Claude 消息接口

• 完全兼容 Anthropic Claude Messages 原生协议（POST /v1/messages）

• 支持多轮对话、流式 SSE、工具调用、extended thinking
• 支持文本、图像等多模态内容
• 响应为上游原样透传，无 {code, data} 外层包装

curl https://aiboxapi.com/v1/messages \
  -H "x-api-key: $API_KEY" \
  -H "anthropic-version: 2025-10-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-sonnet-4-6",
    "max_tokens": 1024,
    "messages": [
      {"role": "user", "content": "你好，世界"}
    ]
  }'

import anthropic

client = anthropic.Anthropic(
    api_key="YOUR_API_KEY",
    base_url="https://aiboxapi.com"
)

message = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[
        {"role": "user", "content": "你好，世界"}
    ]
)

print(message.content)

import Anthropic from '@anthropic-ai/sdk';

const client = new Anthropic({
  apiKey: process.env.API_KEY,
  baseURL: 'https://aiboxapi.com'
});

const message = await client.messages.create({
  model: 'claude-sonnet-4-6',
  max_tokens: 1024,
  messages: [
    { role: 'user', content: '你好，世界' }
  ]
});

console.log(message.content);

package main

import (
    "bytes"
    "encoding/json"
    "fmt"
    "io/ioutil"
    "net/http"
    "os"
)

func main() {
    url := "https://aiboxapi.com/v1/messages"

    payload := map[string]interface{}{
        "model": "claude-sonnet-4-6",
        "max_tokens": 1024,
        "messages": []map[string]string{
            {
                "role":    "user",
                "content": "你好，世界",
            },
        },
    }

    jsonData, _ := json.Marshal(payload)

    req, _ := http.NewRequest("POST", url, bytes.NewBuffer(jsonData))
    req.Header.Set("x-api-key", os.Getenv("API_KEY"))
    req.Header.Set("anthropic-version", "2025-10-01")
    req.Header.Set("Content-Type", "application/json")

    client := &http.Client{}
    resp, err := client.Do(req)
    if err != nil {
        panic(err)
    }
    defer resp.Body.Close()

    body, _ := ioutil.ReadAll(resp.Body)
    fmt.Println(string(body))
}

import java.net.http.HttpClient;
import java.net.http.HttpRequest;
import java.net.http.HttpResponse;
import java.net.URI;

public class Main {
    public static void main(String[] args) throws Exception {
        String url = "https://aiboxapi.com/v1/messages";
        String apiKey = System.getenv("API_KEY");

        String payload = """
        {
          "model": "claude-sonnet-4-6",
          "max_tokens": 1024,
          "messages": [
            {
              "role": "user",
              "content": "你好，世界"
            }
          ]
        }
        """;

        HttpClient client = HttpClient.newHttpClient();
        HttpRequest request = HttpRequest.newBuilder()
            .uri(URI.create(url))
            .header("x-api-key", apiKey)
            .header("anthropic-version", "2025-10-01")
            .header("Content-Type", "application/json")
            .POST(HttpRequest.BodyPublishers.ofString(payload))
            .build();

        HttpResponse response = client.send(request,
            HttpResponse.BodyHandlers.ofString());

        System.out.println(response.body());
    }
}

<?php

$url = "https://aiboxapi.com/v1/messages";
$apiKey = getenv('API_KEY');

$payload = [
    "model" => "claude-sonnet-4-6",
    "max_tokens" => 1024,
    "messages" => [
        [
            "role" => "user",
            "content" => "你好，世界"
        ]
    ]
];

$ch = curl_init($url);
curl_setopt($ch, CURLOPT_RETURNTRANSFER, true);
curl_setopt($ch, CURLOPT_POST, true);
curl_setopt($ch, CURLOPT_POSTFIELDS, json_encode($payload));
curl_setopt($ch, CURLOPT_HTTPHEADER, [
    "x-api-key: " . $apiKey,
    "anthropic-version: 2025-10-01",
    "Content-Type: application/json"
]);

