diff --git a/agent/config_visibility_test.go b/agent/config_visibility_test.go
index a9b41485..c0adc875 100644
--- a/agent/config_visibility_test.go
+++ b/agent/config_visibility_test.go
@@ -246,6 +246,55 @@ func TestToolManageStrategyRejectsFixedMinPositionSizeUpdates(t *testing.T) {
 	}
 }
 
+func TestToolManageStrategyReportsChangedAndRejectedFields(t *testing.T) {
+	dbPath := filepath.Join(t.TempDir(), "strategy-change-summary.db")
+	st, err := store.New(dbPath)
+	if err != nil {
+		t.Fatalf("create store: %v", err)
+	}
+	a := New(nil, st, DefaultConfig(), slog.Default())
+
+	resp := a.toolManageStrategy("default", `{"action":"create","name":"高频-短线ETH","config":{"coin_source":{"source_type":"static","static_coins":["ETHUSDT"]},"indicators":{"klines":{"primary_timeframe":"1m","selected_timeframes":["1m","3m"]}},"order_execution_speed":"fast"}}`)
+	if strings.Contains(resp, `"error"`) {
+		t.Fatalf("expected create to succeed with rejected unknown fields, got: %s", resp)
+	}
+	for _, want := range []string{
+		`"created_strategy_id"`,
+		`"changed_fields"`,
+		`coin_source.source_type`,
+		`indicators.klines.primary_timeframe`,
+		`"rejected_fields"`,
+		`order_execution_speed (not in current strategy config)`,
+		`"unchanged_defaults"`,
+	} {
+		if !strings.Contains(resp, want) {
+			t.Fatalf("expected response to contain %q, got: %s", want, resp)
+		}
+	}
+
+	strategies, err := st.Strategy().List("default")
+	if err != nil {
+		t.Fatalf("list strategies: %v", err)
+	}
+	var created *store.Strategy
+	for _, strategy := range strategies {
+		if strategy.Name == "高频-短线ETH" {
+			created = strategy
+			break
+		}
+	}
+	if created == nil {
+		t.Fatalf("expected strategy to be created")
+	}
+	var cfg map[string]any
+	if err := json.Unmarshal([]byte(created.Config), &cfg); err != nil {
+		t.Fatalf("parse config: %v", err)
+	}
+	if _, ok := cfg["order_execution_speed"]; ok {
+		t.Fatalf("unknown field should not be persisted: %s", created.Config)
+	}
+}
+
 func TestExchangeSkillOptionSummaryMatchesManualPage(t *testing.T) {
 	dbPath := filepath.Join(t.TempDir(), "exchange-options.db")
 	st, err := store.New(dbPath)
diff --git a/agent/llm_flow_extractor.go b/agent/llm_flow_extractor.go
index 15ce752d..79a66987 100644
--- a/agent/llm_flow_extractor.go
+++ b/agent/llm_flow_extractor.go
@@ -252,7 +252,7 @@ func allowedFieldSpecsForSkillSession(session skillSession, lang string) []llmFl
 		add(&out, "show_in_competition", displayCatalogFieldName("show_in_competition", lang), false)
 	case "strategy_management":
 		if session.Action == "create" || session.Action == "update_config" {
-			add(&out, "config_patch", "Partial StrategyConfig JSON patch inferred from the user's strategy intent. Use this for strategy requirements such as target coins, trend style, short/long bias, indicators, risk, timeframes, and prompt sections.", false)
+			add(&out, "config_patch", strategyConfigPatchFieldDescription(lang), false)
 		}
 		if session.Action == "update_prompt" {
 			add(&out, "prompt", "Full strategy prompt text to write into the strategy custom prompt.", false)
@@ -270,6 +270,10 @@ func allowedFieldSpecsForSkillSession(session skillSession, lang string) []llmFl
 	return out
 }
 
+func strategyConfigPatchFieldDescription(lang string) string {
+	return "Partial StrategyConfig JSON patch inferred from the user's strategy intent. Use this for strategy requirements such as target coins, trend style, short/long bias, indicators, risk, timeframes, and prompt sections."
