feat(auth): 添加异步密码哈希和验证函数

refactor(relay): 复用HTTP客户端和请求体序列化结果

feat(kernel): 添加获取单个审批记录的方法

fix(store): 改进SaaS连接错误分类和降级处理

docs: 更新审计文档和系统架构文档

refactor(prompt): 优化SQL查询参数化绑定

refactor(migration): 使用静态SQL和COALESCE更新配置项

feat(commands): 添加审批执行状态追踪和事件通知

chore: 更新启动脚本以支持Admin后台

fix(auth-guard): 优化授权状态管理和错误处理

refactor(db): 使用异步密码哈希函数

refactor(totp): 使用异步密码验证函数

style: 清理无用文件和注释

docs: 更新功能全景和审计文档

refactor(service): 优化HTTP客户端重用和请求处理

fix(connection): 改进SaaS不可用时的降级处理

refactor(handlers): 使用异步密码验证函数

chore: 更新依赖和工具链配置

desktop/src-tauri/src/intelligence/extraction_adapter.rs

@@ -0,0 +1,321 @@
+//! Extraction Adapter - Bridges zclaw_growth::LlmDriverForExtraction with the Kernel's LlmDriver
+//!
+//! Implements the `LlmDriverForExtraction` trait by delegating to the Kernel's
+//! `zclaw_runtime::driver::LlmDriver`, which already handles provider-specific
+//! API calls (OpenAI, Anthropic, Gemini, etc.).
+//!
+//! This enables the Growth system's MemoryExtractor to call the LLM for memory
+//! extraction from conversations.
+
+use std::sync::Arc;
+use zclaw_growth::extractor::{LlmDriverForExtraction, prompts};
+use zclaw_growth::types::{ExtractedMemory, MemoryType};
+use zclaw_runtime::driver::{CompletionRequest, ContentBlock, LlmDriver};
+use zclaw_types::{Message, Result, SessionId};
+
+/// Adapter that wraps the Kernel's `LlmDriver` to implement `LlmDriverForExtraction`.
+///
+/// The adapter translates extraction requests into completion requests that the
+/// Kernel's LLM driver can process, then parses the structured JSON response
+/// back into `ExtractedMemory` objects.
+pub struct TauriExtractionDriver {
+    driver: Arc<dyn LlmDriver>,
+    model: String,
+}
+
+impl TauriExtractionDriver {
+    /// Create a new extraction driver wrapping the given LLM driver.
+    ///
+    /// The `model` parameter specifies which model to use for extraction calls.
+    pub fn new(driver: Arc<dyn LlmDriver>, model: String) -> Self {
+        Self { driver, model }
+    }
+
+    /// Build a completion request from the extraction prompt and conversation messages.
+    fn build_request(
+        &self,
+        messages: &[Message],
+        extraction_type: MemoryType,
+    ) -> CompletionRequest {
+        let extraction_prompt = prompts::get_extraction_prompt(extraction_type);
+
+        // Format conversation for the prompt
+        // Message is an enum with variants: User{content}, Assistant{content, thinking},
+        // System{content}, ToolUse{...}, ToolResult{...}
+        let conversation_text = messages
+            .iter()
+            .filter_map(|msg| {
+                match msg {
+                    Message::User { content } => {
+                        Some(format!("[User]: {}", content))
+                    }
+                    Message::Assistant { content, .. } => {
+                        Some(format!("[Assistant]: {}", content))
+                    }
+                    Message::System { content } => {
+                        Some(format!("[System]: {}", content))
+                    }
+                    // Skip tool use/result messages -- not relevant for memory extraction
+                    Message::ToolUse { .. } | Message::ToolResult { .. } => None,
+                }
+            })
+            .collect::<Vec<_>>()
+            .join("\n\n");
+
+        let full_prompt = format!("{}{}", extraction_prompt, conversation_text);
+
+        CompletionRequest {
+            model: self.model.clone(),
+            system: Some(
+                "You are a memory extraction assistant. Analyze conversations and extract \
+                structured memories in valid JSON format. Always respond with valid JSON only, \
+                no additional text or markdown formatting."
+                    .to_string(),
+            ),
+            messages: vec![Message::user(full_prompt)],
+            tools: Vec::new(),
+            max_tokens: Some(2000),
+            temperature: Some(0.3),
+            stop: Vec::new(),
+            stream: false,
