[Bug] <｜Assistant▁END▁OF▁TOOL_CALLS｜> and <｜Tool｜> tokens leak into user-visible streaming output

[NunoLemos101]

📦 部署方式

Docker

📌 软件版本

4.4.3

💻 系统环境

macOS

📌 系统版本

No response

🐛 问题描述

Sometimes when doing tool calls, the LLM starts streaming the tool result duplicated.

I put Claude Code investigating this issue and it says:

  Root Cause                                                                                                                              
                                                                  
  The regex on line 24 of leaked_output_sanitize.go:                                                                                      
                                                                  
  var leakedMetaMarkerPattern = regexp.MustCompile(`(?i)<[｜\|]\s*(?:assistant|tool|end[_▁]of[_▁]sentence|end[_▁]of[_▁]thinking|end[_▁]of[
  _▁]toolresults|end[_▁]of[_▁]instructions)\s*[｜\|]>`)                                                                                   
   
  Tokens NOT matched: 
  
  <｜Tool｜>[...]: The regex matches bare tool between delimiters, but the actual leaked token includes a JSON payload after closing | >, and the content between <｜Tool｜> and <｜end▁of▁toolresults｜> is not stripped

  <｜Assistant▁END▁OF▁TOOL_CALLS｜>: Compound token not in the alternation group — only simple assistant is matched

📷 复现步骤

No response

🚦 期望结果

I manually edited the HTML to show how expected behavior should look:

📝 补充信息

No response