メニュー
Use when the user wants to view or change the per-phase AI model routing strategy for autopilot — for example choosing a cost-optimized, balanced, or quality-max preset, inspecting the current routing table, or customising an individual phase's model binding.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | autopilot-models |
| description | Use when the user wants to view or change the per-phase AI model routing strategy for autopilot — for example choosing a cost-optimized, balanced, or quality-max preset, inspecting the current routing table, or customising an individual phase's model binding. |
| argument-hint | [cost | balanced | quality | custom | show] |
配置各阶段的 AI 模型路由策略，支持预设方案和逐 Phase 自定义。
根据 $ARGUMENTS 选择模式：
show → 展示当前配置cost → 应用省钱优先策略balanced → 应用平衡策略quality → 应用质量优先策略custom → 逐 Phase 自定义show 模式（默认）读取 .claude/autopilot.config.yaml 的 model_routing 节，展示当前配置：
当前模型路由策略:
Phase 1 (需求分析): {model} [{tier}] effort={effort}
Phase 2 (OpenSpec): {model} [{tier}] effort={effort}
Phase 3 (FF 生成): {model} [{tier}] effort={effort}
Phase 4 (测试设计): {model} [{tier}] effort={effort}
Phase 5 (实施): {model} [{tier}] effort={effort} {escalation}
Phase 6 (报告): {model} [{tier}] effort={effort}
Phase 7 (归档): {model} [{tier}] effort={effort}
默认子 Agent 模型: {default_subagent_model}
Fallback 模型: {fallback_model}
模型优先级（高→低）:
1. AUTOPILOT_PHASE{N}_MODEL 环境变量
2. config per-phase 配置 ← 当前配置
3. .claude/agents/*.md 中的 model frontmatter
4. 继承主会话模型
切换策略: /autopilot-models [cost|balanced|quality|custom]
cost / balanced / quality 模式直接应用对应策略预设，无需交互。
如果 $ARGUMENTS 为空，通过 AskUserQuestion 选择：
"选择 AI 模型路由策略（影响成本和质量）："
选项:
- "Cost-Optimized (省钱优先)" →
适合: 原型项目、个人项目
估算: ~60% cost saving vs 纯 Opus
- "Balanced (推荐)" →
适合: 日常开发
估算: ~50% cost saving vs 纯 Opus
- "Quality-Max (质量优先)" →
适合: 生产级代码、关键功能
估算: ~20% cost saving vs 纯 Opus
- "Custom (自定义)" →
逐 Phase 选择模型和 effort
# ── Cost-Optimized (省钱优先) ──
# Opus 仅用于需求分析（长上下文理解不可替代），其余全部 Sonnet/Haiku
cost_optimized:
default_subagent_model: sonnet
fallback_model: sonnet
phases:
phase_1:
tier: deep # 需求分析需 Opus 长上下文（MRCR 76%）
model: opus
effort: high
phase_2:
tier: fast
model: haiku
effort: low
phase_3:
tier: fast
model: haiku
effort: low
phase_4:
tier: standard # Sonnet 编码≈Opus（SWE-bench 差 1.3pp）
model: sonnet
effort: medium
phase_5:
tier: standard
model: sonnet
effort: medium
phase_6:
tier: fast
model: haiku
effort: low
phase_7:
tier: fast
model: haiku
effort: low
# ── Balanced (推荐) ──
# Opus 用于需求分析和实施，Phase 5 失败自动升级
balanced:
default_subagent_model: sonnet
fallback_model: sonnet
phases:
phase_1:
tier: deep
model: opus
effort: high
phase_2:
tier: fast
model: haiku
effort: low
phase_3:
tier: fast
model: haiku
effort: low
phase_4:
tier: standard
model: sonnet
effort: medium
escalate_on_failure_to: deep
phase_5:
tier: deep
model: opus
effort: high
phase_6:
tier: fast
model: haiku
effort: low
phase_7:
tier: fast
model: haiku
effort: low
# ── Quality-Max (质量优先) ──
# Opus 用于所有需要推理的阶段
quality_max:
default_subagent_model: sonnet
fallback_model: sonnet
phases:
phase_1:
tier: deep
model: opus
effort: high
phase_2:
tier: standard
model: sonnet
effort: medium
phase_3:
tier: standard
