Docs: DashScope thinking mode requires thinking_budget, docs only mention enable_thinking

Problem

When enabling thinking mode on DashScope for qwen3.5-27b, Qwen-Agent fails with a 400 error unless thinking_budget is also passed, but the official docs only mention enable_thinking.

Reproduction

from qwen_agent.agents import Assistant

llm_cfg = {
    "model": "qwen3.5-27b",
    "model_type": "qwenvl_dashscope",
    "api_key": "your-dashscope-api-key",
    "generate_cfg": {
        "enable_thinking": True,
    },
}

agent = Assistant(llm=llm_cfg, function_list=[], system_message="")

Error:

qwen_agent.llm.base.ModelServiceError:
Error code: InvalidParameter. Error message: <400> InternalError.Algo.InvalidParameter:
The thinking_budget parameter must be a positive integer and not greater than 0

The same 400 occurs via the DashScope OpenAI-compatible endpoint:

llm_cfg = {
    "model": "qwen3.5-27b",
    "model_type": "qwenvl_oai",
    "model_server": "https://dashscope.aliyuncs.com/compatible-mode/v1",
    "api_key": "your-dashscope-api-key",
    "generate_cfg": {
        "extra_body": {
            "enable_thinking": True,
        },
    },
}

Root cause

DashScope requires thinking_budget (a positive integer no larger than the model's max thinking length) whenever enable_thinking is true. When it is omitted, the gateway treats it as 0 and rejects the request.

Workaround

Pass thinking_budget explicitly. Verified working for qwen3.5-27b:

"generate_cfg": {
    "enable_thinking": True,
    "thinking_budget": 12000,
}

and for the OpenAI-compatible endpoint:

"generate_cfg": {
    "extra_body": {
        "enable_thinking": True,
        "thinking_budget": 12000,
    },
}

With the workaround, reasoning_content is returned correctly and tool calling still works.

Docs gap

• qwen-agent-docs/website/content/en/guide/get_started/configuration.md only documents enable_thinking for DashScope, DashScope OAI, and vLLM.
• The Chinese readthedocs page (QwenLM/Qwen3/docs/source/framework/qwen_agent.rst) has the same gap.

It would be helpful to add thinking_budget to the DashScope examples, and mention that it must be a positive integer not exceeding the model's max thinking length, so users do not hit this cryptic 400.

Environment

• qwen-agent: 0.0.34 (also checked main)
• DashScope model: qwen3.5-27b (hybrid thinking)