Claude Code 多模型分层：Opus 规划、Haiku 扫描、Sonnet 修复
用 Opus 5.5 写计划、Haiku 子代理并行扫文件、Sonnet 子代理修问题、Haiku 复核，并用 Jev 裁决重试或停止等 yes/no 分支以压低账单。
THIS IS F*CKING GOLD Claude Code can run Opus 5.5 as the boss and a team of cheap agents as everyone else. the split that keeps the bill down: → Opus 5.5 writes the plan and goes quiet → Haiku 5.5 subagents scan the files in parallel → Sonnet 5.5 subagents make the fixes → a second Haiku 5.5 pass checks every fix Opus only wakes up twice: → before the run, to write the plan → after it, to sign off on what passed Opus plans. Haiku reads. Sonnet writes. the setup > main session: Opus 5.5 on xhigh (/effort xhigh) > .claude/agents/scanner.md with model: haiku > .claude/agents/patcher.md with model: sonnet > .claude/agents/checker.md with model: haiku > Jev for the yes/no forks: retry or stop, which lane paste this into Claude Code ↓ "Create three subagents in .claude/agents: scanner (model: haiku, read-only, reports findings), patcher (model: sonnet, fixes one finding at a time), checker (model: haiku, confirms each fix). You stay on Opus 5.5: write the plan, dispatch the subagents in parallel, and show me only the findings the checker confirmed. Before touching any file, show me the diff and wait for my ok." ↳ http://code.claude.com/docs/en/sub-agents save it for your next big refactor
这条是引用着另一条写的：x.com/0xkiyoro/status/2107135100911563017