refactor: 重构异步任务系统和计费服务架构

- 重构任务系统：新增 lifecycle (TaskStatus/BillingStatus)、context、application 模块
- 将 video tasks 泛化为 async tasks，支持更通用的异步任务管理
- 新增 Gemini Files 管理模块和管理界面
- 重构 billing 服务：拆分 schema.py 和 service.py
- 新增 candidate 服务模块用于请求候选管理
- 数据库迁移：添加 billing_status、request_id、gemini_file_mappings 表和索引
- 移除废弃的 video_telemetry、task orchestrator 等模块

src/api/admin/gemini_files.py

@@ -0,0 +1,543 @@
+"""
+Gemini Files 管理 API
+
+提供文件映射的管理功能：
+- 列出所有文件映射
+- 删除文件映射
+- 查看文件映射统计
+- 上传文件到 Gemini
+
+优化：HTTP 上传期间不持有数据库连接，避免阻塞其他请求。
+"""
+
+from __future__ import annotations
+
+import asyncio
+import hashlib
+from dataclasses import dataclass
+from datetime import datetime, timezone
+from typing import Any
+
+from fastapi import APIRouter, Depends, File, HTTPException, Query, UploadFile
+from pydantic import BaseModel
+from sqlalchemy import delete, func
+from sqlalchemy.orm import Session
+
+from src.clients.http_client import HTTPClientPool
+from src.core.crypto import crypto_service
+from src.core.logger import logger
+from src.database import create_session, get_db
+from src.models.database import GeminiFileMapping, ProviderAPIKey, User
+from src.services.gemini_files_mapping import delete_file_key_mapping, store_file_key_mapping
+
+
+@dataclass
+class KeyInfo:
+    """Key 信息（用于 HTTP 上传，不依赖数据库会话）"""
+
+    id: str
+    name: str | None
+    decrypted_api_key: str
+
+
+router = APIRouter(prefix="/api/admin/gemini-files", tags=["Gemini Files Management"])
+
+# Gemini Files API 基础 URL
+GEMINI_FILES_BASE_URL = "https://generativelanguage.googleapis.com"
+
+
+# ============ Schema ============
+
+
+class FileMappingResponse(BaseModel):
+    """文件映射响应"""
+
+    id: str
+    file_name: str
+    key_id: str
+    key_name: str | None = None
+    user_id: str | None = None
+    username: str | None = None
+    display_name: str | None = None
+    mime_type: str | None = None
+    created_at: datetime
+    expires_at: datetime
+    is_expired: bool
+
+
+class FileMappingListResponse(BaseModel):
+    """文件映射列表响应"""
+
+    items: list[FileMappingResponse]
+    total: int
+    page: int
+    page_size: int
+
+
+class FileMappingStatsResponse(BaseModel):
+    """文件映射统计响应"""
+
+    total_mappings: int
+    active_mappings: int
+    expired_mappings: int
+    by_mime_type: dict[str, int]
+    capable_keys_count: int
+
+
+# ============ Routes ============
+
+
+@router.get("/mappings", response_model=FileMappingListResponse)
+async def list_file_mappings(
+    db: Session = Depends(get_db),
+    page: int = Query(1, ge=1),
+    page_size: int = Query(20, ge=1, le=100),
+    include_expired: bool = Query(False),
+    search: str | None = Query(None),
+) -> Any:
+    """
+    列出所有文件映射
+
+    - **page**: 页码
+    - **page_size**: 每页数量
+    - **include_expired**: 是否包含已过期的映射
+    - **search**: 搜索文件名或显示名
+    """
+    now = datetime.now(timezone.utc)
+
+    query = db.query(GeminiFileMapping)
+
+    # 过滤过期
+    if not include_expired:
+        query = query.filter(GeminiFileMapping.expires_at > now)
+
+    # 搜索
+    if search:
+        search_pattern = f"%{search}%"
+        query = query.filter(
+            (GeminiFileMapping.file_name.ilike(search_pattern))
+            | (GeminiFileMapping.display_name.ilike(search_pattern))
+        )
+
+    # 总数
+    total = query.count()
+
+    # 分页
+    offset = (page - 1) * page_size
+    mappings = (
+        query.order_by(GeminiFileMapping.created_at.desc()).offset(offset).limit(page_size).all()
+    )
+
+    # 获取关联的 Key 和 User 信息
+    key_ids = {m.key_id for m in mappings}
+    user_ids = {m.user_id for m in mappings if m.user_id}
+
+    keys_map = {}
+    if key_ids:
+        keys = db.query(ProviderAPIKey).filter(ProviderAPIKey.id.in_(key_ids)).all()
+        keys_map = {str(k.id): k.name for k in keys}
+
+    users_map = {}
+    if user_ids:
+        users = db.query(User).filter(User.id.in_(user_ids)).all()
+        users_map = {str(u.id): u.username for u in users}
+
+    items = []
+    for m in mappings:
+        items.append(
+            FileMappingResponse(
+                id=str(m.id),
+                file_name=m.file_name,
+                key_id=str(m.key_id),
+                key_name=keys_map.get(str(m.key_id)),
+                user_id=str(m.user_id) if m.user_id else None,
+                username=users_map.get(str(m.user_id)) if m.user_id else None,
+                display_name=m.display_name,
+                mime_type=m.mime_type,
+                created_at=m.created_at,
+                expires_at=m.expires_at,
+                is_expired=m.expires_at <= now,
+            )
+        )
+
+    return FileMappingListResponse(
+        items=items,
+        total=total,
+        page=page,
+        page_size=page_size,
+    )
+
+
+@router.get("/stats", response_model=FileMappingStatsResponse)
+async def get_file_mapping_stats(
+    db: Session = Depends(get_db),
+) -> Any:
+    """获取文件映射统计信息"""
+    now = datetime.now(timezone.utc)
+
+    # 总数
+    total_mappings = db.query(func.count(GeminiFileMapping.id)).scalar() or 0
+
+    # 活跃数（未过期）
+    active_mappings = (
+        db.query(func.count(GeminiFileMapping.id))
+        .filter(GeminiFileMapping.expires_at > now)
+        .scalar()
+        or 0
+    )
+
+    # 过期数
+    expired_mappings = total_mappings - active_mappings
+
+    # 按 MIME 类型统计
+    mime_stats = (
+        db.query(GeminiFileMapping.mime_type, func.count(GeminiFileMapping.id))
+        .filter(GeminiFileMapping.expires_at > now)
+        .group_by(GeminiFileMapping.mime_type)
+        .all()
+    )
+    by_mime_type = {(mt or "unknown"): count for mt, count in mime_stats}
+
+    # 有 gemini_files 能力的 Key 数量
+    keys = db.query(ProviderAPIKey).filter(ProviderAPIKey.is_active.is_(True)).all()
+    capable_keys_count = sum(
+        1 for key in keys if key.capabilities and key.capabilities.get("gemini_files", False)
+    )
+
+    return FileMappingStatsResponse(
+        total_mappings=total_mappings,
+        active_mappings=active_mappings,
+        expired_mappings=expired_mappings,
+        by_mime_type=by_mime_type,
+        capable_keys_count=capable_keys_count,
+    )
+
+
+@router.delete("/mappings/{mapping_id}")
+async def delete_mapping(
+    mapping_id: str,
+    db: Session = Depends(get_db),
+) -> Any:
+    """
+    删除指定的文件映射
+
+    注意：这只删除映射记录，不会删除 Gemini 上的实际文件
+    """
+    mapping = db.query(GeminiFileMapping).filter(GeminiFileMapping.id == mapping_id).first()
+
+    if not mapping:
+        raise HTTPException(status_code=404, detail="Mapping not found")
+
+    file_name = mapping.file_name
+
+    # 从数据库删除
+    db.delete(mapping)
+    db.commit()
+
+    # 同时从 Redis 删除
+    await delete_file_key_mapping(file_name)
+
+    return {"message": "Mapping deleted successfully", "file_name": file_name}
+
+
+@router.delete("/mappings")
+async def cleanup_expired_mappings(
+    db: Session = Depends(get_db),
+) -> Any:
+    """清理所有过期的文件映射"""
+    now = datetime.now(timezone.utc)
+
+    result = db.execute(delete(GeminiFileMapping).where(GeminiFileMapping.expires_at <= now))
+    db.commit()
+
+    deleted_count = result.rowcount
+
+    return {
+        "message": f"Cleaned up {deleted_count} expired mappings",
+        "deleted_count": deleted_count,
+    }
+
+
+class CapableKeyResponse(BaseModel):
+    """可用 Key 响应"""
+
+    id: str
+    name: str
+    provider_name: str | None = None
+
+
+class UploadResultItem(BaseModel):
+    """单个 Key 的上传结果"""
+
+    key_id: str
+    key_name: str | None = None
+    success: bool
+    file_name: str | None = None
+    error: str | None = None
+
+
+class UploadResponse(BaseModel):
+    """上传响应"""
+
+    display_name: str
+    mime_type: str
+    size_bytes: int
+    results: list[UploadResultItem]
+    success_count: int
+    fail_count: int
+
+
+@router.get("/capable-keys", response_model=list[CapableKeyResponse])
+async def list_capable_keys(
+    db: Session = Depends(get_db),
+) -> Any:
+    """获取所有具有 gemini_files 能力的 Key 列表"""
+    from src.models.database import Provider
+
+    keys = db.query(ProviderAPIKey).filter(ProviderAPIKey.is_active.is_(True)).all()
+    capable_keys = [
+        key for key in keys if key.capabilities and key.capabilities.get("gemini_files", False)
+    ]
+
+    # 获取 Provider 名称
+    provider_ids = {key.provider_id for key in capable_keys}
+    providers = db.query(Provider).filter(Provider.id.in_(provider_ids)).all()
+    provider_map = {str(p.id): p.name for p in providers}
+
+    return [
+        CapableKeyResponse(
+            id=str(key.id),
+            name=key.name,
+            provider_name=provider_map.get(str(key.provider_id)),
+        )
+        for key in capable_keys
+    ]
+
+
+async def _upload_to_key(
+    key_info: KeyInfo,
+    content: bytes,
+    file_size: int,
+    mime_type: str,
+    display_name: str,
+    source_hash: str,
+) -> UploadResultItem:
+    """上传文件到指定的 Key（不依赖数据库会话）"""
+    api_key = key_info.decrypted_api_key
+    client = await HTTPClientPool.get_default_client_async()
+
+    try:
+        # 第一步：初始化可恢复上传
+        init_url = f"{GEMINI_FILES_BASE_URL}/upload/v1beta/files?key={api_key}"
+        init_headers = {
+            "X-Goog-Upload-Protocol": "resumable",
+            "X-Goog-Upload-Command": "start",
+            "X-Goog-Upload-Header-Content-Length": str(file_size),
+            "X-Goog-Upload-Header-Content-Type": mime_type,
+            "Content-Type": "application/json",
+        }
+        init_body = {"file": {"display_name": display_name}}
+
+        init_response = await client.post(init_url, headers=init_headers, json=init_body)
+
+        if init_response.status_code != 200:
+            logger.error(
+                f"Gemini upload init failed for key {key_info.id}: {init_response.status_code}"
+            )
+            return UploadResultItem(
+                key_id=key_info.id,
+                key_name=key_info.name,
+                success=False,
+                error=f"初始化失败: {init_response.status_code}",
+            )
+
+        # 获取上传 URL
+        upload_url = init_response.headers.get("X-Goog-Upload-URL")
+        if not upload_url:
+            return UploadResultItem(
+                key_id=key_info.id,
+                key_name=key_info.name,
+                success=False,
+                error="未获取到上传 URL",
+            )
+
+        # 第二步：上传文件内容
+        upload_headers = {
+            "X-Goog-Upload-Command": "upload, finalize",
+            "X-Goog-Upload-Offset": "0",
+            "Content-Length": str(file_size),
+            "Content-Type": mime_type,
+        }
+
+        upload_response = await client.post(upload_url, headers=upload_headers, content=content)
+
+        if upload_response.status_code != 200:
+            logger.error(
+                f"Gemini upload failed for key {key_info.id}: {upload_response.status_code}"
+            )
+            return UploadResultItem(
+                key_id=key_info.id,
+                key_name=key_info.name,
+                success=False,
+                error=f"上传失败: {upload_response.status_code}",
+            )
+
+        # 解析响应
+        result = upload_response.json()
+        file_info = result.get("file", {})
+        file_name = file_info.get("name", "")
+        response_display_name = file_info.get("displayName", display_name)
+        response_mime_type = file_info.get("mimeType", mime_type)
+
+        # 存储文件映射（包含源文件哈希，用于关联相同源文件的不同上传）
+        await store_file_key_mapping(
+            file_name=file_name,
+            key_id=key_info.id,
+            user_id=None,
+            display_name=response_display_name,
+            mime_type=response_mime_type,
+            source_hash=source_hash,
+        )
+
+        return UploadResultItem(
+            key_id=key_info.id,
+            key_name=key_info.name,
+            success=True,
+            file_name=file_name,
+        )
+
+    except Exception as exc:
+        logger.error(f"Gemini upload error for key {key_info.id}: {exc}")
+        return UploadResultItem(
+            key_id=key_info.id,
+            key_name=key_info.name,
+            success=False,
+            error=str(exc),
+        )
+
+
+@router.post("/upload", response_model=UploadResponse)
+async def upload_file(
+    file: UploadFile = File(...),
+    key_ids: str = Query(..., description="逗号分隔的 Key ID 列表"),
+) -> Any:
+    """
+    上传文件到 Gemini Files API
+
+    - **file**: 要上传的文件
+    - **key_ids**: 逗号分隔的 Key ID 列表，文件将上传到所有指定的 Key
+
+    支持的文件类型：视频、图片、音频、文档等
+    文件大小限制：2GB
+    文件有效期：48小时
+
+    优化：HTTP 上传期间不持有数据库连接
+    """
+    # 解析 Key IDs
+    key_id_list = [kid.strip() for kid in key_ids.split(",") if kid.strip()]
+    if not key_id_list:
+        raise HTTPException(status_code=400, detail="请至少选择一个 Key")
+
+    # ========== 阶段 1：读取文件内容并计算哈希 ==========
+    content = await file.read()
+    file_size = len(content)
+    mime_type = file.content_type or "application/octet-stream"
+    display_name = file.filename or "uploaded_file"
+
+    # 计算源文件哈希（用于重复检测和关联相同源文件的不同上传）
+    source_hash = hashlib.sha256(content).hexdigest()
+
+    # ========== 阶段 2：查询数据库（短暂持有连接）==========
+    key_infos: list[KeyInfo] = []
+    existing_mappings: dict[str, str] = {}  # key_id -> existing file_name
+
+    with create_session() as db:
+        keys = (
+            db.query(ProviderAPIKey)
+            .filter(
+                ProviderAPIKey.id.in_(key_id_list),
+                ProviderAPIKey.is_active.is_(True),
+            )
+            .all()
+        )
+
+        # 过滤有 gemini_files 能力的 Key，并提取必要信息
+        capable_key_ids = []
+        for key in keys:
+            if key.capabilities and key.capabilities.get("gemini_files", False):
+                try:
+                    decrypted_key = crypto_service.decrypt(key.api_key)
+                    key_infos.append(
+                        KeyInfo(
+                            id=str(key.id),
+                            name=key.name,
+                            decrypted_api_key=decrypted_key,
+                        )
+                    )
+                    capable_key_ids.append(str(key.id))
+                except Exception as exc:
+                    logger.error(f"Failed to decrypt provider key {key.id}: {exc}")
+
+        # 检查是否已存在相同 source_hash 的映射（重复检测）
+        if capable_key_ids:
+            now = datetime.now(timezone.utc)
+            existing = (
+                db.query(GeminiFileMapping)
+                .filter(
+                    GeminiFileMapping.source_hash == source_hash,
+                    GeminiFileMapping.key_id.in_(capable_key_ids),
+                    GeminiFileMapping.expires_at > now,  # 只查未过期的
+                )
+                .all()
+            )
+            for mapping in existing:
+                existing_mappings[str(mapping.key_id)] = mapping.file_name
+
+    if not key_infos:
+        raise HTTPException(
+            status_code=400,
+            detail="选中的 Key 都没有「Gemini 文件 API」能力或解密失败",
+        )
+
+    # ========== 阶段 3：并发上传（跳过已有相同文件的 Key）==========
+    results: list[UploadResultItem] = []
+    keys_to_upload: list[KeyInfo] = []
+
+    for key_info in key_infos:
+        if key_info.id in existing_mappings:
+            # 该 Key 已有相同文件，跳过上传
+            results.append(
+                UploadResultItem(
+                    key_id=key_info.id,
+                    key_name=key_info.name,
+                    success=True,
+                    file_name=existing_mappings[key_info.id],
+                    error=None,
+                )
+            )
+            logger.info(
+                f"跳过重复上传: Key {key_info.id} 已有文件 {existing_mappings[key_info.id]}"
+            )
+        else:
+            keys_to_upload.append(key_info)
+
+    # 只对需要上传的 Key 执行上传
+    if keys_to_upload:
+        tasks = [
+            _upload_to_key(key_info, content, file_size, mime_type, display_name, source_hash)
+            for key_info in keys_to_upload
+        ]
+        upload_results = await asyncio.gather(*tasks)
+        results.extend(upload_results)
+
+    success_count = sum(1 for r in results if r.success)
+    fail_count = len(results) - success_count
+
+    return UploadResponse(
+        display_name=display_name,
+        mime_type=mime_type,
+        size_bytes=file_size,
+        results=results,
+        success_count=success_count,
+        fail_count=fail_count,
+    )

