Update config.yaml

config.yaml

@@ -10,7 +10,7 @@
 # ============================================
 
 # 前端显示的版本号（可选，不填则显示默认版本）
-version: "v1.5.12"
+version: "v1.5.13"
 # 服务器配置
 server:
   host: 0.0.0.0 # 监听地址，0.0.0.0 表示监听所有网络接口
@@ -70,7 +70,7 @@ multi_agent:
   robot_use_multi_agent: true # true 时企业微信/钉钉/飞书机器人也走 Eino 多代理（成本更高）
   batch_use_multi_agent: false # true 时「批量任务」队列中每个子任务也走 Eino 多代理（成本更高）
   max_iteration: 0 # 主代理 / plan_execute 执行器最大轮次，0 表示沿用 agent.max_iterations
-  # plan_execute 专用：execute↔replan 外层循环上限，0 表示 Eino 默认 10。Executor 未暴露 Handlers：patch/reduction/plantask 不作用于 PE，但 tool_search 工具列表拆分仍通过共享 ToolsConfig 作用于执行器。
+  # plan_execute 专用：execute↔replan 外层循环上限，0 表示 Eino 默认 10。当前实现下 Executor 会挂载 patch/reduction/tool_search 等前置中间件。
   plan_execute_loop_max_iterations: 0
   sub_agent_max_iterations: 120
   sub_agent_user_context_max_runes: 0 # 子代理 task 描述中自动注入用户原始请求的字符上限；0=默认2000，负数=禁用
@@ -87,15 +87,25 @@ multi_agent:
   # Eino ADK 中间件与 Deep/Supervisor 调参（结构体见 internal/config/config.go → MultiAgentEinoMiddlewareConfig）
   eino_middleware:
     patch_tool_calls: true # true：修补历史中无 tool_result 的悬空 tool_call（流式中断/重试后更稳）；false：关闭；字段省略时默认等同 true
-    tool_search_enable: false # true：工具数 ≥ min 时启用 tool_search，仅前 N 个工具常驻，其余按正则按需解锁，省 token、减误选；false：全量工具进上下文
+    tool_search_enable: true # true：工具数 ≥ min 时启用 tool_search，仅前 N 个工具常驻，其余按正则按需解锁，省 token、减误选；false：全量工具进上下文
     tool_search_min_tools: 20 # 达到该数量才启用 tool_search（避免工具很少时多此一举）；与 always_visible 配合使用
     tool_search_always_visible: 12 # 始终直接暴露给模型的工具个数（顺序与角色工具列表一致）；其余工具进入动态池，需 tool_search 解锁
+    tool_search_always_visible_tools: [read_file, glob, grep, write_file, edit_file, execute, task, transfer_to_agent, exit, write_todos, skill, tool_search, TaskCreate, TaskGet, TaskUpdate, TaskList, record_vulnerability, list_knowledge_risk_types, search_knowledge_base, webshell_exec, webshell_file_list, webshell_file_read, webshell_file_write, manage_webshell_list, manage_webshell_add, manage_webshell_update, manage_webshell_delete, manage_webshell_test, batch_task_list, batch_task_get, batch_task_start, batch_task_rerun, batch_task_pause, batch_task_update_metadata, batch_task_update_schedule, batch_task_schedule_enabled, batch_task_update_task, batch_task_remove_task, batch_task_delete, batch_task_create, batch_task_add_task, http-framework-test] # 后端内置常驻工具白名单（优先于 always_visible 数量策略）
     plantask_enable: false # true：主代理（Deep / Supervisor 主）挂载 TaskCreate/Get/Update/List；需 eino_skills 可用且 skills_dir 存在，否则仅打日志并跳过
     plantask_rel_dir: .eino/plantask # 结构化任务文件相对 skills_dir 的子目录，其下再按会话 ID 分子目录存放
-    reduction_enable: false # true：大工具输出截断/落盘以控上下文；依赖与 plantask 相同的 eino local 写盘后端，无后端时不挂载
+    reduction_enable: true # true：大工具输出截断/落盘以控上下文；依赖与 plantask 相同的 eino local 写盘后端，无后端时不挂载
+    reduction_max_length_for_trunc: 50000 # 单条工具结果超过该字符数（bytes）时截断并落盘（由 reduction 中间件处理）
+    reduction_max_tokens_for_clear: 160000 # 历史工具结果清理阈值（tokens），超阈值时在模型调用前清理旧结果
     reduction_root_dir: "" # 非空：截断/清理内容落盘根路径；空：使用系统临时目录下按会话隔离的默认路径
     reduction_clear_exclude: [] # 不参与「清理阶段」的工具名额外列表（会与 task/transfer/exit 等内置排除项合并）；需要时用 YAML 列表填写
-    reduction_sub_agents: false # true：子代理也挂 reduction；false：仅编排主代理使用 reduction
+    reduction_sub_agents: true # true：子代理也挂 reduction；false：仅编排主代理使用 reduction
+    summarization_trigger_ratio: 0.8 # summarization 触发比例（max_total_tokens * ratio），建议 0.75~0.85
+    summarization_emit_internal_events: true # true：发出 summarization 内部事件（便于诊断）
+    history_input_budget_ratio: 0.35 # 历史入队预算比例（max_total_tokens * ratio）
+    plan_execute_user_input_budget_ratio: 0.35 # plan_execute 中 userInput 预算比例（planner/replanner/executor 共用）
+    plan_execute_executed_steps_budget_ratio: 0.2 # plan_execute 中 executed_steps 预算比例
+    plan_execute_max_step_result_runes: 4000 # plan_execute 每步结果最大字符数（超出截断）
+    plan_execute_keep_last_steps: 8 # plan_execute 仅保留最近 N 步正文，早期步骤折叠为标题
     checkpoint_dir: "" # 非空：为 adk.NewRunner 启用按会话子目录的文件型 CheckPointStore，便于中断恢复持久化；Resume 的 HTTP/前端流程需另行对接
     deep_output_key: "" # 非空：将最终助手输出写入 adk session 的键名（Deep 与 Supervisor 主代理）；空表示不写入
     deep_model_retry_max_retries: 0 # >0：ChatModel 调用失败时的框架级最大重试次数（Deep 与 Supervisor 主）；0：不重试

internal/agent/agent.go

@@ -39,6 +39,7 @@ type Agent struct {
 	toolNameMapping       map[string]string // 工具名称映射：OpenAI格式 -> 原始格式（用于外部MCP工具）
 	currentConversationID string            // 当前对话ID（用于自动传递给工具）
 	promptBaseDir         string            // 解析 system_prompt_path 时相对路径的基准目录（通常为 config.yaml 所在目录）
+	toolDescriptionMode   string            // 工具描述模式: "short" | "full"，默认 short
 }
 
 // ResultStorage 结果存储接口（直接使用 storage 包的类型）
@@ -162,6 +163,7 @@ func NewAgent(cfg *config.OpenAIConfig, agentCfg *config.AgentConfig, mcpServer
 		resultStorage:        resultStorage,
 		largeResultThreshold: largeResultThreshold,
 		toolNameMapping:      make(map[string]string), // 初始化工具名称映射
+		toolDescriptionMode:  "short",
 	}
 }
 
@@ -922,7 +924,7 @@ func (a *Agent) AgentLoopWithProgress(ctx context.Context, userInput string, his
 }
 
 // getAvailableTools 获取可用工具
-// 从MCP服务器动态获取工具列表，使用简短描述以减少token消耗
+// 从MCP服务器动态获取工具列表，描述模式由 tool_description_mode 控制
 // roleTools: 角色配置的工具列表（toolKey格式），如果为空或nil，则使用所有工具（默认角色）
 func (a *Agent) getAvailableTools(roleTools []string) []Tool {
 	// 构建角色工具集合（用于快速查找）
@@ -946,11 +948,7 @@ func (a *Agent) getAvailableTools(roleTools []string) []Tool {
 				continue // 不在角色工具列表中，跳过
 			}
 		}
-		// 使用简短描述（如果存在），否则使用详细描述
-		description := mcpTool.ShortDescription
-		if description == "" {
-			description = mcpTool.Description
-		}
+		description := a.pickToolDescription(mcpTool.ShortDescription, mcpTool.Description)
 
 		// 转换schema中的类型为OpenAI标准类型
 		convertedSchema := a.convertSchemaTypes(mcpTool.InputSchema)
@@ -1024,11 +1022,7 @@ func (a *Agent) getAvailableTools(roleTools []string) []Tool {
 					continue
 				}
 
-				// 使用简短描述（如果存在），否则使用详细描述
-				description := externalTool.ShortDescription
-				if description == "" {
-					description = externalTool.Description
-				}
+				description := a.pickToolDescription(externalTool.ShortDescription, externalTool.Description)
 
 				// 转换schema中的类型为OpenAI标准类型
 				convertedSchema := a.convertSchemaTypes(externalTool.InputSchema)
@@ -1063,6 +1057,19 @@ func (a *Agent) getAvailableTools(roleTools []string) []Tool {
 	return tools
 }
 
+func (a *Agent) pickToolDescription(shortDesc, fullDesc string) string {
+	a.mu.RLock()
+	mode := strings.TrimSpace(strings.ToLower(a.toolDescriptionMode))
+	a.mu.RUnlock()
+	if mode == "full" {
+		return fullDesc
+	}
+	if shortDesc != "" {
+		return shortDesc
+	}
+	return fullDesc
+}
+
 // convertSchemaTypes 递归转换schema中的类型为OpenAI标准类型
 func (a *Agent) convertSchemaTypes(schema map[string]interface{}) map[string]interface{} {
 	if schema == nil {
@@ -1665,6 +1672,18 @@ func (a *Agent) UpdateMaxIterations(maxIterations int) {
 	}
 }
 
+// UpdateToolDescriptionMode 更新工具描述模式（short/full）
+func (a *Agent) UpdateToolDescriptionMode(mode string) {
+	a.mu.Lock()
+	defer a.mu.Unlock()
+	mode = strings.TrimSpace(strings.ToLower(mode))
+	if mode != "full" {
+		mode = "short"
+	}
+	a.toolDescriptionMode = mode
+	a.logger.Info("Agent工具描述模式已更新", zap.String("tool_description_mode", mode))
+}
+
 // formatToolError 格式化工具错误信息，提供更友好的错误描述
 func (a *Agent) formatToolError(toolName string, args map[string]interface{}, err error) string {
 	errorMsg := fmt.Sprintf(`工具执行失败

internal/app/app.go

@@ -133,6 +133,7 @@ func New(cfg *config.Config, log *logger.Logger) (*App, error) {
 		maxIterations = 30 // 默认值
 	}
 	agent := agent.NewAgent(&cfg.OpenAI, &cfg.Agent, mcpServer, externalMCPMgr, log.Logger, maxIterations)
+	agent.UpdateToolDescriptionMode(cfg.Security.ToolDescriptionMode)
 
