server: respect per-request enable_thinking toggle via extra_body

[pju-hoge]

Overview

Fixes enable_thinking being ignored when set per-request via extra_body or chat_template_kwargs. Previously the server only checked chat_template_kwargs, but OpenAI-compatible clients (and tools like Opencode) send it in extra_body. This caused enable_thinking: false to be silently ignored across all shells.

Changes

tools/server/server-common.cpp

Read enable_thinking from extra_body first (OpenAI-compatible path), then fall back to chat_template_kwargs. Pass the value to the slot so it can override the server default.

common/chat.cpp

supports_thinking now requires both template-level reasoning detection AND enable_thinking == true. Prevents thinking tags from being injected when the user explicitly toggles thinking off.

common/chat-auto-parser-generator.cpp

extract_reasoning also gated by enable_thinking, ensuring the PEG parser does not extract reasoning block markers when thinking is disabled.

Fixes

• Misc. bug: enable_thinking param cannot turn off thinking for qwen3.5 #20182 — enable_thinking param cannot turn off thinking for qwen3.5
• disabling reasoning does not work anymore on certain models #20196 — disabling reasoning does not work anymore on certain models
• Eval bug: Bug: Qwen3.5 enable_thinking=false via --chat-template-kwargs is ignored across all shells (PowerShell/Bash) #20409 — enable_thinking=false ignored across all shells
• Misc. bug: Qwen3.5 with Opencode: Assistant response prefill is incompatible with enable_thinking. #20861 — Qwen3.5 with Opencode assistant prefill incompatible

Testing

• Build verified (Release, Linux, cmake --build . -j$(nproc))
• Existing test-chat and test-chat-peg-parser pass
• No new server integration tests added (change is minimal and targeted)

AI Usage Disclosure

YES — AI (OpenCode with Qwen3.6-35B) assisted with code formatting, searching related issues on GitHub, verifying CI compliance against .github/workflows/server.yml, and drafting this PR description. All code changes, logic design, and review were performed by the contributor.

[ggml-gh-bot]

Hi @pju-hoge, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• AI-generated content: This project does not accept PRs, descriptions or commit messages that are fully or predominantly AI-generated. If you have used AI to assist you in writing code, please make sure to disclose that explicitly.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.

[vanmilleru]

#22323
#22162

[isaac-mcfadyen]

As far as I'm aware, extra_body on the OpenAI clients literally adds the fields to the body. There is not a dedicated field called extra_body actually sent in the request. See:

https://github.com/openai/openai-python/blob/e507a4ebeea4c3f93cd48986014a3e2ca79230c2/src/openai/_base_client.py#L2007-L2045

https://github.com/openai/openai-python/blob/e507a4ebeea4c3f93cd48986014a3e2ca79230c2/src/openai/_base_client.py#L502-L509

https://github.com/openai/openai-python/blob/e507a4ebeea4c3f93cd48986014a3e2ca79230c2/src/openai/_base_client.py#L2183-L2192

Also, the general way to "disable" thinking with reasoning models is to add empty <think></think> tags. I suspect that your change in chat.cpp will not add these tags because the model is no longer marked as supporting thinking (and will severely degrade performance as a result).