Claude Code multi-model split: Jev runs the fork-routing layer
Sonnet 5.5 codes, Haiku 5.5 runs subagents, Opus 5.5 stays on call, while Jev routes each fork by typed answers — which file, which tool, retry or stop — logging 1,331 forks.
WHOEVER LEAKED THIS CLAUDE CODE SUBAGENT TIP SHOULD BE REPLACING YOUR TASK MANAGER I thought running multiple AI models together meant one model in a loop and three browser tabs you forget to close. Apparently not. This setup is much more interesting: → SONNET 5.5 is the LEAD (pid 1201, CPU 93.5%, 0.4G, RUN), the only agent that writes code, runs tests and resumes after a fork lands, no other agent allowed to touch main → HAIKU 5.5 runs the SWARM (3 subagents: search sleep, grep RUN 75%, audit RUN 91%), each in its own worktree, maps the repo at 68% while sonnet writes → OPUS 5.5 is the ADVISOR (ON CALL, pid 1205, CPU 42%, 1.1G), reads 313k tokens of session, 3 calls total, silent on routine turns, never writes code itself → JEV runs the FORK LAYER (which file p=0.94, which tool p=0.94, retry or stop p=0.80), 1,331 forks logged in router.jsonl, every fork routed by the typed answer → SPLIT (pid 1209) kicks any "same error x2" back to sonnet, flags 7 forks whose confidence drifted past the sharp threshold this run so you get: Leads coded. Swarms mapped. Advisors silent. Forks typed. the interesting part isn't more models. it's giving every role one pid the task manager can't blur with the others. Instead of one Opus call eating 313k tokens on every turn, you run sonnet as lead, haiku as swarm, opus as on-call advisor, jev as forker, with 1,331 forks logged and 517 sonnet + 671 haiku + 279 opus tokens per second. Lead with Sonnet. Swarm with Haiku. Fork with JEV.
This post quotes another post: “Long read: how to build a Jev-powered agent stack”