$response = curl_exec($ch);
curl_close($ch);

echo $response;
?>

require 'net/http'
require 'json'
require 'uri'

url = URI("https://aiboxapi.com/v1/messages")
api_key = ENV['API_KEY']

payload = {
  model: "claude-sonnet-4-6",
  max_tokens: 1024,
  messages: [
    {
      role: "user",
      content: "你好，世界"
    }
  ]
}

http = Net::HTTP.new(url.host, url.port)
http.use_ssl = true

request = Net::HTTP::Post.new(url)
request["x-api-key"] = api_key
request["anthropic-version"] = "2025-10-01"
request["Content-Type"] = "application/json"
request.body = payload.to_json

response = http.request(request)
puts response.body

import Foundation

let url = URL(string: "https://aiboxapi.com/v1/messages")!
let apiKey = ProcessInfo.processInfo.environment["API_KEY"] ?? ""

let payload: [String: Any] = [
    "model": "claude-sonnet-4-6",
    "max_tokens": 1024,
    "messages": [
        [
            "role": "user",
            "content": "你好，世界"
        ]
    ]
]

var request = URLRequest(url: url)
request.httpMethod = "POST"
request.setValue(apiKey, forHTTPHeaderField: "x-api-key")
request.setValue("2025-10-01", forHTTPHeaderField: "anthropic-version")
request.setValue("application/json", forHTTPHeaderField: "Content-Type")
request.httpBody = try? JSONSerialization.data(withJSONObject: payload)

let task = URLSession.shared.dataTask(with: request) { data, response, error in
    if let error = error {
        print("Error: \(error)")
        return
    }
    
    if let data = data, let responseString = String(data: data, encoding: .utf8) {
        print(responseString)
    }
}

task.resume()

using System;
using System.Net.Http;
using System.Text;
using System.Threading.Tasks;

class Program
{
    static async Task Main(string[] args)
    {
        var url = "https://aiboxapi.com/v1/messages";
        var apiKey = Environment.GetEnvironmentVariable("API_KEY");

        var payload = @"{
            ""model"": ""claude-sonnet-4-6"",
            ""max_tokens"": 1024,
            ""messages"": [
                {
                    ""role"": ""user"",
                    ""content"": ""你好，世界""
                }
            ]
        }";

        using var client = new HttpClient();
        client.DefaultRequestHeaders.Add("x-api-key", apiKey);
        client.DefaultRequestHeaders.Add("anthropic-version", "2025-10-01");

        var content = new StringContent(payload, Encoding.UTF8, "application/json");
        var response = await client.PostAsync(url, content);
        var result = await response.Content.ReadAsStringAsync();

        Console.WriteLine(result);
    }
}

#include <stdio.h>
#include <curl/curl.h>
#include <stdlib.h>

int main(void) {
    CURL *curl;
    CURLcode res;
    const char *api_key = getenv("API_KEY");

    curl_global_init(CURL_GLOBAL_DEFAULT);
    curl = curl_easy_init();

    if(curl) {
        const char *url = "https://aiboxapi.com/v1/messages";
        const char *payload = "{"
            "\"model\":\"claude-sonnet-4-6\","
            "\"max_tokens\":1024,"
            "\"messages\":[{\"role\":\"user\",\"content\":\"你好，世界\"}]"
        "}";

        char auth_header[256];
        snprintf(auth_header, sizeof(auth_header), "x-api-key: %s", api_key);

        struct curl_slist *headers = NULL;
        headers = curl_slist_append(headers, auth_header);
        headers = curl_slist_append(headers, "anthropic-version: 2025-10-01");
        headers = curl_slist_append(headers, "Content-Type: application/json");

        curl_easy_setopt(curl, CURLOPT_URL, url);
        curl_easy_setopt(curl, CURLOPT_POSTFIELDS, payload);
        curl_easy_setopt(curl, CURLOPT_HTTPHEADER, headers);

        res = curl_easy_perform(curl);

        if(res != CURLE_OK) {
            fprintf(stderr, "curl_easy_perform() failed: %s\n",
                    curl_easy_strerror(res));
        }

        curl_slist_free_all(headers);
        curl_easy_cleanup(curl);
    }

    curl_global_cleanup();
    return 0;
}

#import <Foundation/Foundation.h>

int main(int argc, const char * argv[]) {
    @autoreleasepool {
        NSURL *url = [NSURL URLWithString:@"https://aiboxapi.com/v1/messages"];
        NSString *apiKey = [NSProcessInfo processInfo].environment[@"API_KEY"];
        
