Update version number to v1.7.17

agentic 路径(单代理模式)缺少 toolPairReconciler 和 orphanToolPruner，
当 summarization 截断历史破坏配对后，序列化到 OpenAI API 触发 400
"insufficient tool messages following tool_calls message"。

新增 AgenticMessage 版本的 reconciler 和 pruner，与 classic 路径对齐。

Co-authored-by: temp <temp@tempdeMacBook-Pro.local>
Co-authored-by: Cursor <cursoragent@cursor.com>

config.example.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: "v1.7.15"
+version: "v1.7.17"
 # 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
@@ -304,7 +304,7 @@ multi_agent:
     plan_execute_executed_steps_budget_ratio: 0.2 # plan_execute 中 executed_steps 预算比例
     plan_execute_max_step_result_runes: 4000 # plan_execute 每步结果最大字符数（超出截断）
     plan_execute_keep_last_steps: 8 # plan_execute 仅保留最近 N 步正文，早期步骤折叠为标题
-    checkpoint_dir: data/eino-checkpoints # P0：进程崩溃/OOM 后同会话自动 ADK Resume；正常结束会删 .ckpt；与「中断并继续」(last_react_*) 是两套机制
+    checkpoint_dir: "" # 聊天链路不再使用 ADK checkpoint；跨轮模型态统一走 conversations.last_react_*，便于排查 stale context
     model_retry_max_retries: 0 # Eino 原生 ChatModel retry；408/409/425/429/5xx/网络抖动/空流式输出会重试；0=默认 4（永久性 4xx 不重试）
     model_retry_max_backoff_sec: 0 # Eino 原生 ChatModel retry 单次退避上限秒数；0=默认 30
     model_failover_channels: [] # Eino 原生 ChatModel failover；填写 ai.channels ID，例如 [qwen-plus]；retry 耗尽后按顺序切换

internal/agentfinalizer/decision.go

@@ -65,12 +65,6 @@ func FromRunResult(db *database.DB, result *multiagent.RunResult, in Input) Deci
 		if len(in.MCPExecutionIDs) == 0 {
 			in.MCPExecutionIDs = result.MCPExecutionIDs
 		}
-		if strings.TrimSpace(in.Status) == "" {
-			in.Status = result.Status
-		}
-		if strings.TrimSpace(in.CompletionReason) == "" {
-			in.CompletionReason = result.CompletionReason
-		}
 	}
 	d := Decide(db, in)
 	if result != nil {

internal/agentfinalizer/decision_test.go

@@ -7,6 +7,7 @@ import (
 
 	"cyberstrike-ai/internal/database"
 	"cyberstrike-ai/internal/mcp"
+	"cyberstrike-ai/internal/multiagent"
 
 	"go.uber.org/zap"
 )
@@ -130,3 +131,23 @@ func TestDecideAllowsInformationalAnswerWhenExecutionEvidenceIsNotRequired(t *te
 		t.Fatalf("informational response should finalize when execution evidence is not required: %+v", d)
 	}
 }
+
+func TestFromRunResultDoesNotReusePreviousFinalizationStatusAsRunStatus(t *testing.T) {
+	db := newDecisionTestDB(t)
+	saveDecisionTestExecution(t, db, "run-slow", mcp.ToolExecutionStatusRunning)
+	result := &multiagent.RunResult{
+		Response:        "工具已触发，按用户要求直接总结。",
+		MCPExecutionIDs: []string{"run-slow"},
+	}
+
+	first := FromRunResult(db, result, Input{})
+	if first.Finalizable || first.CompletionReason != ReasonPendingTools || result.Status != StatusInProgress {
+		t.Fatalf("first decision should mark pending and write metadata: decision=%+v result=%+v", first, result)
+	}
+
+	saveDecisionTestExecution(t, db, "run-slow", mcp.ToolExecutionStatusCancelled)
+	second := FromRunResult(db, result, Input{})
+	if !second.Finalizable || !second.Finalized || second.Status != StatusCompleted {
+		t.Fatalf("second decision should ignore previous result status after pending cleanup: decision=%+v result=%+v", second, result)
+	}
+}

internal/app/app.go

@@ -1027,9 +1027,11 @@ func setupRoutes(
 		protected.DELETE("/batch-tasks/:queueId/tasks/:taskId", agentHandler.DeleteBatchTask)
 
 		// 对话历史
+		protected.GET("/usage/tokens", conversationHandler.GetTokenUsageStats)
 		protected.POST("/conversations", conversationHandler.CreateConversation)
 		protected.GET("/conversations", conversationHandler.ListConversations)
 		protected.GET("/conversations/:id", conversationHandler.GetConversation)
+		protected.GET("/conversations/:id/token-usage", conversationHandler.GetConversationTokenUsageStats)
 		protected.GET("/conversations/:id/plan-tasks", conversationHandler.GetConversationPlanTasks)
 		protected.GET("/messages/:id/process-details", conversationHandler.GetMessageProcessDetails)
 		protected.GET("/process-details/:id", conversationHandler.GetProcessDetail)

internal/config/config.go

@@ -298,7 +298,8 @@ type MultiAgentEinoMiddlewareConfig struct {
 	PlanExecuteMaxStepResultRunes int `yaml:"plan_execute_max_step_result_runes,omitempty" json:"plan_execute_max_step_result_runes,omitempty"`
 	// PlanExecuteKeepLastSteps keeps only the tail steps in prompt view (default 8).
 	PlanExecuteKeepLastSteps int `yaml:"plan_execute_keep_last_steps,omitempty" json:"plan_execute_keep_last_steps,omitempty"`
-	// CheckpointDir when non-empty enables adk.Runner CheckPointStore (file-backed) for interrupt/resume persistence.
+	// CheckpointDir is retained for config compatibility. Chat agent runs do
+	// not consume it; cross-turn recovery is centralized in conversations.last_react_*.
 	CheckpointDir string `yaml:"checkpoint_dir,omitempty" json:"checkpoint_dir,omitempty"`
 	// DeepOutputKey passed to deep.Config OutputKey (session final text); empty = off.
 	DeepOutputKey string `yaml:"deep_output_key,omitempty" json:"deep_output_key,omitempty"`
@@ -959,13 +960,12 @@ func (c OpenAIConfig) MaxCompletionTokensEffective() int {
 }
 
 // IsDeepSeekEndpointOrModel reports whether the channel targets DeepSeek's
-// official-compatible API or a DeepSeek model family. This is separate from the
+// official-compatible API endpoint. The historical name is kept for compatibility;
-// reasoning profile: profile controls field mapping, while DeepSeek has provider
+// model names alone are not enough to infer DeepSeek wire behavior behind
-// constraints such as default thinking mode and no tool_choice in thinking mode.
+// OpenAI-compatible gateways.
 func (c OpenAIConfig) IsDeepSeekEndpointOrModel() bool {
 	baseURL := strings.ToLower(strings.TrimSpace(c.BaseURL))
-	model := strings.ToLower(strings.TrimSpace(c.Model))
+	return strings.Contains(baseURL, "deepseek")
-	return strings.Contains(baseURL, "deepseek") || strings.Contains(model, "deepseek")
 }
 
 // OpenAIReasoningConfig 全局默认与网关 profile（对话页可通过 ChatRequest.reasoning 覆盖，受 AllowClientReasoning 约束）。

internal/database/conversation.go

@@ -1350,6 +1350,8 @@ func (db *DB) AddProcessDetailWithID(messageID, conversationID, eventType, messa
 		return "", fmt.Errorf("添加过程详情失败: %w", err)
 	}
 
+	db.maybeRecordModelTokenUsage(messageID, conversationID, id, eventType, data)
+
 	return id, nil
 }
 
@@ -1538,6 +1540,11 @@ LIMIT 1`, messageID).Scan(&terminalEvent, &terminalCreatedAt)
 		return nil, fmt.Errorf("统计工具调用详情失败: %w", err)
 	}
 
+	pendingToolStatus := "result_missing"
+	if summary.Status == "running" {
+		pendingToolStatus = "running"
+	}
+
 	execRows, err := db.Query(
 		"SELECT id, event_type, data FROM process_details WHERE message_id = ? AND event_type IN ('tool_call', 'tool_result') ORDER BY created_at ASC, rowid ASC",
 		messageID,
@@ -1578,10 +1585,10 @@ LIMIT 1`, messageID).Scan(&terminalEvent, &terminalCreatedAt)
 				ProcessDetailID: strings.TrimSpace(detailID),
 				ToolName:        toolName,
 				ToolCallID:      toolCallID,
-				// This summary is reconstructed from persisted history, not live
+				// This summary is reconstructed from persisted history. For an
-				// execution state. Until a matching result is found the honest state
+				// active assistant turn, a missing result means the call is still
-				// is "result_missing", never "running".
+				// pending; after the turn is terminal it is genuinely incomplete.
-				Status: "result_missing",
+				Status: pendingToolStatus,
 			})
 			matchedToolIndexes = append(matchedToolIndexes, false)
 			if toolCallID != "" {
@@ -1636,6 +1643,7 @@ LIMIT 1`, messageID).Scan(&terminalEvent, &terminalCreatedAt)
 		return nil, fmt.Errorf("遍历工具执行摘要失败: %w", err)
 	}
 	execRows.Close()
+	db.applyPersistedToolExecutionStatuses(summary.ToolExecutions)
 
 	rows, err := db.Query(
 		"SELECT data FROM process_details WHERE message_id = ? AND event_type = 'iteration' ORDER BY created_at ASC, rowid ASC",
@@ -1704,6 +1712,24 @@ func toolResultStatusFromPayload(payload map[string]interface{}, eventType strin
 	return "completed"
 }
 
+func (db *DB) applyPersistedToolExecutionStatuses(executions []ProcessDetailsToolExecution) {
+	for i := range executions {
+		execID := strings.TrimSpace(executions[i].ExecutionID)
+		if execID == "" {
+			continue
+		}
+		var status string
+		if err := db.QueryRow(`SELECT status FROM tool_executions WHERE id = ?`, execID).Scan(&status); err != nil {
+			continue
+		}
+		status = strings.ToLower(strings.TrimSpace(status))
+		if status == "" {
+			continue
+		}
+		executions[i].Status = status
+	}
+}
+
 func matchToolExecutionIndex(
 	executions []ProcessDetailsToolExecution,
 	matched []bool,

internal/database/database.go

@@ -216,6 +216,32 @@ func (db *DB) initTables() error {
 		FOREIGN KEY (conversation_id) REFERENCES conversations(id) ON DELETE CASCADE
 	);`
 
+	// 创建模型 Token 用量表：process_details 负责时间线回放，本表负责结构化聚合统计。
+	createModelTokenUsageTable := `
+	CREATE TABLE IF NOT EXISTS model_token_usage (
+		id TEXT PRIMARY KEY,
+		process_detail_id TEXT NOT NULL UNIQUE,
+		message_id TEXT NOT NULL,
+		conversation_id TEXT NOT NULL,
+		project_id TEXT,
+		source TEXT NOT NULL DEFAULT '',
+		orchestration TEXT NOT NULL DEFAULT '',
+		reason TEXT NOT NULL DEFAULT '',
+		model TEXT NOT NULL DEFAULT '',
+		model_calls INTEGER NOT NULL DEFAULT 0,
+		prompt_tokens INTEGER NOT NULL DEFAULT 0,
+		completion_tokens INTEGER NOT NULL DEFAULT 0,
+		total_tokens INTEGER NOT NULL DEFAULT 0,
+		cached_tokens INTEGER NOT NULL DEFAULT 0,
+		reasoning_tokens INTEGER NOT NULL DEFAULT 0,
+		created_at DATETIME NOT NULL,
+		updated_at DATETIME NOT NULL,
+		FOREIGN KEY (process_detail_id) REFERENCES process_details(id) ON DELETE CASCADE,
+		FOREIGN KEY (message_id) REFERENCES messages(id) ON DELETE CASCADE,
+		FOREIGN KEY (conversation_id) REFERENCES conversations(id) ON DELETE CASCADE,
+		FOREIGN KEY (project_id) REFERENCES projects(id) ON DELETE SET NULL
+	);`
+
 	// 创建工具执行记录表
 	createToolExecutionsTable := `
 	CREATE TABLE IF NOT EXISTS tool_executions (
@@ -719,6 +745,10 @@ func (db *DB) initTables() error {
 	CREATE INDEX IF NOT EXISTS idx_conversations_updated_at ON conversations(updated_at);
 	CREATE INDEX IF NOT EXISTS idx_process_details_message_id ON process_details(message_id);
 	CREATE INDEX IF NOT EXISTS idx_process_details_conversation_id ON process_details(conversation_id);
+	CREATE INDEX IF NOT EXISTS idx_model_token_usage_created_at ON model_token_usage(created_at);
+	CREATE INDEX IF NOT EXISTS idx_model_token_usage_conversation ON model_token_usage(conversation_id);
+	CREATE INDEX IF NOT EXISTS idx_model_token_usage_project ON model_token_usage(project_id);
+	CREATE INDEX IF NOT EXISTS idx_model_token_usage_model ON model_token_usage(model);
 	CREATE INDEX IF NOT EXISTS idx_tool_executions_tool_name ON tool_executions(tool_name);
 	CREATE INDEX IF NOT EXISTS idx_tool_executions_start_time ON tool_executions(start_time);
 	CREATE INDEX IF NOT EXISTS idx_tool_executions_status ON tool_executions(status);
@@ -806,6 +836,10 @@ func (db *DB) initTables() error {
 		return fmt.Errorf("创建process_details表失败: %w", err)
 	}
 
+	if _, err := db.Exec(createModelTokenUsageTable); err != nil {
+		return fmt.Errorf("创建model_token_usage表失败: %w", err)
+	}
+
 	if _, err := db.Exec(createToolExecutionsTable); err != nil {
 		return fmt.Errorf("创建tool_executions表失败: %w", err)
 	}
@@ -981,6 +1015,10 @@ func (db *DB) initTables() error {
 	if _, err := db.Exec(createIndexes); err != nil {
 		return fmt.Errorf("创建索引失败: %w", err)
 	}
+
+	if err := db.BackfillModelTokenUsageFromProcessDetails(); err != nil {
+		return fmt.Errorf("回填模型Token用量失败: %w", err)
+	}
 	db.logger.Debug("数据库表初始化完成")
 	return nil
 }

internal/database/model_token_usage.go

@@ -0,0 +1,485 @@
+package database
+
+import (
+	"database/sql"
+	"encoding/json"
+	"fmt"
+	"math"
+	"strconv"
+	"strings"
+	"time"
+
+	"github.com/google/uuid"
+	"go.uber.org/zap"
+)
+
+const modelTokenUsageEventType = "eino_usage_summary"
+
+// ModelTokenUsage records one model-usage summary emitted by an Agent run.
+type ModelTokenUsage struct {
+	ID               string    `json:"id"`
+	ProcessDetailID  string    `json:"processDetailId"`
+	MessageID        string    `json:"messageId"`
+	ConversationID   string    `json:"conversationId"`
+	ProjectID        string    `json:"projectId,omitempty"`
+	Source           string    `json:"source"`
+	Orchestration    string    `json:"orchestration"`
+	Reason           string    `json:"reason"`
+	Model            string    `json:"model,omitempty"`
+	ModelCalls       int64     `json:"modelCalls"`
+	PromptTokens     int64     `json:"promptTokens"`
+	CompletionTokens int64     `json:"completionTokens"`
+	TotalTokens      int64     `json:"totalTokens"`
+	CachedTokens     int64     `json:"cachedTokens"`
+	ReasoningTokens  int64     `json:"reasoningTokens"`
+	CreatedAt        time.Time `json:"createdAt"`
+	UpdatedAt        time.Time `json:"updatedAt"`
+}
+
+// ModelTokenUsageSummary is the aggregate shape used by dashboard and APIs.
+type ModelTokenUsageSummary struct {
+	Events           int64 `json:"events"`
+	ModelCalls       int64 `json:"modelCalls"`
+	PromptTokens     int64 `json:"promptTokens"`
+	CompletionTokens int64 `json:"completionTokens"`
+	TotalTokens      int64 `json:"totalTokens"`
+	CachedTokens     int64 `json:"cachedTokens"`
+	ReasoningTokens  int64 `json:"reasoningTokens"`
+}
+
+// ModelTokenUsageBreakdown is a grouped aggregate row.
+type ModelTokenUsageBreakdown struct {
+	Key              string `json:"key"`
+	Label            string `json:"label,omitempty"`
+	Events           int64  `json:"events"`
+	ModelCalls       int64  `json:"modelCalls"`
+	PromptTokens     int64  `json:"promptTokens"`
+	CompletionTokens int64  `json:"completionTokens"`
+	TotalTokens      int64  `json:"totalTokens"`
+	CachedTokens     int64  `json:"cachedTokens"`
+	ReasoningTokens  int64  `json:"reasoningTokens"`
+}
+
+// ModelTokenUsageStats is a compact API response for usage dashboards.
+type ModelTokenUsageStats struct {
+	Summary         ModelTokenUsageSummary     `json:"summary"`
+	Today           ModelTokenUsageSummary     `json:"today"`
+	ByDay           []ModelTokenUsageBreakdown `json:"byDay"`
+	ByModel         []ModelTokenUsageBreakdown `json:"byModel"`
+	ByOrchestration []ModelTokenUsageBreakdown `json:"byOrchestration"`
+	Recent          []ModelTokenUsage          `json:"recent"`
+}
+
+// ModelTokenUsageFilter scopes usage queries.
+type ModelTokenUsageFilter struct {
+	ConversationID string
+	ProjectID      string
+	Since          time.Time
+	Until          time.Time
+	Days           int
+	Access         RBACListAccess
+	Limit          int
+}
+
+func modelTokenUsageFromProcessDetail(messageID, conversationID, processDetailID string, data interface{}) (ModelTokenUsage, bool) {
+	m := mapFromUsageData(data)
+	if len(m) == 0 {
+		return ModelTokenUsage{}, false
+	}
+	usage := ModelTokenUsage{
+		ID:               uuid.New().String(),
+		ProcessDetailID:  strings.TrimSpace(processDetailID),
+		MessageID:        strings.TrimSpace(messageID),
+		ConversationID:   strings.TrimSpace(conversationID),
+		Source:           strings.TrimSpace(fmt.Sprint(m["source"])),
+		Orchestration:    strings.TrimSpace(fmt.Sprint(m["orchestration"])),
+		Reason:           strings.TrimSpace(fmt.Sprint(m["reason"])),
+		Model:            strings.TrimSpace(fmt.Sprint(m["model"])),
+		ModelCalls:       usageInt64(m["modelCalls"]),
+		PromptTokens:     usageInt64(m["promptTokens"]),
+		CompletionTokens: usageInt64(m["completionTokens"]),
+		TotalTokens:      usageInt64(m["totalTokens"]),
+		CachedTokens:     usageInt64(m["cachedTokens"]),
+		ReasoningTokens:  usageInt64(m["reasoningTokens"]),
+	}
+	if usage.TotalTokens == 0 && (usage.PromptTokens > 0 || usage.CompletionTokens > 0) {
+		usage.TotalTokens = usage.PromptTokens + usage.CompletionTokens
+	}
+	if usage.ProcessDetailID == "" || usage.MessageID == "" || usage.ConversationID == "" {
+		return ModelTokenUsage{}, false
+	}
+	if usage.ModelCalls == 0 && usage.TotalTokens == 0 && usage.PromptTokens == 0 && usage.CompletionTokens == 0 && usage.CachedTokens == 0 && usage.ReasoningTokens == 0 {
+		return ModelTokenUsage{}, false
+	}
+	return usage, true
+}
+
+func mapFromUsageData(data interface{}) map[string]interface{} {
+	switch v := data.(type) {
+	case nil:
+		return nil
+	case map[string]interface{}:
+		return v
+	case string:
+		var m map[string]interface{}
+		if err := json.Unmarshal([]byte(v), &m); err == nil {
+			return m
+		}
+	case []byte:
+		var m map[string]interface{}
+		if err := json.Unmarshal(v, &m); err == nil {
+			return m
+		}
+	default:
+		raw, err := json.Marshal(v)
+		if err == nil {
+			var m map[string]interface{}
+			if err := json.Unmarshal(raw, &m); err == nil {
+				return m
+			}
+		}
+	}
+	return nil
+}
+
+func usageInt64(v interface{}) int64 {
+	switch n := v.(type) {
+	case int:
+		return int64(n)
+	case int8:
+		return int64(n)
+	case int16:
+		return int64(n)
+	case int32:
+		return int64(n)
+	case int64:
+		return n
+	case uint:
+		return int64(n)
+	case uint8:
+		return int64(n)
+	case uint16:
+		return int64(n)
+	case uint32:
+		return int64(n)
+	case uint64:
+		if n > math.MaxInt64 {
+			return math.MaxInt64
+		}
+		return int64(n)
+	case float32:
+		return int64(n)
+	case float64:
+		return int64(n)
+	case json.Number:
+		i, _ := n.Int64()
+		return i
+	case string:
+		i, _ := strconv.ParseInt(strings.TrimSpace(n), 10, 64)
+		return i
+	default:
+		i, _ := strconv.ParseInt(strings.TrimSpace(fmt.Sprint(v)), 10, 64)
+		return i
+	}
+}
+
+func (db *DB) maybeRecordModelTokenUsage(messageID, conversationID, processDetailID, eventType string, data interface{}) {
+	if db == nil || eventType != modelTokenUsageEventType {
+		return
+	}
+	usage, ok := modelTokenUsageFromProcessDetail(messageID, conversationID, processDetailID, data)
+	if !ok {
+		return
+	}
+	if err := db.UpsertModelTokenUsage(usage); err != nil && db.logger != nil {
+		db.logger.Warn("保存模型Token用量失败",
+			zap.String("processDetailId", processDetailID),
+			zap.String("conversationId", conversationID),
+			zap.Error(err))
+	}
+}
+
+// UpsertModelTokenUsage persists usage with process_detail_id idempotency.
+func (db *DB) UpsertModelTokenUsage(usage ModelTokenUsage) error {
+	if db == nil {
+		return fmt.Errorf("database is nil")
+	}
+	now := time.Now()
+	createdAt := usage.CreatedAt
+	if createdAt.IsZero() {
+		createdAt = now
+	}
+	if usage.ID == "" {
+		usage.ID = uuid.New().String()
+	}
+	var projectID sql.NullString
+	if err := db.QueryRow(`SELECT project_id FROM conversations WHERE id = ?`, usage.ConversationID).Scan(&projectID); err != nil && err != sql.ErrNoRows {
+		return fmt.Errorf("查询对话项目失败: %w", err)
+	}
+	projectValue := interface{}(nil)
+	if projectID.Valid && strings.TrimSpace(projectID.String) != "" {
+		projectValue = strings.TrimSpace(projectID.String)
+	}
+	_, err := db.Exec(`
+INSERT INTO model_token_usage (
+	id, process_detail_id, message_id, conversation_id, project_id,
+	source, orchestration, reason, model, model_calls,
+	prompt_tokens, completion_tokens, total_tokens, cached_tokens, reasoning_tokens,
+	created_at, updated_at
+) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+ON CONFLICT(process_detail_id) DO UPDATE SET
+	message_id = excluded.message_id,
+	conversation_id = excluded.conversation_id,
+	project_id = excluded.project_id,
+	source = excluded.source,
+	orchestration = excluded.orchestration,
+	reason = excluded.reason,
+	model = excluded.model,
+	model_calls = excluded.model_calls,
+	prompt_tokens = excluded.prompt_tokens,
+	completion_tokens = excluded.completion_tokens,
+	total_tokens = excluded.total_tokens,
+	cached_tokens = excluded.cached_tokens,
+	reasoning_tokens = excluded.reasoning_tokens,
+	created_at = excluded.created_at,
+	updated_at = excluded.updated_at`,
+		usage.ID, usage.ProcessDetailID, usage.MessageID, usage.ConversationID, projectValue,
+		usage.Source, usage.Orchestration, usage.Reason, usage.Model, usage.ModelCalls,
+		usage.PromptTokens, usage.CompletionTokens, usage.TotalTokens, usage.CachedTokens, usage.ReasoningTokens,
+		createdAt, now,
+	)
+	if err != nil {
+		return fmt.Errorf("写入模型Token用量失败: %w", err)
+	}
+	return nil
+}
+
+// BackfillModelTokenUsageFromProcessDetails makes existing timeline usage events queryable.
+func (db *DB) BackfillModelTokenUsageFromProcessDetails() error {
+	if db == nil {
+		return nil
+	}
+	rows, err := db.Query(`
+SELECT pd.id, pd.message_id, pd.conversation_id, pd.data, pd.created_at
+FROM process_details pd
+LEFT JOIN model_token_usage mtu ON mtu.process_detail_id = pd.id
+WHERE pd.event_type = ?
+	AND (mtu.id IS NULL OR mtu.created_at != pd.created_at)`, modelTokenUsageEventType)
+	if err != nil {
+		return fmt.Errorf("查询历史模型Token用量失败: %w", err)
+	}
+	defer rows.Close()
+	for rows.Next() {
+		var processDetailID, messageID, conversationID string
+		var data sql.NullString
+		var createdAt string
+		if err := rows.Scan(&processDetailID, &messageID, &conversationID, &data, &createdAt); err != nil {
+			return fmt.Errorf("扫描历史模型Token用量失败: %w", err)
+		}
+		if !data.Valid {
+			continue
+		}
+		usage, ok := modelTokenUsageFromProcessDetail(messageID, conversationID, processDetailID, data.String)
+		if !ok {
+			continue
+		}
+		usage.CreatedAt = parseModelTokenUsageTime(createdAt)
+		if err := db.UpsertModelTokenUsage(usage); err != nil {
+			return err
+		}
+	}
+	if err := rows.Err(); err != nil {
+		return fmt.Errorf("遍历历史模型Token用量失败: %w", err)
+	}
+	return nil
+}
+
+func (db *DB) GetModelTokenUsageStats(filter ModelTokenUsageFilter) (*ModelTokenUsageStats, error) {
+	if db == nil {
+		return nil, fmt.Errorf("database is nil")
+	}
+	if filter.Days <= 0 {
+		filter.Days = 7
+	}
+	if filter.Limit <= 0 {
+		filter.Limit = 10
+	}
+	where, args := buildModelTokenUsageWhere(filter, "mtu", "c")
+	summary, err := db.queryModelTokenUsageSummary("SELECT "+modelTokenUsageSummarySelect("mtu")+" FROM model_token_usage mtu JOIN conversations c ON c.id = mtu.conversation_id"+where, args...)
+	if err != nil {
+		return nil, err
+	}
+	todayFilter := filter
+	now := time.Now()
+	todayFilter.Since = time.Date(now.Year(), now.Month(), now.Day(), 0, 0, 0, 0, now.Location())
+	todayWhere, todayArgs := buildModelTokenUsageWhere(todayFilter, "mtu", "c")
+	today, err := db.queryModelTokenUsageSummary("SELECT "+modelTokenUsageSummarySelect("mtu")+" FROM model_token_usage mtu JOIN conversations c ON c.id = mtu.conversation_id"+todayWhere, todayArgs...)
+	if err != nil {
+		return nil, err
+	}
+	byDay, err := db.queryModelTokenUsageBreakdown(
+		"SELECT date(mtu.created_at) AS k, date(mtu.created_at) AS label, "+modelTokenUsageSummarySelect("mtu")+" FROM model_token_usage mtu JOIN conversations c ON c.id = mtu.conversation_id"+where+" GROUP BY date(mtu.created_at) ORDER BY k DESC LIMIT ?",
+		append(args, filter.Days)...,
+	)
+	if err != nil {
+		return nil, err
+	}
+	byModel, err := db.queryModelTokenUsageBreakdown(
+		"SELECT COALESCE(NULLIF(TRIM(mtu.model), ''), 'unknown') AS k, COALESCE(NULLIF(TRIM(mtu.model), ''), 'Unknown') AS label, "+modelTokenUsageSummarySelect("mtu")+" FROM model_token_usage mtu JOIN conversations c ON c.id = mtu.conversation_id"+where+" GROUP BY k ORDER BY SUM(mtu.total_tokens) DESC LIMIT ?",
+		append(args, filter.Limit)...,
+	)
+	if err != nil {
+		return nil, err
+	}
+	byOrch, err := db.queryModelTokenUsageBreakdown(
+		"SELECT COALESCE(NULLIF(TRIM(mtu.orchestration), ''), 'unknown') AS k, COALESCE(NULLIF(TRIM(mtu.orchestration), ''), 'Unknown') AS label, "+modelTokenUsageSummarySelect("mtu")+" FROM model_token_usage mtu JOIN conversations c ON c.id = mtu.conversation_id"+where+" GROUP BY k ORDER BY SUM(mtu.total_tokens) DESC LIMIT ?",
+		append(args, filter.Limit)...,
+	)
+	if err != nil {
+		return nil, err
+	}
+	recent, err := db.ListModelTokenUsage(filter)
+	if err != nil {
+		return nil, err
+	}
+	return &ModelTokenUsageStats{
+		Summary:         summary,
+		Today:           today,
+		ByDay:           byDay,
+		ByModel:         byModel,
+		ByOrchestration: byOrch,
+		Recent:          recent,
+	}, nil
+}
+
+func modelTokenUsageSummarySelect(alias string) string {
+	p := ""
+	if alias != "" {
+		p = alias + "."
+	}
+	return fmt.Sprintf(`COUNT(%sid),
+COALESCE(SUM(%smodel_calls), 0),
+COALESCE(SUM(%sprompt_tokens), 0),
+COALESCE(SUM(%scompletion_tokens), 0),
+COALESCE(SUM(%stotal_tokens), 0),
+COALESCE(SUM(%scached_tokens), 0),
+COALESCE(SUM(%sreasoning_tokens), 0)`, p, p, p, p, p, p, p)
+}
+
+func buildModelTokenUsageWhere(filter ModelTokenUsageFilter, usageAlias, convAlias string) (string, []interface{}) {
+	where := " WHERE 1=1"
+	args := []interface{}{}
+	uPrefix := ""
+	if usageAlias != "" {
+		uPrefix = usageAlias + "."
+	}
+	if cid := strings.TrimSpace(filter.ConversationID); cid != "" {
+		where += " AND " + uPrefix + "conversation_id = ?"
+		args = append(args, cid)
+	}
+	where, args = appendConversationProjectFilter(where, args, filter.ProjectID, usageAlias)
+	if !filter.Since.IsZero() {
+		where += " AND " + uPrefix + "created_at >= ?"
+		args = append(args, filter.Since)
+	}
+	if !filter.Until.IsZero() {
+		where += " AND " + uPrefix + "created_at <= ?"
+		args = append(args, filter.Until)
+	}
+	where, args = appendConversationAccessFilter(where, args, filter.Access.UserID, filter.Access.Scope, convAlias)
+	return where, args
+}
+
+func (db *DB) queryModelTokenUsageSummary(query string, args ...interface{}) (ModelTokenUsageSummary, error) {
+	var s ModelTokenUsageSummary
+	err := db.QueryRow(query, args...).Scan(
+		&s.Events, &s.ModelCalls, &s.PromptTokens, &s.CompletionTokens,
+		&s.TotalTokens, &s.CachedTokens, &s.ReasoningTokens,
+	)
+	if err != nil {
+		return s, fmt.Errorf("查询模型Token用量汇总失败: %w", err)
+	}
+	return s, nil
+}
+
+func (db *DB) queryModelTokenUsageBreakdown(query string, args ...interface{}) ([]ModelTokenUsageBreakdown, error) {
+	rows, err := db.Query(query, args...)
+	if err != nil {
+		return nil, fmt.Errorf("查询模型Token用量分组失败: %w", err)
+	}
+	defer rows.Close()
+	out := []ModelTokenUsageBreakdown{}
+	for rows.Next() {
+		var row ModelTokenUsageBreakdown
+		if err := rows.Scan(
+			&row.Key, &row.Label, &row.Events, &row.ModelCalls, &row.PromptTokens,
+			&row.CompletionTokens, &row.TotalTokens, &row.CachedTokens, &row.ReasoningTokens,
+		); err != nil {
+			return nil, fmt.Errorf("扫描模型Token用量分组失败: %w", err)
+		}
+		out = append(out, row)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, fmt.Errorf("遍历模型Token用量分组失败: %w", err)
+	}
+	return out, nil
+}
+
+func (db *DB) ListModelTokenUsage(filter ModelTokenUsageFilter) ([]ModelTokenUsage, error) {
+	if filter.Limit <= 0 {
+		filter.Limit = 20
+	}
+	if filter.Limit > 500 {
+		filter.Limit = 500
+	}
+	where, args := buildModelTokenUsageWhere(filter, "mtu", "c")
+	args = append(args, filter.Limit)
+	rows, err := db.Query(`
+SELECT mtu.id, mtu.process_detail_id, mtu.message_id, mtu.conversation_id,
+	COALESCE(mtu.project_id, ''), mtu.source, mtu.orchestration, mtu.reason, mtu.model,
+	mtu.model_calls, mtu.prompt_tokens, mtu.completion_tokens, mtu.total_tokens,
+	mtu.cached_tokens, mtu.reasoning_tokens, mtu.created_at, mtu.updated_at
+FROM model_token_usage mtu
+JOIN conversations c ON c.id = mtu.conversation_id`+where+`
+ORDER BY mtu.created_at DESC, mtu.rowid DESC
+LIMIT ?`, args...)
+	if err != nil {
+		return nil, fmt.Errorf("查询模型Token用量明细失败: %w", err)
+	}
+	defer rows.Close()
+	out := []ModelTokenUsage{}
+	for rows.Next() {
+		var u ModelTokenUsage
+		var createdAt, updatedAt string
+		if err := rows.Scan(
+			&u.ID, &u.ProcessDetailID, &u.MessageID, &u.ConversationID, &u.ProjectID,
+			&u.Source, &u.Orchestration, &u.Reason, &u.Model, &u.ModelCalls,
+			&u.PromptTokens, &u.CompletionTokens, &u.TotalTokens, &u.CachedTokens,
+			&u.ReasoningTokens, &createdAt, &updatedAt,
+		); err != nil {
+			return nil, fmt.Errorf("扫描模型Token用量明细失败: %w", err)
+		}
+		u.CreatedAt = parseModelTokenUsageTime(createdAt)
+		u.UpdatedAt = parseModelTokenUsageTime(updatedAt)
+		out = append(out, u)
+	}
+	if err := rows.Err(); err != nil {
+		return nil, fmt.Errorf("遍历模型Token用量明细失败: %w", err)
+	}
+	return out, nil
+}
+
+func parseModelTokenUsageTime(s string) time.Time {
+	for _, layout := range []string{
+		"2006-01-02 15:04:05.999999999-07:00",
+		"2006-01-02 15:04:05.999999-07:00",
+		"2006-01-02 15:04:05",
+		time.RFC3339Nano,
+		time.RFC3339,
+	} {
+		if t, err := time.Parse(layout, strings.TrimSpace(s)); err == nil {
+			return t
+		}
+	}
+	return time.Time{}
+}

