fix: 修复 AI 回复无法送达发送者 + 重复消息 + action角色泄露 + OS环境支持

广播逻辑重构:
- AI 回复 (stream_start/response/stream_segments/multi_message/stream_end) 改用 broadcastToUser 发送给所有客户端
- 用户消息回显保持 broadcastToUserExcept 排除发送者

消息去重与角色修复:
- CacheMessage(user) 移至回复生成后，避免本轮 LLM 调用出现重复用户消息
- action 角色消息在 DB 存储时映射为 assistant，DeepSeek 等模型不支持自定义角色
- stream_end defer 机制确保错误路径也会终止客户端思考指示器

OS 完整环境支持:
- host 包重构为 HostBackend 接口 + Direct/WSL/Docker 三种后端
- 新增 os_exec/os_file/os_system 工具供 AI 在完整 Linux 环境中自由操作

其他:
- 视觉模型注入 + 图片预处理后清空 Images 避免传给 Chat 模型
- 图片 URL 相对路径→绝对 URL 转换
- DevTools 链路追踪页面 + 重启修复
- 记忆搜索模糊匹配增强
- 后台思考定时调度支持
- 管理后台页面 (模型配置/用户管理等)
- docs/api 更新广播机制说明

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

backend/ai-core/internal/llm/openai.go

@@ -4,15 +4,16 @@ import (
 	"bufio"
 	"bytes"
 	"context"
+	"encoding/base64"
 	"encoding/json"
 	"fmt"
 	"io"
-	"github.com/yourname/cyrene-ai/pkg/logger"
 	"net/http"
 	"strings"
 	"time"
 
 	"github.com/yourname/cyrene-ai/ai-core/internal/model"
+	"github.com/yourname/cyrene-ai/pkg/logger"
 )
 
 // OpenAIConfig OpenAI适配器配置
@@ -267,12 +268,13 @@ func (p *OpenAIProvider) doChat(ctx context.Context, messages []model.LLMMessage
 		LogCall(r)
 	}()
 
-	// 转换消息格式
+	// 转换消息格式（先解析图片 URL 为 data URL）
 	oaiMessages := make([]openAIMessage, len(messages))
 	for i, msg := range messages {
+		resolvedImages := p.resolveImages(msg.Images)
 		oaiMsg := openAIMessage{
 			Role:             string(msg.Role),
-			Content:          buildContent(msg.Content, msg.Images),
+			Content:          buildContent(msg.Content, resolvedImages),
 			Name:             msg.Name,
 			ToolCallID:       msg.ToolCallID,
 			ReasoningContent: msg.ReasoningContent,
@@ -377,9 +379,10 @@ func (p *OpenAIProvider) doChat(ctx context.Context, messages []model.LLMMessage
 func (p *OpenAIProvider) doChatStream(ctx context.Context, messages []model.LLMMessage, modelName string, tools []OpenAITool) (*http.Response, error) {
 	oaiMessages := make([]openAIMessage, len(messages))
 	for i, msg := range messages {
+		resolvedImages := p.resolveImages(msg.Images)
 		oaiMsg := openAIMessage{
 			Role:             string(msg.Role),
-			Content:          buildContent(msg.Content, msg.Images),
+			Content:          buildContent(msg.Content, resolvedImages),
 			Name:             msg.Name,
 			ToolCallID:       msg.ToolCallID,
 			ReasoningContent: msg.ReasoningContent,
@@ -455,6 +458,67 @@ func contentString(v interface{}) string {
 	return ""
 }
 
+// resolveImages converts non-data URLs to base64 data URLs so external LLM APIs can access them.
+func (p *OpenAIProvider) resolveImages(images []string) []string {
+	if len(images) == 0 {
+		return images
+	}
+	resolved := make([]string, 0, len(images))
+	for _, img := range images {
+		if strings.HasPrefix(img, "data:") {
+			resolved = append(resolved, img)
+			continue
+		}
+		dataURL, err := p.downloadAsDataURL(img)
+		if err != nil {
+			logger.Printf("[openai] 图片下载失败, 保留原始 URL: %s, err=%v", img, err)
+			resolved = append(resolved, img) // 保留原始 URL 作为 fallback
+			continue
+		}
+		resolved = append(resolved, dataURL)
+	}
+	return resolved
+}
+
+// downloadAsDataURL downloads an image from a URL and returns it as a base64 data URL.
+func (p *OpenAIProvider) downloadAsDataURL(url string) (string, error) {
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	req, err := http.NewRequestWithContext(ctx, "GET", url, nil)
+	if err != nil {
+		return "", fmt.Errorf("创建请求失败: %w", err)
+	}
+
+	resp, err := p.httpClient.Do(req)
+	if err != nil {
+		return "", fmt.Errorf("下载失败: %w", err)
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode != http.StatusOK {
+		return "", fmt.Errorf("HTTP %d", resp.StatusCode)
+	}
+
+	// 限制最大 20MB
+	const maxSize = 20 * 1024 * 1024
+	body, err := io.ReadAll(io.LimitReader(resp.Body, maxSize+1))
+	if err != nil {
+		return "", fmt.Errorf("读取失败: %w", err)
+	}
+	if len(body) > maxSize {
+		return "", fmt.Errorf("图片过大: %d bytes", len(body))
+	}
+
+	mimeType := resp.Header.Get("Content-Type")
+	if mimeType == "" {
+		mimeType = http.DetectContentType(body)
+	}
+
+	b64 := base64.StdEncoding.EncodeToString(body)
+	return fmt.Sprintf("data:%s;base64,%s", mimeType, b64), nil
+}
+
 // buildContent converts text + optional images to API content format.
 // Returns a plain string if no images, or a multimodal array otherwise.
 func buildContent(text string, images []string) interface{} {

backend/ai-core/internal/llm/selector.go

@@ -19,6 +19,7 @@ const (
 	PurposeToolCalling      ModelPurpose = "tool_calling"
 	PurposeMemoryExtraction ModelPurpose = "memory_extraction"
 	PurposeVision           ModelPurpose = "vision"
+	PurposeOCR              ModelPurpose = "ocr"
 )
 
 // ErrModelNotRequired is returned when an optional model is unavailable.