Codex 校核(2026-10-09) Pilotfish 当前 README已有八种角色,默认 executor 为 Sonnet medium、mechanical-executor 为 Sonnet low;原文六角色表与 executor=Opus 应保留为当时版本快照,不能用于当前安装配置。Fable 默认需要显式选择,插件模式标为 beta,legacy 安装另有边界。96% 成功率与 46% 成本下降来自项目特定历史评测,本次没有独立重跑,也不代表所有代码任务的效果。
Pilotfish — Multi-Model Orchestration for Claude Code
GitHub: github.com/Nanako0129/pilotfish Stars: 432 | License: MIT | Type: Claude Code config package Release: v1.1.5 (2026-07)
Core Positioning
Multi-model orchestration layer for Claude Code: frontier model (Claude Fable 5/Opus) plans, decides, reviews; cheaper models (Sonnet/Haiku) execute volume work through subagents. Fresh-context verification guards quality instead of using the biggest model everywhere.
Architecture (3 layers, 3 config files)
| Layer | Files | Job |
|---|---|---|
| Machine | ~/.claude/settings.json | Who orchestrates (best) + automatic fallback chain |
| Roles | ~/.claude/agents/*.md | Six role agents, each pinned to the right model tier |
| Policy | ~/.claude/CLAUDE.md | How to delegate in terms of roles, never model names |
Six Roles
- scout (haiku low) — read-only lookups: where is X, symbol usage, config values
- Explore (haiku low) — overrides the built-in Explore agent
- mech-executor (sonnet low) — mechanical work: pattern refactors, docs, bulk edits
- executor (opus medium) — judgment-needed implementation: features, bug fixes, design-sensitive
- verifier (opus medium) — fresh-context adversarial verification, returns CONFIRMED/REFUTED, never fixes
- security-executor (opus high) — security-sensitive work, deliberately avoids Fable 5
Benchmark Data
- Anthropic official: Fable 5 orchestrator + Sonnet 5 workers = 96% of all-Fable performance at 46% cost
- Community (12-worker audit): 6.10 (Sonnet, 58% savings), $3.70 (Haiku, 74% savings)
Why This Matters
Pilotfish represents an emerging pattern: multi-model orchestration. Instead of using one strongest model for everything, different models specialize per task complexity. Anthropic’s own benchmarks validate this direction. Directly applicable pattern for production agent systems.