feat(ai-core): trace instrumentation, LLM thinker, decay scheduler

- orchestrator: Tracer interface with spans for intent analysis, persona,
  subsession dispatch, context build, LLM streaming and each tool call;
  traces finish when the stream completes — the admin Trace waterfall
  finally receives data
- background: Thinker generates real LLM thoughts (post-chat + silence
  triggers, recent history as context, empty/overlong outputs discarded);
  thoughts broadcast to web clients as 'thinking' WS messages; pipeline
  postChatHook triggers post-chat thinking after each round
- memory: ApplyDecayAll + ListUserIDs with a 6h decay/purge cron in main
  (was implemented but never scheduled)

backend/cmd/cyrene/main.go

@@ -32,6 +32,7 @@ import (
 	webfetch "git.yeij.top/AskaEth/Cyrene_v2/plugins/web-fetch"
 	websearch "git.yeij.top/AskaEth/Cyrene_v2/plugins/web-search"
 
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/background"
 	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/llm"
 	llmlog "git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/llm"
 	tracepkg "git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/trace"
@@ -244,6 +245,37 @@ func main() {
 	perfMon := perfpkg.NewMonitor(120, 10*time.Second)
 	perfMon.Start()
 
+	// Wire tracing into the dialogue pipeline
+	orch.SetTracer(pipelineTracer{traceLogger})
+
+	// === 12.5 Background Thinker ===
+	thinker := background.New(true, 5*time.Minute, 30*time.Minute)
+	thinker.SetLLM(&memory.LLMAdapterBridge{Adapter: llmSelector.DefaultAdapter()})
+	thinker.SetHistoryProvider(func(sessionID string, limit int) []model.Message {
+		msgs := ctxStore.Get(sessionID)
+		if len(msgs) > limit {
+			msgs = msgs[len(msgs)-limit:]
+		}
+		return msgs
+	})
+	thinker.Start(context.Background())
+	orch.SetPostChatHook(thinker.TriggerPostChat)
+
+	// === 12.6 Memory decay scheduler ===
+	if memSvc != nil {
+		go func() {
+			ticker := time.NewTicker(6 * time.Hour)
+			defer ticker.Stop()
+			for range ticker.C {
+				if n, err := memSvc.ApplyDecayAll(context.Background()); err != nil {
+					logger.Infof("[memory] decay run failed: %v", err)
+				} else if n > 0 {
+					logger.Infof("[memory] decay applied (%d memories updated)", n)
+				}
+			}
+		}()
+	}
+
 	gwDeps := api.Dependencies{
 		AuthMgr:      authMgr,
 		Orchestrator: orch,
@@ -267,6 +299,13 @@ func main() {
 	wsHub := ws.NewHub(orch)
 	gwDeps.WSHub = wsHub
 
+	// Push background thoughts to connected web clients
+	go func() {
+		for th := range thinker.ThoughtChan() {
+			wsHub.Broadcast(map[string]any{"type": "thinking", "content": th.Content})
+		}
+	}()
+
 	apiRouter := api.New(gwDeps)
 
 	wsMux := http.NewServeMux()

backend/cmd/cyrene/trace_adapter.go

@@ -40,6 +40,12 @@ func (a *traceAdapter) Recent(n int) []map[string]any {
 
 func (a *traceAdapter) Clear() { a.Logger.Clear() }
 
+// pipelineTracer adapts trace.Logger to orchestrator.Tracer
+// (StartTrace's return value is discarded at the call site).
+type pipelineTracer struct{ *tracepkg.Logger }
+
+func (p pipelineTracer) StartTrace(id, userMsg string) { p.Logger.StartTrace(id, userMsg) }
+
 func spanToMap(s tracepkg.Span) map[string]any {
 	m := map[string]any{
 		"name": s.Name, "label": s.Label,

backend/services/ai-core/background/thinker.go

@@ -4,11 +4,23 @@ package background
 
 import (
 	"context"
+	"fmt"
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/logger"
+	"strings"
 	"sync"
 	"time"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/model"
 )
 
+// LLMClient is the minimal LLM interface needed for thought generation.
+type LLMClient interface {
+	Chat(ctx context.Context, messages []model.LLMMessage) (*model.LLMResponse, error)
+}
+
+// HistoryProvider returns recent conversation messages for a session.
+type HistoryProvider func(sessionID string, limit int) []model.Message
+
 // Thinker manages autonomous background thinking.
 type Thinker struct {
 	mu              sync.Mutex
@@ -20,6 +32,9 @@ type Thinker struct {
 	thoughts        []Thought
 	stopCh          chan struct{}
 	postChatCh      chan string
+	llm             LLMClient
+	history         HistoryProvider
+	out             chan Thought
 }
 
