按决策难度分工：Jev 做反射层，Opus 只做判断
Jev 每百万 token 0.042 美元对 Opus 的 4 美元，于是把选文件、选工具、重试或停这类微决策交给 Jev，大模型只看真正难的决定。
OPUS 5.5 SHOULD NEVER READ YOUR FILES OR DECIDE “RETRY OR STOP” $4 per million input tokens for Opus. $0.042 for Jev. same decision, 95x apart so split the job by what each decision actually needs: Jev: reflexes. which file, which tool, retry or stop. answers in milliseconds Haiku 5.5: eyes. reads and audits every file in parallel Sonnet 5.5: hands. writes each fix Opus 5.5: the head. writes the plan, judges what comes back the trap: a workflow agent runs on your session’s model unless something assigns another one. start Claude Code on Opus and every agent is Opus too the fix is two lines. put model: haiku in your scanner’s agent file and model: sonnet in your editor’s, both in .claude/agents/ claude --model opus --effort ultracode first run on a small slice (/config workflowSizeGuideline=small). with ultracode on, the “Large workflow” warning doesn’t show, so watch /workflows while it runs the biggest model should only see the decisions that are actually hard
这条是引用着另一条写的：x.com/rvaniaaaa/status/2109385796751114346