internal/database/model_token_usage_test.go

@@ -0,0 +1,85 @@
+package database
+
+import (
+	"path/filepath"
+	"testing"
+
+	"go.uber.org/zap"
+)
+
+func TestModelTokenUsagePersistsFromUsageProcessDetail(t *testing.T) {
+	db := newModelTokenUsageTestDB(t)
+	conv, err := db.CreateConversation("usage", ConversationCreateMeta{})
+	if err != nil {
+		t.Fatalf("CreateConversation: %v", err)
+	}
+	msg, err := db.AddMessage(conv.ID, "assistant", "done", nil)
+	if err != nil {
+		t.Fatalf("AddMessage: %v", err)
+	}
+	if err := db.AddProcessDetail(msg.ID, conv.ID, modelTokenUsageEventType, "usage", map[string]interface{}{
+		"source":           "eino",
+		"orchestration":    "deep",
+		"reason":           "final",
+		"model":            "gpt-test",
+		"modelCalls":       2,
+		"promptTokens":     10,
+		"completionTokens": 3,
+		"totalTokens":      13,
+		"cachedTokens":     4,
+		"reasoningTokens":  1,
+	}); err != nil {
+		t.Fatalf("AddProcessDetail: %v", err)
+	}
+
+	stats, err := db.GetModelTokenUsageStats(ModelTokenUsageFilter{})
+	if err != nil {
+		t.Fatalf("GetModelTokenUsageStats: %v", err)
+	}
+	if stats.Summary.Events != 1 || stats.Summary.ModelCalls != 2 || stats.Summary.TotalTokens != 13 || stats.Summary.CachedTokens != 4 || stats.Summary.ReasoningTokens != 1 {
+		t.Fatalf("summary = %#v", stats.Summary)
+	}
+	if len(stats.ByModel) != 1 || stats.ByModel[0].Key != "gpt-test" || stats.ByModel[0].TotalTokens != 13 {
+		t.Fatalf("by model = %#v", stats.ByModel)
+	}
+}
+
+func TestModelTokenUsageBackfillIsIdempotent(t *testing.T) {
+	db := newModelTokenUsageTestDB(t)
+	conv, err := db.CreateConversation("usage", ConversationCreateMeta{})
+	if err != nil {
+		t.Fatalf("CreateConversation: %v", err)
+	}
+	msg, err := db.AddMessage(conv.ID, "assistant", "done", nil)
+	if err != nil {
+		t.Fatalf("AddMessage: %v", err)
+	}
+	if err := db.AddProcessDetail(msg.ID, conv.ID, modelTokenUsageEventType, "usage", map[string]interface{}{
+		"source": "eino", "modelCalls": 1, "promptTokens": 7, "completionTokens": 5, "totalTokens": 12,
+	}); err != nil {
+		t.Fatalf("AddProcessDetail: %v", err)
+	}
+	if err := db.BackfillModelTokenUsageFromProcessDetails(); err != nil {
+		t.Fatalf("Backfill 1: %v", err)
+	}
+	if err := db.BackfillModelTokenUsageFromProcessDetails(); err != nil {
+		t.Fatalf("Backfill 2: %v", err)
+	}
+	stats, err := db.GetModelTokenUsageStats(ModelTokenUsageFilter{})
+	if err != nil {
+		t.Fatalf("GetModelTokenUsageStats: %v", err)
+	}
+	if stats.Summary.Events != 1 || stats.Summary.TotalTokens != 12 {
+		t.Fatalf("summary after backfill = %#v", stats.Summary)
+	}
+}
+
+func newModelTokenUsageTestDB(t *testing.T) *DB {
+	t.Helper()
+	db, err := NewDB(filepath.Join(t.TempDir(), "usage.db"), zap.NewNop())
+	if err != nil {
+		t.Fatalf("NewDB: %v", err)
+	}
+	t.Cleanup(func() { _ = db.Close() })
+	return db
+}

internal/database/process_details_summary_test.go

@@ -165,6 +165,32 @@ func TestProcessDetailsSummaryDoesNotReportPersistedOrphanAsRunning(t *testing.T
 	}
 }
 
+func TestProcessDetailsSummaryReportsUnmatchedToolCallAsRunningForActiveTurn(t *testing.T) {
+	db, conversationID, messageID := setupProcessDetailsSummaryTest(t)
+	if _, err := db.Exec(
+		"UPDATE messages SET content = ?, updated_at = ? WHERE id = ?",
+		"处理中...", "2026-08-10T08:00:00Z", messageID,
+	); err != nil {
+		t.Fatalf("update running message: %v", err)
+	}
+	if err := db.AddProcessDetail(messageID, conversationID, "tool_call", "call", map[string]interface{}{
+		"toolName": "execute", "toolCallId": "pending",
+	}); err != nil {
+		t.Fatalf("AddProcessDetail(tool_call): %v", err)
+	}
+
+	summary, err := db.GetProcessDetailsSummary(messageID)
+	if err != nil {
+		t.Fatalf("GetProcessDetailsSummary: %v", err)
+	}
+	if summary.Status != "running" {
+		t.Fatalf("summary status = %q, want running", summary.Status)
+	}
+	if len(summary.ToolExecutions) != 1 || summary.ToolExecutions[0].Status != "running" {
+		t.Fatalf("tool executions = %#v, want running", summary.ToolExecutions)
+	}
+}
+
 func TestProcessDetailsSummaryIncludesPersistedTurnTiming(t *testing.T) {
 	db, _, messageID := setupProcessDetailsSummaryTest(t)
 	startedAt := "2026-08-10T08:00:00Z"

internal/handler/agent.go

@@ -698,18 +698,19 @@ func (h *AgentHandler) mergeAssistantMessagePartialOnCancel(messageID, partial s
 
 // ChatResponse 聊天响应
 type ChatResponse struct {
-	Response            string    `json:"response"`
+	Response                         string    `json:"response"`
-	MCPExecutionIDs     []string  `json:"mcpExecutionIds,omitempty"` // 本次对话中执行的MCP调用ID列表
+	MCPExecutionIDs                  []string  `json:"mcpExecutionIds,omitempty"` // 本次对话中执行的MCP调用ID列表
-	ConversationID      string    `json:"conversationId"`            // 对话ID
+	ConversationID                   string    `json:"conversationId"`            // 对话ID
-	Time                time.Time `json:"time"`
+	Time                             time.Time `json:"time"`
-	Finalizable         bool      `json:"finalizable"`
+	Finalizable                      bool      `json:"finalizable"`
-	Finalized           bool      `json:"finalized"`
+	Finalized                        bool      `json:"finalized"`
-	Status              string    `json:"status,omitempty"`
+	Status                           string    `json:"status,omitempty"`
-	CompletionReason    string    `json:"completionReason,omitempty"`
+	CompletionReason                 string    `json:"completionReason,omitempty"`
-	EvidenceVerified    bool      `json:"evidenceVerified"`
+	EvidenceVerified                 bool      `json:"evidenceVerified"`
-	EvidenceRefs        []string  `json:"evidenceRefs,omitempty"`
+	EvidenceRefs                     []string  `json:"evidenceRefs,omitempty"`
-	PendingExecutionIDs []string  `json:"pendingExecutionIds,omitempty"`
+	PendingExecutionIDs              []string  `json:"pendingExecutionIds,omitempty"`
-	MissingChecks       []string  `json:"missingChecks,omitempty"`
+	MissingChecks                    []string  `json:"missingChecks,omitempty"`
+	AutoCancelledPendingExecutionIDs []string  `json:"autoCancelledPendingExecutionIds,omitempty"`
 }
 
 func (h *AgentHandler) finalizeRobotAgentError(ctx context.Context, assistantMessageID, conversationID string, resultMA *multiagent.RunResult, errMA error) (string, string, error) {
@@ -724,8 +725,13 @@ func (h *AgentHandler) finalizeRobotAgentError(ctx context.Context, assistantMes
 	return "", conversationID, errMA
 }
 
-func (h *AgentHandler) finalizeRobotAgentSuccess(assistantMessageID, conversationID string, resultMA *multiagent.RunResult) (string, string, error) {
+func (h *AgentHandler) finalizeRobotAgentSuccess(taskCtx context.Context, assistantMessageID, conversationID string, resultMA *multiagent.RunResult) (string, string, error) {
-	decision := h.finalizeAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, "robot", resultMA, resultMA.MCPExecutionIDs, multiagent.AggregatedReasoningFromTraceJSON(resultMA.LastAgentTraceInput), true)
+	reasoningContent := multiagent.AggregatedReasoningFromTraceJSON(resultMA.LastAgentTraceInput)
+	decision := h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, "robot", resultMA, resultMA.MCPExecutionIDs, true)
+	if cancelled := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, conversationID, decision, nil); len(cancelled) > 0 {
+		decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, "robot", resultMA, resultMA.MCPExecutionIDs, true)
+	}
+	h.persistFinalizationDecision(conversationID, assistantMessageID, "robot", resultMA.MCPExecutionIDs, reasoningContent, decision)
 	responseText := decision.FinalText
 	if !decision.Finalizable {
 		responseText = finalizationBlockedMessage(decision)
@@ -758,7 +764,7 @@ func (h *AgentHandler) runRobotEinoSingleWithRetry(
 		*taskStatus = "failed"
 		return h.finalizeRobotAgentError(taskCtx, assistantMessageID, conversationID, resultMA, errMA)
 	}
-	return h.finalizeRobotAgentSuccess(assistantMessageID, conversationID, resultMA)
+	return h.finalizeRobotAgentSuccess(taskCtx, assistantMessageID, conversationID, resultMA)
 }
 
 func (h *AgentHandler) runRobotMultiAgentWithRetry(
@@ -779,7 +785,7 @@ func (h *AgentHandler) runRobotMultiAgentWithRetry(
 		*taskStatus = "failed"
 		return h.finalizeRobotAgentError(taskCtx, assistantMessageID, conversationID, resultMA, errMA)
 	}
-	return h.finalizeRobotAgentSuccess(assistantMessageID, conversationID, resultMA)
+	return h.finalizeRobotAgentSuccess(taskCtx, assistantMessageID, conversationID, resultMA)
 }
 
 // ProcessMessageForRobot 供机器人（企业微信/钉钉/飞书）调用：Eino 单/多代理执行路径（含 progressCallback、过程详情），仅不发送 SSE，最后返回完整回复

internal/handler/batch_queue_executor.go

@@ -281,7 +281,12 @@ func (h *AgentHandler) executeOneBatchSubTask(queueID string, queue *BatchTaskQu
 	if useBatchMulti {
 		agentMode = "batch_eino_" + batchOrch
 	}
-	decision := h.finalizeAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, agentMode, resultMA, mcpIDs, reasoningContent, true)
+	decision := h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, agentMode, resultMA, mcpIDs, true)
+	autoCancelledPendingExecutionIDs := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, conversationID, decision, progressCallback)
+	if len(autoCancelledPendingExecutionIDs) > 0 {
+		decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, agentMode, resultMA, mcpIDs, true)
+	}
+	h.persistFinalizationDecision(conversationID, assistantMessageID, agentMode, mcpIDs, reasoningContent, decision)
 	resText := decision.FinalText
 	if !decision.Finalizable {
 		resText = finalizationBlockedMessage(decision)
@@ -289,14 +294,15 @@ func (h *AgentHandler) executeOneBatchSubTask(queueID string, queue *BatchTaskQu
 		sendEvent("finalization_check", resText, decision)
 	}
 	sendEvent("response", resText, finalizationResponsePayload(decision, map[string]interface{}{
-		"conversationId":   conversationID,
+		"conversationId":                   conversationID,
-		"messageId":        assistantMessageID,
+		"messageId":                        assistantMessageID,
-		"agentMode":        agentMode,
+		"agentMode":                        agentMode,
-		"mcpExecutionIds":  mcpIDs,
+		"mcpExecutionIds":                  mcpIDs,
-		"batchQueueId":     queueID,
+		"batchQueueId":                     queueID,
-		"batchTaskId":      task.ID,
+		"batchTaskId":                      task.ID,
-		"batchTaskStatus":  map[bool]string{true: string(BatchTaskStatusCompleted), false: string(BatchTaskStatusFailed)}[decision.Finalizable],
+		"batchTaskStatus":                  map[bool]string{true: string(BatchTaskStatusCompleted), false: string(BatchTaskStatusFailed)}[decision.Finalizable],
-		"candidatePreview": safeTruncateString(resultMA.Response, 500),
+		"candidatePreview":                 safeTruncateString(resultMA.Response, 500),
+		"autoCancelledPendingExecutionIds": autoCancelledPendingExecutionIDs,
 	}))
 
 	if assistantMessageID == "" {

internal/handler/conversation_process_details_test.go

@@ -76,6 +76,65 @@ func TestProcessDetailsPageIncludesTerminalToolStatusAcrossPageBoundary(t *testi
 	}
 }
 
+func TestProcessDetailsPageUsesPersistedExecutionStatusAfterBackgroundCancel(t *testing.T) {
+	gin.SetMode(gin.TestMode)
+	db, err := database.NewDB(filepath.Join(t.TempDir(), "process-details-cancelled.db"), zap.NewNop())
+	if err != nil {
+		t.Fatalf("NewDB: %v", err)
+	}
+	t.Cleanup(func() { _ = db.Close() })
+	conversation, err := db.CreateConversation("cancelled background", database.ConversationCreateMeta{})
+	if err != nil {
+		t.Fatalf("CreateConversation: %v", err)
+	}
+	message, err := db.AddMessage(conversation.ID, "assistant", "done", nil)
+	if err != nil {
+		t.Fatalf("AddMessage: %v", err)
+	}
+	execID := "exec-cancelled-after-background"
+	if err := db.AddProcessDetail(message.ID, conversation.ID, "tool_call", "call", map[string]interface{}{
+		"toolName": "exec", "toolCallId": "call-cancelled", "index": 1, "total": 1,
+	}); err != nil {
+		t.Fatalf("AddProcessDetail(tool_call): %v", err)
+	}
+	if err := db.AddProcessDetail(message.ID, conversation.ID, "tool_result", "background", map[string]interface{}{
+		"toolName": "exec", "toolCallId": "call-cancelled", "executionId": execID, "status": "background_running", "success": true,
+	}); err != nil {
+		t.Fatalf("AddProcessDetail(tool_result): %v", err)
+	}
+	now := time.Now()
+	if err := db.SaveToolExecution(&mcp.ToolExecution{
+		ID:        execID,
+		ToolName:  "exec",
+		Status:    mcp.ToolExecutionStatusCancelled,
+		StartTime: now,
+		EndTime:   &now,
+	}); err != nil {
+		t.Fatalf("SaveToolExecution: %v", err)
+	}
+
+	w := httptest.NewRecorder()
+	c, _ := gin.CreateTestContext(w)
+	c.Request = httptest.NewRequest("GET", "/api/messages/"+message.ID+"/process-details?limit=10&offset=0", nil)
+	c.Params = gin.Params{{Key: "id", Value: message.ID}}
+	NewConversationHandler(db, zap.NewNop()).GetMessageProcessDetails(c)
+	if w.Code != 200 {
+		t.Fatalf("status = %d: %s", w.Code, w.Body.String())
+	}
+	var response struct {
+		ToolExecutions []database.ProcessDetailsToolExecution `json:"toolExecutions"`
+	}
+	if err := json.Unmarshal(w.Body.Bytes(), &response); err != nil {
+		t.Fatalf("decode response: %v", err)
+	}
+	if len(response.ToolExecutions) != 1 {
+		t.Fatalf("tool executions = %d, want 1", len(response.ToolExecutions))
+	}
+	if got := response.ToolExecutions[0].Status; got != mcp.ToolExecutionStatusCancelled {
+		t.Fatalf("tool execution status = %q, want cancelled", got)
+	}
+}
+
 func TestProcessDetailsFullBackfillsEmptyToolCallArgumentsFromExecution(t *testing.T) {
 	gin.SetMode(gin.TestMode)
 	db, err := database.NewDB(filepath.Join(t.TempDir(), "process-details-args.db"), zap.NewNop())

internal/handler/eino_single_agent.go

@@ -192,6 +192,7 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 	var emptyResponseContinueAttempt int
 	var finalizationAutoContinueAttempt int
 	var decision agentfinalizer.Decision
+	var autoCancelledPendingExecutionIDs []string
 
 	for {
 		segmentMainIterationMax := 0
@@ -268,6 +269,10 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 				continue
 			}
 			decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, "eino_single", result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+			if cancelled := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, conversationID, decision, progressCallback); len(cancelled) > 0 {
+				autoCancelledPendingExecutionIDs = mergeMCPExecutionIDLists(autoCancelledPendingExecutionIDs, cancelled)
+				decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, "eino_single", result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+			}
 			if h.tryAutoContinueAfterFinalization(taskCtx, conversationID, result, decision, &finalizationAutoContinueAttempt, &curHistory, &curFinalMessage, progressCallback) {
 				mainIterationOffset += segmentMainIterationMax
 				timeoutCancel()
@@ -384,6 +389,10 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 
 	if decision.CompletionReason == "" {
 		decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, "eino_single", result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		if cancelled := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, conversationID, decision, nil); len(cancelled) > 0 {
+			autoCancelledPendingExecutionIDs = mergeMCPExecutionIDLists(autoCancelledPendingExecutionIDs, cancelled)
+			decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, "eino_single", result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		}
 	}
 	h.persistFinalizationDecision(conversationID, assistantMessageID, "eino_single", cumulativeMCPExecutionIDs, multiagent.AggregatedReasoningFromTraceJSON(result.LastAgentTraceInput), decision)
 
