The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructions in SKILL.md direct the agent to interpolate raw, unsanitized user input directly into a shell command string (e.g., python ... --picks "<用户原始输入>"). This creates a command injection surface where a user could provide input containing shell metacharacters like double quotes followed by command separators (e.g., "; rm -rf /; ") to execute arbitrary commands.\n
Ingestion points: User-provided numbers or text strings for card selection are captured in the third interaction round and passed as the picks argument.\n
Boundary markers: The command template uses double quotes, which can be escaped or terminated by malicious input.\n
Capability inventory: The skill executes a local Python script (draw_cards.py) and reads local reference markdown files.\n
Sanitization: While the backend script draw_cards.py uses regular expressions to safely extract numbers, the skill's instruction for command construction lacks any sanitization or escaping instructions for the agent to prevent shell breakout.\n- [EXTERNAL_DOWNLOADS]: The skill fetches tarot card images from Wikimedia Commons (commons.wikimedia.org). This is a well-known service used to provide visual aids for the tarot reading and is documented here as part of the intended functionality.

tarot-guide