[Bug] Go models inference slow/stuck on Windows (Desktop + TUI) with existing sessions after v1.17.14

[Kuromi-zyzy]

Description

Description: After updating to v1.17.14 on Windows (both Desktop app and native TUI), all Go subscription models (DeepSeek V4 Flash/Pro, etc.) respond extremely slowly or appear stuck at "thinking" when used in existing sessions. Starting a new session resolves it. On WSL TUI with the same v1.17.14, everything works fine.

Isolated testing confirms:

Not a network issue — same behavior with/without system proxy
DeepSeek direct API works normally on Windows
Zen free tier and other providers (Alibaba Bailian) unaffected
Only Go models are affected
Same Go models work fine via WSL TUI on the same machine
OpenCode version: v1.17.14

Steps to reproduce:

Use OpenCode on Windows (Desktop or TUI) with an existing session, select any Go model (e.g., opencode-go/deepseek-v4-flash)
Send a prompt — response is very slow or stuck at "thinking"
Start a brand new session with the same model — response is normal speed
Run the same Go model via WSL TUI (wsl opencode) on the same machine — also normal
Operating System: Windows 11

Terminal: Windows Terminal (TUI + Desktop)

Plugins

No response

OpenCode version

v1.17.14

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Currently experiencing halted and slow inference in OpenCode Go #34667: Halted and slow inference in OpenCode Go (DeepSeek/Mimo on Windows, thinking blocks stopping halfway)
• DeepSeek Thinking takes an extremely long time to generate responses #34836: DeepSeek Thinking takes an extremely long time to generate responses (Go plan, Windows 11, v1.17.13)

If your issue is different or version-specific, please add details distinguishing it from the above.

[Kuromi-zyzy]

It might not be a version-related issue. I just rolled back the TUI version on Windows, yet the problem still persists. What puzzles me, though, is why the TUI inside WSL runs so smoothly.