 	// 设置结果存储到Agent
 	agent.SetResultStorage(resultStorage)

internal/config/config.go

@@ -72,6 +72,8 @@ type MultiAgentEinoMiddlewareConfig struct {
 	ToolSearchEnable        bool `yaml:"tool_search_enable,omitempty" json:"tool_search_enable,omitempty"`
 	ToolSearchMinTools      int  `yaml:"tool_search_min_tools,omitempty" json:"tool_search_min_tools,omitempty"`           // default 20; applies when len(tools) >= this
 	ToolSearchAlwaysVisible int  `yaml:"tool_search_always_visible,omitempty" json:"tool_search_always_visible,omitempty"` // default 12; first N tools stay always visible
+	// ToolSearchAlwaysVisibleTools keeps specified tool names always visible (never hidden by tool_search).
+	ToolSearchAlwaysVisibleTools []string `yaml:"tool_search_always_visible_tools,omitempty" json:"tool_search_always_visible_tools,omitempty"`
 	// Plantask adds TaskCreate/Get/Update/List (file-backed under skills dir); requires eino_skills + local backend.
 	PlantaskEnable bool `yaml:"plantask_enable,omitempty" json:"plantask_enable,omitempty"`
 	// PlantaskRelDir relative to skills_dir for per-conversation task boards (default .eino/plantask).
@@ -79,8 +81,24 @@ type MultiAgentEinoMiddlewareConfig struct {
 	// Reduction truncates/offloads large tool outputs (requires eino local backend for Write).
 	ReductionEnable       bool     `yaml:"reduction_enable,omitempty" json:"reduction_enable,omitempty"`
 	ReductionRootDir      string   `yaml:"reduction_root_dir,omitempty" json:"reduction_root_dir,omitempty"` // default: os temp + conversation id
+	ReductionMaxLengthForTrunc int `yaml:"reduction_max_length_for_trunc,omitempty" json:"reduction_max_length_for_trunc,omitempty"` // default 12000
+	ReductionMaxTokensForClear int `yaml:"reduction_max_tokens_for_clear,omitempty" json:"reduction_max_tokens_for_clear,omitempty"` // default 50000
 	ReductionClearExclude []string `yaml:"reduction_clear_exclude,omitempty" json:"reduction_clear_exclude,omitempty"`
 	ReductionSubAgents    bool     `yaml:"reduction_sub_agents,omitempty" json:"reduction_sub_agents,omitempty"` // also attach to sub-agents
+	// SummarizationTriggerRatio controls summarization trigger threshold as max_total_tokens * ratio (default 0.8).
+	SummarizationTriggerRatio float64 `yaml:"summarization_trigger_ratio,omitempty" json:"summarization_trigger_ratio,omitempty"`
+	// SummarizationEmitInternalEvents controls middleware internal event emission (default true).
+	SummarizationEmitInternalEvents *bool `yaml:"summarization_emit_internal_events,omitempty" json:"summarization_emit_internal_events,omitempty"`
+	// HistoryInputBudgetRatio caps pre-agent history tokens as max_total_tokens * ratio (default 0.35).
+	HistoryInputBudgetRatio float64 `yaml:"history_input_budget_ratio,omitempty" json:"history_input_budget_ratio,omitempty"`
+	// PlanExecuteUserInputBudgetRatio caps planner/replanner/executor userInput prompt budget ratio (default 0.35).
+	PlanExecuteUserInputBudgetRatio float64 `yaml:"plan_execute_user_input_budget_ratio,omitempty" json:"plan_execute_user_input_budget_ratio,omitempty"`
+	// PlanExecuteExecutedStepsBudgetRatio caps executed_steps prompt budget ratio (default 0.2).
+	PlanExecuteExecutedStepsBudgetRatio float64 `yaml:"plan_execute_executed_steps_budget_ratio,omitempty" json:"plan_execute_executed_steps_budget_ratio,omitempty"`
+	// PlanExecuteMaxStepResultRunes caps each executed step result length for prompt view (default 4000).
+	PlanExecuteMaxStepResultRunes int `yaml:"plan_execute_max_step_result_runes,omitempty" json:"plan_execute_max_step_result_runes,omitempty"`
+	// PlanExecuteKeepLastSteps keeps only the tail steps in prompt view (default 8).
+	PlanExecuteKeepLastSteps int `yaml:"plan_execute_keep_last_steps,omitempty" json:"plan_execute_keep_last_steps,omitempty"`
 	// CheckpointDir when non-empty enables adk.Runner CheckPointStore (file-backed) for interrupt/resume persistence.
 	CheckpointDir string `yaml:"checkpoint_dir,omitempty" json:"checkpoint_dir,omitempty"`
 	// DeepOutputKey passed to deep.Config OutputKey (session final text); empty = off.
@@ -91,6 +109,97 @@ type MultiAgentEinoMiddlewareConfig struct {
 	TaskToolDescriptionPrefix string `yaml:"task_tool_description_prefix,omitempty" json:"task_tool_description_prefix,omitempty"`
 }
 
+func (c MultiAgentEinoMiddlewareConfig) SummarizationTriggerRatioEffective() float64 {
+	v := c.SummarizationTriggerRatio
+	if v <= 0 {
+		return 0.8
+	}
+	if v < 0.5 {
+		return 0.5
+	}
+	if v > 0.95 {
+		return 0.95
+	}
+	return v
+}
+
+func (c MultiAgentEinoMiddlewareConfig) SummarizationEmitInternalEventsEffective() bool {
+	if c.SummarizationEmitInternalEvents != nil {
+		return *c.SummarizationEmitInternalEvents
+	}
+	return true
+}
+
+func (c MultiAgentEinoMiddlewareConfig) HistoryInputBudgetRatioEffective() float64 {
+	v := c.HistoryInputBudgetRatio
+	if v <= 0 {
+		return 0.35
+	}
+	if v < 0.15 {
+		return 0.15
+	}
+	if v > 0.6 {
+		return 0.6
+	}
+	return v
+}
+
+func (c MultiAgentEinoMiddlewareConfig) PlanExecuteUserInputBudgetRatioEffective() float64 {
+	v := c.PlanExecuteUserInputBudgetRatio
+	if v <= 0 {
+		return 0.35
+	}
+	if v < 0.1 {
+		return 0.1
+	}
+	if v > 0.6 {
+		return 0.6
+	}
+	return v
+}
+
+func (c MultiAgentEinoMiddlewareConfig) PlanExecuteExecutedStepsBudgetRatioEffective() float64 {
+	v := c.PlanExecuteExecutedStepsBudgetRatio
+	if v <= 0 {
+		return 0.2
+	}
+	if v < 0.08 {
+		return 0.08
+	}
+	if v > 0.5 {
+		return 0.5
+	}
+	return v
+}
+
+func (c MultiAgentEinoMiddlewareConfig) PlanExecuteMaxStepResultRunesEffective() int {
+	if c.PlanExecuteMaxStepResultRunes > 0 {
+		return c.PlanExecuteMaxStepResultRunes
+	}
+	return 4000
+}
+
+func (c MultiAgentEinoMiddlewareConfig) PlanExecuteKeepLastStepsEffective() int {
+	if c.PlanExecuteKeepLastSteps > 0 {
+		return c.PlanExecuteKeepLastSteps
+	}
+	return 8
+}
+
+func (c MultiAgentEinoMiddlewareConfig) ReductionMaxLengthForTruncEffective() int {
+	if c.ReductionMaxLengthForTrunc > 0 {
+		return c.ReductionMaxLengthForTrunc
+	}
+	return 12000
+}
+
+func (c MultiAgentEinoMiddlewareConfig) ReductionMaxTokensForClearEffective() int {
+	if c.ReductionMaxTokensForClear > 0 {
+		return c.ReductionMaxTokensForClear
+	}
+	return 50000
+}
+
 // MultiAgentEinoSkillsConfig toggles Eino official skill progressive disclosure and host filesystem tools.
 type MultiAgentEinoSkillsConfig struct {
 	// Disable skips skill middleware (and does not attach local FS tools for Deep).
@@ -137,6 +246,8 @@ type MultiAgentPublic struct {
 	SubAgentCount                int    `json:"sub_agent_count"`
 	Orchestration                string `json:"orchestration,omitempty"`
 	PlanExecuteLoopMaxIterations int    `json:"plan_execute_loop_max_iterations"`
+	ToolSearchAlwaysVisibleTools []string `json:"tool_search_always_visible_tools,omitempty"`
+	ToolSearchAlwaysVisibleEffectiveTools []string `json:"tool_search_always_visible_effective_tools,omitempty"`
 }
 
 // NormalizeMultiAgentOrchestration 返回 deep、plan_execute 或 supervisor。
@@ -158,6 +269,7 @@ type MultiAgentAPIUpdate struct {
 	RobotUseMultiAgent           bool `json:"robot_use_multi_agent"`
 	BatchUseMultiAgent           bool `json:"batch_use_multi_agent"`
 	PlanExecuteLoopMaxIterations *int `json:"plan_execute_loop_max_iterations,omitempty"`
+	ToolSearchAlwaysVisibleTools []string `json:"tool_search_always_visible_tools,omitempty"`
 }
 
 // RobotsConfig 机器人配置（企业微信、钉钉、飞书等）

internal/database/conversation.go

@@ -4,6 +4,8 @@ import (
 	"database/sql"
 	"encoding/json"
 	"fmt"
+	"os"
+	"path/filepath"
 	"strings"
 	"time"
 
@@ -416,6 +418,14 @@ func (db *DB) DeleteConversation(id string) error {
 	if err != nil {
 		return fmt.Errorf("删除对话失败: %w", err)
 	}
+	// Best-effort cleanup for conversation-scoped filesystem artifacts
+	// (e.g., summarization transcript, reduction/checkpoint files under conversation_artifacts/<id>).
+	if base := strings.TrimSpace(db.conversationArtifactsDir); base != "" {
+		artDir := filepath.Join(base, id)
+		if rmErr := os.RemoveAll(artDir); rmErr != nil {
+			db.logger.Warn("删除会话 artifacts 目录失败", zap.String("conversationId", id), zap.String("dir", artDir), zap.Error(rmErr))
+		}
+	}
 
 	db.logger.Info("对话及其所有相关数据已删除", zap.String("conversationId", id))
 	return nil

internal/database/database.go

@@ -3,6 +3,8 @@ package database
 import (
 	"database/sql"
 	"fmt"
+	"os"
+	"path/filepath"
 	"strings"
 	"time"
 
