Fable-agents

English · 简体中文

Portable installer for a three-layer Claude Code model-routing setup: a Fable 5 orchestrator that only plans, delegates, and synthesizes, plus three subagents (deep-reasoner on Opus, fast-worker on Sonnet, and gatekeeper on Opus as the acceptance and ship gate) and Codex as an independent peer engineer.

Why route

One Max (20x) subscription, one day of heavy multi-agent work (2026-07-05): 902M tokens across 4,911 requests at a 96.1% cache-hit rate — $731 in API-equivalent usage.

The request split is the routing contract at work: Fable answered only 492 requests (10%) — pure orchestration: plan, delegate, judge, synthesize. Sonnet carried 2,924 execution requests; Opus took 705 judgment and gate runs.

Fable bills against its own weekly pool, separate from the all-models pool. With Fable doing orchestration only, the two pools drain in step — 32% Fable vs 26% all-models mid-week. A single model's weekly usage keeps pace with the rest of the subscription combined, instead of the Fable pool burning out on execution work while the shared pool idles.

Install

Option A — clone and run

git clone https://github.com/Ancienttwo/Fable-agents.git
cd Fable-agents
bash scripts/install.sh

Agents and the routing section install once, globally — every project on the machine picks them up. --project only scopes the headless smoke test to a specific project directory (default: cwd); it does not change where anything is installed. The script is idempotent — safe to re-run.

Flags:

• --project <dir> — project to run the smoke test from (default: cwd)
• --skip-plugin — skip installing/checking the Codex plugin
• --skip-smoke — skip the headless agent smoke test (use in sandboxes without API credentials)

Option B — send to your Claude

Paste this to any Claude that can run commands on your machine (Claude Code CLI, desktop app, or IDE extension):

Install the model-routing setup from https://github.com/Ancienttwo/Fable-agents — clone the repo, run bash scripts/install.sh, then report the per-layer results ([new]/[ok]/[conflict]/[warn]). Never overwrite anything on [conflict].

What it installs

1. agents/fast-worker.md, agents/deep-reasoner.md, and agents/gatekeeper.md into your global Claude home ($CLAUDE_CONFIG_DIR/agents, default ~/.claude/agents) — available in every project
2. The ## Model Routing Hierarchy section into your global CLAUDE.md ($CLAUDE_CONFIG_DIR/CLAUDE.md, default ~/.claude/CLAUDE.md)
3. The codex@openai-codex plugin (marketplace openai/codex-plugin-cc), with a readiness check
4. A headless smoke test confirming all three agents respond

The installer never overwrites a file that differs from the bundled version — it reports [conflict] and exits 3 instead, so you can diff and decide.

The routing model

• Fable 5 (main loop) = orchestrator. Plans, delegates, synthesizes. Never does execution work inline.
• Subagents always carry an explicit model/type. No spawn — Agent tool or Workflow — may silently inherit Fable.
• deep-reasoner (Opus, effort: max) — architecture research, hard reasoning, high-risk judgment calls. Recommends; the orchestrator confirms the final framework. Returns RECOMMENDATION: <one-line> — confidence: HIGH/MEDIUM/LOW.
• fast-worker (Sonnet, effort: max) — implementation, tests, refactors, docs, mechanical execution. Returns RESULT: DONE/PARTIAL/BLOCKED with verification output.
• gatekeeper (Opus, effort: max) — last gate after execution subagents deliver: reviews the diff against the goal, runs the project's real verification, and returns a PASS/FAIL/BLOCKED ship recommendation. Advises, never decides — ship actions run only on the orchestrator's explicit execution order; never fixes code.
• Codex — independent peer engineer, not a mandatory reviewer. Invoked via the codex plugin's /codex:* commands or codex exec.
• High-stakes decisions run dual-track: deep-reasoner and Codex each produce a solution independently; the orchestrator compares and synthesizes rather than picking one.

Structure

SKILL.md                          Claude Code skill manifest (usable as a skill if this repo
                                   is placed/symlinked under ~/.claude/skills/)
assets/deep-reasoner.md           Agent definition installed into ~/.claude/agents (global)
assets/fast-worker.md             Agent definition installed into ~/.claude/agents (global)
assets/gatekeeper.md              Agent definition installed into ~/.claude/agents (global)
assets/model-routing-hierarchy.md Source of the CLAUDE.md section the installer appends
scripts/install.sh                Idempotent installer (see flags above)

Prerequisites

• Claude Code CLI (claude) on PATH for the plugin step and smoke tests
• node for the Codex readiness check
• codex CLI + codex login for a fully verified Codex layer (optional — reported as [warn] if missing, not a hard failure)

An already-running interactive Claude Code session loads agent types at startup, so it won't see newly installed agents until restart or /reload-plugins.