update doc api

Made-with: Cursor

aiagent/builtin_tools.go

@@ -0,0 +1,546 @@
+package aiagent
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/ccfos/nightingale/v6/datasource"
+	"github.com/ccfos/nightingale/v6/dscache"
+	"github.com/ccfos/nightingale/v6/models"
+	"github.com/ccfos/nightingale/v6/pkg/prom"
+	"github.com/toolkits/pkg/logger"
+)
+
+const (
+	// ToolTypeBuiltin 内置工具类型
+	ToolTypeBuiltin = "builtin"
+)
+
+// =============================================================================
+// 数据源获取函数（支持注入，便于测试）
+// =============================================================================
+
+// PromClientGetter Prometheus 客户端获取函数类型
+type PromClientGetter func(dsId int64) prom.API
+
+// SQLDatasourceGetter SQL 数据源获取函数类型
+type SQLDatasourceGetter func(dsType string, dsId int64) (datasource.Datasource, bool)
+
+// 默认使用 GlobalCache，可通过 SetPromClientGetter/SetSQLDatasourceGetter 替换
+var (
+	getPromClientFunc     PromClientGetter     = defaultGetPromClient
+	getSQLDatasourceFunc  SQLDatasourceGetter  = defaultGetSQLDatasource
+)
+
+// SetPromClientGetter 设置 Prometheus 客户端获取函数（用于测试）
+func SetPromClientGetter(getter PromClientGetter) {
+	getPromClientFunc = getter
+}
+
+// SetSQLDatasourceGetter 设置 SQL 数据源获取函数（用于测试）
+func SetSQLDatasourceGetter(getter SQLDatasourceGetter) {
+	getSQLDatasourceFunc = getter
+}
+
+// ResetDatasourceGetters 重置为默认的数据源获取函数
+func ResetDatasourceGetters() {
+	getPromClientFunc = defaultGetPromClient
+	getSQLDatasourceFunc = defaultGetSQLDatasource
+}
+
+func defaultGetPromClient(dsId int64) prom.API {
+	// Default: no PromClient available. Use SetPromClientGetter to inject.
+	return nil
+}
+
+func defaultGetSQLDatasource(dsType string, dsId int64) (datasource.Datasource, bool) {
+	return dscache.DsCache.Get(dsType, dsId)
+}
+
+// BuiltinToolHandler 内置工具处理函数
+type BuiltinToolHandler func(ctx context.Context, wfCtx *models.WorkflowContext, args map[string]interface{}) (string, error)
+
+// BuiltinTool 内置工具定义
+type BuiltinTool struct {
+	Definition AgentTool
+	Handler    BuiltinToolHandler
+}
+
+// builtinTools 内置工具注册表
+var builtinTools = map[string]*BuiltinTool{
+	// Prometheus 相关工具
+	"list_metrics": {
+		Definition: AgentTool{
+			Name:        "list_metrics",
+			Description: "搜索 Prometheus 数据源的指标名称，支持关键词模糊匹配",
+			Type:        ToolTypeBuiltin,
+			Parameters: []ToolParameter{
+				{Name: "keyword", Type: "string", Description: "搜索关键词，模糊匹配指标名", Required: false},
+				{Name: "limit", Type: "integer", Description: "返回数量限制，默认30", Required: false},
+			},
+		},
+		Handler: listMetrics,
+	},
+	"get_metric_labels": {
+		Definition: AgentTool{
+			Name:        "get_metric_labels",
+			Description: "获取 Prometheus 指标的所有标签键及其可选值",
+			Type:        ToolTypeBuiltin,
+			Parameters: []ToolParameter{
+				{Name: "metric", Type: "string", Description: "指标名称", Required: true},
+			},
+		},
+		Handler: getMetricLabels,
+	},
+
+	// SQL 类数据源相关工具
+	"list_databases": {
+		Definition: AgentTool{
+			Name:        "list_databases",
+			Description: "列出 SQL 数据源（MySQL/Doris/ClickHouse/PostgreSQL）中的所有数据库",
+			Type:        ToolTypeBuiltin,
+			Parameters:  []ToolParameter{},
+		},
+		Handler: listDatabases,
+	},
+	"list_tables": {
+		Definition: AgentTool{
+			Name:        "list_tables",
+			Description: "列出指定数据库中的所有表",
+			Type:        ToolTypeBuiltin,
+			Parameters: []ToolParameter{
+				{Name: "database", Type: "string", Description: "数据库名", Required: true},
+			},
+		},
+		Handler: listTables,
+	},
+	"describe_table": {
+		Definition: AgentTool{
+			Name:        "describe_table",
+			Description: "获取表的字段结构（字段名、类型、注释）",
+			Type:        ToolTypeBuiltin,
+			Parameters: []ToolParameter{
+				{Name: "database", Type: "string", Description: "数据库名", Required: true},
+				{Name: "table", Type: "string", Description: "表名", Required: true},
+			},
+		},
+		Handler: describeTable,
+	},
+}
+
+// GetBuiltinToolDef 获取内置工具定义
+func GetBuiltinToolDef(name string) (AgentTool, bool) {
+	if tool, ok := builtinTools[name]; ok {
+		return tool.Definition, true
+	}
+	return AgentTool{}, false
+}
+
+// GetBuiltinToolDefs 获取指定的内置工具定义列表
+func GetBuiltinToolDefs(names []string) []AgentTool {
+	var defs []AgentTool
+	for _, name := range names {
+		if def, ok := GetBuiltinToolDef(name); ok {
+			defs = append(defs, def)
+		}
+	}
+	return defs
+}
+
+// GetAllBuiltinToolDefs 获取所有内置工具定义
+func GetAllBuiltinToolDefs() []AgentTool {
+	defs := make([]AgentTool, 0, len(builtinTools))
+	for _, tool := range builtinTools {
+		defs = append(defs, tool.Definition)
+	}
+	return defs
+}
+
+// ExecuteBuiltinTool 执行内置工具
+// 返回值：result, handled, error
+// handled 表示是否是内置工具（true 表示已处理，false 表示不是内置工具需要继续查找）
+func ExecuteBuiltinTool(ctx context.Context, name string, wfCtx *models.WorkflowContext, argsJSON string) (string, bool, error) {
+	tool, exists := builtinTools[name]
+	if !exists {
+		return "", false, nil
+	}
+
+	// 解析参数
+	var args map[string]interface{}
+	if argsJSON != "" {
+		if err := json.Unmarshal([]byte(argsJSON), &args); err != nil {
+			// 如果不是 JSON，尝试作为简单字符串参数
+			args = map[string]interface{}{"input": argsJSON}
+		}
+	}
+	if args == nil {
+		args = make(map[string]interface{})
+	}
+
+	result, err := tool.Handler(ctx, wfCtx, args)
+	return result, true, err
+}
+
+// getDatasourceId 从 wfCtx.Inputs 中获取 datasource_id
+func getDatasourceId(wfCtx *models.WorkflowContext) int64 {
+	if wfCtx == nil || wfCtx.Inputs == nil {
+		return 0
+	}
+	var dsId int64
+	if dsIdStr, ok := wfCtx.Inputs["datasource_id"]; ok {
+		fmt.Sscanf(dsIdStr, "%d", &dsId)
+	}
+	return dsId
+}
+
+// getDatasourceType 从 wfCtx.Inputs 中获取 datasource_type
+func getDatasourceType(wfCtx *models.WorkflowContext) string {
+	if wfCtx == nil || wfCtx.Inputs == nil {
+		return ""
+	}
+	return wfCtx.Inputs["datasource_type"]
+}
+
+// =============================================================================
+// Prometheus 工具实现
+// =============================================================================
+
+// listMetrics 列出 Prometheus 指标
+func listMetrics(ctx context.Context, wfCtx *models.WorkflowContext, args map[string]interface{}) (string, error) {
+	dsId := getDatasourceId(wfCtx)
+	if dsId == 0 {
+		return "", fmt.Errorf("datasource_id not found in inputs")
+	}
+
+	keyword, _ := args["keyword"].(string)
+	limit := 30
+	if l, ok := args["limit"].(float64); ok && l > 0 {
+		limit = int(l)
+	}
+
+	// 获取 Prometheus 客户端
+	client := getPromClientFunc(dsId)
+	if client == nil {
+		return "", fmt.Errorf("prometheus datasource not found: %d", dsId)
+	}
+
+	// 调用 LabelValues 获取 __name__ 的所有值（即所有指标名）
+	values, _, err := client.LabelValues(ctx, "__name__", nil)
+	if err != nil {
+		return "", fmt.Errorf("failed to get metrics: %v", err)
+	}
+
+	// 过滤和限制
+	result := make([]string, 0)
+	keyword = strings.ToLower(keyword)
+	for _, v := range values {
+		m := string(v)
+		if keyword == "" || strings.Contains(strings.ToLower(m), keyword) {
+			result = append(result, m)
+			if len(result) >= limit {
+				break
+			}
+		}
+	}
+
+	logger.Debugf("list_metrics: found %d metrics (keyword=%s, limit=%d)", len(result), keyword, limit)
+
+	bytes, _ := json.Marshal(result)
+	return string(bytes), nil
+}
+
+// getMetricLabels 获取指标的标签
+func getMetricLabels(ctx context.Context, wfCtx *models.WorkflowContext, args map[string]interface{}) (string, error) {
+	dsId := getDatasourceId(wfCtx)
+	if dsId == 0 {
+		return "", fmt.Errorf("datasource_id not found in inputs")
+	}
+
+	metric, ok := args["metric"].(string)
+	if !ok || metric == "" {
+		return "", fmt.Errorf("metric parameter is required")
+	}
+
+	client := getPromClientFunc(dsId)
+	if client == nil {
+		return "", fmt.Errorf("prometheus datasource not found: %d", dsId)
+	}
+
+	// 使用 Series 接口获取指标的所有 series
+	endTime := time.Now()
+	startTime := endTime.Add(-1 * time.Hour)
+	series, _, err := client.Series(ctx, []string{metric}, startTime, endTime)
+	if err != nil {
+		return "", fmt.Errorf("failed to get metric series: %v", err)
+	}
+
+	// 聚合标签键值
+	labels := make(map[string][]string)
+	seen := make(map[string]map[string]bool)
+
+	for _, s := range series {
+		for k, v := range s {
+			key := string(k)
+			val := string(v)
+			if key == "__name__" {
+				continue
+			}
+			if seen[key] == nil {
+				seen[key] = make(map[string]bool)
+			}
+			if !seen[key][val] {
+				seen[key][val] = true
+				labels[key] = append(labels[key], val)
+			}
+		}
+	}
+
+	logger.Debugf("get_metric_labels: metric=%s, found %d labels", metric, len(labels))
+
+	bytes, _ := json.Marshal(labels)
+	return string(bytes), nil
+}
+
+// =============================================================================
+// SQL 数据源工具实现
+// =============================================================================
+
+// SQLMetadataQuerier SQL 元数据查询接口
+type SQLMetadataQuerier interface {
+	ListDatabases(ctx context.Context) ([]string, error)
+	ListTables(ctx context.Context, database string) ([]string, error)
+	DescribeTable(ctx context.Context, database, table string) ([]map[string]interface{}, error)
+}
+
+// listDatabases 列出数据库
+func listDatabases(ctx context.Context, wfCtx *models.WorkflowContext, args map[string]interface{}) (string, error) {
+	dsId := getDatasourceId(wfCtx)
+	dsType := getDatasourceType(wfCtx)
+	if dsId == 0 {
+		return "", fmt.Errorf("datasource_id not found in inputs")
+	}
+	if dsType == "" {
+		return "", fmt.Errorf("datasource_type not found in inputs")
+	}
+
+	plug, exists := getSQLDatasourceFunc(dsType, dsId)
+	if !exists {
+		return "", fmt.Errorf("datasource not found: %s/%d", dsType, dsId)
+	}
+
+	// 构建查询 SQL
+	var sql string
+	switch dsType {
+	case "mysql", "doris":
+		sql = "SHOW DATABASES"
+	case "ck", "clickhouse":
+		sql = "SHOW DATABASES"
+	case "pgsql", "postgresql":
+		sql = "SELECT datname FROM pg_database WHERE datistemplate = false"
+	default:
+		return "", fmt.Errorf("unsupported datasource type for list_databases: %s", dsType)
+	}
+
+	// 执行查询
+	query := map[string]interface{}{"sql": sql}
+	data, _, err := plug.QueryLog(ctx, query)
+	if err != nil {
+		return "", fmt.Errorf("failed to list databases: %v", err)
+	}
+
+	// 提取数据库名
+	databases := extractColumnValues(data, dsType, "database")
+
+	logger.Debugf("list_databases: dsType=%s, found %d databases", dsType, len(databases))
+
+	bytes, _ := json.Marshal(databases)
+	return string(bytes), nil
+}
+
+// listTables 列出表
+func listTables(ctx context.Context, wfCtx *models.WorkflowContext, args map[string]interface{}) (string, error) {
+	dsId := getDatasourceId(wfCtx)
+	dsType := getDatasourceType(wfCtx)
+	if dsId == 0 {
+		return "", fmt.Errorf("datasource_id not found in inputs")
+	}
+
+	database, ok := args["database"].(string)
+	if !ok || database == "" {
+		return "", fmt.Errorf("database parameter is required")
+	}
+
+	plug, exists := getSQLDatasourceFunc(dsType, dsId)
+	if !exists {
+		return "", fmt.Errorf("datasource not found: %s/%d", dsType, dsId)
+	}
+
+	// 构建查询 SQL
+	var sql string
+	switch dsType {
+	case "mysql", "doris":
+		sql = fmt.Sprintf("SHOW TABLES FROM `%s`", database)
+	case "ck", "clickhouse":
+		sql = fmt.Sprintf("SHOW TABLES FROM `%s`", database)
+	case "pgsql", "postgresql":
+		sql = fmt.Sprintf("SELECT tablename FROM pg_tables WHERE schemaname = 'public'")
+	default:
+		return "", fmt.Errorf("unsupported datasource type for list_tables: %s", dsType)
+	}
+
+	// 执行查询
+	query := map[string]interface{}{"sql": sql, "database": database}
+	data, _, err := plug.QueryLog(ctx, query)
+	if err != nil {
+		return "", fmt.Errorf("failed to list tables: %v", err)
+	}
+
+	// 提取表名
+	tables := extractColumnValues(data, dsType, "table")
+
+	logger.Debugf("list_tables: dsType=%s, database=%s, found %d tables", dsType, database, len(tables))
+
+	bytes, _ := json.Marshal(tables)
+	return string(bytes), nil
+}
+
+// describeTable 获取表结构
+func describeTable(ctx context.Context, wfCtx *models.WorkflowContext, args map[string]interface{}) (string, error) {
+	dsId := getDatasourceId(wfCtx)
+	dsType := getDatasourceType(wfCtx)
+	if dsId == 0 {
+		return "", fmt.Errorf("datasource_id not found in inputs")
+	}
+
+	database, ok := args["database"].(string)
+	if !ok || database == "" {
+		return "", fmt.Errorf("database parameter is required")
+	}
+	table, ok := args["table"].(string)
+	if !ok || table == "" {
+		return "", fmt.Errorf("table parameter is required")
+	}
+
+	plug, exists := getSQLDatasourceFunc(dsType, dsId)
+	if !exists {
+		return "", fmt.Errorf("datasource not found: %s/%d", dsType, dsId)
+	}
+
+	// 构建查询 SQL
+	var sql string
+	switch dsType {
+	case "mysql", "doris":
+		sql = fmt.Sprintf("DESCRIBE `%s`.`%s`", database, table)
+	case "ck", "clickhouse":
+		sql = fmt.Sprintf("DESCRIBE TABLE `%s`.`%s`", database, table)
+	case "pgsql", "postgresql":
+		sql = fmt.Sprintf(`SELECT column_name as "Field", data_type as "Type", is_nullable as "Null", column_default as "Default" FROM information_schema.columns WHERE table_schema = 'public' AND table_name = '%s'`, table)
+	default:
+		return "", fmt.Errorf("unsupported datasource type for describe_table: %s", dsType)
+	}
+
+	// 执行查询
+	query := map[string]interface{}{"sql": sql, "database": database}
+	data, _, err := plug.QueryLog(ctx, query)
+	if err != nil {
+		return "", fmt.Errorf("failed to describe table: %v", err)
+	}
+
+	// 转换为统一的列结构
+	columns := convertToColumnInfo(data, dsType)
+
+	logger.Debugf("describe_table: dsType=%s, table=%s.%s, found %d columns", dsType, database, table, len(columns))
+
+	bytes, _ := json.Marshal(columns)
+	return string(bytes), nil
+}
+
+// ColumnInfo 列信息
+type ColumnInfo struct {
+	Name    string `json:"name"`
+	Type    string `json:"type"`
+	Comment string `json:"comment,omitempty"`
+}
+
+// extractColumnValues 从查询结果中提取列值
+func extractColumnValues(data []interface{}, dsType string, columnType string) []string {
+	result := make([]string, 0)
+	for _, row := range data {
+		if rowMap, ok := row.(map[string]interface{}); ok {
+			// 尝试多种可能的列名
+			var value string
+			for _, key := range getPossibleColumnNames(dsType, columnType) {
+				if v, ok := rowMap[key]; ok {
+					if s, ok := v.(string); ok {
+						value = s
+						break
+					}
+				}
+			}
+			if value != "" {
+				result = append(result, value)
+			}
+		}
+	}
+	return result
+}
+
+// getPossibleColumnNames 获取可能的列名
+func getPossibleColumnNames(dsType string, columnType string) []string {
+	switch columnType {
+	case "database":
+		return []string{"Database", "database", "datname", "name"}
+	case "table":
+		return []string{"Tables_in_", "table", "tablename", "name", "Name"}
+	default:
+		return []string{}
+	}
+}
+
+// convertToColumnInfo 将查询结果转换为统一的列信息格式
+func convertToColumnInfo(data []interface{}, dsType string) []ColumnInfo {
+	result := make([]ColumnInfo, 0)
+	for _, row := range data {
+		if rowMap, ok := row.(map[string]interface{}); ok {
+			col := ColumnInfo{}
+
+			// 提取列名
+			for _, key := range []string{"Field", "field", "column_name", "name"} {
+				if v, ok := rowMap[key]; ok {
+					if s, ok := v.(string); ok {
+						col.Name = s
+						break
+					}
+				}
+			}
+
+			// 提取类型
+			for _, key := range []string{"Type", "type", "data_type"} {
+				if v, ok := rowMap[key]; ok {
+					if s, ok := v.(string); ok {
+						col.Type = s
+						break
+					}
+				}
+			}
+
+			// 提取注释（可选）
+			for _, key := range []string{"Comment", "comment", "column_comment"} {
+				if v, ok := rowMap[key]; ok {
+					if s, ok := v.(string); ok {
+						col.Comment = s
+						break
+					}
+				}
+			}
+
+			if col.Name != "" {
+				result = append(result, col)
+			}
+		}
+	}
+	return result
+}

aiagent/llm/claude.go

@@ -0,0 +1,376 @@
+package llm
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+)
+
+const (
+	DefaultClaudeURL     = "https://api.anthropic.com/v1/messages"
+	ClaudeAPIVersion     = "2023-06-01"
+	DefaultClaudeMaxTokens = 4096
+)
+
+// Claude implements the LLM interface for Anthropic Claude API
+type Claude struct {
+	config *Config
+	client *http.Client
+}
+
+// NewClaude creates a new Claude provider
+func NewClaude(cfg *Config, client *http.Client) (*Claude, error) {
+	if cfg.BaseURL == "" {
+		cfg.BaseURL = DefaultClaudeURL
+	}
+	return &Claude{
+		config: cfg,
+		client: client,
+	}, nil
+}
+
+func (c *Claude) Name() string {
+	return ProviderClaude
+}
+
+// Claude API request/response structures
+type claudeRequest struct {
+	Model       string          `json:"model"`
+	Messages    []claudeMessage `json:"messages"`
+	System      string          `json:"system,omitempty"`
+	MaxTokens   int             `json:"max_tokens"`
+	Temperature float64         `json:"temperature,omitempty"`
+	TopP        float64         `json:"top_p,omitempty"`
+	Stop        []string        `json:"stop_sequences,omitempty"`
+	Stream      bool            `json:"stream,omitempty"`
+	Tools       []claudeTool    `json:"tools,omitempty"`
+}
+
+type claudeMessage struct {
+	Role    string               `json:"role"`
+	Content []claudeContentBlock `json:"content"`
+}
+
+type claudeContentBlock struct {
+	Type      string `json:"type"`
+	Text      string `json:"text,omitempty"`
+	ID        string `json:"id,omitempty"`
+	Name      string `json:"name,omitempty"`
+	Input     any    `json:"input,omitempty"`
+	ToolUseID string `json:"tool_use_id,omitempty"`
+	Content   string `json:"content,omitempty"`
+}
+
+type claudeTool struct {
+	Name        string                 `json:"name"`
+	Description string                 `json:"description"`
+	InputSchema map[string]interface{} `json:"input_schema"`
+}
+
+type claudeResponse struct {
+	ID           string               `json:"id"`
+	Type         string               `json:"type"`
+	Role         string               `json:"role"`
+	Content      []claudeContentBlock `json:"content"`
+	Model        string               `json:"model"`
+	StopReason   string               `json:"stop_reason"`
+	StopSequence string               `json:"stop_sequence,omitempty"`
+	Usage        *struct {
+		InputTokens  int `json:"input_tokens"`
+		OutputTokens int `json:"output_tokens"`
+	} `json:"usage,omitempty"`
+	Error *struct {
+		Type    string `json:"type"`
+		Message string `json:"message"`
+	} `json:"error,omitempty"`
+}
+
+// Claude streaming event types
+type claudeStreamEvent struct {
+	Type         string               `json:"type"`
+	Index        int                  `json:"index,omitempty"`
+	ContentBlock *claudeContentBlock  `json:"content_block,omitempty"`
+	Delta        *claudeStreamDelta   `json:"delta,omitempty"`
+	Message      *claudeResponse      `json:"message,omitempty"`
+	Usage        *claudeStreamUsage   `json:"usage,omitempty"`
+}
+
+type claudeStreamDelta struct {
+	Type         string `json:"type"`
+	Text         string `json:"text,omitempty"`
+	PartialJSON  string `json:"partial_json,omitempty"`
+	StopReason   string `json:"stop_reason,omitempty"`
+}
+
+type claudeStreamUsage struct {
+	OutputTokens int `json:"output_tokens"`
+}
+
+func (c *Claude) Generate(ctx context.Context, req *GenerateRequest) (*GenerateResponse, error) {
+	claudeReq := c.convertRequest(req)
+	claudeReq.Stream = false
+
+	respBody, err := c.doRequest(ctx, claudeReq)
+	if err != nil {
+		return nil, err
+	}
+
+	var claudeResp claudeResponse
+	if err := json.Unmarshal(respBody, &claudeResp); err != nil {
+		return nil, fmt.Errorf("failed to parse response: %w", err)
+	}
+
+	if claudeResp.Error != nil {
+		return nil, fmt.Errorf("Claude API error: %s", claudeResp.Error.Message)
+	}
+
+	return c.convertResponse(&claudeResp), nil
+}
+
+func (c *Claude) GenerateStream(ctx context.Context, req *GenerateRequest) (<-chan StreamChunk, error) {
+	claudeReq := c.convertRequest(req)
+	claudeReq.Stream = true
+
+	jsonData, err := json.Marshal(claudeReq)
+	if err != nil {
+		return nil, fmt.Errorf("failed to marshal request: %w", err)
+	}
+
+	httpReq, err := http.NewRequestWithContext(ctx, "POST", c.config.BaseURL, bytes.NewBuffer(jsonData))
+	if err != nil {
+		return nil, fmt.Errorf("failed to create request: %w", err)
+	}
+
+	c.setHeaders(httpReq)
+
+	resp, err := c.client.Do(httpReq)
+	if err != nil {
+		return nil, fmt.Errorf("failed to send request: %w", err)
+	}
+
+	if resp.StatusCode >= 400 {
+		body, _ := io.ReadAll(resp.Body)
+		resp.Body.Close()
+		return nil, fmt.Errorf("Claude API error (status %d): %s", resp.StatusCode, string(body))
+	}
+
+	ch := make(chan StreamChunk, 100)
+	go c.streamResponse(ctx, resp, ch)
+
+	return ch, nil
+}
+
+func (c *Claude) streamResponse(ctx context.Context, resp *http.Response, ch chan<- StreamChunk) {
+	defer close(ch)
+	defer resp.Body.Close()
+
+	reader := bufio.NewReader(resp.Body)
+	var currentToolCall *ToolCall
+
+	for {
+		select {
+		case <-ctx.Done():
+			ch <- StreamChunk{Done: true, Error: ctx.Err()}
+			return
+		default:
+		}
+
+		line, err := reader.ReadString('\n')
+		if err != nil {
+			if err != io.EOF {
+				ch <- StreamChunk{Done: true, Error: err}
+			} else {
+				ch <- StreamChunk{Done: true}
+			}
+			return
+		}
+
+		line = strings.TrimSpace(line)
+		if line == "" || !strings.HasPrefix(line, "data: ") {
+			continue
+		}
+
+		data := strings.TrimPrefix(line, "data: ")
+
+		var event claudeStreamEvent
+		if err := json.Unmarshal([]byte(data), &event); err != nil {
+			continue
+		}
+
+		switch event.Type {
+		case "content_block_start":
+			if event.ContentBlock != nil && event.ContentBlock.Type == "tool_use" {
+				currentToolCall = &ToolCall{
+					ID:   event.ContentBlock.ID,
+					Name: event.ContentBlock.Name,
+				}
+			}
+
+		case "content_block_delta":
+			if event.Delta != nil {
+				chunk := StreamChunk{}
+
+				switch event.Delta.Type {
+				case "text_delta":
+					chunk.Content = event.Delta.Text
+				case "input_json_delta":
+					if currentToolCall != nil {
+						currentToolCall.Arguments += event.Delta.PartialJSON
+					}
+				}
+
+				if chunk.Content != "" {
+					ch <- chunk
+				}
+			}
+
+		case "content_block_stop":
+			if currentToolCall != nil {
+				ch <- StreamChunk{
+					ToolCalls: []ToolCall{*currentToolCall},
+				}
+				currentToolCall = nil
+			}
+
+		case "message_delta":
+			if event.Delta != nil && event.Delta.StopReason != "" {
+				ch <- StreamChunk{
+					FinishReason: event.Delta.StopReason,
+				}
+			}
+
+		case "message_stop":
+			ch <- StreamChunk{Done: true}
+			return
+
+		case "error":
+			ch <- StreamChunk{Done: true, Error: fmt.Errorf("stream error")}
+			return
+		}
+	}
+}
+
+func (c *Claude) convertRequest(req *GenerateRequest) *claudeRequest {
+	claudeReq := &claudeRequest{
+		Model:       c.config.Model,
+		MaxTokens:   req.MaxTokens,
+		Temperature: req.Temperature,
+		TopP:        req.TopP,
+		Stop:        req.Stop,
+	}
+
+	if claudeReq.MaxTokens <= 0 {
+		claudeReq.MaxTokens = DefaultClaudeMaxTokens
+	}
+
+	// Extract system message and convert other messages
+	for _, msg := range req.Messages {
+		if msg.Role == RoleSystem {
+			claudeReq.System = msg.Content
+			continue
+		}
+
+		// Claude uses content blocks instead of plain strings
+		claudeMsg := claudeMessage{
+			Role: msg.Role,
+			Content: []claudeContentBlock{
+				{Type: "text", Text: msg.Content},
+			},
+		}
+		claudeReq.Messages = append(claudeReq.Messages, claudeMsg)
+	}
+
+	// Convert tools
+	for _, tool := range req.Tools {
+		claudeReq.Tools = append(claudeReq.Tools, claudeTool{
+			Name:        tool.Name,
+			Description: tool.Description,
+			InputSchema: tool.Parameters,
+		})
+	}
+
+	return claudeReq
+}
+
+func (c *Claude) convertResponse(resp *claudeResponse) *GenerateResponse {
+	result := &GenerateResponse{
+		FinishReason: resp.StopReason,
+	}
+
+	// Extract text content and tool calls
+	var textParts []string
+	for _, block := range resp.Content {
+		switch block.Type {
+		case "text":
+			textParts = append(textParts, block.Text)
+		case "tool_use":
+			inputJSON, _ := json.Marshal(block.Input)
+			result.ToolCalls = append(result.ToolCalls, ToolCall{
+				ID:        block.ID,
+				Name:      block.Name,
+				Arguments: string(inputJSON),
+			})
+		}
+	}
+
+	result.Content = strings.Join(textParts, "")
+
+	if resp.Usage != nil {
+		result.Usage = &Usage{
+			PromptTokens:     resp.Usage.InputTokens,
+			CompletionTokens: resp.Usage.OutputTokens,
+			TotalTokens:      resp.Usage.InputTokens + resp.Usage.OutputTokens,
+		}
+	}
+
+	return result
+}
+
+func (c *Claude) doRequest(ctx context.Context, req *claudeRequest) ([]byte, error) {
+	jsonData, err := json.Marshal(req)
+	if err != nil {
+		return nil, fmt.Errorf("failed to marshal request: %w", err)
+	}
+
+	httpReq, err := http.NewRequestWithContext(ctx, "POST", c.config.BaseURL, bytes.NewBuffer(jsonData))
+	if err != nil {
+		return nil, fmt.Errorf("failed to create request: %w", err)
+	}
+
+	c.setHeaders(httpReq)
+
+	resp, err := c.client.Do(httpReq)
+	if err != nil {
+		return nil, fmt.Errorf("failed to send request: %w", err)
+	}
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read response: %w", err)
+	}
+
+	if resp.StatusCode >= 400 {
+		return nil, fmt.Errorf("Claude API error (status %d): %s", resp.StatusCode, string(body))
+	}
+
+	return body, nil
+}
+
+func (c *Claude) setHeaders(req *http.Request) {
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("anthropic-version", ClaudeAPIVersion)
+
+	if c.config.APIKey != "" {
+		req.Header.Set("x-api-key", c.config.APIKey)
+	}
+
+	for k, v := range c.config.Headers {
+		req.Header.Set(k, v)
+	}
+}

