diff --git a/backend/services/ai-core/background/thinker.go b/backend/services/ai-core/background/thinker.go
new file mode 100644
index 0000000..efb20a3
--- /dev/null
+++ b/backend/services/ai-core/background/thinker.go
@@ -0,0 +1,155 @@
+// Package background provides the autonomous thinking engine for Cyrene.
+// It generates thoughts after conversations and during idle periods.
+package background
+
+import (
+	"context"
+	"log"
+	"sync"
+	"time"
+)
+
+// Thinker manages autonomous background thinking.
+type Thinker struct {
+	mu              sync.Mutex
+	enabled         bool
+	thinkInterval   time.Duration
+	silenceTimeout  time.Duration
+	lastActivity    time.Time
+	activeSessionID string
+	thoughts        []Thought
+	stopCh          chan struct{}
+	postChatCh      chan string
+}
+
+// Thought represents an autonomous thought generated by the AI.
+type Thought struct {
+	Content   string    `json:"content"`
+	SessionID string    `json:"session_id,omitempty"`
+	CreatedAt time.Time `json:"created_at"`
+}
+
+// New creates a new background thinker.
+func New(enabled bool, thinkInterval, silenceTimeout time.Duration) *Thinker {
+	return &Thinker{
+		enabled:        enabled,
+		thinkInterval:  thinkInterval,
+		silenceTimeout: silenceTimeout,
+		lastActivity:   time.Now(),
+		thoughts:       make([]Thought, 0),
+		stopCh:         make(chan struct{}),
+		postChatCh:     make(chan string, 10),
+	}
+}
+
+// Start begins the background thinking loops.
+func (t *Thinker) Start(ctx context.Context) {
+	if !t.enabled {
+		log.Println("[thinker] background thinking disabled")
+		return
+	}
+
+	go t.silenceLoop(ctx)
+	go t.postChatLoop(ctx)
+	log.Printf("[thinker] started (interval=%v, silence=%v)", t.thinkInterval, t.silenceTimeout)
+}
+
+// Stop shuts down the thinker.
+func (t *Thinker) Stop() {
+	close(t.stopCh)
+	log.Println("[thinker] stopped")
+}
+
+// TriggerPostChat schedules a post-chat think after a conversation.
+func (t *Thinker) TriggerPostChat(sessionID string) {
+	t.mu.Lock()
+	t.activeSessionID = sessionID
+	t.lastActivity = time.Now()
+	t.mu.Unlock()
+
+	select {
+	case t.postChatCh <- sessionID:
+	default:
+	}
+}
+
+// UpdateActivity updates the last activity timestamp (for silence detection).
+func (t *Thinker) UpdateActivity() {
+	t.mu.Lock()
+	t.lastActivity = time.Now()
+	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
+}
+
+func (t *Thinker) silenceLoop(ctx context.Context) {
+	ticker := time.NewTicker(t.thinkInterval)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-t.stopCh:
+			return
+		case <-ticker.C:
+			t.mu.Lock()
+			elapsed := time.Since(t.lastActivity)
+			t.mu.Unlock()
+
+			if elapsed >= t.silenceTimeout {
+				t.generateThought("silence", t.activeSessionID)
+			}
+		case <-ctx.Done():
+			return
+		}
+	}
+}
+
+func (t *Thinker) postChatLoop(ctx context.Context) {
+	for {
+		select {
+		case <-t.stopCh:
+			return
+		case sessionID := <-t.postChatCh:
+			// Delay post-chat thinking to avoid immediate noise
+			time.Sleep(5 * time.Second)
+			t.generateThought("post_chat", sessionID)
+		case <-ctx.Done():
+			return
+		}
+	}
+}
+
+func (t *Thinker) generateThought(trigger string, sessionID string) {
+	// Placeholder: in production this calls the LLM
+	thought := Thought{
+		Content:   "[思考] 当前无主动想法",
+		SessionID: sessionID,
+		CreatedAt: time.Now(),
+	}
+
+	t.mu.Lock()
+	t.thoughts = append(t.thoughts, thought)
+	if len(t.thoughts) > 100 {
+		t.thoughts = t.thoughts[1:]
+	}
+	t.mu.Unlock()
+
+	log.Printf("[thinker] %s thought generated for session %s", trigger, sessionID)
+}
+
+// RecentThoughts returns the most recent thoughts.
+func (t *Thinker) RecentThoughts(n int) []Thought {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+
+	if n > len(t.thoughts) {
+		n = len(t.thoughts)
+	}
+	return t.thoughts[len(t.thoughts)-n:]
+}
diff --git a/backend/services/ai-core/context/context.go b/backend/services/ai-core/context/context.go
new file mode 100644
index 0000000..3ed355d
--- /dev/null
+++ b/backend/services/ai-core/context/context.go
@@ -0,0 +1,178 @@
+// Package context manages conversation context with sliding window and token budgeting.
+package context
+
+import (
+	"fmt"
+	"strings"
+	"sync"
+	"unicode/utf8"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/model"
+)
+
+// Store holds conversation history with a sliding window.
+type Store struct {
+	mu       sync.RWMutex
+	sessions map[string]*SessionHistory
+	maxTokens int // Maximum tokens to include in context
+	maxTurns  int // Maximum message turns to keep
+}
+
+// SessionHistory holds messages for a single session.
+type SessionHistory struct {
+	Messages []model.Message
+	mu       sync.RWMutex
+}
+
+// NewStore creates a new conversation store.
+func NewStore(maxTokens, maxTurns int) *Store {
+	if maxTokens <= 0 {
+		maxTokens = 4000
+	}
+	if maxTurns <= 0 {
+		maxTurns = 50
+	}
+	return &Store{
+		sessions:  make(map[string]*SessionHistory),
+		maxTokens: maxTokens,
+		maxTurns:  maxTurns,
+	}
+}
+
+// Add appends a message to the session history.
+func (s *Store) Add(sessionID string, msg model.Message) {
+	s.mu.Lock()
+	sh, ok := s.sessions[sessionID]
+	if !ok {
+		sh = &SessionHistory{}
+		s.sessions[sessionID] = sh
+	}
+	s.mu.Unlock()
+
+	sh.mu.Lock()
+	sh.Messages = append(sh.Messages, msg)
+	// Trim to max turns
+	if len(sh.Messages) > s.maxTurns {
+		sh.Messages = sh.Messages[len(sh.Messages)-s.maxTurns:]
+	}
+	sh.mu.Unlock()
+}
+
+// Get retrieves recent messages for a session, trimmed to fit token budget.
