Summary
When an agent reply contains bold written as **标题：**内容 (full-width CJK punctuation immediately inside the closing **, followed by more text), the Telegram channel delivers the message with literal ** instead of bold. Other bold spans in the same message render fine, so messages arrive half-formatted.
Environment
- OpenClaw 2026.6.10 (aa69b12), also reproduced against the published dist of 2026.6.11 and 2026.7.1-beta.2 (same
markdown-it bundle)
- Telegram channel,
richMessages: true, local telegram-bot-api server
Reproduction
Input markdown (very common pattern for Chinese LLM output — inline bold labels):
边界：**社区显示：**Fable 与 **有效**。
Actual HTML produced by the Telegram rich-message renderer (captured on the wire to the bot API):
边界：**社区显示：**Fable 与 <b>有效</b>。
The first pair of ** is left literal; the second converts correctly.
Root cause
The markdown→HTML conversion uses markdown-it with strict CommonMark emphasis rules. Per the CommonMark flanking rules, a closing ** that is preceded by punctuation (full-width "：", "。", "、", "）" etc. are Unicode punctuation) must be followed by whitespace or punctuation to be right-flanking. In CJK text there are no inter-word spaces, so the extremely common pattern **标签：**正文 never qualifies and the delimiters stay literal.
This is the well-known CommonMark-vs-CJK emphasis problem. Chinese/Japanese/Korean model output hits it constantly because LLMs love the **label:** text idiom.
Suggested fix
Wire markdown-it-cjk-friendly (implements the CJK-friendly flanking amendment proposed for CommonMark) into the markdown-it instance used for outbound channel formatting — or apply an equivalent pre-pass before emphasis parsing. This would fix bold/italic rendering for CJK users across all channels that go through the shared renderer, not just Telegram.
Happy to provide more captures if useful.
Summary
When an agent reply contains bold written as
**标题：**内容(full-width CJK punctuation immediately inside the closing**, followed by more text), the Telegram channel delivers the message with literal**instead of bold. Other bold spans in the same message render fine, so messages arrive half-formatted.Environment
markdown-itbundle)richMessages: true, local telegram-bot-api serverReproduction
Input markdown (very common pattern for Chinese LLM output — inline bold labels):
Actual HTML produced by the Telegram rich-message renderer (captured on the wire to the bot API):
The first pair of
**is left literal; the second converts correctly.Root cause
The markdown→HTML conversion uses
markdown-itwith strict CommonMark emphasis rules. Per the CommonMark flanking rules, a closing**that is preceded by punctuation (full-width "：", "。", "、", "）" etc. are Unicode punctuation) must be followed by whitespace or punctuation to be right-flanking. In CJK text there are no inter-word spaces, so the extremely common pattern**标签：**正文never qualifies and the delimiters stay literal.This is the well-known CommonMark-vs-CJK emphasis problem. Chinese/Japanese/Korean model output hits it constantly because LLMs love the
**label:** textidiom.Suggested fix
Wire
markdown-it-cjk-friendly(implements the CJK-friendly flanking amendment proposed for CommonMark) into the markdown-it instance used for outbound channel formatting — or apply an equivalent pre-pass before emphasis parsing. This would fix bold/italic rendering for CJK users across all channels that go through the shared renderer, not just Telegram.Happy to provide more captures if useful.