Telegram Rich Messages (Bot API 10.1+) — edit_message stream destroys rich formatting

[vyunolbek]

Description

Telegram Bot API 10.1 introduced sendRichMessage which supports tables, headings, code blocks, collapsible details, checklists, and deeply nested formatting. The gateway already has _send_rich_text() which calls this API for first-sends — but during streaming, intermediate editMessageText calls destroy the rich formatting because:

Root cause

1. First chunk → adapter.send() → _send_rich_text() → sendRichMessage ✅ (rich, but only ~18 chars — just the cursor + prefix)
2. Intermediate chunks → edit_message(finalize=False) → editMessageText without parse mode → replaces the entire message with plain text ❌ (destroys all ## headings, | tables, etc.)
3. Final chunk → edit_message(finalize=True) → format_message() → MarkdownV2 — but format_message() calls _wrap_markdown_tables() which converts GFM tables to text row groups, and ## Title is downgraded to *Title* (MarkdownV2 bold)

Result: Users see raw ##, |, ** markers instead of rendered rich formatting.

Steps to Reproduce

1. Enable Telegram gateway
2. Send any message with rich Markdown (## heading, | table |, `code`)
3. Observe message appears as plain text with visible markers

Expected Behavior

• Tables render natively
• ## Heading renders as actual section heading
• Code blocks render with monospace and syntax highlighting
• All Rich Message features survive streaming

Actual Behavior

Everything rendered as plain text with raw syntax markers visible.

Proposed Fix

In gateway/platforms/telegram.py, edit_message() method, finalize=True path.

There is no editRichMessage API in Bot API 10.1, so the fix sends the final content as a new Rich Message via _send_rich_text() and deletes the old plain-text preview:

# Try Rich Message finalize first
rich_result = await self._send_rich_text(chat_id=chat_id, content=content)
# Clean up the old plain-text preview
try:
    await self._bot.delete_message(chat_id=chat_id, message_id=old_id)
except Exception:
    pass
return SendResult(success=True, message_id=str(rich_result.message_id))

If _send_rich_text() fails, it falls back to the existing MarkdownV2 edit path.

Additional Changes

• tools/send_message_tool.py: Added rich_message parameter ({"html": "..."} or {"markdown": "..."}) to send_message tool schema — lets the agent explicitly send Rich Messages via sendRichMessage API
• gateway/platforms/telegram.py: _send_rich_text() and send_draft(sendRichMessageDraft) already exist but edit_message() was bypassing them

Testing

Tested locally on a self-hosted Hermes instance with Bot API 10.1:

• ✅ Tables render correctly
• ✅ Headings (## H2, ### H3) render as proper headings
• ✅ Code blocks with syntax highlighting
• ✅ Bold, italic, strikethrough, spoilers
• ✅ Collapsible <details> blocks
• ✅ Fallback to MarkdownV2 works when sendRichMessage is unavailable

Environment

• Hermes Agent: self-hosted (latest from main)
• Bot API version: 10.1+
• python-telegram-bot: 22.6

[teknium1]

Fixed on main in #46890 (a6364bf).

Streamed Telegram replies that contain a table (or task list / <details> / block math) now finalize by editing the existing preview in place via Bot API 10.1's editMessageText rich_message parameter. Previously the finalize edit ran through MarkdownV2 — which has no table syntax — and rewrote pipe tables into bullet lists, producing the "table while streaming, list at the last moment" behavior.

No fresh send + delete, so there is no duplicate-preview flicker. Rich messages are enabled by default; set gateway.platforms.telegram.extra.rich_messages: false to force the legacy MarkdownV2 path. Plain replies stay on MarkdownV2 for consistent rendering. Restart the gateway after updating to pick up the change.