Fix(llm): fix bugs & apikey encrypt & support history for mcp-agent-chat & delete modelname in llmSku (#24149)

* fix(llm): mcp-agent apiKey encrypt

* fix(llm): error-handle when instant-model import

* feature(llm): add get-available-network for llm

* feature(llm): support history for mcp-agent-chat

* fix(llm): llm create with net instead of networktype & networkid

* fix(llm): delete modelName in llmSku & fix some bugs

cmd/climc/shell/llm/llm.go

@@ -18,6 +18,7 @@ func init() {
 	cmd.BatchPerform("start", new(options.LLMStartOptions))
 	cmd.Get("probed-models", new(options.LLMIdOptions))
 	cmd.Get("url", new(options.LLMIdOptions))
+	cmd.Custom(shell.CustomActionGet, "available-network", new(options.LLMAvailableNetworkOptions))
 	cmd.Perform("save-instant-model", new(options.LLMSaveInstantModelOptions))
 	cmd.Perform("quick-models", new(options.LLMQuickModelsOptions))
 }

cmd/climc/shell/llm/mcp_agent.go

@@ -4,7 +4,7 @@ import (
 	"bufio"
 	"fmt"
 	"io"
-	"net/url"
+	"net/http"
 	"strings"
 
 	"yunion.io/x/onecloud/cmd/climc/shell"
@@ -36,15 +36,24 @@ func chatStream(s *mcclient.ClientSession, args *options.MCPAgentMCPAgentRequest
 		return err
 	}
 
-	path := fmt.Sprintf("/mcp_agents/%s/chat-stream?message=%s", id, url.QueryEscape(args.Message))
+	bodyJSON, err := args.Params()
+	if err != nil {
+		return fmt.Errorf("failed to build request params: %v", err)
+	}
 
+	headers := http.Header{}
+	headers.Set("Content-Type", "application/json")
+
+	body := strings.NewReader(bodyJSON.String())
+
+	path := fmt.Sprintf("/mcp_agents/%s/chat-stream", id)
 	resp, err := s.RawVersionRequest(
 		modules.MCPAgent.ServiceType(),
 		modules.MCPAgent.EndpointType(),
-		"GET",
+		"POST",
 		path,
-		nil,
-		nil,
+		headers,
+		body,
 	)
 	if err != nil {
 		return err

pkg/apis/llm/llm.go

@@ -4,6 +4,7 @@ import (
 	"time"
 
 	"yunion.io/x/onecloud/pkg/apis"
+	computeapi "yunion.io/x/onecloud/pkg/apis/compute"
 	"yunion.io/x/onecloud/pkg/cloudcommon/db/taskman"
 )
 
@@ -64,8 +65,7 @@ type LLMBaseCreateInput struct {
 	PreferHost string `json:"prefer_host"`
 	AutoStart  bool   `json:"auto_start"`
 
-	NetworkType string `json:"network_type"`
-	NetworkId   string `json:"network_id"`
+	Nets []*computeapi.NetworkConfig `json:"nets"`
 
 	BandwidthMB   int  `json:"bandwidth_mb"`
 	DebugMode     bool `json:"debug_mode"`

pkg/apis/llm/mcp_agent.go

@@ -98,7 +98,14 @@ type LLMToolRequestInput struct {
 }
 
 type LLMMCPAgentRequestInput struct {
-	Message string `json:"message" help:"message to send to MCP agent"`
+	Message string                `json:"message" help:"message to send to MCP agent"`
+	History []MCPAgentChatMessage `json:"history" help:"chat history"`
+}
+
+type MCPAgentChatMessage struct {
+	Role    string `json:"role"`
+	Content string `json:"content"`
+	// ToolCalls []MCPAgentToolCallRecord `json:"tool_calls,omitempty"`
 }
 
 // MCPAgentResponse 表示 Agent 响应
@@ -115,6 +122,7 @@ type MCPAgentResponse struct {
 
 // MCPAgentToolCallRecord 记录工具调用
 type MCPAgentToolCallRecord struct {
+	Id        string                 `json:"id,omitempty"`
 	ToolName  string                 `json:"tool_name"`
 	Arguments map[string]interface{} `json:"arguments"`
 	Result    string                 `json:"result"`

pkg/apis/llm/ollama_const.go

@@ -17,10 +17,10 @@ const (
 	LLM_OLLAMA_HOST_MANIFESTS_DIR  = "/manifests"
 	LLM_OLLAMA_CACHE_DIR           = "/.llm_ollama_cache"
 	LLM_OLLAMA_CACHE_MOUNT_PATH    = "/usr/local"
-	LLM_OLLAMA_LIBRARY_BASE_URL    = `https://registry.ollama.ai/v2/library/%s`
+	LLM_OLLAMA_LIBRARY_BASE_URL    = `https://registry.ollama.ai/v2/%s`
 	LLM_OLLAMA_BASE_PATH           = "/root/.ollama/models"
 	LLM_OLLAMA_BLOBS_DIR           = "/blobs"
-	LLM_OLLAMA_MANIFESTS_BASE_PATH = "/manifests/registry.ollama.ai/library"
+	LLM_OLLAMA_MANIFESTS_BASE_PATH = "/manifests/registry.ollama.ai"
 )
 
 const (

pkg/apis/llm/sku.go

@@ -166,18 +166,17 @@ type LLMSkuListInput struct {
 
 type LLMSkuCreateInput struct {
 	LLMSKuBaseCreateInput
+	MountedModelResourceCreateInput
 
-	LLMImageId   string `json:"llm_image_id"`
-	LLMType      string `json:"llm_type"`
-	LLMModelName string `json:"llm_model_name"`
+	LLMImageId string `json:"llm_image_id"`
+	LLMType    string `json:"llm_type"`
 }
 
 type LLMSkuUpdateInput struct {
 	LLMSkuBaseUpdateInput
 	MountedModelResourceUpdateInput
 
-	LLMImageId   string `json:"llm_image_id"`
-	LLMModelName string `json:"llm_model_name"`
+	LLMImageId string `json:"llm_image_id"`
 }
 
 // type LLMModelCloneInput struct {

pkg/llm/drivers/llm_client/ollama.go

@@ -220,6 +220,13 @@ func (o *ollama) NewUserMessage(content string) models.ILLMChatMessage {
 	}
 }
 