+        }
+    }
+
+    /// Parse the LLM response text into a list of extracted memories.
+    fn parse_response(
+        &self,
+        response_text: &str,
+        extraction_type: MemoryType,
+    ) -> Vec<ExtractedMemory> {
+        // Strip markdown code fences if present
+        let cleaned = response_text
+            .trim()
+            .trim_start_matches("```json")
+            .trim_start_matches("```")
+            .trim_end_matches("```")
+            .trim();
+
+        // Extract the JSON array from the response
+        let json_str = match (cleaned.find('['), cleaned.rfind(']')) {
+            (Some(start), Some(end)) => &cleaned[start..=end],
+            _ => {
+                tracing::warn!(
+                    "[TauriExtractionDriver] No JSON array found in LLM response"
+                );
+                return Vec::new();
+            }
+        };
+
+        let raw_items: Vec<serde_json::Value> = match serde_json::from_str(json_str) {
+            Ok(items) => items,
+            Err(e) => {
+                tracing::warn!(
+                    "[TauriExtractionDriver] Failed to parse extraction JSON: {}",
+                    e
+                );
+                return Vec::new();
+            }
+        };
+
+        raw_items
+            .into_iter()
+            .filter_map(|item| self.parse_memory_item(&item, extraction_type))
+            .collect()
+    }
+
+    /// Parse a single memory item from JSON.
+    fn parse_memory_item(
+        &self,
+        value: &serde_json::Value,
+        fallback_type: MemoryType,
+    ) -> Option<ExtractedMemory> {
+        let content = value.get("content")?.as_str()?.to_string();
+        let category = value
+            .get("category")
+            .and_then(|v| v.as_str())
+            .unwrap_or("unknown")
+            .to_string();
+
+        let confidence = value
+            .get("confidence")
+            .and_then(|v| v.as_f64())
+            .unwrap_or(0.7) as f32;
+
+        let keywords = value
+            .get("keywords")
+            .and_then(|v| v.as_array())
+            .map(|arr| {
+                arr.iter()
+                    .filter_map(|v| v.as_str().map(String::from))
+                    .collect()
+            })
+            .unwrap_or_default();
+
+        Some(
+            ExtractedMemory::new(fallback_type, category, content, SessionId::new())
+                .with_confidence(confidence)
+                .with_keywords(keywords),
+        )
+    }
+}
+
+#[async_trait::async_trait]
+impl LlmDriverForExtraction for TauriExtractionDriver {
+    async fn extract_memories(
+        &self,
+        messages: &[Message],
+        extraction_type: MemoryType,
+    ) -> Result<Vec<ExtractedMemory>> {
+        let type_name = format!("{}", extraction_type);
+        tracing::debug!(
+            "[TauriExtractionDriver] Extracting {} memories from {} messages",
+            type_name,
+            messages.len()
+        );
+
+        // Skip extraction if there are too few messages
+        if messages.len() < 2 {
+            tracing::debug!(
+                "[TauriExtractionDriver] Too few messages ({}) for extraction, skipping",
+                messages.len()
+            );
+            return Ok(Vec::new());
+        }
+
+        let request = self.build_request(messages, extraction_type);
+
+        let response = self.driver.complete(request).await.map_err(|e| {
+            tracing::error!(
+                "[TauriExtractionDriver] LLM completion failed for {}: {}",
+                type_name,
+                e
+            );
+            e
+        })?;
+
+        // Extract text content from response
+        let response_text: String = response
+            .content
+            .into_iter()
+            .filter_map(|block| match block {
+                ContentBlock::Text { text } => Some(text),
+                _ => None,
+            })
+            .collect::<Vec<_>>()
+            .join("");
+
+        if response_text.is_empty() {
+            tracing::warn!(
+                "[TauriExtractionDriver] Empty response from LLM for {} extraction",
+                type_name
+            );
+            return Ok(Vec::new());
+        }
+
+        let memories = self.parse_response(&response_text, extraction_type);
+
+        tracing::info!(
+            "[TauriExtractionDriver] Extracted {} {} memories",
+            memories.len(),
+            type_name
+        );
+
+        Ok(memories)
+    }
+}
+
+/// Global extraction driver instance (lazy-initialized).
+static EXTRACTION_DRIVER: tokio::sync::OnceCell<Arc<TauriExtractionDriver>> =
+    tokio::sync::OnceCell::const_new();
+
+/// Configure the global extraction driver.
+///
+/// Call this during kernel initialization after the Kernel's LLM driver is available.
+pub fn configure_extraction_driver(driver: Arc<dyn LlmDriver>, model: String) {
+    let adapter = TauriExtractionDriver::new(driver, model);
+    let _ = EXTRACTION_DRIVER.set(Arc::new(adapter));
+    tracing::info!("[ExtractionAdapter] Extraction driver configured");
+}
+
+/// Check if the extraction driver is available.
+#[allow(dead_code)]
+pub fn is_extraction_driver_configured() -> bool {
+    EXTRACTION_DRIVER.get().is_some()
+}
+
+/// Get the global extraction driver.
+///
+/// Returns `None` if not yet configured via `configure_extraction_driver`.
+pub fn get_extraction_driver() -> Option<Arc<TauriExtractionDriver>> {
+    EXTRACTION_DRIVER.get().cloned()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn test_extraction_driver_not_configured_by_default() {
+        assert!(!is_extraction_driver_configured());
+    }
+
+    #[test]
+    fn test_parse_empty_response() {
+        // We cannot create a real LlmDriver easily in tests, so we test the
+        // parsing logic via a minimal helper.
+        struct DummyDriver;
+        impl TauriExtractionDriver {
+            fn parse_response_test(
+                &self,
+                response_text: &str,
+                extraction_type: MemoryType,
+            ) -> Vec<ExtractedMemory> {
+                self.parse_response(response_text, extraction_type)
+            }
+        }
+    }
+
+    #[test]
+    fn test_parse_valid_json_response() {
+        let response = r#"```json
+        [
+            {
+                "category": "communication-style",
+                "content": "User prefers concise replies",
+                "confidence": 0.9,
+                "keywords": ["concise", "style"]
+            },
+            {
+                "category": "language",
+                "content": "User prefers Chinese responses",
+                "confidence": 0.85,
+                "keywords": ["Chinese", "language"]
+            }
+        ]
+        ```"#;
+
+        // Verify the parsing logic works by manually simulating it
+        let cleaned = response
+            .trim()
+            .trim_start_matches("```json")
+            .trim_start_matches("```")
+            .trim_end_matches("```")
+            .trim();
+
+        let json_str = &cleaned[cleaned.find('[').unwrap()..=cleaned.rfind(']').unwrap()];
+        let items: Vec<serde_json::Value> = serde_json::from_str(json_str).unwrap();
+        assert_eq!(items.len(), 2);
+        assert_eq!(
+            items[0].get("category").unwrap().as_str().unwrap(),
+            "communication-style"
+        );
+    }
+
+    #[test]
+    fn test_parse_no_json_array() {
+        let response = "No memories could be extracted from this conversation.";
+        let has_array =
+            response.find('[').is_some() && response.rfind(']').is_some();
+        assert!(!has_array);
+    }
+}

desktop/src-tauri/src/intelligence/mod.rs

@@ -31,6 +31,7 @@ pub mod compactor;
 pub mod reflection;
 pub mod identity;
 pub mod validation;
+pub mod extraction_adapter;
 
 // Re-export main types for convenience
 pub use heartbeat::HeartbeatEngineState;
@@ -40,3 +41,13 @@ pub use reflection::{
 pub use identity::{
     AgentIdentityManager, IdentityManagerState,
 };
+
+// Suppress dead-code warnings for extraction adapter accessors — they are
+// consumed externally via full path (crate::intelligence::extraction_adapter::*).
+#[allow(unused_imports)]
+use extraction_adapter::{
+    configure_extraction_driver as _,
+    is_extraction_driver_configured as _,
+    get_extraction_driver as _,
+    TauriExtractionDriver as _,
+};

desktop/src-tauri/src/kernel_commands.rs

@@ -215,6 +215,24 @@ pub async fn kernel_init(
 
     let agent_count = kernel.list_agents().len();
 
+    // Configure extraction driver so the Growth system can call LLM for memory extraction
+    let driver = kernel.driver();
+    crate::intelligence::extraction_adapter::configure_extraction_driver(
+        driver.clone(),
+        model.clone(),
+    );
+
+    // Configure summary driver so the Growth system can generate L0/L1 summaries
+    if let Some(api_key) = config_request.as_ref().and_then(|r| r.api_key.clone()) {
+        crate::summarizer_adapter::configure_summary_driver(
+            crate::summarizer_adapter::TauriSummaryDriver::new(
+                format!("{}/chat/completions", base_url),
+                api_key,
+                Some(model.clone()),
+            ),
+        );
+    }
+
     *kernel_lock = Some(kernel);
 
     Ok(KernelStatusResponse {
@@ -1251,24 +1269,109 @@ pub async fn approval_list(
 }
 
 /// Respond to an approval
+///
+/// When approved, the kernel's `respond_to_approval` internally spawns the Hand
+/// execution.  We additionally emit Tauri events so the frontend can track when
+/// the execution finishes, since the kernel layer has no access to the AppHandle.
 #[tauri::command]
 pub async fn approval_respond(
+    app: AppHandle,
     state: State<'_, KernelState>,
     id: String,
     approved: bool,
     reason: Option<String>,
 ) -> Result<(), String> {
-    let kernel_lock = state.lock().await;
-    let kernel = kernel_lock.as_ref()
-        .ok_or_else(|| "Kernel not initialized".to_string())?;
+    // Capture hand info before calling respond_to_approval (which mutates the approval)
+    let hand_id = {
+        let kernel_lock = state.lock().await;
+        let kernel = kernel_lock.as_ref()
+            .ok_or_else(|| "Kernel not initialized".to_string())?;
 
-    kernel.respond_to_approval(&id, approved, reason).await
-        .map_err(|e| format!("Failed to respond to approval: {}", e))
+        let approvals = kernel.list_approvals().await;
+        let entry = approvals.iter().find(|a| a.id == id && a.status == "pending")
+            .ok_or_else(|| format!("Approval not found or already resolved: {}", id))?;
+        entry.hand_id.clone()
+    };
+
+    // Call kernel respond_to_approval (this updates status and spawns Hand execution)
+    {
+        let kernel_lock = state.lock().await;
+        let kernel = kernel_lock.as_ref()
+            .ok_or_else(|| "Kernel not initialized".to_string())?;
+
+        kernel.respond_to_approval(&id, approved, reason).await
+            .map_err(|e| format!("Failed to respond to approval: {}", e))?;
+    }
+
+    // When approved, monitor the Hand execution and emit events to the frontend.
+    // The kernel's respond_to_approval changes status to "approved" immediately,
+    // then the spawned task sets it to "completed" or "failed" when done.
+    if approved {
+        let approval_id = id.clone();
+        let kernel_state: KernelState = (*state).clone();
+
+        tokio::spawn(async move {
+            let timeout = tokio::time::Duration::from_secs(300);
+            let poll_interval = tokio::time::Duration::from_millis(500);
+
+            let result = tokio::time::timeout(timeout, async {
+                loop {
+                    tokio::time::sleep(poll_interval).await;
+
+                    let kernel_lock = kernel_state.lock().await;
+                    if let Some(kernel) = kernel_lock.as_ref() {
+                        // Use get_approval to check any status (not just "pending")
+                        if let Some(entry) = kernel.get_approval(&approval_id).await {
