memoryweave/go/internal/api/server.go

// HTTP 服务器 — 多 Agent 架构（FileGraph + NetworkEventBus + 跨Agent缓存失效）
package api

import (
	"bytes"
	"context"
	"encoding/json"
	"fmt"
	"io"
	"log"
	"net/http"
	"os"
	"path/filepath"
	"runtime"
	"sort"
	"strconv"
	"strings"
	"time"

	"github.com/xiaoxue/memoryweave/internal/api/middleware"
	"github.com/xiaoxue/memoryweave/internal/api/routes"
	"github.com/xiaoxue/memoryweave/internal/distill"
	"github.com/xiaoxue/memoryweave/internal/governance"
	"github.com/xiaoxue/memoryweave/internal/metrics"
	"github.com/xiaoxue/memoryweave/internal/models"
	"github.com/xiaoxue/memoryweave/internal/redact"
	"github.com/xiaoxue/memoryweave/internal/selfoptimize"
	"github.com/xiaoxue/memoryweave/internal/storage"
)

// cachedGraphStoreRef 全局图谱缓存引用（用于 graph write 操作时主动失效）
var cachedGraphStoreRef *storage.GraphCacheRef

// skillByNameHandler GET/DELETE /api/v1/skills/{name} 的 method 分支分发器
func skillByNameHandler(w http.ResponseWriter, r *http.Request) {
	switch r.Method {
	case http.MethodGet:
		routes.Skills.Get(w, r)
	case http.MethodDelete:
		routes.Skills.Delete(w, r)
	default:
		http.Error(w, "method not allowed", http.StatusMethodNotAllowed)
	}
}

func NewServer() http.Handler {
	mux := http.NewServeMux()

	// ─── 初始化依赖 ──────────────────────────────
	emb := storage.NewEmbedder(os.Getenv("VLLM_ENDPOINT"))
	rerank := storage.NewReranker(os.Getenv("RERANK_ENDPOINT"), emb)

	// 存储后端
	ldb := initStorageBackend(os.Getenv("STORAGE_BACKEND"), emb)

	// 召回元数据「累积延迟更新」缓冲（2026-09-12 优化B）
	// 读路径不再逐条写 LanceDB（MVCC 每写一行一个版本 → _versions 膨胀根因），
	// 改为窗口内合并 + 单事务批量提交（每批 1 个版本）。
	storage.InitRecallWriteBuffer(ldb)

	// 软删除累积延迟更新缓冲（2026-09-12 优化C）
	// 软删除操作不再逐条写 LanceDB（每写一行一个版本 → _versions 膨胀），
	// 改为窗口内合并 + 单事务批量提交（每批 1 个版本）。
	storage.InitSoftDeleteWriteBuffer(ldb)

	// 启动时数据目录一致性检查（防止路径混乱导致读取废弃数据）
	runStartupChecks(os.Getenv("STORAGE_BACKEND"))

	// G7.2: 注册 LDB getter（供 skill crystallize 使用）
	routes.RegisterLDBGetter(func() storage.LanceDB { return ldb })

	// 冲突检测器（E4.1: Commit 时矛盾检测依赖此实例）
	// 2026-09-20：改用带持久化的构造 —— 此前 active 是纯内存 map，zhiyid 重启即丢
	// （实测 17:17 重启后 pending 冲突全部消失）。落盘失败不致命，降级为内存模式并记日志。
	conflictDetector, cdErr := governance.NewPersistentConflictDetector("/var/lib/memoryweave/conflicts.json")
	if cdErr != nil {
		log.Printf("[conflict] 持久化加载失败，降级为内存模式: %v", cdErr)
	}
	// 注册全局实例，供 selfoptimize 触发器执行器读取（TConflict 判据）
	governance.SetGlobalConflictDetector(conflictDetector)

	// 图谱（带 Navigate 结果缓存：TTL 5min，容量 500，热点实体缓存命中加速）
	graphStore, cachedGraphStoreRef := storage.NewCachedGraphStore(initGraphStore())

	api := routes.NewAPI(ldb, emb, rerank, conflictDetector, graphStore)

	// G1: Recall 管线挂图谱扩展
	api.Pipeline.SetGraphExpander(graphStore)

	// 启动时初始化 Prometheus 指标
	go func() {
		time.Sleep(1 * time.Second) // 等待 Rust sidecar 就绪
		if stats, err := ldb.Stats(); err == nil {
			if total, ok := stats["total_memories"].(float64); ok {
				metrics.TotalMemories.Set(total)
			}
			if total, ok := stats["total_episodes"].(float64); ok {
				metrics.TotalEpisodes.Set(total)
			}
			log.Printf("[metrics] 初始化指标: memories=%.0f", stats["total_memories"])
		}
	}()

	// 图谱（带 Navigate 结果缓存：TTL 5min，容量 500，热点实体缓存命中加速）
	// 已在上面初始化
	graphUpdater := governance.NewAutoGraphUpdater(graphStore)
	graphAPI := routes.NewGraphAPI(graphStore)

	// G4: 自动蒸馏 → 注入图谱更新器
	routes.SetGraphUpdater(graphUpdater)

	// 冲突
	conflictAPI := routes.NewConflictAPI(conflictDetector)

	// ─── 缺口
	gapDetector := selfoptimize.NewGapDetector(emb, ldb)
	gapDetector.EnableGapRedisPersistence()
	selfoptimize.SetGlobalGapDetector(gapDetector) // 注册全局实例供 recall 回调使用

	// 因果追踪持久化
	routes.CascadeR.Tracker().EnableCausalRedisPersistence()
	gapAPI := routes.NewGapAPI(gapDetector)
	routes.InitGapRepair(gapDetector, ldb) // 共享同一个 GapDetector + LanceDB
	// G1a: 挂缺隙记录器（0 结果 → 自动记录 miss）
	api.Pipeline.SetMissRecorder(func(query, namespace string) {
		gapDetector.RecordMiss(query + " [" + namespace + "]")
	})

	feedbackAPI := routes.NewFeedbackAPI(ldb)
	// 遗忘器：支持按 Agent 类型设置衰减率
	forgetter := governance.NewForgetter()
	adminAPI := routes.NewAdminAPI(ldb, forgetter, graphStore)
	agentRegistry := routes.NewAgentRegistry(nil)
	evalAPI := routes.NewEvalAPI(api.Pipeline, ldb)
	l3API := routes.WM
	consolPipe := routes.NewConsolidationPipeline(ldb, graphStore, conflictDetector)
	consolPipe.SetDataDir("/var/lib/memoryweave/lancedb", "/var/lib/memoryweave/graph.db")

	// Obsidian
	obsidian := routes.NewObsidianSyncer("", ldb)

	// ─── 蒸馏引擎 ───────────────────────────────
	llmEndpoint := os.Getenv("LLM_ENDPOINT")
	llmModel := os.Getenv("LLM_MODEL")
	if llmModel == "" {
		llmModel = "deepseek/deepseek-v4-pro"
	}
	llmAPIKey := os.Getenv("LLM_API_KEY")
	if llmAPIKey == "" {
		llmAPIKey = os.Getenv("API_KEY")
	}
	distillEngine := distill.NewEngine(llmEndpoint, llmModel, llmAPIKey)
	routes.DistillEngineRef = distillEngine
	// 蒸馏完成 → 自动图谱更新 + 冲突检测 + 被动验证 + 回写内存
	distill.OnDistillComplete = func(input distill.DistillInput, result distill.DistillResult) {
		entityNames := make([]string, len(result.Entities))
		for i, e := range result.Entities {
			entityNames[i] = e.Name
		}
		graphUpdater.UpdateFromDistill(&governance.DistillInput{
			Content: input.Content, Facts: result.Facts,
			Entities: entityNames, Namespace: input.Namespace,
			EpisodeID: input.EpisodeID,
		})

