diff --git a/controller/app.go b/controller/app.go
index d5da5b1..a94ab10 100644
--- a/controller/app.go
+++ b/controller/app.go
@@ -95,6 +95,14 @@ func (a *App) RevokeAgent(agentID string) error {
 	return a.control.RevokeAgent(agentID)
 }
 
+// RestartAgent sends a paired, authenticated lifecycle command to one online Agent.
+func (a *App) RestartAgent(agentID string) (controlplane.AgentRestart, error) {
+	if a.startErr != nil {
+		return controlplane.AgentRestart{}, a.startErr
+	}
+	return a.control.RestartAgent(agentID)
+}
+
 func (a *App) CreateTask(agentID, instruction string, timeoutSeconds int) (string, error) {
 	if a.startErr != nil {
 		return "", a.startErr
diff --git a/controller/frontend/src/App.tsx b/controller/frontend/src/App.tsx
index 7d507a3..84ad346 100644
--- a/controller/frontend/src/App.tsx
+++ b/controller/frontend/src/App.tsx
@@ -1,6 +1,6 @@
 import {FormEvent, useEffect, useRef, useState} from "react";
 import "./App.css";
-import {AgentDirectory, AgentLLMStatuses, AgentMCPStatuses, AgentOperationalDetails, AgentSkillStatuses, Agents, AuditTrail, CancelTask, ClearTaskHistory, ContinueTaskSession, CreatePairingCode, CreateTaskWithImages, DeleteMCPTool, DeployLLMProfile, DeploySkill, DispatchMentionsWithImages, ImportSkillsFromDirectory, LLMProfiles, MCPTools, OnlineAgents, RecentTasks, RenameAgent, RetryTask, RevokeAgent, SaveLLMProfile, SaveMCPTool, SaveSkill, SetAgentMCPTools, SetAgentSkills, SkillDocument, Skills, Snapshot, TaskDetails, TaskSessionTurns} from "../wailsjs/go/main/App";
+import {AgentDirectory, AgentLLMStatuses, AgentMCPStatuses, AgentOperationalDetails, AgentSkillStatuses, Agents, AuditTrail, CancelTask, ClearTaskHistory, ContinueTaskSession, CreatePairingCode, CreateTaskWithImages, DeleteMCPTool, DeployLLMProfile, DeploySkill, DispatchMentionsWithImages, ImportSkillsFromDirectory, LLMProfiles, MCPTools, OnlineAgents, RecentTasks, RenameAgent, RestartAgent, RetryTask, RevokeAgent, SaveLLMProfile, SaveMCPTool, SaveSkill, SetAgentMCPTools, SetAgentSkills, SkillDocument, Skills, Snapshot, TaskDetails, TaskSessionTurns} from "../wailsjs/go/main/App";
 import {controlplane, main} from "../wailsjs/go/models";
 
 type Page = "overview" | "agents" | "batch" | "direct" | "skills" | "plugins" | "keys" | "audit" | "settings";
@@ -187,6 +187,14 @@ function App() {
         finally { setLoading(false); }
     };
 
+    const restartAgent = async (agent: controlplane.Agent) => {
+        if (agent.status !== "online") { setError("该 Agent 当前离线，无法通过已认证连接发送重启指令。"); return; }
+        if (!window.confirm(`重启“${agent.name}”？该 Agent 正在运行的任务将被安全取消。`)) return;
+        setLoading(true);
+        try { await RestartAgent(agent.id); setNotice(`已请求 ${agent.name} 重启，正在等待其重新连接。`); await refresh(); }
+        catch (reason) { setError(String(reason)); } finally { setLoading(false); }
+    };
+
     const saveLLMProfile = async (input: controlplane.LLMProfileInput) => {
         setLoading(true);
         try {
@@ -277,7 +285,7 @@ function App() {
         {error && <div className="message error">{error}</div>}
         {notice && <div className="message notice">{notice}</div>}
         {page === "overview" && <><Overview snapshot={snapshot} agents={agents} onNavigate={setPage}/><Tasks agents={agents} tasks={tasks} loading={loading} onCancel={(task) => void cancelTask(task)} onRetry={(task) => void retryTask(task)} onClearHistory={() => void clearTaskHistory()} onOpen={(task) => void showTaskDetail(task)}/></>}
-        {page === "agents" && <AgentManagement skills={skills} skillStatuses={agentSkillStatuses} mcpTools={mcpTools} mcpStatuses={agentMCPStatuses} loading={loading} onRename={(agentID, name) => void renameAgent(agentID, name)} onSetSkills={(targetAgentID, skillIDs) => void setAgentSkills(targetAgentID, skillIDs)} onSetMCPTools={(targetAgentID, toolIDs) => void setAgentMCPTools(targetAgentID, toolIDs)} onRevoke={(agent) => void revokeAgent(agent)} onPair={() => setPage("keys")}/>} 
+        {page === "agents" && <AgentManagement skills={skills} skillStatuses={agentSkillStatuses} mcpTools={mcpTools} mcpStatuses={agentMCPStatuses} loading={loading} onRename={(agentID, name) => void renameAgent(agentID, name)} onSetSkills={(targetAgentID, skillIDs) => void setAgentSkills(targetAgentID, skillIDs)} onSetMCPTools={(targetAgentID, toolIDs) => void setAgentMCPTools(targetAgentID, toolIDs)} onRestart={(agent) => void restartAgent(agent)} onRevoke={(agent) => void revokeAgent(agent)} onPair={() => setPage("keys")}/>} 
         {page === "batch" && <MentionDispatch agents={agents} tasks={tasks} loading={loading} onSubmit={dispatchMentions} onOpen={(task) => void showTaskDetail(task)}/>}
         {page === "direct" && <DirectAssignment agents={agents} tasks={tasks} loading={loading} onSubmit={dispatchDirectTask} onOpen={(task) => void showTaskDetail(task)}/>} 
         {page === "keys" && <><Security snapshot={snapshot} pairingCode={pairingCode} loading={loading} onCreate={() => void createPairingCode()}/><LLMConfiguration profiles={llmProfiles} agents={managedAgents} statuses={agentLLMStatuses} loading={loading} onSave={(input) => void saveLLMProfile(input)} onDeploy={(profileID, agentIDs) => void deployLLMProfile(profileID, agentIDs)}/></>} 
@@ -300,7 +308,7 @@ function Overview({snapshot, agents, onNavigate}: {snapshot?: controlplane.Snaps
     </>;
 }
 
-function AgentManagement({skills, skillStatuses, mcpTools, mcpStatuses, loading, onRename, onSetSkills, onSetMCPTools, onRevoke, onPair}: {skills: controlplane.Skill[]; skillStatuses: controlplane.AgentSkillStatus[]; mcpTools: controlplane.MCPTool[]; mcpStatuses: controlplane.AgentMCPStatus[]; loading: boolean; onRename: (agentID: string, name: string) => void; onSetSkills: (agentID: string, skillIDs: string[]) => void; onSetMCPTools: (agentID: string, toolIDs: string[]) => void; onRevoke: (agent: controlplane.Agent) => void; onPair: () => void}) {
+function AgentManagement({skills, skillStatuses, mcpTools, mcpStatuses, loading, onRename, onSetSkills, onSetMCPTools, onRestart, onRevoke, onPair}: {skills: controlplane.Skill[]; skillStatuses: controlplane.AgentSkillStatus[]; mcpTools: controlplane.MCPTool[]; mcpStatuses: controlplane.AgentMCPStatus[]; loading: boolean; onRename: (agentID: string, name: string) => void; onSetSkills: (agentID: string, skillIDs: string[]) => void; onSetMCPTools: (agentID: string, toolIDs: string[]) => void; onRestart: (agent: controlplane.Agent) => void; onRevoke: (agent: controlplane.Agent) => void; onPair: () => void}) {
     const [query, setQuery] = useState("");
     const [connectionStatus, setConnectionStatus] = useState("");
     const [mcpStatus, setMCPStatus] = useState("");
@@ -334,7 +342,7 @@ function AgentManagement({skills, skillStatuses, mcpTools, mcpStatuses, loading,
         <p className="management-intro">名称搜索和所有状态筛选均在总控服务端执行。点击卡片可查看最近任务、审计、部署轨迹和失败原因。</p>
         <div className="agent-directory-toolbar"><label>按名称搜索<input type="search" value={query} onChange={(event) => updateFilter(() => setQuery(event.target.value))} placeholder="输入 Agent 名称"/></label><label>连接<select value={connectionStatus} onChange={event => updateFilter(() => setConnectionStatus(event.target.value))}><option value="">全部</option><option value="online">在线</option><option value="offline">离线</option></select></label><label>MCP<select value={mcpStatus} onChange={event => updateFilter(() => setMCPStatus(event.target.value))}><option value="">全部</option><option value="未启用">未启用</option><option value="pending">等待连接</option><option value="sent">下发中</option><option value="applied">已应用</option><option value="failed">失败</option></select></label><label>LLM<select value={llmStatus} onChange={event => updateFilter(() => setLLMStatus(event.target.value))}><option value="">全部</option><option value="未部署">未部署</option><option value="pending">等待连接</option><option value="sent">下发中</option><option value="applied">已应用</option><option value="failed">失败</option></select></label><label>Skill<select value={skillStatus} onChange={event => updateFilter(() => setSkillStatus(event.target.value))}><option value="">全部</option><option value="未选择">未选择</option><option value="pending">等待连接</option><option value="sent">下发中</option><option value="applied">已应用</option><option value="failed">失败</option></select></label><span>{directoryLoading ? "正在查询…" : `共 ${directory.totalItems} 个 Agent`}</span></div>
         {directoryError && <p className="directory-error">{directoryError}</p>}{directory.totalItems === 0 ? <p className="empty">还没有符合筛选条件的 Agent。</p> : <><div className="agent-card-list">{directory.items.map(agent => <AgentCard key={agent.id} agent={agent} onOpen={() => void openAgent(agent)}/>)}</div><div className="directory-pagination"><button className="quiet" type="button" disabled={directoryLoading || page <= 1} onClick={() => setPage(current => current - 1)}>上一页</button><span>第 {directory.page} / {directory.totalPages} 页</span><button className="quiet" type="button" disabled={directoryLoading || page >= directory.totalPages} onClick={() => setPage(current => current + 1)}>下一页</button></div></>}
-        {selectedAgent && <AgentDetailModal agent={selectedDetail?.agent ?? new controlplane.Agent(selectedAgent)} detail={selectedDetail} skills={skills} selectedSkillIDs={skillStatuses.filter(status => status.agentID === selectedAgent.id).map(status => status.skillID)} deployment={selectedDetail?.deployment} mcpTools={mcpTools} selectedToolIDs={mcpStatuses.filter(status => status.agentID === selectedAgent.id).map(status => status.toolID)} toolStatuses={mcpStatuses.filter(status => status.agentID === selectedAgent.id)} loading={loading || detailLoading} onClose={() => { setSelectedAgent(undefined); setSelectedDetail(undefined); }} onRename={onRename} onSetSkills={onSetSkills} onSetMCPTools={onSetMCPTools} onRevoke={onRevoke}/>} 
+        {selectedAgent && <AgentDetailModal agent={selectedDetail?.agent ?? new controlplane.Agent(selectedAgent)} detail={selectedDetail} skills={skills} selectedSkillIDs={skillStatuses.filter(status => status.agentID === selectedAgent.id).map(status => status.skillID)} deployment={selectedDetail?.deployment} mcpTools={mcpTools} selectedToolIDs={mcpStatuses.filter(status => status.agentID === selectedAgent.id).map(status => status.toolID)} toolStatuses={mcpStatuses.filter(status => status.agentID === selectedAgent.id)} loading={loading || detailLoading} onClose={() => { setSelectedAgent(undefined); setSelectedDetail(undefined); }} onRename={onRename} onSetSkills={onSetSkills} onSetMCPTools={onSetMCPTools} onRestart={onRestart} onRevoke={onRevoke}/>} 
     </section>;
 }
 
