Resumable useChat without Redis

Vercel AI SDK useChat resume often assumes Redis. FluxyChat does not. The session is the room.

What you get

FluxyRoomChatTransport maps sendMessages to invokeAgentRest and reconnectToStream to GET /ai/streams/active then GET /ai/streams/:id/resume. Those routes read RATE_LIMIT_KV or STREAM_RESUME_KV. No Redis.

Live token paint still uses the room WebSocket. The HTTP body is the full reply encoded as text-delta chunks.

If the room Durable Object hibernates mid-call, that LLM request does not resume like Cloudflare Think chatRecovery. Generation for POST /agents/:id/invoke runs on the Worker isolate, and when AGENT is bound it runs on AgentDurableObject id room-run:{project}:{room} so two invokes serialize. Kill the room object; the run object can keep going. Kill the run object and you lose the in-flight model call.

Clone

npx @fluxy-chat/create-fluxy-chat@latest my-room --example shared-ai-room

import { FluxyChatClient, FluxyRoomChatTransport } from "@fluxy-chat/sdk";

const client = new FluxyChatClient({
  baseUrl: "https://your-worker.example.workers.dev",
  userId: "alice",
  token: jwt,
});
const transport = new FluxyRoomChatTransport({
  client,
  roomId: "general",
  agentId: "assistant-id",
});

Pass transport to AI SDK useChat({ transport }). Do not import ai from @fluxy-chat/sdk.

Mid-turn inject: POST /rooms/:roomId/agent-steer with { content }. Stop from another tab: POST /rooms/:roomId/agent-stop. Room stopAgentStream still aborts the Worker fetch.

Dev.to / Show HN posting is on you. This page is the tutorial.