hermes - 💡(How to fix) Fix DeepSeek provider: stream retry credential loss + 600s server-side stream limit [1 participants]

Summary

Two related issues when using DeepSeek provider via Anthropic-compatible endpoint (https://api.deepseek.com/anthropic) with long-running sessions:

Issue 1: Stream retry breaks credential pool

After a ReadTimeout stream drop, retry fails with credential error:

Stream drop on attempt 2/3 -- retrying.
  provider=deepseek base_url=https://api.deepseek.com/anthropic
  error_type=ReadTimeout elapsed=601.51s

Failed to rebuild shared OpenAI client (stream_retry_pool_cleanup)
  error=Missing credentials. Please pass an api_key...

The credential context from the initial connection is lost when rebuilding the client for stream retry.

Issue 2: DeepSeek 600s server-side stream limit

DeepSeek's Anthropic-compatible endpoint drops streams after ~600 seconds:

http_status=200 bytes=775 chunks=4 elapsed=601.51s ttfb=1.50s upstream=[server=openresty]
http_status=200 bytes=58354 chunks=287 elapsed=605.87s ttfb=0.89s upstream=[server=openresty]
http_status=200 bytes=41394 chunks=205 elapsed=604.86s ttfb=0.99s upstream=[server=openresty]

• HTTP 200 returned (connection established successfully)
• TTFB is fast (~1s)
• Connection drops at ~600s regardless of request_timeout_seconds (tested at 1200s)
• Upstream is openresty (nginx), suggesting server-side proxy timeout

Environment

• Hermes Agent v0.13.0
• Provider: deepseek (via Anthropic-compatible endpoint)
• Model: deepseek-v4-pro
• Config: request_timeout_seconds=1200, stale_timeout_seconds=1800, streaming.enabled=true
• macOS 26.2, Intel x86_64

Workaround

Higher request_timeout_seconds does not help (600s limit is server-side). Sessions under 10 minutes work fine. Enabling streaming.enabled=true may help with connection management.