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: "v1.6.6"
+version: "v1.6.8"
 # 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
@@ -41,6 +41,13 @@ openai:
   api_key: sk-xxxxxxx # API 密钥（必填）
   model: qwen3-max # 模型名称（必填）
   max_total_tokens: 120000 # LLM 相关上下文的最大 Token 数限制（内存压缩和攻击链构建会共用此配置）
+  # Eino 路径模型推理：DeepSeek/OpenAI 为 thinking / reasoning_effort 等；provider 为 claude 时合并为 Anthropic 顶层 thinking（extended thinking），mode: off 关闭
+  reasoning:
+    mode: off # auto | on | off；off 时不附加任何推理扩展字段
+    effort: max # low | medium | high | max；空表示不指定（openai_compat 下 auto 且无强度时不发请求扩展）
+    allow_client_reasoning: true # false 时忽略对话请求体 reasoning，仅以下方为准
+    profile: openai_compat # auto | deepseek_compat | openai_compat | output_config_effort
+    # extra_request_fields: {} # 可选：管理员自定义根级 JSON 片段（高级）
 # ============================================
 # 信息收集（FOFA）配置（可选）
 # ============================================

internal/agent/agent.go

@@ -193,6 +193,10 @@ type ChatMessage struct {
 	Content    string     `json:"content,omitempty"`
 	ToolCalls  []ToolCall `json:"tool_calls,omitempty"`
 	ToolCallID string     `json:"tool_call_id,omitempty"`
+	// ToolName 仅 tool 角色：从 Eino/轨迹 JSON 的 name 或 tool_name 恢复，供续跑构造 ToolMessage。
+	ToolName string `json:"tool_name,omitempty"`
+	// ReasoningContent 对应 OpenAI/DeepSeek 的 reasoning_content；思考模式 + 工具调用后续跑须回传（见 DeepSeek 文档）。
+	ReasoningContent string `json:"reasoning_content,omitempty"`
 }
 
 // MarshalJSON 自定义JSON序列化，将tool_calls中的arguments转换为JSON字符串
@@ -206,11 +210,17 @@ func (cm ChatMessage) MarshalJSON() ([]byte, error) {
 	if cm.Content != "" {
 		aux["content"] = cm.Content
 	}
+	if cm.ReasoningContent != "" {
+		aux["reasoning_content"] = cm.ReasoningContent
+	}
 
 	// 添加tool_call_id（如果存在）
 	if cm.ToolCallID != "" {
 		aux["tool_call_id"] = cm.ToolCallID
 	}
+	if cm.ToolName != "" {
+		aux["tool_name"] = cm.ToolName
+	}
 
 	// 转换tool_calls，将arguments转换为JSON字符串
 	if len(cm.ToolCalls) > 0 {
@@ -438,6 +448,7 @@ func (a *Agent) AgentLoopWithProgress(ctx context.Context, userInput string, his
 				Content:    msg.Content,
 				ToolCalls:  msg.ToolCalls,
 				ToolCallID: msg.ToolCallID,
+				ToolName:   msg.ToolName,
 			})
 			addedCount++
 			contentPreview := msg.Content
@@ -657,8 +668,8 @@ func (a *Agent) AgentLoopWithProgress(ctx context.Context, userInput string, his
 
 		// 检查是否有工具调用
 		if len(choice.Message.ToolCalls) > 0 {
-			// 思考内容：如果本轮启用了思考流式增量（thinking_stream_*），前端会去重；
-			// 同时也需要在该“思考阶段结束”时补一条可落库的 thinking（用于刷新后持久化展示）。
+			// ReAct 助手正文流式增量（thinking_stream_*）在 UI 上归为「思考」；若与 streamId 重复则前端会去重。
+			// 该条 thinking 用于刷新后持久化展示（与流式聚合一致）。
 			if choice.Message.Content != "" {
 				sendProgress("thinking", choice.Message.Content, map[string]interface{}{
 					"iteration": i + 1,

internal/attackchain/builder.go

@@ -301,7 +301,7 @@ func (b *Builder) formatProcessDetailsForAttackChain(details []database.ProcessD
 		// 目标：以主 agent（编排器）视角输出整轮迭代
 		// - 保留：编排器工具调用/结果、对子代理的 task 调度、子代理最终回复（不含推理）
 		// - 丢弃：thinking/planning/progress 等噪声、子代理的工具细节与推理过程
-		if d.EventType == "progress" || d.EventType == "thinking" || d.EventType == "planning" {
+		if d.EventType == "progress" || d.EventType == "thinking" || d.EventType == "reasoning_chain" || d.EventType == "planning" {
 			continue
 		}
 

internal/config/config.go

@@ -90,7 +90,8 @@ type MultiAgentEinoMiddlewareConfig struct {
 	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 已不影响 Eino：从 last_react 轨迹转 ADK 消息时**不再**按 token 比例裁剪（完整注入）。
+	// 字段仍保留，便于旧版 config 不报错；新部署可省略。
 	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"`
@@ -345,6 +346,48 @@ type OpenAIConfig struct {
 	BaseURL        string `yaml:"base_url" json:"base_url"`
 	Model          string `yaml:"model" json:"model"`
 	MaxTotalTokens int    `yaml:"max_total_tokens,omitempty" json:"max_total_tokens,omitempty"`
+	// Reasoning 控制 Eino ChatModel 的 thinking / reasoning_effort / output_config 等（仅 Eino 路径生效；原生 ReAct 忽略）。
+	Reasoning OpenAIReasoningConfig `yaml:"reasoning,omitempty" json:"reasoning,omitempty"`
+}
+
+// OpenAIReasoningConfig 全局默认与网关 profile（对话页可通过 ChatRequest.reasoning 覆盖，受 AllowClientReasoning 约束）。
+type OpenAIReasoningConfig struct {
+	// Mode: auto（默认）| on | off | default（与 auto 相同）。off 时不向模型附加推理扩展字段。
+	Mode string `yaml:"mode,omitempty" json:"mode,omitempty"`
+	// Effort: low | medium | high | max；空表示不单独指定强度（各 profile 行为见 internal/reasoning）。
+	Effort string `yaml:"effort,omitempty" json:"effort,omitempty"`
+	// AllowClientReasoning 为 false 时忽略请求体 reasoning；nil 或未设置等同于 true。
+	AllowClientReasoning *bool `yaml:"allow_client_reasoning,omitempty" json:"allow_client_reasoning,omitempty"`
+	// Profile: auto | deepseek_compat | openai_compat | output_config_effort
+	Profile string `yaml:"profile,omitempty" json:"profile,omitempty"`
+	// ExtraRequestFields 合并进 Chat Completions 根 JSON（管理员用；与自动字段同名时后者覆盖）。
+	ExtraRequestFields map[string]interface{} `yaml:"extra_request_fields,omitempty" json:"extra_request_fields,omitempty"`
+}
+
+// ModeEffective returns auto when empty or default.
+func (c OpenAIReasoningConfig) ModeEffective() string {
+	m := strings.ToLower(strings.TrimSpace(c.Mode))
+	if m == "" || m == "default" {
+		return "auto"
+	}
+	return m
+}
+
+// ProfileEffective returns auto when empty.
+func (c OpenAIReasoningConfig) ProfileEffective() string {
+	p := strings.ToLower(strings.TrimSpace(c.Profile))
+	if p == "" {
+		return "auto"
+	}
+	return p
+}
+
+// AllowClientReasoningEffective true when client may send ChatRequest.reasoning.
+func (c OpenAIReasoningConfig) AllowClientReasoningEffective() bool {
+	if c.AllowClientReasoning == nil {
+		return true
+	}
+	return *c.AllowClientReasoning
 }
 
 type FofaConfig struct {

internal/database/conversation.go

@@ -25,14 +25,15 @@ type Conversation struct {
 
 // Message 消息
 type Message struct {
-	ID              string                   `json:"id"`
-	ConversationID  string                   `json:"conversationId"`
-	Role            string                   `json:"role"`
-	Content         string                   `json:"content"`
-	MCPExecutionIDs []string                 `json:"mcpExecutionIds,omitempty"`
-	ProcessDetails  []map[string]interface{} `json:"processDetails,omitempty"`
-	CreatedAt       time.Time                `json:"createdAt"`
-	UpdatedAt       time.Time                `json:"updatedAt"`
+	ID               string                   `json:"id"`
+	ConversationID string                   `json:"conversationId"`
+	Role             string                   `json:"role"`
+	Content          string                   `json:"content"`
+	ReasoningContent string                   `json:"reasoningContent,omitempty"`
+	MCPExecutionIDs  []string                 `json:"mcpExecutionIds,omitempty"`
+	ProcessDetails   []map[string]interface{} `json:"processDetails,omitempty"`
+	CreatedAt        time.Time                `json:"createdAt"`
+	UpdatedAt        time.Time                `json:"updatedAt"`
 }
 
 // CreateConversation 创建新对话
@@ -498,8 +499,8 @@ func (db *DB) AddMessage(conversationID, role, content string, mcpExecutionIDs [
 	}
 
 	_, err := db.Exec(
-		"INSERT INTO messages (id, conversation_id, role, content, mcp_execution_ids, created_at, updated_at) VALUES (?, ?, ?, ?, ?, ?, ?)",
-		id, conversationID, role, content, mcpIDsJSON, now, now,
+		"INSERT INTO messages (id, conversation_id, role, content, reasoning_content, mcp_execution_ids, created_at, updated_at) VALUES (?, ?, ?, ?, ?, ?, ?, ?)",
+		id, conversationID, role, content, "", mcpIDsJSON, now, now,
 	)
 	if err != nil {
 		return nil, fmt.Errorf("添加消息失败: %w", err)
@@ -523,10 +524,30 @@ func (db *DB) AddMessage(conversationID, role, content string, mcpExecutionIDs [
 	return message, nil
 }
 
+// UpdateAssistantMessageFinalize 更新助手消息终态（正文、MCP id、思考链聚合文本，供无轨迹回退时回放）。
+func (db *DB) UpdateAssistantMessageFinalize(messageID, content string, mcpExecutionIDs []string, reasoningContent string) error {
+	var mcpIDsJSON string
+	if len(mcpExecutionIDs) > 0 {
+		jsonData, err := json.Marshal(mcpExecutionIDs)
+		if err != nil {
+			return fmt.Errorf("序列化MCP执行ID失败: %w", err)
+		}
+		mcpIDsJSON = string(jsonData)
+	}
+	_, err := db.Exec(
+		"UPDATE messages SET content = ?, mcp_execution_ids = ?, reasoning_content = ?, updated_at = ? WHERE id = ?",
+		content, mcpIDsJSON, strings.TrimSpace(reasoningContent), time.Now(), messageID,
+	)
+	if err != nil {
+		return fmt.Errorf("更新助手消息失败: %w", err)
+	}
+	return nil
+}
+
 // GetMessages 获取对话的所有消息
 func (db *DB) GetMessages(conversationID string) ([]Message, error) {
 	rows, err := db.Query(
-		"SELECT id, conversation_id, role, content, mcp_execution_ids, created_at, updated_at FROM messages WHERE conversation_id = ? ORDER BY created_at ASC",
+		"SELECT id, conversation_id, role, content, reasoning_content, mcp_execution_ids, created_at, updated_at FROM messages WHERE conversation_id = ? ORDER BY created_at ASC",
 		conversationID,
 	)
 	if err != nil {
@@ -537,13 +558,17 @@ func (db *DB) GetMessages(conversationID string) ([]Message, error) {
 	var messages []Message
 	for rows.Next() {
 		var msg Message
+		var reasoning sql.NullString
 		var mcpIDsJSON sql.NullString
 		var createdAt string
 		var updatedAt sql.NullString
 
-		if err := rows.Scan(&msg.ID, &msg.ConversationID, &msg.Role, &msg.Content, &mcpIDsJSON, &createdAt, &updatedAt); err != nil {
+		if err := rows.Scan(&msg.ID, &msg.ConversationID, &msg.Role, &msg.Content, &reasoning, &mcpIDsJSON, &createdAt, &updatedAt); err != nil {
 			return nil, fmt.Errorf("扫描消息失败: %w", err)
 		}
+		if reasoning.Valid {
+			msg.ReasoningContent = reasoning.String
+		}
 
 		// 尝试多种时间格式解析
 		var err error
@@ -683,7 +708,7 @@ type ProcessDetail struct {
 	ID             string    `json:"id"`
 	MessageID      string    `json:"messageId"`
 	ConversationID string    `json:"conversationId"`
-	EventType      string    `json:"eventType"` // iteration, thinking, tool_calls_detected, tool_call, tool_result, progress, error
+	EventType      string    `json:"eventType"` // iteration, thinking, reasoning_chain, tool_calls_detected, tool_call, tool_result, progress, error
 	Message        string    `json:"message"`
 	Data           string    `json:"data"` // JSON格式的数据
 	CreatedAt      time.Time `json:"createdAt"`

internal/database/database.go

@@ -594,6 +594,25 @@ func (db *DB) migrateMessagesTable() error {
 
 	// 回填已有数据：让 updated_at 至少等于 created_at，避免前端出现空/当前时间回退。
 	_, _ = db.Exec("UPDATE messages SET updated_at = created_at WHERE updated_at IS NULL OR updated_at = ''")
+
+	// reasoning_content：DeepSeek 思考模式 + 工具调用续跑；与 last_react_input 互补，供消息表回退路径回放
+	var rcColCount int
+	errRC := db.QueryRow("SELECT COUNT(*) FROM pragma_table_info('messages') WHERE name='reasoning_content'").Scan(&rcColCount)
+	if errRC != nil {
+		if _, addErr := db.Exec("ALTER TABLE messages ADD COLUMN reasoning_content TEXT"); addErr != nil {
+			errMsg := strings.ToLower(addErr.Error())
+			if !strings.Contains(errMsg, "duplicate column") && !strings.Contains(errMsg, "already exists") {
+				return fmt.Errorf("添加 messages.reasoning_content 字段失败: %w", addErr)
+			}
+		}
+	} else if rcColCount == 0 {
+		if _, err := db.Exec("ALTER TABLE messages ADD COLUMN reasoning_content TEXT"); err != nil {
+			errMsg := strings.ToLower(err.Error())
+			if !strings.Contains(errMsg, "duplicate column") && !strings.Contains(errMsg, "already exists") {
+				return fmt.Errorf("添加 messages.reasoning_content 字段失败: %w", err)
+			}
+		}
+	}
 	return nil
 }
 

internal/handler/agent.go

@@ -19,6 +19,7 @@ import (
 	"cyberstrike-ai/internal/agent"
 	"cyberstrike-ai/internal/config"
 	"cyberstrike-ai/internal/database"
+	"cyberstrike-ai/internal/reasoning"
 	"cyberstrike-ai/internal/mcp"
 	"cyberstrike-ai/internal/mcp/builtin"
 	"cyberstrike-ai/internal/multiagent"
@@ -201,6 +202,14 @@ type ChatAttachment struct {
 	ServerPath string `json:"serverPath,omitempty"` // 已保存在 chat_uploads 下的绝对路径（由 POST /api/chat-uploads 返回）
 }
 
+// ChatReasoningRequest 对话页「模型推理」意图（仅 Eino 路径消费；原生 agent-loop 忽略）。
+type ChatReasoningRequest struct {
+	// Mode: default（跟随系统）| off | on | auto
+	Mode string `json:"mode,omitempty"`
+	// Effort: low | medium | high | max；空表示不指定（由系统默认与各 profile 决定）。
+	Effort string `json:"effort,omitempty"`
+}
+
 // ChatRequest 聊天请求
 type ChatRequest struct {
 	Message              string           `json:"message" binding:"required"`
@@ -209,10 +218,18 @@ type ChatRequest struct {
 	Attachments          []ChatAttachment `json:"attachments,omitempty"`
 	WebShellConnectionID string           `json:"webshellConnectionId,omitempty"` // WebShell 管理 - AI 助手：当前选中的连接 ID，仅使用 webshell_* 工具
 	Hitl                 *HITLRequest     `json:"hitl,omitempty"`
+	Reasoning            *ChatReasoningRequest `json:"reasoning,omitempty"`
 	// Orchestration 仅对 /api/multi-agent、/api/multi-agent/stream：deep | plan_execute | supervisor；空则等同 deep。机器人/批量等无请求体时由服务端默认 deep。/api/eino-agent* 不使用此字段。
 	Orchestration string `json:"orchestration,omitempty"`
 }
 
+func chatReasoningToClientIntent(r *ChatReasoningRequest) *reasoning.ClientIntent {
+	if r == nil {
+		return nil
+	}
+	return &reasoning.ClientIntent{Mode: r.Mode, Effort: r.Effort}
+}
+
 type HITLRequest struct {
 	Enabled        bool     `json:"enabled"`
 	Mode           string   `json:"mode,omitempty"`
@@ -567,14 +584,7 @@ func (h *AgentHandler) AgentLoop(c *gin.Context) {
 			h.logger.Warn("获取历史消息失败", zap.Error(err))
 			agentHistoryMessages = []agent.ChatMessage{}
 		} else {
-			// 将数据库消息转换为Agent消息格式
-			agentHistoryMessages = make([]agent.ChatMessage, 0, len(historyMessages))
-			for _, msg := range historyMessages {
-				agentHistoryMessages = append(agentHistoryMessages, agent.ChatMessage{
-					Role:    msg.Role,
-					Content: msg.Content,
-				})
-			}
+			agentHistoryMessages = dbMessagesToAgentChatMessages(historyMessages)
 			h.logger.Info("从消息表加载历史消息", zap.Int("count", len(agentHistoryMessages)))
 		}
 	} else {
@@ -775,6 +785,7 @@ func (h *AgentHandler) ProcessMessageForRobot(ctx context.Context, conversationI
 			progressCallback,
 			h.agentsMarkdownDir,
 			"deep",
+			nil,
 		)
 		if errMA != nil {
 			if shouldPersistEinoAgentTraceAfterRunError(ctx) {
@@ -788,17 +799,8 @@ func (h *AgentHandler) ProcessMessageForRobot(ctx context.Context, conversationI
 			return "", conversationID, errMA
 		}
 		if assistantMessageID != "" {
-			mcpIDsJSON := ""
-			if len(resultMA.MCPExecutionIDs) > 0 {
-				jsonData, _ := json.Marshal(resultMA.MCPExecutionIDs)
-				mcpIDsJSON = string(jsonData)
-			}
-			_, err = h.db.Exec(
-				"UPDATE messages SET content = ?, mcp_execution_ids = ?, updated_at = ? WHERE id = ?",
-				resultMA.Response, mcpIDsJSON, time.Now(), assistantMessageID,
-			)
-			if err != nil {
-				h.logger.Warn("机器人：更新助手消息失败", zap.Error(err))
+			if errU := h.db.UpdateAssistantMessageFinalize(assistantMessageID, resultMA.Response, resultMA.MCPExecutionIDs, multiagent.AggregatedReasoningFromTraceJSON(resultMA.LastAgentTraceInput)); errU != nil {
+				h.logger.Warn("机器人：更新助手消息失败", zap.Error(errU))
 			}
 		} else {
 			if _, err = h.db.AddMessage(conversationID, "assistant", resultMA.Response, resultMA.MCPExecutionIDs); err != nil {
@@ -823,17 +825,8 @@ func (h *AgentHandler) ProcessMessageForRobot(ctx context.Context, conversationI
 
 	// 更新助手消息内容与 MCP 执行 ID（与 stream 一致）
 	if assistantMessageID != "" {
-		mcpIDsJSON := ""
-		if len(result.MCPExecutionIDs) > 0 {
-			jsonData, _ := json.Marshal(result.MCPExecutionIDs)
-			mcpIDsJSON = string(jsonData)
-		}
-		_, err = h.db.Exec(
-		"UPDATE messages SET content = ?, mcp_execution_ids = ?, updated_at = ? WHERE id = ?",
-		result.Response, mcpIDsJSON, time.Now(), assistantMessageID,
-		)
-		if err != nil {
-			h.logger.Warn("机器人：更新助手消息失败", zap.Error(err))
+		if errU := h.db.UpdateAssistantMessageFinalize(assistantMessageID, result.Response, result.MCPExecutionIDs, multiagent.AggregatedReasoningFromTraceJSON(result.LastAgentTraceInput)); errU != nil {
+			h.logger.Warn("机器人：更新助手消息失败", zap.Error(errU))
 		}
 	} else {
 		if _, err = h.db.AddMessage(conversationID, "assistant", result.Response, result.MCPExecutionIDs); err != nil {
@@ -891,10 +884,12 @@ func (h *AgentHandler) createProgressCallback(runCtx context.Context, cancelRun
 		return ""
 	}
 
-	// thinking_stream_*：不逐条落库，按 streamId 聚合，在后续关键事件前补一条可持久化的 thinking
+	// thinking_stream_*（ReAct 等助手正文流）与 reasoning_chain_stream_*（Eino ReasoningContent）：
+	// 不逐条落库，按 streamId 聚合，flush 时分别落 thinking / reasoning_chain。
 	type thinkingBuf struct {
-		b    strings.Builder
-		meta map[string]interface{}
+		b         strings.Builder
+		meta      map[string]interface{}
+		persistAs string // "thinking" | "reasoning_chain"
 	}
 	thinkingStreams := make(map[string]*thinkingBuf) // streamId -> buf
 	flushedThinking := make(map[string]bool)         // streamId -> flushed
@@ -948,8 +943,12 @@ func (h *AgentHandler) createProgressCallback(runCtx context.Context, cancelRun
 				}
 				data[k] = v
 			}
-			if err := h.db.AddProcessDetail(assistantMessageID, conversationID, "thinking", content, data); err != nil {
-				h.logger.Warn("保存过程详情失败", zap.Error(err), zap.String("eventType", "thinking"))
+			persist := tb.persistAs
+			if persist != "reasoning_chain" {
+				persist = "thinking"
+			}
+			if err := h.db.AddProcessDetail(assistantMessageID, conversationID, persist, content, data); err != nil {
+				h.logger.Warn("保存过程详情失败", zap.Error(err), zap.String("eventType", persist))
 			}
 			flushedThinking[sid] = true
 		}
@@ -1177,14 +1176,20 @@ func (h *AgentHandler) createProgressCallback(runCtx context.Context, cancelRun
 			return
 		}
 
-		// 聚合 thinking_stream_*（ReasoningContent），不逐条落库
-		if eventType == "thinking_stream_start" {
+		// 聚合 thinking_stream_* / reasoning_chain_stream_*，不逐条落库
+		if eventType == "thinking_stream_start" || eventType == "reasoning_chain_stream_start" {
+			persistAs := "thinking"
+			if eventType == "reasoning_chain_stream_start" {
+				persistAs = "reasoning_chain"
+			}
 			if dataMap, ok := data.(map[string]interface{}); ok {
 				if sid, ok2 := dataMap["streamId"].(string); ok2 && sid != "" {
 					tb := thinkingStreams[sid]
 					if tb == nil {
-						tb = &thinkingBuf{meta: map[string]interface{}{}}
+						tb = &thinkingBuf{meta: map[string]interface{}{}, persistAs: persistAs}
 						thinkingStreams[sid] = tb
+					} else {
+						tb.persistAs = persistAs
 					}
 					// 记录元信息（source/einoAgent/einoRole/iteration 等）
 					for k, v := range dataMap {
@@ -1194,15 +1199,21 @@ func (h *AgentHandler) createProgressCallback(runCtx context.Context, cancelRun
 			}
 			return
 		}
-		if eventType == "thinking_stream_delta" {
+		if eventType == "thinking_stream_delta" || eventType == "reasoning_chain_stream_delta" {
+			persistAs := "thinking"
+			if eventType == "reasoning_chain_stream_delta" {
+				persistAs = "reasoning_chain"
+			}
 			if dataMap, ok := data.(map[string]interface{}); ok {
 				if sid, ok2 := dataMap["streamId"].(string); ok2 && sid != "" {
 					tb := thinkingStreams[sid]
 					if tb == nil {
-						tb = &thinkingBuf{meta: map[string]interface{}{}}
+						tb = &thinkingBuf{meta: map[string]interface{}{}, persistAs: persistAs}
 						thinkingStreams[sid] = tb
+					} else if tb.persistAs == "" {
+						tb.persistAs = persistAs
 					}
-					// delta 片段直接拼接；message 本身就是 reasoning content
+					// delta 片段直接拼接
 					tb.b.WriteString(message)
 					// 有时 delta 先到 start 未到，补充元信息
 					for k, v := range dataMap {
@@ -1213,10 +1224,9 @@ func (h *AgentHandler) createProgressCallback(runCtx context.Context, cancelRun
 			return
 		}
 
