Claude Code Orchestration Kit

Claude Code orchestration kit

What this is

By default, every subagent runs on the same model as your main session, inherits its
effort level, and can spawn subagents of its own. Fable spawning Fable to run a grep is
how a week of quota disappears in a day.

This kit fixes that. The main session is the orchestrator: it plans, writes briefs,
reviews, and verifies. It never edits code. Five subagents do the work. Each one has a
pinned model, a pinned effort level, and a fixed tool list, and none of them can spawn
another subagent.

A brief is the written task each subagent receives: what to do, what not to do, and
the exact condition that means it is done.

A bucket is the folder that holds everything about one task: its briefs, reports,
findings, and decisions. Its name is a slug: a short name with no spaces, only
lowercase letters, numbers, and hyphens, so it is safe as a folder name. Example: bug-42.

Install

The Response contract section at the top of core/CLAUDE.md is one person's reply
preferences (answer length, troubleshooting order). Edit it to match yours.

The five agents

The normal loop is: you plan → builder builds → refuter checks → you decide.

None of the five can spawn a subagent. The Agent tool is not in any of their
tool lists, so it does not exist for them. This is enforced, not just requested. For any
other subagent you start, set this once so it cannot spawn either. In a terminal:

In the desktop app there is no terminal, so put it in ~/.claude/settings.json instead:

The extras settings.json sets the same value for a repo. After installing, test it once:
spawn a subagent and ask it whether it has the Agent tool. It should say no.

Your first task

You do not name tasks or run a command for each step. You say what you want, and Claude
keeps the…