src/api/admin/provider_query.py

@@ -20,6 +20,7 @@ from src.core.logger import logger
 from src.database.database import get_db
 from src.models.database import Provider, ProviderEndpoint, User
 from src.services.model.fetch_scheduler import (
+    MODEL_FETCH_HTTP_TIMEOUT,
     get_upstream_models_from_cache,
     set_upstream_models_to_cache,
 )
@@ -135,7 +136,9 @@ async def query_available_models(
             return [], f"Key {api_key.name or api_key.id}: decrypt failed", False
 
         endpoint_configs = build_all_format_configs(api_key_value, format_to_endpoint)
-        models, errors, has_success = await fetch_models_from_endpoints(endpoint_configs)
+        models, errors, has_success = await fetch_models_from_endpoints(
+            endpoint_configs, timeout=MODEL_FETCH_HTTP_TIMEOUT
+        )
 
         # 写入缓存
         if models:
@@ -274,7 +277,9 @@ async def _fetch_models_for_single_key(
         raise HTTPException(status_code=500, detail="Failed to decrypt API key")
 
     endpoint_configs = build_all_format_configs(api_key_value, format_to_endpoint)
-    all_models, errors, has_success = await fetch_models_from_endpoints(endpoint_configs)
+    all_models, errors, has_success = await fetch_models_from_endpoints(
+        endpoint_configs, timeout=MODEL_FETCH_HTTP_TIMEOUT
+    )
 
     # 按 model id 聚合，合并所有 api_format
     unique_models = _aggregate_models_by_id(all_models)

src/api/admin/providers/summary.py

@@ -309,6 +309,7 @@ def _build_provider_summary(db: Session, provider: Provider) -> ProviderWithEndp
         website=provider.website,
         provider_priority=provider.provider_priority,
         keep_priority_on_conversion=provider.keep_priority_on_conversion,
+        enable_format_conversion=provider.enable_format_conversion,
         is_active=provider.is_active,
         billing_type=provider.billing_type.value if provider.billing_type else None,
         monthly_quota_usd=provider.monthly_quota_usd,

src/api/admin/usage/routes.py

@@ -271,6 +271,9 @@ class AdminUsageStatsAdapter(AdminApiAdapter):
         self.end_date = end_date
 
     async def handle(self, context: ApiRequestContext) -> Any:  # type: ignore[override]
+        # Perf: use a single aggregate query (avoid 3 full scans).
+        from sqlalchemy import case
+
         db = context.db
         query = db.query(Usage)
         if self.start_date:
@@ -278,56 +281,53 @@ class AdminUsageStatsAdapter(AdminApiAdapter):
         if self.end_date:
             query = query.filter(Usage.created_at <= self.end_date)
 
-        total_stats = query.with_entities(
+        stats = query.with_entities(
             func.count(Usage.id).label("total_requests"),
             func.sum(Usage.total_tokens).label("total_tokens"),
             func.sum(Usage.total_cost_usd).label("total_cost"),
             func.sum(Usage.actual_total_cost_usd).label("total_actual_cost"),
             func.avg(Usage.response_time_ms).label("avg_response_time_ms"),
-        ).first()
-
-        # 缓存统计
-        cache_stats = query.with_entities(
             func.sum(Usage.cache_creation_input_tokens).label("cache_creation_tokens"),
             func.sum(Usage.cache_read_input_tokens).label("cache_read_tokens"),
             func.sum(Usage.cache_creation_cost_usd).label("cache_creation_cost"),
             func.sum(Usage.cache_read_cost_usd).label("cache_read_cost"),
+            func.sum(
+                case(
+                    (
+                        (Usage.status_code >= 400) | (Usage.error_message.isnot(None)),
+                        1,
+                    ),
+                    else_=0,
+                )
+            ).label("error_count"),
         ).first()
 
-        # 错误统计
-        error_count = query.filter(
-            (Usage.status_code >= 400) | (Usage.error_message.isnot(None))
-        ).count()
-
         context.add_audit_metadata(
             action="usage_stats",
             start_date=self.start_date.isoformat() if self.start_date else None,
             end_date=self.end_date.isoformat() if self.end_date else None,
         )
 
-        total_requests = total_stats.total_requests if total_stats else 0
-        avg_response_time_ms = float(total_stats.avg_response_time_ms or 0) if total_stats else 0
+        total_requests = int(stats.total_requests or 0) if stats else 0
+        avg_response_time_ms = float(stats.avg_response_time_ms or 0) if stats else 0
         avg_response_time = avg_response_time_ms / 1000.0
+        error_count = int(stats.error_count or 0) if stats else 0
 
         return {
             "total_requests": total_requests,
-            "total_tokens": int(total_stats.total_tokens or 0),
-            "total_cost": float(total_stats.total_cost or 0),
-            "total_actual_cost": float(total_stats.total_actual_cost or 0),
+            "total_tokens": int(stats.total_tokens or 0) if stats else 0,
+            "total_cost": float(stats.total_cost or 0) if stats else 0,
+            "total_actual_cost": float(stats.total_actual_cost or 0) if stats else 0,
             "avg_response_time": round(avg_response_time, 2),
             "error_count": error_count,
             "error_rate": (
                 round((error_count / total_requests) * 100, 2) if total_requests > 0 else 0
             ),
             "cache_stats": {
-                "cache_creation_tokens": (
-                    int(cache_stats.cache_creation_tokens or 0) if cache_stats else 0
-                ),
-                "cache_read_tokens": int(cache_stats.cache_read_tokens or 0) if cache_stats else 0,
-                "cache_creation_cost": (
-                    float(cache_stats.cache_creation_cost or 0) if cache_stats else 0
-                ),
-                "cache_read_cost": float(cache_stats.cache_read_cost or 0) if cache_stats else 0,
+                "cache_creation_tokens": (int(stats.cache_creation_tokens or 0) if stats else 0),
+                "cache_read_tokens": int(stats.cache_read_tokens or 0) if stats else 0,
+                "cache_creation_cost": (float(stats.cache_creation_cost or 0) if stats else 0),
+                "cache_read_cost": float(stats.cache_read_cost or 0) if stats else 0,
             },
         }
 
@@ -637,6 +637,7 @@ class AdminUsageRecordsAdapter(AdminApiAdapter):
 
     async def handle(self, context: ApiRequestContext) -> Any:  # type: ignore[override]
         from sqlalchemy import or_
+        from sqlalchemy.orm import load_only
 
         from src.utils.database_helpers import escape_like_pattern, safe_truncate_escaped
 
@@ -677,13 +678,13 @@ class AdminUsageRecordsAdapter(AdminApiAdapter):
             escaped = escape_like_pattern(self.username)
             query = query.filter(User.username.ilike(f"%{escaped}%", escape="\\"))
         if self.model:
-            # 支持模型名模糊搜索
-            escaped = escape_like_pattern(self.model)
-            query = query.filter(Usage.model.ilike(f"%{escaped}%", escape="\\"))
+            # 模型筛选：前端为下拉框精确值，使用精确匹配以启用索引
+            # 如需模糊搜索，请使用 search 参数。
+            query = query.filter(Usage.model == self.model)
         if self.provider:
-            # 支持提供商名称搜索
-            escaped = escape_like_pattern(self.provider)
-            query = query.filter(Provider.name.ilike(f"%{escaped}%", escape="\\"))
+            # 提供商筛选：前端为下拉框精确值，使用精确匹配以启用索引
+            # 如需模糊搜索，请使用 search 参数。
+            query = query.filter(Provider.name == self.provider)
         if self.status:
             # 状态筛选
             # 旧的筛选值（基于 is_stream 和 status_code）：stream, standard, error
@@ -714,7 +715,51 @@ class AdminUsageRecordsAdapter(AdminApiAdapter):
         if self.end_date:
             query = query.filter(Usage.created_at <= self.end_date)
 
-        total = query.count()
+        # Perf: avoid Query.count() building a subquery selecting many columns
+        total = int(query.with_entities(func.count(Usage.id)).scalar() or 0)
+
+        # Perf: do not load large request/response columns for list view
+        query = query.options(
+            load_only(
+                Usage.id,
+                Usage.request_id,
+                Usage.user_id,
+                Usage.api_key_id,
+                Usage.provider_name,
+                Usage.provider_id,
+                Usage.provider_endpoint_id,
+                Usage.provider_api_key_id,
+                Usage.model,
+                Usage.target_model,
+                Usage.input_tokens,
+                Usage.output_tokens,
+                Usage.cache_creation_input_tokens,
+                Usage.cache_read_input_tokens,
+                Usage.total_tokens,
+                Usage.total_cost_usd,
+                Usage.actual_total_cost_usd,
+                Usage.rate_multiplier,
+                Usage.response_time_ms,
+                Usage.first_byte_time_ms,
+                Usage.created_at,
+                Usage.is_stream,
+                Usage.status_code,
+                Usage.error_message,
+                Usage.status,
+                Usage.api_format,
+                Usage.endpoint_api_format,
+                Usage.has_format_conversion,
+                Usage.request_metadata,
+                Usage.input_price_per_1m,
+                Usage.output_price_per_1m,
+                Usage.cache_creation_price_per_1m,
+                Usage.cache_read_price_per_1m,
+            ),
+            load_only(User.id, User.email, User.username),
+            load_only(ProviderEndpoint.id, ProviderEndpoint.api_format),
+            load_only(ProviderAPIKey.id, ProviderAPIKey.name),
+            load_only(ApiKey.id, ApiKey.name, ApiKey.key_encrypted),
+        )
         records = (
             query.order_by(Usage.created_at.desc()).offset(self.offset).limit(self.limit).all()
         )
@@ -779,7 +824,9 @@ class AdminUsageRecordsAdapter(AdminApiAdapter):
         )
 
         # 构建 provider_id -> Provider 名称的映射，避免 N+1 查询
-        provider_ids = [usage.provider_id for usage, _, _, _, _ in records if usage.provider_id]
+        provider_ids = list(
+            {usage.provider_id for usage, _, _, _, _ in records if usage.provider_id}
+        )
         provider_map = {}
         if provider_ids:
             providers_data = (
@@ -788,6 +835,7 @@ class AdminUsageRecordsAdapter(AdminApiAdapter):
             provider_map = {str(p.id): p.name for p in providers_data}
 
         data = []
+        api_key_display_cache: dict[str, str] = {}
         for usage, user, endpoint, provider_api_key, user_api_key in records:
             actual_cost = (
                 float(usage.actual_total_cost_usd)
@@ -829,7 +877,9 @@ class AdminUsageRecordsAdapter(AdminApiAdapter):
                         {
                             "id": user_api_key.id,
                             "name": user_api_key.name,
-                            "display": user_api_key.get_display_key(),
+                            "display": api_key_display_cache.setdefault(
+                                user_api_key.id, user_api_key.get_display_key()
+                            ),
                         }
                         if user_api_key
                         else None

src/api/base/models_service.py

@@ -311,7 +311,7 @@ def get_compatible_provider_formats(
             endpoint_format,
             format_acceptance_config,
             is_stream=False,
-            global_conversion_enabled=global_conversion_enabled,
+            effective_conversion_enabled=global_conversion_enabled,
         )
         if not is_compatible:
             continue

src/api/dashboard/routes.py

@@ -813,23 +813,31 @@ class DashboardRecentRequestsAdapter(DashboardAdapter):
     async def handle(self, context: ApiRequestContext) -> Any:  # type: ignore[override]
         db = context.db
         user = context.user
-        query = db.query(Usage)
+        # Perf: select only required columns (avoid loading large JSON/BLOB fields).
+        query = db.query(
+            Usage.id,
+            Usage.user_id,
+            Usage.model,
+            Usage.total_tokens,
+            Usage.created_at,
+            Usage.is_stream,
+            DBUser.username,
+        ).outerjoin(DBUser, DBUser.id == Usage.user_id)
         if user.role != UserRole.ADMIN:
             query = query.filter(Usage.user_id == user.id)
 
-        recent_requests = query.order_by(Usage.created_at.desc()).limit(self.limit).all()
+        rows = query.order_by(Usage.created_at.desc()).limit(self.limit).all()
 
         results = []
-        for req in recent_requests:
-            owner = db.query(DBUser).filter(DBUser.id == req.user_id).first()
+        for req_id, _user_id, model, total_tokens, created_at, is_stream, username in rows:
             results.append(
                 {
-                    "id": req.id,
-                    "user": owner.username if owner else "Unknown",
-                    "model": req.model or "N/A",
-                    "tokens": req.total_tokens,
-                    "time": req.created_at.strftime("%H:%M") if req.created_at else None,
-                    "is_stream": req.is_stream,
+                    "id": req_id,
+                    "user": username or "Unknown",
+                    "model": model or "N/A",
+                    "tokens": int(total_tokens or 0),
+                    "time": created_at.strftime("%H:%M") if created_at else None,
+                    "is_stream": bool(is_stream),
                 }
             )
 
@@ -848,18 +856,25 @@ class DashboardProviderStatusAdapter(DashboardAdapter):
         providers = db.query(Provider).filter(Provider.is_active.is_(True)).all()
         since = datetime.now(timezone.utc) - timedelta(days=1)
 
