diff --git a/config/config.yaml b/config/config.yaml
index c85ac02..c6945d6 100644
--- a/config/config.yaml
+++ b/config/config.yaml
@@ -9,12 +9,13 @@ ollama:
   timeout: "120s"
   level: "info"
   format: "json"
+ 
 
 sys:
   session_len: 3
   channel_pool_len: 100
   channel_pool_size: 32
-
+  llm_pool_len: 5
 redis:
   host: 47.97.27.195:6379
   type: node
diff --git a/internal/biz/router.go b/internal/biz/router.go
index 6e5fccb..1ed0ed8 100644
--- a/internal/biz/router.go
+++ b/internal/biz/router.go
@@ -16,10 +16,12 @@ import (
 	"encoding/json"
 	"fmt"
 	"strings"
+	"time"
 
 	"gitea.cdlsxd.cn/self-tools/l_request"
 	"github.com/gofiber/fiber/v2/log"
 	"github.com/gofiber/websocket/v2"
+	"github.com/ollama/ollama/api"
 	"github.com/tmc/langchaingo/llms"
 	"xorm.io/builder"
 )
@@ -34,6 +36,7 @@ type AiRouterBiz struct {
 	hisImpl     *impl.ChatImpl
 	conf        *config.Config
 	utilAgent   *utils_ollama.UtilOllama
+	ollama      *utils_ollama.Client
 	channelPool *pkg.SafeChannelPool
 }
 
@@ -48,6 +51,7 @@ func NewAiRouterBiz(
 	conf *config.Config,
 	utilAgent *utils_ollama.UtilOllama,
 	channelPool *pkg.SafeChannelPool,
+	ollama *utils_ollama.Client,
 ) *AiRouterBiz {
 	return &AiRouterBiz{
 		//aiClient:    aiClient,
@@ -59,6 +63,7 @@ func NewAiRouterBiz(
 		taskImpl:    taskImpl,
 		utilAgent:   utilAgent,
 		channelPool: channelPool,
+		ollama:      ollama,
 	}
 }
 
@@ -68,30 +73,52 @@ func (r *AiRouterBiz) Route(ctx context.Context, req *entitys.ChatRequest) (*ent
 	return nil, nil
 }
 
-// Route 执行智能路由
 func (r *AiRouterBiz) RouteWithSocket(c *websocket.Conn, req *entitys.ChatSockRequest) (err error) {
-	ch, err := r.channelPool.Get()
-	if err != nil {
-		return err
-	}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	//ch := r.channelPool.Get()
+	ch := make(chan entitys.ResponseData)
+	done := make(chan struct{})
+
+	go func() {
+		defer close(done)
+		for {
+			select {
+			case v, ok := <-ch:
+				if !ok {
+					return
+				}
+				// 带超时的发送，避免阻塞
+				if err := sendWithTimeout(c, v, 2*time.Second); err != nil {
+					log.Errorf("Send error: %v", err)
+					cancel() // 通知主流程退出
+					return
+				}
+			case <-ctx.Done():
+				return
+			}
+		}
+	}()
+
 	defer func() {
+
 		if err != nil {
-			entitys.MsgSend(c, entitys.ResponseData{
+			_ = entitys.MsgSend(c, entitys.ResponseData{
 				Done:    false,
 				Content: err.Error(),
 				Type:    entitys.ResponseErr,
 			})
 		}
-		entitys.MsgSend(c, entitys.ResponseData{
+		_ = entitys.MsgSend(c, entitys.ResponseData{
 			Done:    true,
 			Content: "",
 			Type:    entitys.ResponseEnd,
 		})
-		err = r.channelPool.Put(ch)
-		if err != nil {
-			close(ch)
-		}
+		//r.channelPool.Put(ch)
+		close(ch)
 	}()
+
 	session := c.Headers("X-Session", "")
 	if len(session) == 0 {
 		return errors.SessionNotFound
@@ -119,46 +146,94 @@ func (r *AiRouterBiz) RouteWithSocket(c *websocket.Conn, req *entitys.ChatSockRe
 	if err != nil {
 		return errors.SystemError
 	}
-	//意图预测
+
 	prompt := r.getPromptLLM(sysInfo, history, req.Text, task)
-	match, err := r.utilAgent.Llm.GenerateContent(context.TODO(), prompt,
-		llms.WithJSONMode(),
-	)
-	if err != nil {
-		return errors.SystemError
-	}
-	log.Info(match.Choices[0].Content)
+
+	AgentClient := r.utilAgent.Get()
+
 	ch <- entitys.ResponseData{
 		Done:    false,
-		Content: match.Choices[0].Content,
+		Content: "准备意图识别",
 		Type:    entitys.ResponseLog,
 	}
-	var matchJson entitys.Match
-	err = json.Unmarshal([]byte(match.Choices[0].Content), &matchJson)
+
+	match, err := AgentClient.Llm.GenerateContent(
+		ctx, // 使用可取消的上下文
+		prompt,
+		llms.WithJSONMode(),
+	)
+	resMsg := match.Choices[0].Content
+
+	r.utilAgent.Put(AgentClient)
+	ch <- entitys.ResponseData{
+		Done:    false,
+		Content: resMsg,
+		Type:    entitys.ResponseLog,
+	}
+	ch <- entitys.ResponseData{
+		Done:    false,
+		Content: "意图识别结束",
+		Type:    entitys.ResponseLog,
+	}
+	//for i := 1; i < 10; i++ {
+	//	ch <- entitys.ResponseData{
+	//		Done:    false,
+	//		Content: fmt.Sprintf("%d", i),
+	//		Type:    entitys.ResponseLog,
+	//	}
+	//	time.Sleep(1 * time.Second)
+	//}
+	//return
+
 	if err != nil {
+		log.Errorf("LLM error: %v", err)
 		return errors.SystemError
 	}
-	go func() {
-		defer func() {
-			if r := recover(); r != nil {
-				err = fmt.Errorf("recovered from panic: %v", r)
-			}
-		}()
-		defer close(ch)
-		err = r.handleMatch(c, ch, &matchJson, task, sysInfo)
-		if err != nil {
-			return
-		}
-	}()
 
