@irisrun/core

The pure durability engine — own your state as a journal that replays byte-for-byte. runTurn drives one turn over an append-only journal: acquire-lease → replay → recover → step/effect/wait/finish, checkpoint before every effect, and re-run replay after every committed step to assert the reconstruction byte-equals live state. No I/O of its own — the store, scheduler, clock, and performers are all injected — and it imports no host, transport, or Node-only API, so the same engine runs unchanged on a server or the edge.

What it is

The journal is the single source of truth (JournalRecord); state is the result of replaying it (replay + Reducer). assertReplayConsistency is the always-on defense against determinism bugs — it throws ReplayDivergenceError on any divergence, never a silent pass (the engine decides whether to call it; the assertion always asserts). The effect engine journals an intent, performs it, and journals the result, so recovery and replay reuse the same machinery. A single-writer lease over StateStore.cas returns a monotonic Version — the fence — and mutual exclusion is enforced by fencing (a taken-over holder is rejected at append with stale_fence), not by acquire-time locking. shouldSnapshot / migrateSession bound replay cost and move sessions forward. Everything rides two ports — StateStore and Scheduler — that host adapters implement. The harness layer adds a pure agent loop: harnessProgram (a kernel encoded as a Program<HarnessState> over the unchanged runTurn), composable seams (composeGate / composeDecideNext / composeAssemble), and a defaultBundle of tactics — a seam consultation is journaled as a tactic effect, so replay never re-invokes a tactic.

Use it

Library-only (no CLI). Implement the two ports for your host, register your performers, and call runTurn; the harness kernel and defaultBundle give you a ready agent loop.

See docs/Architecture and docs/Introduction.

Part of Iris — own, portable, verifiable state.