+        # Avoid N+1: compute 24h request counts for all providers in one GROUP BY query.
+        provider_names = [p.name for p in providers if p and p.name]
+        counts: dict[str, int] = {}
+        if provider_names:
+            rows = (
+                db.query(Usage.provider_name, func.count(Usage.id))
+                .filter(and_(Usage.created_at >= since, Usage.provider_name.in_(provider_names)))
+                .group_by(Usage.provider_name)
+                .all()
+            )
+            counts = {str(name): int(cnt or 0) for name, cnt in rows if name}
+
         entries = []
         for provider in providers:
-            count = (
-                db.query(func.count(Usage.id))
-                .filter(and_(Usage.provider_name == provider.name, Usage.created_at >= since))
-                .scalar()
-            )
             entries.append(
                 {
                     "name": provider.name,
                     "status": "active" if provider.is_active else "inactive",
-                    "requests": count,
+                    "requests": int(counts.get(provider.name, 0)),
                 }
             )
 

src/api/handlers/base/video_adapter_base.py

@@ -79,10 +79,8 @@ class VideoAdapterBase(ApiAdapter):
                 path_params=path_params,
             )
 
-        # Cancel task
-        if method in {"DELETE", "POST"} and (
-            path.endswith("/cancel") or path_params.get("action") == "cancel"
-        ):
+        # Cancel task (POST /videos/{id}/cancel or explicit action=cancel)
+        if (method == "POST" and path.endswith("/cancel")) or path_params.get("action") == "cancel":
             if not task_id:
                 raise HTTPException(
                     status_code=400, detail="Task ID is required for cancel operation"
@@ -95,6 +93,16 @@ class VideoAdapterBase(ApiAdapter):
                 path_params=path_params,
             )
 
+        # Delete task (DELETE /videos/{id})
+        if method == "DELETE" and task_id:
+            return await handler.handle_delete_task(
+                task_id=task_id,
+                http_request=http_request,
+                original_headers=context.original_headers,
+                query_params=context.query_params,
+                path_params=path_params,
+            )
+
         # Remix task
         if method == "POST" and path.endswith("/remix") and task_id:
             return await handler.handle_remix_task(

src/api/handlers/base/video_handler_base.py

@@ -26,7 +26,7 @@ from src.services.cache.aware_scheduler import ProviderCandidate
 if TYPE_CHECKING:
     import httpx
 
-    from src.services.task.orchestrator import SubmitOutcome
+    from src.services.candidate.submit import SubmitOutcome
 
 # 敏感信息匹配正则（预编译提升性能）
 _SENSITIVE_PATTERN = re.compile(
@@ -53,21 +53,39 @@ def sanitize_error_message(message: str, max_length: int = 200) -> str:
     return sanitized[:max_length]
 
 
+def extract_short_id_from_operation(operation_id: str) -> str:
+    """
+    从 operation ID 中提取短 ID
+
+    我们对外暴露的 operation name 格式是:
+    - models/{model}/operations/{short_id}
+
+    此函数提取最后一部分作为 short_id，用于在数据库中查找任务。
+
+    Args:
+        operation_id: 原始 operation ID（如 "models/veo-3.1/operations/abc123"）
+
+    Returns:
+        short_id（如 "abc123"）
+    """
+    # 格式: models/{model}/operations/{short_id}
+    # 或者直接是 short_id
+    if "/" in operation_id:
+        # 提取最后一部分
+        return operation_id.rsplit("/", 1)[-1]
+    return operation_id
+
+
 def normalize_gemini_operation_id(operation_id: str) -> str:
     """
-    规范化 Gemini operation ID，确保以 "operations/" 开头
-
-    Gemini API 返回的任务 ID 格式可能是 "operations/xxx" 或 "xxx"，
-    此函数统一规范化为 "operations/xxx" 格式。
+    规范化 Gemini operation ID（保留用于向后兼容）
 
     Args:
         operation_id: 原始 operation ID
 
     Returns:
-        规范化后的 operation ID
+        规范化后的 operation ID（原样返回）
     """
-    if not operation_id.startswith("operations/"):
-        return f"operations/{operation_id}"
     return operation_id
 
 
@@ -143,6 +161,18 @@ class VideoHandlerBase(ABC):
     ) -> JSONResponse:
         """取消任务"""
 
+    async def handle_delete_task(
+        self,
+        *,
+        task_id: str,
+        http_request: Request,
+        original_headers: dict[str, str],
+        query_params: dict[str, str] | None = None,
+        path_params: dict[str, Any] | None = None,
+    ) -> JSONResponse:
+        """删除已完成或失败的视频任务 - 可选实现"""
+        raise HTTPException(status_code=501, detail="Delete not supported for this provider")
+
     async def handle_remix_task(
         self,
         *,
@@ -205,6 +235,7 @@ class VideoHandlerBase(ABC):
         }
 
     def _get_task(self, task_id: str) -> VideoTask:
+        """通过 UUID 查找任务（OpenAI Sora 风格）"""
         task = (
             self.db.query(VideoTask)
             .filter(VideoTask.id == task_id, VideoTask.user_id == self.user.id)
@@ -229,7 +260,7 @@ class VideoHandlerBase(ABC):
         except ValueError:
             status = VideoStatus.PENDING
         return InternalVideoTask(
-            id=task.id,
+            id=task.id,  # OpenAI Sora 使用 UUID
             external_id=task.external_task_id,
             status=status,
             progress_percent=task.progress_percent or 0,
@@ -243,6 +274,52 @@ class VideoHandlerBase(ABC):
             extra={"model": task.model},
         )
 
+    def _finalize_usage_on_submit_failure(
+        self,
+        candidate_keys: list[dict[str, Any]],
+        status_code: int | None,
+    ) -> None:
+        """
+        提交失败时结算 pending usage（避免遗留 pending 状态）。
+
+        从 candidate_keys 中提取最后尝试的 provider 信息，更新 Usage 记录。
+        """
+        from src.services.usage.service import UsageService
+
+        # 提取 provider 信息：优先取最后一个有 attempt 的候选
+        provider_name = "unknown"
+        provider_id = None
+        endpoint_id = None
+        key_id = None
+
+        for ck in reversed(candidate_keys):
+            if ck.get("attempt_status") or ck.get("selected"):
+                provider_name = ck.get("provider_name") or "unknown"
+                provider_id = ck.get("provider_id")
+                endpoint_id = ck.get("endpoint_id")
+                key_id = ck.get("key_id")
+                break
+
+        try:
+            # 更新 usage 状态并设置 provider 信息
+            UsageService.update_usage_status(
+                self.db,
+                request_id=self.request_id,
+                status="failed",
+                error_message=f"submit_failed (status_code={status_code or 'unknown'})",
+                provider=provider_name,
+                provider_id=provider_id,
+                provider_endpoint_id=endpoint_id,
+                provider_api_key_id=key_id,
+                status_code=status_code,
+            )
+        except Exception as exc:
+            logger.warning(
+                "Failed to finalize usage on submit failure: request_id=%s, error=%s",
+                self.request_id,
+                sanitize_error_message(str(exc)),
+            )
+
     def _build_billing_rule_snapshot(
         self, rule_lookup: BillingRuleLookupResult | None
     ) -> dict[str, Any]:
@@ -290,16 +367,16 @@ class VideoHandlerBase(ABC):
         - 无可用候选 / 全部失败：抛 HTTPException(503)
         """
         # 延迟导入，避免 handler 基类层引入过多依赖导致循环
-        from src.services.task.orchestrator import (
+        from src.services.candidate.service import CandidateService
+        from src.services.candidate.submit import (
             AllCandidatesFailedError,
-            AsyncTaskOrchestrator,
             SubmitOutcome,
             UpstreamClientRequestError,
         )
 
-        orchestrator = AsyncTaskOrchestrator(self.db)
+        candidate_service = CandidateService(self.db)
         try:
-            return await orchestrator.submit_with_failover(
+            return await candidate_service.submit_with_failover(
                 api_format=api_format,
                 model_name=model_name,
                 affinity_key=str(self.api_key.id),
@@ -314,8 +391,12 @@ class VideoHandlerBase(ABC):
                 max_candidates=max_candidates,
             )
         except UpstreamClientRequestError as exc:
+            # 将 pending usage 结算为 failed，并记录 provider 信息
+            self._finalize_usage_on_submit_failure(exc.candidate_keys, exc.response.status_code)
             return self._build_error_response(exc.response)
         except AllCandidatesFailedError as exc:
+            # 将 pending usage 结算为 failed
+            self._finalize_usage_on_submit_failure(exc.candidate_keys, exc.last_status_code)
             detail = "No available provider for video generation"
             if config.billing_require_rule:
                 detail = "No available provider with billing rule for video generation"

src/api/handlers/gemini/adapter.py

@@ -18,7 +18,6 @@ from src.core.api_format import ApiFamily, get_auth_handler
 from src.core.api_format.enums import AuthMethod
 from src.core.logger import logger
 from src.models.gemini import GeminiRequest
-from src.services.gemini_files_mapping import extract_file_names_from_request
 from src.services.provider.transport import redact_url_for_log
 
 
@@ -63,11 +62,9 @@ class GeminiChatAdapter(ChatAdapterBase):
     def detect_capability_requirements(
         self,
         headers: dict[str, str],  # noqa: ARG002 - 预留
-        request_body: dict[str, Any] | None = None,
+        request_body: dict[str, Any] | None = None,  # noqa: ARG002 - 预留
     ) -> dict[str, bool]:
-        """检测是否需要 Gemini Files API 能力"""
-        if request_body and extract_file_names_from_request(request_body):
-            return {"gemini_files_api": True}
+        """Gemini API 无特殊能力要求"""
         return {}
 
     def _merge_path_params(

src/api/handlers/gemini/handler.py

@@ -40,28 +40,31 @@ class GeminiChatHandler(ChatHandlerBase):
         Gemini 文件与上传它的 API Key 绑定，必须使用同一 Key 访问。
         此方法从缓存中查找文件→Key 映射，优先使用正确的 Key。
 
+        当同一源文件被上传到多个 Key 时，会返回所有可用的 Key ID，
+        让系统能够选择任意可用的 Key。
+
         注意事项：
         - 如果映射缺失（缓存过期/重启），会记录警告，请求可能失败
-        - 如果多个文件属于不同 Key，只能使用其中一个，其他文件可能无法访问
+        - 优先返回所有支持该文件的 Key，让调度器选择可用的
         """
         from src.core.logger import logger
         from src.services.gemini_files_mapping import (
             extract_file_names_from_request,
-            get_file_key_mapping,
+            get_all_key_ids_for_file,
         )
 
         file_names = extract_file_names_from_request(request_body or {})
         if not file_names:
             return None
 
-        preferred_key_ids: list[str] = []
-        unmapped_files: list[str] = []  # 记录找不到映射的文件
+        all_key_ids: set[str] = set()
+        unmapped_files: list[str] = []
 
         for file_name in file_names:
-            key_id = await get_file_key_mapping(file_name)
-            if key_id:
-                if key_id not in preferred_key_ids:
-                    preferred_key_ids.append(key_id)
+            # 获取所有支持该文件的 Key（包括通过 source_hash 关联的）
+            key_ids = await get_all_key_ids_for_file(file_name)
+            if key_ids:
+                all_key_ids.update(key_ids)
             else:
                 unmapped_files.append(file_name)
 
@@ -72,14 +75,10 @@ class GeminiChatHandler(ChatHandlerBase):
                 "请求可能失败（文件属于其他 Key 或映射已过期）"
             )
 
-        # 警告：多个文件属于不同 Key
-        if len(preferred_key_ids) > 1:
-            logger.warning(
-                f"[{self.request_id}] 请求使用了多个文件，但它们属于不同的 Key: "
-                f"{preferred_key_ids}，只能使用第一个 Key，其他文件可能无法访问"
-            )
+        if all_key_ids:
+            logger.debug(f"[{self.request_id}] 文件引用可用的 Key: {list(all_key_ids)}")
 
-        return preferred_key_ids or None
+        return list(all_key_ids) if all_key_ids else None
 
     def extract_model_from_request(
         self,

src/api/handlers/gemini/video_handler.py

@@ -4,6 +4,7 @@ Gemini Video Handler - Veo 视频生成实现
 
 from __future__ import annotations
 
+import time
 from datetime import datetime, timedelta, timezone
 from typing import Any, AsyncIterator
 from uuid import uuid4
@@ -34,12 +35,14 @@ from src.core.api_format.conversion.internal_video import (
     VideoStatus,
 )
 from src.core.api_format.conversion.normalizers.gemini import GeminiNormalizer
+from src.core.api_format.conversion.registry import format_conversion_registry
 from src.core.api_format.headers import HOP_BY_HOP_HEADERS
 from src.core.crypto import crypto_service
 from src.core.logger import logger
 from src.models.database import ApiKey, ProviderAPIKey, ProviderEndpoint, User, VideoTask
 from src.services.billing.rule_service import BillingRuleLookupResult, BillingRuleService
 from src.services.cache.aware_scheduler import ProviderCandidate
+from src.services.usage.service import UsageService
 
 
 class GeminiVeoHandler(VideoHandlerBase):
@@ -92,20 +95,93 @@ class GeminiVeoHandler(VideoHandlerBase):
         except ValueError as e:
             raise HTTPException(status_code=400, detail=str(e))
 