@@ -355,12 +363,13 @@ function useCloseOnEscape(onClose: () => void) {
     }, [onClose]);
 }
 
-function AgentDetailModal({agent, detail, skills, selectedSkillIDs, deployment, mcpTools, selectedToolIDs, toolStatuses, loading, onClose, onRename, onSetSkills, onSetMCPTools, onRevoke}: {agent: controlplane.Agent; detail?: controlplane.AgentOperationalDetail; skills: controlplane.Skill[]; selectedSkillIDs: string[]; deployment?: controlplane.AgentDirectoryItem; mcpTools: controlplane.MCPTool[]; selectedToolIDs: string[]; toolStatuses: controlplane.AgentMCPStatus[]; loading: boolean; onClose: () => void; onRename: (agentID: string, name: string) => void; onSetSkills: (agentID: string, skillIDs: string[]) => void; onSetMCPTools: (agentID: string, toolIDs: string[]) => void; onRevoke: (agent: controlplane.Agent) => void}) {
+function AgentDetailModal({agent, detail, skills, selectedSkillIDs, deployment, mcpTools, selectedToolIDs, toolStatuses, loading, onClose, onRename, onSetSkills, onSetMCPTools, onRestart, onRevoke}: {agent: controlplane.Agent; detail?: controlplane.AgentOperationalDetail; skills: controlplane.Skill[]; selectedSkillIDs: string[]; deployment?: controlplane.AgentDirectoryItem; mcpTools: controlplane.MCPTool[]; selectedToolIDs: string[]; toolStatuses: controlplane.AgentMCPStatus[]; loading: boolean; onClose: () => void; onRename: (agentID: string, name: string) => void; onSetSkills: (agentID: string, skillIDs: string[]) => void; onSetMCPTools: (agentID: string, toolIDs: string[]) => void; onRestart: (agent: controlplane.Agent) => void; onRevoke: (agent: controlplane.Agent) => void}) {
     useCloseOnEscape(onClose);
-    return <div className="task-modal-backdrop" role="presentation" onMouseDown={(event) => { if (event.currentTarget === event.target) onClose(); }}><section className="task-modal agent-detail-modal" role="dialog" aria-modal="true" aria-labelledby="agent-detail-title"><div className="task-modal-heading"><div><p className="eyebrow">AGENT DETAIL</p><h2 id="agent-detail-title">{agent.name}</h2></div><button className="quiet" type="button" onClick={onClose}>关闭</button></div>{loading && !detail ? <p>正在读取 Agent 运行详情…</p> : <><AgentRow agent={agent} skills={skills} selectedSkillIDs={selectedSkillIDs} deployment={deployment} mcpTools={mcpTools} selectedToolIDs={selectedToolIDs} toolStatuses={toolStatuses} loading={loading} onRename={onRename} onSetSkills={onSetSkills} onSetMCPTools={onSetMCPTools} onRevoke={onRevoke}/><AgentOperationalSections detail={detail}/></>}</section></div>;
+    const restarting = agent.restartStatus === "restarting" || agent.restartStatus === "sent";
+    return <div className="task-modal-backdrop" role="presentation" onMouseDown={(event) => { if (event.currentTarget === event.target) onClose(); }}><section className="task-modal agent-detail-modal" role="dialog" aria-modal="true" aria-labelledby="agent-detail-title"><div className="task-modal-heading"><div><p className="eyebrow">AGENT DETAIL</p><h2 id="agent-detail-title">{agent.name}</h2></div><button className="quiet" type="button" onClick={onClose}>关闭</button></div>{loading && !detail ? <p>正在读取 Agent 运行详情…</p> : <><div className="agent-actions"><button className="quiet" type="button" disabled={loading || restarting || agent.status !== "online"} title={agent.status !== "online" ? "Agent 离线，无法通过已认证连接重启" : undefined} onClick={() => onRestart(agent)}>{agent.status !== "online" ? "Agent 离线，无法重启" : restarting ? "重启中…" : "重启 Agent"}</button><span>{agent.restartStatus === "reconnected" ? "已重连" : agent.restartStatus === "failed" ? `重启失败：${agent.restartFailure || "请查看审计日志"}` : restarting ? "重启中，正在等待 Agent 重新连接…" : ""}</span></div><AgentRow agent={agent} skills={skills} selectedSkillIDs={selectedSkillIDs} deployment={deployment} mcpTools={mcpTools} selectedToolIDs={selectedToolIDs} toolStatuses={toolStatuses} loading={loading} onRename={onRename} onSetSkills={onSetSkills} onSetMCPTools={onSetMCPTools} onRestart={onRestart} onRevoke={onRevoke}/><AgentOperationalSections detail={detail}/></>}</section></div>;
 }
 
-function AgentRow({agent, skills, selectedSkillIDs, deployment, mcpTools, selectedToolIDs, toolStatuses, loading, onRename, onSetSkills, onSetMCPTools, onRevoke}: {agent: controlplane.Agent; skills: controlplane.Skill[]; selectedSkillIDs: string[]; deployment?: controlplane.AgentDirectoryItem; mcpTools: controlplane.MCPTool[]; selectedToolIDs: string[]; toolStatuses: controlplane.AgentMCPStatus[]; loading: boolean; onRename: (agentID: string, name: string) => void; onSetSkills: (agentID: string, skillIDs: string[]) => void; onSetMCPTools: (agentID: string, toolIDs: string[]) => void; onRevoke: (agent: controlplane.Agent) => void}) {
+function AgentRow({agent, skills, selectedSkillIDs, deployment, mcpTools, selectedToolIDs, toolStatuses, loading, onRename, onSetSkills, onSetMCPTools, onRestart, onRevoke}: {agent: controlplane.Agent; skills: controlplane.Skill[]; selectedSkillIDs: string[]; deployment?: controlplane.AgentDirectoryItem; mcpTools: controlplane.MCPTool[]; selectedToolIDs: string[]; toolStatuses: controlplane.AgentMCPStatus[]; loading: boolean; onRename: (agentID: string, name: string) => void; onSetSkills: (agentID: string, skillIDs: string[]) => void; onSetMCPTools: (agentID: string, toolIDs: string[]) => void; onRestart: (agent: controlplane.Agent) => void; onRevoke: (agent: controlplane.Agent) => void}) {
     const [name, setName] = useState(agent.name);
     const [selected, setSelected] = useState(selectedSkillIDs);
     const [selectedTools, setSelectedTools] = useState(selectedToolIDs);
diff --git a/controller/frontend/wailsjs/go/main/App.d.ts b/controller/frontend/wailsjs/go/main/App.d.ts
index e7e0312..75945a0 100644
--- a/controller/frontend/wailsjs/go/main/App.d.ts
+++ b/controller/frontend/wailsjs/go/main/App.d.ts
@@ -55,6 +55,8 @@ export function RecentTasks():Promise<Array<controlplane.Task>>;
 
 export function RenameAgent(arg1:string,arg2:string):Promise<void>;
 
+export function RestartAgent(arg1:string):Promise<controlplane.AgentRestart>;
+
 export function RetryTask(arg1:string):Promise<string>;
 
 export function RevokeAgent(arg1:string):Promise<void>;
diff --git a/controller/frontend/wailsjs/go/main/App.js b/controller/frontend/wailsjs/go/main/App.js
index 920232f..46817d7 100644
--- a/controller/frontend/wailsjs/go/main/App.js
+++ b/controller/frontend/wailsjs/go/main/App.js
@@ -106,6 +106,10 @@ export function RenameAgent(arg1, arg2) {
   return window['go']['main']['App']['RenameAgent'](arg1, arg2);
 }
 