		// 回写蒸馏产物到记忆库（LanceDB）
		// 2026-09-07 方案B 质量门槛: 碎片(<20字)/噪声句(状态汇报/桥接/reflection)不入库,
		// 与已有记忆完全重复跳过(蒸馏直写不走 commit exact_dup, 需本层查重)。
		for _, fact := range result.Facts {
			// 2026-09-10 P0：直写路径同样脱敏（这里绕过 FilterQualityFacts，
			// 事故里泄漏进镜像的记忆正是从这条路径入库的）。
			fact = redact.RedactSecrets(fact)
			if !distill.IsQualityFact(fact) {
				continue // 碎片/噪声过滤
			}
			// 查重: 同 namespace 已存在**语义等价**记忆 → 跳过
			// 必须用 cosine 近重复判据（≥0.98），不能只比字符串相等——
			// 否则 "牧尘在招商局工作" 与 "牧尘在招商局工作。"（多一个句号）
			// 会被判为两条不同事实双双入库，造成近重复堆积。
			var skip bool
			if qvec, eErr := emb.EncodeSingle(fact); eErr == nil {
				if sim, _ := ldb.Search("memories", qvec, 5, input.Namespace); len(sim) > 0 {
					for _, ex := range sim {
						if ex.Content == fact || cosineSimilarity(qvec, ex.Vector) >= 0.98 {
							skip = true
							break
						}
					}
				}
			}
			if skip {
				continue
			}
			mem := models.MemoryRecord{
				ID:           fmt.Sprintf("mem_%d", time.Now().UnixNano()),
				AgentID:      input.AgentID,
				Namespace:    input.Namespace,
				Content:      fact,
				Category:     "distilled",
				Tier:         "normal",
				QualityScore: result.Overall,
				Version:      1,
				CreatedAt:    time.Now(),
				UpdatedAt:    time.Now(),
				DerivedFrom:  input.EpisodeID,
				Freshness:    "fresh",
				Source:       "llm_distill", // 2026-09-07 方案C provenance: LLM蒸馏推断来源
			}
			if err := ldb.InsertMemory(mem); err != nil {
				log.Printf("[distill] commit fact failed: %v", err)
			}
		}
		// 冲突检测：加载已有记忆进行比较
		// 必须用**新内容的真实向量**做语义检索取近邻；用零向量在语义索引上
		// 返回的只是"范数最小"的任意行，真正的矛盾对象不在比对集里 → 永远检不出冲突。
		existingMems := []models.MemoryRecord{}
		if qv, eErr := emb.EncodeSingle(input.Content); eErr == nil {
			if found, _ := ldb.Search("memories", qv, 100, input.Namespace); len(found) > 0 {
				existingMems = found
			}
		}
		if len(existingMems) == 0 {
			zeroVec := make([]float32, 1024)
			existingMems, _ = ldb.Search("memories", zeroVec, 100, input.Namespace)
		}
		conflictExisting := make([]map[string]interface{}, 0)
		for _, m := range existingMems {
			conflictExisting = append(conflictExisting, map[string]interface{}{
				"content":  m.Content,
				"entities": entityNames, // 用当前内容提取的实体做交叉比较
			})
		}
		for _, c := range conflictDetector.Scan(input.Content, entityNames, conflictExisting) {
			if c.Strategy == "latest_wins" {
				conflictDetector.AutoResolve(c)
			}
		}
		// 被动验证：使用已加载的记忆进行 P1/P2/P3 匹配
		validMems := make([]selfoptimize.MemoryForValidation, len(existingMems))
		for i, m := range existingMems {
			validMems[i] = selfoptimize.MemoryForValidation{
				ID: m.ID, Content: m.Content, QualityScore: m.QualityScore,
			}
		}
		validated := selfoptimize.Validator.Validate(input.Content, validMems)
		// 被动验证结果写回 quality_score
		// 修复（2026-09-20）：原先用 GetRecords() 遍历【全部历史记录】并无条件写回 ——
		// 每来一次输入就重写全库（≈6000 条），每次 Update 新增一个 LanceDB 版本快照
		// （manifest 引用全部 data 文件 ⇒ O(n²)），导致版本数每日 ~1.9 万、占用每日涨十几 G。
		// 现改为：只写【本次真正匹配】且【Confidence 实际发生变化】的记录。
		for _, r := range validated {
			if r.Confidence > 0 && r.Changed {
				_ = ldb.Update("memories", r.MemoryID, map[string]any{
					"quality_score": r.Confidence,
				})
			}
		}
	}

	// ─── CO_OCCURS 追踪器（持久化到 Redis）───
	storage.CoOccurTrackerInstance = storage.NewCoOccurTracker(storage.RedisCoOccurPersist())
	storage.LoadCoOccurFromRedis()

	// ─── Dashboard 持久化到 Redis（restart 不丢指标）───
	selfoptimize.EnableRedisPersistence()

	// ─── VProp 持久化到 Redis（restart 不丢）───
	selfoptimize.VProp.EnableVPropRedisPersistence()

	// ─── Skill 持久化到 Redis（G7: restart 不丢 skills）───
	routes.BayesianSkills.EnableRedisPersistence()

	// ─── V 值传播器（竞争性架构核心）─────────────
	vPropagator := selfoptimize.VProp

	// 注入 V 值查询函数到存储层（召回排序融合）
	storage.VValueProvider = func(memoryID string) float64 {
		return vPropagator.GetMemVValue(memoryID)
	}
	// 在蒸馏完成回调中记录 V 值
	originalOnComplete := distill.OnDistillComplete
	distill.OnDistillComplete = func(input distill.DistillInput, result distill.DistillResult) {
		// 调用已有的回调
		if originalOnComplete != nil {
			originalOnComplete(input, result)
		}
		// 记录 V 值决策：蒸馏成功 = 有用
		if len(result.Facts) > 0 {
			vPropagator.RecordDecision(
				"distill_"+input.EpisodeID,
				[]string{input.EpisodeID},
				"distill", "success", "",
			)
		}
		// 更新仪表盘指标
		selfoptimize.Dash.RecordUseful()
		// 记录蒸馏质量损失（avg_distill_loss = 1 - Overall）
		// 仅记录真实蒸馏：Overall > 0 表示 LLM 成功返回，过滤 fallback (Overall=0)
		if result.Overall > 0 {
			loss := 1.0 - result.Overall
			if loss < 0 {
				loss = 0
			}
			if loss > 1 {
				loss = 1
			}
			selfoptimize.Dash.RecordDistillLoss(loss)
		}
	}

	// ─── 跨 Agent 缓存失效回调 ───────────────────
	// 收到其他 Agent 的广播 → 失效本地缓存
	governance.GlobalEventBus.Subscribe("cache.invalidate", "self")

	// ─── 路由注册 ──────────────────────────────

	// ─── 静态文件服务（Web UI）─────────────────────
	// webUIRoot 在函数顶部声明
	var webUIRoot string
	// 后续在 catch-all 中处理（见文件末尾）

	mux.HandleFunc("/health", routes.HandleHealth)
	mux.HandleFunc("/api/v1/health", routes.HandleHealth)
	// 图谱缓存统计
	mux.HandleFunc("/api/v1/cache/stats", func(w http.ResponseWriter, r *http.Request) {
		if cachedGraphStoreRef != nil {
			respondJSON(w, 200, map[string]interface{}{
				"graph_cache": cachedGraphStoreRef.Stats(),
			})
		} else {
			respondJSON(w, 200, map[string]interface{}{"graph_cache": nil})
		}
	})
	// ─── Prometheus /metrics ─────────────────────
	mux.Handle("/metrics", metrics.Handler())

	// 核心 API
	mux.HandleFunc("/api/v1/commit", func(w http.ResponseWriter, r *http.Request) {
		api.Commit(w, r)
		// 跨 Agent 广播 + 自动蒸馏
		go func() {
			var req struct {
				AgentID   string `json:"agent_id"`
				Namespace string `json:"namespace"`
				Content   string `json:"content"`
			}
			if r.Body != nil {
				body, _ := io.ReadAll(r.Body)
				json.Unmarshal(body, &req)
				r.Body = io.NopCloser(bytes.NewReader(body))
			}
			if req.Namespace != "" {
				// 通知其他 Agent：缓存失效 + 新记忆事件
				governance.PushCacheInvalidate(req.Namespace, "commit")
				governance.PushMemoryCommitted(req.AgentID, req.Namespace, "")
			}
		}()
	})
	mux.HandleFunc("/api/v1/recall", api.Recall)
	mux.HandleFunc("/api/v1/recall/debug", api.RecallDebug)
	mux.HandleFunc("/api/v1/feedback", api.Feedback)
	mux.HandleFunc("/api/v1/bootstrap", api.Bootstrap)
	mux.HandleFunc("/api/v1/stats", api.Stats)
	mux.HandleFunc("/api/v1/batch-commit", api.BatchCommit)
	// 自优化指标（Dashboard，含 avg_distill_loss）
	mux.HandleFunc("/api/v1/metrics", func(w http.ResponseWriter, r *http.Request) {
		m := selfoptimize.Dash.Metrics()
		if stats, err := ldb.Stats(); err == nil {
			if total, ok := stats["total_memories"].(float64); ok {
				m["total_memories"] = total
			}
			if total, ok := stats["total_episodes"].(float64); ok {
				m["total_episodes"] = total
			}
		}
		respondJSON(w, 200, m)
	})
	// ─── WebSocket 实时推送 ──────────────────────
	mux.HandleFunc("/api/v1/ws/", routes.WSHandler)
	// 社区发现（Louvain + 社区摘要）
	mux.HandleFunc("/api/v1/community/summarize", api.HandleCommunitySummarize)
	// 深度推理（人侧数据 → 三类清单）
	mux.HandleFunc("/api/v1/deep-think", api.HandleDeepThink)

	mux.HandleFunc("/api/v1/ws", func(w http.ResponseWriter, r *http.Request) {
		w.Header().Set("Content-Type", "application/json")
		w.Write([]byte(`{"status":"ok","protocol":"websocket","endpoint":"/api/v1/ws/{agent_id}"}`))
	})

