The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates communication between multiple agents, creating a surface for indirect prompt injection where instructions could be embedded in the task or context arguments.
Ingestion points: Data enters the system through the task and context parameters of the delegate_to_choom tool in handler.ts and tools.ts.
Boundary markers: The skill employs structured headers (e.g., ## Your Task, ## Context from orchestrator) and clear RULES sections to delimit instructions from data passed to the target agent.
Capability inventory: The skill can trigger tool execution in other agents via an internal chat API, record interactions in the database, and manage files in the workspace through delegated worker agents.
Sanitization: The implementation uses verifyDelegationClaims to validate that agents actually performed the file operations they claim to have done. It also truncates long responses to prevent agents from being overwhelmed by large context windows.
[DATA_EXFILTRATION]: The list_team tool provides metadata about available agents (name, model, and specialization). This information disclosure is restricted to internal database records and is a standard requirement for agent orchestration functionality.
[COMMAND_EXECUTION]: While the skill orchestrates execution across multiple agents, it does not perform direct shell command execution. Security is maintained by passing a delegation flag to the underlying platform API, which limits the tools and capabilities available to the worker agent.

choom-delegation