@@ -401,10 +410,11 @@ func (h *AgentHandler) EinoSingleAgentLoopStream(c *gin.Context) {
 		h.tasks.UpdateTaskStatus(conversationID, taskStatus)
 	}
 	sendEvent("response", responseText, finalizationResponsePayload(decision, map[string]interface{}{
-		"mcpExecutionIds": cumulativeMCPExecutionIDs,
+		"mcpExecutionIds":                  cumulativeMCPExecutionIDs,
-		"conversationId":  conversationID,
+		"conversationId":                   conversationID,
-		"messageId":       assistantMessageID,
+		"messageId":                        assistantMessageID,
-		"agentMode":       "eino_single",
+		"agentMode":                        "eino_single",
+		"autoCancelledPendingExecutionIds": autoCancelledPendingExecutionIDs,
 	}))
 	sendEvent("done", "", map[string]interface{}{"conversationId": conversationID})
 }
@@ -464,6 +474,7 @@ func (h *AgentHandler) EinoSingleAgentLoop(c *gin.Context) {
 	var emptyResponseContinueAttempt int
 	var finalizationAutoContinueAttempt int
 	var decision agentfinalizer.Decision
+	var autoCancelledPendingExecutionIDs []string
 	for {
 		result, runErr = multiagent.RunEinoSingleChatModelAgent(
 			taskCtx,
@@ -493,6 +504,10 @@ func (h *AgentHandler) EinoSingleAgentLoop(c *gin.Context) {
 			continue
 		}
 		decision = h.decideAgentRunForDeliveryWithPolicy(prep.ConversationID, prep.AssistantMessageID, "eino_single", result, result.MCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		if cancelled := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, prep.ConversationID, decision, progressCallback); len(cancelled) > 0 {
+			autoCancelledPendingExecutionIDs = mergeMCPExecutionIDLists(autoCancelledPendingExecutionIDs, cancelled)
+			decision = h.decideAgentRunForDeliveryWithPolicy(prep.ConversationID, prep.AssistantMessageID, "eino_single", result, result.MCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		}
 		if h.tryAutoContinueAfterFinalization(taskCtx, prep.ConversationID, result, decision, &finalizationAutoContinueAttempt, &curHist, &curMsg, progressCallback) {
 			continue
 		}
@@ -509,18 +524,19 @@ func (h *AgentHandler) EinoSingleAgentLoop(c *gin.Context) {
 		responseText = finalizationBlockedMessage(decision)
 	}
 	c.JSON(http.StatusOK, gin.H{
-		"response":            responseText,
+		"response":                         responseText,
-		"conversationId":      prep.ConversationID,
+		"conversationId":                   prep.ConversationID,
-		"mcpExecutionIds":     result.MCPExecutionIDs,
+		"mcpExecutionIds":                  result.MCPExecutionIDs,
-		"assistantMessageId":  prep.AssistantMessageID,
+		"assistantMessageId":               prep.AssistantMessageID,
-		"agentMode":           "eino_single",
+		"agentMode":                        "eino_single",
-		"finalized":           decision.Finalized,
+		"finalized":                        decision.Finalized,
-		"finalizable":         decision.Finalizable,
+		"finalizable":                      decision.Finalizable,
-		"status":              decision.Status,
+		"status":                           decision.Status,
-		"completionReason":    decision.CompletionReason,
+		"completionReason":                 decision.CompletionReason,
-		"evidenceVerified":    decision.EvidenceVerified,
+		"evidenceVerified":                 decision.EvidenceVerified,
-		"evidenceRefs":        decision.EvidenceRefs,
+		"evidenceRefs":                     decision.EvidenceRefs,
-		"pendingExecutionIds": decision.PendingExecutionIDs,
+		"pendingExecutionIds":              decision.PendingExecutionIDs,
-		"missingChecks":       decision.MissingChecks,
+		"missingChecks":                    decision.MissingChecks,
+		"autoCancelledPendingExecutionIds": autoCancelledPendingExecutionIDs,
 	})
 }

internal/handler/finalization_auto_continue.go

@@ -2,16 +2,21 @@ package handler
 
 import (
 	"context"
+	"strings"
 	"time"
 
 	"cyberstrike-ai/internal/agent"
 	"cyberstrike-ai/internal/agentfinalizer"
+	"cyberstrike-ai/internal/mcp"
 	"cyberstrike-ai/internal/multiagent"
 
 	"go.uber.org/zap"
 )
 
 const finalizationAutoContinueMaxAttempts = 2
+const finalizationPendingToolCancelWait = 2 * time.Second
+const finalizationPendingToolCancelPoll = 50 * time.Millisecond
+const finalizationPendingToolCancelNote = "Agent 迭代已结束，最终回复前自动终止未完成的工具执行"
 
 func shouldAutoContinueAfterFinalization(d agentfinalizer.Decision, attempt int) bool {
 	if d.Finalizable || d.Finalized {
@@ -75,3 +80,105 @@ func finalizationAutoContinueBackoff(attempt int) time.Duration {
 	}
 	return time.Duration(attempt) * time.Second
 }
+
+func (h *AgentHandler) cleanupPendingToolExecutionsAfterIteration(
+	taskCtx context.Context,
+	conversationID string,
+	decision agentfinalizer.Decision,
+	progressCallback func(eventType, message string, data interface{}),
+) []string {
+	if h == nil || h.agent == nil || decision.CompletionReason != agentfinalizer.ReasonPendingTools {
+		return nil
+	}
+	pending := uniqueNonEmptyStrings(decision.PendingExecutionIDs)
+	if len(pending) == 0 {
+		return nil
+	}
+	cancelled := make([]string, 0, len(pending))
+	for _, executionID := range pending {
+		if h.agent.CancelMCPToolExecutionWithNote(executionID, finalizationPendingToolCancelNote) {
+			cancelled = append(cancelled, executionID)
+		} else if h.logger != nil {
+			h.logger.Warn("finalization pending tool cleanup could not cancel execution",
+				zap.String("conversationId", conversationID),
+				zap.String("executionId", executionID))
+		}
+	}
+	if len(cancelled) == 0 {
+		return nil
+	}
+	if progressCallback != nil {
+		progressCallback("finalization_pending_tools_cancelled", "迭代结束，已自动终止仍在运行的工具执行。", map[string]interface{}{
+			"conversationId":                   conversationID,
+			"source":                           "finalizer",
+			"autoCancelledPendingExecutionIds": cancelled,
+			"pendingExecutionIds":              pending,
+			"reason":                           agentfinalizer.ReasonPendingTools,
+		})
+	}
+	h.waitForToolExecutionsToLeavePending(taskCtx, cancelled, finalizationPendingToolCancelWait)
+	return cancelled
+}
+
+func (h *AgentHandler) waitForToolExecutionsToLeavePending(ctx context.Context, executionIDs []string, wait time.Duration) {
+	if h == nil || h.db == nil || len(executionIDs) == 0 || wait <= 0 {
+		return
+	}
+	timer := time.NewTimer(wait)
+	defer timer.Stop()
+	ticker := time.NewTicker(finalizationPendingToolCancelPoll)
+	defer ticker.Stop()
+	for {
+		if !h.hasPendingToolExecutions(executionIDs) {
+			return
+		}
+		select {
+		case <-contextDone(ctx):
+			return
+		case <-timer.C:
+			return
+		case <-ticker.C:
+		}
+	}
+}
+
+func (h *AgentHandler) hasPendingToolExecutions(executionIDs []string) bool {
+	if h == nil || h.db == nil {
+		return false
+	}
+	for _, executionID := range uniqueNonEmptyStrings(executionIDs) {
+		exec, err := h.db.GetToolExecution(executionID)
+		if err != nil || exec == nil {
+			continue
+		}
+		switch strings.TrimSpace(exec.Status) {
+		case mcp.ToolExecutionStatusQueued, mcp.ToolExecutionStatusRunning:
+			return true
+		}
+	}
+	return false
+}
+
+func uniqueNonEmptyStrings(values []string) []string {
+	seen := make(map[string]struct{}, len(values))
+	out := make([]string, 0, len(values))
+	for _, value := range values {
+		value = strings.TrimSpace(value)
+		if value == "" {
+			continue
+		}
+		if _, ok := seen[value]; ok {
+			continue
+		}
+		seen[value] = struct{}{}
+		out = append(out, value)
+	}
+	return out
+}
+
+func contextDone(ctx context.Context) <-chan struct{} {
+	if ctx == nil {
+		return nil
+	}
+	return ctx.Done()
+}

internal/handler/finalization_auto_continue_test.go

@@ -1,9 +1,18 @@
 package handler
 
 import (
+	"context"
+	"path/filepath"
 	"testing"
+	"time"
 
+	agentpkg "cyberstrike-ai/internal/agent"
 	"cyberstrike-ai/internal/agentfinalizer"
+	"cyberstrike-ai/internal/config"
+	"cyberstrike-ai/internal/database"
+	"cyberstrike-ai/internal/mcp"
+
+	"go.uber.org/zap"
 )
 
 func TestShouldAutoContinueAfterFinalization(t *testing.T) {
@@ -57,3 +66,66 @@ func TestRequestRequiresExecutionEvidenceUsesExplicitPolicyOnly(t *testing.T) {
 		t.Fatal("explicit false policy should not require execution evidence")
 	}
 }
+
+func TestCleanupPendingToolExecutionsAfterIterationAllowsFinalization(t *testing.T) {
+	logger := zap.NewNop()
+	db, err := database.NewDB(filepath.Join(t.TempDir(), "cleanup-finalization.db"), logger)
+	if err != nil {
+		t.Fatalf("NewDB: %v", err)
+	}
+	t.Cleanup(func() { _ = db.Close() })
+
+	server := mcp.NewServerWithStorage(logger, db)
+	server.ConfigureToolWaitTimeoutSeconds(1)
+	server.RegisterTool(mcp.Tool{Name: "block", InputSchema: map[string]interface{}{"type": "object"}}, func(ctx context.Context, args map[string]interface{}) (*mcp.ToolResult, error) {
+		<-ctx.Done()
+		return nil, ctx.Err()
+	})
+	ag := agentpkg.NewAgent(&config.OpenAIConfig{}, &config.AgentConfig{}, server, nil, logger, 10)
+	h := &AgentHandler{agent: ag, db: db, logger: logger}
+
+	callCtx := mcp.WithMCPConversationID(context.Background(), "conv-cleanup")
+	result, execID, err := server.CallTool(callCtx, "block", nil)
+	if err != nil {
+		t.Fatalf("CallTool: %v", err)
+	}
+	if result == nil || !result.IsError || execID == "" {
+		t.Fatalf("expected background wait result, result=%#v execID=%q", result, execID)
+	}
+
+	decision := agentfinalizer.Decide(db, agentfinalizer.Input{
+		Response:        "基于已完成信息的阶段性总结。",
+		MCPExecutionIDs: []string{execID},
+	})
+	if decision.CompletionReason != agentfinalizer.ReasonPendingTools {
+		t.Fatalf("decision reason = %s, want pending tools: %+v", decision.CompletionReason, decision)
+	}
+
+	var eventType string
+	cancelled := h.cleanupPendingToolExecutionsAfterIteration(context.Background(), "conv-cleanup", decision, func(et, _ string, _ interface{}) {
+		eventType = et
+	})
+	if len(cancelled) != 1 || cancelled[0] != execID {
+		t.Fatalf("cancelled = %#v, want [%s]", cancelled, execID)
+	}
+	if eventType != "finalization_pending_tools_cancelled" {
+		t.Fatalf("event type = %q", eventType)
+	}
+
+	deadline := time.Now().Add(time.Second)
+	for time.Now().Before(deadline) {
+		exec, err := db.GetToolExecution(execID)
+		if err == nil && exec != nil && exec.Status == mcp.ToolExecutionStatusCancelled {
+			after := agentfinalizer.Decide(db, agentfinalizer.Input{
+				Response:        "基于已完成信息的阶段性总结。",
+				MCPExecutionIDs: []string{execID},
+			})
+			if !after.Finalizable || !after.Finalized {
+				t.Fatalf("decision should finalize after cleanup: %+v", after)
+			}
+			return
+		}
+		time.Sleep(10 * time.Millisecond)
+	}
+	t.Fatal("execution did not become cancelled")
+}

internal/handler/multi_agent.go

@@ -205,6 +205,7 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 	}
 	agentMode := "eino_" + effectiveOrch
 	var decision agentfinalizer.Decision
+	var autoCancelledPendingExecutionIDs []string
 
 	for {
 		segmentMainIterationMax := 0
@@ -282,6 +283,10 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 				continue
 			}
 			decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, agentMode, result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+			if cancelled := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, conversationID, decision, progressCallback); len(cancelled) > 0 {
+				autoCancelledPendingExecutionIDs = mergeMCPExecutionIDLists(autoCancelledPendingExecutionIDs, cancelled)
+				decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, agentMode, result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+			}
 			if h.tryAutoContinueAfterFinalization(taskCtx, conversationID, result, decision, &finalizationAutoContinueAttempt, &curHistory, &curFinalMessage, progressCallback) {
 				mainIterationOffset += segmentMainIterationMax
 				timeoutCancel()
@@ -398,6 +403,10 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 
 	if decision.CompletionReason == "" {
 		decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, agentMode, result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		if cancelled := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, conversationID, decision, nil); len(cancelled) > 0 {
+			autoCancelledPendingExecutionIDs = mergeMCPExecutionIDLists(autoCancelledPendingExecutionIDs, cancelled)
+			decision = h.decideAgentRunForDeliveryWithPolicy(conversationID, assistantMessageID, agentMode, result, cumulativeMCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		}
 	}
 	h.persistFinalizationDecision(conversationID, assistantMessageID, agentMode, cumulativeMCPExecutionIDs, multiagent.AggregatedReasoningFromTraceJSON(result.LastAgentTraceInput), decision)
 
@@ -415,10 +424,11 @@ func (h *AgentHandler) MultiAgentLoopStream(c *gin.Context) {
 		h.tasks.UpdateTaskStatus(conversationID, taskStatus)
 	}
 	sendEvent("response", responseText, finalizationResponsePayload(decision, map[string]interface{}{
-		"mcpExecutionIds": cumulativeMCPExecutionIDs,
+		"mcpExecutionIds":                  cumulativeMCPExecutionIDs,
-		"conversationId":  conversationID,
+		"conversationId":                   conversationID,
-		"messageId":       assistantMessageID,
+		"messageId":                        assistantMessageID,
-		"agentMode":       agentMode,
+		"agentMode":                        agentMode,
+		"autoCancelledPendingExecutionIds": autoCancelledPendingExecutionIDs,
 	}))
 	sendEvent("done", "", map[string]interface{}{"conversationId": conversationID})
 }
@@ -478,6 +488,7 @@ func (h *AgentHandler) MultiAgentLoop(c *gin.Context) {
 	}
 	agentMode := "eino_" + effectiveOrch
 	var decision agentfinalizer.Decision
+	var autoCancelledPendingExecutionIDs []string
 	for {
 		result, runErr = multiagent.RunDeepAgent(
 			taskCtx,
@@ -514,6 +525,10 @@ func (h *AgentHandler) MultiAgentLoop(c *gin.Context) {
 			continue
 		}
 		decision = h.decideAgentRunForDeliveryWithPolicy(prep.ConversationID, prep.AssistantMessageID, agentMode, result, result.MCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		if cancelled := h.cleanupPendingToolExecutionsAfterIteration(taskCtx, prep.ConversationID, decision, progressCallback); len(cancelled) > 0 {
+			autoCancelledPendingExecutionIDs = mergeMCPExecutionIDLists(autoCancelledPendingExecutionIDs, cancelled)
+			decision = h.decideAgentRunForDeliveryWithPolicy(prep.ConversationID, prep.AssistantMessageID, agentMode, result, result.MCPExecutionIDs, requestRequiresExecutionEvidence(&req))
+		}
 		if h.tryAutoContinueAfterFinalization(taskCtx, prep.ConversationID, result, decision, &finalizationAutoContinueAttempt, &curHist, &curMsg, progressCallback) {
 			continue
 		}
@@ -533,18 +548,19 @@ func (h *AgentHandler) MultiAgentLoop(c *gin.Context) {
 		responseText = finalizationBlockedMessage(decision)
 	}
 	c.JSON(http.StatusOK, ChatResponse{
-		Response:            responseText,
+		Response:                         responseText,
-		MCPExecutionIDs:     result.MCPExecutionIDs,
+		MCPExecutionIDs:                  result.MCPExecutionIDs,
-		ConversationID:      prep.ConversationID,
+		ConversationID:                   prep.ConversationID,
-		Time:                time.Now(),
+		Time:                             time.Now(),
-		Finalizable:         decision.Finalizable,
+		Finalizable:                      decision.Finalizable,
-		Finalized:           decision.Finalized,
+		Finalized:                        decision.Finalized,
-		Status:              decision.Status,
+		Status:                           decision.Status,
-		CompletionReason:    decision.CompletionReason,
+		CompletionReason:                 decision.CompletionReason,
-		EvidenceVerified:    decision.EvidenceVerified,
+		EvidenceVerified:                 decision.EvidenceVerified,
-		EvidenceRefs:        decision.EvidenceRefs,
+		EvidenceRefs:                     decision.EvidenceRefs,
-		PendingExecutionIDs: decision.PendingExecutionIDs,
+		PendingExecutionIDs:              decision.PendingExecutionIDs,
-		MissingChecks:       decision.MissingChecks,
+		MissingChecks:                    decision.MissingChecks,
+		AutoCancelledPendingExecutionIDs: autoCancelledPendingExecutionIDs,
 	})
 }
 

internal/handler/token_usage.go

@@ -0,0 +1,92 @@
+package handler
+
+import (
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+
+	"cyberstrike-ai/internal/database"
+	"cyberstrike-ai/internal/security"
+
+	"github.com/gin-gonic/gin"
+	"go.uber.org/zap"
+)
+
+// GetTokenUsageStats returns model token usage aggregates for dashboard views.
+func (h *ConversationHandler) GetTokenUsageStats(c *gin.Context) {
+	filter := tokenUsageFilterFromQuery(c)
+	if session, ok := security.CurrentSession(c); ok {
+		filter.Access = database.RBACListAccess{UserID: session.UserID, Scope: session.Scope}
+	}
+	stats, err := h.db.GetModelTokenUsageStats(filter)
+	if err != nil {
+		h.logger.Error("获取Token用量统计失败", zap.Error(err))
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	c.JSON(http.StatusOK, stats)
+}
+
+// GetConversationTokenUsageStats returns token usage scoped to one conversation.
+func (h *ConversationHandler) GetConversationTokenUsageStats(c *gin.Context) {
+	filter := tokenUsageFilterFromQuery(c)
+	filter.ConversationID = strings.TrimSpace(c.Param("id"))
+	if session, ok := security.CurrentSession(c); ok {
+		filter.Access = database.RBACListAccess{UserID: session.UserID, Scope: session.Scope}
+	}
+	stats, err := h.db.GetModelTokenUsageStats(filter)
+	if err != nil {
+		h.logger.Error("获取对话Token用量统计失败", zap.Error(err), zap.String("conversationId", filter.ConversationID))
+		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
+		return
+	}
+	c.JSON(http.StatusOK, stats)
+}
+
+func tokenUsageFilterFromQuery(c *gin.Context) database.ModelTokenUsageFilter {
+	days, _ := strconv.Atoi(strings.TrimSpace(c.DefaultQuery("days", "7")))
+	if days <= 0 {
+		days = 7
+	}
+	if days > 365 {
+		days = 365
+	}
+	limit, _ := strconv.Atoi(strings.TrimSpace(c.DefaultQuery("limit", "10")))
+	if limit <= 0 {
+		limit = 10
+	}
+	if limit > 500 {
+		limit = 500
+	}
+	filter := database.ModelTokenUsageFilter{
+		ConversationID: strings.TrimSpace(c.Query("conversation_id")),
+		ProjectID:      strings.TrimSpace(c.Query("project_id")),
+		Days:           days,
+		Limit:          limit,
+	}
+	if since := parseTokenUsageQueryTime(c.Query("since")); !since.IsZero() {
+		filter.Since = since
+	} else if days > 0 {
+		now := time.Now()
+		start := time.Date(now.Year(), now.Month(), now.Day(), 0, 0, 0, 0, now.Location()).AddDate(0, 0, -(days - 1))
+		filter.Since = start
+	}
+	if until := parseTokenUsageQueryTime(c.Query("until")); !until.IsZero() {
+		filter.Until = until
+	}
+	return filter
+}
+
+func parseTokenUsageQueryTime(raw string) time.Time {
+	raw = strings.TrimSpace(raw)
+	if raw == "" {
+		return time.Time{}
+	}
+	for _, layout := range []string{time.RFC3339Nano, time.RFC3339, "2006-01-02"} {
+		if t, err := time.Parse(layout, raw); err == nil {
+			return t
+		}
+	}
+	return time.Time{}
+}

internal/multiagent/agentic_orphan_tool_pruner_middleware.go

@@ -0,0 +1,126 @@
+package multiagent
+
+import (
+	"context"
+
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/schema"
+	"go.uber.org/zap"
+)
+
+// agenticOrphanToolPrunerMiddleware is the AgenticMessage equivalent of
+// orphanToolPrunerMiddleware. It removes user-role messages whose content
+// blocks are exclusively FunctionToolResult entries with CallIDs that do not
+// match any FunctionToolCall in the history.
+//
+// This is a defense-in-depth layer after agenticToolPairReconcilerMiddleware;
+// the reconciler handles the common case (assistant followed by its results)
+// while this pruner catches stray results that appear before their assistant
+// or in non-adjacent positions (e.g. after summarization rewriting).
+type agenticOrphanToolPrunerMiddleware struct {
+	*adk.TypedBaseChatModelAgentMiddleware[*schema.AgenticMessage]
+	logger *zap.Logger
+	phase  string
+}
+
+func newAgenticOrphanToolPrunerMiddleware(logger *zap.Logger, phase string) adk.TypedChatModelAgentMiddleware[*schema.AgenticMessage] {
+	return &agenticOrphanToolPrunerMiddleware{
+		TypedBaseChatModelAgentMiddleware: &adk.TypedBaseChatModelAgentMiddleware[*schema.AgenticMessage]{},
+		logger:                            logger,
+		phase:                             phase,
+	}
+}
+
+func (m *agenticOrphanToolPrunerMiddleware) BeforeModelRewriteState(
+	ctx context.Context,
+	state *adk.TypedChatModelAgentState[*schema.AgenticMessage],
+	mc *adk.TypedModelContext[*schema.AgenticMessage],
+) (context.Context, *adk.TypedChatModelAgentState[*schema.AgenticMessage], error) {
+	_ = mc
+	if m == nil || state == nil || len(state.Messages) == 0 {
+		return ctx, state, nil
+	}
+
+	// Pass 1: collect all provided CallIDs from assistant FunctionToolCall blocks.
+	provided := make(map[string]struct{}, 8)
+	for _, msg := range state.Messages {
+		if msg == nil || msg.Role != schema.AgenticRoleTypeAssistant {
+			continue
+		}
+		for _, block := range msg.ContentBlocks {
+			if block != nil && block.FunctionToolCall != nil && block.FunctionToolCall.CallID != "" {
+				provided[block.FunctionToolCall.CallID] = struct{}{}
+			}
+		}
+	}
+
+	// Fast path: check if any orphan exists.
+	hasOrphan := false
+	for _, msg := range state.Messages {
+		if msg == nil || !isPureAgenticToolResult(msg) {
+			continue
+		}
+		for _, id := range agenticToolResultCallIDs(msg) {
+			if _, ok := provided[id]; !ok {
+				hasOrphan = true
+				break
+			}
+		}
+		if hasOrphan {
+			break
+		}
+	}
+	if !hasOrphan {
+		return ctx, state, nil
+	}
+
+	// Pass 2: build pruned list.
+	pruned := make([]*schema.AgenticMessage, 0, len(state.Messages))
+	var droppedIDs []string
+	var droppedNames []string
+	for _, msg := range state.Messages {
+		if msg == nil {
+			continue
+		}
+		if !isPureAgenticToolResult(msg) {
+			pruned = append(pruned, msg)
+			continue
+		}
+		// Check if ALL result call IDs are orphans. If any is matched, keep the
+		// message (the reconciler already handled partial mismatches).
+		allOrphan := true
+		for _, id := range agenticToolResultCallIDs(msg) {
+			if _, ok := provided[id]; ok {
+				allOrphan = false
+				break
+			}
+		}
+		if allOrphan {
+			for _, block := range msg.ContentBlocks {
+				if block != nil && block.FunctionToolResult != nil {
+					droppedIDs = append(droppedIDs, block.FunctionToolResult.CallID)
+					droppedNames = append(droppedNames, block.FunctionToolResult.Name)
+				}
+			}
+			continue
+		}
+		pruned = append(pruned, msg)
+	}
+
+	if len(droppedIDs) == 0 {
+		return ctx, state, nil
+	}
+	if m.logger != nil {
+		m.logger.Warn("agentic orphan tool messages pruned before model call",
+			zap.String("phase", m.phase),
+			zap.Int("dropped_count", len(droppedIDs)),
+			zap.Strings("dropped_tool_call_ids", droppedIDs),
+			zap.Strings("dropped_tool_names", droppedNames),
+			zap.Int("messages_before", len(state.Messages)),
+			zap.Int("messages_after", len(pruned)),
+		)
+	}
+	ns := *state
+	ns.Messages = pruned
+	return ctx, &ns, nil
+}

