hermes - 💡(How to fix) Fix Feature Request: Strip reasoning/<think> blocks from TTS preprocessing

Problem

When /reasoning show is enabled, the model's reasoning/thinking content is included in the final assistant message. TTS (text-to-speech) receives this full text and reads the reasoning aloud, which is undesirable — users want to see reasoning but not hear it spoken.

Current behavior

Root Cause

Neither of the two TTS text-preparation functions strips reasoning/<think> blocks:

1. prepare_tts_text (gateway/platforms/base.py:2264) — strips markdown formatting, truncates to 4000 chars. No reasoning filter.
2. _strip_markdown_for_tts (tools/tts_tool.py:2252) — strips code blocks, links, bold, italic, headers, lists. No reasoning filter.

The <think> regex does exist in the codebase at tools/tts_tool.py:2340, but it is only used in stream_tts_to_speaker() (the ElevenLabs streaming path) — not in the general TTS pipeline.

Proposed Solution

Add <think>...</think> block stripping to _strip_markdown_for_tts() so it applies to all TTS providers uniformly:

# tools/tts_tool.py
_THINK_BLOCK = re.compile(r"<think[\s>].*?</think>", flags=re.DOTALL)

def _strip_markdown_for_tts(text: str) -> str:
    text = _THINK_BLOCK.sub("", text)  # strip reasoning before anything else
    # ... existing markdown stripping ...

This would:

• Apply to all TTS providers (Edge, OpenAI, ElevenLabs, Mistral, etc.)
• Not affect the displayed text (reasoning still visible)
• Only affect the audio output

Environment

• Hermes Agent (git-installed, up to date)
• TTS provider: Edge TTS (zh-CN-XiaoxiaoNeural)
• Platform: Telegram

Additional Context

The same issue applies to non-<think> reasoning formats (e.g., DeepSeek's reasoning_content field, MiniMax thinking blocks). A more comprehensive solution might check for reasoning_content in the API response and exclude it from the TTS feed, but stripping <think> blocks in TTS preprocessing is the simplest first step that covers most cases.