+func (o *ollama) NewAssistantMessage(content string) models.ILLMChatMessage {
+	return &OllamaChatMessage{
+		Role:    "assistant",
+		Content: content,
+	}
+}
+
 func (o *ollama) NewAssistantMessageWithToolCalls(toolCalls []models.ILLMToolCall) models.ILLMChatMessage {
 	// to ollama tool calls
 	ollamaToolCalls := make([]OllamaToolCall, len(toolCalls))

pkg/llm/drivers/llm_client/openai.go

@@ -194,8 +194,12 @@ func (o *openai) doChatStreamRequest(ctx context.Context, mcpAgent *models.SMCPA
 		return errors.Wrap(err, "create request")
 	}
 	httpReq.Header.Set("Content-Type", "application/json")
-	if mcpAgent.ApiKey != "" {
-		httpReq.Header.Set("Authorization", "Bearer "+mcpAgent.ApiKey)
+	apiKey, err := mcpAgent.GetApiKey()
+	if err != nil {
+		return err
+	}
+	if apiKey != "" {
+		httpReq.Header.Set("Authorization", "Bearer "+apiKey)
 	}
 
 	client := &http.Client{
@@ -277,8 +281,12 @@ func (o *openai) doChatRequest(ctx context.Context, mcpAgent *models.SMCPAgent,
 		return nil, errors.Wrap(err, "create request")
 	}
 	httpReq.Header.Set("Content-Type", "application/json")
-	if mcpAgent.ApiKey != "" {
-		httpReq.Header.Set("Authorization", "Bearer "+mcpAgent.ApiKey)
+	apiKey, err := mcpAgent.GetApiKey()
+	if err != nil {
+		return nil, errors.Wrap(err, "get apiKey")
+	}
+	if apiKey != "" {
+		httpReq.Header.Set("Authorization", "Bearer "+apiKey)
 	}
 
 	client := &http.Client{
@@ -319,6 +327,13 @@ func (o *openai) NewUserMessage(content string) models.ILLMChatMessage {
 	}
 }
 
+func (o *openai) NewAssistantMessage(content string) models.ILLMChatMessage {
+	return &OpenAIChatMessage{
+		Role:    "assistant",
+		Content: content,
+	}
+}
+
 func (o *openai) NewAssistantMessageWithToolCalls(toolCalls []models.ILLMToolCall) models.ILLMChatMessage {
 	openaiToolCalls := make([]OpenAIToolCall, len(toolCalls))
 	for i, tc := range toolCalls {

pkg/llm/drivers/llm_container/ollama.go

@@ -48,7 +48,7 @@ func (o *ollama) GetContainerSpec(ctx context.Context, llm *models.SLLM, image *
 		},
 	}
 
-	if len(devices) == 0 && len(*sku.Devices) > 0 {
+	if len(devices) == 0 && (sku.Devices != nil && len(*sku.Devices) > 0) {
 		for i := range *sku.Devices {
 			index := i
 			spec.Devices = append(spec.Devices, &computeapi.ContainerDevice{
@@ -159,7 +159,8 @@ func (o *ollama) GetContainerSpec(ctx context.Context, llm *models.SLLM, image *
 
 func (o *ollama) DownloadModel(ctx context.Context, userCred mcclient.TokenCredential, llm *models.SLLM, tmpDir string, modelName string, modelTag string) (string, []string, error) {
 	// 1. download manifest from registry
-	manifestsUrl := fmt.Sprintf(api.LLM_OLLAMA_LIBRARY_BASE_URL, fmt.Sprintf("%s/manifests/%s", modelName, modelTag))
+	namespace, repo := getNamespaceAndRepo(modelName)
+	manifestsUrl := fmt.Sprintf(api.LLM_OLLAMA_LIBRARY_BASE_URL, fmt.Sprintf("%s/%s/manifests/%s", namespace, repo, modelTag))
 	log.Infof("Downloading manifest from %s", manifestsUrl)
 
 	manifestContent, err := llm.HttpGet(ctx, manifestsUrl)
@@ -187,9 +188,9 @@ func (o *ollama) DownloadModel(ctx context.Context, userCred mcclient.TokenCrede
 
 	// 4. create directory structure
 	// tmpDir/blobs/
-	// tmpDir/manifests/registry.ollama.ai/library/<modelName>/<modelTag>
+	// tmpDir/manifests/registry.ollama.ai/<namespace>/<repo>
 	blobsDir := path.Join(tmpDir, "blobs")
-	manifestsDir := path.Join(tmpDir, api.LLM_OLLAMA_MANIFESTS_BASE_PATH, modelName)
+	manifestsDir := path.Join(tmpDir, api.LLM_OLLAMA_MANIFESTS_BASE_PATH, namespace, repo)
 	if err := os.MkdirAll(blobsDir, 0755); err != nil {
 		return "", nil, errors.Wrapf(err, "failed to create blobs directory %s", blobsDir)
 	}
@@ -210,8 +211,8 @@ func (o *ollama) DownloadModel(ctx context.Context, userCred mcclient.TokenCrede
 		}
 
 		// download blob from registry
-		// URL format: https://registry.ollama.ai/v2/library/<modelName>/blobs/<digest>
-		blobUrl := fmt.Sprintf(api.LLM_OLLAMA_LIBRARY_BASE_URL, fmt.Sprintf("%s/blobs/%s", modelName, digest))
+		// URL format: https://registry.ollama.ai/v2/<namespace>/<repo>/blobs/<digest>
+		blobUrl := fmt.Sprintf(api.LLM_OLLAMA_LIBRARY_BASE_URL, fmt.Sprintf("%s/%s/blobs/%s", namespace, repo, digest))
 		log.Infof("Downloading blob %s from %s", blobFileName, blobUrl)
 
 		if err := llm.HttpDownloadFile(ctx, blobUrl, blobPath); err != nil {
@@ -232,7 +233,7 @@ func (o *ollama) DownloadModel(ctx context.Context, userCred mcclient.TokenCrede
 		blobFileName := strings.Replace(blob, ":", "-", 1)
 		mounts[i] = path.Join(api.LLM_OLLAMA_BASE_PATH, api.LLM_OLLAMA_BLOBS_DIR, blobFileName)
 	}
-	mounts = append(mounts, path.Join(api.LLM_OLLAMA_BASE_PATH, api.LLM_OLLAMA_MANIFESTS_BASE_PATH, modelName, modelTag))
+	mounts = append(mounts, path.Join(api.LLM_OLLAMA_BASE_PATH, api.LLM_OLLAMA_MANIFESTS_BASE_PATH, namespace, repo, modelTag))
 
 	return modelId, mounts, nil
 }
@@ -245,7 +246,8 @@ func (o *ollama) PreInstallModel(ctx context.Context, userCred mcclient.TokenCre
 	}
 