	// 知识图谱
	mux.HandleFunc("/api/v1/graph/stats", graphAPI.Stats)
	mux.HandleFunc("/api/v1/graph/query", graphAPI.Query)
	mux.HandleFunc("/api/v1/graph/navigate", graphAPI.Navigate)
	// P2: 边反馈
	mux.HandleFunc("/api/v1/graph/edge/feedback", api.EdgeFeedback)
	// 新增：添加关系边（POST JSON body: {"from":"实体A","to":"实体B","relation":"关系类型","namespace":""}）
	mux.HandleFunc("/api/v1/graph/edge", func(w http.ResponseWriter, r *http.Request) {
		if r.Method != "POST" {
			http.Error(w, "POST only", 405)
			return
		}
		var req struct {
			From      string `json:"from"`
			To        string `json:"to"`
			Relation  string `json:"relation"`
			Namespace string `json:"namespace"`
		}
		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
			http.Error(w, err.Error(), 400)
			return
		}
		if req.From == "" || req.To == "" || req.Relation == "" {
			http.Error(w, "from, to, relation required", 400)
			return
		}
		ns := req.Namespace
		if ns == "" {
			ns = "default"
		}
		edgeID := fmt.Sprintf("e_%s_%s_%d", req.From, req.Relation, time.Now().UnixNano())
		maxRetries := 5
		var lastErr error
		for attempt := 0; attempt < maxRetries; attempt++ {
			lastErr = graphStore.AddEdge(edgeID, req.From, req.To, req.Relation, ns, 1.0)
			if lastErr == nil {
				break // Success, exit retry loop
			}
			// Check for rate limit error (429) - zhiyid returns rate_limit_exceeded
			if strings.Contains(lastErr.Error(), "rate_limit_exceeded") || strings.Contains(lastErr.Error(), "429") {
				// Exponential backoff: 1s, 2s, 4s, 8s, 16s
				backoff := time.Duration(1<<uint(attempt)) * time.Second
				time.Sleep(backoff)
				continue
			}
			// Any other error - don't retry
			break
		}
		if lastErr != nil {
			// Max retries exceeded - write to disk for later retry
			failedPath := "/tmp/memoryweave/failed_edges.jsonl"
			dir := filepath.Dir(failedPath)
			os.MkdirAll(dir, 0755)
			failedEdge := map[string]interface{}{
				"edge_id":   edgeID,
				"from":      req.From,
				"to":        req.To,
				"relation":  req.Relation,
				"namespace": ns,
				"weight":    1.0,
				"timestamp": time.Now().UnixNano(),
				"error":     lastErr.Error(),
			}
			jsonData, _ := json.Marshal(failedEdge)
			// Append to jsonl file
			f, _ := os.OpenFile(failedPath, os.O_APPEND|os.O_CREATE|os.O_WRONLY, 0644)
			defer f.Close()
			f.Write(jsonData)
			f.WriteString("\n")
			// Return success with warning that edge was queued for later retry
			w.Header().Set("Content-Type", "application/json")
			json.NewEncoder(w).Encode(map[string]interface{}{
				"status":  "queued_for_retry",
				"edge_id": edgeID,
				"message": "Edge write failed after max retries; queued to /tmp/memoryweave/failed_edges.jsonl",
			})
			return
		}
		// 图谱变更：失效 navigate 缓存
		if cachedGraphStoreRef != nil {
			cachedGraphStoreRef.InvalidateAll()
		}
		json.NewEncoder(w).Encode(map[string]interface{}{"status": "ok", "edge_id": edgeID})
	})
	// 新增：pagerank + evidence_count
	mux.HandleFunc("/api/v1/graph/pagerank", func(w http.ResponseWriter, r *http.Request) {
		rank := graphStore.PageRank(0.85, 20)
		data, _ := json.Marshal(map[string]interface{}{
			"pagerank": rank,
			"count":    len(rank),
		})
		w.Header().Set("Content-Type", "application/json")
		w.Write(data)
	})
	mux.HandleFunc("/api/v1/graph/evidence/", func(w http.ResponseWriter, r *http.Request) {
		entity := r.URL.Path[len("/api/v1/graph/evidence/"):]
		count := graphStore.EvidenceCount(entity)
		data, _ := json.Marshal(map[string]interface{}{
			"entity":         entity,
			"evidence_count": count,
		})
		w.Header().Set("Content-Type", "application/json")
		w.Write(data)
	})

	// 图谱可视化导出
	mux.HandleFunc("/api/v1/graph/export", func(w http.ResponseWriter, r *http.Request) {
		ns := r.URL.Query().Get("namespace")
		limitStr := r.URL.Query().Get("limit")
		limit := 0
		if limitStr != "" {
			if l, err := strconv.Atoi(limitStr); err == nil && l > 0 {
				limit = l
			}
		}
		nodes, edges := graphStore.GetGraph(ns, limit)
		respondJSON(w, 200, map[string]interface{}{
			"nodes": nodes,
			"edges": edges,
			"count": map[string]int{"nodes": len(nodes), "edges": len(edges)},
		})
	})

	// 图谱导航 + Obsidian 笔记关联
	mux.HandleFunc("/api/v1/graph/notes", func(w http.ResponseWriter, r *http.Request) {
		entity := r.URL.Query().Get("entity")
		maxHops := 2
		if h := r.URL.Query().Get("max_hops"); h != "" {
			if parsed, err := strconv.Atoi(h); err == nil && parsed > 0 {
				maxHops = parsed
			}
		}
		maxNotes := 10
		if n := r.URL.Query().Get("max_notes"); n != "" {
			if parsed, err := strconv.Atoi(n); err == nil && parsed > 0 {
				maxNotes = parsed
			}
		}

		var noteResults []map[string]interface{}
		var paths []map[string]interface{}
		if entity != "" {
			p, err := graphStore.Navigate(routes.NormalizeEntity(entity), maxHops, "", nil)
			if err == nil {
				paths = p
				entities := []string{routes.NormalizeEntity(entity)}
				for _, path := range p {
					if s, ok := path["source"].(string); ok {
						entities = append(entities, s)
					}
					if t, ok := path["target"].(string); ok {
						entities = append(entities, t)
					}
				}
				if obsidian != nil {
					noteResults = obsidian.SearchNotesByEntities(entities, maxNotes)
				}
			}
		}
		respondJSON(w, 200, map[string]interface{}{
			"entity":      entity,
			"graph_paths": paths,
			"notes":       noteResults,
			"paths_count": len(paths),
			"notes_count": len(noteResults),
		})
	})

	// 自然语言图谱查询 /api/v1/graph/nl_query
	// 输入: {"query": "织忆和牧尘是什么关系"}
	// 返回: 自然语言关系描述
	mux.HandleFunc("/api/v1/graph/nl_query", func(w http.ResponseWriter, r *http.Request) {
		if r.Method != "POST" {
			http.Error(w, "POST only", http.StatusMethodNotAllowed)
			return
		}
		var req struct {
			Query string `json:"query"`
		}
		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
			http.Error(w, "invalid body", http.StatusBadRequest)
			return
		}
		query := strings.TrimSpace(req.Query)
		if query == "" {
			http.Error(w, "query is required", http.StatusBadRequest)
			return
		}

		// 1. 从 query 中提取可能的实体名（去掉问号、语气词）
		cleanQuery := strings.TrimSuffix(strings.TrimSuffix(query, "？"), "?")
		cleanQuery = strings.TrimPrefix(cleanQuery, "请问")
		cleanQuery = strings.TrimPrefix(cleanQuery, "请说")
		cleanQuery = strings.TrimSpace(cleanQuery)

		// 2. 尝试识别实体对（"A和B的关系"、"A与B"）
		var entityA, entityB string
		// 模式1: "A和B是什么关系" / "A和B有什么关系" / "A和B的关系"
		hasAnd := strings.Contains(cleanQuery, "和") || strings.Contains(cleanQuery, "与")
		if hasAnd && (strings.Contains(cleanQuery, "关系") || strings.Contains(cleanQuery, "关联")) {
			delim := "和"
			if strings.Contains(cleanQuery, "与") {
				delim = "与"
			}
			parts := strings.Split(cleanQuery, delim)
			if len(parts) == 2 {
				entityA = strings.TrimSpace(parts[0])
				rawB := strings.TrimSpace(parts[1])
				// 去掉句末的问句模式得到 entityB
				entityB = strings.TrimSuffix(rawB, "是什么关系")
				entityB = strings.TrimSuffix(entityB, "有什么关系")
				entityB = strings.TrimSuffix(entityB, "的")
				entityB = strings.TrimSuffix(entityB, "是什么")
				entityB = strings.TrimSpace(entityB)
				// 如果去掉问句后为空（如"织忆和牧尘" -> entityB=""），则原始 query 可能是"A和B"
				if entityB == "" {
					entityB = rawB
				}
			}
		}

		var result map[string]interface{}
		if entityA != "" && entityB != "" {
			// 双向查询：A → B 的关系
			paths, _ := graphStore.Navigate(routes.NormalizeEntity(entityA), 2, "", nil)
			relB, _ := graphStore.Navigate(routes.NormalizeEntity(entityB), 2, "", nil)
			// 找 A → B 的直接边
			var directPath map[string]interface{}
			for _, p := range paths {
				toNorm := routes.NormalizeEntity(p["to"].(string))
				targetNorm := routes.NormalizeEntity(entityB)
				if toNorm == targetNorm {
					directPath = p
					break
				}
			}
			result = map[string]interface{}{
				"query":        query,
				"entity_a":     entityA,
				"entity_b":     entityB,
				"direct_path":  directPath,
				"paths_from_a": len(paths),
				"paths_from_b": len(relB),
				"summary":      fmt.Sprintf("%s 与 %s 存在 %d 条关联路径", entityA, entityB, len(paths)),
			}
		} else {
			// 单实体查询：找 entity 的关系网
			entity := cleanQuery
			paths, _ := graphStore.Navigate(routes.NormalizeEntity(entity), 2, "", nil)
			// 按 relation 分组
			relCounts := make(map[string]int)
			for _, p := range paths {
				rel := p["relation"].(string)
				if rel == "" {
					rel = "unknown"
				}
				relCounts[rel]++
			}
			// 生成自然语言描述
			var rels []string
			for rel, cnt := range relCounts {
				rels = append(rels, fmt.Sprintf("%s(%d条)", rel, cnt))
			}
			sort.Strings(rels)
			result = map[string]interface{}{
				"query":            query,
				"entity":           entity,
				"total_paths":      len(paths),
				"relation_summary": rels,
				"summary":          fmt.Sprintf("%s 共有 %d 条关联，分布在 %d 种关系类型", entity, len(paths), len(relCounts)),
			}
		}
		respondJSON(w, 200, result)
	})

