The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a routing system that processes user-provided task descriptions and keywords to select appropriate AI models. This design creates a surface for indirect prompt injection where malicious instructions embedded in the task data could influence model selection or the behavior of sub-agents.
Ingestion points: Task prompts and user-supplied keywords are ingested via the routing logic defined in SKILL.md and the configurations in routing-rules.json.
Boundary markers: The skill does not specify the use of delimiters or 'ignore embedded instructions' markers when interpolating untrusted prompts into the Task tool calls.
Capability inventory: The routing logic directs tasks to sub-agents that possess file system access, code review capabilities, and deployment tools.
Sanitization: There is no evidence of input validation or sanitization to prevent adversarial payloads within the task descriptions from affecting the system's routing decisions or subsequent executions.

model-routing