+export function RestartAgent(arg1) {
+  return window['go']['main']['App']['RestartAgent'](arg1);
+}
+
 export function RetryTask(arg1) {
   return window['go']['main']['App']['RetryTask'](arg1);
 }
diff --git a/controller/frontend/wailsjs/go/models.ts b/controller/frontend/wailsjs/go/models.ts
index 0223e73..a538725 100644
--- a/controller/frontend/wailsjs/go/models.ts
+++ b/controller/frontend/wailsjs/go/models.ts
@@ -5,6 +5,8 @@ export namespace controlplane {
 	    name: string;
 	    status: string;
 	    lastSeenAt: string;
+	    restartStatus: string;
+	    restartFailure: string;
 	
 	    static createFrom(source: any = {}) {
 	        return new Agent(source);
@@ -16,6 +18,8 @@ export namespace controlplane {
 	        this.name = source["name"];
 	        this.status = source["status"];
 	        this.lastSeenAt = source["lastSeenAt"];
+	        this.restartStatus = source["restartStatus"];
+	        this.restartFailure = source["restartFailure"];
 	    }
 	}
 	export class AgentDirectoryFilter {
@@ -57,6 +61,8 @@ export namespace controlplane {
 	    mcpStatus: string;
 	    mcpFailure: string;
 	    lastDeploymentAt: string;
+	    restartStatus: string;
+	    restartFailure: string;
 	
 	    static createFrom(source: any = {}) {
 	        return new AgentDirectoryItem(source);
@@ -78,6 +84,8 @@ export namespace controlplane {
 	        this.mcpStatus = source["mcpStatus"];
 	        this.mcpFailure = source["mcpFailure"];
 	        this.lastDeploymentAt = source["lastDeploymentAt"];
+	        this.restartStatus = source["restartStatus"];
+	        this.restartFailure = source["restartFailure"];
 	    }
 	}
 	export class AgentDirectoryPage {
@@ -310,6 +318,28 @@ export namespace controlplane {
 		    return a;
 		}
 	}
+	export class AgentRestart {
+	    requestID: string;
+	    agentID: string;
+	    status: string;
+	    requestedAt: string;
+	    confirmedAt: string;
+	    failure: string;
+	
+	    static createFrom(source: any = {}) {
+	        return new AgentRestart(source);
+	    }
+	
+	    constructor(source: any = {}) {
+	        if ('string' === typeof source) source = JSON.parse(source);
+	        this.requestID = source["requestID"];
+	        this.agentID = source["agentID"];
+	        this.status = source["status"];
+	        this.requestedAt = source["requestedAt"];
+	        this.confirmedAt = source["confirmedAt"];
+	        this.failure = source["failure"];
+	    }
+	}
 	export class AgentSkillStatus {
 	    agentID: string;
 	    agentName: string;
diff --git a/controller/internal/controlplane/agent_directory.go b/controller/internal/controlplane/agent_directory.go
index e5af0f0..f825fbf 100644
--- a/controller/internal/controlplane/agent_directory.go
+++ b/controller/internal/controlplane/agent_directory.go
@@ -35,6 +35,8 @@ type AgentDirectoryItem struct {
 	MCPStatus        string `json:"mcpStatus"`
 	MCPFailure       string `json:"mcpFailure"`
 	LastDeploymentAt string `json:"lastDeploymentAt"`
+	RestartStatus    string `json:"restartStatus"`
+	RestartFailure   string `json:"restartFailure"`
 }
 
 type AgentDirectoryPage struct {
@@ -85,11 +87,13 @@ var agentDirectoryBaseSQL = `WITH skill_summary AS (
 	` + mcpStatusExpression("a.agent_id") + ` AS mcp_status,
 	` + mcpErrorExpression("a.agent_id") + ` AS mcp_failure,
 	COALESCE((SELECT MAX(updated_at) FROM agent_mcp_assignments m WHERE m.agent_id = a.agent_id), '') AS mcp_updated_at,
-	MAX(COALESCE(l.updated_at, ''), COALESCE((SELECT MAX(updated_at) FROM agent_mcp_assignments m WHERE m.agent_id = a.agent_id), ''), COALESCE(s.updated_at, '')) AS last_deployment_at
+	MAX(COALESCE(l.updated_at, ''), COALESCE((SELECT MAX(updated_at) FROM agent_mcp_assignments m WHERE m.agent_id = a.agent_id), ''), COALESCE(s.updated_at, '')) AS last_deployment_at,
+	COALESCE(r.status, '') AS restart_status, COALESCE(r.failure, '') AS restart_failure
 	FROM managed_agents a
 	LEFT JOIN agent_llm_assignments l ON l.agent_id = a.agent_id
 	LEFT JOIN llm_profiles p ON p.profile_id = l.profile_id
 	LEFT JOIN skill_summary s ON s.agent_id = a.agent_id
+	LEFT JOIN agent_restart_operations r ON r.agent_id = a.agent_id
 )
 `
 
@@ -164,7 +168,7 @@ func (c *ControlPlane) AgentDirectory(filter AgentDirectoryFilter) (AgentDirecto
 	}
 	queryArgs := append(append([]any{}, args...), filter.PageSize, (page.Page-1)*filter.PageSize)
 	rows, err := c.db.Query(agentDirectoryBaseSQL+`SELECT agent_id, name, status, last_seen_at, llm_profile, llm_status, llm_failure,
-		skill_total, skill_applied, skill_status, skill_failure, mcp_status, mcp_failure, last_deployment_at
+		skill_total, skill_applied, skill_status, skill_failure, mcp_status, mcp_failure, last_deployment_at, restart_status, restart_failure
 		FROM directory`+where+" ORDER BY CASE status WHEN 'online' THEN 0 ELSE 1 END, name COLLATE NOCASE, agent_id LIMIT ? OFFSET ?", queryArgs...)
 	if err != nil {
 		return page, fmt.Errorf("无法读取 Agent 列表: %w", err)
@@ -172,7 +176,7 @@ func (c *ControlPlane) AgentDirectory(filter AgentDirectoryFilter) (AgentDirecto
 	defer rows.Close()
 	for rows.Next() {
 		var item AgentDirectoryItem
-		if err := rows.Scan(&item.ID, &item.Name, &item.Status, &item.LastSeenAt, &item.LLMProfile, &item.LLMStatus, &item.LLMFailure, &item.SkillTotal, &item.SkillApplied, &item.SkillStatus, &item.SkillFailure, &item.MCPStatus, &item.MCPFailure, &item.LastDeploymentAt); err != nil {
+		if err := rows.Scan(&item.ID, &item.Name, &item.Status, &item.LastSeenAt, &item.LLMProfile, &item.LLMStatus, &item.LLMFailure, &item.SkillTotal, &item.SkillApplied, &item.SkillStatus, &item.SkillFailure, &item.MCPStatus, &item.MCPFailure, &item.LastDeploymentAt, &item.RestartStatus, &item.RestartFailure); err != nil {
 			return page, err
 		}
 		item.LLMFailure, item.SkillFailure, item.MCPFailure = sanitizeDisplayText(item.LLMFailure), sanitizeDisplayText(item.SkillFailure), sanitizeDisplayText(item.MCPFailure)
@@ -189,8 +193,8 @@ func (c *ControlPlane) AgentOperationalDetails(agentID string) (AgentOperational
 	// Fetch by ID directly so duplicate display names cannot select another Agent.
 	var item AgentDirectoryItem
 	err := c.db.QueryRow(agentDirectoryBaseSQL+`SELECT agent_id, name, status, last_seen_at, llm_profile, llm_status, llm_failure,
-		skill_total, skill_applied, skill_status, skill_failure, mcp_status, mcp_failure, last_deployment_at FROM directory WHERE agent_id = ?`, agentID).Scan(
-		&item.ID, &item.Name, &item.Status, &item.LastSeenAt, &item.LLMProfile, &item.LLMStatus, &item.LLMFailure, &item.SkillTotal, &item.SkillApplied, &item.SkillStatus, &item.SkillFailure, &item.MCPStatus, &item.MCPFailure, &item.LastDeploymentAt,
+		skill_total, skill_applied, skill_status, skill_failure, mcp_status, mcp_failure, last_deployment_at, restart_status, restart_failure FROM directory WHERE agent_id = ?`, agentID).Scan(
+		&item.ID, &item.Name, &item.Status, &item.LastSeenAt, &item.LLMProfile, &item.LLMStatus, &item.LLMFailure, &item.SkillTotal, &item.SkillApplied, &item.SkillStatus, &item.SkillFailure, &item.MCPStatus, &item.MCPFailure, &item.LastDeploymentAt, &item.RestartStatus, &item.RestartFailure,
 	)
 	if errors.Is(err, sql.ErrNoRows) {
 		return AgentOperationalDetail{}, errors.New("未找到 Agent")
@@ -198,7 +202,7 @@ func (c *ControlPlane) AgentOperationalDetails(agentID string) (AgentOperational
 	if err != nil {
 		return AgentOperationalDetail{}, fmt.Errorf("无法读取 Agent 详情: %w", err)
 	}
-	detail := AgentOperationalDetail{Agent: Agent{ID: item.ID, Name: item.Name, Status: item.Status, LastSeenAt: item.LastSeenAt}, Deployment: item, RecentTasks: make([]Task, 0), RecentAudit: make([]AuditEvent, 0), DeploymentHistory: make([]AuditEvent, 0), Failures: make([]DeploymentFailure, 0)}
+	detail := AgentOperationalDetail{Agent: Agent{ID: item.ID, Name: item.Name, Status: item.Status, LastSeenAt: item.LastSeenAt, RestartStatus: item.RestartStatus, RestartFailure: item.RestartFailure}, Deployment: item, RecentTasks: make([]Task, 0), RecentAudit: make([]AuditEvent, 0), DeploymentHistory: make([]AuditEvent, 0), Failures: make([]DeploymentFailure, 0)}
 	detail.RecentTasks, err = c.recentTasksForAgent(agentID, 8)
 	if err != nil {
 		return detail, err
diff --git a/controller/internal/controlplane/controlplane.go b/controller/internal/controlplane/controlplane.go
index 4023f1f..329cad7 100644
--- a/controller/internal/controlplane/controlplane.go
+++ b/controller/internal/controlplane/controlplane.go
@@ -50,10 +50,12 @@ type Snapshot struct {
 
 // Agent is a safe summary of a registered Agent.
 type Agent struct {
-	ID         string `json:"id"`
-	Name       string `json:"name"`
-	Status     string `json:"status"`
-	LastSeenAt string `json:"lastSeenAt"`
+	ID             string `json:"id"`
+	Name           string `json:"name"`
+	Status         string `json:"status"`
+	LastSeenAt     string `json:"lastSeenAt"`
+	RestartStatus  string `json:"restartStatus"`
+	RestartFailure string `json:"restartFailure"`
 }
 
