test(handler): add comprehensive stream processor unit tests

- Update usage records table to display first_byte_time_ms metrics
- Improve request timeline visualization for latency tracking
- Extend usage types for new timing information

alembic/versions/20251215_1707_180e63a9c83a_add_first_byte_time_ms_to_usage_table.py

@@ -0,0 +1,28 @@
+"""add first_byte_time_ms to usage table
+
+Revision ID: 180e63a9c83a
+Revises: e9b3d63f0cbf
+Create Date: 2025-12-15 17:07:44.631032+00:00
+
+"""
+from alembic import op
+import sqlalchemy as sa
+
+
+# revision identifiers, used by Alembic.
+revision = '180e63a9c83a'
+down_revision = 'e9b3d63f0cbf'
+branch_labels = None
+depends_on = None
+
+
+def upgrade() -> None:
+    """应用迁移：升级到新版本"""
+    # 添加首字时间字段到 usage 表
+    op.add_column('usage', sa.Column('first_byte_time_ms', sa.Integer(), nullable=True))
+
+
+def downgrade() -> None:
+    """回滚迁移：降级到旧版本"""
+    # 删除首字时间字段
+    op.drop_column('usage', 'first_byte_time_ms')

docker-compose.yml

@@ -12,8 +12,6 @@ services:
       TZ: Asia/Shanghai
     volumes:
       - postgres_data:/var/lib/postgresql/data
-    ports:
-      - "${DB_PORT:-5432}:5432"
     healthcheck:
       test: ["CMD-SHELL", "pg_isready -U postgres"]
       interval: 5s
@@ -27,8 +25,6 @@ services:
     command: redis-server --appendonly yes --requirepass ${REDIS_PASSWORD}
     volumes:
       - redis_data:/data
-    ports:
-      - "${REDIS_PORT:-6379}:6379"
     healthcheck:
       test: ["CMD", "redis-cli", "--raw", "incr", "ping"]
       interval: 5s

frontend/src/api/cache.ts

@@ -290,6 +290,19 @@ export interface UnmappedEntry {
   ttl: number | null
 }
 
+// Provider 模型映射缓存（Redis 缓存）
+export interface ProviderModelMapping {
+  provider_id: string
+  provider_name: string
+  global_model_id: string
+  global_model_name: string
+  global_model_display_name: string | null
+  provider_model_name: string
+  aliases: string[] | null
+  ttl: number | null
+  hit_count: number
+}
+
 export interface ModelMappingCacheStats {
   available: boolean
   message?: string
@@ -303,6 +316,7 @@ export interface ModelMappingCacheStats {
     global_model_resolve: number
   }
   mappings?: ModelMappingItem[]
+  provider_model_mappings?: ProviderModelMapping[] | null
   unmapped?: UnmappedEntry[] | null
 }
 
@@ -337,5 +351,13 @@ export const modelMappingCacheApi = {
   async clearByName(modelName: string): Promise<ClearModelMappingCacheResponse> {
     const response = await api.delete(`/api/admin/monitoring/cache/model-mapping/${encodeURIComponent(modelName)}`)
     return response.data
+  },
+
+  /**
+   * 清除指定 Provider 和 GlobalModel 的映射缓存
+   */
+  async clearProviderModel(providerId: string, globalModelId: string): Promise<ClearModelMappingCacheResponse> {
+    const response = await api.delete(`/api/admin/monitoring/cache/model-mapping/provider/${providerId}/${globalModelId}`)
+    return response.data
   }
 }

frontend/src/features/usage/components/HorizontalRequestTimeline.vue

@@ -479,10 +479,25 @@ const groupedTimeline = computed<NodeGroup[]>(() => {
   return groups
 })
 
-// 计算链路总耗时（从第一个节点开始到最后一个节点结束）
+// 计算链路总耗时（使用成功候选的 latency_ms 字段）
+// 优先使用 latency_ms，因为它与 Usage.response_time_ms 使用相同的时间基准
+// 避免 finished_at - started_at 带来的额外延迟（数据库操作时间）
 const totalTraceLatency = computed(() => {
   if (!timeline.value || timeline.value.length === 0) return 0
 
+  // 查找成功的候选，使用其 latency_ms
+  const successCandidate = timeline.value.find(c => c.status === 'success')
+  if (successCandidate?.latency_ms != null) {
+    return successCandidate.latency_ms
+  }
+
+  // 如果没有成功的候选，查找失败但有 latency_ms 的候选
+  const failedWithLatency = timeline.value.find(c => c.status === 'failed' && c.latency_ms != null)
+  if (failedWithLatency?.latency_ms != null) {
+    return failedWithLatency.latency_ms
+  }
+
+  // 回退：使用 finished_at - started_at 计算
   let earliestStart: number | null = null
   let latestEnd: number | null = null
 

frontend/src/features/usage/components/UsageRecordsTable.vue

@@ -177,8 +177,9 @@
             费用
           </TableHead>
           <TableHead class="h-12 font-semibold w-[70px] text-right">
-            <div class="inline-block max-w-[2rem] leading-tight">
-              响应时间
+            <div class="flex flex-col items-end text-xs gap-0.5">
+              <span>首字</span>
+              <span class="text-muted-foreground font-normal">总耗时</span>
             </div>
           </TableHead>
         </TableRow>
@@ -356,15 +357,28 @@
             </div>
           </TableCell>
           <TableCell class="text-right py-4 w-[70px]">
-            <span
+            <div
               v-if="record.status === 'pending' || record.status === 'streaming'"
-              class="text-primary tabular-nums"
+              class="flex flex-col items-end text-xs gap-0.5"
             >
-              {{ getElapsedTime(record) }}
-            </span>
-            <span v-else-if="record.response_time_ms">
-              {{ (record.response_time_ms / 1000).toFixed(2) }}s
-            </span>
+              <span class="text-primary tabular-nums">
+                {{ getElapsedTime(record) }}
+              </span>
+            </div>
+            <div
+              v-else-if="record.response_time_ms != null"
+              class="flex flex-col items-end text-xs gap-0.5"
+            >
+              <span
+                v-if="record.first_byte_time_ms != null"
+                class="tabular-nums"
+              >{{ (record.first_byte_time_ms / 1000).toFixed(2) }}s</span>
+              <span
+                v-else
+                class="text-muted-foreground"
+              >-</span>
+              <span class="text-muted-foreground tabular-nums">{{ (record.response_time_ms / 1000).toFixed(2) }}s</span>
+            </div>
             <span
               v-else
               class="text-muted-foreground"

frontend/src/features/usage/types.ts

@@ -78,6 +78,7 @@ export interface UsageRecord {
   cost: number
   actual_cost?: number
   response_time_ms?: number
+  first_byte_time_ms?: number  // 首字时间 (TTFB)
   is_stream: boolean
   status_code?: number
   error_message?: string

frontend/src/views/admin/CacheMonitoring.vue

@@ -299,6 +299,26 @@ async function clearModelMappingByName(modelName: string) {
   }
 }
 
+async function clearProviderModelMapping(providerId: string, globalModelId: string, displayName?: string) {
+  const confirmed = await showConfirm({
+    title: '确认清除',
+    message: `确定要清除 ${displayName || 'Provider 模型映射'} 的缓存吗？`,
+    confirmText: '确认清除',
+    variant: 'destructive'
+  })
+
+  if (!confirmed) return
+
+  try {
+    await modelMappingCacheApi.clearProviderModel(providerId, globalModelId)
+    showSuccess('已清除 Provider 模型映射缓存')
+    await fetchModelMappingStats()
+  } catch (error) {
+    showError('清除缓存失败')
+    log.error('清除 Provider 模型映射缓存失败', error)
+  }
+}
+
 function formatTTL(ttl: number | null): string {
   if (ttl === null || ttl < 0) return '-'
   if (ttl < 60) return `${ttl}s`
@@ -872,9 +892,125 @@ onBeforeUnmount(() => {
         </div>
       </div>
 
+      <!-- Provider 模型映射缓存 -->
+      <div
+        v-if="modelMappingStats?.available && modelMappingStats.provider_model_mappings && modelMappingStats.provider_model_mappings.length > 0"
+        class="border-t border-border/40"
+      >
+        <div class="px-6 py-3 text-xs text-muted-foreground border-b border-border/30 bg-muted/20">
+          Provider 模型映射缓存
+        </div>
+        <!-- 桌面端表格 -->
+        <Table class="hidden md:table">
+          <TableHeader>
+            <TableRow>
+              <TableHead class="w-[15%]">
+                提供商
+              </TableHead>
+              <TableHead class="w-[25%]">
+                请求名称
+              </TableHead>
+              <TableHead class="w-8 text-center" />
+              <TableHead class="w-[25%]">
+                映射模型
+              </TableHead>
+              <TableHead class="w-[10%] text-center">
+                剩余
+              </TableHead>
+              <TableHead class="w-[10%] text-center">
+                次数
+              </TableHead>
+              <TableHead class="w-[7%] text-right">
+                操作
+              </TableHead>
+            </TableRow>
+          </TableHeader>
+          <TableBody>
+            <template
+              v-for="(mapping, index) in modelMappingStats.provider_model_mappings"
+              :key="index"
+            >
+              <TableRow
+                v-for="(alias, aliasIndex) in (mapping.aliases || [])"
+                :key="`${index}-${aliasIndex}`"
+              >
+                <TableCell>
+                  <Badge variant="outline" class="text-xs">
+                    {{ mapping.provider_name }}
+                  </Badge>
+                </TableCell>
+                <TableCell>
+                  <span class="text-sm font-mono">{{ alias }}</span>
+                </TableCell>
+                <TableCell class="text-center">
+                  <ArrowRight class="h-4 w-4 text-muted-foreground" />
+                </TableCell>
+                <TableCell>
+                  <span class="text-sm font-mono font-medium">{{ mapping.provider_model_name }}</span>
+                </TableCell>
+                <TableCell class="text-center">
+                  <span class="text-xs text-muted-foreground">{{ formatTTL(mapping.ttl) }}</span>
+                </TableCell>
+                <TableCell class="text-center">
+                  <span class="text-sm">{{ mapping.hit_count || 0 }}</span>
+                </TableCell>
+                <TableCell class="text-right">
+                  <Button
+                    size="icon"
+                    variant="ghost"
+                    class="h-7 w-7 text-muted-foreground/70 hover:text-destructive"
+                    title="清除缓存"
+                    @click="clearProviderModelMapping(mapping.provider_id, mapping.global_model_id, `${mapping.provider_name} - ${alias}`)"
+                  >
+                    <Trash2 class="h-3.5 w-3.5" />
+                  </Button>
+                </TableCell>
+              </TableRow>
+            </template>
+          </TableBody>
+        </Table>
+        <!-- 移动端卡片 -->
+        <div class="md:hidden divide-y divide-border/40">
+          <template
+            v-for="(mapping, index) in modelMappingStats.provider_model_mappings"
+            :key="`m-pm-${index}`"
+          >
+            <div
+              v-for="(alias, aliasIndex) in (mapping.aliases || [])"
+              :key="`m-pm-${index}-${aliasIndex}`"
+              class="p-4 space-y-2"
+            >
+              <div class="flex items-center justify-between">
+                <Badge variant="outline" class="text-xs">
+                  {{ mapping.provider_name }}
+                </Badge>
+                <div class="flex items-center gap-2">
+                  <span class="text-xs text-muted-foreground">{{ formatTTL(mapping.ttl) }}</span>
+                  <span class="text-xs">{{ mapping.hit_count || 0 }}次</span>
+                  <Button
+                    size="icon"
+                    variant="ghost"
+                    class="h-6 w-6 text-muted-foreground/70 hover:text-destructive"
+                    title="清除缓存"
+                    @click="clearProviderModelMapping(mapping.provider_id, mapping.global_model_id, `${mapping.provider_name} - ${alias}`)"
+                  >
+                    <Trash2 class="h-3 w-3" />
+                  </Button>
+                </div>
+              </div>
+              <div class="flex items-center gap-2 text-sm">
+                <span class="font-mono">{{ alias }}</span>
+                <ArrowRight class="h-3.5 w-3.5 shrink-0 text-muted-foreground/60" />
+                <span class="font-mono font-medium">{{ mapping.provider_model_name }}</span>
+              </div>
+            </div>
+          </template>
+        </div>
+      </div>
+
       <!-- 无缓存状态 -->
       <div
-        v-else-if="modelMappingStats?.available && (!modelMappingStats.mappings || modelMappingStats.mappings.length === 0) && (!modelMappingStats.unmapped || modelMappingStats.unmapped.length === 0)"
+        v-else-if="modelMappingStats?.available && (!modelMappingStats.mappings || modelMappingStats.mappings.length === 0) && (!modelMappingStats.unmapped || modelMappingStats.unmapped.length === 0) && (!modelMappingStats.provider_model_mappings || modelMappingStats.provider_model_mappings.length === 0)"
         class="px-6 py-8 text-center text-sm text-muted-foreground"
       >
         暂无模型解析缓存

src/api/admin/monitoring/cache.py

@@ -12,6 +12,7 @@ from fastapi.responses import PlainTextResponse
 from sqlalchemy.orm import Session
 
 from src.api.base.admin_adapter import AdminApiAdapter
+from src.api.base.context import ApiRequestContext
 from src.api.base.pagination import PaginationMeta, build_pagination_payload, paginate_sequence
 from src.api.base.pipeline import ApiRequestPipeline
 from src.clients.redis_client import get_redis_client_sync
@@ -87,19 +88,19 @@ def resolve_user_identifier(db: Session, identifier: str) -> Optional[str]:
     # 2. 尝试作为 Username 查询
     user = db.query(User).filter(User.username == identifier).first()
     if user:
-        logger.debug(f"通过Username解析: {identifier} -> {user.id[:8]}...")
+        logger.debug(f"通过Username解析: {identifier} -> {user.id[:8]}...")  # type: ignore[index]
         return user.id
 
     # 3. 尝试作为 Email 查询
     user = db.query(User).filter(User.email == identifier).first()
     if user:
-        logger.debug(f"通过Email解析: {identifier} -> {user.id[:8]}...")
+        logger.debug(f"通过Email解析: {identifier} -> {user.id[:8]}...")  # type: ignore[index]
         return user.id
 
     # 4. 尝试作为 API Key ID 查询
     api_key = db.query(ApiKey).filter(ApiKey.id == identifier).first()
     if api_key:
-        logger.debug(f"通过API Key ID解析: {identifier[:8]}... -> User ID: {api_key.user_id[:8]}...")
+        logger.debug(f"通过API Key ID解析: {identifier[:8]}... -> User ID: {api_key.user_id[:8]}...")  # type: ignore[index]
         return api_key.user_id
 
     # 无法识别
@@ -111,7 +112,7 @@ def resolve_user_identifier(db: Session, identifier: str) -> Optional[str]:
 async def get_cache_stats(
     request: Request,
     db: Session = Depends(get_db),
-):
+) -> Any:
     """
     获取缓存亲和性统计信息
 
@@ -131,7 +132,7 @@ async def get_user_affinity(
     user_identifier: str,
     request: Request,
     db: Session = Depends(get_db),
-):
+) -> Any:
     """
     查询指定用户的所有缓存亲和性
 
@@ -157,7 +158,7 @@ async def list_affinities(
     limit: int = Query(100, ge=1, le=1000, description="返回数量限制"),
     offset: int = Query(0, ge=0, description="偏移量"),
     db: Session = Depends(get_db),
-):
+) -> Any:
     """
     获取所有缓存亲和性列表，可选按关键词过滤
 
@@ -173,7 +174,7 @@ async def clear_user_cache(
     user_identifier: str,
     request: Request,
     db: Session = Depends(get_db),
-):
+) -> Any:
     """
     Clear cache affinity for a specific user
 
