The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates multi-agent delegation and inter-agent communication, creating a surface for indirect prompt injection if sub-agents process untrusted external data. Ingestion points: Untrusted data enters the agent context through API invocation functions such as invoke_parallel in scripts/claude_client.py and AgentPool.run in scripts/agent_pool.py. Boundary markers: Inter-agent messages in AgentPool are delimited with explicit author attribution tags. Capability inventory: The skill performs network operations via the Anthropic and Gemini APIs and has file-read access to specific project configuration paths for credential management. Sanitization: The skill relies on model-level safety and a 'Review is not delegable' protocol rather than automated filtering.

orchestrating-agents