+        # 异步任务：提前创建 pending usage，便于前端看到“处理中”
+        try:
+            UsageService.create_pending_usage(
+                db=self.db,
+                request_id=self.request_id,
+                user=self.user,
+                api_key=self.api_key,
+                model=internal_request.model,
+                is_stream=False,
+                request_type="video",
+                api_format=self.FORMAT_ID,
+                request_headers=original_headers,
+                request_body=original_request_body,
+            )
+        except Exception as exc:
+            logger.warning(
+                "Failed to create pending usage for video request_id=%s: %s",
+                self.request_id,
+                sanitize_error_message(str(exc)),
+            )
+
+        # 用于跟踪是否发生了格式转换
+        format_conversion_info: dict[str, Any] = {
+            "converted": False,
+            "provider_format": None,
+        }
+
         async def _submit(candidate: ProviderCandidate) -> Any:
             upstream_key, endpoint, _key, auth_info = await self._resolve_upstream_key(candidate)
-            upstream_url = self._build_upstream_url(endpoint.base_url, internal_request.model)
-            headers = self._build_upstream_headers(
-                original_headers, upstream_key, endpoint, auth_info
+
+            # 检测目标格式
+            provider_format = make_signature_key(
+                str(getattr(endpoint, "api_family", "")).strip().lower(),
+                str(getattr(endpoint, "endpoint_kind", "")).strip().lower(),
             )
-            client = await HTTPClientPool.get_default_client_async()
-            return await client.post(upstream_url, headers=headers, json=original_request_body)
+            needs_conversion = provider_format.upper() != self.FORMAT_ID.upper()
+            format_conversion_info["provider_format"] = provider_format
+            format_conversion_info["converted"] = needs_conversion
+
+            if needs_conversion and provider_format.upper().startswith("OPENAI:"):
+                # Gemini -> OpenAI 格式转换
+                converted_body = format_conversion_registry.convert_video_request(
+                    original_request_body,
+                    self.FORMAT_ID,
+                    provider_format,
+                )
+                # 确保 seconds 字段为字符串类型（上游 Go 服务要求 string）
+                if "seconds" in converted_body and converted_body["seconds"] is not None:
+                    converted_body["seconds"] = str(converted_body["seconds"])
+
+                # 构建 OpenAI 风格的 URL
+                upstream_url = self._build_openai_upstream_url(endpoint.base_url)
+
+                # 构建 OpenAI 风格的请求头
+                headers = self._build_openai_upstream_headers(
+                    original_headers, upstream_key, endpoint
+                )
+
+                client = await HTTPClientPool.get_default_client_async()
+                return await client.post(upstream_url, headers=headers, json=converted_body)
+            else:
+                # 原始 Gemini 格式
+                upstream_url = self._build_upstream_url(endpoint.base_url, internal_request.model)
+                headers = self._build_upstream_headers(
+                    original_headers, upstream_key, endpoint, auth_info
+                )
+                client = await HTTPClientPool.get_default_client_async()
+                return await client.post(upstream_url, headers=headers, json=original_request_body)
 
         def _extract_task_id(payload: dict[str, Any]) -> str | None:
-            value = payload.get("name")
-            if not value:
-                return None
-            return normalize_gemini_operation_id(str(value))
+            # 根据响应格式提取 task ID
+            # Gemini: {"name": "operations/..."}
+            # OpenAI: {"id": "..."}
+            if "name" in payload:
+                value = payload.get("name")
+                logger.debug(
+                    "[GeminiVeoHandler] Upstream response name=%s, keys=%s",
+                    value,
+                    list(payload.keys()) if isinstance(payload, dict) else type(payload),
+                )
+                if not value:
+                    return None
+                return normalize_gemini_operation_id(str(value))
+            if "id" in payload:
+                # OpenAI 格式
+                return str(payload["id"])
+            return None
 
         outcome_or_response = await self._submit_with_failover(
             api_format=self.FORMAT_ID,
@@ -114,7 +190,7 @@ class GeminiVeoHandler(VideoHandlerBase):
             submit_func=_submit,
             extract_external_task_id=_extract_task_id,
             supported_auth_types={"api_key", "vertex_ai"},
-            allow_format_conversion=False,
+            allow_format_conversion=True,
             max_candidates=10,
         )
         if isinstance(outcome_or_response, JSONResponse):
@@ -135,35 +211,92 @@ class GeminiVeoHandler(VideoHandlerBase):
 
         external_task_id = outcome.external_task_id
 
+        # 如果发生了格式转换，记录转换后的请求体
+        converted_request_body = original_request_body
+        if format_conversion_info["converted"]:
+            try:
+                converted_request_body = format_conversion_registry.convert_video_request(
+                    original_request_body,
+                    self.FORMAT_ID,
+                    format_conversion_info["provider_format"],
+                )
+            except Exception as e:
+                logger.warning(
+                    "[GeminiVeoHandler] Failed to record converted request: %s",
+                    sanitize_error_message(str(e)),
+                )
+
         task = self._create_task_record(
             external_task_id=external_task_id,
             candidate=outcome.candidate,
             original_request_body=original_request_body,
+            converted_request_body=converted_request_body,
             internal_request=internal_request,
             candidate_keys=outcome.candidate_keys,
             original_headers=original_headers,
             billing_rule_snapshot=billing_rule_snapshot,
+            format_converted=format_conversion_info["converted"],
         )
         try:
             self.db.add(task)
             self.db.flush()  # 先 flush 检测冲突
             self.db.commit()
             self.db.refresh(task)
-            logger.info(
-                f"[GeminiVeoHandler] Task created: id={task.id}, external_task_id={task.external_task_id}, user_id={task.user_id}"
+            logger.debug(
+                "[GeminiVeoHandler] Task created: id=%s, external_task_id=%s",
+                task.id,
+                task.external_task_id,
             )
         except IntegrityError:
             self.db.rollback()
             raise HTTPException(status_code=409, detail="Task already exists")
 
+        # 先构建返回给客户端的响应（使用短 ID 对外暴露）
         internal_task = InternalVideoTask(
-            id=task.id,
+            id=task.short_id,
             external_id=external_task_id,
             status=VideoStatus.SUBMITTED,
             created_at=task.created_at,
             original_request=internal_request,
         )
         response_body = self._normalizer.video_task_from_internal(internal_task)
+
+        # 提交成功后立即结算 Usage（费用暂时为 0，轮询完成后更新）
+        response_time_ms = int((time.time() - self.start_time) * 1000)
+        try:
+            # 构建发送给上游的请求头（脱敏）
+            upstream_request_headers = self._build_upstream_headers(
+                original_headers,
+                "",  # key 不重要，只是用于记录
+                outcome.candidate.endpoint,
+                None,  # auth_info
+            )
+
+            UsageService.finalize_submitted(
+                self.db,
+                request_id=self.request_id,
+                provider_name=outcome.candidate.provider.name,
+                provider_id=outcome.candidate.provider.id,
+                provider_endpoint_id=outcome.candidate.endpoint.id,
+                provider_api_key_id=outcome.candidate.key.id,
+                response_time_ms=response_time_ms,
+                status_code=outcome.upstream_status_code or 200,
+                endpoint_api_format=make_signature_key(
+                    str(getattr(outcome.candidate.endpoint, "api_family", "")).strip().lower(),
+                    str(getattr(outcome.candidate.endpoint, "endpoint_kind", "")).strip().lower(),
+                ),
+                provider_request_headers=upstream_request_headers,
+                response_headers=outcome.upstream_headers,
+                response_body=response_body,  # 使用我们转换后的响应（包含我们的 ID）
+            )
+            self.db.commit()
+        except Exception as exc:
+            logger.warning(
+                "Failed to finalize submitted usage for video request_id=%s: %s",
+                self.request_id,
+                sanitize_error_message(str(exc)),
+            )
+
         return JSONResponse(response_body)
 
     async def handle_get_task(
@@ -234,6 +367,28 @@ class GeminiVeoHandler(VideoHandlerBase):
 
         task.status = VideoStatus.CANCELLED.value
         task.updated_at = datetime.now(timezone.utc)
+
+        # 将 Usage 作废（不收费）
+        # 尝试 finalize_void（处理 pending）和 void_settled（处理已 settled）
+        try:
+            voided = UsageService.finalize_void(
+                self.db,
+                request_id=task.request_id,
+                reason="cancelled_by_user",
+            )
+            if not voided:
+                # pending 状态未找到，尝试处理已 settled 的记录
+                UsageService.void_settled(
+                    self.db,
+                    request_id=task.request_id,
+                    reason="cancelled_by_user",
+                )
+        except Exception as exc:
+            logger.warning(
+                "Failed to void usage for cancelled task=%s: %s",
+                task.id,
+                sanitize_error_message(str(exc)),
+            )
         self.db.commit()
         return JSONResponse({})
 
@@ -275,12 +430,38 @@ class GeminiVeoHandler(VideoHandlerBase):
             if task.video_expires_at < now:
                 raise HTTPException(status_code=410, detail="Video URL has expired")
 
+        # 获取 provider 的认证信息（Gemini 下载视频需要带 API Key）
+        endpoint, key = self._get_endpoint_and_key(task)
+        download_headers: dict[str, str] = {}
+        if key.api_key:
+            try:
+                upstream_key = crypto_service.decrypt(key.api_key)
+                # Gemini API 使用 x-goog-api-key 头进行认证
+                download_headers["x-goog-api-key"] = upstream_key
+
+                # 如果是 Vertex AI，需要使用 OAuth Bearer token
+                auth_info = await get_provider_auth(endpoint, key)
+                if auth_info:
+                    download_headers.pop("x-goog-api-key", None)
+                    download_headers[auth_info.auth_header] = auth_info.auth_value
+            except Exception as exc:
+                logger.warning(
+                    "[VideoDownload] Failed to get auth for download task=%s: %s",
+                    task.id,
+                    sanitize_error_message(str(exc)),
+                )
+                # 继续尝试无认证下载（某些 URL 可能是预签名的）
+
         # 代理下载而非直接重定向，避免暴露上游存储 URL
-        client = await HTTPClientPool.get_default_client_async()
+        # 使用 httpx 支持重定向（Gemini 视频 URL 会重定向到实际存储位置）
+        import httpx
+
         try:
-            request = client.build_request("GET", task.video_url)
-            # 视频下载可能较大，设置 5 分钟超时
-            response = await client.send(request, stream=True, timeout=300.0)
+            # 使用 follow_redirects=True 跟随重定向
+            async with httpx.AsyncClient(
+                follow_redirects=True, timeout=httpx.Timeout(300.0)
+            ) as client:
+                response = await client.get(task.video_url, headers=download_headers)
         except Exception as exc:
             logger.error(
                 "[VideoDownload] Upstream fetch failed user=%s task=%s: %s",
@@ -291,21 +472,14 @@ class GeminiVeoHandler(VideoHandlerBase):
             raise HTTPException(status_code=502, detail="Failed to fetch video")
 
         if response.status_code >= 400:
-            await response.aclose()
             raise HTTPException(status_code=response.status_code, detail="Upstream error")
 
-        async def _iter_bytes() -> AsyncIterator[bytes]:
-            try:
-                async for chunk in response.aiter_bytes():
-                    yield chunk
-            finally:
-                await response.aclose()
-
+        # 返回完整的视频内容（非 streaming，因为需要跟随重定向）
         safe_headers = {
             k: v for k, v in response.headers.items() if k.lower() not in HOP_BY_HOP_HEADERS
         }
-        return StreamingResponse(
-            _iter_bytes(),
+        return Response(
+            content=response.content,
             status_code=response.status_code,
             headers=safe_headers,
             media_type=response.headers.get("content-type", "video/mp4"),
@@ -375,6 +549,36 @@ class GeminiVeoHandler(VideoHandlerBase):
             "status": error.get("status", "BAD_GATEWAY"),
         }
 
+    # ------------------------------------------------------------------
+    # OpenAI format conversion helpers
+    # ------------------------------------------------------------------
+
+    def _build_openai_upstream_url(self, base_url: str | None) -> str:
+        """构建 OpenAI Sora API 的上游 URL"""
+        base = (base_url or "https://api.openai.com").rstrip("/")
+        if base.endswith("/v1"):
+            return f"{base}/videos"
+        return f"{base}/v1/videos"
+
+    def _build_openai_upstream_headers(
+        self,
+        original_headers: dict[str, str],
+        upstream_key: str,
+        endpoint: ProviderEndpoint,
+    ) -> dict[str, str]:
+        """构建 OpenAI 格式的请求头"""
+        extra_headers = get_extra_headers_from_endpoint(endpoint)
+        endpoint_sig = make_signature_key(
+            str(getattr(endpoint, "api_family", "")).strip().lower(),
+            str(getattr(endpoint, "endpoint_kind", "")).strip().lower(),
+        )
+        return build_upstream_headers_for_endpoint(
+            original_headers,
+            endpoint_sig,
+            upstream_key,
+            endpoint_headers=extra_headers,
+        )
+
     def _create_task_record(
         self,
         *,
@@ -385,6 +589,8 @@ class GeminiVeoHandler(VideoHandlerBase):
         candidate_keys: list[dict[str, Any]] | None = None,
         original_headers: dict[str, str] | None = None,
         billing_rule_snapshot: dict[str, Any] | None = None,
+        converted_request_body: dict[str, Any] | None = None,
+        format_converted: bool = False,
     ) -> VideoTask:
         now = datetime.now(timezone.utc)
 
@@ -407,8 +613,14 @@ class GeminiVeoHandler(VideoHandlerBase):
             }
             request_metadata["request_headers"] = safe_headers
 
+        provider_api_format = make_signature_key(
+            str(getattr(candidate.endpoint, "api_family", "")).strip().lower(),
+            str(getattr(candidate.endpoint, "endpoint_kind", "")).strip().lower(),
+        )
+
         return VideoTask(
             id=str(uuid4()),
+            request_id=self.request_id,
             external_task_id=external_task_id,
             user_id=self.user.id,
             api_key_id=self.api_key.id,
@@ -416,15 +628,12 @@ class GeminiVeoHandler(VideoHandlerBase):
             endpoint_id=candidate.endpoint.id,
             key_id=candidate.key.id,
             client_api_format=self.FORMAT_ID,
-            provider_api_format=make_signature_key(
-                str(getattr(candidate.endpoint, "api_family", "")).strip().lower(),
-                str(getattr(candidate.endpoint, "endpoint_kind", "")).strip().lower(),
-            ),
-            format_converted=False,
+            provider_api_format=provider_api_format,
+            format_converted=format_converted,
             model=internal_request.model,
             prompt=internal_request.prompt,
             original_request_body=original_request_body,
-            converted_request_body=original_request_body,
+            converted_request_body=converted_request_body or original_request_body,
             duration_seconds=internal_request.duration_seconds,
             resolution=internal_request.resolution,
             aspect_ratio=internal_request.aspect_ratio,
@@ -438,30 +647,45 @@ class GeminiVeoHandler(VideoHandlerBase):
             request_metadata=request_metadata,
         )
 
-    def _get_task_by_external_id(self, external_id: str) -> VideoTask:
-        """按 external_task_id 查找任务（Gemini 使用 operations/{id} 格式）"""
-        normalized_id = normalize_gemini_operation_id(external_id)
-
-        logger.info(
-            f"[GeminiVeoHandler] Looking for task: normalized_id={normalized_id}, user_id={self.user.id}"
+    def _task_to_internal(self, task: VideoTask) -> InternalVideoTask:
+        """覆盖父类方法，Gemini 使用 short_id 作为对外暴露的 ID"""
+        try:
+            status = VideoStatus(task.status)
+        except ValueError:
+            status = VideoStatus.PENDING
+        return InternalVideoTask(
+            id=task.short_id,  # Gemini 使用短 ID
+            external_id=task.external_task_id,
+            status=status,
+            progress_percent=task.progress_percent or 0,
+            progress_message=task.progress_message,
+            video_url=task.video_url,
+            video_urls=task.video_urls or [],
+            created_at=task.created_at,
+            completed_at=task.completed_at,
+            error_code=task.error_code,
+            error_message=task.error_message,
+            extra={"model": task.model},
         )
 
+    def _get_task_by_external_id(self, external_id: str) -> VideoTask:
+        """按 short_id 查找任务（我们对外暴露的 operation 格式是 models/{model}/operations/{short_id}）"""
+        from src.api.handlers.base.video_handler_base import extract_short_id_from_operation
+
+        short_id = extract_short_id_from_operation(external_id)
+
+        # 通过 short_id 查找任务
         task = (
             self.db.query(VideoTask)
             .filter(
-                VideoTask.external_task_id == normalized_id,
+                VideoTask.short_id == short_id,
                 VideoTask.user_id == self.user.id,
             )
             .first()
         )
         if not task:
-            logger.warning(
-                f"[GeminiVeoHandler] Task not found: normalized_id={normalized_id}, user_id={self.user.id}"
-            )
+            logger.debug("[GeminiVeoHandler] Task not found: short_id=%s", short_id)
             raise HTTPException(status_code=404, detail="Video task not found")
-        logger.info(
-            f"[GeminiVeoHandler] Task found: id={task.id}, external_task_id={task.external_task_id}"
-        )
         return task
 