aiagent/llm/gemini.go

@@ -0,0 +1,376 @@
+package llm
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+)
+
+const (
+	DefaultGeminiURL = "https://generativelanguage.googleapis.com/v1beta/models"
+)
+
+// Gemini implements the LLM interface for Google Gemini API
+type Gemini struct {
+	config *Config
+	client *http.Client
+}
+
+// NewGemini creates a new Gemini provider
+func NewGemini(cfg *Config, client *http.Client) (*Gemini, error) {
+	if cfg.BaseURL == "" {
+		cfg.BaseURL = DefaultGeminiURL
+	}
+	return &Gemini{
+		config: cfg,
+		client: client,
+	}, nil
+}
+
+func (g *Gemini) Name() string {
+	return ProviderGemini
+}
+
+// Gemini API request/response structures
+type geminiRequest struct {
+	Contents          []geminiContent         `json:"contents"`
+	SystemInstruction *geminiContent          `json:"systemInstruction,omitempty"`
+	Tools             []geminiTool            `json:"tools,omitempty"`
+	GenerationConfig  *geminiGenerationConfig `json:"generationConfig,omitempty"`
+}
+
+type geminiContent struct {
+	Role  string       `json:"role,omitempty"`
+	Parts []geminiPart `json:"parts"`
+}
+
+type geminiPart struct {
+	Text             string                  `json:"text,omitempty"`
+	FunctionCall     *geminiFunctionCall     `json:"functionCall,omitempty"`
+	FunctionResponse *geminiFunctionResponse `json:"functionResponse,omitempty"`
+}
+
+type geminiFunctionCall struct {
+	Name string                 `json:"name"`
+	Args map[string]interface{} `json:"args"`
+}
+
+type geminiFunctionResponse struct {
+	Name     string                 `json:"name"`
+	Response map[string]interface{} `json:"response"`
+}
+
+type geminiTool struct {
+	FunctionDeclarations []geminiFunctionDeclaration `json:"functionDeclarations,omitempty"`
+}
+
+type geminiFunctionDeclaration struct {
+	Name        string                 `json:"name"`
+	Description string                 `json:"description"`
+	Parameters  map[string]interface{} `json:"parameters,omitempty"`
+}
+
+type geminiGenerationConfig struct {
+	Temperature     float64  `json:"temperature,omitempty"`
+	TopP            float64  `json:"topP,omitempty"`
+	MaxOutputTokens int      `json:"maxOutputTokens,omitempty"`
+	StopSequences   []string `json:"stopSequences,omitempty"`
+}
+
+type geminiResponse struct {
+	Candidates []struct {
+		Content       geminiContent `json:"content"`
+		FinishReason  string        `json:"finishReason"`
+		SafetyRatings []struct {
+			Category    string `json:"category"`
+			Probability string `json:"probability"`
+		} `json:"safetyRatings,omitempty"`
+	} `json:"candidates"`
+	UsageMetadata *struct {
+		PromptTokenCount     int `json:"promptTokenCount"`
+		CandidatesTokenCount int `json:"candidatesTokenCount"`
+		TotalTokenCount      int `json:"totalTokenCount"`
+	} `json:"usageMetadata,omitempty"`
+	Error *struct {
+		Code    int    `json:"code"`
+		Message string `json:"message"`
+		Status  string `json:"status"`
+	} `json:"error,omitempty"`
+}
+
+func (g *Gemini) Generate(ctx context.Context, req *GenerateRequest) (*GenerateResponse, error) {
+	geminiReq := g.convertRequest(req)
+
+	url := g.buildURL(false)
+	respBody, err := g.doRequest(ctx, url, geminiReq)
+	if err != nil {
+		return nil, err
+	}
+
+	var geminiResp geminiResponse
+	if err := json.Unmarshal(respBody, &geminiResp); err != nil {
+		return nil, fmt.Errorf("failed to parse response: %w", err)
+	}
+
+	if geminiResp.Error != nil {
+		return nil, fmt.Errorf("Gemini API error: %s", geminiResp.Error.Message)
+	}
+
+	return g.convertResponse(&geminiResp), nil
+}
+
+func (g *Gemini) GenerateStream(ctx context.Context, req *GenerateRequest) (<-chan StreamChunk, error) {
+	geminiReq := g.convertRequest(req)
+
+	jsonData, err := json.Marshal(geminiReq)
+	if err != nil {
+		return nil, fmt.Errorf("failed to marshal request: %w", err)
+	}
+
+	url := g.buildURL(true)
+	httpReq, err := http.NewRequestWithContext(ctx, "POST", url, bytes.NewBuffer(jsonData))
+	if err != nil {
+		return nil, fmt.Errorf("failed to create request: %w", err)
+	}
+
+	g.setHeaders(httpReq)
+
+	resp, err := g.client.Do(httpReq)
+	if err != nil {
+		return nil, fmt.Errorf("failed to send request: %w", err)
+	}
+
+	if resp.StatusCode >= 400 {
+		body, _ := io.ReadAll(resp.Body)
+		resp.Body.Close()
+		return nil, fmt.Errorf("Gemini API error (status %d): %s", resp.StatusCode, string(body))
+	}
+
+	ch := make(chan StreamChunk, 100)
+	go g.streamResponse(ctx, resp, ch)
+
+	return ch, nil
+}
+
+func (g *Gemini) streamResponse(ctx context.Context, resp *http.Response, ch chan<- StreamChunk) {
+	defer close(ch)
+	defer resp.Body.Close()
+
+	reader := bufio.NewReader(resp.Body)
+	var buffer strings.Builder
+
+	for {
+		select {
+		case <-ctx.Done():
+			ch <- StreamChunk{Done: true, Error: ctx.Err()}
+			return
+		default:
+		}
+
+		line, err := reader.ReadString('\n')
+		if err != nil {
+			if err != io.EOF {
+				ch <- StreamChunk{Done: true, Error: err}
+			} else {
+				ch <- StreamChunk{Done: true}
+			}
+			return
+		}
+
+		line = strings.TrimSpace(line)
+
+		// Gemini streams JSON objects, accumulate until we have a complete one
+		if line == "" {
+			continue
+		}
+
+		// Handle SSE format if present
+		if strings.HasPrefix(line, "data: ") {
+			line = strings.TrimPrefix(line, "data: ")
+		}
+
+		buffer.WriteString(line)
+
+		// Try to parse accumulated JSON
+		var geminiResp geminiResponse
+		if err := json.Unmarshal([]byte(buffer.String()), &geminiResp); err != nil {
+			// Not complete yet, continue accumulating
+			continue
+		}
+
+		// Reset buffer for next response
+		buffer.Reset()
+
+		if len(geminiResp.Candidates) > 0 {
+			candidate := geminiResp.Candidates[0]
+			chunk := StreamChunk{
+				FinishReason: candidate.FinishReason,
+			}
+
+			for _, part := range candidate.Content.Parts {
+				if part.Text != "" {
+					chunk.Content += part.Text
+				}
+				if part.FunctionCall != nil {
+					argsJSON, _ := json.Marshal(part.FunctionCall.Args)
+					chunk.ToolCalls = append(chunk.ToolCalls, ToolCall{
+						Name:      part.FunctionCall.Name,
+						Arguments: string(argsJSON),
+					})
+				}
+			}
+
+			ch <- chunk
+
+			if candidate.FinishReason != "" && candidate.FinishReason != "STOP" {
+				ch <- StreamChunk{Done: true}
+				return
+			}
+		}
+	}
+}
+
+func (g *Gemini) convertRequest(req *GenerateRequest) *geminiRequest {
+	geminiReq := &geminiRequest{
+		GenerationConfig: &geminiGenerationConfig{
+			Temperature:     req.Temperature,
+			TopP:            req.TopP,
+			MaxOutputTokens: req.MaxTokens,
+			StopSequences:   req.Stop,
+		},
+	}
+
+	// Convert messages
+	for _, msg := range req.Messages {
+		if msg.Role == RoleSystem {
+			geminiReq.SystemInstruction = &geminiContent{
+				Parts: []geminiPart{{Text: msg.Content}},
+			}
+			continue
+		}
+
+		// Map roles
+		role := msg.Role
+		if role == RoleAssistant {
+			role = "model"
+		}
+
+		geminiReq.Contents = append(geminiReq.Contents, geminiContent{
+			Role:  role,
+			Parts: []geminiPart{{Text: msg.Content}},
+		})
+	}
+
+	// Convert tools
+	if len(req.Tools) > 0 {
+		var declarations []geminiFunctionDeclaration
+		for _, tool := range req.Tools {
+			declarations = append(declarations, geminiFunctionDeclaration{
+				Name:        tool.Name,
+				Description: tool.Description,
+				Parameters:  tool.Parameters,
+			})
+		}
+		geminiReq.Tools = []geminiTool{{FunctionDeclarations: declarations}}
+	}
+
+	return geminiReq
+}
+
+func (g *Gemini) convertResponse(resp *geminiResponse) *GenerateResponse {
+	result := &GenerateResponse{}
+
+	if len(resp.Candidates) > 0 {
+		candidate := resp.Candidates[0]
+		result.FinishReason = candidate.FinishReason
+
+		var textParts []string
+		for _, part := range candidate.Content.Parts {
+			if part.Text != "" {
+				textParts = append(textParts, part.Text)
+			}
+			if part.FunctionCall != nil {
+				argsJSON, _ := json.Marshal(part.FunctionCall.Args)
+				result.ToolCalls = append(result.ToolCalls, ToolCall{
+					Name:      part.FunctionCall.Name,
+					Arguments: string(argsJSON),
+				})
+			}
+		}
+		result.Content = strings.Join(textParts, "")
+	}
+
+	if resp.UsageMetadata != nil {
+		result.Usage = &Usage{
+			PromptTokens:     resp.UsageMetadata.PromptTokenCount,
+			CompletionTokens: resp.UsageMetadata.CandidatesTokenCount,
+			TotalTokens:      resp.UsageMetadata.TotalTokenCount,
+		}
+	}
+
+	return result
+}
+
+func (g *Gemini) buildURL(stream bool) string {
+	action := "generateContent"
+	if stream {
+		action = "streamGenerateContent"
+	}
+
+	// Check if baseURL already contains the full path
+	if strings.Contains(g.config.BaseURL, ":generateContent") ||
+		strings.Contains(g.config.BaseURL, ":streamGenerateContent") {
+		return fmt.Sprintf("%s?key=%s", g.config.BaseURL, g.config.APIKey)
+	}
+
+	return fmt.Sprintf("%s/%s:%s?key=%s",
+		g.config.BaseURL,
+		g.config.Model,
+		action,
+		g.config.APIKey,
+	)
+}
+
+func (g *Gemini) doRequest(ctx context.Context, url string, req *geminiRequest) ([]byte, error) {
+	jsonData, err := json.Marshal(req)
+	if err != nil {
+		return nil, fmt.Errorf("failed to marshal request: %w", err)
+	}
+
+	httpReq, err := http.NewRequestWithContext(ctx, "POST", url, bytes.NewBuffer(jsonData))
+	if err != nil {
+		return nil, fmt.Errorf("failed to create request: %w", err)
+	}
+
+	g.setHeaders(httpReq)
+
+	resp, err := g.client.Do(httpReq)
+	if err != nil {
+		return nil, fmt.Errorf("failed to send request: %w", err)
+	}
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read response: %w", err)
+	}
+
+	if resp.StatusCode >= 400 {
+		return nil, fmt.Errorf("Gemini API error (status %d): %s", resp.StatusCode, string(body))
+	}
+
+	return body, nil
+}
+
+func (g *Gemini) setHeaders(req *http.Request) {
+	req.Header.Set("Content-Type", "application/json")
+
+	for k, v := range g.config.Headers {
+		req.Header.Set(k, v)
+	}
+}

aiagent/llm/helper.go

@@ -0,0 +1,135 @@
+package llm
+
+import (
+	"context"
+	"strings"
+)
+
+// Chat is a convenience function for simple chat completions
+func Chat(ctx context.Context, llm LLM, messages []Message) (string, error) {
+	resp, err := llm.Generate(ctx, &GenerateRequest{
+		Messages: messages,
+	})
+	if err != nil {
+		return "", err
+	}
+	return resp.Content, nil
+}
+
+// ChatWithSystem is a convenience function for chat with a system prompt
+func ChatWithSystem(ctx context.Context, llm LLM, systemPrompt string, userMessage string) (string, error) {
+	messages := []Message{
+		{Role: RoleSystem, Content: systemPrompt},
+		{Role: RoleUser, Content: userMessage},
+	}
+	return Chat(ctx, llm, messages)
+}
+
+// NewMessage creates a new message
+func NewMessage(role, content string) Message {
+	return Message{Role: role, Content: content}
+}
+
+// SystemMessage creates a system message
+func SystemMessage(content string) Message {
+	return Message{Role: RoleSystem, Content: content}
+}
+
+// UserMessage creates a user message
+func UserMessage(content string) Message {
+	return Message{Role: RoleUser, Content: content}
+}
+
+// AssistantMessage creates an assistant message
+func AssistantMessage(content string) Message {
+	return Message{Role: RoleAssistant, Content: content}
+}
+
+// DetectProvider attempts to detect the provider from the base URL
+func DetectProvider(baseURL string) string {
+	baseURL = strings.ToLower(baseURL)
+
+	switch {
+	case strings.Contains(baseURL, "anthropic.com"):
+		return ProviderClaude
+	case strings.Contains(baseURL, "generativelanguage.googleapis.com"):
+		return ProviderGemini
+	case strings.Contains(baseURL, "aiplatform.googleapis.com"):
+		return ProviderVertex
+	case strings.Contains(baseURL, "bedrock"):
+		return ProviderBedrock
+	case strings.Contains(baseURL, "localhost:11434"):
+		return ProviderOllama
+	default:
+		// Default to OpenAI-compatible
+		return ProviderOpenAI
+	}
+}
+
+// DetectProviderFromModel attempts to detect the provider from the model name
+func DetectProviderFromModel(model string) string {
+	model = strings.ToLower(model)
+
+	switch {
+	case strings.HasPrefix(model, "claude"):
+		return ProviderClaude
+	case strings.HasPrefix(model, "gemini"):
+		return ProviderGemini
+	case strings.HasPrefix(model, "gpt") || strings.HasPrefix(model, "o1") || strings.HasPrefix(model, "o3"):
+		return ProviderOpenAI
+	case strings.HasPrefix(model, "llama") || strings.HasPrefix(model, "mistral") || strings.HasPrefix(model, "qwen"):
+		return ProviderOllama
+	default:
+		return ProviderOpenAI
+	}
+}
+
+// BuildToolDefinition creates a tool definition with JSON schema parameters
+func BuildToolDefinition(name, description string, properties map[string]interface{}, required []string) ToolDefinition {
+	params := map[string]interface{}{
+		"type":       "object",
+		"properties": properties,
+	}
+	if len(required) > 0 {
+		params["required"] = required
+	}
+
+	return ToolDefinition{
+		Name:        name,
+		Description: description,
+		Parameters:  params,
+	}
+}
+
+// CollectStream collects all chunks from a stream into a single response
+func CollectStream(ch <-chan StreamChunk) (*GenerateResponse, error) {
+	var content strings.Builder
+	var toolCalls []ToolCall
+	var finishReason string
+	var lastErr error
+
+	for chunk := range ch {
+		if chunk.Error != nil {
+			lastErr = chunk.Error
+		}
+		if chunk.Content != "" {
+			content.WriteString(chunk.Content)
+		}
+		if len(chunk.ToolCalls) > 0 {
+			toolCalls = append(toolCalls, chunk.ToolCalls...)
+		}
+		if chunk.FinishReason != "" {
+			finishReason = chunk.FinishReason
+		}
+	}
+
+	if lastErr != nil {
+		return nil, lastErr
+	}
+
+	return &GenerateResponse{
+		Content:      content.String(),
+		ToolCalls:    toolCalls,
+		FinishReason: finishReason,
+	}, nil
+}

aiagent/llm/llm.go

@@ -0,0 +1,193 @@
+// Package llm provides a unified interface for multiple LLM providers.
+// Supports OpenAI-compatible APIs, Claude/Anthropic, and Gemini.
+package llm
+
+import (
+	"context"
+	"crypto/tls"
+	"fmt"
+	"net/http"
+	"net/url"
+	"time"
+)
+
+// Provider types
+const (
+	ProviderOpenAI    = "openai"    // OpenAI and compatible APIs (Azure, vLLM, etc.)
+	ProviderClaude    = "claude"    // Anthropic Claude
+	ProviderGemini    = "gemini"    // Google Gemini
+	ProviderOllama    = "ollama"    // Ollama local models
+	ProviderBedrock   = "bedrock"   // AWS Bedrock
+	ProviderVertex    = "vertex"    // Google Vertex AI
+)
+
+// Role constants
+const (
+	RoleSystem    = "system"
+	RoleUser      = "user"
+	RoleAssistant = "assistant"
+)
+
+// Message represents a chat message
+type Message struct {
+	Role    string `json:"role"`
+	Content string `json:"content"`
+}
+
+// ToolCall represents a tool/function call from the LLM
+type ToolCall struct {
+	ID        string `json:"id"`
+	Name      string `json:"name"`
+	Arguments string `json:"arguments"`
+}
+
+// ToolDefinition defines a tool that the LLM can call
+type ToolDefinition struct {
+	Name        string                 `json:"name"`
+	Description string                 `json:"description"`
+	Parameters  map[string]interface{} `json:"parameters,omitempty"`
+}
+
+// GenerateRequest is the unified request for LLM generation
+type GenerateRequest struct {
+	Messages    []Message         `json:"messages"`
+	Tools       []ToolDefinition  `json:"tools,omitempty"`
+	MaxTokens   int               `json:"max_tokens,omitempty"`
+	Temperature float64           `json:"temperature,omitempty"`
+	TopP        float64           `json:"top_p,omitempty"`
+	Stop        []string          `json:"stop,omitempty"`
+	Stream      bool              `json:"stream,omitempty"`
+}
+
+// GenerateResponse is the unified response from LLM generation
+type GenerateResponse struct {
+	Content      string     `json:"content"`
+	ToolCalls    []ToolCall `json:"tool_calls,omitempty"`
+	FinishReason string     `json:"finish_reason"`
+	Usage        *Usage     `json:"usage,omitempty"`
+}
+
+// Usage represents token usage statistics
+type Usage struct {
+	PromptTokens     int `json:"prompt_tokens"`
+	CompletionTokens int `json:"completion_tokens"`
+	TotalTokens      int `json:"total_tokens"`
+}
+
+// StreamChunk represents a chunk in streaming response
+type StreamChunk struct {
+	Content      string     `json:"content,omitempty"`
+	ToolCalls    []ToolCall `json:"tool_calls,omitempty"`
+	FinishReason string     `json:"finish_reason,omitempty"`
+	Done         bool       `json:"done"`
+	Error        error      `json:"error,omitempty"`
+}
+
+// LLM is the unified interface for all LLM providers
+type LLM interface {
+	// Name returns the provider name
+	Name() string
+
+	// Generate sends a request to the LLM and returns the response
+	Generate(ctx context.Context, req *GenerateRequest) (*GenerateResponse, error)
+
+	// GenerateStream sends a request and returns a channel for streaming responses
+	GenerateStream(ctx context.Context, req *GenerateRequest) (<-chan StreamChunk, error)
+}
+
+// Config is the configuration for creating an LLM provider
+type Config struct {
+	// Provider type: openai, claude, gemini, ollama, bedrock, vertex
+	Provider string `json:"provider"`
+
+	// API endpoint URL
+	BaseURL string `json:"base_url,omitempty"`
+
+	// API key or token
+	APIKey string `json:"api_key,omitempty"`
+
+	// Model name (e.g., "gpt-4", "claude-3-opus", "gemini-pro")
+	Model string `json:"model"`
+
+	// Additional headers for API requests
+	Headers map[string]string `json:"headers,omitempty"`
+
+	// HTTP timeout in milliseconds
+	Timeout int `json:"timeout,omitempty"`
+
+	// Skip SSL verification (for self-signed certs)
+	SkipSSLVerify bool `json:"skip_ssl_verify,omitempty"`
+
+	// HTTP proxy URL
+	Proxy string `json:"proxy,omitempty"`
+
+	// Provider-specific options
+	Options map[string]interface{} `json:"options,omitempty"`
+}
+
+// DefaultConfig returns a config with default values
+func DefaultConfig() *Config {
+	return &Config{
+		Provider: ProviderOpenAI,
+		Timeout:  60000,
+	}
+}
+
+// New creates an LLM instance based on the config
+func New(cfg *Config) (LLM, error) {
+	if cfg == nil {
+		cfg = DefaultConfig()
+	}
+
+	// Create HTTP client
+	client := createHTTPClient(cfg)
+
+	switch cfg.Provider {
+	case ProviderOpenAI, "":
+		return NewOpenAI(cfg, client)
+	case ProviderClaude:
+		return NewClaude(cfg, client)
+	case ProviderGemini:
+		return NewGemini(cfg, client)
+	case ProviderOllama:
+		// Ollama uses OpenAI-compatible API
+		if cfg.BaseURL == "" {
+			cfg.BaseURL = "http://localhost:11434/v1"
+		}
+		return NewOpenAI(cfg, client)
+	default:
+		return nil, fmt.Errorf("unsupported LLM provider: %s", cfg.Provider)
+	}
+}
+
+// createHTTPClient creates an HTTP client with the given config
+func createHTTPClient(cfg *Config) *http.Client {
+	transport := &http.Transport{
+		TLSClientConfig: &tls.Config{
+			InsecureSkipVerify: cfg.SkipSSLVerify,
+		},
+	}
+
+	if cfg.Proxy != "" {
+		if proxyURL, err := url.Parse(cfg.Proxy); err == nil {
+			transport.Proxy = http.ProxyURL(proxyURL)
+		}
+	}
+
+	timeout := cfg.Timeout
+	if timeout <= 0 {
+		timeout = 60000
+	}
+
+	return &http.Client{
+		Timeout:   time.Duration(timeout) * time.Millisecond,
+		Transport: transport,
+	}
+}
+
+// Helper function to convert internal messages to provider-specific format
+func ConvertMessages(messages []Message) []Message {
+	result := make([]Message, len(messages))
+	copy(result, messages)
+	return result
+}

aiagent/llm/openai.go

@@ -0,0 +1,416 @@
+package llm
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"time"
+)
+
+const (
+	DefaultOpenAIURL = "https://api.openai.com/v1/chat/completions"
+
+	// 重试相关配置
+	maxRetries       = 3
+	initialRetryWait = 5 * time.Second  // rate limit 时初始等待 5 秒
+	maxRetryWait     = 60 * time.Second // 最大等待 60 秒
+)
+
+// OpenAI implements the LLM interface for OpenAI and compatible APIs
+type OpenAI struct {
+	config *Config
+	client *http.Client
+}
+
+// NewOpenAI creates a new OpenAI provider
+func NewOpenAI(cfg *Config, client *http.Client) (*OpenAI, error) {
+	if cfg.BaseURL == "" {
+		cfg.BaseURL = DefaultOpenAIURL
+	}
+	return &OpenAI{
+		config: cfg,
+		client: client,
+	}, nil
+}
+
+func (o *OpenAI) Name() string {
+	return ProviderOpenAI
+}
+
+// OpenAI API request/response structures
+type openAIRequest struct {
+	Model       string              `json:"model"`
+	Messages    []openAIMessage     `json:"messages"`
+	Tools       []openAITool        `json:"tools,omitempty"`
+	MaxTokens   int                 `json:"max_tokens,omitempty"`
+	Temperature float64             `json:"temperature,omitempty"`
+	TopP        float64             `json:"top_p,omitempty"`
+	Stop        []string            `json:"stop,omitempty"`
+	Stream      bool                `json:"stream,omitempty"`
+}
+
+type openAIMessage struct {
+	Role       string           `json:"role"`
+	Content    string           `json:"content,omitempty"`
+	ToolCalls  []openAIToolCall `json:"tool_calls,omitempty"`
+	ToolCallID string           `json:"tool_call_id,omitempty"`
+}
+
+type openAITool struct {
+	Type     string           `json:"type"`
+	Function openAIFunction   `json:"function"`
+}
+
+type openAIFunction struct {
+	Name        string                 `json:"name"`
+	Description string                 `json:"description"`
+	Parameters  map[string]interface{} `json:"parameters,omitempty"`
+}
+
+type openAIToolCall struct {
+	ID       string `json:"id"`
+	Type     string `json:"type"`
+	Function struct {
+		Name      string `json:"name"`
+		Arguments string `json:"arguments"`
+	} `json:"function"`
+}
+
+type openAIResponse struct {
+	ID      string `json:"id"`
+	Object  string `json:"object"`
+	Created int64  `json:"created"`
+	Model   string `json:"model"`
+	Choices []struct {
+		Index        int           `json:"index"`
+		Message      openAIMessage `json:"message"`
+		Delta        openAIMessage `json:"delta"`
+		FinishReason string        `json:"finish_reason"`
+	} `json:"choices"`
+	Usage *struct {
+		PromptTokens     int `json:"prompt_tokens"`
+		CompletionTokens int `json:"completion_tokens"`
+		TotalTokens      int `json:"total_tokens"`
+	} `json:"usage,omitempty"`
+	Error *struct {
+		Message string `json:"message"`
+		Type    string `json:"type"`
+		Code    string `json:"code"`
+	} `json:"error,omitempty"`
+}
+
+func (o *OpenAI) Generate(ctx context.Context, req *GenerateRequest) (*GenerateResponse, error) {
+	// Convert to OpenAI format
+	openAIReq := o.convertRequest(req)
+	openAIReq.Stream = false
+
+	// Make request
+	respBody, err := o.doRequest(ctx, openAIReq)
+	if err != nil {
+		return nil, err
+	}
+
+	// Parse response
+	var openAIResp openAIResponse
+	if err := json.Unmarshal(respBody, &openAIResp); err != nil {
+		return nil, fmt.Errorf("failed to parse response: %w", err)
+	}
+
+	if openAIResp.Error != nil {
+		return nil, fmt.Errorf("OpenAI API error: %s", openAIResp.Error.Message)
+	}
+
+	if len(openAIResp.Choices) == 0 {
+		return nil, fmt.Errorf("no response from OpenAI")
+	}
+
+	// Convert to unified response
+	return o.convertResponse(&openAIResp), nil
+}
+
+// isRetryableStatus 检查是否是可重试的 HTTP 状态码
+func isRetryableStatus(statusCode int) bool {
+	switch statusCode {
+	case 429, 500, 502, 503, 504:
+		return true
+	default:
+		return false
+	}
+}
+
+func (o *OpenAI) GenerateStream(ctx context.Context, req *GenerateRequest) (<-chan StreamChunk, error) {
+	// Convert to OpenAI format
+	openAIReq := o.convertRequest(req)
+	openAIReq.Stream = true
+
+	// Create request body
+	jsonData, err := json.Marshal(openAIReq)
+	if err != nil {
+		return nil, fmt.Errorf("failed to marshal request: %w", err)
+	}
+
+	var resp *http.Response
+	var lastErr error
+	retryWait := initialRetryWait
+
+	// 重试循环
+	for attempt := 0; attempt <= maxRetries; attempt++ {
+		if attempt > 0 {
+			// 等待后重试
+			select {
+			case <-ctx.Done():
+				return nil, ctx.Err()
+			case <-time.After(retryWait):
+			}
+			// 指数退避，但不超过最大等待时间
+			retryWait *= 2
+			if retryWait > maxRetryWait {
+				retryWait = maxRetryWait
+			}
+		}
+
+		httpReq, err := http.NewRequestWithContext(ctx, "POST", o.config.BaseURL, bytes.NewBuffer(jsonData))
+		if err != nil {
+			return nil, fmt.Errorf("failed to create request: %w", err)
+		}
+
+		o.setHeaders(httpReq)
+
+		// Make request
+		resp, err = o.client.Do(httpReq)
+		if err != nil {
+			lastErr = fmt.Errorf("failed to send request: %w", err)
+			continue // 网络错误，重试
+		}
+
+		if resp.StatusCode >= 400 {
+			body, _ := io.ReadAll(resp.Body)
+			resp.Body.Close()
+			lastErr = fmt.Errorf("OpenAI API error (status %d): %s", resp.StatusCode, string(body))
+
+			// 检查是否可重试
+			if isRetryableStatus(resp.StatusCode) && attempt < maxRetries {
+				continue // 可重试的错误，继续重试
+			}
+			return nil, lastErr
+		}
+
+		// 成功，跳出循环
+		break
+	}
+
+	if resp == nil {
+		return nil, lastErr
+	}
+
+	// Create channel and start streaming
+	ch := make(chan StreamChunk, 100)
+	go o.streamResponse(ctx, resp, ch)
+
+	return ch, nil
+}
+
+func (o *OpenAI) streamResponse(ctx context.Context, resp *http.Response, ch chan<- StreamChunk) {
+	defer close(ch)
+	defer resp.Body.Close()
+
+	reader := bufio.NewReader(resp.Body)
+
+	for {
+		select {
+		case <-ctx.Done():
+			ch <- StreamChunk{Done: true, Error: ctx.Err()}
+			return
+		default:
+		}
+
+		line, err := reader.ReadString('\n')
+		if err != nil {
+			if err != io.EOF {
+				ch <- StreamChunk{Done: true, Error: err}
+			} else {
+				ch <- StreamChunk{Done: true}
+			}
+			return
+		}
+
+		line = strings.TrimSpace(line)
+		if line == "" {
+			continue
+		}
+
+		if !strings.HasPrefix(line, "data: ") {
+			continue
+		}
+
+		data := strings.TrimPrefix(line, "data: ")
+		if data == "[DONE]" {
+			ch <- StreamChunk{Done: true}
+			return
+		}
+
+		var streamResp openAIResponse
+		if err := json.Unmarshal([]byte(data), &streamResp); err != nil {
+			continue
+		}
+
+		if len(streamResp.Choices) > 0 {
+			delta := streamResp.Choices[0].Delta
+			chunk := StreamChunk{
+				Content:      delta.Content,
+				FinishReason: streamResp.Choices[0].FinishReason,
+			}
+
+			// Handle tool calls in stream
+			if len(delta.ToolCalls) > 0 {
+				for _, tc := range delta.ToolCalls {
+					chunk.ToolCalls = append(chunk.ToolCalls, ToolCall{
+						ID:        tc.ID,
+						Name:      tc.Function.Name,
+						Arguments: tc.Function.Arguments,
+					})
+				}
+			}
+
+			ch <- chunk
+		}
+	}
+}
+
+func (o *OpenAI) convertRequest(req *GenerateRequest) *openAIRequest {
+	openAIReq := &openAIRequest{
+		Model:       o.config.Model,
+		MaxTokens:   req.MaxTokens,
+		Temperature: req.Temperature,
+		TopP:        req.TopP,
+		Stop:        req.Stop,
+	}
+
+	// Convert messages
+	for _, msg := range req.Messages {
+		openAIReq.Messages = append(openAIReq.Messages, openAIMessage{
+			Role:    msg.Role,
+			Content: msg.Content,
+		})
+	}
+
+	// Convert tools
+	for _, tool := range req.Tools {
+		openAIReq.Tools = append(openAIReq.Tools, openAITool{
+			Type: "function",
+			Function: openAIFunction{
+				Name:        tool.Name,
+				Description: tool.Description,
+				Parameters:  tool.Parameters,
+			},
+		})
+	}
+
+	return openAIReq
+}
+
+func (o *OpenAI) convertResponse(resp *openAIResponse) *GenerateResponse {
+	result := &GenerateResponse{}
+
+	if len(resp.Choices) > 0 {
+		choice := resp.Choices[0]
+		result.Content = choice.Message.Content
+		result.FinishReason = choice.FinishReason
+
+		// Convert tool calls
+		for _, tc := range choice.Message.ToolCalls {
+			result.ToolCalls = append(result.ToolCalls, ToolCall{
+				ID:        tc.ID,
+				Name:      tc.Function.Name,
+				Arguments: tc.Function.Arguments,
+			})
+		}
+	}
+
+	if resp.Usage != nil {
+		result.Usage = &Usage{
+			PromptTokens:     resp.Usage.PromptTokens,
+			CompletionTokens: resp.Usage.CompletionTokens,
+			TotalTokens:      resp.Usage.TotalTokens,
+		}
+	}
+
+	return result
+}
+
+func (o *OpenAI) doRequest(ctx context.Context, req *openAIRequest) ([]byte, error) {
+	jsonData, err := json.Marshal(req)
+	if err != nil {
+		return nil, fmt.Errorf("failed to marshal request: %w", err)
+	}
+
+	var lastErr error
+	retryWait := initialRetryWait
+
+	// 重试循环
+	for attempt := 0; attempt <= maxRetries; attempt++ {
+		if attempt > 0 {
+			// 等待后重试
+			select {
+			case <-ctx.Done():
+				return nil, ctx.Err()
+			case <-time.After(retryWait):
+			}
+			// 指数退避
+			retryWait *= 2
+			if retryWait > maxRetryWait {
+				retryWait = maxRetryWait
+			}
+		}
+
+		httpReq, err := http.NewRequestWithContext(ctx, "POST", o.config.BaseURL, bytes.NewBuffer(jsonData))
+		if err != nil {
+			return nil, fmt.Errorf("failed to create request: %w", err)
+		}
+
+		o.setHeaders(httpReq)
+
+		resp, err := o.client.Do(httpReq)
+		if err != nil {
+			lastErr = fmt.Errorf("failed to send request: %w", err)
+			continue // 网络错误，重试
+		}
+
+		body, err := io.ReadAll(resp.Body)
+		resp.Body.Close()
+		if err != nil {
+			lastErr = fmt.Errorf("failed to read response: %w", err)
+			continue
+		}
+
+		if resp.StatusCode >= 400 {
+			lastErr = fmt.Errorf("OpenAI API error (status %d): %s", resp.StatusCode, string(body))
+			// 检查是否可重试
+			if isRetryableStatus(resp.StatusCode) && attempt < maxRetries {
+				continue
+			}
+			return nil, lastErr
+		}
+
+		return body, nil
+	}
+
+	return nil, lastErr
+}
+
+func (o *OpenAI) setHeaders(req *http.Request) {
+	req.Header.Set("Content-Type", "application/json")
+
+	if o.config.APIKey != "" {
+		req.Header.Set("Authorization", "Bearer "+o.config.APIKey)
+	}
+
+	for k, v := range o.config.Headers {
+		req.Header.Set(k, v)
+	}
+}

