mirror of
https://github.com/browseros-ai/BrowserOS.git
synced 2026-05-21 12:55:09 +00:00
0d5fa08063
The conversation page funneled every message through the server-side queue. Two consequences fell out of that: 1. No streaming UX — the queue worker dispatches server-side, the browser only sees status snapshots, and the assistant turn just "pops into existence" when history refetches. Killed the LLM streaming experience entirely. 2. Images take the queue's separate buildMessagePartsFromAttachments path, which is independent of the direct /chat path the user already verified end-to-end. Any breakage there only ever shows up here. Route sends through the existing useAgentConversation.send (direct /chat with token streaming) when nothing's in flight, and only fall back to the queue when there's actually a stream or queued item ahead. The initial-message handoff (?q= from home composer) goes through the same router. Tab-close safety for already-queued work is preserved — only the head-of-queue benefits from streaming, which matches the user's mental model: "type, stream; type again, queue behind it."