+                            match entry.status.as_str() {
+                                "completed" => {
+                                    tracing::info!("[approval_respond] Hand '{}' completed for approval {}", hand_id, approval_id);
+                                    return (true, None::<String>);
+                                }
+                                "failed" => {
+                                    let error_msg = entry.input.get("error")
+                                        .and_then(|v| v.as_str())
+                                        .unwrap_or("Unknown error")
+                                        .to_string();
+                                    tracing::warn!("[approval_respond] Hand '{}' failed for approval {}: {}", hand_id, approval_id, error_msg);
+                                    return (false, Some(error_msg));
+                                }
+                                _ => {} // "approved" = still running
+                            }
+                        } else {
+                            // Entry disappeared entirely — kernel was likely restarted
+                            return (false, Some("Approval entry disappeared".to_string()));
+                        }
+                    } else {
+                        return (false, Some("Kernel not available".to_string()));
+                    }
+                }
+            }).await;
+
+            let (success, error) = match result {
+                Ok((s, e)) => (s, e),
+                Err(_) => (false, Some("Hand execution timed out (5 minutes)".to_string())),
+            };
+
+            let _ = app.emit("hand-execution-complete", serde_json::json!({
+                "approvalId": approval_id,
+                "handId": hand_id,
+                "success": success,
+                "error": error,
+            }));
+        });
+    }
+
+    Ok(())
 }
 
-/// Approve a hand execution (alias for approval_respond with approved=true)
+/// Approve a hand execution
+///
+/// When approved, the kernel's `respond_to_approval` internally spawns the Hand
+/// execution.  We additionally emit Tauri events so the frontend can track when
+/// the execution finishes.
 #[tauri::command]
 pub async fn hand_approve(
+    app: AppHandle,
     state: State<'_, KernelState>,
     hand_name: String,
     run_id: String,
@@ -1301,6 +1404,66 @@ pub async fn hand_approve(
     kernel.respond_to_approval(&run_id, approved, reason).await
         .map_err(|e| format!("Failed to approve hand: {}", e))?;
 
+    // When approved, monitor the Hand execution and emit events to the frontend
+    if approved {
+        let approval_id = run_id.clone();
+        let hand_id = hand_name.clone();
+        let kernel_state: KernelState = (*state).clone();
+
+        tokio::spawn(async move {
+            // Poll the approval status until it transitions from "approved" to
+            // "completed" or "failed" (set by the kernel's spawned task).
+            // Timeout after 5 minutes to avoid hanging forever.
+            let timeout = tokio::time::Duration::from_secs(300);
+            let poll_interval = tokio::time::Duration::from_millis(500);
+
+            let result = tokio::time::timeout(timeout, async {
+                loop {
+                    tokio::time::sleep(poll_interval).await;
+
+                    let kernel_lock = kernel_state.lock().await;
+                    if let Some(kernel) = kernel_lock.as_ref() {
+                        // Use get_approval to check any status (not just "pending")
+                        if let Some(entry) = kernel.get_approval(&approval_id).await {
+                            match entry.status.as_str() {
+                                "completed" => {
+                                    tracing::info!("[hand_approve] Hand '{}' execution completed for approval {}", hand_id, approval_id);
+                                    return (true, None::<String>);
+                                }
+                                "failed" => {
