The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates a System 1 / System 2 loop where environmental observations are processed by an AI model to determine actions. This creates a surface for indirect prompt injection if the environment data contains malicious instructions.
Ingestion points: The Loop class ingests untrusted data through the user-provided observe and images callbacks in src/decision_api/__init__.py.
Boundary markers: The prompt construction in _spawn_prompt lacks explicit delimiters or instructions to the decision model to disregard potentially adversarial content within the observation state.
Capability inventory: The skill possesses the capability to execute actions in the environment via the act callback based on the AI model's output.
Sanitization: No sanitization, escaping, or schema validation is applied to the observation data before it is passed to the decision model.
[EXTERNAL_DOWNLOADS]: The skill documentation and logic reference the Prime Inference API (api.pinference.ai). This is a legitimate vendor resource used to host the decision models required for the skill's primary functionality.

decision-api