跳转到主要内容

17 自定义提供商

用扩展的 pi.registerProvider() 注册自定义提供商、OAuth 登录与自定义流式 API。

扩展可以通过 pi.registerProvider() 注册自定义模型提供商。它可以实现：

  • 代理 —— 将请求路由到企业代理或 API 网关
  • 自定义端点 —— 使用自托管或私有的模型部署
  • OAuth/SSO —— 为企业级提供商添加认证流程
  • 自定义 API —— 为非标准 LLM API 实现流式传输

示例扩展

请参阅以下完整的提供商示例：

目录

快速参考

扩展既可以注册一个完整的 pi-ai Provider，也可以使用旧版的提供商配置形式。当需要自定义认证、过滤、刷新或流式行为时，建议使用完整的 Provider。pi 会将 models.json 覆盖叠加在已注册的原生提供商之上。

import { createProvider, openAICompletionsApi } from "@earendil-works/pi-ai";
import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";

export default function (pi: ExtensionAPI) {
  pi.registerProvider(createProvider({
    id: "native-local",
    name: "Native Local",
    baseUrl: "http://localhost:8080/v1",
    auth: {
      apiKey: {
        name: "Local server API key",
        async login(interaction) {
          return {
            type: "api_key",
            key: await interaction.prompt({ type: "secret", message: "API key" })
          };
        },
        async resolve({ credential }) {
          return credential?.key
            ? { auth: { apiKey: credential.key }, source: "stored API key" }
            : undefined;
        }
      }
    },
    models: [],
    api: openAICompletionsApi()
  }));

  // Legacy provider-config form:
  // Override baseUrl for existing provider
  pi.registerProvider("anthropic", {
    baseUrl: "https://proxy.example.com"
  });

  // Register new provider with models
  pi.registerProvider("my-provider", {
    name: "My Provider",
    baseUrl: "https://api.example.com",
    apiKey: "$MY_API_KEY",
    api: "openai-completions",
    models: [
      {
        id: "my-model",
        name: "My Model",
        reasoning: false,
        input: ["text", "image"],
        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
        contextWindow: 128000,
        maxTokens: 4096
      }
    ]
  });
}

扩展工厂函数也可以是 async 的。要实现动态模型发现，请在工厂函数中（而不是在 session_start 里）获取并注册模型。pi 会在继续启动之前等待工厂函数完成，因此该提供商在交互式启动期间以及 pi --list-models 中都可用。

覆盖已有提供商

最简单的用例：将已有提供商的请求重定向到代理。

// All Anthropic requests now go through your proxy
pi.registerProvider("anthropic", {
  baseUrl: "https://proxy.example.com"
});

// Add custom headers to OpenAI requests
pi.registerProvider("openai", {
  headers: {
    "X-Custom-Header": "value"
  }
});

// Both baseUrl and headers
pi.registerProvider("google", {
  baseUrl: "https://ai-gateway.corp.com/google",
  headers: {
    "X-Corp-Auth": "$CORP_AUTH_TOKEN"  // env var or literal
  }
});

当只提供 baseUrl 和/或 headers（不含 models）时，该提供商的所有既有模型都会保留，并使用新的端点。

注册新提供商

要添加一个全新的提供商，请在必要配置之外指定 models。

如果模型列表来自远程端点，请使用 async 扩展工厂函数：

import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";

export default async function (pi: ExtensionAPI) {
  const response = await fetch("http://localhost:1234/v1/models");
  const payload = (await response.json()) as {
    data: Array<{
      id: string;
      name?: string;
      context_window?: number;
      max_tokens?: number;
    }>;
  };

  pi.registerProvider("local-openai", {
    baseUrl: "http://localhost:1234/v1",
    apiKey: "$LOCAL_OPENAI_API_KEY",
    api: "openai-completions",
    models: payload.data.map((model) => ({
      id: model.id,
      name: model.name ?? model.id,
      reasoning: false,
      input: ["text"],
      cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
      contextWindow: model.context_window ?? 128000,
      maxTokens: model.max_tokens ?? 4096,
    })),
  });
}

