Belay is a durable job engine for Elixir built on Postgres that makes the step — not the whole job — the unit of recovery. Completed steps are journaled and replayed in microseconds on retry, so expensive API calls, LLM invocations, and human-approval waits are never re-executed unnecessarily. Key features include memoized steps with cost tracking, per-job and fleet-wide spend budgets (including token-based rate limiting with true-up), human-in-the-loop via durable await/signal, dynamic fan-out DAGs, event streams, replay debugging, leaderless cron, encrypted inputs, and an embedded dashboard. It requires only Postgrex and Postgres — no Redis, no Ecto, no separate workflow server. The library ships at 1.0.0-rc.5 with formal TLC model verification (187M+ states, zero violations), a 7-hour chaos soak, and a proven migration path from Oban.