@@ -188,7 +189,7 @@ async def clear_user_cache(
 async def clear_all_cache(
     request: Request,
     db: Session = Depends(get_db),
-):
+) -> Any:
     """
     Clear all cache affinities
 
@@ -203,7 +204,7 @@ async def clear_provider_cache(
     provider_id: str,
     request: Request,
     db: Session = Depends(get_db),
-):
+) -> Any:
     """
     Clear cache affinities for a specific provider
 
@@ -218,7 +219,7 @@ async def clear_provider_cache(
 async def get_cache_config(
     request: Request,
     db: Session = Depends(get_db),
-):
+) -> Any:
     """
     获取缓存相关配置
 
@@ -234,7 +235,7 @@ async def get_cache_config(
 async def get_cache_metrics(
     request: Request,
     db: Session = Depends(get_db),
-):
+) -> Any:
     """
     以 Prometheus 文本格式暴露缓存调度指标，方便接入 Grafana。
     """
@@ -246,7 +247,7 @@ async def get_cache_metrics(
 
 
 class AdminCacheStatsAdapter(AdminApiAdapter):
-    async def handle(self, context):  # type: ignore[override]
+    async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
         try:
             redis_client = get_redis_client_sync()
             scheduler = await get_cache_aware_scheduler(redis_client)
@@ -266,7 +267,7 @@ class AdminCacheStatsAdapter(AdminApiAdapter):
 
 
 class AdminCacheMetricsAdapter(AdminApiAdapter):
-    async def handle(self, context):  # type: ignore[override]
+    async def handle(self, context: ApiRequestContext) -> PlainTextResponse:
         try:
             redis_client = get_redis_client_sync()
             scheduler = await get_cache_aware_scheduler(redis_client)
@@ -391,7 +392,7 @@ class AdminCacheMetricsAdapter(AdminApiAdapter):
 class AdminGetUserAffinityAdapter(AdminApiAdapter):
     user_identifier: str
 
-    async def handle(self, context):  # type: ignore[override]
+    async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
         db = context.db
         try:
             user_id = resolve_user_identifier(db, self.user_identifier)
@@ -472,7 +473,7 @@ class AdminListAffinitiesAdapter(AdminApiAdapter):
     limit: int
     offset: int
 
-    async def handle(self, context):  # type: ignore[override]
+    async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
         db = context.db
         redis_client = get_redis_client_sync()
         if not redis_client:
@@ -682,7 +683,7 @@ class AdminListAffinitiesAdapter(AdminApiAdapter):
 class AdminClearUserCacheAdapter(AdminApiAdapter):
     user_identifier: str
 
-    async def handle(self, context):  # type: ignore[override]
+    async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
         db = context.db
         try:
             redis_client = get_redis_client_sync()
@@ -786,7 +787,7 @@ class AdminClearUserCacheAdapter(AdminApiAdapter):
 
 
 class AdminClearAllCacheAdapter(AdminApiAdapter):
-    async def handle(self, context):  # type: ignore[override]
+    async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
         try:
             redis_client = get_redis_client_sync()
             affinity_mgr = await get_affinity_manager(redis_client)
@@ -806,7 +807,7 @@ class AdminClearAllCacheAdapter(AdminApiAdapter):
 class AdminClearProviderCacheAdapter(AdminApiAdapter):
     provider_id: str
 
-    async def handle(self, context):  # type: ignore[override]
+    async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
         try:
             redis_client = get_redis_client_sync()
             affinity_mgr = await get_affinity_manager(redis_client)
@@ -829,7 +830,7 @@ class AdminClearProviderCacheAdapter(AdminApiAdapter):
 
 
 class AdminCacheConfigAdapter(AdminApiAdapter):
-    async def handle(self, context):  # type: ignore[override]
+    async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
         from src.services.cache.affinity_manager import CacheAffinityManager
         from src.services.cache.aware_scheduler import CacheAwareScheduler
         from src.services.rate_limit.adaptive_reservation import get_adaptive_reservation_manager
@@ -878,7 +879,7 @@ class AdminCacheConfigAdapter(AdminApiAdapter):
 async def get_model_mapping_cache_stats(
     request: Request,
     db: Session = Depends(get_db),
-):
+) -> Any:
     """
     获取模型映射缓存统计信息
 
@@ -895,7 +896,7 @@ async def get_model_mapping_cache_stats(
 async def clear_all_model_mapping_cache(
     request: Request,
     db: Session = Depends(get_db),
-):
+) -> Any:
     """
     清除所有模型映射缓存
 
@@ -910,7 +911,7 @@ async def clear_model_mapping_cache_by_name(
     model_name: str,
     request: Request,
     db: Session = Depends(get_db),
-):
+) -> Any:
     """
     清除指定模型名称的映射缓存
 
@@ -921,8 +922,28 @@ async def clear_model_mapping_cache_by_name(
     return await pipeline.run(adapter=adapter, http_request=request, db=db, mode=adapter.mode)
 
 
+@router.delete("/model-mapping/provider/{provider_id}/{global_model_id}")
+async def clear_provider_model_mapping_cache(
+    provider_id: str,
+    global_model_id: str,
+    request: Request,
+    db: Session = Depends(get_db),
+) -> Any:
+    """
+    清除指定 Provider 和 GlobalModel 的模型映射缓存
+
+    参数:
+    - provider_id: Provider ID
+    - global_model_id: GlobalModel ID
+    """
+    adapter = AdminClearProviderModelMappingCacheAdapter(
+        provider_id=provider_id, global_model_id=global_model_id
+    )
+    return await pipeline.run(adapter=adapter, http_request=request, db=db, mode=adapter.mode)
+
+
 class AdminModelMappingCacheStatsAdapter(AdminApiAdapter):
-    async def handle(self, context):  # type: ignore[override]
+    async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
         import json
 
         from src.clients.redis_client import get_redis_client
@@ -955,7 +976,9 @@ class AdminModelMappingCacheStatsAdapter(AdminApiAdapter):
                 if key_str.startswith("model:id:"):
                     model_id_keys.append(key_str)
                 elif key_str.startswith("model:provider_global:"):
-                    provider_global_keys.append(key_str)
+                    # 过滤掉 hits 统计键，只保留实际的缓存键
+                    if not key_str.startswith("model:provider_global:hits:"):
+                        provider_global_keys.append(key_str)
 
             async for key in redis.scan_iter(match="global_model:*", count=100):
                 key_str = key.decode() if isinstance(key, bytes) else key
@@ -1067,6 +1090,85 @@ class AdminModelMappingCacheStatsAdapter(AdminApiAdapter):
             # 按 mapping_name 排序
             mappings.sort(key=lambda x: x["mapping_name"])
 
+            # 3. 解析 provider_global 缓存（Provider 级别的模型解析缓存）
+            provider_model_mappings = []
+            # 预加载 Provider 和 GlobalModel 数据
+            provider_map = {str(p.id): p for p in db.query(Provider).filter(Provider.is_active.is_(True)).all()}
+            global_model_map = {str(gm.id): gm for gm in db.query(GlobalModel).filter(GlobalModel.is_active.is_(True)).all()}
+
+            for key in provider_global_keys[:100]:  # 最多处理 100 个
+                # key 格式: model:provider_global:{provider_id}:{global_model_id}
+                try:
+                    parts = key.replace("model:provider_global:", "").split(":")
+                    if len(parts) != 2:
+                        continue
+                    provider_id, global_model_id = parts
+
+                    cached_value = await redis.get(key)
+                    ttl = await redis.ttl(key)
+
+                    # 获取命中次数
+                    hit_count_key = f"model:provider_global:hits:{provider_id}:{global_model_id}"
+                    hit_count_raw = await redis.get(hit_count_key)
+                    hit_count = int(hit_count_raw) if hit_count_raw else 0
+
+                    if cached_value:
+                        cached_str = (
+                            cached_value.decode()
+                            if isinstance(cached_value, bytes)
+                            else cached_value
+                        )
+                        try:
+                            cached_data = json.loads(cached_str)
+                            provider_model_name = cached_data.get("provider_model_name")
+                            provider_model_aliases = cached_data.get("provider_model_aliases", [])
+
+                            # 获取 Provider 和 GlobalModel 信息
+                            provider = provider_map.get(provider_id)
+                            global_model = global_model_map.get(global_model_id)
+
+                            if provider and global_model:
+                                # 提取别名名称
+                                alias_names = []
+                                if provider_model_aliases:
+                                    for alias_entry in provider_model_aliases:
+                                        if isinstance(alias_entry, dict) and alias_entry.get("name"):
+                                            alias_names.append(alias_entry["name"])
+
+                                # provider_model_name 为空时跳过
+                                if not provider_model_name:
+                                    continue
+
+                                # 只显示有实际映射的条目：
+                                # 1. 全局模型名 != Provider 模型名（模型名称映射）
+                                # 2. 或者有别名配置
+                                has_name_mapping = global_model.name != provider_model_name
+                                has_aliases = len(alias_names) > 0
+
+                                if has_name_mapping or has_aliases:
+                                    # 构建用于展示的别名列表
+                                    # 如果只有名称映射没有别名，则用 global_model_name 作为"请求名称"
+                                    display_aliases = alias_names if alias_names else [global_model.name]
+
+                                    provider_model_mappings.append({
+                                        "provider_id": provider_id,
+                                        "provider_name": provider.display_name or provider.name,
+                                        "global_model_id": global_model_id,
+                                        "global_model_name": global_model.name,
+                                        "global_model_display_name": global_model.display_name,
+                                        "provider_model_name": provider_model_name,
+                                        "aliases": display_aliases,
+                                        "ttl": ttl if ttl > 0 else None,
+                                        "hit_count": hit_count,
+                                    })
+                        except json.JSONDecodeError:
+                            pass
+                except Exception as e:
+                    logger.warning(f"解析 provider_global 缓存键 {key} 失败: {e}")
+
+            # 按 provider_name + global_model_name 排序
+            provider_model_mappings.sort(key=lambda x: (x["provider_name"], x["global_model_name"]))
+
             response_data = {
                 "available": True,
                 "ttl_seconds": CacheTTL.MODEL,
@@ -1079,6 +1181,7 @@ class AdminModelMappingCacheStatsAdapter(AdminApiAdapter):
                     "global_model_resolve": len(global_model_resolve_keys),
                 },
                 "mappings": mappings,
+                "provider_model_mappings": provider_model_mappings if provider_model_mappings else None,
                 "unmapped": unmapped_entries if unmapped_entries else None,
             }
 
@@ -1094,7 +1197,7 @@ class AdminModelMappingCacheStatsAdapter(AdminApiAdapter):
 
 
 class AdminClearAllModelMappingCacheAdapter(AdminApiAdapter):
-    async def handle(self, context):  # type: ignore[override]
+    async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
         from src.clients.redis_client import get_redis_client
 
         try:
@@ -1136,7 +1239,7 @@ class AdminClearAllModelMappingCacheAdapter(AdminApiAdapter):
 class AdminClearModelMappingCacheByNameAdapter(AdminApiAdapter):
     model_name: str
 
-    async def handle(self, context):  # type: ignore[override]
+    async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
         from src.clients.redis_client import get_redis_client
 
         try:
@@ -1176,3 +1279,55 @@ class AdminClearModelMappingCacheByNameAdapter(AdminApiAdapter):
         except Exception as exc:
             logger.exception(f"清除模型映射缓存失败: {exc}")
             raise HTTPException(status_code=500, detail=f"清除失败: {exc}")
+
+
+@dataclass
+class AdminClearProviderModelMappingCacheAdapter(AdminApiAdapter):
+    provider_id: str
+    global_model_id: str
+
+    async def handle(self, context: ApiRequestContext) -> Dict[str, Any]:  # type: ignore[override]
+        from src.clients.redis_client import get_redis_client
+
+        try:
+            redis = await get_redis_client(require_redis=False)
+            if not redis:
+                raise HTTPException(status_code=503, detail="Redis 未启用")
+
+            deleted_keys = []
+
+            # 清除 provider_global 缓存
+            provider_global_key = f"model:provider_global:{self.provider_id}:{self.global_model_id}"
+            if await redis.exists(provider_global_key):
+                await redis.delete(provider_global_key)
+                deleted_keys.append(provider_global_key)
+
+            # 清除对应的 hit_count 缓存
+            hit_count_key = f"model:provider_global:hits:{self.provider_id}:{self.global_model_id}"
+            if await redis.exists(hit_count_key):
+                await redis.delete(hit_count_key)
+                deleted_keys.append(hit_count_key)
+
+            logger.info(
+                f"已清除 Provider 模型映射缓存: provider_id={self.provider_id[:8]}..., "
+                f"global_model_id={self.global_model_id[:8]}..., 删除键={deleted_keys}"
+            )
+            context.add_audit_metadata(
+                action="provider_model_mapping_cache_clear",
+                provider_id=self.provider_id,
+                global_model_id=self.global_model_id,
+                deleted_keys=deleted_keys,
+            )
+            return {
+                "status": "ok",
+                "message": "已清除 Provider 模型映射缓存",
+                "provider_id": self.provider_id,
+                "global_model_id": self.global_model_id,
+                "deleted_keys": deleted_keys,
+            }
+
+        except HTTPException:
+            raise
+        except Exception as exc:
+            logger.exception(f"清除 Provider 模型映射缓存失败: {exc}")
+            raise HTTPException(status_code=500, detail=f"清除失败: {exc}")

src/api/admin/usage/routes.py

@@ -628,6 +628,7 @@ class AdminUsageRecordsAdapter(AdminApiAdapter):
                     "actual_cost": actual_cost,
                     "rate_multiplier": rate_multiplier,
                     "response_time_ms": usage.response_time_ms,
+                    "first_byte_time_ms": usage.first_byte_time_ms,  # 首字时间 (TTFB)
                     "created_at": usage.created_at.isoformat(),
                     "is_stream": usage.is_stream,
                     "input_price_per_1m": usage.input_price_per_1m,
@@ -738,6 +739,7 @@ class AdminUsageDetailAdapter(AdminApiAdapter):
             "status_code": usage_record.status_code,
             "error_message": usage_record.error_message,
             "response_time_ms": usage_record.response_time_ms,
+            "first_byte_time_ms": usage_record.first_byte_time_ms,  # 首字时间 (TTFB)
             "created_at": usage_record.created_at.isoformat() if usage_record.created_at else None,
             "request_headers": usage_record.request_headers,
             "request_body": usage_record.get_request_body(),

src/api/handlers/base/base_handler.py

@@ -100,6 +100,8 @@ class MessageTelemetry:
         cache_read_tokens: int = 0,
         is_stream: bool = False,
         provider_request_headers: Optional[Dict[str, Any]] = None,
+        # 时间指标
+        first_byte_time_ms: Optional[int] = None,  # 首字时间/TTFB
         # Provider 侧追踪信息（用于记录真实成本）
         provider_id: Optional[str] = None,
         provider_endpoint_id: Optional[str] = None,
@@ -133,6 +135,7 @@ class MessageTelemetry:
             api_format=api_format,
             is_stream=is_stream,
             response_time_ms=response_time_ms,
+            first_byte_time_ms=first_byte_time_ms,  # 传递首字时间
             status_code=status_code,
             request_headers=request_headers,
             request_body=request_body,
@@ -395,3 +398,24 @@ class BaseMessageHandler:
 
