feat(proxy): 节点状态简化、连接事件记录、可靠性指标与批量删除

- 移除 UNHEALTHY 中间状态，节点状态简化为 ONLINE/OFFLINE
- 新增 proxy_node_events 表记录 tunnel 连接/断开/错误事件
- 新增 failed_requests/dns_failures/stream_errors 可靠性指标（增量累加）
- tunnel 重连改为固定 1s 延迟，移除指数退避逻辑
- resolver/service 改为以 TunnelManager 内存状态判断节点可用性，避免 DB 竞态
- 修正 Claude cache_control 字段格式，使用 ttl 字段控制缓存时长
- 移除前端手动勾选 capability 的 UI，改为从价格配置自动推断
- 新增全局模型批量删除 API，替换前端并行单个删除

aether-proxy/src/config.rs

@@ -202,19 +202,21 @@ pub struct Config {
     #[arg(long, env = "AETHER_PROXY_LOG_JSON", default_value_t = false)]
     pub log_json: bool,
 
-    /// WebSocket reconnect base delay in milliseconds
+    /// Deprecated: reconnect now uses a fixed 1s delay. Kept for config compatibility.
     #[arg(
         long,
         env = "AETHER_PROXY_TUNNEL_RECONNECT_BASE_MS",
-        default_value_t = 500
+        default_value_t = 500,
+        hide = true
     )]
     pub tunnel_reconnect_base_ms: u64,
 
-    /// WebSocket reconnect max delay in milliseconds
+    /// Deprecated: reconnect now uses a fixed 1s delay. Kept for config compatibility.
     #[arg(
         long,
         env = "AETHER_PROXY_TUNNEL_RECONNECT_MAX_MS",
-        default_value_t = 30000
+        default_value_t = 30000,
+        hide = true
     )]
     pub tunnel_reconnect_max_ms: u64,
 
@@ -288,13 +290,6 @@ impl Config {
         if self.aether_retry_max_attempts == 0 {
             anyhow::bail!("aether_retry_max_attempts must be >= 1");
         }
-        if self.tunnel_reconnect_base_ms > self.tunnel_reconnect_max_ms {
-            anyhow::bail!(
-                "tunnel_reconnect_base_ms ({}) must be <= tunnel_reconnect_max_ms ({})",
-                self.tunnel_reconnect_base_ms,
-                self.tunnel_reconnect_max_ms
-            );
-        }
         if self.upstream_connect_timeout_secs == 0 {
             anyhow::bail!("upstream_connect_timeout_secs must be > 0");
         }

aether-proxy/src/tunnel/client.rs

@@ -1,6 +1,5 @@
 //! WebSocket tunnel client: connect, authenticate, and run the tunnel.
 
-use std::sync::atomic::{AtomicU32, Ordering};
 use std::sync::Arc;
 use std::time::Duration;
 
@@ -226,20 +225,6 @@ pub fn build_tls_config() -> rustls::ClientConfig {
         .with_no_client_auth()
 }
 
-/// Calculate next reconnect delay with exponential backoff + jitter.
-pub fn next_reconnect_delay(state: &Arc<AppState>, reconnect_attempts: &AtomicU32) -> Duration {
-    let attempt = reconnect_attempts.fetch_add(1, Ordering::Relaxed);
-    let base_ms = state.config.tunnel_reconnect_base_ms;
-    let max_ms = state.config.tunnel_reconnect_max_ms;
-
-    let delay_ms = base_ms.saturating_mul(1u64 << attempt.min(10)).min(max_ms);
-
-    let jitter = (delay_ms / 4).max(1);
-    let jitter_ms = rand_u64() % jitter;
-
-    Duration::from_millis(delay_ms + jitter_ms)
-}
-
 fn build_tunnel_url(server: &ServerContext) -> String {
     let base = server.aether_url.trim_end_matches('/');
     let ws_base = if base.starts_with("https://") {
@@ -251,20 +236,3 @@ fn build_tunnel_url(server: &ServerContext) -> String {
     };
     format!("{}/api/internal/proxy-tunnel", ws_base)
 }
-
-/// Simple pseudo-random u64 (no external crate needed).
-fn rand_u64() -> u64 {
-    use std::sync::atomic::{AtomicU64, Ordering};
-    use std::time::{SystemTime, UNIX_EPOCH};
-    static COUNTER: AtomicU64 = AtomicU64::new(0);
-    let seed = SystemTime::now()
-        .duration_since(UNIX_EPOCH)
-        .unwrap_or_default()
-        .as_nanos() as u64;
-    let cnt = COUNTER.fetch_add(1, Ordering::Relaxed);
-    let mut x = seed ^ cnt;
-    x ^= x << 13;
-    x ^= x >> 7;
-    x ^= x << 17;
-    x
-}

aether-proxy/src/tunnel/mod.rs

@@ -5,7 +5,6 @@ pub mod protocol;
 pub mod stream_handler;
 pub mod writer;
 
-use std::sync::atomic::{AtomicU32, Ordering};
 use std::sync::Arc;
 use std::time::Duration;
 
@@ -14,10 +13,9 @@ use tracing::{error, info};
 
 use crate::state::{AppState, ServerContext};
 
-/// Minimum connection duration (seconds) to consider a session "stable".
+/// Fixed reconnect delay -- short enough for fast recovery, long enough to
-/// If a connection lasts shorter than this, the backoff counter is NOT reset,
+/// avoid CPU spin when the network is completely down.
-/// preventing rapid reconnect loops on persistently bad networks.
+const RECONNECT_DELAY: Duration = Duration::from_secs(1);
-const MIN_STABLE_DURATION: Duration = Duration::from_secs(30);
 
 /// Run the tunnel mode main loop (connect, dispatch, reconnect).
 ///
@@ -31,42 +29,17 @@ pub async fn run(
 ) {
     info!(server = %server.server_label, conn = conn_idx, "starting tunnel");
 
-    // Per-connection reconnect counter (avoids N connections interfering
-    // with each other's backoff via the shared ServerContext field).
-    let reconnect_attempts = AtomicU32::new(0);
-
     loop {
-        let connect_start = tokio::time::Instant::now();
-
         match client::connect_and_run(state, server, conn_idx, &mut shutdown).await {
             Ok(client::TunnelOutcome::Shutdown) => {
                 info!(server = %server.server_label, conn = conn_idx, "tunnel shut down gracefully");
                 return;
             }
             Ok(client::TunnelOutcome::Disconnected) => {
-                let duration = connect_start.elapsed();
+                info!(server = %server.server_label, conn = conn_idx, "tunnel disconnected, reconnecting");
-                if duration >= MIN_STABLE_DURATION {
-                    // Stable session -- reset backoff for quick reconnect
-                    reconnect_attempts.store(0, Ordering::Release);
-                    info!(
-                        server = %server.server_label,
-                        conn = conn_idx,
-                        duration_secs = duration.as_secs(),
-                        "tunnel disconnected after stable session"
-                    );
-                } else {
-                    // Short-lived session -- keep backoff increasing
-                    info!(
-                        server = %server.server_label,
-                        conn = conn_idx,
-                        duration_secs = duration.as_secs(),
-                        "tunnel disconnected quickly, increasing backoff"
-                    );
-                }
             }
             Err(e) => {
-                // Connection failed -- keep backoff increasing
+                error!(server = %server.server_label, conn = conn_idx, error = %e, "tunnel connection error, reconnecting");
-                error!(server = %server.server_label, conn = conn_idx, error = %e, "tunnel connection lost");
             }
         }
 