 // Task is a persisted controller task summary. Credentials never appear here.
@@ -323,7 +325,8 @@ func (c *ControlPlane) removeRevokedAgents() {
 }
 
 func (c *ControlPlane) queryAgents(where string) ([]Agent, error) {
-	rows, err := c.db.Query("SELECT agent_id, name, status, last_seen_at FROM managed_agents " + where + " ORDER BY name")
+	where = strings.ReplaceAll(where, "status", "a.status")
+	rows, err := c.db.Query("SELECT a.agent_id, a.name, a.status, a.last_seen_at, COALESCE(r.status, ''), COALESCE(r.failure, '') FROM managed_agents a LEFT JOIN agent_restart_operations r ON r.agent_id = a.agent_id " + where + " ORDER BY a.name")
 	if err != nil {
 		return nil, err
 	}
@@ -331,7 +334,7 @@ func (c *ControlPlane) queryAgents(where string) ([]Agent, error) {
 	agents := make([]Agent, 0)
 	for rows.Next() {
 		var agent Agent
-		if err := rows.Scan(&agent.ID, &agent.Name, &agent.Status, &agent.LastSeenAt); err != nil {
+		if err := rows.Scan(&agent.ID, &agent.Name, &agent.Status, &agent.LastSeenAt, &agent.RestartStatus, &agent.RestartFailure); err != nil {
 			return nil, err
 		}
 		agents = append(agents, agent)
@@ -511,10 +514,20 @@ func (c *ControlPlane) CancelTask(taskID string) error {
 	if client == nil {
 		return errors.New("目标 Agent 当前不在线，暂不能取消")
 	}
+	// Persist the cancellation intent before sending.  An Agent can respond on
+	// another goroutine immediately; recording it first makes every late
+	// completed/failed event ineligible to overwrite the requested outcome.
+	result, err := c.db.Exec("UPDATE task_runs SET status = 'cancelling' WHERE task_id = ? AND status = ?", taskID, status)
+	if err != nil {
+		return errors.New("无法记录取消状态，请刷新后重试")
+	}
+	if changed, _ := result.RowsAffected(); changed != 1 {
+		return errors.New("任务状态已变化，请刷新后重试")
+	}
 	if err := client.send(map[string]any{"type": "task_cancel", "schema_version": protocolVersion, "task_id": taskID, "reason": "由总控取消"}); err != nil {
+		_, _ = c.db.Exec("UPDATE task_runs SET status = ? WHERE task_id = ? AND status = 'cancelling'", status, taskID)
 		return errors.New("取消指令未发送成功，请刷新后重试")
 	}
-	_, _ = c.db.Exec("UPDATE task_runs SET status = 'cancelling' WHERE task_id = ?", taskID)
 	c.recordEvent(taskID, "cancelling", "已向 Agent 发送取消指令。", time.Now().UTC().Format(time.RFC3339Nano))
 	c.recordAudit("task.cancel_requested", "controller", "agent:"+agentID, "cancelling", "已向 Agent 请求取消任务。")
 	return nil
@@ -723,6 +736,10 @@ func (c *ControlPlane) handleMessage(client *connection, message map[string]any)
 		if client.agentID != "" {
 			c.applyMCPAck(client.agentID, message)
 		}
+	case "agent_restart_ack":
+		if client.agentID != "" {
+			c.applyRestartAck(client.agentID, message)
+		}
 	default:
 		_ = client.send(map[string]any{"type": "error", "message": "不支持的消息类型。"})
 	}
@@ -738,6 +755,7 @@ func (c *ControlPlane) attach(client *connection, agentID string) {
 		_ = previous.ws.Close()
 	}
 	c.setAgentStatus(agentID, "online")
+	c.markRestartReconnected(agentID)
 }
 
