Bug Report: stripMarkdown() function incorrectly deletes underscores, breaking programming language identifiers

Description

The stripMarkdown() function in OpenClaw's core text runtime contains a regex bug that incorrectly removes underscores from text. This breaks all programming language code (C, C++, Python, etc.) sent through channels that call this function, such as Weixin channel.

Expected Behavior:
Underscores in text should only be removed when they are part of valid Markdown italic syntax (_text_). Programming language identifiers should be preserved.

Actual Behavior:
All underscores are removed from text, destroying programming language code:

• some_variable_name → somevariablename
• RATE_LIMITER_H → RATELIMITERH
• client_id → clientid

Affected Components

• File: /home/user/.npm-global/lib/node_modules/openclaw/dist/text-runtime-B-kOpuLv.js
• Function: stripMarkdown(text)
• Channels Affected: Any channel that uses stripMarkdown() to convert Markdown to plain text (confirmed: openclaw-weixin)

Note: Telegram and WhatsApp channels are not affected because they use markdownToTelegramHtml() / markdownToWhatsApp() instead of stripMarkdown().

Root Cause

Problematic regex in stripMarkdown():

result = result.replace(/(?<!_)_(?!_)(.+?)(?<!_)_(?!_)/g, "$1");

Analysis:

• This regex attempts to remove Markdown italic syntax: _text_
• The non-greedy pattern (.+?) combined with lookahead/lookbehind assertions causes unexpected backtracking
• For strings with multiple underscores, it matches first _ and last _ (not intended adjacent pair)
• Global flag g causes repeated replacement, eventually deleting ALL underscores

Test Results:

Note: Short identifiers are sometimes preserved due to boundary conditions, but longer identifiers are consistently broken.

Reproduction Steps

1. Send a message containing code with underscores through Weixin channel
2. Example C code:

   RateLimiter* limiter_create(int max_requests, int window_ms);

3. Received message in Weixin shows:

   RateLimiter* limitercreate(int maxrequests, int window_ms);

4. All underscores are deleted, making code invalid

Or reproduce programmatically:

function stripMarkdown(text) {
  let result = text;
  result = result.replace(/\*\*(.+?)\*\*/g, "$1");
  result = result.replace(/__(.+?)__/g, "$1");
  result = result.replace(/(?<!\*)\*(?!\*)(.+?)(?<!\*)\*(?!\*)/g, "$1");
  result = result.replace(/(?<!_)_(?!_)(.+?)(?<!_)_(?!_)/g, "$1");  // BUG
  result = result.replace(/~~(.+?)~~/g, "$1");
  result = result.replace(/^#{1,6}\s+(.+)$/gm, "$1");
  result = result.replace(/^>\s?(.*)$/gm, "$1");
  result = result.replace(/^[-*_]{3,}$/gm, "");
  result = result.replace(/`([^`]+)`/g, "$1");
  result = result.replace(/\n{3,}/g, "\n\n");
  return result.trim();
}

console.log(stripMarkdown("RATE_LIMITER_H"));  // Output: "RATELIMITERH"

Impact

• Code sharing impossible: Cannot share C/C++/Python code via Weixin channel
• Technical discussions broken: Variable names, function names, constants are corrupted
• Severity: High - Complete failure for programming language use cases

Suggested Fix

Option 1: Improve regex with word boundaries

// Only remove Markdown italic with word boundaries
result = result.replace(/\b_([a-zA-Z]+)_\b/g, "$1");

Option 2: Protect code blocks

// Don't process content inside code blocks
const codeBlocks = [];
result = result.replace(/```[\s\S]*?```/g, (match) => {
  codeBlocks.push(match);
  return `__CODE_BLOCK_${codeBlocks.length - 1}__`;
});
// ... process markdown ...
result = result.replace(/__CODE_BLOCK_(\d+)__/g, (_, idx) => codeBlocks[idx]);

Option 3: Use a proper Markdown parser

Replace regex-based approach with a proper Markdown AST parser that can distinguish between Markdown syntax and programming code.

Environment

OpenClaw: 2026.3.24 (cff6dc9)
Node.js: v22.22.2
OS: Oracle Linux Server 10.0
Plugin: @tencent-weixin/openclaw-weixin v2.1.1

Related Code

Weixin channel usage:

// /home/user/.openclaw/extensions/openclaw-weixin/src/messaging/send.ts
import { stripMarkdown } from "openclaw/plugin-sdk/text-runtime";

export function markdownToPlainText(text: string): string {
  let result = text;
  result = result.replace(/```[^\n]*\n?([\s\S]*?)```/g, (_, code: string) => code.trim());
  result = result.replace(/!\[[^\]]*\]\([^)]*\)/g, "");
  result = result.replace(/\[([^\]]+)\]\([^)]*\)/g, "$1");
  result = result.replace(/^\|[\s:|-]+\|$/gm, "");
  result = result.replace(/^\|(.+)\|$/gm, (_, inner: string) =>
    inner.split("|").map((cell) => cell.trim()).join("  "),
  );
  result = stripMarkdown(result);  // ❌ Problematic call
  return result;
}

Telegram channel (NOT affected):

// Telegram uses markdownToTelegramHtml() instead of stripMarkdown()
function renderTelegramHtmlText(text, options = {}) {
  if ((options.textMode ?? "markdown") === "html") return text;
  return markdownToTelegramHtml(text, { tableMode: options.tableMode });
}

Additional Context

• Issue discovered when sharing C language code via Weixin
• Telegram and WhatsApp channels work correctly (they preserve Markdown instead of stripping it)
• This is a core function that may affect other channels that use stripMarkdown()