 	// mkdir llm-registry-base-path / modelname
-	mkdirReigtryBasePaht := fmt.Sprintf("mkdir -p %s", path.Join(api.LLM_OLLAMA_BASE_PATH, api.LLM_OLLAMA_MANIFESTS_BASE_PATH, instMdl.ModelName))
+	namespace, repo := getNamespaceAndRepo(instMdl.ModelName)
+	mkdirReigtryBasePaht := fmt.Sprintf("mkdir -p %s", path.Join(api.LLM_OLLAMA_BASE_PATH, api.LLM_OLLAMA_MANIFESTS_BASE_PATH, namespace, repo))
 	_, err = exec(ctx, lc.CmpId, mkdirReigtryBasePaht, 10)
 	if err != nil {
 		return errors.Wrap(err, "failed to mkdir llm-registry-base-path / modelname")
@@ -296,13 +298,11 @@ func (o *ollama) GetInstantModelIdByPostOverlay(postOverlay *commonapi.Container
 		for k := range postOverlay.Image.PathMap {
 			idx := strings.Index(k, api.LLM_OLLAMA_MANIFESTS_BASE_PATH)
 			if idx != -1 {
-				suffix := k[idx+len(api.LLM_OLLAMA_MANIFESTS_BASE_PATH):]
-				parts := strings.Split(strings.Trim(suffix, "/"), "/")
-				if len(parts) >= 2 {
-					modelName := parts[len(parts)-2]
-					modelTag := parts[len(parts)-1]
-					log.Infof("In GetInstantModelIdByPostOverlay, Extracted modelName: %s, modelTag: %s, Got modelId: %s", modelName, modelTag, mdlNameToId[modelName+":"+modelTag])
-					return mdlNameToId[modelName+":"+modelTag]
+				path := k[idx:]
+				nameTag := parseModelName(path)
+				if nameTag != "" {
+					log.Infof("In GetInstantModelIdByPostOverlay, Extracted nameTag: %s, Got modelId: %s", nameTag, mdlNameToId[nameTag])
+					return mdlNameToId[nameTag]
 				}
 			}
 		}
@@ -320,7 +320,8 @@ func (o *ollama) DetectModelPaths(ctx context.Context, userCred mcclient.TokenCr
 	for idx, blob := range pkgInfo.Blobs {
 		originBlobs[idx] = path.Join(api.LLM_OLLAMA_BASE_PATH, api.LLM_OLLAMA_BLOBS_DIR, blob)
 	}
-	originManifests := path.Join(api.LLM_OLLAMA_BASE_PATH, api.LLM_OLLAMA_MANIFESTS_BASE_PATH, pkgInfo.Name, pkgInfo.Tag)
+	namespace, repo := getNamespaceAndRepo(pkgInfo.Name)
+	originManifests := path.Join(api.LLM_OLLAMA_BASE_PATH, api.LLM_OLLAMA_MANIFESTS_BASE_PATH, namespace, repo, pkgInfo.Tag)
 
 	var checks []string
 	for _, blob := range originBlobs {
@@ -509,7 +510,8 @@ func exec(ctx context.Context, containerId string, cmd string, timeoutSec int64)
 }
 
 func getManifestsPath(modelName, modelTag string) string {
-	return path.Join(api.LLM_OLLAMA_BASE_PATH, api.LLM_OLLAMA_MANIFESTS_BASE_PATH, modelName, modelTag)
+	namespace, repo := getNamespaceAndRepo(modelName)
+	return path.Join(api.LLM_OLLAMA_BASE_PATH, api.LLM_OLLAMA_MANIFESTS_BASE_PATH, namespace, repo, modelTag)
 }
 
 type Layer struct {
@@ -591,6 +593,9 @@ func parseModelName(path string) string {
 		name := model[:lastSlash]
 		tag := model[lastSlash+1:]
 		tag = strings.TrimRight(tag, `\`)
+		if after, ok := strings.CutPrefix(name, "library/"); ok {
+			name = after
+		}
 		return name + ":" + tag
 	}
 	return strings.TrimRight(model, `\`)
@@ -642,3 +647,11 @@ func (o *ollama) GetLLMUrl(ctx context.Context, userCred mcclient.TokenCredentia
 		return fmt.Sprintf("http://%s:%d", server.HostAccessIp, accessInfo.AccessPort), nil
 	}
 }
+
+func getNamespaceAndRepo(modelName string) (string, string) {
+	if strings.Contains(modelName, "/") {
+		parts := strings.Split(modelName, "/")
+		return parts[0], parts[1]
+	}
+	return "library", modelName
+}

pkg/llm/models/dify_sku.go

@@ -41,15 +41,15 @@ type SDifySkuManager struct {
 type SDifySku struct {
 	SLLMSkuBase
 
-	PostgresImageId     string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required"`
-	RedisImageId        string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required"`
-	NginxImageId        string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required"`
-	DifyApiImageId      string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required"`
-	DifyPluginImageId   string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required"`
-	DifyWebImageId      string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required"`
-	DifySandboxImageId  string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required"`
-	DifySSRFImageId     string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required"`
-	DifyWeaviateImageId string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required"`
+	PostgresImageId     string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required" update:"user"`
+	RedisImageId        string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required" update:"user"`
+	NginxImageId        string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required" update:"user"`
+	DifyApiImageId      string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required" update:"user"`
+	DifyPluginImageId   string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required" update:"user"`
+	DifyWebImageId      string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required" update:"user"`
+	DifySandboxImageId  string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required" update:"user"`
+	DifySSRFImageId     string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required" update:"user"`
+	DifyWeaviateImageId string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required" update:"user"`
 }
 
 func (man *SDifySkuManager) ListItemFilter(

pkg/llm/models/llm.go

@@ -361,11 +361,7 @@ func (llm *SLLM) GetLLMSku(skuId string) (*SLLMSku, error) {
 
 func (llm *SLLM) GetLargeLanguageModelName(name string) (modelName string, modelTag string, err error) {
 	if name == "" {
-		sku, err := llm.GetLLMSku("")
-		if err != nil {
-			return "", "", err
-		}
-		name = sku.LLMModelName
+		return "", "", errors.Wrap(errors.ErrInvalidStatus, "model name is empty")
 	}
 	parts := strings.Split(name, ":")
 	modelName = parts[0]
@@ -578,3 +574,33 @@ func fetchNetworks(ctx context.Context, userCred mcclient.TokenCredential, netwo
 	}
 	return networks, nil
 }
+
+func (man *SLLMManager) GetAvailableNetwork(ctx context.Context, userCred mcclient.TokenCredential, query jsonutils.JSONObject) (jsonutils.JSONObject, error) {
+	s := auth.GetSession(ctx, userCred, "")
+
+	ret := jsonutils.NewDict()
+
+	q := jsonutils.NewDict()
+	if query != nil {
+		q.Update(query)
+	}
+	q.Set("server_type", jsonutils.NewString(string(computeapi.NETWORK_TYPE_HOSTLOCAL)))
+	q.Set("is_auto_alloc", jsonutils.NewBool(true))
+	q.Set("status", jsonutils.NewString(computeapi.NETWORK_STATUS_AVAILABLE))
+	q.Set("limit", jsonutils.NewInt(1))
+
+	result, err := compute.Networks.List(s, q)
+	if err == nil && result.Total > 0 {
+		ret.Add(jsonutils.NewInt(int64(result.Total)), "auto_alloc_network_hostlocal_count")
+	}
+
+	q.Set("server_type", jsonutils.NewString(string(computeapi.NETWORK_TYPE_GUEST)))
+	q.Set("vpc_id", jsonutils.NewString(computeapi.DEFAULT_VPC_ID))
+
+	resultGuest, err := compute.Networks.List(s, q)
+	if err == nil && resultGuest.Total > 0 {
+		ret.Add(jsonutils.NewInt(int64(resultGuest.Total)), "auto_alloc_network_guest_count")
+	}
+
+	return ret, nil
+}

pkg/llm/models/llm_base.go

@@ -97,23 +97,58 @@ func (man *SLLMBaseManager) ValidateCreateData(ctx context.Context, userCred mcc
 		input.PreferHost = hostDetails.Id
 	}
 