 // pushConfigurationSnapshots starts only after the authentication acknowledgement
@@ -877,6 +895,12 @@ func (c *ControlPlane) applyTaskEvent(agentID string, event map[string]any) {
 	if c.db.QueryRow("SELECT agent_id, status, timeout_seconds, COALESCE(session_id, ''), instruction FROM task_runs WHERE task_id = ?", taskID).Scan(&expectedAgentID, &currentStatus, &timeoutSeconds, &sessionID, &instruction) != nil || expectedAgentID != agentID || isTerminal(currentStatus) {
 		return
 	}
+	// Cancellation is a write-once intent.  An already-running model or MCP
+	// request may return late, but it must not change the requested terminal
+	// outcome to completed/failed or resurrect the task as running.
+	if currentStatus == "cancelling" && status != "cancelled" {
+		return
+	}
 	message, _ := event["message"].(string)
 	output, _ := event["output"].(string)
 	errText, _ := event["error"].(string)
@@ -1065,6 +1089,7 @@ func initialiseSchema(db *sql.DB) error {
 		"CREATE TABLE IF NOT EXISTS task_sessions (session_id TEXT PRIMARY KEY, agent_id TEXT NOT NULL, turn_limit INTEGER NOT NULL DEFAULT 50, created_at TEXT NOT NULL, updated_at TEXT NOT NULL)",
 		"CREATE TABLE IF NOT EXISTS task_session_messages (session_id TEXT NOT NULL, position INTEGER NOT NULL, role TEXT NOT NULL, content TEXT NOT NULL, created_at TEXT NOT NULL, PRIMARY KEY(session_id, position))",
 		"CREATE TABLE IF NOT EXISTS audit_events (event_id INTEGER PRIMARY KEY AUTOINCREMENT, action TEXT NOT NULL, actor TEXT NOT NULL, target TEXT NOT NULL, outcome TEXT NOT NULL, detail TEXT NOT NULL, created_at TEXT NOT NULL, operator TEXT NOT NULL DEFAULT '', target_agent_id TEXT NOT NULL DEFAULT '', resource_type TEXT NOT NULL DEFAULT 'system', result TEXT NOT NULL DEFAULT '', security_detail TEXT NOT NULL DEFAULT '')",
+		"CREATE TABLE IF NOT EXISTS agent_restart_operations (agent_id TEXT PRIMARY KEY, request_id TEXT NOT NULL, status TEXT NOT NULL, requested_at TEXT NOT NULL, confirmed_at TEXT, failure TEXT NOT NULL DEFAULT '')",
 	}
 	for _, statement := range statements {
 		if _, err := db.Exec(statement); err != nil {
diff --git a/controller/internal/controlplane/restart.go b/controller/internal/controlplane/restart.go
new file mode 100644
index 0000000..a3226dd
--- /dev/null
+++ b/controller/internal/controlplane/restart.go
@@ -0,0 +1,104 @@
+package controlplane
+
+import (
+	"errors"
+	"fmt"
+	"strings"
+	"time"
+)
+
+const restartCooldown = 60 * time.Second
+
+// AgentRestart is the safe lifecycle state rendered by the controller.  It
+// contains no remote process details or credentials.
+type AgentRestart struct {
+	RequestID   string `json:"requestID"`
+	AgentID     string `json:"agentID"`
+	Status      string `json:"status"`
+	RequestedAt string `json:"requestedAt"`
+	ConfirmedAt string `json:"confirmedAt"`
+	Failure     string `json:"failure"`
+}
+
+// RestartAgent sends an authenticated command over the existing paired WSS
+// session.  The controller deliberately never starts or terminates a remote
+// machine process itself.
+func (c *ControlPlane) RestartAgent(agentID string) (AgentRestart, error) {
+	agentID = strings.TrimSpace(agentID)
+	if agentID == "" {
+		return AgentRestart{}, errors.New("Agent 标识无效")
+	}
+	c.expireStaleAgents()
+	var status string
+	if err := c.db.QueryRow("SELECT status FROM managed_agents WHERE agent_id = ?", agentID).Scan(&status); err != nil {
+		return AgentRestart{}, errors.New("未找到 Agent")
+	}
+	client := c.clientFor(agentID)
+	if status != "online" || client == nil {
+		c.recordAudit("remote.restart_rejected", "controller", "agent:"+agentID, "failed", "目标 Agent 当前离线，无法通过已认证连接发送重启指令。")
+		return AgentRestart{}, errors.New("目标 Agent 当前离线，无法重启；请等待其重新连接")
+	}
+	var existing AgentRestart
+	err := c.db.QueryRow("SELECT request_id, agent_id, status, requested_at, COALESCE(confirmed_at, ''), COALESCE(failure, '') FROM agent_restart_operations WHERE agent_id = ?", agentID).Scan(&existing.RequestID, &existing.AgentID, &existing.Status, &existing.RequestedAt, &existing.ConfirmedAt, &existing.Failure)
+	if err == nil {
+		if existing.Status == "restarting" || existing.Status == "sent" {
+			return existing, nil // idempotent repeated click; do not send another command.
+		}
+		if existing.Status == "reconnected" {
+			if requested, parseErr := time.Parse(time.RFC3339Nano, existing.RequestedAt); parseErr == nil && time.Since(requested) < restartCooldown {
+				return existing, errors.New("该 Agent 刚完成重启，请稍后再试")
+			}
+		}
+	}
+	requestID, err := randomHex(16)
+	if err != nil {
+		return AgentRestart{}, err
+	}
+	now := time.Now().UTC().Format(time.RFC3339Nano)
+	op := AgentRestart{RequestID: requestID, AgentID: agentID, Status: "restarting", RequestedAt: now}
+	if _, err := c.db.Exec(`INSERT INTO agent_restart_operations(request_id, agent_id, status, requested_at, confirmed_at, failure)
+		VALUES (?, ?, ?, ?, '', '') ON CONFLICT(agent_id) DO UPDATE SET request_id=excluded.request_id, status=excluded.status, requested_at=excluded.requested_at, confirmed_at='', failure=''`, requestID, agentID, op.Status, now); err != nil {
+		return AgentRestart{}, fmt.Errorf("无法记录重启操作: %w", err)
+	}
+	c.recordAudit("remote.restart_requested", "controller", "agent:"+agentID, "restarting", "已通过认证控制连接请求 Agent 重启。")
+	if err := client.send(map[string]any{"type": "agent_restart", "schema_version": protocolVersion, "request_id": requestID, "sent_at": now}); err != nil {
+		_, _ = c.db.Exec("UPDATE agent_restart_operations SET status='failed', failure=? WHERE agent_id=? AND request_id=?", "重启指令未发送成功。", agentID, requestID)
+		c.recordAudit("remote.restart_delivery", "controller", "agent:"+agentID, "failed", "重启指令未发送成功。")
+		return AgentRestart{}, errors.New("重启指令未发送成功，请刷新后重试")
+	}
+	c.recordAudit("remote.restart_delivery", "controller", "agent:"+agentID, "sent", "重启指令已下发，等待 Agent 确认。")
+	return op, nil
+}
+
+func (c *ControlPlane) applyRestartAck(agentID string, message map[string]any) {
+	requestID, _ := message["request_id"].(string)
+	status, _ := message["status"].(string)
+	errorText, _ := message["error"].(string)
+	if requestID == "" || (status != "accepted" && status != "restarting" && status != "failed") {
+		return
+	}
+	var expected string
+	if c.db.QueryRow("SELECT request_id FROM agent_restart_operations WHERE agent_id=?", agentID).Scan(&expected) != nil || expected != requestID {
+		return
+	}
+	if status == "failed" {
+		detail := limit(sanitizeDisplayText(errorText), 1000)
+		if detail == "" {
+			detail = "Agent 未能启动替代进程。"
+		}
+		_, _ = c.db.Exec("UPDATE agent_restart_operations SET status='failed', confirmed_at=?, failure=? WHERE agent_id=?", time.Now().UTC().Format(time.RFC3339Nano), detail, agentID)
+		c.recordAudit("remote.restart_confirmed", "agent:"+agentID, "agent:"+agentID, "failed", detail)
+		return
+	}
+	_, _ = c.db.Exec("UPDATE agent_restart_operations SET status='restarting', confirmed_at=?, failure='' WHERE agent_id=?", time.Now().UTC().Format(time.RFC3339Nano), agentID)
+	c.recordAudit("remote.restart_confirmed", "agent:"+agentID, "agent:"+agentID, "restarting", "Agent 已确认正在安全结束任务并重启。")
+}
+
+func (c *ControlPlane) markRestartReconnected(agentID string) {
+	result, err := c.db.Exec("UPDATE agent_restart_operations SET status='reconnected', confirmed_at=?, failure='' WHERE agent_id=? AND status IN ('sent', 'restarting')", time.Now().UTC().Format(time.RFC3339Nano), agentID)
+	if err == nil {
+		if changed, _ := result.RowsAffected(); changed == 1 {
+			c.recordAudit("remote.restart_reconnected", "agent:"+agentID, "agent:"+agentID, "reconnected", "替代 Agent 已重新通过认证并恢复连接。")
+		}
+	}
+}
diff --git a/controller/internal/controlplane/restart_test.go b/controller/internal/controlplane/restart_test.go
new file mode 100644
index 0000000..3543097
--- /dev/null
+++ b/controller/internal/controlplane/restart_test.go
@@ -0,0 +1,122 @@
+package controlplane
+
+import (
+	"crypto/tls"
+	"net"
+	"os"
+	"path/filepath"
+	"strconv"
+	"testing"
+	"time"
+
+	"github.com/gorilla/websocket"
+)
+
+func newRestartTestController(t *testing.T) (*ControlPlane, string) {
+	t.Helper()
+	listener, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatal(err)
+	}
+	port := listener.Addr().(*net.TCPAddr).Port
+	_ = listener.Close()
+	configPath := filepath.Join(t.TempDir(), "controller-config.json")
+	if err := os.WriteFile(configPath, []byte(`{"agent_host":"127.0.0.1","agent_port":`+strconv.Itoa(port)+`}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("MULTICLAW_CONTROLLER_DATA_DIR", t.TempDir())
+	controller, err := Open(configPath)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := controller.StartAgentServer(); err != nil {
+		t.Fatal(err)
+	}
+	return controller, "wss://127.0.0.1:" + strconv.Itoa(port)
+}
+
+func pairRestartTestAgent(t *testing.T, controller *ControlPlane, url string) (*websocket.Conn, string, string) {
+	t.Helper()
+	client, _, err := (&websocket.Dialer{TLSClientConfig: &tls.Config{InsecureSkipVerify: true}}).Dial(url, nil) //nolint:gosec -- test-only TLS.
+	if err != nil {
+		t.Fatal(err)
+	}
+	code, err := controller.CreatePairingCode()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := client.WriteJSON(map[string]any{"type": "pair_request", "code": code, "agent_name": "restart-agent"}); err != nil {
+		t.Fatal(err)
+	}
+	var paired map[string]any
+	if err := client.ReadJSON(&paired); err != nil {
+		t.Fatal(err)
+	}
+	return client, paired["agent_id"].(string), paired["credential"].(string)
+}
+
+func TestRestartOnlineAgentIsIdempotentAndReconnects(t *testing.T) {
+	controller, url := newRestartTestController(t)
+	defer controller.Close()
+	client, agentID, credential := pairRestartTestAgent(t, controller, url)
+	defer client.Close()
+
+	first, err := controller.RestartAgent(agentID)
+	if err != nil || first.Status != "restarting" {
+		t.Fatalf("restart request failed: %#v / %v", first, err)
+	}
+	var command map[string]any
+	if err := client.ReadJSON(&command); err != nil || command["type"] != "agent_restart" || command["request_id"] != first.RequestID {
+		t.Fatalf("expected restart command, got %#v / %v", command, err)
+	}
+	second, err := controller.RestartAgent(agentID)
+	if err != nil || second.RequestID != first.RequestID {
+		t.Fatalf("duplicate must reuse operation: %#v / %v", second, err)
+	}
+	if err := client.WriteJSON(map[string]any{"type": "agent_restart_ack", "request_id": first.RequestID, "status": "accepted"}); err != nil {
+		t.Fatal(err)
+	}
+	time.Sleep(25 * time.Millisecond)
+	_ = client.Close()
+
+	replacement, _, err := (&websocket.Dialer{TLSClientConfig: &tls.Config{InsecureSkipVerify: true}}).Dial(url, nil) //nolint:gosec -- test-only TLS.
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer replacement.Close()
+	if err := replacement.WriteJSON(map[string]any{"type": "authenticate", "agent_id": agentID, "credential": credential, "agent_name": "restart-agent"}); err != nil {
+		t.Fatal(err)
+	}
+	var accepted map[string]any
+	if err := replacement.ReadJSON(&accepted); err != nil || accepted["type"] != "auth_accepted" {
+		t.Fatalf("replacement did not reconnect: %#v / %v", accepted, err)
+	}
+	var status string
+	if err := controller.db.QueryRow("SELECT status FROM agent_restart_operations WHERE agent_id=?", agentID).Scan(&status); err != nil || status != "reconnected" {
+		t.Fatalf("restart state = %q / %v", status, err)
+	}
+}
+
+func TestRestartOfflineAgentIsRejectedAndLateCompletionCannotWin(t *testing.T) {
+	controller, _ := newRestartTestController(t)
+	defer controller.Close()
+	now := time.Now().UTC().Format(time.RFC3339Nano)
+	if _, err := controller.db.Exec("INSERT INTO managed_agents(agent_id, name, credential_hash, status, last_seen_at) VALUES ('offline', 'offline', 'hash', 'offline', ?)", now); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := controller.RestartAgent("offline"); err == nil {
+		t.Fatal("offline Agent restart must be rejected")
+	}
+	if _, err := controller.db.Exec("INSERT INTO task_runs(task_id, agent_id, agent_name, instruction, status, created_at, timeout_seconds) VALUES ('cancelled-late', 'offline', 'offline', 'test', 'cancelling', ?, 30)", now); err != nil {
+		t.Fatal(err)
+	}
+	controller.applyTaskEvent("offline", map[string]any{"task_id": "cancelled-late", "status": "completed", "output": "late"})
+	var taskStatus string
+	if err := controller.db.QueryRow("SELECT status FROM task_runs WHERE task_id='cancelled-late'").Scan(&taskStatus); err != nil || taskStatus != "cancelling" {
+		t.Fatalf("late completion overwrote cancellation: %q / %v", taskStatus, err)
+	}
+	controller.applyTaskEvent("offline", map[string]any{"task_id": "cancelled-late", "status": "cancelled", "message": "已取消"})
+	if err := controller.db.QueryRow("SELECT status FROM task_runs WHERE task_id='cancelled-late'").Scan(&taskStatus); err != nil || taskStatus != "cancelled" {
+		t.Fatalf("cancel acknowledgement missing: %q / %v", taskStatus, err)
+	}
+}
diff --git a/controller/internal/controlplane/skills.go b/controller/internal/controlplane/skills.go
index b3643c4..af106ca 100644
--- a/controller/internal/controlplane/skills.go
+++ b/controller/internal/controlplane/skills.go
@@ -15,7 +15,7 @@ import (
 // same limit so content is never silently truncated after deployment.
 const maxSkillContentBytes = 50000
 
-// Skill is a safe Markdown-only capability guide owned by the controller.
+// Skill is a Markdown-only capability guide owned by the controller.
 type Skill struct {
 	ID        string `json:"id"`
 	Name      string `json:"name"`
@@ -114,7 +114,7 @@ func (c *ControlPlane) SaveSkill(input SkillInput) (Skill, error) {
 		if _, err = c.db.Exec("INSERT INTO skill_packages(skill_id, name, version, content, checksum, updated_at) VALUES (?, ?, ?, ?, ?, ?)", skillID, name, version, content, checksum, now); err != nil {
 			return Skill{}, fmt.Errorf("无法保存 Skill: %w", err)
 		}
-		c.recordAudit("skill.package_saved", "controller", "skill:"+skillID, "completed", "已保存经哈希校验的 Skill 包。")
+		c.recordAudit("skill.package_saved", "controller", "skill:"+skillID, "completed", "已保存 Skill 包。")
 	case err != nil:
 		return Skill{}, fmt.Errorf("无法读取 Skill: %w", err)
 	default:
diff --git a/mulit_agent/src/mulit_agent_app/application/agent_service.py b/mulit_agent/src/mulit_agent_app/application/agent_service.py
index 64528e3..e8d06eb 100644
--- a/mulit_agent/src/mulit_agent_app/application/agent_service.py
+++ b/mulit_agent/src/mulit_agent_app/application/agent_service.py
@@ -2,12 +2,11 @@
 
 from __future__ import annotations
 
-import hashlib
 import json
 import re
 import shutil
 import tempfile
-from collections.abc import Iterable
+from collections.abc import Callable, Iterable
 from pathlib import Path
 from threading import Lock, Thread
 from urllib.parse import urlparse
@@ -18,6 +17,7 @@ from mulit_agent_app.infrastructure.controller_session import ControllerSession
 from mulit_agent_app.infrastructure.credential_store import CredentialStore
 from mulit_agent_app.infrastructure.database import AgentDatabase
 from mulit_agent_app.infrastructure.mcp_client import McpClient, McpClientError, open_client
+from mulit_agent_app.infrastructure.process_restart import start_replacement
 from mulit_agent_app.infrastructure.provider_factory import build_provider
 from mulit_agent_app.infrastructure.tool_registry import McpEndpoint
 
