I stopped letting Opus write my code (and my weekly quota survived)

one day claude ate 62% of my weekly quota in a single day. i was letting opus write every line.

so i changed one rule: claude plans and reviews, cheaper models write the code.

the split

• lead (claude): reads the task, splits it into small pieces with the exact files and an acceptance check, then reviews the diff
• workers (codex, opencode, free models): write the code for one piece each
• lead again: runs the tests itself, reads the real diff, sends it back if its off

the trick is the spec. a worker that gets "fix the login bug" burns tokens wandering around the repo. a worker that gets "edit src/auth.ts lines 40-80, expected behaviour X, run npm test auth, short report" finishes fast and cheap.

what broke when i did it by hand

• 5 terminals, copy pasting context between them
• forgetting which agent was on which file
• two agents editing the same file
• a long build dying at the 5 hour limit in the middle of a task

what fixed it

• one git worktree per agent, so they never share a checkout
• one memory file in the repo, so the next session doesnt start from zero
• the lead always looks at the real diff, never trusts the worker's report

you can do all of this with plain terminals and a CLAUDE.md. i ended up building a desktop app for it, foxpit, but the split works without it too.