State is what completed tasks leave behind.
Many agents work one project at the same time. Nobody merges, and nothing is lost — because state is never stored. It is derived, by a deterministic fold, from an append-only record of completed tasks, and the losing versions stay in the record, cloneable, with the reasons they lost.
This page is a projection of that record — a view, not the truth. It can prove it: burn this desk's memory and it rebuilds itself, identically, from the record alone.
The record will appear here, one completion at a time, in the fold's application order.
No state yet — nothing has completed.
Everything the desk holds — the state above, the conflicts, this page's picture of the project — is derived from the record and kept only as a cache. Burn the cache and the desk holds nothing. Rebuild, and if the digest comes back identical, the desk never was a second source of truth. It is safe to try: the record is untouched, and the desk also heals itself on its 60-second cadence.
Digest of the derived state, as the desk now holds it:
—
Digest before burn: —
Digest after rebuild: —
The replay above walks the record in the fold's application order. But agents don't arrive in order — they finish when they finish. So take the same records, scramble the order they arrive in, and fold again, here in your browser, with the very same code. Step through it: the halfway houses may differ from the canonical walk. The destination may not. The fold doesn't care what order work arrived in — the same completions land on the same state.
No shuffled fold yet.
Canonical-order digest (the desk's live digest): —
Shuffled-fold digest (your browser, this arrival order): —
—
—
—
The drafted record appears here — parents set to the current frontier.