The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external documentation from chunks.json and instructs the agent to interpret these findings. Malicious instructions embedded in these documents could influence agent behavior due to the lack of boundary markers or sanitization. * Ingestion points: Data read from chunks.json in query_simple.py. * Boundary markers: Absent in SKILL.md instructions. * Capability inventory: File-writing via add_document.py and shell execution via query_counter.py. * Sanitization: No validation or escaping of retrieved content before prompt interpolation.
[COMMAND_EXECUTION]: The query_counter.py script utilizes subprocess.run to invoke system commands (open or xdg-open) for displaying protocol documents. While documented, this involves active shell command execution on the user's host.
[DYNAMIC_EXECUTION]: The add_document.py utility allows the agent to parse new Markdown documents and persistently modify the skill's local chunks.json database at runtime.

dji-referee