 

src/api/handlers/openai/video_handler.py

@@ -5,6 +5,7 @@ OpenAI Video Handler - Sora 视频生成实现
 from __future__ import annotations
 
 import json
+import time
 from datetime import datetime, timedelta, timezone
 from typing import Any, AsyncIterator
 from uuid import uuid4
@@ -14,6 +15,7 @@ from fastapi.responses import JSONResponse, Response, StreamingResponse
 from sqlalchemy.exc import IntegrityError
 from sqlalchemy.orm import Session
 
+from src.api.handlers.base.request_builder import get_provider_auth
 from src.api.handlers.base.video_handler_base import VideoHandlerBase, sanitize_error_message
 from src.clients.http_client import HTTPClientPool
 from src.config.settings import config
@@ -30,6 +32,7 @@ from src.core.api_format.conversion.internal_video import (
     VideoStatus,
 )
 from src.core.api_format.conversion.normalizers.openai import OpenAINormalizer
+from src.core.api_format.conversion.registry import format_conversion_registry
 from src.core.api_format.headers import HOP_BY_HOP_HEADERS
 from src.core.crypto import crypto_service
 from src.core.logger import logger
@@ -91,16 +94,91 @@ class OpenAIVideoHandler(VideoHandlerBase):
             )
             raise HTTPException(status_code=400, detail=str(e))
 
+        # 异步任务：提前创建 pending usage，便于前端看到“处理中”
+        try:
+            UsageService.create_pending_usage(
+                db=self.db,
+                request_id=self.request_id,
+                user=self.user,
+                api_key=self.api_key,
+                model=internal_request.model,
+                is_stream=False,
+                request_type="video",
+                api_format=self.FORMAT_ID,
+                request_headers=original_headers,
+                request_body=original_request_body,
+            )
+        except Exception as exc:
+            logger.warning(
+                "Failed to create pending usage for video request_id=%s: %s",
+                self.request_id,
+                sanitize_error_message(str(exc)),
+            )
+
+        # 用于跟踪是否发生了格式转换
+        format_conversion_info: dict[str, Any] = {
+            "converted": False,
+            "provider_format": None,
+        }
+
         async def _submit(candidate: ProviderCandidate) -> Any:
             upstream_key, endpoint, _provider_key = await self._resolve_upstream_key(candidate)
-            upstream_url = self._build_upstream_url(endpoint.base_url)
-            headers = self._build_upstream_headers(original_headers, upstream_key, endpoint)
-            client = await HTTPClientPool.get_default_client_async()
-            return await client.post(upstream_url, headers=headers, json=original_request_body)
+
+            # 检测目标格式
+            provider_format = make_signature_key(
+                str(getattr(endpoint, "api_family", "")).strip().lower(),
+                str(getattr(endpoint, "endpoint_kind", "")).strip().lower(),
+            )
+            needs_conversion = provider_format.upper() != self.FORMAT_ID.upper()
+            format_conversion_info["provider_format"] = provider_format
+            format_conversion_info["converted"] = needs_conversion
+
+            # 确保 seconds 字段为字符串类型（上游 Go 服务要求 string）
+            request_body = original_request_body.copy()
+            if "seconds" in request_body and request_body["seconds"] is not None:
+                request_body["seconds"] = str(request_body["seconds"])
+
+            if needs_conversion and provider_format.upper().startswith("GEMINI:"):
+                # OpenAI -> Gemini 格式转换
+                converted_body = format_conversion_registry.convert_video_request(
+                    request_body,
+                    self.FORMAT_ID,
+                    provider_format,
+                )
+                # 如果 model 不在请求体中，从路径或内部请求中获取
+                if "model" not in converted_body:
+                    converted_body["model"] = internal_request.model
+
+                # 构建 Gemini 风格的 URL
+                upstream_url = self._build_gemini_upstream_url(
+                    endpoint.base_url, internal_request.model
+                )
+
+                # 构建 Gemini 风格的请求头
+                auth_info = await get_provider_auth(endpoint, _provider_key)
+                headers = self._build_gemini_upstream_headers(
+                    original_headers, upstream_key, endpoint, auth_info
+                )
+
+                client = await HTTPClientPool.get_default_client_async()
+                return await client.post(upstream_url, headers=headers, json=converted_body)
+            else:
+                # 原始 OpenAI 格式
+                upstream_url = self._build_upstream_url(endpoint.base_url)
+                headers = self._build_upstream_headers(original_headers, upstream_key, endpoint)
+                client = await HTTPClientPool.get_default_client_async()
+                return await client.post(upstream_url, headers=headers, json=request_body)
 
         def _extract_task_id(payload: dict[str, Any]) -> str | None:
-            value = payload.get("id")
-            return str(value) if value else None
+            # 根据响应格式提取 task ID
+            # OpenAI: {"id": "..."}
+            # Gemini: {"name": "operations/..."}
+            if "id" in payload:
+                return str(payload["id"])
+            if "name" in payload:
+                # Gemini 格式
+                return str(payload["name"])
+            return None
 
         # 捕获提交阶段的所有错误，记录失败任务
         try:
@@ -110,8 +188,8 @@ class OpenAIVideoHandler(VideoHandlerBase):
                 task_type="video",
                 submit_func=_submit,
                 extract_external_task_id=_extract_task_id,
-                supported_auth_types={"api_key"},
-                allow_format_conversion=False,
+                supported_auth_types={"api_key", "vertex_ai"},
+                allow_format_conversion=True,
                 max_candidates=10,
             )
         except HTTPException as exc:
@@ -156,14 +234,31 @@ class OpenAIVideoHandler(VideoHandlerBase):
 
         external_task_id = outcome.external_task_id
 
+        # 如果发生了格式转换，记录转换后的请求体
+        converted_request_body = original_request_body
+        if format_conversion_info["converted"]:
+            try:
+                converted_request_body = format_conversion_registry.convert_video_request(
+                    original_request_body,
+                    self.FORMAT_ID,
+                    format_conversion_info["provider_format"],
+                )
+            except Exception as e:
+                logger.warning(
+                    "[OpenAIVideoHandler] Failed to record converted request: %s",
+                    sanitize_error_message(str(e)),
+                )
+
         task = self._create_task_record(
             external_task_id=external_task_id,
             candidate=outcome.candidate,
             original_request_body=original_request_body,
+            converted_request_body=converted_request_body,
             internal_request=internal_request,
             candidate_keys=outcome.candidate_keys,
             original_headers=original_headers,
             billing_rule_snapshot=billing_rule_snapshot,
+            format_converted=format_conversion_info["converted"],
         )
         try:
             self.db.add(task)
@@ -174,6 +269,7 @@ class OpenAIVideoHandler(VideoHandlerBase):
             self.db.rollback()
             raise HTTPException(status_code=409, detail="Task already exists")
 
+        # 先构建返回给客户端的响应（OpenAI Sora 使用 UUID）
         internal_task = InternalVideoTask(
             id=task.id,
             external_id=external_task_id,
@@ -182,6 +278,42 @@ class OpenAIVideoHandler(VideoHandlerBase):
             original_request=internal_request,
         )
         response_body = self._normalizer.video_task_from_internal(internal_task)
+
+        # 提交成功后立即结算 Usage（费用暂时为 0，轮询完成后更新）
+        response_time_ms = int((time.time() - self.start_time) * 1000)
+        try:
+            # 构建发送给上游的请求头（脱敏）
+            upstream_request_headers = self._build_upstream_headers(
+                original_headers,
+                "",  # key 不重要，只是用于记录
+                outcome.candidate.endpoint,
+            )
+
+            UsageService.finalize_submitted(
+                self.db,
+                request_id=self.request_id,
+                provider_name=outcome.candidate.provider.name,
+                provider_id=outcome.candidate.provider.id,
+                provider_endpoint_id=outcome.candidate.endpoint.id,
+                provider_api_key_id=outcome.candidate.key.id,
+                response_time_ms=response_time_ms,
+                status_code=outcome.upstream_status_code or 200,
+                endpoint_api_format=make_signature_key(
+                    str(getattr(outcome.candidate.endpoint, "api_family", "")).strip().lower(),
+                    str(getattr(outcome.candidate.endpoint, "endpoint_kind", "")).strip().lower(),
+                ),
+                provider_request_headers=upstream_request_headers,
+                response_headers=outcome.upstream_headers,
+                response_body=response_body,  # 使用我们转换后的响应（包含我们的 ID）
+            )
+            self.db.commit()
+        except Exception as exc:
+            logger.warning(
+                "Failed to finalize submitted usage for video request_id=%s: %s",
+                self.request_id,
+                sanitize_error_message(str(exc)),
+            )
+
         return JSONResponse(response_body)
 
     async def handle_get_task(
@@ -206,17 +338,59 @@ class OpenAIVideoHandler(VideoHandlerBase):
         query_params: dict[str, str] | None = None,
         path_params: dict[str, Any] | None = None,
     ) -> JSONResponse:
-        tasks = (
-            self.db.query(VideoTask)
-            .filter(VideoTask.user_id == self.user.id)
-            .order_by(VideoTask.created_at.desc())
-            .limit(100)
-            .all()
-        )
+        params = query_params or {}
+
+        # 解析分页参数
+        after = params.get("after")
+        try:
+            limit = min(int(params.get("limit") or 20), 100)  # 默认 20，最大 100
+        except (ValueError, TypeError):
+            limit = 20
+        order = params.get("order", "desc").lower()
+        if order not in ("asc", "desc"):
+            order = "desc"
+
+        # 构建查询
+        query = self.db.query(VideoTask).filter(VideoTask.user_id == self.user.id)
+
+        # 处理游标分页（after 参数使用 UUID）
+        if after:
+            after_task = (
+                self.db.query(VideoTask)
+                .filter(VideoTask.id == after, VideoTask.user_id == self.user.id)
+                .first()
+            )
+            if after_task and after_task.created_at:
+                if order == "desc":
+                    query = query.filter(VideoTask.created_at < after_task.created_at)
+                else:
+                    query = query.filter(VideoTask.created_at > after_task.created_at)
+
+        # 排序
+        if order == "asc":
+            query = query.order_by(VideoTask.created_at.asc())
+        else:
+            query = query.order_by(VideoTask.created_at.desc())
+
+        # 获取 limit + 1 条记录以判断是否有更多数据
+        tasks = query.limit(limit + 1).all()
+        has_more = len(tasks) > limit
+        tasks = tasks[:limit]
+
         items = [
             self._normalizer.video_task_from_internal(self._task_to_internal(t)) for t in tasks
         ]
-        return JSONResponse({"object": "list", "data": items})
+
+        response_data: dict[str, Any] = {
+            "object": "list",
+            "data": items,
+            "has_more": has_more,
+        }
+        # 如果有更多数据，返回最后一条的 ID 作为下一页游标
+        if has_more and tasks:
+            response_data["last_id"] = tasks[-1].id
+
+        return JSONResponse(response_data)
 
     async def handle_cancel_task(
         self,
@@ -245,9 +419,80 @@ class OpenAIVideoHandler(VideoHandlerBase):
 
         task.status = VideoStatus.CANCELLED.value
         task.updated_at = datetime.now(timezone.utc)
+
+        # 将 Usage 作废（不收费）
+        # 尝试 finalize_void（处理 pending）和 void_settled（处理已 settled）
+        try:
+            voided = UsageService.finalize_void(
+                self.db,
+                request_id=task.request_id,
+                reason="cancelled_by_user",
+            )
+            if not voided:
+                # pending 状态未找到，尝试处理已 settled 的记录
+                UsageService.void_settled(
+                    self.db,
+                    request_id=task.request_id,
+                    reason="cancelled_by_user",
+                )
+        except Exception as exc:
+            logger.warning(
+                "Failed to void usage for cancelled task=%s: %s",
+                task.id,
+                sanitize_error_message(str(exc)),
+            )
         self.db.commit()
         return JSONResponse({})
 
+    async def handle_delete_task(
+        self,
+        *,
+        task_id: str,
+        http_request: Request,
+        original_headers: dict[str, str],
+        query_params: dict[str, str] | None = None,
+        path_params: dict[str, Any] | None = None,
+    ) -> JSONResponse:
+        """删除已完成或失败的视频及其存储资源"""
+        task = self._get_task(task_id)
+
+        # 只能删除已完成或失败的视频
+        if task.status not in (VideoStatus.COMPLETED.value, VideoStatus.FAILED.value):
+            raise HTTPException(
+                status_code=400,
+                detail=f"Can only delete completed or failed videos (current status: {task.status})",
+            )
+
+        # 如果有 external_task_id，向上游发送删除请求
+        if task.external_task_id:
+            try:
+                endpoint, key = self._get_endpoint_and_key(task)
+                if key.api_key:
+                    upstream_key = crypto_service.decrypt(key.api_key)
+                    upstream_url = self._build_upstream_url(
+                        endpoint.base_url, task.external_task_id
+                    )
+                    headers = self._build_upstream_headers(original_headers, upstream_key, endpoint)
+
+                    client = await HTTPClientPool.get_default_client_async()
+                    response = await client.delete(upstream_url, headers=headers)
+                    if response.status_code >= 400 and response.status_code != 404:
+                        # 404 表示上游已删除，不算错误
+                        return self._build_error_response(response)
+            except Exception as exc:
+                logger.warning(
+                    "Failed to delete video from upstream task=%s: %s",
+                    task.id,
+                    sanitize_error_message(str(exc)),
+                )
+                # 继续删除本地记录
+
+        # 删除本地任务记录
+        self.db.delete(task)
+        self.db.commit()
+
+        return JSONResponse({"id": task_id, "object": "video", "deleted": True})
+
     async def handle_remix_task(
         self,
         *,
@@ -280,8 +525,13 @@ class OpenAIVideoHandler(VideoHandlerBase):
         )
         headers = self._build_upstream_headers(original_headers, upstream_key, endpoint)
 
+        # 确保 seconds 字段为字符串类型（上游 Go 服务要求 string）
+        request_body = original_request_body.copy()
+        if "seconds" in request_body and request_body["seconds"] is not None:
+            request_body["seconds"] = str(request_body["seconds"])
+
         client = await HTTPClientPool.get_default_client_async()
-        response = await client.post(upstream_url, headers=headers, json=original_request_body)
+        response = await client.post(upstream_url, headers=headers, json=request_body)
 
         if response.status_code >= 400:
             return self._build_error_response(response)
@@ -350,7 +600,7 @@ class OpenAIVideoHandler(VideoHandlerBase):
             raise HTTPException(status_code=409, detail="Task already exists")
 
         internal_task = InternalVideoTask(
-            id=task.id,
+            id=task.id,  # OpenAI Sora 使用 UUID
             external_id=external_task_id,
             status=VideoStatus.SUBMITTED,
             created_at=task.created_at,
@@ -389,15 +639,25 @@ class OpenAIVideoHandler(VideoHandlerBase):
         if task.status == VideoStatus.CANCELLED.value:
             raise HTTPException(status_code=404, detail="Video task was cancelled")
 