+func (s *Store) Get(sessionID string) []model.Message {
+	s.mu.RLock()
+	sh, ok := s.sessions[sessionID]
+	s.mu.RUnlock()
+	if !ok {
+		return nil
+	}
+
+	sh.mu.RLock()
+	defer sh.mu.RUnlock()
+
+	// Return all if within budget
+	tokens := countTokens(sh.Messages)
+	if tokens <= s.maxTokens {
+		// Return a copy
+		result := make([]model.Message, len(sh.Messages))
+		copy(result, sh.Messages)
+		return result
+	}
+
+	// Slide window from the end, keeping most recent messages
+	// Always keep at least the last 4 messages (2 turns)
+	minKeep := 4
+	if len(sh.Messages) < minKeep {
+		minKeep = len(sh.Messages)
+	}
+
+	result := make([]model.Message, 0, minKeep)
+	tokens = 0
+	for i := len(sh.Messages) - 1; i >= 0; i-- {
+		msgTokens := tokenEstimate(sh.Messages[i].Content)
+		if tokens+msgTokens > s.maxTokens && len(result) >= minKeep {
+			break
+		}
+		result = append([]model.Message{sh.Messages[i]}, result...)
+		tokens += msgTokens
+	}
+	return result
+}
+
+// Clear removes all messages for a session.
+func (s *Store) Clear(sessionID string) {
+	s.mu.Lock()
+	delete(s.sessions, sessionID)
+	s.mu.Unlock()
+}
+
+// Builder assembles the full context for an LLM call.
+type Builder struct {
+	store *Store
+}
+
+// NewBuilder creates a context builder.
+func NewBuilder(store *Store) *Builder {
+	return &Builder{store: store}
+}
+
+// Build creates the full message array for an LLM request.
+// systemPrompt is injected as the first system message.
+// session history, subsession results, and IoT context are merged.
+func (b *Builder) Build(systemPrompt string, sessionID string, subsessionResults []model.SubSessionResult, iotContext string) []model.LLMMessage {
+	messages := make([]model.LLMMessage, 0)
+
+	// 1. System prompt
+	if systemPrompt != "" {
+		messages = append(messages, model.LLMMessage{Role: "system", Content: systemPrompt})
+	}
+
+	// 2. Subsession results (injected as system context)
+	for _, r := range subsessionResults {
+		if r.Error != "" {
+			continue
+		}
+		if r.Content != "" {
+			msg := model.LLMMessage{
+				Role:    "system",
+				Content: fmt.Sprintf("[%s context]\n%s", r.Provider, r.Content),
+			}
+			messages = append(messages, msg)
+		}
+	}
+
+	// 3. IoT device context
+	if iotContext != "" {
+		messages = append(messages, model.LLMMessage{
+			Role:    "system",
+			Content: fmt.Sprintf("[IoT devices]\n%s", iotContext),
+		})
+	}
+
+	// 4. Conversation history
+	history := b.store.Get(sessionID)
+	for _, m := range history {
+		messages = append(messages, model.LLMMessage{Role: m.Role, Content: m.Content})
+	}
+
+	return messages
+}
+
+// --- Helpers ---
+
+// tokenEstimate estimates the number of tokens in a string (rough: 1 token ≈ 2 Chinese chars ≈ 0.75 English words).
+func tokenEstimate(text string) int {
+	words := strings.Fields(text)
+	chars := utf8.RuneCountInString(text)
+	// Approximate: Chinese-heavy text ~ chars/2, English-heavy ~ words*1.3
+	return (chars/2 + int(float64(len(words))*1.3)) / 2
+}
+
+// countTokens estimates total tokens in a message slice.
+func countTokens(msgs []model.Message) int {
+	total := 0
+	for _, m := range msgs {
+		total += tokenEstimate(m.Content)
+	}
+	return total
+}
diff --git a/backend/services/ai-core/llm/llm.go b/backend/services/ai-core/llm/llm.go
new file mode 100644
index 0000000..d502aa4
--- /dev/null
+++ b/backend/services/ai-core/llm/llm.go
@@ -0,0 +1,297 @@
+// Package llm provides LLM abstraction with streaming, model routing, and response caching.
+package llm
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"sync"
+	"time"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/model"
+)
+
+// Config holds LLM connection configuration.
+type Config struct {
+	BaseURL       string
+	APIKey        string
+	Model         string
+	FallbackModel string
+	MaxRetries    int
+	Timeout       time.Duration
+}
+
+// DefaultConfig returns a sensible default Config.
+func DefaultConfig() Config {
+	return Config{
+		BaseURL:       "https://api.openai.com/v1",
+		Model:         "gpt-4o",
+		FallbackModel: "gpt-4o-mini",
+		MaxRetries:    3,
+		Timeout:       180 * time.Second,
+	}
+}
+
+// Adapter handles communication with an OpenAI-compatible LLM API.
+type Adapter struct {
+	cfg    Config
+	client *http.Client
+}
+
+// NewAdapter creates a new LLM adapter.
+func NewAdapter(cfg Config) *Adapter {
+	return &Adapter{
+		cfg: cfg,
+		client: &http.Client{
+			Timeout: cfg.Timeout,
+			Transport: &http.Transport{
+				MaxIdleConns:        100,
+				MaxIdleConnsPerHost: 20,
+				IdleConnTimeout:     90 * time.Second,
+			},
+		},
+	}
+}
+
+// ModelName returns the configured model name.
+func (a *Adapter) ModelName() string { return a.cfg.Model }
+
+// Chat sends a chat completion request and returns the full response.
+func (a *Adapter) Chat(ctx context.Context, messages []model.LLMMessage) (*model.LLMResponse, error) {
+	body := map[string]any{
+		"model":    a.cfg.Model,
+		"messages": messages,
+	}
+
+	resp := &model.LLMResponse{Model: a.cfg.Model}
+	err := a.doRequest(ctx, "POST", "/chat/completions", body, func(data map[string]any) error {
+		if choices, ok := data["choices"].([]any); ok && len(choices) > 0 {
+			choice := choices[0].(map[string]any)
+			if msg, ok := choice["message"].(map[string]any); ok {
+				if content, ok := msg["content"].(string); ok {
+					resp.Content = content
+				}
+			}
+			if reason, ok := choice["finish_reason"].(string); ok {
+				resp.FinishReason = reason
+			}
+		}
+		if usage, ok := data["usage"].(map[string]any); ok {
+			if total, ok := usage["total_tokens"].(float64); ok {
+				resp.TokensUsed = int(total)
+			}
+		}
+		return nil
+	})
+
+	return resp, err
+}
+
+// ChatStream sends a chat completion request and returns a token channel for streaming.