internal/multiagent/agentic_tool_pair_reconciler_middleware.go

@@ -0,0 +1,247 @@
+package multiagent
+
+import (
+	"context"
+	"fmt"
+
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/schema"
+	"go.uber.org/zap"
+)
+
+// agenticToolPairReconcilerMiddleware is the AgenticMessage equivalent of
+// toolPairReconcilerMiddleware. It ensures every assistant FunctionToolCall
+// block is followed by a matching FunctionToolResult message, patching or
+// dropping as needed so the downstream model never receives an unpaired
+// tool-call history.
+//
+// In the AgenticMessage protocol:
+//   - Assistant tool calls: Role=AgenticRoleTypeAssistant with FunctionToolCall content blocks.
+//   - Tool results: Role=AgenticRoleTypeUser with FunctionToolResult content blocks.
+//
+// This middleware runs after summarization which may truncate history and
+// break pairings.
+type agenticToolPairReconcilerMiddleware struct {
+	*adk.TypedBaseChatModelAgentMiddleware[*schema.AgenticMessage]
+	logger *zap.Logger
+	phase  string
+}
+
+func newAgenticToolPairReconcilerMiddleware(logger *zap.Logger, phase string) adk.TypedChatModelAgentMiddleware[*schema.AgenticMessage] {
+	return &agenticToolPairReconcilerMiddleware{
+		TypedBaseChatModelAgentMiddleware: &adk.TypedBaseChatModelAgentMiddleware[*schema.AgenticMessage]{},
+		logger:                            logger,
+		phase:                             phase,
+	}
+}
+
+func (m *agenticToolPairReconcilerMiddleware) BeforeModelRewriteState(
+	ctx context.Context,
+	state *adk.TypedChatModelAgentState[*schema.AgenticMessage],
+	mc *adk.TypedModelContext[*schema.AgenticMessage],
+) (context.Context, *adk.TypedChatModelAgentState[*schema.AgenticMessage], error) {
+	_ = mc
+	if m == nil || state == nil || len(state.Messages) == 0 {
+		return ctx, state, nil
+	}
+
+	usedIDs := make(map[string]struct{}, 16)
+	changed := false
+	patched := 0
+	dropped := 0
+	out := make([]*schema.AgenticMessage, 0, len(state.Messages))
+
+	for i := 0; i < len(state.Messages); {
+		msg := state.Messages[i]
+		if msg == nil {
+			changed = true
+			i++
+			continue
+		}
+
+		calls := agenticFunctionToolCalls(msg)
+
+		// Non-assistant or assistant without tool calls — but check for orphan
+		// tool-result messages (user role with only FunctionToolResult blocks).
+		if len(calls) == 0 {
+			if isPureAgenticToolResult(msg) {
+				// Orphan tool result not preceded by its assistant; drop it.
+				changed = true
+				dropped++
+				i++
+				continue
+			}
+			out = append(out, msg)
+			i++
+			continue
+		}
+
+		// Deduplicate / fix empty call IDs.
+		idsChanged := false
+		for ci := range calls {
+			id := calls[ci].CallID
+			_, duplicate := usedIDs[id]
+			if id == "" || duplicate {
+				base := fmt.Sprintf("patched_agentic_call_%d_%d", i, ci)
+				id = base
+				for suffix := 1; ; suffix++ {
+					if _, exists := usedIDs[id]; !exists {
+						break
+					}
+					id = fmt.Sprintf("%s_%d", base, suffix)
+				}
+				calls[ci].CallID = id
+				idsChanged = true
+				changed = true
+			}
+			usedIDs[id] = struct{}{}
+		}
+
+		assistant := msg
+		if idsChanged {
+			assistant = cloneAgenticMessageWithCalls(msg, calls)
+		}
+		out = append(out, assistant)
+
+		// Build expected set.
+		expected := make(map[string]*schema.FunctionToolCall, len(calls))
+		for ci := range calls {
+			expected[calls[ci].CallID] = calls[ci]
+		}
+
+		// Consume following tool-result messages.
+		results := make(map[string]*schema.AgenticMessage, len(calls))
+		j := i + 1
+		for j < len(state.Messages) {
+			next := state.Messages[j]
+			if next == nil {
+				changed = true
+				j++
+				continue
+			}
+			if !isPureAgenticToolResult(next) {
+				break
+			}
+			resultCallIDs := agenticToolResultCallIDs(next)
+			consumed := false
+			for _, rid := range resultCallIDs {
+				if _, wanted := expected[rid]; !wanted {
+					continue
+				}
+				if _, dup := results[rid]; dup {
+					continue
+				}
+				results[rid] = next
+				consumed = true
+			}
+			if !consumed {
+				changed = true
+				dropped++
+			}
+			j++
+		}
+
+		// Emit results in call order, patching missing ones.
+		for _, tc := range calls {
+			if result, ok := results[tc.CallID]; ok {
+				out = append(out, result)
+				continue
+			}
+			out = append(out, makeAgenticPatchedToolResult(tc.CallID, tc.Name))
+			changed = true
+			patched++
+		}
+		i = j
+	}
+
+	if !changed {
+		return ctx, state, nil
+	}
+	if m.logger != nil {
+		m.logger.Warn("agentic tool-call/result pairs reconciled before model call",
+			zap.String("phase", m.phase),
+			zap.Int("patched_results", patched),
+			zap.Int("dropped_results", dropped),
+			zap.Int("messages_before", len(state.Messages)),
+			zap.Int("messages_after", len(out)),
+		)
+	}
+	ns := *state
+	ns.Messages = out
+	return ctx, &ns, nil
+}
+
+// agenticFunctionToolCalls extracts FunctionToolCall pointers from an
+// assistant message's content blocks. Returns nil for non-assistant messages.
+func agenticFunctionToolCalls(msg *schema.AgenticMessage) []*schema.FunctionToolCall {
+	if msg == nil || msg.Role != schema.AgenticRoleTypeAssistant {
+		return nil
+	}
+	var out []*schema.FunctionToolCall
+	for _, block := range msg.ContentBlocks {
+		if block != nil && block.FunctionToolCall != nil {
+			out = append(out, block.FunctionToolCall)
+		}
+	}
+	return out
+}
+
+// isPureAgenticToolResult returns true when the message is a user-role
+// message whose content blocks are exclusively FunctionToolResult entries.
+func isPureAgenticToolResult(msg *schema.AgenticMessage) bool {
+	if msg == nil || msg.Role != schema.AgenticRoleTypeUser || len(msg.ContentBlocks) == 0 {
+		return false
+	}
+	for _, block := range msg.ContentBlocks {
+		if block == nil {
+			continue
+		}
+		if block.FunctionToolResult == nil {
+			return false
+		}
+	}
+	return true
+}
+
+// agenticToolResultCallIDs extracts all CallIDs from FunctionToolResult blocks.
+func agenticToolResultCallIDs(msg *schema.AgenticMessage) []string {
+	if msg == nil {
+		return nil
+	}
+	var ids []string
+	for _, block := range msg.ContentBlocks {
+		if block != nil && block.FunctionToolResult != nil && block.FunctionToolResult.CallID != "" {
+			ids = append(ids, block.FunctionToolResult.CallID)
+		}
+	}
+	return ids
+}
+
+func cloneAgenticMessageWithCalls(msg *schema.AgenticMessage, calls []*schema.FunctionToolCall) *schema.AgenticMessage {
+	cloned := *msg
+	cloned.ContentBlocks = make([]*schema.ContentBlock, 0, len(msg.ContentBlocks))
+	callIdx := 0
+	for _, block := range msg.ContentBlocks {
+		if block != nil && block.FunctionToolCall != nil && callIdx < len(calls) {
+			cloned.ContentBlocks = append(cloned.ContentBlocks, schema.NewContentBlock(calls[callIdx]))
+			callIdx++
+		} else {
+			cloned.ContentBlocks = append(cloned.ContentBlocks, block)
+		}
+	}
+	return &cloned
+}
+
+func makeAgenticPatchedToolResult(callID, name string) *schema.AgenticMessage {
+	return &schema.AgenticMessage{
+		Role: schema.AgenticRoleTypeUser,
+		ContentBlocks: []*schema.ContentBlock{schema.NewContentBlock(&schema.FunctionToolResult{
+			CallID: callID,
+			Name:   name,
+			Content: []*schema.FunctionToolResultContentBlock{{
+				Type: schema.FunctionToolResultContentBlockTypeText,
+				Text: &schema.UserInputText{Text: patchedMissingToolResult},
+			}},
+		})},
+	}
+}

internal/multiagent/agentic_tool_pair_reconciler_middleware_test.go

@@ -0,0 +1,157 @@
+package multiagent
+
+import (
+	"context"
+	"testing"
+
+	"github.com/cloudwego/eino/adk"
+	"github.com/cloudwego/eino/schema"
+)
+
+func TestAgenticToolPairReconcilerPatchesMissing(t *testing.T) {
+	t.Parallel()
+	mw := newAgenticToolPairReconcilerMiddleware(nil, "test")
+	state := &adk.TypedChatModelAgentState[*schema.AgenticMessage]{
+		Messages: []*schema.AgenticMessage{
+			agenticAssistantToolCall("c1", "search", `{"q":"x"}`),
+			agenticAssistantToolCall("c2", "execute", `{"cmd":"ls"}`),
+			// c1 result present, c2 missing
+			agenticToolResult("c1", "search", "found it"),
+		},
+	}
+	_, out, err := mw.BeforeModelRewriteState(context.Background(), state, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	// Expected: assistant(c1) -> result(c1) -> assistant(c2) -> patched_result(c2)
+	if len(out.Messages) != 4 {
+		t.Fatalf("messages = %d, want 4", len(out.Messages))
+	}
+	// c1 assistant
+	if calls := agenticFunctionToolCalls(out.Messages[0]); len(calls) != 1 || calls[0].CallID != "c1" {
+		t.Fatal("msg[0] should be assistant(c1)")
+	}
+	// c1 result
+	if ids := agenticToolResultCallIDs(out.Messages[1]); len(ids) != 1 || ids[0] != "c1" {
+		t.Fatal("msg[1] should be result(c1)")
+	}
+	// c2 assistant
+	if calls := agenticFunctionToolCalls(out.Messages[2]); len(calls) != 1 || calls[0].CallID != "c2" {
+		t.Fatal("msg[2] should be assistant(c2)")
+	}
+	// c2 patched result
+	if ids := agenticToolResultCallIDs(out.Messages[3]); len(ids) != 1 || ids[0] != "c2" {
+		t.Fatal("msg[3] should be patched result(c2)")
+	}
+	resultText := out.Messages[3].ContentBlocks[0].FunctionToolResult.Content[0].Text.Text
+	if resultText != patchedMissingToolResult {
+		t.Fatalf("patched text = %q", resultText)
+	}
+}
+
+func TestAgenticToolPairReconcilerDropsOrphan(t *testing.T) {
+	t.Parallel()
+	mw := newAgenticToolPairReconcilerMiddleware(nil, "test")
+	state := &adk.TypedChatModelAgentState[*schema.AgenticMessage]{
+		Messages: []*schema.AgenticMessage{
+			// Orphan tool result with no preceding assistant
+			agenticToolResult("orphan", "deleted_tool", "stale data"),
+			{Role: schema.AgenticRoleTypeUser, ContentBlocks: []*schema.ContentBlock{
+				schema.NewContentBlock(&schema.UserInputText{Text: "hello"}),
+			}},
+		},
+	}
+	_, out, err := mw.BeforeModelRewriteState(context.Background(), state, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(out.Messages) != 1 {
+		t.Fatalf("messages = %d, want 1 (orphan dropped)", len(out.Messages))
+	}
+	if out.Messages[0].ContentBlocks[0].UserInputText == nil {
+		t.Fatal("remaining message should be the user text")
+	}
+}
+
+func TestAgenticToolPairReconcilerNoopWhenPaired(t *testing.T) {
+	t.Parallel()
+	mw := newAgenticToolPairReconcilerMiddleware(nil, "test")
+	state := &adk.TypedChatModelAgentState[*schema.AgenticMessage]{
+		Messages: []*schema.AgenticMessage{
+			agenticAssistantToolCall("c1", "search", `{}`),
+			agenticToolResult("c1", "search", "ok"),
+		},
+	}
+	_, out, err := mw.BeforeModelRewriteState(context.Background(), state, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	// Should return original state unchanged
+	if &out.Messages[0] == &state.Messages[0] {
+		// pointer equality on slice — state not cloned
+	}
+	if len(out.Messages) != 2 {
+		t.Fatalf("messages = %d, want 2", len(out.Messages))
+	}
+}
+
+func TestAgenticToolPairReconcilerFixesEmptyCallID(t *testing.T) {
+	t.Parallel()
+	mw := newAgenticToolPairReconcilerMiddleware(nil, "test")
+	state := &adk.TypedChatModelAgentState[*schema.AgenticMessage]{
+		Messages: []*schema.AgenticMessage{
+			{
+				Role: schema.AgenticRoleTypeAssistant,
+				ContentBlocks: []*schema.ContentBlock{schema.NewContentBlock(&schema.FunctionToolCall{
+					CallID: "", Name: "search", Arguments: `{}`,
+				})},
+			},
+		},
+	}
+	_, out, err := mw.BeforeModelRewriteState(context.Background(), state, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	calls := agenticFunctionToolCalls(out.Messages[0])
+	if len(calls) != 1 || calls[0].CallID == "" {
+		t.Fatalf("empty call ID should be patched, got %q", calls[0].CallID)
+	}
+}
+
+func TestAgenticOrphanToolPrunerRemovesOrphan(t *testing.T) {
+	t.Parallel()
+	mw := newAgenticOrphanToolPrunerMiddleware(nil, "test")
+	state := &adk.TypedChatModelAgentState[*schema.AgenticMessage]{
+		Messages: []*schema.AgenticMessage{
+			agenticAssistantToolCall("c1", "search", `{}`),
+			agenticToolResult("c1", "search", "ok"),
+			// Orphan: no assistant has call_id "c_orphan"
+			agenticToolResult("c_orphan", "deleted", "stale"),
+		},
+	}
+	_, out, err := mw.BeforeModelRewriteState(context.Background(), state, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(out.Messages) != 2 {
+		t.Fatalf("messages = %d, want 2 (orphan pruned)", len(out.Messages))
+	}
+}
+
+func TestAgenticOrphanToolPrunerNoopWhenClean(t *testing.T) {
+	t.Parallel()
+	mw := newAgenticOrphanToolPrunerMiddleware(nil, "test")
+	state := &adk.TypedChatModelAgentState[*schema.AgenticMessage]{
+		Messages: []*schema.AgenticMessage{
+			agenticAssistantToolCall("c1", "search", `{}`),
+			agenticToolResult("c1", "search", "ok"),
+		},
+	}
+	_, out, err := mw.BeforeModelRewriteState(context.Background(), state, nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(out.Messages) != 2 {
+		t.Fatalf("messages = %d, want 2", len(out.Messages))
+	}
+}

internal/multiagent/eino_agentic_chat_model_tail_middleware.go

@@ -20,8 +20,13 @@ func appendEinoAgenticChatModelTailMiddlewares(
 	handlers = append(handlers, newAgenticSystemMessageNormalizerMiddleware(cfg.logger, cfg.phase))
 	handlers = append(handlers, newAgenticContinuationUserDedupMiddleware(cfg.logger, cfg.phase))
 	if cfg.agenticSummarization != nil {
+		handlers = append(handlers, newAgenticToolPairReconcilerMiddleware(cfg.logger, cfg.phase+"_pre_summarization"))
 		handlers = append(handlers, cfg.agenticSummarization)
 	}
+	handlers = append(handlers, newAgenticToolPairReconcilerMiddleware(cfg.logger, cfg.phase))
+	if !cfg.skipOrphanPruner {
+		handlers = append(handlers, newAgenticOrphanToolPrunerMiddleware(cfg.logger, cfg.phase))
+	}
 	if !cfg.skipTrace && cfg.trace != nil {
 		if capMw := newAgenticModelFacingTraceMiddleware(cfg.trace); capMw != nil {
 			handlers = append(handlers, capMw)

internal/multiagent/eino_agentic_chat_model_tail_middleware_test.go

@@ -106,7 +106,8 @@ func TestAppendEinoAgenticChatModelTailMiddlewares(t *testing.T) {
 		phase: "agentic",
 		trace: holder,
 	})
-	if len(handlers) != 3 {
+	// system + continuation + reconciler + orphan_pruner + trace
-		t.Fatalf("handlers = %d, want system + continuation + trace", len(handlers))
+	if len(handlers) != 5 {
+		t.Fatalf("handlers = %d, want system + continuation + reconciler + orphan_pruner + trace", len(handlers))
 	}
 }

internal/multiagent/eino_exit_fallback_test.go

@@ -55,6 +55,70 @@ func TestEinoExtractFallbackAssistantFromMsgs_prefersToolOverEarlierAssistant(t
 	}
 }
 
+func TestEinoExtractFallbackAssistantFromMsgs_plainAssistant(t *testing.T) {
+	msgs := []*schema.Message{
+		schema.UserMessage("hi"),
+		schema.AssistantMessage("plain answer", nil),
+	}
+	if got := einoExtractFallbackAssistantFromMsgs(msgs); got != "plain answer" {
+		t.Fatalf("got %q", got)
+	}
+}
+
+func TestEinoExtractFallbackAssistantFromMsgs_finalAssistantAfterToolResult(t *testing.T) {
+	msgs := []*schema.Message{
+		schema.UserMessage("hi"),
+		schema.AssistantMessage("", []schema.ToolCall{{
+			ID:   "call-1",
+			Type: "function",
+			Function: schema.FunctionCall{
+				Name:      "execute",
+				Arguments: `{"command":"pwd"}`,
+			},
+		}}),
+		schema.ToolMessage("/tmp", "call-1", schema.WithToolName("execute")),
+		schema.AssistantMessage("final after tool", nil),
+	}
+	if got := einoExtractFallbackAssistantFromMsgs(msgs); got != "final after tool" {
+		t.Fatalf("got %q", got)
+	}
+}
+
+func TestEinoExtractFallbackAssistantFromMsgs_doesNotUseAssistantBeforeUnfinishedToolResult(t *testing.T) {
+	msgs := []*schema.Message{
+		schema.UserMessage("hi"),
+		schema.AssistantMessage("I will inspect that.", nil),
+		schema.AssistantMessage("", []schema.ToolCall{{
+			ID:   "call-1",
+			Type: "function",
+			Function: schema.FunctionCall{
+				Name:      "execute",
+				Arguments: `{"command":"pwd"}`,
+			},
+		}}),
+		schema.ToolMessage("/tmp", "call-1", schema.WithToolName("execute")),
+	}
+	if got := einoExtractFallbackAssistantFromMsgs(msgs); got != "" {
+		t.Fatalf("got %q, want empty", got)
+	}
+}
+
+func TestEinoRunResultBuilderFinalFallsBackToPlainAssistantTrace(t *testing.T) {
+	runMessages := newEinoRunMessageAccumulator(nil)
+	runMessages.Append(schema.UserMessage("hi"))
+	runMessages.Append(schema.AssistantMessage("plain answer", nil))
+
+	got := newEinoRunResultBuilder(einoRunResultBuilderConfig{
+		OrchMode:    "deep",
+		EmptyHint:   "empty",
+		RunMessages: runMessages,
+	}).BuildFinal()
+
+	if got.Response != "plain answer" {
+		t.Fatalf("response = %q, want plain answer", got.Response)
+	}
+}
+
 func toolExitMsg(content, callID string) *schema.Message {
 	m := schema.ToolMessage(content, callID)
 	m.ToolName = "exit"

internal/multiagent/eino_run_error_handler.go

@@ -3,6 +3,8 @@ package multiagent
 import (
 	"context"
 	"errors"
+	"fmt"
+	"strings"
 
 	"github.com/cloudwego/eino/adk"
 )
@@ -82,12 +84,109 @@ func (h *einoRunErrorHandler) emitError(err error, kind string) {
 	if h == nil || h.progress == nil || err == nil {
 		return
 	}
+	userErr := einoUserFacingRunError(err)
 	data := map[string]interface{}{
 		"conversationId": h.conversationID,
 		"source":         "eino",
+		"error":          err.Error(),
 	}
 	if kind != "" {
 		data["errorKind"] = kind
+	} else if userErr.kind != "" {
+		data["errorKind"] = userErr.kind
 	}
-	h.progress("error", err.Error(), data)
+	if userErr.summary != "" {
+		data["errorSummary"] = userErr.summary
+	}
+	if userErr.retryExhausted {
+		data["retryExhausted"] = true
+		if userErr.totalRetries > 0 {
+			data["totalRetries"] = userErr.totalRetries
+		}
+	}
+	if userErr.rawLastError != "" {
+		data["lastError"] = userErr.rawLastError
+	}
+	if userErr.technicalError != "" {
+		data["technicalError"] = userErr.technicalError
+	}
+	if userErr.hasModelOriginalError {
+		data["modelOriginalError"] = userErr.rawLastError
+	} else if userErr.retryExhausted {
+		data["hasModelOriginalError"] = false
+	}
+	message := err.Error()
+	if userErr.message != "" {
+		message = userErr.message
+	}
+	h.progress("error", message, data)
+}
+
+type einoRunUserError struct {
+	message               string
+	kind                  string
+	summary               string
+	rawLastError          string
+	technicalError        string
+	retryExhausted        bool
+	totalRetries          int
+	hasModelOriginalError bool
+}
+
+func einoUserFacingRunError(err error) einoRunUserError {
+	var out einoRunUserError
+	if err == nil {
+		return out
+	}
+	var retryErr *adk.RetryExhaustedError
+	if !errors.As(err, &retryErr) {
+		return out
+	}
+	out.retryExhausted = true
+	out.totalRetries = retryErr.TotalRetries
+	lastErr := retryErr.LastErr
+	if lastErr == nil {
+		out.kind = "model_retry_exhausted"
+		out.summary = "模型调用多次重试后仍未成功。"
+		out.message = out.summary
+		return out
+	}
+	out.rawLastError = strings.TrimSpace(lastErr.Error())
+	if isEinoShouldRetryOutputRejected(lastErr) {
+		out.kind = "model_output_rejected"
+		out.summary = "模型未返回原始错误；输出被重试策略拒绝。"
+		out.technicalError = out.rawLastError
+		out.message = formatEinoRetryExhaustedMessage(out.summary, retryErr.TotalRetries)
+		return out
+	}
+	kind, summary := einoTransientRunErrorUserDetail(lastErr)
+	if strings.TrimSpace(summary) == "" {
+		summary = einoTrimRetryErrorSummary(lastErr.Error())
+	}
+	if kind == "" {
+		kind = "model_retry_exhausted"
+	}
+	out.kind = kind
+	out.summary = summary
+	out.hasModelOriginalError = out.rawLastError != ""
+	out.message = formatEinoRetryExhaustedMessage(summary, retryErr.TotalRetries)
+	return out
+}
+
+func isEinoShouldRetryOutputRejected(err error) bool {
+	if err == nil {
+		return false
+	}
+	return strings.Contains(strings.ToLower(err.Error()), "model output rejected by shouldretry")
+}
+
+func formatEinoRetryExhaustedMessage(summary string, totalRetries int) string {
+	summary = strings.TrimSpace(summary)
+	if summary == "" {
+		summary = "模型调用多次重试后仍未成功。"
+	}
+	if totalRetries > 0 {
+		return fmt.Sprintf("模型调用重试已耗尽（已重试 %d 次）：%s", totalRetries, summary)
+	}
+	return "模型调用重试已耗尽：" + summary
 }

internal/multiagent/eino_run_error_handler_test.go

@@ -3,6 +3,7 @@ package multiagent
 import (
 	"context"
 	"errors"
+	"strings"
 	"testing"
 
 	"github.com/cloudwego/eino/adk"
@@ -61,6 +62,99 @@ func TestEinoRunErrorHandlerTimeoutAndGeneralErrorProgress(t *testing.T) {
 	}
 }
 
+func TestEinoRunErrorHandlerRetryExhaustedEmptyOutputProgress(t *testing.T) {
+	err := &adk.RetryExhaustedError{
+		LastErr:      errors.New("model output rejected by ShouldRetry at attempt 5"),
+		TotalRetries: 4,
+	}
+	var message string
+	var data map[string]interface{}
+
+	got := newEinoRunErrorHandler(einoRunErrorHandlerConfig{
+		ConversationID: "conv-1",
+		Progress: func(eventType, msg string, raw interface{}) {
+			if eventType == "error" {
+				message = msg
+				data, _ = raw.(map[string]interface{})
+			}
+		},
+	}).Handle(err)
+
+	if !errors.Is(got, err) {
+		t.Fatalf("err = %v", got)
+	}
+	if !strings.Contains(message, "模型调用重试已耗尽") ||
+		!strings.Contains(message, "模型未返回原始错误；输出被重试策略拒绝。") ||
+		strings.Contains(message, "model output rejected by ShouldRetry at attempt 5") {
+		t.Fatalf("message = %q", message)
+	}
+	if data["errorKind"] != "model_output_rejected" {
+		t.Fatalf("errorKind = %#v", data["errorKind"])
+	}
+	if data["errorSummary"] != "模型未返回原始错误；输出被重试策略拒绝。" {
+		t.Fatalf("errorSummary = %#v", data["errorSummary"])
+	}
+	if data["hasModelOriginalError"] != false {
+		t.Fatalf("hasModelOriginalError = %#v", data["hasModelOriginalError"])
+	}
+	if data["retryExhausted"] != true || data["totalRetries"] != 4 {
+		t.Fatalf("retry metadata = %#v", data)
+	}
+	if data["lastError"] != "model output rejected by ShouldRetry at attempt 5" {
+		t.Fatalf("lastError = %#v", data["lastError"])
+	}
+	if data["technicalError"] != "model output rejected by ShouldRetry at attempt 5" {
+		t.Fatalf("technicalError = %#v", data["technicalError"])
+	}
+	if _, ok := data["modelOriginalError"]; ok {
+		t.Fatalf("modelOriginalError should be absent for ShouldRetry rejection, got %#v", data["modelOriginalError"])
+	}
+	if data["error"] != err.Error() {
+		t.Fatalf("raw error = %#v, want %#v", data["error"], err.Error())
+	}
+}
+
+func TestEinoRunErrorHandlerRetryExhaustedOriginalErrorProgress(t *testing.T) {
+	err := &adk.RetryExhaustedError{
+		LastErr:      errors.New("HTTP 429 Too Many Requests"),
+		TotalRetries: 3,
+	}
+	var message string
+	var data map[string]interface{}
+
+	got := newEinoRunErrorHandler(einoRunErrorHandlerConfig{
+		ConversationID: "conv-1",
+		Progress: func(eventType, msg string, raw interface{}) {
+			if eventType == "error" {
+				message = msg
+				data, _ = raw.(map[string]interface{})
+			}
+		},
+	}).Handle(err)
+
+	if !errors.Is(got, err) {
+		t.Fatalf("err = %v", got)
+	}
+	if !strings.Contains(message, "HTTP 429 Too Many Requests") {
+		t.Fatalf("message = %q", message)
+	}
+	if data["errorKind"] != "rate_limit" {
+		t.Fatalf("errorKind = %#v", data["errorKind"])
+	}
+	if data["errorSummary"] != "HTTP 429 Too Many Requests" {
+		t.Fatalf("errorSummary = %#v", data["errorSummary"])
+	}
+	if data["lastError"] != "HTTP 429 Too Many Requests" {
+		t.Fatalf("lastError = %#v", data["lastError"])
+	}
+	if data["modelOriginalError"] != "HTTP 429 Too Many Requests" {
+		t.Fatalf("modelOriginalError = %#v", data["modelOriginalError"])
+	}
+	if _, ok := data["hasModelOriginalError"]; ok {
+		t.Fatalf("hasModelOriginalError should be absent when original error is present, got %#v", data["hasModelOriginalError"])
+	}
+}
+
 func TestEinoRunErrorHandlerIterationLimitProgress(t *testing.T) {
 	var events []string
 	var errorKind interface{}

internal/multiagent/eino_run_message_accumulator.go

@@ -57,6 +57,16 @@ func (a *einoRunMessageAccumulator) Messages() []adk.Message {
 	return a.msgs
 }
 