-	if len(input.NetworkType) > 0 && !api.IsLLMSkuBaseNetworkType(input.NetworkType) {
-		return input, errors.Wrapf(httperrors.ErrInputParameter, "invalid network type %s", input.NetworkType)
-	}
-
-	if len(input.NetworkId) > 0 {
-		s := auth.GetSession(ctx, userCred, "")
-		netObj, err := compute.Networks.Get(s, input.NetworkId, nil)
-		if err != nil {
-			return input, errors.Wrapf(httperrors.ErrInputParameter, "invalid network_id %s", input.NetworkId)
+	// 处理网络配置
+	var firstNet *computeapi.NetworkConfig
+	if len(input.Nets) > 0 {
+		firstNet = input.Nets[0]
+		firstNet.Index = 0
+		if len(string(firstNet.NetType)) > 0 && !api.IsLLMSkuBaseNetworkType(string(firstNet.NetType)) {
+			return input, errors.Wrapf(httperrors.ErrInputParameter, "invalid network type %s", firstNet.NetType)
 		}
-		input.NetworkId, _ = netObj.GetString("id")
-		input.NetworkType, _ = netObj.GetString("server_type")
+		if len(firstNet.Network) > 0 {
+			s := auth.GetSession(ctx, userCred, "")
+			netObj, err := compute.Networks.Get(s, firstNet.Network, nil)
+			if err != nil {
+				return input, errors.Wrapf(httperrors.ErrInputParameter, "invalid network_id %s", firstNet.Network)
+			}
+			netId, _ := netObj.GetString("id")
+			netType, _ := netObj.GetString("server_type")
+			firstNet.Network = netId
+			if len(string(firstNet.NetType)) == 0 {
+				firstNet.NetType = computeapi.TNetworkType(netType)
+			}
+		}
+	} else {
+		return input, errors.Wrap(httperrors.ErrInputParameter, "nets cannot be empty")
 	}
 
 	return input, nil
 }
 
+func (llmBase *SLLMBase) CustomizeCreate(ctx context.Context, userCred mcclient.TokenCredential, ownerId mcclient.IIdentityProvider, query jsonutils.JSONObject, data jsonutils.JSONObject) error {
+	err := llmBase.SVirtualResourceBase.CustomizeCreate(ctx, userCred, ownerId, query, data)
+	if err != nil {
+		return errors.Wrap(err, "SVirtualResourceBase.CustomizeCreate")
+	}
+
+	var input api.LLMBaseCreateInput
+	if err := data.Unmarshal(&input); err != nil {
+		return errors.Wrap(err, "unmarshal LLMBaseCreateInput")
+	}
+
+	if len(input.Nets) > 0 {
+		firstNet := input.Nets[0]
+		if len(string(firstNet.NetType)) > 0 {
+			llmBase.NetworkType = string(firstNet.NetType)
+		}
+		if len(firstNet.Network) > 0 {
+			llmBase.NetworkId = firstNet.Network
+		}
+	}
+
+	return nil
+}
+
 func GetServerIdsByHost(ctx context.Context, userCred mcclient.TokenCredential, hostId string) ([]string, error) {
 	s := auth.GetSession(ctx, userCred, options.Options.Region)
 	params := computeoptions.ServerListOptions{}

pkg/llm/models/llm_base_pod.go

@@ -109,21 +109,27 @@ func GetLLMBasePodCreateInput(
 			})
 		}
 	}
-	bandwidth := llmBase.BandwidthMb
-	if bandwidth == 0 {
-		bandwidth = skuBase.Bandwidth
+	var network *computeapi.NetworkConfig
+	if len(input.Nets) > 0 {
+		network = input.Nets[0]
+		networkCopy := *network
+		network = &networkCopy
+		network.Index = 0
 	}
 
-	network := &computeapi.NetworkConfig{
-		BwLimit: bandwidth,
-		NetType: computeapi.TNetworkType(llmBase.NetworkType),
+	bandwidth := input.BandwidthMB
+	if bandwidth == 0 && network.BwLimit != 0 {
+		bandwidth = network.BwLimit
 	}
-	if llmBase.NetworkType == string(computeapi.NETWORK_TYPE_HOSTLOCAL) {
+	if bandwidth == 0 && skuBase.Bandwidth != 0 {
+		bandwidth = skuBase.Bandwidth
+	}
+	network.BwLimit = bandwidth
+
+	networkType := string(network.NetType)
+	if networkType == string(computeapi.NETWORK_TYPE_HOSTLOCAL) {
 		network.PortMappings = portMappings
 	}
-	if len(llmBase.NetworkId) > 0 {
-		network.Network = llmBase.NetworkId
-	}
 
 	data.Networks = []*computeapi.NetworkConfig{
 		network,

pkg/llm/models/llm_client_driver.go

@@ -57,6 +57,7 @@ type ILLMClient interface {
 	ChatStream(ctx context.Context, mcpAgent *SMCPAgent, messages interface{}, tools interface{}, onChunk func(ILLMChatResponse) error) error
 
 	NewUserMessage(content string) ILLMChatMessage
+	NewAssistantMessage(content string) ILLMChatMessage
 	NewAssistantMessageWithToolCalls(toolCalls []ILLMToolCall) ILLMChatMessage
 	NewToolMessage(toolId string, toolName string, content string) ILLMChatMessage
 	NewSystemMessage(content string) ILLMChatMessage
@@ -64,6 +65,32 @@ type ILLMClient interface {
 	ConvertMCPTools(mcpTools []mcp.Tool) []ILLMTool
 }
 