-	for v := range ch {
-		if err := entitys.MsgSend(c, v); err != nil {
-			return err
-		}
+	//msg, err := r.ollama.ToolSelect(ctx, r.getPromptOllama(sysInfo, history, req.Text), []api.Tool{})
+	//if err != nil {
+	//	return
+	//}
+	//resMsg := msg.Message.Content
+	select {
+	case ch <- entitys.ResponseData{
+		Done:    false,
+		Content: resMsg,
+		Type:    entitys.ResponseLog,
+	}:
+	case <-ctx.Done():
+		return ctx.Err()
 	}
-	return
+
+	var matchJson entitys.Match
+	if err := json.Unmarshal([]byte(resMsg), &matchJson); err != nil {
+		return errors.SystemError
+	}
+
+	if err := r.handleMatch(ctx, c, ch, &matchJson, task, sysInfo); err != nil {
+		return err
+	}
+
+	return nil
 }
 
+// 辅助函数：带超时的 WebSocket 发送
+func sendWithTimeout(c *websocket.Conn, data entitys.ResponseData, timeout time.Duration) error {
+	sendCtx, cancel := context.WithTimeout(context.Background(), timeout)
+	defer cancel()
+
+	done := make(chan error, 1)
+	go func() {
+		done <- entitys.MsgSend(c, data)
+	}()
+
+	select {
+	case err := <-done:
+		return err
+	case <-sendCtx.Done():
+		return sendCtx.Err()
+	}
+}
 func (r *AiRouterBiz) handleOtherTask(c *websocket.Conn, ch chan entitys.ResponseData, matchJson *entitys.Match) (err error) {
 	ch <- entitys.ResponseData{
 		Done:    false,
@@ -169,7 +244,7 @@ func (r *AiRouterBiz) handleOtherTask(c *websocket.Conn, ch chan entitys.Respons
 	return
 }
 
-func (r *AiRouterBiz) handleMatch(c *websocket.Conn, ch chan entitys.ResponseData, matchJson *entitys.Match, tasks []model.AiTask, sysInfo model.AiSy) (err error) {
+func (r *AiRouterBiz) handleMatch(ctx context.Context, c *websocket.Conn, ch chan entitys.ResponseData, matchJson *entitys.Match, tasks []model.AiTask, sysInfo model.AiSy) (err error) {
 
 	if !matchJson.IsMatch {
 		ch <- entitys.ResponseData{
@@ -404,6 +479,23 @@ func (r *AiRouterBiz) getPrompt(sysInfo model.AiSy, history []model.AiChatHi, re
 	return prompt
 }
 
+func (r *AiRouterBiz) getPromptOllama(sysInfo model.AiSy, history []model.AiChatHi, reqInput string) []api.Message {
+	var (
+		prompt = make([]api.Message, 0)
+	)
+	prompt = append(prompt, api.Message{
+		Role:    "system",
+		Content: r.buildSystemPrompt(sysInfo.SysPrompt),
+	}, api.Message{
+		Role:    "assistant",
+		Content: pkg.JsonStringIgonErr(r.buildAssistant(history)),
+	}, api.Message{
+		Role:    "user",
+		Content: reqInput,
+	})
+	return prompt
+}
+
 func (r *AiRouterBiz) getPromptLLM(sysInfo model.AiSy, history []model.AiChatHi, reqInput string, tasks []model.AiTask) []llms.MessageContent {
 	var (
 		prompt = make([]llms.MessageContent, 0)
diff --git a/internal/config/config.go b/internal/config/config.go
index 6a0bcc4..2e80b18 100644
--- a/internal/config/config.go
+++ b/internal/config/config.go
@@ -28,6 +28,7 @@ type SysConfig struct {
 	SessionLen      int `mapstructure:"session_len"`
 	ChannelPoolLen  int `mapstructure:"channel_pool_len"`
 	ChannelPoolSize int `mapstructure:"channel_pool_size"`
+	LlmPoolLen      int `mapstructure:"llm_pool_len"`
 }
 