+func (a *einoRunMessageAccumulator) NewMessages() []adk.Message {
+	if a == nil {
+		return nil
+	}
+	if a.baseCount < 0 || a.baseCount >= len(a.msgs) {
+		return nil
+	}
+	return a.msgs[a.baseCount:]
+}
+
 func (a *einoRunMessageAccumulator) BaseCount() int {
 	if a == nil {
 		return 0

internal/multiagent/eino_run_message_accumulator_test.go

@@ -27,6 +27,10 @@ func TestEinoRunMessageAccumulatorTracksBaseAndAppends(t *testing.T) {
 	if len(msgs) != 2 || msgs[1].Role != schema.Assistant || msgs[1].Content != "hello" {
 		t.Fatalf("messages = %#v", msgs)
 	}
+	newMsgs := acc.NewMessages()
+	if len(newMsgs) != 1 || newMsgs[0].Content != "hello" {
+		t.Fatalf("new messages = %#v", newMsgs)
+	}
 }
 
 func TestEinoRunMessageAccumulatorToolMessage(t *testing.T) {

internal/multiagent/eino_run_result_builder.go

@@ -45,7 +45,7 @@ func (b *einoRunResultBuilder) BuildFinal() *RunResult {
 func (b *einoRunResultBuilder) build(partial bool) *RunResult {
 	var runMsgs []adk.Message
 	if b.cfg.RunMessages != nil {
-		runMsgs = b.cfg.RunMessages.Messages()
+		runMsgs = b.cfg.RunMessages.NewMessages()
 	}
 	var lastAssistant string
 	var lastPlanExecuteExecutor string
@@ -107,6 +107,9 @@ func buildEinoRunResultFromAccumulated(
 	if cleaned == "" {
 		if fb := strings.TrimSpace(einoExtractFallbackAssistantFromMsgs(runAccumulatedMsgs)); fb != "" {
 			cleaned = fb
+			if orchMode == "plan_execute" {
+				cleaned = UnwrapPlanExecuteUserText(cleaned)
+			}
 		}
 	}
 	cleaned = dedupeRepeatedParagraphs(cleaned, 80)
@@ -146,32 +149,38 @@ func markModelFacingTraceForPersistence(msgs []adk.Message) []adk.Message {
 	return out
 }
 
-// einoExtractFallbackAssistantFromMsgs 在「主通道未产出助手正文」时，从 Eino ADK 轨迹中回填用户可见回复。
+// einoExtractFallbackAssistantFromMsgs 在「主通道未产出助手正文」时，从 Eino ADK
-// 典型场景：监督者仅调用 exit（final_result 落在 Tool 消息中），或工具结果已写入历史但 lastAssistant 未更新。
+// 原生消息轨迹中回填用户可见回复。这里保持克制：只采纳倒序最近的可交付终态，
+// 避免把工具调用前的过渡语或子任务过程误升为最终回复。
 //
-// 优先级：最后一次 exit 工具输出 → 最后一条含 exit 的助手 tool_calls 参数中的 final_result。
+// 可交付终态：
+// - exit 工具输出；
+// - assistant 调用 exit 时 arguments.final_result；
+// - 没有后续普通工具结果截断的纯 assistant 正文。
 func einoExtractFallbackAssistantFromMsgs(msgs []adk.Message) string {
 	for i := len(msgs) - 1; i >= 0; i-- {
 		m := msgs[i]
-		if m == nil || m.Role != schema.Tool {
+		if m == nil {
 			continue
 		}
-		if !strings.EqualFold(strings.TrimSpace(m.ToolName), adk.ToolInfoExit.Name) {
+		switch m.Role {
-			continue
+		case schema.Tool:
-		}
+			if strings.EqualFold(strings.TrimSpace(m.ToolName), adk.ToolInfoExit.Name) {
-		content := strings.TrimSpace(m.Content)
+				content := strings.TrimSpace(m.Content)
-		if content == "" || strings.HasPrefix(content, einomcp.ToolErrorPrefix) {
+				if content != "" && !strings.HasPrefix(content, einomcp.ToolErrorPrefix) {
-			continue
+					return content
-		}
+				}
-		return content
+			}
-	}
+			return ""
-	for i := len(msgs) - 1; i >= 0; i-- {
+		case schema.Assistant:
-		m := msgs[i]
+			if s := einoExtractExitFinalFromAssistantToolCalls(m); s != "" {
-		if m == nil || m.Role != schema.Assistant {
+				return s
-			continue
+			}
-		}
+			if len(m.ToolCalls) == 0 {
-		if s := einoExtractExitFinalFromAssistantToolCalls(m); s != "" {
+				if content := strings.TrimSpace(m.Content); content != "" {
-			return s
+					return content
+				}
+			}
 		}
 	}
 	return ""

internal/multiagent/eino_run_result_builder_test.go

@@ -55,6 +55,24 @@ func TestEinoRunResultBuilderFinalUsesSnapshots(t *testing.T) {
 	}
 }
 
+func TestEinoRunResultBuilderFallbackIgnoresBaseHistory(t *testing.T) {
+	runMessages := newEinoRunMessageAccumulator([]adk.Message{
+		schema.UserMessage("previous request"),
+		schema.AssistantMessage("previous answer", nil),
+		schema.UserMessage("new request"),
+	})
+
+	got := newEinoRunResultBuilder(einoRunResultBuilderConfig{
+		OrchMode:    "deep",
+		EmptyHint:   "empty",
+		RunMessages: runMessages,
+	}).BuildFinal()
+
+	if got.Response != "empty" {
+		t.Fatalf("response = %q, want empty hint", got.Response)
+	}
+}
+
 func TestEinoRunResultBuilderPlanExecutePrefersExecutorOutput(t *testing.T) {
 	runMessages := newEinoRunMessageAccumulator(nil)
 	runMessages.Append(schema.AssistantMessage(`{"response":"planner text"}`, nil))
@@ -73,3 +91,18 @@ func TestEinoRunResultBuilderPlanExecutePrefersExecutorOutput(t *testing.T) {
 		t.Fatalf("response = %q, want executor text", got.Response)
 	}
 }
+
+func TestEinoRunResultBuilderPlanExecuteUnwrapsFallbackAssistant(t *testing.T) {
+	runMessages := newEinoRunMessageAccumulator(nil)
+	runMessages.Append(schema.AssistantMessage(`{"response":"fallback executor text"}`, nil))
+
+	got := newEinoRunResultBuilder(einoRunResultBuilderConfig{
+		OrchMode:    "plan_execute",
+		EmptyHint:   "empty",
+		RunMessages: runMessages,
+	}).BuildFinal()
+
+	if got.Response != "fallback executor text" {
+		t.Fatalf("response = %q, want fallback executor text", got.Response)
+	}
+}

internal/multiagent/eino_run_runtime_session.go

@@ -371,5 +371,9 @@ func (s *einoRunRuntimeSession) emitUsageSummary(reason string) bool {
 	if s == nil || s.usage == nil {
 		return false
 	}
-	return s.usage.EmitOnce(s.conversationID, s.orchMode, reason, s.progress, s.logger)
+	modelName := ""
+	if s.args != nil {
+		modelName = s.args.ModelName
+	}
+	return s.usage.EmitOnce(s.conversationID, s.orchMode, reason, modelName, s.progress, s.logger)
 }

internal/multiagent/eino_run_usage_accumulator.go

@@ -61,6 +61,7 @@ func (a *einoRunUsageAccumulator) EmitOnce(
 	conversationID string,
 	orchestration string,
 	reason string,
+	modelName string,
 	progress func(eventType, message string, data interface{}),
 	logger *zap.Logger,
 ) bool {
@@ -81,6 +82,7 @@ func (a *einoRunUsageAccumulator) EmitOnce(
 		"source":           "eino",
 		"orchestration":    orchestration,
 		"reason":           reason,
+		"model":            modelName,
 		"modelCalls":       s.ModelCalls,
 		"promptTokens":     s.PromptTokens,
 		"completionTokens": s.CompletionTokens,
@@ -96,6 +98,7 @@ func (a *einoRunUsageAccumulator) EmitOnce(
 			zap.String("conversationId", conversationID),
 			zap.String("orchestration", orchestration),
 			zap.String("reason", reason),
+			zap.String("model", modelName),
 			zap.Int("modelCalls", s.ModelCalls),
 			zap.Int("promptTokens", s.PromptTokens),
 			zap.Int("completionTokens", s.CompletionTokens),

internal/multiagent/eino_run_usage_accumulator_test.go

@@ -49,16 +49,16 @@ func TestEinoRunUsageAccumulatorEmitOnce(t *testing.T) {
 		}
 	}
 
-	if !acc.EmitOnce("conv-1", "deep", "final", progress, nil) {
+	if !acc.EmitOnce("conv-1", "deep", "final", "gpt-test", progress, nil) {
 		t.Fatal("first emit should return true")
 	}
-	if acc.EmitOnce("conv-1", "deep", "partial", progress, nil) {
+	if acc.EmitOnce("conv-1", "deep", "partial", "gpt-test", progress, nil) {
 		t.Fatal("second emit should return false")
 	}
 	if len(events) != 1 {
 		t.Fatalf("events = %#v, want one usage summary", events)
 	}
-	if events[0]["conversationId"] != "conv-1" || events[0]["orchestration"] != "deep" || events[0]["reason"] != "final" || events[0]["totalTokens"] != 3 {
+	if events[0]["conversationId"] != "conv-1" || events[0]["orchestration"] != "deep" || events[0]["reason"] != "final" || events[0]["model"] != "gpt-test" || events[0]["totalTokens"] != 3 {
 		t.Fatalf("event = %#v", events[0])
 	}
 }

internal/multiagent/eino_single_runner.go

@@ -203,15 +203,18 @@ func RunEinoSingleChatModelAgent(
 	}
 
 	return runEinoADKAgentLoop(ctx, &einoADKRunLoopArgs{
-		OrchMode:                "eino_single",
+		OrchMode:             "eino_single",
-		OrchestratorName:        einoSingleAgentName,
+		OrchestratorName:     einoSingleAgentName,
-		ConversationID:          conversationID,
+		ConversationID:       conversationID,
-		Progress:                progress,
+		Progress:             progress,
-		Logger:                  logger,
+		Logger:               logger,
-		SnapshotMCPIDs:          snapshotMCPIDs,
+		SnapshotMCPIDs:       snapshotMCPIDs,
-		StreamsMainAssistant:    streamsMainAssistant,
+		StreamsMainAssistant: streamsMainAssistant,
-		EinoRoleTag:             einoRoleTag,
+		EinoRoleTag:          einoRoleTag,
-		CheckpointDir:           ma.EinoMiddleware.CheckpointDir,
+		// Chat history recovery is intentionally centralized in last_react_*.
+		// ADK checkpoints are a second persisted model-state channel and make
+		// stale-context bugs hard to reason about across user turns.
+		CheckpointDir:           "",
 		RunRetryMaxAttempts:     RunRetryMaxAttemptsFromConfig(&ma.EinoMiddleware),
 		RunRetryMaxBackoffSec:   int(einoRunRetryMaxBackoffFromConfig(&ma.EinoMiddleware).Seconds()),
 		McpIDsMu:                &mcpIDsMu,

internal/multiagent/runner.go

@@ -614,15 +614,18 @@ func RunDeepAgent(
 	}
 
 	return runEinoADKAgentLoop(ctx, &einoADKRunLoopArgs{
-		OrchMode:                orchMode,
+		OrchMode:             orchMode,
-		OrchestratorName:        orchestratorName,
+		OrchestratorName:     orchestratorName,
-		ConversationID:          conversationID,
+		ConversationID:       conversationID,
-		Progress:                progress,
+		Progress:             progress,
-		Logger:                  logger,
+		Logger:               logger,
-		SnapshotMCPIDs:          snapshotMCPIDs,
+		SnapshotMCPIDs:       snapshotMCPIDs,
-		StreamsMainAssistant:    streamsMainAssistant,
+		StreamsMainAssistant: streamsMainAssistant,
-		EinoRoleTag:             einoRoleTag,
+		EinoRoleTag:          einoRoleTag,
-		CheckpointDir:           ma.EinoMiddleware.CheckpointDir,
+		// Chat history recovery is intentionally centralized in last_react_*.
+		// ADK checkpoints are a second persisted model-state channel and make
+		// stale-context bugs hard to reason about across user turns.
+		CheckpointDir:           "",
 		RunRetryMaxAttempts:     RunRetryMaxAttemptsFromConfig(&ma.EinoMiddleware),
 		RunRetryMaxBackoffSec:   int(einoRunRetryMaxBackoffFromConfig(&ma.EinoMiddleware).Seconds()),
 		McpIDsMu:                &mcpIDsMu,

internal/openai/reasoning_payload_test.go

@@ -205,7 +205,7 @@ func TestReasoningToolChoiceCompatRoundTripperDeepSeek(t *testing.T) {
 	}
 }
 
-func TestReasoningToolChoiceCompatRoundTripperDeepSeekEndpointWinsOverProfile(t *testing.T) {
+func TestReasoningToolChoiceCompatRoundTripperOpenAIProfileWinsOverDeepSeekEndpoint(t *testing.T) {
 	var gotBody string
 	rt := &reasoningToolChoiceCompatRoundTripper{
 		cfg: &config.OpenAIConfig{
@@ -235,11 +235,11 @@ func TestReasoningToolChoiceCompatRoundTripperDeepSeekEndpointWinsOverProfile(t
 	if err != nil {
 		t.Fatal(err)
 	}
-	if strings.Contains(gotBody, "tool_choice") {
+	if !strings.Contains(gotBody, "tool_choice") {
-		t.Fatalf("expected DeepSeek tool_choice stripped despite openai_compat profile, got %s", gotBody)
+		t.Fatalf("expected tool_choice preserved for explicit openai_compat profile, got %s", gotBody)
 	}
 	if !strings.Contains(gotBody, "tools") {
-		t.Fatalf("expected tools preserved for DeepSeek, got %s", gotBody)
+		t.Fatalf("expected tools preserved for explicit openai_compat profile, got %s", gotBody)
 	}
 }
 

internal/openai/reasoning_tool_choice_compat.go

@@ -55,9 +55,6 @@ func isDeepSeekToolChoiceCompatProfile(cfg *config.OpenAIConfig) bool {
 	if cfg == nil {
 		return false
 	}
-	if cfg.IsDeepSeekEndpointOrModel() {
-		return true
-	}
 	profile := strings.ToLower(strings.TrimSpace(cfg.Reasoning.ProfileEffective()))
 	if profile == "deepseek" || profile == "deepseek_compat" {
 		return true
@@ -65,5 +62,5 @@ func isDeepSeekToolChoiceCompatProfile(cfg *config.OpenAIConfig) bool {
 	if profile != "" && profile != "auto" {
 		return false
 	}
-	return false
+	return cfg.IsDeepSeekEndpointOrModel()
 }

internal/reasoning/eino.go

@@ -36,7 +36,7 @@ func ApplyPlanExecutePlannerModelConfig(cfg *einoopenai.ChatModelConfig, oa *con
 	}
 	mergeExtraRequestFields(cfg, oa.Reasoning.ExtraRequestFields)
 	clearReasoningFromChatModelConfig(cfg)
-	if resolveWireProfile(oa, &oa.Reasoning) == wireDeepseek || oa.IsDeepSeekEndpointOrModel() {
+	if resolveWireProfile(oa, &oa.Reasoning) == wireDeepseek {
 		// DeepSeek enables thinking by default, so omission would not actually
 		// disable it for the planner's forced tool-choice requests.
 		applyThinkingDisabled(cfg)
@@ -88,9 +88,9 @@ func ApplyToEinoChatModelConfig(cfg *einoopenai.ChatModelConfig, oa *config.Open
 		clearReasoningFromChatModelConfig(cfg)
 		// Strict OpenAI endpoints reject unknown `thinking` fields, whereas the
 		// DeepSeek API enables thinking by default and requires an explicit
-		// thinking.type=disabled switch. Detect the actual DeepSeek target even
+		// thinking.type=disabled switch. The configured profile is authoritative;
-		// when the configured reasoning profile was left as openai_compat.
+		// auto-detection only happens inside resolveWireProfile for profile=auto.
-		if resolveWireProfile(oa, sr) == wireDeepseek || oa.IsDeepSeekEndpointOrModel() {
+		if resolveWireProfile(oa, sr) == wireDeepseek {
 			applyThinkingDisabled(cfg)
 		}
 		return
@@ -132,7 +132,7 @@ func AgenticOpenAIExtraFields(oa *config.OpenAIConfig, client *ClientIntent) map
 	fields := cloneExtraRequestFields(sr.ExtraRequestFields)
 	if mode == "off" {
 		clearReasoningExtraFields(fields)
-		if resolveWireProfile(oa, sr) == wireDeepseek || oa.IsDeepSeekEndpointOrModel() {
+		if resolveWireProfile(oa, sr) == wireDeepseek {
 			if fields == nil {
 				fields = make(map[string]any)
 			}
@@ -194,7 +194,7 @@ func AgenticOpenAIPlannerExtraFields(oa *config.OpenAIConfig) map[string]any {
 	}
 	fields := cloneExtraRequestFields(oa.Reasoning.ExtraRequestFields)
 	clearReasoningExtraFields(fields)
-	if resolveWireProfile(oa, &oa.Reasoning) == wireDeepseek || oa.IsDeepSeekEndpointOrModel() {
+	if resolveWireProfile(oa, &oa.Reasoning) == wireDeepseek {
 		if fields == nil {
 			fields = make(map[string]any)
 		}

internal/reasoning/eino_test.go

@@ -140,7 +140,7 @@ func TestAgenticOpenAIPlannerExtraFields_deepseekDisablesThinking(t *testing.T)
 	}
 }
 
-func TestAgenticOpenAIPlannerExtraFields_deepseekEndpointWinsOverOpenAIProfile(t *testing.T) {
+func TestAgenticOpenAIPlannerExtraFields_openAIProfileWinsOverDeepseekEndpoint(t *testing.T) {
 	oa := &config.OpenAIConfig{
 		BaseURL: "https://api.deepseek.com/v1",
 		Model:   "deepseek-v4-flash",
@@ -155,12 +155,10 @@ func TestAgenticOpenAIPlannerExtraFields_deepseekEndpointWinsOverOpenAIProfile(t
 		},
 	}
 	got := AgenticOpenAIPlannerExtraFields(oa)
-	if _, ok := got["reasoning_effort"]; ok {
+	for _, key := range reasoningPayloadKeysForTest {
-		t.Fatalf("planner should strip reasoning_effort: %#v", got)
+		if _, ok := got[key]; ok {
-	}
+			t.Fatalf("planner fields unexpectedly contain %q: %#v", key, got)
-	thinking, ok := got["thinking"].(map[string]any)
+		}
-	if !ok || thinking["type"] != "disabled" {
-		t.Fatalf("expected deepseek thinking disabled despite openai_compat profile, got %#v", got)
 	}
 	if got["vendor_option"] != true {
 		t.Fatalf("vendor option not preserved: %#v", got)
@@ -189,7 +187,7 @@ func TestApplyPlanExecutePlannerModelConfig_stripsReasoningWhenGlobalOn(t *testi
 	}
 }
 
-func TestApplyPlanExecutePlannerModelConfig_deepseekEndpointWinsOverOpenAIProfile(t *testing.T) {
+func TestApplyPlanExecutePlannerModelConfig_openAIProfileWinsOverDeepseekEndpoint(t *testing.T) {
 	cfg := &einoopenai.ChatModelConfig{ExtraFields: map[string]any{
 		"thinking":         map[string]any{"type": "enabled"},
 		"reasoning_effort": "high",
@@ -205,16 +203,7 @@ func TestApplyPlanExecutePlannerModelConfig_deepseekEndpointWinsOverOpenAIProfil
 		},
 	}
 	ApplyPlanExecutePlannerModelConfig(cfg, oa)
-	if cfg.ReasoningEffort != "" {
+	assertNoReasoningFields(t, cfg)
-		t.Fatalf("expected ReasoningEffort omitted, got %q", cfg.ReasoningEffort)
-	}
-	if _, ok := cfg.ExtraFields["reasoning_effort"]; ok {
-		t.Fatalf("expected reasoning_effort omitted, got %#v", cfg.ExtraFields)
-	}
-	thinking, ok := cfg.ExtraFields["thinking"].(map[string]any)
-	if !ok || thinking["type"] != "disabled" {
-		t.Fatalf("expected deepseek thinking disabled despite openai_compat profile, got %#v", cfg.ExtraFields)
-	}
 	if cfg.ExtraFields["vendor_option"] != true {
 		t.Fatalf("expected unrelated extra field preserved, got %#v", cfg.ExtraFields)
 	}
@@ -246,6 +235,89 @@ func TestApplyReasoningOff_omitsAllReasoningFields(t *testing.T) {
 	}
 }
 
+func TestApplyReasoningOff_openAICompatDeepseekModelOmitsAllReasoningFields(t *testing.T) {
+	allowClient := false
+	cfg := &einoopenai.ChatModelConfig{ExtraFields: map[string]any{
+		"thinking":         map[string]any{"type": "enabled"},
+		"reasoning_effort": "high",
+	}}
+	oa := &config.OpenAIConfig{
+		Provider: "openai_compatible",
+		BaseURL:  "http://your-gateway:port/v1",
+		Model:    "deepseek-v4-flash-0731",
+		Reasoning: config.OpenAIReasoningConfig{
+			Mode:                 "off",
+			Effort:               "high",
+			Profile:              "openai_compat",
+			AllowClientReasoning: &allowClient,
+			ExtraRequestFields: map[string]interface{}{
+				"thinking":      map[string]any{"type": "disabled"},
+				"output_config": map[string]any{"effort": "high"},
+				"vendor_option": true,
+			},
+		},
+	}
+	ApplyToEinoChatModelConfig(cfg, oa, nil)
+	assertNoReasoningFields(t, cfg)
+	if cfg.ExtraFields["vendor_option"] != true {
+		t.Fatalf("expected unrelated extra field preserved, got %#v", cfg.ExtraFields)
+	}
+}
+
+func TestAgenticOpenAIExtraFields_openAICompatDeepseekModelOmitsAllReasoningFields(t *testing.T) {
+	oa := &config.OpenAIConfig{
+		Provider: "openai_compatible",
+		BaseURL:  "http://your-gateway:port/v1",
+		Model:    "deepseek-v4-flash-0731",
+		Reasoning: config.OpenAIReasoningConfig{
+			Mode:    "off",
+			Effort:  "high",
+			Profile: "openai_compat",
+			ExtraRequestFields: map[string]interface{}{
+				"thinking":         map[string]any{"type": "disabled"},
+				"reasoning_effort": "high",
+				"vendor_option":    true,
+			},
+		},
+	}
+	got := AgenticOpenAIExtraFields(oa, nil)
+	for _, key := range reasoningPayloadKeysForTest {
+		if _, ok := got[key]; ok {
+			t.Fatalf("agentic fields unexpectedly contain %q: %#v", key, got)
+		}
+	}
+	if got["vendor_option"] != true {
+		t.Fatalf("vendor option not preserved: %#v", got)
+	}
+}
+
+func TestAgenticOpenAIPlannerExtraFields_openAICompatDeepseekModelOmitsAllReasoningFields(t *testing.T) {
+	oa := &config.OpenAIConfig{
+		Provider: "openai_compatible",
+		BaseURL:  "http://your-gateway:port/v1",
+		Model:    "deepseek-v4-flash-0731",
+		Reasoning: config.OpenAIReasoningConfig{
+			Mode:    "on",
+			Effort:  "high",
+			Profile: "openai_compat",
+			ExtraRequestFields: map[string]interface{}{
+				"thinking":         map[string]any{"type": "enabled"},
+				"reasoning_effort": "high",
+				"vendor_option":    true,
+			},
+		},
+	}
+	got := AgenticOpenAIPlannerExtraFields(oa)
+	for _, key := range reasoningPayloadKeysForTest {
+		if _, ok := got[key]; ok {
+			t.Fatalf("planner fields unexpectedly contain %q: %#v", key, got)
+		}
+	}
+	if got["vendor_option"] != true {
+		t.Fatalf("vendor option not preserved: %#v", got)
+	}
+}
+
 func TestApplyReasoningOff_clientOverrideOmit(t *testing.T) {
 	cfg := &einoopenai.ChatModelConfig{}
 	oa := &config.OpenAIConfig{Reasoning: config.OpenAIReasoningConfig{
@@ -256,7 +328,7 @@ func TestApplyReasoningOff_clientOverrideOmit(t *testing.T) {
 }
 
 func TestApplyReasoningOff_deepseekExplicitlyDisablesDefaultThinking(t *testing.T) {
-	for _, profile := range []string{"deepseek_compat", "auto", "openai_compat"} {
+	for _, profile := range []string{"deepseek_compat", "auto"} {
 		t.Run(profile, func(t *testing.T) {
 			cfg := &einoopenai.ChatModelConfig{ExtraFields: map[string]any{
 				"reasoning_effort": "high",
@@ -287,6 +359,25 @@ func TestApplyReasoningOff_deepseekExplicitlyDisablesDefaultThinking(t *testing.
 	}
 }
 