@@ -75,11 +48,8 @@ pub async fn run(
             return;
         }
 
-        let delay = client::next_reconnect_delay(state, &reconnect_attempts);
-        info!(server = %server.server_label, conn = conn_idx, delay_ms = delay.as_millis(), "reconnecting tunnel");
-
         tokio::select! {
-            _ = tokio::time::sleep(delay) => {}
+            _ = tokio::time::sleep(RECONNECT_DELAY) => {}
             _ = shutdown.changed() => {
                 info!(server = %server.server_label, conn = conn_idx, "shutdown requested during reconnect wait");
                 return;

alembic/versions/20260228_1238_48afe197cc15_proxy_node_metrics_and_events.py

@@ -0,0 +1,114 @@
+"""proxy_node_metrics_and_events
+
+Revision ID: 48afe197cc15
+Revises: b2c3d4e5f6a7
+Create Date: 2026-02-28 04:33:11.201185+00:00
+
+"""
+
+import sqlalchemy as sa
+from sqlalchemy import inspect
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision = "48afe197cc15"
+down_revision = "b2c3d4e5f6a7"
+branch_labels = None
+depends_on = None
+
+
+def _column_exists(table_name: str, column_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    columns = [c["name"] for c in insp.get_columns(table_name)]
+    return column_name in columns
+
+
+def _table_exists(table_name: str) -> bool:
+    bind = op.get_bind()
+    insp = inspect(bind)
+    return table_name in insp.get_table_names()
+
+
+def upgrade() -> None:
+    # proxy_nodes: 新增错误指标字段
+    if not _column_exists("proxy_nodes", "failed_requests"):
+        op.add_column(
+            "proxy_nodes",
+            sa.Column(
+                "failed_requests",
+                sa.BigInteger(),
+                nullable=False,
+                server_default="0",
+                comment="累计失败请求数",
+            ),
+        )
+    if not _column_exists("proxy_nodes", "dns_failures"):
+        op.add_column(
+            "proxy_nodes",
+            sa.Column(
+                "dns_failures",
+                sa.BigInteger(),
+                nullable=False,
+                server_default="0",
+                comment="累计 DNS 失败数",
+            ),
+        )
+    if not _column_exists("proxy_nodes", "stream_errors"):
+        op.add_column(
+            "proxy_nodes",
+            sa.Column(
+                "stream_errors",
+                sa.BigInteger(),
+                nullable=False,
+                server_default="0",
+                comment="累计流错误数",
+            ),
+        )
+
+    # proxy_node_events: 连接事件表
+    if not _table_exists("proxy_node_events"):
+        op.create_table(
+            "proxy_node_events",
+            sa.Column("id", sa.BigInteger(), autoincrement=True, nullable=False),
+            sa.Column("node_id", sa.String(length=36), nullable=False),
+            sa.Column(
+                "event_type",
+                sa.String(length=20),
+                nullable=False,
+                comment="事件类型: connected, disconnected, error",
+            ),
+            sa.Column(
+                "detail",
+                sa.String(length=500),
+                nullable=True,
+                comment="事件详情（如断开原因）",
+            ),
+            sa.Column("created_at", sa.DateTime(timezone=True), nullable=False),
+            sa.ForeignKeyConstraint(["node_id"], ["proxy_nodes.id"], ondelete="CASCADE"),
+            sa.PrimaryKeyConstraint("id"),
+        )
+        op.create_index(
+            "idx_proxy_node_events_node_created",
+            "proxy_node_events",
+            ["node_id", "created_at"],
+        )
+        op.create_index(
+            op.f("ix_proxy_node_events_node_id"),
+            "proxy_node_events",
+            ["node_id"],
+        )
+
+
+def downgrade() -> None:
+    if _table_exists("proxy_node_events"):
+        op.drop_index(op.f("ix_proxy_node_events_node_id"), table_name="proxy_node_events")
+        op.drop_index("idx_proxy_node_events_node_created", table_name="proxy_node_events")
+        op.drop_table("proxy_node_events")
+    if _column_exists("proxy_nodes", "stream_errors"):
+        op.drop_column("proxy_nodes", "stream_errors")
+    if _column_exists("proxy_nodes", "dns_failures"):
+        op.drop_column("proxy_nodes", "dns_failures")
+    if _column_exists("proxy_nodes", "failed_requests"):
+        op.drop_column("proxy_nodes", "failed_requests")

frontend/src/api/endpoints/global-models.ts

@@ -68,6 +68,16 @@ export async function deleteGlobalModel(
   await client.delete(`/api/admin/models/global/${id}`, { params: { force } })
 }
 
+/**
+ * 批量删除 GlobalModel
+ */
+export async function batchDeleteGlobalModels(
+  ids: string[]
+): Promise<{ success_count: number; failed: Array<{ id: string; error: string }> }> {
+  const response = await client.post('/api/admin/models/global/batch-delete', { ids })
+  return response.data
+}
+
 /**
  * 批量为 GlobalModel 添加关联提供商
  */

frontend/src/api/proxy-nodes.ts

@@ -13,7 +13,7 @@ export interface ProxyNode {
   ip: string
   port: number
   region: string | null
-  status: 'online' | 'unhealthy' | 'offline'
+  status: 'online' | 'offline'
   is_manual: boolean
   tunnel_mode: boolean
   tunnel_connected: boolean
@@ -34,10 +34,20 @@ export interface ProxyNode {
   active_connections: number
   total_requests: number
   avg_latency_ms: number | null
+  failed_requests: number
+  dns_failures: number
+  stream_errors: number
   created_at: string
   updated_at: string
 }
 
