Description

When using Claude Opus 4.5 with <thinking_mode>interleaved</thinking_mode> in the system prompt, the thinking/reasoning content is being sent to the channel (Feishu in my case) even when Think: off is set in the session.

Expected Behavior

When Think: off is configured, the gateway should filter out all thinking/reasoning content before sending to the channel. Only the final response should be delivered.

Actual Behavior

Internal thinking content like:

• "让我想想..."
• "用户问的是..."
• "已发送的有 5 条了。让我选一条新的..."

...is being sent to the user along with the actual response.

Environment

• OpenClaw version: 2026.2.19-2 (45d9b20)
• Model: anthropic/claude-opus-4-5-20251101
• Channel: Feishu
• Session status shows: Think: off

Steps to Reproduce

1. Use Claude Opus 4.5 with extended thinking enabled
2. Set Think: off in session
3. Send a message that triggers internal reasoning
4. Observe that thinking content appears in channel output

Possible Cause

The filtering logic for <thinking> blocks may not be applied when the model uses interleaved thinking mode, or the filter runs before the full response is assembled.

Workaround

Currently manually avoiding writing thinking-style content in responses, but this is not ideal.