aiops/backend/pkg/agentcli/llm/anthropic.go

// anthropic.go - Anthropic Messages 协议 LLM 客户端实现 // Logan.Li
// 支持官方 Claude（api.anthropic.com）及智谱 GLM 编码套餐（open.bigmodel.cn/api/anthropic）等
// 所有 Anthropic /v1/messages 兼容端点。与 openai.go 实现同一 Client 接口，由 client.go 按 Protocol 路由。
// 鉴权统一用 Authorization: Bearer（智谱编码套餐实测通过；官方 Claude 较新版本亦支持）+ anthropic-version。
package llm

import (
	"bufio"
	"bytes"
	"context"
	"encoding/json"
	"fmt"
	"io"
	"net/http"
	"strings"
	"time"
)

// AnthropicClient Anthropic Messages 协议客户端 // Logan.Li
type AnthropicClient struct {
	cfg    Config
	client *http.Client
}

// NewAnthropicClient 创建 Anthropic 兼容客户端 // Logan.Li
func NewAnthropicClient(cfg Config) Client {
	timeout := cfg.Timeout
	if timeout <= 0 {
		timeout = 180 * time.Second
	}
	return &AnthropicClient{cfg: cfg, client: &http.Client{Timeout: timeout}}
}

// Name 返回客户端名称 // Logan.Li
func (c *AnthropicClient) Name() string { return "anthropic(" + c.cfg.Model + ")" }

// ---- 请求体结构（Anthropic Messages 格式） // Logan.Li ----

// anthropicMessage Anthropic 消息：content 可为字符串或 block 数组，用 RawMessage 原样嵌入。
type anthropicMessage struct {
	Role    string          `json:"role"`
	Content json.RawMessage `json:"content"`
}

// anthropicThinking 启用 Claude thinking 模式（Claude 3.7 Sonnet 支持） // Logan.Li
type anthropicThinking struct {
	Type         string `json:"type"`
	BudgetTokens int    `json:"budget_tokens"`
}

// anthropicRequest Anthropic /v1/messages 请求体 // Logan.Li
type anthropicRequest struct {
	Model       string             `json:"model"`
	MaxTokens   int                `json:"max_tokens"`
	System      string             `json:"system,omitempty"` // system 抽到顶层（不在 messages 内）
	Messages    []anthropicMessage `json:"messages"`
	Tools       []anthropicTool    `json:"tools,omitempty"`
	Temperature float64            `json:"temperature,omitempty"`
	Stream      bool               `json:"stream,omitempty"`
	Thinking    *anthropicThinking `json:"thinking,omitempty"` // Claude 3.7 Sonnet 推理模式 // Logan.Li
}

// anthropicTool Anthropic 工具定义（input_schema，对应 OpenAI 的 parameters） // Logan.Li
type anthropicTool struct {
	Name        string                 `json:"name"`
	Description string                 `json:"description"`
	InputSchema map[string]interface{} `json:"input_schema"`
}

// 响应 content block：text 或 tool_use（互斥字段） // Logan.Li
type anthropicBlock struct {
	Type  string          `json:"type"` // text | tool_use
	Text  string          `json:"text,omitempty"`
	ID    string          `json:"id,omitempty"`    // tool_use
	Name  string          `json:"name,omitempty"`  // tool_use
	Input json.RawMessage `json:"input,omitempty"` // tool_use
}

// 请求侧 content block（assistant 输出 tool_use / user 输出 tool_result） // Logan.Li
type textBlock struct {
	Type string `json:"type"` // 固定 "text"
	Text string `json:"text"`
}
type toolUseBlock struct {
	Type  string          `json:"type"` // 固定 "tool_use"
	ID    string          `json:"id"`
	Name  string          `json:"name"`
	Input json.RawMessage `json:"input"`
}
type toolResultBlock struct {
	Type      string `json:"type"` // 固定 "tool_result"
	ToolUseID string `json:"tool_use_id"`
	Content   string `json:"content"`
}

// anthropicResponse Anthropic /v1/messages 响应体 // Logan.Li
type anthropicResponse struct {
	ID      string           `json:"id"`
	Model   string           `json:"model"`
	Content []anthropicBlock `json:"content"`
	Usage   struct {
		InputTokens  int `json:"input_tokens"`
		OutputTokens int `json:"output_tokens"`
	} `json:"usage"`
	Error *struct {
		Type    string `json:"type"`
		Message string `json:"message"`
	} `json:"error,omitempty"`
}