+func (a *Adapter) ChatStream(ctx context.Context, messages []model.LLMMessage) (<-chan model.StreamToken, error) {
+	body := map[string]any{
+		"model":    a.cfg.Model,
+		"messages": messages,
+		"stream":   true,
+	}
+
+	ch := make(chan model.StreamToken, 64)
+
+	go func() {
+		defer close(ch)
+
+		err := a.doStreamRequest(ctx, "POST", "/chat/completions", body, ch)
+		if err != nil {
+			select {
+			case ch <- model.StreamToken{Content: fmt.Sprintf("[LLM error: %v]", err), Done: true}:
+			default:
+			}
+		}
+	}()
+
+	return ch, nil
+}
+
+// doRequest sends a non-streaming request and parses the response.
+func (a *Adapter) doRequest(ctx context.Context, method, path string, body map[string]any, parser func(map[string]any) error) error {
+	jsonBody, _ := json.Marshal(body)
+
+	req, err := http.NewRequestWithContext(ctx, method, a.cfg.BaseURL+path, bytes.NewReader(jsonBody))
+	if err != nil {
+		return fmt.Errorf("create request: %w", err)
+	}
+	req.Header.Set("Authorization", "Bearer "+a.cfg.APIKey)
+	req.Header.Set("Content-Type", "application/json")
+
+	resp, err := a.client.Do(req)
+	if err != nil {
+		return fmt.Errorf("http request: %w", err)
+	}
+	defer resp.Body.Close()
+
+	respBytes, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return fmt.Errorf("read response: %w", err)
+	}
+
+	if resp.StatusCode != http.StatusOK {
+		return fmt.Errorf("API error %d: %s", resp.StatusCode, string(respBytes[:min(len(respBytes), 200)]))
+	}
+
+	var data map[string]any
+	if err := json.Unmarshal(respBytes, &data); err != nil {
+		return fmt.Errorf("parse response: %w", err)
+	}
+	return parser(data)
+}
+
+// doStreamRequest sends a streaming request and sends tokens to the channel.
+func (a *Adapter) doStreamRequest(ctx context.Context, method, path string, body map[string]any, ch chan<- model.StreamToken) error {
+	jsonBody, _ := json.Marshal(body)
+
+	req, err := http.NewRequestWithContext(ctx, method, a.cfg.BaseURL+path, bytes.NewReader(jsonBody))
+	if err != nil {
+		return err
+	}
+	req.Header.Set("Authorization", "Bearer "+a.cfg.APIKey)
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Accept", "text/event-stream")
+
+	resp, err := a.client.Do(req)
+	if err != nil {
+		return err
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode != http.StatusOK {
+		errMsg, _ := io.ReadAll(resp.Body)
+		return fmt.Errorf("stream API error %d: %s", resp.StatusCode, string(errMsg[:min(len(errMsg), 200)]))
+	}
+
+	scanner := bufio.NewScanner(resp.Body)
+	for scanner.Scan() {
+		line := strings.TrimSpace(scanner.Text())
+		if line == "" || !strings.HasPrefix(line, "data: ") {
+			continue
+		}
+		data := strings.TrimPrefix(line, "data: ")
+		if data == "[DONE]" {
+			ch <- model.StreamToken{Done: true}
+			return nil
+		}
+
+		var frame map[string]any
+		if err := json.Unmarshal([]byte(data), &frame); err != nil {
+			continue
+		}
+
+		if choices, ok := frame["choices"].([]any); ok && len(choices) > 0 {
+			choice := choices[0].(map[string]any)
+			if delta, ok := choice["delta"].(map[string]any); ok {
+				if content, ok := delta["content"].(string); ok {
+					ch <- model.StreamToken{Content: content}
+				}
+			}
+			if reason, ok := choice["finish_reason"].(string); ok && reason != "" {
+				ch <- model.StreamToken{FinishReason: reason, Done: true}
+				return nil
+			}
+		}
+	}
+	return scanner.Err()
+}
+
+func min(a, b int) int {
+	if a < b {
+		return a
+	}
+	return b
+}
+
+// ModelSelector routes requests to different models based on purpose.
+type ModelSelector struct {
+	adapters map[model.LLMPurpose]*Adapter
+	default_ *Adapter
+}
+
+// NewModelSelector creates a selector with purpose-specific adapters.
+func NewModelSelector(cfg Config) *ModelSelector {
+	base := NewAdapter(cfg)
+	return &ModelSelector{
+		adapters: map[model.LLMPurpose]*Adapter{},
+		default_: base,
+	}
+}
+
+// SetAdapter assigns an adapter for a specific purpose.
+func (s *ModelSelector) SetAdapter(purpose model.LLMPurpose, a *Adapter) {
+	s.adapters[purpose] = a
+}
+
+// Select returns the adapter for the given purpose, falling back to default.
+func (s *ModelSelector) Select(purpose model.LLMPurpose) *Adapter {
+	if a, ok := s.adapters[purpose]; ok {
+		return a
+	}
+	return s.default_
+}
+
+// DefaultAdapter returns the default adapter.
+func (s *ModelSelector) DefaultAdapter() *Adapter {
+	return s.default_
+}
+
+// SemanticCache caches common greetings to avoid redundant LLM calls.
+type SemanticCache struct {
+	entries map[string]CachedResponse
+	mu      sync.RWMutex
+}
+
+// CachedResponse holds a pre-computed response for a specific input.
+type CachedResponse struct {
+	Content  string               `json:"content"`
+	Segments []model.ReviewSegment `json:"segments"`
+	ExpireAt time.Time            `json:"expire_at"`
+}
+
+// NewSemanticCache creates a cache preloaded with common greetings.
+func NewSemanticCache() *SemanticCache {
+	c := &SemanticCache{entries: make(map[string]CachedResponse)}
+	ttl := 24 * time.Hour
+
+	c.Set("你好", "你好呀~ 有什么可以帮你的吗？", nil, ttl)
+	c.Set("早", "早上好♪ 今天有什么计划吗？", nil, ttl)
+	c.Set("晚安", "晚安~ 做个好梦！", nil, ttl)
+	c.Set("谢谢", "不客气~ 随时可以找我帮忙哦", nil, ttl)
+	c.Set("hello", "Hello~ How can I help you today?", nil, ttl)
+
+	return c
+}
+
+// Set adds a cached response.