        NSDictionary *payload = @{
            @"model": @"claude-sonnet-4-6",
            @"max_tokens": @1024,
            @"messages": @[
                @{
                    @"role": @"user",
                    @"content": @"你好，世界"
                }
            ]
        };
        
        NSError *error;
        NSData *jsonData = [NSJSONSerialization dataWithJSONObject:payload
                                                          options:0
                                                            error:&error];
        
        NSMutableURLRequest *request = [NSMutableURLRequest requestWithURL:url];
        [request setHTTPMethod:@"POST"];
        [request setValue:apiKey forHTTPHeaderField:@"x-api-key"];
        [request setValue:@"2025-10-01" forHTTPHeaderField:@"anthropic-version"];
        [request setValue:@"application/json" forHTTPHeaderField:@"Content-Type"];
        [request setHTTPBody:jsonData];
        
        NSURLSessionDataTask *task = [[NSURLSession sharedSession] 
            dataTaskWithRequest:request
            completionHandler:^(NSData *data, NSURLResponse *response, NSError *error) {
                if (error) {
                    NSLog(@"Error: %@", error);
                    return;
                }
                NSString *result = [[NSString alloc] initWithData:data 
                                                        encoding:NSUTF8StringEncoding];
                NSLog(@"%@", result);
            }];
        
        [task resume];
        [[NSRunLoop mainRunLoop] run];
    }
    return 0;
}

(* Requires cohttp and yojson libraries *)
open Lwt
open Cohttp
open Cohttp_lwt_unix

let url = "https://aiboxapi.com/v1/messages"
let api_key = Sys.getenv "API_KEY"

let payload = {|{
  "model": "claude-sonnet-4-6",
  "max_tokens": 1024,
  "messages": [
    {
      "role": "user",
      "content": "你好，世界"
    }
  ]
}|}

let () =
  let headers = Header.init ()
    |> fun h -> Header.add h "x-api-key" api_key
    |> fun h -> Header.add h "anthropic-version" "2025-10-01"
    |> fun h -> Header.add h "Content-Type" "application/json"
  in
  let body = Cohttp_lwt.Body.of_string payload in
  
  let response = Client.post ~headers ~body (Uri.of_string url) >>= fun (resp, body) ->
    body |> Cohttp_lwt.Body.to_string >|= fun body_str ->
    print_endline body_str
  in
  Lwt_main.run response

import 'dart:convert';
import 'dart:io';
import 'package:http/http.dart' as http;

void main() async {
  final url = Uri.parse('https://aiboxapi.com/v1/messages');
  final apiKey = Platform.environment['API_KEY'];
  
  final payload = {
    'model': 'claude-sonnet-4-6',
    'max_tokens': 1024,
    'messages': [
      {
        'role': 'user',
        'content': '你好，世界'
      }
    ]
  };
  
  final response = await http.post(
    url,
    headers: {
      'x-api-key': apiKey!,
      'anthropic-version': '2025-10-01',
      'Content-Type': 'application/json',
    },
    body: jsonEncode(payload),
  );
  
  print(response.body);
}

library(httr)
library(jsonlite)

url <- "https://aiboxapi.com/v1/messages"
api_key <- Sys.getenv("API_KEY")

payload <- list(
  model = "claude-sonnet-4-6",
  max_tokens = 1024,
  messages = list(
    list(
      role = "user",
      content = "你好，世界"
    )
  )
)

response <- POST(
  url,
  add_headers(
    `x-api-key` = api_key,
    `anthropic-version` = "2025-10-01",
    `Content-Type` = "application/json"
  ),
  body = toJSON(payload, auto_unbox = TRUE),
  encode = "raw"
)

cat(content(response, "text"))