+export interface ProxyNodeEvent {
+  id: number
+  event_type: 'connected' | 'disconnected' | 'error'
+  detail: string | null
+  created_at: string
+}
+
 export interface ProxyNodeListResponse {
   items: ProxyNode[]
   total: number
@@ -103,4 +113,9 @@ export const proxyNodesApi = {
     const response = await apiClient.post<ProxyNodeTestResult>('/api/admin/proxy-nodes/test-url', data)
     return response.data
   },
+
+  async listNodeEvents(nodeId: string, limit = 50): Promise<{ items: ProxyNodeEvent[] }> {
+    const response = await apiClient.get<{ items: ProxyNodeEvent[] }>(`/api/admin/proxy-nodes/${nodeId}/events`, { params: { limit } })
+    return response.data
+  },
 }

frontend/src/features/models/components/GlobalModelFormDialog.vue

@@ -179,31 +179,6 @@
             </div>
           </section>
 
-          <!-- Key 能力配置 -->
-          <section
-            v-if="availableCapabilities.length > 0"
-            class="space-y-2"
-          >
-            <h4 class="font-medium text-sm">
-              模型偏好
-            </h4>
-            <div class="flex flex-wrap gap-2">
-              <label
-                v-for="cap in availableCapabilities"
-                :key="cap.name"
-                class="flex items-center gap-2 px-2.5 py-1 rounded-md border bg-muted/30 cursor-pointer text-sm"
-              >
-                <input
-                  type="checkbox"
-                  :checked="form.supported_capabilities?.includes(cap.name)"
-                  class="rounded"
-                  @change="toggleCapability(cap.name)"
-                >
-                <span>{{ cap.display_name }}</span>
-              </label>
-            </div>
-          </section>
-
           <!-- 价格配置 -->
           <section class="space-y-3">
             <h4 class="font-medium text-sm">
@@ -212,7 +187,7 @@
             <TieredPricingEditor
               ref="tieredPricingEditorRef"
               v-model="tieredPricing"
-              :show-cache1h="form.supported_capabilities?.includes('cache_1h')"
+              :show-cache1h="true"
             />
             <div class="flex items-center gap-3 pt-2 border-t">
               <Label class="text-xs whitespace-nowrap">按次计费</Label>
@@ -353,7 +328,7 @@
 </template>
 
 <script setup lang="ts">
-import { ref, computed, watch, onMounted } from 'vue'
+import { ref, computed, watch } from 'vue'
 import {
   Loader2, Layers, SquarePen,
   Search, ChevronRight, Plus, Trash2
@@ -378,7 +353,6 @@ import {
   type GlobalModelUpdate,
 } from '@/api/global-models'
 import type { TieredPricingConfig } from '@/api/endpoints/types'
-import { getAllCapabilities, type CapabilityDefinition } from '@/api/endpoints'
 
 const props = defineProps<{
   open: boolean
@@ -665,34 +639,6 @@ function fillVideoResolutionPricePreset(preset: 'common' | 'sora' | 'veo') {
   }))
 }
 
-// Key 能力选项
-const availableCapabilities = ref<CapabilityDefinition[]>([])
-
-// 加载可用能力列表
-async function loadCapabilities() {
-  try {
-    availableCapabilities.value = await getAllCapabilities()
-  } catch (err) {
-    log.error('Failed to load capabilities:', err)
-  }
-}
-
-// 切换能力
-function toggleCapability(capName: string) {
-  if (!form.value.supported_capabilities) {
-    form.value.supported_capabilities = []
-  }
-  const index = form.value.supported_capabilities.indexOf(capName)
-  if (index >= 0) {
-    form.value.supported_capabilities.splice(index, 1)
-  } else {
-    form.value.supported_capabilities.push(capName)
-  }
-}
-
-onMounted(() => {
-  loadCapabilities()
-})
 
 // 加载模型列表
 async function loadModels() {
@@ -829,6 +775,19 @@ async function handleSubmit() {
   // Apply billing (video) pricing into config before cleaning/submitting.
   applyVideoPricingToConfig()
 
+  // Auto-infer supported_capabilities from tiered pricing config
+  const caps = new Set(form.value.supported_capabilities || [])
+  const has1hPricing = finalTieredPricing?.tiers?.some(
+    (t: Record<string, unknown>) => Array.isArray(t.cache_ttl_pricing)
+      && (t.cache_ttl_pricing as Array<Record<string, unknown>>).some(c => c.ttl_minutes === 60)
+  )
+  if (has1hPricing) {
+    caps.add('cache_1h')
+  } else {
+    caps.delete('cache_1h')
+  }
+  form.value.supported_capabilities = caps.size > 0 ? [...caps] : []
+
   // 清理空的 config
   const cleanConfig = form.value.config && Object.keys(form.value.config).length > 0
     ? form.value.config

frontend/src/features/models/components/ModelDetailDrawer.vue

@@ -182,7 +182,7 @@
                     v-if="getFirst1hCachePrice(model.default_tiered_pricing) !== '-'"
                     class="flex items-center gap-3 p-3 rounded-lg border bg-muted/20"
                   >
-                    <Label class="text-xs text-muted-foreground whitespace-nowrap">1h 缓存创建</Label>
+                    <Label class="text-xs text-muted-foreground whitespace-nowrap">1h 缓存</Label>
                     <span class="text-sm font-mono">{{ getFirst1hCachePrice(model.default_tiered_pricing) }}</span>
                   </div>
                   <!-- 按次计费 -->

frontend/src/features/models/components/TieredPricingEditor.vue

@@ -112,7 +112,7 @@
           v-if="showCache1h"
           class="space-y-1"
         >
-          <Label class="text-xs text-muted-foreground">1h 缓存创建</Label>
+          <Label class="text-xs text-muted-foreground">1h 缓存</Label>
           <Input
             :model-value="getCache1hDisplay(index)"
             type="number"
@@ -209,26 +209,6 @@ watch(
   { immediate: true }
 )
 
-// 监听 showCache1h 变化
-watch(
-  () => props.showCache1h,
-  (newValue, oldValue) => {
-    if (oldValue === true && newValue === false) {
-      // 取消勾选时，清除本地的 1h 缓存数据和手动设置标记
-      localTiers.value.forEach((tier, i) => {
-        tier.cache_ttl_pricing = undefined
-        if (cacheManuallySet[i]) {
-          cacheManuallySet[i].cache1h = false
-        }
-      })
-      syncToParent()
-    } else if (oldValue === false && newValue === true) {
-      // 勾选时，同步自动计算的价格到父组件
-      syncToParent()
-    }
-  }
-)
-
 // 验证错误
 const validationError = computed(() => {
   if (localTiers.value.length === 0) {

frontend/src/features/providers/components/ProviderModelFormDialog.vue

@@ -245,18 +245,8 @@ const tieredPricingEditorRef = ref<InstanceType<typeof TieredPricingEditor> | nu
 
 const isEditing = computed(() => !!props.editingModel)
 