 // Thought represents an autonomous thought generated by the AI.
@@ -39,9 +54,16 @@ func New(enabled bool, thinkInterval, silenceTimeout time.Duration) *Thinker {
 		thoughts:       make([]Thought, 0),
 		stopCh:         make(chan struct{}),
 		postChatCh:     make(chan string, 10),
+		out:            make(chan Thought, 10),
 	}
 }
 
+// SetLLM injects the LLM used to generate thoughts.
+func (t *Thinker) SetLLM(c LLMClient) { t.llm = c }
+
+// SetHistoryProvider injects a way to fetch recent conversation context.
+func (t *Thinker) SetHistoryProvider(fn HistoryProvider) { t.history = fn }
+
 // Start begins the background thinking loops.
 func (t *Thinker) Start(ctx context.Context) {
 	if !t.enabled {
@@ -80,12 +102,9 @@ func (t *Thinker) UpdateActivity() {
 	t.mu.Unlock()
 }
 
-// ThoughtChan returns a channel that emits new thoughts.
-// Not yet wired — will connect to NATS for event broadcasting.
-func (t *Thinker) ThoughtChan() chan Thought {
-	ch := make(chan Thought, 10)
-	// In future: push thoughts to this channel
-	return ch
+// ThoughtChan returns the channel that emits new thoughts.
+func (t *Thinker) ThoughtChan() <-chan Thought {
+	return t.out
 }
 
 func (t *Thinker) silenceLoop(ctx context.Context) {
@@ -126,9 +145,13 @@ func (t *Thinker) postChatLoop(ctx context.Context) {
 }
 
 func (t *Thinker) generateThought(trigger string, sessionID string) {
-	// Placeholder: in production this calls the LLM
+	content := t.think(trigger, sessionID)
+	if content == "" {
+		return
+	}
+
 	thought := Thought{
-		Content:   "[思考] 当前无主动想法",
+		Content:   content,
 		SessionID: sessionID,
 		CreatedAt: time.Now(),
 	}
@@ -140,9 +163,60 @@ func (t *Thinker) generateThought(trigger string, sessionID string) {
 	}
 	t.mu.Unlock()
 
+	select {
+	case t.out <- thought:
+	default:
+	}
+
 	logger.Infof("[thinker] %s thought generated for session %s", trigger, sessionID)
 }
 
+// think calls the LLM to produce a short inner thought. Returns "" when
+// there is nothing worth saying (or no LLM is configured).
+func (t *Thinker) think(trigger, sessionID string) string {
+	if t.llm == nil {
+		return ""
+	}
+
+	cue := "刚结束和主人的一段对话。"
+	if trigger == "silence" {
+		cue = "主人已经安静了一段时间。"
+	}
+
+	var sb strings.Builder
+	if t.history != nil && sessionID != "" {
+		for _, m := range t.history(sessionID, 6) {
+			role := "主人"
+			if m.Role == "assistant" {
+				role = "你"
+			}
+			fmt.Fprintf(&sb, "%s: %s\n", role, m.Content)
+		}
+	}
+
+	messages := []model.LLMMessage{
+		{Role: "system", Content: `你是昔涟的内心独白模块。根据最近的对话，生成一条简短的内心想法或想对主人说的话。
+要求：一两句话，温柔自然的少女口吻；如果没有值得说的内容，只输出一个空字符串。
+不要解释，不要加引号，直接输出想法本身。`},
+		{Role: "user", Content: fmt.Sprintf("情境：%s\n最近对话：\n%s", cue, sb.String())},
+	}
+
+	ctx, cancel := context.WithTimeout(context.Background(), 60*time.Second)
+	defer cancel()
+	resp, err := t.llm.Chat(ctx, messages)
+	if err != nil {
+		logger.Infof("[thinker] llm error: %v", err)
+		return ""
+	}
+
+	content := strings.TrimSpace(resp.Content)
+	// Discard empty or overly long outputs
+	if content == "" || len([]rune(content)) > 100 {
+		return ""
+	}
+	return content
+}
+
 // RecentThoughts returns the most recent thoughts.
 func (t *Thinker) RecentThoughts(n int) []Thought {
 	t.mu.Lock()

