REACT · server-threads

Real OpenAI text through the actual Hashbrown SDK. Unchanged provider events are paced at 20 ms for filming. Footage is slowed to half speed; no narration. The loopback demo session and process-memory repository are application integrations, not production authentication or durable agent checkpoints.

Start a local demo session. Identity stays on the server.

Load Lighthouse before sending a request.

OpenAI streams the named plan into text chat.

Save completed text with the loaded server version.

A browser refresh starts a fresh capture.

Load the saved thread. No model request.

The next request includes the saved conversation.

Another client saves a newer server version.

Save returns 409. Unsaved text stays visible.

Reload explicitly adopts the newer server snapshot.

Review the actual HTTP operations and SDK messages.

Replay makes no requests or writes.

Back live.

Server saves are explicit completed-text snapshots with owner and version checks. No tools, generated UI, shared state or paused checkpoints are restored. The competing client makes a real independent HTTP write; a stale save returns 409. Anonymous and foreign-owner checks stop before provider invocation. Refresh clears SDK sessions and captures; loading server text starts no generation. Replay makes no requests or writes.