aiagent/llm/prompt.go

@@ -0,0 +1,133 @@
+package llm
+
+import (
+	"fmt"
+	"runtime"
+	"strings"
+	"time"
+)
+
+// ToolInfo 工具信息（用于提示词构建）
+type ToolInfo struct {
+	Name        string
+	Description string
+	Parameters  []ToolParamInfo
+}
+
+// ToolParamInfo 工具参数信息
+type ToolParamInfo struct {
+	Name        string
+	Type        string
+	Description string
+	Required    bool
+}
+
+// PromptData 提示词模板数据
+type PromptData struct {
+	Platform string // 操作系统
+	Date     string // 当前日期
+}
+
+// BuildToolsSection 构建工具描述段落
+func BuildToolsSection(tools []ToolInfo) string {
+	if len(tools) == 0 {
+		return ""
+	}
+
+	var sb strings.Builder
+	sb.WriteString("## Available Tools\n\n")
+
+	for _, tool := range tools {
+		sb.WriteString(fmt.Sprintf("### %s\n", tool.Name))
+		sb.WriteString(fmt.Sprintf("%s\n", tool.Description))
+
+		if len(tool.Parameters) > 0 {
+			sb.WriteString("Parameters:\n")
+			for _, param := range tool.Parameters {
+				required := ""
+				if param.Required {
+					required = " (required)"
+				}
+				sb.WriteString(fmt.Sprintf("- %s (%s)%s: %s\n", param.Name, param.Type, required, param.Description))
+			}
+		}
+		sb.WriteString("\n")
+	}
+
+	return sb.String()
+}
+
+// BuildToolsListBrief 构建简洁的工具列表（用于 Plan 模式）
+func BuildToolsListBrief(tools []ToolInfo) string {
+	if len(tools) == 0 {
+		return ""
+	}
+
+	var sb strings.Builder
+	sb.WriteString("## Available Tools\n\n")
+
+	for _, tool := range tools {
+		sb.WriteString(fmt.Sprintf("- **%s**: %s\n", tool.Name, tool.Description))
+	}
+
+	return sb.String()
+}
+
+// BuildEnvSection 构建环境信息段落
+func BuildEnvSection() string {
+	var sb strings.Builder
+	sb.WriteString("## Environment\n\n")
+	sb.WriteString(fmt.Sprintf("- Platform: %s\n", runtime.GOOS))
+	sb.WriteString(fmt.Sprintf("- Date: %s\n", time.Now().Format("2006-01-02")))
+	return sb.String()
+}
+
+// BuildSkillsSection 构建技能指导段落
+func BuildSkillsSection(skillContents []string) string {
+	if len(skillContents) == 0 {
+		return ""
+	}
+
+	var sb strings.Builder
+	sb.WriteString("## 专项技能指导\n\n")
+
+	if len(skillContents) == 1 {
+		sb.WriteString("你已被加载以下专项技能，请参考技能中的流程：\n\n")
+		sb.WriteString(skillContents[0])
+		sb.WriteString("\n\n")
+	} else {
+		sb.WriteString("你已被加载以下专项技能，请参考技能中的流程来制定执行计划：\n\n")
+		for i, content := range skillContents {
+			sb.WriteString(fmt.Sprintf("### 技能 %d\n\n", i+1))
+			sb.WriteString(content)
+			sb.WriteString("\n\n")
+		}
+	}
+
+	return sb.String()
+}
+
+// BuildPreviousFindingsSection 构建之前发现段落
+func BuildPreviousFindingsSection(findings []string) string {
+	if len(findings) == 0 {
+		return ""
+	}
+
+	var sb strings.Builder
+	sb.WriteString("## Previous Findings\n\n")
+	for _, finding := range findings {
+		sb.WriteString(fmt.Sprintf("- %s\n", finding))
+	}
+	sb.WriteString("\n")
+
+	return sb.String()
+}
+
+// BuildCurrentStepSection 构建当前步骤段落
+func BuildCurrentStepSection(goal, approach string) string {
+	var sb strings.Builder
+	sb.WriteString("## Current Step\n\n")
+	sb.WriteString(fmt.Sprintf("**Goal**: %s\n", goal))
+	sb.WriteString(fmt.Sprintf("**Approach**: %s\n\n", approach))
+	return sb.String()
+}

aiagent/mcp.go

@@ -0,0 +1,571 @@
+package aiagent
+
+import (
+	"bytes"
+	"context"
+	"crypto/tls"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/url"
+	"os"
+	"os/exec"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/modelcontextprotocol/go-sdk/mcp"
+	"github.com/toolkits/pkg/logger"
+)
+
+const (
+	// MCP 传输类型
+	MCPTransportStdio = "stdio" // 标准输入/输出传输
+	MCPTransportSSE   = "sse"   // HTTP Server-Sent Events 传输
+
+	// 默认超时
+	DefaultMCPTimeout        = 30000 // 30 秒
+	DefaultMCPConnectTimeout = 10000 // 10 秒
+)
+
+// MCPConfig MCP 服务器配置（在 AIAgentConfig 中使用）
+type MCPConfig struct {
+	// MCP 服务器列表
+	Servers []MCPServerConfig `json:"servers"`
+}
+
+// MCPServerConfig 单个 MCP 服务器配置
+type MCPServerConfig struct {
+	// 服务器名称（唯一标识）
+	Name string `json:"name"`
+
+	// 传输类型：stdio 或 sse
+	Transport string `json:"transport"`
+
+	// === stdio 传输配置 ===
+	Command string            `json:"command,omitempty"` // 启动命令
+	Args    []string          `json:"args,omitempty"`    // 命令参数
+	Env     map[string]string `json:"env,omitempty"`     // 环境变量（支持 ${VAR} 从系统环境变量读取）
+
+	// === SSE 传输配置 ===
+	URL           string            `json:"url,omitempty"`             // SSE 服务器 URL
+	Headers       map[string]string `json:"headers,omitempty"`         // 请求头（支持 ${VAR} 从系统环境变量读取）
+	SkipSSLVerify bool              `json:"skip_ssl_verify,omitempty"` // 跳过 SSL 验证
+
+	// === 鉴权配置（SSE 传输）===
+	// 便捷鉴权配置，会自动设置对应的 Header
+	AuthType string `json:"auth_type,omitempty"` // 鉴权类型：bearer, api_key, basic
+	APIKey   string `json:"api_key,omitempty"`   // API Key（支持 ${VAR} 从系统环境变量读取）
+	Username string `json:"username,omitempty"`  // Basic Auth 用户名
+	Password string `json:"password,omitempty"`  // Basic Auth 密码（支持 ${VAR}）
+
+	// 通用配置
+	Timeout        int `json:"timeout,omitempty"`         // 工具调用超时（毫秒）
+	ConnectTimeout int `json:"connect_timeout,omitempty"` // 连接超时（毫秒）
+}
+
+// MCPToolConfig MCP 工具配置（在 AgentTool 中使用）
+type MCPToolConfig struct {
+	// MCP 服务器名称（引用 MCPConfig.Servers 中的配置）
+	ServerName string `json:"server_name"`
+
+	// 工具名称（MCP 服务器返回的工具名）
+	ToolName string `json:"tool_name"`
+}
+
+// MCPTool MCP 工具定义（用于内部表示）
+type MCPTool struct {
+	Name        string                 `json:"name"`
+	Description string                 `json:"description,omitempty"`
+	InputSchema map[string]interface{} `json:"inputSchema,omitempty"`
+}
+
+// MCPToolsCallResult 工具调用结果
+type MCPToolsCallResult struct {
+	Content []MCPContent `json:"content"`
+	IsError bool         `json:"isError,omitempty"`
+}
+
+// MCPContent 工具返回内容
+type MCPContent struct {
+	Type     string `json:"type"`
+	Text     string `json:"text,omitempty"`
+	Data     string `json:"data,omitempty"`
+	MimeType string `json:"mimeType,omitempty"`
+}
+
+// MCPClient MCP 客户端（基于官方 go-sdk）
+type MCPClient struct {
+	config *MCPServerConfig
+
+	// SDK 客户端和会话（stdio 传输）
+	client  *mcp.Client
+	session *mcp.ClientSession
+
+	// SSE 传输（SDK 暂不支持 SSE 客户端，保留自定义实现）
+	httpClient *http.Client
+	sseURL     string
+
+	// 通用
+	mu          sync.Mutex
+	initialized bool
+	tools       []MCPTool // 缓存的工具列表
+}
+
+// expandEnvVars 展开字符串中的环境变量引用
+func expandEnvVars(s string) string {
+	return os.ExpandEnv(s)
+}
+
+// NewMCPClient 创建 MCP 客户端
+func NewMCPClient(config *MCPServerConfig) (*MCPClient, error) {
+	client := &MCPClient{
+		config: config,
+	}
+
+	return client, nil
+}
+
+// Connect 连接到 MCP 服务器
+func (c *MCPClient) Connect(ctx context.Context) error {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	if c.initialized {
+		return nil
+	}
+
+	var err error
+	switch c.config.Transport {
+	case MCPTransportStdio:
+		err = c.connectStdio(ctx)
+	case MCPTransportSSE:
+		err = c.connectSSE(ctx)
+	default:
+		return fmt.Errorf("unsupported MCP transport: %s", c.config.Transport)
+	}
+
+	if err != nil {
+		return err
+	}
+
+	c.initialized = true
+	return nil
+}
+
+// connectStdio 通过 stdio 连接（使用官方 SDK）
+func (c *MCPClient) connectStdio(ctx context.Context) error {
+	if c.config.Command == "" {
+		return fmt.Errorf("stdio transport requires command")
+	}
+
+	// 准备环境变量
+	env := os.Environ()
+	for k, v := range c.config.Env {
+		expandedValue := expandEnvVars(v)
+		env = append(env, fmt.Sprintf("%s=%s", k, expandedValue))
+	}
+
+	// 创建 exec.Cmd
+	cmd := exec.CommandContext(ctx, c.config.Command, c.config.Args...)
+	cmd.Env = env
+
+	// 使用官方 SDK 的 CommandTransport
+	transport := &mcp.CommandTransport{
+		Command: cmd,
+	}
+
+	// 创建 MCP 客户端
+	c.client = mcp.NewClient(
+		&mcp.Implementation{
+			Name:    "nightingale-aiagent",
+			Version: "1.0.0",
+		},
+		nil,
+	)
+
+	// 连接并初始化（Connect 会自动进行 initialize 握手）
+	session, err := c.client.Connect(ctx, transport, nil)
+	if err != nil {
+		return fmt.Errorf("failed to connect MCP client: %v", err)
+	}
+	c.session = session
+
+	logger.Infof("MCP stdio server started: %s", c.config.Name)
+	return nil
+}
+
+// connectSSE 通过 SSE 连接（保留自定义实现，SDK 暂不支持 SSE 客户端）
+func (c *MCPClient) connectSSE(ctx context.Context) error {
+	if c.config.URL == "" {
+		return fmt.Errorf("SSE transport requires URL")
+	}
+
+	// 创建 HTTP 客户端
+	transport := &http.Transport{
+		TLSClientConfig: &tls.Config{InsecureSkipVerify: c.config.SkipSSLVerify},
+	}
+
+	timeout := c.config.ConnectTimeout
+	if timeout <= 0 {
+		timeout = DefaultMCPConnectTimeout
+	}
+
+	c.httpClient = &http.Client{
+		Timeout:   time.Duration(timeout) * time.Millisecond,
+		Transport: transport,
+	}
+
+	c.sseURL = c.config.URL
+
+	logger.Infof("MCP SSE client configured: %s", c.config.Name)
+	return nil
+}
+
+// ListTools 获取工具列表
+func (c *MCPClient) ListTools(ctx context.Context) ([]MCPTool, error) {
+	c.mu.Lock()
+	if len(c.tools) > 0 {
+		tools := c.tools
+		c.mu.Unlock()
+		return tools, nil
+	}
+	c.mu.Unlock()
+
+	var tools []MCPTool
+
+	switch c.config.Transport {
+	case MCPTransportStdio:
+		// 使用官方 SDK
+		if c.session == nil {
+			return nil, fmt.Errorf("MCP session not initialized")
+		}
+
+		result, err := c.session.ListTools(ctx, nil)
+		if err != nil {
+			return nil, fmt.Errorf("failed to list tools: %v", err)
+		}
+
+		// 转换为内部格式
+		for _, tool := range result.Tools {
+			inputSchema := make(map[string]interface{})
+			if tool.InputSchema != nil {
+				// 将 SDK 的 InputSchema 转换为 map
+				schemaBytes, err := json.Marshal(tool.InputSchema)
+				if err == nil {
+					json.Unmarshal(schemaBytes, &inputSchema)
+				}
+			}
+
+			tools = append(tools, MCPTool{
+				Name:        tool.Name,
+				Description: tool.Description,
+				InputSchema: inputSchema,
+			})
+		}
+
+	case MCPTransportSSE:
+		// 使用自定义 HTTP 实现
+		var err error
+		tools, err = c.listToolsSSE(ctx)
+		if err != nil {
+			return nil, err
+		}
+	}
+
+	c.mu.Lock()
+	c.tools = tools
+	c.mu.Unlock()
+
+	return tools, nil
+}
+
+// listToolsSSE 通过 SSE 获取工具列表
+func (c *MCPClient) listToolsSSE(ctx context.Context) ([]MCPTool, error) {
+	req := map[string]interface{}{
+		"jsonrpc": "2.0",
+		"id":      1,
+		"method":  "tools/list",
+	}
+
+	resp, err := c.sendSSERequest(ctx, req)
+	if err != nil {
+		return nil, err
+	}
+
+	resultBytes, _ := json.Marshal(resp["result"])
+	var result struct {
+		Tools []MCPTool `json:"tools"`
+	}
+	if err := json.Unmarshal(resultBytes, &result); err != nil {
+		return nil, fmt.Errorf("failed to parse tools list: %v", err)
+	}
+
+	return result.Tools, nil
+}
+
+// CallTool 调用工具
+func (c *MCPClient) CallTool(ctx context.Context, name string, arguments map[string]interface{}) (*MCPToolsCallResult, error) {
+	switch c.config.Transport {
+	case MCPTransportStdio:
+		return c.callToolStdio(ctx, name, arguments)
+	case MCPTransportSSE:
+		return c.callToolSSE(ctx, name, arguments)
+	default:
+		return nil, fmt.Errorf("unsupported transport: %s", c.config.Transport)
+	}
+}
+
+// callToolStdio 通过 stdio 调用工具（使用官方 SDK）
+func (c *MCPClient) callToolStdio(ctx context.Context, name string, arguments map[string]interface{}) (*MCPToolsCallResult, error) {
+	if c.session == nil {
+		return nil, fmt.Errorf("MCP session not initialized")
+	}
+
+	// 调用工具
+	result, err := c.session.CallTool(ctx, &mcp.CallToolParams{
+		Name:      name,
+		Arguments: arguments,
+	})
+	if err != nil {
+		return nil, fmt.Errorf("tool call failed: %v", err)
+	}
+
+	// 转换结果
+	mcpResult := &MCPToolsCallResult{
+		IsError: result.IsError,
+	}
+
+	for _, content := range result.Content {
+		mc := MCPContent{}
+
+		// 根据具体类型提取内容
+		switch c := content.(type) {
+		case *mcp.TextContent:
+			mc.Type = "text"
+			mc.Text = c.Text
+		case *mcp.ImageContent:
+			mc.Type = "image"
+			mc.Data = string(c.Data)
+			mc.MimeType = c.MIMEType
+		case *mcp.AudioContent:
+			mc.Type = "audio"
+			mc.Data = string(c.Data)
+			mc.MimeType = c.MIMEType
+		case *mcp.EmbeddedResource:
+			mc.Type = "resource"
+			if c.Resource != nil {
+				if c.Resource.Text != "" {
+					mc.Text = c.Resource.Text
+				} else if c.Resource.Blob != nil {
+					mc.Data = string(c.Resource.Blob)
+				}
+				mc.MimeType = c.Resource.MIMEType
+			}
+		case *mcp.ResourceLink:
+			mc.Type = "resource_link"
+			mc.Text = c.URI
+		default:
+			// 尝试通过 JSON 序列化获取内容
+			if data, err := json.Marshal(content); err == nil {
+				mc.Type = "unknown"
+				mc.Text = string(data)
+			}
+		}
+
+		mcpResult.Content = append(mcpResult.Content, mc)
+	}
+
+	return mcpResult, nil
+}
+
+// callToolSSE 通过 SSE 调用工具
+func (c *MCPClient) callToolSSE(ctx context.Context, name string, arguments map[string]interface{}) (*MCPToolsCallResult, error) {
+	req := map[string]interface{}{
+		"jsonrpc": "2.0",
+		"id":      1,
+		"method":  "tools/call",
+		"params": map[string]interface{}{
+			"name":      name,
+			"arguments": arguments,
+		},
+	}
+
+	resp, err := c.sendSSERequest(ctx, req)
+	if err != nil {
+		return nil, err
+	}
+
+	if errObj, ok := resp["error"].(map[string]interface{}); ok {
+		return nil, fmt.Errorf("MCP error: %v", errObj["message"])
+	}
+
+	resultBytes, _ := json.Marshal(resp["result"])
+	var result MCPToolsCallResult
+	if err := json.Unmarshal(resultBytes, &result); err != nil {
+		return nil, fmt.Errorf("failed to parse tool call result: %v", err)
+	}
+
+	return &result, nil
+}
+
+// setAuthHeaders 设置鉴权请求头
+func (c *MCPClient) setAuthHeaders(req *http.Request) {
+	cfg := c.config
+
+	if cfg.AuthType == "" && cfg.APIKey == "" {
+		return
+	}
+
+	apiKey := expandEnvVars(cfg.APIKey)
+	username := expandEnvVars(cfg.Username)
+	password := expandEnvVars(cfg.Password)
+
+	switch strings.ToLower(cfg.AuthType) {
+	case "bearer":
+		if apiKey != "" {
+			req.Header.Set("Authorization", "Bearer "+apiKey)
+		}
+	case "api_key", "apikey":
+		if apiKey != "" {
+			req.Header.Set("X-API-Key", apiKey)
+		}
+	case "basic":
+		if username != "" {
+			req.SetBasicAuth(username, password)
+		}
+	default:
+		if apiKey != "" {
+			req.Header.Set("Authorization", "Bearer "+apiKey)
+		}
+	}
+}
+
+// sendSSERequest 通过 HTTP 发送请求
+func (c *MCPClient) sendSSERequest(ctx context.Context, req map[string]interface{}) (map[string]interface{}, error) {
+	baseURL := c.sseURL
+	if !strings.HasSuffix(baseURL, "/") {
+		baseURL += "/"
+	}
+
+	postURL := baseURL + "message"
+	if _, err := url.Parse(postURL); err != nil {
+		return nil, fmt.Errorf("invalid URL: %v", err)
+	}
+
+	data, err := json.Marshal(req)
+	if err != nil {
+		return nil, fmt.Errorf("failed to marshal request: %v", err)
+	}
+
+	httpReq, err := http.NewRequestWithContext(ctx, "POST", postURL, bytes.NewBuffer(data))
+	if err != nil {
+		return nil, fmt.Errorf("failed to create HTTP request: %v", err)
+	}
+
+	httpReq.Header.Set("Content-Type", "application/json")
+	c.setAuthHeaders(httpReq)
+
+	for k, v := range c.config.Headers {
+		httpReq.Header.Set(k, expandEnvVars(v))
+	}
+
+	resp, err := c.httpClient.Do(httpReq)
+	if err != nil {
+		return nil, fmt.Errorf("HTTP request failed: %v", err)
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode >= 400 {
+		body, _ := io.ReadAll(resp.Body)
+		return nil, fmt.Errorf("HTTP error %d: %s", resp.StatusCode, string(body))
+	}
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read response: %v", err)
+	}
+
+	var result map[string]interface{}
+	if err := json.Unmarshal(body, &result); err != nil {
+		return nil, fmt.Errorf("failed to parse response: %v", err)
+	}
+
+	return result, nil
+}
+
+// Close 关闭连接
+func (c *MCPClient) Close() error {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	if c.session != nil {
+		c.session.Close()
+		c.session = nil
+	}
+
+	c.client = nil
+	c.initialized = false
+	logger.Infof("MCP client closed: %s", c.config.Name)
+	return nil
+}
+
+// MCPClientManager MCP 客户端管理器
+type MCPClientManager struct {
+	clients map[string]*MCPClient
+	mu      sync.RWMutex
+}
+
+// NewMCPClientManager 创建 MCP 客户端管理器
+func NewMCPClientManager() *MCPClientManager {
+	return &MCPClientManager{
+		clients: make(map[string]*MCPClient),
+	}
+}
+
+// GetOrCreateClient 获取或创建 MCP 客户端
+func (m *MCPClientManager) GetOrCreateClient(ctx context.Context, config *MCPServerConfig) (*MCPClient, error) {
+	m.mu.RLock()
+	client, ok := m.clients[config.Name]
+	m.mu.RUnlock()
+
+	if ok {
+		return client, nil
+	}
+
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	// 再次检查（double-check locking）
+	if client, ok := m.clients[config.Name]; ok {
+		return client, nil
+	}
+
+	// 创建新客户端
+	client, err := NewMCPClient(config)
+	if err != nil {
+		return nil, err
+	}
+
+	// 连接
+	if err := client.Connect(ctx); err != nil {
+		return nil, err
+	}
+
+	m.clients[config.Name] = client
+	return client, nil
+}
+
+// CloseAll 关闭所有客户端
+func (m *MCPClientManager) CloseAll() {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	for name, client := range m.clients {
+		if err := client.Close(); err != nil {
+			logger.Warningf("Failed to close MCP client %s: %v", name, err)
+		}
+	}
+	m.clients = make(map[string]*MCPClient)
+}

aiagent/prompts/embed.go

@@ -0,0 +1,30 @@
+package prompts
+
+import (
+	_ "embed"
+)
+
+// ReAct 模式系统提示词
+//
+//go:embed react_system.md
+var ReactSystemPrompt string
+
+// Plan+ReAct 模式规划阶段系统提示词
+//
+//go:embed plan_system.md
+var PlanSystemPrompt string
+
+// 步骤执行提示词
+//
+//go:embed step_execution.md
+var StepExecutionPrompt string
+
+// 综合分析提示词
+//
+//go:embed synthesis.md
+var SynthesisPrompt string
+
+// 用户提示词默认模板
+//
+//go:embed user_default.md
+var UserDefaultTemplate string

aiagent/prompts/plan_system.md

@@ -0,0 +1,65 @@
+You are an intelligent AI Agent capable of analyzing tasks, creating execution plans, and solving complex problems.
+
+Your role is to understand user requests and create structured, actionable execution plans.
+
+## Core Capabilities
+
+- **Alert Analysis**: Analyze alerts, investigate root causes, correlate events
+- **Data Analysis**: Analyze batch data, identify patterns, generate insights
+- **SQL Generation**: Convert natural language to SQL queries
+- **General Problem Solving**: Break down complex tasks into actionable steps
+
+## Planning Principles
+
+1. **Understand First**: Carefully analyze what the user is asking for
+2. **Identify Key Areas**: Determine which domains, systems, or aspects are involved
+3. **Create Logical Steps**: Order steps by priority or logical sequence
+4. **Be Specific**: Each step should have a clear goal and concrete approach
+5. **Reference Tools**: Consider available tools when designing your approach
+
+## Response Format
+
+You must respond in the following JSON format:
+
+```json
+{
+  "task_summary": "Brief summary of the input/request",
+  "goal": "The overall goal of this task",
+  "focus_areas": ["area1", "area2", "area3"],
+  "steps": [
+    {
+      "step_number": 1,
+      "goal": "What to accomplish in this step",
+      "approach": "How to accomplish it (which tools/methods to use)"
+    },
+    {
+      "step_number": 2,
+      "goal": "...",
+      "approach": "..."
+    }
+  ]
+}
+```
+
+## Focus Areas by Task Type
+
+**Alert/Incident Analysis:**
+- Network: latency, packet loss, DNS resolution
+- Database: query performance, connections, locks, replication
+- Application: error rates, response times, resource usage
+- Infrastructure: CPU, memory, disk I/O, network throughput
+
+**Batch Alert Analysis:**
+- Pattern recognition: common labels, time correlation
+- Aggregation: group by severity, source, category
+- Trend analysis: frequency, escalation patterns
+
+**SQL Generation:**
+- Schema understanding: tables, columns, relationships
+- Query optimization: indexes, join strategies
+- Data validation: constraints, data types
+
+**General Analysis:**
+- Data collection: gather relevant information
+- Processing: transform, filter, aggregate
+- Output: format results appropriately

aiagent/prompts/react_system.md

@@ -0,0 +1,42 @@
+You are an intelligent AI Agent capable of analyzing tasks, creating execution plans, and solving complex problems.
+
+Your capabilities include but are not limited to:
+- **Root Cause Analysis**: Analyze alerts, investigate incidents, identify root causes
+- **Data Analysis**: Query and analyze metrics, logs, traces, and other data sources
+- **SQL Generation**: Convert natural language queries to SQL statements
+- **Information Synthesis**: Summarize and extract insights from complex data
+- **Content Generation**: Generate titles, summaries, and structured reports
+
+## Core Principles
+
+1. **Systematic Analysis**: Gather sufficient information before making conclusions
+2. **Evidence-Based**: Support conclusions with specific data from tool outputs
+3. **Tool Efficiency**: Use tools wisely, avoid redundant calls
+4. **Clear Communication**: Keep responses focused and actionable
+5. **Adaptability**: Adjust your approach based on the task type
+
+## Response Format
+
+You must respond in the following format:
+
+```
+Thought: [Your reasoning about the current situation and what to do next]
+Action: [The tool name to use, or 'Final Answer' if you have enough information]
+Action Input: [The input to the action - for tools, provide JSON parameters; for Final Answer, provide your result]
+```
+
+## Task Guidelines
+
+1. **Understand the request**: Carefully analyze what the user is asking for
+2. **Choose appropriate tools**: Select tools that best fit the task requirements
+3. **Iterate as needed**: Gather additional information if initial results are insufficient
+4. **Validate results**: Verify your conclusions before providing the final answer
+5. **Be concise**: Provide clear, well-structured responses
+
+## Final Answer Requirements
+
+Your Final Answer should:
+- Directly address the user's request
+- Be well-structured and easy to understand
+- Include supporting evidence or reasoning when applicable
+- Provide actionable recommendations if relevant

aiagent/prompts/step_execution.md

@@ -0,0 +1,35 @@
+You are an intelligent AI Agent executing a specific step as part of a larger execution plan.
+
+## Your Task
+
+Focus on completing the current step efficiently and thoroughly. Use the available tools to gather information, process data, or generate results as needed to achieve the step's goal.
+
+## Response Format
+
+Respond in this format:
+
+```
+Thought: [Your reasoning about what to do for this step]
+Action: [Tool name or 'Step Complete' when done]
+Action Input: [Tool parameters as JSON, or step summary for 'Step Complete']
+```
+
+## Step Execution Guidelines
+
+1. **Stay Focused**: Only work on the current step's goal
+2. **Be Thorough**: Gather enough information to achieve the goal
+3. **Document Progress**: Note important findings in your thoughts
+4. **Know When to Stop**: Complete the step when you have sufficient results
+5. **Handle Failures**: If a tool fails, try alternatives or note the limitation
+
+## When to Mark Step Complete
+
+Mark the step as complete when:
+- You have achieved the step's goal
+- You have gathered sufficient information or generated the required output
+- Further work would be outside the step's scope
+
+Your step summary should include:
+- Key results or findings relevant to the step's goal
+- Tools used and their outputs
+- Any limitations or issues encountered

aiagent/prompts/synthesis.md

@@ -0,0 +1,37 @@
+You are an intelligent AI Agent synthesizing results from multiple execution steps into a comprehensive final output.
+
+## Your Task
+
+Review all the results from the completed steps and provide a unified, well-structured response that addresses the original request.
+
+## Response Guidelines
+
+Based on the task type, structure your response appropriately:
+
+**For Root Cause Analysis:**
+- Summary of the root cause
+- Supporting evidence from investigation
+- Impact assessment
+- Recommended actions
+
+**For Data Analysis / SQL Generation:**
+- Query results or generated SQL
+- Key insights from the data
+- Any caveats or limitations
+
+**For Information Synthesis:**
+- Structured summary of findings
+- Key insights and patterns
+- Relevant conclusions
+
+**For Content Generation:**
+- Generated content (title, summary, etc.)
+- Alternative options if applicable
+
+## Synthesis Principles
+
+1. **Integrate Results**: Combine findings from all completed steps coherently
+2. **Prioritize Relevance**: Focus on the most important information
+3. **Be Structured**: Organize output in a clear, logical format
+4. **Be Concise**: Avoid unnecessary verbosity while ensuring completeness
+5. **Address the Request**: Ensure the final output directly answers the original task

aiagent/prompts/user_default.md

@@ -0,0 +1,7 @@
+## Alert Information
+
+{{.AlertContent}}
+
+## Analysis Request
+
+Please analyze this alert and identify the root cause. Provide evidence-based conclusions and actionable recommendations.