-// 计算是否显示 1h 缓存输入框
+// 1h 缓存定价始终显示
-const showCache1h = computed(() => {
+const showCache1h = true
-  if (isEditing.value) {
-    // 编辑模式：检查当前配置是否有 1h 缓存配置（从 tiered_pricing 或 effective_tiered_pricing 中检测）
-    const pricing = props.editingModel?.tiered_pricing || props.editingModel?.effective_tiered_pricing
-    return pricing?.tiers?.some(t => t.cache_ttl_pricing?.some(c => c.ttl_minutes === 60)) ?? false
-  } else {
-    // 添加模式：从选中的全局模型中读取 supported_capabilities
-    const selectedModel = availableGlobalModels.value.find(m => m.id === form.value.global_model_id)
-    return selectedModel?.supported_capabilities?.includes('cache_1h') ?? false
-  }
-})
 
 // 表单状态
 const submitting = ref(false)

frontend/src/views/admin/ModelManagement.vue

@@ -656,6 +656,7 @@ import {
   getGlobalModel,
   updateGlobalModel,
   deleteGlobalModel,
+  batchDeleteGlobalModels,
   batchAssignToProviders,
   getGlobalModelProviders,
   type GlobalModelResponse,
@@ -1276,15 +1277,13 @@ async function confirmBatchDeleteModels() {
   submittingBatchManage.value = true
   try {
     const ids = Array.from(selectedBatchManageModelIds.value)
-    const results = await Promise.allSettled(ids.map(id => deleteGlobalModel(id)))
+    const result = await batchDeleteGlobalModels(ids)
-    const successCount = results.filter(r => r.status === 'fulfilled').length
-    const failCount = results.filter(r => r.status === 'rejected').length
 
-    if (successCount > 0) {
+    if (result.success_count > 0) {
-      success(`成功删除 ${successCount} 个模型`)
+      success(`成功删除 ${result.success_count} 个模型`)
     }
-    if (failCount > 0) {
+    if (result.failed.length > 0) {
-      showError(`${failCount} 个模型删除失败`, '部分失败')
+      showError(`${result.failed.length} 个模型删除失败`, '部分失败')
     }
 
     // 清除选中的已删除模型

frontend/src/views/admin/ProxyNodes.vue

@@ -48,9 +48,6 @@
                 <SelectItem value="online">
                   在线
                 </SelectItem>
-                <SelectItem value="unhealthy">
-                  异常
-                </SelectItem>
                 <SelectItem value="offline">
                   离线
                 </SelectItem>
@@ -86,9 +83,6 @@
                 <SelectItem value="online">
                   在线
                 </SelectItem>
-                <SelectItem value="unhealthy">
-                  异常
-                </SelectItem>
                 <SelectItem value="offline">
                   离线
                 </SelectItem>
@@ -237,6 +231,16 @@
                   >
                     <Settings class="h-4 w-4" />
                   </Button>
+                  <Button
+                    v-if="!node.is_manual"
+                    variant="ghost"
+                    size="icon"
+                    class="h-8 w-8"
+                    title="连接事件"
+                    @click="handleViewEvents(node)"
+                  >
+                    <History class="h-4 w-4" />
+                  </Button>
                   <Button
                     variant="ghost"
                     size="icon"
@@ -548,6 +552,77 @@
         </Button>
       </template>
     </Dialog>
+
+    <!-- 连接事件对话框 -->
+    <Dialog
+      :open="showEventsDialog"
+      title="连接事件"
+      :description="eventsNode ? `${eventsNode.name} 的连接历史` : ''"
+      size="lg"
+      @update:open="(v: boolean) => { if (!v) { showEventsDialog = false; eventsNode = null; nodeEvents = [] } }"
+    >
+      <div class="space-y-3">
+        <!-- 可靠性指标摘要 -->
+        <div
+          v-if="eventsNode"
+          class="grid grid-cols-3 gap-3 text-sm"
+        >
+          <div class="bg-muted/40 rounded-lg px-3 py-2 text-center">
+            <span class="block text-foreground/60 text-xs">失败请求</span>
+            <span class="tabular-nums font-medium">{{ formatNumber(eventsNode.failed_requests || 0) }}</span>
+          </div>
+          <div class="bg-muted/40 rounded-lg px-3 py-2 text-center">
+            <span class="block text-foreground/60 text-xs">DNS 失败</span>
+            <span class="tabular-nums font-medium">{{ formatNumber(eventsNode.dns_failures || 0) }}</span>
+          </div>
+          <div class="bg-muted/40 rounded-lg px-3 py-2 text-center">
+            <span class="block text-foreground/60 text-xs">流错误</span>
+            <span class="tabular-nums font-medium">{{ formatNumber(eventsNode.stream_errors || 0) }}</span>
+          </div>
+        </div>
+
+        <!-- 事件列表 -->
+        <div
+          v-if="loadingEvents"
+          class="py-8 text-center text-muted-foreground text-sm"
+        >
+          加载中...
+        </div>
+        <div
+          v-else-if="nodeEvents.length === 0"
+          class="py-8 text-center text-muted-foreground text-sm"
+        >
+          暂无连接事件记录
+        </div>
+        <div
+          v-else
+          class="max-h-80 overflow-y-auto space-y-1.5"
+        >
+          <div
+            v-for="event in nodeEvents"
+            :key="event.id"
+            class="flex items-center gap-2 px-3 py-2 rounded-lg bg-muted/30 text-sm"
+          >
+            <Badge
+              :variant="eventTypeVariant(event.event_type)"
+              class="text-[10px] px-1.5 py-0 shrink-0"
+            >
+              {{ eventTypeLabel(event.event_type) }}
+            </Badge>
+            <span class="text-muted-foreground truncate flex-1">{{ event.detail || '-' }}</span>
+            <span class="text-xs text-muted-foreground/70 tabular-nums shrink-0">{{ formatTime(event.created_at) }}</span>
+          </div>
+        </div>
+      </div>
+      <template #footer>
+        <Button
+          variant="outline"
+          @click="showEventsDialog = false; eventsNode = null; nodeEvents = []"
+        >
+          关闭
+        </Button>
+      </template>
+    </Dialog>
   </div>
 </template>
 
@@ -556,7 +631,7 @@ import { ref, computed, onMounted, watch } from 'vue'
 import { useProxyNodesStore } from '@/stores/proxy-nodes'
 import { useToast } from '@/composables/useToast'
 import { useConfirm } from '@/composables/useConfirm'
-import { proxyNodesApi, type ProxyNode, type ProxyNodeRemoteConfig } from '@/api/proxy-nodes'
+import { proxyNodesApi, type ProxyNode, type ProxyNodeRemoteConfig, type ProxyNodeEvent } from '@/api/proxy-nodes'
 
 import {
   Card,
@@ -580,7 +655,7 @@ import {
   Dialog,
 } from '@/components/ui'
 