@@ -21,7 +23,8 @@ func configureDBPool(db *sql.DB) {
 // DB 数据库连接
 type DB struct {
 	*sql.DB
-	logger *zap.Logger
+	logger                   *zap.Logger
+	conversationArtifactsDir string
 }
 
 // NewDB 创建数据库连接
@@ -41,6 +44,13 @@ func NewDB(dbPath string, logger *zap.Logger) (*DB, error) {
 		DB:     db,
 		logger: logger,
 	}
+	// Keep conversation-scoped artifacts near database files, so cleanup can follow conversation lifecycle.
+	baseDir := filepath.Join(filepath.Dir(dbPath), "conversation_artifacts")
+	if mkErr := os.MkdirAll(baseDir, 0o755); mkErr == nil {
+		database.conversationArtifactsDir = baseDir
+	} else if logger != nil {
+		logger.Warn("创建 conversation artifacts 目录失败", zap.String("dir", baseDir), zap.Error(mkErr))
+	}
 
 	// 初始化表
 	if err := database.initTables(); err != nil {

internal/handler/config.go

@@ -17,6 +17,7 @@ import (
 	"cyberstrike-ai/internal/config"
 	"cyberstrike-ai/internal/knowledge"
 	"cyberstrike-ai/internal/mcp"
+	"cyberstrike-ai/internal/mcp/builtin"
 	"cyberstrike-ai/internal/openai"
 	"cyberstrike-ai/internal/security"
 
@@ -90,6 +91,7 @@ type AttackChainUpdater interface {
 type AgentUpdater interface {
 	UpdateConfig(cfg *config.OpenAIConfig)
 	UpdateMaxIterations(maxIterations int)
+	UpdateToolDescriptionMode(mode string)
 }
 
 // NewConfigHandler 创建新的配置处理器
@@ -232,13 +234,7 @@ func (h *ConfigHandler) GetConfig(c *gin.Context) {
 			if configToolMap[mcpTool.Name] {
 				continue
 			}
-			description := mcpTool.ShortDescription
-			if description == "" {
-				description = mcpTool.Description
-			}
-			if len(description) > 10000 {
-				description = description[:10000] + "..."
-			}
+			description := h.pickToolDescription(mcpTool.ShortDescription, mcpTool.Description)
 			tools = append(tools, ToolConfigInfo{
 				Name:        mcpTool.Name,
 				Description: description,
@@ -275,6 +271,11 @@ func (h *ConfigHandler) GetConfig(c *gin.Context) {
 		SubAgentCount:                subAgentCount,
 		Orchestration:                config.NormalizeMultiAgentOrchestration(h.config.MultiAgent.Orchestration),
 		PlanExecuteLoopMaxIterations: h.config.MultiAgent.PlanExecuteLoopMaxIterations,
+		ToolSearchAlwaysVisibleTools: append([]string(nil), h.config.MultiAgent.EinoMiddleware.ToolSearchAlwaysVisibleTools...),
+		ToolSearchAlwaysVisibleEffectiveTools: mergeToolNameLists(
+			h.config.MultiAgent.EinoMiddleware.ToolSearchAlwaysVisibleTools,
+			builtin.GetAllBuiltinTools(),
+		),
 	}
 
 	c.JSON(http.StatusOK, GetConfigResponse{
@@ -430,13 +431,7 @@ func (h *ConfigHandler) GetTools(c *gin.Context) {
 				continue
 			}
 
-			description := mcpTool.ShortDescription
-			if description == "" {
-				description = mcpTool.Description
-			}
-			if len(description) > 10000 {
-				description = description[:10000] + "..."
-			}
+			description := h.pickToolDescription(mcpTool.ShortDescription, mcpTool.Description)
 
 			toolInfo := ToolConfigInfo{
 				Name:        mcpTool.Name,
@@ -689,11 +684,13 @@ func (h *ConfigHandler) UpdateConfig(c *gin.Context) {
 		if req.MultiAgent.PlanExecuteLoopMaxIterations != nil {
 			h.config.MultiAgent.PlanExecuteLoopMaxIterations = *req.MultiAgent.PlanExecuteLoopMaxIterations
 		}
+		h.config.MultiAgent.EinoMiddleware.ToolSearchAlwaysVisibleTools = dedupeToolNameList(req.MultiAgent.ToolSearchAlwaysVisibleTools)
 		h.logger.Info("更新多代理配置",
 			zap.Bool("enabled", h.config.MultiAgent.Enabled),
 			zap.Bool("robot_use_multi_agent", h.config.MultiAgent.RobotUseMultiAgent),
 			zap.Bool("batch_use_multi_agent", h.config.MultiAgent.BatchUseMultiAgent),
 			zap.Int("plan_execute_loop_max_iterations", h.config.MultiAgent.PlanExecuteLoopMaxIterations),
+			zap.Int("tool_search_always_visible_tools", len(h.config.MultiAgent.EinoMiddleware.ToolSearchAlwaysVisibleTools)),
 		)
 	}
 
@@ -1061,6 +1058,7 @@ func (h *ConfigHandler) ApplyConfig(c *gin.Context) {
 	if h.agent != nil {
 		h.agent.UpdateConfig(&h.config.OpenAI)
 		h.agent.UpdateMaxIterations(h.config.Agent.MaxIterations)
+		h.agent.UpdateToolDescriptionMode(h.config.Security.ToolDescriptionMode)
 		h.logger.Info("Agent配置已更新")
 	}
 
@@ -1383,6 +1381,33 @@ func updateMultiAgentConfig(doc *yaml.Node, cfg config.MultiAgentConfig) {
 	setBoolInMap(maNode, "robot_use_multi_agent", cfg.RobotUseMultiAgent)
 	setBoolInMap(maNode, "batch_use_multi_agent", cfg.BatchUseMultiAgent)
 	setIntInMap(maNode, "plan_execute_loop_max_iterations", cfg.PlanExecuteLoopMaxIterations)
+	mwNode := ensureMap(maNode, "eino_middleware")
+	setFlowStringSliceInMap(mwNode, "tool_search_always_visible_tools", dedupeToolNameList(cfg.EinoMiddleware.ToolSearchAlwaysVisibleTools))
+}
+
+func dedupeToolNameList(in []string) []string {
+	if len(in) == 0 {
+		return []string{}
+	}
+	seen := make(map[string]struct{}, len(in))
+	out := make([]string, 0, len(in))
+	for _, name := range in {
+		n := strings.TrimSpace(name)
+		if n == "" {
+			continue
+		}
+		key := strings.ToLower(n)
+		if _, ok := seen[key]; ok {
+			continue
+		}
+		seen[key] = struct{}{}
+		out = append(out, n)
+	}
+	return out
+}
+
+func mergeToolNameLists(a, b []string) []string {
+	return dedupeToolNameList(append(append([]string{}, a...), b...))
 }
 
 func ensureMap(parent *yaml.Node, path ...string) *yaml.Node {

internal/multiagent/eino_input_telemetry.go

@@ -0,0 +1,133 @@
+package multiagent
+
+import (
+	"context"
+	"strings"
+
+	"cyberstrike-ai/internal/agent"
+
+	"github.com/bytedance/sonic"
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/schema"
+	"go.uber.org/zap"
+)
+
+type einoModelInputTelemetryMiddleware struct {
+	adk.BaseChatModelAgentMiddleware
+	logger         *zap.Logger
+	modelName      string
+	conversationID string
+	phase          string
+}
+
+func newEinoModelInputTelemetryMiddleware(
+	logger *zap.Logger,
+	modelName string,
+	conversationID string,
+	phase string,
+) adk.ChatModelAgentMiddleware {
+	if logger == nil {
+		return nil
+	}
+	return &einoModelInputTelemetryMiddleware{
+		logger:         logger,
+		modelName:      strings.TrimSpace(modelName),
+		conversationID: strings.TrimSpace(conversationID),
+		phase:          strings.TrimSpace(phase),
+	}
+}
+
+func (m *einoModelInputTelemetryMiddleware) BeforeModelRewriteState(
+	ctx context.Context,
+	state *adk.ChatModelAgentState,
+	mc *adk.ModelContext,
+) (context.Context, *adk.ChatModelAgentState, error) {
+	if m == nil || m.logger == nil || state == nil {
+		return ctx, state, nil
+	}
+	tokens := estimateTokensForMessagesAndTools(ctx, m.modelName, state.Messages, mcTools(mc))
+	m.logger.Info("eino model input estimated",
+		zap.String("phase", m.phase),
+		zap.String("conversation_id", m.conversationID),
+		zap.Int("messages", len(state.Messages)),
+		zap.Int("tools", len(mcTools(mc))),
+		zap.Int("input_tokens_estimated", tokens),
+	)
+	return ctx, state, nil
+}
+
+func mcTools(mc *adk.ModelContext) []*schema.ToolInfo {
+	if mc == nil || len(mc.Tools) == 0 {
+		return nil
+	}
+	return mc.Tools
+}
+
+func estimateTokensForMessagesAndTools(
+	_ context.Context,
+	modelName string,
+	messages []adk.Message,
+	tools []*schema.ToolInfo,
+) int {
+	var sb strings.Builder
+	for _, msg := range messages {
+		if msg == nil {
+			continue
+		}
+		sb.WriteString(string(msg.Role))
+		sb.WriteByte('\n')
+		sb.WriteString(msg.Content)
+		sb.WriteByte('\n')
+		if msg.ReasoningContent != "" {
+			sb.WriteString(msg.ReasoningContent)
+			sb.WriteByte('\n')
+		}
+		if len(msg.ToolCalls) > 0 {
+			if b, err := sonic.Marshal(msg.ToolCalls); err == nil {
+				sb.Write(b)
+				sb.WriteByte('\n')
+			}
+		}
+	}
+	for _, tl := range tools {
+		if tl == nil {
+			continue
+		}
+		cp := *tl
+		cp.Extra = nil
+		if text, err := sonic.MarshalString(cp); err == nil {
+			sb.WriteString(text)
+			sb.WriteByte('\n')
+		}
+	}
+	text := sb.String()
+	if text == "" {
+		return 0
+	}
+	tc := agent.NewTikTokenCounter()
+	if n, err := tc.Count(modelName, text); err == nil {
+		return n
+	}
+	return (len(text) + 3) / 4
+}
+
+func logPlanExecuteModelInputEstimate(
+	logger *zap.Logger,
+	modelName string,
+	conversationID string,
+	phase string,
+	msgs []adk.Message,
+) {
+	if logger == nil {
+		return
+	}
+	tokens := estimateTokensForMessagesAndTools(context.Background(), modelName, msgs, nil)
+	logger.Info("eino model input estimated",
+		zap.String("phase", phase),
+		zap.String("conversation_id", strings.TrimSpace(conversationID)),
+		zap.Int("messages", len(msgs)),
+		zap.Int("tools", 0),
+		zap.Int("input_tokens_estimated", tokens),
+	)
+}
+

internal/multiagent/eino_middleware.go

@@ -8,6 +8,7 @@ import (
 	"strings"
 
