Random occasional stalls when using local vLLM provider

Description

I'm using local vLLM with Qwen3.8-27B. The only MCP I have configured is a local Context7, but I've seen this happen even before adding the MCP. I've seen this happen a couple times though I have not identified any specific pattern to suggest what may be the cause. The symptom is that in the middle of thinking, opencode would just stall. There is no indication in the UI that it hit some sort of problem. There is no error on vLLM side. In fact, vLLM is healthy since the local health check is working just fine. It would appear that opencode just stopped sending further requests to vLLM. The only way to notice this looking at the vLLM logs. If for an extended amount of time while opencode thinks it's still working no further requests were made, I'd have to cancel out, then tell it continue where it left off.

(APIServer pid=1) INFO 09-02 18:41:33 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 3.59, Accepted throughput: 21.50 tokens/s, Drafted throughput: 58.10 tokens/s, Accepted: 215 tokens, Drafted: 581 tokens, Per-position acceptance rate: 0.723, 0.602, 0.482, 0.337, 0.217, 0.133, 0.096, Avg Draft acceptance rate: 37.0%
(APIServer pid=1) INFO:     127.0.0.1:50018 - "GET /health HTTP/1.1" 200 OK
(APIServer pid=1) INFO 09-02 18:41:43 [loggers.py:310] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 9.4 tokens/s, Running: 0 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 96.5%
(APIServer pid=1) INFO 09-02 18:41:43 [metrics.py:120] SpecDecoding metrics: Mean acceptance length: 4.57, Accepted throughput: 7.50 tokens/s, Drafted throughput: 14.70 tokens/s, Accepted: 75 tokens, Drafted: 147 tokens, Per-position acceptance rate: 0.714, 0.714, 0.476, 0.429, 0.429, 0.429, 0.381, Avg Draft acceptance rate: 51.0%
(APIServer pid=1) INFO 09-02 18:41:53 [loggers.py:310] Engine 000: Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 0.0 tokens/s, Running: 0 reqs, Waiting: 0 reqs, GPU KV cache usage: 0.0%, Prefix cache hit rate: 96.5%
(APIServer pid=1) INFO:     127.0.0.1:40896 - "GET /health HTTP/1.1" 200 OK
(APIServer pid=1) INFO:     127.0.0.1:45562 - "GET /health HTTP/1.1" 200 OK
(APIServer pid=1) INFO:     127.0.0.1:51508 - "GET /health HTTP/1.1" 200 OK
(APIServer pid=1) INFO:     127.0.0.1:41964 - "GET /health HTTP/1.1" 200 OK

Plugins

No response

OpenCode version

1.18.26

Steps to reproduce

1. Connect to local vLLM provider
2. Prompt opencode to work on something
3. Tail vLLM logs and catch it when it stalls
4. Interrupt opencode
5. Prompt opencode to resume

Screenshot and/or share link

No response

Operating System

Windows 11, WSL2: Ubuntu 24.04 LTS

Terminal

VSCode Terminal