	// 图谱脏数据清理 /api/v1/graph/cleanup
	mux.HandleFunc("/api/v1/graph/cleanup", func(w http.ResponseWriter, r *http.Request) {
		if r.Method != "POST" && r.Method != "GET" {
			http.Error(w, "GET or POST only", http.StatusMethodNotAllowed)
			return
		}
		dryRun := r.URL.Query().Get("dry_run") != "false" // 默认为 dryRun
		count, ids, err := graphStore.CleanupNoiseNodes(dryRun)
		if err != nil {
			http.Error(w, "cleanup failed: "+err.Error(), http.StatusInternalServerError)
			return
		}
		// 实际执行后失效缓存（dry_run=true 时不执行，无需失效）
		if !dryRun && cachedGraphStoreRef != nil && count > 0 {
			cachedGraphStoreRef.InvalidateAll()
		}
		respondJSON(w, 200, map[string]interface{}{
			"dry_run":  dryRun,
			"removed":  count,
			"node_ids": ids,
			"message":  fmt.Sprintf("Found %d noise nodes", count),
		})
	})

	// 图谱多 agent 测试污染清理（namespace + 名称子串圈定，不碰全表编码规则）
	// 2026-09-08 卫安/A03/A04 371 节点：openclaw-main/a06-main/hermes-main 中测试 agent 产生的概念/事实节点
	// GET/POST /api/v1/graph/cleanup/scoped?dry_run=true&namespace=a,b,c&name=a03,a04,卫安
	mux.HandleFunc("/api/v1/graph/cleanup/scoped", func(w http.ResponseWriter, r *http.Request) {
		if r.Method != "POST" && r.Method != "GET" {
			http.Error(w, "GET or POST only", http.StatusMethodNotAllowed)
			return
		}
		dryRun := r.URL.Query().Get("dry_run") != "false" // 默认为 dryRun
		nsParam := r.URL.Query().Get("namespace")
		nameParam := r.URL.Query().Get("name")
		if nsParam == "" || nameParam == "" {
			http.Error(w, "namespace and name (comma separated) are required", http.StatusBadRequest)
			return
		}
		namespaces := strings.Split(nsParam, ",")
		nameContains := strings.Split(nameParam, ",")

		type scopedCleaner interface {
			CleanupScopedNodes(dryRun bool, namespaces, nameContains []string) (int, []string, error)
		}
		cleaner, ok := graphStore.(scopedCleaner)
		if !ok {
			http.Error(w, "scoped cleanup not supported by current graph backend", http.StatusNotImplemented)
			return
		}
		count, ids, err := cleaner.CleanupScopedNodes(dryRun, namespaces, nameContains)
		if err != nil {
			http.Error(w, "cleanup failed: "+err.Error(), http.StatusInternalServerError)
			return
		}
		// 实际执行后失效缓存（dry_run=true 时不执行，无需失效）
		if !dryRun && cachedGraphStoreRef != nil && count > 0 {
			cachedGraphStoreRef.InvalidateAll()
		}
		respondJSON(w, 200, map[string]interface{}{
			"dry_run":       dryRun,
			"removed":       count,
			"node_ids":      ids,
			"namespaces":    namespaces,
			"name_contains": nameContains,
			"message":       fmt.Sprintf("Found %d scoped noise nodes", count),
		})
	})

	// 静态文件服务（知识图谱可视化 HTML）
	staticDir := os.Getenv("STATIC_DIR")
	if staticDir == "" {
		staticDir = "/home/muc/projects/memoryweave/go/static"
	}
	fileServer := http.FileServer(http.Dir(staticDir))
	mux.Handle("/static/", http.StripPrefix("/static/", fileServer))

	// 冲突
	mux.HandleFunc("/api/v1/conflicts", func(w http.ResponseWriter, r *http.Request) {
		if r.Method == "GET" {
			conflictAPI.List(w, r)
		} else {
			http.NotFound(w, r)
		}
	})
	mux.HandleFunc("/api/v1/conflicts/resolve", conflictAPI.Resolve)

	// 反馈 — G5: 修正记忆时填充 version_history
	mux.HandleFunc("/api/v1/feedback/useful", func(w http.ResponseWriter, r *http.Request) {
		feedbackAPI.MarkUseful(w, r)
		// 质量监控：useful → 质量上升
		var req struct {
			MemoryID string `json:"memory_id"`
		}
		if r.Body != nil {
			body, _ := io.ReadAll(r.Body)
			json.Unmarshal(body, &req)
			r.Body = io.NopCloser(bytes.NewReader(body))
			if req.MemoryID != "" {
				selfoptimize.Dash.RecordUseful()
				// V 值：useful = success
				selfoptimize.VProp.RecordDecision(
					"useful_"+req.MemoryID,
					[]string{req.MemoryID},
					"user_feedback", "success", "",
				)
				selfoptimize.QualityMonitor.Check(req.MemoryID, 0, 5) // 清除潜在告警
			}
		}
	})
	mux.HandleFunc("/api/v1/feedback/not-useful", func(w http.ResponseWriter, r *http.Request) {
		feedbackAPI.MarkNotUseful(w, r)
		// 质量监控：not-useful → 检查是否需要降权
		var req struct {
			MemoryID string `json:"memory_id"`
		}
		if r.Body != nil {
			body, _ := io.ReadAll(r.Body)
			json.Unmarshal(body, &req)
			r.Body = io.NopCloser(bytes.NewReader(body))
			if req.MemoryID != "" {
				selfoptimize.Dash.RecordNotUseful()
				// V 值：not-useful = failure
				selfoptimize.VProp.RecordDecision(
					"notuseful_"+req.MemoryID,
					[]string{req.MemoryID},
					"user_feedback", "failure", "",
				)
				score := selfoptimize.Dash.QualityScore()
				if record := selfoptimize.QualityMonitor.Check(req.MemoryID, score, 5); record != nil {
					_ = record // WebSocket 通知
				}
			}
		}
	})
	mux.HandleFunc("/api/v1/feedback/deprecate", feedbackAPI.Deprecate)
	mux.HandleFunc("/api/v1/feedback/correct", func(w http.ResponseWriter, r *http.Request) {
		feedbackAPI.Correct(w, r)
		var req struct {
			MemoryID   string `json:"memory_id"`
			NewContent string `json:"new_content"`
			Source     string `json:"source"`
		}
		if r.Body != nil {
			body, _ := io.ReadAll(r.Body)
			json.Unmarshal(body, &req)
			r.Body = io.NopCloser(bytes.NewReader(body))
			if req.MemoryID != "" && req.NewContent != "" {
				routes.FillVersionHistory(req.MemoryID, "", req.NewContent, req.Source)
				// 跨 Agent 通知
				governance.PushMemoryUpdated(req.MemoryID, 0, "corrected")
			}
		}
	})

	// 缺口
	mux.HandleFunc("/api/v1/gaps", gapAPI.List)
	mux.HandleFunc("/api/v1/gaps/detect", gapAPI.Detect)
	mux.HandleFunc("/api/v1/gaps/close/", func(w http.ResponseWriter, r *http.Request) {
		gapAPI.Close(w, r)
		topic := r.URL.Path[len("/api/v1/gaps/close/"):]
		routes.PushGapFilled(topic, 1)
		governance.PushGapFilled(topic, 1)
	})
	mux.HandleFunc("/api/v1/gaps/repair", routes.GapRepair.FullRepairHandler)

	// Agent 注册
	mux.HandleFunc("/api/v1/agents/register", agentRegistry.Register)
	mux.HandleFunc("/api/v1/agents", agentRegistry.List)