model: sonnet
effort: medium
phase_4:
tier: deep
model: opus
effort: high
phase_5:
tier: deep
model: opus
effort: high
phase_6:
tier: fast
model: haiku
effort: low
phase_7:
tier: fast
model: haiku
effort: low
Step 1: 读取当前 .claude/autopilot.config.yaml
Step 2: 用选定策略的 model_routing 值替换 config 中的 model_routing 节
保留 enabled: true
Step 3: 写入更新后的 config
Step 4: 运行 validate-config.sh 验证
Step 5: 输出结果
✓ 模型路由策略已更新: {strategy_name}
Phase 1 (需求分析): opus [deep] effort=high
Phase 2 (OpenSpec): haiku [fast] effort=low
Phase 3 (FF 生成): haiku [fast] effort=low
Phase 4 (测试设计): sonnet [std] effort=medium ↑opus
Phase 5 (实施): opus [deep] effort=high
Phase 6 (报告): haiku [fast] effort=low
Phase 7 (归档): haiku [fast] effort=low
单次实验覆盖: AUTOPILOT_PHASE{N}_MODEL=<model> /autopilot <需求>
custom 模式逐 Phase 通过 AskUserQuestion 选择模型和 effort：
对 Phase 1-7 逐个询问：
"Phase {N} ({phase_name}) 选择模型："
选项:
- "opus (深度推理，最贵)" → tier=deep, effort=high
- "sonnet (编码平衡，推荐)" → tier=standard, effort=medium
- "haiku (快速机械，最便宜)" → tier=fast, effort=low
完成后写入 config 并输出结果表格。
Phase 6B（代码审查）不含 autopilot-phase 标记，dispatch 时传入 critical=true。
resolve-model-routing.sh 的 critical 升级逻辑自动将 fast/standard 升级到 deep/opus。
因此无论选择哪种策略，Phase 6B 代码审查始终使用 Opus 级别模型。 这是 P0 质量保障——代码审查是发现 memory leak/async bug 等隐藏缺陷的最后防线。
| 场景 | 推荐模型 | 数据依据 |
|---|---|---|
| 需求分析 | Opus | MRCR v2 长上下文: Opus 76% vs Sonnet 18.5% |
| 测试设计 | Sonnet | SWE-bench: Sonnet 79.6% vs Opus 80.9%（仅差 1.3pp） |
| 代码实现 | Opus/Sonnet | Opus 能发现 Sonnet 遗漏的 memory leak/async bug |
| 模板操作 | Haiku | 机械性操作无需深度推理，成本低 5x |
| 代码审查 | Opus | 社区报告: Opus 级审查发现隐藏缺陷率显著高于 Sonnet |
Use when the user invokes /slim-task or requests structured multi-phase task execution (structured task, SOP execution, parallel sub-agent dispatch) with blind-audit quality review.
Use when autopilot orchestrator enters Phase 1 to understand requirements, drive multi-round clarification, and produce the structured requirement packet checkpoint.
Use when the autopilot orchestrator enters Phase 5 after Phase 4 test case design has been approved and is about to choose an implementation path (parallel, serial, or TDD) and dispatch implementation work.
Use when the autopilot orchestrator reaches Phase 7 after Phase 6 tri-path reports are available and must finalize the delivery by collecting summaries, extracting knowledge, previewing Allure, and evaluating archive readiness.
Translates Figma designs into production-ready application code with 1:1 visual fidelity. Use when implementing UI code from Figma files, when user mentions "implement design", "generate code", "implement component", "figma to code", "figma 转代码", "figma 还原", "实现设计稿", provides Figma URLs, or asks to build components matching Figma specs. For Figma canvas writes via `use_figma`, use `figma-use`. For full-screen design generation in Figma, use `figma-generate-design`.
Use when the autopilot orchestrator main thread is about to dispatch a phase-specific sub-agent via the Task tool and needs the JSON envelope contract, explicit path injection, and parameterized prompt templates. Not for direct user invocation; skip when the current context is not the autopilot orchestrator.