I stumbled up on https://github.com/powersync-community/powersync-convex-todolist-demo which is an experimental way of integrating a Local sqlLite with Convex. Maybe the Power Sync community could also help to Build a spacetimeDB Integration 🤔

I'd like to propose a scope for this that keeps SpacetimeDB out of the "sync engine" business: ship a small set of mechanisms and leave the policy (what an intent means when it lands late) to the developer, who knows their domain.

Opt-in per reducer. Something like #[reducer(offline)] marks a reducer as safe to apply late. Only those get queued; everything else keeps today's fail-fast behaviour, so existing apps are unaffected.

The mechanisms:

1. Idempotent reducer calls. Calls carry a (client_stream, seq) id; the server keeps an applied set per stream and stores results, including errors, so a retry returns the stored outcome instead of running the reducer again. This is the same shape as Async IDC's st_inbound_stream / st_inbound_msg. It's useful on its own: it would let auto-reconnect (#5927) safely resend calls that currently settle with UnknownCallResultError. I'd suggest it as the first step.
2. Intent context on the reducer context. The client timestamp, the last transaction offset the client had observed, and a replay flag, so a reducer can tell that an intent is stale and apply its own policy (reject, merge, last-write-wins).
3. Resumable subscriptions. Subscribe from a transaction offset, falling back to a full resend when it's too old, so reconnecting after days doesn't re-download everything.
4. A persistable client cache and a durable outbox in the SDKs, with hooks such as onApplied / onRefused.

Left to the developer: conflict semantics, how refusals are surfaced in the UI, what data is kept offline, and platform storage persistence. The docs would need to state the risks of applying intents late plainly.

Prior art: spacetimedb-predict covers the latency side (in-memory only, with a 10 s timeout); Replicache uses the same split between mutators and server-authoritative rebase.