-		// 当 Agent 同时发送 thinking_stream_* 和 thinking（带同一 streamId）时，
-		// thinking_stream_* 已经会在 flushThinkingStreams() 聚合落库；
-		// 这里跳过同 streamId 的 thinking，避免 processDetails 双份展示。
-		if eventType == "thinking" {
+		// 当 Agent 同时发送 *_stream_* 与同名 streamId 的 thinking/reasoning_chain 时，
+		// 流式聚合已会在 flushThinkingStreams() 落库；此处跳过逐条重复。
+		if eventType == "thinking" || eventType == "reasoning_chain" {
 			if dataMap, ok := data.(map[string]interface{}); ok {
 				if sid, ok2 := dataMap["streamId"].(string); ok2 && sid != "" {
 					if tb, exists := thinkingStreams[sid]; exists && tb != nil {
@@ -1245,7 +1255,7 @@ func (h *AgentHandler) createProgressCallback(runCtx context.Context, cancelRun
 			if eventType == "tool_result" {
 				discardPlanningIfEchoesToolResult(&respPlan, data)
 			}
-			// 在关键过程事件落库前，先把「规划中」与 thinking_stream 落库
+			// 在关键过程事件落库前，先把「规划中」与聚合中的 thinking / reasoning_chain 流落库
 			flushResponsePlan()
 			flushThinkingStreams()
 			if err := h.db.AddProcessDetail(assistantMessageID, conversationID, eventType, message, data); err != nil {
@@ -1427,14 +1437,7 @@ func (h *AgentHandler) AgentLoopStream(c *gin.Context) {
 			h.logger.Warn("获取历史消息失败", zap.Error(err))
 			agentHistoryMessages = []agent.ChatMessage{}
 		} else {
-			// 将数据库消息转换为Agent消息格式
-			agentHistoryMessages = make([]agent.ChatMessage, 0, len(historyMessages))
-			for _, msg := range historyMessages {
-				agentHistoryMessages = append(agentHistoryMessages, agent.ChatMessage{
-					Role:    msg.Role,
-					Content: msg.Content,
-				})
-			}
+			agentHistoryMessages = dbMessagesToAgentChatMessages(historyMessages)
 			h.logger.Info("从消息表加载历史消息", zap.Int("count", len(agentHistoryMessages)))
 		}
 	} else {
@@ -1727,20 +1730,8 @@ func (h *AgentHandler) AgentLoopStream(c *gin.Context) {
 
 	// 更新助手消息内容
 	if assistantMsg != nil {
-		_, err = h.db.Exec(
-			"UPDATE messages SET content = ?, mcp_execution_ids = ?, updated_at = ? WHERE id = ?",
-			result.Response,
-			func() string {
-				if len(result.MCPExecutionIDs) > 0 {
-					jsonData, _ := json.Marshal(result.MCPExecutionIDs)
-					return string(jsonData)
-				}
-				return ""
-			}(),
-			time.Now(), assistantMessageID,
-		)
-		if err != nil {
-			h.logger.Error("更新助手消息失败", zap.Error(err))
+		if errU := h.db.UpdateAssistantMessageFinalize(assistantMessageID, result.Response, result.MCPExecutionIDs, multiagent.AggregatedReasoningFromTraceJSON(result.LastAgentTraceInput)); errU != nil {
+			h.logger.Error("更新助手消息失败", zap.Error(errU))
 		}
 	} else {
 		// 如果之前创建失败，现在创建
@@ -1789,27 +1780,51 @@ func (h *AgentHandler) CancelAgentLoop(c *gin.Context) {
 			return
 		}
 		execID := h.tasks.ActiveMCPExecutionID(req.ConversationID)
-		if execID == "" {
-			c.JSON(http.StatusBadRequest, gin.H{"error": "当前没有正在执行的 MCP 工具（例如模型尚在推理、尚未发起工具调用）。请等待工具开始执行后再试，或使用「彻底停止」结束整轮任务。"})
-			return
-		}
 		note := strings.TrimSpace(req.Reason)
-		if !h.agent.CancelMCPToolExecutionWithNote(execID, note) {
-			c.JSON(http.StatusNotFound, gin.H{"error": "未找到进行中的工具执行或该调用已结束"})
+		if execID != "" {
+			if !h.agent.CancelMCPToolExecutionWithNote(execID, note) {
+				c.JSON(http.StatusNotFound, gin.H{"error": "未找到进行中的工具执行或该调用已结束"})
+				return
+			}
+			h.logger.Info("对话页仅终止当前 MCP 工具",
+				zap.String("conversationId", req.ConversationID),
+				zap.String("executionId", execID),
+				zap.Bool("hasNote", note != ""),
+			)
+			c.JSON(http.StatusOK, gin.H{
+				"status":              "tool_abort_requested",
+				"conversationId":      req.ConversationID,
+				"executionId":         execID,
+				"message":             "已请求终止当前工具调用；工具返回后本轮推理将继续（与 MCP 监控页终止一致）。",
+				"continueAfter":       true,
+				"interruptWithNote":   note != "",
+				"continueWithoutTool": false,
+			})
 			return
 		}
-		h.logger.Info("对话页仅终止当前 MCP 工具",
+		// 无进行中的 MCP 工具（模型纯推理/流式输出阶段）：取消当前上下文并由 Eino 流式处理器合并用户补充后自动续跑。
+		h.tasks.SetInterruptContinueNote(req.ConversationID, note)
+		ok, err := h.tasks.CancelTask(req.ConversationID, multiagent.ErrInterruptContinue)
+		if err != nil {
+			h.logger.Error("中断并继续（无工具）失败", zap.Error(err))
+			c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+			return
+		}
+		if !ok {
+			c.JSON(http.StatusNotFound, gin.H{"error": "未找到正在执行的任务"})
+			return
+		}
+		h.logger.Info("对话页中断并继续（无 MCP 工具，将自动续跑）",
 			zap.String("conversationId", req.ConversationID),
-			zap.String("executionId", execID),
 			zap.Bool("hasNote", note != ""),
 		)
 		c.JSON(http.StatusOK, gin.H{
-			"status":            "tool_abort_requested",
-			"conversationId":    req.ConversationID,
-			"executionId":       execID,
-			"message":           "已请求终止当前工具调用；工具返回后本轮推理将继续（与 MCP 监控页终止一致）。",
-			"continueAfter":     true,
-			"interruptWithNote": note != "",
+			"status":              "interrupt_continue_scheduled",
+			"conversationId":      req.ConversationID,
+			"message":             "已请求暂停当前推理；用户补充将合并到上下文并自动继续执行（无需整轮停止）。",
+			"continueAfter":       true,
+			"interruptWithNote":   note != "",
+			"continueWithoutTool": true,
 		})
 		return
 	}
@@ -2640,12 +2655,12 @@ func (h *AgentHandler) executeBatchQueue(queueID string) {
 			var runErr error
 			switch {
 			case useBatchMulti:
-				resultMA, runErr = multiagent.RunDeepAgent(taskCtx, h.config, &h.config.MultiAgent, h.agent, h.logger, conversationID, finalMessage, []agent.ChatMessage{}, roleTools, progressCallback, h.agentsMarkdownDir, batchOrch)
+				resultMA, runErr = multiagent.RunDeepAgent(taskCtx, h.config, &h.config.MultiAgent, h.agent, h.logger, conversationID, finalMessage, []agent.ChatMessage{}, roleTools, progressCallback, h.agentsMarkdownDir, batchOrch, nil)
 			case useEinoSingle:
 				if h.config == nil {
 					runErr = fmt.Errorf("服务器配置未加载")
 				} else {
-					resultMA, runErr = multiagent.RunEinoSingleChatModelAgent(taskCtx, h.config, &h.config.MultiAgent, h.agent, h.logger, conversationID, finalMessage, []agent.ChatMessage{}, roleTools, progressCallback)
+					resultMA, runErr = multiagent.RunEinoSingleChatModelAgent(taskCtx, h.config, &h.config.MultiAgent, h.agent, h.logger, conversationID, finalMessage, []agent.ChatMessage{}, roleTools, progressCallback, nil)
 				}
 			default:
 				result, runErr = h.agent.AgentLoopWithProgress(taskCtx, finalMessage, []agent.ChatMessage{}, conversationID, progressCallback, roleTools)
@@ -2744,17 +2759,7 @@ func (h *AgentHandler) executeBatchQueue(queueID string) {
 
 			// 更新助手消息内容
 			if assistantMessageID != "" {
-				mcpIDsJSON := ""
-				if len(mcpIDs) > 0 {
-					jsonData, _ := json.Marshal(mcpIDs)
-					mcpIDsJSON = string(jsonData)
-				}
-				if _, updateErr := h.db.Exec(
-					"UPDATE messages SET content = ?, mcp_execution_ids = ?, updated_at = ? WHERE id = ?",
-					resText,
-					mcpIDsJSON,
-					time.Now(), assistantMessageID,
-				); updateErr != nil {
+				if updateErr := h.db.UpdateAssistantMessageFinalize(assistantMessageID, resText, mcpIDs, multiagent.AggregatedReasoningFromTraceJSON(lastIn)); updateErr != nil {
 					h.logger.Warn("更新助手消息失败", zap.String("queueId", queueID), zap.String("taskId", task.ID), zap.Error(updateErr))
 					// 如果更新失败，尝试创建新消息
 					_, err = h.db.AddMessage(conversationID, "assistant", resText, mcpIDs)
@@ -2846,6 +2851,10 @@ func (h *AgentHandler) loadHistoryFromAgentTrace(conversationID string) ([]agent
 		if content, ok := msgMap["content"].(string); ok {
 			msg.Content = content
 		}
+		// DeepSeek 思考模式：含工具调用的 assistant 须在后续请求中回传 reasoning_content
+		if rc, ok := msgMap["reasoning_content"].(string); ok && strings.TrimSpace(rc) != "" {
+			msg.ReasoningContent = rc
+		}
 
 		// 解析tool_calls（如果存在）
 		if toolCallsRaw, ok := msgMap["tool_calls"]; ok && toolCallsRaw != nil {
@@ -2901,6 +2910,11 @@ func (h *AgentHandler) loadHistoryFromAgentTrace(conversationID string) ([]agent
 		if toolCallID, ok := msgMap["tool_call_id"].(string); ok {
 			msg.ToolCallID = toolCallID
 		}
+		if tn, ok := msgMap["tool_name"].(string); ok && strings.TrimSpace(tn) != "" {
+			msg.ToolName = strings.TrimSpace(tn)
+		} else if tn, ok := msgMap["name"].(string); ok && strings.TrimSpace(tn) != "" && strings.EqualFold(msg.Role, "tool") {
+			msg.ToolName = strings.TrimSpace(tn)
+		}
 
 		agentMessages = append(agentMessages, msg)
 	}
@@ -2946,3 +2960,18 @@ func (h *AgentHandler) loadHistoryFromAgentTrace(conversationID string) ([]agent
 	)
 	return agentMessages, nil
 }
+
+// dbMessagesToAgentChatMessages maps DB rows to agent ChatMessage for history fallback
+// (includes reasoning_content for DeepSeek thinking + tool replay).
+func dbMessagesToAgentChatMessages(msgs []database.Message) []agent.ChatMessage {
+	out := make([]agent.ChatMessage, 0, len(msgs))
+	for i := range msgs {
+		m := msgs[i]
+		out = append(out, agent.ChatMessage{
+			Role:             m.Role,
+			Content:          m.Content,
+			ReasoningContent: m.ReasoningContent,
+		})
+	}
+	return out
+}

internal/handler/config.go

@@ -1312,6 +1312,19 @@ func updateOpenAIConfig(doc *yaml.Node, cfg config.OpenAIConfig) {
 	if cfg.MaxTotalTokens > 0 {
 		setIntInMap(openaiNode, "max_total_tokens", cfg.MaxTotalTokens)
 	}
+	rn := ensureMap(openaiNode, "reasoning")
+	if strings.TrimSpace(cfg.Reasoning.Mode) != "" {
+		setStringInMap(rn, "mode", cfg.Reasoning.Mode)
+	}
+	if strings.TrimSpace(cfg.Reasoning.Effort) != "" {
+		setStringInMap(rn, "effort", cfg.Reasoning.Effort)
+	}
+	if cfg.Reasoning.AllowClientReasoning != nil {
+		setBoolInMap(rn, "allow_client_reasoning", *cfg.Reasoning.AllowClientReasoning)
+	}
+	if strings.TrimSpace(cfg.Reasoning.Profile) != "" {
+		setStringInMap(rn, "profile", cfg.Reasoning.Profile)
+	}
 }
 
 func updateFOFAConfig(doc *yaml.Node, cfg config.FofaConfig) {

internal/handler/eino_single_agent.go

@@ -46,7 +46,7 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 	sendEvent := func(eventType, message string, data interface{}) {
 		if eventType == "error" && baseCtx != nil {
 			cause := context.Cause(baseCtx)
-			if errors.Is(cause, ErrTaskCancelled) {
+			if errors.Is(cause, ErrTaskCancelled) || errors.Is(cause, multiagent.ErrInterruptContinue) {
 				return
 			}
 		}
@@ -175,29 +175,69 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 	}
 	taskOwned = true
 
-	progressCallback := h.createProgressCallback(taskCtx, cancelWithCause, conversationID, assistantMessageID, sendEvent)
-	taskCtx = mcp.WithMCPConversationID(taskCtx, conversationID)
-	taskCtx = mcp.WithToolRunRegistry(taskCtx, h.tasks)
-	taskCtx = multiagent.WithHITLToolInterceptor(taskCtx, func(ctx context.Context, toolName, arguments string) (string, error) {
-		return h.interceptHITLForEinoTool(ctx, cancelWithCause, conversationID, assistantMessageID, sendEvent, toolName, arguments)
-	})
+	var cumulativeMCPExecutionIDs []string
 
-	result, runErr = multiagent.RunEinoSingleChatModelAgent(
-		taskCtx,
-		h.config,
-		&h.config.MultiAgent,
-		h.agent,
-		h.logger,
-		conversationID,
-		curFinalMessage,
-		curHistory,
-		roleTools,
-		progressCallback,
-	)
-	timeoutCancel()
+	for {
+		progressCallback := h.createProgressCallback(taskCtx, cancelWithCause, conversationID, assistantMessageID, sendEvent)
+		taskCtxLoop := mcp.WithMCPConversationID(taskCtx, conversationID)
+		taskCtxLoop = mcp.WithToolRunRegistry(taskCtxLoop, h.tasks)
+		taskCtxLoop = multiagent.WithHITLToolInterceptor(taskCtxLoop, func(ctx context.Context, toolName, arguments string) (string, error) {
+			return h.interceptHITLForEinoTool(ctx, cancelWithCause, conversationID, assistantMessageID, sendEvent, toolName, arguments)
+		})
+
+		result, runErr = multiagent.RunEinoSingleChatModelAgent(
+			taskCtxLoop,
+			h.config,
+			&h.config.MultiAgent,
+			h.agent,
+			h.logger,
+			conversationID,
+			curFinalMessage,
+			curHistory,
+			roleTools,
+			progressCallback,
+			chatReasoningToClientIntent(req.Reasoning),
+		)
+		timeoutCancel()
+
+		if result != nil && len(result.MCPExecutionIDs) > 0 {
+			cumulativeMCPExecutionIDs = mergeMCPExecutionIDLists(cumulativeMCPExecutionIDs, result.MCPExecutionIDs)
+		}
+
+		if runErr == nil {
+			break
+		}
 
-	if runErr != nil {
 		cause := context.Cause(baseCtx)
+		if errors.Is(cause, multiagent.ErrInterruptContinue) {
+			if shouldPersistEinoAgentTraceAfterRunError(baseCtx) {
+				h.persistEinoAgentTraceForResume(conversationID, result)
+			}
+			note := h.tasks.TakeInterruptContinueNote(conversationID)
+			icSummary := interruptContinueTimelineSummary(note)
+			progressCallback("user_interrupt_continue", icSummary, map[string]interface{}{
+				"conversationId": conversationID,
+				"rawReason":      strings.TrimSpace(note),
+				"emptyReason":    strings.TrimSpace(note) == "",
+				"kind":           "no_active_mcp_tool",
+			})
+			inject := formatInterruptContinueUserMessage(note)
+			// 不写入 messages 表为 user 气泡：避免主对话流出现大段模板；说明已由 user_interrupt_continue 记入助手 process_details（迭代详情）。
+			if hist, err := h.loadHistoryFromAgentTrace(conversationID); err == nil && len(hist) > 0 {
+				curHistory = hist
+			}
+			curFinalMessage = inject
+			sendEvent("progress", "已合并用户补充与最新轨迹，正在继续推理…", map[string]interface{}{
+				"conversationId": conversationID,
+				"source":         "interrupt_continue",
+			})
+			h.tasks.UpdateTaskStatus(conversationID, "running")
+			baseCtx, cancelWithCause = context.WithCancelCause(context.Background())
+			h.tasks.BindTaskCancel(conversationID, cancelWithCause)
+			taskCtx, timeoutCancel = context.WithTimeout(baseCtx, 600*time.Minute)
+			continue
+		}
+
 		if shouldPersistEinoAgentTraceAfterRunError(baseCtx) {
 			h.persistEinoAgentTraceForResume(conversationID, result)
 		}
@@ -258,18 +298,7 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 	}
 
 	if assistantMessageID != "" {
-		mcpIDsJSON := ""
-		if len(result.MCPExecutionIDs) > 0 {
-			jsonData, _ := json.Marshal(result.MCPExecutionIDs)
-			mcpIDsJSON = string(jsonData)
-		}
-		_, _ = h.db.Exec(
-			"UPDATE messages SET content = ?, mcp_execution_ids = ?, updated_at = ? WHERE id = ?",
-			result.Response,
-			mcpIDsJSON,
-			time.Now(),
-			assistantMessageID,
-		)
+		_ = h.db.UpdateAssistantMessageFinalize(assistantMessageID, result.Response, cumulativeMCPExecutionIDs, multiagent.AggregatedReasoningFromTraceJSON(result.LastAgentTraceInput))
 	}
 
 	if result.LastAgentTraceInput != "" || result.LastAgentTraceOutput != "" {
@@ -279,7 +308,7 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 	}
 
 	sendEvent("response", result.Response, map[string]interface{}{
-		"mcpExecutionIds": result.MCPExecutionIDs,
+		"mcpExecutionIds": cumulativeMCPExecutionIDs,
 		"conversationId":  conversationID,
 		"messageId":       assistantMessageID,
 		"agentMode":       "eino_single",
@@ -337,6 +366,7 @@ func (h *AgentHandler) EinoSingleAgentLoop(c *gin.Context) {
 		prep.History,
 		prep.RoleTools,
 		progressCallback,
+		chatReasoningToClientIntent(req.Reasoning),
 	)
 	if runErr != nil {
 		if shouldPersistEinoAgentTraceAfterRunError(baseCtx) {
@@ -347,18 +377,7 @@ func (h *AgentHandler) EinoSingleAgentLoop(c *gin.Context) {
 	}
 
 	if prep.AssistantMessageID != "" {
-		mcpIDsJSON := ""
-		if len(result.MCPExecutionIDs) > 0 {
-			jsonData, _ := json.Marshal(result.MCPExecutionIDs)
-			mcpIDsJSON = string(jsonData)
-		}
-		_, _ = h.db.Exec(
-			"UPDATE messages SET content = ?, mcp_execution_ids = ?, updated_at = ? WHERE id = ?",
-			result.Response,
-			mcpIDsJSON,
-			time.Now(),
-			prep.AssistantMessageID,
-		)
+		_ = h.db.UpdateAssistantMessageFinalize(prep.AssistantMessageID, result.Response, result.MCPExecutionIDs, multiagent.AggregatedReasoningFromTraceJSON(result.LastAgentTraceInput))
 	}
 	if result.LastAgentTraceInput != "" || result.LastAgentTraceOutput != "" {
 		_ = h.db.SaveAgentTrace(prep.ConversationID, result.LastAgentTraceInput, result.LastAgentTraceOutput)

internal/handler/multi_agent.go

@@ -63,7 +63,7 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 		// 为避免 UI 看到“取消错误 + cancelled 文案”两条回复，这里直接丢弃取消对应的 error。
 		if eventType == "error" && baseCtx != nil {
 			cause := context.Cause(baseCtx)
-			if errors.Is(cause, ErrTaskCancelled) {
+			if errors.Is(cause, ErrTaskCancelled) || errors.Is(cause, multiagent.ErrInterruptContinue) {
 				return
 			}
 		}
@@ -184,31 +184,72 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 	}
 	taskOwned = true
 
-	progressCallback := h.createProgressCallback(taskCtx, cancelWithCause, conversationID, assistantMessageID, sendEvent)
-	taskCtx = mcp.WithMCPConversationID(taskCtx, conversationID)
-	taskCtx = mcp.WithToolRunRegistry(taskCtx, h.tasks)
-	taskCtx = multiagent.WithHITLToolInterceptor(taskCtx, func(ctx context.Context, toolName, arguments string) (string, error) {
-		return h.interceptHITLForEinoTool(ctx, cancelWithCause, conversationID, assistantMessageID, sendEvent, toolName, arguments)
-	})
+	// 同一 HTTP 流内多段 Run（如中断并继续）合并 MCP execution id，供最终 response / 库表与工具芯片展示完整列表
+	var cumulativeMCPExecutionIDs []string
 
-	result, runErr = multiagent.RunDeepAgent(
-		taskCtx,
-		h.config,
-		&h.config.MultiAgent,
-		h.agent,
-		h.logger,
-		conversationID,
-		curFinalMessage,
-		curHistory,
-		roleTools,
-		progressCallback,
-		h.agentsMarkdownDir,
-		orch,
-	)
-	timeoutCancel()
+	for {
+		progressCallback := h.createProgressCallback(taskCtx, cancelWithCause, conversationID, assistantMessageID, sendEvent)
+		taskCtxLoop := mcp.WithMCPConversationID(taskCtx, conversationID)
+		taskCtxLoop = mcp.WithToolRunRegistry(taskCtxLoop, h.tasks)
+		taskCtxLoop = multiagent.WithHITLToolInterceptor(taskCtxLoop, func(ctx context.Context, toolName, arguments string) (string, error) {
+			return h.interceptHITLForEinoTool(ctx, cancelWithCause, conversationID, assistantMessageID, sendEvent, toolName, arguments)
+		})
+
+		result, runErr = multiagent.RunDeepAgent(
+			taskCtxLoop,
+			h.config,
+			&h.config.MultiAgent,
+			h.agent,
+			h.logger,
+			conversationID,
+			curFinalMessage,
+			curHistory,
+			roleTools,
+			progressCallback,
+			h.agentsMarkdownDir,
+			orch,
+			chatReasoningToClientIntent(req.Reasoning),
+		)
+		timeoutCancel()
+
+		if result != nil && len(result.MCPExecutionIDs) > 0 {
+			cumulativeMCPExecutionIDs = mergeMCPExecutionIDLists(cumulativeMCPExecutionIDs, result.MCPExecutionIDs)
+		}
+
+		if runErr == nil {
+			break
+		}
 