这会在启动完成之前注册获取到的模型。

pi.registerProvider("my-llm", {
  baseUrl: "https://api.my-llm.com/v1",
  apiKey: "$MY_LLM_API_KEY",  // env var reference
  api: "openai-completions",  // which streaming API to use
  models: [
    {
      id: "my-llm-large",
      name: "My LLM Large",
      reasoning: true,        // supports extended thinking
      input: ["text", "image"],
      cost: {
        input: 3.0,           // $/million tokens
        output: 15.0,
        cacheRead: 0.3,
        cacheWrite: 3.75
      },
      contextWindow: 200000,
      maxTokens: 16384
    }
  ]
});

提供 models 时，它会替换该提供商的所有既有模型。

apiKey 与自定义请求头的取值使用与 models.json 相同的配置值语法：值以 !command 开头时将整个值作为命令执行；$ENV_VAR 和 ${ENV_VAR} 插值环境变量；$$ 输出字面量 $；$! 输出字面量 !。

注销提供商

使用 pi.unregisterProvider(name) 移除之前通过 pi.registerProvider(name, ...) 注册的提供商：

// Register
pi.registerProvider("my-llm", {
  baseUrl: "https://api.my-llm.com/v1",
  apiKey: "$MY_LLM_API_KEY",
  api: "openai-completions",
  models: [
    {
      id: "my-llm-large",
      name: "My LLM Large",
      reasoning: true,
      input: ["text", "image"],
      cost: { input: 3.0, output: 15.0, cacheRead: 0.3, cacheWrite: 3.75 },
      contextWindow: 200000,
      maxTokens: 16384
    }
  ]
});

// Later, remove it
pi.unregisterProvider("my-llm");

注销会移除该提供商的动态模型、API 密钥回退、OAuth 提供商注册以及自定义流处理器注册。之前被覆盖的内置模型或提供商行为都会恢复。

在扩展初始加载阶段之后发起的调用会立即生效，因此无需 /reload。

API 类型

api 字段决定使用哪个流式实现：

API适用对象
anthropic-messagesAnthropic Claude API 及其兼容实现
openai-completionsOpenAI Chat Completions API 及其兼容实现
openai-responsesOpenAI Responses API
azure-openai-responsesAzure OpenAI Responses API
openai-codex-responsesOpenAI Codex Responses API
mistral-conversationsMistral 原生 Chat Completions 流式
google-generative-aiGoogle Generative AI API
google-vertexGoogle Vertex AI API
bedrock-converse-streamAmazon Bedrock Converse API

大多数 OpenAI 兼容提供商都可以使用 openai-completions。模型特有的思考级别用模型级别的 thinkingLevelMap 设置，提供商的怪癖用 compat 处理。xhigh 和 max 级别需要显式启用，要求映射条目为非 null 值，且中间可能隔着不支持的空档：

models: [{
  id: "custom-model",
  // ...
  reasoning: true,
  thinkingLevelMap: {              // map pi levels to provider values; null hides unsupported levels
    minimal: null,
    low: null,
    medium: null,
    high: "default",
    xhigh: null,
    max: "max"
  },
  compat: {
    supportsDeveloperRole: false,   // use "system" instead of "developer"
    supportsReasoningEffort: true,
    maxTokensField: "max_tokens",   // instead of "max_completion_tokens"
    requiresToolResultName: true,   // tool results need name field
    thinkingFormat: "qwen",        // top-level enable_thinking: true
    cacheControlFormat: "anthropic" // Anthropic-style cache_control markers
  }
}]

OpenRouter 风格的 reasoning: { effort } 控制使用 openrouter。Together 风格的 reasoning: { enabled } 控制使用 together；配合 supportsReasoningEffort 时，它还会发送 reasoning_effort。读取 chat_template_kwargs.enable_thinking 且需要 preserve_thinking 的本地 Qwen 兼容服务器使用 qwen-chat-template。 对于通过在系统提示词、最后一个工具定义以及最后一条 user、assistant 或工具结果文本内容上添加 cache_control 来暴露 Anthropic 风格提示词缓存的 OpenAI 兼容提供商，使用 cacheControlFormat: "anthropic"。

对于使用 api: "anthropic-messages" 的 Anthropic 兼容提供商，若其上游模型需要自适应思考（thinking.type: "adaptive" 加 output_config.effort），请在相应的模型或提供商上设置 compat.forceAdaptiveThinking: true。内置的自适应 Claude 模型会自动设置。compat.allowEmptySignature: true 只应用于会输出空思考签名并期望回放时传回 signature: "" 的提供商。