         # 创建后台任务，不阻塞当前流
         asyncio.create_task(_do_update())
+
+    def _log_request_error(self, message: str, error: Exception) -> None:
+        """记录请求错误日志，对业务异常不打印堆栈
+
+        Args:
+            message: 错误消息前缀
+            error: 异常对象
+        """
+        from src.core.exceptions import (
+            ProviderException,
+            QuotaExceededException,
+            RateLimitException,
+            ModelNotSupportedException,
+        )
+
+        if isinstance(error, (ProviderException, QuotaExceededException, RateLimitException, ModelNotSupportedException)):
+            # 业务异常：简洁日志，不打印堆栈
+            logger.error(f"{message}: [{type(error).__name__}] {error}")
+        else:
+            # 未知异常：完整堆栈
+            logger.exception(f"{message}: {error}")

src/api/handlers/base/chat_handler_base.py

@@ -34,6 +34,7 @@ from src.api.handlers.base.response_parser import ResponseParser
 from src.api.handlers.base.stream_context import StreamContext
 from src.api.handlers.base.stream_processor import StreamProcessor
 from src.api.handlers.base.stream_telemetry import StreamTelemetryRecorder
+from src.api.handlers.base.utils import build_sse_headers
 from src.config.settings import config
 from src.core.exceptions import (
     EmbeddedErrorException,
@@ -365,7 +366,7 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
                 ctx,
                 original_headers,
                 original_request_body,
-                self.elapsed_ms(),
+                self.start_time,  # 传入开始时间，让 telemetry 在流结束后计算响应时间
             )
 
             # 创建监控流
@@ -378,11 +379,12 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
             return StreamingResponse(
                 monitored_stream,
                 media_type="text/event-stream",
+                headers=build_sse_headers(),
                 background=background_tasks,
             )
 
         except Exception as e:
-            logger.exception(f"流式请求失败: {e}")
+            self._log_request_error("流式请求失败", e)
             await self._record_stream_failure(ctx, e, original_headers, original_request_body)
             raise
 
@@ -473,12 +475,13 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
 
             stream_response.raise_for_status()
 
-            # 创建行迭代器
-            line_iterator = stream_response.aiter_lines()
+            # 使用字节流迭代器（避免 aiter_lines 的性能问题）
+            # aiter_raw() 返回原始数据块，无缓冲，实现真正的流式传输
+            byte_iterator = stream_response.aiter_raw()
 
             # 预读检测嵌套错误
-            prefetched_lines = await stream_processor.prefetch_and_check_error(
-                line_iterator,
+            prefetched_chunks = await stream_processor.prefetch_and_check_error(
+                byte_iterator,
                 provider,
                 endpoint,
                 ctx,
@@ -503,13 +506,14 @@ class ChatHandlerBase(BaseMessageHandler, ABC):
             await http_client.aclose()
             raise
 
-        # 创建流生成器
+        # 创建流生成器（传入字节流迭代器）
         return stream_processor.create_response_stream(
             ctx,
-            line_iterator,
+            byte_iterator,
             response_ctx,
             http_client,
-            prefetched_lines,
+            prefetched_chunks,
+            start_time=self.start_time,
         )
 
     async def _record_stream_failure(

src/api/handlers/base/cli_handler_base.py

@@ -11,17 +11,15 @@ CLI Message Handler 通用基类
 """
 
 import asyncio
+import codecs
 import json
 import time
-from abc import abstractmethod
-from dataclasses import dataclass, field
 from typing import (
     Any,
     AsyncGenerator,
     Callable,
     Dict,
     Optional,
-    Tuple,
 )
 
 import httpx
@@ -35,6 +33,8 @@ from src.api.handlers.base.base_handler import (
 )
 from src.api.handlers.base.parsers import get_parser_for_format
 from src.api.handlers.base.request_builder import PassthroughRequestBuilder
+from src.api.handlers.base.stream_context import StreamContext
+from src.api.handlers.base.utils import build_sse_headers
 
 # 直接从具体模块导入，避免循环依赖
 from src.api.handlers.base.response_parser import (
@@ -61,63 +61,6 @@ from src.services.provider.transport import build_provider_url
 from src.utils.sse_parser import SSEEventParser
 
 
-
-@dataclass
-class StreamContext:
-    """流式请求的上下文信息"""
-
-    # 请求信息
-    model: str = "unknown"  # 用户请求的原始模型名
-    mapped_model: Optional[str] = None  # 映射后的目标模型名（如果发生了映射）
-    api_format: str = ""
-    request_id: str = ""
-
-    # 用户信息（提前提取避免 Session detached）
-    user_id: int = 0
-    api_key_id: int = 0
-
-    # 统计信息
-    input_tokens: int = 0
-    output_tokens: int = 0
-    cached_tokens: int = 0  # cache_read_input_tokens
-    cache_creation_tokens: int = 0  # cache_creation_input_tokens
-    collected_text: str = ""
-    response_id: Optional[str] = None
-    final_usage: Optional[Dict[str, Any]] = None
-    final_response: Optional[Dict[str, Any]] = None
-    parsed_chunks: list = field(default_factory=list)
-
-    # 流状态
-    start_time: float = field(default_factory=time.time)
-    chunk_count: int = 0
-    data_count: int = 0
-    has_completion: bool = False
-
-    # 响应信息
-    status_code: int = 200
-    response_headers: Dict[str, str] = field(default_factory=dict)
-
-    # 请求信息（发送给 Provider 的）
-    provider_request_headers: Dict[str, str] = field(default_factory=dict)
-    provider_request_body: Optional[Dict[str, Any]] = None  # 实际发送的请求体
-
-    # Provider 信息
-    provider_name: Optional[str] = None
-    provider_id: Optional[str] = None  # Provider ID（用于记录真实成本）
-    endpoint_id: Optional[str] = None
-    key_id: Optional[str] = None
-    attempt_id: Optional[str] = None
-    attempt_synced: bool = False
-    error_message: Optional[str] = None
-
-    # 格式转换信息
-    provider_api_format: str = ""  # Provider 的 API 格式（用于响应转换）
-    client_api_format: str = ""  # 客户端请求的 API 格式
-
-    # Provider 响应元数据（存储 provider 返回的额外信息，如 Gemini 的 modelVersion）
-    response_metadata: Dict[str, Any] = field(default_factory=dict)
-
-
 class CliMessageHandlerBase(BaseMessageHandler):
     """
     CLI 格式消息处理器基类
@@ -409,24 +352,12 @@ class CliMessageHandlerBase(BaseMessageHandler):
             return StreamingResponse(
                 monitored_stream,
                 media_type="text/event-stream",
+                headers=build_sse_headers(),
                 background=background_tasks,
             )
 
         except Exception as e:
-            # 对于已知的业务异常，只记录简洁的错误信息，不输出完整堆栈
-            from src.core.exceptions import (
-                ProviderException,
-                QuotaExceededException,
-                RateLimitException,
-                ModelNotSupportedException,
-            )
-
-            if isinstance(e, (ProviderException, QuotaExceededException, RateLimitException, ModelNotSupportedException)):
-                # 业务异常：简洁日志
-                logger.error(f"流式请求失败: [{type(e).__name__}] {e}")
-            else:
-                # 未知异常：完整堆栈
-                logger.exception(f"流式请求失败: {e}")
+            self._log_request_error("流式请求失败", e)
             await self._record_stream_failure(ctx, e, original_headers, original_request_body)
             raise
 
@@ -446,7 +377,7 @@ class CliMessageHandlerBase(BaseMessageHandler):
         ctx.chunk_count = 0
         ctx.data_count = 0
         ctx.has_completion = False
-        ctx.collected_text = ""
+        ctx._collected_text_parts = []  # 重置文本收集
         ctx.input_tokens = 0
         ctx.output_tokens = 0
         ctx.cached_tokens = 0
@@ -534,12 +465,12 @@ class CliMessageHandlerBase(BaseMessageHandler):
 
             stream_response.raise_for_status()
 
-            # 创建行迭代器（只创建一次，后续会继续使用）
-            line_iterator = stream_response.aiter_lines()
+            # 使用字节流迭代器（避免 aiter_lines 的性能问题）
+            byte_iterator = stream_response.aiter_raw()
 
             # 预读第一个数据块，检测嵌套错误（HTTP 200 但响应体包含错误）
-            prefetched_lines = await self._prefetch_and_check_embedded_error(
-                line_iterator, provider, endpoint, ctx
+            prefetched_chunks = await self._prefetch_and_check_embedded_error(
+                byte_iterator, provider, endpoint, ctx
             )
 
         except httpx.HTTPStatusError as e:
@@ -564,10 +495,10 @@ class CliMessageHandlerBase(BaseMessageHandler):
         # 创建流生成器（带预读数据，使用同一个迭代器）
         return self._create_response_stream_with_prefetch(
             ctx,
-            line_iterator,
+            byte_iterator,
             response_ctx,
             http_client,
-            prefetched_lines,
+            prefetched_chunks,
         )
 
     async def _create_response_stream(
@@ -577,58 +508,75 @@ class CliMessageHandlerBase(BaseMessageHandler):
         response_ctx: Any,
         http_client: httpx.AsyncClient,
     ) -> AsyncGenerator[bytes, None]:
-        """创建响应流生成器"""
+        """创建响应流生成器（使用字节流）"""
         try:
             sse_parser = SSEEventParser()
             last_data_time = time.time()
             streaming_status_updated = False
+            buffer = b""
+            # 使用增量解码器处理跨 chunk 的 UTF-8 字符
+            decoder = codecs.getincrementaldecoder("utf-8")(errors="replace")
 
             # 检查是否需要格式转换
             needs_conversion = self._needs_format_conversion(ctx)
 
-            async for line in stream_response.aiter_lines():
+            async for chunk in stream_response.aiter_raw():
                 # 在第一次输出数据前更新状态为 streaming
                 if not streaming_status_updated:
                     self._update_usage_to_streaming(ctx.request_id)
                     streaming_status_updated = True
 
-                normalized_line = line.rstrip("\r")
-                events = sse_parser.feed_line(normalized_line)
-
-                if normalized_line == "":
-                    for event in events:
-                        self._handle_sse_event(
-                            ctx,
-                            event.get("event"),
-                            event.get("data") or "",
+                buffer += chunk
+                # 处理缓冲区中的完整行
+                while b"\n" in buffer:
+                    line_bytes, buffer = buffer.split(b"\n", 1)
+                    try:
+                        # 使用增量解码器，可以正确处理跨 chunk 的多字节字符
+                        line = decoder.decode(line_bytes + b"\n", False).rstrip("\n")
+                    except Exception as e:
+                        logger.warning(
+                            f"[{self.request_id}] UTF-8 解码失败: {e}, "
+                            f"bytes={line_bytes[:50]!r}"
                         )
-                    yield b"\n"
-                    continue
+                        continue
 
-                ctx.chunk_count += 1
+                    normalized_line = line.rstrip("\r")
+                    events = sse_parser.feed_line(normalized_line)
 
-                # 空流检测：超过阈值且无数据，发送错误事件并结束
-                if ctx.chunk_count > self.EMPTY_CHUNK_THRESHOLD and ctx.data_count == 0:
-                    elapsed = time.time() - last_data_time
-                    if elapsed > self.DATA_TIMEOUT:
-                        logger.warning(f"提供商 '{ctx.provider_name}' 流超时且无数据")
-                        error_event = {
-                            "type": "error",
-                            "error": {
-                                "type": "empty_stream_timeout",
-                                "message": f"提供商 '{ctx.provider_name}' 流超时且未返回有效数据",
-                            },
-                        }
-                        yield f"event: error\ndata: {json.dumps(error_event)}\n\n".encode("utf-8")
-                        return  # 结束生成器
+                    if normalized_line == "":
+                        for event in events:
+                            self._handle_sse_event(
+                                ctx,
+                                event.get("event"),
+                                event.get("data") or "",
+                            )
+                        yield b"\n"
+                        continue
 
-                # 格式转换或直接透传
-                if needs_conversion:
-                    converted_line = self._convert_sse_line(ctx, line, events)
-                    if converted_line:
-                        yield (converted_line + "\n").encode("utf-8")
-                else:
-                    yield (line + "\n").encode("utf-8")
+                    ctx.chunk_count += 1
+
+                    # 空流检测：超过阈值且无数据，发送错误事件并结束
+                    if ctx.chunk_count > self.EMPTY_CHUNK_THRESHOLD and ctx.data_count == 0:
+                        elapsed = time.time() - last_data_time
+                        if elapsed > self.DATA_TIMEOUT:
+                            logger.warning(f"提供商 '{ctx.provider_name}' 流超时且无数据")
+                            error_event = {
+                                "type": "error",
+                                "error": {
+                                    "type": "empty_stream_timeout",
+                                    "message": f"提供商 '{ctx.provider_name}' 流超时且未返回有效数据",
+                                },
+                            }
+                            yield f"event: error\ndata: {json.dumps(error_event)}\n\n".encode("utf-8")
+                            return  # 结束生成器
+
+                    # 格式转换或直接透传
+                    if needs_conversion:
+                        converted_line = self._convert_sse_line(ctx, line, events)
+                        if converted_line:
+                            yield (converted_line + "\n").encode("utf-8")
+                    else:
+                        yield (line + "\n").encode("utf-8")
 
                 for event in events:
                     self._handle_sse_event(
@@ -702,7 +650,7 @@ class CliMessageHandlerBase(BaseMessageHandler):
 
     async def _prefetch_and_check_embedded_error(
         self,
-        line_iterator: Any,
+        byte_iterator: Any,
         provider: Provider,
         endpoint: ProviderEndpoint,
         ctx: StreamContext,
@@ -716,20 +664,25 @@ class CliMessageHandlerBase(BaseMessageHandler):
         同时检测 HTML 响应（通常是 base_url 配置错误导致返回网页）。
 
         Args:
-            line_iterator: 行迭代器（aiter_lines() 返回的迭代器）
+            byte_iterator: 字节流迭代器
             provider: Provider 对象
             endpoint: Endpoint 对象
             ctx: 流上下文
 
         Returns:
-            预读的行列表（需要在后续流中先输出）
+            预读的字节块列表（需要在后续流中先输出）
 
         Raises:
             EmbeddedErrorException: 如果检测到嵌套错误
             ProviderNotAvailableException: 如果检测到 HTML 响应（配置错误）
         """
-        prefetched_lines: list = []
+        prefetched_chunks: list = []
         max_prefetch_lines = 5  # 最多预读5行来检测错误
+        buffer = b""
+        line_count = 0
+        should_stop = False
+        # 使用增量解码器处理跨 chunk 的 UTF-8 字符
+        decoder = codecs.getincrementaldecoder("utf-8")(errors="replace")
 
         try:
             # 获取对应格式的解析器
@@ -742,69 +695,86 @@ class CliMessageHandlerBase(BaseMessageHandler):
             else:
                 provider_parser = self.parser
 
-            line_count = 0
-            async for line in line_iterator:
-                prefetched_lines.append(line)
-                line_count += 1
+            async for chunk in byte_iterator:
+                prefetched_chunks.append(chunk)
+                buffer += chunk
 
-                # 解析数据
-                normalized_line = line.rstrip("\r")
+                # 尝试按行解析缓冲区
+                while b"\n" in buffer:
+                    line_bytes, buffer = buffer.split(b"\n", 1)
+                    try:
+                        # 使用增量解码器，可以正确处理跨 chunk 的多字节字符
+                        line = decoder.decode(line_bytes + b"\n", False).rstrip("\n")
+                    except Exception as e:
+                        logger.warning(
+                            f"[{self.request_id}] 预读时 UTF-8 解码失败: {e}, "
+                            f"bytes={line_bytes[:50]!r}"
+                        )
+                        continue
 