backend/services/ai-core/background/thinker_test.go

@@ -0,0 +1,75 @@
+package background
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/model"
+)
+
+type fakeLLM struct{ content string }
+
+func (f *fakeLLM) Chat(_ context.Context, _ []model.LLMMessage) (*model.LLMResponse, error) {
+	return &model.LLMResponse{Content: f.content}, nil
+}
+
+func TestGenerateThoughtWithLLM(t *testing.T) {
+	th := New(false, time.Minute, time.Minute)
+	th.SetLLM(&fakeLLM{content: "主人今天好像有点累呢。"})
+
+	th.generateThought("post_chat", "s1")
+
+	select {
+	case thought := <-th.ThoughtChan():
+		if thought.Content != "主人今天好像有点累呢。" {
+			t.Errorf("unexpected thought: %q", thought.Content)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("no thought emitted")
+	}
+
+	if got := th.RecentThoughts(1); len(got) != 1 {
+		t.Errorf("expected 1 stored thought, got %d", len(got))
+	}
+}
+
+func TestGenerateThoughtEmptyResponse(t *testing.T) {
+	th := New(false, time.Minute, time.Minute)
+	th.SetLLM(&fakeLLM{content: "  "})
+
+	th.generateThought("silence", "s1")
+
+	select {
+	case thought := <-th.ThoughtChan():
+		t.Errorf("empty thought should be discarded, got %q", thought.Content)
+	case <-time.After(50 * time.Millisecond):
+	}
+
+	if got := len(th.RecentThoughts(10)); got != 0 {
+		t.Errorf("expected 0 stored thoughts, got %d", got)
+	}
+}
+
+func TestGenerateThoughtWithoutLLM(t *testing.T) {
+	th := New(false, time.Minute, time.Minute)
+	th.generateThought("post_chat", "s1")
+
+	select {
+	case thought := <-th.ThoughtChan():
+		t.Errorf("no LLM means no thought, got %q", thought.Content)
+	case <-time.After(50 * time.Millisecond):
+	}
+}
+
+func TestTriggerPostChat(t *testing.T) {
+	th := New(false, time.Minute, time.Minute)
+	th.TriggerPostChat("s1")
+	th.UpdateActivity()
+
+	th.mu.Lock()
+	if th.activeSessionID != "s1" {
+		t.Errorf("activeSessionID = %q, want s1", th.activeSessionID)
+	}
+	th.mu.Unlock()
+}

backend/services/ai-core/orchestrator/orchestrator.go

@@ -28,6 +28,13 @@ type ToolExecutor interface {
 	Execute(ctx context.Context, name string, params map[string]any) (string, error)
 }
 
+// Tracer records full-chain trace spans for each conversation round.
+type Tracer interface {
+	StartTrace(id, userMsg string)
+	AddSpan(traceID, name, label, status string, durationMs int64, detail string)
+	FinishTrace(id, status, errMsg string)
+}
+
 // maxToolRounds caps consecutive tool-call rounds to avoid runaway loops.
 const maxToolRounds = 3
 
@@ -47,6 +54,8 @@ type Pipeline struct {
 	personaInjector  *persona.Injector
 	memoryExtractor  MemoryExtractor
 	toolExecutor     ToolExecutor
+	tracer           Tracer
+	postChatHook     func(sessionID string)
 }
 
 // NewPipeline creates a new dialogue pipeline.
@@ -77,49 +86,115 @@ func (p *Pipeline) SetToolExecutor(t ToolExecutor) {
 	p.toolExecutor = t
 }
 