迁移说明：Mistral 已从 openai-completions 迁移到 mistral-conversations。 原生 Mistral 模型请使用 mistral-conversations。 如果你有意让 Mistral 兼容/自定义端点走 openai-completions，请按需显式设置 compat 标志。

认证请求头

如果提供商期望 Authorization: Bearer <key> 但并不使用标准 API，请设置 authHeader: true：

pi.registerProvider("custom-api", {
  baseUrl: "https://api.example.com",
  apiKey: "$MY_API_KEY",
  authHeader: true,  // adds Authorization: Bearer header
  api: "openai-completions",
  models: [...]
});

密钥会按请求逐次解析。显式设置的请求 Authorization 头优先于生成的值。

OAuth 支持

添加可与 /login 集成的 OAuth/SSO 认证：

import type { OAuthCredentials, OAuthLoginCallbacks } from "@earendil-works/pi-ai";

pi.registerProvider("corporate-ai", {
  baseUrl: "https://ai.corp.com/v1",
  api: "openai-responses",
  models: [...],
  oauth: {
    name: "Corporate AI (SSO)",

    async login(callbacks: OAuthLoginCallbacks): Promise<OAuthCredentials> {
      const method = await callbacks.onSelect({
        message: "Select login method:",
        options: [
          { id: "browser", label: "Browser OAuth" },
          { id: "device", label: "Device code" }
        ]
      });
      if (!method) throw new Error("Login cancelled");

      let code: string;
      if (method === "device") {
        callbacks.onDeviceCode({
          userCode: "ABCD-1234",
          verificationUri: "https://sso.corp.com/device",
          intervalSeconds: 5,
          expiresInSeconds: 900
        });
        code = await pollDeviceCodeUntilComplete();
      } else {
        callbacks.onAuth({ url: "https://sso.corp.com/authorize?..." });
        code = await callbacks.onPrompt({ message: "Enter SSO code:" });
      }

      // Exchange for tokens (your implementation)
      const tokens = await exchangeCodeForTokens(code);

      return {
        refresh: tokens.refreshToken,
        access: tokens.accessToken,
        expires: Date.now() + tokens.expiresIn * 1000
      };
    },

    async refreshToken(credentials: OAuthCredentials, signal: AbortSignal): Promise<OAuthCredentials> {
      const tokens = await refreshAccessToken(credentials.refresh, signal);
      return {
        refresh: tokens.refreshToken ?? credentials.refresh,
        access: tokens.accessToken,
        expires: Date.now() + tokens.expiresIn * 1000
      };
    },

    getApiKey(credentials: OAuthCredentials): string {
      return credentials.access;
    }
  }
});

注册后，用户即可通过 /login corporate-ai 完成认证。

OAuthLoginCallbacks

callbacks 对象为提供商自己的登录流程提供与 UI 无关的交互方式：

interface OAuthLoginCallbacks {
  // Open URL in browser (for OAuth redirects)
  onAuth(params: { url: string }): void;

  // Show device code (for device authorization flow)
  onDeviceCode(params: {
    userCode: string;
    verificationUri: string;
    intervalSeconds?: number;
    expiresInSeconds?: number;
  }): void;

  // Show transient progress
  onProgress?(message: string): void;

  // Prompt user for input (for manual token entry)
  onPrompt(params: { message: string }): Promise<string>;

  // Show an interactive selector, e.g. to choose browser OAuth vs device code
  onSelect(params: {
    message: string;
    options: { id: string; label: string }[];
  }): Promise<string | undefined>;
}

OAuthCredentials

凭据会持久化到 ~/.pi/agent/auth.json：

interface OAuthCredentials {
  refresh: string;   // Refresh token (for refreshToken())
  access: string;    // Access token (returned by getApiKey())
  expires: number;   // Expiration timestamp in milliseconds
}

自定义流式 API

对于使用非标准 API 的提供商，可实现 streamSimple。编写自己的实现之前，请先研究现有的 API 实现：

参考实现：

流式模式

所有提供商都遵循相同的模式：

import {
  type AssistantMessage,
  type AssistantMessageEventStream,
  type Context,
  type Model,
  type SimpleStreamOptions,
  calculateCost,
  createAssistantMessageEventStream,
} from "@earendil-works/pi-ai";

function streamMyProvider(
  model: Model<any>,
  context: Context,
  options?: SimpleStreamOptions
): AssistantMessageEventStream {
  const stream = createAssistantMessageEventStream();