aiagent/skill.go

@@ -0,0 +1,573 @@
+package aiagent
+
+import (
+	"bufio"
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"time"
+
+	"github.com/toolkits/pkg/logger"
+	"gopkg.in/yaml.v3"
+)
+
+const (
+	// SkillFileName 技能主文件名
+	SkillFileName = "SKILL.md"
+	// SkillToolsDir 技能工具目录名
+	SkillToolsDir = "skill_tools"
+
+	// 默认配置
+	DefaultMaxSkills = 2
+)
+
+// SkillConfig 技能配置（在 AIAgentConfig 中使用）
+// 技能目录路径通过全局配置 Plus.AIAgentSkillsPath 设置
+type SkillConfig struct {
+	// 技能选择配置（优先级：SkillNames > LLM 选择 > DefaultSkills）
+	AutoSelect    bool     `json:"auto_select,omitempty"`    // 是否让 LLM 自动选择技能（默认 true）
+	SkillNames    []string `json:"skill_names,omitempty"`    // 直接指定技能名列表（手动模式）
+	MaxSkills     int      `json:"max_skills,omitempty"`     // LLM 最多选择几个技能（默认 2）
+	DefaultSkills []string `json:"default_skills,omitempty"` // 默认技能列表（LLM 无法选择时使用）
+}
+
+// SkillMetadata 技能元数据（Level 1 - 总是在内存中）
+type SkillMetadata struct {
+	// 核心字段（与 Anthropic 官方一致）
+	Name        string `yaml:"name" json:"name"`
+	Description string `yaml:"description" json:"description"`
+
+	// 可选扩展字段
+	RecommendedTools []string `yaml:"recommended_tools,omitempty" json:"recommended_tools,omitempty"`
+	BuiltinTools     []string `yaml:"builtin_tools,omitempty" json:"builtin_tools,omitempty"` // 内置工具列表
+
+	// 内部字段
+	Path     string    `json:"-"` // 技能目录路径
+	LoadedAt time.Time `json:"-"` // 加载时间
+}
+
+// SkillContent 技能内容（Level 2 - 匹配时加载）
+type SkillContent struct {
+	Metadata    *SkillMetadata `json:"metadata"`
+	MainContent string         `json:"main_content"` // SKILL.md 正文
+}
+
+// SkillTool Skill 专用工具（Level 3 - 按需加载）
+type SkillTool struct {
+	Name        string                 `yaml:"name" json:"name"`               // 工具名称
+	Type        string                 `yaml:"type" json:"type"`               // 处理器类型：annotation_qd, script, callback 等
+	Description string                 `yaml:"description" json:"description"` // 工具描述
+	Config      map[string]interface{} `yaml:"config" json:"config"`           // 处理器配置
+
+	// 参数定义（可选）
+	Parameters []ToolParameter `yaml:"parameters,omitempty" json:"parameters,omitempty"`
+}
+
+// SkillResources 技能扩展资源（Level 3 - 按需加载）
+type SkillResources struct {
+	SkillTools map[string]*SkillTool `json:"skill_tools"` // 工具名 -> 工具定义
+	References map[string]string     `json:"references"`  // 引用文件内容
+}
+
+// SkillRegistry 技能注册表
+type SkillRegistry struct {
+	skillsPath   string                           // 技能目录路径
+	skills       map[string]*SkillMetadata        // name -> metadata
+	contentCache map[string]*SkillContent         // name -> content (LRU cache)
+	toolsCache   map[string]map[string]*SkillTool // skillName -> toolName -> tool
+	mu           sync.RWMutex
+}
+
+// NewSkillRegistry 创建技能注册表
+func NewSkillRegistry(skillsPath string) *SkillRegistry {
+	registry := &SkillRegistry{
+		skillsPath:   skillsPath,
+		skills:       make(map[string]*SkillMetadata),
+		contentCache: make(map[string]*SkillContent),
+		toolsCache:   make(map[string]map[string]*SkillTool),
+	}
+
+	// 初始加载所有技能元数据
+	if err := registry.loadAllMetadata(); err != nil {
+		logger.Warningf("Failed to load skill metadata: %v", err)
+	}
+
+	return registry
+}
+
+// loadAllMetadata 加载所有技能的元数据（Level 1）
+func (r *SkillRegistry) loadAllMetadata() error {
+	if r.skillsPath == "" {
+		return nil
+	}
+
+	// 检查目录是否存在
+	if _, err := os.Stat(r.skillsPath); os.IsNotExist(err) {
+		logger.Debugf("Skills directory does not exist: %s", r.skillsPath)
+		return nil
+	}
+
+	// 遍历技能目录
+	entries, err := os.ReadDir(r.skillsPath)
+	if err != nil {
+		return fmt.Errorf("failed to read skills directory: %v", err)
+	}
+
+	r.mu.Lock()
+	defer r.mu.Unlock()
+
+	for _, entry := range entries {
+		if !entry.IsDir() {
+			continue
+		}
+
+		skillPath := filepath.Join(r.skillsPath, entry.Name())
+		skillFile := filepath.Join(skillPath, SkillFileName)
+
+		// 检查 SKILL.md 是否存在
+		if _, err := os.Stat(skillFile); os.IsNotExist(err) {
+			continue
+		}
+
+		// 加载元数据
+		metadata, err := r.loadMetadataFromFile(skillFile)
+		if err != nil {
+			logger.Warningf("Failed to load skill metadata from %s: %v", skillFile, err)
+			continue
+		}
+
+		metadata.Path = skillPath
+		metadata.LoadedAt = time.Now()
+		r.skills[metadata.Name] = metadata
+
+		logger.Debugf("Loaded skill metadata: %s from %s", metadata.Name, skillPath)
+	}
+
+	logger.Infof("Loaded %d skills from %s", len(r.skills), r.skillsPath)
+	return nil
+}
+
+// 从 SKILL.md 文件加载元数据
+func (r *SkillRegistry) loadMetadataFromFile(filePath string) (*SkillMetadata, error) {
+	file, err := os.Open(filePath)
+	if err != nil {
+		return nil, fmt.Errorf("failed to open file: %v", err)
+	}
+	defer file.Close()
+
+	// 解析 YAML frontmatter
+	scanner := bufio.NewScanner(file)
+	var inFrontmatter bool
+	var frontmatterLines []string
+
+	for scanner.Scan() {
+		line := strings.TrimSpace(scanner.Text())
+
+		if line == "---" {
+			if !inFrontmatter {
+				inFrontmatter = true
+				continue
+			} else {
+				// frontmatter 结束
+				break
+			}
+		}
+
+		if inFrontmatter {
+			frontmatterLines = append(frontmatterLines, line)
+		}
+	}
+
+	if err := scanner.Err(); err != nil {
+		return nil, fmt.Errorf("failed to scan file: %v", err)
+	}
+
+	if len(frontmatterLines) == 0 {
+		return nil, fmt.Errorf("no frontmatter found in %s", filePath)
+	}
+
+	// 解析 YAML
+	frontmatter := strings.Join(frontmatterLines, "\n")
+	var metadata SkillMetadata
+	if err := yaml.Unmarshal([]byte(frontmatter), &metadata); err != nil {
+		return nil, fmt.Errorf("failed to parse frontmatter: %v", err)
+	}
+
+	if metadata.Name == "" {
+		return nil, fmt.Errorf("skill name is required in frontmatter")
+	}
+
+	return &metadata, nil
+}
+
+// GetByName 根据名称获取技能元数据
+func (r *SkillRegistry) GetByName(name string) *SkillMetadata {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+
+	return r.skills[name]
+}
+
+// ListAll 列出所有技能元数据
+func (r *SkillRegistry) ListAll() []*SkillMetadata {
+	r.mu.RLock()
+	defer r.mu.RUnlock()
+
+	result := make([]*SkillMetadata, 0, len(r.skills))
+	for _, metadata := range r.skills {
+		result = append(result, metadata)
+	}
+	return result
+}
+
+// LoadContent 加载技能内容（Level 2）
+func (r *SkillRegistry) LoadContent(metadata *SkillMetadata) (*SkillContent, error) {
+	if metadata == nil {
+		return nil, fmt.Errorf("metadata is nil")
+	}
+
+	// 检查缓存
+	r.mu.RLock()
+	if cached, ok := r.contentCache[metadata.Name]; ok {
+		r.mu.RUnlock()
+		return cached, nil
+	}
+	r.mu.RUnlock()
+
+	// 加载内容
+	skillFile := filepath.Join(metadata.Path, SkillFileName)
+	content, err := r.loadContentFromFile(skillFile)
+	if err != nil {
+		return nil, err
+	}
+
+	skillContent := &SkillContent{
+		Metadata:    metadata,
+		MainContent: content,
+	}
+
+	// 缓存
+	r.mu.Lock()
+	r.contentCache[metadata.Name] = skillContent
+	r.mu.Unlock()
+
+	return skillContent, nil
+}
+
+// loadContentFromFile 从 SKILL.md 文件加载正文内容
+func (r *SkillRegistry) loadContentFromFile(filePath string) (string, error) {
+	file, err := os.Open(filePath)
+	if err != nil {
+		return "", fmt.Errorf("failed to open file: %v", err)
+	}
+	defer file.Close()
+
+	scanner := bufio.NewScanner(file)
+	var inFrontmatter bool
+	var frontmatterEnded bool
+	var contentLines []string
+
+	for scanner.Scan() {
+		line := scanner.Text()
+
+		if strings.TrimSpace(line) == "---" {
+			if !inFrontmatter {
+				inFrontmatter = true
+				continue
+			} else {
+				frontmatterEnded = true
+				continue
+			}
+		}
+
+		if frontmatterEnded {
+			contentLines = append(contentLines, line)
+		}
+	}
+
+	if err := scanner.Err(); err != nil {
+		return "", fmt.Errorf("failed to scan file: %v", err)
+	}
+
+	return strings.TrimSpace(strings.Join(contentLines, "\n")), nil
+}
+
+// LoadSkillTool 加载单个 skill_tool（Level 3 - 完整配置）
+func (r *SkillRegistry) LoadSkillTool(skillName, toolName string) (*SkillTool, error) {
+	// 检查缓存
+	r.mu.RLock()
+	if skillTools, ok := r.toolsCache[skillName]; ok {
+		if tool, ok := skillTools[toolName]; ok {
+			r.mu.RUnlock()
+			return tool, nil
+		}
+	}
+	r.mu.RUnlock()
+
+	// 获取技能元数据
+	metadata := r.GetByName(skillName)
+	if metadata == nil {
+		return nil, fmt.Errorf("skill '%s' not found", skillName)
+	}
+
+	// 加载工具
+	toolFile := filepath.Join(metadata.Path, SkillToolsDir, toolName+".yaml")
+	tool, err := r.loadToolFromFile(toolFile)
+	if err != nil {
+		return nil, err
+	}
+
+	// 缓存
+	r.mu.Lock()
+	if r.toolsCache[skillName] == nil {
+		r.toolsCache[skillName] = make(map[string]*SkillTool)
+	}
+	r.toolsCache[skillName][toolName] = tool
+	r.mu.Unlock()
+
+	return tool, nil
+}
+
+// LoadSkillToolDescription 加载 skill_tool 的描述信息（轻量级，只读取 name 和 description）
+func (r *SkillRegistry) LoadSkillToolDescription(skillName, toolName string) (string, error) {
+	// 检查缓存 - 如果已经加载了完整工具，直接返回 description
+	r.mu.RLock()
+	if skillTools, ok := r.toolsCache[skillName]; ok {
+		if tool, ok := skillTools[toolName]; ok {
+			r.mu.RUnlock()
+			return tool.Description, nil
+		}
+	}
+	r.mu.RUnlock()
+
+	// 获取技能元数据
+	metadata := r.GetByName(skillName)
+	if metadata == nil {
+		return "", fmt.Errorf("skill '%s' not found", skillName)
+	}
+
+	// 只加载 description（不缓存完整工具，保持延迟加载特性）
+	toolFile := filepath.Join(metadata.Path, SkillToolsDir, toolName+".yaml")
+	tool, err := r.loadToolFromFile(toolFile)
+	if err != nil {
+		return "", err
+	}
+
+	return tool.Description, nil
+}
+
+// LoadAllSkillToolDescriptions 加载技能目录下所有 skill_tools 的描述
+func (r *SkillRegistry) LoadAllSkillToolDescriptions(skillName string) (map[string]string, error) {
+	metadata := r.GetByName(skillName)
+	if metadata == nil {
+		return nil, fmt.Errorf("skill '%s' not found", skillName)
+	}
+
+	toolsDir := filepath.Join(metadata.Path, SkillToolsDir)
+
+	// 检查目录是否存在
+	if _, err := os.Stat(toolsDir); os.IsNotExist(err) {
+		return make(map[string]string), nil
+	}
+
+	// 遍历 skill_tools 目录
+	entries, err := os.ReadDir(toolsDir)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read skill_tools directory: %v", err)
+	}
+
+	descriptions := make(map[string]string)
+	for _, entry := range entries {
+		if entry.IsDir() {
+			continue
+		}
+
+		// 只处理 .yaml 文件
+		name := entry.Name()
+		if !strings.HasSuffix(name, ".yaml") && !strings.HasSuffix(name, ".yml") {
+			continue
+		}
+
+		toolFile := filepath.Join(toolsDir, name)
+		tool, err := r.loadToolFromFile(toolFile)
+		if err != nil {
+			logger.Warningf("Failed to load skill tool %s: %v", toolFile, err)
+			continue
+		}
+
+		descriptions[tool.Name] = tool.Description
+	}
+
+	return descriptions, nil
+}
+
+// loadToolFromFile 从文件加载工具定义
+func (r *SkillRegistry) loadToolFromFile(filePath string) (*SkillTool, error) {
+	data, err := os.ReadFile(filePath)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read tool file: %v", err)
+	}
+
+	var tool SkillTool
+	if err := yaml.Unmarshal(data, &tool); err != nil {
+		return nil, fmt.Errorf("failed to parse tool file: %v", err)
+	}
+
+	return &tool, nil
+}
+
+// LoadReference 加载引用文件（Level 3）
+func (r *SkillRegistry) LoadReference(metadata *SkillMetadata, refName string) (string, error) {
+	if metadata == nil {
+		return "", fmt.Errorf("metadata is nil")
+	}
+
+	refFile := filepath.Join(metadata.Path, refName)
+	data, err := os.ReadFile(refFile)
+	if err != nil {
+		return "", fmt.Errorf("failed to read reference file: %v", err)
+	}
+
+	return string(data), nil
+}
+
+// Reload 重新加载所有技能元数据
+func (r *SkillRegistry) Reload() error {
+	r.mu.Lock()
+	r.skills = make(map[string]*SkillMetadata)
+	r.contentCache = make(map[string]*SkillContent)
+	r.toolsCache = make(map[string]map[string]*SkillTool)
+	r.mu.Unlock()
+
+	return r.loadAllMetadata()
+}
+
+// SkillSelector 技能选择器接口
+type SkillSelector interface {
+	// SelectMultiple 让 LLM 根据任务内容选择最合适的技能（可多选）
+	SelectMultiple(ctx context.Context, taskContext string, availableSkills []*SkillMetadata, maxSkills int) ([]*SkillMetadata, error)
+}
+
+// LLMSkillSelector 基于 LLM 的技能选择器
+type LLMSkillSelector struct {
+	llmCaller func(ctx context.Context, messages []ChatMessage) (string, error)
+}
+
+// NewLLMSkillSelector 创建 LLM 技能选择器
+func NewLLMSkillSelector(llmCaller func(ctx context.Context, messages []ChatMessage) (string, error)) *LLMSkillSelector {
+	return &LLMSkillSelector{
+		llmCaller: llmCaller,
+	}
+}
+
+// SelectMultiple 使用 LLM 选择技能
+func (s *LLMSkillSelector) SelectMultiple(ctx context.Context, taskContext string, availableSkills []*SkillMetadata, maxSkills int) ([]*SkillMetadata, error) {
+	if len(availableSkills) == 0 {
+		return nil, nil
+	}
+
+	if maxSkills <= 0 {
+		maxSkills = DefaultMaxSkills
+	}
+
+	// 构建提示词
+	systemPrompt := s.buildSelectionPrompt(availableSkills, maxSkills)
+
+	messages := []ChatMessage{
+		{Role: "system", Content: systemPrompt},
+		{Role: "user", Content: taskContext},
+	}
+
+	// 调用 LLM
+	response, err := s.llmCaller(ctx, messages)
+	if err != nil {
+		return nil, fmt.Errorf("LLM call failed: %v", err)
+	}
+
+	// 解析响应
+	selectedNames := s.parseSelectionResponse(response)
+	if len(selectedNames) == 0 {
+		return nil, nil
+	}
+
+	// 限制数量
+	if len(selectedNames) > maxSkills {
+		selectedNames = selectedNames[:maxSkills]
+	}
+
+	// 转换为 SkillMetadata
+	skillMap := make(map[string]*SkillMetadata)
+	for _, skill := range availableSkills {
+		skillMap[skill.Name] = skill
+	}
+
+	var result []*SkillMetadata
+	for _, name := range selectedNames {
+		if skill, ok := skillMap[name]; ok {
+			result = append(result, skill)
+		}
+	}
+
+	return result, nil
+}
+
+// buildSelectionPrompt 构建技能选择提示词
+func (s *LLMSkillSelector) buildSelectionPrompt(availableSkills []*SkillMetadata, maxSkills int) string {
+	var sb strings.Builder
+
+	sb.WriteString(fmt.Sprintf(`你是一个技能选择器。根据以下任务上下文，选择最合适的技能（可选择 1-%d 个）。
+
+## 可用技能
+
+`, maxSkills))
+
+	for i, skill := range availableSkills {
+		sb.WriteString(fmt.Sprintf("%d. **%s**\n", i+1, skill.Name))
+		sb.WriteString(fmt.Sprintf("   %s\n\n", skill.Description))
+	}
+
+	sb.WriteString(`## 输出格式
+
+请以 JSON 数组格式返回选中的技能名称，例如：
+` + "```json\n" + `["skill-name-1", "skill-name-2"]
+` + "```" + `
+
+## 选择原则
+
+1. 选择与任务最相关的技能
+2. 如果任务涉及多个领域，可以选择多个技能
+3. 优先选择更具体、更专业的技能
+4. 如果没有合适的技能，返回空数组 []
+
+请返回技能名称数组：`)
+
+	return sb.String()
+}
+
+// parseSelectionResponse 解析 LLM 的选择响应
+func (s *LLMSkillSelector) parseSelectionResponse(response string) []string {
+	// 尝试从 JSON 代码块中提取
+	response = strings.TrimSpace(response)
+
+	// 查找 JSON 数组
+	start := strings.Index(response, "[")
+	end := strings.LastIndex(response, "]")
+
+	if start < 0 || end <= start {
+		return nil
+	}
+
+	jsonStr := response[start : end+1]
+
+	var skillNames []string
+	if err := json.Unmarshal([]byte(jsonStr), &skillNames); err != nil {
+		logger.Warningf("Failed to parse skill selection response: %v", err)
+		return nil
+	}
+
+	return skillNames
+}

center/cconf/conf.go

@@ -21,6 +21,12 @@ type Center struct {
 	CleanPipelineExecutionDay int
 	MigrateBusiGroupLabel     bool
 	RSA                       httpx.RSAConfig
+	AIAgent                   AIAgent
+}
+
+type AIAgent struct {
+	Enable     bool   `toml:"Enable"`
+	SkillsPath string `toml:"SkillsPath"`
 }
 
 type Plugin struct {

center/cconf/ops.go

@@ -300,6 +300,14 @@ ops:
       cname: View Alerting Engines
     - name: /system/version
       cname: View Product Version
+    - name: /ai-config/agents
+      cname: AI Config - Agents
+    - name: /ai-config/llm-configs
+      cname: AI Config - LLM Configs
+    - name: /ai-config/skills
+      cname: AI Config - Skills
+    - name: /ai-config/mcp-servers
+      cname: AI Config - MCP Servers
 
 `
 )

center/router/router.go

@@ -520,6 +520,50 @@ func (rt *Router) Config(r *gin.Engine) {
 		pages.PUT("/config", rt.auth(), rt.admin(), rt.configPutByKey)
 		pages.GET("/site-info", rt.siteInfo)
 
+		// AI Config management
+		pages.GET("/ai-agents", rt.auth(), rt.admin(), rt.aiAgentGets)
+		pages.GET("/ai-agent/:id", rt.auth(), rt.admin(), rt.aiAgentGet)
+		pages.POST("/ai-agents", rt.auth(), rt.admin(), rt.aiAgentAdd)
+		pages.PUT("/ai-agent/:id", rt.auth(), rt.admin(), rt.aiAgentPut)
+		pages.DELETE("/ai-agent/:id", rt.auth(), rt.admin(), rt.aiAgentDel)
+
+		pages.GET("/ai-llm-configs", rt.auth(), rt.admin(), rt.aiLLMConfigGets)
+		pages.GET("/ai-llm-config/:id", rt.auth(), rt.admin(), rt.aiLLMConfigGet)
+		pages.POST("/ai-llm-configs", rt.auth(), rt.admin(), rt.aiLLMConfigAdd)
+		pages.PUT("/ai-llm-config/:id", rt.auth(), rt.admin(), rt.aiLLMConfigPut)
+		pages.DELETE("/ai-llm-config/:id", rt.auth(), rt.admin(), rt.aiLLMConfigDel)
+		pages.POST("/ai-llm-config/test", rt.auth(), rt.admin(), rt.aiLLMConfigTest)
+
+		pages.GET("/ai-skills", rt.auth(), rt.admin(), rt.aiSkillGets)
+		pages.GET("/ai-skill/:id", rt.auth(), rt.admin(), rt.aiSkillGet)
+		pages.POST("/ai-skills", rt.auth(), rt.admin(), rt.aiSkillAdd)
+		pages.PUT("/ai-skill/:id", rt.auth(), rt.admin(), rt.aiSkillPut)
+		pages.DELETE("/ai-skill/:id", rt.auth(), rt.admin(), rt.aiSkillDel)
+		pages.POST("/ai-skills/import", rt.auth(), rt.admin(), rt.aiSkillImport)
+		pages.POST("/ai-skill/:id/files", rt.auth(), rt.admin(), rt.aiSkillFileAdd)
+		pages.GET("/ai-skill-file/:fileId", rt.auth(), rt.admin(), rt.aiSkillFileGet)
+		pages.DELETE("/ai-skill-file/:fileId", rt.auth(), rt.admin(), rt.aiSkillFileDel)
+
+		pages.GET("/mcp-servers", rt.auth(), rt.admin(), rt.mcpServerGets)
+		pages.GET("/mcp-server/:id", rt.auth(), rt.admin(), rt.mcpServerGet)
+		pages.POST("/mcp-servers", rt.auth(), rt.admin(), rt.mcpServerAdd)
+		pages.PUT("/mcp-server/:id", rt.auth(), rt.admin(), rt.mcpServerPut)
+		pages.DELETE("/mcp-server/:id", rt.auth(), rt.admin(), rt.mcpServerDel)
+		pages.POST("/ai-agent/:id/test", rt.auth(), rt.admin(), rt.aiAgentTest)
+		pages.POST("/mcp-server/test", rt.auth(), rt.admin(), rt.mcpServerTest)
+		pages.GET("/mcp-server/:id/tools", rt.auth(), rt.admin(), rt.mcpServerTools)
+
+		// AI Conversations
+		pages.GET("/ai-conversations", rt.auth(), rt.user(), rt.aiConversationGets)
+		pages.POST("/ai-conversations", rt.auth(), rt.user(), rt.aiConversationAdd)
+		pages.GET("/ai-conversation/:id", rt.auth(), rt.user(), rt.aiConversationGet)
+		pages.PUT("/ai-conversation/:id", rt.auth(), rt.user(), rt.aiConversationPut)
+		pages.DELETE("/ai-conversation/:id", rt.auth(), rt.user(), rt.aiConversationDel)
+		pages.POST("/ai-conversation/:id/messages", rt.auth(), rt.user(), rt.aiConversationMessageAdd)
+
+		// AI chat (SSE), dispatches by action_key
+		pages.POST("/ai-chat", rt.auth(), rt.user(), rt.aiChat)
+
 		// source token 相关路由
 		pages.POST("/source-token", rt.auth(), rt.user(), rt.sourceTokenAdd)
 