+type SLLMToolCall struct {
+	Id       string
+	Function SLLMFunctionCall
+}
+
+func (tc *SLLMToolCall) GetId() string {
+	return tc.Id
+}
+
+func (tc *SLLMToolCall) GetFunction() ILLMFunctionCall {
+	return &tc.Function
+}
+
+type SLLMFunctionCall struct {
+	Name      string
+	Arguments map[string]interface{}
+}
+
+func (fc *SLLMFunctionCall) GetName() string {
+	return fc.Name
+}
+
+func (fc *SLLMFunctionCall) GetArguments() map[string]interface{} {
+	return fc.Arguments
+}
+
 var (
 	llmClientDrivers = newDrivers()
 )

pkg/llm/models/llm_save_instant_model.go

@@ -207,6 +207,9 @@ func (llm *SLLM) HttpGet(ctx context.Context, url string) ([]byte, error) {
 	defer resp.Body.Close()
 
 	if resp.StatusCode != http.StatusOK {
+		if resp.StatusCode == http.StatusNotFound {
+			return nil, httperrors.NewResourceNotFoundError("url %s not found", url)
+		}
 		return nil, errors.Errorf("unexpected status code: %d", resp.StatusCode)
 	}
 
@@ -231,6 +234,9 @@ func (llm *SLLM) HttpDownloadFile(ctx context.Context, url string, filePath stri
 	defer resp.Body.Close()
 
 	if resp.StatusCode != http.StatusOK {
+		if resp.StatusCode == http.StatusNotFound {
+			return errors.Wrapf(httperrors.ErrResourceNotFound, "url %s not found", url)
+		}
 		return errors.Errorf("unexpected status code: %d", resp.StatusCode)
 	}
 

pkg/llm/models/llm_sku.go

@@ -52,9 +52,8 @@ type SLLMSku struct {
 	SLLMSkuBase
 	SMountedModelsResource
 
-	LLMImageId   string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required"`
-	LLMType      string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required"`
-	LLMModelName string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required"`
+	LLMImageId string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required" update:"user"`
+	LLMType    string `width:"128" charset:"ascii" nullable:"false" list:"user" create:"required"`
 }
 
 func (man *SLLMSkuManager) ListItemFilter(
@@ -161,10 +160,29 @@ func (man *SLLMSkuManager) ValidateCreateData(ctx context.Context, userCred mccl
 		return input, errors.Wrap(httperrors.ErrInputParameter, "llm_type must be one of "+strings.Join(api.LLM_CONTAINER_TYPES.List(), ","))
 	}
 
-	_, err = validators.ValidateModel(ctx, userCred, GetLLMImageManager(), &input.LLMImageId)
+	imgObj, err := validators.ValidateModel(ctx, userCred, GetLLMImageManager(), &input.LLMImageId)
 	if err != nil {
 		return input, errors.Wrapf(err, "validate image_id %s", input.LLMImageId)
 	}
+	llmImage := imgObj.(*SLLMImage)
+	if llmImage.LLMType != input.LLMType {
+		return input, errors.Wrapf(httperrors.ErrInvalidStatus, "image %s is not of type %s", input.LLMImageId, input.LLMType)
+	}
+	input.LLMImageId = llmImage.Id
+
+	if input.MountedModels != nil {
+		for i, mdl := range input.MountedModels {
+			instMdl, err := GetInstantModelManager().FetchByIdOrName(ctx, userCred, mdl)
+			if err != nil {
+				return input, errors.Wrapf(err, "validate mounted model %s", mdl)
+			}
+			instantModle := instMdl.(*SInstantModel)
+			if instantModle.LlmType != input.LLMType {
+				return input, errors.Wrapf(httperrors.ErrInvalidStatus, "mounted model %s is not of type %s", mdl, input.LLMType)
+			}
+			input.MountedModels[i] = instantModle.GetId()
+		}
+	}
 
 	input.Status = api.STATUS_READY
 	return input, nil
@@ -200,7 +218,12 @@ func (sku *SLLMSku) ValidateUpdateData(ctx context.Context, userCred mcclient.To
 		if err != nil {
 			return input, errors.Wrapf(err, "validate image_id %s", input.LLMImageId)
 		}
-		input.LLMImageId = imgObj.GetId()
+		llmImage := imgObj.(*SLLMImage)
+		if llmImage.LLMType != sku.LLMType {
+			return input, errors.Wrapf(httperrors.ErrInvalidStatus, "image %s is not of type %s", input.LLMImageId, sku.LLMType)
+		}
+		input.LLMImageId = llmImage.Id
+		log.Infof("update llm_image_id %s to %s", sku.LLMImageId, input.LLMImageId)
 	}
 
 	return input, nil

pkg/llm/models/mcp_agent.go

@@ -10,6 +10,7 @@ import (
 	"yunion.io/x/jsonutils"
 	"yunion.io/x/log"
 	"yunion.io/x/pkg/errors"
+	seclib "yunion.io/x/pkg/utils"
 	"yunion.io/x/sqlchemy"
 
 	api "yunion.io/x/onecloud/pkg/apis/llm"
@@ -68,6 +69,48 @@ type SMCPAgent struct {
 	McpServer string `width:"512" charset:"utf8" nullable:"false" list:"user" create:"optional" update:"user"`
 }
 
+func (mcp *SMCPAgent) BeforeInsert() {
+	if len(mcp.Id) == 0 {
+		mcp.Id = db.DefaultUUIDGenerator()
+	}
+	if len(mcp.ApiKey) > 0 {
+		sec, err := seclib.EncryptAESBase64(mcp.Id, mcp.ApiKey)
+		if err != nil {
+			log.Errorf("EncryptAESBase64 fail %s", err)
+		} else {
+			mcp.ApiKey = sec
+		}
+	}
+	mcp.SSharableVirtualResourceBase.BeforeInsert()
+}
+
+func (mcp *SMCPAgent) BeforeUpdate() {
+	if len(mcp.ApiKey) > 0 {
+		// heuristic to check if it is plaintext
+		_, err := seclib.DescryptAESBase64(mcp.Id, mcp.ApiKey)
+		if err != nil {
+			sec, err := seclib.EncryptAESBase64(mcp.Id, mcp.ApiKey)
+			if err != nil {
+				log.Errorf("EncryptAESBase64 fail %s", err)
+			} else {
+				mcp.ApiKey = sec
+			}
+		}
+	}
+}
+
+func (mcp *SMCPAgent) GetApiKey() (string, error) {
+	if len(mcp.ApiKey) == 0 {
+		return "", nil
+	}
+	// try decrypt
+	key, err := seclib.DescryptAESBase64(mcp.Id, mcp.ApiKey)
+	if err == nil {
+		return key, nil
+	}
+	return mcp.ApiKey, nil
+}
+
 func (man *SMCPAgentManager) CustomizeHandlerInfo(info *appsrv.SHandlerInfo) {
 	man.SSharableVirtualResourceBaseManager.CustomizeHandlerInfo(info)
 
@@ -100,12 +143,20 @@ func (man *SMCPAgentManager) ValidateCreateData(ctx context.Context, userCred mc
 		}
 		input.LLMUrl = llmUrl
 
