Stop burning Fable 5 tokens on find-and-replace

What it does

It drops three config files into ~/.claude/ that split Claude Code into eight role-based subagents, each locked to a specific model tier. The main session uses a frontier model—Fable 5 or Opus—to plan, decide, and review, while cheaper models handle lookups, bulk edits, and verification in the background. Everything is global, so one install covers every project, and a fallback chain keeps you working when the frontier model quota runs dry.

The interesting bit

The orchestration policy is written in terms of roles rather than model names, so the system delegates by job function instead of vendor branding. The README cites Anthropic’s own benchmark data to back the economics: a Fable 5 orchestrator paired with Sonnet workers reportedly scores 96% of an all-Fable setup at 46% of the cost.

Key highlights

• Eight strictly bounded roles, from read-only scout (Haiku) to security-executor (Opus)
• Phase-gated workflow with tool-enforced read-only plan and security reviews before approval
• Fresh-context verifier subagents attempt to refute completed work instead of relying on self-critique
• Automatic fallback chain when frontier models become unavailable
• Overrides the built-in Explore subagent to Haiku, preventing background searches from silently burning Opus-tier tokens

Caveats

• Requires Claude Code 2.1.207 or newer; the tool-enforced read-only roles rely on agent allowlists introduced in that version
• Native Windows without WSL may hit POSIX assumptions in the install runbook
• No runtime code of its own; if Claude Code’s subagent system breaks, pilotfish breaks with it

Verdict

Install it if you treat Claude Code as a daily driver and routinely exhaust weekly model quotas. Skip it if you already micromanage every subagent call or rarely trigger subscription limits.