+func TestApplyReasoningOff_openAIProfileWinsOverDeepseekEndpoint(t *testing.T) {
+	cfg := &einoopenai.ChatModelConfig{ExtraFields: map[string]any{
+		"reasoning_effort": "high",
+		"vendor_option":    true,
+	}}
+	oa := &config.OpenAIConfig{
+		BaseURL: "https://api.deepseek.com",
+		Model:   "deepseek-v4-pro",
+		Reasoning: config.OpenAIReasoningConfig{
+			Mode: "off", Effort: "high", Profile: "openai_compat",
+		},
+	}
+	ApplyToEinoChatModelConfig(cfg, oa, nil)
+	assertNoReasoningFields(t, cfg)
+	if cfg.ExtraFields["vendor_option"] != true {
+		t.Fatalf("expected unrelated extra field preserved, got %#v", cfg.ExtraFields)
+	}
+}
+
 func TestApplyReasoningOff_wirePayloadOmitsThinking(t *testing.T) {
 	var requestBody map[string]any
 	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {

internal/security/rbac_middleware.go

@@ -118,6 +118,8 @@ func permissionForRequest(method, fullPath string) string {
 		return "hitl:write"
 	case strings.HasPrefix(path, "/agent-loop"), strings.HasPrefix(path, "/batch-tasks"):
 		return crudPermission(method, "tasks")
+	case path == "/usage/tokens":
+		return "dashboard:read"
 	case strings.HasPrefix(path, "/conversations"), strings.HasPrefix(path, "/messages"), strings.HasPrefix(path, "/process-details"):
 		return crudPermission(method, "chat")
 	case strings.HasPrefix(path, "/groups"):

internal/security/rbac_middleware_test.go

@@ -119,6 +119,12 @@ func TestRBACResourcePickerRequiresWritePermission(t *testing.T) {
 	}
 }
 
+func TestRBACMiddlewareMapsTokenUsageStatsToDashboardRead(t *testing.T) {
+	if got := permissionForRequest(http.MethodGet, "/api/usage/tokens"); got != "dashboard:read" {
+		t.Fatalf("token usage permission = %q, want dashboard:read", got)
+	}
+}
+
 func TestMCPInvocationPermissionIsSeparateFromMCPAdministration(t *testing.T) {
 	if got := permissionForRequest(http.MethodPost, "/api/mcp"); got != "mcp:execute" {
 		t.Fatalf("MCP invocation permission = %q, want mcp:execute", got)

web/static/css/style.css

@@ -858,7 +858,7 @@ html[data-theme="dark"] .vulnerability-alert-switch input:disabled + .vulnerabil
 }
 
 .conversation-sidebar {
-    width: 280px;
+    width: 320px;
     background: linear-gradient(180deg, #ffffff 0%, #fafbfc 100%);
     border-right: 1px solid var(--border-color);
     display: flex;
@@ -4107,62 +4107,50 @@ html[data-theme="dark"] .sidebar-content:hover::-webkit-scrollbar-thumb:hover {
     position: relative;
 }
 
-/* 消息复制按钮 - 位于消息气泡右下角 */
+/* 消息复制按钮 - 与时间戳同一行 */
 .message-copy-btn {
-    position: absolute;
+    position: static;
-    bottom: 12px;
+    display: inline-flex;
-    right: 12px;
-    display: flex;
     align-items: center;
     justify-content: center;
-    gap: 6px;
+    width: 28px;
-    padding: 8px 14px;
+    height: 28px;
-    background: #ffffff;
+    padding: 0;
-    border: 1px solid rgba(0, 0, 0, 0.08);
+    background: transparent;
-    border-radius: 20px;
+    border: 1px solid transparent;
-    color: #666;
+    border-radius: 6px;
-    font-size: 0.8125rem;
+    color: var(--text-secondary, #888);
     cursor: pointer;
-    transition: all 0.25s cubic-bezier(0.4, 0, 0.2, 1);
+    opacity: 0.72;
-    box-shadow: 0 2px 8px rgba(0, 0, 0, 0.08), 0 1px 2px rgba(0, 0, 0, 0.04);
+    flex-shrink: 0;
-    z-index: 10;
+    transition: opacity 0.2s ease, color 0.2s ease, background 0.2s ease, border-color 0.2s ease;
-    opacity: 0;
-    transform: translateY(4px);
-    font-weight: 500;
-}
-
-.message-bubble:hover .message-copy-btn {
-    opacity: 1;
-    transform: translateY(0);
 }
 
 .message-copy-btn:hover {
-    background: rgba(255, 255, 255, 1);
+    color: var(--accent-color, #0066ff);
-    border-color: rgba(0, 102, 255, 0.2);
+    background: rgba(0, 102, 255, 0.07);
-    color: #0066ff;
+    border-color: rgba(0, 102, 255, 0.14);
-    box-shadow: 0 4px 12px rgba(0, 102, 255, 0.15), 0 2px 4px rgba(0, 0, 0, 0.08);
+    opacity: 1;
-    transform: translateY(-1px);
 }
 
 .message-copy-btn:active {
-    transform: translateY(0) scale(0.98);
+    background: rgba(0, 102, 255, 0.11);
-    box-shadow: 0 2px 6px rgba(0, 102, 255, 0.12), 0 1px 2px rgba(0, 0, 0, 0.06);
 }
 
 .message-copy-btn svg {
     width: 16px;
     height: 16px;
     flex-shrink: 0;
-    transition: transform 0.2s ease;
 }
 
-.message-copy-btn:hover svg {
+.message-copy-btn:focus-visible {
-    transform: scale(1.1);
+    opacity: 1;
+    outline: 2px solid var(--accent-color, #0066ff);
+    outline-offset: 1px;
 }
 
 .message-copy-btn span {
-    font-weight: 500;
+    display: none;
-    letter-spacing: 0.01em;
 }
 
 /* 时间戳 + 删除本轮（与气泡分离，和「展开详情」同一视觉层级） */
@@ -24302,11 +24290,15 @@ tr.mcp-stats-tool-row[data-tool-name]:focus-visible {
 /* 第一行：核心 KPI（视觉层次 + 色块强调） */
 .dashboard-kpi-row {
     display: grid;
-    grid-template-columns: repeat(4, 1fr);
+    grid-template-columns: repeat(5, minmax(0, 1fr));
     gap: 16px;
     margin-bottom: 24px;
 }
 
+@media (max-width: 1200px) {
+    .dashboard-kpi-row { grid-template-columns: repeat(3, 1fr); }
+}
+
 @media (max-width: 900px) {
     .dashboard-kpi-row { grid-template-columns: repeat(2, 1fr); }
 }
@@ -24466,6 +24458,7 @@ tr.mcp-stats-tool-row[data-tool-name]:focus-visible {
 .dashboard-kpi-card:nth-child(2) { background: linear-gradient(145deg, #fff 0%, #fef2f2 100%); }
 .dashboard-kpi-card:nth-child(3) { background: linear-gradient(145deg, #fff 0%, #f0fdf4 100%); }
 .dashboard-kpi-card:nth-child(4) { background: linear-gradient(145deg, #fff 0%, #f0fdfa 100%); }
+.dashboard-kpi-card:nth-child(5) { background: linear-gradient(145deg, #fff 0%, #f8fafc 100%); }
 
 .dashboard-kpi-card:hover {
     transform: translateY(-3px);
@@ -24494,6 +24487,7 @@ tr.mcp-stats-tool-row[data-tool-name]:focus-visible {
 .dashboard-kpi-icon-vuln { background: rgba(239, 68, 68, 0.1); color: #ef4444; }
 .dashboard-kpi-icon-calls { background: rgba(34, 197, 94, 0.1); color: #22c55e; }
 .dashboard-kpi-icon-rate { background: rgba(20, 184, 166, 0.1); color: #14b8a6; }
+.dashboard-kpi-icon-tokens { background: rgba(99, 102, 241, 0.1); color: #6366f1; }
 
 .dashboard-kpi-value {
     font-size: 1.875rem;
@@ -28734,12 +28728,54 @@ html[data-theme="dark"] .vulnerability-details code.vuln-detail-field-value {
 }
 
 .role-selector-icon {
-    font-size: 1rem;
-    line-height: 1;
     flex-shrink: 0;
-    display: flex;
+}
+
+.agent-mode-logo {
+    --agent-logo-a: #858d98;
+    --agent-logo-b: #858d98;
+    display: inline-flex;
     align-items: center;
     justify-content: center;
+    width: 18px;
+    height: 18px;
+    border: 0;
+    border-radius: 0;
+    background: transparent;
+    color: var(--agent-logo-a);
+    box-shadow: none;
+    overflow: visible;
+    flex: 0 0 auto;
+}
+
+.agent-mode-logo--deep {
+    --agent-logo-a: #858d98;
+    --agent-logo-b: #858d98;
+}
+
+.agent-mode-logo--plan {
+    --agent-logo-a: #858d98;
+    --agent-logo-b: #858d98;
+}
+
+.agent-mode-logo--supervisor {
+    --agent-logo-a: #858d98;
+    --agent-logo-b: #858d98;
+}
+
+.agent-mode-logo--default {
+    --agent-logo-a: #858d98;
+    --agent-logo-b: #858d98;
+}
+
+.agent-mode-logo__svg {
+    width: 100%;
+    height: 100%;
+    display: block;
+    stroke: currentColor;
+    stroke-width: 1.9;
+    stroke-linecap: round;
+    stroke-linejoin: round;
 }
 
 .role-selector-text {
@@ -29049,6 +29085,16 @@ html[data-theme="dark"] .vulnerability-details code.vuln-detail-field-value {
     border-color: rgba(138, 43, 226, 0.3);
     box-shadow: 0 2px 6px rgba(138, 43, 226, 0.2);
 }
+
+.role-selection-item-main .role-selection-item-icon-main.agent-mode-logo,
+.role-selection-item-main:hover .role-selection-item-icon-main.agent-mode-logo,
+.role-selection-item-main.selected .role-selection-item-icon-main.agent-mode-logo {
+    width: 28px;
+    height: 28px;
+    border-color: transparent;
+    background: transparent;
+    box-shadow: none;
+}
 .role-selection-item-content-main {
     display: flex;
     flex-direction: column;
@@ -36309,7 +36355,8 @@ html[data-theme="dark"] .conversation-reasoning-card .hitl-reviewer-toggle-btn.i
 html[data-theme="dark"] .dashboard-kpi-card:nth-child(1),
 html[data-theme="dark"] .dashboard-kpi-card:nth-child(2),
 html[data-theme="dark"] .dashboard-kpi-card:nth-child(3),
-html[data-theme="dark"] .dashboard-kpi-card:nth-child(4) {
+html[data-theme="dark"] .dashboard-kpi-card:nth-child(4),
+html[data-theme="dark"] .dashboard-kpi-card:nth-child(5) {
     background: linear-gradient(145deg, #111827 0%, #172033 100%);
 }
 
@@ -36797,21 +36844,17 @@ html[data-theme="dark"] .webshell-ai-msg.assistant.webshell-ai-candidate-output
 }
 
 html[data-theme="dark"] .message-copy-btn {
-    background: #1f2937;
-    border-color: #334155;
     color: var(--text-secondary);
-    box-shadow: 0 2px 8px rgba(0, 0, 0, 0.28);
 }
 
 html[data-theme="dark"] .message-copy-btn:hover {
-    background: #263244;
+    background: rgba(96, 165, 250, 0.12);
     border-color: rgba(96, 165, 250, 0.45);
     color: var(--accent-hover);
-    box-shadow: 0 4px 12px rgba(96, 165, 250, 0.18);
 }
 
 html[data-theme="dark"] .message-copy-btn:active {
-    box-shadow: 0 2px 6px rgba(96, 165, 250, 0.14);
+    background: rgba(96, 165, 250, 0.18);
 }
 
 html[data-theme="dark"] .message.user .message-bubble {
@@ -37364,6 +37407,16 @@ html[data-theme="dark"] .role-selection-item-icon-main {
     border-color: rgba(148, 163, 184, 0.12) !important;
 }
 
+html[data-theme="dark"] .role-selection-item-icon-main.agent-mode-logo,
+html[data-theme="dark"] .project-conversation-preview-mode-icon.agent-mode-logo,
+html[data-theme="dark"] .role-selector-icon.agent-mode-logo {
+    --agent-logo-a: #94a3b8;
+    --agent-logo-b: #94a3b8;
+    border-color: transparent !important;
+    background: transparent !important;
+    box-shadow: none !important;
+}
+
 html[data-theme="dark"] .sidebar-list-pagination,
 html[data-theme="dark"] .sidebar-list-pagination-inner,
 html[data-theme="dark"] .conversation-sidebar-pagination,
@@ -44995,7 +45048,7 @@ html[data-theme="dark"] .project-folder-preview-edit:focus-visible {
 .project-conversation-preview {
     position: fixed;
     z-index: 1200;
-    width: min(300px, calc(100vw - 32px));
+    width: min(340px, calc(100vw - 32px));
     padding: 12px 14px 11px;
     border: 1px solid rgba(30, 41, 59, 0.15);
     border-radius: 14px;
@@ -45011,11 +45064,12 @@ html[data-theme="dark"] .project-folder-preview-edit:focus-visible {
 }
 
 .project-conversation-preview-header {
-    display: grid;
+    display: flex;
-    grid-template-columns: minmax(0, 1fr) auto;
+    flex-direction: column;
-    align-items: baseline;
+    align-items: flex-start;
-    gap: 10px;
+    gap: 3px;
     margin-bottom: 8px;
+    min-width: 0;
 }
 
 .project-conversation-preview-title {
@@ -45025,8 +45079,11 @@ html[data-theme="dark"] .project-folder-preview-edit:focus-visible {
     font-size: 0.9rem;
     font-weight: 650;
     line-height: 1.35;
-    text-overflow: ellipsis;
+    display: -webkit-box;
-    white-space: nowrap;
+    -webkit-line-clamp: 2;
+    line-clamp: 2;
+    -webkit-box-orient: vertical;
+    overflow-wrap: anywhere;
 }
 
 .project-conversation-preview-age {
@@ -45058,6 +45115,13 @@ html[data-theme="dark"] .project-folder-preview-edit:focus-visible {
     color: #858d98;
 }
 
+.project-conversation-preview-mode-icon.agent-mode-logo {
+    width: 16px;
+    height: 16px;
+    margin-right: 0;
+    color: #858d98;
+}
+
 .project-conversation-preview-project,
 .project-conversation-preview-mode {
     min-width: 0;
@@ -46346,15 +46410,6 @@ html[data-theme="dark"] .chat-composer-surface > .chat-session-settings-popover
     margin-bottom: 0;
 }
 
-.message.assistant-turn-with-process .message-copy-btn {
-    right: 0;
-    bottom: -34px;
-    padding: 5px 9px;
-    border: 0;
-    background: transparent;
-    box-shadow: none;
-}
-
 .message.assistant-turn-with-process .mcp-call-section {
     width: 100%;
     margin: 0;
@@ -46404,8 +46459,9 @@ html[data-theme="dark"] .chat-composer-surface > .chat-session-settings-popover
 .turn-process-leading {
     display: inline-flex;
     align-items: center;
-    gap: 9px;
+    gap: 8px;
     min-width: 0;
+    flex-wrap: wrap;
 }
 
 .turn-process-status-dot {
@@ -46422,6 +46478,26 @@ html[data-theme="dark"] .chat-composer-surface > .chat-session-settings-popover
     animation: codex-turn-pulse 1.55s ease-in-out infinite;
 }
 
+.turn-process-token-chip {
+    display: inline;
+    padding: 0;
+    border: 0;
+    border-radius: 0;
+    background: transparent;
+    color: inherit;
+    font-size: inherit;
+    font-weight: inherit;
+    line-height: inherit;
+    opacity: 0.72;
+    white-space: nowrap;
+}
+
+.turn-process-token-chip::before {
+    content: "· ";
+    color: inherit;
+    font-weight: inherit;
+}
+
 @keyframes codex-turn-pulse {
     0%, 100% { opacity: 0.55; transform: scale(0.88); }
     50% { opacity: 1; transform: scale(1); }

web/static/i18n/en-US.json

@@ -127,6 +127,9 @@
     "vulnTotal": "Total vulnerabilities",
     "toolCalls": "Tool invocations",
     "successRate": "Tool success rate",
+    "tokenUsage": "Token usage",
+    "tokenUsageSub": "Last 7 days {{calls}} calls · Today {{today}}",
+    "noTokenUsageYet": "No usage yet",
     "clickToViewTasks": "Click to view tasks",
     "clickToViewChat": "Click to view conversations",
     "clickToViewVuln": "Click to view vulnerabilities",
@@ -633,6 +636,8 @@
     "turnDurationMinutes": "{{minutes}} min {{seconds}} sec",
     "turnDurationHours": "{{hours}} hr {{minutes}} min",
     "turnProcessAria": "{{state}}; expand or collapse execution details",
+    "turnTokenUsageLabel": "{{tokens}} tokens",
+    "turnTokenUsageTitle": "Token usage: {{total}} (input {{prompt}}, output {{completion}}, cached {{cached}}, reasoning {{reasoning}}, {{calls}} calls)",
     "turnNumber": "Turn {{number}}",
     "turnPending": "Processing…",
     "expandDetailLazyHint": "Expand details (loads iteration details on click)",

web/static/i18n/zh-CN.json

@@ -127,6 +127,9 @@
     "vulnTotal": "漏洞总数",
     "toolCalls": "工具调用次数",
     "successRate": "工具执行成功率",
+    "tokenUsage": "Token 用量",
+    "tokenUsageSub": "近 7 天 {{calls}} 次调用 · 今日 {{today}}",
+    "noTokenUsageYet": "暂无用量",
     "clickToViewTasks": "点击查看任务管理",
     "clickToViewChat": "点击查看对话",
     "clickToViewVuln": "点击查看漏洞管理",
@@ -621,6 +624,8 @@
     "turnDurationMinutes": "{{minutes}} 分钟 {{seconds}} 秒",
     "turnDurationHours": "{{hours}} 小时 {{minutes}} 分钟",
     "turnProcessAria": "{{state}}，展开或收起执行过程",
+    "turnTokenUsageLabel": "{{tokens}} tokens",
+    "turnTokenUsageTitle": "Token 用量：{{total}}（输入 {{prompt}}，输出 {{completion}}，缓存 {{cached}}，推理 {{reasoning}}，调用 {{calls}} 次）",
     "turnNumber": "第 {{number}} 轮",
     "turnPending": "正在处理…",
     "expandDetailLazyHint": "展开详情（点击后加载迭代详情）",

web/static/js/chat-message-copy.test.cjs

@@ -0,0 +1,35 @@
+const fs = require('node:fs');
+const test = require('node:test');
+const assert = require('node:assert/strict');
+
+const chat = fs.readFileSync('web/static/js/chat.js', 'utf8');
+const monitor = fs.readFileSync('web/static/js/monitor.js', 'utf8');
+
+function functionSource(source, name, nextName) {
+    const start = source.indexOf(`function ${name}(`);
+    const end = source.indexOf(`function ${nextName}(`, start);
+    assert.notEqual(start, -1, `${name} should exist`);
+    assert.notEqual(end, -1, `${nextName} should follow ${name}`);
+    return source.slice(start, end);
+}
+
+test('用户和助手消息使用同一复制按钮入口', () => {
+    const helperSource = functionSource(chat, 'appendMessageCopyButton', 'addMessage');
+    const addMessageSource = functionSource(chat, 'addMessage', 'copyMessageToClipboard');
+
+    assert.match(helperSource, /classList\.contains\('assistant'\)[\s\S]*classList\.contains\('user'\)/);
+    assert.match(helperSource, /const footer = ensureMessageMetaFooter\(content\)/);
+    assert.match(helperSource, /message-bubble \.message-copy-btn/);
+    assert.match(helperSource, /copyMessageToClipboard\(messageDiv, this\)/);
+    assert.match(addMessageSource, /role === 'assistant' \|\| role === 'user'[\s\S]*messageDiv\.dataset\.originalContent = content/);
+    assert.match(addMessageSource, /metaFooter\.appendChild\(timeDiv\)/);
+    assert.match(addMessageSource, /role === 'assistant' \|\| role === 'user'[\s\S]*appendMessageCopyButton\(messageDiv\)/);
+});
+
+test('刷新消息内容时会保留或补回复制按钮', () => {
+    const refreshSource = functionSource(chat, 'refreshSystemReadyMessageBubbles', 'appendMessageCopyButton');
+    const updateSource = functionSource(monitor, 'updateAssistantBubbleContent', 'isConversationTaskRunning');
+
+    assert.match(refreshSource, /appendMessageCopyButton\(messageDiv\)/);
+    assert.match(updateSource, /window\.appendMessageCopyButton\(assistantElement\)/);
+});

web/static/js/chat.js

@@ -1001,23 +1001,30 @@ function getAgentModeLabelForValue(mode) {
     }
 }
 
-function getAgentModeIconForValue(mode) {
+function getAgentModeIconClassForValue(mode) {
     switch (mode) {
-        case CHAT_AGENT_MODE_EINO_SINGLE: return '⚡';
+        case CHAT_AGENT_MODE_EINO_SINGLE: return 'eino';
-        case 'deep': return '🧩';
+        case 'deep': return 'deep';
-        case 'plan_execute': return '📋';
+        case 'plan_execute': return 'plan';
-        case 'supervisor': return '🎯';
+        case 'supervisor': return 'supervisor';
-        default: return '🤖';
+        default: return 'default';
     }
 }
 
+function renderAgentModeLogoMarkup() {
+    return '<svg class="agent-mode-logo__svg" viewBox="0 0 24 24" fill="none" aria-hidden="true"><rect x="3" y="11" width="18" height="10" rx="2"/><circle cx="12" cy="5" r="2"/><path d="M12 7v4"/><path d="M8 16h.01"/><path d="M16 16h.01"/></svg>';
+}
+
 function syncAgentModeFromValue(value) {
     const hid = document.getElementById('agent-mode-select');
     const label = document.getElementById('agent-mode-text');
     const icon = document.getElementById('agent-mode-icon');
     if (hid) hid.value = value;
     if (label) label.textContent = getAgentModeLabelForValue(value);
-    if (icon) icon.textContent = getAgentModeIconForValue(value);
+    if (icon) {
+        icon.className = 'role-selector-icon agent-mode-logo agent-mode-logo--' + getAgentModeIconClassForValue(value);
+        icon.innerHTML = renderAgentModeLogoMarkup();
+    }
     document.querySelectorAll('.agent-mode-option').forEach(function (el) {
         const v = el.getAttribute('data-value');
         el.classList.toggle('selected', v === value);
@@ -3507,9 +3514,60 @@ function refreshSystemReadyMessageBubbles() {
         bubble.innerHTML = formattedContent;
         if (typeof wrapTablesInBubble === 'function') wrapTablesInBubble(bubble);
         messageDiv.dataset.originalContent = text;
+        appendMessageCopyButton(messageDiv);
     });
 }
 
+function ensureMessageMetaFooter(content) {
+    if (!content) return null;
+    let footer = content.querySelector('.message-meta-footer');
+    if (footer) return footer;
+    const timeDiv = content.querySelector('.message-time');
+    footer = document.createElement('div');
+    footer.className = 'message-meta-footer';
+    if (timeDiv && timeDiv.parentNode === content) {
+        timeDiv.parentNode.insertBefore(footer, timeDiv);
+        footer.appendChild(timeDiv);
+    } else {
+        content.appendChild(footer);
+    }
+    return footer;
+}
+
+function appendMessageCopyButton(messageDiv) {
+    if (!messageDiv) return null;
+    if (!messageDiv.classList || (!messageDiv.classList.contains('assistant') && !messageDiv.classList.contains('user'))) {
+        return null;
+    }
+    const content = messageDiv.querySelector('.message-content');
+    const footer = ensureMessageMetaFooter(content);
+    if (!footer) return null;
+
+    messageDiv.querySelectorAll('.message-bubble .message-copy-btn').forEach((btn) => btn.remove());
+    let copyBtn = footer.querySelector('.message-copy-btn');
+    if (copyBtn) return copyBtn;
+
+    copyBtn = document.createElement('button');
+    copyBtn.type = 'button';
+    copyBtn.className = 'message-copy-btn';
+    copyBtn.innerHTML = '<svg width="16" height="16" viewBox="0 0 24 24" fill="none" xmlns="http://www.w3.org/2000/svg"><rect x="9" y="9" width="13" height="13" rx="2" ry="2" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round" fill="none"/><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round" fill="none"/></svg><span>' + (typeof window.t === 'function' ? window.t('common.copy') : '复制') + '</span>';
+    copyBtn.title = typeof window.t === 'function' ? window.t('chat.copyMessageTitle') : '复制消息内容';
+    copyBtn.setAttribute('aria-label', copyBtn.title);
+    copyBtn.onclick = function(e) {
+        e.stopPropagation();
+        copyMessageToClipboard(messageDiv, this);
+    };
+    const deleteBtn = footer.querySelector('.message-delete-turn-btn');
+    if (deleteBtn) {
+        footer.insertBefore(copyBtn, deleteBtn);
+    } else {
+        footer.appendChild(copyBtn);
+    }
+    return copyBtn;
+}
+window.appendMessageCopyButton = appendMessageCopyButton;
+window.ensureMessageMetaFooter = ensureMessageMetaFooter;
+
 // 添加消息（options.systemReadyMessage 为 true 时，语言切换会刷新该条文案）
 function addMessage(role, content, mcpExecutionIds = null, progressId = null, createdAt = null, options = null) {
     const messagesDiv = document.getElementById('chat-messages');
@@ -3581,23 +3639,10 @@ function addMessage(role, content, mcpExecutionIds = null, progressId = null, cr
     contentWrapper.appendChild(bubble);
     
