@irisrun/store-fs

The serverless host that proves portability across machines. A node:fs adapter that holds no long-lived handle — every method opens, reads/writes, and returns, so a fresh instance over the same root behaves identically to a reused one (the cold-per-turn invariant). It enforces the same CAS / fencing / high-water-mark / snapshot invariants as the sqlite, memory, and DO stores — the invariants that make replay byte-identical — so the same image resumes the same session on a different machine by construction. This is host B for the cross-host portability proof.

What it is

FsStateStore + FsScheduler implement the two ports (StateStore + Scheduler) over atomic filesystem primitives. CAS is an O_EXCL create of <expected+1>.json — a true atomic decision point, no read-modify-write window — and append is fenced and dense via the same O_EXCL check, with all-or-nothing rollback. The hwm is derived from the snapshot + journal directory (no mutable sentinel to race), and writeSnapshot seeds it by raising snapshots/<upToSeq>. FsScheduler keeps timers/signals as durable files under _wake/, so a cold instance sees prior wakeups; dueWakeups peeks, confirmWoken consumes only after the resumed turn commits. Node-only, so it lives in a host adapter, never in core. Not a determinism bypass — it is a real port impl.

Use it

node --conditions=iris-src examples/portability-demo.ts   # resume on a different host

See docs/Deploy and docs/Architecture.

Part of Iris — own, portable, verifiable state.