[Bug] WebSocket heartbeat defaults (2s interval + 2 missed) tear down connections on mobile / high-latency networks → ~12s reconnect loop

[tianxiaz]

TL;DR

关键点	内容
根因	dsh-api-gateway 心跳默认 2s 一次、连丢 2 个 pong 就 socket.terminate()
影响	移动 / 高延迟网络下健康连接被误杀 → ~12s 周期性重连
修复建议	默认间隔改 30s；MAX_MISSED_HEARTBEATS 改为可配置；terminate 前加日志
已验证 workaround	配置 websocketHeartbeatIntervalMs: 60000

Summary

dsh-api-gateway's WebSocket heartbeat defaults are too aggressive for mobile and high-latency networks. The server terminates a healthy connection after roughly 4–6 seconds of PONG delay — which mobile networks routinely exceed — producing a permanent reconnect loop (~12s cycle) for remote/domain-access users.

Environment

• DeepSeek Harness: 0.2.0-rc.2
• Access: reverse-proxied domain (https://… → nginx → 127.0.0.1:3080)
• Network: China Mobile 5G (high latency + intermittent packet loss)
• Browser: macOS Safari / Chrome

Root cause

In packages/dsh-api-gateway/lib/index.js (compiled 0.2.0-rc.2):

const DEFAULT_WEBSOCKET_HEARTBEAT_INTERVAL_MS = 2e3;   // line 551 — 2 seconds
const MAX_MISSED_HEARTBEATS = 2;                        // line 172 — hard-coded

RemoteStreamMuxServer.startHeartbeat() pings every websocketHeartbeatIntervalMs (default 2s) and calls socket.terminate() once MAX_MISSED_HEARTBEATS (2) consecutive PONGs are missed (line 244). A single PONG round-trip delay of >~4s — routine on mobile — therefore tears down an otherwise-healthy connection.

Observed behavior

• nginx error log, repeating every ~10–14s:
  • recv() failed (104: Connection reset by peer) while proxying upgraded connection … /api/remote.mux
  • send() failed (32: Broken pipe) while proxying upgraded connection … /api/remote.mux
• nginx access log shows /api/remote.mux reconnecting every ~10–14s in bursts.
• The DSH server journal is silent — the heartbeat termination is not logged, which made this very hard to diagnose (the disconnect looks like a client-side RST, not a server decision).

Suggested fix

1. Raise DEFAULT_WEBSOCKET_HEARTBEAT_INTERVAL_MS from 2e3 to 30000 (industry-standard keepalive is 20–60s).
2. Make MAX_MISSED_HEARTBEATS configurable instead of a hard-coded constant.
3. Log a warning before socket.terminate() (missed count + interval) so this class of issue is diagnosable.

Verified workaround

Setting websocketHeartbeatIntervalMs: 60000 on the typert-gateway (@deepseek-ai/dsh-api-gateway) config eliminates the reconnect loop:

- id: typert-gateway
  name: "@deepseek-ai/dsh-api-gateway"
  config:
    websocketHeartbeatIntervalMs: 60000

Related

• Discussion [Bug] Question/approval cards missing on mobile until page refresh — WebSocket silently suspended, no heartbeat / no visibilitychange reconnect #2842 — WebSocket silently suspended on mobile
• Discussion [Bug Report] Web UI 在后端进程退出后无任何可见提醒（ConnectionBanner 组件存在但从未挂载） #916 — ConnectionBanner never mounted

---

中文一句话简述：

「dsh-api-gateway 的 WebSocket 心跳默认值过于激进：默认 2 秒 ping 一次、连丢 2 个 pong 就 socket.terminate()。移动/高延迟网络下 pong 单程超过约 4 秒即触发误杀，导致 ~12 秒一次的重连循环；且断开时无日志、难以排查。建议：默认间隔改 30s、MAX_MISSED_HEARTBEATS 改为可配置、terminate 前加日志。」