+        # 支持 variant 查询参数: video (默认), thumbnail, spritesheet
+        variant = (query_params or {}).get("variant", "video")
+
+        # 如果 video_url 是完整的 HTTP URL，直接代理该 URL（适用于不支持 /content 端点的上游如 API易）
+        # 保持流式代理而非重定向，确保客户端行为与官方 OpenAI 一致
+        if variant == "video" and task.video_url and task.video_url.startswith("http"):
+            logger.debug(
+                "[VideoDownload] Proxying direct URL task=%s url=%s",
+                task_id,
+                task.video_url,
+            )
+            return await self._proxy_direct_url(task.video_url, task_id)
+
         if not task.external_task_id:
             raise HTTPException(status_code=500, detail="Task missing external_task_id")
         endpoint, key = self._get_endpoint_and_key(task)
         if not key.api_key:
             raise HTTPException(status_code=500, detail="Provider key not configured")
         upstream_key = crypto_service.decrypt(key.api_key)
-
-        # 支持 variant 查询参数: video (默认), thumbnail, spritesheet
-        variant = (query_params or {}).get("variant", "video")
         if variant not in {"video", "thumbnail", "spritesheet"}:
             raise HTTPException(
                 status_code=400,
@@ -412,6 +672,12 @@ class OpenAIVideoHandler(VideoHandlerBase):
         headers = self._build_upstream_headers(original_headers, upstream_key, endpoint)
 
         client = await HTTPClientPool.get_default_client_async()
+        logger.debug(
+            "[VideoDownload] Requesting upstream url=%s task=%s external_task_id=%s",
+            upstream_url,
+            task_id,
+            task.external_task_id,
+        )
         try:
             # 使用 httpx 的 stream 方法并正确管理上下文
             # 视频下载可能较大，设置 5 分钟超时
@@ -419,8 +685,9 @@ class OpenAIVideoHandler(VideoHandlerBase):
             response = await client.send(request, stream=True, timeout=300.0)
         except Exception as exc:
             logger.warning(
-                "[VideoDownload] Upstream connection failed task=%s: %s",
+                "[VideoDownload] Upstream connection failed task=%s url=%s: %s",
                 task_id,
+                upstream_url,
                 sanitize_error_message(str(exc)),
             )
             raise HTTPException(status_code=502, detail="Upstream connection failed") from exc
@@ -483,6 +750,46 @@ class OpenAIVideoHandler(VideoHandlerBase):
             raise HTTPException(status_code=500, detail="Failed to decrypt provider key")
         return upstream_key, candidate.endpoint, candidate.key
 
+    async def _proxy_direct_url(self, url: str, task_id: str) -> Response | StreamingResponse:
+        """代理直接的视频 URL（如 CDN URL），保持与官方 API 一致的流式返回行为"""
+        client = await HTTPClientPool.get_default_client_async()
+        try:
+            request = client.build_request("GET", url)
+            response = await client.send(request, stream=True, timeout=300.0)
+        except Exception as exc:
+            logger.warning(
+                "[VideoDownload] Direct URL connection failed task=%s url=%s: %s",
+                task_id,
+                url,
+                sanitize_error_message(str(exc)),
+            )
+            raise HTTPException(status_code=502, detail="Video download failed") from exc
+
+        if response.status_code >= 400:
+            await response.aread()  # consume body before closing
+            await response.aclose()
+            return JSONResponse(
+                status_code=response.status_code,
+                content={"error": {"type": "upstream_error", "message": "Video not available"}},
+            )
+
+        async def _iter_bytes() -> AsyncIterator[bytes]:
+            try:
+                async for chunk in response.aiter_bytes():
+                    yield chunk
+            finally:
+                await response.aclose()
+
+        safe_headers = {
+            k: v for k, v in response.headers.items() if k.lower() not in HOP_BY_HOP_HEADERS
+        }
+        return StreamingResponse(
+            _iter_bytes(),
+            status_code=response.status_code,
+            headers=safe_headers,
+            media_type=response.headers.get("content-type", "video/mp4"),
+        )
+
     def _build_upstream_url(self, base_url: str | None, suffix: str | None = None) -> str:
         base = (base_url or self.DEFAULT_BASE_URL).rstrip("/")
         if base.endswith("/v1"):
@@ -508,6 +815,42 @@ class OpenAIVideoHandler(VideoHandlerBase):
             endpoint_headers=extra_headers,
         )
 
+    # ------------------------------------------------------------------
+    # Gemini format conversion helpers
+    # ------------------------------------------------------------------
+
+    def _build_gemini_upstream_url(self, base_url: str | None, model: str) -> str:
+        """构建 Gemini Veo API 的上游 URL"""
+        base = (base_url or "https://generativelanguage.googleapis.com").rstrip("/")
+        if base.endswith("/v1beta"):
+            base = base[: -len("/v1beta")]
+        return f"{base}/v1beta/models/{model}:predictLongRunning"
+
+    def _build_gemini_upstream_headers(
+        self,
+        original_headers: dict[str, str],
+        upstream_key: str,
+        endpoint: ProviderEndpoint,
+        auth_info: Any | None,
+    ) -> dict[str, str]:
+        """构建 Gemini 格式的请求头"""
+        extra_headers = get_extra_headers_from_endpoint(endpoint)
+        endpoint_sig = make_signature_key(
+            str(getattr(endpoint, "api_family", "")).strip().lower(),
+            str(getattr(endpoint, "endpoint_kind", "")).strip().lower(),
+        )
+        headers = build_upstream_headers_for_endpoint(
+            original_headers,
+            endpoint_sig,
+            upstream_key,
+            endpoint_headers=extra_headers,
+        )
+        if auth_info:
+            # 覆盖为 OAuth2 Bearer（Vertex AI）
+            headers.pop("x-goog-api-key", None)
+            headers[auth_info.auth_header] = auth_info.auth_value
+        return headers
+
     # _build_error_response 继承自基类 VideoHandlerBase
 
     def _create_task_record(
@@ -520,6 +863,8 @@ class OpenAIVideoHandler(VideoHandlerBase):
         candidate_keys: list[dict[str, Any]] | None = None,
         original_headers: dict[str, str] | None = None,
         billing_rule_snapshot: dict[str, Any] | None = None,
+        converted_request_body: dict[str, Any] | None = None,
+        format_converted: bool = False,
     ) -> VideoTask:
         now = datetime.now(timezone.utc)
         size = internal_request.extra.get("original_size")
@@ -543,8 +888,14 @@ class OpenAIVideoHandler(VideoHandlerBase):
             }
             request_metadata["request_headers"] = safe_headers
 
+        provider_api_format = make_signature_key(
+            str(getattr(candidate.endpoint, "api_family", "")).strip().lower(),
+            str(getattr(candidate.endpoint, "endpoint_kind", "")).strip().lower(),
+        )
+
         return VideoTask(
             id=str(uuid4()),
+            request_id=self.request_id,
             external_task_id=external_task_id,
             user_id=self.user.id,
             api_key_id=self.api_key.id,
@@ -552,15 +903,12 @@ class OpenAIVideoHandler(VideoHandlerBase):
             endpoint_id=candidate.endpoint.id,
             key_id=candidate.key.id,
             client_api_format=self.FORMAT_ID,
-            provider_api_format=make_signature_key(
-                str(getattr(candidate.endpoint, "api_family", "")).strip().lower(),
-                str(getattr(candidate.endpoint, "endpoint_kind", "")).strip().lower(),
-            ),
-            format_converted=False,
+            provider_api_format=provider_api_format,
+            format_converted=format_converted,
             model=internal_request.model,
             prompt=internal_request.prompt,
             original_request_body=original_request_body,
-            converted_request_body=original_request_body,
+            converted_request_body=converted_request_body or original_request_body,
             duration_seconds=internal_request.duration_seconds,
             resolution=internal_request.resolution,
             aspect_ratio=internal_request.aspect_ratio,
@@ -580,8 +928,23 @@ class OpenAIVideoHandler(VideoHandlerBase):
             status = VideoStatus(task.status)
         except ValueError:
             status = VideoStatus.PENDING
+
+        # 构建 extra 字段
+        extra: dict[str, Any] = {
+            "model": task.model,
+            "size": task.size,
+            "seconds": str(task.duration_seconds) if task.duration_seconds else None,
+            "prompt": task.prompt,
+        }
+
+        # 检查是否是 remix 视频
+        if task.original_request_body and isinstance(task.original_request_body, dict):
+            remixed_from = task.original_request_body.get("remix_video_id")
+            if remixed_from:
+                extra["remixed_from_video_id"] = remixed_from
+
         return InternalVideoTask(
-            id=task.id,
+            id=task.id,  # OpenAI Sora 使用 UUID
             external_id=task.external_task_id,
             status=status,
             progress_percent=task.progress_percent or 0,
@@ -596,7 +959,7 @@ class OpenAIVideoHandler(VideoHandlerBase):
             expires_at=task.video_expires_at,
             error_code=task.error_code,
             error_message=task.error_message,
-            extra={"model": task.model, "size": task.size, "seconds": task.duration_seconds},
+            extra=extra,
         )
 
     async def _record_failed_usage(
@@ -609,8 +972,6 @@ class OpenAIVideoHandler(VideoHandlerBase):
         original_headers: dict[str, str],
     ) -> None:
         """记录失败请求的使用记录（无任务记录）"""
-        import time
-
         response_time_ms = int((time.time() - self.start_time) * 1000)
         safe_headers = {
             k: v
@@ -669,8 +1030,6 @@ class OpenAIVideoHandler(VideoHandlerBase):
         candidate_keys: list[dict[str, Any]] | None = None,
     ) -> None:
         """创建失败的任务记录和使用记录"""
-        import time
-
         now = datetime.now(timezone.utc)
         response_time_ms = int((time.time() - self.start_time) * 1000)
 
@@ -696,6 +1055,7 @@ class OpenAIVideoHandler(VideoHandlerBase):
         # 创建失败的任务记录
         task = VideoTask(
             id=str(uuid4()),
+            request_id=self.request_id,
             external_task_id=None,
             user_id=self.user.id,
             api_key_id=self.api_key.id,

src/api/public/__init__.py

@@ -14,13 +14,14 @@ from .system_catalog import router as system_catalog_router
 from .videos import router as videos_router
 
 router = APIRouter()
-# Models API 需要在最前面注册，避免被其他路由的 path 参数捕获
+# Video API 路由需要在 Models API 之前注册，因为 Models API 有 /v1beta/models/{path} 通配符路由
+# 会错误匹配 /v1beta/models/{model}/operations/{id}/content 等视频路由
+router.include_router(videos_router, tags=["Video Generation"])
 router.include_router(models_router)
 router.include_router(claude_router, tags=["Claude API"])
 router.include_router(openai_router)
 router.include_router(gemini_router, tags=["Gemini API"])
 router.include_router(gemini_files_router, tags=["Gemini Files API"])
-router.include_router(videos_router, tags=["Video Generation"])
 router.include_router(system_catalog_router, tags=["System Catalog"])
 router.include_router(catalog_router)
 router.include_router(capabilities_router)

src/api/public/gemini_files.py

@@ -15,14 +15,17 @@ Gemini Files API 代理端点
 
 参考文档：
 https://ai.google.dev/api/files
+
+优化：HTTP 代理请求期间不持有数据库连接，避免阻塞其他请求。
 """
 
 from __future__ import annotations
 
+from dataclasses import dataclass
 from typing import Any, Dict, Optional, Tuple
 from urllib.parse import urlencode
 
-from fastapi import APIRouter, Depends, HTTPException, Request, Response
+from fastapi import APIRouter, HTTPException, Request, Response
 from fastapi.responses import JSONResponse
 from sqlalchemy.orm import Session
 
@@ -30,20 +33,30 @@ from src.clients.http_client import HTTPClientPool
 from src.core.api_format import get_auth_handler, get_default_auth_method_for_endpoint
 from src.core.crypto import crypto_service
 from src.core.logger import logger
-from src.database import get_db
+from src.database import create_session
 from src.models.database import ApiKey, GlobalModel, Model, Provider, ProviderEndpoint, User
 from src.services.auth.service import AuthService
 from src.services.cache.aware_scheduler import CacheAwareScheduler, ProviderCandidate
 from src.services.gemini_files_mapping import delete_file_key_mapping, store_file_key_mapping
 from src.services.provider.transport import redact_url_for_log
 
+
+@dataclass
+class UpstreamContext:
+    """上游请求上下文（不依赖数据库会话）"""
+
+    upstream_key: str
+    base_url: str
+    file_key_id: str
+    user_id: str
+
+
 router = APIRouter(tags=["Gemini Files API"])
 
 # Gemini Files API 基础 URL
 GEMINI_FILES_BASE_URL = "https://generativelanguage.googleapis.com"
 
-# Gemini Files API 能力标签
-REQUIRED_CAPABILITIES = {"gemini_files_api": True}
+# Gemini Files API 无能力限制（任何 Gemini key 都可用）
 
 # 需要从客户端请求中移除的头部（这些会由代理重新设置或不应转发）
 HEADERS_TO_REMOVE = frozenset(
@@ -184,9 +197,25 @@ async def _select_provider_candidate(
     db: Session,
     user_api_key: ApiKey,
     model_name: str,
+    require_files_capability: bool = True,
 ) -> ProviderCandidate | None:
-    """选择支持 Files API 的 Provider/Endpoint/Key 组合"""
+    """
+    选择可用的 Provider/Endpoint/Key 组合
+
+    Args:
+        db: 数据库会话
+        user_api_key: 用户 API Key
+        model_name: 模型名称
+        require_files_capability: 是否要求 gemini_files 能力（默认 True）
+
+    Returns:
+        匹配的候选，如果没有则返回 None
+    """
     scheduler = CacheAwareScheduler()
+
+    # 要求 gemini_files 能力：只有 Google 官方 API 才支持 Files API
+    capability_requirements = {"gemini_files": True} if require_files_capability else None
+
     candidates, _global_model_id = await scheduler.list_all_candidates(
         db=db,
         api_format="gemini:chat",
@@ -194,7 +223,7 @@ async def _select_provider_candidate(
         affinity_key=str(user_api_key.id),
         user_api_key=user_api_key,
         max_candidates=10,
-        capability_requirements=REQUIRED_CAPABILITIES,
+        capability_requirements=capability_requirements,
     )
     for candidate in candidates:
         auth_type = getattr(candidate.key, "auth_type", "api_key") or "api_key"
@@ -206,12 +235,20 @@ async def _select_provider_candidate(
 async def _resolve_upstream_context(
     request: Request,
     db: Session,
-) -> tuple[str, str, str]:
+) -> tuple[str, str, str, str]:
     """
-    解析上游 Key 与 Base URL
+    解析上游 Key 与 Base URL（需要外部提供 db session）
 
-    仅允许系统 API Key，通过能力标签选择支持 Files API 的 Provider Key。
+    仅允许系统 API Key，选择可用的 Gemini Provider Key（无能力限制）。
+
+    Args:
+        request: HTTP 请求
+        db: 数据库会话
+
+    Returns:
+        (upstream_key, base_url, key_id, user_id)
     """
+
     client_key = _extract_gemini_api_key(request)
     if not client_key:
         raise HTTPException(
@@ -252,14 +289,19 @@ async def _resolve_upstream_context(
             },
         )
 
