The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is architected to process and respond to data received from external, potentially untrusted agents via the A2A 1.0 protocol.
Ingestion points: Processes external data from a2a-message.json, a2a-task.json, and a2a-artifact.json payloads, as well as webhook push notification callbacks.
Boundary markers: Employs schema validation at every communication boundary and includes a specific rule to cross-check received messages against the original task description (ASI01 Goal Hijack prevention).
Capability inventory: The skill has access to sensitive tools including execute_command, write_file, and create_file through its defined toolset.
Sanitization: Utilizes grammar-constrained decoding for artifact validation and schema-error reflection to handle malformed inputs, though it relies on standard LLM safety for natural language parsing.

agent-a2a-protocol