使用GLM5.1,模型反馈接收不到信息。 || Using GLM5.1, model feedback cannot receive information.

[Saro50]

What version of Kimi Code CLI is running?

kimi, version 1.27.0

Which open platform/subscription were you using?

/setup

Which model were you using?

glm-5.1

What platform is your computer?

Darwin 25.3.0 arm64 arm

What issue are you seeing?

VIS上看，实际工具有把信息返回
{
"user_input": [
{
"type": "text",
"text": "@README.md 了解下这个项目，把你收到的工具响应全部输出，无论是否有内容，不需要分析说明，仅透传工具响应"
}
]
}

{
"type": "text",
"text": "工具响应为空（<tools></tools> 标签内无内容）。README.md 文件可能为空或读取未返回内容。让我用 Shell 确认一下："
}

What steps can reproduce the bug?

配置config ，让模型读文件就会出现，切换为glm5,就是正常的，但glm5.1在claude code内能正确读取文件

What is the expected behavior?

期望能正确读取文件

Additional information

暂无

---

What version of Kimi Code CLI is running?

kimi, version 1.27.0

Which open platform/subscription were you using?

/setup

Which model were you using?

glm-5.1

What platform is your computer?

Darwin 25.3.0 arm64 arm

What issue are you seeing?

Looking at VIS, the actual tool returns information.
{
"user_input": [
{
"type": "text",
"text": "@README.md Learn about this project and output all the tool responses you receive, regardless of whether there is content or not. No analysis or explanation is required. Only the tool responses are transparently transmitted."
}
]
}

{
"type": "text",
"text": "The tool response is empty (there is no content in the <tools></tools> tag). The README.md file may be empty or the read does not return content. Let me confirm with Shell:"
}

What steps can reproduce the bug?

Configure config and let the model read the file and it will appear. If you switch to glm5, it will be normal, but glm5.1 can read the file correctly in the claude code.

What is the expected behavior?

Expect the file to be read correctly

Additional information

Not yet

[Saro50]

补充使用的https://open.bigmodel.cn/api/coding/paas/v4 无法使用，更换为https://open.bigmodel.cn/api/anthropic 有时能读出信息，表现不稳定

---

The supplementary https://open.bigmodel.cn/api/coding/paas/v4 cannot be used. Replace it with https://open.bigmodel.cn/api/anthropic. Sometimes the information can be read and the performance is unstable.

[loongxjin]

1. tool_result_to_message 对成功的 shell 命令生成了 2 个 TextPart：
   • TextPart(text='Command executed successfully.')
   • TextPart(text='main\n')
2. Message._serialize_content 在有多个 TextPart 时，序列化为 [{"type": "text", "text": "..."},
   ...]（list 格式）
3. tool message content 字段通常期望 str，传 list
   格式时，部分模型无法正确解析，导致 LLM 只看到 标签部分，看不到实际输出。
4. LLM 反复认为 "工具似乎没有返回输出"，于是不断换格式重试。
   修复方案是把 tool result 的 message 和 output 合并到一个 TextPart，确保序列化后 content 是纯
   str.

[RealKai42]

感谢反馈。这个问题的根因是 GLM API 的 tool message content 字段只接受 string 类型，而 OpenAI 官方规范定义的类型是 Union[str, Iterable[ChatCompletionContentPartTextParam]]（即 string 或 array 都合法）。

当工具执行后返回多段文本时（例如状态信息 + 实际输出），我们会生成多个 TextPart，序列化后变成 array 格式。OpenAI 官方 API 能正常处理，但 GLM 的实现会将其视为空内容。

这属于 GLM 侧对 OpenAI 兼容协议的实现问题，建议向智谱反馈。

我会关闭这个 issue，如果依旧认为是 bug，可以重新 open