Imagine you have a really smart helper (Claude or Codex) that can read code and write code. But it forgets everything after each conversation. So we built a system where:
- A notebook (
_overview.md) keeps track of what's been done and what's next - A simple loop (Python script) wakes up the AI CLI, says "read the notebook and do the next thing", then waits
- The AI reads the notebook, does one piece of work, writes what happened back in the notebook, and goes to sleep
- The loop wakes it up again, and repeats until the job is done
- You (through a parent session) watch the progress and answer questions when needed
That's it. The Python loop is intentionally dumb; the child agent makes all the decisions.