-                # 检测 HTML 响应（base_url 配置错误的常见症状）
-                lower_line = normalized_line.lower()
-                if lower_line.startswith("<!doctype") or lower_line.startswith("<html"):
-                    logger.error(
-                        f"  [{self.request_id}] 检测到 HTML 响应，可能是 base_url 配置错误: "
-                        f"Provider={provider.name}, Endpoint={endpoint.id[:8]}..., "
-                        f"base_url={endpoint.base_url}"
-                    )
-                    raise ProviderNotAvailableException(
-                        f"提供商 '{provider.name}' 返回了 HTML 页面而非 API 响应，请检查 endpoint 的 base_url 配置是否正确"
-                    )
+                    line_count += 1
+                    normalized_line = line.rstrip("\r")
 
-                if not normalized_line or normalized_line.startswith(":"):
-                    # 空行或注释行，继续预读
-                    if line_count >= max_prefetch_lines:
+                    # 检测 HTML 响应（base_url 配置错误的常见症状）
+                    lower_line = normalized_line.lower()
+                    if lower_line.startswith("<!doctype") or lower_line.startswith("<html"):
+                        logger.error(
+                            f"  [{self.request_id}] 检测到 HTML 响应，可能是 base_url 配置错误: "
+                            f"Provider={provider.name}, Endpoint={endpoint.id[:8]}..., "
+                            f"base_url={endpoint.base_url}"
+                        )
+                        raise ProviderNotAvailableException(
+                            f"提供商 '{provider.name}' 返回了 HTML 页面而非 API 响应，请检查 endpoint 的 base_url 配置是否正确"
+                        )
+
+                    if not normalized_line or normalized_line.startswith(":"):
+                        # 空行或注释行，继续预读
+                        if line_count >= max_prefetch_lines:
+                            break
+                        continue
+
+                    # 尝试解析 SSE 数据
+                    data_str = normalized_line
+                    if normalized_line.startswith("data: "):
+                        data_str = normalized_line[6:]
+
+                    if data_str == "[DONE]":
+                        should_stop = True
                         break
-                    continue
 
-                # 尝试解析 SSE 数据
-                data_str = normalized_line
-                if normalized_line.startswith("data: "):
-                    data_str = normalized_line[6:]
+                    try:
+                        data = json.loads(data_str)
+                    except json.JSONDecodeError:
+                        # 不是有效 JSON，可能是部分数据，继续
+                        if line_count >= max_prefetch_lines:
+                            break
+                        continue
 
-                if data_str == "[DONE]":
+                    # 使用解析器检查是否为错误响应
+                    if isinstance(data, dict) and provider_parser.is_error_response(data):
+                        # 提取错误信息
+                        parsed = provider_parser.parse_response(data, 200)
+                        logger.warning(f"  [{self.request_id}] 检测到嵌套错误: "
+                            f"Provider={provider.name}, "
+                            f"error_type={parsed.error_type}, "
+                            f"message={parsed.error_message}")
+                        raise EmbeddedErrorException(
+                            provider_name=str(provider.name),
+                            error_code=(
+                                int(parsed.error_type)
+                                if parsed.error_type and parsed.error_type.isdigit()
+                                else None
+                            ),
+                            error_message=parsed.error_message,
+                            error_status=parsed.error_type,
+                        )
+
+                    # 预读到有效数据，没有错误，停止预读
+                    should_stop = True
                     break
 
-                try:
-                    data = json.loads(data_str)
-                except json.JSONDecodeError:
-                    # 不是有效 JSON，可能是部分数据，继续
-                    if line_count >= max_prefetch_lines:
-                        break
-                    continue
-
-                # 使用解析器检查是否为错误响应
-                if isinstance(data, dict) and provider_parser.is_error_response(data):
-                    # 提取错误信息
-                    parsed = provider_parser.parse_response(data, 200)
-                    logger.warning(f"  [{self.request_id}] 检测到嵌套错误: "
-                        f"Provider={provider.name}, "
-                        f"error_type={parsed.error_type}, "
-                        f"message={parsed.error_message}")
-                    raise EmbeddedErrorException(
-                        provider_name=str(provider.name),
-                        error_code=(
-                            int(parsed.error_type)
-                            if parsed.error_type and parsed.error_type.isdigit()
-                            else None
-                        ),
-                        error_message=parsed.error_message,
-                        error_status=parsed.error_type,
-                    )
-
-                # 预读到有效数据，没有错误，停止预读
-                break
+                if should_stop or line_count >= max_prefetch_lines:
+                    break
 
         except EmbeddedErrorException:
             # 重新抛出嵌套错误
@@ -813,112 +783,168 @@ class CliMessageHandlerBase(BaseMessageHandler):
             # 其他异常（如网络错误）在预读阶段发生，记录日志但不中断
             logger.debug(f"  [{self.request_id}] 预读流时发生异常: {e}")
 
-        return prefetched_lines
+        return prefetched_chunks
 
     async def _create_response_stream_with_prefetch(
         self,
         ctx: StreamContext,
-        line_iterator: Any,
+        byte_iterator: Any,
         response_ctx: Any,
         http_client: httpx.AsyncClient,
-        prefetched_lines: list,
+        prefetched_chunks: list,
     ) -> AsyncGenerator[bytes, None]:
-        """创建响应流生成器（带预读数据）"""
+        """创建响应流生成器（带预读数据，使用字节流）"""
         try:
             sse_parser = SSEEventParser()
             last_data_time = time.time()
+            buffer = b""
+            first_yield = True  # 标记是否是第一次 yield
+            # 使用增量解码器处理跨 chunk 的 UTF-8 字符
+            decoder = codecs.getincrementaldecoder("utf-8")(errors="replace")
 
             # 检查是否需要格式转换
             needs_conversion = self._needs_format_conversion(ctx)
 
             # 在第一次输出数据前更新状态为 streaming
-            if prefetched_lines:
+            if prefetched_chunks:
                 self._update_usage_to_streaming(ctx.request_id)
 
-            # 先处理预读的数据
-            for line in prefetched_lines:
-                normalized_line = line.rstrip("\r")
-                events = sse_parser.feed_line(normalized_line)
+            # 先处理预读的字节块
+            for chunk in prefetched_chunks:
+                buffer += chunk
+                # 处理缓冲区中的完整行
+                while b"\n" in buffer:
+                    line_bytes, buffer = buffer.split(b"\n", 1)
+                    try:
+                        # 使用增量解码器，可以正确处理跨 chunk 的多字节字符
+                        line = decoder.decode(line_bytes + b"\n", False).rstrip("\n")
+                    except Exception as e:
+                        logger.warning(
+                            f"[{self.request_id}] UTF-8 解码失败: {e}, "
+                            f"bytes={line_bytes[:50]!r}"
+                        )
+                        continue
+
+                    normalized_line = line.rstrip("\r")
+                    events = sse_parser.feed_line(normalized_line)
+
+                    if normalized_line == "":
+                        for event in events:
+                            self._handle_sse_event(
+                                ctx,
+                                event.get("event"),
+                                event.get("data") or "",
+                            )
+                        # 记录首字时间 (第一次 yield)
+                        if first_yield:
+                            ctx.record_first_byte_time(self.start_time)
+                            first_yield = False
+                        yield b"\n"
+                        continue
+
+                    ctx.chunk_count += 1
+
+                    # 格式转换或直接透传
+                    if needs_conversion:
+                        converted_line = self._convert_sse_line(ctx, line, events)
+                        if converted_line:
+                            # 记录首字时间 (第一次 yield)
+                            if first_yield:
+                                ctx.record_first_byte_time(self.start_time)
+                                first_yield = False
+                            yield (converted_line + "\n").encode("utf-8")
+                    else:
+                        # 记录首字时间 (第一次 yield)
+                        if first_yield:
+                            ctx.record_first_byte_time(self.start_time)
+                            first_yield = False
+                        yield (line + "\n").encode("utf-8")
 
-                if normalized_line == "":
                     for event in events:
                         self._handle_sse_event(
                             ctx,
                             event.get("event"),
                             event.get("data") or "",
                         )
-                    yield b"\n"
-                    continue
 
-                ctx.chunk_count += 1
-
-                # 格式转换或直接透传
-                if needs_conversion:
-                    converted_line = self._convert_sse_line(ctx, line, events)
-                    if converted_line:
-                        yield (converted_line + "\n").encode("utf-8")
-                else:
-                    yield (line + "\n").encode("utf-8")
-
-                for event in events:
-                    self._handle_sse_event(
-                        ctx,
-                        event.get("event"),
-                        event.get("data") or "",
-                    )
-
-                if ctx.data_count > 0:
-                    last_data_time = time.time()
+                    if ctx.data_count > 0:
+                        last_data_time = time.time()
 
             # 继续处理剩余的流数据（使用同一个迭代器）
-            async for line in line_iterator:
-                normalized_line = line.rstrip("\r")
-                events = sse_parser.feed_line(normalized_line)
+            async for chunk in byte_iterator:
+                buffer += chunk
+                # 处理缓冲区中的完整行
+                while b"\n" in buffer:
+                    line_bytes, buffer = buffer.split(b"\n", 1)
+                    try:
+                        # 使用增量解码器，可以正确处理跨 chunk 的多字节字符
+                        line = decoder.decode(line_bytes + b"\n", False).rstrip("\n")
+                    except Exception as e:
+                        logger.warning(
+                            f"[{self.request_id}] UTF-8 解码失败: {e}, "
+                            f"bytes={line_bytes[:50]!r}"
+                        )
+                        continue
+
+                    normalized_line = line.rstrip("\r")
+                    events = sse_parser.feed_line(normalized_line)
+
+                    if normalized_line == "":
+                        for event in events:
+                            self._handle_sse_event(
+                                ctx,
+                                event.get("event"),
+                                event.get("data") or "",
+                            )
+                        # 记录首字时间 (第一次 yield) - 如果预读数据为空
+                        if first_yield:
+                            ctx.record_first_byte_time(self.start_time)
+                            first_yield = False
+                        yield b"\n"
+                        continue
+
+                    ctx.chunk_count += 1
+
+                    # 空流检测：超过阈值且无数据，发送错误事件并结束
+                    if ctx.chunk_count > self.EMPTY_CHUNK_THRESHOLD and ctx.data_count == 0:
+                        elapsed = time.time() - last_data_time
+                        if elapsed > self.DATA_TIMEOUT:
+                            logger.warning(f"提供商 '{ctx.provider_name}' 流超时且无数据")
+                            error_event = {
+                                "type": "error",
+                                "error": {
+                                    "type": "empty_stream_timeout",
+                                    "message": f"提供商 '{ctx.provider_name}' 流超时且未返回有效数据",
+                                },
+                            }
+                            yield f"event: error\ndata: {json.dumps(error_event)}\n\n".encode("utf-8")
+                            return
+
+                    # 格式转换或直接透传
+                    if needs_conversion:
+                        converted_line = self._convert_sse_line(ctx, line, events)
+                        if converted_line:
+                            # 记录首字时间 (第一次 yield) - 如果预读数据为空
+                            if first_yield:
+                                ctx.record_first_byte_time(self.start_time)
+                                first_yield = False
+                            yield (converted_line + "\n").encode("utf-8")
+                    else:
+                        # 记录首字时间 (第一次 yield) - 如果预读数据为空
+                        if first_yield:
+                            ctx.record_first_byte_time(self.start_time)
+                            first_yield = False
+                        yield (line + "\n").encode("utf-8")
 
-                if normalized_line == "":
                     for event in events:
                         self._handle_sse_event(
                             ctx,
                             event.get("event"),
                             event.get("data") or "",
                         )
-                    yield b"\n"
-                    continue
 
-                ctx.chunk_count += 1
-
-                # 空流检测：超过阈值且无数据，发送错误事件并结束
-                if ctx.chunk_count > self.EMPTY_CHUNK_THRESHOLD and ctx.data_count == 0:
-                    elapsed = time.time() - last_data_time
-                    if elapsed > self.DATA_TIMEOUT:
-                        logger.warning(f"提供商 '{ctx.provider_name}' 流超时且无数据")
-                        error_event = {
-                            "type": "error",
-                            "error": {
-                                "type": "empty_stream_timeout",
-                                "message": f"提供商 '{ctx.provider_name}' 流超时且未返回有效数据",
-                            },
-                        }
-                        yield f"event: error\ndata: {json.dumps(error_event)}\n\n".encode("utf-8")
-                        return
-
-                # 格式转换或直接透传
-                if needs_conversion:
-                    converted_line = self._convert_sse_line(ctx, line, events)
-                    if converted_line:
-                        yield (converted_line + "\n").encode("utf-8")
-                else:
-                    yield (line + "\n").encode("utf-8")
-
-                for event in events:
-                    self._handle_sse_event(
-                        ctx,
-                        event.get("event"),
-                        event.get("data") or "",
-                    )
-
-                if ctx.data_count > 0:
-                    last_data_time = time.time()
+                    if ctx.data_count > 0:
+                        last_data_time = time.time()
 
             # 处理剩余事件
             flushed_events = sse_parser.flush()
@@ -1047,7 +1073,7 @@ class CliMessageHandlerBase(BaseMessageHandler):
         # 提取文本内容
         text = self.parser.extract_text_content(data)
         if text:
-            ctx.collected_text += text
+            ctx.append_text(text)
 
         # 检查完成事件
         if event_type in ("response.completed", "message_stop"):
@@ -1099,9 +1125,11 @@ class CliMessageHandlerBase(BaseMessageHandler):
     ) -> None:
         """在流完成后记录统计信息"""
         try:
-            await asyncio.sleep(0.1)
+            # 使用 self.start_time 作为时间基准，与首字时间保持一致
+            # 注意：不要把统计延迟算进响应时间里
+            response_time_ms = int((time.time() - self.start_time) * 1000)
 
-            response_time_ms = int((time.time() - ctx.start_time) * 1000)
+            await asyncio.sleep(0.1)
 
             if not ctx.provider_name:
                 logger.warning(f"[{ctx.request_id}] 流式请求失败，未选中提供商")
@@ -1181,6 +1209,7 @@ class CliMessageHandlerBase(BaseMessageHandler):
                         input_tokens=actual_input_tokens,
                         output_tokens=ctx.output_tokens,
                         response_time_ms=response_time_ms,
+                        first_byte_time_ms=ctx.first_byte_time_ms,  # 传递首字时间
                         status_code=ctx.status_code,
                         request_headers=original_headers,
                         request_body=actual_request_body,
@@ -1201,9 +1230,18 @@ class CliMessageHandlerBase(BaseMessageHandler):
                         response_metadata=ctx.response_metadata if ctx.response_metadata else None,
                     )
                     logger.debug(f"{self.FORMAT_ID} 流式响应完成")
-                    # 简洁的请求完成摘要
-                    logger.info(f"[OK] {self.request_id[:8]} | {ctx.model} | {ctx.provider_name} | {response_time_ms}ms | "
-                        f"in:{ctx.input_tokens or 0} out:{ctx.output_tokens or 0}")
+                    # 简洁的请求完成摘要（两行格式）
+                    line1 = (
+                        f"[OK] {self.request_id[:8]} | {ctx.model} | {ctx.provider_name}"
+                    )
+                    if ctx.first_byte_time_ms:
+                        line1 += f" | TTFB: {ctx.first_byte_time_ms}ms"
+
+                    line2 = (
+                        f"      Total: {response_time_ms}ms | "
+                        f"in:{ctx.input_tokens or 0} out:{ctx.output_tokens or 0}"
+                    )
+                    logger.info(f"{line1}\n{line2}")
 