     // 保存原始内容到消息元素，用于复制功能
-    if (role === 'assistant') {
+    if (role === 'assistant' || role === 'user') {
         messageDiv.dataset.originalContent = content;
     }
     
-    // 为助手消息添加复制按钮（复制整个回复内容）- 放在消息气泡右下角
-    if (role === 'assistant') {
-        const copyBtn = document.createElement('button');
-        copyBtn.className = 'message-copy-btn';
-        copyBtn.innerHTML = '<svg width="16" height="16" viewBox="0 0 24 24" fill="none" xmlns="http://www.w3.org/2000/svg"><rect x="9" y="9" width="13" height="13" rx="2" ry="2" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round" fill="none"/><path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1" stroke="currentColor" stroke-width="1.5" stroke-linecap="round" stroke-linejoin="round" fill="none"/></svg><span>' + (typeof window.t === 'function' ? window.t('common.copy') : '复制') + '</span>';
-        copyBtn.title = typeof window.t === 'function' ? window.t('chat.copyMessageTitle') : '复制消息内容';
-        copyBtn.onclick = function(e) {
-            e.stopPropagation();
-            copyMessageToClipboard(messageDiv, this);
-        };
-        bubble.appendChild(copyBtn);
-    }
-    
     // 添加时间戳
     const timeDiv = document.createElement('div');
     timeDiv.className = 'message-time';
@@ -3626,8 +3671,16 @@ function addMessage(role, content, mcpExecutionIds = null, progressId = null, cr
     try {
         timeDiv.dataset.messageTime = messageTime.toISOString();
     } catch (e) { /* ignore */ }
-    contentWrapper.appendChild(timeDiv);
+    const metaFooter = document.createElement('div');
+    metaFooter.className = 'message-meta-footer';
+    metaFooter.appendChild(timeDiv);
+    contentWrapper.appendChild(metaFooter);
     messageDiv.appendChild(contentWrapper);
+
+    // 为用户和助手消息添加复制按钮（复制整条消息内容）
+    if (role === 'assistant' || role === 'user') {
+        appendMessageCopyButton(messageDiv);
+    }
     
     // 有 MCP 执行记录且非流式占位消息时展示调用按钮；带 progressId 的流式占位不挂此条（与进度卡片一致，结束时 integrate 再创建）
     if (role === 'assistant' && (mcpExecutionIds && Array.isArray(mcpExecutionIds) && mcpExecutionIds.length > 0) && !progressId) {
@@ -4132,6 +4185,10 @@ function renderProcessDetails(messageId, processDetails, options) {
         detailsContainer.dataset.lazyNotLoaded = '0';
         detailsContainer.dataset.loaded = '1';
     }
+    const turnUsageFromDetails = extractAssistantTurnTokenUsage(processDetails);
+    if (turnUsageFromDetails) {
+        setAssistantTurnTokenUsage(messageElement, turnUsageFromDetails);
+    }
     processDetails = mergeMessageReasoningContentIntoProcessDetails(processDetails, reasoningFromMessage);
     processDetails = filterNoiseProcessDetails(processDetails);
     processDetails = dedupeConsecutiveProcessDetailRows(processDetails);
@@ -4428,7 +4485,7 @@ function renderProcessDetails(messageId, processDetails, options) {
                 }
             }
         }
-        if (!timelineOpts.toolStatus && eventType === 'tool_call' && detail.id && toolStatusByProcessDetailId.has(String(detail.id))) {
+        if (eventType === 'tool_call' && detail.id && toolStatusByProcessDetailId.has(String(detail.id))) {
             timelineOpts.toolStatus = toolStatusByProcessDetailId.get(String(detail.id));
         }
         const itemId = addTimelineItem(timeline, eventType, timelineOpts);
@@ -4887,6 +4944,134 @@ function formatAssistantTurnDuration(durationMs) {
         : seconds + ' 秒';
 }
 
+function assistantTurnUsageNumber(value) {
+    const n = Number(value);
+    return Number.isFinite(n) && n > 0 ? Math.round(n) : 0;
+}
+
+function normalizeAssistantTurnTokenUsage(data) {
+    const source = data && typeof data === 'object' ? data : {};
+    const usage = {
+        modelCalls: assistantTurnUsageNumber(source.modelCalls),
+        promptTokens: assistantTurnUsageNumber(source.promptTokens),
+        completionTokens: assistantTurnUsageNumber(source.completionTokens),
+        totalTokens: assistantTurnUsageNumber(source.totalTokens),
+        cachedTokens: assistantTurnUsageNumber(source.cachedTokens),
+        reasoningTokens: assistantTurnUsageNumber(source.reasoningTokens),
+        model: source.model != null ? String(source.model).trim() : ''
+    };
+    if (usage.totalTokens <= 0 && (usage.promptTokens > 0 || usage.completionTokens > 0)) {
+        usage.totalTokens = usage.promptTokens + usage.completionTokens;
+    }
+    return usage.totalTokens > 0 ? usage : null;
+}
+
+function mergeAssistantTurnTokenUsage(target, usage) {
+    if (!usage) return target || null;
+    const out = target || {
+        modelCalls: 0,
+        promptTokens: 0,
+        completionTokens: 0,
+        totalTokens: 0,
+        cachedTokens: 0,
+        reasoningTokens: 0,
+        model: ''
+    };
+    out.modelCalls += usage.modelCalls || 0;
+    out.promptTokens += usage.promptTokens || 0;
+    out.completionTokens += usage.completionTokens || 0;
+    out.totalTokens += usage.totalTokens || 0;
+    out.cachedTokens += usage.cachedTokens || 0;
+    out.reasoningTokens += usage.reasoningTokens || 0;
+    if (!out.model && usage.model) out.model = usage.model;
+    return out.totalTokens > 0 ? out : null;
+}
+
+function extractAssistantTurnTokenUsage(processDetails) {
+    if (!Array.isArray(processDetails)) return null;
+    let total = null;
+    processDetails.forEach((detail) => {
+        if (!detail || String(detail.eventType || '').trim() !== 'eino_usage_summary') return;
+        const usage = normalizeAssistantTurnTokenUsage(detail.data);
+        total = mergeAssistantTurnTokenUsage(total, usage);
+    });
+    return total;
+}
+
+function setAssistantTurnTokenUsage(messageElementOrId, usage) {
+    const messageElement = typeof messageElementOrId === 'string'
+        ? document.getElementById(messageElementOrId)
+        : messageElementOrId;
+    if (!messageElement || !messageElement.dataset) return;
+    const normalized = normalizeAssistantTurnTokenUsage(usage);
+    if (!normalized) {
+        delete messageElement.dataset.turnModelCalls;
+        delete messageElement.dataset.turnPromptTokens;
+        delete messageElement.dataset.turnCompletionTokens;
+        delete messageElement.dataset.turnTotalTokens;
+        delete messageElement.dataset.turnCachedTokens;
+        delete messageElement.dataset.turnReasoningTokens;
+        delete messageElement.dataset.turnModel;
+        syncAssistantTurnSummary(messageElement);
+        return;
+    }
+    messageElement.dataset.turnModelCalls = String(normalized.modelCalls || 0);
+    messageElement.dataset.turnPromptTokens = String(normalized.promptTokens || 0);
+    messageElement.dataset.turnCompletionTokens = String(normalized.completionTokens || 0);
+    messageElement.dataset.turnTotalTokens = String(normalized.totalTokens || 0);
+    messageElement.dataset.turnCachedTokens = String(normalized.cachedTokens || 0);
+    messageElement.dataset.turnReasoningTokens = String(normalized.reasoningTokens || 0);
+    if (normalized.model) {
+        messageElement.dataset.turnModel = normalized.model;
+    } else {
+        delete messageElement.dataset.turnModel;
+    }
+    syncAssistantTurnSummary(messageElement);
+}
+
+function getAssistantTurnTokenUsage(messageElement) {
+    if (!messageElement || !messageElement.dataset) return null;
+    return normalizeAssistantTurnTokenUsage({
+        modelCalls: messageElement.dataset.turnModelCalls,
+        promptTokens: messageElement.dataset.turnPromptTokens,
+        completionTokens: messageElement.dataset.turnCompletionTokens,
+        totalTokens: messageElement.dataset.turnTotalTokens,
+        cachedTokens: messageElement.dataset.turnCachedTokens,
+        reasoningTokens: messageElement.dataset.turnReasoningTokens,
+        model: messageElement.dataset.turnModel
+    });
+}
+
+function formatAssistantTurnTokenCount(value) {
+    const n = assistantTurnUsageNumber(value);
+    if (n >= 1000000) return (n / 1000000).toFixed(n >= 10000000 ? 0 : 1).replace(/\.0$/, '') + 'M';
+    if (n >= 1000) return (n / 1000).toFixed(n >= 100000 ? 0 : 1).replace(/\.0$/, '') + 'K';
+    return String(n);
+}
+
+function formatAssistantTurnTokenUsageLabel(usage) {
+    const tokens = formatAssistantTurnTokenCount(usage && usage.totalTokens);
+    return typeof window.t === 'function'
+        ? window.t('chat.turnTokenUsageLabel', { tokens: tokens })
+        : tokens + ' tokens';
+}
+
+function formatAssistantTurnTokenUsageTitle(usage) {
+    const safeUsage = usage || {};
+    const values = {
+        total: formatAssistantTurnTokenCount(safeUsage.totalTokens),
+        prompt: formatAssistantTurnTokenCount(safeUsage.promptTokens),
+        completion: formatAssistantTurnTokenCount(safeUsage.completionTokens),
+        cached: formatAssistantTurnTokenCount(safeUsage.cachedTokens),
+        reasoning: formatAssistantTurnTokenCount(safeUsage.reasoningTokens),
+        calls: formatAssistantTurnTokenCount(safeUsage.modelCalls),
+        model: safeUsage.model || ''
+    };
+    return typeof window.t === 'function'
+        ? window.t('chat.turnTokenUsageTitle', values)
+        : 'Token usage: ' + values.total + ' (input ' + values.prompt + ', output ' + values.completion + ')';
+}
+
 function assistantTurnTimestamp(value) {
     if (value == null || value === '') return NaN;
     const n = new Date(value).getTime();
@@ -4985,6 +5170,12 @@ function syncAssistantTurnSummary(messageElementOrId) {
             : 0;
     }
     const duration = formatAssistantTurnDuration(durationMs);
+    const tokenUsage = getAssistantTurnTokenUsage(messageElement);
+    const tokenUsageHtml = tokenUsage
+        ? '<span class="turn-process-token-chip" title="' + escapeHtml(formatAssistantTurnTokenUsageTitle(tokenUsage)) + '">' +
+            escapeHtml(formatAssistantTurnTokenUsageLabel(tokenUsage)) +
+          '</span>'
+        : '';
     let text;
     if (status === 'running') {
         text = typeof window.t === 'function' ? window.t('chat.turnElapsedRunning', { duration: duration }) : '已处理 ' + duration;
@@ -5001,6 +5192,7 @@ function syncAssistantTurnSummary(messageElementOrId) {
         <span class="turn-process-leading">
             <span class="turn-process-status-dot${status === 'running' ? ' is-running' : ''}" aria-hidden="true"></span>
             <span class="turn-process-summary-text">${escapeHtml(text)}</span>
+            ${tokenUsageHtml}
         </span>
         <svg class="turn-process-chevron" viewBox="0 0 20 20" fill="none" aria-hidden="true"><path d="M7.5 5.5L12 10l-4.5 4.5" stroke="currentColor" stroke-width="1.6" stroke-linecap="round" stroke-linejoin="round"/></svg>
     `;
@@ -5012,8 +5204,10 @@ function syncAssistantTurnSummary(messageElementOrId) {
 }
 
 window.setAssistantTurnTiming = setAssistantTurnTiming;
+window.setAssistantTurnTokenUsage = setAssistantTurnTokenUsage;
 window.syncAssistantTurnSummary = syncAssistantTurnSummary;
 window.formatAssistantTurnDuration = formatAssistantTurnDuration;
+window.extractAssistantTurnTokenUsage = extractAssistantTurnTokenUsage;
 
 /** 渗透测试区：工具栏（展开详情 | N次工具执行）+ 独立工具列表 + 迭代时间线 */
 function ensureMcpCallSectionChrome(messageElement, messageId) {
@@ -6156,6 +6350,43 @@ async function prefetchLastAssistantProcessDetails() {
     }
 }
 
+async function hydrateConversationTokenUsage(conversationId, expectedSeq, signal) {
+    const id = String(conversationId || '').trim();
+    if (!id || typeof apiFetch !== 'function' || typeof window.setAssistantTurnTokenUsage !== 'function') return;
+    if (signal && signal.aborted) return;
+    const params = new URLSearchParams();
+    params.set('since', '1970-01-01');
+    params.set('limit', '500');
+    const res = await apiFetch(
+        '/api/conversations/' + encodeURIComponent(id) + '/token-usage?' + params.toString(),
+        signal ? { signal: signal } : undefined
+    );
+    const payload = await res.json().catch(() => ({}));
+    if (!res.ok || (signal && signal.aborted)) return;
+    if (expectedSeq != null && expectedSeq !== loadConversationRequestSeq) return;
+    if (currentConversationId !== id) return;
+    const rows = Array.isArray(payload && payload.recent) ? payload.recent : [];
+    if (rows.length === 0) return;
+    const byMessage = new Map();
+    rows.forEach((row) => {
+        const messageId = row && row.messageId != null ? String(row.messageId).trim() : '';
+        if (!messageId) return;
+        const usage = normalizeAssistantTurnTokenUsage(row);
+        if (!usage) return;
+        byMessage.set(messageId, mergeAssistantTurnTokenUsage(byMessage.get(messageId) || null, usage));
+    });
+    if (byMessage.size === 0) return;
+    document.querySelectorAll('#chat-messages .message.assistant[data-backend-message-id]').forEach((messageElement) => {
+        const backendMessageId = messageElement && messageElement.dataset
+            ? String(messageElement.dataset.backendMessageId || '').trim()
+            : '';
+        const usage = backendMessageId ? byMessage.get(backendMessageId) : null;
+        if (usage) {
+            window.setAssistantTurnTokenUsage(messageElement, usage);
+        }
+    });
+}
+
 async function loadConversation(conversationId) {
     conversationId = String(conversationId || '').trim();
     if (!conversationId) return;
@@ -6458,6 +6689,11 @@ async function loadConversation(conversationId) {
             if (seq !== loadConversationRequestSeq) {
                 return;
             }
+            hydrateConversationTokenUsage(conversationId, seq, conversationLoadController.signal).catch((e) => {
+                if (!e || e.name !== 'AbortError') {
+                    console.warn('hydrateConversationTokenUsage failed', e);
+                }
+            });
             if (currentConversationId === conversationId && typeof window.restoreHitlInlineForConversation === 'function') {
                 await window.restoreHitlInlineForConversation(conversationId);
             }

web/static/js/dashboard.js

@@ -66,10 +66,12 @@ async function refreshDashboard() {
         setDashboardOverviewPlaceholder('…');
         setEl('dashboard-kpi-tools-calls', '…');
         setEl('dashboard-kpi-success-rate', '…');
+        setEl('dashboard-kpi-token-usage', '…');
         setKpiSubText('dashboard-kpi-tasks-sub-text', '…');
         setKpiSubText('dashboard-kpi-vuln-sub-text', '…');
         setKpiSubText('dashboard-kpi-tools-sub-text', '…');
         setKpiSubText('dashboard-kpi-rate-sub-text', '…');
+        setKpiSubText('dashboard-kpi-token-sub-text', '…');
         hideEl('dashboard-kpi-vuln-critical-badge');
         hideEl('dashboard-alert-banner');
         setRecentVulnsLoading();
@@ -127,7 +129,7 @@ async function refreshDashboard() {
             hitlPendingRes, notificationsRes, externalMcpStatsRes,
             webshellRes,
             c2ListenersRes, c2SessionsRes, c2TasksRes,
-            projectSummaryRes, severityFilteredStatsRes
+            projectSummaryRes, severityFilteredStatsRes, tokenUsageRes
         ] = await Promise.all([
             fetchJson('/api/agent-loop/tasks'),
             fetchJson('/api/vulnerabilities/stats'),
@@ -159,7 +161,8 @@ async function refreshDashboard() {
             fetchJson(dashboardProjectScopedUrl('/api/c2/sessions?limit=500')),
             fetchJson(dashboardProjectScopedUrl('/api/c2/tasks?page=1&page_size=1')),
             fetchJson('/api/projects/dashboard-summary?fact_limit=10'),
-            selectedSeverityStatus ? fetchJson('/api/vulnerabilities/stats?status=' + encodeURIComponent(selectedSeverityStatus)) : Promise.resolve(null)
+            selectedSeverityStatus ? fetchJson('/api/vulnerabilities/stats?status=' + encodeURIComponent(selectedSeverityStatus)) : Promise.resolve(null),
+            fetchJson(dashboardProjectScopedUrl('/api/usage/tokens?days=7&limit=5'))
         ]);
 
         // 如果在 await 期间 controller 已被 abort，说明又有新刷新启动了，丢弃本次结果
@@ -330,6 +333,8 @@ async function refreshDashboard() {
             renderDashboardToolsBar(null);
         }
 
+        renderDashboardTokenUsage(tokenUsageRes);
+
         // 「能力总览 → MCP 工具」用配置总数（包含未被调用过的工具）；专项接口失败时回落到 monitor 的 names.length
         if (toolsConfigRes && typeof toolsConfigRes.total === 'number') {
             setEl('dashboard-resource-tools', formatNumber(toolsConfigRes.total));
@@ -435,10 +440,12 @@ async function refreshDashboard() {
         setDashboardOverviewPlaceholder('-');
         setEl('dashboard-kpi-success-rate', '-');
         setEl('dashboard-kpi-tools-calls', '-');
+        setEl('dashboard-kpi-token-usage', '-');
         setKpiSubText('dashboard-kpi-tasks-sub-text', '-');
         setKpiSubText('dashboard-kpi-vuln-sub-text', '-');
         setKpiSubText('dashboard-kpi-tools-sub-text', '-');
         setKpiSubText('dashboard-kpi-rate-sub-text', '-');
+        setKpiSubText('dashboard-kpi-token-sub-text', '-');
         ['tools', 'skills', 'knowledge', 'roles', 'agents'].forEach(function (k) {
             setEl('dashboard-resource-' + k, '-');
         });
@@ -700,6 +707,41 @@ function setKpiRateBadge(id, rate, failedCount) {
     }
 }
 
+function renderDashboardTokenUsage(res) {
+    const summary = res && res.summary ? res.summary : null;
+    if (!summary) {
+        setEl('dashboard-kpi-token-usage', '-');
+        setKpiSubText('dashboard-kpi-token-sub-text', '-');
+        return;
+    }
+    const total = Number(summary.totalTokens || 0);
+    const calls = Number(summary.modelCalls || 0);
+    const today = res && res.today ? Number(res.today.totalTokens || 0) : 0;
+    if (!Number.isFinite(total) || total <= 0) {
+        setEl('dashboard-kpi-token-usage', '0');
+        setKpiSubText('dashboard-kpi-token-sub-text', dt('dashboard.noTokenUsageYet', null, '暂无用量'));
+        return;
+    }
+    setEl('dashboard-kpi-token-usage', formatTokenUsageCompact(total));
+    setKpiSubText('dashboard-kpi-token-sub-text',
+        dt('dashboard.tokenUsageSub', {
+            today: formatTokenUsageCompact(today),
+            calls: Number.isFinite(calls) ? calls : 0
+        }, '近 7 天 ' + (Number.isFinite(calls) ? calls : 0) + ' 次调用 · 今日 ' + formatTokenUsageCompact(today)));
+}
+
+function formatTokenUsageCompact(num) {
+    const n = Number(num || 0);
+    if (!Number.isFinite(n) || n <= 0) return '0';
+    if (n >= 1000000) {
+        return (n / 1000000).toFixed(n >= 10000000 ? 0 : 1).replace(/\.0$/, '') + 'M';
+    }
+    if (n >= 1000) {
+        return (n / 1000).toFixed(n >= 10000 ? 0 : 1).replace(/\.0$/, '') + 'K';
+    }
+    return String(Math.trunc(n));
+}
+
 // sessionStorage：告警条「×」忽略记录 + 最近一次**实际展示过**的 reason 片段（不含 level），
 // 用于在「问题从多变少」（如审完 HITL 后只剩严重漏洞）时，避免误用更早对「仅子集」的忽略。
 var DASH_SESSION_ALERT_DISMISSED = 'dashboard.dismissedAlert';

web/static/js/monitor.js

@@ -1044,7 +1044,7 @@ function updateAssistantBubbleContent(assistantMessageId, content, renderMarkdow
     const bubble = assistantElement.querySelector('.message-bubble');
     if (!bubble) return;
 
-    // 保留复制按钮：addMessage 会把按钮 append 在 message-bubble 里
+    // 清理旧版本可能残留在气泡内的复制按钮；新版按钮统一在时间行。
     const copyBtn = bubble.querySelector('.message-copy-btn');
     if (copyBtn) copyBtn.remove();
 
@@ -1066,7 +1066,9 @@ function updateAssistantBubbleContent(assistantMessageId, content, renderMarkdow
     if (typeof wrapTablesInBubble === 'function') {
         wrapTablesInBubble(bubble);
     }
-    if (copyBtn) bubble.appendChild(copyBtn);
+    if (typeof window.appendMessageCopyButton === 'function') {
+        window.appendMessageCopyButton(assistantElement);
+    }
 
     if (typeof window.csMarkdownSanitize !== 'undefined') {
         window.csMarkdownSanitize.stripSuspiciousImages(bubble);
@@ -3372,6 +3374,18 @@ function handleStreamEvent(event, progressElement, progressId,
             });
             break;
 
+        case 'finalization_pending_tools_cancelled': {
+            const d = event.data || {};
+            markToolExecutionItemsCancelled(timeline, autoCancelledExecutionIdsFromData(d));
+            addTimelineItem(timeline, 'progress', {
+                title: '工具执行已收尾',
+                message: event.message,
+                data: d,
+                expanded: false
+            });
+            break;
+        }
+
         case 'hitl_audit_agent_started': {
             const auditData = Object.assign({}, event.data || {}, {
                 reviewer: 'audit_agent',
@@ -3984,6 +3998,7 @@ function handleStreamEvent(event, progressElement, progressId,
             const responseData = event.data || {};
             const mcpIds = mergeMcpExecutionIDLists(typeof getMcpIds === 'function' ? (getMcpIds() || []) : [], responseData.mcpExecutionIds || []);
             setMcpIds(mcpIds);
+            markToolExecutionItemsCancelled(timeline, autoCancelledExecutionIdsFromData(responseData));
 
             // 更新对话ID
             if (responseData.conversationId) {
@@ -5886,6 +5901,9 @@ function getToolResultDisplayState(data, opts) {
         }
         return { kind: 'background_running', isError: false, success: false };
     }
+    if (explicitStatus === 'cancelled' || explicitStatus === 'canceled') {
+        return { kind: 'cancelled', isError: true, success: false };
+    }
     const parts = [];
     if (opts.rawText != null) parts.push(String(opts.rawText));
     collectToolResultTextParts(data.result, parts, 0);
@@ -5917,6 +5935,13 @@ function getBackgroundRunningToolLabel() {
     return '后台执行中';
 }
 
+function toolDisplayStatusFromState(displayState) {
+    if (!displayState) return 'completed';
+    if (displayState.kind === 'background_running') return 'background_running';
+    if (displayState.kind === 'cancelled') return 'cancelled';
+    return displayState.isError ? 'failed' : 'completed';
+}
+
 function buildToolResultSectionHtml(data, opts) {
     opts = opts || {};
     const _t = function (k, o) {
@@ -6158,7 +6183,10 @@ function mergeToolResultIntoCallItem(item, data, options) {
         }
         item.dataset.toolResultMerged = '1';
         item.dataset.toolSuccess = (!displayState.isError && !backgroundRunning) ? '1' : '0';
-        item.dataset.toolDisplayStatus = backgroundRunning ? 'background_running' : (displayState.isError ? 'failed' : 'completed');
+        item.dataset.toolDisplayStatus = toolDisplayStatusFromState(displayState);
+        if (data.executionId != null && String(data.executionId).trim() !== '') {
+            item.dataset.toolExecutionId = String(data.executionId).trim();
+        }
         item.classList.remove('tool-call-running', 'tool-call-completed', 'tool-call-failed');
         item.classList.add(backgroundRunning ? 'tool-call-running' : (displayState.isError ? 'tool-call-failed' : 'tool-call-completed'));
         applyToolCallStatus(item, item.dataset.toolDisplayStatus);
@@ -6198,7 +6226,10 @@ function mergeToolResultIntoCallItem(item, data, options) {
 
     item.dataset.toolResultMerged = '1';
     item.dataset.toolSuccess = (!displayState.isError && !backgroundRunning) ? '1' : '0';
-    item.dataset.toolDisplayStatus = backgroundRunning ? 'background_running' : (displayState.isError ? 'failed' : 'completed');
+    item.dataset.toolDisplayStatus = toolDisplayStatusFromState(displayState);
+    if (data.executionId != null && String(data.executionId).trim() !== '') {
+        item.dataset.toolExecutionId = String(data.executionId).trim();
+    }
     item.classList.remove('tool-call-running', 'tool-call-completed', 'tool-call-failed');
     item.classList.add(backgroundRunning ? 'tool-call-running' : (displayState.isError ? 'tool-call-failed' : 'tool-call-completed'));
     applyToolCallStatus(item, item.dataset.toolDisplayStatus);
@@ -6361,6 +6392,9 @@ function getToolCallStatusPresentation(status) {
     if (normalized === 'failed') {
         return { status: normalized, itemClass: 'tool-call-failed', badgeClass: 'tool-status-failed', label: translate('timeline.execFailed', '执行失败'), icon: '❌ ' };
     }
+    if (normalized === 'cancelled' || normalized === 'canceled') {
+        return { status: 'cancelled', itemClass: 'tool-call-failed', badgeClass: 'tool-status-failed', label: translate('tasks.statusCancelled', '已取消'), icon: '⛔ ' };
+    }
     if (normalized === 'result_missing') {
         return { status: normalized, itemClass: 'tool-call-incomplete', badgeClass: 'tool-status-incomplete', label: translate('timeline.resultMissing', '结果记录缺失'), icon: '⚠️ ' };
     }
@@ -6401,6 +6435,52 @@ function updateToolCallStatus(progressId, toolCallId, status) {
     applyToolCallStatus(item, status);
 }
 
