The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted Word documents (.docx) through several modes. Content from these files could contain instructions designed to manipulate the agent's behavior during analysis or when performing edits.
Ingestion points: Word documents loaded via scripts/review_pass.py for reading and annotation.
Boundary markers: Document content is rendered to a simplified text/markdown format in read mode before processing.
Capability inventory: The skill possesses file write capabilities (annotate, track, revise modes) and indirect shell execution through Word COM automation.
Sanitization: No explicit sanitization of document content is mentioned in the driver; logic resides in the ane_package.officeops library.
[DYNAMIC_EXECUTION]: The skill's Word COM automation (used in revise mode) relies on dynamic script generation. It creates temporary PowerShell scripts to interface with Microsoft Word on Windows systems.
Evidence: The API documentation in references/officeops-api.md explicitly notes that COM surfaces drive PowerShell 5.1 using ASCII script files with UTF-8 payloads.
[COMMAND_EXECUTION]: The skill's test infrastructure uses the subprocess module to run the Python interpreter and verify the CLI driver's behavior.
Evidence: scripts/test_review_pass.py contains a helper function _run that calls subprocess.run([sys.executable, ...]) to execute the driver script during testing.

office-review-pass