// buildMessages 把 OpenAI 风格 Message[] 转成 Anthropic messages + 顶层 system // Logan.Li
// - system role 抽到顶层 system 字段
// - tool role（工具结果）→ user 消息的 tool_result 块
// - assistant 带 tool_calls → content 含 tool_use 块
func (c *AnthropicClient) buildMessages(req *CompletionRequest) (string, []anthropicMessage) {
	var system string
	var msgs []anthropicMessage
	for _, m := range req.Messages {
		switch m.Role {
		case "system":
			if system == "" {
				system = m.Content
			} else {
				system += "\n" + m.Content
			}
		case "tool":
			raw, _ := json.Marshal([]toolResultBlock{{
				Type: "tool_result", ToolUseID: m.ToolCallID, Content: m.Content,
			}})
			msgs = append(msgs, anthropicMessage{Role: "user", Content: raw})
		case "assistant":
			// assistant 消息可能含文本 + tool_use 计划，组装成 content block 数组
			blocks := make([]interface{}, 0, 1+len(m.ToolCalls))
			if m.Content != "" {
				blocks = append(blocks, textBlock{Type: "text", Text: m.Content})
			}
			for _, tc := range m.ToolCalls {
				args := json.RawMessage("{}")
				if strings.TrimSpace(tc.Function.Arguments) != "" {
					args = json.RawMessage(tc.Function.Arguments)
				}
				blocks = append(blocks, toolUseBlock{
					Type: "tool_use", ID: tc.ID, Name: tc.Function.Name, Input: args,
				})
			}
			raw, _ := json.Marshal(blocks)
			msgs = append(msgs, anthropicMessage{Role: "assistant", Content: raw})
		default: // user
			raw, _ := json.Marshal(m.Content)
			msgs = append(msgs, anthropicMessage{Role: "user", Content: raw})
		}
	}
	return system, msgs
}

// buildTools OpenAI Tool[](parameters) → Anthropic [](input_schema) // Logan.Li
func buildTools(tools []Tool) []anthropicTool {
	out := make([]anthropicTool, 0, len(tools))
	for _, t := range tools {
		out = append(out, anthropicTool{
			Name: t.Function.Name, Description: t.Function.Description, InputSchema: t.Function.Parameters,
		})
	}
	return out
}

// isAnthropicThinkingModel 判断是否为支持 thinking 的 Claude 模型 // Logan.Li
func isAnthropicThinkingModel(model string) bool {
	m := strings.ToLower(model)
	return strings.Contains(m, "claude-3-7") || strings.Contains(m, "claude-3.7")
}

// buildRequest 构造 Anthropic 请求体（Complete 与 Stream 共用） // Logan.Li
func (c *AnthropicClient) buildRequest(req *CompletionRequest, stream bool) ([]byte, error) {
	system, msgs := c.buildMessages(req)
	payload := anthropicRequest{
		Model: c.cfg.Model, MaxTokens: c.cfg.MaxTokens, System: system,
		Messages: msgs, Tools: buildTools(req.Tools),
		Temperature: c.cfg.Temperature, Stream: stream,
	}
	if payload.MaxTokens <= 0 {
		payload.MaxTokens = 4096
	}
	// 启用 Claude thinking 模式时，temperature 必须固定为 1，max_tokens 需大于 budget_tokens // Logan.Li
	enabled := req.ReasoningEnabled || c.cfg.ReasoningEnabled
	if enabled && isAnthropicThinkingModel(payload.Model) {
		payload.Temperature = 1
		budget := payload.MaxTokens / 2
		if budget < 1024 {
			budget = 1024
		}
		if payload.MaxTokens <= budget {
			payload.MaxTokens = budget + 1024
		}
		payload.Thinking = &anthropicThinking{Type: "enabled", BudgetTokens: budget}
	}
	body, err := json.Marshal(payload)
	if err != nil {
		return nil, fmt.Errorf("序列化请求失败: %w", err)
	}
	return body, nil
}

// newHTTPRequest 构造 POST /v1/messages 请求（Bearer + anthropic-version） // Logan.Li
func (c *AnthropicClient) newHTTPRequest(ctx context.Context, body []byte) (*http.Request, error) {
	httpReq, err := http.NewRequestWithContext(ctx, http.MethodPost, c.cfg.Endpoint+"/v1/messages", bytes.NewReader(body))
	if err != nil {
		return nil, fmt.Errorf("构造请求失败: %w", err)
	}
	httpReq.Header.Set("Content-Type", "application/json")
	httpReq.Header.Set("Authorization", "Bearer "+c.cfg.APIKey)
	httpReq.Header.Set("anthropic-version", "2023-06-01")
	return httpReq, nil
}

