Description

When using free NVIDIA models (e.g. z-ai/glm-5.1, moonshotai/kimi-k2.6) for extended tasks (30+ minutes), the UI freezes and requires manual cancellation. There are two root causes:

1. Connection hang: NVIDIA accepts the TCP connection but never sends a response. After 10+ minutes the server drops the connection with ECONNRESET or HTTP 504. During this time the session appears frozen with no feedback.
2. EngineCore crash: NVIDIA's inference engine crashes mid-stream and emits { message: "EngineCore encountered an issue. See stack trace (above) for the root cause." } as a plain object inside the SSE stream. This error is not recognised as retryable so the session stops with a permanent error.

Neither case shows a retry — the user must manually cancel and re-prompt.

Plugins

None

OpenCode version

1.16.2

Steps to reproduce

1. Configure opencode with a free NVIDIA provider model (z-ai/glm-5.1 or moonshotai/kimi-k2.6)
2. Start a long-running agentic task (code generation, multi-step build, etc.)
3. Run for 30 minutes or more
4. Observe the UI freezes with no spinner or retry message
5. The only recovery is pressing the cancel keybind and re-sending the prompt manually

Screenshot and/or share link

No response

Operating System

macOS 25.5.0 (Darwin arm64)

Terminal

iTerm