+func (c *SemanticCache) Set(input, content string, segments []model.ReviewSegment, ttl time.Duration) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	c.entries[strings.ToLower(strings.TrimSpace(input))] = CachedResponse{
+		Content:  content,
+		Segments: segments,
+		ExpireAt: time.Now().Add(ttl),
+	}
+}
+
+// Get retrieves a cached response, or nil if not found/expired.
+func (c *SemanticCache) Get(input string) *CachedResponse {
+	c.mu.RLock()
+	defer c.mu.RUnlock()
+
+	key := strings.ToLower(strings.TrimSpace(input))
+	if entry, ok := c.entries[key]; ok && time.Now().Before(entry.ExpireAt) {
+		return &entry
+	}
+	return nil
+}
diff --git a/backend/services/ai-core/model/model.go b/backend/services/ai-core/model/model.go
new file mode 100644
index 0000000..803847b
--- /dev/null
+++ b/backend/services/ai-core/model/model.go
@@ -0,0 +1,95 @@
+// Package model defines shared data types for the AI-Core engine.
+package model
+
+import "time"
+
+// Message represents a single chat message in a conversation.
+type Message struct {
+	ID        string    `json:"id"`
+	SessionID string    `json:"session_id"`
+	Role      string    `json:"role"` // "user" | "assistant" | "action" | "system"
+	Content   string    `json:"content"`
+	CreatedAt time.Time `json:"created_at"`
+}
+
+// Session represents a conversation session.
+type Session struct {
+	ID        string    `json:"id"`
+	UserID    string    `json:"user_id"`
+	Title     string    `json:"title"`
+	State     string    `json:"state"` // "active" | "idle"
+	CreatedAt time.Time `json:"created_at"`
+	UpdatedAt time.Time `json:"updated_at"`
+}
+
+// SubSessionResult holds the result from a single subsession provider.
+type SubSessionResult struct {
+	Provider string `json:"provider"` // "general" | "memory" | "iot" | "review"
+	Content  string `json:"content"`
+	Error    string `json:"error,omitempty"`
+}
+
+// Intent represents the detected intent of a user message.
+type Intent struct {
+	Type       string   `json:"type"`       // "chat" | "iot_control" | "iot_query" | "greeting" | "memory_trigger" | "knowledge" | "task"
+	Confidence float64  `json:"confidence"` // 0.0 - 1.0
+	Entities   []string `json:"entities"`   // Detected entities (device names, etc.)
+	RawText    string   `json:"raw_text"`
+}
+
+// LLMMessage is the standardized format sent to LLM APIs.
+type LLMMessage struct {
+	Role    string `json:"role"`    // "system" | "user" | "assistant" | "tool"
+	Content string `json:"content"`
+	Name    string `json:"name,omitempty"`
+}
+
+// LLMResponse represents a non-streaming LLM response.
+type LLMResponse struct {
+	Content      string `json:"content"`
+	FinishReason string `json:"finish_reason"` // "stop" | "length" | "tool_calls"
+	Model        string `json:"model"`
+	TokensUsed   int    `json:"tokens_used"`
+}
+
+// StreamToken is a single token from a streaming LLM response.
+type StreamToken struct {
+	Content      string `json:"content"`
+	FinishReason string `json:"finish_reason,omitempty"`
+	Done         bool   `json:"done"`
+}
+
+// ReviewSegment represents a parsed action/chat segment from LLM output.
+type ReviewSegment struct {
+	Type    string `json:"type"` // "action" | "chat"
+	Content string `json:"content"`
+}
+
+// ToolCall represents a tool call requested by the LLM.
+type ToolCall struct {
+	ID        string         `json:"id"`
+	Name      string         `json:"name"`
+	Arguments map[string]any `json:"arguments"`
+}
+
+// Intent types
+const (
+	IntentChat          = "chat"
+	IntentIoTControl    = "iot_control"
+	IntentIoTQuery      = "iot_query"
+	IntentGreeting      = "greeting"
+	IntentMemoryTrigger = "memory_trigger"
+	IntentKnowledge     = "knowledge"
+	IntentTask          = "task"
+)
+
+// LLM purposes for model routing
+type LLMPurpose string
+
+const (
+	PurposeChat              LLMPurpose = "chat"
+	PurposeIntentAnalysis    LLMPurpose = "intent_analysis"
+	PurposeDeepThinking      LLMPurpose = "deep_thinking"
+	PurposeMemoryExtraction  LLMPurpose = "memory_extraction"
+	PurposeToolCalling       LLMPurpose = "tool_calling"
+)
diff --git a/backend/services/ai-core/orchestrator/orchestrator.go b/backend/services/ai-core/orchestrator/orchestrator.go
new file mode 100644
index 0000000..7aad318
--- /dev/null
+++ b/backend/services/ai-core/orchestrator/orchestrator.go
@@ -0,0 +1,230 @@
+// Package orchestrator implements the main dialogue pipeline for Cyrene.
+// It coordinates intent analysis, subsession dispatch, LLM synthesis, and review parsing.
+package orchestrator
+
+import (
+	"context"
+	"fmt"
+	"strings"
+	"time"
+
+	ctxpkg "git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/context"
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/llm"
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/model"
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/persona"
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/subsession"
+)
+
+// Pipeline orchestrates the full dialogue flow.
+type Pipeline struct {
+	llmSelector   *llm.ModelSelector
+	cache         *llm.SemanticCache
+	ctxStore      *ctxpkg.Store
+	ctxBuilder    *ctxpkg.Builder
+	subMgr        *subsession.Manager
+	personaInjector *persona.Injector
+}
+
+// NewPipeline creates a new dialogue pipeline.