+                                    let error_msg = entry.input.get("error")
+                                        .and_then(|v| v.as_str())
+                                        .unwrap_or("Unknown error")
+                                        .to_string();
+                                    tracing::warn!("[hand_approve] Hand '{}' execution failed for approval {}: {}", hand_id, approval_id, error_msg);
+                                    return (false, Some(error_msg));
+                                }
+                                _ => {} // still running (status is "approved")
+                            }
+                        } else {
+                            // Entry disappeared entirely — kernel was likely restarted
+                            return (false, Some("Approval entry disappeared".to_string()));
+                        }
+                    } else {
+                        return (false, Some("Kernel not available".to_string()));
+                    }
+                }
+            }).await;
+
+            let (success, error) = match result {
+                Ok((s, e)) => (s, e),
+                Err(_) => (false, Some("Hand execution timed out (5 minutes)".to_string())),
+            };
+
+            let _ = app.emit("hand-execution-complete", serde_json::json!({
+                "approvalId": approval_id,
+                "handId": hand_id,
+                "success": success,
+                "error": error,
+            }));
+        });
+    }
+
     Ok(serde_json::json!({
         "status": if approved { "approved" } else { "rejected" },
         "hand_name": hand_name,

desktop/src/components/AuditLogsPanel.tsx

@@ -633,7 +633,8 @@ export function AuditLogsPanel() {
     setVerificationResult(null);
 
     try {
-      // Call ZCLAW API to verify the chain
+      // Call ZCLAW API to verify the chain (only available on GatewayClient)
+      if (!('verifyAuditLogChain' in client)) throw new Error('KernelClient does not support chain verification');
       const result = await client.verifyAuditLogChain(log.id);
 
       const verification: MerkleVerificationResult = {

desktop/src/components/LoginPage.tsx

@@ -86,6 +86,7 @@ function ZclawLogo({ className }: { className?: string }) {
 
 /** 根据运行环境自动选择 SaaS 服务器地址 */
 function getSaasUrl(): string {
+  if (import.meta.env.DEV) return DEV_SAAS_URL;
   return isTauriRuntime() ? PRODUCTION_SAAS_URL : DEV_SAAS_URL;
 }
 

desktop/src/components/SaaS/SaaSSettings.tsx

@@ -139,12 +139,12 @@ export function SaaSSettings() {
               <CloudFeatureRow
                 name="团队协作"
                 description="与团队成员共享 Agent 和技能"
-                status={account?.role === 'admin' || account?.role === 'pro' ? 'active' : 'inactive'}
+                status={account?.role === 'admin' || account?.role === 'super_admin' ? 'active' : 'inactive'}
               />
               <CloudFeatureRow
                 name="高级分析"
                 description="使用统计和用量分析仪表板"
-                status={account?.role === 'admin' || account?.role === 'pro' ? 'active' : 'inactive'}
+                status={account?.role === 'admin' || account?.role === 'super_admin' ? 'active' : 'inactive'}
               />
             </div>
           </div>

desktop/src/lib/saas-client.ts

@@ -23,8 +23,8 @@ export interface SaaSAccountInfo {
   username: string;
   email: string;
   display_name: string;
-  role: string;
-  status: string;
+  role: 'super_admin' | 'admin' | 'user';
+  status: 'active' | 'disabled' | 'suspended';
   totp_enabled: boolean;
   created_at: string;
 }
@@ -64,6 +64,7 @@ export interface SaaSErrorResponse {
 /** Login response from POST /api/v1/auth/login */
 export interface SaaSLoginResponse {
   token: string;
+  refresh_token: string;
   account: SaaSAccountInfo;
 }
 
@@ -322,8 +323,8 @@ export interface AccountPublic {
   username: string;
   email: string;
   display_name: string;
-  role: string;
-  status: string;
+  role: 'super_admin' | 'admin' | 'user';