  (async () => {
    // Initialize output message
    const output: AssistantMessage = {
      role: "assistant",
      content: [],
      api: model.api,
      provider: model.provider,
      model: model.id,
      usage: {
        input: 0,
        output: 0,
        cacheRead: 0,
        cacheWrite: 0,
        totalTokens: 0,
        cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
      },
      stopReason: "pending",
      timestamp: Date.now(),
    };

    try {
      // Push start event
      stream.push({ type: "start", partial: output });

      // Make API request and process response...
      // Push content events as they arrive and set stopReason from the terminal event.
      if (output.stopReason === "pending") {
        throw new Error("Provider stream ended without a stop reason");
      }
      if (output.stopReason === "error" || output.stopReason === "aborted") {
        throw new Error(output.errorMessage || "An unknown error occurred");
      }

      // Push done event
      stream.push({
        type: "done",
        reason: output.stopReason,
        message: output
      });
      stream.end();
    } catch (error) {
      output.stopReason = options?.signal?.aborted ? "aborted" : "error";
      output.errorMessage = error instanceof Error ? error.message : String(error);
      stream.push({ type: "error", reason: output.stopReason, error: output });
      stream.end();
    }
  })();

  return stream;
}

事件类型

按以下顺序通过 stream.push() 推送事件：

  1. { type: "start", partial: output } - 流开始

  2. 内容事件（可重复，为每个块跟踪 contentIndex）：

    • { type: "text_start", contentIndex, partial } - 文本块开始
    • { type: "text_delta", contentIndex, delta, partial } - 文本分片
    • { type: "text_end", contentIndex, content, partial } - 文本块结束
    • { type: "thinking_start", contentIndex, partial } - 思考开始
    • { type: "thinking_delta", contentIndex, delta, partial } - 思考分片
    • { type: "thinking_end", contentIndex, content, partial } - 思考结束
    • { type: "toolcall_start", contentIndex, partial } - 工具调用开始
    • { type: "toolcall_delta", contentIndex, delta, partial } - 工具调用 JSON 分片
    • { type: "toolcall_end", contentIndex, toolCall, partial } - 工具调用结束

  3. { type: "done", reason, message } 或 { type: "error", reason, error } - 流结束

每个事件的 partial 字段包含当前 AssistantMessage 的状态。收到数据时更新 output.content，然后把 output 作为 partial 传入。

内容块

内容块到达时将其添加到 output.content：

// Text block
output.content.push({ type: "text", text: "" });
stream.push({ type: "text_start", contentIndex: output.content.length - 1, partial: output });

// As text arrives
const block = output.content[contentIndex];
if (block.type === "text") {
  block.text += delta;
  stream.push({ type: "text_delta", contentIndex, delta, partial: output });
}

// When block completes
stream.push({ type: "text_end", contentIndex, content: block.text, partial: output });

工具调用

工具调用需要累积 JSON 并解析：

// Start tool call
output.content.push({
  type: "toolCall",
  id: toolCallId,
  name: toolName,
  arguments: {}
});
stream.push({ type: "toolcall_start", contentIndex: output.content.length - 1, partial: output });

// Accumulate JSON
let partialJson = "";
partialJson += jsonDelta;
try {
  block.arguments = JSON.parse(partialJson);
} catch {}
stream.push({ type: "toolcall_delta", contentIndex, delta: jsonDelta, partial: output });

// Complete
stream.push({
  type: "toolcall_end",
  contentIndex,
  toolCall: { type: "toolCall", id, name, arguments: block.arguments },
  partial: output
});

用量与费用

根据 API 响应更新用量并计算费用：

output.usage.input = response.usage.input_tokens;
output.usage.output = response.usage.output_tokens;
output.usage.cacheRead = response.usage.cache_read_tokens ?? 0;
output.usage.cacheWrite = response.usage.cache_write_tokens ?? 0;
output.usage.totalTokens = output.usage.input + output.usage.output +
                           output.usage.cacheRead + output.usage.cacheWrite;
calculateCost(model, output.usage);