-import { Search, Trash2, Plus, SquarePen, Activity, Loader2, Settings } from 'lucide-vue-next'
+import { Search, Trash2, Plus, SquarePen, Activity, Loader2, Settings, History } from 'lucide-vue-next'
 import { parseApiError } from '@/utils/errorParser'
 import { formatRegion } from '@/utils/region'
 import HardwareTooltip from './components/HardwareTooltip.vue'
@@ -616,6 +691,12 @@ const configForm = ref({
   heartbeat_interval: '30',
 })
 
+// 连接事件对话框
+const showEventsDialog = ref(false)
+const eventsNode = ref<ProxyNode | null>(null)
+const nodeEvents = ref<ProxyNodeEvent[]>([])
+const loadingEvents = ref(false)
+
 // 测试连通性
 const testingNodes = ref(new Set<string>())
 const testingUrl = ref(false)
@@ -834,10 +915,41 @@ async function handleTest(node: ProxyNode) {
   }
 }
 
+async function handleViewEvents(node: ProxyNode) {
+  eventsNode.value = node
+  showEventsDialog.value = true
+  loadingEvents.value = true
+  try {
+    const res = await proxyNodesApi.listNodeEvents(node.id, 50)
+    nodeEvents.value = res.items
+  } catch (err: unknown) {
+    toastError(parseApiError(err, '加载事件失败'))
+  } finally {
+    loadingEvents.value = false
+  }
+}
+
+function eventTypeLabel(type: string) {
+  switch (type) {
+    case 'connected': return '连接'
+    case 'disconnected': return '断开'
+    case 'error': return '错误'
+    default: return type
+  }
+}
+
+function eventTypeVariant(type: string) {
+  switch (type) {
+    case 'connected': return 'success' as const
+    case 'disconnected': return 'destructive' as const
+    case 'error': return 'destructive' as const
+    default: return 'secondary' as const
+  }
+}
+
 function statusVariant(status: string) {
   switch (status) {
     case 'online': return 'success' as const
-    case 'unhealthy': return 'secondary' as const
     case 'offline': return 'destructive' as const
     default: return 'secondary' as const
   }
@@ -846,7 +958,6 @@ function statusVariant(status: string) {
 function statusLabel(status: string) {
   switch (status) {
     case 'online': return '在线'
-    case 'unhealthy': return '异常'
     case 'offline': return '离线'
     default: return status
   }

frontend/src/views/user/components/UserModelDetailDrawer.vue

@@ -192,7 +192,7 @@
                   v-if="getFirst1hCachePrice(model.default_tiered_pricing) !== '-'"
                   class="flex items-center gap-3 p-3 rounded-lg border bg-muted/20"
                 >
-                  <Label class="text-xs text-muted-foreground whitespace-nowrap">1h 缓存创建</Label>
+                  <Label class="text-xs text-muted-foreground whitespace-nowrap">1h 缓存</Label>
                   <span class="text-sm font-mono">{{ getFirst1hCachePrice(model.default_tiered_pricing) }}</span>
                 </div>
                 <!-- 按次计费 -->

src/api/admin/models/global_models.py

@@ -9,7 +9,7 @@ from __future__ import annotations
 from dataclasses import dataclass
 from typing import Any
 
-from fastapi import APIRouter, Depends, Query, Request, Response
+from fastapi import APIRouter, Body, Depends, Query, Request, Response
 from sqlalchemy.orm import Session
 
 from src.api.base.admin_adapter import AdminApiAdapter
@@ -181,6 +181,21 @@ async def delete_global_model(
     return Response(status_code=204)
 
 
+@router.post("/batch-delete")
+async def batch_delete_global_models(
+    request: Request,
+    ids: list[str] = Body(..., embed=True, max_length=100),
+    db: Session = Depends(get_db),
+) -> dict:
+    """
+    批量删除 GlobalModel
+
+    顺序删除多个 GlobalModel（每个独立提交），避免并行删除导致的锁竞争。
+    """
+    adapter = AdminBatchDeleteGlobalModelsAdapter(ids=ids)
+    return await pipeline.run(adapter=adapter, http_request=request, db=db, mode=adapter.mode)
+
+
 @router.post(
     "/{global_model_id}/assign-to-providers", response_model=BatchAssignToProvidersResponse
 )
@@ -517,6 +532,50 @@ class AdminDeleteGlobalModelAdapter(AdminApiAdapter):
         return None
 
 
+@dataclass
+class AdminBatchDeleteGlobalModelsAdapter(AdminApiAdapter):
+    """批量删除多个 GlobalModel（顺序执行，每个删除独立提交）"""
+
+    ids: list[str]
+
+    async def handle(self, context: ApiRequestContext) -> Any:  # type: ignore[override]
+        from src.core.exceptions import NotFoundException
+        from src.models.database import GlobalModel
+
+        success_count = 0
+        failed: list[dict] = []
+        deleted_names: list[tuple[str, str]] = []  # (name, id)
+
+        for gm_id in self.ids:
+            try:
+                gm = context.db.query(GlobalModel).filter(GlobalModel.id == gm_id).first()
+                if gm:
+                    name = gm.name
+                    mid = gm.id
+                    GlobalModelService.delete_global_model(context.db, gm_id)
+                    deleted_names.append((name, mid))
+                    success_count += 1
+                else:
+                    failed.append({"id": gm_id, "error": "not found"})
+            except NotFoundException:
+                failed.append({"id": gm_id, "error": "not found"})
+            except Exception as e:
+                context.db.rollback()
+                failed.append({"id": gm_id, "error": str(e)})
+
+        # 批量失效缓存
+        if deleted_names:
+            from src.services.cache.invalidation import get_cache_invalidation_service
+
+            cache_service = get_cache_invalidation_service()
+            for name, mid in deleted_names:
+                await cache_service.on_global_model_changed(name, mid)
+
+        logger.info("批量删除 GlobalModel: success={}, failed={}", success_count, len(failed))
+
+        return {"success_count": success_count, "failed": failed}
+
+
 @dataclass
 class AdminBatchAssignToProvidersAdapter(AdminApiAdapter):
     """批量为 Provider 添加 GlobalModel 实现"""

src/api/admin/proxy_nodes/routes.py

@@ -221,6 +221,17 @@ async def update_proxy_node_config(
     return await pipeline.run(adapter=adapter, http_request=request, db=db, mode=adapter.mode)
 