                 # 更新候选记录的最终状态和延迟时间
                 # 注意：RequestExecutor 会在流开始时过早地标记成功（只记录了连接建立的时间）
@@ -1255,7 +1293,8 @@ class CliMessageHandlerBase(BaseMessageHandler):
         original_request_body: Dict[str, Any],
     ) -> None:
         """记录流式请求失败"""
-        response_time_ms = int((time.time() - ctx.start_time) * 1000)
+        # 使用 self.start_time 作为时间基准，与首字时间保持一致
+        response_time_ms = int((time.time() - self.start_time) * 1000)
 
         status_code = 503
         if isinstance(error, ProviderAuthException):

src/api/handlers/base/parsers.py

@@ -13,6 +13,7 @@ from src.api.handlers.base.response_parser import (
     ResponseParser,
     StreamStats,
 )
+from src.api.handlers.base.utils import extract_cache_creation_tokens
 
 
 def _check_nested_error(response: Dict[str, Any]) -> Tuple[bool, Optional[Dict[str, Any]]]:
@@ -252,7 +253,7 @@ class ClaudeResponseParser(ResponseParser):
         usage = response.get("usage", {})
         result.input_tokens = usage.get("input_tokens", 0)
         result.output_tokens = usage.get("output_tokens", 0)
-        result.cache_creation_tokens = usage.get("cache_creation_input_tokens", 0)
+        result.cache_creation_tokens = extract_cache_creation_tokens(usage)
         result.cache_read_tokens = usage.get("cache_read_input_tokens", 0)
 
         # 检查错误（支持嵌套错误格式）
@@ -265,11 +266,16 @@ class ClaudeResponseParser(ResponseParser):
         return result
 
     def extract_usage_from_response(self, response: Dict[str, Any]) -> Dict[str, int]:
+        # 对于 message_start 事件，usage 在 message.usage 路径下
+        # 对于其他响应，usage 在顶层
         usage = response.get("usage", {})
+        if not usage and "message" in response:
+            usage = response.get("message", {}).get("usage", {})
+
         return {
             "input_tokens": usage.get("input_tokens", 0),
             "output_tokens": usage.get("output_tokens", 0),
-            "cache_creation_tokens": usage.get("cache_creation_input_tokens", 0),
+            "cache_creation_tokens": extract_cache_creation_tokens(usage),
             "cache_read_tokens": usage.get("cache_read_input_tokens", 0),
         }
 

src/api/handlers/base/stream_context.py

@@ -8,6 +8,7 @@
 - 请求/响应数据
 """
 
+import time
 from dataclasses import dataclass, field
 from typing import Any, Dict, List, Optional
 
@@ -25,12 +26,18 @@ class StreamContext:
     model: str
     api_format: str
 
+    # 请求标识信息（CLI handler 需要）
+    request_id: str = ""
+    user_id: int = 0
+    api_key_id: int = 0
+
     # Provider 信息（在请求执行时填充）
     provider_name: Optional[str] = None
     provider_id: Optional[str] = None
     endpoint_id: Optional[str] = None
     key_id: Optional[str] = None
     attempt_id: Optional[str] = None
+    attempt_synced: bool = False
     provider_api_format: Optional[str] = None  # Provider 的响应格式
 
     # 模型映射
@@ -43,7 +50,14 @@ class StreamContext:
     cache_creation_tokens: int = 0
 
     # 响应内容
-    collected_text: str = ""
+    _collected_text_parts: List[str] = field(default_factory=list, repr=False)
+    response_id: Optional[str] = None
+    final_usage: Optional[Dict[str, Any]] = None
+    final_response: Optional[Dict[str, Any]] = None
+
+    # 时间指标
+    first_byte_time_ms: Optional[int] = None  # 首字时间 (TTFB - Time To First Byte)
+    start_time: float = field(default_factory=time.time)
 
     # 响应状态
     status_code: int = 200
@@ -55,6 +69,12 @@ class StreamContext:
     provider_request_headers: Dict[str, str] = field(default_factory=dict)
     provider_request_body: Optional[Dict[str, Any]] = None
 
+    # 格式转换信息（CLI handler 需要）
+    client_api_format: str = ""
+
+    # Provider 响应元数据（CLI handler 需要）
+    response_metadata: Dict[str, Any] = field(default_factory=dict)
+
     # 流式处理统计
     data_count: int = 0
     chunk_count: int = 0
@@ -71,16 +91,30 @@ class StreamContext:
         self.chunk_count = 0
         self.data_count = 0
         self.has_completion = False
-        self.collected_text = ""
+        self._collected_text_parts = []
         self.input_tokens = 0
         self.output_tokens = 0
         self.cached_tokens = 0
         self.cache_creation_tokens = 0
         self.error_message = None
         self.status_code = 200
+        self.first_byte_time_ms = None
         self.response_headers = {}
         self.provider_request_headers = {}
         self.provider_request_body = None
+        self.response_id = None
+        self.final_usage = None
+        self.final_response = None
+
+    @property
+    def collected_text(self) -> str:
+        """已收集的文本内容（按需拼接，避免在流式过程中频繁做字符串拷贝）"""
+        return "".join(self._collected_text_parts)
+
+    def append_text(self, text: str) -> None:
+        """追加文本内容（仅在需要收集文本时调用）"""
+        if text:
+            self._collected_text_parts.append(text)
 
     def update_provider_info(
         self,
@@ -104,14 +138,40 @@ class StreamContext:
         cached_tokens: Optional[int] = None,
         cache_creation_tokens: Optional[int] = None,
     ) -> None:
-        """更新 Token 使用统计"""
-        if input_tokens is not None:
+        """
+        更新 Token 使用统计
+
+        采用防御性更新策略：只有当新值 > 0 或当前值为 0 时才更新，避免用 0 覆盖已有的正确值。
+
+        设计原理：
+        - 在流式响应中，某些事件可能不包含完整的 usage 信息（字段为 0 或不存在）
+        - 后续事件可能会提供完整的统计数据
+        - 通过这种策略，确保一旦获得非零值就保留它，不会被后续的 0 值覆盖
+
+        示例场景：
+        - message_start 事件：input_tokens=100, output_tokens=0
+        - message_delta 事件：input_tokens=0, output_tokens=50
+        - 最终结果：input_tokens=100, output_tokens=50
+
+        注意事项：
+        - 此策略假设初始值为 0 是正确的默认状态
+        - 如果需要将已有值重置为 0，请直接修改实例属性（不使用此方法）
+
+        Args:
+            input_tokens: 输入 tokens 数量
+            output_tokens: 输出 tokens 数量
+            cached_tokens: 缓存命中 tokens 数量
+            cache_creation_tokens: 缓存创建 tokens 数量
+        """
+        if input_tokens is not None and (input_tokens > 0 or self.input_tokens == 0):
             self.input_tokens = input_tokens
-        if output_tokens is not None:
+        if output_tokens is not None and (output_tokens > 0 or self.output_tokens == 0):
             self.output_tokens = output_tokens
-        if cached_tokens is not None:
+        if cached_tokens is not None and (cached_tokens > 0 or self.cached_tokens == 0):
             self.cached_tokens = cached_tokens
-        if cache_creation_tokens is not None:
+        if cache_creation_tokens is not None and (
+            cache_creation_tokens > 0 or self.cache_creation_tokens == 0
+        ):
             self.cache_creation_tokens = cache_creation_tokens
 
     def mark_failed(self, status_code: int, error_message: str) -> None:
@@ -119,6 +179,19 @@ class StreamContext:
         self.status_code = status_code
         self.error_message = error_message
 
+    def record_first_byte_time(self, start_time: float) -> None:
+        """
+        记录首字时间 (TTFB - Time To First Byte)
+
+        应在第一次向客户端发送数据时调用。
+        如果已记录过,则不会覆盖(避免重试时重复记录)。
+
+        Args:
+            start_time: 请求开始时间 (time.time())
+        """
+        if self.first_byte_time_ms is None:
+            self.first_byte_time_ms = int((time.time() - start_time) * 1000)
+
     def is_success(self) -> bool:
         """检查请求是否成功"""
         return self.status_code < 400
@@ -145,10 +218,22 @@ class StreamContext:
         获取日志摘要
 
         用于请求完成/失败时的日志输出。
+        包含首字时间 (TTFB) 和总响应时间,分两行显示。
         """
         status = "OK" if self.is_success() else "FAIL"
-        return (
+
+        # 第一行:基本信息 + 首字时间
+        line1 = (
             f"[{status}] {request_id[:8]} | {self.model} | "
-            f"{self.provider_name or 'unknown'} | {response_time_ms}ms | "
+            f"{self.provider_name or 'unknown'}"
+        )
+        if self.first_byte_time_ms is not None:
+            line1 += f" | TTFB: {self.first_byte_time_ms}ms"
+
+        # 第二行:总响应时间 + tokens
+        line2 = (
+            f"      Total: {response_time_ms}ms | "
             f"in:{self.input_tokens} out:{self.output_tokens}"
         )
+
+        return f"{line1}\n{line2}"

src/api/handlers/base/stream_processor.py

@@ -9,7 +9,9 @@
 """
 
 import asyncio
+import codecs
 import json
+import time
 from typing import Any, AsyncGenerator, Callable, Optional
 
 import httpx
@@ -36,6 +38,8 @@ class StreamProcessor:
         request_id: str,
         default_parser: ResponseParser,
         on_streaming_start: Optional[Callable[[], None]] = None,
+        *,
+        collect_text: bool = False,
     ):
         """
         初始化流处理器
@@ -48,6 +52,7 @@ class StreamProcessor:
         self.request_id = request_id
         self.default_parser = default_parser
         self.on_streaming_start = on_streaming_start
+        self.collect_text = collect_text
 
     def get_parser_for_provider(self, ctx: StreamContext) -> ResponseParser:
         """
@@ -112,9 +117,10 @@ class StreamProcessor:
             )
 
         # 提取文本
-        text = parser.extract_text_content(data)
-        if text:
-            ctx.collected_text += text
+        if self.collect_text:
+            text = parser.extract_text_content(data)
+            if text:
+                ctx.append_text(text)
 
         # 检查完成
         event_type = event_name or data.get("type", "")
@@ -123,7 +129,7 @@ class StreamProcessor:
 
     async def prefetch_and_check_error(
         self,
-        line_iterator: Any,
+        byte_iterator: Any,
         provider: Provider,
         endpoint: ProviderEndpoint,
         ctx: StreamContext,
@@ -136,97 +142,126 @@ class StreamProcessor:
         这种情况需要在流开始输出之前检测，以便触发重试逻辑。
 
         Args:
-            line_iterator: 行迭代器
+            byte_iterator: 字节流迭代器
             provider: Provider 对象
             endpoint: Endpoint 对象
             ctx: 流式上下文
             max_prefetch_lines: 最多预读行数
 
         Returns:
-            预读的行列表
+            预读的字节块列表
 
         Raises:
             EmbeddedErrorException: 如果检测到嵌套错误
         """
-        prefetched_lines: list = []
+        prefetched_chunks: list = []
         parser = self.get_parser_for_provider(ctx)
+        buffer = b""
+        line_count = 0
+        should_stop = False
+        # 使用增量解码器处理跨 chunk 的 UTF-8 字符
+        decoder = codecs.getincrementaldecoder("utf-8")(errors="replace")
 
         try:
-            line_count = 0
-            async for line in line_iterator:
-                prefetched_lines.append(line)
-                line_count += 1
+            async for chunk in byte_iterator:
+                prefetched_chunks.append(chunk)
+                buffer += chunk
 
-                normalized_line = line.rstrip("\r")
-                if not normalized_line or normalized_line.startswith(":"):
-                    if line_count >= max_prefetch_lines:
+                # 尝试按行解析缓冲区
+                while b"\n" in buffer:
+                    line_bytes, buffer = buffer.split(b"\n", 1)
+                    try:
+                        # 使用增量解码器，可以正确处理跨 chunk 的多字节字符
+                        line = decoder.decode(line_bytes + b"\n", False).rstrip("\r\n")
+                    except Exception as e:
+                        logger.warning(
+                            f"[{self.request_id}] 预读时 UTF-8 解码失败: {e}, "
+                            f"bytes={line_bytes[:50]!r}"
+                        )
+                        continue
+
+                    line_count += 1
+
+                    # 跳过空行和注释行
+                    if not line or line.startswith(":"):
+                        if line_count >= max_prefetch_lines:
+                            should_stop = True
+                            break
+                        continue
+
+                    # 尝试解析 SSE 数据
+                    data_str = line
+                    if line.startswith("data: "):
+                        data_str = line[6:]
+
+                    if data_str == "[DONE]":
+                        should_stop = True
                         break
-                    continue
 
-                # 尝试解析 SSE 数据
-                data_str = normalized_line
-                if normalized_line.startswith("data: "):
-                    data_str = normalized_line[6:]
+                    try:
+                        data = json.loads(data_str)
+                    except json.JSONDecodeError:
+                        if line_count >= max_prefetch_lines:
+                            should_stop = True
+                            break
+                        continue
 
-                if data_str == "[DONE]":
+                    # 使用解析器检查是否为错误响应
+                    if isinstance(data, dict) and parser.is_error_response(data):
+                        parsed = parser.parse_response(data, 200)
+                        logger.warning(
+                            f"  [{self.request_id}] 检测到嵌套错误: "
+                            f"Provider={provider.name}, "
+                            f"error_type={parsed.error_type}, "
+                            f"message={parsed.error_message}"
+                        )
+                        raise EmbeddedErrorException(
+                            provider_name=str(provider.name),
+                            error_code=(
+                                int(parsed.error_type)
+                                if parsed.error_type and parsed.error_type.isdigit()
+                                else None
+                            ),
+                            error_message=parsed.error_message,
+                            error_status=parsed.error_type,
+                        )
+
+                    # 预读到有效数据，没有错误，停止预读
+                    should_stop = True
                     break
 
-                try:
-                    data = json.loads(data_str)
-                except json.JSONDecodeError:
-                    if line_count >= max_prefetch_lines:
-                        break
-                    continue
-
-                # 使用解析器检查是否为错误响应
-                if isinstance(data, dict) and parser.is_error_response(data):
-                    parsed = parser.parse_response(data, 200)
-                    logger.warning(
-                        f"  [{self.request_id}] 检测到嵌套错误: "
-                        f"Provider={provider.name}, "
-                        f"error_type={parsed.error_type}, "
-                        f"message={parsed.error_message}"
-                    )
-                    raise EmbeddedErrorException(
-                        provider_name=str(provider.name),
-                        error_code=(
-                            int(parsed.error_type)
-                            if parsed.error_type and parsed.error_type.isdigit()
-                            else None
-                        ),
-                        error_message=parsed.error_message,
-                        error_status=parsed.error_type,
-                    )
-
-                # 预读到有效数据，没有错误，停止预读
-                break
+                if should_stop or line_count >= max_prefetch_lines:
+                    break
 
         except EmbeddedErrorException:
             raise
         except Exception as e:
             logger.debug(f"  [{self.request_id}] 预读流时发生异常: {e}")
 
-        return prefetched_lines
+        return prefetched_chunks
 
     async def create_response_stream(
         self,
         ctx: StreamContext,
-        line_iterator: Any,
+        byte_iterator: Any,
         response_ctx: Any,
         http_client: httpx.AsyncClient,
-        prefetched_lines: Optional[list] = None,
+        prefetched_chunks: Optional[list] = None,
+        *,
+        start_time: Optional[float] = None,
     ) -> AsyncGenerator[bytes, None]:
         """
         创建响应流生成器
 
-        统一的流生成器，支持带预读数据和不带预读数据两种情况。
+        从字节流中解析 SSE 数据并转发，支持预读数据。
 
         Args:
             ctx: 流式上下文
-            line_iterator: 行迭代器
+            byte_iterator: 字节流迭代器
             response_ctx: HTTP 响应上下文管理器
             http_client: HTTP 客户端
-            prefetched_lines: 预读的行列表（可选）
+            prefetched_chunks: 预读的字节块列表（可选）
+            start_time: 请求开始时间,用于计算 TTFB（可选）
 