-		sku, err := llm.GetLLMSku("")
-		if err != nil {
-			return input, errors.Wrapf(err, "get LLM Sku from LLM %s", input.LLMId)
-		}
 		if len(input.Model) == 0 {
-			input.Model = sku.LLMModelName
+			mdlInfos, err := llm.getProbedInstantModelsExt(ctx, userCred)
+			if err != nil {
+				return input, errors.Wrap(err, "get probed models from LLM instance")
+			}
+			if len(mdlInfos) == 0 {
+				return input, httperrors.NewBadRequestError("no available models found in LLM instance %s", input.LLMId)
+			}
+			var firstModel api.LLMInternalInstantMdlInfo
+			for _, mdlInfo := range mdlInfos {
+				firstModel = mdlInfo
+				break
+			}
+			input.Model = fmt.Sprintf("%s:%s", firstModel.Name, firstModel.Tag)
 		}
 	}
 
@@ -159,11 +210,22 @@ func (man *SMCPAgentManager) ValidateUpdateData(ctx context.Context, userCred mc
 		}
 		input.LLMUrl = &llmUrl
 
-		sku, err := llm.GetLLMSku("")
-		if err != nil {
-			return input, errors.Wrapf(err, "get LLM Sku from LLM %s", *input.LLMId)
+		if input.Model == nil || len(*input.Model) == 0 {
+			mdlInfos, err := llm.getProbedInstantModelsExt(ctx, userCred)
+			if err != nil {
+				return input, errors.Wrap(err, "get probed models from LLM instance")
+			}
+			if len(mdlInfos) == 0 {
+				return input, httperrors.NewBadRequestError("no available models found in LLM instance %s", *input.LLMId)
+			}
+			var firstModel api.LLMInternalInstantMdlInfo
+			for _, mdlInfo := range mdlInfos {
+				firstModel = mdlInfo
+				break
+			}
+			modelStr := fmt.Sprintf("%s:%s", firstModel.Name, firstModel.Tag)
+			input.Model = &modelStr
 		}
-		input.Model = &sku.LLMModelName
 	}
 
 	// 如果更新 llm_driver，验证其值
@@ -294,9 +356,10 @@ func (mcp *SMCPAgent) GetDetailsToolRequest(
 // 	return jsonutils.Marshal(result), nil
 // }
 
-func (mcp *SMCPAgent) GetDetailsChatStream(
+func (mcp *SMCPAgent) PerformChatStream(
 	ctx context.Context,
 	userCred mcclient.TokenCredential,
+	query jsonutils.JSONObject,
 	input api.LLMMCPAgentRequestInput,
 ) (jsonutils.JSONObject, error) {
 	appParams := appsrv.AppContextGetParams(ctx)
@@ -361,11 +424,22 @@ func (mcp *SMCPAgent) process(ctx context.Context, userCred mcclient.TokenCreden
 	systemPrompt := buildSystemPrompt()
 
 	// 初始化消息历史
-	messages := []ILLMChatMessage{
-		llmClient.NewSystemMessage(systemPrompt),
-		llmClient.NewUserMessage(req.Message),
+	messages := make([]ILLMChatMessage, 0)
+	messages = append(messages, llmClient.NewSystemMessage(systemPrompt))
+
+	// 处理历史消息
+	if len(req.History) > 0 {
+		historyMessages := processHistoryMessages(
+			req.History,
+			llmClient,
+			options.Options.MCPAgentUserCharLimit,
+			options.Options.MCPAgentAssistantCharLimit,
+		)
+		messages = append(messages, historyMessages...)
 	}
 
+	messages = append(messages, llmClient.NewUserMessage(req.Message))
+
 	// 记录工具调用
 	var toolCallRecords []api.MCPAgentToolCallRecord
 
@@ -406,37 +480,14 @@ func (mcp *SMCPAgent) process(ctx context.Context, userCred mcclient.TokenCreden
 	toolCalls := resp.GetToolCalls()
 	log.Infof("Got %d tool calls from Phase 1", len(toolCalls))
 
-	// 将助手决定调用工具的消息加入历史
-	messages = append(messages, llmClient.NewAssistantMessageWithToolCalls(toolCalls))
-
-	// 执行每个工具调用
-	for _, tc := range toolCalls {
-		fc := tc.GetFunction()
-		toolName := fc.GetName()
-		arguments := fc.GetArguments()
-
-		if arguments == nil {
-			arguments = make(map[string]interface{})
-		}
-
-		log.Infof("Calling tool: %s with arguments: %v", toolName, arguments)
-
-		// 调用 MCP 工具
-		result, err := mcpClient.CallTool(ctx, toolName, arguments)
-		resultText := utils.FormatToolResult(toolName, result, err)
-		log.Infoln("Get result from mcp query", resultText)
-
-		// 记录
-		toolCallRecords = append(toolCallRecords, api.MCPAgentToolCallRecord{
-			ToolName:  toolName,
-			Arguments: arguments,
-			Result:    resultText,
-		})
-
-		// 将工具执行结果加入历史
-		messages = append(messages, llmClient.NewToolMessage(tc.GetId(), toolName, resultText))
+	toolCallRecords, toolMessages, err := processToolCalls(ctx, toolCalls, mcpClient, llmClient)
+	if err != nil {
+		return nil, errors.Wrap(err, "process tool calls")
 	}
 
+	// 将工具调用相关的消息加入历史
+	messages = append(messages, toolMessages...)
+
 	log.Infof("Phase 2: Streaming Response...")
 