+func NewPipeline(
+	llmSel *llm.ModelSelector,
+	cache *llm.SemanticCache,
+	ctxStore *ctxpkg.Store,
+	subMgr *subsession.Manager,
+	personaInjector *persona.Injector,
+) *Pipeline {
+	return &Pipeline{
+		llmSelector:     llmSel,
+		cache:           cache,
+		ctxStore:        ctxStore,
+		ctxBuilder:      ctxpkg.NewBuilder(ctxStore),
+		subMgr:          subMgr,
+		personaInjector: personaInjector,
+	}
+}
+
+// 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, userMessage string) (<-chan model.StreamToken, *model.Intent, error) {
+	// 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)
+
+		intent := &model.Intent{Type: model.IntentGreeting, Confidence: 1.0, RawText: userMessage}
+		return ch, intent, nil
+	}
+
+	// 2. Intent analysis
+	intent := p.analyzeIntent(userMessage)
+
+	// 3. Build system prompt
+	systemPrompt := p.personaInjector.BuildSystemPrompt()
+
+	// 4. Sub-session dispatch (parallel)
+	subInput := subsession.ProviderInput{
+		UserMessage: userMessage,
+		History:     p.ctxStore.Get(sessionID),
+		Intent:      intent,
+	}
+	subResults := p.subMgr.Dispatch(ctx, subInput, 500*time.Millisecond)
+
+	// 5. Build LLM context
+	llmMessages := p.ctxBuilder.Build(systemPrompt, sessionID, subResults, "")
+	llmMessages = append(llmMessages, model.LLMMessage{Role: "user", Content: userMessage})
+
+	// 6. LLM synthesis (streaming)
+	adapter := p.llmSelector.Select(model.PurposeChat)
+	tokenCh, err := adapter.ChatStream(ctx, llmMessages)
+	if err != nil {
+		return nil, intent, err
+	}
+
+	// 7. Store user message
+	p.ctxStore.Add(sessionID, model.Message{Role: "user", Content: userMessage})
+
+	return tokenCh, intent, nil
+}
+
+// --- Intent Analysis ---
+
+// controlWords are common IoT control verbs.
+var controlWords = []string{"打开", "关闭", "关掉", "关上", "调到", "设置", "开关", "调节", "调高", "调低"}
+
+// deviceWords are common IoT device names.
+var deviceWords = []string{"灯", "空调", "窗帘", "电视", "风扇", "加湿器", "插座", "门锁", "传感器"}
+
+// greetingWords are common greetings.
+var greetingWords = []string{"你好", "早", "晚安", "嗨", "hello", "hi", "在吗", "在不在"}
+
+func (p *Pipeline) analyzeIntent(message string) *model.Intent {
+	msg := strings.ToLower(message)
+
+	// Fast path: IoT command
+	hasControl := false
+	for _, w := range controlWords {
+		if strings.Contains(msg, w) {
+			hasControl = true
+			break
+		}
+	}
+	hasDevice := false
+	var entities []string
+	for _, w := range deviceWords {
+		if strings.Contains(msg, w) {
+			hasDevice = true
+			entities = append(entities, w)
+		}
+	}
+	if hasControl && hasDevice {
+		// Check for "开" ambiguity: "开心", "开始", "离开" should NOT trigger IoT
+		if strings.Contains(msg, "开") {
+			ambiguous := false
+			for _, prefix := range []string{"开心", "开始", "离开", "展开", "打开"} {
+				if prefix == "打开" {
+					continue
+				}
+				if strings.Contains(msg, prefix) {
+					ambiguous = true
+					break
+				}
+			}
+			if !ambiguous {
+				return &model.Intent{Type: model.IntentIoTControl, Confidence: 0.95, Entities: entities, RawText: message}
+			}
+		}
+		return &model.Intent{Type: model.IntentIoTControl, Confidence: 0.9, Entities: entities, RawText: message}
+	}
+	if hasDevice {
+		return &model.Intent{Type: model.IntentIoTQuery, Confidence: 0.85, Entities: entities, RawText: message}
+	}
+
+	// Fast path: Greeting
+	for _, w := range greetingWords {
+		if strings.Contains(msg, w) || msg == w {
+			return &model.Intent{Type: model.IntentGreeting, Confidence: 0.95, RawText: message}
+		}
+	}
+
+	// Fast path: Memory trigger
+	memoryKeywords := []string{"记得", "还记得", "以前", "上次", "之前", "过去"}
+	for _, w := range memoryKeywords {
+		if strings.Contains(msg, w) {
+			return &model.Intent{Type: model.IntentMemoryTrigger, Confidence: 0.7, RawText: message}
+		}
+	}
+
+	// Default: chat
+	return &model.Intent{Type: model.IntentChat, Confidence: 0.6, RawText: message}
+}
+
+// --- Helpers ---
+
+// BuildStreamResponse accumulates streaming tokens and stores the final response.
+func BuildStreamResponse(ctx context.Context, tokenCh <-chan model.StreamToken, store *ctxpkg.Store, sessionID string) (string, error) {
+	var fullContent strings.Builder
+
+	for {
+		select {
+		case <-ctx.Done():
+			return fullContent.String(), ctx.Err()
+		case token, ok := <-tokenCh:
+			if !ok {
+				return fullContent.String(), nil
+			}
+			if token.Done {
+				// Store the complete response
+				content := fullContent.String()
+				store.Add(sessionID, model.Message{Role: "assistant", Content: content})
+				return content, nil
+			}
+			fullContent.WriteString(token.Content)
+		}
+	}
+}
+
+// IsStrongIoTCommand checks if the message is unambiguously an IoT command.
+func IsStrongIoTCommand(message string) bool {
+	hasControl := false
+	for _, cw := range controlWords {
+		if strings.Contains(message, cw) {
+			hasControl = true
+			break
+		}
+	}
+	hasDevice := false
+	for _, dw := range deviceWords {
+		if strings.Contains(message, dw) {
+			hasDevice = true
+			break
+		}
+	}
+	return hasControl && hasDevice
+}
+
+// BuildIoTContext creates a device context string for the LLM prompt.
+func BuildIoTContext(devices []DeviceState) string {
+	if len(devices) == 0 {
+		return ""
+	}
+	var sb strings.Builder
+	sb.WriteString("当前智能家居设备状态:\n")
+	for _, d := range devices {
+		state := "关"
+		if power, ok := d.State["power"]; ok {
+			if p, ok := power.(bool); ok && p {
+				state = "开"
+			}
+		}
+		sb.WriteString(fmt.Sprintf("- %s: %s\n", d.Name, state))
+	}
+	return sb.String()
+}
+
+// DeviceState holds a simplified device state for context building.