+}
+
 func currentFieldValuesForSkillSession(session skillSession) map[string]string {
 	values := map[string]string{}
 	for key, value := range session.Fields {
diff --git a/agent/llm_skill_router.go b/agent/llm_skill_router.go
index 829ae300..5dfc27f5 100644
--- a/agent/llm_skill_router.go
+++ b/agent/llm_skill_router.go
@@ -20,6 +20,7 @@ type unifiedTurnDecision struct {
 	TopicIntent      string         `json:"topic_intent,omitempty"`
 	BusinessAction   string         `json:"business_action,omitempty"`
 	TargetSkill      string         `json:"target_skill,omitempty"`
+	Tasks            []WorkflowTask `json:"tasks,omitempty"`
 	TargetSnapshotID string         `json:"target_snapshot_id,omitempty"`
 	ContextMode      string         `json:"context_mode,omitempty"`
 	ExtractedData    map[string]any `json:"extracted_data,omitempty"`
@@ -117,6 +118,7 @@ func normalizeUnifiedTurnDecision(decision unifiedTurnDecision) unifiedTurnDecis
 	decision.TargetSnapshotID = strings.TrimSpace(decision.TargetSnapshotID)
 	decision.ContextMode = strings.TrimSpace(strings.ToLower(decision.ContextMode))
 	decision.ReplyToUser = strings.TrimSpace(decision.ReplyToUser)
+	decision.Tasks = normalizeWorkflowDecomposition(workflowDecomposition{Tasks: decision.Tasks}).Tasks
 	if decision.ExtractedData == nil {
 		decision.ExtractedData = map[string]any{}
 	}
@@ -134,7 +136,7 @@ func normalizeUnifiedTurnDecision(decision unifiedTurnDecision) unifiedTurnDecis
 		decision.TopicIntent = ""
 	}
 	switch decision.BusinessAction {
-	case "direct_answer", "new_skill", "continue_skill", "planned_agent", "none":
+	case "direct_answer", "new_skill", "skill_tasks", "continue_skill", "planned_agent", "none":
 	default:
 		decision.BusinessAction = ""
 	}
@@ -157,8 +159,13 @@ func (d unifiedTurnDecision) reliable() bool {
 	case "direct_answer":
 		return strings.TrimSpace(d.ReplyToUser) != ""
 	case "new_skill":
+		if len(d.Tasks) > 0 {
+			return true
+		}
 		skill, _ := parseTargetSkill(d.TargetSkill)
 		return skill != ""
+	case "skill_tasks":
+		return len(d.Tasks) > 0
 	case "continue_skill":
 		return d.TopicIntent == "continue_active"
 	case "planned_agent", "none":
@@ -234,12 +241,20 @@ topic_intent values:
 
 business_action values:
 - "direct_answer": reply_to_user is the final answer; do not change state
-- "new_skill": start a management/diagnosis skill; target_skill is required
+- "skill_tasks": start one or more management/diagnosis skill tasks; tasks is required
+- "new_skill": legacy single-skill route; target_skill is required if tasks is empty
 - "continue_skill": continue the active skill session
 - "planned_agent": hand off to the execution planner/tools
 - "none": only valid with cancel when no more action is needed
 
-target_skill format for new_skill:
+tasks format for skill_tasks:
+- id: "task_1", "task_2", ...
+- skill: one available skill name
+- action: one available action
+- request: the self-contained user-readable subtask
+- depends_on: array of task ids, empty when independent
+
+target_skill format for legacy new_skill:
 skill_name:action, for example "trader_management:create".
 Available skills:
 trader_management, exchange_management, model_management, strategy_management,
@@ -262,7 +277,9 @@ Rules:
 - If the user answers the previous assistant question, choose continue_active.
 - If the user only says "你好", "hi", "谢谢", "收到", choose instant_reply + direct_answer unless it clearly answers a pending task.
 - If the user asks a read-only management query, prefer planned_agent unless the answer is already fully available in the provided context.
-- Use new_skill for clear management tasks such as creating/updating/deleting/configuring trader/model/exchange/strategy.
+- Use skill_tasks for clear management tasks such as creating/updating/deleting/configuring trader/model/exchange/strategy.
+- If the user request contains multiple management operations, include multiple tasks and depends_on where a later task needs an earlier result.
+- If the request contains exactly one management operation, include exactly one task.
 - Use planned_agent for multi-step, tool-heavy, market/account, diagnosis, or ambiguous tasks.
 - For model_management, "provider" means AI vendor, never an exchange.
 - Current references are context only. Do not copy them into extracted_data unless the user explicitly says this/current/that previous one.
@@ -271,7 +288,7 @@ Rules:
 - confidence should reflect how safe it is to execute this decision without the old router fallback.