 	"cyberstrike-ai/internal/config"
+	"cyberstrike-ai/internal/mcp/builtin"
 
 	localbk "github.com/cloudwego/eino-ext/adk/backend/local"
 	"github.com/cloudwego/eino/adk"
@@ -65,6 +66,66 @@ func splitToolsForToolSearch(all []tool.BaseTool, alwaysVisible int) (static []t
 	return append([]tool.BaseTool(nil), all[:alwaysVisible]...), append([]tool.BaseTool(nil), all[alwaysVisible:]...), true
 }
 
+func splitToolsForToolSearchByNames(all []tool.BaseTool, names []string, fallbackAlwaysVisible int) (static []tool.BaseTool, dynamic []tool.BaseTool, ok bool) {
+	nameSet := make(map[string]struct{}, len(names))
+	for _, n := range names {
+		n = strings.TrimSpace(strings.ToLower(n))
+		if n == "" {
+			continue
+		}
+		nameSet[n] = struct{}{}
+	}
+	if len(nameSet) == 0 {
+		return splitToolsForToolSearch(all, fallbackAlwaysVisible)
+	}
+	static = make([]tool.BaseTool, 0, len(all))
+	dynamic = make([]tool.BaseTool, 0, len(all))
+	for _, t := range all {
+		if t == nil {
+			continue
+		}
+		info, err := t.Info(context.Background())
+		name := ""
+		if err == nil && info != nil {
+			name = strings.TrimSpace(strings.ToLower(info.Name))
+		}
+		if _, keep := nameSet[name]; keep {
+			static = append(static, t)
+			continue
+		}
+		dynamic = append(dynamic, t)
+	}
+	if len(static) == 0 || len(dynamic) == 0 {
+		// fallback: preserve previous behavior when whitelist misses all or includes all.
+		return splitToolsForToolSearch(all, fallbackAlwaysVisible)
+	}
+	return static, dynamic, true
+}
+
+func mergeAlwaysVisibleToolNames(configured []string) []string {
+	merged := make([]string, 0, len(configured)+32)
+	seen := make(map[string]struct{}, len(configured)+32)
+	add := func(name string) {
+		n := strings.TrimSpace(strings.ToLower(name))
+		if n == "" {
+			return
+		}
+		if _, ok := seen[n]; ok {
+			return
+		}
+		seen[n] = struct{}{}
+		merged = append(merged, n)
+	}
+	for _, n := range configured {
+		add(n)
+	}
+	// Always include hardcoded backend builtin MCP tools from constants.
+	for _, n := range builtin.GetAllBuiltinTools() {
+		add(n)
+	}
+	return merged
+}
+
 func buildReductionMiddleware(ctx context.Context, mw config.MultiAgentEinoMiddlewareConfig, convID string, loc *localbk.Local, logger *zap.Logger) (adk.ChatModelAgentMiddleware, error) {
 	if loc == nil {
 		return nil, fmt.Errorf("reduction: local backend nil")
@@ -87,6 +148,8 @@ func buildReductionMiddleware(ctx context.Context, mw config.MultiAgentEinoMiddl
 		RootDir:           root,
 		ReadFileToolName:  "read_file",
 		ClearExcludeTools: excl,
+		MaxLengthForTrunc: mw.ReductionMaxLengthForTruncEffective(),
+		MaxTokensForClear: int64(mw.ReductionMaxTokensForClearEffective()),
 	})
 	if err != nil {
 		return nil, err
@@ -142,7 +205,7 @@ func prependEinoMiddlewares(
 		alwaysVis = 12
 	}
 	if mw.ToolSearchEnable && len(tools) >= minTools {
-		static, dynamic, split := splitToolsForToolSearch(tools, alwaysVis)
+		static, dynamic, split := splitToolsForToolSearchByNames(tools, mergeAlwaysVisibleToolNames(mw.ToolSearchAlwaysVisibleTools), alwaysVis)
 		if split && len(dynamic) > 0 {
 			ts, terr := toolsearch.New(ctx, &toolsearch.Config{DynamicTools: dynamic})
 			if terr != nil {

internal/multiagent/eino_model_rewrite_pipeline.go

@@ -0,0 +1,38 @@
+package multiagent
+
+import (
+	"context"
+	"fmt"
+
+	"github.com/cloudwego/eino/adk"
+)
+
+func applyBeforeModelRewriteHandlers(
+	ctx context.Context,
+	msgs []adk.Message,
+	handlers []adk.ChatModelAgentMiddleware,
+) ([]adk.Message, error) {
+	if len(msgs) == 0 || len(handlers) == 0 {
+		return msgs, nil
+	}
+	state := &adk.ChatModelAgentState{Messages: msgs}
+	modelCtx := &adk.ModelContext{}
+	curCtx := ctx
+	for _, h := range handlers {
+		if h == nil {
+			continue
+		}
+		nextCtx, nextState, err := h.BeforeModelRewriteState(curCtx, state, modelCtx)
+		if err != nil {
+			return nil, fmt.Errorf("before model rewrite: %w", err)
+		}
+		if nextCtx != nil {
+			curCtx = nextCtx
+		}
+		if nextState != nil {
+			state = nextState
+		}
+	}
+	return state.Messages, nil
+}
+

internal/multiagent/eino_orchestration.go

@@ -5,6 +5,7 @@ import (
 	"fmt"
 	"strings"
 
+	"cyberstrike-ai/internal/agent"
 	"cyberstrike-ai/internal/config"
 
 	"github.com/cloudwego/eino-ext/components/model/openai"
@@ -25,7 +26,12 @@ type PlanExecuteRootArgs struct {
 	LoopMaxIter          int
 	// AppCfg / Logger 非空时为 Executor 挂载与 Deep/Supervisor 一致的 Eino summarization 中间件。
 	AppCfg *config.Config
+	MwCfg  *config.MultiAgentEinoMiddlewareConfig
+	// ConversationID is used for transcript/isolation paths in middleware.
+	ConversationID string
 	Logger *zap.Logger
+	// ModelName is used for model input token estimation logs.
+	ModelName string
 	// ExecPreMiddlewares 是由 prependEinoMiddlewares 构建的前置中间件（patchtoolcalls, reduction, toolsearch, plantask），
 	// 与 Deep/Supervisor 主代理的 mainOrchestratorPre 一致。
 	ExecPreMiddlewares []adk.ChatModelAgentMiddleware
@@ -33,6 +39,8 @@ type PlanExecuteRootArgs struct {
 	SkillMiddleware adk.ChatModelAgentMiddleware
 	// FilesystemMiddleware 是 Eino filesystem 中间件，当 eino_skills.filesystem_tools 启用时提供本机文件读写与 Shell 能力（可选）。
 	FilesystemMiddleware adk.ChatModelAgentMiddleware
+	// PlannerReplannerRewriteHandlers applies BeforeModelRewriteState pipeline for planner/replanner input.
+	PlannerReplannerRewriteHandlers []adk.ChatModelAgentMiddleware
 }
 
 // NewPlanExecuteRoot 返回 plan → execute → replan 预置编排根节点（与 Deep / Supervisor 并列）。
@@ -50,7 +58,7 @@ func NewPlanExecuteRoot(ctx context.Context, a *PlanExecuteRootArgs) (adk.Resuma
 	plannerCfg := &planexecute.PlannerConfig{
 		ToolCallingChatModel: tcm,
 	}
-	if fn := planExecutePlannerGenInput(a.OrchInstruction); fn != nil {
+	if fn := planExecutePlannerGenInput(a.OrchInstruction, a.AppCfg, a.MwCfg, a.Logger, a.ModelName, a.ConversationID, a.PlannerReplannerRewriteHandlers); fn != nil {
 		plannerCfg.GenInputFn = fn
 	}
 	planner, err := planexecute.NewPlanner(ctx, plannerCfg)
@@ -59,7 +67,7 @@ func NewPlanExecuteRoot(ctx context.Context, a *PlanExecuteRootArgs) (adk.Resuma
 	}
 	replanner, err := planexecute.NewReplanner(ctx, &planexecute.ReplannerConfig{
 		ChatModel:  tcm,
-		GenInputFn: planExecuteReplannerGenInput(a.OrchInstruction),
+		GenInputFn: planExecuteReplannerGenInput(a.OrchInstruction, a.AppCfg, a.MwCfg, a.Logger, a.ModelName, a.ConversationID, a.PlannerReplannerRewriteHandlers),
 	})
 	if err != nil {
 		return nil, fmt.Errorf("plan_execute replanner: %w", err)
@@ -81,17 +89,20 @@ func NewPlanExecuteRoot(ctx context.Context, a *PlanExecuteRootArgs) (adk.Resuma
 	}
 	// 4. summarization（最后，与 Deep/Supervisor 一致）
 	if a.AppCfg != nil {
-		sumMw, sumErr := newEinoSummarizationMiddleware(ctx, a.ExecModel, a.AppCfg, a.Logger)
+		sumMw, sumErr := newEinoSummarizationMiddleware(ctx, a.ExecModel, a.AppCfg, a.MwCfg, a.ConversationID, a.Logger)
 		if sumErr != nil {
 			return nil, fmt.Errorf("plan_execute executor summarization: %w", sumErr)
 		}
 		execHandlers = append(execHandlers, sumMw)
 	}
+	if teleMw := newEinoModelInputTelemetryMiddleware(a.Logger, a.ModelName, a.ConversationID, "plan_execute_executor"); teleMw != nil {
+		execHandlers = append(execHandlers, teleMw)
+	}
 	executor, err := newPlanExecuteExecutor(ctx, &planexecute.ExecutorConfig{
 		Model:         a.ExecModel,
 		ToolsConfig:   a.ToolsCfg,
 		MaxIterations: a.ExecMaxIter,
-		GenInputFn:    planExecuteExecutorGenInput(a.OrchInstruction),
+		GenInputFn:    planExecuteExecutorGenInput(a.OrchInstruction, a.AppCfg, a.MwCfg, a.Logger, a.ModelName, a.ConversationID),
 	}, execHandlers)
 	if err != nil {
 		return nil, fmt.Errorf("plan_execute executor: %w", err)
@@ -110,20 +121,42 @@ func NewPlanExecuteRoot(ctx context.Context, a *PlanExecuteRootArgs) (adk.Resuma
 