         Yields:
             编码后的响应数据块
@@ -234,25 +269,82 @@ class StreamProcessor:
         try:
             sse_parser = SSEEventParser()
             streaming_started = False
+            buffer = b""
+            # 使用增量解码器处理跨 chunk 的 UTF-8 字符
+            decoder = codecs.getincrementaldecoder("utf-8")(errors="replace")
 
             # 处理预读数据
-            if prefetched_lines:
+            if prefetched_chunks:
                 if not streaming_started and self.on_streaming_start:
                     self.on_streaming_start()
                     streaming_started = True
 
-                for line in prefetched_lines:
-                    for chunk in self._process_line(ctx, sse_parser, line):
-                        yield chunk
+                for chunk in prefetched_chunks:
+                    # 记录首字时间 (TTFB) - 在 yield 之前记录
+                    if start_time is not None:
+                        ctx.record_first_byte_time(start_time)
+                        start_time = None  # 只记录一次
+
+                    # 把原始数据转发给客户端
+                    yield chunk
+
+                    buffer += chunk
+                    # 处理缓冲区中的完整行
+                    while b"\n" in buffer:
+                        line_bytes, buffer = buffer.split(b"\n", 1)
+                        try:
+                            # 使用增量解码器，可以正确处理跨 chunk 的多字节字符
+                            line = decoder.decode(line_bytes + b"\n", False)
+                            self._process_line(ctx, sse_parser, line)
+                        except Exception as e:
+                            # 解码失败，记录警告但继续处理
+                            logger.warning(
+                                f"[{self.request_id}] UTF-8 解码失败: {e}, "
+                                f"bytes={line_bytes[:50]!r}"
+                            )
+                            continue
 
             # 处理剩余的流数据
-            async for line in line_iterator:
+            async for chunk in byte_iterator:
                 if not streaming_started and self.on_streaming_start:
                     self.on_streaming_start()
                     streaming_started = True
 
-                for chunk in self._process_line(ctx, sse_parser, line):
-                    yield chunk
+                # 记录首字时间 (TTFB) - 在 yield 之前记录（如果预读数据为空）
+                if start_time is not None:
+                    ctx.record_first_byte_time(start_time)
+                    start_time = None  # 只记录一次
+
+                # 原始数据透传
+                yield chunk
+
+                buffer += chunk
+                # 处理缓冲区中的完整行
+                while b"\n" in buffer:
+                    line_bytes, buffer = buffer.split(b"\n", 1)
+                    try:
+                        # 使用增量解码器，可以正确处理跨 chunk 的多字节字符
+                        line = decoder.decode(line_bytes + b"\n", False)
+                        self._process_line(ctx, sse_parser, line)
+                    except Exception as e:
+                        # 解码失败，记录警告但继续处理
+                        logger.warning(
+                            f"[{self.request_id}] UTF-8 解码失败: {e}, "
+                            f"bytes={line_bytes[:50]!r}"
+                        )
+                        continue
+
+            # 处理剩余的缓冲区数据（如果有未完成的行）
+            if buffer:
+                try:
+                    # 使用 final=True 处理最后的不完整字符
+                    line = decoder.decode(buffer, True)
+                    self._process_line(ctx, sse_parser, line)
+                except Exception as e:
+                    logger.warning(
+                        f"[{self.request_id}] 处理剩余缓冲区失败: {e}, "
+                        f"bytes={buffer[:50]!r}"
+                    )
 
             # 处理剩余事件
             for event in sse_parser.flush():
@@ -268,7 +360,7 @@ class StreamProcessor:
         ctx: StreamContext,
         sse_parser: SSEEventParser,
         line: str,
-    ) -> list[bytes]:
+    ) -> None:
         """
         处理单行数据
 
@@ -276,26 +368,17 @@ class StreamProcessor:
             ctx: 流式上下文
             sse_parser: SSE 解析器
             line: 原始行数据
-
-        Returns:
-            要发送的数据块列表
         """
-        result: list[bytes] = []
-        normalized_line = line.rstrip("\r")
+        # SSEEventParser 以“去掉换行符”的单行文本作为输入；这里统一剔除 CR/LF，
+        # 避免把空行误判成 "\n" 并导致事件边界解析错误。
+        normalized_line = line.rstrip("\r\n")
         events = sse_parser.feed_line(normalized_line)
 
-        if normalized_line == "":
-            for event in events:
-                self.handle_sse_event(ctx, event.get("event"), event.get("data") or "")
-            result.append(b"\n")
-        else:
+        if normalized_line != "":
             ctx.chunk_count += 1
-            result.append((line + "\n").encode("utf-8"))
 
-            for event in events:
-                self.handle_sse_event(ctx, event.get("event"), event.get("data") or "")
-
-        return result
+        for event in events:
+            self.handle_sse_event(ctx, event.get("event"), event.get("data") or "")
 
     async def create_monitored_stream(
         self,
@@ -317,16 +400,26 @@ class StreamProcessor:
             响应数据块
         """
         try:
+            # 断连检查频率：每次 await 都会引入调度开销，过于频繁会让流式"发一段停一段"
+            # 这里按时间间隔节流，兼顾及时停止上游读取与吞吐平滑性。
+            next_disconnect_check_at = 0.0
+            disconnect_check_interval_s = 0.25
+
             async for chunk in stream_generator:
-                if await is_disconnected():
-                    logger.warning(f"ID:{self.request_id} | Client disconnected")
-                    ctx.status_code = 499  # Client Closed Request
-                    ctx.error_message = "client_disconnected"
-                    break
+                now = time.monotonic()
+                if now >= next_disconnect_check_at:
+                    next_disconnect_check_at = now + disconnect_check_interval_s
+                    if await is_disconnected():
+                        logger.warning(f"ID:{self.request_id} | Client disconnected")
+                        ctx.status_code = 499  # Client Closed Request
+                        ctx.error_message = "client_disconnected"
+
+                        break
                 yield chunk
         except asyncio.CancelledError:
             ctx.status_code = 499
             ctx.error_message = "client_disconnected"
+
             raise
         except Exception as e:
             ctx.status_code = 500

src/api/handlers/base/stream_telemetry.py

@@ -8,6 +8,7 @@
 """
 
 import asyncio
+import time
 from typing import Any, Dict, Optional
 
 from sqlalchemy.orm import Session
@@ -57,7 +58,7 @@ class StreamTelemetryRecorder:
         ctx: StreamContext,
         original_headers: Dict[str, str],
         original_request_body: Dict[str, Any],
-        response_time_ms: int,
+        start_time: float,
     ) -> None:
         """
         记录流式统计信息
@@ -66,11 +67,15 @@ class StreamTelemetryRecorder:
             ctx: 流式上下文
             original_headers: 原始请求头
             original_request_body: 原始请求体
-            response_time_ms: 响应时间（毫秒）
+            start_time: 请求开始时间 (time.time())
         """
         bg_db = None
 
         try:
+            # 在流结束后计算响应时间，与首字时间使用相同的时间基准
+            # 注意：不要把统计延迟（stream_stats_delay）算进响应时间里
+            response_time_ms = int((time.time() - start_time) * 1000)
+
             await asyncio.sleep(config.stream_stats_delay)  # 等待流完全关闭
 
             if not ctx.provider_name:
@@ -155,6 +160,7 @@ class StreamTelemetryRecorder:
             input_tokens=ctx.input_tokens,
             output_tokens=ctx.output_tokens,
             response_time_ms=response_time_ms,
+            first_byte_time_ms=ctx.first_byte_time_ms,  # 传递首字时间
             status_code=ctx.status_code,
             request_headers=original_headers,
             request_body=actual_request_body,

src/api/handlers/base/utils.py

@@ -0,0 +1,55 @@
+"""
+Handler 基础工具函数
+"""
+
+from typing import Any, Dict, Optional
+
+
+def extract_cache_creation_tokens(usage: Dict[str, Any]) -> int:
+    """
+    提取缓存创建 tokens（兼容新旧格式）
+
+    Claude API 在不同版本中使用了不同的字段名来表示缓存创建 tokens：
+    - 新格式（2024年后）：使用 claude_cache_creation_5_m_tokens 和
+      claude_cache_creation_1_h_tokens 分别表示 5 分钟和 1 小时缓存
+    - 旧格式：使用 cache_creation_input_tokens 表示总的缓存创建 tokens
+
+    此函数自动检测并适配两种格式，优先使用新格式。
+
+    Args:
+        usage: API 响应中的 usage 字典
+
+    Returns:
+        缓存创建 tokens 总数
+    """
+    # 检查新格式字段是否存在（而非值是否为 0）
+    # 如果字段存在，即使值为 0 也是合法的，不应 fallback 到旧格式
+    has_new_format = (
+        "claude_cache_creation_5_m_tokens" in usage
+        or "claude_cache_creation_1_h_tokens" in usage
+    )
+
+    if has_new_format:
+        cache_5m = usage.get("claude_cache_creation_5_m_tokens", 0)
+        cache_1h = usage.get("claude_cache_creation_1_h_tokens", 0)
+        return int(cache_5m) + int(cache_1h)
+
+    # 回退到旧格式
+    return int(usage.get("cache_creation_input_tokens", 0))
+
+
+def build_sse_headers(extra_headers: Optional[Dict[str, str]] = None) -> Dict[str, str]:
+    """
+    构建 SSE（text/event-stream）推荐响应头，用于减少代理缓冲带来的卡顿/成段输出。
+
+    说明：
+    - Cache-Control: no-transform 可避免部分代理对流做压缩/改写导致缓冲
+    - X-Accel-Buffering: no 可显式提示 Nginx 关闭缓冲（即使全局已关闭也无害）
+    """
+    headers: Dict[str, str] = {
+        "Cache-Control": "no-cache, no-transform",
+        "X-Accel-Buffering": "no",
+    }
+    if extra_headers:
+        headers.update(extra_headers)
+    return headers

src/api/handlers/claude/handler.py

@@ -8,6 +8,7 @@ Claude Chat Handler - 基于通用 Chat Handler 基类的简化实现
 from typing import Any, Dict, Optional
 
 from src.api.handlers.base.chat_handler_base import ChatHandlerBase
+from src.api.handlers.base.utils import extract_cache_creation_tokens
 
 
 class ClaudeChatHandler(ChatHandlerBase):
@@ -63,7 +64,7 @@ class ClaudeChatHandler(ChatHandlerBase):
         result["model"] = mapped_model
         return result
 
-    async def _convert_request(self, request):
+    async def _convert_request(self, request: Any) -> Any:
         """
         将请求转换为 Claude 格式
 
@@ -109,30 +110,18 @@ class ClaudeChatHandler(ChatHandlerBase):
         Claude 格式使用：
         - input_tokens / output_tokens
         - cache_creation_input_tokens / cache_read_input_tokens
+        - 新格式：claude_cache_creation_5_m_tokens / claude_cache_creation_1_h_tokens
         """
         usage = response.get("usage", {})
 
-        input_tokens = usage.get("input_tokens", 0)
-        output_tokens = usage.get("output_tokens", 0)
-        cache_creation_input_tokens = usage.get("cache_creation_input_tokens", 0)
-        cache_read_input_tokens = usage.get("cache_read_input_tokens", 0)
-
-        # 处理新的 cache_creation 格式
-        if "cache_creation" in usage:
-            cache_creation_data = usage.get("cache_creation", {})
-            if not cache_creation_input_tokens:
-                cache_creation_input_tokens = cache_creation_data.get(
-                    "ephemeral_5m_input_tokens", 0
-                ) + cache_creation_data.get("ephemeral_1h_input_tokens", 0)
-
         return {
-            "input_tokens": input_tokens,
-            "output_tokens": output_tokens,
-            "cache_creation_input_tokens": cache_creation_input_tokens,
-            "cache_read_input_tokens": cache_read_input_tokens,
+            "input_tokens": usage.get("input_tokens", 0),
+            "output_tokens": usage.get("output_tokens", 0),
+            "cache_creation_input_tokens": extract_cache_creation_tokens(usage),
+            "cache_read_input_tokens": usage.get("cache_read_input_tokens", 0),
         }
 
-    def _normalize_response(self, response: Dict) -> Dict:
+    def _normalize_response(self, response: Dict[str, Any]) -> Dict[str, Any]:
         """
         规范化 Claude 响应
 
@@ -143,8 +132,9 @@ class ClaudeChatHandler(ChatHandlerBase):
             规范化后的响应
         """
         if self.response_normalizer and self.response_normalizer.should_normalize(response):
-            return self.response_normalizer.normalize_claude_response(
+            result: Dict[str, Any] = self.response_normalizer.normalize_claude_response(
                 response_data=response,
                 request_id=self.request_id,
             )
+            return result
         return response

src/api/handlers/claude/stream_parser.py

@@ -9,6 +9,8 @@ from __future__ import annotations
 import json
 from typing import Any, Dict, List, Optional
 
+from src.api.handlers.base.utils import extract_cache_creation_tokens
+
 
 class ClaudeStreamParser:
     """
@@ -193,7 +195,7 @@ class ClaudeStreamParser:
                 return {
                     "input_tokens": usage.get("input_tokens", 0),
                     "output_tokens": usage.get("output_tokens", 0),
-                    "cache_creation_tokens": usage.get("cache_creation_input_tokens", 0),
+                    "cache_creation_tokens": extract_cache_creation_tokens(usage),
                     "cache_read_tokens": usage.get("cache_read_input_tokens", 0),
                 }
 
@@ -204,7 +206,7 @@ class ClaudeStreamParser:
                 return {
                     "input_tokens": usage.get("input_tokens", 0),
                     "output_tokens": usage.get("output_tokens", 0),
-                    "cache_creation_tokens": usage.get("cache_creation_input_tokens", 0),
+                    "cache_creation_tokens": extract_cache_creation_tokens(usage),
                     "cache_read_tokens": usage.get("cache_read_input_tokens", 0),
                 }
 

src/api/handlers/claude_cli/handler.py

@@ -11,6 +11,7 @@ from src.api.handlers.base.cli_handler_base import (
     CliMessageHandlerBase,
     StreamContext,
 )
+from src.api.handlers.base.utils import extract_cache_creation_tokens
 
 
 class ClaudeCliMessageHandler(CliMessageHandlerBase):
@@ -95,11 +96,12 @@ class ClaudeCliMessageHandler(CliMessageHandlerBase):
             usage = message.get("usage", {})
             if usage:
                 ctx.input_tokens = usage.get("input_tokens", 0)
-                # Claude 的缓存 tokens 使用不同的字段名
+
                 cache_read = usage.get("cache_read_input_tokens", 0)
                 if cache_read:
                     ctx.cached_tokens = cache_read
-                cache_creation = usage.get("cache_creation_input_tokens", 0)
+
+                cache_creation = extract_cache_creation_tokens(usage)
                 if cache_creation:
                     ctx.cache_creation_tokens = cache_creation
 
@@ -109,7 +111,7 @@ class ClaudeCliMessageHandler(CliMessageHandlerBase):
             if delta.get("type") == "text_delta":
                 text = delta.get("text", "")
                 if text:
-                    ctx.collected_text += text
+                    ctx.append_text(text)
 
         # 处理消息增量（包含最终 usage）
         elif event_type == "message_delta":
@@ -119,11 +121,15 @@ class ClaudeCliMessageHandler(CliMessageHandlerBase):
                     ctx.input_tokens = usage["input_tokens"]
                 if "output_tokens" in usage:
                     ctx.output_tokens = usage["output_tokens"]
-                # 更新缓存 tokens
+
+                # 更新缓存读取 tokens
                 if "cache_read_input_tokens" in usage:
                     ctx.cached_tokens = usage["cache_read_input_tokens"]
-                if "cache_creation_input_tokens" in usage:
-                    ctx.cache_creation_tokens = usage["cache_creation_input_tokens"]
+
+                # 更新缓存创建 tokens
+                cache_creation = extract_cache_creation_tokens(usage)
+                if cache_creation > 0:
+                    ctx.cache_creation_tokens = cache_creation
 
