The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines an agent orchestration pattern that ingests untrusted user input into the planning context, which creates a potential vulnerability surface.
Ingestion points: The query parameter in the Agent.run method in SKILL.md allows external data to enter the agent's memory.
Boundary markers: The AgentState.get_context function in SKILL.md uses basic role-based formatting without strong delimiters or instructions for the model to ignore potentially malicious embedded content.
Capability inventory: The orchestration framework explicitly supports executing sensitive tools such as search_web, query_database, and call_api as defined in the TOOL_ORDER mapping in SKILL.md.
Sanitization: No sanitization, escaping, or schema validation logic for tool outputs or user inputs is present in the provided architectural templates.

agent-orchestration-planner