The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes local system binaries including LibreOffice, pdftoppm, and the unshare utility for document conversion and rendering.
Evidence: Subprocess calls found in scripts/render_docx.py and scripts/check_runtime.py.
Context: These operations are core to the skill's function and are performed with security-focused parameters, such as using network namespaces (unshare --net) to prevent outgoing network access during untrusted file processing.
[DYNAMIC_EXECUTION]: The skill performs dynamic module inspection to detect the presence of environment dependencies.
Evidence: Use of importlib.import_module in scripts/check_runtime.py to check for libraries like python-docx and jsonschema.
Context: This is a benign runtime environmental check. The skill explicitly forbids the automatic installation of missing dependencies, requiring manual user intervention for setup.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted document content and metadata which could contain malicious instructions.
Ingestion points: System reads DOCX structure and text via inspect_docx.py, edit_docx.py, and compare_docx.py.
Boundary markers: Instructions in SKILL.md and references/safety.md strictly define that all document data must be treated as untrusted input and never interpreted as instructions.
Capability inventory: Capability is limited to deterministic OOXML manipulation via the python-docx library and isolated rendering.
Sanitization: Implements schema validation for all structured inputs and a 'package admission' layer in lib/package_safety.py that verifies ZIP integrity and blocks active content like macros or OLE objects before the agent or parsers interact with the file.

krt-word-illuminator