	// ─── 多 Agent EventBus 管理 ──────────────────
	mux.HandleFunc("/api/v1/eventbus/subscribe", func(w http.ResponseWriter, r *http.Request) {
		var req struct {
			EventType string `json:"event_type"`
			Callback  string `json:"callback_url"`
		}
		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
			http.Error(w, "invalid request body", http.StatusBadRequest)
			return
		}
		governance.GlobalEventBus.Subscribe(req.EventType, req.Callback)
		respondJSON(w, 200, map[string]string{"status": "subscribed"})
	})
	mux.HandleFunc("/api/v1/eventbus/unsubscribe", func(w http.ResponseWriter, r *http.Request) {
		var req struct {
			EventType string `json:"event_type"`
			Callback  string `json:"callback_url"`
		}
		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
			http.Error(w, "invalid request body", http.StatusBadRequest)
			return
		}
		governance.GlobalEventBus.Unsubscribe(req.EventType, req.Callback)
		respondJSON(w, 200, map[string]string{"status": "unsubscribed"})
	})
	mux.HandleFunc("/api/v1/eventbus/list", func(w http.ResponseWriter, r *http.Request) {
		subs := governance.GlobalEventBus.ListSubscribers()
		respondJSON(w, 200, subs)
	})

	// 跨 Agent 缓存失效接收端点（供其他 Agent 实例 HTTP 回调）
	mux.HandleFunc("/api/v1/cache/invalidate", func(w http.ResponseWriter, r *http.Request) {
		var evt struct {
			Type    string            `json:"type"`
			Payload map[string]string `json:"payload"`
		}
		if err := json.NewDecoder(r.Body).Decode(&evt); err != nil {
			http.Error(w, "invalid request body", http.StatusBadRequest)
			return
		}
		ns := evt.Payload["namespace"]
		storage.SearchCacheInstance.Invalidate(ns)
		log.Printf("[cache] 跨 Agent 失效接收: %s", ns)
		respondJSON(w, 200, map[string]string{"status": "invalidated"})
	})

	// 管理
	// ?task=full 触发完整整合（DBSCAN + 衰减校准 + 质量回溯），默认 cluster_only
	mux.HandleFunc("/api/v1/admin/consolidate", func(w http.ResponseWriter, r *http.Request) {
		task := r.URL.Query().Get("task")
		if task == "" {
			task = "cluster_only"
		}
		// full 模式给 5 分钟；cluster_only 保持 30 秒
		timeout := 30 * time.Second
		if task == "full" {
			timeout = 5 * time.Minute
		}
		ctx, cancel := context.WithTimeout(r.Context(), timeout)
		defer cancel()
		done := make(chan *routes.ConsolidationReport, 1)
		go func() {
			report, _ := consolPipe.RunWithMode(task)
			done <- report
		}()
		select {
		case <-ctx.Done():
			respondJSON(w, 408, map[string]string{"error": "consolidation timeout", "task": task})
			return
		case report := <-done:
			if report == nil {
				respondJSON(w, 500, map[string]string{"error": "consolidation failed", "task": task})
				return
			}
			respondJSON(w, 200, report)
			go governance.PushDistillationComplete(report.Merged, report.ConflictsFound)
		}
	})
	mux.HandleFunc("/api/v1/admin/forget", adminAPI.Forget)
	mux.HandleFunc("/api/v1/admin/dedup", api.Dedup)
	mux.HandleFunc("/api/v1/admin/backup", adminAPI.Backup)
	mux.HandleFunc("/api/v1/admin/restore", adminAPI.Restore)
	mux.HandleFunc("/api/v1/admin/backups", adminAPI.ListBackups)
	mux.HandleFunc("/api/v1/admin/audit", adminAPI.Audit)
	// 遗忘器类型管理
	mux.HandleFunc("/api/v1/admin/forgetter/type", func(w http.ResponseWriter, r *http.Request) {
		var req struct {
			AgentType string `json:"agent_type"`
		}
		if r.Method == "POST" {
			if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
				http.Error(w, "invalid request body", http.StatusBadRequest)
				return
			}
			forgetter.SetAgentType(req.AgentType)
		}
		respondJSON(w, 200, map[string]interface{}{
			"agent_type": forgetter.AgentType(),
			"decay_rate": AgentTypeDecayOrDefault(forgetter.AgentType()),
		})
	})
	mux.HandleFunc("/api/v1/admin/distill/force", func(w http.ResponseWriter, r *http.Request) {
		if routes.DistillEngineRef == nil {
			respondJSON(w, 400, map[string]string{"error": "distill engine not initialized"})
			return
		}
		var req struct {
			Content   string `json:"content"`
			Category  string `json:"category"`
			Namespace string `json:"namespace"`
			AgentID   string `json:"agent_id"`
		}
		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
			http.Error(w, "invalid request body", http.StatusBadRequest)
			return
		}
		if req.Content == "" {
			respondJSON(w, 400, map[string]string{"error": "content required"})
			return
		}
		if req.Namespace == "" {
			req.Namespace = "hermes-main"
		}
		if req.AgentID == "" {
			req.AgentID = "hermes-a06"
		}
		epID := fmt.Sprintf("ep_manual_%d", time.Now().UnixNano())
		routes.AutoDistillTrigger(epID, req.Content, req.Category, req.Namespace, req.AgentID)
		respondJSON(w, 200, map[string]string{"status": "queued", "episode_id": epID})
	})
	// 蒸馏队列状态端点（G8-G9 配套）
	mux.HandleFunc("/api/v1/distill/status", func(w http.ResponseWriter, r *http.Request) {
		if routes.DistillEngineRef == nil {
			respondJSON(w, 200, map[string]interface{}{"status": "not_initialized", "queue_len": 0})
			return
		}
		respondJSON(w, 200, routes.DistillEngineRef.GetStatus())
	})
	mux.HandleFunc("/api/v1/distill/queue", func(w http.ResponseWriter, r *http.Request) {
		if routes.DistillEngineRef == nil {
			respondJSON(w, 200, map[string]interface{}{"queue": []interface{}{}, "count": 0})
			return
		}
		items := routes.DistillEngineRef.QueueItems()
		respondJSON(w, 200, map[string]interface{}{"queue": items, "count": len(items)})
	})
	mux.HandleFunc("/api/v1/distill/quota", func(w http.ResponseWriter, r *http.Request) {
		if routes.DistillEngineRef == nil {
			respondJSON(w, 200, map[string]interface{}{"remaining": 0, "used": 0, "limit": 0, "status": "not_initialized"})
			return
		}
		respondJSON(w, 200, routes.DistillEngineRef.GetQuota())
	})
	mux.HandleFunc("/api/v1/distilled/{id}", func(w http.ResponseWriter, r *http.Request) {
		if r.Method == "DELETE" {
			adminAPI.DeleteDistilled(w, r)
		} else {
			http.NotFound(w, r)
		}
	})
	// P2: 记忆离线整合（LightMem UPDATE_PROMPT）— 手动触发
	// POST /api/v1/consolidate/memory {"namespace":"hermes-main"}
	mux.HandleFunc("/api/v1/consolidate/memory", func(w http.ResponseWriter, r *http.Request) {
		if r.Method != "POST" {
			http.Error(w, "POST only", 405)
			return
		}
		if routes.DistillEngineRef == nil {
			respondJSON(w, 400, map[string]string{"error": "distill engine not initialized"})
			return
		}
		var req struct {
			Namespace string `json:"namespace"`
			Limit     int    `json:"limit"`
		}
		_ = json.NewDecoder(r.Body).Decode(&req)
		if req.Namespace == "" {
			req.Namespace = "hermes-main"
		}
		if req.Limit <= 0 {
			req.Limit = 50
		}

		// 取全部记忆（零向量搜索）
		zeroVec := make([]float32, 1024)
		mems, err := ldb.Search("memories", zeroVec, req.Limit, req.Namespace)
		if err != nil || len(mems) == 0 {
			respondJSON(w, 200, map[string]interface{}{"status": "ok", "processed": 0, "message": "no memories to consolidate"})
			return
		}

		// 两两比较相似度（简单 Jaccard/词重叠启发式，避免调用嵌入）
		// 真实场景：应使用向量相似度，此处用词重叠近似
		processed := 0
		updated := 0
		deleted := 0
		ignored := 0
		errors_ := 0

		// 按目标分组整合：外层 i 为目标，内层收集所有高相似候选，
		// 每个目标只调一次 LLM（UpdatePrompt 原生支持多候选），避免 O(n²) LLM 调用风暴。
		// 2026-09-06 fix(t_04 ④): 原实现每对相似记忆调一次 LLM → limit=100 产生 4950 对
		// → 跑 223s+ 未完成 → 每天 processed=0（不是 namespace 过滤 bug）。
		for i := 0; i < len(mems); i++ {
			// 收集所有与 mems[i] 高相似的候选（j > i 且未处理）
			var cands []distill.MemoryCandidate
			for j := i + 1; j < len(mems); j++ {
				if mems[j].Content == "" {
					continue
				}
				sim := distill.TextSimilarity(mems[i].Content, mems[j].Content)
				if sim < 0.5 {
					continue
				}
				cands = append(cands, distill.MemoryCandidate{ID: mems[j].ID, Content: mems[j].Content})
			}
			if len(cands) == 0 {
				continue
			}