上下文溢出错误

当请求超出模型的上下文窗口时，pi 可以通过对会话进行上下文压缩并重试来自动恢复。只有当 pi 把失败识别为溢出时，该恢复机制才会生效。

检测在定稿的 assistant 消息上进行：

如果你的提供商返回的溢出错误消息 pi 无法识别，请在注册该提供商的同一个扩展中规范化错误。使用 message_end 处理器改写 assistant 消息，使其 errorMessage 以 pi 能识别的短语开头。通用回退值 context_length_exceeded 是最安全的选择。

const MY_PROVIDER_OVERFLOW_PATTERN = /your provider's overflow phrase/i;

export default function (pi: ExtensionAPI) {
  pi.registerProvider("my-provider", { /* ... */ });

  pi.on("message_end", (event, ctx) => {
    const message = event.message;
    if (message.role !== "assistant") return;
    if (message.stopReason !== "error") return;
    if (
      message.provider !== "my-provider" &&
      ctx.model?.provider !== "my-provider"
    )
      return;

    const errorMessage = message.errorMessage ?? "";
    if (errorMessage.includes("context_length_exceeded")) return;
    if (!MY_PROVIDER_OVERFLOW_PATTERN.test(errorMessage)) return;

    return {
      message: {
        ...message,
        errorMessage: `context_length_exceeded: ${errorMessage}`,
      },
    };
  });
}

message_end 在 pi 为自动上下文压缩记录该 assistant 消息之前运行，因此 pi 检查的是改写后的 errorMessage。有了这一处理，pi 将：

  1. 从 errorMessage 检测溢出。
  2. 将失败的 assistant 消息从活跃上下文中丢弃。
  3. 执行上下文压缩。
  4. 重试该请求一次。

请谨慎地对改写加保护条件：

  • 将作用域限定在你的提供商（message.provider 与 ctx.model?.provider），避免影响其他提供商的无关错误。
  • 匹配提供商特有的模式，而不是 pi 的通用溢出模式。如果改写了限流或节流错误（rate limit、too many requests），会错误地触发上下文压缩，而不是走 pi 正常的退避重试路径。
  • 当 errorMessage 已包含 context_length_exceeded 时跳过，以保证处理器的幂等性。

注册

注册你的流式函数：

pi.registerProvider("my-provider", {
  baseUrl: "https://api.example.com",
  apiKey: "$MY_API_KEY",
  api: "my-custom-api",
  models: [...],
  streamSimple: streamMyProvider
});

测试你的实现

使用内置提供商所用的同一套测试来测试你的提供商。从 packages/ai/test/ 复制并改造以下测试文件：

测试用途
stream.test.ts基础流式、文本输出
tokens.test.tstoken 计数与用量
abort.test.tsAbortSignal 处理
empty.test.ts空/最小响应
context-overflow.test.ts上下文窗口限制
image-limits.test.ts图像输入处理
unicode-surrogate.test.tsUnicode 边界情况
tool-call-without-result.test.ts工具调用边界情况
image-tool-result.test.ts工具结果中的图像
total-tokens.test.ts总 token 计算
cross-provider-handoff.test.ts提供商之间的上下文交接

用你的提供商/模型组合运行测试以验证兼容性。

配置参考

interface ProviderConfig {
  /** Display name for the provider in UI such as /login. */
  name?: string;

  /** API endpoint URL. Required when defining models. */
  baseUrl?: string;

  /** API key literal, env interpolation ($ENV_VAR or ${ENV_VAR}), or !command. Required when defining models (unless oauth). */
  apiKey?: string;

  /** API type for streaming. Required at provider or model level when defining models. */
  api?: Api;

  /** Custom streaming implementation for non-standard APIs. */
  streamSimple?: (
    model: Model<Api>,
    context: Context,
    options?: SimpleStreamOptions
  ) => AssistantMessageEventStream;

  /** Custom headers to include in requests. Values use the same resolution syntax as apiKey. */
  headers?: Record<string, string>;

  /** If true, adds Authorization: Bearer header with the resolved API key. */
  authHeader?: boolean;

  /** Models to register. If provided, replaces all existing models for this provider. */
  models?: ProviderModelConfig[];

  /** OAuth provider for /login support. */
  oauth?: {
    name: string;
    login(callbacks: OAuthLoginCallbacks): Promise<OAuthCredentials>;
    refreshToken(credentials: OAuthCredentials, signal: AbortSignal): Promise<OAuthCredentials>;
    getApiKey(credentials: OAuthCredentials): string;
  };
}

模型定义参考

interface ProviderModelConfig {
  /** Model ID (e.g., "claude-sonnet-4-20250514"). */
  id: string;

  /** Display name (e.g., "Claude 4 Sonnet"). */
  name: string;

  /** API type override for this specific model. */
  api?: Api;

