Summary

When using Gemini models (e.g., gemini-3-pro-high), the <think>...</think> and <final>...</final> tags are sent directly to the user instead of being filtered out. Claude models do not have this issue.

Expected Behavior

Only the content inside <final> should be visible to the user. The <think> block should be hidden/stripped by the Gateway before sending to the channel.

Actual Behavior

The entire response including <think> and <final> tags is sent to the user via Telegram.

Environment

• OpenClaw version: 2026.1.30
• Channel: Telegram
• Model: google-antigravity/gemini-3-pro-high
• Provider: google-antigravity (OAuth)

Additional Context

Claude models (e.g., claude-opus-4-5-thinking) correctly hide the <think> block and only output <final> content. This suggests the Gateway may need model-specific post-processing for Gemini.