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/subsession/memory_provider.go

@@ -2,12 +2,15 @@ package subsession
 
 import (
 	"context"
+	"encoding/json"
 	"fmt"
-	"github.com/yourname/cyrene-ai/pkg/logger"
+	"strings"
 	"time"
 
+	"github.com/yourname/cyrene-ai/ai-core/internal/llm"
 	"github.com/yourname/cyrene-ai/ai-core/internal/memory"
 	"github.com/yourname/cyrene-ai/ai-core/internal/model"
+	"github.com/yourname/cyrene-ai/pkg/logger"
 )
 
 // MemoryRetriever 记忆检索接口
@@ -16,9 +19,12 @@ type MemoryRetriever interface {
 }
 
 // MemoryProvider 记忆检索子会话提供者
-// 职责：检索与当前对话相关的用户记忆，排序去重，返回结构化摘要
+// 职责：检索与当前对话相关的用户记忆，排序去重，返回结构化摘要。
+// 支持 LLM 驱动的模糊关键词扩展搜索。
 type MemoryProvider struct {
-	retriever MemoryRetriever
+	retriever  MemoryRetriever
+	llmAdapter *llm.Adapter
+	memClient  *memory.Client
 }
 
 // NewMemoryProvider 创建记忆检索子会话提供者
@@ -28,6 +34,12 @@ func NewMemoryProvider(retriever MemoryRetriever) *MemoryProvider {
 	}
 }
 
+// SetFuzzySearch enables LLM-driven fuzzy keyword expansion for broader memory retrieval.
+func (p *MemoryProvider) SetFuzzySearch(llmAdapter *llm.Adapter, memClient *memory.Client) {
+	p.llmAdapter = llmAdapter
+	p.memClient = memClient
+}
+
 func (p *MemoryProvider) Type() model.SubSessionType {
 	return model.SubSessionMemory
 }
@@ -93,6 +105,7 @@ func (p *MemoryProvider) Execute(ctx context.Context, subCtx []model.LLMMessage)
 		return result, nil
 	}
 
+	// Phase 1: exact/keyword retrieval
 	memories, err := p.retriever.Retrieve(ctx, userID, userMessage)
 	if err != nil {
 		logger.Printf("[memory-subsession] 记忆检索失败: %v", err)
@@ -101,6 +114,20 @@ func (p *MemoryProvider) Execute(ctx context.Context, subCtx []model.LLMMessage)
 		return result, nil
 	}
 
+	seen := make(map[string]bool)
+	for _, m := range memories {
+		seen[m.ID] = true
+	}
+
+	// Phase 2: LLM-driven fuzzy keyword expansion + semantic search
+	fuzzyMemories := p.fuzzySearch(ctx, userID, userMessage)
+	for _, m := range fuzzyMemories {
+		if !seen[m.ID] {
+			seen[m.ID] = true
+			memories = append(memories, m)
+		}
+	}
+
 	// 转换为 MemorySnippet
 	snippets := make([]model.MemorySnippet, 0, len(memories))
 	for _, m := range memories {
@@ -117,7 +144,7 @@ func (p *MemoryProvider) Execute(ctx context.Context, subCtx []model.LLMMessage)
 	if len(snippets) == 0 {
 		result.Summary = "（没有找到相关记忆）"
 	} else {
-		result.Summary = fmt.Sprintf("检索到 %d 条相关记忆", len(snippets))
+		result.Summary = fmt.Sprintf("检索到 %d 条相关记忆（含模糊匹配）", len(snippets))
 		// 按重要性列出前几条
 		topCount := len(snippets)
 		if topCount > 3 {
@@ -138,6 +165,74 @@ func (p *MemoryProvider) Execute(ctx context.Context, subCtx []model.LLMMessage)
 	}
 
 	result.Memories = snippets
-	logger.Printf("[memory-subsession] 完成: %s", result.Summary)
+	logger.Printf("[memory-subsession] 完成: %s (精确=%d, 模糊=%d)", result.Summary, len(memories)-len(fuzzyMemories), len(fuzzyMemories))
 	return result, nil
 }
+
+// fuzzySearch expands the user message into fuzzy keywords via LLM and performs semantic search.
+func (p *MemoryProvider) fuzzySearch(ctx context.Context, userID, userMessage string) []memory.MemoryEntry {
+	if p.llmAdapter == nil || p.memClient == nil {
+		return nil
+	}
+
+	keywords := p.expandKeywords(ctx, userMessage)
+	if len(keywords) == 0 {
+		return nil
+	}
+
+	logger.Printf("[memory-subsession] 模糊关键词: %v", keywords)
+
+	var allResults []memory.MemoryEntry
+	seen := make(map[string]bool)
+
+	for _, kw := range keywords {
+		results, err := p.memClient.QueryByText(ctx, userID, kw, "", 0, 5)
+		if err != nil {
+			logger.Printf("[memory-subsession] 模糊搜索 '%s' 失败: %v", kw, err)
+			continue
+		}
+		for _, m := range results {
+			if !seen[m.ID] {
+				seen[m.ID] = true
+				allResults = append(allResults, m)
+			}
+		}
+	}
+
+	return allResults
+}
+
+// expandKeywords uses LLM to generate fuzzy/related search keywords from the user message.
+func (p *MemoryProvider) expandKeywords(ctx context.Context, message string) []string {
+	prompt := fmt.Sprintf(
+		"从以下对话消息中提取 3-5 个可用于模糊搜索记忆的关键词。这些关键词应该是：\n"+
+			"- 与话题相关的抽象概念\n- 同义词和相关词\n- 更宽泛或更具体的相关概念\n"+
+			"- 不要包含消息中已经出现的原词\n\n"+
+			"用户消息：「%s」\n\n"+
+			"只输出 JSON 字符串数组，例如：[\"关键词1\",\"关键词2\"]", message)
+
+	resp, err := p.llmAdapter.Chat(ctx, []model.LLMMessage{
+		{Role: model.RoleSystem, Content: "你是记忆搜索专家。输出 JSON 字符串数组。"},
+		{Role: model.RoleUser, Content: prompt},
+	})
+	if err != nil {
+		logger.Printf("[memory-subsession] 关键词扩展失败: %v", err)
+		return nil
+	}
+
+	text := strings.TrimSpace(resp.Content)
+	// Extract JSON array
+	if idx := strings.Index(text, "["); idx >= 0 {
+		if end := strings.LastIndex(text, "]"); end > idx {
+			text = text[idx : end+1]
+		}
+	}
+
+	var keywords []string
+	if err := json.Unmarshal([]byte(text), &keywords); err != nil {
+		logger.Printf("[memory-subsession] 解析关键词 JSON 失败: %v (raw=%s)", err, resp.Content)
+		return nil
+	}
+
+	return keywords
+}