Claude Code 多模型分工 + JEV 路由微决策

delost @thedelost原推发布于 2026.10.11

在 X 上打开原推

this is pure f*cking treasure Claude Code tip: once Opus 5.5 is your main model, stop babysitting it turn after turn and stop burning Opus tokens on work Sonnet 5 can swarm give it a finish line with /goal claude -p "/goal all tests pass and lint is clean, or stop after 20 turns" Opus 5.5 keeps writing the code turn after turn Sonnet 5 swarms the routine work at medium Haiku reads the transcript after every turn and answers one question: is the goal met? → no: the reason goes back and Opus starts the next turn → yes: the goal clears and an achieved entry is logged → never: your "or stop after 20 turns" clause ends it Haiku grades. Opus ships JEV engineering is the same move one layer down: the forks that need no thinker (which file, which tool, retry or halt) go to JEV in milliseconds, and the big models only see the ones that split the full tree > Opus 5.5 on high runs the main session > worker edits and patches code on Sonnet 5 > explorer maps files and call sites on Sonnet 5 > verifier runs tests and linters on Sonnet 5 > all three subagents on medium effort > Haiku grades the /goal after every turn > JEV routes micro-forks in milliseconds paste the tree and this prompt into Claude Code ↓ "Reconfigure my Claude Code setup around this tree: 1. Set the main session to Opus 5.5 on high effort. 2. Create three Sonnet 5 subagents on medium effort: worker (edits code), explorer (maps files and call sites, read-only), verifier (runs tests and linters, read-only). 3. The /goal evaluator only reads the transcript, so make the verifier print the real test and lint output before every turn ends. 4. Route fast deterministic forks (tool choice, path selection, retries) through JEV. 5. Add one rule to CLAUDE.md: run long tasks with claude -p '/goal <criteria>, or stop after 20 turns' Show me every change as a diff first. No edits until I say go." ↳ http://code.claude.com/docs/en/goal

这条是引用着另一条写的：x.com/hanakoxbt/status/2101709924828934222