center/router/router_ai_config.go

@@ -0,0 +1,747 @@
+package router
+
+import (
+	"bytes"
+	"crypto/tls"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/url"
+	"path/filepath"
+	"strings"
+	"time"
+
+	"github.com/ccfos/nightingale/v6/models"
+	"github.com/ccfos/nightingale/v6/pkg/ginx"
+
+	"github.com/gin-gonic/gin"
+	"gopkg.in/yaml.v3"
+)
+
+// ========================
+// AI Agent handlers
+// ========================
+
+func (rt *Router) aiAgentGets(c *gin.Context) {
+	lst, err := models.AIAgentGets(rt.Ctx)
+	ginx.Dangerous(err)
+	ginx.NewRender(c).Data(lst, nil)
+}
+
+func (rt *Router) aiAgentGet(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.AIAgentGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "ai agent not found")
+	}
+	ginx.NewRender(c).Data(obj, nil)
+}
+
+func (rt *Router) aiAgentAdd(c *gin.Context) {
+	var obj models.AIAgent
+	ginx.BindJSON(c, &obj)
+	ginx.Dangerous(obj.Verify())
+
+	me := c.MustGet("user").(*models.User)
+
+	ginx.Dangerous(obj.Create(rt.Ctx, me.Username))
+	ginx.NewRender(c).Data(obj.Id, nil)
+}
+
+func (rt *Router) aiAgentPut(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.AIAgentGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "ai agent not found")
+	}
+
+	var ref models.AIAgent
+	ginx.BindJSON(c, &ref)
+	ginx.Dangerous(ref.Verify())
+
+	me := c.MustGet("user").(*models.User)
+
+	ginx.NewRender(c).Message(obj.Update(rt.Ctx, me.Username, ref))
+}
+
+func (rt *Router) aiAgentDel(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.AIAgentGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "ai agent not found")
+	}
+	ginx.NewRender(c).Message(obj.Delete(rt.Ctx))
+}
+
+// ========================
+// AI Skill handlers
+// ========================
+
+func (rt *Router) aiSkillGets(c *gin.Context) {
+	search := ginx.QueryStr(c, "search", "")
+	lst, err := models.AISkillGets(rt.Ctx, search)
+	ginx.Dangerous(err)
+	ginx.NewRender(c).Data(lst, nil)
+}
+
+func (rt *Router) aiSkillGet(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.AISkillGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "ai skill not found")
+	}
+
+	// Include associated files (without content)
+	files, err := models.AISkillFileGets(rt.Ctx, id)
+	ginx.Dangerous(err)
+	obj.Files = files
+
+	ginx.NewRender(c).Data(obj, nil)
+}
+
+func (rt *Router) aiSkillAdd(c *gin.Context) {
+	var obj models.AISkill
+	ginx.BindJSON(c, &obj)
+	ginx.Dangerous(obj.Verify())
+
+	me := c.MustGet("user").(*models.User)
+	obj.CreatedBy = me.Username
+	obj.UpdatedBy = me.Username
+
+	ginx.Dangerous(obj.Create(rt.Ctx))
+	ginx.NewRender(c).Data(obj.Id, nil)
+}
+
+func (rt *Router) aiSkillPut(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.AISkillGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "ai skill not found")
+	}
+
+	var ref models.AISkill
+	ginx.BindJSON(c, &ref)
+	ginx.Dangerous(ref.Verify())
+
+	me := c.MustGet("user").(*models.User)
+	ref.UpdatedBy = me.Username
+
+	ginx.NewRender(c).Message(obj.Update(rt.Ctx, ref))
+}
+
+func (rt *Router) aiSkillDel(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.AISkillGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "ai skill not found")
+	}
+
+	// Cascade delete skill files
+	ginx.Dangerous(models.AISkillFileDeleteBySkillId(rt.Ctx, id))
+	ginx.NewRender(c).Message(obj.Delete(rt.Ctx))
+}
+
+func (rt *Router) aiSkillImport(c *gin.Context) {
+	file, header, err := c.Request.FormFile("file")
+	ginx.Dangerous(err)
+	defer file.Close()
+
+	ext := strings.ToLower(filepath.Ext(header.Filename))
+	if ext != ".md" {
+		ginx.Bomb(http.StatusBadRequest, "only .md files are supported")
+	}
+
+	content, err := io.ReadAll(file)
+	ginx.Dangerous(err)
+
+	meta, instructions := parseSkillMarkdown(string(content), header.Filename, ext)
+	me := c.MustGet("user").(*models.User)
+
+	skill := models.AISkill{
+		Name:          meta.Name,
+		Description:   meta.Description,
+		Instructions:  instructions,
+		License:       meta.License,
+		Compatibility: meta.Compatibility,
+		Metadata:      meta.Metadata,
+		AllowedTools:  meta.AllowedTools,
+		CreatedBy:     me.Username,
+		UpdatedBy:     me.Username,
+	}
+	ginx.Dangerous(skill.Create(rt.Ctx))
+	ginx.NewRender(c).Data(skill.Id, nil)
+}
+
+// parseSkillMarkdown parses a SKILL.md file with optional YAML frontmatter.
+// Frontmatter format:
+//
+//	---
+//	name: my-skill
+//	description: what this skill does
+//	---
+//	# Actual instructions content...
+type skillFrontmatter struct {
+	Name          string            `yaml:"name"`
+	Description   string            `yaml:"description"`
+	License       string            `yaml:"license"`
+	Compatibility string            `yaml:"compatibility"`
+	Metadata      map[string]string `yaml:"metadata"`
+	AllowedTools  string            `yaml:"allowed-tools"`
+}
+
+func parseSkillMarkdown(content, filename, ext string) (meta skillFrontmatter, instructions string) {
+	text := strings.TrimSpace(content)
+
+	// Try to parse YAML frontmatter (between --- delimiters)
+	if strings.HasPrefix(text, "---") {
+		endIdx := strings.Index(text[3:], "\n---")
+		if endIdx >= 0 {
+			frontmatter := text[3 : 3+endIdx]
+			body := strings.TrimSpace(text[3+endIdx+4:]) // skip past closing ---
+
+			if yaml.Unmarshal([]byte(frontmatter), &meta) == nil && meta.Name != "" {
+				return meta, body
+			}
+		}
+	}
+
+	// No valid frontmatter, fallback: filename as name, entire content as instructions
+	meta.Name = strings.TrimSuffix(filename, ext)
+	return meta, content
+}
+
+// ========================
+// AI Skill File handlers
+// ========================
+
+func (rt *Router) aiSkillFileAdd(c *gin.Context) {
+	skillId := ginx.UrlParamInt64(c, "id")
+
+	// Verify skill exists
+	skill, err := models.AISkillGetById(rt.Ctx, skillId)
+	ginx.Dangerous(err)
+	if skill == nil {
+		ginx.Bomb(http.StatusNotFound, "ai skill not found")
+	}
+
+	file, header, err := c.Request.FormFile("file")
+	ginx.Dangerous(err)
+	defer file.Close()
+
+	// Validate file extension
+	ext := strings.ToLower(filepath.Ext(header.Filename))
+	allowed := map[string]bool{".md": true, ".txt": true, ".json": true, ".yaml": true, ".yml": true, ".csv": true}
+	if !allowed[ext] {
+		ginx.Bomb(http.StatusBadRequest, "file type not allowed, only .md/.txt/.json/.yaml/.csv")
+	}
+
+	// Validate file size (2MB max)
+	if header.Size > 2*1024*1024 {
+		ginx.Bomb(http.StatusBadRequest, "file size exceeds 2MB limit")
+	}
+
+	content, err := io.ReadAll(file)
+	ginx.Dangerous(err)
+
+	me := c.MustGet("user").(*models.User)
+	skillFile := models.AISkillFile{
+		SkillId:   skillId,
+		Name:      header.Filename,
+		Content:   string(content),
+		CreatedBy: me.Username,
+	}
+	ginx.Dangerous(skillFile.Create(rt.Ctx))
+	ginx.NewRender(c).Data(skillFile.Id, nil)
+}
+
+func (rt *Router) aiSkillFileGet(c *gin.Context) {
+	fileId := ginx.UrlParamInt64(c, "fileId")
+	obj, err := models.AISkillFileGetById(rt.Ctx, fileId)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "file not found")
+	}
+	ginx.NewRender(c).Data(obj, nil)
+}
+
+func (rt *Router) aiSkillFileDel(c *gin.Context) {
+	fileId := ginx.UrlParamInt64(c, "fileId")
+	obj, err := models.AISkillFileGetById(rt.Ctx, fileId)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "file not found")
+	}
+	ginx.NewRender(c).Message(obj.Delete(rt.Ctx))
+}
+
+// ========================
+// MCP Server handlers
+// ========================
+
+func (rt *Router) mcpServerGets(c *gin.Context) {
+	lst, err := models.MCPServerGets(rt.Ctx)
+	ginx.Dangerous(err)
+	ginx.NewRender(c).Data(lst, nil)
+}
+
+func (rt *Router) mcpServerGet(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.MCPServerGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "mcp server not found")
+	}
+	ginx.NewRender(c).Data(obj, nil)
+}
+
+func (rt *Router) mcpServerAdd(c *gin.Context) {
+	var obj models.MCPServer
+	ginx.BindJSON(c, &obj)
+	ginx.Dangerous(obj.Verify())
+
+	me := c.MustGet("user").(*models.User)
+	obj.CreatedBy = me.Username
+	obj.UpdatedBy = me.Username
+
+	ginx.Dangerous(obj.Create(rt.Ctx))
+	ginx.NewRender(c).Data(obj.Id, nil)
+}
+
+func (rt *Router) mcpServerPut(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.MCPServerGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "mcp server not found")
+	}
+
+	var ref models.MCPServer
+	ginx.BindJSON(c, &ref)
+	ginx.Dangerous(ref.Verify())
+
+	me := c.MustGet("user").(*models.User)
+	ref.UpdatedBy = me.Username
+
+	ginx.NewRender(c).Message(obj.Update(rt.Ctx, ref))
+}
+
+func (rt *Router) mcpServerDel(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.MCPServerGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "mcp server not found")
+	}
+	ginx.NewRender(c).Message(obj.Delete(rt.Ctx))
+}
+
+// ========================
+// AI LLM Config handlers
+// ========================
+
+func (rt *Router) aiLLMConfigGets(c *gin.Context) {
+	lst, err := models.AILLMConfigGets(rt.Ctx)
+	ginx.Dangerous(err)
+	ginx.NewRender(c).Data(lst, nil)
+}
+
+func (rt *Router) aiLLMConfigGet(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.AILLMConfigGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "ai llm config not found")
+	}
+	ginx.NewRender(c).Data(obj, nil)
+}
+
+func (rt *Router) aiLLMConfigAdd(c *gin.Context) {
+	var obj models.AILLMConfig
+	ginx.BindJSON(c, &obj)
+	ginx.Dangerous(obj.Verify())
+
+	me := c.MustGet("user").(*models.User)
+
+	ginx.Dangerous(obj.Create(rt.Ctx, me.Username))
+	ginx.NewRender(c).Data(obj.Id, nil)
+}
+
+func (rt *Router) aiLLMConfigPut(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.AILLMConfigGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "ai llm config not found")
+	}
+
+	var ref models.AILLMConfig
+	ginx.BindJSON(c, &ref)
+	ginx.Dangerous(ref.Verify())
+
+	me := c.MustGet("user").(*models.User)
+
+	ginx.NewRender(c).Message(obj.Update(rt.Ctx, me.Username, ref))
+}
+
+func (rt *Router) aiLLMConfigDel(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.AILLMConfigGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "ai llm config not found")
+	}
+	ginx.NewRender(c).Message(obj.Delete(rt.Ctx))
+}
+
+func (rt *Router) aiLLMConfigTest(c *gin.Context) {
+	var body struct {
+		APIType     string                `json:"api_type"`
+		APIURL      string                `json:"api_url"`
+		APIKey      string                `json:"api_key"`
+		Model       string                `json:"model"`
+		ExtraConfig models.LLMExtraConfig `json:"extra_config"`
+	}
+	ginx.BindJSON(c, &body)
+
+	if body.APIType == "" || body.APIURL == "" || body.APIKey == "" || body.Model == "" {
+		ginx.Bomb(http.StatusBadRequest, "api_type, api_url, api_key, model are required")
+	}
+
+	obj := &models.AILLMConfig{
+		APIType:     body.APIType,
+		APIURL:      body.APIURL,
+		APIKey:      body.APIKey,
+		Model:       body.Model,
+		ExtraConfig: body.ExtraConfig,
+	}
+
+	start := time.Now()
+	testErr := testAIAgent(obj)
+	durationMs := time.Since(start).Milliseconds()
+
+	result := gin.H{
+		"success":     testErr == nil,
+		"duration_ms": durationMs,
+	}
+	ginx.NewRender(c).Data(result, testErr)
+}
+
+// ========================
+// AI Agent test
+// ========================
+
+func (rt *Router) aiAgentTest(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+
+	agent, err := models.AIAgentGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if agent == nil {
+		ginx.Bomb(http.StatusNotFound, "ai agent not found")
+	}
+
+	llmCfg, err := models.AILLMConfigGetById(rt.Ctx, agent.LLMConfigId)
+	ginx.Dangerous(err)
+	if llmCfg == nil {
+		ginx.Bomb(http.StatusBadRequest, "referenced LLM config not found")
+	}
+
+	start := time.Now()
+	testErr := testAIAgent(llmCfg)
+	durationMs := time.Since(start).Milliseconds()
+
+	result := gin.H{
+		"success":     testErr == nil,
+		"duration_ms": durationMs,
+	}
+	if testErr != nil {
+		result["error"] = testErr.Error()
+	}
+	ginx.NewRender(c).Data(result, nil)
+}
+
+func testAIAgent(p *models.AILLMConfig) error {
+	extra := p.ExtraConfig
+
+	// Build HTTP client with ExtraConfig settings
+	timeout := 30 * time.Second
+	if extra.TimeoutSeconds > 0 {
+		timeout = time.Duration(extra.TimeoutSeconds) * time.Second
+	}
+
+	transport := &http.Transport{}
+	if extra.SkipTLSVerify {
+		transport.TLSClientConfig = &tls.Config{InsecureSkipVerify: true}
+	}
+	if extra.Proxy != "" {
+		if proxyURL, err := url.Parse(extra.Proxy); err == nil {
+			transport.Proxy = http.ProxyURL(proxyURL)
+		}
+	}
+
+	client := &http.Client{Timeout: timeout, Transport: transport}
+
+	var reqURL string
+	var reqBody []byte
+	hdrs := map[string]string{"Content-Type": "application/json"}
+
+	switch p.APIType {
+	case "openai":
+		base := strings.TrimRight(p.APIURL, "/")
+		if strings.HasSuffix(base, "/chat/completions") {
+			reqURL = base
+		} else {
+			reqURL = base + "/chat/completions"
+		}
+		reqBody, _ = json.Marshal(map[string]interface{}{
+			"model":      p.Model,
+			"messages":   []map[string]string{{"role": "user", "content": "Hi"}},
+			"max_tokens": 5,
+		})
+		hdrs["Authorization"] = "Bearer " + p.APIKey
+	case "claude":
+		reqURL = strings.TrimRight(p.APIURL, "/") + "/v1/messages"
+		reqBody, _ = json.Marshal(map[string]interface{}{
+			"model":      p.Model,
+			"messages":   []map[string]string{{"role": "user", "content": "Hi"}},
+			"max_tokens": 5,
+		})
+		hdrs["x-api-key"] = p.APIKey
+		hdrs["anthropic-version"] = "2023-06-01"
+	case "gemini":
+		reqURL = strings.TrimRight(p.APIURL, "/") + "/v1beta/models/" + p.Model + ":generateContent?key=" + p.APIKey
+		reqBody, _ = json.Marshal(map[string]interface{}{
+			"contents": []map[string]interface{}{
+				{"parts": []map[string]string{{"text": "Hi"}}},
+			},
+		})
+	default:
+		return fmt.Errorf("unsupported api_type: %s", p.APIType)
+	}
+
+	req, err := http.NewRequest("POST", reqURL, bytes.NewReader(reqBody))
+	if err != nil {
+		return err
+	}
+	for k, v := range hdrs {
+		req.Header.Set(k, v)
+	}
+	// Apply custom headers from ExtraConfig
+	for k, v := range extra.CustomHeaders {
+		req.Header.Set(k, v)
+	}
+
+	resp, err := client.Do(req)
+	if err != nil {
+		return err
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode >= 400 {
+		body, _ := io.ReadAll(resp.Body)
+		if len(body) > 500 {
+			body = body[:500]
+		}
+		return fmt.Errorf("HTTP %d: %s", resp.StatusCode, string(body))
+	}
+	return nil
+}
+
+// ========================
+// MCP Server test & tools
+// ========================
+
+func (rt *Router) mcpServerTest(c *gin.Context) {
+	var body struct {
+		URL     string            `json:"url"`
+		Headers map[string]string `json:"headers"`
+	}
+	ginx.BindJSON(c, &body)
+
+	if body.URL == "" {
+		ginx.Bomb(http.StatusBadRequest, "url is required")
+	}
+
+	obj := &models.MCPServer{
+		URL:     body.URL,
+		Headers: body.Headers,
+	}
+
+	start := time.Now()
+	tools, testErr := listMCPTools(obj)
+	durationMs := time.Since(start).Milliseconds()
+
+	result := gin.H{
+		"success":     testErr == nil,
+		"duration_ms": durationMs,
+		"tool_count":  len(tools),
+	}
+	if testErr != nil {
+		result["error"] = testErr.Error()
+	}
+	ginx.NewRender(c).Data(result, nil)
+}
+
+func (rt *Router) mcpServerTools(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.MCPServerGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "mcp server not found")
+	}
+
+	tools, err := listMCPTools(obj)
+	ginx.Dangerous(err)
+	ginx.NewRender(c).Data(tools, nil)
+}
+
+type mcpTool struct {
+	Name        string `json:"name"`
+	Description string `json:"description"`
+}
+
+func listMCPTools(s *models.MCPServer) ([]mcpTool, error) {
+	client := &http.Client{Timeout: 30 * time.Second}
+
+	hdrs := s.Headers
+
+	// Step 1: Initialize
+	initResp, initSessionID, err := sendMCPRPC(client, s.URL, hdrs, "", 1, "initialize", map[string]interface{}{
+		"protocolVersion": "2024-11-05",
+		"capabilities":    map[string]interface{}{},
+		"clientInfo":      map[string]interface{}{"name": "nightingale", "version": "1.0.0"},
+	})
+	if err != nil {
+		return nil, fmt.Errorf("initialize: %v", err)
+	}
+	_ = initResp
+
+	// Send initialized notification
+	sendMCPRPC(client, s.URL, hdrs, initSessionID, 0, "notifications/initialized", map[string]interface{}{})
+
+	// Step 2: List tools
+	toolsResp, _, err := sendMCPRPC(client, s.URL, hdrs, initSessionID, 2, "tools/list", map[string]interface{}{})
+	if err != nil {
+		return nil, fmt.Errorf("tools/list: %v", err)
+	}
+
+	if toolsResp == nil || toolsResp.Result == nil {
+		return []mcpTool{}, nil
+	}
+
+	toolsRaw, ok := toolsResp.Result["tools"]
+	if !ok {
+		return []mcpTool{}, nil
+	}
+
+	toolsJSON, _ := json.Marshal(toolsRaw)
+	var tools []mcpTool
+	json.Unmarshal(toolsJSON, &tools)
+	return tools, nil
+}
+
+type jsonRPCResponse struct {
+	JSONRPC string                 `json:"jsonrpc"`
+	ID      interface{}            `json:"id"`
+	Result  map[string]interface{} `json:"result"`
+	Error   *jsonRPCError          `json:"error"`
+}
+
+type jsonRPCError struct {
+	Code    int    `json:"code"`
+	Message string `json:"message"`
+}
+
+func sendMCPRPC(client *http.Client, serverURL string, hdrs map[string]string, sessionID string, id int, method string, params interface{}) (*jsonRPCResponse, string, error) {
+	body := map[string]interface{}{
+		"jsonrpc": "2.0",
+		"method":  method,
+		"params":  params,
+	}
+	if id > 0 {
+		body["id"] = id
+	}
+
+	reqBody, _ := json.Marshal(body)
+	req, err := http.NewRequest("POST", serverURL, bytes.NewReader(reqBody))
+	if err != nil {
+		return nil, "", err
+	}
+
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("Accept", "application/json, text/event-stream")
+	if sessionID != "" {
+		req.Header.Set("Mcp-Session-Id", sessionID)
+	}
+	for k, v := range hdrs {
+		req.Header.Set(k, v)
+	}
+
+	resp, err := client.Do(req)
+	if err != nil {
+		return nil, "", err
+	}
+	defer resp.Body.Close()
+
+	newSessionID := resp.Header.Get("Mcp-Session-Id")
+	if newSessionID == "" {
+		newSessionID = sessionID
+	}
+
+	// Notification (no id) - no response body expected
+	if id <= 0 {
+		return nil, newSessionID, nil
+	}
+
+	if resp.StatusCode >= 400 {
+		respBody, _ := io.ReadAll(resp.Body)
+		if len(respBody) > 500 {
+			respBody = respBody[:500]
+		}
+		return nil, newSessionID, fmt.Errorf("HTTP %d: %s", resp.StatusCode, string(respBody))
+	}
+
+	respBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, newSessionID, err
+	}
+
+	// Handle SSE response
+	contentType := resp.Header.Get("Content-Type")
+	if strings.Contains(contentType, "text/event-stream") {
+		for _, line := range strings.Split(string(respBody), "\n") {
+			if strings.HasPrefix(line, "data: ") {
+				data := strings.TrimPrefix(line, "data: ")
+				var rpcResp jsonRPCResponse
+				if json.Unmarshal([]byte(data), &rpcResp) == nil && (rpcResp.Result != nil || rpcResp.Error != nil) {
+					if rpcResp.Error != nil {
+						return &rpcResp, newSessionID, fmt.Errorf("RPC error %d: %s", rpcResp.Error.Code, rpcResp.Error.Message)
+					}
+					return &rpcResp, newSessionID, nil
+				}
+			}
+		}
+		return nil, newSessionID, fmt.Errorf("no valid JSON-RPC response in SSE stream")
+	}
+
+	// Handle JSON response
+	var rpcResp jsonRPCResponse
+	if err := json.Unmarshal(respBody, &rpcResp); err != nil {
+		if len(respBody) > 200 {
+			respBody = respBody[:200]
+		}
+		return nil, newSessionID, fmt.Errorf("invalid response: %s", string(respBody))
+	}
+
+	if rpcResp.Error != nil {
+		return &rpcResp, newSessionID, fmt.Errorf("RPC error %d: %s", rpcResp.Error.Code, rpcResp.Error.Message)
+	}
+
+	return &rpcResp, newSessionID, nil
+}

center/router/router_ai_conversation.go

@@ -0,0 +1,114 @@
+package router
+
+import (
+	"net/http"
+
+	"github.com/ccfos/nightingale/v6/models"
+	"github.com/ccfos/nightingale/v6/pkg/ginx"
+
+	"github.com/gin-gonic/gin"
+)
+
+func (rt *Router) aiConversationGets(c *gin.Context) {
+	me := c.MustGet("user").(*models.User)
+	lst, err := models.AIConversationGetsByUserId(rt.Ctx, me.Id)
+	ginx.Dangerous(err)
+	ginx.NewRender(c).Data(lst, nil)
+}
+
+func (rt *Router) aiConversationAdd(c *gin.Context) {
+	var obj models.AIConversation
+	ginx.BindJSON(c, &obj)
+
+	me := c.MustGet("user").(*models.User)
+	obj.UserId = me.Id
+	ginx.Dangerous(obj.Verify())
+	ginx.Dangerous(obj.Create(rt.Ctx))
+	ginx.NewRender(c).Data(obj, nil)
+}
+
+func (rt *Router) aiConversationGet(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.AIConversationGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "conversation not found")
+	}
+
+	me := c.MustGet("user").(*models.User)
+	if obj.UserId != me.Id {
+		ginx.Bomb(http.StatusForbidden, "forbidden")
+	}
+
+	messages, err := models.AIConversationMessageGetsByConversationId(rt.Ctx, id)
+	ginx.Dangerous(err)
+
+	ginx.NewRender(c).Data(gin.H{
+		"conversation": obj,
+		"messages":     messages,
+	}, nil)
+}
+
+func (rt *Router) aiConversationPut(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.AIConversationGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "conversation not found")
+	}
+
+	me := c.MustGet("user").(*models.User)
+	if obj.UserId != me.Id {
+		ginx.Bomb(http.StatusForbidden, "forbidden")
+	}
+
+	var body struct {
+		Title string `json:"title"`
+	}
+	ginx.BindJSON(c, &body)
+
+	ginx.NewRender(c).Message(obj.Update(rt.Ctx, body.Title))
+}
+
+func (rt *Router) aiConversationDel(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.AIConversationGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "conversation not found")
+	}
+
+	me := c.MustGet("user").(*models.User)
+	if obj.UserId != me.Id {
+		ginx.Bomb(http.StatusForbidden, "forbidden")
+	}
+
+	ginx.NewRender(c).Message(obj.Delete(rt.Ctx))
+}
+
+func (rt *Router) aiConversationMessageAdd(c *gin.Context) {
+	id := ginx.UrlParamInt64(c, "id")
+	obj, err := models.AIConversationGetById(rt.Ctx, id)
+	ginx.Dangerous(err)
+	if obj == nil {
+		ginx.Bomb(http.StatusNotFound, "conversation not found")
+	}
+
+	me := c.MustGet("user").(*models.User)
+	if obj.UserId != me.Id {
+		ginx.Bomb(http.StatusForbidden, "forbidden")
+	}
+
+	var msgs []models.AIConversationMessage
+	ginx.BindJSON(c, &msgs)
+
+	for i := range msgs {
+		msgs[i].ConversationId = id
+		ginx.Dangerous(msgs[i].Create(rt.Ctx))
+	}
+
+	// Update conversation timestamp
+	obj.UpdateTime(rt.Ctx)
+
+	ginx.NewRender(c).Message(nil)
+}

center/router/router_aiagent.go

@@ -0,0 +1,345 @@
+package router
+
+import (
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"time"
+
+	"github.com/ccfos/nightingale/v6/aiagent"
+	"github.com/ccfos/nightingale/v6/models"
+	"github.com/ccfos/nightingale/v6/pkg/ginx"
+	"github.com/ccfos/nightingale/v6/pkg/prom"
+
+	"github.com/gin-gonic/gin"
+	"github.com/toolkits/pkg/logger"
+)
+
+// AIChatRequest is the generic chat request dispatched by action_key.
+type AIChatRequest struct {
+	ActionKey string                `json:"action_key"` // e.g. "query_generator"
+	UserInput string                `json:"user_input"`
+	History   []aiagent.ChatMessage `json:"history,omitempty"`
+	Context   map[string]interface{} `json:"context,omitempty"` // action-specific params
+}
+
+// actionHandler defines how each action_key is processed.
+type actionHandler struct {
+	useCase     string // maps to AIAgent.UseCase for finding the right agent config
+	validate    func(req *AIChatRequest) error
+	selectTools func(req *AIChatRequest) []string
+	buildPrompt func(req *AIChatRequest) string
+	buildInputs func(req *AIChatRequest) map[string]string
+}
+
+var actionRegistry = map[string]*actionHandler{
+	"query_generator": {
+		useCase:     "chat",
+		validate:    validateQueryGenerator,
+		selectTools: selectQueryGeneratorTools,
+		buildPrompt: buildQueryGeneratorPrompt,
+		buildInputs: buildQueryGeneratorInputs,
+	},
+}
+
+// --- query_generator action ---
+
+func ctxStr(ctx map[string]interface{}, key string) string {
+	if v, ok := ctx[key]; ok {
+		if s, ok := v.(string); ok {
+			return s
+		}
+	}
+	return ""
+}
+
+func ctxInt64(ctx map[string]interface{}, key string) int64 {
+	if v, ok := ctx[key]; ok {
+		switch n := v.(type) {
+		case float64:
+			return int64(n)
+		case int64:
+			return n
+		case json.Number:
+			i, _ := n.Int64()
+			return i
+		}
+	}
+	return 0
+}
+
+func validateQueryGenerator(req *AIChatRequest) error {
+	dsType := ctxStr(req.Context, "datasource_type")
+	dsID := ctxInt64(req.Context, "datasource_id")
+	if dsType == "" {
+		return fmt.Errorf("context.datasource_type is required")
+	}
+	if dsID == 0 {
+		return fmt.Errorf("context.datasource_id is required")
+	}
+	return nil
+}
+
+func selectQueryGeneratorTools(req *AIChatRequest) []string {
+	dsType := ctxStr(req.Context, "datasource_type")
+	switch dsType {
+	case "prometheus":
+		return []string{"list_metrics", "get_metric_labels"}
+	case "mysql", "doris", "ck", "clickhouse", "pgsql", "postgresql":
+		return []string{"list_databases", "list_tables", "describe_table"}
+	default:
+		return nil
+	}
+}
+
+func buildQueryGeneratorPrompt(req *AIChatRequest) string {
+	dsType := ctxStr(req.Context, "datasource_type")
+	dbName := ctxStr(req.Context, "database_name")
+	tableName := ctxStr(req.Context, "table_name")
+
+	switch dsType {
+	case "prometheus":
+		return fmt.Sprintf(`You are a PromQL expert. The user wants to query Prometheus metrics.
+
+User request: %s
+
+Please use the available tools to explore the metrics and generate the correct PromQL query.
+- First use list_metrics to find relevant metrics
+- Then use get_metric_labels to understand the label structure
+- Finally provide the PromQL query as your Final Answer
+
+Your Final Answer MUST be a valid JSON object with these fields:
+{"query": "<the PromQL query>", "explanation": "<brief explanation in the user's language>"}`, req.UserInput)
+
+	default: // SQL-based datasources
+		dbContext := ""
+		if dbName != "" {
+			dbContext += fmt.Sprintf("\nTarget database: %s", dbName)
+		}
+		if tableName != "" {
+			dbContext += fmt.Sprintf("\nTarget table: %s", tableName)
+		}
+
+		return fmt.Sprintf(`You are a SQL expert for %s databases. The user wants to query data.
+%s
+User request: %s
+
+Please use the available tools to explore the database schema and generate the correct SQL query.
+- Use list_databases to see available databases
+- Use list_tables to see tables in the target database
+- Use describe_table to understand the table structure
+- Finally provide the SQL query as your Final Answer
+
+Your Final Answer MUST be a valid JSON object with these fields:
+{"query": "<the SQL query>", "explanation": "<brief explanation in the user's language>"}`, dsType, dbContext, req.UserInput)
+	}
+}
+
+func buildQueryGeneratorInputs(req *AIChatRequest) map[string]string {
+	inputs := map[string]string{
+		"user_input": req.UserInput,
+	}
+	for _, key := range []string{"datasource_type", "datasource_id", "database_name", "table_name"} {
+		if v := ctxStr(req.Context, key); v != "" {
+			inputs[key] = v
+		}
+	}
+	// datasource_id may be a number in JSON
+	if inputs["datasource_id"] == "" {
+		if id := ctxInt64(req.Context, "datasource_id"); id > 0 {
+			inputs["datasource_id"] = fmt.Sprintf("%d", id)
+		}
+	}
+	return inputs
+}
+
+// --- generic handler ---
+
+func (rt *Router) aiChat(c *gin.Context) {
+	if !rt.Center.AIAgent.Enable {
+		ginx.Bomb(http.StatusServiceUnavailable, "AI Agent is not enabled")
+		return
+	}
+
+	var req AIChatRequest
+	ginx.BindJSON(c, &req)
+
+	if req.UserInput == "" {
+		ginx.Bomb(http.StatusBadRequest, "user_input is required")
+		return
+	}
+	if req.ActionKey == "" {
+		ginx.Bomb(http.StatusBadRequest, "action_key is required")
+		return
+	}
+	if req.Context == nil {
+		req.Context = make(map[string]interface{})
+	}
+
+	handler, ok := actionRegistry[req.ActionKey]
+	if !ok {
+		ginx.Bomb(http.StatusBadRequest, "unsupported action_key: %s", req.ActionKey)
+		return
+	}
+
+	logger.Infof("[AIChat] action=%s, user_input=%q", req.ActionKey, truncStr(req.UserInput, 100))
+
+	// Action-specific validation
+	if handler.validate != nil {
+		if err := handler.validate(&req); err != nil {
+			ginx.Bomb(http.StatusBadRequest, err.Error())
+			return
+		}
+	}
+
+	// Find AI agent by use_case
+	agent, err := models.AIAgentGetByUseCase(rt.Ctx, handler.useCase)
+	if err != nil || agent == nil {
+		ginx.Bomb(http.StatusBadRequest, "no AI agent configured for use_case=%s", handler.useCase)
+		return
+	}
+
+	// Resolve LLM config
+	llmCfg, err := models.AILLMConfigGetById(rt.Ctx, agent.LLMConfigId)
+	if err != nil || llmCfg == nil {
+		ginx.Bomb(http.StatusBadRequest, "referenced LLM config not found")
+		return
+	}
+	agent.LLMConfig = llmCfg
+
+	// Select tools
+	var tools []aiagent.AgentTool
+	if handler.selectTools != nil {
+		toolNames := handler.selectTools(&req)
+		if toolNames != nil {
+			tools = aiagent.GetBuiltinToolDefs(toolNames)
+		}
+	}
+
+	// Parse extra config
+	extraConfig := llmCfg.ExtraConfig
+
+	timeout := 120000
+	if extraConfig.TimeoutSeconds > 0 {
+		timeout = extraConfig.TimeoutSeconds * 1000
+	}
+
+	// Build prompt
+	userPrompt := ""
+	if handler.buildPrompt != nil {
+		userPrompt = handler.buildPrompt(&req)
+	}
+
+	// Build workflow inputs
+	inputs := map[string]string{"user_input": req.UserInput}
+	if handler.buildInputs != nil {
+		inputs = handler.buildInputs(&req)
+	}
+
+	// Create agent
+	agentCfg := aiagent.NewAgent(&aiagent.AIAgentConfig{
+		Provider:           llmCfg.APIType,
+		LLMURL:             llmCfg.APIURL,
+		Model:              llmCfg.Model,
+		APIKey:             llmCfg.APIKey,
+		Headers:            extraConfig.CustomHeaders,
+		AgentMode:          aiagent.AgentModeReAct,
+		Tools:              tools,
+		Timeout:            timeout,
+		Stream:             true,
+		UserPromptTemplate: userPrompt,
+		SkipSSLVerify:      extraConfig.SkipTLSVerify,
+		Proxy:              extraConfig.Proxy,
+		Temperature:        extraConfig.Temperature,
+		MaxTokens:          extraConfig.MaxTokens,
+	})
+
+	// Inject PromClient getter
+	aiagent.SetPromClientGetter(func(dsId int64) prom.API {
+		return rt.PromClients.GetCli(dsId)
+	})
+
+	// Streaming setup
+	streamChan := make(chan *models.StreamChunk, 100)
+	wfCtx := &models.WorkflowContext{
+		Stream:     true,
+		StreamChan: streamChan,
+		Inputs:     inputs,
+	}
+
+	c.Header("Content-Type", "text/event-stream")
+	c.Header("Cache-Control", "no-cache")
+	c.Header("Connection", "keep-alive")
+	c.Header("X-Accel-Buffering", "no")
+
+	startTime := time.Now()
+
+	go func() {
+		defer func() {
+			if r := recover(); r != nil {
+				logger.Errorf("[AIChat] PANIC in agent goroutine: %v", r)
+				streamChan <- &models.StreamChunk{
+					Type:      models.StreamTypeError,
+					Content:   fmt.Sprintf("internal error: %v", r),
+					Done:      true,
+					Timestamp: time.Now().UnixMilli(),
+				}
+				close(streamChan)
+			}
+		}()
+		_, _, err := agentCfg.Process(rt.Ctx, wfCtx)
+		if err != nil {
+			logger.Errorf("[AIChat] agent Process error: %v", err)
+		}
+	}()
+
+	// Stream SSE events
+	var accumulatedMessage string
+	c.Stream(func(w io.Writer) bool {
+		chunk, ok := <-streamChan
+		if !ok {
+			return false
+		}
+
+		data, _ := json.Marshal(chunk)
+
+		if chunk.Type == models.StreamTypeText || chunk.Type == models.StreamTypeThinking {
+			if chunk.Delta != "" {
+				accumulatedMessage += chunk.Delta
+			} else if chunk.Content != "" {
+				accumulatedMessage += chunk.Content
+			}
+		}
+
+		if chunk.Type == models.StreamTypeError {
+			fmt.Fprintf(w, "event: error\ndata: %s\n\n", data)
+			c.Writer.Flush()
+			return false
+		}
+
+		if chunk.Done || chunk.Type == models.StreamTypeDone {
+			doneData := map[string]interface{}{
+				"type":        "done",
+				"duration_ms": time.Since(startTime).Milliseconds(),
+				"message":     accumulatedMessage,
+				"response":    chunk.Content,
+			}
+			finalData, _ := json.Marshal(doneData)
+			fmt.Fprintf(w, "event: done\ndata: %s\n\n", finalData)
+			c.Writer.Flush()
+			return false
+		}
+
+		fmt.Fprintf(w, "event: chunk\ndata: %s\n\n", data)
+		c.Writer.Flush()
+		return true
+	})
+}
+
+func truncStr(s string, maxLen int) string {
+	if len(s) <= maxLen {
+		return s
+	}
+	return s[:maxLen] + "..."
+}

center/router/router_alert_rule.go

@@ -13,10 +13,10 @@ import (
 
 	"github.com/ccfos/nightingale/v6/alert/mute"
 	"github.com/ccfos/nightingale/v6/models"
+	"github.com/ccfos/nightingale/v6/pkg/ginx"
 	"github.com/ccfos/nightingale/v6/pkg/strx"
 	"github.com/ccfos/nightingale/v6/pushgw/pconf"
 	"github.com/ccfos/nightingale/v6/pushgw/writer"
-	"github.com/ccfos/nightingale/v6/pkg/ginx"
 
