Fable Agents

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

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  — 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…