The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external tools for document processing and validation.
ooxml/scripts/pack.py uses soffice (LibreOffice) in headless mode to validate document integrity by converting them to HTML.
ooxml/scripts/validation/redlining.py uses git diff to provide precise character-level comparison between original and modified documents to ensure tracked changes are correctly implemented.
[EXTERNAL_DOWNLOADS]: The skill requires several well-known document processing utilities.
Documentation in SKILL.md instructs the environment setup for pandoc, docx (NPM), LibreOffice, Poppler, and defusedxml (PyPI).
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill processes document content which could contain sensitive information.
The logic is designed to work within a workspace directory and uses tempfile for intermediate operations.
No network-bound exfiltration patterns or hardcoded credentials were detected.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from .docx files provided by users.
Ingestion points: Files are unpacked via ooxml/scripts/unpack.py and read into the agent's context for editing and analysis.
Boundary markers: The skill does not explicitly define delimiters for ingested text when read into the LLM context, relying on the agent's internal reasoning.
Capability inventory: The skill has read/write access to the filesystem and calls subprocesses for validation (soffice, git).
Sanitization: The skill uses defusedxml (in scripts/document.py, scripts/utilities.py, and ooxml/scripts/pack.py) to parse XML content, which effectively mitigates XML-based attacks like XXE (XML External Entity).
[DYNAMIC_EXECUTION]: The skill's workflow involves the agent generating Python scripts to manipulate OOXML using the provided Document library.
This is the intended architecture of the skill and uses the internal library rather than executing untrusted external code.

docx