 
+@router.get("/{node_id}/events")
+async def list_proxy_node_events(
+    node_id: str,
+    request: Request,
+    limit: int = Query(50, ge=1, le=200),
+    db: Session = Depends(get_db),
+) -> Any:
+    adapter = AdminListProxyNodeEventsAdapter(node_id=node_id, limit=limit)
+    return await pipeline.run(adapter=adapter, http_request=request, db=db, mode=adapter.mode)
+
+
 # ---------------------------------------------------------------------------
 # 辅助函数
 # ---------------------------------------------------------------------------
@@ -517,3 +528,34 @@ class AdminTestProxyUrlAdapter(AdminApiAdapter):
             username=req.username,
             password=req.password,
         )
+
+
+@dataclass
+class AdminListProxyNodeEventsAdapter(AdminApiAdapter):
+    """查询代理节点连接事件（连接/断开/错误历史）"""
+
+    name: str = "admin_list_proxy_node_events"
+    node_id: str = ""
+    limit: int = 50
+
+    async def handle(self, context: ApiRequestContext) -> Any:
+        from src.models.database import ProxyNodeEvent
+
+        events = (
+            context.db.query(ProxyNodeEvent)
+            .filter(ProxyNodeEvent.node_id == self.node_id)
+            .order_by(ProxyNodeEvent.created_at.desc())
+            .limit(self.limit)
+            .all()
+        )
+        return {
+            "items": [
+                {
+                    "id": e.id,
+                    "event_type": e.event_type,
+                    "detail": e.detail,
+                    "created_at": e.created_at,
+                }
+                for e in events
+            ],
+        }

src/api/admin/proxy_tunnel.py

@@ -119,6 +119,7 @@ async def proxy_tunnel_ws(ws: WebSocket) -> None:
     # 启动服务端 ping 任务，防止中间代理因空闲超时关闭连接
     ping_task = asyncio.create_task(_ping_loop(conn))
 
+    disconnect_reason: str | None = None
     try:
         oversized_count = 0
         while True:
@@ -126,6 +127,7 @@ async def proxy_tunnel_ws(ws: WebSocket) -> None:
                 data = await asyncio.wait_for(ws.receive_bytes(), timeout=_IDLE_TIMEOUT)
             except asyncio.TimeoutError:
                 logger.warning("tunnel idle timeout for node_id={}", node_id)
+                disconnect_reason = "idle timeout"
                 await ws.close(code=4004, reason="idle timeout")
                 break
             if len(data) > _MAX_FRAME_SIZE:
@@ -133,6 +135,7 @@ async def proxy_tunnel_ws(ws: WebSocket) -> None:
                 logger.warning("tunnel frame too large from {}: {} bytes", node_id, len(data))
                 if oversized_count >= 5:
                     logger.warning("too many oversized frames from {}, closing", node_id)
+                    disconnect_reason = "too many oversized frames"
                     await ws.close(code=4003, reason="too many oversized frames")
                     break
                 continue
@@ -146,14 +149,16 @@ async def proxy_tunnel_ws(ws: WebSocket) -> None:
             await manager.handle_incoming_frame(conn, frame)
 
     except WebSocketDisconnect:
+        disconnect_reason = "WebSocket disconnected"
         logger.info("tunnel WebSocket disconnected: node_id={}", node_id)
     except Exception as e:
+        disconnect_reason = f"error: {e}"
         logger.error("tunnel WebSocket error for node_id={}: {}", node_id, e)
     finally:
         ping_task.cancel()
         manager.unregister(conn)
         if not manager.has_tunnel(node_id):
-            await _update_tunnel_status(node_id, connected=False)
+            await _update_tunnel_status(node_id, connected=False, detail=disconnect_reason)
         else:
             logger.info("tunnel connection closed but pool still active: node_id={}", node_id)
 
@@ -174,14 +179,16 @@ async def _ping_loop(conn: TunnelConnection) -> None:
         pass
 
 
-async def _update_tunnel_status(node_id: str, *, connected: bool) -> None:
+async def _update_tunnel_status(
-    """更新 ProxyNode 的 tunnel 连接状态（在线程池中执行，避免阻塞 event loop）"""
+    node_id: str, *, connected: bool, detail: str | None = None
+) -> None:
+    """更新 ProxyNode 的 tunnel 连接状态并记录事件（在线程池中执行）"""
 
     def _sync_update() -> None:
         from datetime import datetime, timezone
 
         from src.database import create_session
-        from src.models.database import ProxyNode, ProxyNodeStatus
+        from src.models.database import ProxyNode, ProxyNodeEvent, ProxyNodeStatus
 
         db = create_session()
         try:
@@ -193,9 +200,16 @@ async def _update_tunnel_status(node_id: str, *, connected: bool) -> None:
                     node.tunnel_connected_at = now
                     node.status = ProxyNodeStatus.ONLINE
                 else:
-                    # 记录断开时刻，供 health_scheduler 计算 UNHEALTHY 缓冲期
                     node.tunnel_connected_at = now
-                    node.status = ProxyNodeStatus.UNHEALTHY
+                    node.status = ProxyNodeStatus.OFFLINE
+
+                # 记录连接事件
+                event = ProxyNodeEvent(
+                    node_id=node_id,
+                    event_type="connected" if connected else "disconnected",
+                    detail=detail,
+                )
+                db.add(event)
                 db.commit()
         finally:
             db.close()

src/api/handlers/claude/adapter.py

@@ -131,7 +131,7 @@ class ClaudeChatAdapter(ChatAdapterBase):
         request_body: dict[str, Any] | None = None,
     ) -> dict[str, bool]:
         """检测 Claude 请求中隐含的能力需求"""
-        return ClaudeCapabilityDetector.detect_from_headers(headers)
+        return ClaudeCapabilityDetector.detect_from_headers(headers, request_body)
 
     # =========================================================================
     # Claude 特定的计费逻辑

src/models/database.py

@@ -874,7 +874,6 @@ class ProxyNodeStatus(PyEnum):
     """代理节点状态"""
 
     ONLINE = "online"
-    UNHEALTHY = "unhealthy"
     OFFLINE = "offline"
 
 
@@ -919,6 +918,9 @@ class ProxyNode(Base):
     active_connections = Column(Integer, default=0, nullable=False)
     total_requests = Column(BigInteger, default=0, nullable=False)
     avg_latency_ms = Column(Float, nullable=True)
+    failed_requests = Column(BigInteger, default=0, nullable=False, comment="累计失败请求数")
+    dns_failures = Column(BigInteger, default=0, nullable=False, comment="累计 DNS 失败数")
+    stream_errors = Column(BigInteger, default=0, nullable=False, comment="累计流错误数")
 
     # 硬件信息（注册时上报，JSON 可扩展）
     hardware_info = Column(
@@ -962,6 +964,33 @@ class ProxyNode(Base):
     __table_args__ = (UniqueConstraint("ip", "port", name="uq_proxy_node_ip_port"),)
 