 // ServerConfig 服务器配置
diff --git a/internal/pkg/channel_pool.go b/internal/pkg/channel_pool.go
index 5fa2137..eda85fa 100644
--- a/internal/pkg/channel_pool.go
+++ b/internal/pkg/channel_pool.go
@@ -3,7 +3,6 @@ package pkg
 import (
 	"ai_scheduler/internal/config"
 	"ai_scheduler/internal/entitys"
-	"errors"
 	"sync"
 )
 
@@ -25,29 +24,29 @@ func NewSafeChannelPool(c *config.Config) (*SafeChannelPool, func()) {
 }
 
 // 从池中获取 channel（若无空闲则创建新 channel）
-func (p *SafeChannelPool) Get() (chan entitys.ResponseData, error) {
+func (p *SafeChannelPool) Get() chan entitys.ResponseData {
 	p.mu.Lock()
 	defer p.mu.Unlock()
 
 	if p.closed {
-		return nil, errors.New("pool is closed")
+		return make(chan entitys.ResponseData, p.bufSize)
 	}
 
 	select {
 	case ch := <-p.pool: // 从池中取
-		return ch, nil
+		return ch
 	default: // 池为空，创建新 channel
-		return make(chan entitys.ResponseData, p.bufSize), nil
+		return make(chan entitys.ResponseData, p.bufSize)
 	}
 }
 
 // 将 channel 放回池中（必须确保 channel 已清空！）
-func (p *SafeChannelPool) Put(ch chan entitys.ResponseData) error {
+func (p *SafeChannelPool) Put(ch chan entitys.ResponseData) {
 	p.mu.Lock()
 	defer p.mu.Unlock()
 
 	if p.closed {
-		return errors.New("pool is closed")
+		return
 	}
 
 	// 清空 channel（防止复用时读取旧数据）
@@ -62,7 +61,7 @@ func (p *SafeChannelPool) Put(ch chan entitys.ResponseData) error {
 	default: // 池已满，直接关闭 channel（避免泄漏）
 		close(ch)
 	}
-	return nil
+	return
 }
 
 // 关闭池（释放所有资源）
diff --git a/internal/pkg/func.go b/internal/pkg/func.go
index a31a818..20d9d7c 100644
--- a/internal/pkg/func.go
+++ b/internal/pkg/func.go
@@ -1,8 +1,23 @@
 package pkg
 
-import "encoding/json"
+import (
+	"ai_scheduler/internal/entitys"
+	"encoding/json"
+)
 
 func JsonStringIgonErr(data interface{}) string {
 	dataByte, _ := json.Marshal(data)
 	return string(dataByte)
 }
+
+// IsChannelClosed 检查给定的 channel 是否已经关闭
+// 参数 ch: 要检查的 channel，类型为 chan entitys.ResponseData
+// 返回值: bool 类型，true 表示 channel 已关闭，false 表示未关闭
+func IsChannelClosed(ch chan entitys.ResponseData) bool {
+	select {
+	case _, ok := <-ch: // 尝试从 channel 中读取数据
+		return !ok // 如果 ok=false，说明 channel 已关闭
+	default: // 如果 channel 暂时无数据可读（但不一定关闭）
+		return false // channel 未关闭（但可能有数据未读取）
+	}
+}
diff --git a/internal/pkg/utils_ollama/client.go b/internal/pkg/utils_ollama/client.go
index 390caa8..9d9ac82 100644
--- a/internal/pkg/utils_ollama/client.go
+++ b/internal/pkg/utils_ollama/client.go
@@ -46,7 +46,7 @@ func (c *Client) ToolSelect(ctx context.Context, messages []api.Message, tools [
 		Messages: messages,
 		Stream:   new(bool), // 设置为false，不使用流式响应
 		Think:    &api.ThinkValue{Value: true},
-		Tools:    tools,
+		//Tools:    tools,
 	}
 
 	err = c.client.Chat(ctx, req, func(resp api.ChatResponse) error {
diff --git a/internal/pkg/utils_ollama/ollama.go b/internal/pkg/utils_ollama/ollama.go
index 65388a3..6121838 100644
--- a/internal/pkg/utils_ollama/ollama.go
+++ b/internal/pkg/utils_ollama/ollama.go
@@ -2,40 +2,112 @@ package utils_ollama
 
 import (
 	"ai_scheduler/internal/config"
+	"math/rand"
+	"net"
 	"net/http"
 	"os"
+	"sync"
+	"time"
 
