I was burning 300k tokens on one ticket and stalling at 60k so I tried this (Foreman: tiered, budgeted, checkpointed)https://www.reddit.com/r/programming/comments/1vzyohi/i_was_burning_300k_tokens_on_one_ticket_and
Every framework let one model do everything until it stalls at 55k with cacheRead≈0. Foreman is different: Hierarchy, not flat: director → lead (premium, 90k ctxCap) → coder/tester (standard, 60k) → drone (economy, 40k). routing[implement]=coder , routing[plan]=lead . Premium never implements — it judges from 6k briefs. Memory, not amnesia: every run has steps/outTokens/ctxCap/ctxKill/usd enforced mid-run ( lib.mjs:checkBudget() one deep seam, pure, testable). Past ctxCap , --continue is refused — next run starts fresh from a deterministic 6k checkpoint brief (git diff + REPORT +…