+type DeviceState struct {
+	ID    string
+	Name  string
+	State map[string]any
+}
diff --git a/backend/services/ai-core/orchestrator/orchestrator_test.go b/backend/services/ai-core/orchestrator/orchestrator_test.go
new file mode 100644
index 0000000..16fbe67
--- /dev/null
+++ b/backend/services/ai-core/orchestrator/orchestrator_test.go
@@ -0,0 +1,147 @@
+package orchestrator
+
+import (
+	"testing"
+
+	ctxpkg "git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/context"
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/llm"
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/model"
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/persona"
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/subsession"
+)
+
+func setupTestPipeline() *Pipeline {
+	return NewPipeline(
+		llm.NewModelSelector(llm.DefaultConfig()),
+		llm.NewSemanticCache(),
+		ctxpkg.NewStore(4000, 50),
+		subsession.NewManager(&subsession.GeneralProvider{}, &subsession.MemoryProvider{}),
+		persona.NewInjector(persona.BuiltinDefault()),
+	)
+}
+
+func TestIntentAnalysis_IoT(t *testing.T) {
+	p := setupTestPipeline()
+
+	tests := []struct {
+		input    string
+		wantType string
+	}{
+		{"打开客厅灯", model.IntentIoTControl},
+		{"关掉卧室灯", model.IntentIoTControl},
+		{"把空调调到26度", model.IntentIoTControl},
+		{"关闭窗帘", model.IntentIoTControl},
+		{"灯开了吗", model.IntentIoTQuery},
+		{"传感器温度多少", model.IntentIoTQuery},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.input, func(t *testing.T) {
+			intent := p.analyzeIntent(tt.input)
+			if intent.Type != tt.wantType {
+				t.Errorf("analyzeIntent(%q).Type = %q, want %q", tt.input, intent.Type, tt.wantType)
+			}
+		})
+	}
+}
+
+func TestIntentAnalysis_Greeting(t *testing.T) {
+	p := setupTestPipeline()
+
+	tests := []string{"你好", "早", "晚安", "在吗"}
+
+	for _, input := range tests {
+		t.Run(input, func(t *testing.T) {
+			intent := p.analyzeIntent(input)
+			if intent.Type != model.IntentGreeting {
+				t.Errorf("analyzeIntent(%q).Type = %q, want %s", input, intent.Type, model.IntentGreeting)
+			}
+		})
+	}
+}
+
+func TestIntentAnalysis_Chat(t *testing.T) {
+	p := setupTestPipeline()
+
+	intent := p.analyzeIntent("今天天气怎么样")
+	if intent.Type != model.IntentChat {
+		t.Errorf("expected chat intent, got %s", intent.Type)
+	}
+}
+
+func TestIntentAnalysis_MemoryTrigger(t *testing.T) {
+	p := setupTestPipeline()
+
+	intent := p.analyzeIntent("你还记得我喜欢什么吗")
+	if intent.Type != model.IntentMemoryTrigger {
+		t.Errorf("expected memory_trigger, got %s", intent.Type)
+	}
+}
+
+func TestIntentAnalysis_AmbiguousOpen(t *testing.T) {
+	p := setupTestPipeline()
+
+	// "开心" and "开始" should NOT trigger IoT intent
+	intent := p.analyzeIntent("我今天很开心")
+	if intent.Type == model.IntentIoTControl {
+		t.Errorf("'开心' should not trigger IoT, got %s", intent.Type)
+	}
+
+	intent = p.analyzeIntent("开始工作")
+	if intent.Type == model.IntentIoTControl {
+		t.Errorf("'开始' should not trigger IoT, got %s", intent.Type)
+	}
+}
+
+func TestIsStrongIoTCommand(t *testing.T) {
+	if !IsStrongIoTCommand("打开客厅灯") {
+		t.Error("expected true for '打开客厅灯'")
+	}
+	if !IsStrongIoTCommand("关闭空调") {
+		t.Error("expected true for '关闭空调'")
+	}
+	if IsStrongIoTCommand("今天天气真好") {
+		t.Error("expected false for non-IoT message")
+	}
+}
+
+func TestBuildIoTContext(t *testing.T) {
+	devices := []DeviceState{
+		{Name: "客厅灯", State: map[string]any{"power": true}},
+		{Name: "卧室灯", State: map[string]any{"power": false}},
+	}
+
+	ctx := BuildIoTContext(devices)
+	if ctx == "" {
+		t.Error("expected non-empty context")
+	}
+	if len(ctx) < 10 {
+		t.Errorf("context too short: %s", ctx)
+	}
+}
+
+func TestSemanticCache(t *testing.T) {
+	p := setupTestPipeline()
+
+	cached := p.cache.Get("你好")
+	if cached == nil {
+		t.Error("'你好' should be cached")
+	}
+	if cached != nil && cached.Content == "" {
+		t.Error("cached content should not be empty")
+	}
+
+	notCached := p.cache.Get("随机字符串不会命中缓存")
+	if notCached != nil {
+		t.Error("random string should not be cached")
+	}
+}
+
+func TestModelConstants(t *testing.T) {
+	if model.IntentChat != "chat" {
+		t.Errorf("IntentChat = %s", model.IntentChat)
+	}
+	if model.IntentIoTControl != "iot_control" {
+		t.Errorf("IntentIoTControl = %s", model.IntentIoTControl)
+	}
+}
diff --git a/backend/services/ai-core/persona/persona.go b/backend/services/ai-core/persona/persona.go
new file mode 100644
index 0000000..2ae7255
--- /dev/null
+++ b/backend/services/ai-core/persona/persona.go
@@ -0,0 +1,185 @@
+// Package persona manages AI character loading and system prompt injection.
+package persona
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+)
+
+// Persona defines an AI character's configuration.
+type Persona struct {
+	Name         string   `yaml:"name" json:"name"`
+	Description  string   `yaml:"description" json:"description"`
+	Greeting     string   `yaml:"greeting" json:"greeting"`
+	Traits       []string `yaml:"traits" json:"traits"`
+	Style        string   `yaml:"style" json:"style"`
+	Constraints  []string `yaml:"constraints" json:"constraints"`
+
+	// IoT-specific personality
+	IoTResponseStyle string `yaml:"iot_response_style" json:"iot_response_style,omitempty"`
+}
+
+// Loader loads persona configurations from a directory.
+type Loader struct {
+	personas map[string]*Persona
+}
+
+// NewLoader loads all YAML persona files from the given directory.