 // planExecutePlannerGenInput 将 orchestrator instruction 作为 SystemMessage 注入 planner 输入。
 // 返回 nil 时 Eino 使用内置默认 planner prompt。
-func planExecutePlannerGenInput(orchInstruction string) planexecute.GenPlannerModelInputFn {
+func planExecutePlannerGenInput(
+	orchInstruction string,
+	appCfg *config.Config,
+	mwCfg *config.MultiAgentEinoMiddlewareConfig,
+	logger *zap.Logger,
+	modelName string,
+	conversationID string,
+	rewriteHandlers []adk.ChatModelAgentMiddleware,
+) planexecute.GenPlannerModelInputFn {
 	oi := strings.TrimSpace(orchInstruction)
-	if oi == "" {
+	if oi == "" && appCfg == nil {
 		return nil
 	}
 	return func(ctx context.Context, userInput []adk.Message) ([]adk.Message, error) {
+		userInput = capPlanExecuteUserInputMessages(userInput, appCfg, mwCfg)
 		msgs := make([]adk.Message, 0, 1+len(userInput))
-		msgs = append(msgs, schema.SystemMessage(oi))
+		if oi != "" {
+			msgs = append(msgs, schema.SystemMessage(oi))
+		}
 		msgs = append(msgs, userInput...)
+		if rewritten, rerr := applyBeforeModelRewriteHandlers(ctx, msgs, rewriteHandlers); rerr == nil && len(rewritten) > 0 {
+			msgs = rewritten
+		}
+		logPlanExecuteModelInputEstimate(logger, modelName, conversationID, "plan_execute_planner", msgs)
 		return msgs, nil
 	}
 }
 
-func planExecuteExecutorGenInput(orchInstruction string) planexecute.GenModelInputFn {
+func planExecuteExecutorGenInput(
+	orchInstruction string,
+	appCfg *config.Config,
+	mwCfg *config.MultiAgentEinoMiddlewareConfig,
+	logger *zap.Logger,
+	modelName string,
+	conversationID string,
+) planexecute.GenModelInputFn {
 	oi := strings.TrimSpace(orchInstruction)
 	return func(ctx context.Context, in *planexecute.ExecutionContext) ([]adk.Message, error) {
 		planContent, err := in.Plan.MarshalJSON()
@@ -131,9 +164,9 @@ func planExecuteExecutorGenInput(orchInstruction string) planexecute.GenModelInp
 			return nil, err
 		}
 		userMsgs, err := planexecute.ExecutorPrompt.Format(ctx, map[string]any{
-			"input":          planExecuteFormatInput(in.UserInput),
+			"input":          planExecuteFormatInput(capPlanExecuteUserInputMessages(in.UserInput, appCfg, mwCfg)),
 			"plan":           string(planContent),
-			"executed_steps": planExecuteFormatExecutedSteps(in.ExecutedSteps),
+			"executed_steps": planExecuteFormatExecutedSteps(in.ExecutedSteps, appCfg, mwCfg),
 			"step":           in.Plan.FirstStep(),
 		})
 		if err != nil {
@@ -142,6 +175,7 @@ func planExecuteExecutorGenInput(orchInstruction string) planexecute.GenModelInp
 		if oi != "" {
 			userMsgs = append([]adk.Message{schema.SystemMessage(oi)}, userMsgs...)
 		}
+		logPlanExecuteModelInputEstimate(logger, modelName, conversationID, "plan_execute_executor_gen_input", userMsgs)
 		return userMsgs, nil
 	}
 }
@@ -155,18 +189,22 @@ func planExecuteFormatInput(input []adk.Message) string {
 	return sb.String()
 }
 
-func planExecuteFormatExecutedSteps(results []planexecute.ExecutedStep) string {
-	capped := capPlanExecuteExecutedSteps(results)
-	var sb strings.Builder
-	for _, result := range capped {
-		sb.WriteString(fmt.Sprintf("Step: %s\nResult: %s\n\n", result.Step, result.Result))
-	}
-	return sb.String()
+func planExecuteFormatExecutedSteps(results []planexecute.ExecutedStep, appCfg *config.Config, mwCfg *config.MultiAgentEinoMiddlewareConfig) string {
+	capped := capPlanExecuteExecutedStepsWithConfig(results, mwCfg)
+	return renderPlanExecuteStepsByBudget(capped, appCfg, mwCfg)
 }
 
 // planExecuteReplannerGenInput 与 Eino 默认 Replanner 输入一致，但 executed_steps 经 cap 后再写入 prompt，
 // 且在 orchInstruction 非空时 prepend SystemMessage 使 replanner 也能接收全局指令。
-func planExecuteReplannerGenInput(orchInstruction string) planexecute.GenModelInputFn {
+func planExecuteReplannerGenInput(
+	orchInstruction string,
+	appCfg *config.Config,
+	mwCfg *config.MultiAgentEinoMiddlewareConfig,
+	logger *zap.Logger,
+	modelName string,
+	conversationID string,
+	rewriteHandlers []adk.ChatModelAgentMiddleware,
+) planexecute.GenModelInputFn {
 	oi := strings.TrimSpace(orchInstruction)
 	return func(ctx context.Context, in *planexecute.ExecutionContext) ([]adk.Message, error) {
 		planContent, err := in.Plan.MarshalJSON()
@@ -175,8 +213,8 @@ func planExecuteReplannerGenInput(orchInstruction string) planexecute.GenModelIn
 		}
 		msgs, err := planexecute.ReplannerPrompt.Format(ctx, map[string]any{
 			"plan":           string(planContent),
-			"input":          planExecuteFormatInput(in.UserInput),
-			"executed_steps": planExecuteFormatExecutedSteps(in.ExecutedSteps),
+			"input":          planExecuteFormatInput(capPlanExecuteUserInputMessages(in.UserInput, appCfg, mwCfg)),
+			"executed_steps": planExecuteFormatExecutedSteps(in.ExecutedSteps, appCfg, mwCfg),
 			"plan_tool":      planexecute.PlanToolInfo.Name,
 			"respond_tool":   planexecute.RespondToolInfo.Name,
 		})
@@ -186,10 +224,120 @@ func planExecuteReplannerGenInput(orchInstruction string) planexecute.GenModelIn
 		if oi != "" {
 			msgs = append([]adk.Message{schema.SystemMessage(oi)}, msgs...)
 		}
+		if rewritten, rerr := applyBeforeModelRewriteHandlers(ctx, msgs, rewriteHandlers); rerr == nil && len(rewritten) > 0 {
+			msgs = rewritten
+		}
+		logPlanExecuteModelInputEstimate(logger, modelName, conversationID, "plan_execute_replanner", msgs)
 		return msgs, nil
 	}
 }
 
+func capPlanExecuteUserInputMessages(input []adk.Message, appCfg *config.Config, mwCfg *config.MultiAgentEinoMiddlewareConfig) []adk.Message {
+	if len(input) == 0 {
+		return input
+	}
+	maxTotal := 120000
+	modelName := "gpt-4o"
+	if appCfg != nil {
+		if appCfg.OpenAI.MaxTotalTokens > 0 {
+			maxTotal = appCfg.OpenAI.MaxTotalTokens
+		}
+		if m := strings.TrimSpace(appCfg.OpenAI.Model); m != "" {
+			modelName = m
+		}
+	}
+	// Reserve most tokens for planner/replanner prompt and tool schema.
+	ratio := 0.35
+	if mwCfg != nil {
+		ratio = mwCfg.PlanExecuteUserInputBudgetRatioEffective()
+	}
+	budget := int(float64(maxTotal) * ratio)
+	if budget < 4096 {
+		budget = 4096
+	}
+	tc := agent.NewTikTokenCounter()
+	out := make([]adk.Message, 0, len(input))
+	used := 0
+	for i := len(input) - 1; i >= 0; i-- {
+		msg := input[i]
+		if msg == nil {
+			continue
+		}
+		n, err := tc.Count(modelName, string(msg.Role)+"\n"+msg.Content)
+		if err != nil {
+			n = (len(msg.Content) + 3) / 4
+		}
+		if n <= 0 {
+			n = 1
+		}
+		if used+n > budget {
+			break
+		}
+		used += n
+		out = append(out, msg)
+	}
+	for i, j := 0, len(out)-1; i < j; i, j = i+1, j-1 {
+		out[i], out[j] = out[j], out[i]
+	}
+	if len(out) == 0 {
+		// Keep the latest user message at least.
+		return []adk.Message{input[len(input)-1]}
+	}
+	return out
+}
+
+func renderPlanExecuteStepsByBudget(steps []planexecute.ExecutedStep, appCfg *config.Config, mwCfg *config.MultiAgentEinoMiddlewareConfig) string {
+	if len(steps) == 0 {
+		return ""
+	}
+	maxTotal := 120000
+	modelName := "gpt-4o"
+	if appCfg != nil {
+		if appCfg.OpenAI.MaxTotalTokens > 0 {
+			maxTotal = appCfg.OpenAI.MaxTotalTokens
+		}
+		if m := strings.TrimSpace(appCfg.OpenAI.Model); m != "" {
+			modelName = m
+		}
+	}
+	ratio := 0.2
+	if mwCfg != nil {
+		ratio = mwCfg.PlanExecuteExecutedStepsBudgetRatioEffective()
+	}
+	budget := int(float64(maxTotal) * ratio)
+	if budget < 3072 {
+		budget = 3072
+	}
+	tc := agent.NewTikTokenCounter()
+	var kept []string
+	used := 0
+	skipped := 0
+	for i := len(steps) - 1; i >= 0; i-- {
+		block := fmt.Sprintf("Step: %s\nResult: %s\n\n", steps[i].Step, steps[i].Result)
+		n, err := tc.Count(modelName, block)
+		if err != nil {
+			n = (len(block) + 3) / 4
+		}
+		if n <= 0 {
+			n = 1
+		}
+		if used+n > budget {
+			skipped = i + 1
+			break
+		}
+		used += n
+		kept = append(kept, block)
+	}
+	var sb strings.Builder
+	if skipped > 0 {
+		sb.WriteString(fmt.Sprintf("Earlier executed steps omitted due to context budget: %d steps.\n\n", skipped))
+	}
+	for i := len(kept) - 1; i >= 0; i-- {
+		sb.WriteString(kept[i])
+	}
+	return sb.String()
+}
+
 // planExecuteStreamsMainAssistant 将规划/执行/重规划各阶段助手流式输出映射到主对话区。
 func planExecuteStreamsMainAssistant(agent string) bool {
 	if agent == "" {

internal/multiagent/eino_single_runner.go

@@ -125,7 +125,7 @@ func RunEinoSingleChatModelAgent(
 		return nil, fmt.Errorf("eino single 模型: %w", err)
 	}
 
