbitrary command execution (subprocess.run) and runtime compilation (gcc).
Sanitization: Strong measures are in place, including the use of defusedxml to prevent XXE attacks and a safe_extract helper that validates zip entries against path traversal and symlink attacks.

docx