{
  "model": "claude-sonnet-4-6",
  "id": "msg_011CdfeHuC728oxqaLrRNbcB",
  "type": "message",
  "role": "assistant",
  "content": [
    {
      "type": "text",
      "text": "你好！我是Claude。很高兴见到你。"
    }
  ],
  "stop_reason": "end_turn",
  "stop_sequence": null,
  "stop_details": null,
  "usage": {
    "input_tokens": 12,
    "cache_creation_input_tokens": 0,
    "cache_read_input_tokens": 0,
    "cache_creation": {
      "ephemeral_5m_input_tokens": 0,
      "ephemeral_1h_input_tokens": 0
    },
    "output_tokens": 18,
    "service_tier": "standard",
    "inference_geo": "global"
  }
}

{
  "error": {
    "code": "model_not_found",
    "message": "model not found (request id: 20260803181903172761862QzohKPXF)",
    "param": "",
    "type": "aibox_error"
  }
}

{
  "error": {
    "code": "",
    "message": "无效的API密钥 (request id: 20260803181903172761862QzohKPXF)",
    "param": "",
    "type": "aibox_error"
  }
}

{
  "error": {
    "code": "",
    "message": "账户余额不足 (request id: 20260803181903172761862QzohKPXF)",
    "param": "",
    "type": "aibox_error"
  }
}

{
  "error": {
    "code": "",
    "message": "请求过于频繁 (request id: 20260803181903172761862QzohKPXF)",
    "param": "",
    "type": "aibox_error"
  }
}

{
  "error": {
    "code": "",
    "message": "服务器内部错误 (request id: 20260803181903172761862QzohKPXF)",
    "param": "",
    "type": "aibox_error"
  }
}

Authorizations

鉴权支持两种方式，任选其一：

x-api-key: YOUR_API_KEY

Authorization: Bearer YOUR_API_KEY

Body

  角色类型

  可选值：`user`（用户消息）、`assistant`（AI回复，用于多轮对话和预填充）

  注：Claude API 的 system 提示词使用单独的 `system` 参数，不在 messages 中

  消息内容

  填写消息的文本内容

单条用户消息示例：

[{"role": "user", "content": "你好，Claude"}]

多轮对话示例：

[
  {"role": "user", "content": "你好"},
  {"role": "assistant", "content": "你好！我是Claude。"},
  {"role": "user", "content": "能解释一下AI吗？"}
]

预填充助手回复：

[
  {"role": "user", "content": "太阳的希腊名称是？(A) Sol (B) Helios (C) Sun"},
  {"role": "assistant", "content": "答案是 ("}
]

{
  "system": "你是一位专业的Python编程导师"
}

{
  "system": [
    {
      "type": "text",
      "text": "你是一位专业的Python编程导师"
    }
  ]
}

{
  "tools": [
    {
      "name": "get_weather",
      "description": "获取指定位置的当前天气",
      "input_schema": {
        "type": "object",
        "properties": {
          "location": {
            "type": "string",
            "description": "城市和省份，例如：北京"
          },
          "unit": {
            "type": "string",
            "enum": ["celsius", "fahrenheit"],
            "description": "温度单位"
          }
        },
        "required": ["location"]
      }
    }
  ]
}

Response

{ "type": "text", "text": "OK" }

{
  "type": "tool_use",
  "id": "toolu_01QgsazxKXSfQVj9Q1XxjYXo",
  "name": "get_weather",
  "input": { "city": "Beijing" },
  "caller": { "type": "direct" }
}

`caller` 是上游新增字段，官方文档尚未收录，解析时忽略即可。

{
  "type": "thinking",
  "thinking": "推理过程文本...",
  "signature": "<约 500+ 字符的签名串>"
}

多轮对话回传 thinking 块时，必须**原样带回** `signature`，否则上游会拒绝。

**不要假设 `content[0]` 就是文本。** 开启 thinking 时 `content[0]` 可能是 thinking 块。应遍历筛选：

```python theme={null}
text = "".join(b.text for b in resp.content if b.type == "text")
```

  输入 token 数

  输出 token 数（**已包含** thinking tokens，计费勿重复累加）

  缓存写入 token

  缓存命中 token

  `{ ephemeral_5m_input_tokens, ephemeral_1h_input_tokens }`

  如 `"standard"`

  如 `"global"`

  开启 thinking 时可能出现：`{ thinking_tokens: int }`

使用示例

基础对话

import anthropic

client = anthropic.Anthropic(
    api_key="YOUR_API_KEY",
    base_url="https://aiboxapi.com"
)

message = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[
        {"role": "user", "content": "解释量子计算的基本原理"}
    ]
)

print(message.content[0].text)