             # 检查是否结束
             delta = data.get("delta", {})

src/api/handlers/gemini_cli/handler.py

@@ -160,7 +160,7 @@ class GeminiCliMessageHandler(CliMessageHandlerBase):
             parts = content.get("parts", [])
             for part in parts:
                 if "text" in part:
-                    ctx.collected_text += part["text"]
+                    ctx.append_text(part["text"])
 
             # 检查结束原因
             finish_reason = candidate.get("finishReason")

src/api/handlers/openai_cli/handler.py

@@ -94,9 +94,9 @@ class OpenAICliMessageHandler(CliMessageHandlerBase):
         if event_type in ["response.output_text.delta", "response.outtext.delta"]:
             delta = data.get("delta")
             if isinstance(delta, str):
-                ctx.collected_text += delta
+                ctx.append_text(delta)
             elif isinstance(delta, dict) and "text" in delta:
-                ctx.collected_text += delta["text"]
+                ctx.append_text(delta["text"])
 
         # 处理完成事件
         elif event_type == "response.completed":
@@ -124,7 +124,7 @@ class OpenAICliMessageHandler(CliMessageHandlerBase):
                             if content_item.get("type") == "output_text":
                                 text = content_item.get("text", "")
                                 if text:
-                                    ctx.collected_text += text
+                                    ctx.append_text(text)
 
         # 备用：从顶层 usage 提取
         usage_obj = data.get("usage")

src/core/cache_service.py

@@ -120,6 +120,33 @@ class CacheService:
             logger.warning(f"缓存检查失败: {key} - {e}")
             return False
 
+    @staticmethod
+    async def incr(key: str, ttl_seconds: Optional[int] = None) -> int:
+        """
+        递增缓存值
+
+        Args:
+            key: 缓存键
+            ttl_seconds: 可选，如果提供则刷新 TTL
+
+        Returns:
+            递增后的值，如果失败返回 0
+        """
+        try:
+            redis = await get_redis_client(require_redis=False)
+            if not redis:
+                return 0
+
+            result = await redis.incr(key)
+            # 如果提供了 TTL，刷新过期时间
+            if ttl_seconds is not None:
+                await redis.expire(key, ttl_seconds)
+            return result
+
+        except Exception as e:
+            logger.warning(f"缓存递增失败: {key} - {e}")
+            return 0
+
 
 # 缓存键前缀
 class CacheKeys:

src/models/database.py

@@ -307,7 +307,8 @@ class Usage(Base):
     is_stream = Column(Boolean, default=False)  # 是否为流式请求
     status_code = Column(Integer)
     error_message = Column(Text, nullable=True)
-    response_time_ms = Column(Integer)  # 响应时间（毫秒）
+    response_time_ms = Column(Integer)  # 总响应时间（毫秒）
+    first_byte_time_ms = Column(Integer, nullable=True)  # 首字时间/TTFB（毫秒）
 
     # 请求状态追踪
     # pending: 请求开始处理中

src/services/cache/model_cache.py

@@ -2,11 +2,9 @@
 Model 映射缓存服务 - 减少模型查询
 """
 
-import json
 import time
-from typing import Optional
+from typing import List, Optional
 
-from sqlalchemy.exc import OperationalError, ProgrammingError
 from sqlalchemy.orm import Session
 
 from src.config.constants import CacheTTL
@@ -106,6 +104,7 @@ class ModelCacheService:
             Model 对象或 None
         """
         cache_key = f"model:provider_global:{provider_id}:{global_model_id}"
+        hit_count_key = f"model:provider_global:hits:{provider_id}:{global_model_id}"
 
         # 1. 尝试从缓存获取
         cached_data = await CacheService.get(cache_key)
@@ -113,6 +112,8 @@ class ModelCacheService:
             logger.debug(
                 f"Model 缓存命中(provider+global): {provider_id[:8]}...+{global_model_id[:8]}..."
             )
+            # 递增命中计数，同时刷新 TTL
+            await CacheService.incr(hit_count_key, ttl_seconds=ModelCacheService.CACHE_TTL)
             return ModelCacheService._dict_to_model(cached_data)
 
         # 2. 缓存未命中，查询数据库
@@ -130,6 +131,8 @@ class ModelCacheService:
         if model:
             model_dict = ModelCacheService._model_to_dict(model)
             await CacheService.set(cache_key, model_dict, ttl_seconds=ModelCacheService.CACHE_TTL)
+            # 重置命中计数（新缓存从1开始）
+            await CacheService.set(hit_count_key, 1, ttl_seconds=ModelCacheService.CACHE_TTL)
             logger.debug(
                 f"Model 已缓存(provider+global): {provider_id[:8]}...+{global_model_id[:8]}..."
             )
@@ -189,9 +192,10 @@ class ModelCacheService:
         # 清除 model:id 缓存
         await CacheService.delete(f"model:id:{model_id}")
 
-        # 清除 provider_global 缓存（如果提供了必要参数）
+        # 清除 provider_global 缓存及其命中计数（如果提供了必要参数）
         if provider_id and global_model_id:
             await CacheService.delete(f"model:provider_global:{provider_id}:{global_model_id}")
+            await CacheService.delete(f"model:provider_global:hits:{provider_id}:{global_model_id}")
             logger.debug(
                 f"Model 缓存已清除: {model_id}, provider_global:{provider_id[:8]}...:{global_model_id[:8]}..."
             )
@@ -230,16 +234,20 @@ class ModelCacheService:
         db: Session, model_name: str
     ) -> Optional[GlobalModel]:
         """
-        通过名称或映射解析 GlobalModel（带缓存，支持映射匹配）
+        通过名称解析 GlobalModel（带缓存）
 
         查找顺序：
         1. 检查缓存
-        2. 通过映射匹配（查询 Model 表的 provider_model_name 和 provider_model_aliases）
+        2. 通过 provider_model_name 匹配（查询 Model 表）
         3. 直接匹配 GlobalModel.name（兜底）
 
+        注意：此方法不使用 provider_model_aliases 进行全局解析。
+        provider_model_aliases 是 Provider 级别的别名配置，只在特定 Provider 上下文中生效，
+        由 resolve_provider_model() 处理。
+
         Args:
             db: 数据库会话
-            model_name: 模型名称（可以是 GlobalModel.name 或映射名称）
+            model_name: 模型名称（可以是 GlobalModel.name 或 provider_model_name）
 
         Returns:
             GlobalModel 对象或 None
@@ -273,116 +281,53 @@ class ModelCacheService:
                     logger.debug(f"GlobalModel 缓存命中(映射解析): {normalized_name}")
                     return ModelCacheService._dict_to_global_model(cached_data)
 
-            # 2. 优先通过 provider_model_name 和映射名称匹配（Provider 配置优先级最高）
-            from sqlalchemy import or_
-
+            # 2. 通过 provider_model_name 匹配（不考虑 provider_model_aliases）
+            # 重要：provider_model_aliases 是 Provider 级别的别名配置，只在特定 Provider 上下文中生效
+            # 全局解析不应该受到某个 Provider 别名配置的影响
+            # 例如：Provider A 把 "haiku" 映射到 "sonnet"，不应该影响 Provider B 的 "haiku" 解析
             from src.models.database import Provider
 
-            # 构建精确的映射匹配条件
-            # 注意：provider_model_aliases 是 JSONB 数组，需要使用 PostgreSQL 的 JSONB 操作符
-            # 对于 SQLite，会在 Python 层面进行过滤
-            try:
-                # 尝试使用 PostgreSQL 的 JSONB 查询（更高效）
-                # 使用 json.dumps 确保正确转义特殊字符，避免 SQL 注入
-                jsonb_pattern = json.dumps([{"name": normalized_name}])
-                models_with_global = (
-                    db.query(Model, GlobalModel)
-                    .join(Provider, Model.provider_id == Provider.id)
-                    .join(GlobalModel, Model.global_model_id == GlobalModel.id)
-                    .filter(
-                        Provider.is_active == True,
-                        Model.is_active == True,
-                        GlobalModel.is_active == True,
-                        or_(
-                            Model.provider_model_name == normalized_name,
-                            # PostgreSQL JSONB 查询：检查数组中是否有包含 {"name": "xxx"} 的元素
-                            Model.provider_model_aliases.op("@>")(jsonb_pattern),
-                        ),
-                    )
-                    .all()
-                )
-            except (OperationalError, ProgrammingError) as e:
-                # JSONB 操作符不支持（如 SQLite），回退到加载匹配 provider_model_name 的 Model
-                # 并在 Python 层过滤 aliases
-                logger.debug(
-                    f"JSONB 查询失败，回退到 Python 过滤: {e}",
-                )
-                # 优化：先用 provider_model_name 缩小范围，再加载其他可能匹配的记录
-                models_with_global = (
-                    db.query(Model, GlobalModel)
-                    .join(Provider, Model.provider_id == Provider.id)
-                    .join(GlobalModel, Model.global_model_id == GlobalModel.id)
-                    .filter(
-                        Provider.is_active == True,
-                        Model.is_active == True,
-                        GlobalModel.is_active == True,
-                    )
-                    .all()
+            models_with_global = (
+                db.query(Model, GlobalModel)
+                .join(Provider, Model.provider_id == Provider.id)
+                .join(GlobalModel, Model.global_model_id == GlobalModel.id)
+                .filter(
+                    Provider.is_active == True,
+                    Model.is_active == True,
+                    GlobalModel.is_active == True,
+                    Model.provider_model_name == normalized_name,
                 )
+                .all()
+            )
 
-            # 用于存储匹配结果：{(model_id, global_model_id): (GlobalModel, match_type, priority)}
-            # 使用字典去重，同一个 Model 只保留优先级最高的匹配
-            matched_models_dict = {}
-
-            # 遍历查询结果进行匹配
+            # 收集匹配的 GlobalModel（只通过 provider_model_name 匹配）
+            matched_global_models: List[GlobalModel] = []
+            seen_global_model_ids: set[str] = set()
             for model, gm in models_with_global:
-                key = (model.id, gm.id)
-
-                # 检查 provider_model_aliases 是否匹配（优先级更高）
-                if model.provider_model_aliases:
-                    for alias_entry in model.provider_model_aliases:
-                        if isinstance(alias_entry, dict):
-                            alias_name = alias_entry.get("name", "").strip()
-                            if alias_name == normalized_name:
-                                # alias 优先级为 0（最高），覆盖任何已存在的匹配
-                                matched_models_dict[key] = (gm, "alias", 0)
-                                logger.debug(
-                                    f"模型名称 '{normalized_name}' 通过映射名称匹配到 "
-                                    f"GlobalModel: {gm.name} (Model: {model.id[:8]}...)"
-                                )
-                                break
-
-                # 如果还没有匹配（或只有 provider_model_name 匹配），检查 provider_model_name
-                if key not in matched_models_dict or matched_models_dict[key][1] != "alias":
-                    if model.provider_model_name == normalized_name:
-                        # provider_model_name 优先级为 1（兜底），只在没有 alias 匹配时使用
-                        if key not in matched_models_dict:
-                            matched_models_dict[key] = (gm, "provider_model_name", 1)
-                            logger.debug(
-                                f"模型名称 '{normalized_name}' 通过 provider_model_name 匹配到 "
-                                f"GlobalModel: {gm.name} (Model: {model.id[:8]}...)"
-                            )
-
-            # 如果通过 provider_model_name/alias 找到了，直接返回
-            if matched_models_dict:
-                # 转换为列表并排序：按 priority（alias=0 优先）、然后按 GlobalModel.name
-                matched_global_models = [
-                    (gm, match_type) for gm, match_type, priority in matched_models_dict.values()
-                ]
-                matched_global_models.sort(
-                    key=lambda item: (
-                        0 if item[1] == "alias" else 1,  # alias 优先
-                        item[0].name  # 同优先级按名称排序（确定性）
+                if gm.id not in seen_global_model_ids:
+                    seen_global_model_ids.add(gm.id)
+                    matched_global_models.append(gm)
+                    logger.debug(
+                        f"模型名称 '{normalized_name}' 通过 provider_model_name 匹配到 "
+                        f"GlobalModel: {gm.name} (Model: {model.id[:8]}...)"
                     )
-                )
 
-                # 记录解析方式
-                resolution_method = matched_global_models[0][1]
+            # 如果通过 provider_model_name 找到了，返回
+            if matched_global_models:
+                resolution_method = "provider_model_name"
 
                 if len(matched_global_models) > 1:
-                    # 检测到冲突
-                    unique_models = {gm.id: gm for gm, _ in matched_global_models}
-                    if len(unique_models) > 1:
-                        model_names = [gm.name for gm in unique_models.values()]
-                        logger.warning(
-                            f"模型映射冲突: 名称 '{normalized_name}' 匹配到多个不同的 GlobalModel: "
-                            f"{', '.join(model_names)}，使用第一个匹配结果"
-                        )
-                        # 记录冲突指标
-                        model_mapping_conflict_total.inc()
+                    # 检测到冲突（多个不同的 GlobalModel 有相同的 provider_model_name）
+                    model_names = [gm.name for gm in matched_global_models if gm.name]
+                    logger.warning(
+                        f"模型映射冲突: 名称 '{normalized_name}' 匹配到多个不同的 GlobalModel: "
+                        f"{', '.join(model_names)}，使用第一个匹配结果"
+                    )
+                    # 记录冲突指标
+                    model_mapping_conflict_total.inc()
 
                 # 返回第一个匹配的 GlobalModel
-                result_global_model: GlobalModel = matched_global_models[0][0]
+                result_global_model = matched_global_models[0]
                 global_model_dict = ModelCacheService._global_model_to_dict(result_global_model)
                 await CacheService.set(
                     cache_key, global_model_dict, ttl_seconds=ModelCacheService.CACHE_TTL

src/services/provider/transport.py

@@ -6,7 +6,7 @@
 - 根据 API 格式或端点配置生成请求 URL
 """
 
-from typing import Any, Dict, Optional
+from typing import TYPE_CHECKING, Any, Dict, Optional
 from urllib.parse import urlencode
 
 from src.core.api_format_metadata import get_auth_config, get_default_path, resolve_api_format
@@ -14,11 +14,14 @@ from src.core.crypto import crypto_service
 from src.core.enums import APIFormat
 from src.core.logger import logger
 
+if TYPE_CHECKING:
+    from src.models.database import ProviderAPIKey, ProviderEndpoint
+
 
 
 def build_provider_headers(
-    endpoint,
-    key,
+    endpoint: "ProviderEndpoint",
+    key: "ProviderAPIKey",
     original_headers: Optional[Dict[str, str]] = None,
     *,
     extra_headers: Optional[Dict[str, str]] = None,
@@ -28,7 +31,8 @@ def build_provider_headers(
     """
     headers: Dict[str, str] = {}
 
-    decrypted_key = crypto_service.decrypt(key.api_key)
+    # api_key 在数据库中是 NOT NULL，类型标注为 Optional 是 SQLAlchemy 限制
+    decrypted_key = crypto_service.decrypt(key.api_key)  # type: ignore[arg-type]
 
     # 根据 API 格式自动选择认证头
     api_format = getattr(endpoint, "api_format", None)
@@ -68,8 +72,32 @@ def build_provider_headers(
     return headers
 