 	var finalAnswer strings.Builder
@@ -472,3 +523,90 @@ func (mcp *SMCPAgent) process(ctx context.Context, userCred mcclient.TokenCreden
 func buildSystemPrompt() string {
 	return api.MCP_AGENT_SYSTEM_PROMPT
 }
+
+func processHistoryMessages(
+	history []api.MCPAgentChatMessage,
+	llmClient ILLMClient,
+	maxUserChars int,
+	maxAssistantChars int,
+) []ILLMChatMessage {
+	if len(history) == 0 {
+		return []ILLMChatMessage{}
+	}
+
+	var userChars, assistantChars int
+	processedMessages := make([]ILLMChatMessage, 0)
+
+	// 从最新的消息开始遍历，保留最新消息，丢弃最旧消息
+	for i := len(history) - 1; i >= 0; i-- {
+		msg := history[i]
+		msgChars := len(msg.Content)
+
+		switch msg.Role {
+		case "user":
+			if userChars+msgChars > maxUserChars {
+				break
+			}
+			userChars += msgChars
+			processedMessages = append(processedMessages, llmClient.NewUserMessage(msg.Content))
+		case "assistant":
+			if assistantChars+msgChars > maxAssistantChars {
+				break
+			}
+			assistantChars += msgChars
+
+			if len(msg.Content) > 0 {
+				processedMessages = append(processedMessages, llmClient.NewAssistantMessage(msg.Content))
+			}
+		}
+	}
+
+	for i, j := 0, len(processedMessages)-1; i < j; i, j = i+1, j-1 {
+		processedMessages[i], processedMessages[j] = processedMessages[j], processedMessages[i]
+	}
+
+	return processedMessages
+}
+
+// processToolCalls 处理工具调用
+func processToolCalls(
+	ctx context.Context,
+	toolCalls []ILLMToolCall,
+	mcpClient *utils.MCPClient,
+	llmClient ILLMClient,
+) ([]api.MCPAgentToolCallRecord, []ILLMChatMessage, error) {
+	toolCallRecords := make([]api.MCPAgentToolCallRecord, 0)
+	messagesToAdd := make([]ILLMChatMessage, 0)
+
+	messagesToAdd = append(messagesToAdd, llmClient.NewAssistantMessageWithToolCalls(toolCalls))
+
+	// 执行每个工具调用
+	for _, tc := range toolCalls {
+		fc := tc.GetFunction()
+		toolName := fc.GetName()
+		arguments := fc.GetArguments()
+
+		if arguments == nil {
+			arguments = make(map[string]interface{})
+		}
+
+		log.Infof("Calling tool: %s with arguments: %v", toolName, arguments)
+
+		// 调用 MCP 工具
+		result, err := mcpClient.CallTool(ctx, toolName, arguments)
+		resultText := utils.FormatToolResult(toolName, result, err)
+		log.Infoln("Get result from mcp query", resultText)
+
+		toolCallRecords = append(toolCallRecords, api.MCPAgentToolCallRecord{
+			Id:        tc.GetId(),
+			ToolName:  toolName,
+			Arguments: arguments,
+			Result:    resultText,
+		})
+
+		// 将工具执行结果加入历史
+		messagesToAdd = append(messagesToAdd, llmClient.NewToolMessage(tc.GetId(), toolName, resultText))
+	}
+
+	return toolCallRecords, messagesToAdd, nil
+}

pkg/llm/options/option.go

@@ -30,6 +30,9 @@ type LLMOptions struct {
 	// MCP Agent 配置
 	MCPServerURL    string `help:"MCP Server URL" default:"http://default-mcp-server:30876"`
 	MCPAgentTimeout int    `help:"MCP Agent request timeout in seconds" default:"120"`
+
+	MCPAgentUserCharLimit      int `help:"MCP Agent user char limit" default:"3200"`
+	MCPAgentAssistantCharLimit int `help:"MCP Agent assistant char limit" default:"6400"`
 }
 
 var (

pkg/llm/service/handler.go

@@ -2,13 +2,18 @@ package service
 
 import (
 	"context"
+	"fmt"
 	"net/http"
 
+	"yunion.io/x/jsonutils"
 	"yunion.io/x/onecloud/pkg/appsrv"
 	"yunion.io/x/onecloud/pkg/appsrv/dispatcher"
 	"yunion.io/x/onecloud/pkg/cloudcommon/db"
 	"yunion.io/x/onecloud/pkg/cloudcommon/db/taskman"
+	"yunion.io/x/onecloud/pkg/cloudcommon/policy"
+	"yunion.io/x/onecloud/pkg/httperrors"
 	"yunion.io/x/onecloud/pkg/llm/models"
+	"yunion.io/x/onecloud/pkg/mcclient/auth"
 )
 
 func handleOllamaRegistryYAML(ctx context.Context, w http.ResponseWriter, r *http.Request) {
@@ -17,6 +22,34 @@ func handleOllamaRegistryYAML(ctx context.Context, w http.ResponseWriter, r *htt
 	appsrv.Send(w, yamlContent)
 }
 
+func AddAvailableNetworkHandler(prefix string, app *appsrv.Application) {
+	app.AddHandler2("GET", fmt.Sprintf("%s/available-network", prefix), auth.Authenticate(handleLLMAvailableNetwork), nil, "get_llm_available_network", nil)
+}
+
+func handleLLMAvailableNetwork(ctx context.Context, w http.ResponseWriter, r *http.Request) {
+	userCred := auth.FetchUserCredential(ctx, policy.FilterPolicyCredential)
+	if userCred == nil {
+		httperrors.UnauthorizedError(ctx, w, "Unauthorized")
+		return
+	}
+	query, err := jsonutils.ParseQueryString(r.URL.RawQuery)
+	if err != nil {
+		httperrors.InvalidInputError(ctx, w, "Parse query string %q: %v", r.URL.RawQuery, err)
+		return
+	}
+	ret, err := models.GetLLMManager().GetAvailableNetwork(ctx, userCred, query)
+	if err != nil {
+		httperrors.GeneralServerError(ctx, w, err)
+		return
+	}
+
+	wrapped := jsonutils.NewDict()
+	if ret != nil {
+		wrapped.Add(ret, "llm")
+	}
+	appsrv.SendJSON(w, wrapped)
+}
+
 func InitHandlers(app *appsrv.Application, isSlave bool) {
 	db.InitAllManagers()
 	db.RegistUserCredCacheUpdater()
@@ -25,6 +58,8 @@ func InitHandlers(app *appsrv.Application, isSlave bool) {
 
 	app.AddHandler("GET", "/ollama-registry.yaml", handleOllamaRegistryYAML)
 
+	AddAvailableNetworkHandler(models.GetLLMManager().KeywordPlural(), app)
+
 	for _, manager := range []db.IModelManager{
 		taskman.TaskManager,
 		taskman.SubTaskManager,

pkg/mcclient/modules/llm/mod_llm.go

@@ -1,6 +1,11 @@
 package llm
 
 import (
+	"fmt"
+
+	"yunion.io/x/jsonutils"
+
+	"yunion.io/x/onecloud/pkg/mcclient"
 	"yunion.io/x/onecloud/pkg/mcclient/modulebase"
 	"yunion.io/x/onecloud/pkg/mcclient/modules"
 )
@@ -21,3 +26,14 @@ func init() {
 type LLMManager struct {
 	modulebase.ResourceManager
 }
+
+func (this *LLMManager) GetAvailableNetwork(session *mcclient.ClientSession, params jsonutils.JSONObject) (jsonutils.JSONObject, error) {
+	path := fmt.Sprintf("/%s/available-network", this.KeywordPlural)
+	if params != nil {
+		qs := params.QueryString()
+		if len(qs) > 0 {
+			path = fmt.Sprintf("%s?%s", path, qs)
+		}
+	}
+	return modulebase.Get(this.ResourceManager, session, path, this.Keyword)
+}

pkg/mcclient/options/llm/dify.go

@@ -2,7 +2,10 @@ package llm
 
 import (
 	"yunion.io/x/jsonutils"
+	"yunion.io/x/pkg/errors"
 
+	computeapi "yunion.io/x/onecloud/pkg/apis/compute"
+	"yunion.io/x/onecloud/pkg/cloudcommon/cmdline"
 	"yunion.io/x/onecloud/pkg/mcclient/options"
 )
 
