Intermittent 60s+ delay with "timeout" status on local Ollama models (direct Ollama: 1.3s) #52431

Description

Environment:
OpenCode 1.18.30, opencode serve on MacBook Pro M2 Max (32GB)
Ollama with qwen3-coder:30b (21GB, 100% GPU, loaded with "Forever" keepalive)
Client via tunnel proxy to opencode serve --port 4096

Symptom:
Simple prompt ("What is 2+2? Reply with just the number.") via OpenCode to local Ollama model:
Takes 67-70 seconds
Client reports "status": "timeout"
But the answer is correct in transcript_tail (model did respond)

Same prompt:
Direct via ollama run qwen3-coder:30b: 1.354 seconds
Via OpenCode to cloud model (Zen/muse-spark-1.3-contributor-free): 11 seconds, clean "status": "idle"
What we ruled out:
Not model-specific (both qwen3-coder:30b and qwen3.6:35b-a3b affected)
Not network (cloud models work fine via same path)
Not system resources (CPU 85% idle, memory 12GB/32GB, GPU 100%)
Not server logs (clean, no errors)
Not plugins (tested with Superpowers and Ponytail removed individually and together - no change)

Plugin context:
Issue occurred with two project-scoped plugins installed:
Superpowers: ctx.session.hook("context") on every session
Ponytail: experimental.chat.system.transform on every message
Removed each individually - no change. Removed both - still slow initially. Cleared ~/.cache/opencode (5MB models.json) + restarted Ollama + restarted serve - issue became intermittent (one fast at 7s, one slow at 67s, then consistently fast at 6-17s).

Current status: Working (6-17s, idle), but root cause unknown.

Question: What in OpenCode's Ollama provider could add 60+ seconds intermittently, with the client timing out but the response eventually completing?

Bug #46506 is related

Plugins

No response

OpenCode version

No response

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response