Problem

When users execute /think off, they still receive reasoning_content from reasoning-capable models (e.g., GLM-4.7, GLM-4.6, Kimi K2.5, MiniMax-M2.1).

Expected behavior: /think off should completely disable reasoning content.
Actual behavior: Reasoning content is still returned.

Steps to Reproduce

1. Use a reasoning-capable model (e.g., GLM-4.7)
2. Execute /reasoning on
3. Execute /think off
4. Send a message (e.g., "1+1")
5. Observe that reasoning_content is still returned

Impact

• Affects all users using /think off with reasoning-capable models
• High priority - breaks expected command behavior