React Todo tests were doing asynchronous work outside the wrapper React expects, so they printed warnings even when the tests passed. I put the waits inside the right wrapper and ran the browser files one at a time because they share temporary server state, leaving the tests green without the noisy output.
Moon Lander had moved to a newer way of saving data, but its test helper still pretended the old methods existed. I updated that helper to follow the current insert-and-update flow, so the four tests can reach their real assertions again.
I made Better Auth work when user IDs are ordinary strings instead of UUIDs. Browser and server code now turn the same ID into the same internal identity, including for WebSocket connections, while keeping the original ID available and protecting built-in session fields.
I fixed pending transaction overlays that were keyed only by row ID. They now also include the table name, so two tables can safely have the same row ID without one table's pending value hiding the other's.
I renamed the public Rust API's batch terminology to transaction terminology. Internal query-engine and wire-format names stay unchanged where they are implementation details, while the public API, bindings, examples, and tests now use the clearer vocabulary.
I tightened the configuration types for inserts and authentication so incompatible options cannot be combined accidentally. I also made local-first authentication return its new token through configuration inspection without exposing the original secret.
I removed two old configuration options that had newer replacements: the authentication secret storage key and the combined permissions setting. The affected example now uses the current, separate options.
I added focused tests for two synchronisation guarantees in the new core. They prove that a rejected update does not block a later fresh commit, and that global reads eventually use the authoritative remote value while a local write remains pending.
I made the database recover cleanly when a storage commit fails before anything is durably written. The in-memory state and sequence clock are restored, so an exact retry can succeed instead of leaving the database permanently poisoned.
I changed single-row update and delete policy checks to read only the requested row. They no longer load every row in the table just to find one matching identifier.
I fixed subscriptions losing local inserts, updates, or deletes while they were applying a server snapshot. Pending changes now remain visible until they are accepted or rejected, while genuine server changes still reconcile normally.
I added a local-first authentication mode to React's JazzProvider. It loads the saved secret before creating the client, shows the configured fallback while that happens, and keeps the existing authentication modes unchanged.
I fixed query paths that treated every column named id as Jazz's internal row identifier. When an application declares its own id, filters and joins now use that application value, while tables without one keep the existing behaviour.
I renamed Vue query subscriptions' reactive loading property to isLoading. The behaviour stays the same, but the name now makes it clearer that the value is a boolean state.
I renamed Svelte query subscriptions' reactive loading property to isLoading. Existing loading, success, and error transitions continue to work while the public name better matches the value.
I removed the old propagate query option so propagation is the single public way to control forwarding. Applications that need a query to stay local can now use propagation: "local-only" consistently.
I added reactive one-row helpers for React, Vue, Solid, and Svelte. They keep the query live while making loading, empty, and populated results clear without making developers pull the first item out of an array themselves.
I changed upsert to take the table, row ID, data, and optional settings in that order. This keeps the row ID separate from mutation settings and matches the shape used by the other row mutation methods.
I made global sequence allocation fail safely when the counter reaches its maximum value. The final value can be used once, but any later allocation now returns a clear error instead of wrapping around and reusing an old sequence number, and every current allocation path uses the same guard.
I replaced loosely shaped subscription messages with typed events that the native and TypeScript layers can agree on. The subscription manager now prepares the exact row layout once, so each update can decode fixed positions instead of guessing from field names, while unknown or redefined layouts fail clearly.