Problem

SSE connections through the proxy chain (Cloudflare → Envoy → Node.js) are killed after ~125 seconds of idle time because no keepalive mechanism exists.

Evidence from Production Logs

Envoy access logs show ALL SSE GET requests dying at exactly ~125s with DR (Downstream Reset) flag and 0 response body bytes:

[2026-01-23T10:23:54] "GET / HTTP/2" 200 DR 0 0 125032 1 "claude-code/2.1.17"
[2026-01-23T10:26:01] "GET / HTTP/2" 200 DR 0 0 125014 2 "claude-code/2.1.17"
[2026-01-23T10:28:07] "GET / HTTP/2" 200 DR 0 0 125030 2 "claude-code/2.1.17"
[2026-01-23T10:30:13] "GET / HTTP/2" 200 DR 0 0 125009 1 "claude-code/2.1.17"

Proxy Chain Configuration (Verified)

Client → Cloudflare (idle timeout ~100-125s) → Envoy (lb1) → Node.js (192.168.95.53:3333)

Envoy config is correct (no timeouts for this route):

• Route: timeout: 0s, idle_timeout: 0s
• HCM: stream_idle_timeout: 86400s
• Cluster: TCP keepalive configured (300s/30s/5 probes)

Cloudflare is the bottleneck — kills connections with no data after ~100-125s.

Sub-Problems

1. No SSE Heartbeat/Keepalive

The MCP SDK's StreamableHTTPServerTransport has no built-in ping/heartbeat mechanism. During long tool calls (up to 47s with retries in enhancedFetch), the SSE stream is completely idle.

SSE spec supports comment-based keepalives:

: ping\n\n

These are ignored by clients but keep the connection alive through proxies.

2. Node.js HTTP Server Default Timeouts

No explicit timeout configuration on the HTTP server:

• keepAliveTimeout defaults to 5000ms (Node.js 24)
• headersTimeout defaults to 60000ms

Between tool calls on the same HTTP/1.1 connection, a 5-second gap causes connection reset.

Proposed Solution

SSE Heartbeat (every 30s)

After SSE stream is established, send periodic keepalive comments:

// In server.ts, after SSE connection is established
const heartbeatInterval = setInterval(() => {
  try {
    controller.enqueue(encoder.encode(": ping\n\n"));
  } catch {
    clearInterval(heartbeatInterval);
  }
}, 30000); // Every 30 seconds — well under Cloudflare's 100s limit

// Clean up on stream close
stream.on('close', () => clearInterval(heartbeatInterval));

HTTP Server Timeouts

const httpServer = http.createServer(app);
httpServer.keepAliveTimeout = 620000; // 620s — above Cloudflare's max (600s Enterprise)
httpServer.headersTimeout = 625000;   // Must be > keepAliveTimeout
httpServer.timeout = 0;               // No socket timeout for streaming

Fix Checklist

• Add SSE keepalive ping (:ping\n\n every 30s) to Streamable HTTP transport
• Configure keepAliveTimeout on HTTP server (620s for proxy compatibility)
• Configure headersTimeout on HTTP server (> keepAliveTimeout)
• Set server.timeout = 0 for SSE streaming support
• Make heartbeat interval configurable via env var (GITLAB_SSE_HEARTBEAT_MS)
• Add integration test: verify SSE stream survives > 125s with heartbeat
• Document proxy chain timeout requirements

Impact

Serious — All SSE connections die after ~2 minutes regardless of transport routing fix (#138). This can be developed in parallel with #138.

Files to Modify

• src/server.ts — HTTP server timeout configuration + SSE heartbeat setup

Relationship

Independent of #138 (transport routing). Can be developed and tested in parallel. Both fixes are required for stable production operation.