The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes scripts/drawio_cli.py to interact with the locally installed Draw.io desktop application via subprocess.run. The command-line arguments are programmatically controlled to ensure only valid conversion and export flags are passed to the binary.
[SAFE]: The file scripts/inspect_vsdx.py implements defensive ZIP parsing to prevent security risks associated with Office Open XML files. It includes checks to prevent path traversal, limits the total number of entries, and enforces a maximum uncompressed size of 100MB to mitigate decompression-bomb (zip-bomb) vulnerabilities.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a surface for indirect prompt injection by extracting text content from external .vsdx and .drawio diagrams and presenting it to the agent.
Ingestion points: Text extraction occurs in inspect_vsdx.py and inspect_drawio.py when parsing shape labels and diagram metadata.
Boundary markers: The extracted text is cleaned of HTML tags but does not use explicit delimiters or isolation markers when passed to the LLM context.
Capability inventory: The skill provides capabilities to write new files to the filesystem and execute the local drawio CLI.
Sanitization: Ingested content undergoes basic regex cleaning and XML validation, though linguistic instructions embedded in diagrams are not explicitly filtered.
[SAFE]: No network communication patterns were detected. The skill processes all diagram data locally, and the documentation explicitly states that enterprise architecture and business data are treated as private, avoiding uploads to third-party services.

soia-dev-drawio-visio-diagrams