-	if runErr != nil {
 		cause := context.Cause(baseCtx)
+		if errors.Is(cause, multiagent.ErrInterruptContinue) {
+			if shouldPersistEinoAgentTraceAfterRunError(baseCtx) {
+				h.persistEinoAgentTraceForResume(conversationID, result)
+			}
+			note := h.tasks.TakeInterruptContinueNote(conversationID)
+			icSummary := interruptContinueTimelineSummary(note)
+			progressCallback("user_interrupt_continue", icSummary, map[string]interface{}{
+				"conversationId": conversationID,
+				"rawReason":      strings.TrimSpace(note),
+				"emptyReason":    strings.TrimSpace(note) == "",
+				"kind":           "no_active_mcp_tool",
+			})
+			inject := formatInterruptContinueUserMessage(note)
+			// 不写入 messages 表为 user 气泡：避免主对话流出现大段模板；说明已由 user_interrupt_continue 记入助手 process_details（迭代详情）。
+			if hist, err := h.loadHistoryFromAgentTrace(conversationID); err == nil && len(hist) > 0 {
+				curHistory = hist
+			}
+			curFinalMessage = inject
+			sendEvent("progress", "已合并用户补充与最新轨迹，正在继续推理…", map[string]interface{}{
+				"conversationId": conversationID,
+				"source":         "interrupt_continue",
+			})
+			h.tasks.UpdateTaskStatus(conversationID, "running")
+			baseCtx, cancelWithCause = context.WithCancelCause(context.Background())
+			h.tasks.BindTaskCancel(conversationID, cancelWithCause)
+			taskCtx, timeoutCancel = context.WithTimeout(baseCtx, 600*time.Minute)
+			continue
+		}
+
 		if shouldPersistEinoAgentTraceAfterRunError(baseCtx) {
 			h.persistEinoAgentTraceForResume(conversationID, result)
 		}
@@ -269,18 +310,7 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 	}
 
 	if assistantMessageID != "" {
-		mcpIDsJSON := ""
-		if len(result.MCPExecutionIDs) > 0 {
-			jsonData, _ := json.Marshal(result.MCPExecutionIDs)
-			mcpIDsJSON = string(jsonData)
-		}
-		_, _ = h.db.Exec(
-			"UPDATE messages SET content = ?, mcp_execution_ids = ?, updated_at = ? WHERE id = ?",
-			result.Response,
-			mcpIDsJSON,
-			time.Now(),
-			assistantMessageID,
-		)
+		_ = h.db.UpdateAssistantMessageFinalize(assistantMessageID, result.Response, cumulativeMCPExecutionIDs, multiagent.AggregatedReasoningFromTraceJSON(result.LastAgentTraceInput))
 	}
 
 	if result.LastAgentTraceInput != "" || result.LastAgentTraceOutput != "" {
@@ -294,7 +324,7 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 		effectiveOrch = config.NormalizeMultiAgentOrchestration(o)
 	}
 	sendEvent("response", result.Response, map[string]interface{}{
-		"mcpExecutionIds": result.MCPExecutionIDs,
+		"mcpExecutionIds": cumulativeMCPExecutionIDs,
 		"conversationId":  conversationID,
 		"messageId":       assistantMessageID,
 		"agentMode":       "eino_" + effectiveOrch,
@@ -350,6 +380,7 @@ func (h *AgentHandler) MultiAgentLoop(c *gin.Context) {
 		progressCallback,
 		h.agentsMarkdownDir,
 		strings.TrimSpace(req.Orchestration),
+		chatReasoningToClientIntent(req.Reasoning),
 	)
 	if runErr != nil {
 		if shouldPersistEinoAgentTraceAfterRunError(baseCtx) {
@@ -365,18 +396,7 @@ func (h *AgentHandler) MultiAgentLoop(c *gin.Context) {
 	}
 
 	if prep.AssistantMessageID != "" {
-		mcpIDsJSON := ""
-		if len(result.MCPExecutionIDs) > 0 {
-			jsonData, _ := json.Marshal(result.MCPExecutionIDs)
-			mcpIDsJSON = string(jsonData)
-		}
-		_, _ = h.db.Exec(
-			"UPDATE messages SET content = ?, mcp_execution_ids = ?, updated_at = ? WHERE id = ?",
-			result.Response,
-			mcpIDsJSON,
-			time.Now(),
-			prep.AssistantMessageID,
-		)
+		_ = h.db.UpdateAssistantMessageFinalize(prep.AssistantMessageID, result.Response, result.MCPExecutionIDs, multiagent.AggregatedReasoningFromTraceJSON(result.LastAgentTraceInput))
 	}
 
 	if result.LastAgentTraceInput != "" || result.LastAgentTraceOutput != "" {
@@ -406,6 +426,52 @@ func (h *AgentHandler) persistEinoAgentTraceForResume(conversationID string, res
 	}
 }
 
+// mergeMCPExecutionIDLists 去重合并多段 Run 的 MCP execution id（顺序：先 dst 后 more）。
+func mergeMCPExecutionIDLists(dst []string, more []string) []string {
+	seen := make(map[string]struct{}, len(dst)+len(more))
+	out := make([]string, 0, len(dst)+len(more))
+	add := func(ids []string) {
+		for _, id := range ids {
+			id = strings.TrimSpace(id)
+			if id == "" {
+				continue
+			}
+			if _, ok := seen[id]; ok {
+				continue
+			}
+			seen[id] = struct{}{}
+			out = append(out, id)
+		}
+	}
+	add(dst)
+	add(more)
+	return out
+}
+
+// interruptContinueTimelineSummary 时间线 / process_details 中展示的简短正文（完整模板已写入另一条用户消息）。
+func interruptContinueTimelineSummary(note string) string {
+	note = strings.TrimSpace(note)
+	if note == "" {
+		return "用户选择「中断并继续」，未填写说明；已按默认渗透补充模板合并上下文并续跑。"
+	}
+	return "用户中断说明（原文）：\n\n" + note
+}
+
+// formatInterruptContinueUserMessage 将「中断并继续」弹窗中的说明格式化为新一轮 user 消息（渗透场景下强调路径补充与端口复扫）。
+func formatInterruptContinueUserMessage(note string) string {
+	var b strings.Builder
+	b.WriteString("【用户补充 / 中断后继续】\n")
+	if s := strings.TrimSpace(note); s != "" {
+		b.WriteString(s)
+		b.WriteString("\n\n")
+	}
+	b.WriteString("【请在本轮落实】\n")
+	b.WriteString("- 将用户提供的接口路径、参数、业务变化纳入后续测试与推理。\n")
+	b.WriteString("- 若资产或目标信息有更新，请对目标重新执行端口/服务探测，再基于新结果规划下一步。\n")
+	b.WriteString("- 在已有轨迹基础上推进，避免无意义重复已完成的步骤。\n")
+	return strings.TrimSpace(b.String())
+}
+
 func multiAgentHTTPErrorStatus(err error) (int, string) {
 	msg := err.Error()
 	switch {

internal/handler/multi_agent_prepare.go

@@ -55,13 +55,7 @@ func (h *AgentHandler) prepareMultiAgentSession(req *ChatRequest) (*multiAgentPr
 		if getErr != nil {
 			agentHistoryMessages = []agent.ChatMessage{}
 		} else {
-			agentHistoryMessages = make([]agent.ChatMessage, 0, len(historyMessages))
-			for _, msg := range historyMessages {
-				agentHistoryMessages = append(agentHistoryMessages, agent.ChatMessage{
-					Role:    msg.Role,
-					Content: msg.Content,
-				})
-			}
+			agentHistoryMessages = dbMessagesToAgentChatMessages(historyMessages)
 		}
 	}
 

internal/handler/task_manager.go

@@ -6,6 +6,8 @@ import (
 	"strings"
 	"sync"
 	"time"
+
+	"cyberstrike-ai/internal/multiagent"
 )
 
 // ErrTaskCancelled 用户取消任务的错误
@@ -32,6 +34,9 @@ type AgentTask struct {
 	// ActiveMCPExecutionID 当前正在执行的 MCP 工具 executionId（仅内存，供「中断并继续」= 仅掐当前工具）
 	ActiveMCPExecutionID string `json:"-"`
 
+	// InterruptContinueNote 无 MCP 时「中断并继续」由用户在弹窗中填写的补充说明（Cancel 前写入，续跑轮次读取后清空）
+	InterruptContinueNote string `json:"-"`
+
 	cancel func(error)
 }
 
@@ -65,6 +70,50 @@ func (m *AgentTaskManager) UnregisterRunningTool(conversationID, executionID str
 	}
 }
 
+// SetInterruptContinueNote 在发起 ErrInterruptContinue 取消前写入用户补充说明（仅内存）。
+func (m *AgentTaskManager) SetInterruptContinueNote(conversationID, note string) {
+	conversationID = strings.TrimSpace(conversationID)
+	if conversationID == "" {
+		return
+	}
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	if t, ok := m.tasks[conversationID]; ok && t != nil {
+		t.InterruptContinueNote = note
+	}
+}
+
+// TakeInterruptContinueNote 读取并清空补充说明（续跑开始时调用一次）。
+func (m *AgentTaskManager) TakeInterruptContinueNote(conversationID string) string {
+	conversationID = strings.TrimSpace(conversationID)
+	if conversationID == "" {
+		return ""
+	}
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	if t, ok := m.tasks[conversationID]; ok && t != nil {
+		n := t.InterruptContinueNote
+		t.InterruptContinueNote = ""
+		return n
+	}
+	return ""
+}
+
+// BindTaskCancel 在同一运行任务内替换与 context 绑定的 cancel 函数（用于中断后继续时换新 baseCtx）。
+func (m *AgentTaskManager) BindTaskCancel(conversationID string, cancel context.CancelCauseFunc) {
+	conversationID = strings.TrimSpace(conversationID)
+	if conversationID == "" || cancel == nil {
+		return
+	}
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	if t, ok := m.tasks[conversationID]; ok && t != nil {
+		t.cancel = func(err error) {
+			cancel(err)
+		}
+	}
+}
+
 // ActiveMCPExecutionID 返回当前会话进行中的工具 executionId，无则空串。
 func (m *AgentTaskManager) ActiveMCPExecutionID(conversationID string) string {
 	conversationID = strings.TrimSpace(conversationID)
@@ -210,8 +259,16 @@ func (m *AgentTaskManager) CancelTask(conversationID string, cause error) (bool,
 		return true, nil
 	}
 
-	task.Status = "cancelling"
-	task.CancellingAt = time.Now()
+	// ErrInterruptContinue：仅掐断当前推理步骤，随后由处理器续跑，不进入长时间「取消中」态。
+	if cause != nil && errors.Is(cause, multiagent.ErrInterruptContinue) {
+		task.Status = "running"
+	} else {
+		task.Status = "cancelling"
+		task.CancellingAt = time.Now()
+	}
+	if cause != nil && errors.Is(cause, ErrTaskCancelled) {
+		task.InterruptContinueNote = ""
+	}
 	cancel := task.cancel
 	m.mu.Unlock()
 

internal/multiagent/eino_adk_run_loop.go

@@ -13,7 +13,9 @@ import (
 	"sync/atomic"
 	"unicode/utf8"
 
+	"cyberstrike-ai/internal/agent"
 	"cyberstrike-ai/internal/einomcp"
+	"cyberstrike-ai/internal/openai"
 
 	"github.com/cloudwego/eino/adk"
 	"github.com/cloudwego/eino/schema"
@@ -40,6 +42,13 @@ func normalizeStreamingDelta(current, incoming string) (next, delta string) {
 	return current + incoming, incoming
 }
 
+func isInterruptContinue(ctx context.Context) bool {
+	if ctx == nil {
+		return false
+	}
+	return errors.Is(context.Cause(ctx), ErrInterruptContinue)
+}
+
 func isEinoIterationLimitError(err error) bool {
 	if err == nil {
 		return false
@@ -70,6 +79,11 @@ type einoADKRunLoopArgs struct {
 	McpIDsMu *sync.Mutex
 	McpIDs   *[]string
 
+	// FilesystemMonitorAgent / FilesystemMonitorRecord 非 nil 时，将 Eino ADK filesystem 中间件工具（ls/read_file/write_file/edit_file/glob/grep）
+	// 在完成时写入 MCP 监控；execute 仍由 eino_execute_monitor 记录，此处跳过。
+	FilesystemMonitorAgent  *agent.Agent
+	FilesystemMonitorRecord einomcp.ExecutionRecorder
+
 	// ToolInvokeNotify 与 einomcp.ToolsFromDefinitions 共享：run loop 在迭代前 Set，MCP 桥 Fire 以补全 tool_result。
 	ToolInvokeNotify *einomcp.ToolInvokeNotifyHolder
 
@@ -77,6 +91,10 @@ type einoADKRunLoopArgs struct {
 
 	// EmptyResponseMessage 当未捕获到助手正文时的占位（多代理与单代理文案不同）。
 	EmptyResponseMessage string
+
+	// ModelFacingTrace 可选：由各 ChatModelAgent Handlers 链末尾中间件写入「即将送入模型」的消息快照；
+	// 非空时优先用于 LastAgentTraceInput 序列化，使续跑与 summarization/reduction 后的上下文一致。
+	ModelFacingTrace *modelFacingTraceHolder
 }
 
 func runEinoADKAgentLoop(ctx context.Context, args *einoADKRunLoopArgs, baseMsgs []adk.Message) (*RunResult, error) {
@@ -399,7 +417,8 @@ func runEinoADKAgentLoop(ctx context.Context, args *einoADKRunLoopArgs, baseMsgs
 		}
 		ids := snapshotMCPIDs()
 		return buildEinoRunResultFromAccumulated(
-			orchMode, runAccumulatedMsgs, lastAssistant, lastPlanExecuteExecutor, emptyHint, ids, true,
+			orchMode, runAccumulatedMsgs, persistTraceSource(args, runAccumulatedMsgs),
+			lastAssistant, lastPlanExecuteExecutor, emptyHint, ids, true,
 		), runErr
 	}
 
@@ -409,10 +428,18 @@ func runEinoADKAgentLoop(ctx context.Context, args *einoADKRunLoopArgs, baseMsgs
 		case <-ctx.Done():
 			flushAllPendingAsFailed(ctx.Err())
 			if progress != nil {
-				progress("error", "Request cancelled / 请求已取消", map[string]interface{}{
-					"conversationId": conversationID,
-					"source":         "eino",
-				})
+				if isInterruptContinue(ctx) {
+					progress("progress", "已暂停当前输出，正在合并用户补充并继续…", map[string]interface{}{
+						"conversationId": conversationID,
+						"source":         "eino",
+						"kind":           "interrupt_continue",
+					})
+				} else {
+					progress("error", "Request cancelled / 请求已取消", map[string]interface{}{
+						"conversationId": conversationID,
+						"source":         "eino",
+					})
+				}
 			}
 			return takePartial(ctx.Err())
 		default:
@@ -426,10 +453,18 @@ func runEinoADKAgentLoop(ctx context.Context, args *einoADKRunLoopArgs, baseMsgs
 			if ctxErr := ctx.Err(); ctxErr != nil {
 				flushAllPendingAsFailed(ctxErr)
 				if progress != nil {
-					progress("error", ctxErr.Error(), map[string]interface{}{
-						"conversationId": conversationID,
-						"source":         "eino",
-					})
+					if isInterruptContinue(ctx) {
+						progress("progress", "已暂停当前输出，正在合并用户补充并继续…", map[string]interface{}{
+							"conversationId": conversationID,
+							"source":         "eino",
+							"kind":           "interrupt_continue",
+						})
+					} else {
+						progress("error", ctxErr.Error(), map[string]interface{}{
+							"conversationId": conversationID,
+							"source":         "eino",
+						})
+					}
 				}
 				return takePartial(ctxErr)
 			}
@@ -516,104 +551,140 @@ func runEinoADKAgentLoop(ctx context.Context, args *einoADKRunLoopArgs, baseMsgs
 			var mainAssistantBuf string
 			var mainAssistDupTarget string // 非空表示本段主助手流需缓冲至 EOF，与 execute 输出比对去重
 			var reasoningBuf string
+			var prevReasoningDisplay string // UI 用：剥离 Claude 内部 signature 尾缀后的累计展示
 			var streamRecvErr error
+			type streamMsg struct {
+				chunk *schema.Message
+				err   error
+			}
+			recvCh := make(chan streamMsg, 8)
+			go func() {
+				defer close(recvCh)
+				for {
+					ch, rerr := mv.MessageStream.Recv()
+					recvCh <- streamMsg{chunk: ch, err: rerr}
+					if rerr != nil {
+						return
+					}
+				}
+			}()
+		streamRecvLoop:
 			for {
-				chunk, rerr := mv.MessageStream.Recv()
-				if rerr != nil {
-					if errors.Is(rerr, io.EOF) {
-						break
+				select {
+				case <-ctx.Done():
+					streamRecvErr = ctx.Err()
+					break streamRecvLoop
+				case sm, ok := <-recvCh:
+					if !ok {
+						break streamRecvLoop
 					}
-					if logger != nil {
-						logger.Warn("eino stream recv error, flushing incomplete stream",
-							zap.Error(rerr),
-							zap.String("agent", ev.AgentName),
-							zap.Int("toolFragments", len(toolStreamFragments)))
-					}
-					streamRecvErr = rerr
-					break
-				}
-				if chunk == nil {
-					continue
-				}
-				if progress != nil && strings.TrimSpace(chunk.ReasoningContent) != "" {
-					var reasoningDelta string
-					reasoningBuf, reasoningDelta = normalizeStreamingDelta(reasoningBuf, chunk.ReasoningContent)
-					if reasoningDelta != "" {
-						if reasoningStreamID == "" {
-							reasoningStreamID = fmt.Sprintf("eino-reasoning-%s-%d", conversationID, atomic.AddInt64(&reasoningStreamSeq, 1))
-							progress("thinking_stream_start", " ", map[string]interface{}{
-								"streamId":      reasoningStreamID,
-								"source":        "eino",
-								"einoAgent":     ev.AgentName,
-								"einoRole":      einoRoleTag(ev.AgentName),
-								"orchestration": orchMode,
-							})
+					chunk, rerr := sm.chunk, sm.err
+					if rerr != nil {
+						if errors.Is(rerr, io.EOF) {
+							break streamRecvLoop
 						}
-						progress("thinking_stream_delta", reasoningDelta, map[string]interface{}{
-							"streamId": reasoningStreamID,
-						})
+						if logger != nil {
+							logger.Warn("eino stream recv error, flushing incomplete stream",
+								zap.Error(rerr),
+								zap.String("agent", ev.AgentName),
+								zap.Int("toolFragments", len(toolStreamFragments)))
+						}
+						streamRecvErr = rerr
+						break streamRecvLoop
 					}
-				}
-				if chunk.Content != "" {
-					if progress != nil && streamsMainAssistant(ev.AgentName) {
-						var contentDelta string
-						mainAssistantBuf, contentDelta = normalizeStreamingDelta(mainAssistantBuf, chunk.Content)
-						if contentDelta != "" {
-							if mainAssistDupTarget == "" {
-								executeStdoutDupMu.Lock()
-								if pendingExecuteStdoutDup != "" {
-									mainAssistDupTarget = pendingExecuteStdoutDup
-								}
-								executeStdoutDupMu.Unlock()
-							}
-							if mainAssistDupTarget != "" {
-								// 已展示过 tool_result，缓冲全文；EOF 后与 execute 输出相同则不再发助手流
+					if chunk == nil {
+						continue
+					}
+					if progress != nil && strings.TrimSpace(chunk.ReasoningContent) != "" {
+						var reasoningDelta string
+						reasoningBuf, reasoningDelta = normalizeStreamingDelta(reasoningBuf, chunk.ReasoningContent)
+						if reasoningDelta != "" {
+							fullDisplay := openai.DisplayReasoningContent(reasoningBuf)
+							var displayDelta string
+							if strings.HasPrefix(fullDisplay, prevReasoningDisplay) {
+								displayDelta = fullDisplay[len(prevReasoningDisplay):]
 							} else {
-								if !streamHeaderSent {
-									progress("response_start", "", map[string]interface{}{
-										"conversationId":     conversationID,
-										"mcpExecutionIds":    snapshotMCPIDs(),
-										"messageGeneratedBy": "eino:" + ev.AgentName,
-										"einoRole":           "orchestrator",
-										"einoAgent":          ev.AgentName,
-										"orchestration":      orchMode,
-									})
-									streamHeaderSent = true
-								}
-								progress("response_delta", contentDelta, map[string]interface{}{
-									"conversationId":  conversationID,
-									"mcpExecutionIds": snapshotMCPIDs(),
-									"einoRole":        "orchestrator",
-									"einoAgent":       ev.AgentName,
-									"orchestration":   orchMode,
-								})
+								displayDelta = fullDisplay
 							}
-						}
-					} else if !streamsMainAssistant(ev.AgentName) {
-						var subDelta string
-						subAssistantBuf, subDelta = normalizeStreamingDelta(subAssistantBuf, chunk.Content)
-						if subDelta != "" {
-							if progress != nil {
-								if subReplyStreamID == "" {
-									subReplyStreamID = fmt.Sprintf("eino-sub-reply-%s-%d", conversationID, atomic.AddInt64(&einoSubReplyStreamSeq, 1))
-									progress("eino_agent_reply_stream_start", "", map[string]interface{}{
-										"streamId":       subReplyStreamID,
-										"einoAgent":      ev.AgentName,
-										"einoRole":       "sub",
-										"conversationId": conversationID,
-										"source":         "eino",
+							prevReasoningDisplay = fullDisplay
+							if displayDelta != "" {
+								if reasoningStreamID == "" {
+									reasoningStreamID = fmt.Sprintf("eino-reasoning-%s-%d", conversationID, atomic.AddInt64(&reasoningStreamSeq, 1))
+									progress("reasoning_chain_stream_start", " ", map[string]interface{}{
+										"streamId":      reasoningStreamID,
+										"source":        "eino",
+										"einoAgent":     ev.AgentName,
+										"einoRole":      einoRoleTag(ev.AgentName),
+										"orchestration": orchMode,
 									})
 								}
-								progress("eino_agent_reply_stream_delta", subDelta, map[string]interface{}{
-									"streamId":       subReplyStreamID,
-									"conversationId": conversationID,
+								progress("reasoning_chain_stream_delta", displayDelta, map[string]interface{}{
+									"streamId": reasoningStreamID,
 								})
 							}
 						}
 					}
-				}
-				if len(chunk.ToolCalls) > 0 {
-					toolStreamFragments = append(toolStreamFragments, chunk.ToolCalls...)
+					if chunk.Content != "" {
+						if progress != nil && streamsMainAssistant(ev.AgentName) {
+							var contentDelta string
+							mainAssistantBuf, contentDelta = normalizeStreamingDelta(mainAssistantBuf, chunk.Content)
+							if contentDelta != "" {
+								if mainAssistDupTarget == "" {
+									executeStdoutDupMu.Lock()
+									if pendingExecuteStdoutDup != "" {
+										mainAssistDupTarget = pendingExecuteStdoutDup
+									}
+									executeStdoutDupMu.Unlock()
+								}
+								if mainAssistDupTarget != "" {
+									// 已展示过 tool_result，缓冲全文；EOF 后与 execute 输出相同则不再发助手流
+								} else {
+									if !streamHeaderSent {
+										progress("response_start", "", map[string]interface{}{
+											"conversationId":     conversationID,
+											"mcpExecutionIds":    snapshotMCPIDs(),
+											"messageGeneratedBy": "eino:" + ev.AgentName,
+											"einoRole":           "orchestrator",
+											"einoAgent":          ev.AgentName,
+											"orchestration":      orchMode,
+										})
+										streamHeaderSent = true
+									}
+									progress("response_delta", contentDelta, map[string]interface{}{
+										"conversationId":  conversationID,
+										"mcpExecutionIds": snapshotMCPIDs(),
+										"einoRole":        "orchestrator",
+										"einoAgent":       ev.AgentName,
+										"orchestration":   orchMode,
+									})
+								}
+							}
+						} else if !streamsMainAssistant(ev.AgentName) {
+							var subDelta string
+							subAssistantBuf, subDelta = normalizeStreamingDelta(subAssistantBuf, chunk.Content)
+							if subDelta != "" {
+								if progress != nil {
+									if subReplyStreamID == "" {
+										subReplyStreamID = fmt.Sprintf("eino-sub-reply-%s-%d", conversationID, atomic.AddInt64(&einoSubReplyStreamSeq, 1))
+										progress("eino_agent_reply_stream_start", "", map[string]interface{}{
+											"streamId":       subReplyStreamID,
+											"einoAgent":      ev.AgentName,
+											"einoRole":       "sub",
+											"conversationId": conversationID,
+											"source":         "eino",
+										})
+									}
+									progress("eino_agent_reply_stream_delta", subDelta, map[string]interface{}{
+										"streamId":       subReplyStreamID,
+										"conversationId": conversationID,
+									})
+								}
+							}
+						}
+					}
+					if len(chunk.ToolCalls) > 0 {
+						toolStreamFragments = append(toolStreamFragments, chunk.ToolCalls...)
+					}
 				}
 			}
 			if streamsMainAssistant(ev.AgentName) {
@@ -683,10 +754,17 @@ func runEinoADKAgentLoop(ctx context.Context, args *einoADKRunLoopArgs, baseMsgs
 			}
 			var lastToolChunk *schema.Message
 			if merged := mergeStreamingToolCallFragments(toolStreamFragments); len(merged) > 0 {
-				lastToolChunk = &schema.Message{ToolCalls: merged}
+				lastToolChunk = mergeMessageToolCalls(&schema.Message{ToolCalls: merged})
 			}
 			tryEmitToolCallsOnce(lastToolChunk, ev.AgentName, orchestratorName, conversationID, progress, toolEmitSeen, subAgentToolStep, markPending)
+			// 流式路径此前只把 tool_calls 推给进度 UI，未写入 runAccumulatedMsgs；落库后 loadHistory→RepairOrphan 会删掉全部 tool 结果，表现为「续跑/下轮失忆」。
+			if lastToolChunk != nil && len(lastToolChunk.ToolCalls) > 0 {
+				runAccumulatedMsgs = append(runAccumulatedMsgs, schema.AssistantMessage("", lastToolChunk.ToolCalls))
+			}
 			if streamRecvErr != nil {
+				if isInterruptContinue(ctx) {
+					return takePartial(streamRecvErr)
+				}
 				if progress != nil {
 					progress("eino_stream_error", streamRecvErr.Error(), map[string]interface{}{
 						"conversationId": conversationID,
@@ -711,7 +789,7 @@ func runEinoADKAgentLoop(ctx context.Context, args *einoADKRunLoopArgs, baseMsgs
 
 		if mv.Role == schema.Assistant {
 			if progress != nil && strings.TrimSpace(msg.ReasoningContent) != "" {
-				progress("thinking", strings.TrimSpace(msg.ReasoningContent), map[string]interface{}{
+				progress("reasoning_chain", openai.DisplayReasoningContent(strings.TrimSpace(msg.ReasoningContent)), map[string]interface{}{
 					"conversationId": conversationID,
 					"source":         "eino",
 					"einoAgent":      ev.AgentName,
@@ -817,11 +895,15 @@ func runEinoADKAgentLoop(ctx context.Context, args *einoADKRunLoopArgs, baseMsgs
 			if toolCallID != "" {
 				removePendingByID(toolCallID)
 				if _, loaded := toolResultSent.LoadOrStore(toolCallID, struct{}{}); loaded {
+					// ToolInvokeNotify 可能已推过 tool_result（如 execute 流式包装里 Fire 仅携带截断后的 stdout），
+					// 此处仍应用 ADK Tool 消息中的完整内容刷新去重基准，避免模型复述全文时与截断串比对失败而重复展示「助手输出」。
+					recordPendingExecuteStdoutDup(toolName, content, isErr)
 					continue
 				}
 				data["toolCallId"] = toolCallID
 			}
 			recordPendingExecuteStdoutDup(toolName, content, isErr)
+			recordEinoADKFilesystemToolMonitor(args.FilesystemMonitorAgent, args.FilesystemMonitorRecord, toolName, toolCallID, runAccumulatedMsgs, content, isErr)
 			progress("tool_result", fmt.Sprintf("工具结果 (%s)", toolName), data)
 		}
 	}
@@ -831,11 +913,21 @@ func runEinoADKAgentLoop(ctx context.Context, args *einoADKRunLoopArgs, baseMsgs
 	mcpIDsMu.Unlock()
 
 	out := buildEinoRunResultFromAccumulated(
-		orchMode, runAccumulatedMsgs, lastAssistant, lastPlanExecuteExecutor, emptyHint, ids, false,
+		orchMode, runAccumulatedMsgs, persistTraceSource(args, runAccumulatedMsgs),
+		lastAssistant, lastPlanExecuteExecutor, emptyHint, ids, false,
 	)
 	return out, nil
 }
 
