feat: persist agent stream on browser refresh

[Bl3f]

Problem

When a user refreshes the browser while an agent is running, the stream connection drops. The agent continues running on the backend, but the frontend loses its connection — the stream stops, the sidebar loader disappears, and the response appears frozen.

Solution

Leverage the Vercel AI SDK's built-in reconnectToStream / resumeStream support to resume the stream after refresh, and keep the sidebar showing a spinner for running agents.

Backend

• StreamBuffer (apps/backend/src/lib/stream-buffer.ts) — A new utility that consumes a ReadableStream, buffers all chunks in memory, and allows creating multiple independent readers. Each reader replays buffered content from the start and then continues with live data. This enables stream resumption without losing any data.

• AgentManager changes — The agent's stream() method now wraps the output through a StreamBuffer, and a new resumeStream() method returns a reader that filters out initial-setup events (data-newChat, data-newUserMessage) so they don't replay on reconnect.

• GET /api/agent/:chatId/stream — New resume endpoint that returns a createUIMessageStreamResponse from the buffer, or 204 if no agent is running.

• chat.activeChats tRPC query — Returns the list of chat IDs with actively running agents for the current user.

Frontend

• useActiveAgents hook — Polls chat.activeChats on app load (and every 10s while agents are active). Sets chatActivityStore.running = true for active agents so the sidebar shows spinners immediately after refresh. Also clears stale running state when agents finish between polls.

• useAgent hook — Configures prepareReconnectToStreamRequest on the transport so the AI SDK's reconnectToStream hits the correct URL. Computes a shouldResume flag based on: chat exists, DB messages loaded, activity store shows running, and no frontend stream is already active. Passes this as the resume option to useChat.

How it works

1. User refreshes the browser while an agent is running
2. On load, useActiveAgents queries chat.activeChats → backend returns running chat IDs
3. chatActivityStore is updated → sidebar shows spinners for running chats
4. When the user is on (or navigates to) a running chat, useAgent detects shouldResume = true
5. useChat({ resume: true }) triggers the AI SDK's resumeStream() → GET /api/agent/:chatId/stream
6. Backend returns the full buffered stream (replay + live) → frontend processes it seamlessly
7. If the agent finishes before the user reconnects, the endpoint returns 204 and the sidebar clears

  

[github-actions]

🚀 Preview Deployment

URL	https://pr-482-747f906.preview.getnao.io
Commit	747f906

⚠️ No LLM API keys configured - you'll see the API key setup flow when trying to chat.

---

_{Preview will be automatically removed when this PR is closed.}