bug: Streaming markdown formatting breaks code fences and whitespace #421
Description
问题描述
When using streaming responses, markdown formatting can break in a few ways:
- fenced code blocks may lose their expected structure
- whitespace-only tokens are dropped, which removes meaningful spaces/newlines/indentation
- some plain text may get rendered as part of a code block
This is especially visible when asking the model to output multiple markdown code block examples.
复现步骤
A simple prompt that reproduces the issue:
‘‘‘最常见的ai回复的代码块有哪些，举例输出5个’’’
预期行为
- Python（最常见，AI 经常输出算法/脚本）
print("Hello, World!")- JavaScript（前端/Node.js 常用）
console.log("Hello, World!");- HTML（网页结构最常见）
<!DOCTYPE html>
<html>
<body>
<h1>Hello, World!</h1>
</body>
</html>- CSS（样式代码高频出现）
body {
color: blue;
font-family: Arial;
}- SQL（数据库查询最常见）
SELECT * FROM users WHERE id = 1;实际行为
- Python代码块```pythondef
greet(name):
return f"Hello, {name}!"
print(greet("AI"))
2. JavaScript代码块```javascriptconst
message = "Hello from JS!";
console.log(message);
- HTML代码块```html
AI回复示例
- CSS代码块```css.example
{ color: blue;
font-size:16px;
}
5. SQL查询代码块```sqlSELECT
* FROM usersWHERE created_at > '2025-01-01';
部署方式
Docker Compose
版本/提交号
相关日志/报错
The main issue appears to be that whitespace-only tokens are discarded in streaming mode:
<PYTHON>
if not in_think and not filtered.strip():
continue
This removes meaningful spaces/newlines/indentation and breaks markdown/code formatting.
There was also a related issue where tag filtering could remove surrounding text too aggressively.请求样例（可选）
Fix direction
The following changes fixed the issue in my local test:
Preserve whitespace-only tokens in stream output
Strip only the tag itself instead of dropping the whole token
Repair fragmented fenced code block starts in streaming mode