+func persistTraceSource(args *einoADKRunLoopArgs, fallback []adk.Message) []adk.Message {
+	if args != nil && args.ModelFacingTrace != nil {
+		if snap := args.ModelFacingTrace.Snapshot(); len(snap) > 0 {
+			return snap
+		}
+	}
+	return fallback
+}
+
 func einoPartialRunLastOutputHint() string {
 	return "[执行未正常结束（用户停止、超时或异常）。续跑时请基于上文已产生的工具与结果继续，勿重复已完成步骤。]\n" +
 		"[Run ended abnormally; continue from the trace above without repeating completed steps.]"
@@ -844,13 +936,18 @@ func einoPartialRunLastOutputHint() string {
 func buildEinoRunResultFromAccumulated(
 	orchMode string,
 	runAccumulatedMsgs []adk.Message,
+	persistMsgs []adk.Message,
 	lastAssistant string,
 	lastPlanExecuteExecutor string,
 	emptyHint string,
 	mcpIDs []string,
 	partial bool,
 ) *RunResult {
-	histJSON, _ := json.Marshal(runAccumulatedMsgs)
+	traceForJSON := persistMsgs
+	if len(traceForJSON) == 0 {
+		traceForJSON = runAccumulatedMsgs
+	}
+	histJSON, _ := json.Marshal(traceForJSON)
 	cleaned := strings.TrimSpace(lastAssistant)
 	if orchMode == "plan_execute" {
 		if e := strings.TrimSpace(lastPlanExecuteExecutor); e != "" {

internal/multiagent/eino_filesystem_tool_monitor.go

@@ -0,0 +1,101 @@
+package multiagent
+
+import (
+	"encoding/json"
+	"errors"
+	"strings"
+
+	"cyberstrike-ai/internal/agent"
+	"cyberstrike-ai/internal/einomcp"
+
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/schema"
+)
+
+// einoADKFilesystemToolNames 与 cloudwego/eino/adk/middlewares/filesystem 默认 ToolName* 一致。
+// execute 已由 eino_execute_monitor 落库，此处不包含。
+var einoADKFilesystemToolNames = map[string]struct{}{
+	"ls":         {},
+	"read_file":  {},
+	"write_file": {},
+	"edit_file":  {},
+	"glob":       {},
+	"grep":       {},
+}
+
+func isBuiltinEinoADKFilesystemToolName(name string) bool {
+	n := strings.ToLower(strings.TrimSpace(name))
+	_, ok := einoADKFilesystemToolNames[n]
+	return ok
+}
+
+func toolCallArgsFromAccumulated(msgs []adk.Message, toolCallID, expectToolName string) map[string]interface{} {
+	tid := strings.TrimSpace(toolCallID)
+	expect := strings.TrimSpace(expectToolName)
+	for i := len(msgs) - 1; i >= 0; i-- {
+		m := msgs[i]
+		if m == nil || m.Role != schema.Assistant || len(m.ToolCalls) == 0 {
+			continue
+		}
+		for j := len(m.ToolCalls) - 1; j >= 0; j-- {
+			tc := m.ToolCalls[j]
+			if tid != "" && strings.TrimSpace(tc.ID) != tid {
+				continue
+			}
+			fn := strings.TrimSpace(tc.Function.Name)
+			if expect != "" && !strings.EqualFold(fn, expect) {
+				continue
+			}
+			raw := strings.TrimSpace(tc.Function.Arguments)
+			if raw == "" {
+				return map[string]interface{}{}
+			}
+			var args map[string]interface{}
+			if err := json.Unmarshal([]byte(raw), &args); err != nil {
+				return map[string]interface{}{"arguments_raw": raw}
+			}
+			if args == nil {
+				return map[string]interface{}{}
+			}
+			return args
+		}
+	}
+	return map[string]interface{}{}
+}
+
+// recordEinoADKFilesystemToolMonitor 将 Eino ADK filesystem 中间件工具结果写入 MCP 监控（与 execute / MCP 桥芯片一致）。
+func recordEinoADKFilesystemToolMonitor(
+	ag *agent.Agent,
+	rec einomcp.ExecutionRecorder,
+	toolName string,
+	toolCallID string,
+	msgs []adk.Message,
+	resultText string,
+	isErr bool,
+) {
+	if ag == nil || rec == nil {
+		return
+	}
+	name := strings.TrimSpace(toolName)
+	if name == "" || strings.EqualFold(name, "execute") {
+		return
+	}
+	if !isBuiltinEinoADKFilesystemToolName(name) {
+		return
+	}
+	args := toolCallArgsFromAccumulated(msgs, toolCallID, name)
+	storedName := "eino_fs::" + strings.ToLower(name)
+	var invErr error
+	if isErr {
+		t := strings.TrimSpace(resultText)
+		if t == "" {
+			invErr = errors.New("tool error")
+		} else {
+			invErr = errors.New(t)
+		}
+	}
+	id := ag.RecordLocalToolExecution(storedName, args, resultText, invErr)
+	if id != "" {
+		rec(id)
+	}
+}

internal/multiagent/eino_model_facing_trace.go

@@ -0,0 +1,84 @@
+package multiagent
+
+import (
+	"context"
+	"encoding/json"
+	"sync"
+
+	"github.com/cloudwego/eino/adk"
+)
+
+// modelFacingTraceHolder 保存「即将送入 ChatModel」的消息快照（已走 summarization / reduction / orphan 修剪等），
+// 用于 last_react_input 落库，使续跑与「上下文压缩后」的模型视角一致，而非仅依赖事件流 append 的 runAccumulatedMsgs。
+type modelFacingTraceHolder struct {
+	mu sync.Mutex
+	// msgs 为深拷贝后的切片，避免框架后续原地修改污染快照
+	msgs []adk.Message
+}
+
+func newModelFacingTraceHolder() *modelFacingTraceHolder {
+	return &modelFacingTraceHolder{}
+}
+
+// Snapshot 返回当前快照的再一次深拷贝（供序列化落库，避免与 holder 互斥长期持锁）。
+func (h *modelFacingTraceHolder) Snapshot() []adk.Message {
+	if h == nil {
+		return nil
+	}
+	h.mu.Lock()
+	defer h.mu.Unlock()
+	return cloneADKMessagesForTrace(h.msgs)
+}
+
+func (h *modelFacingTraceHolder) storeFromState(state *adk.ChatModelAgentState) {
+	if h == nil || state == nil || len(state.Messages) == 0 {
+		return
+	}
+	cloned := cloneADKMessagesForTrace(state.Messages)
+	if len(cloned) == 0 {
+		return
+	}
+	h.mu.Lock()
+	h.msgs = cloned
+	h.mu.Unlock()
+}
+
+func cloneADKMessagesForTrace(msgs []adk.Message) []adk.Message {
+	if len(msgs) == 0 {
+		return nil
+	}
+	b, err := json.Marshal(msgs)
+	if err != nil {
+		return nil
+	}
+	var out []adk.Message
+	if err := json.Unmarshal(b, &out); err != nil {
+		return nil
+	}
+	return out
+}
+
+// modelFacingTraceMiddleware 必须在 Handlers 链中处于 **BeforeModel 最后**（telemetry 之后），
+// 此时 state.Messages 即为本次 LLM 调用的最终入参。
+type modelFacingTraceMiddleware struct {
+	adk.BaseChatModelAgentMiddleware
+	holder *modelFacingTraceHolder
+}
+
+func newModelFacingTraceMiddleware(holder *modelFacingTraceHolder) adk.ChatModelAgentMiddleware {
+	if holder == nil {
+		return nil
+	}
+	return &modelFacingTraceMiddleware{holder: holder}
+}
+
+func (m *modelFacingTraceMiddleware) BeforeModelRewriteState(
+	ctx context.Context,
+	state *adk.ChatModelAgentState,
+	mc *adk.ModelContext,
+) (context.Context, *adk.ChatModelAgentState, error) {
+	if m.holder != nil && state != nil {
+		m.holder.storeFromState(state)
+	}
+	return ctx, state, nil
+}

internal/multiagent/eino_orchestration.go

@@ -41,6 +41,8 @@ type PlanExecuteRootArgs struct {
 	FilesystemMiddleware adk.ChatModelAgentMiddleware
 	// PlannerReplannerRewriteHandlers applies BeforeModelRewriteState pipeline for planner/replanner input.
 	PlannerReplannerRewriteHandlers []adk.ChatModelAgentMiddleware
+	// ModelFacingTrace 可选：由 Executor Handlers 链末尾写入，供 last_react 与 summarization 后上下文对齐。
+	ModelFacingTrace *modelFacingTraceHolder
 }
 
 // NewPlanExecuteRoot 返回 plan → execute → replan 预置编排根节点（与 Deep / Supervisor 并列）。
@@ -101,6 +103,11 @@ func NewPlanExecuteRoot(ctx context.Context, a *PlanExecuteRootArgs) (adk.Resuma
 	if teleMw := newEinoModelInputTelemetryMiddleware(a.Logger, a.ModelName, a.ConversationID, "plan_execute_executor"); teleMw != nil {
 		execHandlers = append(execHandlers, teleMw)
 	}
+	if a.ModelFacingTrace != nil {
+		if capMw := newModelFacingTraceMiddleware(a.ModelFacingTrace); capMw != nil {
+			execHandlers = append(execHandlers, capMw)
+		}
+	}
 	executor, err := newPlanExecuteExecutor(ctx, &planexecute.ExecutorConfig{
 		Model:         a.ExecModel,
 		ToolsConfig:   a.ToolsCfg,

internal/multiagent/eino_single_runner.go

@@ -13,6 +13,7 @@ import (
 	"cyberstrike-ai/internal/config"
 	"cyberstrike-ai/internal/einomcp"
 	"cyberstrike-ai/internal/openai"
+	"cyberstrike-ai/internal/reasoning"
 
 	einoopenai "github.com/cloudwego/eino-ext/components/model/openai"
 	"github.com/cloudwego/eino/adk"
@@ -37,6 +38,7 @@ func RunEinoSingleChatModelAgent(
 	history []agent.ChatMessage,
 	roleTools []string,
 	progress func(eventType, message string, data interface{}),
+	reasoningClient *reasoning.ClientIntent,
 ) (*RunResult, error) {
 	if appCfg == nil || ag == nil {
 		return nil, fmt.Errorf("eino single: 配置或 Agent 为空")
@@ -121,6 +123,7 @@ func RunEinoSingleChatModelAgent(
 		Model:      appCfg.OpenAI.Model,
 		HTTPClient: httpClient,
 	}
+	reasoning.ApplyToEinoChatModelConfig(baseModelCfg, &appCfg.OpenAI, reasoningClient)
 
 	mainModel, err := einoopenai.NewChatModel(ctx, baseModelCfg)
 	if err != nil {
@@ -132,7 +135,9 @@ func RunEinoSingleChatModelAgent(
 		return nil, fmt.Errorf("eino single summarization: %w", err)
 	}
 
-	handlers := make([]adk.ChatModelAgentMiddleware, 0, 4)
+	modelFacingTrace := newModelFacingTraceHolder()
+
+	handlers := make([]adk.ChatModelAgentMiddleware, 0, 8)
 	if len(mainOrchestratorPre) > 0 {
 		handlers = append(handlers, mainOrchestratorPre...)
 	}
@@ -150,6 +155,9 @@ func RunEinoSingleChatModelAgent(
 	if teleMw := newEinoModelInputTelemetryMiddleware(logger, appCfg.OpenAI.Model, conversationID, "eino_single"); teleMw != nil {
 		handlers = append(handlers, teleMw)
 	}
+	if capMw := newModelFacingTraceMiddleware(modelFacingTrace); capMw != nil {
+		handlers = append(handlers, capMw)
+	}
 
 	maxIter := ma.MaxIteration
 	if maxIter <= 0 {
@@ -164,7 +172,7 @@ func RunEinoSingleChatModelAgent(
 			Tools:               mainToolsForCfg,
 			UnknownToolsHandler: einomcp.UnknownToolReminderHandler(),
 			ToolCallMiddlewares: []compose.ToolMiddleware{
-				{Invokable: hitlToolCallMiddleware()},
+				hitlToolCallMiddleware(),
 				{Invokable: softRecoveryToolCallMiddleware()},
 			},
 		},
@@ -223,19 +231,22 @@ func RunEinoSingleChatModelAgent(
 	}
 
 	return runEinoADKAgentLoop(ctx, &einoADKRunLoopArgs{
-		OrchMode:             "eino_single",
-		OrchestratorName:     einoSingleAgentName,
-		ConversationID:       conversationID,
-		Progress:             progress,
-		Logger:               logger,
-		SnapshotMCPIDs:       snapshotMCPIDs,
-		StreamsMainAssistant: streamsMainAssistant,
-		EinoRoleTag:          einoRoleTag,
-		CheckpointDir:        ma.EinoMiddleware.CheckpointDir,
-		McpIDsMu:             &mcpIDsMu,
-		McpIDs:               &mcpIDs,
-		ToolInvokeNotify:     toolInvokeNotify,
-		DA:                   chatAgent,
+		OrchMode:                "eino_single",
+		OrchestratorName:        einoSingleAgentName,
+		ConversationID:          conversationID,
+		Progress:                progress,
+		Logger:                  logger,
+		SnapshotMCPIDs:          snapshotMCPIDs,
+		StreamsMainAssistant:    streamsMainAssistant,
+		EinoRoleTag:             einoRoleTag,
+		CheckpointDir:           ma.EinoMiddleware.CheckpointDir,
+		McpIDsMu:                &mcpIDsMu,
+		McpIDs:                  &mcpIDs,
+		FilesystemMonitorAgent:  ag,
+		FilesystemMonitorRecord: recorder,
+		ToolInvokeNotify:        toolInvokeNotify,
+		DA:                      chatAgent,
+		ModelFacingTrace:        modelFacingTrace,
 		EmptyResponseMessage: "(Eino ADK single-agent session completed but no assistant text was captured. Check process details or logs.) " +
 			"（Eino ADK 单代理会话已完成，但未捕获到助手文本输出。请查看过程详情或日志。）",
 	}, baseMsgs)

internal/multiagent/eino_summarize.go

@@ -214,7 +214,7 @@ func summarizeFinalizeWithRecentAssistantToolTrail(
 		selectedCount++
 	}
 
-	// 还原时间顺序
+	// 还原时间顺序。round 内为原始 *schema.Message 指针，保留 ReasoningContent（DeepSeek 工具续跑所必需）。
 	selectedMsgs := make([]adk.Message, 0, 8)
 	for i := len(selectedRoundsReverse) - 1; i >= 0; i-- {
 		selectedMsgs = append(selectedMsgs, selectedRoundsReverse[i].messages...)

internal/multiagent/hitl_middleware.go

@@ -8,6 +8,7 @@ import (
 
 	"github.com/cloudwego/eino/adk"
 	"github.com/cloudwego/eino/compose"
+	"github.com/cloudwego/eino/schema"
 )
 
 type hitlInterceptorKey struct{}
@@ -41,7 +42,31 @@ func WithHITLToolInterceptor(ctx context.Context, fn HITLToolInterceptor) contex
 	return context.WithValue(ctx, hitlInterceptorKey{}, fn)
 }
 
-func hitlToolCallMiddleware() compose.InvokableToolMiddleware {
+// hitlToolCallMiddleware 同时注册 Invokable 与 Streamable。
+// Eino filesystem 的 execute 为流式工具（StreamableTool），仅挂 Invokable 时人机协同不会拦截，会直接执行。
+func hitlToolCallMiddleware() compose.ToolMiddleware {
+	return compose.ToolMiddleware{
+		Invokable:  hitlInvokableToolCallMiddleware(),
+		Streamable: hitlStreamableToolCallMiddleware(),
+	}
+}
+
+func hitlClearReturnDirectlyIfTransfer(ctx context.Context, toolName string) {
+	if !strings.EqualFold(strings.TrimSpace(toolName), adk.TransferToAgentToolName) {
+		return
+	}
+	_ = compose.ProcessState[*adk.State](ctx, func(_ context.Context, st *adk.State) error {
+		if st == nil {
+			return nil
+		}
+		st.ReturnDirectlyToolCallID = ""
+		st.HasReturnDirectly = false
+		st.ReturnDirectlyEvent = nil
+		return nil
+	})
+}
+
+func hitlInvokableToolCallMiddleware() compose.InvokableToolMiddleware {
 	return func(next compose.InvokableToolEndpoint) compose.InvokableToolEndpoint {
 		return func(ctx context.Context, input *compose.ToolInput) (*compose.ToolOutput, error) {
 			if input != nil {
@@ -55,17 +80,7 @@ func hitlToolCallMiddleware() compose.InvokableToolMiddleware {
 							// transfer_to_agent 在 Eino 中标记为 returnDirectly：工具成功后 ReAct 子图会直接 END，
 							// 并依赖真实工具内的 SendToolGenAction 触发移交。HITL 拒绝时不会执行真实工具，
 							// 若仍走 returnDirectly 分支，监督者会在无 Transfer 动作的情况下结束，模型不再迭代。
-							if strings.EqualFold(strings.TrimSpace(input.Name), adk.TransferToAgentToolName) {
-								_ = compose.ProcessState[*adk.State](ctx, func(_ context.Context, st *adk.State) error {
-									if st == nil {
-										return nil
-									}
-									st.ReturnDirectlyToolCallID = ""
-									st.HasReturnDirectly = false
-									st.ReturnDirectlyEvent = nil
-									return nil
-								})
-							}
+							hitlClearReturnDirectlyIfTransfer(ctx, input.Name)
 							return &compose.ToolOutput{Result: msg}, nil
 						}
 						return nil, err
@@ -79,3 +94,30 @@ func hitlToolCallMiddleware() compose.InvokableToolMiddleware {
 		}
 	}
 }
+
+func hitlStreamableToolCallMiddleware() compose.StreamableToolMiddleware {
+	return func(next compose.StreamableToolEndpoint) compose.StreamableToolEndpoint {
+		return func(ctx context.Context, input *compose.ToolInput) (*compose.StreamToolOutput, error) {
+			if input != nil {
+				if fn, ok := ctx.Value(hitlInterceptorKey{}).(HITLToolInterceptor); ok && fn != nil {
+					edited, err := fn(ctx, input.Name, input.Arguments)
+					if err != nil {
+						if IsHumanRejectError(err) {
+							msg := fmt.Sprintf("[HITL Reject] Tool '%s' was rejected by human reviewer. Reason: %s\nPlease adjust parameters/plan and continue without this call.",
+								input.Name, strings.TrimSpace(err.Error()))
+							hitlClearReturnDirectlyIfTransfer(ctx, input.Name)
+							return &compose.StreamToolOutput{
+								Result: schema.StreamReaderFromArray([]string{msg}),
+							}, nil
+						}
+						return nil, err
+					}
+					if edited != "" {
+						input.Arguments = edited
+					}
+				}
+			}
+			return next(ctx, input)
+		}
+	}
+}

internal/multiagent/interrupt.go

@@ -0,0 +1,7 @@
+package multiagent
+
+import "errors"
+
+// ErrInterruptContinue 作为 context.CancelCause 使用：用户选择「中断并继续」且当前无进行中的 MCP 工具时，
+// 取消当前推理/流式输出，并在同一会话任务内携带用户补充说明自动续跑下一轮（类似 Hermes 式人机回合）。
+var ErrInterruptContinue = errors.New("agent interrupt: continue with user-supplied context")

internal/multiagent/reasoning_trace.go

@@ -0,0 +1,52 @@
+package multiagent
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+)
+
+// AggregatedReasoningFromTraceJSON concatenates non-empty assistant `reasoning_content`
+// fields from last_react-style JSON (slice of message objects) in document order.
+// Used to persist on the single assistant bubble row for audit and for GetMessages fallback
+// when the full trace JSON is unavailable. For strict per-message replay, prefer last_react_input.
+func AggregatedReasoningFromTraceJSON(traceJSON string) string {
+	traceJSON = strings.TrimSpace(traceJSON)
+	if traceJSON == "" {
+		return ""
+	}
+	var arr []map[string]interface{}
+	if err := json.Unmarshal([]byte(traceJSON), &arr); err != nil {
+		return ""
+	}
+	var b strings.Builder
+	for _, m := range arr {
+		role, _ := m["role"].(string)
+		if !strings.EqualFold(strings.TrimSpace(role), "assistant") {
+			continue
+		}
+		rc := reasoningContentFromMessageMap(m)
+		if rc == "" {
+			continue
+		}
+		if b.Len() > 0 {
+			b.WriteByte('\n')
+		}
+		b.WriteString(rc)
+	}
+	return b.String()
+}
+
+func reasoningContentFromMessageMap(m map[string]interface{}) string {
+	if m == nil {
+		return ""
+	}
+	switch v := m["reasoning_content"].(type) {
+	case string:
+		return strings.TrimSpace(v)
+	case nil:
+		return ""
+	default:
+		return strings.TrimSpace(fmt.Sprint(v))
+	}
+}

internal/multiagent/reasoning_trace_test.go

@@ -0,0 +1,20 @@
+package multiagent
+
+import "testing"
+
+func TestAggregatedReasoningFromTraceJSON(t *testing.T) {
+	const j = `[
+{"role":"user","content":"hi"},
+{"role":"assistant","content":"c1","reasoning_content":"r1","tool_calls":[{"id":"1","type":"function","function":{"name":"f","arguments":"{}"}}]},
+{"role":"tool","tool_call_id":"1","content":"out"},
+{"role":"assistant","content":"c2","reasoning_content":"r2"}
+]`
+	got := AggregatedReasoningFromTraceJSON(j)
+	want := "r1\nr2"
+	if got != want {
+		t.Fatalf("got %q want %q", got, want)
+	}
+	if AggregatedReasoningFromTraceJSON("") != "" || AggregatedReasoningFromTraceJSON("[]") != "" {
+		t.Fatal("empty expected")
+	}
+}

internal/multiagent/runner.go

@@ -17,6 +17,7 @@ import (
 	"cyberstrike-ai/internal/config"
 	"cyberstrike-ai/internal/einomcp"
 	"cyberstrike-ai/internal/openai"
+	"cyberstrike-ai/internal/reasoning"
 
 	einoopenai "github.com/cloudwego/eino-ext/components/model/openai"
 	"github.com/cloudwego/eino/adk"
@@ -48,6 +49,7 @@ type toolCallPendingInfo struct {
 
 // RunDeepAgent 使用 Eino 多代理预置编排执行一轮对话（deep / plan_execute / supervisor；流式事件通过 progress 回调输出）。
 // orchestrationOverride 非空时优先（如聊天/WebShell 请求体）；否则用 multi_agent.orchestration（遗留 yaml）；皆空则按 deep。
+// reasoningClient 来自 ChatRequest.reasoning；可为 nil（机器人/批量等走全局 openai.reasoning）。
 func RunDeepAgent(
 	ctx context.Context,
 	appCfg *config.Config,
@@ -61,6 +63,7 @@ func RunDeepAgent(
 	progress func(eventType, message string, data interface{}),
 	agentsMarkdownDir string,
 	orchestrationOverride string,
+	reasoningClient *reasoning.ClientIntent,
 ) (*RunResult, error) {
 	if appCfg == nil || ma == nil || ag == nil {
 		return nil, fmt.Errorf("multiagent: 配置或 Agent 为空")
@@ -163,6 +166,7 @@ func RunDeepAgent(
 		Model:      appCfg.OpenAI.Model,
 		HTTPClient: httpClient,
 	}
+	reasoning.ApplyToEinoChatModelConfig(baseModelCfg, &appCfg.OpenAI, reasoningClient)
 
 	deepMaxIter := ma.MaxIteration
 	if deepMaxIter <= 0 {
@@ -285,7 +289,7 @@ func RunDeepAgent(
 						Tools:               subToolsForCfg,
 						UnknownToolsHandler: einomcp.UnknownToolReminderHandler(),
 						ToolCallMiddlewares: []compose.ToolMiddleware{
-							{Invokable: hitlToolCallMiddleware()},
+							hitlToolCallMiddleware(),
 							{Invokable: softRecoveryToolCallMiddleware()},
 						},
 					},
@@ -311,6 +315,8 @@ func RunDeepAgent(
 		return nil, fmt.Errorf("多代理主 summarization 中间件: %w", err)
 	}
 
