The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted document data (.docx) provided by users, which creates a surface for indirect prompt injection if documents contain malicious instructions.
Ingestion points: ooxml/scripts/unpack.py and scripts/document.py read and parse OOXML files into the agent's context.
Capability inventory: The skill has filesystem write access and can execute specific system utilities (soffice, git).
Sanitization: The skill implements strong security mitigations by using the defusedxml library for all XML parsing in scripts/utilities.py and scripts/document.py, which protects against XML External Entity (XXE) and billion laughs attacks.
[COMMAND_EXECUTION]: System commands are executed via subprocess.run to perform document validation and version comparisons.
ooxml/scripts/pack.py executes soffice to validate document integrity by performing a headless conversion to HTML.
ooxml/scripts/validation/redlining.py executes git diff to provide character-level or word-level comparisons during tracked changes validation.
All executions use static command lists and do not invoke a shell, preventing command injection.
[EXTERNAL_DOWNLOADS]: The skill documentation provides instructions for installing necessary system and language-specific dependencies from well-known repositories.
System utilities: pandoc, libreoffice, and poppler-utils via the standard system package manager (apt-get).
JavaScript library: docx via the NPM registry.
Python library: defusedxml via the PyPI registry.

docx