The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a framework for interactive widgets that process runtime state data, which presents an attack surface for indirect prompt injection.
Ingestion points: The generated JavaScript for widgets interacts with window.opensquillaVisualize.getState() as specified in SKILL.md to restore user session state.
Boundary markers: The skill provides explicit instructions to validate all saved values before rendering and to avoid default overwrites during state restoration.
Capability inventory: The skill utilizes the show_widget and edit_widget tools to render and update visualization code directly within the conversation.
Sanitization: Guidelines in SKILL.md and references/parameter-simulation.txt require strict range validation, type checking, and schema verification for all external state data before it is used in calculations or rendering.
[SAFE]: The skill establishes a highly restricted execution environment for all generated visual components to prevent data exfiltration and unauthorized execution.
Evidence: The documentation in SKILL.md and references/guidance-base.txt explicitly forbids the use of fetch, XHR, WebSockets, eval, and new Function. It also prohibits loading external libraries, CDNs, or remote assets, requiring all data and code to be self-contained within the HTML fragment and subject to a 1 MiB size limit.

visualize