Using Jev as a micro-decision layer to split Claude Code across worktrees
Jev acts as a routing layer that uses sharpness scores of 0.54, 0.83 and 0.89 to decide which file to edit, which agent to dispatch, and whether to retry or stop, alongside three isolated worktree agents editing code in parallel.
this is pure f*cking gold for anyone still running one Claude Code agent at a time and wondering why one feature takes a whole day the Claude Code worktree tip hit the front page yesterday. one shell flag stacks Sonnet, JEV and three worktree agents, each one doing only the job it's best at three things it cracks open in a single session: > Sonnet 5.5 as main session: reviews, never edits. runs on high effort, --worktree feature-auth, 12 commits merged by hand across the session > JEV fork layer: 4,295 micro-forks per session with a sharpness score. 0.54 which file, 0.83 which agent, 0.89 retry or stop. sharp runs land in code, split goes back to Sonnet > 3 worktree agents in isolation: refactorer, migrator and tests running in parallel branches (+184 -34, +0 -0, +125 -19), each one locked to its own checkout while Sonnet reviews none of it is a guess every fork logs a decision: 17:04:42 jev retry or stop → stop + ask, p=0.89, sharp to code - getting going > one shell flag connects the whole tree: claude --worktree feature-auth > paste the worktree spec and the 3-rule isolation prompt into CLAUDE.md, both are in the article > run it once and watch the session log. your main checkout stays untouched on 100% of writes, by refusal - the boundary > Sonnet reviews. JEV routes. the three worktree agents edit. main stays yours > no agent touches another agent's worktree, no model does another model's job, which is the whole reason your bill drops without the output dropping give this setup one evening, and running Claude Code on one agent at a time stops making any sense
This post quotes another post: x.com/hanakoxbt/status/2110067711363457244