			// 一次 LLM 决策：目标 mems[i] + 全部高相似候选
			result, err := routes.DistillEngineRef.ConsolidateMemory(
				distill.MemoryCandidate{ID: mems[i].ID, Content: mems[i].Content},
				cands,
			)
			if err != nil {
				errors_++
				continue
			}
			processed++
			switch result.Action {
			case "update":
				if result.NewMemory != "" {
					if err := ldb.UpdateMemoryContent(mems[i].ID, result.NewMemory, "consolidate"); err != nil {
						log.Printf("[consolidate] update failed %s: %v", mems[i].ID, err)
					} else {
						updated++
					}
				}
			case "delete":
				if err := ldb.SoftDelete(mems[i].ID, "consolidate: conflict"); err != nil {
					log.Printf("[consolidate] delete failed %s: %v", mems[i].ID, err)
				} else {
					deleted++
				}
			default:
				ignored++
			}
		}

		respondJSON(w, 200, map[string]interface{}{
			"status":         "ok",
			"processed":      processed,
			"updated":        updated,
			"deleted":        deleted,
			"ignored":        ignored,
			"errors":         errors_,
			"total_memories": len(mems),
		})
	})
	// GET /api/v1/memories — list all memories (zero-vector search, for plugin compat)
	mux.HandleFunc("/api/v1/memories", adminAPI.ListMemories)
	mux.HandleFunc("/api/v1/memories/export", adminAPI.ExportMemoriesMD)
	// /api/v1/memory/{id}/versions — existing version history endpoint
	mux.HandleFunc("/api/v1/memory/", func(w http.ResponseWriter, r *http.Request) {
		if r.Method == "GET" {
			adminAPI.Versions(w, r)
		} else {
			http.NotFound(w, r)
		}
	})

	// L3
	mux.HandleFunc("/api/v1/l3/worldmodel", func(w http.ResponseWriter, r *http.Request) {
		if r.Method == "GET" {
			l3API.GetHandler(w, r)
		} else if r.Method == "POST" {
			l3API.UpdateHandler(w, r)
		} else {
			http.NotFound(w, r)
		}
	})

	// 触发器
	mux.HandleFunc("/api/v1/triggers", routes.Triggers.List)
	mux.HandleFunc("/api/v1/triggers/fire", routes.Triggers.Fire)
	// 触发器 pause/resume — 手动解析路径
	mux.HandleFunc("/api/v1/admin/triggers/", func(w http.ResponseWriter, r *http.Request) {
		path := r.URL.Path
		// 格式: /api/v1/admin/triggers/{id}/{action}
		parts := strings.Split(strings.TrimPrefix(path, "/api/v1/admin/triggers/"), "/")
		if len(parts) != 2 {
			http.NotFound(w, r)
			return
		}
		id, action := parts[0], parts[1]
		// 构造一个新请求，让 Pause/Resume 能通过 PathValue 读取
		r.SetPathValue("id", id)
		switch action {
		case "pause":
			routes.Triggers.Pause(w, r)
		case "resume":
			routes.Triggers.Resume(w, r)
		default:
			http.NotFound(w, r)
		}
	})

	// Skills（G7）
	mux.HandleFunc("/api/v1/skills", routes.Skills.List)
	mux.HandleFunc("POST /api/v1/skills", routes.Skills.Register) // 注册新 skill
	mux.HandleFunc("/api/v1/skills/stats", routes.Skills.Stats)   // skill 统计
	mux.HandleFunc("/api/v1/skills/bayes", func(w http.ResponseWriter, r *http.Request) {
		list := routes.BayesianSkills.List()
		respondJSON(w, 200, list)
	})
	mux.HandleFunc("/api/v1/skills/{name}", skillByNameHandler) // GET/DELETE /api/v1/skills/{name}
	mux.HandleFunc("/api/v1/skills/{name}/trial", routes.Skills.Trial)
	mux.HandleFunc("/api/v1/skills/{name}/execute", routes.ExecuteSkill) // G7.3 skill执行
	// G7.2: 结晶路由
	mux.HandleFunc("/api/v1/crystallize/candidates", routes.GetSkillCandidates) // 获取候选
	mux.HandleFunc("/api/v1/crystallize/memory/{id}", routes.CrystallizeSkill)  // 对记忆执行结晶

	// 评估 + 自调参
	mux.HandleFunc("/api/v1/eval/run", evalAPI.Run)
	mux.HandleFunc("/api/v1/eval/history", evalAPI.History)
	mux.HandleFunc("/api/v1/eval/generate", evalAPI.Generate)
	mux.HandleFunc("/api/v1/tuning/status", routes.Tuner.StatusHandler)
	mux.HandleFunc("/api/v1/tuning/run", routes.Tuner.RunHandler)
	mux.HandleFunc("/api/v1/tuning/analytics", routes.Tuner.AnalyticsHandler)

	// 仪表盘 + 验证 + V值 + 缓存 + 流水线
	mux.HandleFunc("/api/v1/metrics/self", func(w http.ResponseWriter, r *http.Request) {
		metrics := selfoptimize.Dash.Metrics()
		respondJSON(w, 200, metrics)
	})
	mux.HandleFunc("/api/v1/validate/passive", func(w http.ResponseWriter, r *http.Request) {
		records := selfoptimize.Validator.GetRecords()
		respondJSON(w, 200, records)
	})
	mux.HandleFunc("/api/v1/vvalue/decisions", func(w http.ResponseWriter, r *http.Request) {
		decisions := selfoptimize.VProp.ListRecentDecisions(20)
		respondJSON(w, 200, decisions)
	})
	// G5.3: V 值查询 + 触发反向传播
	mux.HandleFunc("/api/v1/vvalue/memory/", func(w http.ResponseWriter, r *http.Request) {
		memID := strings.TrimPrefix(r.URL.Path, "/api/v1/vvalue/memory/")
		if memID == "" {
			respondJSON(w, 400, map[string]string{"error": "memory_id required"})
			return
		}
		v := selfoptimize.VProp.GetMemVValue(memID)
		respondJSON(w, 200, map[string]interface{}{"memory_id": memID, "v_value": v})
	})
	mux.HandleFunc("/api/v1/trace/vprop", func(w http.ResponseWriter, r *http.Request) {
		var req struct {
			MemoryIDs []string `json:"memory_ids"`
			Action    string   `json:"action"`
			Outcome   string   `json:"outcome"`
			ParentID  string   `json:"parent_id"`
		}
		if err := json.NewDecoder(r.Body).Decode(&req); err == nil && len(req.MemoryIDs) > 0 {
			id := fmt.Sprintf("trace_%d", time.Now().UnixNano())
			selfoptimize.VProp.RecordDecision(id, req.MemoryIDs, req.Action, req.Outcome, req.ParentID)
			// 更新相关记忆的 quality_score
			for _, mid := range req.MemoryIDs {
				v := selfoptimize.VProp.GetMemVValue(mid)
				lamb := 0.1
				if ldb, ok := ldb.(interface {
					UpdateQualityScore(id string, delta float64) error
				}); ok {
					_ = ldb.UpdateQualityScore(mid, lamb*v)
				}
			}
			respondJSON(w, 200, map[string]string{"status": "propagated"})
			return
		}
		respondJSON(w, 400, map[string]string{"error": "memory_ids and outcome required"})
	})
	mux.HandleFunc("/api/v1/admin/cache", func(w http.ResponseWriter, r *http.Request) {
		stats := storage.SearchCacheInstance.Stats()
		respondJSON(w, 200, stats)
	})
	mux.HandleFunc("/api/v1/admin/pipeline", func(w http.ResponseWriter, r *http.Request) {
		stats := selfoptimize.Flow.Stats()
		respondJSON(w, 200, stats)
	})

	// G6: ephemeral namespace 管理
	mux.HandleFunc("/api/v1/admin/ephemeral/clean", func(w http.ResponseWriter, r *http.Request) {
		now := time.Now().UnixMilli()
		agents := agentRegistry.ListAgents()
		var cleaned []string
		for _, agent := range agents {
			if now-agent.LastSeen > 5*60*1000 {
				ns := agent.AgentID + "-ephemeral"
				zeroVec := make([]float32, 1024)
				memories, _ := ldb.Search("memories", zeroVec, 1000, ns)
				for _, mem := range memories {
					if err := ldb.SoftDelete(mem.ID, "ephemeral_expired"); err != nil {
						log.Printf("[softdelete] failed id: %v", err)
					}
				}
				if len(memories) > 0 {
					cleaned = append(cleaned, ns)
					log.Printf("[ephemeral] cleaned %d memories from %s (inactive for %ds)",
						len(memories), ns, (now-agent.LastSeen)/1000)
				}
			}
		}
		respondJSON(w, 200, map[string]interface{}{
			"status":             "cleaned",
			"cleaned_namespaces": cleaned,
		})
	})

	// Obsidian
	mux.HandleFunc("/api/v1/obsidian/push", obsidian.PushHandler)
	mux.HandleFunc("/api/v1/obsidian/pull", obsidian.PullHandler)
	mux.HandleFunc("/api/v1/obsidian/status", obsidian.StatusHandler)

