[Bug] [thinking] enabled=false 对第三方 OpenAI 兼容供应商不生效（DeepSeek 仍默认思考） || [Bug] [thinking] enabled=false does not take effect for third-party OpenAI compatible vendors (DeepSeek still defaults to thinking)

[lin200083]

问题描述

当在 中配置第三方 OpenAI 兼容供应商（如 DeepSeek V4 Flash via Sensenova）时， 无法关闭模型的思考模式。模型仍然会输出推理过程。

复现步骤

1. 在 config.toml 中配置第三方 OpenAI 兼容供应商：

   [providers.my-provider]
   type = "openai"
   base_url = "https://xxx/v1"
   api_key = "sk-xxx"
   
   [models.my-model]
   provider = "my-provider"
   model = "deepseek-v4-flash"
   max_context_size = 1048576
   capabilities = ["thinking"]
   
   [thinking]
   enabled = false

2. 启动 CLI，向模型发送消息

3. 模型仍然输出推理内容（reasoning_content）

根因分析

DeepSeek V4 Flash 默认启用思考模式，其 API 参数 reasoning_effort 缺省值为 "medium"。当 CLI 的 [thinking] enabled = false 时，CLI 只是不发送 reasoning_effort 参数，而非显式传递 reasoning_effort: "none"。结果 DeepSeek 服务端收不到该参数，就按自身默认值继续思考。

尝试过的方案（均无效）

1. 设置 effort = "none" — CLI 未将 "none" 透传给 API
2. 移除 capabilities = ["thinking"]" + enabled = false` — CLI 完全不参与思考参数，DeepSeek 仍按默认值思考
3. 设置 support_efforts = ["low", "medium", "high", "none"]" + default_effort = "none"` — 同样未生效

期望行为

当 [thinking] enabled = false 时，对于 OpenAI 兼容供应商，CLI 应在请求体中显式传递 reasoning_effort: "none"，而非省略该参数。

或者，当 effort = "none" 时，CLI 应将 "none" 值透传给供应商 API。

环境信息

• Kimi Code CLI 版本：latest
• 供应商：Sensenova（OpenAI 兼容）→ DeepSeek V4 Flash
• 操作系统：Windows

---

Problem description

When configuring a third-party OpenAI compatible vendor (such as DeepSeek V4 Flash via Sensenova) in , it is not possible to turn off the think mode of the model. The model still outputs the inference process.

Reproduction steps

1. Configure third-party OpenAI compatible providers in config.toml:

   [providers.my-provider]
   type = "openai"
   base_url = "https://xxx/v1"
   api_key = "sk-xxx"
   
   [models.my-model]
   provider = "my-provider"
   model = "deepseek-v4-flash"
   max_context_size = 1048576
   capabilities = ["thinking"]
   
   [thinking]
   enabled = false

2. Start the CLI and send messages to the model

3. The model still outputs reasoning content (reasoning_content)

Root cause analysis

DeepSeek V4 Flash enables thinking mode by default, and its API parameter reasoning_effort has a default value of "medium". When the CLI's [thinking] enabled = false, the CLI simply doesn't send the reasoning_effort parameter** instead of explicitly passing reasoning_effort: "none". As a result, the DeepSeek server did not receive this parameter and continued to think according to its default value.

Tried solutions (all invalid)

1. Set effort = "none" — CLI does not pass "none" through to API
2. Remove capabilities = ["thinking"]" + enabled = false` - CLI does not participate in thinking parameters at all, DeepSeek still thinks according to the default value
3. Set support_efforts = ["low", "medium", "high", "none"]" + default_effort = "none"` — also does not take effect

Desired behavior

When [thinking] enabled = false, for OpenAI compatible vendors, the CLI should explicitly pass reasoning_effort: "none" in the request body instead of omitting this parameter.

Alternatively, when effort = "none", the CLI should transparently pass the "none" value to the vendor API.

Environment information

• Kimi Code CLI version: latest
• Supplier: Sensenova (OpenAI compatible) → DeepSeek V4 Flash
• Operating system: Windows

[lin200083]

错误仓库，已迁移到 MoonshotAI/kimi-code

---

Error repository, migrated to MoonshotAI/kimi-code