+	modelFacingTrace := newModelFacingTraceHolder()
+
 	// 与 deep.Config.Name / supervisor 主代理 Name 一致。
 	orchestratorName := "cyberstrike-deep"
 	orchDescription := "Coordinates specialist agents and MCP tools for authorized security testing."
@@ -407,6 +413,9 @@ func RunDeepAgent(
 	if teleMw := newEinoModelInputTelemetryMiddleware(logger, appCfg.OpenAI.Model, conversationID, "deep_orchestrator"); teleMw != nil {
 		deepHandlers = append(deepHandlers, teleMw)
 	}
+	if capMw := newModelFacingTraceMiddleware(modelFacingTrace); capMw != nil {
+		deepHandlers = append(deepHandlers, capMw)
+	}
 
 	supHandlers := []adk.ChatModelAgentMiddleware{}
 	if len(mainOrchestratorPre) > 0 {
@@ -420,13 +429,16 @@ func RunDeepAgent(
 	if teleMw := newEinoModelInputTelemetryMiddleware(logger, appCfg.OpenAI.Model, conversationID, "supervisor_orchestrator"); teleMw != nil {
 		supHandlers = append(supHandlers, teleMw)
 	}
+	if capMw := newModelFacingTraceMiddleware(modelFacingTrace); capMw != nil {
+		supHandlers = append(supHandlers, capMw)
+	}
 
 	mainToolsCfg := adk.ToolsConfig{
 		ToolsNodeConfig: compose.ToolsNodeConfig{
 			Tools:               mainToolsForCfg,
 			UnknownToolsHandler: einomcp.UnknownToolReminderHandler(),
 			ToolCallMiddlewares: []compose.ToolMiddleware{
-				{Invokable: hitlToolCallMiddleware()},
+				hitlToolCallMiddleware(),
 				{Invokable: softRecoveryToolCallMiddleware()},
 			},
 		},
@@ -465,6 +477,7 @@ func RunDeepAgent(
 			ExecPreMiddlewares:   mainOrchestratorPre,
 			SkillMiddleware:      einoSkillMW,
 			FilesystemMiddleware: peFsMw,
+			ModelFacingTrace:     modelFacingTrace,
 			PlannerReplannerRewriteHandlers: []adk.ChatModelAgentMiddleware{
 				mainSumMw,
 				// 孤儿 tool 消息兜底：必须挂在 summarization 之后、telemetry 之前。
@@ -556,96 +569,99 @@ func RunDeepAgent(
 	}
 
 	return runEinoADKAgentLoop(ctx, &einoADKRunLoopArgs{
-		OrchMode:             orchMode,
-		OrchestratorName:     orchestratorName,
-		ConversationID:       conversationID,
-		Progress:             progress,
-		Logger:               logger,
-		SnapshotMCPIDs:       snapshotMCPIDs,
-		StreamsMainAssistant: streamsMainAssistant,
-		EinoRoleTag:          einoRoleTag,
-		CheckpointDir:        ma.EinoMiddleware.CheckpointDir,
-		McpIDsMu:             &mcpIDsMu,
-		McpIDs:               &mcpIDs,
-		ToolInvokeNotify:     toolInvokeNotify,
-		DA:                   da,
+		OrchMode:                orchMode,
+		OrchestratorName:        orchestratorName,
+		ConversationID:          conversationID,
+		Progress:                progress,
+		Logger:                  logger,
+		SnapshotMCPIDs:          snapshotMCPIDs,
+		StreamsMainAssistant:    streamsMainAssistant,
+		EinoRoleTag:             einoRoleTag,
+		CheckpointDir:           ma.EinoMiddleware.CheckpointDir,
+		McpIDsMu:                &mcpIDsMu,
+		McpIDs:                  &mcpIDs,
+		FilesystemMonitorAgent:  ag,
+		FilesystemMonitorRecord: recorder,
+		ToolInvokeNotify:        toolInvokeNotify,
+		DA:                      da,
+		ModelFacingTrace:        modelFacingTrace,
 		EmptyResponseMessage: "(Eino multi-agent orchestration completed but no assistant text was captured. Check process details or logs.) " +
 			"（Eino 多代理编排已完成，但未捕获到助手文本输出。请查看过程详情或日志。）",
 	}, baseMsgs)
 }
 
+func chatToolCallsToSchema(tcs []agent.ToolCall) []schema.ToolCall {
+	if len(tcs) == 0 {
+		return nil
+	}
+	out := make([]schema.ToolCall, 0, len(tcs))
+	for _, tc := range tcs {
+		if strings.TrimSpace(tc.ID) == "" {
+			continue
+		}
+		argsStr := ""
+		if tc.Function.Arguments != nil {
+			b, err := json.Marshal(tc.Function.Arguments)
+			if err == nil {
+				argsStr = string(b)
+			}
+		}
+		typ := tc.Type
+		if typ == "" {
+			typ = "function"
+		}
+		out = append(out, schema.ToolCall{
+			ID:   tc.ID,
+			Type: typ,
+			Function: schema.FunctionCall{
+				Name:      tc.Function.Name,
+				Arguments: argsStr,
+			},
+		})
+	}
+	return out
+}
+
+// historyToMessages 将轨迹恢复的 ChatMessage 转为 Eino ADK 消息：**不裁剪条数、不按 token 预算截断**，
+// 并保留 user / assistant（含仅 tool_calls）/ tool，与库中 last_react 轨迹一致。
 func historyToMessages(history []agent.ChatMessage, appCfg *config.Config, mwCfg *config.MultiAgentEinoMiddlewareConfig) []adk.Message {
+	_ = appCfg
+	_ = mwCfg
 	if len(history) == 0 {
 		return nil
 	}
-	// Keep a bounded tail first; then enforce a token budget.
-	const maxHistoryMessages = 200
-	start := 0
-	if len(history) > maxHistoryMessages {
-		start = len(history) - maxHistoryMessages
-	}
-	raw := make([]adk.Message, 0, len(history[start:]))
-	for _, h := range history[start:] {
-		switch h.Role {
+	raw := make([]adk.Message, 0, len(history))
+	for _, h := range history {
+		role := strings.ToLower(strings.TrimSpace(h.Role))
+		switch role {
 		case "user":
 			if strings.TrimSpace(h.Content) != "" {
 				raw = append(raw, schema.UserMessage(h.Content))
 			}
 		case "assistant":
-			if strings.TrimSpace(h.Content) == "" && len(h.ToolCalls) > 0 {
+			toolSchema := chatToolCallsToSchema(h.ToolCalls)
+			hasRC := strings.TrimSpace(h.ReasoningContent) != ""
+			if len(toolSchema) > 0 || strings.TrimSpace(h.Content) != "" || hasRC {
+				am := schema.AssistantMessage(h.Content, toolSchema)
+				if hasRC {
+					am.ReasoningContent = strings.TrimSpace(h.ReasoningContent)
+				}
+				raw = append(raw, am)
+			}
+		case "tool":
+			if strings.TrimSpace(h.ToolCallID) == "" && strings.TrimSpace(h.Content) == "" {
 				continue
 			}
-			if strings.TrimSpace(h.Content) != "" {
-				raw = append(raw, schema.AssistantMessage(h.Content, nil))
+			var opts []schema.ToolMessageOption
+			if tn := strings.TrimSpace(h.ToolName); tn != "" {
+				opts = append(opts, schema.WithToolName(tn))
 			}
+			raw = append(raw, schema.ToolMessage(h.Content, h.ToolCallID, opts...))
 		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
+	return raw
 }
 
 // mergeStreamingToolCallFragments 将流式多帧的 ToolCall 按 index 合并 arguments（与 schema.concatToolCalls 行为一致）。

internal/multiagent/runner_reasoning_history_test.go

@@ -0,0 +1,22 @@
+package multiagent
+
+import (
+	"testing"
+
+	"cyberstrike-ai/internal/agent"
+)
+
+func TestHistoryToMessagesPreservesReasoningContent(t *testing.T) {
+	h := []agent.ChatMessage{
+		{Role: "user", Content: "u"},
+		{Role: "assistant", Content: "c", ReasoningContent: "r1", ToolCalls: []agent.ToolCall{{ID: "t1", Type: "function", Function: agent.FunctionCall{Name: "f", Arguments: map[string]interface{}{}}}}},
+	}
+	msgs := historyToMessages(h, nil, nil)
+	if len(msgs) != 2 {
+		t.Fatalf("len=%d", len(msgs))
+	}
+	am := msgs[1]
+	if am.ReasoningContent != "r1" || am.Content != "c" {
+		t.Fatalf("got reasoning=%q content=%q", am.ReasoningContent, am.Content)
+	}
+}

internal/openai/claude_bridge.go

@@ -9,6 +9,9 @@ package openai
 //   Stream:   Claude SSE (event: content_block_delta / message_delta) → OpenAI SSE 格式
 //   Auth:     Bearer → x-api-key
 //   Tools:    OpenAI tools[] → Claude tools[] (input_schema)
+//
+// Extended thinking: 顶层 `thinking` 从 OpenAI 请求体透传；响应中 `thinking` block 映射为
+// `reasoning_content`（可读前缀 + 内部 JSON 尾缀以保留 signature，供多轮工具续跑；UI 用 openai.DisplayReasoningContent 剥离）。
 
 import (
 	"bufio"
@@ -38,6 +41,7 @@ type claudeRequest struct {
 	Messages  []claudeMessage `json:"messages"`
 	Tools     []claudeTool    `json:"tools,omitempty"`
 	Stream    bool            `json:"stream,omitempty"`
+	Thinking  json.RawMessage `json:"thinking,omitempty"`
 }
 
 type claudeMessage struct {
@@ -76,6 +80,10 @@ type claudeContentBlock struct {
 	// text block
 	Text string `json:"text,omitempty"`
 
+	// thinking block (extended thinking)
+	Thinking  string `json:"thinking,omitempty"`
+	Signature string `json:"signature,omitempty"`
+
 	// tool_use block (assistant 返回)
 	ID    string          `json:"id,omitempty"`
 	Name  string          `json:"name,omitempty"`
@@ -176,7 +184,13 @@ func convertOpenAIToClaude(payload interface{}) (*claudeRequest, error) {
 
 		// tool_calls (assistant 消息中包含工具调用)
 		if role == "assistant" {
+			rc, _ := mm["reasoning_content"].(string)
+			_, thinkingReplay := parseClaudeReasoningAssistantBlocks(rc)
+
 			var blocks []claudeContentBlock
+			for _, tb := range thinkingReplay {
+				blocks = append(blocks, tb)
+			}
 			if content != "" {
 				blocks = append(blocks, claudeContentBlock{Type: "text", Text: content})
 			}
@@ -290,6 +304,13 @@ func convertOpenAIToClaude(payload interface{}) (*claudeRequest, error) {
 		}
 	}
 
+	// Extended thinking (Anthropic top-level); merged from Eino ExtraFields / admin extras.
+	if th, ok := oai["thinking"]; ok && th != nil {
+		if raw, err := json.Marshal(th); err == nil && len(raw) > 0 && string(raw) != "null" {
+			req.Thinking = json.RawMessage(raw)
+		}
+	}
+
 	return req, nil
 }
 
@@ -318,9 +339,12 @@ func claudeToOpenAIResponseJSON(claudeBody []byte) ([]byte, error) {
 
 	var textContent string
 	var toolCalls []interface{}
+	var thinkingBlocks []claudeContentBlock
 
 	for _, block := range cr.Content {
 		switch block.Type {
+		case "thinking":
+			thinkingBlocks = append(thinkingBlocks, block)
 		case "text":
 			textContent += block.Text
 		case "tool_use":
@@ -344,6 +368,18 @@ func claudeToOpenAIResponseJSON(claudeBody []byte) ([]byte, error) {
 	if len(toolCalls) > 0 {
 		message["tool_calls"] = toolCalls
 	}
+	if len(thinkingBlocks) > 0 {
+		var parts []string
+		for _, tb := range thinkingBlocks {
+			if strings.TrimSpace(tb.Thinking) != "" {
+				parts = append(parts, tb.Thinking)
+			}
+		}
+		rc := appendClaudeReasoningRoundTrip(strings.Join(parts, "\n\n"), thinkingBlocks)
+		if rc != "" {
+			message["reasoning_content"] = rc
+		}
+	}
 
 	choice := map[string]interface{}{
 		"index":         0,
@@ -901,8 +937,16 @@ func (rt *claudeRoundTripper) RoundTrip(req *http.Request) (*http.Response, erro
 
 		reader := bufio.NewReader(resp.Body)
 		blockToToolIndex := make(map[int]int)
+		blockIndexToType := make(map[int]string)
 		nextToolIndex := 0
 
+		type thinkingAcc struct {
+			text strings.Builder
+			sig  strings.Builder
+		}
+		thinkingByIndex := make(map[int]*thinkingAcc)
+		var finishedThinking []claudeContentBlock
+
 		for {
 			line, readErr := reader.ReadString('\n')
 			if readErr != nil {
@@ -947,6 +991,11 @@ func (rt *claudeRoundTripper) RoundTrip(req *http.Request) (*http.Response, erro
 				blockIdx := int(blockIdxFlt)
 				cb, _ := event["content_block"].(map[string]interface{})
 				bt, _ := cb["type"].(string)
+				blockIndexToType[blockIdx] = bt
+
+				if bt == "thinking" {
+					thinkingByIndex[blockIdx] = &thinkingAcc{}
+				}
 
 				if bt == "tool_use" {
 					id, _ := cb["id"].(string)
@@ -986,7 +1035,35 @@ func (rt *claudeRoundTripper) RoundTrip(req *http.Request) (*http.Response, erro
 				delta, _ := event["delta"].(map[string]interface{})
 				dt, _ := delta["type"].(string)
 
-				if dt == "text_delta" {
+				if dt == "thinking_delta" {
+					tPart, _ := delta["thinking"].(string)
+					if tPart != "" {
+						if acc := thinkingByIndex[blockIdx]; acc != nil {
+							acc.text.WriteString(tPart)
+						}
+						oaiChunk := map[string]interface{}{
+							"choices": []map[string]interface{}{
+								{
+									"delta": map[string]interface{}{
+										"reasoning_content": tPart,
+									},
+								},
+							},
+						}
+						b, _ := json.Marshal(oaiChunk)
+						if !writeLine("data: " + string(b) + "\n\n") {
+							pw.Close()
+							return
+						}
+					}
+				} else if dt == "signature_delta" {
+					sigPart, _ := delta["signature"].(string)
+					if sigPart != "" {
+						if acc := thinkingByIndex[blockIdx]; acc != nil {
+							acc.sig.WriteString(sigPart)
+						}
+					}
+				} else if dt == "text_delta" {
 					text, _ := delta["text"].(string)
 					oaiChunk := map[string]interface{}{
 						"choices": []map[string]interface{}{
@@ -1031,6 +1108,21 @@ func (rt *claudeRoundTripper) RoundTrip(req *http.Request) (*http.Response, erro
 					}
 				}
 
+			case "content_block_stop":
+				blockIdxFlt, _ := event["index"].(float64)
+				blockIdx := int(blockIdxFlt)
+				bt := blockIndexToType[blockIdx]
+				if bt == "thinking" {
+					if acc := thinkingByIndex[blockIdx]; acc != nil {
+						finishedThinking = append(finishedThinking, claudeContentBlock{
+							Type:      "thinking",
+							Thinking:  acc.text.String(),
+							Signature: acc.sig.String(),
+						})
+						delete(thinkingByIndex, blockIdx)
+					}
+				}
+
 			case "message_delta":
 				d, _ := event["delta"].(map[string]interface{})
 				if sr, ok := d["stop_reason"].(string); ok {
@@ -1051,6 +1143,25 @@ func (rt *claudeRoundTripper) RoundTrip(req *http.Request) (*http.Response, erro
 				}
 
 			case "message_stop":
+				if len(finishedThinking) > 0 {
+					suffix := appendClaudeReasoningRoundTrip("", finishedThinking)
+					if strings.TrimSpace(suffix) != "" {
+						oaiChunk := map[string]interface{}{
+							"choices": []map[string]interface{}{
+								{
+									"delta": map[string]interface{}{
+										"reasoning_content": suffix,
+									},
+								},
+							},
+						}
+						b, _ := json.Marshal(oaiChunk)
+						if !writeLine("data: " + string(b) + "\n\n") {
+							pw.Close()
+							return
+						}
+					}
+				}
 				writeLine("data: [DONE]\n\n")
 				pw.Close()
 				return

internal/openai/claude_reasoning_roundtrip.go

@@ -0,0 +1,81 @@
+package openai
+
+import (
+	"encoding/json"
+	"strings"
+)
+
+// claudeReasoningRoundTripSep separates human-readable reasoning from a JSON payload of
+// Anthropic thinking blocks (with signatures) for multi-turn extended thinking + tools.
+// Not shown in UI (see DisplayReasoningContent).
+const claudeReasoningRoundTripSep = "\n---CSAI_CLAUDE_THINKING_BLOCKS---\n"
+
+// DisplayReasoningContent returns reasoning text suitable for the UI (strips internal
+// Claude round-trip JSON suffix). Safe for DeepSeek/plain reasoning strings (no-op).
+func DisplayReasoningContent(s string) string {
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return ""
+	}
+	i := strings.LastIndex(s, claudeReasoningRoundTripSep)
+	if i < 0 {
+		return s
+	}
+	return strings.TrimSpace(s[:i])
+}
+
+func appendClaudeReasoningRoundTrip(display string, blocks []claudeContentBlock) string {
+	var payload []map[string]string
+	for _, b := range blocks {
+		if b.Type != "thinking" {
+			continue
+		}
+		payload = append(payload, map[string]string{
+			"type":      b.Type,
+			"thinking":  b.Thinking,
+			"signature": b.Signature,
+		})
+	}
+	if len(payload) == 0 {
+		return strings.TrimSpace(display)
+	}
+	js, err := json.Marshal(payload)
+	if err != nil {
+		return strings.TrimSpace(display)
+	}
+	d := strings.TrimSpace(display)
+	if d == "" {
+		return claudeReasoningRoundTripSep + string(js)
+	}
+	return d + claudeReasoningRoundTripSep + string(js)
+}
+
+// parseClaudeReasoningAssistantBlocks extracts Anthropic thinking blocks from an OpenAI-style
+// reasoning_content string. When no suffix is present, blocks is nil (caller must not invent signatures).
+func parseClaudeReasoningAssistantBlocks(reasoningContent string) (display string, blocks []claudeContentBlock) {
+	reasoningContent = strings.TrimSpace(reasoningContent)
+	if reasoningContent == "" {
+		return "", nil
+	}
+	idx := strings.LastIndex(reasoningContent, claudeReasoningRoundTripSep)
+	if idx < 0 {
+		return reasoningContent, nil
+	}
+	display = strings.TrimSpace(reasoningContent[:idx])
+	jsonPart := strings.TrimSpace(reasoningContent[idx+len(claudeReasoningRoundTripSep):])
+	var arr []struct {
+		Type      string `json:"type"`
+		Thinking  string `json:"thinking"`
+		Signature string `json:"signature"`
+	}
+	if err := json.Unmarshal([]byte(jsonPart), &arr); err != nil {
+		return reasoningContent, nil
+	}
+	for _, x := range arr {
+		if x.Type != "thinking" {
+			continue
+		}
+		blocks = append(blocks, claudeContentBlock{Type: "thinking", Thinking: x.Thinking, Signature: x.Signature})
+	}
+	return display, blocks
+}

internal/openai/claude_reasoning_roundtrip_test.go

@@ -0,0 +1,102 @@
+package openai
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+func TestDisplayReasoningContent(t *testing.T) {
+	raw := "hello" + claudeReasoningRoundTripSep + `[{"type":"thinking","thinking":"x","signature":"sig"}]`
+	if d := DisplayReasoningContent(raw); d != "hello" {
+		t.Fatalf("got %q", d)
+	}
+	if DisplayReasoningContent("plain") != "plain" {
+		t.Fatal()
+	}
+}
+
+func TestAppendParseClaudeReasoningRoundTrip(t *testing.T) {
+	blocks := []claudeContentBlock{
+		{Type: "thinking", Thinking: "a", Signature: "s1"},
+		{Type: "thinking", Thinking: "b", Signature: "s2"},
+	}
+	s := appendClaudeReasoningRoundTrip("sum", blocks)
+	if !strings.Contains(s, claudeReasoningRoundTripSep) {
+		t.Fatal("missing sep")
+	}
+	display, back := parseClaudeReasoningAssistantBlocks(s)
+	if display != "sum" || len(back) != 2 {
+		t.Fatalf("display=%q len=%d", display, len(back))
+	}
+	if back[0].Signature != "s1" || back[1].Thinking != "b" {
+		t.Fatalf("%+v", back)
+	}
+}
+
+func TestConvertOpenAIToClaude_AssistantReasoningReplay(t *testing.T) {
+	rc := appendClaudeReasoningRoundTrip("vis", []claudeContentBlock{
+		{Type: "thinking", Thinking: "t1", Signature: "sig1"},
+	})
+	payload := map[string]interface{}{
+		"model": "claude-3-5-sonnet-latest",
+		"messages": []interface{}{
+			map[string]interface{}{
+				"role":              "assistant",
+				"content":           "out",
+				"reasoning_content": rc,
+			},
+		},
+	}
+	req, err := convertOpenAIToClaude(payload)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(req.Messages) != 1 {
+		t.Fatalf("messages=%d", len(req.Messages))
+	}
+	blocks := req.Messages[0].Content.Blocks
+	if len(blocks) < 2 {
+		t.Fatalf("blocks=%d", len(blocks))
+	}
+	if blocks[0].Type != "thinking" || blocks[0].Signature != "sig1" {
+		t.Fatalf("first block %+v", blocks[0])
+	}
+	foundText := false
+	for _, b := range blocks {
+		if b.Type == "text" && b.Text == "out" {
+			foundText = true
+		}
+	}
+	if !foundText {
+		t.Fatalf("blocks=%+v", blocks)
+	}
+}
+
+func TestClaudeToOpenAIResponseJSON_Thinking(t *testing.T) {
+	claudeBody := []byte(`{
+		"id":"msg_1","type":"message","role":"assistant","model":"x","stop_reason":"end_turn",
+		"content":[
+			{"type":"thinking","thinking":"step","signature":"sigx"},
+			{"type":"text","text":"hi"}
+		]
+	}`)
+	oai, err := claudeToOpenAIResponseJSON(claudeBody)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var wrap map[string]interface{}
+	if err := json.Unmarshal(oai, &wrap); err != nil {
+		t.Fatal(err)
+	}
+	choices := wrap["choices"].([]interface{})
+	ch0 := choices[0].(map[string]interface{})
+	msg := ch0["message"].(map[string]interface{})
+	rc, _ := msg["reasoning_content"].(string)
+	if !strings.Contains(rc, "step") || !strings.Contains(rc, claudeReasoningRoundTripSep) {
+		t.Fatalf("reasoning_content=%q", rc)
+	}
+	if msg["content"] != "hi" {
+		t.Fatal()
+	}
+}