@@ -106,7 +106,7 @@ def _validated_tool_arguments(value: object) -> list[str]:
 
 
 class AgentService:
-    def __init__(self, config: AgentConfig) -> None:
+    def __init__(self, config: AgentConfig, restart_process: Callable[[], bool] = start_replacement, on_restarted: Callable[[], None] | None = None) -> None:
         self._config = config
         self._credential_store = CredentialStore()
         self.session = ControllerSession(config, self._credential_store)
@@ -114,11 +114,16 @@ class AgentService:
         self.task_executor = TaskExecutor(self._database)
         self._mcp_lock = Lock()
         self._mcp_sequence = 0
+        self._restart_lock = Lock()
+        self._restart_request_id = ""
+        self._restart_process = restart_process
+        self._on_restarted = on_restarted
         self._restore_controller_llm_config()
         self._restore_skill_context()
         self._restore_mcp_tools()
         self.session.task_dispatched.connect(self.task_executor.submit)
         self.session.task_cancelled.connect(self.task_executor.cancel)
+        self.session.restart_requested.connect(self.restart)
         self.session.config_snapshot_received.connect(self.apply_config_snapshot)
         self.session.skill_snapshot_received.connect(self.apply_skill_snapshot)
         self.session.mcp_snapshot_received.connect(self.apply_mcp_snapshot)
@@ -168,11 +173,48 @@ class AgentService:
     def reconnect(self) -> None:
         self.session.reconnect()
 
+    def set_restart_callback(self, callback: Callable[[], None]) -> None:
+        """Install the UI shutdown callback used after a replacement starts."""
+        self._on_restarted = callback
+
     def close(self) -> None:
         """Stop the session and release every deployed MCP connection."""
         self.session.close()
         self.task_executor.close()
 
+    def restart(self, request_id: object) -> None:
+        """Perform an authenticated, idempotent local restart sequence."""
+        if not isinstance(request_id, str) or not request_id:
+            return
+        with self._restart_lock:
+            if self._restart_request_id:
+                # The controller may retry the same message while the first is
+                # in flight.  Acknowledge it without spawning another process.
+                self.session.send_restart_ack(request_id, "restarting")
+                return
+            self._restart_request_id = request_id
+        Thread(target=self._restart_worker, args=(request_id,), name="agent-restart", daemon=True).start()
+
+    def _restart_worker(self, request_id: str) -> None:
+        reason = "Agent 正在重启，任务已中止。"
+        self.task_executor.cancel_all(reason)
+        # Cancellation events are durable in the outbox; make a best effort to
+        # deliver them before the old socket is closed.
+        self.session.send_restart_ack(request_id, "accepted")
+        self.task_executor.close()
+        self.session.close()
+        if not self._restart_process():
+            # The old process remains alive. Reconnect it so the durable failure
+            # acknowledgement reaches the controller instead of leaving a
+            # permanently ambiguous "重启中" state.
+            self.session.resume_after_failed_restart()
+            self.session.send_restart_ack(request_id, "failed", "无法启动替代 Agent 进程。")
+            with self._restart_lock:
+                self._restart_request_id = ""
+            return
+        if self._on_restarted is not None:
+            self._on_restarted()
+
     def apply_config_snapshot(self, payload: object) -> None:
         if not isinstance(payload, dict):
             return
@@ -235,12 +277,11 @@ class AgentService:
                 name = self._required_text(item, "name")
                 version = self._required_text(item, "version")
                 content = self._required_text(item, "content")
-                checksum = self._required_text(item, "checksum")
-                if len(content) > MAX_SKILL_CONTENT_CHARS or not re.fullmatch(r"[a-f0-9]{64}", checksum):
-                    raise ValueError("Skill 内容或校验值无效。")
-                actual_checksum = hashlib.sha256(content.encode("utf-8")).hexdigest()
-                if actual_checksum != checksum:
-                    raise ValueError("Skill 校验失败。")
+                if len(content) > MAX_SKILL_CONTENT_CHARS:
+                    raise ValueError("Skill 内容过长。")
+                # Skill 内容以控制端本次快照为准，直接原子覆盖本地文件。
+                # 不校验 checksum：本地允许人工调整，且控制端下次推送应能
+                # 无条件覆盖，而不因换行或本地修改而拒绝新内容。
                 destination = self._config.data_dir / "skills" / skill_id
                 destination.mkdir(parents=True, exist_ok=True)
                 target = destination / "SKILL.md"
@@ -250,7 +291,6 @@ class AgentService:
                 temporary_path.replace(target)
                 self._database.set_meta(f"skill.{skill_id}.name", name[:80])
                 self._database.set_meta(f"skill.{skill_id}.version", version)
-                self._database.set_meta(f"skill.{skill_id}.checksum", checksum)
             except (OSError, RuntimeError, ValueError) as error:
                 if isinstance(skill_id, str):
                     self.session.send_skill_ack(skill_id, "failed", str(error))
diff --git a/mulit_agent/src/mulit_agent_app/application/task_executor.py b/mulit_agent/src/mulit_agent_app/application/task_executor.py
index ef19bbe..72b19ee 100644
--- a/mulit_agent/src/mulit_agent_app/application/task_executor.py
+++ b/mulit_agent/src/mulit_agent_app/application/task_executor.py
@@ -150,6 +150,7 @@ class TaskExecutor:
             self._full_access_tool,
             tuple(endpoint.bridge for endpoint in endpoints),
             on_progress=self._report_progress(task_id),
+            is_active=lambda: self._database.task_status(task_id) == "running",
         )
 
         try:
@@ -205,10 +206,26 @@ class TaskExecutor:
         return report
 
     def cancel(self, task_id: str, reason: str = "任务已由总控取消。") -> None:
-        if self._database.task_status(task_id) not in {"accepted", "running"}:
-            return
-        self._full_access_tool.cancel(task_id)
-        self._finish(task_id, "cancelled", reason)
+        with self._lock:
+            current = self._database.task_status(task_id)
+            if current is None:
+                # Dispatch and cancel may cross on a reconnect.  The controller
+                # owns the task record, so explicitly acknowledge its terminal
+                # cancellation even when this Agent has not accepted it yet.
+                self._database.cancel_unreceived_task(task_id, reason, self._now())
+                self.task_event.emit({"task_id": task_id, "status": "cancelled", "message": reason})
+                return
+            if current not in {"accepted", "running"}:
+                return
+            self._full_access_tool.cancel(task_id)
+            self._finish(task_id, "cancelled", reason)
+
+    def cancel_all(self, reason: str = "Agent 正在重启，任务已中止。") -> list[str]:
+        """End every local task before lifecycle resources are closed."""
+        task_ids = self._database.active_task_ids()
+        for task_id in task_ids:
+            self.cancel(task_id, reason)
+        return task_ids
 
     def _timeout(self, task_id: str) -> None:
         if self._database.task_status(task_id) not in {"accepted", "running"}:
@@ -224,11 +241,16 @@ class TaskExecutor:
         error: str | None = None,
         execution_source: str | None = None,
     ) -> None:
-        self._cancel_timers(task_id)
-        self._cancel_deadline_timer(task_id)
         with self._lock:
+            # Terminal state is write-once.  Serialising the status check and
+            # write also prevents a late model result racing a user cancel.
+            current = self._database.task_status(task_id)
+            if current in {"completed", "failed", "cancelled", "timed_out"}:
+                return
+            self._cancel_timers(task_id)
+            self._cancel_deadline_timer(task_id)
             self._deadline_windows.pop(task_id, None)
-        self._database.update_task(task_id, status, message, self._now(), output=output, error=error, execution_source=execution_source)
+            self._database.update_task(task_id, status, message, self._now(), output=output, error=error, execution_source=execution_source)
         event: dict[str, object] = {"task_id": task_id, "status": status, "message": message}
         if output is not None:
             event["output"] = output
diff --git a/mulit_agent/src/mulit_agent_app/infrastructure/builtin_tools.py b/mulit_agent/src/mulit_agent_app/infrastructure/builtin_tools.py
index efc06c0..cbf42da 100644
--- a/mulit_agent/src/mulit_agent_app/infrastructure/builtin_tools.py
+++ b/mulit_agent/src/mulit_agent_app/infrastructure/builtin_tools.py
@@ -95,6 +95,9 @@ def _run_powershell(script: str, environment: dict[str, str] | None = None) -> s
         timeout=_POWERSHELL_TIMEOUT,
         check=False,
         env={**os.environ, **(environment or {})},
+        # Built-in desktop helpers also use PowerShell.  They must remain
+        # invisible just like model-requested full-access commands.
+        creationflags=getattr(subprocess, "CREATE_NO_WINDOW", 0),
     )
     output = "\n".join(part for part in ((process.stdout or "").strip(), (process.stderr or "").strip()) if part)
     if process.returncode:
diff --git a/mulit_agent/src/mulit_agent_app/infrastructure/controller_session.py b/mulit_agent/src/mulit_agent_app/infrastructure/controller_session.py
index 1e8edf0..299aa22 100644
--- a/mulit_agent/src/mulit_agent_app/infrastructure/controller_session.py
+++ b/mulit_agent/src/mulit_agent_app/infrastructure/controller_session.py
@@ -54,6 +54,7 @@ class ControllerSession:
         self.config_snapshot_received = EventHook()
         self.skill_snapshot_received = EventHook()
         self.mcp_snapshot_received = EventHook()
+        self.restart_requested = EventHook()
 
     def update_agent_name(self, agent_name: str) -> None:
         self._config = replace(self._config, agent_name=agent_name)
@@ -90,6 +91,12 @@ class ControllerSession:
         self._cancel_reconnect()
         self._close_socket()
 
+    def resume_after_failed_restart(self) -> None:
+        """Re-open the authenticated session when no replacement was created."""
+        self._stopping.clear()
+        self._reconnect_disabled = False
+        self.reconnect()
+
     def _close_socket(self) -> None:
         self._heartbeat_stop.set()
         with self._socket_lock:
@@ -177,6 +184,8 @@ class ControllerSession:
             self._handle_task_dispatch(payload)
         elif message_type == "task_cancel":
             self._handle_task_cancel(payload)
+        elif message_type == "agent_restart":
+            self._handle_restart(payload)
         elif message_type == "config_snapshot":
             self.config_snapshot_received.emit(payload)
         elif message_type == "skill_snapshot":
@@ -335,6 +344,13 @@ class ControllerSession:
             return
         self.task_cancelled.emit(task_id, reason if isinstance(reason, str) else "任务已由总控取消。")
 