 	"github.com/gofiber/fiber/v2/log"
 	"github.com/tmc/langchaingo/llms/ollama"
 )
 
 type UtilOllama struct {
-	Llm *ollama.LLM
+	LlmClientPool *sync.Pool
+	poolSize      int // 记录池大小，用于调试
+	model         string
+	serverURL     string
+	c             *config.Config
+}
+
+type LlmObj struct {
+	Number string
+	Llm    *ollama.LLM
 }
 
 func NewUtilOllama(c *config.Config, logger log.AllLogger) *UtilOllama {
-	llm, err := ollama.New(
-		ollama.WithModel(c.Ollama.Model),
-		ollama.WithHTTPClient(http.DefaultClient),
-		ollama.WithServerURL(getUrl(c)),
-		ollama.WithKeepAlive("-1s"),
-	)
-	if err != nil {
-		logger.Fatal(err)
-		panic(err)
+	poolSize := c.Sys.LlmPoolLen
+	if poolSize <= 0 {
+		poolSize = 10 // 默认值
+		logger.Warnf("LlmPoolLen not set, using default: %d", poolSize)
+	}
+
+	// 初始化 Pool
+	pool := &sync.Pool{
+		New: func() interface{} {
+			llm, err := ollama.New(
+				ollama.WithModel(c.Ollama.Model),
+				ollama.WithHTTPClient(http.DefaultClient),
+				ollama.WithServerURL(getUrl(c)),
+				ollama.WithKeepAlive("-1s"),
+			)
+			if err != nil {
+				logger.Fatalf("Failed to create Ollama client: %v", err)
+				panic(err) // 或者返回 nil + 错误处理
+			}
+			number := randStr(5)
+			log.Info(number)
+			return &LlmObj{
+				Number: number,
+				Llm:    llm,
+			}
+		},
+	}
+
+	// 预填充 Pool
+	for i := 0; i < poolSize; i++ {
+		pool.Put(pool.New())
 	}
 
 	return &UtilOllama{
-		Llm: llm,
+		LlmClientPool: pool,
+		poolSize:      poolSize,
+		model:         c.Ollama.Model,
+		serverURL:     getUrl(c),
 	}
+
 }
 
-//func (o *UtilOllama) a() {
-//	var agent agents.Agent
-//	agent = agents.NewOneShotAgent(llm, tools, opts...)
-//
-//	agents.NewExecutor()
-//}
+func (o *UtilOllama) NewClient() *ollama.LLM {
+	llm, _ := ollama.New(
+		ollama.WithModel(o.c.Ollama.Model),
+		ollama.WithHTTPClient(&http.Client{
+			Transport: &http.Transport{
+				MaxIdleConns:        100,              // 最大空闲连接数（默认 2，太小）
+				MaxIdleConnsPerHost: 100,              // 每个 Host 的最大空闲连接数（默认 2）
+				IdleConnTimeout:     90 * time.Second, // 空闲连接超时时间
+				DialContext: (&net.Dialer{
+					Timeout:   30 * time.Second, // 连接超时
+					KeepAlive: 30 * time.Second, // TCP Keep-Alive
+				}).DialContext,
+			},
+			Timeout: 60 * time.Second, // 整体请求超时（避免无限等待）
+		}),
+		ollama.WithServerURL(getUrl(o.c)),
+		ollama.WithKeepAlive("-1s"),
+	)
+	return llm
+}
+
+// Get 返回一个可用的 LLM 客户端
+func (o *UtilOllama) Get() *LlmObj {
+	client := o.LlmClientPool.Get().(*LlmObj)
+	return client
+}
+
+// Put 归还客户端（可选：检查是否仍可用）
+func (o *UtilOllama) Put(llm *LlmObj) {
+	if llm == nil {
+		return
+	}
+	o.LlmClientPool.Put(llm)
+}
+
+// Stats 返回池的统计信息（用于监控）
+func (o *UtilOllama) Stats() (current, max int) {
+	return o.poolSize, o.poolSize
+}
 
 func getUrl(c *config.Config) string {
 	baseURL := c.Ollama.BaseURL
@@ -45,3 +117,13 @@ func getUrl(c *config.Config) string {
 	}
 	return baseURL
 }
+
+var letters = []rune("abcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ")
+
+func randStr(n int) string {
+	b := make([]rune, n)
+	for i := range b {
+		b[i] = letters[rand.Intn(len(letters))]
+	}
+	return string(b)
+}