Why do I often face "Streaming response failed" when using free Nemotron 3 Ultra in OpenCode Zen?

Description

What happened?
When using the free Nemotron 3 Ultra model provided by OpenCode Zen (via the built-in model list, NOT configured with my own NVIDIA API key), the streaming response frequently fails in the middle of a task. This happens almost every time, especially when the task involves multiple tool calls or generates longer outputs.

The error message appears as:

"Streaming response failed"

After the failure, I have to manually type "continue" or "继续" to resume, but sometimes the model repeats already-generated content, which wastes tokens and time.

What did you expect?
I expected the streaming response to complete without interruption, or at least have automatic retry/resume mechanism built into the free tier.

Steps to reproduce

1. Open OpenCode (using the desktop app, not CLI)
2. Select the free Nemotron 3 Ultra model from the OpenCode Zen model list
3. Give it a moderately complex task, e.g.:
   • "Analyze the structure of my project and suggest improvements"
   • Or any task that requires multiple file reads or code generation
4. Wait for the response — at some point, Streaming response failed appears
5. The session is interrupted; I have to manually type "continue" to resume

Plugins I'm using

• opencode-auto-resume (installed via npm, but it doesn't seem to prevent or recover from this specific failure)
• No other custom plugins

OpenCode version

• The latest stable desktop version (downloaded from opencode.ai)

Additional context

• I do NOT have my own NVIDIA API key configured — I'm relying purely on the free OpenCode Zen offering.
• I've seen discussions in the community suggesting that the thinking parameter might cause compatibility issues with NVIDIA's API. However, since I'm using the OpenCode Zen gateway, I don't have access to modify the request parameters directly.
• The frequent interruptions make the model almost unusable for production or serious development work.

Suggestion
Could the OpenCode Zen gateway be updated to:

1. Handle the thinking parameter compatibility issue with NVIDIA models, OR
2. Implement automatic retry/reconnection logic for streaming failures on free models?

Thank you for your great work on OpenCode! 🙏

Plugins

opencode-auto-resume

OpenCode version

v1.18.3

Steps to reproduce

No response

Screenshot and/or share link

Operating System

macOS 26.5.2

Terminal

No response