The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes a system for processing dialogue data where narrative logic is stored as strings within JSON files, presenting a potential attack surface.\n
Ingestion points: JSON dialogue graph files (referenced in references/runner.md).\n
Boundary markers: The design relies on a restricted interpreter rather than string delimiters to isolate logic.\n
Capability inventory: The implementation in references/runner.md uses the ast module for expression parsing.\n
Sanitization: The evaluate function in references/runner.md implements a strict whitelist of safe AST nodes (constants, arithmetic, comparisons) to prevent arbitrary code execution from the narrative data.\n- [DYNAMIC_EXECUTION]: The skill includes a reference implementation for a custom evaluator that parses and executes logic strings at runtime.\n
Evidence: references/runner.md features an evaluate function that uses ast.parse to interpret narrative conditions and variable assignments in a controlled, sandboxed manner.

dialogue-systems