Bug: openai-compatible deepseek/glm/minimax models lost reasoning variants after qwen exclusion

[elonazoulay]

Description

ProviderTransform.variants() should return the standard low / medium / high reasoning variants for reasoning-capable deepseek, glm, and minimax models when they are routed through @ai-sdk/openai-compatible.

On current dev, those model families are being excluded by the same blanket guard that also excludes qwen, kimi, and related models, so they incorrectly return {} instead.

This issue is intentionally narrow: qwen, kimi, and k2p5 should remain excluded. This is a follow-up to PR #21212, not a blanket revert.

Plugins

OpenCode version

1.4.11

Steps to reproduce

1. Inspect packages/opencode/src/provider/transform.ts on current dev.
2. Call ProviderTransform.variants() with a reasoning-capable deepseek, glm, or minimax model using @ai-sdk/openai-compatible.
3. Observe that it returns {} instead of low / medium / high.

Screenshot and/or share link

Operating System

macOS

Terminal

zsh

[github-actions]

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

• GLM models excluded from reasoning variants — no way to toggle thinking at runtime #18598: GLM models excluded from reasoning variants — this issue covers the same ProviderTransform.variants() exclusion for GLM models that you describe, and may overlap significantly with the GLM portion of this report.
• Bug: capabilities overridden for @ai-sdk/openai-compatible providers #22036: capabilities overridden for @ai-sdk/openai-compatible providers — touches related behavior where model capabilities (including reasoning) are not correctly applied for openai-compatible providers like minimax.

Please review the above to see if your report is already covered or if this is a distinct enough issue to track separately.