 
 Return JSON with this exact shape:
-{"topic_intent":"continue_active|start_new|resume_snapshot|cancel|instant_reply","business_action":"direct_answer|new_skill|continue_skill|planned_agent|none","target_skill":"","target_snapshot_id":"","context_mode":"use_current|fresh_context|resume_snapshot","extracted_data":{},"reply_to_user":"","confidence":0.0}`)
+{"topic_intent":"continue_active|start_new|resume_snapshot|cancel|instant_reply","business_action":"direct_answer|skill_tasks|new_skill|continue_skill|planned_agent|none","target_skill":"","tasks":[{"id":"task_1","skill":"","action":"","request":"","depends_on":[]}],"target_snapshot_id":"","context_mode":"use_current|fresh_context|resume_snapshot","extracted_data":{},"reply_to_user":"","confidence":0.0}`)
 
 	userPrompt := fmt.Sprintf("Language: %s\nUser message: %s\n\nPrevious assistant reply:\n%s\n\nCurrent reference summary:\n%s\n\nActive flow summary:\n%s\n\nSuspended snapshots JSON:\n%s\n\nRecent conversation:\n%s\n\nManagement domain primer:\n%s\n\nActive task details:\n%s\n",
 		lang,
@@ -334,6 +351,9 @@ func (a *Agent) executeUnifiedTurnDecision(ctx context.Context, storeUserID stri
 		a.runPostResponseMaintenanceAsync(userID)
 		return decision.ReplyToUser, true, nil
 	case "new_skill":
+		if len(decision.Tasks) > 0 {
+			return a.executeUnifiedSkillTasks(ctx, storeUserID, userID, lang, text, decision, onEvent)
+		}
 		skill, action := parseTargetSkill(decision.TargetSkill)
 		if skill == "" {
 			return "", false, nil
@@ -351,6 +371,8 @@ func (a *Agent) executeUnifiedTurnDecision(ctx context.Context, storeUserID stri
 		decision.ExtractedData = filterExtractedDataForActiveSession(session, decision.ExtractedData, lang)
 		mergeExtractedData(&session, decision.ExtractedData)
 		return a.driveActiveSession(ctx, storeUserID, userID, lang, text, session, onEvent)
+	case "skill_tasks":
+		return a.executeUnifiedSkillTasks(ctx, storeUserID, userID, lang, text, decision, onEvent)
 	case "continue_skill":
 		activeSession, hasActive := a.getActiveSkillSession(userID)
 		if !hasActive {
@@ -373,6 +395,39 @@ func (a *Agent) executeUnifiedTurnDecision(ctx context.Context, storeUserID stri
 	}
 }
 
+func (a *Agent) executeUnifiedSkillTasks(ctx context.Context, storeUserID string, userID int64, lang, text string, decision unifiedTurnDecision, onEvent func(event, data string)) (string, bool, error) {
+	tasks := normalizeWorkflowDecomposition(workflowDecomposition{Tasks: decision.Tasks}).Tasks
+	if len(tasks) == 0 {
+		return "", false, nil
+	}
+	if a.hasAnyActiveContext(userID) && decision.ContextMode == "fresh_context" {
+		if !a.suspendActiveContexts(userID, lang) {
+			a.clearSkillSession(userID)
+			a.clearWorkflowSession(userID)
+			a.clearExecutionState(userID)
+		}
+		a.clearActiveSkillSession(userID)
+	}
+	if len(tasks) == 1 {
+		task := tasks[0]
+		session := newActiveSkillSession(userID, task.Skill, task.Action)
+		session.Goal = defaultIfEmpty(strings.TrimSpace(task.Request), strings.TrimSpace(text))
+		decision.ExtractedData = filterExtractedDataForActiveSession(session, decision.ExtractedData, lang)
+		mergeExtractedData(&session, decision.ExtractedData)
+		return a.driveActiveSession(ctx, storeUserID, userID, lang, defaultIfEmpty(task.Request, text), session, onEvent)
+	}
+	session := normalizeWorkflowSession(WorkflowSession{
+		UserID:          userID,
+		OriginalRequest: strings.TrimSpace(text),
+		Tasks:           tasks,
+	})
+	if len(session.Tasks) == 0 {
+		return "", false, nil
+	}
+	a.saveWorkflowSession(userID, session)
+	return a.maybeAdvanceWorkflow(ctx, storeUserID, userID, lang, session, onEvent)
+}
+
 func parseLLMSkillRouteDecision(raw string) (llmSkillRouteDecision, error) {
 	raw = strings.TrimSpace(raw)
 	raw = strings.TrimPrefix(raw, "```json")
diff --git a/agent/planner_runtime.go b/agent/planner_runtime.go
index 198d6ec4..3a8401da 100644
--- a/agent/planner_runtime.go
+++ b/agent/planner_runtime.go
@@ -2823,7 +2823,7 @@ type nextStepDecision struct {
 }
 
 func (a *Agent) decideNextStep(ctx context.Context, userID int64, lang string, state ExecutionState) (nextStepDecision, error) {
-	toolDefs, _ := json.Marshal(agentTools())
+	toolDefs, _ := json.Marshal(plannerToolsForText(state.Goal))
 	obsJSON, _ := json.Marshal(buildObservationContext(state))
 	recentlyFetchedJSON, _ := json.Marshal(buildRecentlyFetchedData(state, time.Now().UTC()))
 	currentTurnCtx := a.buildCurrentTurnContext(userID, lang, state.Goal)
@@ -3010,7 +3010,7 @@ func (a *Agent) buildRecentConversationContext(userID int64, currentUserText str
 }
 
 func (a *Agent) createExecutionPlan(ctx context.Context, userID int64, lang, userText string, state ExecutionState) (executionPlan, error) {
-	toolDefs, _ := json.Marshal(agentTools())
+	toolDefs, _ := json.Marshal(plannerToolsForText(userText))
 	currentTurnCtx := a.buildCurrentTurnContext(userID, lang, userText)
 	activeTaskCtx := a.buildActiveTaskStateContext(userID, lang)
 	currentReferenceSummary := buildCurrentReferenceSummary(lang, a.semanticCurrentReferences(userID))
diff --git a/agent/planner_tools_test.go b/agent/planner_tools_test.go