+  status: 'active' | 'disabled' | 'suspended';
   totp_enabled: boolean;
   last_login_at: string | null;
   created_at: string;

desktop/src/store/connectionStore.ts

@@ -352,8 +352,11 @@ export const useConnectionStore = create<ConnectionStore>((set, get) => {
 
         // === SaaS Relay Mode ===
         // Check connection mode from localStorage (set by saasStore).
-        // This takes priority over Tauri/Gateway when the user has selected SaaS mode.
+        // When SaaS is unreachable, gracefully degrade to local kernel mode
+        // so the desktop app remains functional.
         const savedMode = localStorage.getItem('zclaw-connection-mode');
+        let saasDegraded = false;
+
         if (savedMode === 'saas') {
           const { loadSaaSSession, saasClient } = await import('../lib/saas-client');
           const session = loadSaaSSession();
@@ -379,13 +382,26 @@ export const useConnectionStore = create<ConnectionStore>((set, get) => {
               useSaaSStore.getState().logout();
               throw new Error('SaaS 会话已过期，请重新登录');
             }
+
+            // SaaS unreachable — degrade to local kernel mode
             const errMsg = err instanceof Error ? err.message : String(err);
-            throw new Error(`SaaS 平台连接失败: ${errMsg}`);
+            log.warn(`SaaS 平台连接失败: ${errMsg} — 降级到本地 Kernel 模式`);
+
+            // Mark SaaS as unreachable in store
+            try {
+              const { useSaaSStore } = await import('./saasStore');
+              useSaaSStore.setState({ saasReachable: false });
+            } catch { /* non-critical */ }
+
+            saasDegraded = true;
           }
 
-          set({ connectionState: 'connected', gatewayVersion: 'saas-relay' });
-          log.debug('Connected to SaaS relay');
-          return;
+          if (!saasDegraded) {
+            set({ connectionState: 'connected', gatewayVersion: 'saas-relay' });
+            log.debug('Connected to SaaS relay');
+            return;
+          }
+          // Fall through to Tauri Kernel / Gateway mode
         }
 
         // === Internal Kernel Mode (Tauri) ===

desktop/src/store/saasStore.ts

@@ -97,7 +97,9 @@ export type SaaSStore = SaaSStateSlice & SaaSActionsSlice;
 
 // === Constants ===
 
-const DEFAULT_SAAS_URL = 'https://saas.zclaw.com';
+const DEFAULT_SAAS_URL = import.meta.env.DEV
+  ? 'http://127.0.0.1:8080'
+  : 'https://saas.zclaw.com';
 
 // === Helpers ===
 
@@ -218,14 +220,21 @@ export const useSaaSStore = create<SaaSStore>((set, get) => {
             ? err.message
             : String(err);
 
-        const isNetworkError = message.includes('Failed to fetch')
+        const isTimeout = message.includes('signal timed out')
+          || message.includes('Timeout')
+          || message.includes('timed out')
+          || message.includes('AbortError');
+
+        const isConnectionRefused = message.includes('Failed to fetch')
           || message.includes('NetworkError')
           || message.includes('ECONNREFUSED')
-          || message.includes('timeout');
+          || message.includes('connection refused');
 
-        const userMessage = isNetworkError
-          ? `无法连接到 SaaS 服务器: ${requestUrl}`
-          : message;
+        const userMessage = isTimeout
+          ? `连接 SaaS 服务器超时，请确认后端服务正在运行: ${requestUrl}`
+          : isConnectionRefused
+            ? `无法连接到 SaaS 服务器，请确认后端服务已启动: ${requestUrl}`
+            : message;
 
         set({ isLoading: false, error: userMessage });
         throw new Error(userMessage);
@@ -491,7 +500,6 @@ export const useSaaSStore = create<SaaSStore>((set, get) => {
           const existing = localStorage.getItem(storageKey);
 
           // Diff check: skip if local was modified since last pull
-          const lastPullKey = `zclaw-config-pull-ts.${config.category}.${config.key}`;
           const dirtyKey = `zclaw-config-dirty.${config.category}.${config.key}`;
           const lastPulledValue = localStorage.getItem(`zclaw-config-pulled.${config.category}.${config.key}`);