Fix Qwen3 content extraction breaking code formatting

[iSevenDays]

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

Summary

Fixes a bug in Qwen3 content extraction where aggressive whitespace cleanup was breaking proper code formatting in non-tool-call streaming output.

Problem

The qwen3::extract_content_during_parsing() function was using an aggressive regex pattern that collapsed multiple newlines:

content = std::regex_replace(content, std::regex(R"(\n\s*\n)"), "\n");

This broke:

• PEP 8 compliance: 2 empty lines between Python functions were collapsed to 1
• Code formatting: Any intentional whitespace structure was lost
• Streaming output: Users requesting properly formatted code received malformed results

Root Cause

The regex R"(\n\s*\n)" replaces any sequence of newline + whitespace + newline with a single newline, which is overly aggressive for content that should preserve formatting.

Solution

✅ Replace aggressive cleanup with gentle trimming

• Remove std::regex_replace(content, std::regex(R"(\n\s*\n)"), "\n")
• Replace with string_strip(content) following original llama.cpp patterns
• Only trim leading/trailing whitespace, preserve internal formatting

✅ Add proper include

• Add #include "../../common/common.h" for string_strip access

✅ Add regression test

• New test_qwen3_whitespace_preservation() validates PEP 8 compliance
• Ensures 2 empty lines between functions are preserved

Testing

• ✅ PEP 8 compliance: 2 empty lines between functions preserved
• ✅ Tool call parsing: All Qwen3 tests continue to pass
• ✅ No regressions: Existing functionality maintained
• ✅ Follows patterns: Matches original llama.cpp whitespace handling

Example

Before (broken):

def celsius_to_fahrenheit(celsius):
    return celsius * 9/5 + 32
def fahrenheit_to_celsius(fahrenheit):  # Missing empty lines\!
    return (fahrenheit - 32) * 5/9

After (fixed):

def celsius_to_fahrenheit(celsius):
    return celsius * 9/5 + 32


def fahrenheit_to_celsius(fahrenheit):  # Proper PEP 8 spacing preserved\!
    return (fahrenheit - 32) * 5/9

Files Changed

• examples/server/parsers/qwen3_parser.hpp - Fix whitespace handling
• tests/test-function-calls.cpp - Add regression test

Test Results

🧹 Testing Qwen3 Whitespace Preservation Fix:
✅ PASS: Qwen3: PEP 8 double empty lines preserved
✅ PASS: Qwen3: Content not empty after processing
✅ PASS: Qwen3: Function content preserved
✅ PASS: Qwen3: Second function preserved

[gopinath87607]

hi i tested this pr in vs code continue extension with recent unsolth quant qwen 30b a3b coder but its still breaking when its calling second time tool and main repo not even working

update

looks like model itself there is a error let me check again

[gopinath87607]

https://huggingface.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/discussions/1

kindly check this discussion

update this pull
ggml-org/llama.cpp#14962

[iSevenDays]

@gopinath87607 here is my latest branch updates
#670
I found some issues and differences between llama.cpp and ik_llama.cpp and fixed them. That's probably not everything, but works much stable.

[Danmoreng]

Apparently qwen3 coder uses a different tool calling structure with XML instead of json, llama.cpp currently ads the functionality here: ggml-org/llama.cpp#15012

[gopinath87607]

@gopinath87607 here is my latest branch updates #670 I found some issues and differences between llama.cpp and ik_llama.cpp and fixed them. That's probably not everything, but works much stable.

thanks mate i will check it do we need to use any special tag for tool activation like jinja

[ubergarm]

Some more discussion on similar qwen with toolcalling stuff here: https://huggingface.co/bartowski/Qwen_Qwen3-235B-A22B-Instruct-2507-GGUF/discussions/1