Verified against Semitexa Ultimate 2026.10.03.1952

Live by Default

A page is live without publish code. A database write through the ORM publishes an invalidation; every open feed and island that watches it re-runs, for its own visitor, in whichever worker holds the page's stream. This page is the map of the pieces. Each links to its own page.

One stream per page

A page has one KISS stream (/__semitexa_kiss). Everything the page sends goes through HUG (/__semitexa_hug), and everything the server pushes comes back on KISS. There is no per-feed connection and no WebSocket server. See SSE Stream.

• Resumable. Every data frame carries an id. A reconnect names the last one it got and is given what it missed, or told to reset. The lifecycle frames (connected, close) carry no id, and deferred blocks keep their own numeric ids, which the replay ring does not hold.
• Prompt. A frame queued for a stream on the same worker is written at once. Deferred content measured 26 ms after DOMContentLoaded; it was about 380 ms while the stream loop slept out its tick.
• As the visitor. Every re-run, deferred region and island is rendered as the visitor of that page: re-resolved from its session each time, so a visitor who signs out stops receiving data.

Lists: keyed patches

A grid subscribes to its feed through HUG and says it applies patches. When a write re-runs the feed, the server sends only the rows that changed against the page it last sent (ui.collection.patch: upsert, remove, order). The grid keeps every unchanged row's element, so focus, a checked box and an open menu survive.

Components: islands

A component that implements UiIslandInterface::watches() is drawn again when what it reads is written, and morphed in place. platform.dashboard is one. See Components.

State on the server

#[UiState] properties are a component's state, kept in the shared store under a key only the visitor's session can name. A handler changes them, and the component is saved and drawn again. See Components.

Deferred, concurrently

Deferred components render in parallel, each in its own coroutine, and arrive as they finish. Measured with two widgets that each take 0.4 s: about 0.4 s together, not 0.8. See Deferred Blocks.

Answers at once

ui-optimistic draws the answer an action expects the moment it fires, and undoes it if the server refuses. A grid's delete is optimistic by default. See UI Events.

Watching it

In development, every page keeps a timeline:

• which component sent which event, how long it took, and the effects that came back;
• every frame its stream wrote;
• each subscription and re-run, with what the re-run sent.

bin/semitexa ai:observe timeline --id=sse_…

The same timeline is at /__observatory/timeline. See AI Tooling.

Proof in the Playground