 
+def _normalize_base_url(base_url: str, path: str) -> str:
+    """
+    规范化 base_url，去除末尾的斜杠和可能与 path 重复的版本前缀。
+
+    只有当 path 以版本前缀开头时，才从 base_url 中移除该前缀，
+    避免拼接出 /v1/v1/messages 这样的重复路径。
+
+    兼容用户填写的各种格式：
+    - https://api.example.com
+    - https://api.example.com/
+    - https://api.example.com/v1
+    - https://api.example.com/v1/
+    """
+    base = base_url.rstrip("/")
+    # 只在 path 以版本前缀开头时才去除 base_url 中的该前缀
+    # 例如：base="/v1", path="/v1/messages" -> 去除 /v1
+    # 例如：base="/v1", path="/chat/completions" -> 不去除（用户可能期望保留）
+    for suffix in ("/v1beta", "/v1", "/v2", "/v3"):
+        if base.endswith(suffix) and path.startswith(suffix):
+            base = base[: -len(suffix)]
+            break
+    return base
+
+
 def build_provider_url(
-    endpoint,
+    endpoint: "ProviderEndpoint",
     *,
     query_params: Optional[Dict[str, Any]] = None,
     path_params: Optional[Dict[str, Any]] = None,
@@ -88,8 +116,6 @@ def build_provider_url(
         path_params: 路径模板参数 (如 {model})
         is_stream: 是否为流式请求，用于 Gemini API 选择正确的操作方法
     """
-    base = endpoint.base_url.rstrip("/")
-
     # 准备路径参数，添加 Gemini API 所需的 action 参数
     effective_path_params = dict(path_params) if path_params else {}
 
@@ -123,6 +149,9 @@ def build_provider_url(
     if not path.startswith("/"):
         path = f"/{path}"
 
+    # 先确定 path，再根据 path 规范化 base_url
+    # base_url 在数据库中是 NOT NULL，类型标注为 Optional 是 SQLAlchemy 限制
+    base = _normalize_base_url(endpoint.base_url, path)  # type: ignore[arg-type]
     url = f"{base}{path}"
 
     # 添加查询参数
@@ -134,7 +163,7 @@ def build_provider_url(
     return url
 
 
-def _resolve_default_path(api_format) -> str:
+def _resolve_default_path(api_format: Optional[str]) -> str:
     """
     根据 API 格式返回默认路径
     """

src/services/usage/service.py

@@ -157,6 +157,7 @@ class UsageService:
         api_format: Optional[str] = None,
         is_stream: bool = False,
         response_time_ms: Optional[int] = None,
+        first_byte_time_ms: Optional[int] = None,  # 首字时间 (TTFB)
         status_code: int = 200,
         error_message: Optional[str] = None,
         metadata: Optional[Dict[str, Any]] = None,
@@ -368,6 +369,7 @@ class UsageService:
             status_code=status_code,
             error_message=error_message,
             response_time_ms=response_time_ms,
+            first_byte_time_ms=first_byte_time_ms,  # 首字时间 (TTFB)
             status=status,  # 请求状态追踪
             request_metadata=metadata,
             request_headers=processed_request_headers,
@@ -419,6 +421,7 @@ class UsageService:
         api_format: Optional[str] = None,
         is_stream: bool = False,
         response_time_ms: Optional[int] = None,
+        first_byte_time_ms: Optional[int] = None,  # 首字时间 (TTFB)
         status_code: int = 200,
         error_message: Optional[str] = None,
         metadata: Optional[Dict[str, Any]] = None,
@@ -629,6 +632,7 @@ class UsageService:
             status_code=status_code,
             error_message=error_message,
             response_time_ms=response_time_ms,
+            first_byte_time_ms=first_byte_time_ms,  # 首字时间 (TTFB)
             status=status,  # 请求状态追踪
             request_metadata=metadata,
             request_headers=processed_request_headers,
@@ -649,6 +653,7 @@ class UsageService:
             existing_usage.status_code = status_code
             existing_usage.error_message = error_message
             existing_usage.response_time_ms = response_time_ms
+            existing_usage.first_byte_time_ms = first_byte_time_ms  # 更新首字时间
             # 更新请求头和请求体（如果有新值）
             if processed_request_headers is not None:
                 existing_usage.request_headers = processed_request_headers
@@ -1315,11 +1320,11 @@ class UsageService:
         default_timeout_seconds: int = 300,
     ) -> List[Dict[str, Any]]:
         """
-        获取活跃请求状态（用于前端轮询），并自动清理超时的 pending 请求
+        获取活跃请求状态（用于前端轮询），并自动清理超时的 pending/streaming 请求
 
         与 get_active_requests 不同，此方法：
         1. 返回轻量级的状态字典而非完整 Usage 对象
-        2. 自动检测并清理超时的 pending 请求
+        2. 自动检测并清理超时的 pending/streaming 请求
         3. 支持按 ID 列表查询特定请求
 
         Args:
@@ -1343,6 +1348,7 @@ class UsageService:
             Usage.output_tokens,
             Usage.total_cost_usd,
             Usage.response_time_ms,
+            Usage.first_byte_time_ms,  # 首字时间 (TTFB)
             Usage.created_at,
             Usage.provider_endpoint_id,
             ProviderEndpoint.timeout.label("endpoint_timeout"),
@@ -1361,10 +1367,10 @@ class UsageService:
 
         records = query.all()
 
-        # 检查超时的 pending 请求
+        # 检查超时的 pending/streaming 请求
         timeout_ids = []
         for r in records:
-            if r.status == "pending" and r.created_at:
+            if r.status in ("pending", "streaming") and r.created_at:
                 # 使用端点配置的超时时间，若无则使用默认值
                 timeout_seconds = r.endpoint_timeout or default_timeout_seconds
 
@@ -1392,6 +1398,7 @@ class UsageService:
                 "output_tokens": r.output_tokens,
                 "cost": float(r.total_cost_usd) if r.total_cost_usd else 0,
                 "response_time_ms": r.response_time_ms,
+                "first_byte_time_ms": r.first_byte_time_ms,  # 首字时间 (TTFB)
             }
             for r in records
         ]

tests/__init__.py

@@ -0,0 +1 @@
+"""测试模块"""

tests/api/handlers/base/test_stream_context.py

@@ -0,0 +1,117 @@
+from src.api.handlers.base import stream_context
+from src.api.handlers.base.stream_context import StreamContext
+
+
+def test_collected_text_append_and_property() -> None:
+    ctx = StreamContext(model="test-model", api_format="OPENAI")
+    assert ctx.collected_text == ""
+
+    ctx.append_text("hello")
+    ctx.append_text(" ")
+    ctx.append_text("world")
+    assert ctx.collected_text == "hello world"
+
+
+def test_reset_for_retry_clears_state() -> None:
+    ctx = StreamContext(model="test-model", api_format="OPENAI")
+    ctx.append_text("x")
+    ctx.update_usage(input_tokens=10, output_tokens=5)
+    ctx.parsed_chunks.append({"type": "chunk"})
+    ctx.chunk_count = 3
+    ctx.data_count = 2
+    ctx.has_completion = True
+    ctx.status_code = 418
+    ctx.error_message = "boom"
+
+    ctx.reset_for_retry()
+
+    assert ctx.collected_text == ""
+    assert ctx.input_tokens == 0
+    assert ctx.output_tokens == 0
+    assert ctx.parsed_chunks == []
+    assert ctx.chunk_count == 0
+    assert ctx.data_count == 0
+    assert ctx.has_completion is False
+    assert ctx.status_code == 200
+    assert ctx.error_message is None
+
+
+def test_record_first_byte_time(monkeypatch) -> None:
+    """测试记录首字时间"""
+    ctx = StreamContext(model="claude-3", api_format="claude_messages")
+    start_time = 100.0
+    monkeypatch.setattr(stream_context.time, "time", lambda: 100.0123)  # 12.3ms
+
+    # 记录首字时间
+    ctx.record_first_byte_time(start_time)
+
+    # 验证首字时间已记录
+    assert ctx.first_byte_time_ms == 12
+
+
+def test_record_first_byte_time_idempotent(monkeypatch) -> None:
+    """测试首字时间只记录一次"""
+    ctx = StreamContext(model="claude-3", api_format="claude_messages")
+    start_time = 100.0
+
+    # 第一次记录
+    monkeypatch.setattr(stream_context.time, "time", lambda: 100.010)
+    ctx.record_first_byte_time(start_time)
+    first_value = ctx.first_byte_time_ms
+
+    # 第二次记录（应该被忽略）
+    monkeypatch.setattr(stream_context.time, "time", lambda: 100.020)
+    ctx.record_first_byte_time(start_time)
+    second_value = ctx.first_byte_time_ms
+
+    # 验证值没有改变
+    assert first_value == second_value
+
+
+def test_reset_for_retry_clears_first_byte_time(monkeypatch) -> None:
+    """测试重试时清除首字时间"""
+    ctx = StreamContext(model="claude-3", api_format="claude_messages")
+    start_time = 100.0
+
+    # 记录首字时间
+    monkeypatch.setattr(stream_context.time, "time", lambda: 100.010)
+    ctx.record_first_byte_time(start_time)
+    assert ctx.first_byte_time_ms is not None
+
+    # 重置
+    ctx.reset_for_retry()
+
+    # 验证首字时间已清除
+    assert ctx.first_byte_time_ms is None
+
+
+def test_get_log_summary_with_first_byte_time() -> None:
+    """测试日志摘要包含首字时间"""
+    ctx = StreamContext(model="claude-3", api_format="claude_messages")
+    ctx.provider_name = "anthropic"
+    ctx.input_tokens = 100
+    ctx.output_tokens = 50
+    ctx.first_byte_time_ms = 123
+
+    summary = ctx.get_log_summary("request-id-123", 456)
+
+    # 验证包含首字时间和总时间（大写格式）
+    assert "TTFB: 123ms" in summary
+    assert "Total: 456ms" in summary
+    assert "in:100 out:50" in summary
+
+
+def test_get_log_summary_without_first_byte_time() -> None:
+    """测试日志摘要在没有首字时间时的格式"""
+    ctx = StreamContext(model="claude-3", api_format="claude_messages")
+    ctx.provider_name = "anthropic"
+    ctx.input_tokens = 100
+    ctx.output_tokens = 50
+    # first_byte_time_ms 保持为 None
+
+    summary = ctx.get_log_summary("request-id-123", 456)
+
+    # 验证不包含首字时间标记,但有总时间（使用大写 TTFB 和 Total）
+    assert "TTFB:" not in summary
+    assert "Total: 456ms" in summary
+    assert "in:100 out:50" in summary

tests/api/handlers/base/test_stream_processor.py

@@ -0,0 +1,32 @@
+from typing import Any, Dict, Optional
+
+from src.api.handlers.base.response_parser import ParsedChunk, ParsedResponse, ResponseParser, StreamStats
+from src.api.handlers.base.stream_context import StreamContext
+from src.api.handlers.base.stream_processor import StreamProcessor
+from src.utils.sse_parser import SSEEventParser
+
+
+class DummyParser(ResponseParser):
+    def parse_sse_line(self, line: str, stats: StreamStats) -> Optional[ParsedChunk]:
+        return None
+
+    def parse_response(self, response: Dict[str, Any], status_code: int) -> ParsedResponse:
+        return ParsedResponse(raw_response=response, status_code=status_code)
+
+    def extract_usage_from_response(self, response: Dict[str, Any]) -> Dict[str, int]:
+        return {}
+
+    def extract_text_content(self, response: Dict[str, Any]) -> str:
+        return ""
+
+
+def test_process_line_strips_newlines_and_finalizes_event() -> None:
+    ctx = StreamContext(model="test-model", api_format="OPENAI")
+    processor = StreamProcessor(request_id="test-request", default_parser=DummyParser())
+    sse_parser = SSEEventParser()
+
+    processor._process_line(ctx, sse_parser, 'data: {"type":"response.completed"}\n')
+    processor._process_line(ctx, sse_parser, "\n")
+
+    assert ctx.has_completion is True
+

tests/api/handlers/base/test_utils.py

@@ -0,0 +1,104 @@
+"""测试 handler 基础工具函数"""
+
+import pytest
+
+from src.api.handlers.base.utils import build_sse_headers, extract_cache_creation_tokens
+
+
+class TestExtractCacheCreationTokens:
+    """测试 extract_cache_creation_tokens 函数"""
+
+    def test_new_format_only(self) -> None:
+        """测试只有新格式字段"""
+        usage = {
+            "claude_cache_creation_5_m_tokens": 100,
+            "claude_cache_creation_1_h_tokens": 200,
+        }
+        assert extract_cache_creation_tokens(usage) == 300
+
+    def test_new_format_5m_only(self) -> None:
+        """测试只有 5 分钟缓存"""
+        usage = {
+            "claude_cache_creation_5_m_tokens": 150,
+            "claude_cache_creation_1_h_tokens": 0,
+        }
+        assert extract_cache_creation_tokens(usage) == 150
+
+    def test_new_format_1h_only(self) -> None:
+        """测试只有 1 小时缓存"""
+        usage = {
+            "claude_cache_creation_5_m_tokens": 0,
+            "claude_cache_creation_1_h_tokens": 250,
+        }
+        assert extract_cache_creation_tokens(usage) == 250
+
+    def test_old_format_only(self) -> None:
+        """测试只有旧格式字段"""
+        usage = {
+            "cache_creation_input_tokens": 500,
+        }
+        assert extract_cache_creation_tokens(usage) == 500
+
+    def test_both_formats_prefers_new(self) -> None:
+        """测试同时存在时优先使用新格式"""
+        usage = {
+            "claude_cache_creation_5_m_tokens": 100,
+            "claude_cache_creation_1_h_tokens": 200,
+            "cache_creation_input_tokens": 999,  # 应该被忽略
+        }
+        assert extract_cache_creation_tokens(usage) == 300
+
+    def test_empty_usage(self) -> None:
+        """测试空字典"""
+        usage = {}
+        assert extract_cache_creation_tokens(usage) == 0
+
+    def test_all_zeros(self) -> None:
+        """测试所有字段都为 0"""
+        usage = {
+            "claude_cache_creation_5_m_tokens": 0,
+            "claude_cache_creation_1_h_tokens": 0,
+            "cache_creation_input_tokens": 0,
+        }
+        assert extract_cache_creation_tokens(usage) == 0
+
+    def test_partial_new_format_with_old_format_fallback(self) -> None:
+        """测试新格式字段不存在时回退到旧格式"""
+        usage = {
+            "cache_creation_input_tokens": 123,
+        }
+        assert extract_cache_creation_tokens(usage) == 123
+
+    def test_new_format_zero_should_not_fallback(self) -> None:
+        """测试新格式字段存在但为 0 时，不应 fallback 到旧格式"""
+        usage = {
+            "claude_cache_creation_5_m_tokens": 0,
+            "claude_cache_creation_1_h_tokens": 0,
+            "cache_creation_input_tokens": 456,
+        }
+        # 新格式字段存在，即使值为 0 也应该使用新格式（返回 0）
+        # 而不是 fallback 到旧格式（返回 456）
+        assert extract_cache_creation_tokens(usage) == 0
+
+    def test_unrelated_fields_ignored(self) -> None:
+        """测试忽略无关字段"""
+        usage = {
+            "input_tokens": 1000,
+            "output_tokens": 2000,
+            "cache_read_input_tokens": 300,
+            "claude_cache_creation_5_m_tokens": 50,
+            "claude_cache_creation_1_h_tokens": 75,
+        }
+        assert extract_cache_creation_tokens(usage) == 125
+
+
+class TestBuildSSEHeaders:
+    def test_default_headers(self) -> None:
+        headers = build_sse_headers()
+        assert headers["Cache-Control"] == "no-cache, no-transform"
+        assert headers["X-Accel-Buffering"] == "no"
+
+    def test_merge_extra_headers(self) -> None:
+        headers = build_sse_headers({"X-Test": "1", "Cache-Control": "custom"})
+        assert headers["X-Test"] == "1"
+        assert headers["Cache-Control"] == "custom"