+func NewLoader(dir string) (*Loader, error) {
+	l := &Loader{personas: make(map[string]*Persona)}
+
+	entries, err := os.ReadDir(dir)
+	if err != nil {
+		// Return empty loader if directory doesn't exist
+		if os.IsNotExist(err) {
+			return l, nil
+		}
+		return nil, fmt.Errorf("read persona dir %s: %w", dir, err)
+	}
+
+	for _, entry := range entries {
+		if entry.IsDir() || !strings.HasSuffix(entry.Name(), ".yaml") && !strings.HasSuffix(entry.Name(), ".yml") {
+			continue
+		}
+		path := filepath.Join(dir, entry.Name())
+		p, err := loadFile(path)
+		if err != nil {
+			return nil, fmt.Errorf("load persona %s: %w", entry.Name(), err)
+		}
+		name := strings.TrimSuffix(entry.Name(), filepath.Ext(entry.Name()))
+		p.Name = name
+		l.personas[name] = p
+	}
+
+	return l, nil
+}
+
+// Get returns a persona by name, or nil if not found.
+func (l *Loader) Get(name string) *Persona {
+	return l.personas[name]
+}
+
+// List returns all loaded persona names.
+func (l *Loader) List() []string {
+	names := make([]string, 0, len(l.personas))
+	for name := range l.personas {
+		names = append(names, name)
+	}
+	return names
+}
+
+// Count returns the number of loaded personas.
+func (l *Loader) Count() int {
+	return len(l.personas)
+}
+
+// Default returns the first persona, or a built-in default.
+func (l *Loader) Default() *Persona {
+	for _, p := range l.personas {
+		return p
+	}
+	return BuiltinDefault()
+}
+
+// loadFile reads a single persona YAML file.
+func loadFile(path string) (*Persona, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return nil, err
+	}
+	return parseYAML(data)
+}
+
+// parseYAML parses persona YAML content (simple key-value parser).
+func parseYAML(data []byte) (*Persona, error) {
+	p := &Persona{}
+	lines := strings.Split(string(data), "\n")
+	var currentList *[]string
+
+	for _, line := range lines {
+		line = strings.TrimSpace(line)
+		if line == "" || strings.HasPrefix(line, "#") {
+			continue
+		}
+
+		if strings.HasPrefix(line, "- ") && currentList != nil {
+			*currentList = append(*currentList, strings.TrimPrefix(line, "- "))
+			continue
+		}
+		currentList = nil
+
+		parts := strings.SplitN(line, ":", 2)
+		if len(parts) != 2 {
+			continue
+		}
+		key := strings.TrimSpace(parts[0])
+		val := strings.TrimSpace(parts[1])
+
+		switch key {
+		case "name":
+			p.Name = val
+		case "description":
+			p.Description = val
+		case "greeting":
+			p.Greeting = val
+		case "style":
+			p.Style = val
+		case "iot_response_style":
+			p.IoTResponseStyle = val
+		case "traits":
+			currentList = &p.Traits
+		case "constraints":
+			currentList = &p.Constraints
+		}
+	}
+	return p, nil
+}
+
+// BuiltinDefault returns a hardcoded default persona.
+func BuiltinDefault() *Persona {
+	return &Persona{
+		Name:        "default",
+		Description: "一个友好的 AI 助手",
+		Greeting:    "你好呀~ 有什么可以帮你的吗？",
+		Traits:      []string{"友善", "细心", "幽默"},
+		Style:       "用轻松自然的中文交流，偶尔加一点俏皮表情",
+		Constraints: []string{"不编造事实", "不提供医疗法律建议", "保持礼貌"},
+	}
+}
+
+// Injector builds the system prompt from a persona.
+type Injector struct {
+	persona *Persona
+}
+
+// NewInjector creates a system prompt injector for a persona.
+func NewInjector(p *Persona) *Injector {
+	if p == nil {
+		p = BuiltinDefault()
+	}
+	return &Injector{persona: p}
+}
+
+// BuildSystemPrompt constructs the full system prompt string.
+func (inj *Injector) BuildSystemPrompt() string {
+	var sb strings.Builder
+
+	sb.WriteString(fmt.Sprintf("你是%s，%s。\n", inj.persona.Name, inj.persona.Description))
+
+	if len(inj.persona.Traits) > 0 {
+		sb.WriteString(fmt.Sprintf("性格特点: %s\n", strings.Join(inj.persona.Traits, "、")))
+	}
+	if inj.persona.Style != "" {
+		sb.WriteString(fmt.Sprintf("说话风格: %s\n", inj.persona.Style))
+	}
+	if len(inj.persona.Constraints) > 0 {
+		sb.WriteString("遵守规则:\n")
+		for _, c := range inj.persona.Constraints {
+			sb.WriteString(fmt.Sprintf("- %s\n", c))
+		}
+	}
+
+	return sb.String()
+}
diff --git a/backend/services/ai-core/subsession/subsession.go b/backend/services/ai-core/subsession/subsession.go
new file mode 100644
index 0000000..526f614
--- /dev/null
+++ b/backend/services/ai-core/subsession/subsession.go
@@ -0,0 +1,163 @@
+// Package subsession manages parallel sub-session providers for dialogue enrichment.
+package subsession
+
+import (
+	"context"
+	"fmt"
+	"strings"
+	"sync"
+	"time"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/model"
+)
+
+// Provider executes a specific sub-session task in parallel with others.
+type Provider interface {
+	Name() string
+	Execute(ctx context.Context, input ProviderInput) (*model.SubSessionResult, error)
+}
+
+// ProviderInput holds the context for sub-session execution.
+type ProviderInput struct {
+	UserMessage string
+	History     []model.Message
+	Intent      *model.Intent
+}
+
+// Manager orchestrates parallel sub-session execution.
+type Manager struct {
+	providers []Provider
+}
+
+// NewManager creates a sub-session manager.
+func NewManager(providers ...Provider) *Manager {
+	return &Manager{providers: providers}
+}
+
+// Dispatch runs all providers in parallel and returns results within the timeout.
+func (m *Manager) Dispatch(ctx context.Context, input ProviderInput, timeout time.Duration) []model.SubSessionResult {
+	ctx, cancel := context.WithTimeout(ctx, timeout)
+	defer cancel()
+
+	var wg sync.WaitGroup
+	results := make([]model.SubSessionResult, 0, len(m.providers))
+	var mu sync.Mutex
+
+	for _, p := range m.providers {
+		wg.Add(1)
+		go func(provider Provider) {
+			defer wg.Done()
+			result, err := provider.Execute(ctx, input)
+			mu.Lock()
+			defer mu.Unlock()
+			if err != nil {
+				results = append(results, model.SubSessionResult{
+					Provider: provider.Name(),
+					Error:    err.Error(),
+				})
+			} else if result != nil {
+				results = append(results, *result)
+			}
+		}(p)
+	}
+
+	wg.Wait()
+	return results
+}
+
+// --- Built-in Providers ---
+
+// GeneralProvider handles general conversation intent.