-	mainSumMw, err := newEinoSummarizationMiddleware(ctx, mainModel, appCfg, logger)
+	mainSumMw, err := newEinoSummarizationMiddleware(ctx, mainModel, appCfg, &ma.EinoMiddleware, conversationID, logger)
 	if err != nil {
 		return nil, fmt.Errorf("eino single summarization: %w", err)
 	}
@@ -145,6 +145,9 @@ func RunEinoSingleChatModelAgent(
 		handlers = append(handlers, einoSkillMW)
 	}
 	handlers = append(handlers, mainSumMw)
+	if teleMw := newEinoModelInputTelemetryMiddleware(logger, appCfg.OpenAI.Model, conversationID, "eino_single"); teleMw != nil {
+		handlers = append(handlers, teleMw)
+	}
 
 	maxIter := ma.MaxIteration
 	if maxIter <= 0 {
@@ -165,11 +168,29 @@ func RunEinoSingleChatModelAgent(
 		},
 		EmitInternalEvents: true,
 	}
+	ins := injectToolNamesOnlyInstruction(ctx, ag.EinoSingleAgentSystemInstruction(), mainTools)
+	if logger != nil {
+		names := collectToolNames(ctx, mainTools)
+		mountedNames := collectToolNames(ctx, mainToolsForCfg)
+		hasToolSearch := false
+		for _, n := range names {
+			if strings.EqualFold(strings.TrimSpace(n), "tool_search") {
+				hasToolSearch = true
+				break
+			}
+		}
+		logger.Info("eino tool-name injection",
+			zap.String("scope", "eino_single"),
+			zap.Int("tool_names", len(names)),
+			zap.Int("mounted_tool_names", len(mountedNames)),
+			zap.Bool("has_tool_search", hasToolSearch),
+		)
+	}
 
 	chatCfg := &adk.ChatModelAgentConfig{
 		Name:          einoSingleAgentName,
 		Description:   "Eino ADK ChatModelAgent with MCP tools for authorized security testing.",
-		Instruction:   ag.EinoSingleAgentSystemInstruction(),
+		Instruction:   ins,
 		Model:         mainModel,
 		ToolsConfig:   mainToolsCfg,
 		MaxIterations: maxIter,
@@ -188,7 +209,7 @@ func RunEinoSingleChatModelAgent(
 		return nil, fmt.Errorf("eino single NewChatModelAgent: %w", err)
 	}
 
-	baseMsgs := historyToMessages(history)
+	baseMsgs := historyToMessages(history, appCfg, &ma.EinoMiddleware)
 	baseMsgs = append(baseMsgs, schema.UserMessage(userMessage))
 
 	streamsMainAssistant := func(agent string) bool {

internal/multiagent/eino_summarize.go

@@ -3,6 +3,8 @@ package multiagent
 import (
 	"context"
 	"fmt"
+	"os"
+	"path/filepath"
 	"strings"
 
 	"cyberstrike-ai/internal/agent"
@@ -32,6 +34,8 @@ func newEinoSummarizationMiddleware(
 	ctx context.Context,
 	summaryModel model.BaseChatModel,
 	appCfg *config.Config,
+	mwCfg *config.MultiAgentEinoMiddlewareConfig,
+	conversationID string,
 	logger *zap.Logger,
 ) (adk.ChatModelAgentMiddleware, error) {
 	if summaryModel == nil || appCfg == nil {
@@ -41,7 +45,14 @@ func newEinoSummarizationMiddleware(
 	if maxTotal <= 0 {
 		maxTotal = 120000
 	}
-	trigger := int(float64(maxTotal) * 0.9)
+	triggerRatio := 0.8
+	emitInternalEvents := true
+	if mwCfg != nil {
+		triggerRatio = mwCfg.SummarizationTriggerRatioEffective()
+		emitInternalEvents = mwCfg.SummarizationEmitInternalEventsEffective()
+	}
+	// Keep enough safety margin for tokenizer/model-side accounting mismatch.
+	trigger := int(float64(maxTotal) * triggerRatio)
 	if trigger < 4096 {
 		trigger = maxTotal
 		if trigger < 4096 {
@@ -65,6 +76,18 @@ func newEinoSummarizationMiddleware(
 	if recentTrailMax > trigger/2 {
 		recentTrailMax = trigger / 2
 	}
+	transcriptPath := ""
+	if conv := strings.TrimSpace(conversationID); conv != "" {
+		baseRoot := filepath.Join(os.TempDir(), "cyberstrike-summarization")
+		if dbPath := strings.TrimSpace(appCfg.Database.Path); dbPath != "" {
+			// Persist with the same lifecycle as local conversation storage.
+			baseRoot = filepath.Join(filepath.Dir(dbPath), "conversation_artifacts", sanitizeEinoPathSegment(conv), "summarization")
+		}
+		base := baseRoot
+		if mkErr := os.MkdirAll(base, 0o755); mkErr == nil {
+			transcriptPath = filepath.Join(base, "transcript.txt")
+		}
+	}
 
 	mw, err := summarization.New(ctx, &summarization.Config{
 		Model: summaryModel,
@@ -73,7 +96,8 @@ func newEinoSummarizationMiddleware(
 		},
 		TokenCounter:       tokenCounter,
 		UserInstruction:    einoSummarizeUserInstruction,
-		EmitInternalEvents: false,
+		EmitInternalEvents: emitInternalEvents,
+		TranscriptFilePath: transcriptPath,
 		PreserveUserMessages: &summarization.PreserveUserMessages{
 			Enabled:   true,
 			MaxTokens: preserveMax,
@@ -85,11 +109,16 @@ func newEinoSummarizationMiddleware(
 			if logger == nil {
 				return nil
 			}
+			beforeTokens, _ := tokenCounter(ctx, &summarization.TokenCounterInput{Messages: before.Messages})
+			afterTokens, _ := tokenCounter(ctx, &summarization.TokenCounterInput{Messages: after.Messages})
 			logger.Info("eino summarization 已压缩上下文",
 				zap.Int("messages_before", len(before.Messages)),
 				zap.Int("messages_after", len(after.Messages)),
+				zap.Int("tokens_before_estimated", beforeTokens),
+				zap.Int("tokens_after_estimated", afterTokens),
 				zap.Int("max_total_tokens", maxTotal),
 				zap.Int("trigger_context_tokens", trigger),
+				zap.String("transcript_file", transcriptPath),
 			)
 			return nil
 		},

internal/multiagent/eino_tool_name_injection.go

@@ -0,0 +1,73 @@
+package multiagent
+
+import (
+	"context"
+	"strings"
+
+	"github.com/cloudwego/eino/components/tool"
+)
+
+// injectToolNamesOnlyInstruction prepends a compact tool-name-only section into
+// the system instruction so the model can reference current callable names.
+func injectToolNamesOnlyInstruction(ctx context.Context, instruction string, tools []tool.BaseTool) string {
+	names := collectToolNames(ctx, tools)
+	if len(names) == 0 {
+		return strings.TrimSpace(instruction)
+	}
+	hasToolSearch := false
+	for _, n := range names {
+		if strings.EqualFold(strings.TrimSpace(n), "tool_search") {
+			hasToolSearch = true
+			break
+		}
+	}
+
+	var sb strings.Builder
+	sb.WriteString("以下是当前会话中可调用的工具名称列表（仅名称，无参数定义）：\n")
+	for _, name := range names {
+		sb.WriteString("- ")
+		sb.WriteString(name)
+		sb.WriteByte('\n')
+	}
+	sb.WriteString("\n使用规则：\n")
+	sb.WriteString("1) 上述仅为名称列表，不包含参数定义。\n")
+	if hasToolSearch {
+		sb.WriteString("2) 在调用具体工具前，应先使用 tool_search 查看工具详情与参数要求，再发起调用。\n")
+	} else {
+		sb.WriteString("2) 调用具体工具前，请先确认该工具的参数要求；不确定时先澄清再调用。\n")
+	}
+	sb.WriteString("3) 不要臆造不存在的工具名。\n\n")
+	if s := strings.TrimSpace(instruction); s != "" {
+		sb.WriteString(s)
+	}
+	return sb.String()
+}
+
+func collectToolNames(ctx context.Context, tools []tool.BaseTool) []string {
+	if len(tools) == 0 {
+		return nil
+	}
+	seen := make(map[string]struct{}, len(tools))
+	out := make([]string, 0, len(tools))
+	for _, t := range tools {
+		if t == nil {
+			continue
+		}
+		info, err := t.Info(ctx)
+		if err != nil || info == nil {
+			continue
+		}
+		name := strings.TrimSpace(info.Name)
+		if name == "" {
+			continue
+		}
+		key := strings.ToLower(name)
+		if _, ok := seen[key]; ok {
+			continue
+		}
+		seen[key] = struct{}{}
+		out = append(out, name)
+	}
+	return out
+}
+

internal/multiagent/plan_execute_executor.go

@@ -71,7 +71,7 @@ func planExecuteDefaultGenExecutorInput(ctx context.Context, in *planexecute.Exe
 	return planexecute.ExecutorPrompt.Format(ctx, map[string]any{
 		"input":          planExecuteFormatInput(in.UserInput),
 		"plan":           string(planContent),
-		"executed_steps": planExecuteFormatExecutedSteps(in.ExecutedSteps),
+		"executed_steps": planExecuteFormatExecutedSteps(in.ExecutedSteps, nil, nil),
 		"step":           in.Plan.FirstStep(),
 	})
 }

internal/multiagent/plan_execute_steps_cap.go

@@ -5,6 +5,8 @@ import (
 	"strings"
 	"unicode/utf8"
 
+	"cyberstrike-ai/internal/config"
+
 	"github.com/cloudwego/eino/adk/prebuilt/planexecute"
 )
 
@@ -12,8 +14,11 @@ import (
 // 此处仅约束「写入模型 prompt 的视图」，不修改 Eino session 中的原始 ExecutedSteps。
 
 const (
-	planExecuteMaxStepResultRunes = 12000
-	planExecuteKeepLastSteps      = 16
+	defaultPlanExecuteMaxStepResultRunes = 4000
+	defaultPlanExecuteKeepLastSteps      = 8
+	// Backward-compatible aliases for tests and existing references.
+	planExecuteMaxStepResultRunes = defaultPlanExecuteMaxStepResultRunes
+	planExecuteKeepLastSteps      = defaultPlanExecuteKeepLastSteps
 )
 
 func truncateRunesWithSuffix(s string, maxRunes int, suffix string) string {
@@ -29,16 +34,26 @@ func truncateRunesWithSuffix(s string, maxRunes int, suffix string) string {
 
