The Agent Skills Directory

[COMMAND_EXECUTION]: The engine.py script provides a command-line interface for the agent to manage game cycles, such as initialization and phase transitions, which is a standard and safe use of CLI tools.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill allows the configuration of an audit log path via the --audit flag and creates local directories for storage. This operation is restricted to the local filesystem and used for legitimate game history logging without exposing sensitive system data or credentials.
[INDIRECT_PROMPT_INJECTION]: The skill processes player-provided data through JSON payloads, creating a surface for potential indirect injection if malicious instructions are embedded in player text.
Ingestion points: Player replies are ingested as JSON payloads via the --payload argument in engine.py.
Boundary markers: Structured JSON parsing is used, and the skill documentation emphasizes protocol validation.
Capability inventory: Local filesystem directory creation and file writing for audit logs.
Sanitization: Input is parsed using json.loads and validated by the referenced werewolf_core package.

werewolf-judge