-    candidate = await _select_provider_candidate(db, user_api_key, model_name)
+    # 选择可用的 provider candidate（要求 gemini_files 能力）
+    candidate = await _select_provider_candidate(
+        db, user_api_key, model_name, require_files_capability=True
+    )
+
     if not candidate:
         raise HTTPException(
             status_code=503,
             detail={
                 "error": {
                     "code": 503,
-                    "message": "No available key with gemini_files_api capability",
+                    "message": "No available Gemini key with 'gemini_files' capability. "
+                    "Please ensure at least one Provider Key has the 'gemini_files' capability enabled.",
                     "status": "UNAVAILABLE",
                 }
             },
@@ -268,7 +310,7 @@ async def _resolve_upstream_context(
     try:
         upstream_key = crypto_service.decrypt(candidate.key.api_key)
     except Exception as exc:
-        logger.error(f"Failed to decrypt provider key for Gemini Files API: {exc}")
+        logger.error("Failed to decrypt provider key for Gemini Files API: %s", exc)
         raise HTTPException(
             status_code=500,
             detail={
@@ -281,7 +323,29 @@ async def _resolve_upstream_context(
         )
 
     base_url = candidate.endpoint.base_url or GEMINI_FILES_BASE_URL
-    return upstream_key, base_url, str(candidate.key.id)
+    return upstream_key, base_url, str(candidate.key.id), str(user.id)
+
+
+async def _resolve_upstream_context_standalone(request: Request) -> UpstreamContext:
+    """
+    解析上游上下文（自管理数据库连接，适用于 HTTP 代理场景）
+
+    优化：在返回上下文后立即释放数据库连接，HTTP 请求期间不持有连接。
+
+    Args:
+        request: HTTP 请求
+
+    Returns:
+        UpstreamContext: 包含所有必要信息的上下文对象
+    """
+    with create_session() as db:
+        upstream_key, base_url, file_key_id, user_id = await _resolve_upstream_context(request, db)
+        return UpstreamContext(
+            upstream_key=upstream_key,
+            base_url=base_url,
+            file_key_id=file_key_id,
+            user_id=user_id,
+        )
 
 
 async def _proxy_request(
@@ -291,6 +355,7 @@ async def _proxy_request(
     content: bytes | None = None,
     json_body: dict[str, Any] | None = None,
     file_key_id: str | None = None,
+    user_id: str | None = None,
 ) -> Response:
     """
     代理请求到上游 Gemini API
@@ -302,6 +367,7 @@ async def _proxy_request(
         content: 原始请求体（二进制）
         json_body: JSON 请求体
         file_key_id: 上游 Provider Key ID，用于成功响应时存储 file→key 映射
+        user_id: 用户 ID，用于文件映射的权限验证
 
     Returns:
         FastAPI Response 对象
@@ -338,12 +404,28 @@ async def _proxy_request(
             try:
                 payload = response.json()
                 file_name = None
+                file_obj = None
+
                 if isinstance(payload, dict):
+                    # 单文件上传响应
                     file_name = payload.get("name")
+                    file_obj = payload
+
+                    # 嵌套格式：{"file": {...}}
                     if not file_name and isinstance(payload.get("file"), dict):
                         file_name = payload["file"].get("name")
-                    if file_name:
-                        await store_file_key_mapping(file_name, file_key_id)
+                        file_obj = payload["file"]
+
+                    if file_name and file_obj:
+                        display_name = file_obj.get("displayName") or file_obj.get("display_name")
+                        mime_type = file_obj.get("mimeType") or file_obj.get("mime_type")
+                        await store_file_key_mapping(
+                            file_name,
+                            file_key_id,
+                            user_id=user_id,
+                            display_name=display_name,
+                            mime_type=mime_type,
+                        )
                         logger.debug(
                             f"Gemini file→key 映射已存储: {file_name} → key_id={file_key_id}"
                         )
@@ -355,14 +437,26 @@ async def _proxy_request(
                         mapped_count = 0
                         for item in files_list:
                             if isinstance(item, dict) and item.get("name"):
-                                await store_file_key_mapping(item["name"], file_key_id)
+                                item_display_name = item.get("displayName") or item.get(
+                                    "display_name"
+                                )
+                                item_mime_type = item.get("mimeType") or item.get("mime_type")
+                                await store_file_key_mapping(
+                                    item["name"],
+                                    file_key_id,
+                                    user_id=user_id,
+                                    display_name=item_display_name,
+                                    mime_type=item_mime_type,
+                                )
                                 mapped_count += 1
                         if mapped_count > 0:
                             logger.debug(
-                                f"Gemini list_files 批量映射已存储: {mapped_count} 个文件 → key_id={file_key_id}"
+                                "Gemini list_files 批量映射已存储: %d 个文件 → key_id=%s",
+                                mapped_count,
+                                file_key_id,
                             )
             except (ValueError, KeyError) as e:
-                logger.debug(f"Failed to store Gemini file mapping: {e}")
+                logger.debug("Failed to store Gemini file mapping: %s", e)
 
         return Response(
             content=response.content,
@@ -373,7 +467,7 @@ async def _proxy_request(
 
     except Exception as e:
         sanitized_error = redact_url_for_log(str(e))
-        logger.error(f"Gemini Files API proxy error: {sanitized_error}")
+        logger.error("Gemini Files API proxy error: %s", sanitized_error)
         return JSONResponse(
             status_code=502,
             content={
@@ -394,7 +488,6 @@ async def _proxy_request(
 @router.post("/upload/v1beta/files")
 async def upload_file(
     request: Request,
-    db: Session = Depends(get_db),
 ) -> Any:
     """
     上传文件到 Gemini Files API
@@ -421,26 +514,34 @@ async def upload_file(
         }
     }
     ```
-    """
-    upstream_key, base_url, file_key_id = await _resolve_upstream_context(request, db)
 
-    # 读取请求体
+    优化：HTTP 代理期间不持有数据库连接
+    """
+    # 阶段 1：解析上下文（短暂持有数据库连接）
+    ctx = await _resolve_upstream_context_standalone(request)
+
+    # 阶段 2：读取请求体
     body = await request.body()
 
-    # 构建上游请求
+    # 阶段 3：代理请求（不持有数据库连接）
     upstream_url = _build_upstream_url(
-        base_url,
+        ctx.base_url,
         "/v1beta/files",
         dict(request.query_params),
         is_upload=True,
     )
 
-    headers = _build_upstream_headers(dict(request.headers), upstream_key)
+    headers = _build_upstream_headers(dict(request.headers), ctx.upstream_key)
 
-    logger.debug(f"Gemini Files upload proxy: POST {redact_url_for_log(upstream_url)}")
+    logger.debug("Gemini Files upload proxy: POST %s", redact_url_for_log(upstream_url))
 
     return await _proxy_request(
-        "POST", upstream_url, headers, content=body, file_key_id=file_key_id
+        "POST",
+        upstream_url,
+        headers,
+        content=body,
+        file_key_id=ctx.file_key_id,
+        user_id=ctx.user_id,
     )
 
 
@@ -452,13 +553,14 @@ async def upload_file(
 @router.get("/v1beta/files")
 async def list_files(
     request: Request,
-    db: Session = Depends(get_db),
     pageSize: int | None = None,
     pageToken: str | None = None,
 ) -> Any:
     """
     列出已上传的文件
 
+    优化：HTTP 代理期间不持有数据库连接
+
     **认证方式**:
     - `x-goog-api-key` 请求头，或
     - `?key=` URL 参数
@@ -488,22 +590,222 @@ async def list_files(
     }
     ```
     """
-    upstream_key, base_url, file_key_id = await _resolve_upstream_context(request, db)
+    # 阶段 1：解析上下文（短暂持有数据库连接）
+    ctx = await _resolve_upstream_context_standalone(request)
 
-    # 构建查询参数
+    # 阶段 2：代理请求（不持有数据库连接）
     query_params = dict(request.query_params)
     if pageSize is not None:
         query_params["pageSize"] = pageSize
     if pageToken is not None:
         query_params["pageToken"] = pageToken
 
-    # 构建上游请求
-    upstream_url = _build_upstream_url(base_url, "/v1beta/files", query_params)
+    upstream_url = _build_upstream_url(ctx.base_url, "/v1beta/files", query_params)
+    headers = _build_upstream_headers(dict(request.headers), ctx.upstream_key)
+
+    logger.debug("Gemini Files list proxy: GET %s", redact_url_for_log(upstream_url))
+
+    return await _proxy_request(
+        "GET", upstream_url, headers, file_key_id=ctx.file_key_id, user_id=ctx.user_id
+    )
+
+
+# ==============================================================================
+# 下载文件内容端点（用于视频等媒体文件）
+# 注意：必须在 /v1beta/files/{file_name:path} 之前注册，否则会被通配符路由捕获
+# ==============================================================================
+
+
+async def _find_video_task_by_id(
+    db: Session, short_id: str, user_id: str
+) -> tuple[str | None, str | None]:
+    """
+    通过短 ID 查找视频任务，返回其 provider key 和 video_url
+
+    Args:
+        db: 数据库会话
+        short_id: 视频任务的短 ID（VideoTask.short_id，Gemini 风格）
+        user_id: 用户 ID（用于权限验证）
+
+    Returns:
+        (upstream_key, video_url) - 如果找到任务返回 key 和 url，否则返回 (None, None)
+    """
+    from src.models.database import ProviderAPIKey, VideoTask
+
+    logger.debug(
+        "[Files Download] Searching video task: short_id=%s, user_id=%s", short_id, user_id
+    )
+
+    # 通过 short_id 查找，同时验证用户权限
+    task = (
+        db.query(VideoTask)
+        .filter(VideoTask.short_id == short_id, VideoTask.user_id == user_id)
+        .first()
+    )
+
+    if not task:
+        logger.debug("[Files Download] No video task found: short_id=%s", short_id)
+        return None, None
+
+    if not task.video_url:
+        logger.debug("[Files Download] Task found but no video_url: short_id=%s", short_id)
+        return None, None
+
+    if not task.key_id:
+        logger.debug("[Files Download] Task found but no key_id: short_id=%s", short_id)
+        return None, task.video_url
+
+    # 获取 provider key
+    provider_key = db.query(ProviderAPIKey).filter(ProviderAPIKey.id == task.key_id).first()
+    if not provider_key or not provider_key.api_key:
+        logger.debug("[Files Download] Provider key not found: key_id=%s", task.key_id)
+        return None, task.video_url
+
+    try:
+        upstream_key = crypto_service.decrypt(provider_key.api_key)
+        logger.debug("[Files Download] Found key for task: short_id=%s", short_id)
+        return upstream_key, task.video_url
+    except Exception as e:
+        logger.error("[Files Download] Failed to decrypt key: %s", e)
+        return None, task.video_url
+
+
+@router.get("/v1beta/files/{file_id}:download")
+async def download_file(
+    file_id: str,
+    request: Request,
+) -> Any:
+    """
+    下载文件（官方 Gemini API 格式）
+
+    **认证方式**:
+    - `x-goog-api-key` 请求头，或
+    - `?key=` URL 参数
+
+    **路径参数**:
+    - `file_id`: 文件 ID
+      - 以 `aev_` 开头：视频任务下载（如 `aev_sknuzqlo8sds`，Gemini 风格短 ID）
+      - 其他：普通 Gemini 文件下载（透传到上游）
+
+    **查询参数**:
+    - `alt=media`: 可选，保持与官方 API 兼容
+
+    **示例**:
+    ```
+    GET /v1beta/files/aev_{short_id}:download?alt=media  # 视频任务
+    GET /v1beta/files/{gemini_file_id}:download?alt=media  # 普通文件
+    ```
+
+    优化：HTTP 下载期间不持有数据库连接
+    """
+    import httpx
+    from fastapi import HTTPException
+    from fastapi.responses import JSONResponse, Response
+
+    # ========== 阶段 1：数据库操作（短暂持有连接）==========
+    client_key = _extract_gemini_api_key(request)
+    if not client_key:
+        raise HTTPException(
+            status_code=401,
+            detail={
+                "error": {"code": 401, "message": "API key required", "status": "UNAUTHENTICATED"}
+            },
+        )
+
+    # 在数据库会话内完成所有查询
+    with create_session() as db:
+        auth_result = AuthService.authenticate_api_key(db, client_key)
+        if not auth_result:
+            raise HTTPException(
+                status_code=401,
+                detail={
+                    "error": {
+                        "code": 401,
+                        "message": "API key not valid",
+                        "status": "UNAUTHENTICATED",
+                    }
+                },
+            )
+
+        user, _user_api_key = auth_result
+
+        # 根据前缀判断处理方式
+        if file_id.startswith("aev_"):
+            # 视频任务下载：使用短 ID 查找
+            short_id = file_id[4:]  # 去掉 "aev_" 前缀
+            logger.debug("[Files Download] Video task: short_id=%s, user_id=%s", short_id, user.id)
+            upstream_key, video_url = await _find_video_task_by_id(db, short_id, user.id)
+            if not upstream_key or not video_url:
+                raise HTTPException(
+                    status_code=404,
+                    detail={
+                        "error": {
+                            "code": 404,
+                            "message": f"Video not found or not ready: {file_id}",
+                            "status": "NOT_FOUND",
+                        }
+                    },
+                )
+            upstream_url = video_url
+        else:
+            # 普通文件下载：透传到 Gemini
+            try:
+                upstream_key, base_url, _file_key_id, _user_id = await _resolve_upstream_context(
+                    request, db
+                )
+            except HTTPException:
+                raise HTTPException(
+                    status_code=404,
+                    detail={
+                        "error": {
+                            "code": 404,
+                            "message": f"File not found: {file_id}",
+                            "status": "NOT_FOUND",
+                        }
+                    },
+                )
+            file_name = f"files/{file_id}" if not file_id.startswith("files/") else file_id
+            upstream_url = _build_upstream_url(
+                base_url,
+                f"/v1beta/{file_name}:download",
+                dict(request.query_params),
+            )
+
+    # ========== 阶段 2：HTTP 下载（不持有数据库连接）==========
     headers = _build_upstream_headers(dict(request.headers), upstream_key)
 
-    logger.debug(f"Gemini Files list proxy: GET {redact_url_for_log(upstream_url)}")
+    logger.debug("Gemini Files download proxy: GET %s", redact_url_for_log(upstream_url))
 
-    return await _proxy_request("GET", upstream_url, headers, file_key_id=file_key_id)
+    # 使用 follow_redirects=True 跟随重定向（Gemini 文件下载会重定向）
+    try:
+        async with httpx.AsyncClient(follow_redirects=True, timeout=httpx.Timeout(300.0)) as client:
+            response = await client.get(upstream_url, headers=headers)
+    except Exception as exc:
+        logger.error("Gemini Files download failed: %s", exc)
+        raise HTTPException(status_code=502, detail="Failed to download file")
+
+    if response.status_code >= 400:
+        content: dict[str, Any]
+        if response.headers.get("content-type", "").startswith("application/json"):
+            try:
+                content = response.json()
+            except Exception:
+                content = {"error": response.text}
+        else:
+            content = {"error": response.text}
+        return JSONResponse(content=content, status_code=response.status_code)
+
+    # 返回文件内容
+    return Response(
+        content=response.content,
+        status_code=response.status_code,
+        headers={
+            k: v
+            for k, v in response.headers.items()
+            if k.lower() not in {"transfer-encoding", "connection", "keep-alive"}
+        },
+        media_type=response.headers.get("content-type", "application/octet-stream"),
+    )
 