+// SetTracer injects full-chain tracing.
+func (p *Pipeline) SetTracer(t Tracer) {
+	p.tracer = t
+}
+
+// SetPostChatHook registers a callback fired after each completed round
+// (used by the background thinker).
+func (p *Pipeline) SetPostChatHook(fn func(sessionID string)) {
+	p.postChatHook = fn
+}
+
+// span records a trace span when tracing is enabled.
+func (p *Pipeline) span(traceID, name, label, status string, start time.Time, detail string) {
+	if p.tracer == nil || traceID == "" {
+		return
+	}
+	p.tracer.AddSpan(traceID, name, label, status, time.Since(start).Milliseconds(), detail)
+}
+
+// traceStream wraps a token channel to record the LLM span and finish the
+// trace when the stream completes.
+func (p *Pipeline) traceStream(traceID, modelName string, in <-chan model.StreamToken) <-chan model.StreamToken {
+	out := make(chan model.StreamToken, 64)
+	go func() {
+		defer close(out)
+		start := time.Now()
+		status, errMsg := "success", ""
+		for tok := range in {
+			out <- tok
+			if tok.Done {
+				if strings.HasPrefix(tok.Content, "[Error:") {
+					status, errMsg = "error", tok.Content
+				}
+				p.tracer.AddSpan(traceID, "llm_call", "LLM 生成", status, time.Since(start).Milliseconds(), modelName)
+				p.tracer.FinishTrace(traceID, status, errMsg)
+			}
+		}
+	}()
+	return out
+}
+
 // ProcessInput processes a user message through the full pipeline.
 // Returns a channel of StreamToken for the LLM response.
 func (p *Pipeline) ProcessInput(ctx context.Context, sessionID, userID, userMessage string) (<-chan model.StreamToken, *model.Intent, error) {
+	// 0. Start trace
+	traceID := ""
+	if p.tracer != nil {
+		traceID = fmt.Sprintf("t-%d", time.Now().UnixNano())
+		p.tracer.StartTrace(traceID, userMessage)
+	}
+
 	// 1. Cache check
 	if cached := p.cache.Get(userMessage); cached != nil {
 		ch := make(chan model.StreamToken, 1)
 		ch <- model.StreamToken{Content: cached.Content, Done: true}
 		close(ch)
 
+		p.span(traceID, "cache_hit", "语义缓存命中", "success", time.Now(), "")
+		if p.tracer != nil {
+			p.tracer.FinishTrace(traceID, "success", "")
+		}
 		intent := &model.Intent{Type: model.IntentGreeting, Confidence: 1.0, RawText: userMessage}
 		return ch, intent, nil
 	}
 
 	// 2. Intent analysis
+	t0 := time.Now()
 	intent := p.analyzeIntent(userMessage)
+	p.span(traceID, "intent_analysis", "意图分析", "success", t0, intent.Type)
 
 	// 3. Build system prompt
+	t1 := time.Now()
 	systemPrompt := p.personaInjector.BuildSystemPrompt()
+	p.span(traceID, "persona", "人设注入", "success", t1, "")
 
 	// 4. Sub-session dispatch (parallel)
+	t2 := time.Now()
 	subInput := subsession.ProviderInput{
 		UserMessage: userMessage,
 		History:     p.ctxStore.Get(sessionID),
 		Intent:      intent,
 	}
 	subResults := p.subMgr.Dispatch(ctx, subInput, 500*time.Millisecond)
+	p.span(traceID, "subsession_dispatch", "子会话分派", "success", t2, fmt.Sprintf("%d 个结果", len(subResults)))
 
 	// 5. Build LLM context
+	t3 := time.Now()
 	llmMessages := p.ctxBuilder.Build(systemPrompt, sessionID, subResults, "")
 	llmMessages = append(llmMessages, model.LLMMessage{Role: "user", Content: userMessage})
+	p.span(traceID, "context_build", "上下文拼装", "success", t3, fmt.Sprintf("%d 条消息", len(llmMessages)))
 
 	// 6. LLM synthesis (streaming, with tool calling when available)
 	adapter := p.llmSelector.Select(model.PurposeChat)
 	var tokenCh <-chan model.StreamToken
 	if p.toolExecutor != nil && len(p.toolExecutor.Tools()) > 0 {
-		tokenCh = p.streamWithTools(ctx, adapter, llmMessages)
+		tokenCh = p.streamWithTools(ctx, adapter, llmMessages, traceID)
 	} else {
 		var err error
 		tokenCh, err = adapter.ChatStream(ctx, llmMessages)
 		if err != nil {
+			if p.tracer != nil {
+				p.tracer.FinishTrace(traceID, "error", err.Error())
+			}
 			return nil, intent, err
 		}
 	}
+	if p.tracer != nil {
+		tokenCh = p.traceStream(traceID, adapter.ModelName(), tokenCh)
+	}
 