// Complete 一次性补全（Anthropic /v1/messages） // Logan.Li
// 响应 content blocks：text 拼为 Content，tool_use 转为 ToolCalls；usage 映射。
func (c *AnthropicClient) Complete(ctx context.Context, req *CompletionRequest) (*CompletionResponse, error) {
	body, err := c.buildRequest(req, false)
	if err != nil {
		return nil, err
	}
	httpReq, err := c.newHTTPRequest(ctx, body)
	if err != nil {
		return nil, err
	}
	resp, err := c.client.Do(httpReq)
	if err != nil {
		return nil, fmt.Errorf("LLM 连接失败: %w", err)
	}
	defer resp.Body.Close()
	raw, _ := io.ReadAll(resp.Body)
	if resp.StatusCode != http.StatusOK {
		return nil, fmt.Errorf("LLM 返回错误 (HTTP %d): %s", resp.StatusCode, extractErrorMessage(raw))
	}
	var parsed anthropicResponse
	if err := json.Unmarshal(raw, &parsed); err != nil {
		return nil, fmt.Errorf("解析 LLM 响应失败: %w", err)
	}
	if parsed.Error != nil && parsed.Error.Message != "" {
		return nil, fmt.Errorf("LLM 错误: %s", parsed.Error.Message)
	}

	var textBuf strings.Builder
	var toolCalls []ToolCall
	for _, b := range parsed.Content {
		switch b.Type {
		case "text", "thinking":
			textBuf.WriteString(b.Text)
		case "tool_use":
			args := string(b.Input)
			if strings.TrimSpace(args) == "" {
				args = "{}"
			}
			toolCalls = append(toolCalls, ToolCall{
				ID: b.ID, Type: "function",
				Function: FunctionCall{Name: b.Name, Arguments: args},
			})
		}
	}
	return &CompletionResponse{
		ID:    parsed.ID,
		Model: parsed.Model,
		Message: Message{
			Role:      "assistant",
			Content:   textBuf.String(),
			ToolCalls: toolCalls,
		},
		Usage: Usage{
			PromptTokens:     parsed.Usage.InputTokens,
			CompletionTokens: parsed.Usage.OutputTokens,
			TotalTokens:      parsed.Usage.InputTokens + parsed.Usage.OutputTokens,
		},
	}, nil
}

// Stream 流式补全（Anthropic SSE） // Logan.Li
// 逐 token 回调 text_delta；usage 从 message_start(input) + message_delta(output) 取。
// 注：流式不组装 tool_use，仅用于智能助手纯对话；Agent CLI ReAct 循环必须使用 Complete。
func (c *AnthropicClient) Stream(ctx context.Context, req *CompletionRequest, onDelta func(string)) (*CompletionResponse, error) {
	body, err := c.buildRequest(req, true)
	if err != nil {
		return nil, err
	}
	httpReq, err := c.newHTTPRequest(ctx, body)
	if err != nil {
		return nil, err
	}
	httpReq.Header.Set("Accept", "text/event-stream")

	streamClient := &http.Client{Timeout: 0} // 流式由 ctx 控制超时
	resp, err := streamClient.Do(httpReq)
	if err != nil {
		return nil, fmt.Errorf("LLM 连接失败: %w", err)
	}
	defer resp.Body.Close()
	if resp.StatusCode != http.StatusOK {
		raw, _ := io.ReadAll(resp.Body)
		return nil, fmt.Errorf("LLM 返回错误 (HTTP %d): %s", resp.StatusCode, extractErrorMessage(raw))
	}

	var contentBuilder strings.Builder
	var usage Usage
	var eventType string
	reader := bufio.NewReader(resp.Body)
	for {
		line, readErr := reader.ReadString('\n')
		trimmed := strings.TrimSpace(line)
		switch {
		case strings.HasPrefix(trimmed, "event:"):
			eventType = strings.TrimSpace(strings.TrimPrefix(trimmed, "event:"))
		case strings.HasPrefix(trimmed, "data:"):
			data := strings.TrimSpace(strings.TrimPrefix(trimmed, "data:"))
			if data != "" && data != "[DONE]" {
				switch eventType {
				case "content_block_delta":
					var d struct {
						Delta struct {
							Type string `json:"type"`
							Text string `json:"text"`
						} `json:"delta"`
					}
					if json.Unmarshal([]byte(data), &d) == nil && d.Delta.Type == "text_delta" && d.Delta.Text != "" {
						contentBuilder.WriteString(d.Delta.Text)
						if onDelta != nil {
							onDelta(d.Delta.Text)
						}
					}
				case "message_start":
					var d struct {
						Message struct {
							Usage struct {
								InputTokens int `json:"input_tokens"`
							} `json:"usage"`
						} `json:"message"`
					}
					if json.Unmarshal([]byte(data), &d) == nil {
						usage.PromptTokens = d.Message.Usage.InputTokens
					}
				case "message_delta":
					var d struct {
						Usage struct {
							OutputTokens int `json:"output_tokens"`
						} `json:"usage"`
					}
					if json.Unmarshal([]byte(data), &d) == nil {
						usage.CompletionTokens = d.Usage.OutputTokens
						usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
					}
				}
			}
			eventType = "" // 消费完 data 重置，防默认事件串用
		}
		if readErr != nil {
			break
		}
	}
	return &CompletionResponse{
		Model:  c.cfg.Model,
		Message: Message{Role: "assistant", Content: contentBuilder.String()},
		Usage:  usage,
	}, nil
}