Streaming ECONNRESET / "socket connection was closed unexpectedly" with SGLang OpenAI-compatible provider during tool-using agent runs (Qwen3.5 + reasoning content) #22803

Description

What happened?

When using opencode with a self-hosted SGLang server serving Qwen3.5-27B-FP8 as a custom OpenAI-compatible provider, streaming chat completions fail intermittently during agent/build-mode runs with:

• ECONNRESET on POST /v1/chat/completions
• UI message: Cannot connect to API: The socket connection was closed unexpectedly. For more i...
• Auto-retry loops indefinitely (Connection reset by server, retrying in 25s - attempt #21)

There is no nginx / reverse proxy / load balancer between opencode and SGLang — it is a direct TCP connection from the opencode sidecar to http://<host>:8000/v1/chat/completions.

Sidecar error logs

ERROR service=llm providerID=aifactory modelID=qwen3.5-27b-fp8
  sessionID=... agent=build mode=primary
  error={"error":{"code":"ECONNRESET","path":"http://<host>:8000/v1/chat/completions","errno":0}}
  stream error

ERROR service=session.processor
  error=The socket connection was closed unexpectedly.
  For more information, pass `verbose: true` in the second argument to fetch()

Plugins

none

OpenCode version

1.4.3

Steps to reproduce

1. Start SGLang serving Qwen3.5-27B-FP8 with an OpenAI-compatible endpoint:

   python3 -m sglang.launch_server \
     --model-path Qwen/Qwen3.5-27B-FP8 \
     --host 0.0.0.0 --port 8000 \
     --tp-size 2 \
     --mem-fraction-static 0.9 \
     --context-length 92000 \
     --chunked-prefill-size 8192 \
     --max-running-requests 12 \
     --reasoning-parser qwen3 \
     --tool-call-parser qwen3_coder \
     --api-key sk-

2. Configure opencode with a custom OpenAI-compatible provider pointing directly at http://<host>:8000 (no proxy in between), model id qwen3.5-27b-fp8.

3. Start a build-mode / agent session and give it a task that triggers tool use, e.g. "continue working on the frontend — explore the existing code structure first".

4. Observe:

   • First few tool rounds may succeed.
   • After 1–3 rounds the stream drops with ECONNRESET / "socket connection was closed unexpectedly".
   • opencode enters an infinite retry loop (Connection reset by server, retrying in 25s - attempt #21).

Workarounds we tried

• Switching SGLang to --tool-call-parser qwen3 (instead of qwen3_coder) changes the parser routing and appears to reduce the frequency of drops, but the issue still occurs on longer agent runs.
• Disabling reasoning entirely (either by removing --reasoning-parser qwen3 from SGLang, or by sending chat_template_kwargs: {"enable_thinking": false} in the request) makes the ECONNRESET go away — the stream completes cleanly every time.
  However, this is not a real solution for our use case: without reasoning enabled, the model does not reliably plan or chain tool calls anymore, i.e. it is no longer agentic. Build-mode sessions become single-shot replies instead of multi-step tool-using runs, which defeats the purpose of using opencode with a reasoning model. So reasoning has to stay on, and the crash has to be fixed on the streaming-client side.

Screenshot and/or share link

Operating System

Server (SGLang host): Ubuntu 22.04, 2× NVIDIA GPUs, SGLang in a venv, launched via systemd - Client (opencode host): Windows 11

Terminal

No response