The Agent Skills Directory

[COMMAND_EXECUTION]: The skill orchestrates parallel agents by spawning independent claude CLI processes using tmux split-window or it2 commands. To prevent shell injection vulnerabilities, the system implements strict regex validation (^[A-Za-z0-9_-]+$) for all agent and team identifiers used in command construction.
[INDIRECT_PROMPT_INJECTION]: As a coordination framework, the skill has a data ingestion surface involving task descriptions and inter-agent messages stored in ~/.claude/teams/ JSON files.
Ingestion points: Reads task and message data from local filesystem queues.
Boundary markers: Uses structured JSON schemas to delimit instructions from metadata.
Capability inventory: Spawns sub-agents, executes shell commands via tmux, and performs atomic filesystem writes.
Sanitization: Employs role-based capability matrices and risk-tier classification to restrict agents from executing high-risk tasks without oversight.
[DYNAMIC_EXECUTION]: The skill dynamically generates and executes agent configurations at runtime, assigning specific roles (implementer, researcher, tester, reviewer, architect) that constrain tool access and execution permissions based on the task's risk profile.

agent-teams