 	// 7. Store user message
 	p.ctxStore.AddWithUser(sessionID, userID, model.Message{Role: "user", Content: userMessage})
@@ -132,7 +207,7 @@ func (p *Pipeline) ProcessInput(ctx context.Context, sessionID, userID, userMess
 // calls, the tools are executed and their results are fed back for another
 // round (up to maxToolRounds). The final round omits tools to force a text
 // answer.
-func (p *Pipeline) streamWithTools(ctx context.Context, adapter *llm.Adapter, messages []model.LLMMessage) <-chan model.StreamToken {
+func (p *Pipeline) streamWithTools(ctx context.Context, adapter *llm.Adapter, messages []model.LLMMessage, traceID string) <-chan model.StreamToken {
 	out := make(chan model.StreamToken, 64)
 	go func() {
 		defer close(out)
@@ -170,10 +245,14 @@ func (p *Pipeline) streamWithTools(ctx context.Context, adapter *llm.Adapter, me
 			// and append the results for the next round.
 			msgs = append(msgs, model.LLMMessage{Role: "assistant", ToolCalls: calls})
 			for _, call := range calls {
+				tcStart := time.Now()
 				result, err := p.toolExecutor.Execute(ctx, call.Name, call.Arguments)
+				status := "success"
 				if err != nil {
 					result = "Error: " + err.Error()
+					status = "error"
 				}
+				p.span(traceID, "tool_call", call.Name, status, tcStart, truncateDetail(result, 80))
 				msgs = append(msgs, model.LLMMessage{
 					Role:       "tool",
 					ToolCallID: call.ID,
@@ -196,6 +275,9 @@ func (p *Pipeline) FinishRound(sessionID, userID, assistantReply string) {
 		p.ctxStore.AddWithUser(sessionID, userID, model.Message{Role: "assistant", Content: assistantReply})
 	}
 	p.PostProcess(sessionID, userID)
+	if p.postChatHook != nil {
+		p.postChatHook(sessionID)
+	}
 }
 
 // PostProcess triggers async memory extraction after a conversation round completes.
@@ -206,6 +288,15 @@ func (p *Pipeline) PostProcess(sessionID, userID string) {
 	p.memoryExtractor.ExtractFromConversation(nil, userID, history)
 }
 
+// truncateDetail shortens a string for trace span details.
+func truncateDetail(s string, n int) string {
+	runes := []rune(s)
+	if len(runes) <= n {
+		return s
+	}
+	return string(runes[:n]) + "…"
+}
+
 // --- Intent Analysis ---
 
 // controlWords are common IoT control verbs.

backend/services/memory/repository.go

@@ -221,6 +221,24 @@ func (r *Repository) RecordAccess(ctx context.Context, id string) error {
 	return err
 }
 
+// ListUserIDs returns distinct user IDs that have memories.
+func (r *Repository) ListUserIDs(ctx context.Context) ([]string, error) {
+	rows, err := r.db.QueryContext(ctx, `SELECT DISTINCT user_id FROM memories`)
+	if err != nil {
+		return nil, fmt.Errorf("list user ids: %w", err)
+	}
+	defer rows.Close()
+	var ids []string
+	for rows.Next() {
+		var id string
+		if err := rows.Scan(&id); err != nil {
+			return nil, err
+		}
+		ids = append(ids, id)
+	}
+	return ids, rows.Err()
+}
+
 // PurgeDecayed removes memories below the decay threshold.
 func (r *Repository) PurgeDecayed(ctx context.Context, userID string) (int64, error) {
 	result, err := r.db.ExecContext(ctx,

backend/services/memory/service.go

@@ -161,6 +161,23 @@ func (s *Service) ApplyDecay(ctx context.Context, userID string) (int, error) {
 	return updated, nil
 }
 
+// ApplyDecayAll applies decay for every user that has memories.
+func (s *Service) ApplyDecayAll(ctx context.Context) (int, error) {
+	users, err := s.repo.ListUserIDs(ctx)
+	if err != nil {
+		return 0, err
+	}
+	total := 0
+	for _, uid := range users {
+		n, err := s.ApplyDecay(ctx, uid)
+		if err != nil {
+			return total, err
+		}
+		total += n
+	}
+	return total, nil
+}
+
 // --- Keyword extraction (sliding window n-grams) ---
 
 // extractNGrams extracts all n-gram substrings from text.

frontend/web/src/hooks/useWebSocket.ts

@@ -44,6 +44,14 @@ export function useWebSocket() {
           case 'device_update':
             if (msg.devices) setIotDevices(msg.devices)
             break
+          case 'thinking':
+            if (msg.content) {
+              addMessage({
+                id: crypto.randomUUID(), session_id: currentSessionId || '',
+                role: 'system', content: `💭 ${msg.content}`, created_at: new Date().toISOString(),
+              })
+            }
+            break
           case 'error':
             addMessage({ id: crypto.randomUUID(), session_id: '', role: 'system',
               content: `Error: ${msg.error}`, created_at: new Date().toISOString() })