 // capPlanExecuteExecutedSteps 折叠较早步骤、截断单步过长结果，供 prompt 使用。
 func capPlanExecuteExecutedSteps(steps []planexecute.ExecutedStep) []planexecute.ExecutedStep {
+	return capPlanExecuteExecutedStepsWithConfig(steps, nil)
+}
+
+func capPlanExecuteExecutedStepsWithConfig(steps []planexecute.ExecutedStep, mwCfg *config.MultiAgentEinoMiddlewareConfig) []planexecute.ExecutedStep {
 	if len(steps) == 0 {
 		return steps
 	}
+	maxStepResultRunes := defaultPlanExecuteMaxStepResultRunes
+	keepLastSteps := defaultPlanExecuteKeepLastSteps
+	if mwCfg != nil {
+		maxStepResultRunes = mwCfg.PlanExecuteMaxStepResultRunesEffective()
+		keepLastSteps = mwCfg.PlanExecuteKeepLastStepsEffective()
+	}
 	out := make([]planexecute.ExecutedStep, 0, len(steps)+1)
 	start := 0
-	if len(steps) > planExecuteKeepLastSteps {
-		start = len(steps) - planExecuteKeepLastSteps
+	if len(steps) > keepLastSteps {
+		start = len(steps) - keepLastSteps
 		var b strings.Builder
 		b.WriteString(fmt.Sprintf("（上文已完成 %d 步；此处仅保留步骤标题以节省上下文，完整输出已省略。后续 %d 步仍保留正文。）\n",
-			start, planExecuteKeepLastSteps))
+			start, keepLastSteps))
 		for i := 0; i < start; i++ {
 			b.WriteString(fmt.Sprintf("- %s\n", steps[i].Step))
 		}
@@ -50,8 +65,8 @@ func capPlanExecuteExecutedSteps(steps []planexecute.ExecutedStep) []planexecute
 	suffix := "\n…[step result truncated]"
 	for i := start; i < len(steps); i++ {
 		e := steps[i]
-		if utf8.RuneCountInString(e.Result) > planExecuteMaxStepResultRunes {
-			e.Result = truncateRunesWithSuffix(e.Result, planExecuteMaxStepResultRunes, suffix)
+		if utf8.RuneCountInString(e.Result) > maxStepResultRunes {
+			e.Result = truncateRunesWithSuffix(e.Result, maxStepResultRunes, suffix)
 		}
 		out = append(out, e)
 	}

internal/multiagent/runner.go

@@ -237,7 +237,7 @@ func RunDeepAgent(
 				subMax = subDefaultIter
 			}
 
-			subSumMw, err := newEinoSummarizationMiddleware(ctx, subModel, appCfg, logger)
+			subSumMw, err := newEinoSummarizationMiddleware(ctx, subModel, appCfg, &ma.EinoMiddleware, conversationID, logger)
 			if err != nil {
 				return nil, fmt.Errorf("子代理 %q summarization 中间件: %w", id, err)
 			}
@@ -257,11 +257,33 @@ func RunDeepAgent(
 				subHandlers = append(subHandlers, einoSkillMW)
 			}
 			subHandlers = append(subHandlers, subSumMw)
+			if teleMw := newEinoModelInputTelemetryMiddleware(logger, appCfg.OpenAI.Model, conversationID, "sub_agent"); teleMw != nil {
+				subHandlers = append(subHandlers, teleMw)
+			}
 
+			subInstrFinal := injectToolNamesOnlyInstruction(ctx, instr, subTools)
+			if logger != nil {
+				subNames := collectToolNames(ctx, subTools)
+				mountedNames := collectToolNames(ctx, subToolsForCfg)
+				hasToolSearch := false
+				for _, n := range subNames {
+					if strings.EqualFold(strings.TrimSpace(n), "tool_search") {
+						hasToolSearch = true
+						break
+					}
+				}
+				logger.Info("eino tool-name injection",
+					zap.String("scope", "sub_agent"),
+					zap.String("agent", id),
+					zap.Int("tool_names", len(subNames)),
+					zap.Int("mounted_tool_names", len(mountedNames)),
+					zap.Bool("has_tool_search", hasToolSearch),
+				)
+			}
 			sa, err := adk.NewChatModelAgent(ctx, &adk.ChatModelAgentConfig{
 				Name:        id,
 				Description: desc,
-				Instruction: instr,
+				Instruction: subInstrFinal,
 				Model:       subModel,
 				ToolsConfig: adk.ToolsConfig{
 					ToolsNodeConfig: compose.ToolsNodeConfig{
@@ -289,7 +311,7 @@ func RunDeepAgent(
 		return nil, fmt.Errorf("多代理主模型: %w", err)
 	}
 
-	mainSumMw, err := newEinoSummarizationMiddleware(ctx, mainModel, appCfg, logger)
+	mainSumMw, err := newEinoSummarizationMiddleware(ctx, mainModel, appCfg, &ma.EinoMiddleware, conversationID, logger)
 	if err != nil {
 		return nil, fmt.Errorf("多代理主 summarization 中间件: %w", err)
 	}
@@ -313,6 +335,25 @@ func RunDeepAgent(
 			orchDescription = d
 		}
 	}
+	orchInstruction = injectToolNamesOnlyInstruction(ctx, orchInstruction, mainTools)
+	if logger != nil {
+		mainNames := collectToolNames(ctx, mainTools)
+		mountedNames := collectToolNames(ctx, mainToolsForCfg)
+		hasToolSearch := false
+		for _, n := range mainNames {
+			if strings.EqualFold(strings.TrimSpace(n), "tool_search") {
+				hasToolSearch = true
+				break
+			}
+		}
+		logger.Info("eino tool-name injection",
+			zap.String("scope", "orchestrator"),
+			zap.String("orchestration", orchMode),
+			zap.Int("tool_names", len(mainNames)),
+			zap.Int("mounted_tool_names", len(mountedNames)),
+			zap.Bool("has_tool_search", hasToolSearch),
+		)
+	}
 
 	supInstr := strings.TrimSpace(orchInstruction)
 	if orchMode == "supervisor" {
@@ -352,6 +393,9 @@ func RunDeepAgent(
 		deepHandlers = append(deepHandlers, einoSkillMW)
 	}
 	deepHandlers = append(deepHandlers, mainSumMw)
+	if teleMw := newEinoModelInputTelemetryMiddleware(logger, appCfg.OpenAI.Model, conversationID, "deep_orchestrator"); teleMw != nil {
+		deepHandlers = append(deepHandlers, teleMw)
+	}
 
 	supHandlers := []adk.ChatModelAgentMiddleware{}
 	if len(mainOrchestratorPre) > 0 {
@@ -361,6 +405,9 @@ func RunDeepAgent(
 		supHandlers = append(supHandlers, einoSkillMW)
 	}
 	supHandlers = append(supHandlers, mainSumMw)
+	if teleMw := newEinoModelInputTelemetryMiddleware(logger, appCfg.OpenAI.Model, conversationID, "supervisor_orchestrator"); teleMw != nil {
+		supHandlers = append(supHandlers, teleMw)
+	}
 
 	mainToolsCfg := adk.ToolsConfig{
 		ToolsNodeConfig: compose.ToolsNodeConfig{
@@ -399,10 +446,17 @@ func RunDeepAgent(
 			ExecMaxIter:          deepMaxIter,
 			LoopMaxIter:          ma.PlanExecuteLoopMaxIterations,
 			AppCfg:               appCfg,
+			MwCfg:                &ma.EinoMiddleware,
+			ConversationID:       conversationID,
 			Logger:               logger,
+			ModelName:            appCfg.OpenAI.Model,
 			ExecPreMiddlewares:   mainOrchestratorPre,
 			SkillMiddleware:      einoSkillMW,
 			FilesystemMiddleware: peFsMw,
+			PlannerReplannerRewriteHandlers: []adk.ChatModelAgentMiddleware{
+				mainSumMw,
+				newEinoModelInputTelemetryMiddleware(logger, appCfg.OpenAI.Model, conversationID, "plan_execute_planner_replanner_rewrite"),
+			},
 		})
 		if perr != nil {
 			return nil, perr
@@ -468,7 +522,7 @@ func RunDeepAgent(
 		da = dDeep
 	}
 
-	baseMsgs := historyToMessages(history)
+	baseMsgs := historyToMessages(history, appCfg, &ma.EinoMiddleware)
 	baseMsgs = append(baseMsgs, schema.UserMessage(userMessage))
 
 	streamsMainAssistant := func(agent string) bool {
@@ -505,34 +559,77 @@ func RunDeepAgent(
 	}, baseMsgs)
 }
 
-func historyToMessages(history []agent.ChatMessage) []adk.Message {
+func historyToMessages(history []agent.ChatMessage, appCfg *config.Config, mwCfg *config.MultiAgentEinoMiddlewareConfig) []adk.Message {
 	if len(history) == 0 {
 		return nil
 	}
-	// 放宽条数上限：跨轮历史交给 Eino Summarization（阈值对齐 openai.max_total_tokens）在调用模型前压缩，避免在入队前硬截断为 40 条。
-	const maxHistoryMessages = 300
+	// Keep a bounded tail first; then enforce a token budget.
+	const maxHistoryMessages = 200
 	start := 0
 	if len(history) > maxHistoryMessages {
 		start = len(history) - maxHistoryMessages
 	}
-	out := make([]adk.Message, 0, len(history[start:]))
+	raw := make([]adk.Message, 0, len(history[start:]))
 	for _, h := range history[start:] {
 		switch h.Role {
 		case "user":
 			if strings.TrimSpace(h.Content) != "" {
-				out = append(out, schema.UserMessage(h.Content))
+				raw = append(raw, schema.UserMessage(h.Content))
 			}
 		case "assistant":
 			if strings.TrimSpace(h.Content) == "" && len(h.ToolCalls) > 0 {
 				continue
 			}
 			if strings.TrimSpace(h.Content) != "" {
-				out = append(out, schema.AssistantMessage(h.Content, nil))
+				raw = append(raw, schema.AssistantMessage(h.Content, nil))
 			}
 		default:
 			continue
 		}
 	}
+	if len(raw) == 0 {
+		return raw
+	}
+	maxTotal := 120000
+	modelName := "gpt-4o"
+	if appCfg != nil {
+		if appCfg.OpenAI.MaxTotalTokens > 0 {
+			maxTotal = appCfg.OpenAI.MaxTotalTokens
+		}
+		if m := strings.TrimSpace(appCfg.OpenAI.Model); m != "" {
+			modelName = m
+		}
+	}
+	ratio := 0.35
+	if mwCfg != nil {
+		ratio = mwCfg.HistoryInputBudgetRatioEffective()
+	}
+	budget := int(float64(maxTotal) * ratio)
+	if budget < 4096 {
+		budget = 4096
+	}
+	tc := agent.NewTikTokenCounter()
+	outRev := make([]adk.Message, 0, len(raw))
+	used := 0
+	for i := len(raw) - 1; i >= 0; i-- {
+		msg := raw[i]
+		n, err := tc.Count(modelName, string(msg.Role)+"\n"+msg.Content)
+		if err != nil {
+			n = (len(msg.Content) + 3) / 4
+		}
+		if n <= 0 {
+			n = 1
+		}
+		if used+n > budget {
+			break
+		}
+		used += n
+		outRev = append(outRev, msg)
+	}
+	out := make([]adk.Message, 0, len(outRev))
+	for i := len(outRev) - 1; i >= 0; i-- {
+		out = append(out, outRev[i])
+	}
 	return out
 }
 

