The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external content (textbooks, papers, web content) and ingests it into a long-term memory graph using the memocli tool.
Ingestion points: External data is ingested during 'Sub-process 1' in SKILL.md and read by the visualization server from a user-specified directory.
Boundary markers: No explicit boundary markers or instructions to ignore embedded commands are present in the ingestion flow.
Capability inventory: The skill uses shell commands via memocli and executes a local Python server for visualization.
Sanitization: While the Python server (server.py) uses yaml.safe_load() for file parsing, the visualization frontend (index.html) uses D3.js's .html() method to render entity observations directly. This lack of output sanitization creates a potential Cross-Site Scripting (XSS) surface if the agent extracts and stores malicious script or HTML tags from source materials.
[COMMAND_EXECUTION]: The skill relies on executing memocli commands via shell to manage the hierarchical knowledge graph. It also requires the user to manually execute a Python script (server.py) to provide the interactive dashboard.
[EXTERNAL_DOWNLOADS]: The visualization dashboard (index.html) fetches the D3.js library from d3js.org. This is a well-known and trusted source for data visualization software and does not escalate the security verdict.

subject-learning-assistant