+    def _handle_restart(self, payload: dict[str, object]) -> None:
+        request_id = payload.get("request_id")
+        if not isinstance(request_id, str) or not request_id.strip() or len(request_id) > 128:
+            self.connection_error.emit("总控下发了无效重启指令，已忽略。")
+            return
+        self.restart_requested.emit(request_id)
+
     def send_task_event(self, event: object) -> None:
         if not isinstance(event, dict):
             return
@@ -369,6 +385,16 @@ class ControllerSession:
             payload["error"] = error[:1000]
         self._send(payload)
 
+    def send_restart_ack(self, request_id: str, status: str, error: str = "") -> None:
+        payload: dict[str, object] = {"type": "agent_restart_ack", "request_id": request_id, "status": status}
+        if error:
+            payload["error"] = error[:1000]
+        if not self._send(payload):
+            # A failed replacement attempt happens after the old socket was
+            # deliberately closed.  Preserve the acknowledgement and replay it
+            # after this still-running Agent reconnects.
+            self._database.enqueue_event(json.dumps(payload, ensure_ascii=False, separators=(",", ":")), datetime.now(UTC).isoformat())
+
     def _flush_pending_task_events(self) -> None:
         for event_id, raw in self._database.pending_events():
             try:
diff --git a/mulit_agent/src/mulit_agent_app/infrastructure/database.py b/mulit_agent/src/mulit_agent_app/infrastructure/database.py
index f011834..475b65f 100644
--- a/mulit_agent/src/mulit_agent_app/infrastructure/database.py
+++ b/mulit_agent/src/mulit_agent_app/infrastructure/database.py
@@ -71,6 +71,14 @@ class AgentDatabase:
             row = connection.execute("SELECT status FROM task_runs WHERE task_id = ?", (task_id,)).fetchone()
         return str(row[0]) if row is not None else None
 
+    def active_task_ids(self) -> list[str]:
+        """Return only work which can still be stopped by this Agent."""
+        with sqlite3.connect(self._database_path) as connection:
+            rows = connection.execute(
+                "SELECT task_id FROM task_runs WHERE status IN ('accepted', 'running')"
+            ).fetchall()
+        return [str(row[0]) for row in rows]
+
     def create_task(self, task_id: str, instruction: str, received_at: str) -> None:
         with sqlite3.connect(self._database_path) as connection:
             connection.execute(
@@ -79,6 +87,15 @@ class AgentDatabase:
             )
             self._record_event(connection, task_id, "accepted", "任务已接收。", received_at)
 
+    def cancel_unreceived_task(self, task_id: str, reason: str, occurred_at: str) -> None:
+        """Store a cancellation tombstone when cancel overtakes dispatch."""
+        with sqlite3.connect(self._database_path) as connection:
+            connection.execute(
+                "INSERT OR IGNORE INTO task_runs(task_id, instruction, status, received_at, finished_at, error) VALUES (?, '', 'cancelled', ?, ?, ?)",
+                (task_id, occurred_at, occurred_at, reason),
+            )
+            self._record_event(connection, task_id, "cancelled", reason, occurred_at)
+
     def get_meta(self, key: str) -> str:
         with sqlite3.connect(self._database_path) as connection:
             row = connection.execute("SELECT value FROM application_meta WHERE key = ?", (key,)).fetchone()
diff --git a/mulit_agent/src/mulit_agent_app/infrastructure/full_access_tool.py b/mulit_agent/src/mulit_agent_app/infrastructure/full_access_tool.py
index cc86b0f..619ffe2 100644
--- a/mulit_agent/src/mulit_agent_app/infrastructure/full_access_tool.py
+++ b/mulit_agent/src/mulit_agent_app/infrastructure/full_access_tool.py
@@ -53,6 +53,10 @@ class FullAccessTool:
             text=True,
             encoding="utf-8",
             errors="replace",
+            # The Agent is a desktop application, not a terminal application.
+            # Keep model-requested PowerShell work in the background so each
+            # tool call does not steal focus or flash a console window.
+            creationflags=getattr(subprocess, "CREATE_NO_WINDOW", 0),
         )
         with self._lock:
             self._processes[task_id] = process
diff --git a/mulit_agent/src/mulit_agent_app/infrastructure/openai_compatible_provider.py b/mulit_agent/src/mulit_agent_app/infrastructure/openai_compatible_provider.py
index b7b3a20..967b073 100644
--- a/mulit_agent/src/mulit_agent_app/infrastructure/openai_compatible_provider.py
+++ b/mulit_agent/src/mulit_agent_app/infrastructure/openai_compatible_provider.py
@@ -19,6 +19,11 @@ class ProviderError(RuntimeError):
 _THINK_BLOCK = re.compile(r"<think(?:\s[^>]*)?>.*?</think\s*>", re.IGNORECASE | re.DOTALL)
 _UNCLOSED_THINK = re.compile(r"<think(?:\s[^>]*)?>.*$", re.IGNORECASE | re.DOTALL)
 
+# Desktop UI workflows often require several observe → act → verify cycles.
+# Keep a finite guard against runaway model loops, while allowing complete
+# multi-step workflows such as image generation to finish.
+_MAX_TOOL_ROUNDS = 200
+
 
 def visible_model_output(content: str) -> str:
     """Remove provider reasoning traces before any task result leaves the Agent."""
@@ -66,8 +71,14 @@ class OpenAICompatibleProvider:
                 for spec in toolbox.specs()
             ]
             request_body["tool_choice"] = "auto"
-        for _ in range(12):
+        for _ in range(_MAX_TOOL_ROUNDS):
+            if toolbox is not None and not toolbox.is_active():
+                # A request in flight cannot always be aborted by its provider,
+                # but it must never lead to a further request or tool call.
+                return ""
             response_payload = self._request_json(request_body, api_key)
+            if toolbox is not None and not toolbox.is_active():
+                return ""
             try:
                 message = response_payload["choices"][0]["message"]
             except (KeyError, IndexError, TypeError) as error:
@@ -90,6 +101,8 @@ class OpenAICompatibleProvider:
                 }
             )
             for call in tool_calls:
+                if not toolbox.is_active():
+                    return ""
                 if not isinstance(call, dict):
                     continue
                 call_id = call.get("id")
diff --git a/mulit_agent/src/mulit_agent_app/infrastructure/process_restart.py b/mulit_agent/src/mulit_agent_app/infrastructure/process_restart.py
new file mode 100644
index 0000000..c3e8145
--- /dev/null
+++ b/mulit_agent/src/mulit_agent_app/infrastructure/process_restart.py
@@ -0,0 +1,36 @@
+"""Start a replacement Agent process without exposing a console window."""
+
+from __future__ import annotations
+
+import os
+import subprocess
+import sys
+from collections.abc import Callable
+from pathlib import Path
+
+
+def start_replacement(
+    popen: Callable[..., object] = subprocess.Popen,
+    executable: str | None = None,
+    arguments: list[str] | None = None,
+) -> bool:
+    """Return only after the replacement process has been created successfully.
+
+    The controller never calls this: it runs solely inside the authenticated
+    Agent.  ``CREATE_NO_WINDOW`` prevents a console flash on Windows.
+    """
+    program = executable or sys.executable
+    argv = [program, *(arguments if arguments is not None else sys.argv[1:])]
+    try:
+        popen(
+            argv,
+            cwd=str(Path.cwd()),
+            stdin=subprocess.DEVNULL,
+            stdout=subprocess.DEVNULL,
+            stderr=subprocess.DEVNULL,
+            close_fds=os.name != "nt",
+            creationflags=getattr(subprocess, "CREATE_NO_WINDOW", 0),
+        )
+    except OSError:
+        return False
+    return True
diff --git a/mulit_agent/src/mulit_agent_app/infrastructure/provider_factory.py b/mulit_agent/src/mulit_agent_app/infrastructure/provider_factory.py
index 4265726..a381513 100644
--- a/mulit_agent/src/mulit_agent_app/infrastructure/provider_factory.py
+++ b/mulit_agent/src/mulit_agent_app/infrastructure/provider_factory.py
@@ -14,7 +14,9 @@ from mulit_agent_app.infrastructure.openai_compatible_provider import (
 )
 from mulit_agent_app.infrastructure.tool_registry import as_toolbox, system_prompt
 
-_MAX_TOOL_ROUNDS = 12
+# Desktop UI workflows often require several observe → act → verify cycles.
+# This remains finite to protect against runaway model loops.
+_MAX_TOOL_ROUNDS = 200
 
 
 def build_provider(provider: str, base_url: str, model: str, credential_store: CredentialStore):
diff --git a/mulit_agent/src/mulit_agent_app/infrastructure/tool_registry.py b/mulit_agent/src/mulit_agent_app/infrastructure/tool_registry.py
index a7e81c7..e283554 100644
--- a/mulit_agent/src/mulit_agent_app/infrastructure/tool_registry.py
+++ b/mulit_agent/src/mulit_agent_app/infrastructure/tool_registry.py
@@ -127,10 +127,12 @@ class Toolbox:
         bridges: tuple[McpToolBridge, ...] = (),
         with_builtin: bool = True,
         on_progress: Callable[[str], None] | None = None,
+        is_active: Callable[[], bool] | None = None,
     ) -> None:
         self._task_id = task_id
         self._full_access = full_access
         self._on_progress = on_progress
+        self._is_active = is_active or (lambda: True)
         self.used_full_access = False
         self.used_mcp = False
         self.used_builtin = False
@@ -168,6 +170,10 @@ class Toolbox:
 
     def call(self, name: str, arguments: object) -> str:
         """Dispatch one model tool call. Tool failures never take down the task."""
+        # A model response may arrive after cancellation.  Do not turn that
+        # late response into another local/MCP side effect.
+        if not self.is_active():
+            return "任务已取消；忽略迟到的工具调用。"
         tool = self._tools.get(name)
         handler = self._handlers.get(name)
         if tool is None or handler is None:
@@ -184,6 +190,12 @@ class Toolbox:
         self._report(f"{name} 返回：{_summarise_result(result)}")
         return result
 
