Qwen3.6 tool calls loop with empty arguments: qwen-chat-template missing preserve_thinking

[0xbentang]

What happened?

Qwen3.6 loops emitting tool calls with empty {} arguments on multi-turn conversations via LM Studio + thinkingFormat: "qwen-chat-template". After 2–3 turns the model reasons correctly ("I need to pass title") but every call arrives with empty args.

Cause: Qwen3.6's chat template drops prior-turn <think> blocks unless preserve_thinking=true is passed in chat_template_kwargs. pi only sets enable_thinking at packages/ai/src/providers/openai-completions.ts:424. Background: https://www.reddit.com/r/LocalLLaMA/comments/1sne4gh/psa_qwen36_ships_with_preserve_thinking_make_sure

This makes local Qwen3.6 GGUF unusable for agent loops via qwen-chat-template.

Steps to reproduce

1. Load unsloth/Qwen3.6-35B-A3B-GGUF in LM Studio.
2. Register it in models.json with "thinkingFormat": "qwen-chat-template" and "reasoning": true.
3. Run any task that calls the same tool across ≥3 turns. After turn 2 every tool call has arguments: {}.

Expected behavior

Prior thinking preserved across turns. One-line fix:

(params as any).chat_template_kwargs = {
    enable_thinking: !!options?.reasoningEffort,
    preserve_thinking: true,
};

Version

0.67.6

[github-actions]

This issue was auto-closed. All issues from new contributors are auto-closed by default.

Maintainers review auto-closed issues daily and reopen worthwhile ones. Issues that do not meet the quality bar in CONTRIBUTING.md will not be reopened or receive a reply.

If a maintainer replies lgtmi on one of your issues, your future issues will stay open. If a maintainer replies lgtm, your future issues and PRs will stay open.

See CONTRIBUTING.md.

[badlogic]

Implemented.

qwen-chat-template now sends chat_template_kwargs.preserve_thinking: true alongside enable_thinking, so prior Qwen thinking is preserved across turns and multi-turn tool calls no longer degrade to empty {} arguments.

npm run check passed.

[montvid]

Starting only from qwen3.6 qwen supports preserve_thinking. All other qwen models do not - this could be not a good fix. Anyway, these 2 settings can be changed server side.