佬可以参照官方文档补充下?
一个更通用的配置可能如下,包括思考等级的配置:
(deepseek虽然默认开启思考模式,但思考等级如 high 和 max 的切换还是需要显式配置和指定的)
{
"name": "xxx",
"vendor": "customendpoint",
"apiKey": "${input:chat.lm.secret.-6046024e}",
"apiType": "chat-completions",
"models": [
{
"id": "xxx",
"name": "xxx",
"url": "xxx",
"toolCalling": true,
"vision": true,
"maxInputTokens": 200000,
"maxOutputTokens": 62144,
"thinking": true,
"streaming": true, //流式传输
"reasoningEffortFormat": "chat-completions",
"supportsReasoningEffort": ["none", "low", "medium", "high", "xhigh"] //按需调整
}
]
}
给出一个 deepseek 的参考示例,没使用 Anthropic 端口的原因主要在于,reasoningEffortFormat没有给出对应 anthropic 的参数,不确定是否支持相应的切换
{
"name": "DeepSeek",
"vendor": "customendpoint",
"apiKey": "${input:chat.lm.secret.-384b5a08}",
"apiType": "chat-completions",
"models": [
{
"id": "deepseek-v4-pro",
"name": "Deepseek-V4-Pro",
"url": "https://api.deepseek.com",
"toolCalling": true,
"vision": false,
"streaming": true,
"thinking": true,
"maxInputTokens": 700000,
"maxOutputTokens": 300000,
"reasoningEffortFormat": "chat-completions",
"supportsReasoningEffort": ["high", "max"]
},
{
"id": "deepseek-v4-flash",
"name": "Deepseek-V4-Flash",
"url": "https://api.deepseek.com",
"toolCalling": true,
"vision": false,
"streaming": true,
"thinking": true,
"maxInputTokens": 700000,
"maxOutputTokens": 300000,
"reasoningEffortFormat": "chat-completions",
"supportsReasoningEffort": ["high", "max"]
}
]
}
顺便,vscode 的上下文大小似乎是按maxInputTokens + maxOutputTokens算的(错了当我没说
所以不要让两者之和超过模型上下文大小

不过还是会推荐用插件,能提供更细颗粒度的配置和行为操控
但是第三方插件不能在新的 agent windows 里使用) ^-^