多轮对话

messages = [
    {"role": "user", "content": "什么是机器学习？"},
    {"role": "assistant", "content": "机器学习是人工智能的一个分支..."},
    {"role": "user", "content": "能举个实际应用的例子吗？"}
]

message = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=messages
)

使用系统提示词

message = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    system="你是一位资深的Python开发专家，擅长代码审查和优化建议。",
    messages=[
        {"role": "user", "content": "如何优化这段代码？\n\n[代码]"}
    ]
)

流式响应

with client.messages.stream(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[
        {"role": "user", "content": "写一篇关于AI的短文"}
    ]
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)

工具使用

tools = [
    {
        "name": "get_stock_price",
        "description": "获取股票的实时价格",
        "input_schema": {
            "type": "object",
            "properties": {
                "ticker": {
                    "type": "string",
                    "description": "股票代码，例如：AAPL"
                }
            },
            "required": ["ticker"]
        }
    }
]

message = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    tools=tools,
    messages=[
        {"role": "user", "content": "特斯拉的股价是多少？"}
    ]
)

# 处理工具调用
if message.stop_reason == "tool_use":
    tool_use = next(block for block in message.content if block.type == "tool_use")
    print(f"调用工具: {tool_use.name}")
    print(f"参数: {tool_use.input}")

视觉理解

message = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image",
                    "source": {
                        "type": "url",
                        "url": "https://example.com/image.jpg"
                    }
                },
                {
                    "type": "text",
                    "text": "描述这张图片"
                }
            ]
        }
    ]
)

Base64图像

import base64

with open("image.jpg", "rb") as image_file:
    image_data = base64.b64encode(image_file.read()).decode("utf-8")

message = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[
        {
            "role": "user",
            "content": [
                {
                    "type": "image",
                    "source": {
                        "type": "base64",
                        "media_type": "image/jpeg",
                        "data": image_data
                    }
                },
                {
                    "type": "text",
                    "text": "分析这张图片"
                }
            ]
        }
    ]
)

最佳实践

1. 提示词工程

清晰的角色定义：

system = """你是一位经验丰富的数据科学家，专长包括：
- 统计分析和数据可视化
- 机器学习模型开发
- Python和R编程
请提供专业、准确的建议。"""

结构化输出：

message = "请以JSON格式返回分析结果，包含summary、key_findings和recommendations字段。"

2. 错误处理

from anthropic import APIError, RateLimitError

try:
    message = client.messages.create(
        model="claude-sonnet-4-6",
        max_tokens=1024,
        messages=[{"role": "user", "content": "你好"}]
    )
except RateLimitError:
    print("速率限制，请稍后重试")
except APIError as e:
    print(f"API错误: {e}")

3. Token优化

# 使用更短的提示词
messages = [
    {"role": "user", "content": "总结要点：\n\n[长文本]"}
]

# 限制输出长度
message = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=500,  # 限制输出
    messages=messages
)

4. 预填充响应

# 引导模型以特定格式回复
messages = [
    {"role": "user", "content": "列出5个Python最佳实践"},
    {"role": "assistant", "content": "以下是5个Python最佳实践：\n\n1."}
]

message = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=messages
)

流式响应处理

Python流式示例

import anthropic

client = anthropic.Anthropic(
    api_key="YOUR_API_KEY",
    base_url="https://aiboxapi.com"
)

with client.messages.stream(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[
        {"role": "user", "content": "写一个Python装饰器示例"}
    ]
) as stream:
    for text in stream.text_stream:
        print(text, end="", flush=True)

JavaScript流式示例

import Anthropic from '@anthropic-ai/sdk';

const client = new Anthropic({
  apiKey: process.env.API_KEY,
  baseURL: 'https://aiboxapi.com'
});

const stream = await client.messages.stream({
  model: 'claude-sonnet-4-6',
  max_tokens: 1024,
  messages: [
    { role: 'user', content: '写一个React组件示例' }
  ]
});

for await (const chunk of stream) {
  if (chunk.type === 'content_block_delta' && 
      chunk.delta.type === 'text_delta') {
    process.stdout.write(chunk.delta.text);
  }
}