 	"github.com/gin-gonic/gin"
 	"github.com/jinzhu/copier"
@@ -886,6 +886,7 @@ func (rt *Router) batchAlertRuleClone(c *gin.Context) {
 func (rt *Router) timezonesGet(c *gin.Context) {
 	// 返回常用时区列表（按时差去重，每个时差只保留一个代表性时区）
 	timezones := []string{
+		"Local",
 		"UTC",
 		"Asia/Shanghai",       // UTC+8 (代表 Asia/Hong_Kong, Asia/Singapore 等)
 		"Asia/Tokyo",          // UTC+9 (代表 Asia/Seoul 等)

doc/api/ai-agent.md

@@ -0,0 +1,251 @@
+# AI Agent API
+
+所有接口需要管理员权限（`auth` + `admin`）。
+
+## 数据结构
+
+### AIAgent
+
+| 字段 | 类型 | 必填 | 说明 |
+|------|------|------|------|
+| id | int64 | - | 主键，自增 |
+| name | string | 是 | Agent 名称 |
+| description | string | 否 | 描述 |
+| use_case | string | 否 | 用途场景，如 `chat` |
+| llm_config_id | int64 | 是 | 关联的 LLM 配置 ID |
+| skill_ids | int64[] | 否 | 关联的 Skill ID 列表 |
+| mcp_server_ids | int64[] | 否 | 关联的 MCP Server ID 列表 |
+| enabled | int | 否 | 是否启用，默认 1 |
+| created_at | int64 | - | 创建时间（Unix 时间戳） |
+| created_by | string | - | 创建人 |
+| updated_at | int64 | - | 更新时间（Unix 时间戳） |
+| updated_by | string | - | 更新人 |
+| llm_config | object | - | 运行时字段，关联的 LLM 配置对象（不存储） |
+
+---
+
+## 获取 Agent 列表
+
+```
+GET /api/n9e/ai-agents
+```
+
+### 响应
+
+```json
+{
+  "dat": [
+    {
+      "id": 1,
+      "name": "chat-agent",
+      "description": "AI 对话 Agent",
+      "use_case": "chat",
+      "llm_config_id": 1,
+      "skill_ids": [1, 2],
+      "mcp_server_ids": [1],
+      "enabled": 1,
+      "created_at": 1710000000,
+      "created_by": "admin",
+      "updated_at": 1710000000,
+      "updated_by": "admin"
+    }
+  ],
+  "err": ""
+}
+```
+
+---
+
+## 获取 Agent 详情
+
+```
+GET /api/n9e/ai-agent/:id
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | Agent ID |
+
+### 响应
+
+```json
+{
+  "dat": {
+    "id": 1,
+    "name": "chat-agent",
+    "description": "AI 对话 Agent",
+    "use_case": "chat",
+    "llm_config_id": 1,
+    "skill_ids": [1, 2],
+    "mcp_server_ids": [1],
+    "enabled": 1,
+    "created_at": 1710000000,
+    "created_by": "admin",
+    "updated_at": 1710000000,
+    "updated_by": "admin"
+  },
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` Agent 不存在
+
+---
+
+## 创建 Agent
+
+```
+POST /api/n9e/ai-agents
+```
+
+### 请求体
+
+```json
+{
+  "name": "chat-agent",
+  "description": "AI 对话 Agent",
+  "use_case": "chat",
+  "llm_config_id": 1,
+  "skill_ids": [1, 2],
+  "mcp_server_ids": [1],
+  "enabled": 1
+}
+```
+
+### 校验规则
+
+- `name` 必填
+- `llm_config_id` 必填，且大于 0
+
+### 响应
+
+```json
+{
+  "dat": 1,
+  "err": ""
+}
+```
+
+返回新创建的 Agent ID。
+
+---
+
+## 更新 Agent
+
+```
+PUT /api/n9e/ai-agent/:id
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | Agent ID |
+
+### 请求体
+
+```json
+{
+  "name": "chat-agent-v2",
+  "description": "更新后的描述",
+  "use_case": "chat",
+  "llm_config_id": 2,
+  "skill_ids": [1, 3],
+  "mcp_server_ids": [],
+  "enabled": 1
+}
+```
+
+### 校验规则
+
+同创建接口。
+
+### 响应
+
+```json
+{
+  "dat": "",
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` Agent 不存在
+
+---
+
+## 删除 Agent
+
+```
+DELETE /api/n9e/ai-agent/:id
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | Agent ID |
+
+### 响应
+
+```json
+{
+  "dat": "",
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` Agent 不存在
+
+---
+
+## 测试 Agent
+
+通过 Agent 关联的 LLM 配置发送测试请求，验证连通性。
+
+```
+POST /api/n9e/ai-agent/:id/test
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | Agent ID |
+
+### 响应
+
+```json
+{
+  "dat": {
+    "success": true,
+    "duration_ms": 1234
+  },
+  "err": ""
+}
+```
+
+失败时：
+
+```json
+{
+  "dat": {
+    "success": false,
+    "duration_ms": 5000,
+    "error": "HTTP 401: Unauthorized"
+  },
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` Agent 不存在
+- `400` 关联的 LLM 配置不存在

doc/api/ai-llm-config.md

@@ -0,0 +1,290 @@
+# AI LLM Config API
+
+所有接口需要管理员权限（`auth` + `admin`）。
+
+## 数据结构
+
+### AILLMConfig
+
+| 字段 | 类型 | 必填 | 说明 |
+|------|------|------|------|
+| id | int64 | - | 主键，自增 |
+| name | string | 是 | 配置名称 |
+| description | string | 否 | 描述 |
+| api_type | string | 是 | 提供商类型：`openai`、`claude`、`gemini` |
+| api_url | string | 是 | API 地址 |
+| api_key | string | 是 | API 密钥 |
+| model | string | 是 | 模型名称 |
+| extra_config | object | 否 | 高级配置，见 LLMExtraConfig |
+| enabled | int | 否 | 是否启用，默认 1 |
+| created_at | int64 | - | 创建时间（Unix 时间戳） |
+| created_by | string | - | 创建人 |
+| updated_at | int64 | - | 更新时间（Unix 时间戳） |
+| updated_by | string | - | 更新人 |
+
+### LLMExtraConfig
+
+| 字段 | 类型 | 说明 |
+|------|------|------|
+| timeout_seconds | int | 请求超时时间（秒），默认 30 |
+| skip_tls_verify | bool | 跳过 TLS 证书校验 |
+| proxy | string | HTTP 代理地址 |
+| custom_headers | map[string]string | 自定义请求头 |
+| custom_params | map[string]any | 自定义请求参数 |
+| temperature | float64 | 生成温度（可选） |
+| max_tokens | int | 最大输出 Token 数（可选） |
+| context_length | int | 上下文窗口大小（可选） |
+
+---
+
+## 获取 LLM 配置列表
+
+```
+GET /api/n9e/ai-llm-configs
+```
+
+### 响应
+
+```json
+{
+  "dat": [
+    {
+      "id": 1,
+      "name": "gpt-4o",
+      "description": "OpenAI GPT-4o",
+      "api_type": "openai",
+      "api_url": "https://api.openai.com",
+      "api_key": "sk-xxx",
+      "model": "gpt-4o",
+      "extra_config": {
+        "temperature": 0.7,
+        "max_tokens": 4096
+      },
+      "enabled": 1,
+      "created_at": 1710000000,
+      "created_by": "admin",
+      "updated_at": 1710000000,
+      "updated_by": "admin"
+    }
+  ],
+  "err": ""
+}
+```
+
+---
+
+## 获取 LLM 配置详情
+
+```
+GET /api/n9e/ai-llm-config/:id
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | LLM 配置 ID |
+
+### 响应
+
+```json
+{
+  "dat": {
+    "id": 1,
+    "name": "gpt-4o",
+    "description": "OpenAI GPT-4o",
+    "api_type": "openai",
+    "api_url": "https://api.openai.com",
+    "api_key": "sk-xxx",
+    "model": "gpt-4o",
+    "extra_config": {
+      "temperature": 0.7,
+      "max_tokens": 4096
+    },
+    "enabled": 1,
+    "created_at": 1710000000,
+    "created_by": "admin",
+    "updated_at": 1710000000,
+    "updated_by": "admin"
+  },
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` LLM 配置不存在
+
+---
+
+## 创建 LLM 配置
+
+```
+POST /api/n9e/ai-llm-configs
+```
+
+### 请求体
+
+```json
+{
+  "name": "gpt-4o",
+  "description": "OpenAI GPT-4o",
+  "api_type": "openai",
+  "api_url": "https://api.openai.com",
+  "api_key": "sk-xxx",
+  "model": "gpt-4o",
+  "extra_config": {
+    "timeout_seconds": 60,
+    "temperature": 0.7,
+    "max_tokens": 4096,
+    "custom_headers": {
+      "X-Custom": "value"
+    }
+  },
+  "enabled": 1
+}
+```
+
+### 校验规则
+
+- `name`、`api_type`、`api_url`、`api_key`、`model` 均为必填
+
+### 响应
+
+```json
+{
+  "dat": 1,
+  "err": ""
+}
+```
+
+返回新创建的配置 ID。
+
+---
+
+## 更新 LLM 配置
+
+```
+PUT /api/n9e/ai-llm-config/:id
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | LLM 配置 ID |
+
+### 请求体
+
+同创建接口。**注意：如果 `api_key` 为空，则保留原值不更新。**
+
+### 校验规则
+
+同创建接口。
+
+### 响应
+
+```json
+{
+  "dat": "",
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` LLM 配置不存在
+
+---
+
+## 删除 LLM 配置
+
+```
+DELETE /api/n9e/ai-llm-config/:id
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | LLM 配置 ID |
+
+### 响应
+
+```json
+{
+  "dat": "",
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` LLM 配置不存在
+
+---
+
+## 测试 LLM 连接
+
+无需先创建配置，直接传入连接参数进行连通性测试。
+
+```
+POST /api/n9e/ai-llm-config/test
+```
+
+### 请求体
+
+```json
+{
+  "api_type": "openai",
+  "api_url": "https://api.openai.com",
+  "api_key": "sk-xxx",
+  "model": "gpt-4o",
+  "extra_config": {
+    "timeout_seconds": 30,
+    "skip_tls_verify": false,
+    "proxy": "",
+    "custom_headers": {}
+  }
+}
+```
+
+### 校验规则
+
+- `api_type`、`api_url`、`api_key`、`model` 均为必填
+
+### 测试行为
+
+根据 `api_type` 向对应的 API 发送一个最小请求（"Hi"，max_tokens=5）：
+
+| api_type | 请求地址 | 认证方式 |
+|----------|---------|---------|
+| openai | `{api_url}/chat/completions` | `Authorization: Bearer {api_key}` |
+| claude | `{api_url}/v1/messages` | `x-api-key: {api_key}` |
+| gemini | `{api_url}/v1beta/models/{model}:generateContent?key={api_key}` | URL 参数 |
+
+### 响应
+
+成功：
+
+```json
+{
+  "dat": {
+    "success": true,
+    "duration_ms": 856
+  },
+  "err": ""
+}
+```
+
+失败：
+
+```json
+{
+  "dat": {
+    "success": false,
+    "duration_ms": 5000
+  },
+  "err": "HTTP 401: {\"error\": \"invalid api key\"}"
+}
+```

doc/api/ai-mcp-server.md

@@ -0,0 +1,292 @@
+# MCP Server API
+
+所有接口需要管理员权限（`auth` + `admin`）。
+
+## 数据结构
+
+### MCPServer
+
+| 字段 | 类型 | 必填 | 说明 |
+|------|------|------|------|
+| id | int64 | - | 主键，自增 |
+| name | string | 是 | 名称 |
+| url | string | 是 | MCP Server 地址 |
+| headers | map[string]string | 否 | 自定义 HTTP 请求头，用于认证等 |
+| description | string | 否 | 描述 |
+| enabled | int | 否 | 是否启用，默认 1 |
+| created_at | int64 | - | 创建时间（Unix 时间戳） |
+| created_by | string | - | 创建人 |
+| updated_at | int64 | - | 更新时间（Unix 时间戳） |
+| updated_by | string | - | 更新人 |
+
+---
+
+## 获取 MCP Server 列表
+
+```
+GET /api/n9e/mcp-servers
+```
+
+### 响应
+
+```json
+{
+  "dat": [
+    {
+      "id": 1,
+      "name": "my-mcp-server",
+      "url": "https://mcp.example.com/sse",
+      "headers": {
+        "Authorization": "Bearer xxx"
+      },
+      "description": "示例 MCP Server",
+      "enabled": 1,
+      "created_at": 1710000000,
+      "created_by": "admin",
+      "updated_at": 1710000000,
+      "updated_by": "admin"
+    }
+  ],
+  "err": ""
+}
+```
+
+---
+
+## 获取 MCP Server 详情
+
+```
+GET /api/n9e/mcp-server/:id
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | MCP Server ID |
+
+### 响应
+
+```json
+{
+  "dat": {
+    "id": 1,
+    "name": "my-mcp-server",
+    "url": "https://mcp.example.com/sse",
+    "headers": {
+      "Authorization": "Bearer xxx"
+    },
+    "description": "示例 MCP Server",
+    "enabled": 1,
+    "created_at": 1710000000,
+    "created_by": "admin",
+    "updated_at": 1710000000,
+    "updated_by": "admin"
+  },
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` MCP Server 不存在
+
+---
+
+## 创建 MCP Server
+
+```
+POST /api/n9e/mcp-servers
+```
+
+### 请求体
+
+```json
+{
+  "name": "my-mcp-server",
+  "url": "https://mcp.example.com/sse",
+  "headers": {
+    "Authorization": "Bearer xxx"
+  },
+  "description": "示例 MCP Server",
+  "enabled": 1
+}
+```
+
+### 校验规则
+
+- `name` 必填（自动 trim）
+- `url` 必填（自动 trim）
+
+### 响应
+
+```json
+{
+  "dat": 1,
+  "err": ""
+}
+```
+
+返回新创建的 MCP Server ID。
+
+---
+
+## 更新 MCP Server
+
+```
+PUT /api/n9e/mcp-server/:id
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | MCP Server ID |
+
+### 请求体
+
+同创建接口。
+
+### 可更新字段
+
+`name`、`url`、`headers`、`description`、`enabled`。
+
+### 响应
+
+```json
+{
+  "dat": "",
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` MCP Server 不存在
+
+---
+
+## 删除 MCP Server
+
+```
+DELETE /api/n9e/mcp-server/:id
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | MCP Server ID |
+
+### 响应
+
+```json
+{
+  "dat": "",
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` MCP Server 不存在
+
+---
+
+## 测试 MCP Server 连接
+
+无需先创建，直接传入连接参数进行连通性测试。通过 MCP 协议初始化握手并获取工具列表来验证连通性。
+
+```
+POST /api/n9e/mcp-server/test
+```
+
+### 请求体
+
+```json
+{
+  "url": "https://mcp.example.com/sse",
+  "headers": {
+    "Authorization": "Bearer xxx"
+  }
+}
+```
+
+### 校验规则
+
+- `url` 必填
+
+### 测试行为
+
+1. 发送 `initialize` 请求（协议版本 `2024-11-05`）
+2. 发送 `notifications/initialized` 通知
+3. 发送 `tools/list` 请求获取工具列表
+
+支持 JSON 和 SSE（`text/event-stream`）两种响应格式。
+
+### 响应
+
+成功：
+
+```json
+{
+  "dat": {
+    "success": true,
+    "duration_ms": 320,
+    "tool_count": 5
+  },
+  "err": ""
+}
+```
+
+失败：
+
+```json
+{
+  "dat": {
+    "success": false,
+    "duration_ms": 5000,
+    "tool_count": 0,
+    "error": "initialize: HTTP 403: Forbidden"
+  },
+  "err": ""
+}
+```
+
+---
+
+## 获取 MCP Server 工具列表
+
+获取已创建的 MCP Server 提供的工具列表。
+
+```
+GET /api/n9e/mcp-server/:id/tools
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | MCP Server ID |
+
+### 响应
+
+```json
+{
+  "dat": [
+    {
+      "name": "query_database",
+      "description": "Execute a SQL query against the database"
+    },
+    {
+      "name": "search_logs",
+      "description": "Search through application logs"
+    }
+  ],
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` MCP Server 不存在

doc/api/ai-skill.md

@@ -0,0 +1,409 @@
+# AI Skill API
+
+所有接口需要管理员权限（`auth` + `admin`）。
+
+## 数据结构
+
+### AISkill
+
+| 字段 | 类型 | 必填 | 说明 |
+|------|------|------|------|
+| id | int64 | - | 主键，自增 |
+| name | string | 是 | Skill 名称 |
+| description | string | 否 | 描述，建议说明用途和触发场景 |
+| instructions | string | 是 | 提示词指令，支持 Markdown |
+| license | string | 否 | 许可证，如 `MIT`、`Apache-2.0` |
+| compatibility | string | 否 | 兼容性说明，如环境依赖、网络需求等 |
+| metadata | map[string]string | 否 | 扩展元数据，如 `{"author": "org", "version": "1.0"}` |
+| allowed_tools | string | 否 | 预授权工具列表，空格分隔，如 `Bash(git:*) Read` |
+| enabled | int | 否 | 是否启用，默认 1 |
+| created_at | int64 | - | 创建时间（Unix 时间戳） |
+| created_by | string | - | 创建人 |
+| updated_at | int64 | - | 更新时间（Unix 时间戳） |
+| updated_by | string | - | 更新人 |
+| files | AISkillFile[] | - | 关联的资源文件列表（仅详情接口返回，不含 content） |
+
+> `license`、`compatibility`、`metadata`、`allowed_tools` 字段参考 [Agent Skills Specification](https://agentskills.io/specification)。
+
+### AISkillFile
+
+| 字段 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | 主键，自增 |
+| skill_id | int64 | 关联的 Skill ID |
+| name | string | 文件名 |
+| content | string | 文件内容（仅文件详情接口返回） |
+| size | int64 | 文件大小（字节），创建时自动计算 |
+| created_at | int64 | 创建时间（Unix 时间戳） |
+| created_by | string | 创建人 |
+
+---
+
+## 获取 Skill 列表
+
+```
+GET /api/n9e/ai-skills
+```
+
+### 查询参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| search | string | 可选，按 name 或 description 模糊搜索 |
+
+### 响应
+
+```json
+{
+  "dat": [
+    {
+      "id": 1,
+      "name": "query-generator",
+      "description": "生成 PromQL/SQL 查询语句",
+      "instructions": "# Query Generator\n...",
+      "license": "Apache-2.0",
+      "compatibility": "Requires network access",
+      "metadata": {
+        "author": "nightingale",
+        "version": "1.0"
+      },
+      "allowed_tools": "Bash(git:*) Read",
+      "enabled": 1,
+      "created_at": 1710000000,
+      "created_by": "admin",
+      "updated_at": 1710000000,
+      "updated_by": "admin"
+    }
+  ],
+  "err": ""
+}
+```
+
+> 列表接口不返回 `files` 字段。
+
+---
+
+## 获取 Skill 详情
+
+```
+GET /api/n9e/ai-skill/:id
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | Skill ID |
+
+### 响应
+
+返回 Skill 完整信息，并包含关联的资源文件列表（不含文件 content）。
+
+```json
+{
+  "dat": {
+    "id": 1,
+    "name": "query-generator",
+    "description": "生成 PromQL/SQL 查询语句",
+    "instructions": "# Query Generator\n...",
+    "license": "Apache-2.0",
+    "compatibility": "Requires network access",
+    "metadata": {
+      "author": "nightingale",
+      "version": "1.0"
+    },
+    "allowed_tools": "Bash(git:*) Read",
+    "enabled": 1,
+    "created_at": 1710000000,
+    "created_by": "admin",
+    "updated_at": 1710000000,
+    "updated_by": "admin",
+    "files": [
+      {
+        "id": 10,
+        "skill_id": 1,
+        "name": "reference.md",
+        "size": 2048,
+        "created_at": 1710000000,
+        "created_by": "admin"
+      }
+    ]
+  },
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` Skill 不存在
+
+---
+
+## 创建 Skill
+
+```
+POST /api/n9e/ai-skills
+```
+
+### 请求体
+
+```json
+{
+  "name": "query-generator",
+  "description": "生成 PromQL/SQL 查询语句",
+  "instructions": "# Query Generator\n根据用户输入生成查询语句...",
+  "license": "Apache-2.0",
+  "compatibility": "Requires network access",
+  "metadata": {
+    "author": "nightingale",
+    "version": "1.0"
+  },
+  "allowed_tools": "Bash(git:*) Read",
+  "enabled": 1
+}
+```
+
+### 校验规则
+
+- `name` 必填（自动 trim）
+- `instructions` 必填（自动 trim）
+
+### 响应
+
+```json
+{
+  "dat": 1,
+  "err": ""
+}
+```
+
+返回新创建的 Skill ID。
+
+---
+
+## 更新 Skill
+
+```
+PUT /api/n9e/ai-skill/:id
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | Skill ID |
+
+### 请求体
+
+同创建接口。
+
+### 可更新字段
+
+`name`、`description`、`instructions`、`license`、`compatibility`、`metadata`、`allowed_tools`、`enabled`。
+
+### 响应
+
+```json
+{
+  "dat": "",
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` Skill 不存在
+
+---
+
+## 删除 Skill
+
+删除 Skill 时会级联删除关联的所有资源文件。
+
+```
+DELETE /api/n9e/ai-skill/:id
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | Skill ID |
+
+### 响应
+
+```json
+{
+  "dat": "",
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` Skill 不存在
+
+---
+
+## 导入 Skill
+
+从 `.md` 文件导入 Skill，支持 YAML frontmatter 格式。
+
+```
+POST /api/n9e/ai-skills/import
+```
+
+### 请求格式
+
+`multipart/form-data`
+
+| 字段 | 类型 | 说明 |
+|------|------|------|
+| file | file | `.md` 文件 |
+
+### 文件格式
+
+支持标准的 YAML frontmatter + Markdown body：
+
+```markdown
+---
+name: my-skill
+description: 技能描述
+license: MIT
+compatibility: Requires git, docker
+metadata:
+  author: my-org
+  version: "1.0"
+allowed-tools: Bash(git:*) Read
+---
+
+# Skill 指令内容
+
+这里是 instructions 部分...
+```
+
+- 如果文件包含有效的 frontmatter，则从中提取 `name`、`description`、`license`、`compatibility`、`metadata`、`allowed-tools`
+- 如果没有 frontmatter，则以文件名作为 `name`，全部内容作为 `instructions`
+
+### 响应
+
+```json
+{
+  "dat": 1,
+  "err": ""
+}
+```
+
+返回新创建的 Skill ID。
+
+### 错误
+
+- `400` 仅支持 `.md` 文件
+
+---
+
+## 上传 Skill 资源文件
+
+```
+POST /api/n9e/ai-skill/:id/files
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| id | int64 | Skill ID |
+
+### 请求格式
+
+`multipart/form-data`
+
+| 字段 | 类型 | 说明 |
+|------|------|------|
+| file | file | 资源文件 |
+
+### 限制
+
+- 允许的文件类型：`.md`、`.txt`、`.json`、`.yaml`、`.yml`、`.csv`
+- 单文件最大 2MB
+- 每个 Skill 最多 20 个资源文件
+
+### 响应
+
+```json
+{
+  "dat": 10,
+  "err": ""
+}
+```
+
+返回新创建的文件 ID。
+
+### 错误
+
+- `404` Skill 不存在
+- `400` 文件类型不支持 / 文件超过 2MB / 文件数量超过 20
+
+---
+
+## 获取资源文件详情
+
+获取单个资源文件的完整内容。
+
+```
+GET /api/n9e/ai-skill-file/:fileId
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| fileId | int64 | 文件 ID |
+
+### 响应
+
+```json
+{
+  "dat": {
+    "id": 10,
+    "skill_id": 1,
+    "name": "reference.md",
+    "content": "# Reference\n文件完整内容...",
+    "size": 2048,
+    "created_at": 1710000000,
+    "created_by": "admin"
+  },
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` 文件不存在
+
+---
+
+## 删除资源文件
+
+```
+DELETE /api/n9e/ai-skill-file/:fileId
+```
+
+### 路径参数
+
+| 参数 | 类型 | 说明 |
+|------|------|------|
+| fileId | int64 | 文件 ID |
+
+### 响应
+
+```json
+{
+  "dat": "",
+  "err": ""
+}
+```
+
+### 错误
+
+- `404` 文件不存在

dskit/doris/doris.go

@@ -201,7 +201,7 @@ func (d *Doris) NewWriteConn(ctx context.Context, database string) (*sql.DB, err
 func (d *Doris) createTimeoutContext(ctx context.Context) (context.Context, context.CancelFunc) {
 	timeout := d.Timeout
 	if timeout == 0 {
-		timeout = 60
+		timeout = 60000
 	}
 	return context.WithTimeout(ctx, time.Duration(timeout)*time.Millisecond)
 }

