[Stream Error] DeepSeek v4.1 Flash drops mid-stream after ~3.5 min: {"type":"server_error","message":"Network connection lost."} #1124

Summary

When using deepseek/deepseek-v4.1-flash on the /alpha/generate streaming endpoint, long-running requests (e.g. multi-turn tool calling with reasoning) can hang for ~3.5 minutes (208,974 ms) and then abruptly terminate with a mid-stream error event:

{"type":"error","error":{"type":"server_error","message":"Network connection lost."}}

Steps to Reproduce

1. Start a coding session generating a complex response or multi-turn tool execution with deepseek/deepseek-v4.1-flash.
2. Keep the stream connection open while the model reasons/generates tokens.
3. After ~3.5 minutes (observed at ~209s), the stream receives:

   {"type":"error","error":{"type":"server_error","message":"Network connection lost."}}

Observed Log Trace

[18:11:55] POST .../alpha/generate -> stream started
[18:15:24] ✗ ERROR: [CommandCode error: {"type":"server_error","message":"Network connection lost."}] (duration: 208974ms)

Expected Behavior

• The upstream connection between CommandCode's gateway and the DeepSeek backend cluster should maintain keep-alive/heartbeat or handle transient drops gracefully instead of terminating with server_error: Network connection lost. after 3.5 minutes.
• If an upstream drop occurs, provide retryable metadata or connection recovery.

Environment

• CLI / Wire Protocol: 1.88.5
• Model: deepseek/deepseek-v4.1-flash
• OS: Linux x64