The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation recommends installing standard developer utilities such as tmux and the it2 (iTerm2 CLI) tool for visualizing sub-agent processes. These tools are sourced from reputable repositories and are used as intended for process management.\n- [DATA_EXFILTRATION]: The skill manages agent orchestration by reading and writing to the ~/.claude/ directory. This is the native operational path for Claude Code to handle agent configurations, inboxes, and task tracking. Access is restricted to these coordination files and does not involve unauthorized access to sensitive user data.\n- [PROMPT_INJECTION]: The orchestration logic facilitates the transfer of data between agents through local inboxes and task updates. Ingestion points: Agent inboxes (~/.claude/teams/{name}/inboxes/) and shared task descriptions. Boundary markers: None are explicitly defined in the provided patterns. Capability inventory: Agents are equipped with Bash, filesystem access, and general-purpose capabilities. Sanitization: No input validation or instruction-filtering is applied to the inter-agent messages. This creates a surface for indirect prompt injection, which is a standard consideration for multi-agent coordination systems.

orchestrating-swarms