etc/skills/ops-troubleshooting/SKILL.md

@@ -0,0 +1,691 @@
+---
+name: ops-troubleshooting
+description: This skill should be used when the user asks to "troubleshoot", "diagnose", "debug alert", "investigate incident", "故障定位", "告警排查", "问题诊断", "排障", "查告警", "分析告警", "根因分析", "灭火图", "北极星", "tracing", "链路追踪", "日志分析", or discusses monitoring/alerting/observability issues in Flashcat platform (n9e-plus, fc-insight).
+version: 5.0.0
+---
+
+# Flashcat 故障定位专家 (SRE Troubleshooting Expert)
+
+你是一位拥有 10 年以上经验的资深 SRE，专门负责使用 Flashcat 可观测性平台进行故障定位和根因分析。
+
+---
+
+## 核心原则
+
+1. **证据链驱动**: 每个推断都要有页面数据支撑
+2. **按需查询**: 根据分析需要逐步查询，不盲目拉取所有数据
+3. **最小权限**: 只查询必要的数据，避免敏感信息泄露
+4. **时间线优先**: 关注故障发生的时序关系，寻找根因
+5. **定位直接原因**: 不追求 100% 覆盖根因，聚焦于定位直接原因和止损依据
+6. **聚焦故障的时间**：如果跳到其他页面，没找到线索，重新回到故障链接的页面分析
+
+---
+
+## 数据获取方式：Playwright 访问页面
+
+使用 Playwright MCP 工具直接访问 Flashcat Web 页面，可视化查看数据。
+
+**适用场景**：
+
+- 查看灭火图全局视图，识别飘红/飘黄卡片
+- 浏览北极星业务线指标
+- 查看仪表盘和图表趋势
+- 分析 Trace 瀑布图
+- 需要截图作为证据
+- 探索性分析，不确定需要哪些数据
+
+**使用方法**：
+
+```
+# 1. 启动浏览器（如果尚未启动）
+mcp__playwright__browser_navigate to the target URL
+
+# 2. 常用页面导航
+- 灭火图首页: {BASE_URL}/firemap
+- 灭火图卡片详情: {BASE_URL}/firemap/function/{card_id}
+- 北极星: {BASE_URL}/polaris
+- 链路追踪: {BASE_URL}/trace
+- 活跃告警: {BASE_URL}/alert-cur-events
+- 事件墙: {BASE_URL}/events
+
+# 3. 截图保存证据
+mcp__playwright__browser_screenshot
+
+# 4. 页面交互
+- 点击卡片查看详情
+- 切换时间范围
+- 展开/折叠分组
+- 下钻到关联数据
+```
+
+**Playwright 操作示例**：
+
+```
+# 访问灭火图，查看异常卡片
+1. browser_navigate: https://demo.flashcat.cloud
+2. browser_screenshot: 保存当前状态
+3. browser_click: 点击飘红卡片进入详情
+4. browser_screenshot: 保存卡片详情
+5. browser_click: 点击"关联链路"下钻到 Trace
+```
+
+---
+
+## 灭火图核心理念
+
+灭火图是 Flashcat 平台的核心创新，理解其设计理念对故障排查至关重要。
+
+### 定义
+
+> 灭火图：以服务/模块/组件/基础设施等为维度，以聚合视角实时度量特定维度的可用性。
+
+### 推荐层级架构（C端服务）
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│  Layer 1: 接口层                                             │
+│  - 核心接口的成功率、延迟、流量                               │
+│  - 如：下单接口、支付接口、查询接口                           │
+└─────────────────────────────────────────────────────────────┘
+                              ↓
+┌─────────────────────────────────────────────────────────────┐
+│  Layer 2: 微服务层                                           │
+│  - 各微服务的 RED 指标                                       │
+│  - 如：订单服务、用户服务、库存服务                           │
+└─────────────────────────────────────────────────────────────┘
+                              ↓
+┌─────────────────────────────────────────────────────────────┐
+│  Layer 3: 组件层                                             │
+│  - 中间件、缓存、消息队列等                                   │
+│  - 如：Redis、MySQL、Kafka                                   │
+└─────────────────────────────────────────────────────────────┘
+                              ↓
+┌─────────────────────────────────────────────────────────────┐
+│  Layer 4: 基础设施层                                         │
+│  - 主机、容器、网络                                          │
+│  - 如：CPU、内存、磁盘、网络带宽                              │
+└─────────────────────────────────────────────────────────────┘
+```
+
+### 核心指标维度 (USE/RED)
+
+灭火图卡片通常基于以下指标自动设定阈值：
+
+- **时延 (Duration)**: P50/P90/P99 延迟
+- **流量 (Rate)**: 请求量/QPS
+- **错误 (Error)**: 错误率/成功率
+- **饱和度 (Saturation)**: 资源使用率
+
+---
+
+## 排查流程决策树
+
+```
+┌─────────────────────────────────────────────────────────────┐
+│                    故障排查入口                              │
+└─────────────────────────────────────────────────────────────┘
+                              │
+                              ▼
+┌─────────────────────────────────────────────────────────────┐
+│  Step 1: 明确问题                                           │
+│  用户描述了什么？                                            │
+├─────────────────────────────────────────────────────────────┤
+│  A. 具体告警 ID → 走【告警分析流程】                          │
+│  B. 服务/接口名 → 走【服务分析流程】                          │
+│  C. 业务异常   → 走【灭火图分析流程】                         │
+│  D. 时间段异常 → 走【事件墙分析流程】                         │
+│  E. 不确定    → 走【全局概览流程】                           │
+└─────────────────────────────────────────────────────────────┘
+```
+
+### 常见问题快速定位表
+
+
+| 问题类型     | 首选流程 | 访问页面                  |
+| -------- | ---- | --------------------- |
+| 收到告警通知   | 流程 A | /alert-cur-events     |
+| 服务响应慢    | 流程 B | /trace                |
+| 接口报错     | 流程 B | /trace, /log/explorer |
+| 业务指标异常   | 流程 C | /firemap, /polaris    |
+| 灭火图飘红/飘黄 | 流程 C | /firemap              |
+| 发布后出问题   | 流程 D | /events               |
+| 不确定哪里有问题 | 流程 E | /firemap (全局概览)       |
+
+
+---
+
+## 平台架构
+
+Flashcat 可观测性平台包含以下核心组件：
+
+
+| 组件             | 说明                          | 默认端口  |
+| -------------- | --------------------------- | ----- |
+| **n9e-plus**   | 夜莺增强版，告警管理、指标查询、日志查询        | 17000 |
+| **fc-insight** | 可观测分析平台，灭火图、北极星、Tracing、事件墙 | 8080  |
+
+
+## 核心能力
+
+### n9e-plus 能力
+
+1. **告警事件管理** - 当前/历史告警查询、确认、静默
+2. **指标数据查询** - Prometheus、ES、SLS、CLS、TLS 等多数据源
+3. **日志分析** - 多种日志源的统一查询
+4. **采集配置** - 采集任务管理、网络设备监控
+
+### fc-insight 能力
+
+1. **灭火图 (Firemap)** - 业务健康度可视化、SLO 管理
+2. **北极星 (Northstar)** - 业务指标管理、异常检测、告警规则
+3. **事件墙 (Event Wall)** - 告警/变更事件聚合、关联分析
+4. **Tracing** - 分布式链路追踪、服务拓扑
+5. **日志分析** - 多维分析、特征提取
+
+---
+
+## 页面导航与功能说明
+
+### 北极星 (Polaris) - `/polaris`
+
+业务指标的核心观测平台，用于监控业务健康状态。
+
+
+| 页面路径                     | 功能说明                |
+| ------------------------ | ------------------- |
+| `/polaris`               | 北极星首页，业务线指标总览       |
+| `/polaris/indicator`     | 指标池，所有北极星指标的管理列表    |
+| `/polaris/statuspage`    | 状态概览页，业务状态汇总展示      |
+| `/polaris/stability`     | 稳定性看板，SLO/SLA 可用性统计 |
+| `/polaris/screen`        | 大屏管理，北极星大屏配置        |
+| `/polaris/{business_id}` | 业务线详情页，查看具体业务线的指标   |
+
+
+**关键功能**:
+
+- 业务线管理：按业务维度组织指标
+- 指标异常检测：动态阈值、同环比、智能基线
+- 告警规则配置：基于指标配置告警策略
+- 稳定性看板：SLO 目标设定和达成率统计
+
+### 灭火图 (Firemap) - `/firemap`
+
+层级化的业务健康度可视化，支持从全局到细节的下钻分析。
+
+
+| 页面路径                         | 功能说明               |
+| ---------------------------- | ------------------ |
+| `/firemap`                   | 灭火图首页，系统级健康度总览     |
+| `/firemap/card-pool`         | 卡片池，所有卡片的管理列表      |
+| `/firemap/statuspage`        | 状态概览页，全局状态汇总       |
+| `/firemap/slo`               | SLO报表，卡片级别的SLO统计   |
+| `/firemap/screens`           | 拓扑画布，自定义拓扑图配置      |
+| `/firemap/{level}/{card_id}` | 卡片详情页，查看具体卡片的指标和下钻 |
+
+
+**层级结构**:
+
+- **首页层级**: 系统级卡片，如"订单系统"、"支付系统"
+- **分组**: 系统内的子模块分组
+- **详情卡片**: 具体的观测对象（接口、实例、服务等）
+
+**关键功能**:
+
+- 飘红/飘黄告警：视觉化异常标识
+- 卡片快照：查看卡片在特定时间点的状态
+- 下钻关联：从卡片跳转到日志、Trace、仪表盘
+- 规则管理：自动创建和管理卡片的规则
+
+### 事件墙 (Event Wall) - `/events`
+
+告警和变更事件的聚合时间线，用于故障时间线分析。
+
+
+| 页面路径         | 功能说明              |
+| ------------ | ----------------- |
+| `/events`    | 事件墙首页（新版），事件时间线展示 |
+| `/events-v1` | 事件墙（旧版）           |
+
+
+**事件类型**:
+
+- **告警事件**: 来自各告警源的告警
+- **变更事件**: 发布、配置变更、数据库变更等
+- **Incident**: 故障事件聚合
+
+**关键功能**:
+
+- 时间线视图：按时间顺序展示事件
+- 事件聚合：相似事件自动聚合
+- 关联分析：分析告警与变更的时间关系
+
+### 链路分析 (Tracing) - `/trace`
+
+分布式链路追踪，用于服务调用链分析。
+
+
+| 页面路径              | 功能说明             |
+| ----------------- | ---------------- |
+| `/trace`          | 链路检索，按条件查询 Trace |
+| `/trace/topology` | 拓扑分析，服务依赖拓扑图     |
+| `/trace/services` | 应用列表，服务 RED 指标概览 |
+| `/trace/database` | 数据库分析，DB 调用分析    |
+
+
+**关键功能**:
+
+- Trace 检索：按 TraceID、服务、时间范围查询
+- 瀑布图：展示 Span 调用链
+- 服务拓扑：可视化服务依赖关系
+- RED 指标：Rate（请求量）、Error（错误率）、Duration（延迟）
+
+### 日志分析 (Multi-Dimension) - `/multi-dimension`
+
+日志的多维度特征分析，支持异常特征识别。
+
+
+| 页面路径                  | 功能说明           |
+| --------------------- | -------------- |
+| `/log/explorer`       | 日志检索，原始日志查询    |
+| `/multi-dimension`    | 特征分析，日志多维度统计   |
+| `/log-extraction`     | 日志转换，日志字段提取配置  |
+| `/tracing-multiquery` | 联合查询，跨日志源的串联查询 |
+
+
+**关键功能**:
+
+- 主题管理：日志源的字段配置
+- 特征分析：按维度统计日志特征
+- Bubble Up：异常根因分析
+- 联合查询：跨多个日志源的 Trace 串联
+
+### 告警管理 - `/alert-*`
+
+告警规则和事件的管理。
+
+
+| 页面路径                | 功能说明           |
+| ------------------- | -------------- |
+| `/alert-rules`      | 告警规则列表         |
+| `/alert-mutes`      | 屏蔽规则列表         |
+| `/alert-subscribes` | 订阅规则列表         |
+| `/alert-cur-events` | 活跃告警（当前未恢复的告警） |
+| `/alert-his-events` | 历史告警（所有告警记录）   |
+
+
+### 仪表盘 - `/dashboards`
+
+监控仪表盘的管理和展示。
+
+
+| 页面路径                | 功能说明   |
+| ------------------- | ------ |
+| `/dashboards`       | 仪表盘列表  |
+| `/dashboards/{id}`  | 仪表盘详情页 |
+| `/reports`          | 报表管理   |
+| `/template/screens` | 模板大屏管理 |
+
+
+### 基础设施 - `/targets`, `/collects`
+
+监控对象和采集配置管理。
+
+
+| 页面路径                          | 功能说明        |
+| ----------------------------- | ----------- |
+| `/targets`                    | 机器列表，监控对象管理 |
+| `/collects`                   | 数据采集，采集任务配置 |
+| `/settings/source/timeseries` | 数据集成，数据源管理  |
+
+
+### 容器平台 - `/kubernetes`
+
+
+| 页面路径          | 功能说明            |
+| ------------- | --------------- |
+| `/kubernetes` | Kubernetes 集群监控 |
+
+
+### 空间管理 - `/space`
+
+
+| 页面路径     | 功能说明            |
+| -------- | --------------- |
+| `/space` | 空间管理，工作空间的创建和配置 |
+
+
+**空间概念**: 空间是 Flashcat 的多租户隔离单位，不同空间的数据相互隔离。灭火图、北极星等模块都基于空间组织数据。
+
+---
+
+## 故障定位流程详解
+
+### 流程 A: 告警分析流程
+
+**入口条件**: 用户提供了具体的告警 ID 或告警名称
+
+**操作步骤**:
+
+1. 访问 `/alert-cur-events` 或 `/alert-his-events` 查看告警详情
+2. 查看告警规则、触发条件、标签
+3. 点击关联告警查看同一 target 的其他告警
+4. 查看告警相关的指标曲线
+5. 检查是否有关联的灭火图卡片或变更事件
+
+### 流程 B: 服务分析流程
+
+**入口条件**: 用户提到具体服务名或接口
+
+**操作步骤**:
+
+1. 访问 `/trace/services` 查看服务 RED 指标
+2. 访问 `/trace` 搜索错误或慢 Trace
+3. 访问 `/trace/topology` 查看上下游依赖
+4. 访问 `/log/explorer` 查询 ERROR 级别日志
+
+### 流程 C: 灭火图分析流程
+
+**入口条件**: 用户提到业务健康度、灭火图、卡片飘红
+
+**操作步骤**:
+
+1. 访问 `/firemap` 查看全局快照，找出异常卡片
+2. 点击异常卡片查看详情和历史曲线
+3. 检查卡片配置的告警规则
+4. 点击下钻关联到日志/Trace
+
+### 流程 D: 事件墙分析流程
+
+**入口条件**: 用户提到某个时间段有问题，需要分析时间线
+
+**操作步骤**:
+
+1. 访问 `/events` 查看时间段内的事件时间线
+2. 按服务或严重程度筛选事件
+3. 分析变更时间与告警时间的关系
+
+### 流程 E: 全局概览流程
+
+**入口条件**: 用户不确定问题在哪，需要先了解整体状况
+
+**操作步骤**:
+
+1. 访问 `/alert-cur-events` 了解当前活跃告警
+2. 访问 `/firemap` 查看灭火图健康状态
+3. 访问 `/events` 查看最近变更事件
+
+---
+
+## 排查示例
+
+以下是使用 Playwright 进行故障排查的完整示例。
+
+### 场景
+
+用户反馈：灭火图上有卡片飘红，需要定位问题原因。
+
+### Step 1: 访问灭火图首页
+
+**观察要点**：
+
+- 识别飘红（critical）和飘黄（warning）的卡片
+- 记录异常卡片的名称和位置
+- 截图保存当前状态
+
+### Step 2: 点击异常卡片进入详情
+
+**观察要点**：
+
+- 查看卡片的具体指标值（成功率、延迟等）
+- 查看历史趋势曲线
+- 识别异常开始的时间点
+
+### Step 3: 下钻到子卡片
+
+**按照灭火图层级逐层下钻**：
+
+```
+系统级卡片 → 子系统分组 → 具体模块卡片
+```
+
+每层都观察哪些卡片是红色/黄色的，逐步缩小范围。
+
+### Step 4: 查看关联数据
+
+在卡片详情页，点击关联的数据源：
+
+```
+# 查看关联 Trace
+# 查看关联日志
+# 查看关联仪表盘
+```
+
+### Step 5: 分析下钻
+
+---
+
+### 输出分析结论
+
+**直接原因**: 
+
+**证据链**：
+
+## **止损建议**:
+
+## Playwright 常用操作速查
+
+
+| 操作   | 工具调用                          | 说明        |
+| ---- | ----------------------------- | --------- |
+| 打开页面 | `browser_navigate(url)`       | 导航到指定 URL |
+| 截图   | `browser_screenshot()`        | 保存当前页面截图  |
+| 点击   | `browser_click(element)`      | 点击页面元素    |
+| 输入   | `browser_type(element, text)` | 在输入框中输入文本 |
+| 等待   | `browser_wait(selector)`      | 等待元素出现    |
+| 获取文本 | `browser_get_text(selector)`  | 获取元素文本内容  |
+
+
+---
+
+## 分析输出模板
+
+每次排查完成后，按以下格式输出分析报告：
+
+```markdown
+## 故障分析报告
+
+### 1. 问题概述
+- **问题描述**: [用户原始描述]
+- **分析时间**: [开始时间 - 结束时间]
+- **影响范围**: [服务/业务/实例]
+
+### 2. 关键发现
+
+#### 2.1 特征分析
+- **时间区间**: [开始时间] → [结束时间]
+- **筛选条件**: [url.path=/xxx, http.response.status_code=xxx 等]
+- **响应码分布**: [状态码] 共 [数量] 次
+- **关注维度**: [维度名称]
+- **异常 Trace 数**: [数量] 条
+
+#### 2.3 链路分析 (如有)
+- **错误 Trace**: [数量]
+- **错误类型**: [类型]
+- **关键 Span**: [Span 名称]
+
+#### 2.4 日志分析 (如有)
+- **ERROR 日志数量**: [数量]
+- **关键错误信息**: [错误摘要]
+
+#### 2.5 变更关联
+- **可疑变更**: [有/无]
+- **变更内容**: [描述]
+- **变更时间**: [时间] (告警前 [X] 分钟)
+
+### 3. 根因分析
+- **根本原因**: [原因描述]
+- **证据链**:
+  1. [证据1]
+  2. [证据2]
+  3. [证据3]
+
+### 4. 建议措施
+- **立即行动**: [紧急措施]
+- **后续跟进**: [长期改进]
+```
+
+---
+
+## 安全注意事项
+
+1. **最小查询**: 只查询必要数据，避免拉取全量数据
+2. **输出脱敏**: 报告中不要包含密码、密钥等敏感信息
+3. **权限控制**: 确保使用的账号有适当的权限范围
+
+---
+
+## 实战案例：订单提交服务成功率归零排查
+
+以下是一个完整的实战排查案例，展示从问题发现到根因定位的全过程。
+
+### 案例背景
+
+- **问题现象**: 电商系统灭火图卡片飘红，订单提交功能异常
+- **问题链接**: `https://demo.flashcat.cloud/firemap/function/706?end=1768480361&spaceId=848859344386&start=1768437161&time=1768475160`
+- **平台**: Flashcat Demo 环境
+
+### Step 1: 登录平台，访问问题页面
+
+```
+操作: browser_navigate → 登录页面 → 填写凭据 → 点击登录
+观察: 成功进入北极星首页
+```
+
+### Step 2: 导航到异常灭火图卡片
+
+```
+操作: browser_navigate → 异常链接
+观察:
+- 系统: 电商
+- 子系统: 订单子系统 (1/10 异常)
+- 异常卡片: "订单提交" - 成功率 0% ❌
+- 其他卡片状态正常 (99%~100%)
+```
+
+**关键发现**: 只有"订单提交"功能异常，其他功能正常，说明问题范围收敛到该接口。
+
+### Step 3: 进入卡片详情，查看指标趋势
+
+```
+操作: browser_click → 订单提交卡片
+观察:
+- 标签: demo=demo-01
+- 成功率图表: 有异常标记 (exclamation-circle)
+- 响应码统计: 有异常标记
+- 延时指标: 正常
+```
+
+**关键发现**: 成功率和响应码都异常，但延时正常，说明不是性能问题，而是功能性错误。
+
+### Step 4: 查看特征分析
+
+```
+操作: browser_click → 特征分析 标签
+观察:
+- 特征维度，出现非 200 状态码，
+- 点击非 200 状态码
+- 点击日志表格中的 trace 链接
+```
+
+### Step 5: 进去 trace 页，分析调用链 Span
+
+```
+操作: browser_click → Trace链接 → 查看 Span 瀑布图
+观察调用链:
+  /demoAddOrder (入口) ❌
+    └── process demoAddOrder
+          └── call kv interface ❌
+                └── /kv ❌
+                      └── process KV
+                            └── kv update
+                                  └── redis setex ❌ ← 根因所在
+```
+
+**关键发现**: 
+
+### Step 6: 查看 Span 详情，定位根因
+
+```
+操作: browser_click → redis setex span → 展开 Events
+观察:
+- Attributes:
+  - command=setex
+  - key=key1
+  - value=0
+- Events (1):
+  - name=exception
+```
+
+**根因定位**: Redis 连接失败！服务无法连接到 `10.201.0.210:6379`。
+
+### 分析报告
+
+```markdown
+## 故障分析报告
+
+### 1. 问题概述
+- **问题描述**: 电商系统订单提交功能成功率降至 0%
+- **分析时间**: 2026-01-15 19:03 ~ 19:07
+- **影响范围**: 电商系统 - 订单子系统 - 订单提交接口
+
+### 2. 关键发现
+
+#### 2.1 灭火图分析
+- **异常卡片**: 订单提交
+- **成功率**: 0%（阈值: 99%）
+- **其他卡片**: 正常
+
+#### 2.2 链路分析
+- **错误 Trace**: 20 条
+- **调用链深度**: 4 层，7 个 Span
+- **错误 Span**: redis setex
+
+#### 2.3 根因 Span 详情
+- **Service**: demo-dianshang
+- **Operation**: redis setex
+- **异常类型**: *net.OpError
+- **异常信息**: dial tcp 10.201.0.210:6379: connect: connection refused
+
+### 3. 根因分析
+- **根本原因**: Redis 服务不可用
+- **证据链**:
+  1. 灭火图显示订单提交成功率 0%
+  2. Tracing 显示 error=true，HTTP 200
+  3. 调用链定位到 redis setex 操作失败
+  4. 异常信息明确指出连接被拒绝
+
+### 4. 建议措施
+- **立即行动**:
+- **后续跟进**:
+```
+
+### 排查要点总结
+
+
+| 步骤  | 工具/页面      | 关键观察点                 |
+| --- | ---------- | --------------------- |
+| 1   | 灭火图首页      | 识别飘红卡片，确定异常范围         |
+| 2   | 卡片详情       | 查看成功率、响应码、延时趋势        |
+| 3   | Tracing 列表 | 查看 error=true 的 Trace |
+| 4   | Span 瀑布图   | 找到带 ❌ 标记的最底层 Span     |
+| 5   | Span 详情    | 展开 Events 查看具体异常信息    |
+
+
+---
+
+## 其他注意事项
+
+1. **时间范围**: 查询时注意时间范围，避免查询过多数据
+2. **Playwright 登录**: 使用 Playwright 前可能需要先登录，登录后 session 会保持
+3. **截图证据**: 重要发现务必截图保存作为证据
+

etc/skills/promql-generator/SKILL.md

@@ -0,0 +1,94 @@
+---
+name: promql-generator
+description: 根据自然语言生成 PromQL 查询语句
+builtin_tools:
+  - list_metrics
+  - get_metric_labels
+---
+
+# PromQL 生成专家
+
+你是一个 PromQL 专家，根据用户的自然语言描述生成正确的 PromQL 查询语句。
+
+## 工作流程
+
+1. **理解用户意图**：分析用户想要查询什么（指标、条件、聚合方式、时间范围等）
+2. **搜索相关指标**：使用 `list_metrics` 工具搜索可能相关的指标名称
+3. **了解指标结构**：使用 `get_metric_labels` 工具获取指标的标签键值，了解可用的过滤维度
+4. **构建 PromQL**：基于获取的元数据，构建准确的 PromQL 查询语句
+
+## 可用工具
+
+### list_metrics
+搜索 Prometheus 指标名称，支持关键词模糊匹配。
+- `keyword`: 搜索关键词（可选）
+- `limit`: 返回数量限制，默认30
+
+### get_metric_labels
+获取指定指标的所有标签键及其可选值。
+- `metric`: 指标名称（必填）
+
+## PromQL 语法要点
+
+### 选择器
+- 即时向量：`metric_name{label="value"}`
+- 范围向量：`metric_name{label="value"}[5m]`
+- 标签匹配：`=`（精确）, `!=`（不等于）, `=~`（正则）, `!~`（正则否定）
+
+### 聚合操作
+- `sum`, `avg`, `max`, `min`, `count`, `stddev`, `stdvar`
+- `topk(n, metric)`, `bottomk(n, metric)`
+- `by (label)` 或 `without (label)` 进行分组
+
+### 常用函数
+- `rate(metric[5m])` - Counter 类型的每秒增长率
+- `increase(metric[1h])` - Counter 类型的增量
+- `irate(metric[5m])` - 瞬时增长率
+- `histogram_quantile(0.95, metric)` - 分位数计算
+- `avg_over_time(metric[1h])` - 时间范围内平均值
+- `absent(metric)` - 检测指标是否存在
+
+### 运算符
+- 算术：`+`, `-`, `*`, `/`, `%`, `^`
+- 比较：`==`, `!=`, `>`, `<`, `>=`, `<=`
+- 逻辑：`and`, `or`, `unless`
+
+## 输出格式
+
+最终答案必须是 JSON 格式：
+
+```json
+{
+    "query": "生成的 PromQL 语句",
+    "explanation": "查询逻辑的简要说明"
+}
+```
+
+## 注意事项
+
+1. **必须使用工具确认**：不要凭空猜测指标名和标签，必须先用工具确认存在
+2. **rate() 的使用**：`rate()` 只能用于 Counter 类型指标（通常以 `_total`, `_count`, `_sum` 结尾）
+3. **时间窗口选择**：
+   - 短时间窗口（1m-5m）：适合实时监控
+   - 中等窗口（15m-1h）：适合趋势分析
+   - 长时间窗口（1h-24h）：适合容量规划
+4. **找不到指标**：如果搜索不到相关指标，说明原因并建议用户检查指标是否存在或提供更多信息
+
+## 示例
+
+### 用户输入
+"查询 CPU 使用率超过 80% 的机器"
+
+### 工作流程
+1. 使用 `list_metrics` 搜索 "cpu" 相关指标
+2. 找到 `node_cpu_seconds_total`，使用 `get_metric_labels` 查看标签
+3. 发现有 `mode`（包含 idle, user, system 等）和 `instance` 标签
+4. 构建 PromQL：计算 CPU 使用率 = 1 - idle 占比
+
+### 输出
+```json
+{
+    "query": "100 - avg by(instance)(rate(node_cpu_seconds_total{mode=\"idle\"}[5m])) * 100 > 80",
+    "explanation": "计算每台机器的 CPU 使用率（100% 减去空闲占比），筛选超过 80% 的实例"
+}
+```

etc/skills/sql-generator/SKILL.md

@@ -0,0 +1,129 @@
+---
+name: sql-generator
+description: 根据自然语言生成 SQL 查询语句（支持 MySQL/Doris/ClickHouse/PostgreSQL）
+builtin_tools:
+  - list_databases
+  - list_tables
+  - describe_table
+---
+
+# SQL 生成专家
+
+你是一个 SQL 专家，根据用户的自然语言描述生成正确的 SQL 查询语句。支持 MySQL、Doris、ClickHouse、PostgreSQL 等数据库。
+
+## 工作流程
+
+1. **理解用户意图**：分析用户想要查询什么数据、什么条件、什么排序
+2. **探索数据库结构**：使用 `list_databases` 查看可用数据库
+3. **查看表列表**：使用 `list_tables` 查看数据库中的表
+4. **了解表结构**：使用 `describe_table` 获取表的字段信息
+5. **构建 SQL**：基于表结构构建准确的 SQL 查询语句
+
+## 可用工具
+
+### list_databases
+列出数据源中的所有数据库。
+- 无参数
+
+### list_tables
+列出指定数据库中的所有表。
+- `database`: 数据库名（必填）
+
+### describe_table
+获取表的字段结构（字段名、类型、注释）。
+- `database`: 数据库名（必填）
+- `table`: 表名（必填）
+
+## SQL 语法要点
+
+### 基础查询
+```sql
+SELECT column1, column2 FROM database.table WHERE condition;
+```
+
+### 聚合函数
+- `COUNT(*)`, `COUNT(DISTINCT column)`
+- `SUM(column)`, `AVG(column)`
+- `MAX(column)`, `MIN(column)`
+
+### 分组和排序
+```sql
+SELECT column, COUNT(*) as cnt
+FROM table
+GROUP BY column
+HAVING cnt > 10
+ORDER BY cnt DESC
+LIMIT 100;
+```
+
+### 时间处理
+- MySQL: `DATE(column)`, `DATE_SUB(NOW(), INTERVAL 7 DAY)`
+- ClickHouse: `toDate(column)`, `now() - INTERVAL 7 DAY`
+- Doris: `DATE(column)`, `DATE_SUB(NOW(), INTERVAL 7 DAY)`
+
+### 连接查询
+```sql
+SELECT a.*, b.name
+FROM table_a a
+LEFT JOIN table_b b ON a.id = b.a_id;
+```
+
+## 不同数据库的差异
+
+### MySQL
+- 字符串连接：`CONCAT(a, b)`
+- 分页：`LIMIT offset, count` 或 `LIMIT count OFFSET offset`
+
+### ClickHouse
+- 字符串连接：`concat(a, b)`
+- 分页：`LIMIT count OFFSET offset`
+- 近似去重：`uniqExact(column)`
+- 时间函数：`toStartOfHour()`, `toStartOfDay()`
+
+### Doris
+- 类似 MySQL 语法
+- 支持 `LIMIT offset, count`
+
+### PostgreSQL
+- 字符串连接：`a || b` 或 `CONCAT(a, b)`
+- 分页：`LIMIT count OFFSET offset`
+- 类型转换：`column::type`
+
+## 输出格式
+
+最终答案必须是 JSON 格式：
+
+```json
+{
+    "query": "生成的 SQL 语句",
+    "explanation": "查询逻辑的简要说明"
+}
+```
+
+## 注意事项
+
+1. **必须使用工具确认**：不要凭空猜测表名和字段名，必须先用工具确认存在
+2. **完整表名**：使用 `database.table` 格式指定表名
+3. **大表查询**：对于大表，建议加上 `LIMIT` 限制返回行数
+4. **时间过滤**：有时间字段时优先使用时间条件过滤，提高查询效率
+5. **找不到表**：如果找不到相关表，说明原因并建议用户检查表是否存在或提供更多信息
+6. **SQL 注入**：生成的 SQL 应该使用参数化查询的思路，不要拼接用户输入
+
+## 示例
+
+### 用户输入
+"查询最近7天每天的订单金额"
+
+### 工作流程
+1. 使用 `list_databases` 找到业务数据库
+2. 使用 `list_tables` 找到订单表
+3. 使用 `describe_table` 查看订单表结构，找到金额字段和时间字段
+4. 构建 SQL
+
+### 输出
+```json
+{
+    "query": "SELECT DATE(created_at) as date, SUM(amount) as total_amount FROM business.orders WHERE created_at >= DATE_SUB(CURDATE(), INTERVAL 7 DAY) GROUP BY DATE(created_at) ORDER BY date",
+    "explanation": "按天分组统计最近7天的订单金额总和，按日期排序"
+}
+```

go.mod

@@ -55,7 +55,7 @@ require (
 	github.com/tidwall/gjson v1.14.2
 	github.com/toolkits/pkg v1.3.8
 	golang.org/x/exp v0.0.0-20231006140011-7918f672742d
-	golang.org/x/oauth2 v0.27.0
+	golang.org/x/oauth2 v0.34.0
 	gopkg.in/gomail.v2 v2.0.0-20160411212932-81ebce5c23df
 	gopkg.in/yaml.v2 v2.4.0
 	gorm.io/driver/clickhouse v0.6.1
@@ -93,6 +93,7 @@ require (
 	github.com/eapache/go-xerial-snappy v0.0.0-20230731223053-c322873962e3 // indirect
 	github.com/eapache/queue v1.1.0 // indirect
 	github.com/glebarez/go-sqlite v1.21.2 // indirect
+	github.com/google/jsonschema-go v0.4.2 // indirect
 	github.com/hashicorp/errwrap v1.1.0 // indirect
 	github.com/hashicorp/go-multierror v1.1.1 // indirect
 	github.com/hashicorp/go-uuid v1.0.3 // indirect
@@ -103,15 +104,18 @@ require (
 	github.com/jcmturner/gofork v1.7.6 // indirect
 	github.com/jcmturner/gokrb5/v8 v8.4.4 // indirect
 	github.com/jcmturner/rpc/v2 v2.0.3 // indirect
+	github.com/modelcontextprotocol/go-sdk v1.4.0 // indirect
 	github.com/pingcap/errors v0.11.5-0.20250523034308-74f78ae071ee // indirect
 	github.com/pingcap/failpoint v0.0.0-20240528011301-b51a646c7c86 // indirect
 	github.com/pingcap/log v1.1.0 // indirect
 	github.com/rcrowley/go-metrics v0.0.0-20201227073835-cf1acfcdf475 // indirect
 	github.com/remyoudompheng/bigfft v0.0.0-20230129092748-24d4a6f8daec // indirect
 	github.com/rogpeppe/go-internal v1.13.1 // indirect
+	github.com/segmentio/encoding v0.5.3 // indirect
 	github.com/tjfoc/gmsm v1.4.1 // indirect
 	github.com/valyala/fastrand v1.1.0 // indirect
 	github.com/valyala/histogram v1.2.0 // indirect
+	github.com/yosida95/uritemplate/v3 v3.0.2 // indirect
 	github.com/yuin/gopher-lua v1.1.1 // indirect
 	go.uber.org/multierr v1.11.0 // indirect
 	go.uber.org/zap v1.27.0 // indirect
@@ -171,7 +175,7 @@ require (
 	golang.org/x/crypto v0.45.0 // indirect
 	golang.org/x/image v0.18.0 // indirect
 	golang.org/x/net v0.47.0 // indirect
-	golang.org/x/sys v0.38.0 // indirect
+	golang.org/x/sys v0.40.0 // indirect
 	golang.org/x/text v0.31.0 // indirect
 	google.golang.org/protobuf v1.35.1 // indirect
 	gopkg.in/alexcesaro/quotedprintable.v3 v3.0.0-20150716171945-2caba252f4dc // indirect

