Claude Code 分层用模型：Sonnet 主会话、Haiku 子 agent、Jev 管分支

Woody @woody_research原推发布于 2026.10.20

在 X 上打开原推

YOU'RE PAYING OPUS 5.5 TO DO THE EASY WORK. CLAUDE CODE CAN STOP THAT every search, every log dump, every test run lands in the most expensive context you have the fix is a split by job, not by vibe: → Sonnet 5.5 holds the main session: plans the steps, writes the code, applies the fixes → Haiku 5.5 subagents do the reading in parallel: repo map, call sites, docs, test runs → Opus 5.5 stays out of the room until a call is expensive to get wrong why it matters: a single test run can dump thousands of lines into the main session. after that, every answer Sonnet gives is paying to re-read your logs the switch: CLAUDE_CODE_SUBAGENT_MODEL=haiku claude --model sonnet the catch: that variable alone skips the built-in Explore and Plan agents. add CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1 and every subagent follows it the three moments Opus is worth its price: → right before a plan gets locked → the second time the same error shows up → the last look at the diff before you merge Sonnet builds. Haiku reads. Opus signs. the setup > main session: Sonnet 5.5 > .claude/agents/search.md → model: haiku, read-only > .claude/agents/callsites.md → model: haiku > .claude/agents/tests.md → model: haiku, effort: low > .claude/agents/audit.md → model: haiku, isolation: worktree > Opus 5.5 as the advisor for the hard calls > Jev for the yes/no forks: which file, retry or stop paste this into Claude Code ↓ "Move every repo search, call-site lookup and test run into Haiku 5.5 subagents in .claude/agents. Keep the main session on Sonnet 5.5. Call Opus 5.5 only before a plan locks, after a repeated error, and before the final diff. Add one CLAUDE.md line: raw logs never go back to the main session, only summaries. Show me the diff of every file first and wait." ↳ http://code.claude.com/docs/en/sub-agents pay Opus for judgment. pay Haiku for reading

这条是引用着另一条写的：x.com/woody_research/status/2108159353047920705