Chat stream keeps disconnecting

The chat panel shows "reconnecting" every minute, or your custom SDK consumer drops without an error. Find the cause.

The goal

A stable WebSocket that survives long turns and ordinary network blips.

Steps

1. Check the proxy timeout.

   Most "60-second drop" issues are a reverse proxy with a 60-second idle timeout (Nginx default). Raise to at least 5 minutes:

   platos.example.com {
     reverse_proxy webapp:3030 {
       transport http {
         dial_timeout 30s
         response_header_timeout 5m
         read_timeout 5m
         write_timeout 5m
       }
     }
   }
   

2. Force WebSocket transport.

   Some networks force long-polling fallback. The SDK respects transports:

   const stream = await platos.threads.stream({ threadId, transports: ["websocket"] });
   

   Or in the chat UI's URL ?transport=websocket.

3. Check the early-message buffer.

   The race-fix invariant in tool-sync-ws.service.ts:130-135, 281-284 buffers early messages during the auth handshake. A custom client that bypasses it can drop the first frames after reconnect. If you wrote your own client, mirror the buffer.

4. Check the SDK reconnect backoff.

   @platosdev/client reconnects with exponential backoff (1s, 2s, 4s, ..., max 30s). After 30 seconds disconnected, it re-fetches message history rather than replay. If you see "lost" messages on a long disconnect, this is why; bridge through messages.list.

5. Check the audit log.

   monitoring/admin-audit shows auth.session_token.expired events. Session tokens are 5 minutes; long streams need refresh pings. The SDK does this automatically; raw WebSocket clients must implement.

Verify

• Long turns (>1 minute) complete without UI flicker.
• The browser's network tab shows the WebSocket as Upgraded and persistent.
• A simulated network drop reconnects within 5 seconds.

Common findings

• Cloudflare or similar in front with default 100-second timeout. Configure WebSocket-friendly timeouts.
• Corporate proxy that strips Upgrade headers; transport falls back to long-polling and adds latency.
• A misbehaving load balancer health check that closes the socket on health probe.

Next steps

• Trace a single turn to confirm the turn itself completed engine-side.
• Connect an entity (TypeScript) for the entity-side equivalent of the same patterns.