Nap is a durable coding-agent runtime for long-running software work. It builds applications inside isolated sandboxes and treats "done" as a claim rather than a fact: every turn is committed, then verified against the project's own checks. A failure opens a bounded repair turn instead of handing back broken code, and a commit becomes the "last known-good" checkpoint only once it passes. Durability is the design constraint. Events append to Postgres with monotonic sequencing before any client fanout, so job state is a fold over one event log and a reconnect an hour later is the same operation as a reconnect a second later. Idle projects are bundled to object storage and their sandbox destroyed, then restored on the next message rather than left billing. Execution lives on workers, not in the request: turns are enqueued in Postgres under a per-session lease and claimed by a worker, with a reaper running the idle sweep. The three processes run on Kubernetes with KEDA on queue depth and an HPA on open sockets — a k6 ramp to 100 concurrent turns across nine pods finished 2,310 turns with zero sequence gaps, duplicates or socket failures.