+    def is_active(self) -> bool:
+        try:
+            return bool(self._is_active())
+        except Exception:  # noqa: BLE001 - fail closed around cancellation.
+            return False
+
     def _report(self, message: str) -> None:
         """Stream one step to the controller. Reporting must never break a task."""
         if self._on_progress is None:
diff --git a/mulit_agent/src/mulit_agent_app/main.py b/mulit_agent/src/mulit_agent_app/main.py
index bae90fe..34943a3 100644
--- a/mulit_agent/src/mulit_agent_app/main.py
+++ b/mulit_agent/src/mulit_agent_app/main.py
@@ -12,6 +12,8 @@ from mulit_agent_app.presentation.tray import AgentTray
 def main() -> int:
     config = AgentConfig.load()
     configure_logging(config.log_path)
+    # The service starts the replacement first; only then does this callback
+    # close the old UI/process.  Tk work stays on its own event loop.
     service = AgentService(config)
     window = ConnectWindow(service)
     tray = AgentTray()
@@ -25,6 +27,8 @@ def main() -> int:
         tray.stop()
         window.root.after(0, window.root.destroy)
 
+    service.set_restart_callback(quit_application)
+
     tray.open_settings_requested.connect(lambda: window.root.after(0, window.show))
     tray.reconnect_requested.connect(service.reconnect)
     tray.quit_requested.connect(quit_application)
diff --git a/mulit_agent/tests/test_agent_restart.py b/mulit_agent/tests/test_agent_restart.py
new file mode 100644
index 0000000..4b70b47
--- /dev/null
+++ b/mulit_agent/tests/test_agent_restart.py
@@ -0,0 +1,34 @@
+from threading import Event
+from time import monotonic, sleep
+
+from mulit_agent_app.application.agent_service import AgentService
+from mulit_agent_app.config import AgentConfig
+
+
+def _config(tmp_path) -> AgentConfig:
+    data = tmp_path / "agent"
+    return AgentConfig(data, data / "settings.json", data / "agent.sqlite3", data / "agent.log", agent_name="restart-test")
+
+
+def test_duplicate_restart_requests_start_only_one_replacement_and_cancel_work(tmp_path):
+    started, old_exited = [], Event()
+    service = AgentService(_config(tmp_path), restart_process=lambda: started.append("replacement") or True, on_restarted=old_exited.set)
+    service.task_executor.submit("restart-task", "应因 Agent 重启中止")
+
+    service.restart("restart-1")
+    service.restart("restart-1")
+
+    assert old_exited.wait(2)
+    assert started == ["replacement"]
+    assert service.task_executor._database.task_status("restart-task") == "cancelled"
+
+
+def test_restart_failure_does_not_exit_old_agent_or_start_another_process(tmp_path):
+    old_exited = Event()
+    service = AgentService(_config(tmp_path), restart_process=lambda: False, on_restarted=old_exited.set)
+    service.restart("restart-fail")
+    deadline = monotonic() + 2
+    while service._restart_request_id and monotonic() < deadline:
+        sleep(0.01)
+    assert not old_exited.is_set()
+    assert service._restart_request_id == ""
diff --git a/mulit_agent/tests/test_agent_service_skills.py b/mulit_agent/tests/test_agent_service_skills.py
index 0725678..b3dc5ff 100644
--- a/mulit_agent/tests/test_agent_service_skills.py
+++ b/mulit_agent/tests/test_agent_service_skills.py
@@ -1,5 +1,3 @@
-import hashlib
-
 from mulit_agent_app.application.agent_service import AgentService
 from mulit_agent_app.config import AgentConfig
 from mulit_agent_app.infrastructure.database import AgentDatabase
@@ -40,7 +38,8 @@ def test_chinese_skill_name_installs_and_empty_snapshot_removes_it(tmp_path):
                     "name": "图生图工具",
                     "version": "0.3.2",
                     "content": content,
-                    "checksum": hashlib.sha256(content.encode("utf-8")).hexdigest(),
+                    # 校验值只供协议兼容；Agent 必须接受控制端内容并覆盖本地文件。
+                    "checksum": "outdated-checksum",
                 }
             ]
         }
@@ -50,6 +49,27 @@ def test_chinese_skill_name_installs_and_empty_snapshot_removes_it(tmp_path):
     assert service._database.get_meta(f"skill.{skill_id}.name") == "图生图工具"
     assert service.session.acks == [(skill_id, "applied", "")]
 
+    # 人工本地改动不是错误；下一份控制端快照应直接覆盖它。
+    target = config.data_dir / "skills" / skill_id / "SKILL.md"
+    target.write_text("# 本地临时修改", encoding="utf-8")
+    replacement = "# 总控下发版本\n\n应覆盖本地内容。"
+    service.apply_skill_snapshot(
+        {
+            "skills": [
+                {
+                    "id": skill_id,
+                    "name": "图生图工具",
+                    "version": "0.3.3",
+                    "content": replacement,
+                    "checksum": "another-outdated-checksum",
+                }
+            ]
+        }
+    )
+
+    assert target.read_text(encoding="utf-8") == replacement
+    assert service.session.acks[-1] == (skill_id, "applied", "")
+
     service.apply_skill_snapshot({"skills": []})
 
     assert not (config.data_dir / "skills" / skill_id).exists()
diff --git a/mulit_agent/tests/test_builtin_tools.py b/mulit_agent/tests/test_builtin_tools.py
index 7e3eda2..44ce7cb 100644
--- a/mulit_agent/tests/test_builtin_tools.py
+++ b/mulit_agent/tests/test_builtin_tools.py
@@ -15,6 +15,25 @@ from mulit_agent_app.infrastructure.tool_registry import Toolbox
 _HANDLERS = {spec.name: spec.handler for spec in specs()}
 
 
+def test_builtin_powershell_helpers_hide_the_console_window(monkeypatch):
+    calls: list[dict[str, object]] = []
+
+    class Result:
+        stdout = "done"
+        stderr = ""
+        returncode = 0
+
+    def fake_run(*_args, **kwargs):
+        calls.append(kwargs)
+        return Result()
+
+    monkeypatch.setattr(builtin_tools.subprocess, "run", fake_run)
+    monkeypatch.setattr(builtin_tools.subprocess, "CREATE_NO_WINDOW", 42, raising=False)
+
+    assert builtin_tools._run_powershell("Write-Output done") == "done"
+    assert calls[0]["creationflags"] == 42
+
+
 def test_preset_tool_set_covers_file_open_time_process_screenshot_and_clipboard():
     names = set(_HANDLERS)
     assert {
diff --git a/mulit_agent/tests/test_full_access_tool.py b/mulit_agent/tests/test_full_access_tool.py
index dd83b71..557e084 100644
--- a/mulit_agent/tests/test_full_access_tool.py
+++ b/mulit_agent/tests/test_full_access_tool.py
@@ -18,6 +18,26 @@ def test_full_access_tool_redacts_secrets_from_command_output():
     assert "[已脱敏]" in result
 
 
+def test_full_access_tool_hides_the_powershell_window(monkeypatch):
+    calls: list[dict[str, object]] = []
+
+    class Process:
+        returncode = 0
+
+        def communicate(self):
+            return "done", ""
+
+    def fake_popen(*_args, **kwargs):
+        calls.append(kwargs)
+        return Process()
+
+    monkeypatch.setattr("mulit_agent_app.infrastructure.full_access_tool.subprocess.Popen", fake_popen)
+    monkeypatch.setattr("mulit_agent_app.infrastructure.full_access_tool.subprocess.CREATE_NO_WINDOW", 42, raising=False)
+
+    assert FullAccessTool().execute("task-hidden", "Write-Output done") == "done"
+    assert calls[0]["creationflags"] == 42
+
+
 def test_full_access_tool_stages_and_removes_task_attachments():
     from mulit_agent_app.domain.attachments import TaskAttachment
 
diff --git a/mulit_agent/tests/test_openai_compatible_provider.py b/mulit_agent/tests/test_openai_compatible_provider.py
index a07b942..3fea579 100644
--- a/mulit_agent/tests/test_openai_compatible_provider.py
+++ b/mulit_agent/tests/test_openai_compatible_provider.py
@@ -1,6 +1,12 @@
 import json
 
-from mulit_agent_app.infrastructure.openai_compatible_provider import OpenAICompatibleProvider
+from mulit_agent_app.infrastructure.openai_compatible_provider import (
+    _MAX_TOOL_ROUNDS,
+    OpenAICompatibleProvider,
+)
+from mulit_agent_app.infrastructure.provider_factory import (
+    _MAX_TOOL_ROUNDS as PROVIDER_FACTORY_MAX_TOOL_ROUNDS,
+)
 
 
 class _Credentials:
@@ -19,6 +25,11 @@ class _Response:
         return json.dumps({"choices": [{"message": {"content": "  模型结果  "}}]}).encode("utf-8")
 
 
+def test_all_provider_tool_round_limits_allow_long_desktop_workflows():
+    assert _MAX_TOOL_ROUNDS == 200
+    assert PROVIDER_FACTORY_MAX_TOOL_ROUNDS == 200
+
+
 def test_provider_posts_a_chat_completion(monkeypatch, tmp_path):
     captured = {}
 
diff --git a/mulit_agent/tests/test_task_executor.py b/mulit_agent/tests/test_task_executor.py
index dfd1694..4b83080 100644
--- a/mulit_agent/tests/test_task_executor.py
+++ b/mulit_agent/tests/test_task_executor.py
@@ -266,6 +266,31 @@ def test_executor_ignores_progress_after_the_task_ended(tmp_path):
     assert len(events) == reported, "结束后的任务不应再产生中间步骤"
 
 
+def test_late_model_result_cannot_overwrite_a_cancelled_terminal_state(tmp_path):
+    database = AgentDatabase(tmp_path / "agent.sqlite3")
+    started, release, returned, cancelled = Event(), Event(), Event(), Event()
+
+    def runner(_instruction: str) -> str:
+        started.set()
+        assert release.wait(2)
+        returned.set()
+        return "这是取消后才返回的模型结果"
+
+    executor = TaskExecutor(database, runner)
+    executor.task_event.connect(
+        lambda event: cancelled.set()
+        if isinstance(event, dict) and event.get("status") == "cancelled"
+        else None
+    )
+    executor.submit("late-result", "等待模型返回")
+    assert started.wait(2)
+    executor.cancel("late-result", "测试取消")
+    assert cancelled.wait(1)
+    release.set()
+    assert returned.wait(1)
+    assert database.task_status("late-result") == "cancelled"
+
+
 def test_executor_passes_the_conversation_history_to_the_provider(tmp_path):
     database = AgentDatabase(tmp_path / "agent.sqlite3")
     received: list[tuple[tuple[str, str], ...]] = []