internal/reasoning/eino.go

@@ -0,0 +1,250 @@
+// Package reasoning maps user/config intent to CloudWeGo Eino OpenAI ChatModel fields
+// (ReasoningEffort, ExtraFields such as thinking / reasoning_effort / output_config).
+package reasoning
+
+import (
+	"strings"
+
+	"cyberstrike-ai/internal/config"
+
+	einoopenai "github.com/cloudwego/eino-ext/components/model/openai"
+)
+
+// ClientIntent is optional per-request override from ChatRequest.reasoning.
+type ClientIntent struct {
+	Mode   string
+	Effort string
+}
+
+type wireProfile int
+
+const (
+	wireNone wireProfile = iota
+	wireClaude
+	wireDeepseek
+	wireOpenAI
+	wireOutputConfig
+)
+
+// ApplyToEinoChatModelConfig merges reasoning-related options into cfg.
+// Precondition: cfg already has APIKey, BaseURL, Model, HTTPClient set.
+func ApplyToEinoChatModelConfig(cfg *einoopenai.ChatModelConfig, oa *config.OpenAIConfig, client *ClientIntent) {
+	if cfg == nil || oa == nil {
+		return
+	}
+	sr := &oa.Reasoning
+	allowClient := sr.AllowClientReasoningEffective()
+	mode := effectiveMode(sr, client, allowClient)
+
+	// Claude (Anthropic): merge admin extras first; optional extended thinking maps to top-level `thinking`
+	// (see internal/openai convertOpenAIToClaude). DeepSeek/OpenAI-style fields are not sent.
+	if strings.EqualFold(strings.TrimSpace(oa.Provider), "claude") ||
+		strings.EqualFold(strings.TrimSpace(oa.Provider), "anthropic") {
+		if len(sr.ExtraRequestFields) > 0 {
+			if cfg.ExtraFields == nil {
+				cfg.ExtraFields = make(map[string]any)
+			}
+			for k, v := range sr.ExtraRequestFields {
+				cfg.ExtraFields[k] = v
+			}
+		}
+		if mode == "off" {
+			return
+		}
+		applyClaudeExtendedThinking(cfg, mode, effectiveEffort(sr, client, allowClient), oa.Model)
+		return
+	}
+
+	if mode == "off" {
+		return
+	}
+	effort := effectiveEffort(sr, client, allowClient)
+	prof := resolveWireProfile(oa, sr)
+
+	// Admin-defined extra root fields (merged first; automatic keys may follow).
+	if len(sr.ExtraRequestFields) > 0 {
+		if cfg.ExtraFields == nil {
+			cfg.ExtraFields = make(map[string]any)
+		}
+		for k, v := range sr.ExtraRequestFields {
+			cfg.ExtraFields[k] = v
+		}
+	}
+
+	switch prof {
+	case wireClaude, wireNone:
+		return
+	case wireDeepseek:
+		applyDeepseek(cfg, mode, effort)
+	case wireOutputConfig:
+		applyOutputConfigEffort(cfg, mode, effort)
+	default: // wireOpenAI
+		applyOpenAICompat(cfg, mode, effort)
+	}
+}
+
+// applyClaudeExtendedThinking sets Anthropic Messages API `thinking` when absent from ExtraRequestFields.
+// Uses adaptive + summarized display by default (per Anthropic guidance for Claude 4.x); Sonnet 3.7 uses enabled+budget.
+func applyClaudeExtendedThinking(cfg *einoopenai.ChatModelConfig, mode, effort, model string) {
+	if cfg == nil || mode == "off" {
+		return
+	}
+	if cfg.ExtraFields == nil {
+		cfg.ExtraFields = make(map[string]any)
+	}
+	if _, exists := cfg.ExtraFields["thinking"]; exists {
+		return
+	}
+	m := strings.ToLower(strings.TrimSpace(model))
+	thinking := map[string]any{
+		"type":    "adaptive",
+		"display": "summarized",
+	}
+	// Sonnet 3.7: manual extended thinking is the documented path.
+	if strings.Contains(m, "claude-3-7-sonnet") || strings.Contains(m, "3-7-sonnet") || strings.Contains(m, "sonnet-3.7") {
+		thinking = map[string]any{
+			"type":          "enabled",
+			"budget_tokens": 10000,
+			"display":       "summarized",
+		}
+	}
+	// Opus 4.7+: manual enabled+budget rejected — keep adaptive only.
+	if strings.Contains(m, "opus-4-7") || strings.Contains(m, "opus-4.7") {
+		thinking = map[string]any{
+			"type":    "adaptive",
+			"display": "summarized",
+		}
+	}
+	_ = effort // reserved: map to Anthropic effort / output_config when API stabilizes in one place
+	cfg.ExtraFields["thinking"] = thinking
+}
+
+func effectiveMode(sr *config.OpenAIReasoningConfig, client *ClientIntent, allowClient bool) string {
+	server := strings.ToLower(strings.TrimSpace(sr.ModeEffective()))
+	if server == "" || server == "default" {
+		server = "auto"
+	}
+	if !allowClient || client == nil {
+		return server
+	}
+	cm := strings.ToLower(strings.TrimSpace(client.Mode))
+	if cm == "" || cm == "default" {
+		return server
+	}
+	return cm
+}
+
+func effectiveEffort(sr *config.OpenAIReasoningConfig, client *ClientIntent, allowClient bool) string {
+	se := normalizeEffort(sr.Effort)
+	if !allowClient || client == nil {
+		return se
+	}
+	ce := normalizeEffort(client.Effort)
+	if ce != "" {
+		return ce
+	}
+	return se
+}
+
+func normalizeEffort(s string) string {
+	e := strings.ToLower(strings.TrimSpace(s))
+	switch e {
+	case "low", "medium", "high", "max":
+		return e
+	default:
+		return ""
+	}
+}
+
+func resolveWireProfile(oa *config.OpenAIConfig, sr *config.OpenAIReasoningConfig) wireProfile {
+	if strings.EqualFold(strings.TrimSpace(oa.Provider), "claude") {
+		return wireClaude
+	}
+	p := strings.ToLower(strings.TrimSpace(sr.ProfileEffective()))
+	switch p {
+	case "output_config", "output_config_effort":
+		return wireOutputConfig
+	case "openai", "openai_compat":
+		return wireOpenAI
+	case "deepseek", "deepseek_compat":
+		return wireDeepseek
+	case "auto", "":
+		bu := strings.ToLower(oa.BaseURL)
+		mo := strings.ToLower(oa.Model)
+		if strings.Contains(bu, "deepseek") || strings.Contains(mo, "deepseek") {
+			return wireDeepseek
+		}
+		return wireOpenAI
+	default:
+		return wireOpenAI
+	}
+}
+
+func applyDeepseek(cfg *einoopenai.ChatModelConfig, mode, effort string) {
+	// auto: enable thinking for DeepSeek line; on: same; auto without effort still opens thinking.
+	if mode == "off" {
+		return
+	}
+	if mode == "auto" || mode == "on" {
+		if cfg.ExtraFields == nil {
+			cfg.ExtraFields = make(map[string]any)
+		}
+		cfg.ExtraFields["thinking"] = map[string]any{"type": "enabled"}
+	}
+	if effort != "" {
+		if cfg.ExtraFields == nil {
+			cfg.ExtraFields = make(map[string]any)
+		}
+		cfg.ExtraFields["reasoning_effort"] = effortStringForAPI(effort)
+	}
+}
+
+func applyOpenAICompat(cfg *einoopenai.ChatModelConfig, mode, effort string) {
+	if mode == "auto" && effort == "" {
+		return
+	}
+	e := effort
+	if mode == "on" && e == "" {
+		e = "medium"
+	}
+	if e == "" {
+		return
+	}
+	if e == "max" {
+		if cfg.ExtraFields == nil {
+			cfg.ExtraFields = make(map[string]any)
+		}
+		cfg.ExtraFields["reasoning_effort"] = "max"
+		return
+	}
+	switch e {
+	case "low":
+		cfg.ReasoningEffort = einoopenai.ReasoningEffortLevelLow
+	case "medium":
+		cfg.ReasoningEffort = einoopenai.ReasoningEffortLevelMedium
+	case "high":
+		cfg.ReasoningEffort = einoopenai.ReasoningEffortLevelHigh
+	}
+}
+
+func applyOutputConfigEffort(cfg *einoopenai.ChatModelConfig, mode, effort string) {
+	if mode == "auto" && effort == "" {
+		return
+	}
+	e := effort
+	if mode == "on" && e == "" {
+		e = "high"
+	}
+	if e == "" {
+		return
+	}
+	if cfg.ExtraFields == nil {
+		cfg.ExtraFields = make(map[string]any)
+	}
+	cfg.ExtraFields["output_config"] = map[string]any{"effort": effortStringForAPI(e)}
+}
+
+func effortStringForAPI(e string) string {
+	// Gateways expect lowercase strings; "max" kept as max.
+	return strings.ToLower(strings.TrimSpace(e))
+}

web/static/css/style.css

@@ -2391,7 +2391,118 @@ header {
     box-sizing: border-box;
 }
 
-.chat-input-container > .chat-input-with-files {
+.chat-input-primary-row {
+    display: flex;
+    flex-direction: row;
+    align-items: flex-end;
+    gap: 8px;
+    flex: 1;
+    min-width: 0;
+    width: 100%;
+}
+
+.chat-input-leading {
+    display: flex;
+    flex-direction: row;
+    align-items: flex-end;
+    gap: 8px;
+    flex-shrink: 0;
+}
+
+/* Eino：模型推理收进浮层，保持主输入行简洁 */
+.chat-reasoning-wrapper {
+    flex-shrink: 0;
+}
+
+.chat-reasoning-inner {
+    position: relative;
+}
+
+.chat-reasoning-btn {
+    max-width: 10.5rem;
+    padding-left: 0.5rem;
+    padding-right: 0.45rem;
+}
+
+.chat-reasoning-btn .chat-reasoning-btn-icon {
+    flex-shrink: 0;
+    font-size: 0.95rem;
+    line-height: 1;
+    opacity: 0.95;
+}
+
+.chat-reasoning-btn.active .chat-reasoning-btn-icon {
+    opacity: 1;
+}
+
+.chat-reasoning-btn .chat-reasoning-btn-summary {
+    max-width: 7.6rem;
+    overflow: hidden;
+    text-overflow: ellipsis;
+    white-space: nowrap;
+}
+
+.chat-reasoning-btn.active {
+    border-color: rgba(49, 130, 206, 0.45);
+    background: rgba(49, 130, 206, 0.06);
+}
+
+.chat-reasoning-panel {
+    position: absolute;
+    bottom: calc(100% + 8px);
+    left: 0;
+    width: 288px;
+    max-width: calc(100vw - 32px);
+    background: #ffffff;
+    border: 1px solid rgba(0, 0, 0, 0.08);
+    border-radius: 16px;
+    padding: 12px;
+    box-shadow: 0 8px 32px rgba(0, 0, 0, 0.12), 0 4px 16px rgba(0, 0, 0, 0.08), 0 0 0 1px rgba(0, 0, 0, 0.04);
+    z-index: 1000;
+    display: flex;
+    flex-direction: column;
+    gap: 10px;
+    text-align: left;
+}
+
+.chat-reasoning-panel-header {
+    margin-bottom: 0;
+}
+
+.chat-reasoning-panel-hint {
+    font-size: 0.75rem;
+    color: var(--text-muted, #718096);
+    margin: 0;
+    line-height: 1.45;
+}
+
+.chat-reasoning-fields {
+    display: flex;
+    flex-direction: column;
+    gap: 12px;
+}
+
+.chat-reasoning-field-label {
+    display: block;
+    font-size: 0.75rem;
+    font-weight: 600;
+    color: var(--text-muted, #718096);
+    margin-bottom: 6px;
+}
+
+.chat-reasoning-select {
+    width: 100%;
+    box-sizing: border-box;
+    padding: 0.45rem 0.6rem;
+    font-size: 0.8125rem;
+    border: 1px solid var(--border-color, #e2e8f0);
+    border-radius: 8px;
+    background: var(--card-bg, #fff);
+    color: var(--text-color, #2d3748);
+}
+
+.chat-input-container .chat-input-with-files,
+.chat-input-primary-row .chat-input-with-files {
     flex: 1;
     display: flex;
     flex-direction: column;
@@ -2399,7 +2510,8 @@ header {
     gap: 6px;
 }
 
-.chat-input-container > .chat-input-field {
+.chat-input-container > .chat-input-field,
+.chat-input-primary-row .chat-input-field {
     flex: 1;
     display: flex;
     min-width: 0;
@@ -3568,6 +3680,11 @@ header {
     background: rgba(156, 39, 176, 0.05);
 }
 
+.timeline-item-reasoning_chain {
+    border-left-color: #5c6bc0;
+    background: rgba(92, 107, 192, 0.06);
+}
+
 .timeline-item-tool_call {
     border-left-color: #ff9800;
     background: rgba(255, 152, 0, 0.05);
@@ -3593,6 +3710,11 @@ header {
     background: rgba(255, 112, 67, 0.12);
 }
 
+.timeline-item-user_interrupt_continue {
+    border-left-color: #d97706;
+    background: rgba(217, 119, 6, 0.08);
+}
+
 .timeline-item-header {
     display: flex;
     align-items: center;
@@ -3623,6 +3745,12 @@ header {
     line-height: 1.6;
 }
 
+/* 流式增量阶段纯文本展示（避免半段 Markdown 反复解析） */
+.timeline-item-content.timeline-stream-plain {
+    white-space: pre-wrap;
+    word-break: break-word;
+}
+
 .tool-details {
     display: flex;
     flex-direction: column;
@@ -12283,6 +12411,9 @@ header {
 .webshell-ai-process-block .webshell-ai-timeline-thinking {
     border-left-color: #9c27b0;
 }
+.webshell-ai-process-block .webshell-ai-timeline-reasoning_chain {
+    border-left-color: #5c6bc0;
+}
 .webshell-ai-process-block .webshell-ai-timeline-tool_call,
 .webshell-ai-process-block .webshell-ai-timeline-tool_calls_detected {
     border-left-color: #ff9800;
@@ -18300,6 +18431,10 @@ button.chat-files-dropdown-item:hover:not(:disabled) {
     transform: translateX(-50%) translateY(0);
 }
 
+.chat-files-toast.chat-toast--error {
+    background: #b91c1c;
+}
+
 /* 对话附件读取 / 文件管理上传 进度条 */
 /* [hidden] 默认会被本类的 display:flex 覆盖，须显式隐藏否则空闲时仍露出灰条 */
 .chat-upload-progress-row[hidden] {

web/static/i18n/en-US.json

@@ -277,6 +277,7 @@
     "planExecuteStreamPhase": "Phase output",
     "einoSubAgentStep": "Sub-agent {{agent}} · step {{n}}",
     "aiThinking": "AI thinking",
+    "reasoningChain": "Reasoning process",
     "planning": "Planning",
     "assistantStreamPhase": "Assistant output",
     "toolCallsDetected": "Detected {{count}} tool call(s)",
@@ -288,6 +289,7 @@
     "error": "Error",
     "streamNetworkErrorHint": "Connection lost ({{detail}}). A long task may still be running on the server; check running tasks at the top or refresh this conversation later.",
     "taskCancelled": "Task cancelled",
+    "userInterruptContinueTitle": "⏸️ User interrupt & continue",
     "unknownTool": "Unknown tool",
     "einoAgentReplyTitle": "Sub-agent reply",
     "einoStreamErrorTitle": "⚠️ Eino stream interrupted ({{agent}})",
@@ -328,6 +330,19 @@
     "agentModeMulti": "Multi-agent",
     "agentModeSingleHint": "Single-model ReAct loop for chat and tool use",
     "agentModeMultiHint": "Eino prebuilt orchestration (deep / plan_execute / supervisor) for complex tasks",
+    "reasoningModeLabel": "Model reasoning",
+    "reasoningEffortLabel": "Reasoning effort",
+    "reasoningModeDefault": "Use system default",
+    "reasoningModeOff": "Off",
+    "reasoningModeOn": "On",
+    "reasoningModeAuto": "Auto",
+    "reasoningEffortUnset": "Unspecified",
+    "reasoningCompactLabel": "Reasoning",
+    "reasoningCompactAria": "Open model reasoning options",
+    "reasoningPanelTitle": "Model reasoning",
+    "reasoningPanelHint": "Only Eino single- and multi-agent requests use these; merged with defaults in Settings.",
+    "reasoningSummaryFollow": "System",
+    "reasoningSummaryDash": "—",
     "agentModeOrchPlanExecute": "Plan-Exec",
     "agentModeOrchSupervisor": "Supervisor",
     "hitlTitle": "Human-in-the-loop",
@@ -396,7 +411,7 @@
     "stopTask": "Stop task",
     "interruptModalTitle": "Interrupt current step",
     "interruptReasonLabel": "Interrupt note",
-    "interruptModalHint": "Same as MCP monitor \"Stop tool\": ends only the in-flight tool call; the conversation and this run continue. Optional note is merged into the tool result (bilingual USER INTERRUPT NOTE, not raw CLI). Leave empty for a plain stop. If no tool is running yet (model still thinking), wait for a tool call or use \"Stop completely\".",
+    "interruptModalHint": "When a tool is running: same as MCP monitor \"Stop tool\" — only that call is stopped and the run continues; your note can be merged into the tool result (USER INTERRUPT NOTE). When no tool is running (model thinking/streaming only): \"Interrupt & continue\" still works — current output pauses, your note is merged into context and the run resumes automatically; the progress timeline shows a \"User interrupt & continue\" entry. Use this instead of a full stop when you only want to steer; use \"Stop completely\" to end the whole task.",
     "interruptReasonPlaceholder": "e.g. Tool is too slow—skip and summarize…",
     "interruptReasonRequired": "Please enter a short note so the model can continue accordingly.",
     "interruptSubmitting": "Submitting...",
@@ -1591,6 +1606,10 @@
     "maxTotalTokens": "Max Context Tokens",
     "maxTotalTokensPlaceholder": "120000",
     "maxTotalTokensHint": "Shared by memory compression and attack chain building. Default: 120000",
+    "openaiReasoningTitle": "Model reasoning (Eino)",
+    "openaiReasoningHint": "Applies to Eino single-agent and multi-agent only; works with chat-page reasoning controls.",
+    "openaiReasoningProfile": "Wire profile",
+    "openaiReasoningAllowClient": "Allow chat page to override reasoning options",
     "fofaBaseUrlPlaceholder": "https://fofa.info/api/v1/search/all (optional)",
     "fofaBaseUrlHint": "Leave empty for default.",
     "email": "Email",

web/static/i18n/zh-CN.json

@@ -266,6 +266,7 @@
     "planExecuteStreamPhase": "阶段输出",
     "einoSubAgentStep": "子代理 {{agent}} · 第 {{n}} 步",
     "aiThinking": "AI思考",
+    "reasoningChain": "推理过程",
     "planning": "规划中",
     "assistantStreamPhase": "助手输出",
     "toolCallsDetected": "检测到 {{count}} 个工具调用",
@@ -277,6 +278,7 @@
     "error": "错误",
     "streamNetworkErrorHint": "连接已中断（{{detail}}）。长时间任务可能仍在后端执行，请查看顶部「运行中」任务或稍后刷新本对话。",
     "taskCancelled": "任务已取消",
+    "userInterruptContinueTitle": "⏸️ 用户中断并继续",
     "unknownTool": "未知工具",
     "einoAgentReplyTitle": "子代理回复",
     "einoStreamErrorTitle": "⚠️ Eino 流式中断（{{agent}}）",
@@ -317,6 +319,19 @@
     "agentModeMulti": "多代理",
     "agentModeSingleHint": "单模型 ReAct 循环，适合常规对话与工具调用",
     "agentModeMultiHint": "Eino 预置编排（deep / plan_execute / supervisor），适合复杂任务",
+    "reasoningModeLabel": "模型推理",
+    "reasoningEffortLabel": "推理强度",
+    "reasoningModeDefault": "跟随系统",
+    "reasoningModeOff": "关闭",
+    "reasoningModeOn": "开启",
+    "reasoningModeAuto": "自动",
+    "reasoningEffortUnset": "不指定",
+    "reasoningCompactLabel": "推理",
+    "reasoningCompactAria": "打开模型推理选项",
+    "reasoningPanelTitle": "模型推理",
+    "reasoningPanelHint": "仅 Eino 单代理与多代理请求会带上这些参数；与系统设置中的默认值合并。",
+    "reasoningSummaryFollow": "系统",
+    "reasoningSummaryDash": "—",
     "agentModeOrchPlanExecute": "Plan-Exec",
     "agentModeOrchSupervisor": "Supervisor",
     "hitlTitle": "人机协同",
@@ -385,7 +400,7 @@
     "stopTask": "停止任务",
     "interruptModalTitle": "中断当前步骤",
     "interruptReasonLabel": "中断说明",
-    "interruptModalHint": "与 MCP 监控页「终止工具」一致：仅结束当前这一次工具调用，整条对话与本轮推理会继续；工具返回中可附带说明（中英 USER INTERRUPT NOTE 块，与命令行原文区分）。留空则等同仅终止工具。若当前没有工具在执行（模型尚在思考），请等待工具开始或改用「彻底停止」。",
+    "interruptModalHint": "有工具在执行时：与 MCP 监控页「终止工具」一致，仅结束当前这一次工具调用，本轮推理会继续；说明可写入工具返回（USER INTERRUPT NOTE）。无工具在执行时（模型纯思考/流式输出）：仍可「中断并继续」——会暂停当前输出，把你的说明合并进上下文并自动续跑；进度详情时间线会出现「用户中断并继续」条目。不需要整轮停止时请优先用本按钮；要结束整条任务请用「彻底停止」。",
     "interruptReasonPlaceholder": "例如：工具耗时过长，请先跳过并总结当前结果…",
     "interruptReasonRequired": "请填写中断说明，以便模型根据你的意图继续。",
     "interruptSubmitting": "提交中...",
@@ -1580,6 +1595,10 @@
     "maxTotalTokens": "最大上下文 Token 数",
     "maxTotalTokensPlaceholder": "120000",
     "maxTotalTokensHint": "内存压缩和攻击链构建共用此配置，默认 120000",
+    "openaiReasoningTitle": "模型推理（Eino）",
+    "openaiReasoningHint": "仅 Eino 单代理与多代理请求生效；与对话页「模型推理」下拉配合使用。",
+    "openaiReasoningProfile": "线路 profile",
+    "openaiReasoningAllowClient": "允许对话页覆盖推理选项",
     "fofaBaseUrlPlaceholder": "https://fofa.info/api/v1/search/all（可选）",
     "fofaBaseUrlHint": "留空则使用默认地址。",
     "email": "Email",

web/static/js/chat.js

@@ -26,6 +26,11 @@ const DRAFT_SAVE_DELAY = 500; // 500ms防抖延迟
 // 对话文件上传相关（后端会拼接路径与内容发给大模型，前端不再重复发文件列表）
 const MAX_CHAT_FILES = 10;
 const CHAT_FILE_DEFAULT_PROMPT = '请根据上传的文件内容进行分析。';
+/** 与 handler.formatInterruptContinueUserMessage 首段一致；主对话不展示，仅迭代详情（user_interrupt_continue） */
+const CHAT_INTERRUPT_CONTINUE_USER_PREFIX = '【用户补充 / 中断后继续】';
+function isInterruptContinueInjectChatMessage(content) {
+    return typeof content === 'string' && content.trimStart().startsWith(CHAT_INTERRUPT_CONTINUE_USER_PREFIX);
+}
 /**
  * 对话附件：选文件后异步 POST /api/chat-uploads，发送时只传 serverPath（绝对路径），请求体不再内联大文件内容。
  * @type {{ id: number, fileName: string, mimeType: string, serverPath: string|null, uploading: boolean, uploadPercent: number, uploadPromise: Promise<void>|null, uploadError: string|null }[]}
@@ -35,6 +40,8 @@ let chatAttachmentSeq = 0;
 
 // 对话模式：react = 原生 ReAct（/agent-loop）；eino_single = Eino ADK 单代理（/api/eino-agent/stream）；deep / plan_execute / supervisor = Eino 多代理（/api/multi-agent/stream，请求体 orchestration）
 const AGENT_MODE_STORAGE_KEY = 'cyberstrike-chat-agent-mode';
+const REASONING_MODE_LS = 'cyberstrike-chat-reasoning-mode';
+const REASONING_EFFORT_LS = 'cyberstrike-chat-reasoning-effort';
 const CHAT_AGENT_MODE_REACT = 'react';
 const CHAT_AGENT_MODE_EINO_SINGLE = 'eino_single';
 const CHAT_AGENT_EINO_MODES = ['deep', 'plan_execute', 'supervisor'];
@@ -51,6 +58,28 @@ const HITL_MODE_REVIEW_EDIT = 'review_edit';
 const HITL_MODE_OPTIONS = [HITL_MODE_OFF, HITL_MODE_APPROVAL, HITL_MODE_REVIEW_EDIT];
 let hitlApplyFeedbackTimer = null;
 
