The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes external tools using the subprocess module to perform document validation and comparison. \n
In ooxml/scripts/pack.py, it runs soffice (LibreOffice) in headless mode to validate the document structure by attempting a conversion. \n
In ooxml/scripts/validation/redlining.py, it invokes git diff with the --no-index flag to generate detailed character-level diffs between document versions. \n
These operations use fixed command templates and target well-known, legitimate software utilities. \n- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data in the form of .docx files, which are unpacked and parsed. \n
Ingestion points: Untrusted content is ingested in ooxml/scripts/unpack.py and scripts/document.py, where the ZIP archive is extracted and the internal XML files are parsed. \n
Boundary markers: The skill operates on the structured XML DOM level rather than treating the document content as raw instruction text, which provides inherent structural boundaries. \n
Capability inventory: The skill has the ability to write to the local filesystem and execute the commands mentioned above. \n
Sanitization: Security risks associated with processing untrusted XML are mitigated by the consistent use of the defusedxml library. Additionally, the skill performs comprehensive validation of the XML against official OOXML XSD schemas in ooxml/scripts/validation/ to ensure document integrity and detect corruption.

docx