Agent processes die mid-task: the laptop sleeps, a container redeploys, something runs out of memory. The usual fix is to look at what happened and tell it to continue. Pi Durable treats that as a storage problem. Every step is a task that checkpoints before moving on, and a new process opens the same storage and carries on.
The detail I’d build on is replay. A tool reruns after a crash only if it declares that safe; an interrupted deploy is reported to the model, never repeated. That is the right default for anything that touches real hardware. The SQLite storage uses no Node APIs, so with a small adapter it runs inside a Cloudflare Durable Object. That’s where I’d try it first.
The story — Earendil and the Pi community shipped Pi 1.0 alongside an experimental package, Pi Durable: a framework for long-running agents that does not replace the Pi coding agent. A harness opens over memory, SQLite or JSONL storage, runs many conversations concurrently, and lets conversations fork without copying history. Unfinished tasks resume from their last checkpoint after a crash. The source is about 15,000 lines without tests. (Source)