@@ -38,7 +41,21 @@ type DifyCreateOptions struct {
 }
 
 func (o *DifyCreateOptions) Params() (jsonutils.JSONObject, error) {
-	return jsonutils.Marshal(o), nil
+	params := jsonutils.Marshal(o)
+
+	if len(o.Net) > 0 {
+		nets := make([]*computeapi.NetworkConfig, 0)
+		for i, n := range o.Net {
+			net, err := cmdline.ParseNetworkConfig(n, i)
+			if err != nil {
+				return nil, errors.Wrapf(err, "parse network config %s", n)
+			}
+			nets = append(nets, net)
+		}
+		params.(*jsonutils.JSONDict).Add(jsonutils.Marshal(nets), "nets")
+	}
+
+	return params, nil
 }
 
 func (o *DifyCreateOptions) GetCountParam() int {

pkg/mcclient/options/llm/llm.go

@@ -2,8 +2,11 @@ package llm
 
 import (
 	"yunion.io/x/jsonutils"
+	"yunion.io/x/pkg/errors"
 
+	computeapi "yunion.io/x/onecloud/pkg/apis/compute"
 	api "yunion.io/x/onecloud/pkg/apis/llm"
+	"yunion.io/x/onecloud/pkg/cloudcommon/cmdline"
 	"yunion.io/x/onecloud/pkg/mcclient/options"
 )
 
@@ -55,8 +58,7 @@ type LLMBaseCreateOptions struct {
 	ProjectId  string
 	PreferHost string
 
-	NETWORK_TYPE string `json:"network_type" choices:"guest|hostlocal"`
-	NetworkId    string `help:"id of network" json:"network_id"`
+	Net []string `help:"Network descriptions"`
 
 	BandwidthMb int
 
@@ -70,7 +72,21 @@ type LLMCreateOptions struct {
 }
 
 func (o *LLMCreateOptions) Params() (jsonutils.JSONObject, error) {
-	return jsonutils.Marshal(o), nil
+	params := jsonutils.Marshal(o)
+
+	if len(o.Net) > 0 {
+		nets := make([]*computeapi.NetworkConfig, 0)
+		for i, n := range o.Net {
+			net, err := cmdline.ParseNetworkConfig(n, i)
+			if err != nil {
+				return nil, errors.Wrapf(err, "parse network config %s", n)
+			}
+			nets = append(nets, net)
+		}
+		params.(*jsonutils.JSONDict).Add(jsonutils.Marshal(nets), "nets")
+	}
+
+	return params, nil
 }
 
 func (o *LLMCreateOptions) GetCountParam() int {
@@ -117,6 +133,19 @@ func (opts *LLMIdOptions) Params() (jsonutils.JSONObject, error) {
 	return jsonutils.Marshal(opts), nil
 }
 
+type LLMAvailableNetworkOptions struct {
+	NetworkType string `help:"network server_type filter, e.g. guest|hostlocal"`
+	VpcId       string `help:"vpc id filter"`
+}
+
+func (opts *LLMAvailableNetworkOptions) GetId() string {
+	return ""
+}
+
+func (opts *LLMAvailableNetworkOptions) Params() (jsonutils.JSONObject, error) {
+	return options.StructToParams(opts)
+}
+
 type LLMSaveInstantModelOptions struct {
 	LLMIdOptions
 

pkg/mcclient/options/llm/llm_sku.go

@@ -27,9 +27,10 @@ func (o *LLMSkuShowOptions) Params() (jsonutils.JSONObject, error) {
 type LLMSkuCreateOptions struct {
 	LLMSkuBaseCreateOptions
 
-	LLM_IMAGE_ID   string `json:"llm_image_id"`
-	LLM_TYPE       string `json:"llm_type" choices:"ollama"`
-	LLM_MODEL_NAME string `help:"specific model of large language model, for example: qwen3:32b" json:"llm_model_name"`
+	MountedModels []string `help:"mounted models, <model_id> e.g. qwen2:0.5b-dup" json:"mounted_models"`
+
+	LLM_IMAGE_ID string `json:"llm_image_id"`
+	LLM_TYPE     string `json:"llm_type" choices:"ollama"`
 }
 
 func (o *LLMSkuCreateOptions) Params() (jsonutils.JSONObject, error) {
@@ -38,6 +39,7 @@ func (o *LLMSkuCreateOptions) Params() (jsonutils.JSONObject, error) {
 	obj.Unmarshal(dict)
 
 	o.LLMSkuBaseCreateOptions.Params(dict)
+	fetchMountedModels(o.MountedModels, dict)
 	return dict, nil
 }
 
@@ -56,10 +58,9 @@ func (o *LLMSkuDeleteOptions) Params() (jsonutils.JSONObject, error) {
 type LLMSkuUpdateOptions struct {
 	LLMSkuBaseUpdateOptions
 
-	MountedModels []string `help:"mounted models, <model_id>@<model_name>:<model_tag> e.g. 6f48b936a09f@qwen2:0.5b" json:"mounted_models"`
+	MountedModels []string `help:"mounted models, <model_id> e.g. qwen2:0.5b-dup" json:"mounted_models"`
 
-	LlmImageId   string
-	LlmModelName string
+	LlmImageId string `json:"llm_image_id"`
 }
 
 func (o *LLMSkuUpdateOptions) GetId() string {

pkg/mcclient/options/llm/mcp_agent.go

@@ -1,6 +1,7 @@
 package llm
 
 import (
+	"fmt"
 	"strings"
 
 	"yunion.io/x/jsonutils"
@@ -143,12 +144,28 @@ func (opts *MCPAgentToolRequestOptions) Params() (jsonutils.JSONObject, error) {
 type MCPAgentMCPAgentRequestOptions struct {
 	MCPAgentIdOptions
 
-	Message string `help:"message to send to MCP agent" json:"message"`
+	MESSAGE string `help:"message to send to MCP agent" json:"message"`
+	History string `help:"chat history as JSON string, e.g. '[{\"role\":\"user\",\"content\":\"hello\"}]'" json:"history,omitempty"`
 }
 
 func (opts *MCPAgentMCPAgentRequestOptions) Params() (jsonutils.JSONObject, error) {
 	input := api.LLMMCPAgentRequestInput{
-		Message: opts.Message,
+		Message: opts.MESSAGE,
+		History: []api.MCPAgentChatMessage{},
 	}
+
+	if len(opts.History) > 0 {
+		historyJSON, err := jsonutils.ParseString(opts.History)
+		if err != nil {
+			return nil, fmt.Errorf("failed to parse history JSON: %v", err)
+		}
+		if historyJSON != nil {
+			err = historyJSON.Unmarshal(&input.History)
+			if err != nil {
+				return nil, fmt.Errorf("failed to unmarshal history: %v", err)
+			}
+		}
+	}
+
 	return jsonutils.Marshal(input), nil
 }