Name and Version

llama-server --version
version: 9211 (5bf468a2f)
built with GNU 13.3.0 for Linux x86_64

Operating systems

Linux

GGML backends

CUDA

Hardware

NVIDIA GeForce RTX 3090 (4x)

Models

unsloth/Qwen3.6-27B-GGUF

Problem description & steps to reproduce

When using reasoning_format: "none" in streaming mode, the <think> opening tag is stripped from the content deltas in builds ≥b9211. The thinking content flows through content deltas without the opening tag, making it impossible for clients to detect where reasoning content begins.

Expected behavior (build b9191 and earlier):
First content delta with reasoning_format: "none":

{"content": "</think>

Here's a thinking process:

Actual behavior (build b9211 and later):
First content delta:

{"content": "Here's a thinking process:

The </think> tag is stripped but the reasoning content is NOT sent as separate reasoning_content deltas - it's all in content deltas without the tag.

Non-streaming mode works correctly - the full response includes the </think> tag as expected.

Steps to reproduce

1. Start llama-server with a Qwen3.6 model
2. Send streaming request with reasoning_format: "none":

curl -s -X POST http://localhost:8082/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "unsloth/Qwen3.6-27B-GGUF",
    "messages": [{"role": "user", "content": "who are you"}],
    "stream": true,
    "reasoning_format": "none",
    "max_tokens": 30
  }'

3. Observe that the first content delta does NOT include the </think> tag

First Bad Commit

Build b9227 (autoparser PR) introduced this behavior. Build b9191 works correctly.

Relevant log output

N/A - this is observable in the SSE delta output