+function normalizeExecutionIdList(value) {
+    const input = Array.isArray(value) ? value : (value == null ? [] : [value]);
+    const seen = new Set();
+    const out = [];
+    input.forEach(function (v) {
+        const s = String(v == null ? '' : v).trim();
+        if (!s || seen.has(s)) return;
+        seen.add(s);
+        out.push(s);
+    });
+    return out;
+}
+
+function autoCancelledExecutionIdsFromData(data) {
+    data = data || {};
+    return normalizeExecutionIdList(data.autoCancelledPendingExecutionIds || data.autoCancelledExecutionIds || []);
+}
+
+function markToolExecutionItemsCancelled(root, executionIds) {
+    const ids = normalizeExecutionIdList(executionIds);
+    if (!root || ids.length === 0) return 0;
+    const idSet = new Set(ids);
+    let count = 0;
+    root.querySelectorAll('.timeline-item[data-tool-execution-id]').forEach(function (item) {
+        const execId = String(item.dataset.toolExecutionId || '').trim();
+        if (!execId || !idSet.has(execId)) return;
+        item.dataset.toolSuccess = '0';
+        item.dataset.toolDisplayStatus = 'cancelled';
+        item.classList.remove('tool-call-running', 'tool-call-completed', 'tool-call-incomplete');
+        item.classList.add('tool-call-failed');
+        const state = toolCallDetailStateByItemId.get(item.id);
+        if (state && state.resultData && typeof state.resultData === 'object') {
+            state.resultData = Object.assign({}, state.resultData, {
+                status: 'cancelled',
+                success: false,
+                isError: true
+            });
+            state.pending = false;
+            setToolCallDetailState(item, state);
+        }
+        applyToolCallStatus(item, 'cancelled');
+        count++;
+    });
+    return count;
+}
+
 // 添加时间线项目
 function buildWorkflowConditionResultHtml(data) {
     const output = (data && data.output) || {};
@@ -6547,17 +6627,23 @@ function addTimelineItem(timeline, type, options) {
         const mergedDisplayState = merged ? getToolResultDisplayState(merged) : null;
         const mergedBackgroundRunning = mergedDisplayState && mergedDisplayState.kind === 'background_running';
         const terminalStatus = String(options.toolStatus || '').toLowerCase();
+        const forcedStatus = (terminalStatus === 'completed' || terminalStatus === 'failed' || terminalStatus === 'cancelled' || terminalStatus === 'canceled')
+            ? (terminalStatus === 'canceled' ? 'cancelled' : terminalStatus)
+            : '';
         if (merged) {
             item.dataset.toolResultMerged = '1';
-            item.dataset.toolSuccess = (!mergedDisplayState.isError && !mergedBackgroundRunning) ? '1' : '0';
+            item.dataset.toolSuccess = forcedStatus ? (forcedStatus === 'completed' ? '1' : '0') : ((!mergedDisplayState.isError && !mergedBackgroundRunning) ? '1' : '0');
-            item.dataset.toolDisplayStatus = mergedBackgroundRunning ? 'background_running' : (mergedDisplayState.isError ? 'failed' : 'completed');
+            item.dataset.toolDisplayStatus = forcedStatus || toolDisplayStatusFromState(mergedDisplayState);
-            item.classList.add(mergedBackgroundRunning ? 'tool-call-running' : (mergedDisplayState.isError ? 'tool-call-failed' : 'tool-call-completed'));
+            if (merged.executionId != null && String(merged.executionId).trim() !== '') {
+                item.dataset.toolExecutionId = String(merged.executionId).trim();
+            }
+            item.classList.add(item.dataset.toolDisplayStatus === 'background_running' ? 'tool-call-running' : (item.dataset.toolDisplayStatus === 'completed' ? 'tool-call-completed' : 'tool-call-failed'));
             if (d._mergedResultDetailId) {
                 item.dataset.toolResultDetailId = String(d._mergedResultDetailId);
             }
-        } else if (terminalStatus === 'completed' || terminalStatus === 'failed') {
+        } else if (terminalStatus === 'completed' || terminalStatus === 'failed' || terminalStatus === 'cancelled' || terminalStatus === 'canceled') {
             item.dataset.toolSuccess = terminalStatus === 'completed' ? '1' : '0';
-            item.dataset.toolDisplayStatus = terminalStatus;
+            item.dataset.toolDisplayStatus = terminalStatus === 'canceled' ? 'cancelled' : terminalStatus;
             item.classList.add(terminalStatus === 'completed' ? 'tool-call-completed' : 'tool-call-failed');
         } else if (terminalStatus === 'result_missing') {
             item.dataset.toolDisplayStatus = 'result_missing';
@@ -6588,7 +6674,10 @@ function addTimelineItem(timeline, type, options) {
         }
         item.dataset.toolName = (d.toolName != null && d.toolName !== '') ? String(d.toolName) : '';
         item.dataset.toolSuccess = (!displayState.isError && displayState.kind !== 'background_running') ? '1' : '0';
-        item.dataset.toolDisplayStatus = displayState.kind === 'background_running' ? 'background_running' : (displayState.isError ? 'failed' : 'completed');
+        item.dataset.toolDisplayStatus = toolDisplayStatusFromState(displayState);
+        if (d.executionId != null && String(d.executionId).trim() !== '') {
+            item.dataset.toolExecutionId = String(d.executionId).trim();
+        }
     }
     if (type === 'eino_usage_summary' && options.data) {
         const d = options.data;
@@ -6658,10 +6747,14 @@ function addTimelineItem(timeline, type, options) {
         const mergedDisplayState = merged ? getToolResultDisplayState(merged) : null;
         const mergedBackgroundRunning = mergedDisplayState && mergedDisplayState.kind === 'background_running';
         const terminalStatus = String(options.toolStatus || '').toLowerCase();
-        const hasTerminalStatus = terminalStatus === 'completed' || terminalStatus === 'failed';
+        const forcedStatus = (terminalStatus === 'completed' || terminalStatus === 'failed' || terminalStatus === 'cancelled' || terminalStatus === 'canceled')
+            ? (terminalStatus === 'canceled' ? 'cancelled' : terminalStatus)
+            : '';
+        const hasTerminalStatus = terminalStatus === 'completed' || terminalStatus === 'failed' || terminalStatus === 'cancelled' || terminalStatus === 'canceled';
         const hasHistoricalStatus = hasTerminalStatus || terminalStatus === 'result_missing';
         if (merged) {
-            item.classList.add(mergedBackgroundRunning ? 'tool-call-running' : (mergedDisplayState.isError ? 'tool-call-failed' : 'tool-call-completed'));
+            const statusForClass = forcedStatus || toolDisplayStatusFromState(mergedDisplayState);
+            item.classList.add(statusForClass === 'background_running' ? 'tool-call-running' : (statusForClass === 'completed' ? 'tool-call-completed' : 'tool-call-failed'));
         } else if (hasTerminalStatus) {
             item.classList.add(terminalStatus === 'completed' ? 'tool-call-completed' : 'tool-call-failed');
         } else if (terminalStatus === 'result_missing') {
@@ -6671,7 +6764,7 @@ function addTimelineItem(timeline, type, options) {
         }
         setToolCallDetailState(item, {
             args: args,
-            resultData: merged || null,
+            resultData: (merged && forcedStatus) ? Object.assign({}, merged, { status: forcedStatus, success: forcedStatus === 'completed', isError: forcedStatus !== 'completed' }) : (merged || null),
             pending: !merged && !hasHistoricalStatus && !options.skipPendingResult,
             processDetailId: options.processDetailId || '',
             resultDetailId: data._mergedResultDetailId || (merged && merged.processDetailId) || '',
@@ -6723,7 +6816,10 @@ function addTimelineItem(timeline, type, options) {
             payloadDeferred: data._payloadDeferred === true,
             payloadLoaded: data._payloadDeferred !== true
         });
-        item.dataset.toolDisplayStatus = displayState.kind === 'background_running' ? 'background_running' : (displayState.isError ? 'failed' : 'completed');
+        item.dataset.toolDisplayStatus = toolDisplayStatusFromState(displayState);
+        if (data.executionId != null && String(data.executionId).trim() !== '') {
+            item.dataset.toolExecutionId = String(data.executionId).trim();
+        }
         item.classList.add(displayState.kind === 'background_running' ? 'tool-call-running' : (displayState.isError ? 'tool-call-failed' : 'tool-call-completed'));
     } else if (type === 'cancelled') {
         const taskCancelledLabel = typeof window.t === 'function' ? window.t('chat.taskCancelled') : '任务已取消';

web/static/js/project-folder-preview.test.cjs

@@ -18,6 +18,12 @@ function functionSource(source, name, nextName) {
     return source.slice(start, end);
 }
 
+function cssBlock(source, selector) {
+    const match = source.match(new RegExp(`${selector.replace(/[.*+?^${}()|[\]\\]/g, '\\$&')}\\s*\\{[^}]*\\}`));
+    assert.ok(match, `${selector} style block should exist`);
+    return match[0];
+}
+
 test('无项目文件夹与普通项目共用悬浮和键盘聚焦预览', () => {
     const source = functionSource(projects, 'appendChatProjectFolderItem', 'appendChatProjectConversationItem');
 
@@ -122,3 +128,27 @@ test('对话悬浮预览显示本地年月日时分', () => {
     assert.match(zh, /"conversationPreviewDateTime": "\{\{year\}\}年\{\{month\}\}月\{\{day\}\}日 \{\{hour\}\}:\{\{minute\}\}"/);
     assert.match(en, /"conversationPreviewDateTime": "\{\{year\}\}-\{\{month\}\}-\{\{day\}\} \{\{hour\}\}:\{\{minute\}\}"/);
 });
+
+test('对话悬浮预览标题与时间分行显示并保留更多标题内容', () => {
+    const titleStyles = cssBlock(styles, '.project-conversation-preview-title');
+
+    assert.match(styles, /\.conversation-sidebar\s*\{[\s\S]*?width: 320px;/);
+    assert.match(styles, /\.project-conversation-preview\s*\{[\s\S]*?width: min\(340px, calc\(100vw - 32px\)\);/);
+    assert.match(styles, /\.project-conversation-preview-header\s*\{[\s\S]*?flex-direction: column;[\s\S]*?align-items: flex-start;/);
+    assert.match(titleStyles, /-webkit-line-clamp: 2;/);
+    assert.doesNotMatch(titleStyles, /white-space: nowrap;/);
+});
+
+test('对话悬浮预览使用美化后的代理模式徽标', () => {
+    const projects = fs.readFileSync('web/static/js/projects.js', 'utf8');
+
+    assert.match(projects, /function getProjectConversationModeIconClass\(conversation\)/);
+    assert.match(projects, /project-conversation-preview-mode-icon agent-mode-logo agent-mode-logo--default/);
+    assert.match(projects, /agent-mode-logo__svg/);
+    assert.match(projects, /<rect x="3" y="11" width="18" height="10" rx="2"/);
+    assert.match(projects, /agent-mode-logo--' \+ getProjectConversationModeIconClass\(conversation\)/);
+    assert.match(styles, /\.agent-mode-logo\s*\{[\s\S]*?background: transparent;/);
+    assert.match(styles, /\.agent-mode-logo__svg\s*\{[\s\S]*?stroke: currentColor;[\s\S]*?stroke-width: 1\.9;/);
+    assert.match(styles, /\.project-conversation-preview-mode-icon\.agent-mode-logo\s*\{[\s\S]*?width: 16px;/);
+    assert.doesNotMatch(cssBlock(styles, '.agent-mode-logo'), /linear-gradient|box-shadow: 0 5px/);
+});

web/static/js/projects.js

@@ -3084,6 +3084,15 @@ function getProjectConversationModeLabel(conversation) {
     );
 }
 
+function getProjectConversationModeIconClass(conversation) {
+    const mode = String(conversation?.agentMode || conversation?.agent_mode || '').trim().toLowerCase();
+    if (mode === 'eino_single') return 'eino';
+    if (mode === 'deep') return 'deep';
+    if (mode === 'plan_execute') return 'plan';
+    if (mode === 'supervisor') return 'supervisor';
+    return 'default';
+}
+
 function ensureProjectConversationPreview() {
     let preview = document.getElementById('project-conversation-preview');
     if (preview) return preview;
@@ -3102,7 +3111,7 @@ function ensureProjectConversationPreview() {
             <span class="project-conversation-preview-project"></span>
         </div>
         <div class="project-conversation-preview-meta">
-            <svg width="18" height="18" viewBox="0 0 24 24" fill="none" aria-hidden="true"><circle cx="6" cy="5" r="2" stroke="currentColor" stroke-width="1.7"/><circle cx="6" cy="19" r="2" stroke="currentColor" stroke-width="1.7"/><circle cx="18" cy="9" r="2" stroke="currentColor" stroke-width="1.7"/><path d="M6 7v10M8 15c5 0 3-6 8-6" stroke="currentColor" stroke-width="1.7" stroke-linecap="round"/></svg>
+            <span class="project-conversation-preview-mode-icon agent-mode-logo agent-mode-logo--default" aria-hidden="true"><svg class="agent-mode-logo__svg" viewBox="0 0 24 24" fill="none" aria-hidden="true"><rect x="3" y="11" width="18" height="10" rx="2"/><circle cx="12" cy="5" r="2"/><path d="M12 7v4"/><path d="M8 16h.01"/><path d="M16 16h.01"/></svg></span>
             <span class="project-conversation-preview-mode"></span>
             <span class="project-conversation-preview-separator" aria-hidden="true">·</span>
             <span class="project-conversation-preview-status"></span>
@@ -3165,6 +3174,10 @@ function showProjectConversationPreview(conversation, project, row) {
     ageEl.hidden = !ageEl.textContent;
     preview.querySelector('.project-conversation-preview-project').textContent = project?.name
         || pickerMessage(tp, 'chat.conversationPreviewNoProject', '未绑定项目');
+    const modeIcon = preview.querySelector('.project-conversation-preview-mode-icon');
+    if (modeIcon) {
+        modeIcon.className = 'project-conversation-preview-mode-icon agent-mode-logo agent-mode-logo--' + getProjectConversationModeIconClass(conversation);
+    }
     preview.querySelector('.project-conversation-preview-mode').textContent = getProjectConversationModeLabel(conversation);
     statusEl.textContent = status;
     statusEl.className = 'project-conversation-preview-status'

web/static/js/turn-duration-refresh.test.cjs

@@ -66,7 +66,7 @@ function createHarness(nowMs) {
         clearInterval() {},
     };
     vm.runInNewContext(
-        `${timingSource(chat)}; this.setAssistantTurnTiming = setAssistantTurnTiming;`,
+        `${timingSource(chat)}; this.setAssistantTurnTiming = setAssistantTurnTiming; this.setAssistantTurnTokenUsage = setAssistantTurnTokenUsage; this.extractAssistantTurnTokenUsage = extractAssistantTurnTokenUsage;`,
         context
     );
     return context;
@@ -103,6 +103,45 @@ test('已完成任务仍优先使用持久化耗时', () => {
     assert.match(message.label.innerHTML, /耗时 1 分钟 5 秒/);
 });
 
+test('助手轮次摘要会显示持久化 token 用量', () => {
+    const context = createHarness(Date.parse('2026-08-12T02:05:00.000Z'));
+    const message = createMessage();
+
+    context.setAssistantTurnTiming(message, {
+        startedAt: '2026-08-12T02:00:00.000Z',
+        completedAt: '2026-08-12T02:00:03.000Z',
+        durationMs: 3000,
+        status: 'completed',
+    });
+    context.setAssistantTurnTokenUsage(message, {
+        promptTokens: 1200,
+        completionTokens: 34,
+        cachedTokens: 200,
+        reasoningTokens: 12,
+        modelCalls: 1,
+        model: 'deepseek-v3',
+    });
+
+    assert.equal(message.dataset.turnTotalTokens, '1234');
+    assert.match(message.label.innerHTML, /1\.2K tokens/);
+    assert.match(message.label.innerHTML, /turn-process-token-chip/);
+});
+
+test('助手轮次可从 Eino usage summary 过程详情提取 token 用量', () => {
+    const context = createHarness(Date.parse('2026-08-12T02:05:00.000Z'));
+
+    const usage = context.extractAssistantTurnTokenUsage([
+        { eventType: 'progress', data: { totalTokens: 9999 } },
+        { eventType: 'eino_usage_summary', data: { promptTokens: 400, completionTokens: 100, modelCalls: 1 } },
+        { eventType: 'eino_usage_summary', data: { totalTokens: 25, modelCalls: 1 } },
+    ]);
+
+    assert.equal(usage.totalTokens, 525);
+    assert.equal(usage.promptTokens, 400);
+    assert.equal(usage.completionTokens, 100);
+    assert.equal(usage.modelCalls, 2);
+});
+
 test('已中断任务使用固定终态耗时且不再按当前时间增长', () => {
     const context = createHarness(Date.parse('2026-08-13T12:00:00.000Z'));
     const message = createMessage();

web/templates/index.html

@@ -529,6 +529,16 @@
                                         <span class="dashboard-kpi-sub-text" id="dashboard-kpi-rate-sub-text" data-i18n="dashboard.healthyStatus">运行平稳</span>
                                     </div>
                                 </div>
+                                <div class="dashboard-kpi-card" role="button" tabindex="0" onclick="switchPage('chat')" onkeydown="if(event.key==='Enter'||event.key===' ') { event.preventDefault(); switchPage('chat'); }" data-i18n="dashboard.clickToViewChat" data-i18n-attr="title" title="点击查看对话">
+                                    <div class="dashboard-kpi-head">
+                                        <div class="dashboard-kpi-label" data-i18n="dashboard.tokenUsage">Token 用量</div>
+                                        <span class="dashboard-kpi-icon dashboard-kpi-icon-tokens" aria-hidden="true"><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M4 7h16"/><path d="M4 12h16"/><path d="M4 17h16"/><path d="M8 3 6 21"/><path d="m18 3-2 18"/></svg></span>
+                                    </div>
+                                    <div class="dashboard-kpi-value" id="dashboard-kpi-token-usage">-</div>
+                                    <div class="dashboard-kpi-sub">
+                                        <span class="dashboard-kpi-sub-text" id="dashboard-kpi-token-sub-text">-</span>
+                                    </div>
+                                </div>
                             </div>
                             <!-- 两列主内容区 -->
                             <div class="dashboard-grid">
@@ -1213,7 +1223,7 @@
                                 <div id="agent-mode-wrapper" class="agent-mode-wrapper" style="display: none;">
                                     <div class="agent-mode-inner">
                                         <button type="button" id="agent-mode-btn" class="role-selector-btn agent-mode-btn" onclick="toggleAgentModePanel()" data-i18n="chat.agentModeSelectAria" data-i18n-attr="aria-label,title" data-i18n-skip-text="true" aria-label="选择对话执行模式" aria-haspopup="listbox" aria-expanded="false" title="选择对话执行模式">
-                                            <span id="agent-mode-icon" class="role-selector-icon" aria-hidden="true">🤖</span>
+                                            <span id="agent-mode-icon" class="role-selector-icon agent-mode-logo agent-mode-logo--default" aria-hidden="true"><svg class="agent-mode-logo__svg" viewBox="0 0 24 24" fill="none" aria-hidden="true"><rect x="3" y="11" width="18" height="10" rx="2"/><circle cx="12" cy="5" r="2"/><path d="M12 7v4"/><path d="M8 16h.01"/><path d="M16 16h.01"/></svg></span>
                                             <span id="agent-mode-text" class="role-selector-text">单代理</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"/>
@@ -1230,7 +1240,7 @@
                                             </div>
                                             <div class="agent-mode-options">
                                                 <button type="button" class="role-selection-item-main agent-mode-option" data-value="eino_single" role="option" onclick="selectAgentMode('eino_single')" data-agent-mode-detail="CloudWeGo Eino ChatModelAgent + Runner，MCP 工具（/api/eino-agent）" data-i18n="chat.agentModeEinoSingleHint" data-i18n-attr="data-agent-mode-detail" data-i18n-skip-text="true">
-                                                    <div class="role-selection-item-icon-main" aria-hidden="true">⚡</div>
+                                                    <div class="role-selection-item-icon-main agent-mode-logo agent-mode-logo--eino" aria-hidden="true"><svg class="agent-mode-logo__svg" viewBox="0 0 24 24" fill="none" aria-hidden="true"><rect x="3" y="11" width="18" height="10" rx="2"/><circle cx="12" cy="5" r="2"/><path d="M12 7v4"/><path d="M8 16h.01"/><path d="M16 16h.01"/></svg></div>
                                                     <div class="role-selection-item-content-main">
                                                         <div class="role-selection-item-name-main" data-i18n="chat.agentModeEinoSingle">Eino 单代理（ADK）</div>
                                                         <div class="role-selection-item-description-main" data-i18n="chat.agentModeEinoSingleHint">CloudWeGo Eino ChatModelAgent + Runner，MCP 工具（/api/eino-agent）</div>
@@ -1238,7 +1248,7 @@
                                                     <div class="role-selection-checkmark-main agent-mode-check" data-agent-mode-check="eino_single">✓</div>
                                                 </button>
                                                 <button type="button" class="role-selection-item-main agent-mode-option" data-value="deep" role="option" onclick="selectAgentMode('deep')" data-agent-mode-detail="Eino DeepAgent，适合复杂安全测试、多阶段 task 子代理委派与汇总" data-i18n="chat.agentModeDeepHint" data-i18n-attr="data-agent-mode-detail" data-i18n-skip-text="true">
-                                                    <div class="role-selection-item-icon-main" aria-hidden="true">🧩</div>
+                                                    <div class="role-selection-item-icon-main agent-mode-logo agent-mode-logo--deep" aria-hidden="true"><svg class="agent-mode-logo__svg" viewBox="0 0 24 24" fill="none" aria-hidden="true"><rect x="3" y="11" width="18" height="10" rx="2"/><circle cx="12" cy="5" r="2"/><path d="M12 7v4"/><path d="M8 16h.01"/><path d="M16 16h.01"/></svg></div>
                                                     <div class="role-selection-item-content-main">
                                                         <div class="role-selection-item-name-main" data-i18n="chat.agentModeDeep">Deep（DeepAgent）</div>
                                                         <div class="role-selection-item-description-main" data-i18n="chat.agentModeDeepHint">Eino DeepAgent，适合复杂安全测试、多阶段 task 子代理委派与汇总</div>
@@ -1246,7 +1256,7 @@
                                                     <div class="role-selection-checkmark-main agent-mode-check" data-agent-mode-check="deep">✓</div>
                                                 </button>
                                                 <button type="button" class="role-selection-item-main agent-mode-option" data-value="plan_execute" role="option" onclick="selectAgentMode('plan_execute')" data-agent-mode-detail="规划 → 执行 → 重规划（单执行器工具链）" data-i18n="chat.agentModePlanExecuteHint" data-i18n-attr="data-agent-mode-detail" data-i18n-skip-text="true">
-                                                    <div class="role-selection-item-icon-main" aria-hidden="true">📋</div>
+                                                    <div class="role-selection-item-icon-main agent-mode-logo agent-mode-logo--plan" aria-hidden="true"><svg class="agent-mode-logo__svg" viewBox="0 0 24 24" fill="none" aria-hidden="true"><rect x="3" y="11" width="18" height="10" rx="2"/><circle cx="12" cy="5" r="2"/><path d="M12 7v4"/><path d="M8 16h.01"/><path d="M16 16h.01"/></svg></div>
                                                     <div class="role-selection-item-content-main">
                                                         <div class="role-selection-item-name-main" data-i18n="chat.agentModePlanExecuteLabel">Plan-Execute</div>
                                                         <div class="role-selection-item-description-main" data-i18n="chat.agentModePlanExecuteHint">规划 → 执行 → 重规划（单执行器工具链）</div>
@@ -1254,7 +1264,7 @@
                                                     <div class="role-selection-checkmark-main agent-mode-check" data-agent-mode-check="plan_execute">✓</div>
                                                 </button>
                                                 <button type="button" class="role-selection-item-main agent-mode-option" data-value="supervisor" role="option" onclick="selectAgentMode('supervisor')" data-agent-mode-detail="专家路由场景：监督者通过 transfer 动态分派多个专业子代理" data-i18n="chat.agentModeSupervisorHint" data-i18n-attr="data-agent-mode-detail" data-i18n-skip-text="true">
-                                                    <div class="role-selection-item-icon-main" aria-hidden="true">🎯</div>
+                                                    <div class="role-selection-item-icon-main agent-mode-logo agent-mode-logo--supervisor" aria-hidden="true"><svg class="agent-mode-logo__svg" viewBox="0 0 24 24" fill="none" aria-hidden="true"><rect x="3" y="11" width="18" height="10" rx="2"/><circle cx="12" cy="5" r="2"/><path d="M12 7v4"/><path d="M8 16h.01"/><path d="M16 16h.01"/></svg></div>
                                                     <div class="role-selection-item-content-main">
                                                         <div class="role-selection-item-name-main" data-i18n="chat.agentModeSupervisorLabel">Supervisor（专家路由）</div>
                                                         <div class="role-selection-item-description-main" data-i18n="chat.agentModeSupervisorHint">专家路由场景：监督者通过 transfer 动态分派多个专业子代理</div>