sub-agents-execution

What problem does it solve? Delegating isolated subtasks to external AI CLI agents requires consistent invocation, permission control, and result parsing across many backends, which is error-prone to do by hand. ## Core Features & Use Cases - Multi-Backend Delegation: Launch bounded tasks on Codex, Claude, Gemini, Grok, Cursor, GLM, Kimi, or OpenCode CLIs through one unified script. - Role Composition: Apply zero, one, or multiple compatible role definitions from a roles catalog, with the most restrictive permission automatically enforced. - Structured Results: Receive normalized JSON responses with status, result, exit code, and error fields, including timeout and partial-output handling. - Use Case: While working in a large repository, delegate an isolated audit task such as "inspect the active plan and report unresolved path references" to a read-only sub-agent and integrate its validated findings. ## Quick Start Ask the agent to delegate a bounded task to a sub-agent by running scripts/run_subagent.py with a prompt and absolute working directory, optionally naming one or more roles.