Claude Code 三模型分工：Sonnet 主控、Haiku 集群、Opus 待命
用 claude --advisor 加三个 Haiku 子 agent 拆分搜索/审计，并把工具选择、重试等确定性分支交给 Jev，省下大模型 token。
HOW TO RUN SONNET AS LEAD HAIKU AS SWARM AND OPUS AS ADVISOR ON CALL Anthropic just shipped the cleanest multi-model split in Claude Code. one command, three models, way fewer tokens. the play: sonnet 5.5 drives the session > haiku 5.5 swarms search/grep/audit in parallel > opus 5.5 sits on call and only reviews the final diff steps: 1. open terminal and run claude --model sonnet --advisor opus https://code.claude.com/docs/en/sub-agents 2. create three haiku 5.5 subagents in .claude/agents: - search (maps the repo) - grep (finds call sites) - audit (checks the diff in a throwaway worktree) 3. add one rule to CLAUDE.md: send every repo search and test run to the haiku 5.5 subagents so the main context stays small 4. route fast deterministic forks (which file, which tool, retry or stop) through JEV so the big models never see them 5. paste this prompt so Claude rebuilds the tree and shows every change as a diff first: Rebuild my Claude Code setup around this tree: 1. Run the main session on Sonnet 5.5 with Opus 5.5 as advisor. 2. Create three Haiku 5.5 subagents in .claude/agents: search (maps the repo), grep (finds call sites), audit (checks the diff). 3. Give audit isolation: worktree so its checks never touch my working branch. 4. Add one rule to CLAUDE.md: send every repo search and test run to the Haiku 5.5 subagents so the main context stays small. 5. Route fast deterministic forks (tool choice, path selection, retries) through JEV. Show me every change as a diff first. Nothing gets edited until I say go. > the catch opus still burns tokens if you call it on every small decision. keep it strictly on-call. > what not to do do not run the whole session on opus or let haiku write code. wrong model = wasted context. sonnet builds. haiku swarms. opus reviews. bookmark this and reply with your current claude setup if you want the exact agents folder structure.