  /** API endpoint URL override for this specific model. */
  baseUrl?: string;

  /** Whether the model supports extended thinking. */
  reasoning: boolean;

  /** Maps pi thinking levels to provider/model-specific values; null marks a level unsupported. */
  thinkingLevelMap?: Partial<Record<"off" | "minimal" | "low" | "medium" | "high" | "xhigh" | "max", string | null>>;

  /** Supported input types. */
  input: ("text" | "image")[];

  /** Cost per million tokens (for usage tracking). */
  cost: {
    input: number;
    output: number;
    cacheRead: number;
    cacheWrite: number;
  };

  /** Maximum context window size in tokens. */
  contextWindow: number;

  /** Maximum output tokens. */
  maxTokens: number;

  /** Custom headers for this specific model. */
  headers?: Record<string, string>;

  /** Compatibility settings for the selected API. */
  compat?: {
    // openai-completions
    supportsStore?: boolean;
    supportsDeveloperRole?: boolean;
    supportsReasoningEffort?: boolean;
    supportsUsageInStreaming?: boolean;
    supportsFinishReason?: boolean;
    supportsStrictMode?: boolean;
    supportsOpenAIGrammarTools?: boolean; // openai-completions/openai-responses; false falls back to normal function tools
    maxTokensField?: "max_completion_tokens" | "max_tokens";
    requiresToolResultName?: boolean;
    requiresAssistantAfterToolResult?: boolean;
    requiresThinkingAsText?: boolean;
    requiresReasoningContentOnAssistantMessages?: boolean;
    thinkingFormat?: "openai" | "openrouter" | "deepseek" | "together" | "baseten" | "zai" | "qwen" | "chat-template" | "qwen-chat-template" | "string-thinking" | "ant-ling";
    chatTemplateKwargs?: Record<string, string | number | boolean | null | { "$var": "thinking.enabled" | "thinking.effort" | "thinking.budget"; omitWhenOff?: boolean }>;
    chatTemplateArgs?: Record<string, string | number | boolean | null | { "$var": "thinking.enabled" | "thinking.effort" | "thinking.budget"; omitWhenOff?: boolean }>;
    thinkingTokenBudgetField?: "thinking_token_budget" | "thinking_budget" | "thinking_budget_tokens";
    supportsThinkingTokenBudget?: boolean;
    cacheControlFormat?: "anthropic";
    sessionAffinityFormat?: "openai" | "openai-nosession" | "openrouter";
    sendSessionAffinityHeaders?: boolean;

    // anthropic-messages
    supportsEagerToolInputStreaming?: boolean;
    supportsLongCacheRetention?: boolean;
    sendSessionAffinityHeaders?: boolean;
    supportsCacheControlOnTools?: boolean;
    forceAdaptiveThinking?: boolean;
    allowEmptySignature?: boolean;
    supportsStrictTools?: boolean;
  };
}

openrouter 发送 reasoning: { effort }。deepseek 发送 thinking: { type: "enabled" | "disabled" }，启用时还会发送 reasoning_effort。together 发送 reasoning: { enabled }，并在启用 supportsReasoningEffort 时同时发送 reasoning_effort。qwen 用于 DashScope 风格的顶层 enable_thinking。读取 chat_template_kwargs.enable_thinking 且需要 preserve_thinking 的本地 Qwen 兼容服务器使用 qwen-chat-template。chat-template 用于可配置的 chat_template_kwargs，例如由 vLLM 提供服务的 DeepSeek V3.x 可用 chatTemplateKwargs: { "thinking": { "$var": "thinking.enabled" } }。当提供商期望在 chat_template_args 下传递开关取值并可选支持顶层 reasoning_effort 时，将 thinkingFormat: "baseten" 与 chatTemplateArgs 搭配使用。 thinkingTokenBudgetField 将按级别钳制的思考预算作为顶层请求字段发送（vLLM 用 thinking_token_budget，Qwen/SGLang 用 thinking_budget，llama.cpp 用 thinking_budget_tokens）。supportsThinkingTokenBudget: true 是 vLLM 字段名的别名。不要在 DashScope 的 Qwen 模型上将它与 reasoning_effort 组合使用。 cacheControlFormat: "anthropic" 会将 Anthropic 风格的 cache_control 标记应用到系统提示词、最后一个工具定义以及最后一条 user、assistant 或工具结果文本内容上。

最后修改 2026-10-16