 
+class ProxyNodeEvent(Base):
+    """代理节点连接事件表 -- 记录 tunnel 连接/断开/错误事件，用于连接稳定性分析"""
+
+    __tablename__ = "proxy_node_events"
+
+    id = Column(BigInteger, primary_key=True, autoincrement=True)
+    node_id = Column(
+        String(36),
+        ForeignKey("proxy_nodes.id", ondelete="CASCADE"),
+        nullable=False,
+        index=True,
+    )
+    event_type = Column(
+        String(20),
+        nullable=False,
+        comment="事件类型: connected, disconnected, error",
+    )
+    detail = Column(String(500), nullable=True, comment="事件详情（如断开原因）")
+    created_at = Column(
+        DateTime(timezone=True),
+        default=lambda: datetime.now(timezone.utc),
+        nullable=False,
+    )
+
+    __table_args__ = (Index("idx_proxy_node_events_node_created", "node_id", "created_at"),)
+
+
 class GlobalModel(ExportMixin, Base):
     """全局统一模型定义 - 包含价格和能力配置
 

src/modules/proxy_nodes/__init__.py

@@ -21,7 +21,7 @@ if TYPE_CHECKING:
 
 
 def _reset_tunnel_connected_on_startup() -> None:
-    """服务端启动时将所有 tunnel_connected=True 的节点重置为 False/UNHEALTHY。
+    """服务端启动时将所有 tunnel_connected=True 的节点重置为 False/OFFLINE。
 
     服务端重启后 TunnelManager 内存状态丢失，但 DB 中可能残留
     tunnel_connected=True 的记录。如果不重置，health_scheduler 会错误地
@@ -48,7 +48,7 @@ def _reset_tunnel_connected_on_startup() -> None:
             for node in stale_nodes:
                 node.tunnel_connected = False
                 node.tunnel_connected_at = now
-                node.status = ProxyNodeStatus.UNHEALTHY
+                node.status = ProxyNodeStatus.OFFLINE
                 node.updated_at = now
             db.commit()
             logger.info(

src/services/provider/adapters/claude_code/envelope.py

@@ -224,15 +224,32 @@ _VALID_CACHE_TTL_TARGETS = {"ephemeral", "1h"}
 
 
 def _override_cache_control_in_blocks(blocks: list[Any], target: str) -> int:
-    """Override cache_control in a list of content blocks. Returns count of overrides."""
+    """Override cache_control TTL in a list of content blocks. Returns count of overrides.
+
+    According to Anthropic API docs, cache_control format is:
+        {"type": "ephemeral", "ttl": "5m" | "1h"}
+    ``type`` is always "ephemeral"; the ``ttl`` field controls the actual duration.
+    When ttl is absent, the default is 5m (ephemeral).
+    """
     count = 0
     for block in blocks:
         if not isinstance(block, dict):
             continue
         cc = block.get("cache_control")
-        if isinstance(cc, dict):
+        if not isinstance(cc, dict):
-            if cc.get("type") != target:
+            continue
-                cc["type"] = target
+        # Ensure type is always "ephemeral"
+        if cc.get("type") != "ephemeral":
+            cc["type"] = "ephemeral"
+        if target == "ephemeral":
+            # Target is 5m (default) -- remove explicit ttl so it falls back to default
+            if "ttl" in cc:
+                del cc["ttl"]
+                count += 1
+        else:
+            # Target is "1h" -- set ttl explicitly
+            if cc.get("ttl") != target:
+                cc["ttl"] = target
                 count += 1
     return count
 

src/services/proxy_node/health_scheduler.py

@@ -2,9 +2,9 @@
 ProxyNode 心跳检测调度器
 
 定期检查 proxy_nodes 的 tunnel 连接状态，更新节点状态：
-- tunnel 实际连接中      -> ONLINE（包括从 OFFLINE 恢复的情况）
+- tunnel 实际连接中  -> ONLINE
-- tunnel 刚断开 (<60s)   -> UNHEALTHY（缓冲期，避免正在进行的请求被立即切走）
+- tunnel 未连接      -> OFFLINE
-- tunnel 断开超过 60s    -> OFFLINE
+以 TunnelManager 内存中的实际连接状态为准。
 """
 
 from __future__ import annotations
@@ -17,12 +17,18 @@ from src.database import create_session
 from src.models.database import ProxyNode, ProxyNodeStatus
 from src.services.system.scheduler import get_scheduler
 
+# 事件保留天数
+_EVENT_RETENTION_DAYS = 30
+# 每隔多少次心跳检测执行一次事件清理（15s * 240 = 1h）
+_EVENT_CLEANUP_INTERVAL = 240
+
 
 class ProxyNodeHealthScheduler:
     """代理节点心跳检测调度器"""
 
     def __init__(self) -> None:
         self.running = False
+        self._check_count = 0
 
     async def start(self) -> Any:
         if self.running:
@@ -51,6 +57,9 @@ class ProxyNodeHealthScheduler:
 
     async def _scheduled_check(self) -> None:
         await self._check_heartbeats()
+        self._check_count = (self._check_count + 1) % _EVENT_CLEANUP_INTERVAL
+        if self._check_count == 0:
+            await self._cleanup_old_events()
 
     async def _check_heartbeats(self) -> None:
         from src.services.proxy_node.tunnel_manager import get_tunnel_manager
@@ -87,16 +96,9 @@ class ProxyNodeHealthScheduler:
                         node.tunnel_connected_at = now
                     changed += 1
 
-                if actually_connected:
+                new_status = (
-                    new_status = ProxyNodeStatus.ONLINE
+                    ProxyNodeStatus.ONLINE if actually_connected else ProxyNodeStatus.OFFLINE
-                elif node.tunnel_connected_at:
+                )
-                    # tunnel 刚断开：给 60s 缓冲期标记为 UNHEALTHY
-                    elapsed = (now - node.tunnel_connected_at).total_seconds()
-                    new_status = (
-                        ProxyNodeStatus.UNHEALTHY if elapsed < 60 else ProxyNodeStatus.OFFLINE
-                    )
-                else:
-                    new_status = ProxyNodeStatus.OFFLINE
 
                 if node.status != new_status:
                     node.status = new_status
@@ -115,6 +117,42 @@ class ProxyNodeHealthScheduler:
         finally:
             db.close()
 
+    async def _cleanup_old_events(self) -> None:
+        """清理超过保留期的连接事件记录（在线程池中执行，避免阻塞事件循环）"""
+        import asyncio
+
+        def _sync_cleanup() -> None:
+            from datetime import timedelta
+
+            from src.models.database import ProxyNodeEvent
+
+            db = create_session()
+            try:
+                cutoff = datetime.now(timezone.utc) - timedelta(days=_EVENT_RETENTION_DAYS)
+                deleted = (
+                    db.query(ProxyNodeEvent)
+                    .filter(ProxyNodeEvent.created_at < cutoff)
+                    .delete(synchronize_session=False)
+                )
+                if deleted:
+                    db.commit()
+                    logger.info(
+                        "清理 {} 条过期代理节点事件 (>{} 天)", deleted, _EVENT_RETENTION_DAYS
+                    )
+            except Exception as e:
+                try:
+                    db.rollback()
+                except Exception:
+                    pass
+                logger.warning("清理代理节点事件失败: {}", e)
+            finally:
+                db.close()
+
+        try:
+            await asyncio.to_thread(_sync_cleanup)
+        except Exception as e:
+            logger.warning("清理代理节点事件线程执行失败: {}", e)
+
 