new file mode 100644
index 00000000..955efee2
--- /dev/null
+++ b/agent/planner_tools_test.go
@@ -0,0 +1,84 @@
+package agent
+
+import (
+	"encoding/json"
+	"testing"
+
+	"nofx/mcp"
+)
+
+func TestPlannerToolsForMarketIntentAreTrimmed(t *testing.T) {
+	tools := plannerToolsForText("看一下 BTCUSDT 行情和 K线")
+	names := toolNamesForTest(tools)
+
+	for _, expected := range []string{"get_market_snapshot", "get_market_price", "get_kline"} {
+		if !containsString(names, expected) {
+			t.Fatalf("expected market tool %q in %v", expected, names)
+		}
+	}
+	for _, unexpected := range []string{"manage_strategy", "manage_trader", "manage_exchange_config", "manage_model_config"} {
+		if containsString(names, unexpected) {
+			t.Fatalf("did not expect management tool %q in market tools %v", unexpected, names)
+		}
+	}
+}
+
+func TestPlannerToolsForExchangeIntentAreTrimmed(t *testing.T) {
+	tools := plannerToolsForText("帮我添加 okx 交易所 API key")
+	names := toolNamesForTest(tools)
+
+	if len(names) != 2 {
+		t.Fatalf("expected two exchange tools, got %v", names)
+	}
+	for _, expected := range []string{"get_exchange_configs", "manage_exchange_config"} {
+		if !containsString(names, expected) {
+			t.Fatalf("expected exchange tool %q in %v", expected, names)
+		}
+	}
+}
+
+func TestPlannerToolsUseCompactManageStrategyForReadIntent(t *testing.T) {
+	tools := plannerToolsForText("列出我的策略")
+	tool := findToolForTest(tools, "manage_strategy")
+	if tool == nil {
+		t.Fatalf("expected manage_strategy in strategy tools")
+	}
+
+	raw, _ := json.Marshal(tool.Function.Parameters)
+	if len(raw) > 900 {
+		t.Fatalf("expected compact strategy schema, got %d bytes", len(raw))
+	}
+	if string(raw) == "" || !json.Valid(raw) {
+		t.Fatalf("expected valid strategy schema JSON")
+	}
+}
+
+func TestPlannerToolsKeepFullManageStrategyForMutationIntent(t *testing.T) {
+	tools := plannerToolsForText("创建一个 BTC 网格策略")
+	tool := findToolForTest(tools, "manage_strategy")
+	if tool == nil {
+		t.Fatalf("expected manage_strategy in strategy tools")
+	}
+
+	raw, _ := json.Marshal(tool.Function.Parameters)
+	if len(raw) < 1500 {
+		t.Fatalf("expected full strategy schema for mutation intent, got %d bytes", len(raw))
+	}
+}
+
+func toolNamesForTest(tools []mcp.Tool) []string {
+	names := make([]string, 0, len(tools))
+	for _, tool := range tools {
+		names = append(names, tool.Function.Name)
+	}
+	return names
+}
+
+func findToolForTest(tools []mcp.Tool, name string) *mcp.Tool {
+	for i := range tools {
+		if tools[i].Function.Name == name {
+			return &tools[i]
+		}
+	}
+	return nil
+}
diff --git a/agent/skill_execution_handlers.go b/agent/skill_execution_handlers.go
index 3659dfba..33779fae 100644
--- a/agent/skill_execution_handlers.go
+++ b/agent/skill_execution_handlers.go
@@ -2836,12 +2836,52 @@ func (a *Agent) persistStrategyConfigUpdate(storeUserID string, userID int64, la
 			enMsg += "\n\nAdjusted to stay within safe limits:\n- " + strings.Join(warnings, "\n- ")
 		}
 	}
+	if summary := parseStrategyToolChangeSummary(resp); len(summary.ChangedFields) > 0 || len(summary.RejectedFields) > 0 || len(summary.UnchangedDefaults) > 0 {
+		if lang == "zh" {
+			if len(summary.ChangedFields) > 0 {
+				zhMsg += "\n- 实际写入配置：" + strings.Join(summary.ChangedFields, "、")
+			}
+			if len(summary.RejectedFields) > 0 {
+				zhMsg += "\n- 未写入字段：" + strings.Join(summary.RejectedFields, "、")
+			}
+			if len(summary.UnchangedDefaults) > 0 {
+				zhMsg += "\n- 仍使用默认值：" + strings.Join(summary.UnchangedDefaults, "、")
+			}
+		} else {
+			if len(summary.ChangedFields) > 0 {
+				enMsg += "\n- Config fields written: " + strings.Join(summary.ChangedFields, ", ")
+			}
+			if len(summary.RejectedFields) > 0 {
+				enMsg += "\n- Rejected fields: " + strings.Join(summary.RejectedFields, ", ")
+			}
+			if len(summary.UnchangedDefaults) > 0 {
+				enMsg += "\n- Defaults still in use: " + strings.Join(summary.UnchangedDefaults, ", ")
+			}
+		}
+	}
 	if lang == "zh" {
 		return zhMsg
 	}
 	return enMsg
 }
 
+type strategyToolChangeSummary struct {
+	CreatedStrategyID string   `json:"created_strategy_id"`
+	StrategyID        string   `json:"strategy_id"`
+	ChangedFields     []string `json:"changed_fields"`
+	UnchangedDefaults []string `json:"unchanged_defaults"`
+	RejectedFields    []string `json:"rejected_fields"`
+}
+
+func parseStrategyToolChangeSummary(raw string) strategyToolChangeSummary {
+	var payload strategyToolChangeSummary
+	_ = json.Unmarshal([]byte(raw), &payload)
+	payload.ChangedFields = cleanStringList(payload.ChangedFields)
+	payload.UnchangedDefaults = cleanStringList(payload.UnchangedDefaults)
+	payload.RejectedFields = cleanStringList(payload.RejectedFields)
+	return payload
+}
+
 func parseToolWarnings(raw string) []string {
 	var payload struct {
 		Warnings []string `json:"warnings"`
diff --git a/agent/skill_outcome.go b/agent/skill_outcome.go
index 8922ad2d..383ef010 100644
--- a/agent/skill_outcome.go
+++ b/agent/skill_outcome.go
@@ -152,6 +152,8 @@ Rules:
 - Use route "replan" when the user's task is not complete yet and the planner should continue from the new skill outcome.
 - Prefer route "replan" for recoverable errors, unmet goals, missing prerequisites, or cases where another skill/tool sequence may help.
 - If you choose "complete", produce the final user-facing answer in the user's language.
+- For strategy_management create/update outcomes, only mention config fields present in changed_fields, unchanged_defaults, rejected_fields, warnings, or user_message. Do not add strategy settings that are not in the structured skill outcome.
+- If a strategy field is not in the current StrategyConfig or appears in rejected_fields, say it was not written / is not in the current strategy config. Do not use trading common sense to invent fields.
 - ` + cleanUserFacingReplyInstruction + `
 
 Return JSON with this exact shape:
diff --git a/agent/tools.go b/agent/tools.go
index c1705385..e381451d 100644
--- a/agent/tools.go
+++ b/agent/tools.go
@@ -43,6 +43,243 @@ var (
 // agentTools returns the tools available to the LLM for autonomous action.
 func agentTools() []mcp.Tool { return cachedTools }
 
+func plannerToolsForText(text string) []mcp.Tool {
+	domain := plannerToolDomainForText(text)
+	compactStrategy := !looksLikeStrategyMutationIntent(text)
+	names := plannerToolNamesForDomain(domain)
+	return toolsByName(names, compactStrategy)
+}
+
+func plannerToolDomainForText(text string) string {
+	lower := strings.ToLower(strings.TrimSpace(text))
+	if lower == "" {
+		return "general"
+	}
+	if containsAny(lower, []string{"诊断", "排查", "为什么", "为啥", "失败", "报错", "异常", "停止", "没下单", "failed", "error", "diagnose", "debug", "logs", "stopped", "not trading"}) {
+		return "diagnosis"
+	}
+	if hasExplicitManagementDomainCue(text, "exchange") || containsAny(lower, []string{"交易所", "exchange", "apikey", "secret", "passphrase", "wallet address", "api凭证"}) {
+		return "exchange"
+	}
+	if hasExplicitManagementDomainCue(text, "model") || containsAny(lower, []string{"ai model", "模型", "provider", "api key", "custom_model", "custom api"}) {
+		return "model"
+	}
+	if hasExplicitManagementDomainCue(text, "strategy") || containsAny(lower, []string{"策略", "strategy", "选币", "止盈", "止损", "杠杆", "风控", "risk control"}) {
+		return "strategy"
+	}
+	if hasExplicitManagementDomainCue(text, "trader") || containsAny(lower, []string{"交易员", "trader", "启动", "停止交易员", "扫描间隔", "竞技场"}) {
+		return "trader"
+	}
+	if containsAny(lower, []string{"余额", "资产", "仓位", "持仓", "订单", "成交", "交易历史", "balance", "position", "positions", "trade history", "account"}) {
+		return "account"
+	}
+	if containsAny(lower, []string{"行情", "价格", "k线", "kline", "market", "price", "btc", "eth", "sol", "usdt", "股票", "stock"}) {
+		return "market"
+	}
+	return "general"
+}
+
+func plannerToolNamesForDomain(domain string) []string {
+	switch domain {
+	case "market":
+		return []string{"get_market_snapshot", "get_market_price", "get_kline", "search_stock"}
+	case "account":
+		return []string{"get_balance", "get_positions", "get_trade_history"}
+	case "trader":
+		return []string{"get_model_configs", "get_exchange_configs", "get_strategies", "manage_trader"}
+	case "model":
+		return []string{"get_model_configs", "manage_model_config"}
+	case "exchange":
+		return []string{"get_exchange_configs", "manage_exchange_config"}
+	case "strategy":
+		return []string{"get_strategies", "manage_strategy"}
+	case "diagnosis":
+		return []string{"get_backend_logs", "get_model_configs", "get_exchange_configs", "get_strategies", "manage_trader"}
+	default:
+		return []string{
+			"get_preferences", "manage_preferences",
+			"get_backend_logs",
+			"get_exchange_configs", "manage_exchange_config",
+			"get_model_configs", "manage_model_config",
+			"get_strategies", "manage_strategy",
+			"manage_trader",
+			"get_balance", "get_positions", "get_trade_history",
+			"get_market_snapshot", "get_market_price", "get_kline", "search_stock",
+		}
+	}
+}
+
+func toolsByName(names []string, compactStrategy bool) []mcp.Tool {
+	if len(names) == 0 {
+		return nil
+	}
+	byName := make(map[string]mcp.Tool, len(cachedTools))
+	for _, tool := range cachedTools {
+		byName[tool.Function.Name] = tool
+	}
+	out := make([]mcp.Tool, 0, len(names))
+	seen := make(map[string]bool, len(names))
+	for _, name := range names {
+		if seen[name] {
+			continue
+		}
+		seen[name] = true
+		tool, ok := byName[name]
+		if !ok {
+			continue
+		}
+		if compactStrategy && name == "manage_strategy" {
+			tool = compactManageStrategyTool(tool)
+		}
+		out = append(out, tool)
+	}
+	return out
+}
+
+func compactManageStrategyTool(tool mcp.Tool) mcp.Tool {
+	tool.Function.Description = "List, query, delete, activate, duplicate, create, or update strategy templates. Planning schema is compact; use action plus strategy_id/name/description/lang/is_public/config_visible, and include config only when the user explicitly provides strategy config fields."
+	tool.Function.Parameters = map[string]any{
+		"type": "object",
+		"properties": map[string]any{
+			"action":         map[string]any{"type": "string", "enum": []string{"list", "create", "update", "delete", "activate", "duplicate", "get_default_config"}},
+			"strategy_id":    map[string]any{"type": "string"},
+			"name":           map[string]any{"type": "string"},
+			"description":    map[string]any{"type": "string"},
+			"lang":           map[string]any{"type": "string", "enum": []string{"zh", "en"}},
+			"is_public":      map[string]any{"type": "boolean"},
+			"config_visible": map[string]any{"type": "boolean"},
+			"config":         map[string]any{"type": "object", "description": "Strategy config patch. Use precise StrategyConfig field paths/objects from the user request; grid risk fields such as max_drawdown_pct, stop_loss_pct, and daily_loss_limit_pct belong under grid_config. Omit when listing/querying/deleting/activating/duplicating."},
+		},
+		"required": []string{"action"},
+	}
+	return tool
+}
+
+func looksLikeStrategyMutationIntent(text string) bool {
+	lower := strings.ToLower(strings.TrimSpace(text))
+	return hasExplicitManagementDomainCue(text, "strategy") &&
+		containsAny(lower, []string{"创建", "新建", "创一个", "创个", "建一个", "修改", "更新", "编辑", "调整", "配置", "create", "new", "update", "edit", "configure"})
+}
+
+type strategyConfigPatchValidation struct {
+	Config            map[string]any
+	ChangedFields     []string
+	UnchangedDefaults []string
+	RejectedFields    []string
+}
+
+func validateStrategyConfigPatch(config map[string]any) strategyConfigPatchValidation {
+	out := strategyConfigPatchValidation{
+		Config: map[string]any{},
+	}
+	if len(config) == 0 {
+		out.UnchangedDefaults = defaultStrategyConfigSections()
+		return out
+	}
+	schema := strategyConfigSchema()
+	props, _ := schema["properties"].(map[string]any)
+	for key, value := range config {
+		key = strings.TrimSpace(key)
+		if key == "" {
+			continue
+		}
+		prop, ok := props[key]
+		if !ok {
+			out.RejectedFields = append(out.RejectedFields, key+" (not in current strategy config)")
+			continue
+		}
+		cleaned, changed, rejected := sanitizeStrategyConfigValue(key, value, prop)
+		out.RejectedFields = append(out.RejectedFields, rejected...)
+		if len(changed) == 0 {
+			continue
+		}
+		out.Config[key] = cleaned
+		out.ChangedFields = append(out.ChangedFields, changed...)
+	}
+	out.UnchangedDefaults = unchangedStrategyDefaults(out.ChangedFields)
+	sort.Strings(out.ChangedFields)
+	sort.Strings(out.UnchangedDefaults)
+	sort.Strings(out.RejectedFields)
+	return out
+}
+
+func sanitizeStrategyConfigValue(path string, value any, schema any) (any, []string, []string) {
+	schemaMap, _ := schema.(map[string]any)
+	if schemaMap == nil {
+		return value, []string{path}, nil
+	}
+	if strings.EqualFold(strings.TrimSpace(fmt.Sprint(schemaMap["type"])), "object") {
+		props, _ := schemaMap["properties"].(map[string]any)
+		if len(props) == 0 {
+			return value, []string{path}, nil
+		}
+		valueMap, ok := value.(map[string]any)
+		if !ok {
+			if typed, ok := value.(map[string]string); ok {
+				valueMap = make(map[string]any, len(typed))
+				for k, v := range typed {
+					valueMap[k] = v
+				}
+				ok = true
+			}
+		}
+		if !ok {
+			return nil, nil, []string{path + " (expected object)"}
+		}
+		out := make(map[string]any, len(valueMap))
+		var changed []string
+		var rejected []string
+		for key, nestedValue := range valueMap {
+			key = strings.TrimSpace(key)
+			if key == "" {
+				continue
+			}
+			nestedPath := path + "." + key
+			prop, ok := props[key]
+			if !ok {
+				rejected = append(rejected, nestedPath+" (not in current strategy config)")
+				continue
+			}
+			cleaned, nestedChanged, nestedRejected := sanitizeStrategyConfigValue(nestedPath, nestedValue, prop)
+			rejected = append(rejected, nestedRejected...)
+			if len(nestedChanged) == 0 {
+				continue
+			}
+			out[key] = cleaned
+			changed = append(changed, nestedChanged...)
+		}
+		if len(out) == 0 {
+			return nil, nil, rejected
+		}
+		return out, changed, rejected
+	}
+	return value, []string{path}, nil
+}
+
+func defaultStrategyConfigSections() []string {
+	return []string{"strategy_type", "language", "coin_source", "indicators", "custom_prompt", "risk_control", "prompt_sections", "grid_config"}
+}
+
+func unchangedStrategyDefaults(changedFields []string) []string {
+	changedTop := make(map[string]bool, len(changedFields))
+	for _, field := range changedFields {
+		top := strings.TrimSpace(field)
+		if idx := strings.Index(top, "."); idx >= 0 {
+			top = top[:idx]
+		}
+		if top != "" {
+			changedTop[top] = true
+		}
+	}
+	out := make([]string, 0, len(defaultStrategyConfigSections()))
+	for _, section := range defaultStrategyConfigSections() {
+		if !changedTop[section] {
+			out = append(out, section)
+		}
+	}
+	return out
+}
+
 func normalizedEntityName(value string) string {
 	return strings.ToLower(strings.TrimSpace(value))
 }
@@ -1775,14 +2012,15 @@ func (a *Agent) toolManageStrategy(storeUserID, argsJSON string) string {
 		if lockedField, ok := strategyConfigContainsLockedField(args.Config); ok {
 			return fmt.Sprintf(`{"error":"%s"}`, strategyLockedFieldError("zh", lockedField))
 		}
+		validation := validateStrategyConfigPatch(args.Config)
 		if err := a.ensureUniqueStrategyName(storeUserID, name, ""); err != nil {
 			return fmt.Sprintf(`{"error":"%s"}`, err)
 		}
 		defaultConfig := store.GetDefaultStrategyConfig(strings.TrimSpace(args.Lang))
 		var cfg any = defaultConfig
 		var warnings []string
-		if len(args.Config) > 0 {
-			merged, err := store.MergeStrategyConfig(defaultConfig, args.Config)
+		if len(validation.Config) > 0 {
+			merged, err := store.MergeStrategyConfig(defaultConfig, validation.Config)
 			if err != nil {
 				return fmt.Sprintf(`{"error":"invalid strategy config: %s"}`, err)
 			}
@@ -1813,10 +2051,14 @@ func (a *Agent) toolManageStrategy(storeUserID, argsJSON string) string {
 			return fmt.Sprintf(`{"error":"failed to create strategy: %s"}`, err)
 		}
 		payload, _ := json.Marshal(map[string]any{
-			"status":   "ok",
-			"action":   "create",
-			"strategy": safeStrategyForTool(record),
-			"warnings": warnings,
+			"status":              "ok",
+			"action":              "create",
+			"created_strategy_id": record.ID,
+			"strategy":            safeStrategyForTool(record),
+			"changed_fields":      validation.ChangedFields,
+			"unchanged_defaults":  validation.UnchangedDefaults,
+			"rejected_fields":     validation.RejectedFields,
+			"warnings":            warnings,
 		})
 		return string(payload)
 	case "update":
@@ -1827,6 +2069,7 @@ func (a *Agent) toolManageStrategy(storeUserID, argsJSON string) string {
 		if lockedField, ok := strategyConfigContainsLockedField(args.Config); ok {
 			return fmt.Sprintf(`{"error":"%s"}`, strategyLockedFieldError("zh", lockedField))
 		}
+		validation := validateStrategyConfigPatch(args.Config)
 		existing, err := a.store.Strategy().Get(storeUserID, strategyID)
 		if err != nil {
 			return fmt.Sprintf(`{"error":"failed to load strategy: %s"}`, err)
@@ -1855,16 +2098,29 @@ func (a *Agent) toolManageStrategy(storeUserID, argsJSON string) string {
 		if args.ConfigVisible != nil {
 			configVisible = *args.ConfigVisible
 		}
+		metadataChanged := make([]string, 0, 4)
+		if !sameEntityName(name, existing.Name) {
+			metadataChanged = append(metadataChanged, "name")
+		}
+		if description != existing.Description {
+			metadataChanged = append(metadataChanged, "description")
+		}
+		if isPublic != existing.IsPublic {
+			metadataChanged = append(metadataChanged, "is_public")
+		}
+		if configVisible != existing.ConfigVisible {
+			metadataChanged = append(metadataChanged, "config_visible")
+		}
 		configJSON := existing.Config
 		var warnings []string
-		if len(args.Config) > 0 {
+		if len(validation.Config) > 0 {
 			var existingConfig store.StrategyConfig
 			if strings.TrimSpace(existing.Config) != "" {
 				if err := json.Unmarshal([]byte(existing.Config), &existingConfig); err != nil {
 					return fmt.Sprintf(`{"error":"failed to load existing strategy config: %s"}`, err)
 				}
 			}
-			merged, err := store.MergeStrategyConfig(existingConfig, args.Config)
+			merged, err := store.MergeStrategyConfig(existingConfig, validation.Config)
 			if err != nil {
 				return fmt.Sprintf(`{"error":"invalid strategy config: %s"}`, err)
 			}
@@ -1896,11 +2152,18 @@ func (a *Agent) toolManageStrategy(storeUserID, argsJSON string) string {
 		if err != nil {
 			return fmt.Sprintf(`{"error":"strategy updated but failed to reload: %s"}`, err)
 		}
+		changedFields := append([]string{}, metadataChanged...)
