The Agent Skills Directory

[DYNAMIC_EXECUTION]: The dialogue runner implementation provided in references/runner.md uses a custom expression evaluator built with the ast module. This is a secure design choice that avoids the use of unsafe functions like eval() by whitelisting specific AST nodes (Name, Constant, BinOp, Compare, BoolOp), ensuring that dialogue logic cannot execute arbitrary system commands.
[INDIRECT_PROMPT_INJECTION]: The skill's architecture for processing branching dialogue creates a surface for indirect prompt injection via the dialogue data itself. The risk is mitigated by the following factors documented in the files: 1. Ingestion points: The DialogueRunner class in references/runner.md processes a graph dictionary derived from external JSON data. 2. Boundary markers: Data is ingested into specific structured fields (line, choices, text) rather than a free-form stream. 3. Capability inventory: The runner is limited to UI interactions (show_line, show_choices) and local variable manipulation; it does not possess subprocess execution or network access capabilities. 4. Sanitization: The evaluate function provides strict sanitization of expression strings using a whitelist-based AST walker.

dialogue-systems