平台差异与对接注意

响应无包装

POST /v1/messages 成功时直接返回 Anthropic message 对象，没有 {code, data} 外层。官方 SDK、Claude Code、Cline 等才能 1:1 兼容。

错误格式（与官方唯一实质差异）

{
  "error": {
    "code": "model_not_found",
    "message": "... (request id: ...)",
    "param": "",
    "type": "aibox_error"
  }
}

相对 Anthropic 官方：顶层缺少 "type": "error"；error.type 固定为 aibox_error，而非 invalid_request_error 等语义化类型。

对接建议：不要依赖 error.type 做重试分支，改用 HTTP 状态码 + error.code：

报障请提供：error.message 末尾的 request id，以及响应头 x-oneapi-request-id。

流式 SSE

请求加 "stream": true。事件序列与官方一致：

message_start → content_block_start → ping → content_block_delta（多次）→ content_block_stop → message_delta → message_stop

⚠️ 流式与非流式的 usage 结构不同：message_delta.usage 通常只有 4 个 token 字段，没有 cache_creation、service_tier、inference_geo。请分开解析或全部设为可选。

未实现接口

POST /v1/messages/count_tokens 未实现，返回 404。官方 SDK 的 client.messages.count_tokens() 会失败。需要预估 token 时请在本地估算，或读取响应中的 usage.input_tokens。

必须忽略未知字段

本接口对上游透传，Anthropic 可能随时新增字段（如 stop_details、inference_geo、caller、output_tokens_details）。请勿开启严格 schema：

• Go：不要用 DisallowUnknownFields()
• Pydantic：不要 extra="forbid"
• TypeScript / Zod：用 .passthrough() 而非 .strict()

模型名建议

带 -thinking 后缀的同名模型是平台扩展别名。推荐用不带后缀的标准模型名 + 请求体 thinking 参数，便于迁移官方端点。

请求体其它字段与官方一致：model、messages、max_tokens（必填）、system、temperature、top_p、top_k、stop_sequences、stream、tools、tool_choice、thinking、metadata。语义以 Anthropic Messages API 为准。

注意事项

1. API 密钥安全：

   • 使用环境变量存储 API 密钥
   • 不要在代码中硬编码密钥
   • 定期轮换密钥

2. 速率限制：

   • 注意 API 的速率限制
   • 实现重试机制（按 HTTP 状态码）
   • 使用指数退避策略

3. Token 管理：

   • 监控 token 使用量（读 usage）
   • 优化提示词长度
   • 使用适当的 max_tokens 值
   • 开启 thinking 时 output_tokens 已含 thinking，勿重复计费

4. 模型选择：

   • Opus: 复杂任务、需要深度思考
   • Sonnet: 平衡性能和成本
   • Haiku: 快速响应、简单任务

5. 内容解析：

   • 遍历 content 取 type == "text"，不要写死 content[0].text
   • 模型若返回 Markdown 代码块包裹的 JSON，属模型输出而非接口包装（见下方 FAQ）

6. 内容过滤：

   • 验证用户输入
   • 过滤敏感信息
   • 实现内容审核机制

FAQ

响应里 content 的 text 是 ```json ... ``` 代码块，怎么去掉？

这不是接口结构问题。text 字段装的是模型生成的原始内容：模型判断你想要 JSON，就用 Markdown 代码块包起来了。接口不会也不应该改写模型输出。

想拿到干净的结构化数据，有三种正确做法（推荐程度从高到低）：

1. 用 tools 强制结构化输出——最可靠，input 字段直接就是解析好的对象：

{
  "tools": [{
    "name": "emit_result",
    "input_schema": {
      "type": "object",
      "properties": { "answer": { "type": "string" } }
    }
  }],
  "tool_choice": { "type": "tool", "name": "emit_result" }
}

2. prefill 助手消息，让模型从 { 接着写：

{
  "messages": [
    { "role": "user", "content": "..." },
    { "role": "assistant", "content": "{" }
  ]
}

3. 在 system prompt 里明确要求「只输出 JSON，不要加 Markdown 代码块」。

不推荐用正则去剥 code fence——模型偶尔不加围栏时会解析失败。