+		changedFields = append(changedFields, validation.ChangedFields...)
+		sort.Strings(changedFields)
 		payload, _ := json.Marshal(map[string]any{
-			"status":   "ok",
-			"action":   "update",
-			"strategy": safeStrategyForTool(updated),
-			"warnings": warnings,
+			"status":             "ok",
+			"action":             "update",
+			"strategy_id":        updated.ID,
+			"strategy":           safeStrategyForTool(updated),
+			"changed_fields":     changedFields,
+			"unchanged_defaults": validation.UnchangedDefaults,
+			"rejected_fields":    validation.RejectedFields,
+			"warnings":           warnings,
 		})
 		return string(payload)
 	case "delete":
diff --git a/agent/unified_turn_router_test.go b/agent/unified_turn_router_test.go
index 79edc1d3..214e22e0 100644
--- a/agent/unified_turn_router_test.go
+++ b/agent/unified_turn_router_test.go
@@ -34,6 +34,53 @@ func TestParseUnifiedTurnDecisionNormalizesContextPolicy(t *testing.T) {
 	}
 }
 
+func TestParseUnifiedTurnDecisionAcceptsSkillTaskList(t *testing.T) {
+	raw := `{
+		"topic_intent": "start_new",
+		"business_action": "skill_tasks",
+		"context_mode": "fresh_context",
+		"tasks": [
+			{"id":"task_1","skill":"strategy_management","action":"create","request":"创建高频交易策略","depends_on":[]},
+			{"id":"task_2","skill":"trader_management","action":"configure_strategy","request":"绑定到交易员","depends_on":["task_1"]}
+		],
+		"confidence": 0.86
+	}`
+
+	decision, err := parseUnifiedTurnDecision(raw)
+	if err != nil {
+		t.Fatalf("parse unified decision: %v", err)
+	}
+	if decision.BusinessAction != "skill_tasks" {
+		t.Fatalf("expected skill_tasks, got %q", decision.BusinessAction)
+	}
+	if len(decision.Tasks) != 2 {
+		t.Fatalf("expected 2 tasks, got %+v", decision.Tasks)
+	}
+	if decision.Tasks[0].Skill != "strategy_management" || decision.Tasks[0].Action != "create" {
+		t.Fatalf("unexpected first task: %+v", decision.Tasks[0])
+	}
+	if !decision.reliable() {
+		t.Fatalf("expected task-list decision to be reliable: %+v", decision)
+	}
+}
+
+func TestUnifiedTurnDecisionNewSkillCanUseSingleTask(t *testing.T) {
+	decision := normalizeUnifiedTurnDecision(unifiedTurnDecision{
+		TopicIntent:    "start_new",
+		BusinessAction: "new_skill",
+		ContextMode:    "fresh_context",
+		Tasks: []WorkflowTask{{
+			Skill:   "strategy_management",
+			Action:  "create",
+			Request: "创建高频交易策略",
+		}},
+		Confidence: 0.9,
+	})
+	if !decision.reliable() {
+		t.Fatalf("expected new_skill with task list to be reliable: %+v", decision)
+	}
+}
+
 func TestUnifiedTurnDecisionRejectsLowConfidenceAndIncompleteDirectAnswer(t *testing.T) {
 	lowConfidence := unifiedTurnDecision{
 		TopicIntent:    "start_new",
@@ -99,6 +146,8 @@ func TestBuildUnifiedTurnRouterPromptNamesContextPolicy(t *testing.T) {
 		"context_mode values",
 		"fresh_context",
 		"downstream modules",
+		"tasks format",
+		"skill_tasks",
 	} {
 		if !strings.Contains(systemPrompt, want) {
 			t.Fatalf("expected system prompt to contain %q", want)