What problem does it solve? Delegating multi-step coding work to CLI agents like Claude Code or Codex is unreliable when you supervise by watching a terminal screen: output vanishes on restart, status is hard to parse, and stalled agents go unnoticed. This Skill provides a file-driven collaboration pattern where a task file and a work file act as the durable interface between you and the delegated coding agent. ## Core Features & Use Cases - Spawn scripts for both CLIs: spawn_claude.py and spawn_codex.py pre-trust the workspace, create the task/work files, launch a persistent shell session, and deliver the collaboration contract automatically. - Status-based supervision: A reference watcher polls the work file's STATUS: line (WORKING, DONE, NEED_INPUT, HANDOFF) with stall detection and heartbeat wake-ups, so you only act when there is something to do. - Context management and handoff: Read context usage via /status or /context, compact in place with /compact, or trigger a CHECKPOINT handoff to a fresh agent instance with zero context loss. - Use Case: You need a bug fix spanning 15 files with test iterations. Spawn a Claude Code session in a dedicated worktree, append the task to the task file, launch the watcher, and get woken only when the agent reports DONE or NEED_INPUT. ## Quick Start Run the spawn script for your chosen CLI with a workspace directory, then append your task to the printed tasks_file path and let the watcher notify you when the agent finishes.