 _proxy_node_health_scheduler: ProxyNodeHealthScheduler | None = None
 

src/services/proxy_node/resolver.py

@@ -60,17 +60,39 @@ def _get_proxy_node_info(node_id: str) -> dict[str, Any] | None:
     db = create_session()
     try:
         node = db.query(ProxyNode).filter(ProxyNode.id == node_id).first()
-        if not node or node.status != ProxyNodeStatus.ONLINE:
+        if not node:
             _proxy_node_cache[node_id] = (None, now + _PROXY_NODE_CACHE_NEGATIVE_TTL_SECONDS)
             return None
 
-        # tunnel 模式节点必须 tunnel 已连接才可用
+        # tunnel 模式节点：以 TunnelManager 内存中的实际连接状态为准，
-        if node.tunnel_mode and not node.tunnel_connected:
+        # 而非依赖 DB 的 status/tunnel_connected 字段（可能因竞态不同步）。
+        if node.tunnel_mode and not node.is_manual:
+            from src.services.proxy_node.tunnel_manager import get_tunnel_manager
+
+            manager = get_tunnel_manager()
+            if not manager.has_tunnel(node_id):
+                _proxy_node_cache[node_id] = (
+                    None,
+                    now + _PROXY_NODE_CACHE_NEGATIVE_TTL_SECONDS,
+                )
+                return None
+            value: dict[str, Any] = {
+                "name": node.name,
+                "ip": node.ip,
+                "port": node.port,
+                "tunnel_mode": True,
+                "tunnel_connected": True,
+            }
+            _proxy_node_cache[node_id] = (value, now + _PROXY_NODE_CACHE_TTL_SECONDS)
+            return value
+
+        # 手动节点 / 非 tunnel 节点：仍依赖 DB status
+        if node.status != ProxyNodeStatus.ONLINE:
             _proxy_node_cache[node_id] = (None, now + _PROXY_NODE_CACHE_NEGATIVE_TTL_SECONDS)
             return None
 
         if node.is_manual:
-            value: dict[str, Any] = {
+            value = {
                 "is_manual": True,
                 "name": node.name,
                 "proxy_url": node.proxy_url,

src/services/proxy_node/service.py

@@ -58,6 +58,9 @@ def node_to_dict(node: ProxyNode) -> dict[str, Any]:
         "active_connections": node.active_connections,
         "total_requests": node.total_requests,
         "avg_latency_ms": node.avg_latency_ms,
+        "failed_requests": node.failed_requests,
+        "dns_failures": node.dns_failures,
+        "stream_errors": node.stream_errors,
         "hardware_info": node.hardware_info,
         "estimated_max_concurrency": node.estimated_max_concurrency,
         "remote_config": node.remote_config,
@@ -283,7 +286,7 @@ class ProxyNodeService:
                 port=port,
                 region=region,
                 # 新节点：等 tunnel 连接后才上线
-                status=ProxyNodeStatus.UNHEALTHY,
+                status=ProxyNodeStatus.OFFLINE,
                 registered_by=registered_by,
                 last_heartbeat_at=now,
                 heartbeat_interval=heartbeat_interval,
@@ -311,8 +314,16 @@ class ProxyNodeService:
         active_connections: int | None = None,
         total_requests: int | None = None,
         avg_latency_ms: float | None = None,
+        failed_requests: int | None = None,
+        dns_failures: int | None = None,
+        stream_errors: int | None = None,
     ) -> ProxyNode:
-        """处理节点心跳（仅 tunnel 模式节点，更新指标并修正状态不一致）"""
+        """处理节点心跳（仅 tunnel 模式节点，更新指标并修正状态不一致）
+
+        注意: total_requests / failed_requests / dns_failures / stream_errors
+        来自 Rust 端的区间增量（swap(0) 后上报），需要累加到 DB 而非覆盖。
+        active_connections 和 avg_latency_ms 是实时快照，直接覆盖。
+        """
         node = db.query(ProxyNode).filter(ProxyNode.id == node_id).first()
         if not node:
             raise NotFoundException(f"ProxyNode {node_id} 不存在", "proxy_node")
@@ -332,13 +343,23 @@ class ProxyNodeService:
         node.last_heartbeat_at = now
         if heartbeat_interval is not None:
             node.heartbeat_interval = heartbeat_interval
+
+        # 实时快照指标 -- 直接覆盖
         if active_connections is not None:
             node.active_connections = active_connections
-        if total_requests is not None:
-            node.total_requests = total_requests
         if avg_latency_ms is not None:
             node.avg_latency_ms = avg_latency_ms
 
+        # 区间增量指标 -- 累加到累计值
+        if total_requests is not None and total_requests > 0:
+            node.total_requests = (node.total_requests or 0) + total_requests
+        if failed_requests is not None and failed_requests > 0:
+            node.failed_requests = (node.failed_requests or 0) + failed_requests
+        if dns_failures is not None and dns_failures > 0:
+            node.dns_failures = (node.dns_failures or 0) + dns_failures
+        if stream_errors is not None and stream_errors > 0:
+            node.stream_errors = (node.stream_errors or 0) + stream_errors
+
         db.commit()
         db.refresh(node)
         return node
@@ -530,7 +551,12 @@ class ProxyNodeService:
 
         # tunnel 节点：通过 WebSocket tunnel 测试
         if not node.is_manual:
-            if not node.tunnel_connected:
+            # 以 TunnelManager 内存中的实际连接状态为准（与 health_scheduler 一致），
+            # 而非仅依赖 DB 的 tunnel_connected 字段，避免竞态导致误判。
+            from src.services.proxy_node.tunnel_manager import get_tunnel_manager
+
+            manager = get_tunnel_manager()
+            if not manager.has_tunnel(node.id):
                 return {
                     "success": False,
                     "latency_ms": None,

src/services/proxy_node/tunnel_manager.py

@@ -402,6 +402,9 @@ class TunnelManager:
                     active_connections=data.get("active_connections"),
                     total_requests=data.get("total_requests"),
                     avg_latency_ms=data.get("avg_latency_ms"),
+                    failed_requests=data.get("failed_requests"),
+                    dns_failures=data.get("dns_failures"),
+                    stream_errors=data.get("stream_errors"),
                 )
                 result: dict[str, Any] = {}
                 if node.remote_config: