fable-codex-setup

name	fable-codex-setup
description	Bootstrap a Claude Code project with the "Fable 5 orchestrator + Codex executor" architecture — verify/install the OpenAI Codex plugin, set Fable 5 as the main model, create deep-reasoner (Opus), fast-worker (Sonnet), and codex-executor sub-agents, and write the orchestration workflow into CLAUDE.md. Use when the user wants to set up a new project for token-efficient multi-model orchestration, or says "fable codex setup", "setup orchestrator", "初始化 Fable/Codex 專案", or "set up this project with Codex sub-agents".
license	MIT

Fable + Codex Project Setup

Configure the current project so that Fable 5 plans, decomposes, and synthesizes, while cheaper executors burn the tokens:

Role	Model	Job
Orchestrator	Fable 5 (max reasoning)	Plan, decompose, delegate, synthesize. Keep own context lean.
deep-reasoner	Opus	Reasoning-heavy phases: architecture, complex debugging, algorithm design.
fast-worker	Sonnet	Mechanical work: boilerplate, tests, formatting, simple edits.
codex-executor	Haiku wrapper → Codex CLI	Bulk implementation & fresh-perspective problems. A peer senior engineer, not a reviewer.

Architecture credit: @diegocabezas01 — see SOURCE.md. Companion skill: multi-model-orchestrator — the runtime orchestration workflow this setup enables.

All file templates live in this skill's templates/ directory. Prompt patterns for the user live in PROMPTS.md (same directory as this file).

Step 1 — Preflight: Codex CLI + plugin

1. Check the Codex CLI: command -v codex && codex --version.
   • If missing, tell the user to run: npm i -g @openai/codex then codex login (needs ChatGPT Plus/Pro or an OpenAI API key).
2. Check whether the Codex plugin is already installed: look for codex:* entries in your available agent types or skills (e.g. codex:codex-rescue, codex:setup).
   • If absent, the plugin can only be installed via interactive slash commands — print these for the user to type in Claude Code (do NOT attempt them via Bash):

     /plugin marketplace add openai/codex-plugin-cc
     /plugin install codex@openai-codex
     /codex:setup
     

   • Continue with the remaining steps regardless; the codex-executor sub-agent works with the bare Codex CLI even without the plugin.

Step 2 — Fable 5 as the brain

1. Create or update .claude/settings.json in the project root, adding "model": "claude-fable-5". Merge with existing keys — never clobber a settings file that already has content.
2. Tell the user to run these once in their session (they cannot be automated):

   /model    → Fable 5
   /effort   → max
   

Step 3 — Create the sub-agents

Copy the three templates from templates/agents/ into the project's .claude/agents/:

• deep-reasoner.md (model: opus)
• fast-worker.md (model: sonnet)
• codex-executor.md (model: haiku; shells out to codex exec)

If a target file already exists, do not overwrite it — report the conflict and let the user decide.

Step 4 — Orchestration workflow in CLAUDE.md

Append the contents of templates/CLAUDE-orchestration.md to the project's CLAUDE.md (create the file if it doesn't exist). If CLAUDE.md already contains a ## Orchestration workflow heading, skip this step and say so.

Step 5 — Verify and report

1. List every file created/modified.
2. Restate the manual steps still owed by the user (plugin slash commands from Step 1, /model + /effort from Step 2).
3. Print the kickoff prompt from PROMPTS.md § "Tech-lead kickoff" so the user can start driving tasks immediately.

Guardrails

• Surgical changes only: never overwrite existing agents, CLAUDE.md sections, or settings keys without telling the user first.
• Everything in Steps 2–4 is plain file writing — do it directly; do not ask permission step-by-step.
• Slash commands (/plugin, /model, /effort, /codex:setup) are user-interactive: always hand them to the user, never simulate them.