What happened?

When using Qwen Code with glm-5.2 through the OpenAI-compatible provider, thinking text is sometimes emitted as normal assistant output. The visible response includes internal reasoning followed by a stray </think> closing tag, for example:

The user is asking a simple logic puzzle. ... The answer is straightforward.</think>LEAK_CHECK_FINAL: 奖品在 B

This is easy to reproduce on Qwen Code 0.19.3 with the globally installed CLI.

Reproduction command:

qwen "请回答这个测试问题：三个盒子 A、B、C，只有一个有奖品。线索：A 是空的；B 有奖品；C 是空的。最终答案只输出一行：LEAK_CHECK_FINAL: 奖品在 B。" \
  --auth-type openai \
  --model glm-5.2 \
  --approval-mode yolo \
  --output-format stream-json

Observed structured output:

{
  "type": "assistant",
  "message": {
    "content": [
      {
        "type": "text",
        "text": "The user is asking a simple logic puzzle. Three boxes A, B, C, only one has a prize. Clues: A is empty, B has the prize, C is empty. The answer is straightforward.</think>LEAK_CHECK_FINAL: 奖品在 B"
      }
    ]
  }
}

API logging shows that the default request sends max_tokens: 131072. In that failing case, reasoning_content is empty and the reasoning text is returned in normal content.

The same prompt works correctly when the output token cap is lowered:

QWEN_CODE_MAX_OUTPUT_TOKENS=8000 qwen "请回答这个测试问题：三个盒子 A、B、C，只有一个有奖品。线索：A 是空的；B 有奖品；C 是空的。最终答案只输出一行：LEAK_CHECK_FINAL: 奖品在 B。" \
  --auth-type openai \
  --model glm-5.2 \
  --approval-mode yolo \
  --output-format stream-json

With QWEN_CODE_MAX_OUTPUT_TOKENS=8000, API logging shows max_tokens: 8000, the reasoning is returned separately, and the normal text output is only:

LEAK_CHECK_FINAL: 奖品在 B

What did you expect to happen?

Thinking text should not be displayed as normal assistant output.

For models/endpoints that expose reasoning separately, Qwen Code should keep that content in the thinking channel and only show the final answer as normal text. If a provider returns a malformed stream or changes behavior at large max_tokens, Qwen Code should avoid the unsafe request shape or otherwise prevent the leaked reasoning from being displayed as final output.

Client information

Qwen Code: 0.19.3
Runtime: Node.js v22.12.0 / npm 10.9.0
Platform: macOS Darwin 25.5.0 arm64
Auth type: openai
Model: glm-5.2

Login information

Using API key authentication with the OpenAI-compatible provider:

• --auth-type openai
• --model glm-5.2
• Endpoint: DashScope / ModelStudio OpenAI-compatible configuration

Anything else we need to know?

This appears to be max_tokens dependent:

• max_tokens=131072: reproduces; thinking text is returned as normal content, ending with </think>.
• max_tokens=8000: does not reproduce in the same test; thinking is separated and final text is clean.

This may be related to the recent change that made the default request use the model output limit, combined with the GLM-5 family output limit being detected as 128K. It is probably not a TUI rendering-only issue: the leak is visible in headless stream-json output and in the raw logged response shape.

The user is asking a simple logic puzzle. ... The answer is straightforward.</think>LEAK_CHECK_FINAL: 奖品在 B

qwen "请回答这个测试问题：三个盒子 A、B、C，只有一个有奖品。线索：A 是空的；B 有奖品；C 是空的。最终答案只输出一行：LEAK_CHECK_FINAL: 奖品在 B。" \
  --auth-type openai \
  --model glm-5.2 \
  --approval-mode yolo \
  --output-format stream-json

{
  "type": "assistant",
  "message": {
    "content": [
      {
        "type": "text",
        "text": "The user is asking a simple logic puzzle. Three boxes A, B, C, only one has a prize. Clues: A is empty, B has the prize, C is empty. The answer is straightforward.</think>LEAK_CHECK_FINAL: 奖品在 B"
      }
    ]
  }
}

QWEN_CODE_MAX_OUTPUT_TOKENS=8000 qwen "请回答这个测试问题：三个盒子 A、B、C，只有一个有奖品。线索：A 是空的；B 有奖品；C 是空的。最终答案只输出一行：LEAK_CHECK_FINAL: 奖品在 B。" \
  --auth-type openai \
  --model glm-5.2 \
  --approval-mode yolo \
  --output-format stream-json

LEAK_CHECK_FINAL: 奖品在 B

Qwen Code: 0.19.3
Runtime: Node.js v22.12.0 / npm 10.9.0
Platform: macOS Darwin 25.5.0 arm64
Auth type: openai
Model: glm-5.2