fix - fix qwen35 moe reasoning parser

[alibaba-miji]

default set thinking mode = True for qwen35 moe

[LLLLKKKK]

Code Review: PR #812 — fix - fix qwen35 moe reasoning parser

Thanks for the fix! Qwen3.5 MoE models definitely need reasoning parsing enabled by default. Here are my findings:

1. [Bug] Accidental trailing characters (reasoning_tool_base_renderer.py:119)

The comment change adds two trailing middle-dot characters (··, U+00B7):

-            # logprobs模式下使用普通StreamStatus
+            # logprobs模式下使用普通StreamStatus··

This looks unintentional — should be reverted.

2. [Performance] render_chat() called N times redundantly

_create_reasoning_parser() calls self.render_chat(request) which does full Jinja2 rendering + tokenizer encoding. In _create_status_list, this runs inside a list comprehension for _ in range(n), so when n > 1 the same expensive call repeats with identical results. Consider hoisting it out of the loop:

detector = self._create_detector(request)
reasoning_parser = self._create_reasoning_parser(request)
return [
    ReasoningToolStreamStatus(request, detector, reasoning_parser)
    for _ in range(n)
]

3. [Correctness] Exception fallback may be too conservative

except Exception as e:
    logging.error(f"Failed to render chat in _create_reasoning_parser: {e}")
return ReasoningParser(model_type="qwen3")

Since Qwen3.5 templates "always append <think>", the safer fallback on render failure would be qwen3-thinking (force_reasoning=True). With qwen3 as fallback, if rendering fails the <think>...</think> output won't be parsed as reasoning and will leak into normal response content.

4. [Testing] No tests added

No unit tests for the new in_think_mode() or _create_reasoning_parser() overrides. Recommend adding tests to verify:

• in_think_mode returns True by default, False when thinking disabled
• _create_reasoning_parser returns correct parser type based on rendered prompt
• Fallback behavior on exception

Positive notes

• @override decorator correctly used on both new methods
• in_think_mode pattern is consistent with QwenRenderer.in_think_mode
• Core logic is sound — the dynamic qwen3-thinking vs qwen3 selection based on prompt suffix is a clean approach

---

Automated review by CI bot — 2026-03-21

[LLLLKKKK]

🤖 AI Code Review — PR #812 fix - fix qwen35 moe reasoning parser

概述

为 Qwen3CoderRenderer（服务 qwen35_moe / qwen35_dense / qwen3_coder_moe）添加 reasoning parser 支持。Qwen3.5 chat template 默认追加 <think>，但之前 renderer 缺少对应的 reasoning 解析逻辑，导致 thinking 内容无法正确提取。

改动包括：

• 新增 in_think_mode override：除非 disable_thinking() 显式为 True，否则默认开启 thinking
• 新增 _create_reasoning_parser override：根据渲染后的 prompt 是否以 think_start_tag 结尾，选择 qwen3-thinking（force_reasoning）或 qwen3 模式

建议改进

P1: 缺少单元测试

PR 没有为新增的 in_think_mode 和 _create_reasoning_parser 添加测试。建议至少覆盖：

• in_think_mode 在 disable_thinking=True/False/None 时的返回值
• _create_reasoning_parser 在 think 模式下返回正确的 model_type（qwen3 vs qwen3-thinking）
• _create_reasoning_parser 在 render_chat 抛异常时的 fallback 行为

P2: 代码重复

_create_reasoning_parser 与 QwenReasoningToolRenderer._create_reasoning_parser（qwen_reasoning_tool_renderer.py:44-59）几乎完全相同。建议提取到基类 ReasoningToolBaseRenderer 作为默认实现，或提取为共享 helper，避免两处维护。

P2: in_think_mode 忽略 self.think_mode 配置

新的 in_think_mode 完全忽略了 self.think_mode（来自 generate_env_config）。如果运维通过配置显式关闭 think_mode，该 override 仍会返回 True。这与 QwenRenderer.in_think_mode 是同样的模式，属于项目惯例，风险可控，但建议在 docstring 中明确说明。

总结

小而必要的 bugfix，逻辑正确，与已有的 QwenReasoningToolRenderer 模式一致。主要建议是补充测试和减少代码重复。