+type GeneralProvider struct{}
+
+func (p *GeneralProvider) Name() string { return "general" }
+func (p *GeneralProvider) Execute(_ context.Context, input ProviderInput) (*model.SubSessionResult, error) {
+	return &model.SubSessionResult{
+		Provider: "general",
+		Content:  fmt.Sprintf("Intent: %s (confidence: %.2f)", input.Intent.Type, input.Intent.Confidence),
+	}, nil
+}
+
+// MemoryProvider retrieves relevant memories (placeholder — wired to Memory Service later).
+type MemoryProvider struct{}
+
+func (p *MemoryProvider) Name() string { return "memory" }
+func (p *MemoryProvider) Execute(_ context.Context, _ ProviderInput) (*model.SubSessionResult, error) {
+	return &model.SubSessionResult{Provider: "memory", Content: "no relevant memories"}, nil
+}
+
+// IoTProvider executes IoT device queries and commands.
+type IoTProvider struct {
+	// GetDeviceState is injected by the main app to query IoT devices.
+	GetDeviceState func(ctx context.Context) (string, error)
+	// ControlDevice is injected to execute IoT commands.
+	ControlDevice func(ctx context.Context, device, action string) (string, error)
+}
+
+func (p *IoTProvider) Name() string { return "iot" }
+func (p *IoTProvider) Execute(ctx context.Context, input ProviderInput) (*model.SubSessionResult, error) {
+	if input.Intent == nil || (input.Intent.Type != model.IntentIoTControl && input.Intent.Type != model.IntentIoTQuery) {
+		return &model.SubSessionResult{Provider: "iot", Content: ""}, nil
+	}
+
+	if p.GetDeviceState != nil {
+		state, err := p.GetDeviceState(ctx)
+		if err == nil {
+			return &model.SubSessionResult{Provider: "iot", Content: state}, nil
+		}
+	}
+	return &model.SubSessionResult{Provider: "iot", Content: "iot devices available"}, nil
+}
+
+// ReviewProvider parses LLM output into action/chat segments.
+type ReviewProvider struct{}
+
+func (p *ReviewProvider) Name() string { return "review" }
+func (p *ReviewProvider) Execute(_ context.Context, _ ProviderInput) (*model.SubSessionResult, error) {
+	return &model.SubSessionResult{Provider: "review", Content: ""}, nil
+}
+
+// ParseReviewMessages splits a message with action markers into segments.
+// Input: "(歪着头看) 客厅灯已经开啦♪"
+// Output: [{type:"action", content:"歪着头看"}, {type:"chat", content:"客厅灯已经开啦♪"}]
+func ParseReviewMessages(content string) []model.ReviewSegment {
+	var segments []model.ReviewSegment
+	var current strings.Builder
+	inAction := false
+
+	runes := []rune(content)
+	for i := 0; i < len(runes); i++ {
+		r := runes[i]
+
+		// Chinese bracket detection
+		if r == '(' || r == '（' {
+			if current.Len() > 0 && !inAction {
+				segments = append(segments, model.ReviewSegment{Type: "chat", Content: current.String()})
+				current.Reset()
+			}
+			inAction = true
+			continue
+		}
+
+		if r == ')' || r == '）' {
+			if inAction && current.Len() > 0 {
+				segments = append(segments, model.ReviewSegment{Type: "action", Content: current.String()})
+				current.Reset()
+			}
+			inAction = false
+			continue
+		}
+
+		current.WriteRune(r)
+	}
+
+	if current.Len() > 0 {
+		segType := "chat"
+		if inAction {
+			segType = "action"
+		}
+		segments = append(segments, model.ReviewSegment{Type: segType, Content: current.String()})
+	}
+
+	return segments
+}
diff --git a/backend/services/ai-core/subsession/subsession_test.go b/backend/services/ai-core/subsession/subsession_test.go
new file mode 100644
index 0000000..c96de61
--- /dev/null
+++ b/backend/services/ai-core/subsession/subsession_test.go
@@ -0,0 +1,79 @@
+package subsession
+
+import (
+	"testing"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/services/ai-core/model"
+)
+
+func TestParseReviewMessages(t *testing.T) {
+	tests := []struct {
+		input    string
+		expected []model.ReviewSegment
+	}{
+		{
+			"(歪着头看你) 叶酱，客厅灯早就开着啦♪",
+			[]model.ReviewSegment{
+				{Type: "action", Content: "歪着头看你"},
+				{Type: "chat", Content: " 叶酱，客厅灯早就开着啦♪"},
+			},
+		},
+		{
+			"你好呀~ 今天有什么需要帮忙的吗？",
+			[]model.ReviewSegment{
+				{Type: "chat", Content: "你好呀~ 今天有什么需要帮忙的吗？"},
+			},
+		},
+		{
+			"（思考中）嗯...这个问题比较复杂",
+			[]model.ReviewSegment{
+				{Type: "action", Content: "思考中"},
+				{Type: "chat", Content: "嗯...这个问题比较复杂"},
+			},
+		},
+		{
+			"",
+			nil,
+		},
+		{
+			"(only action)",
+			[]model.ReviewSegment{
+				{Type: "action", Content: "only action"},
+			},
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.input, func(t *testing.T) {
+			got := ParseReviewMessages(tt.input)
+			if len(got) != len(tt.expected) {
+				t.Fatalf("expected %d segments, got %d: %+v", len(tt.expected), len(got), got)
+			}
+			for i, seg := range got {
+				if seg.Type != tt.expected[i].Type {
+					t.Errorf("segment[%d].Type = %q, want %q", i, seg.Type, tt.expected[i].Type)
+				}
+				if seg.Content != tt.expected[i].Content {
+					t.Errorf("segment[%d].Content = %q, want %q", i, seg.Content, tt.expected[i].Content)
+				}
+			}
+		})
+	}
+}
+
+func TestProviderNames(t *testing.T) {
+	p := &GeneralProvider{}
+	if p.Name() != "general" {
+		t.Errorf("got %s", p.Name())
+	}
+
+	m := &MemoryProvider{}
+	if m.Name() != "memory" {
+		t.Errorf("got %s", m.Name())
+	}
+
+	i := &IoTProvider{}
+	if i.Name() != "iot" {
+		t.Errorf("got %s", i.Name())
+	}
+}