go.sum

@@ -236,7 +236,10 @@ github.com/google/go-cmp v0.5.5/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/
 github.com/google/go-cmp v0.5.8/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
 github.com/google/go-cmp v0.6.0 h1:ofyhxvXcZhMsU5ulbFiLKl/XBFqE1GSq7atu8tAmTRI=
 github.com/google/go-cmp v0.6.0/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
+github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=
 github.com/google/gofuzz v1.0.0/go.mod h1:dBl0BpW6vV/+mYPU4Po3pmUjxk6FQPldtuIdl/M65Eg=
+github.com/google/jsonschema-go v0.4.2 h1:tmrUohrwoLZZS/P3x7ex0WAVknEkBZM46iALbcqoRA8=
+github.com/google/jsonschema-go v0.4.2/go.mod h1:r5quNTdLOYEz95Ru18zA0ydNbBuYoo9tgaYcxEYhJVE=
 github.com/google/pprof v0.0.0-20230705174524-200ffdc848b8 h1:n6vlPhxsA+BW/XsS5+uqi7GyzaLa5MH7qlSLBZtRdiA=
 github.com/google/pprof v0.0.0-20230705174524-200ffdc848b8/go.mod h1:Jh3hGz2jkYak8qXPD19ryItVnUgpgeqzdkY/D0EaeuA=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
@@ -335,6 +338,8 @@ github.com/mattn/go-sqlite3 v1.14.17 h1:mCRHCLDUBXgpKAqIKsaAaAsrAlbkeomtRFKXh2L6
 github.com/mattn/go-sqlite3 v1.14.17/go.mod h1:2eHXhiwb8IkHr+BDWZGa96P6+rkvnG63S2DGjv9HUNg=
 github.com/mitchellh/mapstructure v1.5.0 h1:jeMsZIYE/09sWLaz43PL7Gy6RuMjD2eJVyuac5Z2hdY=
 github.com/mitchellh/mapstructure v1.5.0/go.mod h1:bFUtVrKA4DC2yAKiSyO/QUcy7e+RRV2QTWOzhPopBRo=
+github.com/modelcontextprotocol/go-sdk v1.4.0 h1:u0kr8lbJc1oBcawK7Df+/ajNMpIDFE41OEPxdeTLOn8=
+github.com/modelcontextprotocol/go-sdk v1.4.0/go.mod h1:Nxc2n+n/GdCebUaqCOhTetptS17SXXNu9IfNTaLDi1E=
 github.com/modern-go/concurrent v0.0.0-20180228061459-e0a39a4cb421/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=
 github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd h1:TRLaZ9cD/w8PVh93nsPXa1VrQ6jlwL5oN8l14QlcNfg=
 github.com/modern-go/concurrent v0.0.0-20180306012644-bacd9c7ef1dd/go.mod h1:6dJC0mAP4ikYIbvyc7fijjWJddQyLn8Ig3JB5CqoB9Q=
@@ -420,6 +425,8 @@ github.com/rogpeppe/go-internal v1.13.1/go.mod h1:uMEvuHeurkdAXX61udpOXGD/AzZDWN
 github.com/scylladb/termtables v0.0.0-20191203121021-c4c0b6d42ff4/go.mod h1:C1a7PQSMz9NShzorzCiG2fk9+xuCgLkPeCvMHYR2OWg=
 github.com/segmentio/asm v1.2.0 h1:9BQrFxC+YOHJlTlHGkTrFWf59nbL3XnCoFLTwDCI7ys=
 github.com/segmentio/asm v1.2.0/go.mod h1:BqMnlJP91P8d+4ibuonYZw9mfnzI9HfxselHZr5aAcs=
+github.com/segmentio/encoding v0.5.3 h1:OjMgICtcSFuNvQCdwqMCv9Tg7lEOXGwm1J5RPQccx6w=
+github.com/segmentio/encoding v0.5.3/go.mod h1:HS1ZKa3kSN32ZHVZ7ZLPLXWvOVIiZtyJnO1gPH1sKt0=
 github.com/shopspring/decimal v1.4.0 h1:bxl37RwXBklmTi0C79JfXCEBD1cqqHt0bbgBAGFp81k=
 github.com/shopspring/decimal v1.4.0/go.mod h1:gawqmDU56v4yIKSwfBSFip1HdCCXN8/+DMd9qYNcwME=
 github.com/smartystreets/assertions v0.0.0-20180927180507-b2de0cb4f26d/go.mod h1:OnSkiWE9lh6wB0YB77sQom3nweQdgAjqCqsofrRNTgc=
@@ -469,6 +476,8 @@ github.com/valyala/histogram v1.2.0/go.mod h1:Hb4kBwb4UxsaNbbbh+RRz8ZR6pdodR57tz
 github.com/xdg-go/pbkdf2 v1.0.0/go.mod h1:jrpuAogTd400dnrH08LKmI/xc1MbPOebTwRqcT5RDeI=
 github.com/xdg-go/scram v1.1.1/go.mod h1:RaEWvsqvNKKvBPvcKeFjrG2cJqOkHTiyTpzz23ni57g=
 github.com/xdg-go/stringprep v1.0.3/go.mod h1:W3f5j4i+9rC0kuIEJL0ky1VpHXQU3ocBgklLGvcBnW8=
+github.com/yosida95/uritemplate/v3 v3.0.2 h1:Ed3Oyj9yrmi9087+NczuL5BwkIc4wvTb5zIM+UJPGz4=
+github.com/yosida95/uritemplate/v3 v3.0.2/go.mod h1:ILOh0sOhIJR3+L/8afwt/kE++YT040gmv5BQTMR2HP4=
 github.com/youmark/pkcs8 v0.0.0-20181117223130-1be2e3e5546d/go.mod h1:rHwXgn7JulP+udvsHwJoVG1YGAP6VLg4y9I5dyZdqmA=
 github.com/yuin/goldmark v1.1.27/go.mod h1:3hX8gzYuyVAZsxl0MRgGTJEmQBFcNTphYh9decYSb74=
 github.com/yuin/goldmark v1.1.30/go.mod h1:3hX8gzYuyVAZsxl0MRgGTJEmQBFcNTphYh9decYSb74=
@@ -571,6 +580,8 @@ golang.org/x/net v0.47.0/go.mod h1:/jNxtkgq5yWUGYkaZGqo27cfGZ1c5Nen03aYrrKpVRU=
 golang.org/x/oauth2 v0.0.0-20180821212333-d2e6202438be/go.mod h1:N/0e6XlmueqKjAGxoOufVs8QHGRruUQn6yWY3a++T0U=
 golang.org/x/oauth2 v0.27.0 h1:da9Vo7/tDv5RH/7nZDz1eMGS/q1Vv1N/7FCrBhI9I3M=
 golang.org/x/oauth2 v0.27.0/go.mod h1:onh5ek6nERTohokkhCD/y2cV4Do3fxFHFuAejCkRWT8=
+golang.org/x/oauth2 v0.34.0 h1:hqK/t4AKgbqWkdkcAeI8XLmbK+4m4G5YeQRrmiotGlw=
+golang.org/x/oauth2 v0.34.0/go.mod h1:lzm5WQJQwKZ3nwavOZ3IS5Aulzxi68dUSgRHujetwEA=
 golang.org/x/sync v0.0.0-20180314180146-1d60e4601c6f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20181108010431-42b317875d0f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
@@ -613,6 +624,8 @@ golang.org/x/sys v0.20.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=
 golang.org/x/sys v0.21.0/go.mod h1:/VUhepiaJMQUp4+oa/7Zr1D23ma6VTLIYjOOTFZPUcA=
 golang.org/x/sys v0.38.0 h1:3yZWxaJjBmCWXqhN1qh02AkOnCQ1poK6oF+a7xWL6Gc=
 golang.org/x/sys v0.38.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=
+golang.org/x/sys v0.40.0 h1:DBZZqJ2Rkml6QMQsZywtnjnnGvHza6BTfYFWY9kjEWQ=
+golang.org/x/sys v0.40.0/go.mod h1:OgkHotnGiDImocRcuBABYBEXf8A9a87e/uXjp9XT3ks=
 golang.org/x/telemetry v0.0.0-20240228155512-f48c80bd79b2/go.mod h1:TeRTkGYfJXctD9OcfyVLyj2J3IxLnKwHJR8f4D8a3YE=
 golang.org/x/term v0.0.0-20201126162022-7de9c90e9dd1/go.mod h1:bj7SfCRtBDWHUb9snDiAeCFNEtKQo2Wmx5Cou7ajbmo=
 golang.org/x/term v0.0.0-20210927222741-03fcf44c2211/go.mod h1:jbD1KX2456YbFQfuXm/mYQcufACuNUgVhRMnK/tPxf8=

models/ai_agent.go

@@ -0,0 +1,96 @@
+package models
+
+import (
+	"fmt"
+	"time"
+
+	"github.com/ccfos/nightingale/v6/pkg/ctx"
+)
+
+type AIAgent struct {
+	Id           int64   `json:"id" gorm:"primaryKey;autoIncrement"`
+	Name         string  `json:"name"`
+	Description  string  `json:"description"`
+	UseCase      string  `json:"use_case"`
+	LLMConfigId  int64   `json:"llm_config_id"`
+	SkillIds     []int64 `json:"skill_ids" gorm:"serializer:json"`
+	MCPServerIds []int64 `json:"mcp_server_ids" gorm:"serializer:json"`
+	Enabled      int     `json:"enabled"`
+	CreatedAt    int64   `json:"created_at"`
+	CreatedBy    string  `json:"created_by"`
+	UpdatedAt    int64   `json:"updated_at"`
+	UpdatedBy    string  `json:"updated_by"`
+
+	// Runtime: resolved from LLMConfigId, not stored in DB
+	LLMConfig *AILLMConfig `json:"llm_config,omitempty" gorm:"-"`
+}
+
+func (a *AIAgent) TableName() string {
+	return "ai_agent"
+}
+
+func (a *AIAgent) Verify() error {
+	if a.Name == "" {
+		return fmt.Errorf("name is required")
+	}
+	if a.LLMConfigId <= 0 {
+		return fmt.Errorf("llm_config_id is required")
+	}
+	return nil
+}
+
+func AIAgentGets(c *ctx.Context) ([]*AIAgent, error) {
+	var lst []*AIAgent
+	err := DB(c).Order("id").Find(&lst).Error
+	return lst, err
+}
+
+func AIAgentGet(c *ctx.Context, where string, args ...interface{}) (*AIAgent, error) {
+	var obj AIAgent
+	err := DB(c).Where(where, args...).First(&obj).Error
+	if err != nil {
+		if err.Error() == "record not found" {
+			return nil, nil
+		}
+		return nil, err
+	}
+	return &obj, nil
+}
+
+func AIAgentGetById(c *ctx.Context, id int64) (*AIAgent, error) {
+	return AIAgentGet(c, "id = ?", id)
+}
+
+func (a *AIAgent) Create(c *ctx.Context, username string) error {
+	now := time.Now().Unix()
+	a.CreatedAt = now
+	a.UpdatedAt = now
+	a.CreatedBy = username
+	a.UpdatedBy = username
+	if a.Enabled == 0 {
+		a.Enabled = 1
+	}
+
+	return Insert(c, a)
+}
+
+func (a *AIAgent) Update(c *ctx.Context, username string, data AIAgent) error {
+	data.UpdatedAt = time.Now().Unix()
+	data.UpdatedBy = username
+
+	return DB(c).Model(a).Select("name", "description", "use_case", "llm_config_id",
+		"skill_ids", "mcp_server_ids",
+		"enabled", "updated_at", "updated_by").Updates(data).Error
+}
+
+func (a *AIAgent) Delete(c *ctx.Context) error {
+	return DB(c).Where("id = ?", a.Id).Delete(&AIAgent{}).Error
+}
+
+func AIAgentGetByUseCase(c *ctx.Context, useCase string) (*AIAgent, error) {
+	return AIAgentGet(c, "use_case = ? and enabled = 1", useCase)
+}
+
+func AIAgentStatistics(c *ctx.Context) (*Statistics, error) {
+	return StatisticsGet(c, &AIAgent{})
+}

models/ai_conversation.go

@@ -0,0 +1,104 @@
+package models
+
+import (
+	"fmt"
+	"time"
+
+	"github.com/ccfos/nightingale/v6/pkg/ctx"
+)
+
+type AIConversation struct {
+	Id        int64  `json:"id" gorm:"primaryKey;autoIncrement"`
+	Title     string `json:"title"`
+	UserId    int64  `json:"user_id"`
+	Context   string `json:"context" gorm:"type:text"` // JSON, page-specific context (datasource, alert rule, etc.)
+	CreatedAt int64  `json:"created_at"`
+	UpdatedAt int64  `json:"updated_at"`
+}
+
+func (c *AIConversation) TableName() string {
+	return "ai_conversation"
+}
+
+func AIConversationGetsByUserId(c *ctx.Context, userId int64) ([]*AIConversation, error) {
+	var lst []*AIConversation
+	err := DB(c).Where("user_id = ?", userId).Order("updated_at desc").Find(&lst).Error
+	return lst, err
+}
+
+func AIConversationGetById(c *ctx.Context, id int64) (*AIConversation, error) {
+	var obj AIConversation
+	err := DB(c).Where("id = ?", id).First(&obj).Error
+	if err != nil {
+		if err.Error() == "record not found" {
+			return nil, nil
+		}
+		return nil, err
+	}
+	return &obj, nil
+}
+
+func (a *AIConversation) Create(c *ctx.Context) error {
+	now := time.Now().Unix()
+	a.CreatedAt = now
+	a.UpdatedAt = now
+	return Insert(c, a)
+}
+
+func (a *AIConversation) Update(c *ctx.Context, title string) error {
+	a.Title = title
+	a.UpdatedAt = time.Now().Unix()
+	return DB(c).Model(a).Select("title", "updated_at").Updates(a).Error
+}
+
+func (a *AIConversation) UpdateTime(c *ctx.Context) error {
+	a.UpdatedAt = time.Now().Unix()
+	return DB(c).Model(a).Select("updated_at").Updates(a).Error
+}
+
+func (a *AIConversation) Delete(c *ctx.Context) error {
+	// Cascade delete messages
+	if err := DB(c).Where("conversation_id = ?", a.Id).Delete(&AIConversationMessage{}).Error; err != nil {
+		return err
+	}
+	return DB(c).Where("id = ?", a.Id).Delete(&AIConversation{}).Error
+}
+
+func (a *AIConversation) Verify() error {
+	if a.UserId == 0 {
+		return fmt.Errorf("user_id is required")
+	}
+	return nil
+}
+
+type AIConversationMessage struct {
+	Id             int64  `json:"id" gorm:"primaryKey;autoIncrement"`
+	ConversationId int64  `json:"conversation_id"`
+	Role           string `json:"role"`
+	Content        string `json:"content" gorm:"type:text"`
+	Thinking       string `json:"thinking" gorm:"type:text"`
+	ToolCalls      string `json:"tool_calls" gorm:"type:text"`
+	Query          string `json:"query" gorm:"type:text"`
+	Explanation    string `json:"explanation" gorm:"type:text"`
+	Error          string `json:"error" gorm:"type:text"`
+	CreatedAt      int64  `json:"created_at"`
+}
+
+func (m *AIConversationMessage) TableName() string {
+	return "ai_conversation_message"
+}
+
+func AIConversationMessageGetsByConversationId(c *ctx.Context, conversationId int64) ([]*AIConversationMessage, error) {
+	var lst []*AIConversationMessage
+	err := DB(c).Where("conversation_id = ?", conversationId).Order("id asc").Find(&lst).Error
+	return lst, err
+}
+
+func (m *AIConversationMessage) Create(c *ctx.Context) error {
+	m.CreatedAt = time.Now().Unix()
+	return Insert(c, m)
+}
+
+func AIConversationMessageDeleteByConversationId(c *ctx.Context, conversationId int64) error {
+	return DB(c).Where("conversation_id = ?", conversationId).Delete(&AIConversationMessage{}).Error
+}

models/ai_llm_config.go

@@ -0,0 +1,115 @@
+package models
+
+import (
+	"fmt"
+	"time"
+
+	"github.com/ccfos/nightingale/v6/pkg/ctx"
+)
+
+type LLMExtraConfig struct {
+	TimeoutSeconds int                `json:"timeout_seconds,omitempty"`
+	SkipTLSVerify  bool               `json:"skip_tls_verify,omitempty"`
+	Proxy          string             `json:"proxy,omitempty"`
+	CustomHeaders  map[string]string  `json:"custom_headers,omitempty"`
+	CustomParams   map[string]any     `json:"custom_params,omitempty"`
+	Temperature    *float64           `json:"temperature,omitempty"`
+	MaxTokens      *int               `json:"max_tokens,omitempty"`
+	ContextLength  *int               `json:"context_length,omitempty"`
+}
+
+type AILLMConfig struct {
+	Id          int64          `json:"id" gorm:"primaryKey;autoIncrement"`
+	Name        string         `json:"name"`
+	Description string         `json:"description"`
+	APIType     string         `json:"api_type"`
+	APIURL      string         `json:"api_url"`
+	APIKey      string         `json:"api_key"`
+	Model       string         `json:"model"`
+	ExtraConfig LLMExtraConfig `json:"extra_config" gorm:"serializer:json"`
+	Enabled     int            `json:"enabled"`
+	CreatedAt   int64          `json:"created_at"`
+	CreatedBy   string         `json:"created_by"`
+	UpdatedAt   int64          `json:"updated_at"`
+	UpdatedBy   string         `json:"updated_by"`
+}
+
+func (a *AILLMConfig) TableName() string {
+	return "ai_llm_config"
+}
+
+func (a *AILLMConfig) Verify() error {
+	if a.Name == "" {
+		return fmt.Errorf("name is required")
+	}
+	if a.APIType == "" {
+		return fmt.Errorf("api_type is required")
+	}
+	if a.APIURL == "" {
+		return fmt.Errorf("api_url is required")
+	}
+	if a.APIKey == "" {
+		return fmt.Errorf("api_key is required")
+	}
+	if a.Model == "" {
+		return fmt.Errorf("model is required")
+	}
+	return nil
+}
+
+func AILLMConfigGets(c *ctx.Context) ([]*AILLMConfig, error) {
+	var lst []*AILLMConfig
+	err := DB(c).Order("id").Find(&lst).Error
+	return lst, err
+}
+
+func AILLMConfigGet(c *ctx.Context, where string, args ...interface{}) (*AILLMConfig, error) {
+	var obj AILLMConfig
+	err := DB(c).Where(where, args...).First(&obj).Error
+	if err != nil {
+		if err.Error() == "record not found" {
+			return nil, nil
+		}
+		return nil, err
+	}
+	return &obj, nil
+}
+
+func AILLMConfigGetById(c *ctx.Context, id int64) (*AILLMConfig, error) {
+	return AILLMConfigGet(c, "id = ?", id)
+}
+
+func AILLMConfigGetEnabled(c *ctx.Context) ([]*AILLMConfig, error) {
+	var lst []*AILLMConfig
+	err := DB(c).Where("enabled = 1").Order("id").Find(&lst).Error
+	return lst, err
+}
+
+func (a *AILLMConfig) Create(c *ctx.Context, username string) error {
+	now := time.Now().Unix()
+	a.CreatedAt = now
+	a.UpdatedAt = now
+	a.CreatedBy = username
+	a.UpdatedBy = username
+	if a.Enabled == 0 {
+		a.Enabled = 1
+	}
+	return Insert(c, a)
+}
+
+func (a *AILLMConfig) Update(c *ctx.Context, username string, data AILLMConfig) error {
+	data.UpdatedAt = time.Now().Unix()
+	data.UpdatedBy = username
+
+	// If api_key is empty, keep the original
+	if data.APIKey == "" {
+		data.APIKey = a.APIKey
+	}
+
+	return DB(c).Model(a).Select("name", "description", "api_type", "api_url", "api_key", "model",
+		"extra_config", "enabled", "updated_at", "updated_by").Updates(data).Error
+}
+
+func (a *AILLMConfig) Delete(c *ctx.Context) error {
+	return DB(c).Where("id = ?", a.Id).Delete(&AILLMConfig{}).Error
+}

models/ai_skill.go

@@ -0,0 +1,91 @@
+package models
+
+import (
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/ccfos/nightingale/v6/pkg/ctx"
+)
+
+type AISkill struct {
+	Id            int64             `json:"id" gorm:"primaryKey;autoIncrement"`
+	Name          string            `json:"name"`
+	Description   string            `json:"description"`
+	Instructions  string            `json:"instructions" gorm:"type:text"`
+	License       string            `json:"license,omitempty"`
+	Compatibility string            `json:"compatibility,omitempty"`
+	Metadata      map[string]string `json:"metadata,omitempty" gorm:"serializer:json"`
+	AllowedTools  string            `json:"allowed_tools,omitempty"`
+	Enabled       int               `json:"enabled"`
+	CreatedAt     int64             `json:"created_at"`
+	CreatedBy     string            `json:"created_by"`
+	UpdatedAt     int64             `json:"updated_at"`
+	UpdatedBy     string            `json:"updated_by"`
+
+	// Runtime fields, not stored in DB
+	Files []*AISkillFile `json:"files,omitempty" gorm:"-"`
+}
+
+func (s *AISkill) TableName() string {
+	return "ai_skill"
+}
+
+func (s *AISkill) Verify() error {
+	s.Name = strings.TrimSpace(s.Name)
+	if s.Name == "" {
+		return fmt.Errorf("name is required")
+	}
+	s.Instructions = strings.TrimSpace(s.Instructions)
+	if s.Instructions == "" {
+		return fmt.Errorf("instructions is required")
+	}
+	return nil
+}
+
+func AISkillGets(c *ctx.Context, search string) ([]*AISkill, error) {
+	var lst []*AISkill
+	session := DB(c).Order("id")
+	if search != "" {
+		session = session.Where("name like ? or description like ?", "%"+search+"%", "%"+search+"%")
+	}
+	err := session.Find(&lst).Error
+	return lst, err
+}
+
+func AISkillGet(c *ctx.Context, where string, args ...interface{}) (*AISkill, error) {
+	var obj AISkill
+	err := DB(c).Where(where, args...).First(&obj).Error
+	if err != nil {
+		if err.Error() == "record not found" {
+			return nil, nil
+		}
+		return nil, err
+	}
+	return &obj, nil
+}
+
+func AISkillGetById(c *ctx.Context, id int64) (*AISkill, error) {
+	return AISkillGet(c, "id = ?", id)
+}
+
+func (s *AISkill) Create(c *ctx.Context) error {
+	now := time.Now().Unix()
+	s.CreatedAt = now
+	s.UpdatedAt = now
+	if s.Enabled == 0 {
+		s.Enabled = 1
+	}
+	return Insert(c, s)
+}
+
+func (s *AISkill) Update(c *ctx.Context, ref AISkill) error {
+	ref.UpdatedAt = time.Now().Unix()
+	return DB(c).Model(s).Select("name", "description", "instructions",
+		"license", "compatibility", "metadata", "allowed_tools",
+		"enabled", "updated_at", "updated_by").Updates(ref).Error
+}
+
+func (s *AISkill) Delete(c *ctx.Context) error {
+	return DB(c).Where("id = ?", s.Id).Delete(&AISkill{}).Error
+}

models/ai_skill_file.go

@@ -0,0 +1,72 @@
+package models
+
+import (
+	"fmt"
+	"time"
+
+	"github.com/ccfos/nightingale/v6/pkg/ctx"
+)
+
+type AISkillFile struct {
+	Id        int64  `json:"id" gorm:"primaryKey;autoIncrement"`
+	SkillId   int64  `json:"skill_id"`
+	Name      string `json:"name"`
+	Content   string `json:"content" gorm:"type:mediumtext"`
+	Size      int64  `json:"size"`
+	CreatedAt int64  `json:"created_at"`
+	CreatedBy string `json:"created_by"`
+}
+
+func (f *AISkillFile) TableName() string {
+	return "ai_skill_file"
+}
+
+func AISkillFileGets(c *ctx.Context, skillId int64) ([]*AISkillFile, error) {
+	var lst []*AISkillFile
+	err := DB(c).Select("id, skill_id, name, size, created_at, created_by").Where("skill_id = ?", skillId).Order("id").Find(&lst).Error
+	return lst, err
+}
+
+func AISkillFileGet(c *ctx.Context, where string, args ...interface{}) (*AISkillFile, error) {
+	var obj AISkillFile
+	err := DB(c).Where(where, args...).First(&obj).Error
+	if err != nil {
+		if err.Error() == "record not found" {
+			return nil, nil
+		}
+		return nil, err
+	}
+	return &obj, nil
+}
+
+func AISkillFileGetById(c *ctx.Context, id int64) (*AISkillFile, error) {
+	return AISkillFileGet(c, "id = ?", id)
+}
+
+func (f *AISkillFile) Create(c *ctx.Context) error {
+	f.Size = int64(len(f.Content))
+	f.CreatedAt = time.Now().Unix()
+
+	// Check file count limit per skill (max 20)
+	var count int64
+	DB(c).Model(&AISkillFile{}).Where("skill_id = ?", f.SkillId).Count(&count)
+	if count >= 20 {
+		return fmt.Errorf("max 20 files per skill")
+	}
+
+	return Insert(c, f)
+}
+
+func (f *AISkillFile) Delete(c *ctx.Context) error {
+	return DB(c).Where("id = ?", f.Id).Delete(&AISkillFile{}).Error
+}
+
+func AISkillFileDeleteBySkillId(c *ctx.Context, skillId int64) error {
+	return DB(c).Where("skill_id = ?", skillId).Delete(&AISkillFile{}).Error
+}
+
+func AISkillFileGetContents(c *ctx.Context, skillId int64) ([]*AISkillFile, error) {
+	var lst []*AISkillFile
+	err := DB(c).Where("skill_id = ?", skillId).Find(&lst).Error
+	return lst, err
+}

models/mcp_server.go

@@ -0,0 +1,86 @@
+package models
+
+import (
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/ccfos/nightingale/v6/pkg/ctx"
+)
+
+type MCPServer struct {
+	Id          int64             `json:"id" gorm:"primaryKey;autoIncrement"`
+	Name        string            `json:"name"`
+	URL         string            `json:"url"`
+	Headers     map[string]string `json:"headers" gorm:"serializer:json"`
+	Description string            `json:"description" gorm:"type:text"`
+	Enabled     int               `json:"enabled"`
+	CreatedAt   int64             `json:"created_at"`
+	CreatedBy   string            `json:"created_by"`
+	UpdatedAt   int64             `json:"updated_at"`
+	UpdatedBy   string            `json:"updated_by"`
+}
+
+func (s *MCPServer) TableName() string {
+	return "mcp_server"
+}
+
+func (s *MCPServer) Verify() error {
+	s.Name = strings.TrimSpace(s.Name)
+	if s.Name == "" {
+		return fmt.Errorf("name is required")
+	}
+	s.URL = strings.TrimSpace(s.URL)
+	if s.URL == "" {
+		return fmt.Errorf("url is required")
+	}
+	return nil
+}
+
+func MCPServerGets(c *ctx.Context) ([]*MCPServer, error) {
+	var lst []*MCPServer
+	err := DB(c).Order("id").Find(&lst).Error
+	return lst, err
+}
+
+func MCPServerGet(c *ctx.Context, where string, args ...interface{}) (*MCPServer, error) {
+	var obj MCPServer
+	err := DB(c).Where(where, args...).First(&obj).Error
+	if err != nil {
+		if err.Error() == "record not found" {
+			return nil, nil
+		}
+		return nil, err
+	}
+	return &obj, nil
+}
+
+func MCPServerGetById(c *ctx.Context, id int64) (*MCPServer, error) {
+	return MCPServerGet(c, "id = ?", id)
+}
+
+func (s *MCPServer) Create(c *ctx.Context) error {
+	now := time.Now().Unix()
+	s.CreatedAt = now
+	s.UpdatedAt = now
+	if s.Enabled == 0 {
+		s.Enabled = 1
+	}
+	return Insert(c, s)
+}
+
+func (s *MCPServer) Update(c *ctx.Context, ref MCPServer) error {
+	ref.UpdatedAt = time.Now().Unix()
+	return DB(c).Model(s).Select("name", "url", "headers", "description",
+		"enabled", "updated_at", "updated_by").Updates(ref).Error
+}
+
+func (s *MCPServer) Delete(c *ctx.Context) error {
+	return DB(c).Where("id = ?", s.Id).Delete(&MCPServer{}).Error
+}
+
+func MCPServerGetEnabled(c *ctx.Context) ([]*MCPServer, error) {
+	var lst []*MCPServer
+	err := DB(c).Where("enabled = 1").Order("id").Find(&lst).Error
+	return lst, err
+}

models/migrate/migrate.go

@@ -69,7 +69,9 @@ func MigrateTables(db *gorm.DB) error {
 		&models.MetricFilter{}, &models.NotificationRecord{}, &models.TargetBusiGroup{},
 		&models.UserToken{}, &models.DashAnnotation{}, MessageTemplate{}, NotifyRule{}, NotifyChannelConfig{}, &EsIndexPatternMigrate{},
 		&models.EventPipeline{}, &models.EventPipelineExecution{}, &models.EmbeddedProduct{}, &models.SourceToken{},
-		&models.SavedView{}, &models.UserViewFavorite{}}
+		&models.SavedView{}, &models.UserViewFavorite{},
+		&models.AILLMConfig{}, &models.AIAgent{}, &models.AISkill{}, &models.AISkillFile{}, &models.MCPServer{},
+		&models.AIConversation{}, &models.AIConversationMessage{}}
 
 	if isPostgres(db) {
 		dts = append(dts, &models.PostgresBuiltinComponent{})