+/** 非阻塞提示（与 chat-files-toast 样式共用） */
+function showChatToast(message, type) {
+    const text = message == null ? '' : String(message);
+    if (!text) return;
+    const el = document.createElement('div');
+    el.className = 'chat-files-toast' + (type === 'error' ? ' chat-toast--error' : '');
+    el.setAttribute('role', 'status');
+    el.textContent = text;
+    document.body.appendChild(el);
+    requestAnimationFrame(function () {
+        el.classList.add('chat-files-toast-visible');
+    });
+    const hideMs = type === 'error' ? 4500 : 2600;
+    setTimeout(function () {
+        el.classList.remove('chat-files-toast-visible');
+        setTimeout(function () { el.remove(); }, 300);
+    }, hideMs);
+}
+if (typeof window !== 'undefined') {
+    window.showChatToast = showChatToast;
+}
+
 function normalizeOrchestrationClient(s) {
     const v = String(s || '').trim().toLowerCase().replace(/-/g, '_');
     if (v === 'plan_execute' || v === 'planexecute' || v === 'pe') return 'plan_execute';
@@ -293,7 +322,7 @@ function showHitlApplyFeedback(text, isError, partial) {
     }
     if (!el) {
         if (text && isError) {
-            alert(text);
+            showChatToast(text, 'error');
         }
         return;
     }
@@ -465,6 +494,132 @@ function syncAgentModeFromValue(value) {
         const v = el.getAttribute('data-value');
         el.classList.toggle('selected', v === value);
     });
+    syncReasoningRowVisibility(value);
+}
+
+function syncReasoningRowVisibility(modeVal) {
+    const wrap = document.getElementById('chat-reasoning-wrapper');
+    if (!wrap) return;
+    const show = modeVal === CHAT_AGENT_MODE_EINO_SINGLE || (multiAgentAPIEnabled && chatAgentModeIsEino(modeVal));
+    wrap.style.display = show ? '' : 'none';
+    if (!show) {
+        closeChatReasoningPanel();
+    } else {
+        updateChatReasoningSummary();
+    }
+}
+
+function reasoningSummaryModeLabel(mode) {
+    const m = (mode || 'default').trim();
+    const t = (typeof window.t === 'function') ? window.t : function (k) { return k; };
+    switch (m) {
+        case 'off': return t('chat.reasoningModeOff');
+        case 'on': return t('chat.reasoningModeOn');
+        case 'auto': return t('chat.reasoningModeAuto');
+        default: return t('chat.reasoningSummaryFollow');
+    }
+}
+
+function updateChatReasoningSummary() {
+    const el = document.getElementById('chat-reasoning-summary');
+    const modeEl = document.getElementById('chat-reasoning-mode');
+    const effEl = document.getElementById('chat-reasoning-effort');
+    if (!el || !modeEl) return;
+    const mode = (modeEl.value || 'default').trim();
+    const effort = effEl && effEl.value ? String(effEl.value).trim() : '';
+    const t = (typeof window.t === 'function') ? window.t : function (k) { return k; };
+    const modePart = reasoningSummaryModeLabel(mode);
+    const effPart = effort || t('chat.reasoningSummaryDash');
+    el.textContent = modePart + ' / ' + effPart;
+}
+
+function closeChatReasoningPanel() {
+    const panel = document.getElementById('chat-reasoning-panel');
+    const btn = document.getElementById('chat-reasoning-btn');
+    if (panel) panel.style.display = 'none';
+    if (btn) {
+        btn.classList.remove('active');
+        btn.setAttribute('aria-expanded', 'false');
+    }
+}
+
+function toggleChatReasoningPanel() {
+    const panel = document.getElementById('chat-reasoning-panel');
+    const btn = document.getElementById('chat-reasoning-btn');
+    if (!panel || !btn) return;
+    const isOpen = panel.style.display === 'flex';
+    if (isOpen) {
+        closeChatReasoningPanel();
+        return;
+    }
+    if (typeof closeAgentModePanel === 'function') {
+        closeAgentModePanel();
+    }
+    if (typeof closeRoleSelectionPanel === 'function') {
+        closeRoleSelectionPanel();
+    }
+    updateChatReasoningSummary();
+    panel.style.display = 'flex';
+    btn.classList.add('active');
+    btn.setAttribute('aria-expanded', 'true');
+}
+
+function restoreChatReasoningControlsFromStorage() {
+    try {
+        const m = document.getElementById('chat-reasoning-mode');
+        const e = document.getElementById('chat-reasoning-effort');
+        if (m) {
+            const v = localStorage.getItem(REASONING_MODE_LS);
+            if (v && ['default', 'off', 'on', 'auto'].indexOf(v) !== -1) {
+                m.value = v;
+            }
+        }
+        if (e) {
+            const v = localStorage.getItem(REASONING_EFFORT_LS);
+            if (v !== null && ['', 'low', 'medium', 'high', 'max'].indexOf(v) !== -1) {
+                e.value = v;
+            }
+        }
+        updateChatReasoningSummary();
+    } catch (err) { /* ignore */ }
+}
+
+function persistChatReasoningPrefs() {
+    try {
+        const m = document.getElementById('chat-reasoning-mode');
+        const elEff = document.getElementById('chat-reasoning-effort');
+        if (m) localStorage.setItem(REASONING_MODE_LS, m.value || 'default');
+        if (elEff) localStorage.setItem(REASONING_EFFORT_LS, elEff.value || '');
+        updateChatReasoningSummary();
+    } catch (err) { /* ignore */ }
+}
+
+/** 供 WebShell 等复用：在 Eino 路径下返回 reasoning 请求片段或 undefined */
+function buildReasoningRequestPayload() {
+    const wrap = document.getElementById('chat-reasoning-wrapper');
+    if (!wrap || wrap.style.display === 'none') {
+        return undefined;
+    }
+    const modeEl = document.getElementById('chat-reasoning-mode');
+    const effEl = document.getElementById('chat-reasoning-effort');
+    if (!modeEl) return undefined;
+    const mode = (modeEl.value || 'default').trim();
+    const effort = effEl && effEl.value ? String(effEl.value).trim() : '';
+    if (mode === 'default' && !effort) {
+        return undefined;
+    }
+    const o = {};
+    if (mode !== 'default') o.mode = mode;
+    if (effort) o.effort = effort;
+    return Object.keys(o).length ? o : undefined;
+}
+
+if (typeof window !== 'undefined') {
+    window.persistChatReasoningPrefs = persistChatReasoningPrefs;
+    window.buildReasoningRequestPayload = buildReasoningRequestPayload;
+    window.closeChatReasoningPanel = closeChatReasoningPanel;
+    window.toggleChatReasoningPanel = toggleChatReasoningPanel;
+    window.updateChatReasoningSummary = updateChatReasoningSummary;
 }
 
 function closeAgentModePanel() {
@@ -486,6 +641,9 @@ function toggleAgentModePanel() {
         closeAgentModePanel();
         return;
     }
+    if (typeof closeChatReasoningPanel === 'function') {
+        closeChatReasoningPanel();
+    }
     if (typeof closeRoleSelectionPanel === 'function') {
         closeRoleSelectionPanel();
     }
@@ -536,6 +694,8 @@ async function initChatAgentModeFromConfig() {
         } catch (e) { /* ignore */ }
         sel.value = stored;
         syncAgentModeFromValue(stored);
+        restoreChatReasoningControlsFromStorage();
+        syncReasoningRowVisibility(stored);
     } catch (e) {
         console.warn('initChatAgentModeFromConfig', e);
     }
@@ -548,6 +708,9 @@ document.addEventListener('languagechange', function () {
     if (v === CHAT_AGENT_MODE_REACT || chatAgentModeIsEinoSingle(v) || chatAgentModeIsEino(v)) {
         syncAgentModeFromValue(v);
     }
+    if (typeof updateChatReasoningSummary === 'function') {
+        updateChatReasoningSummary();
+    }
 });
 
 // 保存输入框草稿到localStorage（防抖版本）
@@ -733,6 +896,10 @@ async function sendMessage() {
             serverPath: a.serverPath
         }));
     }
+    const reasoningPayload = buildReasoningRequestPayload();
+    if (reasoningPayload) {
+        body.reasoning = reasoningPayload;
+    }
     // 发送后清空附件列表
     chatAttachments = [];
     renderChatFileChips();
@@ -2201,6 +2368,8 @@ function renderProcessDetails(messageId, processDetails) {
             }
         } else if (eventType === 'thinking') {
             itemTitle = agPx + '🤔 ' + (typeof window.t === 'function' ? window.t('chat.aiThinking') : 'AI思考');
+        } else if (eventType === 'reasoning_chain') {
+            itemTitle = agPx + '🔗 ' + (typeof window.t === 'function' ? window.t('chat.reasoningChain') : '推理过程');
         } else if (eventType === 'planning') {
             if (typeof window.einoMainStreamPlanningTitle === 'function') {
                 itemTitle = window.einoMainStreamPlanningTitle(data);
@@ -2237,6 +2406,10 @@ function renderProcessDetails(messageId, processDetails) {
             itemTitle = agPx + '🧑‍⚖️ HITL · ' + hitlMsg;
         } else if (eventType === 'progress') {
             itemTitle = typeof window.translateProgressMessage === 'function' ? window.translateProgressMessage(detail.message || '') : (detail.message || '');
+        } else if (eventType === 'user_interrupt_continue') {
+            itemTitle = typeof window.t === 'function'
+                ? window.t('chat.userInterruptContinueTitle')
+                : '⏸️ 用户中断并继续';
         }
         
         addTimelineItem(timeline, eventType, {
@@ -2853,7 +3026,7 @@ async function loadConversation(conversationId) {
         const conversation = await response.json();
         
         if (!response.ok) {
-            alert('加载对话失败: ' + (conversation.error || '未知错误'));
+            showChatToast('加载对话失败: ' + (conversation.error || '未知错误'), 'error');
             return;
         }
         if (seq !== loadConversationRequestSeq) {
@@ -2953,6 +3126,9 @@ async function loadConversation(conversationId) {
 
             // 渲染单条消息的辅助函数
             const renderOneMessage = (msg) => {
+                if (msg.role === 'user' && isInterruptContinueInjectChatMessage(msg.content)) {
+                    return;
+                }
                 let displayContent = msg.content;
                 if (msg.role === 'assistant' && msg.content === '处理中...' && msg.processDetails && msg.processDetails.length > 0) {
                     for (let i = msg.processDetails.length - 1; i >= 0; i--) {
@@ -3061,7 +3237,7 @@ async function loadConversation(conversationId) {
         }
     } catch (error) {
         console.error('加载对话失败:', error);
-        alert('加载对话失败: ' + error.message);
+        showChatToast('加载对话失败: ' + (error && error.message ? error.message : String(error)), 'error');
     }
 }
 
@@ -6617,6 +6793,9 @@ function formatConversationAsMarkdown(conversation, options = {}) {
     }
 
     messages.forEach((msg, index) => {
+        if (msg && msg.role === 'user' && isInterruptContinueInjectChatMessage(msg.content)) {
+            return;
+        }
         const role = getConversationRoleLabel(msg && msg.role);
         const timestamp = formatConversationDateForMarkdown(msg && msg.createdAt);
         const content = msg && typeof msg.content === 'string' ? msg.content : '';
@@ -7196,6 +7375,14 @@ document.addEventListener('click', function(event) {
             closeAgentModePanel();
         }
     }
+
+    const reasoningWrap = document.getElementById('chat-reasoning-wrapper');
+    const reasoningPanel = document.getElementById('chat-reasoning-panel');
+    if (reasoningWrap && reasoningPanel && reasoningPanel.style.display === 'flex') {
+        if (!reasoningWrap.contains(event.target)) {
+            closeChatReasoningPanel();
+        }
+    }
 });
 
 // 创建分组

web/static/js/monitor.js

@@ -273,6 +273,47 @@ function escapeHtmlLocal(text) {
     return div.innerHTML;
 }
 
+/**
+ * 与 internal/openai.normalizeStreamingDelta 一致：兼容网关/模型返回「累计全文」或整包重发，
+ * 避免前端 buffer += chunk 与后端已归一化的增量叠加导致逐段重复（如「响应中显示了响应中显示了」）。
+ * @returns {[string, string]} [nextBuffer, effectiveDelta]
+ */
+function normalizeStreamingDeltaJs(current, incoming) {
+    const cur = current == null ? '' : String(current);
+    const inc = incoming == null ? '' : String(incoming);
+    if (inc === '') {
+        return [cur, ''];
+    }
+    if (cur === '') {
+        return [inc, inc];
+    }
+    if (inc.startsWith(cur) && inc.length > cur.length) {
+        return [inc, inc.slice(cur.length)];
+    }
+    const runeCount = Array.from(cur).length;
+    if (inc === cur && runeCount > 1) {
+        return [cur, ''];
+    }
+    return [cur + inc, inc];
+}
+if (typeof window !== 'undefined') {
+    window.normalizeStreamingDeltaJs = normalizeStreamingDeltaJs;
+}
+
+/** 流式 delta：纯文本，避免每条全量 marked + DOMPurify */
+function setTimelineItemContentStreamPlain(contentEl, text) {
+    if (!contentEl) return;
+    contentEl.classList.add('timeline-stream-plain');
+    contentEl.textContent = text == null ? '' : String(text);
+}
+
+/** 流结束或非流式：富文本（已消毒的 HTML 字符串） */
+function setTimelineItemContentStreamRich(contentEl, html) {
+    if (!contentEl) return;
+    contentEl.classList.remove('timeline-stream-plain');
+    contentEl.innerHTML = html;
+}
+
 function formatAssistantMarkdownContent(text) {
     const raw = text == null ? '' : String(text);
     if (typeof marked !== 'undefined') {
@@ -743,19 +784,33 @@ function integrateProgressToMCPSection(progressId, assistantMessageId, mcpExecut
         mcpSection.appendChild(buttonsContainer);
     }
 
-    const hasExecBtns = buttonsContainer.querySelector('.mcp-detail-btn:not(.process-detail-btn)');
-    if (mcpIds.length > 0 && !hasExecBtns) {
-        mcpIds.forEach((execId, index) => {
+    let maxExecIndex = 0;
+    const existingExecBtns = buttonsContainer.querySelectorAll('.mcp-detail-btn:not(.process-detail-btn)');
+    existingExecBtns.forEach(function (btn) {
+        const n = parseInt(btn.dataset.execIndex, 10);
+        if (!isNaN(n) && n > maxExecIndex) maxExecIndex = n;
+    });
+    const seenExec = new Set();
+    existingExecBtns.forEach(function (btn) {
+        if (btn.dataset.execId) seenExec.add(String(btn.dataset.execId).trim());
+    });
+    let appendedAny = false;
+    if (mcpIds.length > 0) {
+        mcpIds.forEach(function (execId) {
+            const id = execId != null ? String(execId).trim() : '';
+            if (!id || seenExec.has(id)) return;
+            seenExec.add(id);
+            maxExecIndex += 1;
+            appendedAny = true;
             const detailBtn = document.createElement('button');
             detailBtn.className = 'mcp-detail-btn';
-            detailBtn.dataset.execId = execId;
-            detailBtn.dataset.execIndex = String(index + 1);
-            detailBtn.innerHTML = '<span>' + (typeof window.t === 'function' ? window.t('chat.callNumber', { n: index + 1 }) : '调用 #' + (index + 1)) + '</span>';
-            detailBtn.onclick = () => showMCPDetail(execId);
+            detailBtn.dataset.execId = id;
+            detailBtn.dataset.execIndex = String(maxExecIndex);
+            detailBtn.innerHTML = '<span>' + (typeof window.t === 'function' ? window.t('chat.callNumber', { n: maxExecIndex }) : '调用 #' + maxExecIndex) + '</span>';
+            detailBtn.onclick = function () { showMCPDetail(id); };
             buttonsContainer.appendChild(detailBtn);
         });
-        // 使用批量 API 一次性获取所有工具名称（消除 N 次单独请求）
-        if (typeof batchUpdateButtonToolNames === 'function') {
+        if (appendedAny && typeof batchUpdateButtonToolNames === 'function') {
             batchUpdateButtonToolNames(buttonsContainer, mcpIds);
         }
     }
@@ -1038,6 +1093,24 @@ function resolveStreamTimeline(progressId) {
     return timeline;
 }
 
+/** 去重合并 MCP execution id（顺序：先 prev 后 next），用于多段 Run / 多次 SSE 同一任务。 */
+function mergeMcpExecutionIDLists(prev, next) {
+    const seen = new Set();
+    const out = [];
+    const add = function (arr) {
+        if (!Array.isArray(arr)) return;
+        for (let i = 0; i < arr.length; i++) {
+            const s = arr[i] != null ? String(arr[i]).trim() : '';
+            if (!s || seen.has(s)) continue;
+            seen.add(s);
+            out.push(s);
+        }
+    };
+    add(prev);
+    add(next);
+    return out;
+}
+
 // 处理流式事件
 function handleStreamEvent(event, progressElement, progressId, 
                           getAssistantId, setAssistantId, getMcpIds, setMcpIds) {
@@ -1150,20 +1223,38 @@ function handleStreamEvent(event, progressElement, progressId,
             break;
         }
             
-        case 'thinking_stream_start': {
+        case 'thinking_stream_start':
+        case 'reasoning_chain_stream_start': {
             const d = event.data || {};
             const streamId = d.streamId || null;
             if (!streamId) break;
 
+            const timelineType = event.type === 'reasoning_chain_stream_start' ? 'reasoning_chain' : 'thinking';
+
             let state = thinkingStreamStateByProgressId.get(progressId);
             if (!state) {
                 state = new Map();
                 thinkingStreamStateByProgressId.set(progressId, state);
             }
-            // 若已存在，重置 buffer
-            const thinkBase = typeof window.t === 'function' ? window.t('chat.aiThinking') : 'AI思考';
-            const title = timelineAgentBracketPrefix(d) + '🤔 ' + thinkBase;
-            const itemId = addTimelineItem(timeline, 'thinking', {
+            // 同一 streamId 重复 start：复用已有条目，避免孤儿卡片 + 新条目重复收 delta
+            if (state.has(streamId)) {
+                const ex = state.get(streamId);
+                ex.buffer = '';
+                const existingItem = document.getElementById(ex.itemId);
+                if (existingItem) {
+                    const contentEl = existingItem.querySelector('.timeline-item-content');
+                    if (contentEl) {
+                        setTimelineItemContentStreamPlain(contentEl, '');
+                    }
+                }
+                break;
+            }
+            const labelBase = typeof window.t === 'function'
+                ? window.t(timelineType === 'reasoning_chain' ? 'chat.reasoningChain' : 'chat.aiThinking')
+                : (timelineType === 'reasoning_chain' ? '推理过程' : 'AI思考');
+            const emoji = timelineType === 'reasoning_chain' ? '🔗' : '🤔';
+            const title = timelineAgentBracketPrefix(d) + emoji + ' ' + labelBase;
+            const itemId = addTimelineItem(timeline, timelineType, {
                 title: title,
                 message: ' ',
                 data: d
@@ -1172,7 +1263,8 @@ function handleStreamEvent(event, progressElement, progressId,
             break;
         }
 
-        case 'thinking_stream_delta': {
+        case 'thinking_stream_delta':
+        case 'reasoning_chain_stream_delta': {
             const d = event.data || {};
             const streamId = d.streamId || null;
             if (!streamId) break;
@@ -1182,24 +1274,23 @@ function handleStreamEvent(event, progressElement, progressId,
             const s = state.get(streamId);
 
             const delta = event.message || '';
-            s.buffer += delta;
+            const merged = normalizeStreamingDeltaJs(s.buffer, delta);
+            s.buffer = merged[0];
 
             const item = document.getElementById(s.itemId);
             if (item) {
                 const contentEl = item.querySelector('.timeline-item-content');
                 if (contentEl) {
-                    if (typeof formatMarkdown === 'function') {
-                        contentEl.innerHTML = formatMarkdown(s.buffer);
-                    } else {
-                        contentEl.textContent = s.buffer;
-                    }
+                    setTimelineItemContentStreamPlain(contentEl, s.buffer);
                 }
             }
             break;
         }
 
         case 'thinking':
-            // 如果本 thinking 是由 thinking_stream_* 聚合出来的（带 streamId），避免重复创建 timeline item
+        case 'reasoning_chain': {
+            const timelineType = event.type === 'reasoning_chain' ? 'reasoning_chain' : 'thinking';
+            // 若已由 *_stream_* 聚合（带 streamId），避免重复创建 timeline item
             if (event.data && event.data.streamId) {
                 const streamId = event.data.streamId;
                 const state = thinkingStreamStateByProgressId.get(progressId);
@@ -1210,11 +1301,10 @@ function handleStreamEvent(event, progressElement, progressId,
                     if (item) {
                         const contentEl = item.querySelector('.timeline-item-content');
                         if (contentEl) {
-                            // contentEl.innerHTML 用于兼容 Markdown 展示
                             if (typeof formatMarkdown === 'function') {
-                                contentEl.innerHTML = formatMarkdown(s.buffer);
+                                setTimelineItemContentStreamRich(contentEl, formatMarkdown(s.buffer));
                             } else {
-                                contentEl.textContent = s.buffer;
+                                setTimelineItemContentStreamPlain(contentEl, s.buffer);
                             }
                         }
                     }
@@ -1222,12 +1312,17 @@ function handleStreamEvent(event, progressElement, progressId,
                 }
             }
 
-            addTimelineItem(timeline, 'thinking', {
-                title: timelineAgentBracketPrefix(event.data) + '🤔 ' + (typeof window.t === 'function' ? window.t('chat.aiThinking') : 'AI思考'),
+            const labelBase = typeof window.t === 'function'
+                ? window.t(timelineType === 'reasoning_chain' ? 'chat.reasoningChain' : 'chat.aiThinking')
+                : (timelineType === 'reasoning_chain' ? '推理过程' : 'AI思考');
+            const emoji = timelineType === 'reasoning_chain' ? '🔗' : '🤔';
+            addTimelineItem(timeline, timelineType, {
+                title: timelineAgentBracketPrefix(event.data) + emoji + ' ' + labelBase,
                 message: event.message,
                 data: event.data
             });
             break;
+        }
             
         case 'tool_calls_detected':
             addTimelineItem(timeline, 'tool_calls_detected', {
@@ -1271,6 +1366,19 @@ function handleStreamEvent(event, progressElement, progressId,
             });
             break;
 
+        case 'user_interrupt_continue': {
+            const d = event.data || {};
+            const titleBase = typeof window.t === 'function'
+                ? window.t('chat.userInterruptContinueTitle')
+                : '⏸️ 用户中断并继续';
+            addTimelineItem(timeline, 'user_interrupt_continue', {
+                title: titleBase,
+                message: event.message || '',
+                data: d
+            });
+            break;
+        }
+
         case 'eino_stream_error': {
             const d = event.data || {};
             const agent = d.einoAgent ? String(d.einoAgent) : '';
@@ -1456,6 +1564,18 @@ function handleStreamEvent(event, progressElement, progressId,
                 stateMap = new Map();
                 einoAgentReplyStreamStateByProgressId.set(progressId, stateMap);
             }
+            if (stateMap.has(streamId)) {
+                const ex = stateMap.get(streamId);
+                ex.buffer = '';
+                const existingItem = document.getElementById(ex.itemId);
+                if (existingItem) {
+                    let contentEl = existingItem.querySelector('.timeline-item-content');
+                    if (contentEl) {
+                        setTimelineItemContentStreamPlain(contentEl, '');
+                    }
+                }
+                break;
+            }
             const streamingLabel = typeof window.t === 'function' ? window.t('timeline.running') : '执行中...';
             const replyTitleBase = typeof window.t === 'function' ? window.t('chat.einoAgentReplyTitle') : '子代理回复';
             const itemId = addTimelineItem(timeline, 'eino_agent_reply', {
@@ -1477,7 +1597,8 @@ function handleStreamEvent(event, progressElement, progressId,
             const stateMap = einoAgentReplyStreamStateByProgressId.get(progressId);
             if (!stateMap || !stateMap.has(streamId)) break;
             const s = stateMap.get(streamId);
-            s.buffer += delta;
+            const merged = normalizeStreamingDeltaJs(s.buffer, delta);
+            s.buffer = merged[0];
             const item = document.getElementById(s.itemId);
             if (item) {
                 let contentEl = item.querySelector('.timeline-item-content');
@@ -1490,11 +1611,7 @@ function handleStreamEvent(event, progressElement, progressId,
                     }
                 }
                 if (contentEl) {
-                    if (typeof formatMarkdown === 'function') {
-                        contentEl.innerHTML = formatMarkdown(s.buffer);
-                    } else {
-                        contentEl.textContent = s.buffer;
-                    }
+                    setTimelineItemContentStreamPlain(contentEl, s.buffer);
                 }
             }
             break;
@@ -1522,9 +1639,9 @@ function handleStreamEvent(event, progressElement, progressId,
                         item.appendChild(contentEl);
                     }
                     if (typeof formatMarkdown === 'function') {
-                        contentEl.innerHTML = formatMarkdown(full);
+                        setTimelineItemContentStreamRich(contentEl, formatMarkdown(full));
                     } else {
-                        contentEl.textContent = full;
+                        setTimelineItemContentStreamPlain(contentEl, full);
                     }
                     if (d.einoAgent != null && String(d.einoAgent).trim() !== '') {
                         item.dataset.einoAgent = String(d.einoAgent).trim();
@@ -1614,7 +1731,7 @@ function handleStreamEvent(event, progressElement, progressId,
 
             const responseData = event.data || {};
             const mcpIds = responseData.mcpExecutionIds || [];
-            setMcpIds(mcpIds);
+            setMcpIds(mergeMcpExecutionIDLists(typeof getMcpIds === 'function' ? (getMcpIds() || []) : [], mcpIds));
 
             if (responseData.conversationId) {
                 // 如果用户已经开始了新对话（currentConversationId 为 null），且这个事件来自旧对话，则忽略
@@ -1665,7 +1782,8 @@ function handleStreamEvent(event, progressElement, progressId,
             }
 
             const deltaContent = event.message || '';
-            state.buffer += deltaContent;
+            const mergedResp = normalizeStreamingDeltaJs(state.buffer, deltaContent);
+            state.buffer = mergedResp[0];
 