	// 内部分支：跨 Agent 缓存失效回调
	mux.HandleFunc("/_internal/cache/invalidate", func(w http.ResponseWriter, r *http.Request) {
		var evt struct {
			Payload map[string]string `json:"payload"`
		}
		if err := json.NewDecoder(r.Body).Decode(&evt); err != nil {
			http.Error(w, "invalid request body", http.StatusBadRequest)
			return
		}
		ns := evt.Payload["namespace"]
		storage.SearchCacheInstance.Invalidate(ns)
		log.Printf("[cache] 内部跨 Agent 失效: %s", ns)
		w.WriteHeader(200)
	})

	// ─── Redis 组件 ─────────────────────────────
	heartbeat := storage.NewHeartbeat("zhiyid-primary")
	_ = heartbeat // 后台 goroutine 自动心跳

	// ─── 后台引擎启动 ──────────────────────────
	selfoptimize.RegisterCommitFlow(selfoptimize.Flow)
	selfoptimize.RegisterRecallFlow(selfoptimize.Flow)
	selfoptimize.RegisterGapFlow(selfoptimize.Flow)
	selfoptimize.RegisterCorrectFlow(selfoptimize.Flow)
	selfoptimize.RegisterConsolidateFlow(selfoptimize.Flow)
	// 注册真正的 consolidate 处理器：调用 ConsolidationPipeline
	selfoptimize.Flow.Register("consolidate", func(task *selfoptimize.PipelineTask) error {
		report, err := consolPipe.RunWithMode("full")
		if err == nil && report != nil {
			// 注入 consolidate 报告 → Dashboard
			if report.Merged > 0 {
				// 合并启发式: 按合并数比例缩放, 但限定单次贡献 ≤ 0.05, 防止污染 avg_distill_loss
				loss := float64(report.Merged) * 0.002
				if loss > 0.05 {
					loss = 0.05
				}
				selfoptimize.Dash.RecordDistillLoss(loss)
			}
			if report.ConflictsFound > 0 {
				selfoptimize.Dash.RecordConflictResolved(true)
			}
			for _, p := range report.Patterns {
				if len(p) > 13 && p[:13] == "quality_score" {
					// 解析 quality_score=0.50 → 注入 avg_distill_loss
					var qs float64
					if _, e := fmt.Sscanf(p, "quality_score=%f", &qs); e == nil {
						selfoptimize.Dash.RecordDistillLoss(1.0 - qs)
					}
				}
			}
		}
		return err
	})
	// 注册 conflict_muchen 处理器：自动裁决低信任差异冲突
	selfoptimize.Flow.Register("conflict_muchen", func(task *selfoptimize.PipelineTask) error {
		conflicts := conflictAPI.Detector.ListActive()
		for _, c := range conflicts {
			if c.Status != "pending" {
				continue
			}
			// 按策略自动裁决：latest_wins / primary_wins
			resolution := conflictAPI.Detector.AutoResolve(c)
			if resolution == "pending" {
				// 2026-09-20 修复（链路空转）：**不再静默 dismiss**。
				//
				// 原实现把未匹配策略的冲突改写成 "dismiss" 再 Resolve，且记成
				// RecordConflictResolved(true) —— 后果是每条 pending 冲突最多活 10 分钟
				// 即消失、auto_resolve_rate 指标虚高，冲突从不进入任何人视野
				// （实测：17:19 新建冲突，同进程几分钟后 pending:0，未重启）。
				//
				// 现改为：72 小时内保留 pending 供人工处置；超期才按 latest_wins 自动收敛
				// （这是真裁决，才记指标）。保留窗口内不再有任何自动处置。
				if time.Since(c.CreatedAt) < 72*time.Hour {
					continue
				}
				resolution = "latest_wins"
			}
			if err := conflictAPI.Detector.Resolve(c.ID, resolution, ""); err == nil {
				selfoptimize.Dash.RecordConflictResolved(true)
			}
		}
		return nil
	})
	go selfoptimize.Flow.Start()
	selfoptimize.Executor.Start(selfoptimize.Flow)

	// G6: ephemeral 定时清理（每 5 分钟扫描一次，清理断开超过 5 分钟的 agent 的 ephemeral 记忆）
	go func() {
		for {
			time.Sleep(5 * time.Minute)
			now := time.Now().UnixMilli()
			agents := agentRegistry.ListAgents()
			for _, agent := range agents {
				if now-agent.LastSeen > 5*60*1000 {
					ns := agent.AgentID + "-ephemeral"
					zeroVec := make([]float32, 1024)
					memories, _ := ldb.Search("memories", zeroVec, 1000, ns)
					for _, mem := range memories {
						if err := ldb.SoftDelete(mem.ID, "ephemeral_expired"); err != nil {
							log.Printf("[softdelete] failed id: %v", err)
						}
					}
					if len(memories) > 0 {
						log.Printf("[ephemeral] cleaned %d memories from %s (inactive for %ds)",
							len(memories), ns, (now-agent.LastSeen)/1000)
					}
				}
			}
		}
	}()

	// 质量下降监控：每 30 分钟扫描所有低质量记忆
	go func() {
		for {
			time.Sleep(30 * time.Minute)
			zeroVec := make([]float32, 1024)
			memories, err := ldb.Search("memories", zeroVec, 2000, "")
			if err != nil || len(memories) == 0 {
				continue
			}
			alerts := 0
			for _, m := range memories {
				if m.QualityScore > 0 && m.QualityScore < 0.3 {
					if record := selfoptimize.QualityMonitor.Check(m.ID, m.QualityScore, 2); record != nil {
						alerts++
						if record.Status == "deprecating" {
							log.Printf("[quality] memory %s quality=%.2f deprecating", m.ID[:20], m.QualityScore)
						}
					}
				}
			}
			if alerts > 0 {
				log.Printf("[quality] scan: %d low-quality memories flagged", alerts)
			}
		}
	}()

	// 图持久化 — 确保启动后写入
	if saver, ok := graphStore.(interface{ Save() error }); ok {
		saver.Save()
	}

	// ─── 触发器自动执行循环（每 30 秒检查一次）────────
	go func() {
		ticker := time.NewTicker(30 * time.Second)
		defer ticker.Stop()
		for range ticker.C {
			for _, t := range []struct{ id, action string }{
				{"t_distill", "distill"},
				{"t_merge", "merge"},
				{"t_prune", "prune"},
				{"t_decay", "decay"},
				{"t_backtrack", "backtrack"},
				{"t_gap", "gap_scan"},
				{"t_consolidation", "consolidation"},
			} {
				if !routes.Triggers.CanFire(t.id) {
					continue
				}
				routes.Triggers.RecordFire(t.id)
				go func(triggerID, action string) {
					var err error
					switch action {
					case "distill":
						// distill 保持 cluster_only（高频触发，只做快速聚类）
						_, err = consolPipe.Run()
					case "consolidation", "backtrack":
								// full 模式：含 prune + decay + 质量回溯（Step 4）
								// 48h cooldown 保护，不会真正每分钟跑 full（cooldown 内 CanFire 仍返回 false）
								_, err = consolPipe.RunWithMode("full")
								// 2026-09-21 (t_4ce53193)：冲突去重 + 保留策略
								// 去重：同 (L,R) 对只留最早一条，累加 dup_count，归档 + 审计
								conflictDetector.DedupeActive(governance.DefaultConflictRetentionConfig())
								// 保留：pending > 7d 无人处置 → 归档（不删除）
								conflictDetector.RetainAndArchive(governance.DefaultConflictRetentionConfig())
					case "merge":
						_, err = consolPipe.RunMerge()
					case "prune":
						graphStore.Prune(0.15)
					case "decay":
						// 扫描所有记忆，按衰减率计算 freshness
						// 2026-09-06 P1 fix: 每 tick 限量评估（maxDecayBatch），优先处理
						// 最久未访问的（last_recalled_at 最早 → 最该遗忘），防全表单次拉取
						// 内存翻倍 + 长时间阻塞触发器循环。
						memories, e := ldb.GetCandidatesForForgetting()
						decayScanned := 0
						decayForgotten := 0
						if e == nil {
							const maxDecayBatch = 2000
							if len(memories) > maxDecayBatch {
								sort.SliceStable(memories, func(a, b int) bool {
									return memoryTimeVal(memories[a], "last_recalled_at", "created_at") <
										memoryTimeVal(memories[b], "last_recalled_at", "created_at")
								})
								memories = memories[:maxDecayBatch]
							}
							for _, m := range memories {
								decayScanned++
								// 🔒 2026-09-07 修复(误删事故): episodes(原始对话)不参与遗忘——
								// 它们是蒸馏原料/审计历史, 不是可遗忘的记忆。曾因未过滤 category
								// 导致 67 条 8月对话被 auto_forget 清掉(已恢复)。仅 distilled/general 等记忆可遗忘。
								if cat, ok := m["category"].(string); ok && cat == "episodes" {
									continue
								}
								// 🔒 长内容保护: distilled >200 字(有实质信息)不参与 auto 遗忘
								// (曾误删 405 字 CNB 修复经验; 长记忆应由蒸馏/整合管理, 非时间遗忘)
								if cat, _ := m["category"].(string); cat != "episodes" {
									if cs, ok := m["content"].(string); ok && len([]rune(cs)) > 200 {
										continue
									}
								}
								// 解析 last_recalled_at
								lastAccessed := time.Now().Add(-30 * 24 * time.Hour) // 默认30天前
								if t, ok := m["last_recalled_at"].(string); ok && t != "" {
									if parsed, err := time.Parse(time.RFC3339, t); err == nil {
										lastAccessed = parsed
									}
								}
								// 2026-09-07 方案A 碎片快速道: <30 字且 >20 天未访问 → 直接遗忘
								// (绕过 ShouldForget 的 recallCount 保命——碎片 recall_count 可能虚高)。
								// 碎片无记忆价值(审计: 30.9% <20字 + 43% <50字 = 蒸馏无门槛产物)。
								contentStr := ""
								if cs, ok := m["content"].(string); ok {
									contentStr = cs
								}
								if len([]rune(contentStr)) < 30 && time.Since(lastAccessed).Hours()/24 > 20 {
									if id, ok := m["id"].(string); ok {
										_ = ldb.SoftDelete(id, "auto_forget_fragment")
										decayForgotten++
									}
									continue
								}
								recallCount := 0
								if rc, ok := m["recall_count"].(int); ok {
									recallCount = rc
								}
								tier := "normal"
								if ts, ok := m["tier"].(string); ok {
									tier = ts
								}
								// E4.3: 获取实体图谱度（namespace 作为实体）
								entity := ""
								if ns, ok := m["namespace"].(string); ok {
									entity = ns
								}
								degree := 0
								if entity != "" {
									degree = graphStore.GetEntityDegree(entity)
								}
								if forgetter.ShouldForget(lastAccessed, recallCount, tier, degree) {
									if id, ok := m["id"].(string); ok {
										forgetter.ScanAndForget(id)
										// 实际执行软删除（持久化到 LanceDB）
										_ = ldb.SoftDelete(id, "auto_forget")
										decayForgotten++
									}
								}
							}
						}
						// AC-1 验收日志（P1 修复 type 断言后 last_recalled_at 可解析；batch<=2000）
						if decayScanned > 0 {
							log.Printf("[decay] scanned=%d forgotten=%d (batch<=2000, 最久未访问优先)", decayScanned, decayForgotten)
						}
					case "gap_scan":
						// 检查已有缺口：过期 7 天的自动关闭
						gaps := gapDetector.List()
						for _, g := range gaps {
							if !g.Closed && time.Since(g.CreatedAt) > 7*24*time.Hour {
								gapDetector.Close(g.Topic)
							}
						}
						// 自动修复开放的缺口
						for _, g := range gaps {
							if !g.Closed && routes.GapRepair != nil {
								if routes.GapRepair.AutoRepair(g) {
									selfoptimize.Dash.RecordGapClosed()
								}
							}
						}
					}
					if err != nil {
						routes.Triggers.RecordFail(triggerID)
					}
					// Skill 结晶：每次蒸馏成功后记录
					if action == "distill" || action == "consolidation" {
						routes.Skills.RecordTrial("auto_distill", true)
					}
					routes.WSBus.Broadcast("trigger.fired", map[string]string{
						"trigger_id": triggerID,
						"action":     action,
					})
				}(t.id, t.action)
			}
		}
	}()