web/static/css/style.css

@@ -4472,6 +4472,25 @@ header {
     flex-wrap: wrap;
 }
 
+.tool-resident-toggle {
+    display: inline-flex;
+    align-items: center;
+    gap: 4px;
+    font-size: 0.75rem;
+    color: var(--text-secondary);
+    border: 1px solid var(--border-color);
+    border-radius: 12px;
+    padding: 2px 8px;
+    background: var(--bg-secondary);
+    cursor: pointer;
+}
+
+.tool-resident-toggle input[type="checkbox"] {
+    width: 14px;
+    height: 14px;
+    margin: 0;
+}
+
 .external-tool-badge {
     display: inline-flex;
     align-items: center;

web/static/i18n/en-US.json

@@ -576,6 +576,10 @@
     "addExternal": "Add external MCP",
     "toolConfig": "MCP tool config",
     "saveToolConfig": "Save tool config",
+    "alwaysVisibleLabel": "Pinned",
+    "alwaysVisibleHint": "Always keep visible in Tool Search results",
+    "alwaysVisibleBuiltinLabel": "Builtin default",
+    "alwaysVisibleBuiltinHint": "Backend builtin tool is pinned by default and cannot be disabled",
     "externalConfig": "External MCP config",
     "loadingTools": "Loading tools...",
     "loadToolsTimeout": "Tools load timeout. External MCP may be slow. Click Refresh to retry or check connection.",

web/static/i18n/zh-CN.json

@@ -576,6 +576,10 @@
     "addExternal": "添加外部MCP",
     "toolConfig": "MCP 工具配置",
     "saveToolConfig": "保存工具配置",
+    "alwaysVisibleLabel": "常驻",
+    "alwaysVisibleHint": "始终常驻在 Tool Search 可见列表（不被 tool_search 隐藏）",
+    "alwaysVisibleBuiltinLabel": "内置默认",
+    "alwaysVisibleBuiltinHint": "后端内置工具默认常驻，不可关闭",
     "externalConfig": "外部 MCP 配置",
     "loadingTools": "正在加载工具列表...",
     "loadToolsTimeout": "加载工具列表超时，可能是外部MCP连接较慢。请点击\"刷新\"按钮重试，或检查外部MCP连接状态。",

web/static/js/router.js

@@ -301,26 +301,38 @@ async function initPage(pageId) {
             break;
         case 'mcp-management':
             // 初始化MCP管理
+            const startLoadMcpTools = () => {
+                // 加载工具列表（MCP工具配置已移到MCP管理页面）
+                // 使用异步加载，避免阻塞页面渲染
+                if (typeof loadToolsList === 'function') {
+                    // 确保工具分页设置已初始化
+                    if (typeof getToolsPageSize === 'function' && typeof toolsPagination !== 'undefined') {
+                        toolsPagination.pageSize = getToolsPageSize();
+                    }
+                    // 延迟加载，让页面先渲染
+                    setTimeout(() => {
+                        loadToolsList(1, '').catch(err => {
+                            console.error('加载工具列表失败:', err);
+                        });
+                    }, 100);
+                }
+            };
+            // 先拉取全局配置，确保 tool_search 常驻状态按后端生效集合展示
+            if (typeof loadConfig === 'function') {
+                loadConfig(false)
+                    .catch(err => {
+                        console.warn('加载配置失败（将继续加载工具列表）:', err);
+                    })
+                    .finally(startLoadMcpTools);
+            } else {
+                startLoadMcpTools();
+            }
             // 先加载外部MCP列表（快速），然后加载工具列表
             if (typeof loadExternalMCPs === 'function') {
                 loadExternalMCPs().catch(err => {
                     console.warn('加载外部MCP列表失败:', err);
                 });
             }
-            // 加载工具列表（MCP工具配置已移到MCP管理页面）
-            // 使用异步加载，避免阻塞页面渲染
-            if (typeof loadToolsList === 'function') {
-                // 确保工具分页设置已初始化
-                if (typeof getToolsPageSize === 'function' && typeof toolsPagination !== 'undefined') {
-                    toolsPagination.pageSize = getToolsPageSize();
-                }
-                // 延迟加载，让页面先渲染
-                setTimeout(() => {
-                    loadToolsList(1, '').catch(err => {
-                        console.error('加载工具列表失败:', err);
-                    });
-                }, 100);
-            }
             break;
         case 'vulnerabilities':
             // 初始化漏洞管理页面

web/static/js/settings.js

@@ -1,6 +1,8 @@
 // 设置相关功能
 let currentConfig = null;
 let allTools = [];
+let alwaysVisibleToolNames = new Set();
+let alwaysVisibleBuiltinToolNames = new Set();
 // 全局工具状态映射，用于保存用户在所有页面的修改
 // key: 唯一工具标识符（toolKey），value: { enabled: boolean, is_external: boolean, external_mcp: string }
 let toolStateMap = new Map();
@@ -100,6 +102,14 @@ async function loadConfig(loadTools = true) {
         }
         
         currentConfig = await response.json();
+        const alwaysVisibleList = currentConfig?.multi_agent?.tool_search_always_visible_effective_tools;
+        const alwaysVisibleConfigured = currentConfig?.multi_agent?.tool_search_always_visible_tools;
+        alwaysVisibleToolNames = new Set(Array.isArray(alwaysVisibleList) ? alwaysVisibleList.filter(Boolean) : []);
+        alwaysVisibleBuiltinToolNames = new Set(
+            alwaysVisibleToolNames.size > 0 && Array.isArray(alwaysVisibleConfigured)
+                ? Array.from(alwaysVisibleToolNames).filter(name => !alwaysVisibleConfigured.includes(name))
+                : []
+        );
         
         // 填充OpenAI配置
         const providerEl = document.getElementById('openai-provider');
@@ -498,6 +508,8 @@ function renderToolsList() {
             is_external: tool.is_external || false,
             external_mcp: tool.external_mcp || ''
         };
+        const alwaysVisibleChecked = alwaysVisibleToolNames.has(tool.name);
+        const alwaysVisibleLocked = alwaysVisibleBuiltinToolNames.has(tool.name);
         
         // 外部工具标签，显示来源信息（可点击跳转到对应 MCP 卡片）
         let externalBadge = '';
@@ -521,6 +533,11 @@ function renderToolsList() {
                 <div class="tool-item-name">
                     ${escapeHtml(tool.name)}
                     ${externalBadge}
+                    <label class="tool-resident-toggle" title="${typeof window.t === 'function' ? window.t('mcp.alwaysVisibleHint') : '始终常驻在 Tool Search 可见列表'}" onclick="event.stopPropagation()">
+                        <input type="checkbox" ${alwaysVisibleChecked ? 'checked' : ''} ${alwaysVisibleLocked ? 'disabled' : ''} onchange="handleToolAlwaysVisibleChange('${escapeHtml(tool.name)}', this.checked)" />
+                        <span>${typeof window.t === 'function' ? window.t('mcp.alwaysVisibleLabel') : '常驻'}</span>
+                    </label>
+                    ${alwaysVisibleLocked ? `<span class="external-tool-badge" title="${typeof window.t === 'function' ? window.t('mcp.alwaysVisibleBuiltinHint') : '后端内置工具默认常驻，不可关闭'}">${typeof window.t === 'function' ? window.t('mcp.alwaysVisibleBuiltinLabel') : '内置默认'}</span>` : ''}
                     <span class="tool-expand-icon">▶</span>
                 </div>
                 <div class="tool-item-desc">${escapeHtml(tool.description || (typeof window.t === 'function' ? window.t('mcp.noDescription') : '无描述'))}</div>
@@ -716,6 +733,16 @@ function handleToolCheckboxChange(toolKey, enabled) {
     updateToolsStats();
 }
 
+function handleToolAlwaysVisibleChange(toolName, alwaysVisible) {
+    const name = (toolName || '').trim();
+    if (!name) return;
+    if (alwaysVisible) {
+        alwaysVisibleToolNames.add(name);
+    } else {
+        alwaysVisibleToolNames.delete(name);
+    }
+}
+
 // 全选工具
 function selectAllTools() {
     document.querySelectorAll('#tools-list input[type="checkbox"]').forEach(checkbox => {
@@ -886,9 +913,11 @@ async function updateToolsStats() {
     }
     
     const tStats = typeof window.t === 'function' ? window.t : (k) => k;
+    const pinnedCount = alwaysVisibleToolNames.size;
     statsEl.innerHTML = `
         <span title="${tStats('mcp.currentPageEnabled')}">✅ ${tStats('mcp.currentPageEnabled')}: <strong>${currentPageEnabled}</strong> / ${currentPageTotal}</span>
         <span title="${tStats('mcp.totalEnabled')}">📊 ${tStats('mcp.totalEnabled')}: <strong>${totalEnabled}</strong> / ${totalTools}</span>
+        <span title="${tStats('mcp.alwaysVisibleHint')}">📌 ${tStats('mcp.alwaysVisibleLabel')}: <strong>${pinnedCount}</strong></span>
     `;
 }
 
@@ -1230,6 +1259,13 @@ async function saveToolsConfig() {
         const config = {
             openai: currentConfig.openai || {},
             agent: currentConfig.agent || {},
+            multi_agent: {
+                enabled: currentConfig?.multi_agent?.enabled === true,
+                robot_use_multi_agent: currentConfig?.multi_agent?.robot_use_multi_agent === true,
+                batch_use_multi_agent: currentConfig?.multi_agent?.batch_use_multi_agent === true,
+                plan_execute_loop_max_iterations: Number(currentConfig?.multi_agent?.plan_execute_loop_max_iterations || 0),
+                tool_search_always_visible_tools: Array.from(alwaysVisibleToolNames).filter(name => !alwaysVisibleBuiltinToolNames.has(name))
+            },
             tools: []
         };