Core concepts

These six pages are the mental model. Read them in order — each builds on the last, and together they cover everything a fresh reader (human or AI) tends to get wrong.

1. The two paths — read and write are separate and asymmetric.
2. The write path — stage locally, flush a batch, apply in the database.
3. The read path — shapes stream Postgres → Electric → PGlite, via a proxy.
4. The Electric subquery requirement — the mandatory flag, and why it fails closed.
5. Timestamps — microsecond integers carried as decimal strings.
6. Local schema & DDL parity — what local PGlite does and does not replicate.

Once you have the model, the Registry entry options page is the field-by-field reference for configuring a sync table — every option, with a short example, what it achieves, and when to use it. And Worker mode covers the browser topology in which tabs attach through a SharedWorker and capability placement chooses a Safari SW-direct or elected Chromium/Firefox engine (defineSyncWorker / attachSyncClient) instead of the calling thread.

Not everything is sync state. The event lane is the second lane beside the sync rail, for high-volume append-only facts that nothing ever reads back down: appendEvent into a local Outbox, flushed to an ingestion endpoint, delivered through a queue to a consumer callback you write.

The canonical vocabulary for all of these lives in the repository’s CONTEXT.md.