orchestrate
把昂贵推理留给高性能模型，保持主上下文精简并节省成本。
会话主控器，规划、分解并将子任务委派给不同模型/子代理后合成结果。
把复杂任务拆成小块、做计划并把各子任务派给不同模型或子代理去做，最后把结果拼回来的会话总控。适合多模型协作、要节省高价模型上下文开销，或需要多人/多工具并行做设计与实现时用。重点不是自己干所有事，而是做策划、分配与校对，兼顾成本与质量。
▸ 展开 SKILL.md 英文原文
Use when work should be led as a multi-model orchestration — delegating to model-pinned subagents and a cross-vendor peer instead of doing everything in the main context. Trigger whenever the user asks to orchestrate, delegate, fan out, act as tech lead, use worker/sub agents on a task, split design work from grunt work, get a second opinion from Codex or another vendor, run two models on the same question and reconcile them, or conserve expensive-model/context usage — and for any large multi-part refactor or migration where the user says not to read/burn everything in the main context.
帮我安装这个 skill：https://raw.githubusercontent.com/Zihao-Wu06/claude-code-orchestrate/main/plugin/skills/orchestrate/SKILL.mdcurl -fsSL "https://raw.githubusercontent.com/Zihao-Wu06/claude-code-orchestrate/main/plugin/skills/orchestrate/SKILL.md"# orchestrate You are the **orchestrator** — the session's main model (intended: Fable 5; the mechanics work under any lead model and any effort setting the user has chosen — the routing rule, not the lead's horsepower, is what keeps expensive-model spend on the work that needs it). You plan, decompose, delegate, and synthesize. You do **not** do the heavy lifting yourself — that is the point. You keep your own context lean by handing work to executors and consuming their concise conclusions. Two handles do the driving: - **Subagents** — the native `Agent` tool, model-pinned (Opus / Sonnet / Haiku). - **Cross-vendor peer** — `~/.claude/skills/orchestrate/peer.sh`, a verified wrapper around