openclaw - 💡(How to fix) Fix bug(models): Google provider fails to parse <think>/<final> tags, leaking reasoning and truncating replies

Description

When using Google Gemini models (such as gemini-3.1-pro-preview or gemini-3.5-flash), OpenClaw's output parser fails to correctly parse and strip the <think>...</think> and <final>...</final> tag structure.

This results in severe UX issues where internal reasoning is leaked, and the actual final response is often truncated or completely hidden.

Symptoms

1. Reasoning Leak: The content within <think>...</think> is printed directly as visible assistant message text in WebChat, instead of being flagged as isReasoning: true or stripped.
2. Tag Leak: The closing </final> tag is printed at the end of the assistant's message.
3. Severe Truncation: Because the model outputs a long thinking process inside <think>, the total message length easily hits WebChat's size bounds (e.g., chatHistoryMaxChars limit), causing the message to be truncated before the <final> block is reached. The user never gets to see the actual answer!

Cause

While OpenClaw correctly flags and hides native reasoning for providers like DeepSeek, or applies robust tag-stripping for some models, the Google Gemini provider adapter does not seem to recognize or split the custom <think> and <final> XML-like blocks from the stream.

As a result, the entire raw output (including <think> and the <final> tags) is treated as the visible assistant text.

Suggested Fix

Update the Google/Gemini provider parser (or the global stream parser) to:

1. Detect <think>...</think> blocks in the incoming stream from Gemini models and correctly map them to the isReasoning: true metadata (excluding them from the visible message content).
2. Strip <final> and </final> tags from the final visible WebChat display payload so they do not leak into the UI.