             // 更新时间线条目内容
             if (state.itemId) {
@@ -1675,11 +1793,7 @@ function handleStreamEvent(event, progressElement, progressId,
                     if (contentEl) {
                         const meta = state.streamMeta || responseData;
                         const body = formatTimelineStreamBody(state.buffer, meta);
-                        if (typeof formatMarkdown === 'function') {
-                            contentEl.innerHTML = formatMarkdown(body);
-                        } else {
-                            contentEl.textContent = body;
-                        }
+                        setTimelineItemContentStreamPlain(contentEl, body);
                     }
                 }
             }
@@ -1693,7 +1807,7 @@ function handleStreamEvent(event, progressElement, progressId,
 
             // 先更新 mcp ids
             const responseData = event.data || {};
-            const mcpIds = responseData.mcpExecutionIds || [];
+            const mcpIds = mergeMcpExecutionIDLists(typeof getMcpIds === 'function' ? (getMcpIds() || []) : [], responseData.mcpExecutionIds || []);
             setMcpIds(mcpIds);
 
             // 更新对话ID
@@ -2217,7 +2331,7 @@ async function attachRunningTaskEventStream(conversationId) {
                     if (line.indexOf('data: ') === 0) {
                         try {
                             const eventData = JSON.parse(line.slice(6));
-                            handleStreamEvent(eventData, null, progressId, getAssistantIdFn, setAssistantIdFn, function () { return mcpIds; }, function (ids) { mcpIds = ids; });
+                            handleStreamEvent(eventData, null, progressId, getAssistantIdFn, setAssistantIdFn, function () { return mcpIds; }, function (ids) { mcpIds = mergeMcpExecutionIDLists(mcpIds, ids || []); });
                         } catch (e) {
                             console.error('task-events parse', e);
                         }
@@ -2375,7 +2489,7 @@ function addTimelineItem(timeline, type, options) {
     `;
     
     // 根据类型添加详细内容
-    if ((type === 'thinking' || type === 'planning') && options.message) {
+    if ((type === 'thinking' || type === 'reasoning_chain' || type === 'planning') && options.message) {
         const streamBody = typeof formatTimelineStreamBody === 'function'
             ? formatTimelineStreamBody(options.message, options.data)
             : options.message;
@@ -2430,6 +2544,11 @@ function addTimelineItem(timeline, type, options) {
                 ${escapeHtml(options.message || taskCancelledLabel)}
             </div>
         `;
+    } else if (type === 'user_interrupt_continue' && options.message) {
+        const streamBody = typeof formatTimelineStreamBody === 'function'
+            ? formatTimelineStreamBody(options.message, options.data)
+            : options.message;
+        content += `<div class="timeline-item-content">${formatMarkdown(streamBody)}</div>`;
     }
 
     item.innerHTML = content;
@@ -3305,6 +3424,8 @@ function refreshProgressAndTimelineI18n() {
             } else {
                 titleSpan.textContent = ap + '\uD83E\uDD14 ' + _t('chat.aiThinking');
             }
+        } else if (type === 'reasoning_chain') {
+            titleSpan.textContent = ap + '\uD83D\uDD17 ' + _t('chat.reasoningChain');
         } else if (type === 'planning') {
             if (item.dataset.orchestration && typeof einoMainStreamPlanningTitle === 'function') {
                 titleSpan.textContent = einoMainStreamPlanningTitle({
@@ -3331,6 +3452,8 @@ function refreshProgressAndTimelineI18n() {
             titleSpan.textContent = ap + '\uD83D\uDCAC ' + _t('chat.einoAgentReplyTitle');
         } else if (type === 'cancelled') {
             titleSpan.textContent = '\u26D4 ' + _t('chat.taskCancelled');
+        } else if (type === 'user_interrupt_continue') {
+            titleSpan.textContent = _t('chat.userInterruptContinueTitle');
         } else if (type === 'progress' && item.dataset.progressMessage !== undefined) {
             titleSpan.textContent = typeof window.translateProgressMessage === 'function' ? window.translateProgressMessage(item.dataset.progressMessage) : item.dataset.progressMessage;
         }

web/static/js/roles.js

@@ -256,6 +256,9 @@ function toggleRoleSelectionPanel() {
         if (typeof closeAgentModePanel === 'function') {
             closeAgentModePanel();
         }
+        if (typeof closeChatReasoningPanel === 'function') {
+            closeChatReasoningPanel();
+        }
         panel.style.display = 'flex'; // 使用flex布局
         // 添加打开状态的视觉反馈
         if (roleSelectorBtn) {

web/static/js/settings.js

@@ -159,6 +159,27 @@ async function loadConfig(loadTools = true) {
         if (maxTokensEl) {
             maxTokensEl.value = currentConfig.openai.max_total_tokens || 120000;
         }
+        const orm = currentConfig.openai && currentConfig.openai.reasoning ? currentConfig.openai.reasoning : {};
+        const orModeEl = document.getElementById('openai-reasoning-mode');
+        if (orModeEl) {
+            const mv = (orm.mode || 'auto').toString().trim().toLowerCase();
+            orModeEl.value = ['auto', 'on', 'off'].includes(mv) ? mv : 'auto';
+        }
+        const orEffEl = document.getElementById('openai-reasoning-effort');
+        if (orEffEl) {
+            const ev = (orm.effort || '').toString().trim().toLowerCase();
+            orEffEl.value = ['', 'low', 'medium', 'high', 'max'].includes(ev) ? ev : '';
+        }
+        const orProfEl = document.getElementById('openai-reasoning-profile');
+        if (orProfEl) {
+            const pv = (orm.profile || 'auto').toString().trim().toLowerCase();
+            const ok = ['auto', 'deepseek_compat', 'openai_compat', 'output_config_effort'];
+            orProfEl.value = ok.includes(pv) ? pv : 'auto';
+        }
+        const orAllowEl = document.getElementById('openai-reasoning-allow-client');
+        if (orAllowEl) {
+            orAllowEl.checked = orm.allow_client_reasoning !== false;
+        }
 
         // 填充FOFA配置
         const fofa = currentConfig.fofa || {};
@@ -1065,13 +1086,22 @@ async function applySettings() {
         };
         
         const wecomAgentIdVal = document.getElementById('robot-wecom-agent-id')?.value.trim();
+        const prevOpenai = (currentConfig && currentConfig.openai) ? currentConfig.openai : {};
         const config = {
             openai: {
+                ...prevOpenai,
                 provider: provider,
                 api_key: apiKey,
                 base_url: baseUrl,
                 model: model,
-                max_total_tokens: parseInt(document.getElementById('openai-max-total-tokens')?.value) || 120000
+                max_total_tokens: parseInt(document.getElementById('openai-max-total-tokens')?.value) || 120000,
+                reasoning: {
+                    ...(prevOpenai.reasoning || {}),
+                    mode: document.getElementById('openai-reasoning-mode')?.value || 'auto',
+                    effort: (document.getElementById('openai-reasoning-effort')?.value || '').trim(),
+                    profile: document.getElementById('openai-reasoning-profile')?.value || 'auto',
+                    allow_client_reasoning: document.getElementById('openai-reasoning-allow-client')?.checked !== false
+                }
             },
             fofa: {
                 email: document.getElementById('fofa-email')?.value.trim() || '',

web/static/js/webshell.js

@@ -1658,6 +1658,8 @@ function buildWebshellTimelineItemFromDetail(detail) {
         title = ap + ((typeof window.t === 'function') ? window.t('chat.iterationRound', { n: data.iteration || 1 }) : ('第 ' + (data.iteration || 1) + ' 轮迭代'));
     } else if (eventType === 'thinking') {
         title = ap + '🤔 ' + ((typeof window.t === 'function') ? window.t('chat.aiThinking') : 'AI 思考');
+    } else if (eventType === 'reasoning_chain') {
+        title = ap + '🔗 ' + ((typeof window.t === 'function') ? window.t('chat.reasoningChain') : '推理过程');
     } else if (eventType === 'tool_calls_detected') {
         title = ap + '🔧 ' + ((typeof window.t === 'function') ? window.t('chat.toolCallsDetected', { count: data.count || 0 }) : ('检测到 ' + (data.count || 0) + ' 个工具调用'));
     } else if (eventType === 'tool_call') {
@@ -2847,6 +2849,12 @@ function runWebshellAiSend(conn, inputEl, sendBtn, messagesContainer) {
         if (info && info.orchestration) {
             body.orchestration = info.orchestration;
         }
+        if (typeof window.buildReasoningRequestPayload === 'function') {
+            var rp = window.buildReasoningRequestPayload();
+            if (rp) {
+                body.reasoning = rp;
+            }
+        }
         return apiFetch(info.path, {
             method: 'POST',
             headers: { 'Content-Type': 'application/json' },
@@ -2898,7 +2906,10 @@ function runWebshellAiSend(conn, inputEl, sendBtn, messagesContainer) {
                     } else if (_et === 'response_delta') {
                         var deltaText = (_em != null && _em !== '') ? String(_em) : '';
                         if (deltaText) {
-                            streamingTarget += deltaText;
+                            var normR = (typeof window.normalizeStreamingDeltaJs === 'function')
+                                ? window.normalizeStreamingDeltaJs(streamingTarget, deltaText)
+                                : [streamingTarget + deltaText, deltaText];
+                            streamingTarget = normR[0];
                             webshellStreamingTypingId += 1;
                             streamingTypingId = webshellStreamingTypingId;
                             runWebshellAiStreamingTyping(assistantDiv, streamingTarget, streamingTypingId, messagesContainer);
@@ -2950,23 +2961,33 @@ function runWebshellAiSend(conn, inputEl, sendBtn, messagesContainer) {
                         appendTimelineItem('iteration', '🔍 ' + iterTitle, iterMessage, _ed);
                         if (!streamingTarget) assistantDiv.textContent = '…';
 
-                    // ─── Thinking (non-stream + stream) ───
-                    } else if (_et === 'thinking_stream_start' && _ed.streamId) {
-                        var thinkSLabel = wsTOr('chat.aiThinking', 'AI 思考');
+                    // ─── Thinking / reasoning_chain（推理过程，reasoning_content） ───
+                    } else if ((_et === 'thinking_stream_start' || _et === 'reasoning_chain_stream_start') && _ed.streamId) {
+                        var isRcStart = _et === 'reasoning_chain_stream_start';
+                        if (wsThinkingStreams.has(_ed.streamId)) {
+                            var tsExist = wsThinkingStreams.get(_ed.streamId);
+                            tsExist.buf = '';
+                            if (tsExist.body) tsExist.body.textContent = '';
+                        } else {
+                        var thinkSLabel = wsTOr(isRcStart ? 'chat.reasoningChain' : 'chat.aiThinking', isRcStart ? '推理过程' : 'AI 思考');
+                        var thinkEmoji = isRcStart ? '🔗' : '🤔';
                         var thinkSItem = document.createElement('div');
-                        thinkSItem.className = 'webshell-ai-timeline-item webshell-ai-timeline-thinking';
-                        thinkSItem.innerHTML = '<span class="webshell-ai-timeline-title">' + escapeHtml(webshellAgentPx(_ed) + '🤔 ' + thinkSLabel) + '</span>';
+                        thinkSItem.className = 'webshell-ai-timeline-item webshell-ai-timeline-' + (isRcStart ? 'reasoning_chain' : 'thinking');
+                        thinkSItem.innerHTML = '<span class="webshell-ai-timeline-title">' + escapeHtml(webshellAgentPx(_ed) + thinkEmoji + ' ' + thinkSLabel) + '</span>';
                         var thinkSPre = document.createElement('div');
                         thinkSPre.className = 'webshell-ai-timeline-msg webshell-thinking-stream-body';
                         thinkSItem.appendChild(thinkSPre);
                         timelineContainer.appendChild(thinkSItem);
                         timelineContainer.classList.add('has-items');
                         wsThinkingStreams.set(_ed.streamId, { el: thinkSItem, body: thinkSPre, buf: '' });
+                        }
                         if (!streamingTarget) assistantDiv.textContent = '…';
-                    } else if (_et === 'thinking_stream_delta' && _ed.streamId) {
+                    } else if ((_et === 'thinking_stream_delta' || _et === 'reasoning_chain_stream_delta') && _ed.streamId) {
                         var tsD = wsThinkingStreams.get(_ed.streamId);
                         if (tsD) {
-                            tsD.buf += (_em || '');
+                            var normT = (typeof window.normalizeStreamingDeltaJs === 'function')
+                                ? window.normalizeStreamingDeltaJs(tsD.buf, _em || '') : [tsD.buf + (_em || ''), _em || ''];
+                            tsD.buf = normT[0];
                             if (typeof formatMarkdown === 'function') {
                                 tsD.body.innerHTML = formatMarkdown(tsD.buf);
                             } else {
@@ -2974,7 +2995,7 @@ function runWebshellAiSend(conn, inputEl, sendBtn, messagesContainer) {
                             }
                         }
                         if (!streamingTarget) assistantDiv.textContent = '…';
-                    } else if (_et === 'thinking_stream_end' && _ed.streamId) {
+                    } else if ((_et === 'thinking_stream_end' || _et === 'reasoning_chain_stream_end') && _ed.streamId) {
                         var tsE = wsThinkingStreams.get(_ed.streamId);
                         if (tsE) {
                             var fullThink = (_em != null && _em !== '') ? String(_em) : tsE.buf;
@@ -2985,13 +3006,15 @@ function runWebshellAiSend(conn, inputEl, sendBtn, messagesContainer) {
                             }
                             wsThinkingStreams.delete(_ed.streamId);
                         }
-                    } else if (_et === 'thinking' && _em) {
+                    } else if ((_et === 'thinking' || _et === 'reasoning_chain') && _em) {
                         // 如果有 streamId 且已存在流式条目，跳过避免重复
                         if (_ed.streamId && wsThinkingStreams.has(_ed.streamId)) {
-                            // 已由 thinking_stream_* 处理
+                            // 已由 *_stream_* 处理
                         } else {
-                            var thinkLabel = wsTOr('chat.aiThinking', 'AI 思考');
-                            appendTimelineItem('thinking', webshellAgentPx(_ed) + '🤔 ' + thinkLabel, _em, _ed);
+                            var isRc = _et === 'reasoning_chain';
+                            var thinkLabel = wsTOr(isRc ? 'chat.reasoningChain' : 'chat.aiThinking', isRc ? '推理过程' : 'AI 思考');
+                            var thinkEm = isRc ? '🔗' : '🤔';
+                            appendTimelineItem(isRc ? 'reasoning_chain' : 'thinking', webshellAgentPx(_ed) + thinkEm + ' ' + thinkLabel, _em, _ed);
                         }
                         if (!streamingTarget) assistantDiv.textContent = '…';
 
@@ -3076,6 +3099,12 @@ function runWebshellAiSend(conn, inputEl, sendBtn, messagesContainer) {
 
                     // ─── Eino sub-agent reply streaming ───
                     } else if (_et === 'eino_agent_reply_stream_start' && _ed.streamId) {
+                        if (einoSubReplyStreams.has(_ed.streamId)) {
+                            var stExist = einoSubReplyStreams.get(_ed.streamId);
+                            stExist.buf = '';
+                            var preExist = stExist.el && stExist.el.querySelector('.webshell-eino-reply-stream-body');
+                            if (preExist) preExist.textContent = '';
+                        } else {
                         var repTS = wsTOr('chat.einoAgentReplyTitle', '子代理回复');
                         var runTS = wsTOr('timeline.running', '执行中...');
                         var itemS = document.createElement('div');
@@ -3084,11 +3113,14 @@ function runWebshellAiSend(conn, inputEl, sendBtn, messagesContainer) {
                         timelineContainer.appendChild(itemS);
                         timelineContainer.classList.add('has-items');
                         einoSubReplyStreams.set(_ed.streamId, { el: itemS, buf: '' });
+                        }
                         if (!streamingTarget) assistantDiv.textContent = '…';
                     } else if (_et === 'eino_agent_reply_stream_delta' && _ed.streamId) {
                         var stD = einoSubReplyStreams.get(_ed.streamId);
                         if (stD) {
-                            stD.buf += (_em || '');
+                            var normS = (typeof window.normalizeStreamingDeltaJs === 'function')
+                                ? window.normalizeStreamingDeltaJs(stD.buf, _em || '') : [stD.buf + (_em || ''), _em || ''];
+                            stD.buf = normS[0];
                             var preD = stD.el.querySelector('.webshell-eino-reply-stream-body');
                             if (!preD) {
                                 preD = document.createElement('pre');

web/templates/index.html

@@ -894,6 +894,8 @@
                             <div id="active-tasks-bar" class="active-tasks-bar"></div>
                             <div id="chat-messages" class="chat-messages"></div>
                             <div id="chat-input-container" class="chat-input-container">
+                                <div class="chat-input-primary-row">
+                                <div class="chat-input-leading">
                                 <div class="role-selector-wrapper">
                                     <button id="role-selector-btn" class="role-selector-btn" onclick="toggleRoleSelectionPanel()" data-i18n="chat.selectRole" data-i18n-attr="title" title="选择角色">
                                         <span id="role-selector-icon" class="role-selector-icon">🔵</span>
@@ -979,6 +981,50 @@
                                     </div>
                                     <input type="hidden" id="agent-mode-select" value="react" autocomplete="off">
                                 </div>
+                                <div id="chat-reasoning-wrapper" class="chat-reasoning-wrapper" style="display: none;">
+                                    <div class="chat-reasoning-inner">
+                                        <button type="button" id="chat-reasoning-btn" class="role-selector-btn chat-reasoning-btn" onclick="toggleChatReasoningPanel()" aria-expanded="false" aria-haspopup="dialog" aria-controls="chat-reasoning-panel" data-i18n="chat.reasoningCompactAria" data-i18n-attr="aria-label,title" data-i18n-skip-text="true" aria-label="模型推理选项" title="模型推理选项">
+                                            <span class="chat-reasoning-btn-icon" aria-hidden="true">🔎</span>
+                                            <span id="chat-reasoning-summary" class="role-selector-text chat-reasoning-btn-summary"></span>
+                                            <svg class="role-selector-arrow" width="10" height="10" viewBox="0 0 24 24" fill="none" xmlns="http://www.w3.org/2000/svg" aria-hidden="true">
+                                                <path d="M6 9l6 6 6-6" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"/>
+                                            </svg>
+                                        </button>
+                                        <div id="chat-reasoning-panel" class="chat-reasoning-panel" style="display: none;" role="dialog" aria-labelledby="chat-reasoning-panel-title">
+                                            <div class="role-selection-panel-header chat-reasoning-panel-header">
+                                                <h3 id="chat-reasoning-panel-title" class="role-selection-panel-title" data-i18n="chat.reasoningPanelTitle">模型推理</h3>
+                                                <button type="button" class="role-selection-panel-close" onclick="closeChatReasoningPanel()" data-i18n="common.close" data-i18n-attr="title" title="关闭">
+                                                    <svg width="16" height="16" viewBox="0 0 24 24" fill="none" xmlns="http://www.w3.org/2000/svg">
+                                                        <path d="M18 6L6 18M6 6l12 12" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"/>
+                                                    </svg>
+                                                </button>
+                                            </div>
+                                            <p class="chat-reasoning-panel-hint" data-i18n="chat.reasoningPanelHint">仅 Eino 请求生效，与系统设置中的默认值合并。</p>
+                                            <div class="chat-reasoning-fields">
+                                                <div class="chat-reasoning-field">
+                                                    <label class="chat-reasoning-field-label" for="chat-reasoning-mode"><span data-i18n="chat.reasoningModeLabel">模式</span></label>
+                                                    <select id="chat-reasoning-mode" class="chat-reasoning-select" onchange="persistChatReasoningPrefs()">
+                                                        <option value="default" data-i18n="chat.reasoningModeDefault">跟随系统</option>
+                                                        <option value="off" data-i18n="chat.reasoningModeOff">关闭</option>
+                                                        <option value="on" data-i18n="chat.reasoningModeOn">开启</option>
+                                                        <option value="auto" data-i18n="chat.reasoningModeAuto">自动</option>
+                                                    </select>
+                                                </div>
+                                                <div class="chat-reasoning-field">
+                                                    <label class="chat-reasoning-field-label" for="chat-reasoning-effort"><span data-i18n="chat.reasoningEffortLabel">推理强度</span></label>
+                                                    <select id="chat-reasoning-effort" class="chat-reasoning-select" onchange="persistChatReasoningPrefs()">
+                                                        <option value="" data-i18n="chat.reasoningEffortUnset">不指定</option>
+                                                        <option value="low">low</option>
+                                                        <option value="medium">medium</option>
+                                                        <option value="high">high</option>
+                                                        <option value="max">max</option>
+                                                    </select>
+                                                </div>
+                                            </div>
+                                        </div>
+                                    </div>
+                                </div>
+                                </div>
                                 <div class="chat-input-with-files">
                                     <div id="chat-file-list" class="chat-file-list" aria-label="已选文件列表"></div>
                                     <div id="chat-attachment-progress" class="chat-upload-progress-row" hidden role="status" aria-live="polite">
@@ -1002,6 +1048,7 @@
                                         <path d="M5 12h14M12 5l7 7-7 7" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"/>
                                     </svg>
                                 </button>
+                                </div>
                             </div>
                         </div>
                     </div>
@@ -1989,6 +2036,38 @@
                                             <input type="number" id="openai-max-total-tokens" data-i18n="settingsBasic.maxTotalTokensPlaceholder" data-i18n-attr="placeholder" placeholder="120000" min="1000" step="1000" />
                                             <small style="color: var(--text-muted, #718096); font-size: 0.75rem;" data-i18n="settingsBasic.maxTotalTokensHint">内存压缩和攻击链构建共用此配置，默认 120000</small>
                                         </div>
+                                        <div class="form-group">
+                                            <label data-i18n="settingsBasic.openaiReasoningTitle">模型推理（Eino）</label>
+                                            <small class="form-hint" data-i18n="settingsBasic.openaiReasoningHint">仅影响 Eino 单代理与多代理；对话页可覆盖（见下方「允许对话覆盖」）。</small>
+                                            <div style="display: flex; flex-wrap: wrap; gap: 10px; margin-top: 8px; align-items: center;">
+                                                <label for="openai-reasoning-mode" style="font-size: 0.8125rem;" data-i18n="chat.reasoningModeLabel">模式</label>
+                                                <select id="openai-reasoning-mode" style="min-width: 120px; padding: 0.35rem 0.5rem; border-radius: 6px; border: 1px solid var(--border-color, #e2e8f0);">
+                                                    <option value="auto" data-i18n="chat.reasoningModeAuto">自动</option>
+                                                    <option value="on" data-i18n="chat.reasoningModeOn">开启</option>
+                                                    <option value="off" data-i18n="chat.reasoningModeOff">关闭</option>
+                                                </select>
+                                                <label for="openai-reasoning-effort" style="font-size: 0.8125rem;" data-i18n="chat.reasoningEffortLabel">强度</label>
+                                                <select id="openai-reasoning-effort" style="min-width: 100px; padding: 0.35rem 0.5rem; border-radius: 6px; border: 1px solid var(--border-color, #e2e8f0);">
+                                                    <option value="" data-i18n="chat.reasoningEffortUnset">不指定</option>
+                                                    <option value="low">low</option>
+                                                    <option value="medium">medium</option>
+                                                    <option value="high">high</option>
+                                                    <option value="max">max</option>
+                                                </select>
+                                                <label for="openai-reasoning-profile" style="font-size: 0.8125rem;" data-i18n="settingsBasic.openaiReasoningProfile">线路</label>
+                                                <select id="openai-reasoning-profile" style="min-width: 140px; padding: 0.35rem 0.5rem; border-radius: 6px; border: 1px solid var(--border-color, #e2e8f0);">
+                                                    <option value="auto">auto</option>
+                                                    <option value="deepseek_compat">deepseek_compat</option>
+                                                    <option value="openai_compat">openai_compat</option>
+                                                    <option value="output_config_effort">output_config_effort</option>
+                                                </select>
+                                            </div>
+                                            <label class="checkbox-label" style="margin-top: 8px;">
+                                                <input type="checkbox" id="openai-reasoning-allow-client" class="modern-checkbox" checked />
+                                                <span class="checkbox-custom"></span>
+                                                <span class="checkbox-text" data-i18n="settingsBasic.openaiReasoningAllowClient">允许对话页覆盖推理选项</span>
+                                            </label>
+                                        </div>
                                         <div style="display: flex; align-items: center; gap: 8px; margin-top: 2px;">
                                             <a href="javascript:void(0)" id="test-openai-btn" onclick="testOpenAIConnection()" style="font-size: 0.8125rem; color: var(--accent-color, #3182ce); text-decoration: none; cursor: pointer; user-select: none;" data-i18n="settingsBasic.testConnection">测试连接</a>
                                             <span id="test-openai-result" style="font-size: 0.8125rem;"></span>