	// 自优化指标定时采集（每 30 分钟）
	go func() {
		for {
			time.Sleep(30 * time.Minute)
			m := selfoptimize.Dash.Metrics()
			date := time.Now().Format("2006-01-02")
			for k, v := range m {
				storage.GlobalMetricsStore.Set(date, k, v)
			}
		}
	}()

	// 首次指标采集
	m := selfoptimize.Dash.Metrics()
	date := time.Now().Format("2006-01-02")
	for k, v := range m {
		storage.GlobalMetricsStore.Set(date, k, v)
	}

	// Prometheus 指标定时同步（每 15s 采集一次系统指标）
	go func() {
		for {
			time.Sleep(15 * time.Second)
			// 同步 Dashboard → Prometheus gauges
			metrics.SyncFromDashboard(selfoptimize.Dash.Metrics())
			// 采集 SQLite 数据库文件大小（Linux only）
			if runtime.GOOS != "windows" {
				gPath := os.Getenv("GRAPH_PATH")
				if gPath == "" {
					gPath = "/var/lib/memoryweave/graph.db"
				}
				if fi, err := os.Stat(gPath); err == nil {
					metrics.SQLiteDBSizeBytes.Set(float64(fi.Size()))
				}
			}
			// 采集进程 RSS 内存
			if data, err := os.ReadFile("/proc/self/status"); err == nil {
				for _, line := range strings.Split(string(data), "\n") {
					if strings.HasPrefix(line, "VmRSS:") {
						// VmRSS:    12345 kB
						f := strings.Fields(line)
						if len(f) >= 2 {
							if kb, err := strconv.ParseFloat(f[1], 64); err == nil {
								metrics.MemoryRSSBytes.Set(kb * 1024)
							}
						}
						break
					}
				}
			}
		}
	}()

	log.Println("[zhiyid] 多 Agent 架构 — Redis + FileGraph + EventBus + vLLM — 已启动")

	// Web UI 静态文件根目录
	webUIRoot = os.Getenv("ZHIYI_WEB_UI_ROOT")
	if webUIRoot == "" {
		webUIRoot = "../web-ui"
	}

	// Catch-all: 优先检查静态文件，再 JSON
	mux.HandleFunc("/", func(w http.ResponseWriter, r *http.Request) {
		p := r.URL.Path

		// 静态文件服务（安全检查，不允许 .. 遍历）
		if p != "/" && !strings.Contains(p, "..") {
			filePath := webUIRoot + p
			if _, err := os.Stat(filePath); err == nil {
				http.ServeFile(w, r, filePath)
				return
			}
		}

		// 根路径 → index.html
		if p == "/" || p == "" {
			if _, err := os.Stat(webUIRoot + "/index.html"); err == nil {
				http.ServeFile(w, r, webUIRoot+"/index.html")
				return
			}
			respondJSON(w, 200, map[string]interface{}{
				"service": "zhiyid", "status": "ok", "version": "0.1.0",
			})
			return
		}
		respondJSON(w, 404, map[string]interface{}{
			"error": "endpoint not found: " + p,
		})
	})

	return middleware.CORS()(middleware.Auth(mux))
}

// ─── 辅助 ──────────────────────────────────────────

func respondJSON(w http.ResponseWriter, code int, data interface{}) {
	w.Header().Set("Content-Type", "application/json")
	w.WriteHeader(code)
	json.NewEncoder(w).Encode(data)
}

func AgentTypeDecayOrDefault(agentType string) float64 {
	if rate, ok := governance.AgentTypeDecay[agentType]; ok {
		return rate
	}
	return 0.015
}

// memoryTimeVal 取记忆时间戳用于排序（主 key 优先，空则回退副 key；均空返回 0）。
// 2026-09-06 P1: decay 分批按"最久未访问优先"排序需要；兼容 RFC3339 string / time.Time / int。
func memoryTimeVal(m map[string]interface{}, primary, fallback string) int64 {
	for _, key := range []string{primary, fallback} {
		v, ok := m[key]
		if !ok || v == nil {
			continue
		}
		switch val := v.(type) {
		case string:
			if val == "" {
				continue
			}
			if t, err := time.Parse(time.RFC3339, val); err == nil {
				return t.Unix()
			}
			if n, err := strconv.ParseInt(val, 10, 64); err == nil {
				return n
			}
		case time.Time:
			if !val.IsZero() {
				return val.Unix()
			}
		case float64:
			return int64(val)
		case int64:
			return val
		case int:
			return int64(val)
		}
	}
	return 0
}

// runStartupChecks 启动时数据目录一致性检查
// 检测废弃路径（如 /home/muc/data）并警告，防止数据源混乱
func runStartupChecks(backend string) {
	canonicalDataDir := "/var/lib/memoryweave"
	deprecatedPaths := []string{
		"/home/muc/data",
		"/home/muc/.local/share/memoryweave",
	}

	for _, dep := range deprecatedPaths {
		if _, err := os.Stat(dep); err == nil {
			log.Printf("[WARN] 检测到废弃数据目录 %s，仍有数据残留（当前 canonical: %s）", dep, canonicalDataDir)
		}
	}

	// LanceDB 后端：检查 Rust sidecar socket 是否可达
	if backend == "lancedb" {
		sockPath := os.Getenv("LANCEDB_SOCKET")
		if sockPath == "" {
			sockPath = "/tmp/zhiyi-ipc.sock"
		}
		if _, err := os.Stat(sockPath); os.IsNotExist(err) {
			log.Printf("[WARN] LanceDB socket 不存在 (%s)，Rust sidecar 可能未运行", sockPath)
		} else {
			log.Printf("[startup] LanceDB socket 就绪: %s", sockPath)
		}
	}

	// canonical 目录存在性检查
	if _, err := os.Stat(canonicalDataDir); os.IsNotExist(err) {
		log.Printf("[WARN] canonical 数据目录不存在: %s（首次部署？）", canonicalDataDir)
	} else {
		log.Printf("[startup] 数据目录正常: %s", canonicalDataDir)
	}
}