 
 # ==============================================================================
@@ -515,7 +817,6 @@ async def list_files(
 async def get_file(
     file_name: str,
     request: Request,
-    db: Session = Depends(get_db),
 ) -> Any:
     """
     获取指定文件的元数据
@@ -542,24 +843,29 @@ async def get_file(
         "state": "ACTIVE"
     }
     ```
-    """
-    upstream_key, base_url, file_key_id = await _resolve_upstream_context(request, db)
 
+    优化：HTTP 代理期间不持有数据库连接
+    """
+    # 阶段 1：解析上下文（短暂持有数据库连接）
+    ctx = await _resolve_upstream_context_standalone(request)
+
+    # 阶段 2：代理请求（不持有数据库连接）
     # 规范化文件名（确保以 files/ 开头）
     if not file_name.startswith("files/"):
         file_name = f"files/{file_name}"
 
-    # 构建上游请求
     upstream_url = _build_upstream_url(
-        base_url,
+        ctx.base_url,
         f"/v1beta/{file_name}",
         dict(request.query_params),
     )
-    headers = _build_upstream_headers(dict(request.headers), upstream_key)
+    headers = _build_upstream_headers(dict(request.headers), ctx.upstream_key)
 
-    logger.debug(f"Gemini Files get proxy: GET {redact_url_for_log(upstream_url)}")
+    logger.debug("Gemini Files get proxy: GET %s", redact_url_for_log(upstream_url))
 
-    return await _proxy_request("GET", upstream_url, headers, file_key_id=file_key_id)
+    return await _proxy_request(
+        "GET", upstream_url, headers, file_key_id=ctx.file_key_id, user_id=ctx.user_id
+    )
 
 
 # ==============================================================================
@@ -571,7 +877,6 @@ async def get_file(
 async def delete_file(
     file_name: str,
     request: Request,
-    db: Session = Depends(get_db),
 ) -> Any:
     """
     删除指定文件
@@ -585,29 +890,31 @@ async def delete_file(
 
     **响应格式**:
     成功时返回空 JSON 对象：`{}`
-    """
-    upstream_key, base_url, _file_key_id = await _resolve_upstream_context(request, db)
 
+    优化：HTTP 代理期间不持有数据库连接
+    """
+    # 阶段 1：解析上下文（短暂持有数据库连接）
+    ctx = await _resolve_upstream_context_standalone(request)
+
+    # 阶段 2：代理请求（不持有数据库连接）
     # 规范化文件名（确保以 files/ 开头）
     if not file_name.startswith("files/"):
         file_name = f"files/{file_name}"
 
-    # 构建上游请求
     upstream_url = _build_upstream_url(
-        base_url,
+        ctx.base_url,
         f"/v1beta/{file_name}",
         dict(request.query_params),
     )
-    headers = _build_upstream_headers(dict(request.headers), upstream_key)
+    headers = _build_upstream_headers(dict(request.headers), ctx.upstream_key)
 
-    logger.debug(f"Gemini Files delete proxy: DELETE {redact_url_for_log(upstream_url)}")
+    logger.debug("Gemini Files delete proxy: DELETE %s", redact_url_for_log(upstream_url))
 
-    del _file_key_id  # 显式标记：delete 端点不需要存储映射
     response = await _proxy_request("DELETE", upstream_url, headers)
     if response.status_code < 300:
         await delete_file_key_mapping(file_name)
     else:
         logger.debug(
-            f"Gemini Files delete failed, skip mapping cleanup: status={response.status_code}"
+            "Gemini Files delete failed, skip mapping cleanup: status=%s", response.status_code
         )
     return response

src/api/public/videos.py

@@ -61,9 +61,10 @@ async def list_video_tasks_sora(http_request: Request, db: Session = Depends(get
 
 
 @router.delete("/v1/videos/{task_id}")
-async def cancel_video_task_sora(
+async def delete_video_task_sora(
     task_id: str, http_request: Request, db: Session = Depends(get_db)
 ) -> Any:
+    """删除已完成或失败的视频及其存储资源"""
     adapter = OpenAIVideoAdapter()
     return await pipeline.run(
         adapter=adapter,
@@ -71,7 +72,7 @@ async def cancel_video_task_sora(
         db=db,
         mode=adapter.mode,
         api_format_hint=adapter.allowed_api_formats[0],
-        path_params={"task_id": task_id, "action": "cancel"},
+        path_params={"task_id": task_id},
     )
 
 
@@ -121,6 +122,47 @@ async def create_video_veo(model: str, http_request: Request, db: Session = Depe
     )
 
 
+# Gemini Veo operation routes - support both formats:
+# 1. models/{model}/operations/{id} (official Gemini Veo format)
+# 2. operations/{...} (legacy format for compatibility)
+
+
+@router.get("/v1beta/models/{model}/operations/{operation_id}")
+async def get_video_veo_by_model(
+    model: str, operation_id: str, http_request: Request, db: Session = Depends(get_db)
+) -> Any:
+    """Get video task status (Gemini Veo format: models/{model}/operations/{id})"""
+    adapter = GeminiVeoAdapter()
+    # Reconstruct full operation name
+    full_operation_name = f"models/{model}/operations/{operation_id}"
+    return await pipeline.run(
+        adapter=adapter,
+        http_request=http_request,
+        db=db,
+        mode=adapter.mode,
+        api_format_hint=adapter.allowed_api_formats[0],
+        path_params={"task_id": full_operation_name},
+    )
+
+
+@router.post("/v1beta/models/{model}/operations/{operation_id}:cancel")
+async def cancel_video_veo_by_model(
+    model: str, operation_id: str, http_request: Request, db: Session = Depends(get_db)
+) -> Any:
+    """Cancel video task (Gemini Veo format: models/{model}/operations/{id}:cancel)"""
+    adapter = GeminiVeoAdapter()
+    full_operation_name = f"models/{model}/operations/{operation_id}"
+    return await pipeline.run(
+        adapter=adapter,
+        http_request=http_request,
+        db=db,
+        mode=adapter.mode,
+        api_format_hint=adapter.allowed_api_formats[0],
+        path_params={"task_id": full_operation_name, "action": "cancel"},
+    )
+
+
+# Legacy routes for backward compatibility
 @router.get("/v1beta/operations/{operation_id:path}")
 async def get_video_veo(
     operation_id: str, http_request: Request, db: Session = Depends(get_db)
@@ -163,16 +205,4 @@ async def cancel_video_veo(
     )
 
 
-@router.get("/v1beta/operations/{operation_id:path}/content")
-async def download_video_content_veo(
-    operation_id: str, http_request: Request, db: Session = Depends(get_db)
-) -> Any:
-    adapter = GeminiVeoAdapter()
-    return await pipeline.run(
-        adapter=adapter,
-        http_request=http_request,
-        db=db,
-        mode=adapter.mode,
-        api_format_hint=adapter.allowed_api_formats[0],
-        path_params={"task_id": operation_id},
-    )
+# Video download is now handled by /v1beta/files/{task_id}:download in gemini_files.py

src/api/user_me/routes.py

@@ -736,6 +736,7 @@ class GetUsageAdapter(AuthenticatedApiAdapter):
 
     async def handle(self, context: ApiRequestContext) -> Any:  # type: ignore[override]
         from sqlalchemy import or_
+        from sqlalchemy.orm import load_only
 
         from src.models.database import ProviderEndpoint
         from src.utils.database_helpers import escape_like_pattern, safe_truncate_escaped
@@ -877,7 +878,44 @@ class GetUsageAdapter(AuthenticatedApiAdapter):
                 )
 
         # 计算总数用于分页
-        total_records = query.count()
+        # Perf: avoid Query.count() building a subquery selecting many columns
+        total_records = int(query.with_entities(func.count(Usage.id)).scalar() or 0)
+
+        # Perf: do not load large request/response columns for list view
+        query = query.options(
+            load_only(
+                Usage.id,
+                Usage.user_id,
+                Usage.api_key_id,
+                Usage.provider_name,
+                Usage.model,
+                Usage.target_model,
+                Usage.input_tokens,
+                Usage.output_tokens,
+                Usage.total_tokens,
+                Usage.total_cost_usd,
+                Usage.response_time_ms,
+                Usage.first_byte_time_ms,
+                Usage.is_stream,
+                Usage.status,
+                Usage.created_at,
+                Usage.cache_creation_input_tokens,
+                Usage.cache_read_input_tokens,
+                Usage.status_code,
+                Usage.error_message,
+                Usage.api_format,
+                Usage.endpoint_api_format,
+                Usage.has_format_conversion,
+                Usage.input_price_per_1m,
+                Usage.output_price_per_1m,
+                Usage.cache_creation_price_per_1m,
+                Usage.cache_read_price_per_1m,
+                Usage.actual_total_cost_usd,
+                Usage.rate_multiplier,
+            ),
+            load_only(ApiKey.id, ApiKey.name, ApiKey.key_encrypted),
+            load_only(ProviderEndpoint.id, ProviderEndpoint.api_format),
+        )
         usage_records = (
             query.order_by(Usage.created_at.desc()).offset(self.offset).limit(self.limit).all()
         )
@@ -1231,7 +1269,7 @@ class ListAvailableModelsAdapter(AuthenticatedApiAdapter):
                     endpoint_format,
                     format_acceptance_config,
                     is_stream=False,
-                    global_conversion_enabled=global_conversion_enabled,
+                    effective_conversion_enabled=global_conversion_enabled,
                 )
                 if is_compatible:
                     provider_to_formats.setdefault(provider_id, set()).add(endpoint_format)

src/core/api_format/conversion/compatibility.py

@@ -3,11 +3,20 @@
 
 用于候选筛选时判断端点是否可以处理客户端请求格式。
 
+三层开关优先级（从高到低）：
+1. 全局开关 ON → 强制允许（跳过后续检查）
+2. 全局开关 OFF → 看提供商开关
+   - 提供商开关 ON → 强制允许（跳过端点检查）
+   - 提供商开关 OFF → 看端点配置
+3. 端点配置（format_acceptance_config）
+   - enabled=true + 白名单/黑名单检查 → 允许
+   - enabled=false 或未配置 → 禁止
+
 转换逻辑：
 1. 格式完全匹配 -> 透传（无需转换）
-2. 格式不同 -> 需要检查全局开关 + 端点开关
-   - data_format_id 相同 -> 可透传（无需数据转换），但需全局开关 + 端点开关
-   - data_format_id 不同 -> 需要转换，检查全局开关 + 端点配置 + 转换器能力
+2. 格式不同 -> 需要检查三层开关
+   - data_format_id 相同 -> 可透传（无需数据转换）
+   - data_format_id 不同 -> 需要转换，检查转换器能力
 """
 
 from __future__ import annotations
@@ -28,8 +37,10 @@ def is_format_compatible(
     endpoint_api_format: str,
     endpoint_format_acceptance_config: dict | None,
     is_stream: bool,
-    global_conversion_enabled: bool,
+    effective_conversion_enabled: bool,
     registry: FormatConversionRegistry | None = None,
+    *,
+    skip_endpoint_check: bool = False,
 ) -> tuple[bool, bool, str | None]:
     """
     检查端点是否兼容客户端格式
@@ -39,8 +50,9 @@ def is_format_compatible(
         endpoint_api_format: 端点的 API 格式
         endpoint_format_acceptance_config: 端点的格式接受配置
         is_stream: 是否是流式请求
-        global_conversion_enabled: 全局格式转换开关（来自环境变量 FORMAT_CONVERSION_ENABLED，默认 True）
+        effective_conversion_enabled: 有效格式转换开关（全局 OR 提供商）
         registry: 转换器注册表（可选，默认使用全局单例）
+        skip_endpoint_check: 是否跳过端点配置检查（当全局或提供商开关为 ON 时设为 True）
 
     Returns:
         (is_compatible, needs_conversion, skip_reason)
@@ -66,30 +78,36 @@ def is_format_compatible(
     if provider_key == client_key:
         return True, False, None
 
-    # 2. 格式不同 -> 需要检查全局格式转换开关
-    # 即使 data_format_id 相同（如 claude:chat / claude:cli），也需要全局开关启用
-    if not global_conversion_enabled:
-        return False, False, "全局格式转换未启用（环境变量 FORMAT_CONVERSION_ENABLED=false）"
+    # 2. 格式不同 -> 需要检查格式转换开关
+    # 如果有效开关为 False（全局 OFF 且提供商 OFF），直接拒绝
+    if not effective_conversion_enabled:
+        return False, False, "格式转换已禁用（全局和提供商开关均为关闭）"
 
-    # 3. 格式不同时，统一检查端点配置（核心控制）
-    if endpoint_format_acceptance_config is None:
-        return False, False, "端点未配置格式接受策略"
+    # 3. 如果全局或提供商开关为 ON，跳过端点配置检查
+    if not skip_endpoint_check:
+        # 检查端点配置（第三层开关）
+        if endpoint_format_acceptance_config is None:
+            return False, False, "端点未配置格式接受策略"
 
-    config = endpoint_format_acceptance_config
-    if not isinstance(config, dict):
-        return False, False, "端点格式配置无效"
-    if not config.get("enabled", False):
-        return False, False, "端点格式接受未启用"
+        config = endpoint_format_acceptance_config
+        if not isinstance(config, dict):
+            return False, False, "端点格式配置无效"
+        if not config.get("enabled", False):
+            return False, False, "端点格式接受未启用"
 
-    # 检查 reject_formats（优先）
-    reject_formats = config.get("reject_formats", [])
-    if client_key in [f.upper() for f in reject_formats]:
-        return False, False, f"端点拒绝 {client_format} 格式"
+        # 检查 reject_formats（优先）
+        reject_formats = config.get("reject_formats", [])
+        if client_key in [f.upper() for f in reject_formats]:
+            return False, False, f"端点拒绝 {client_format} 格式"
 
-    # 检查 accept_formats
-    accept_formats = config.get("accept_formats", [])
-    if accept_formats and client_key not in [f.upper() for f in accept_formats]:
-        return False, False, f"端点不接受 {client_format} 格式"
+        # 检查 accept_formats
+        accept_formats = config.get("accept_formats", [])
+        if accept_formats and client_key not in [f.upper() for f in accept_formats]:
+            return False, False, f"端点不接受 {client_format} 格式"
+
+        # 检查流式转换
+        if is_stream and not config.get("stream_conversion", True):
+            return False, False, "端点不支持流式格式转换"
 
     # 4. 检查是否可以透传（data_format_id 相同）
     # 例如：claude:chat / claude:cli 的 data_format_id 都是 "claude"，数据格式相同可透传
@@ -99,11 +117,7 @@ def is_format_compatible(
         return True, False, None
 
     # 5. 需要数据转换的情况（data_format_id 不同）
-    # 检查流式转换
-    if is_stream and not config.get("stream_conversion", True):
-        return False, False, "端点不支持流式格式转换"
-
-    # 6. 检查转换器能力
+    # 检查转换器能力
     if not registry.can_convert_full(
         client_key,
         provider_key,

src/core/api_format/conversion/normalizers/gemini.py

@@ -736,10 +736,17 @@ class GeminiNormalizer(FormatNormalizer):
         instance = instances[0] if isinstance(instances[0], dict) else {}
         params = request.get("parameters") or {}
 
+        # 解析 image（用于 image-to-video 或第一帧）
+        # 官方格式: {"image": {"inlineData": {"mimeType": "image/png", "data": "base64..."}}}
         image = instance.get("image") if isinstance(instance, dict) else None
         image_ref = None
         if isinstance(image, dict):
-            image_ref = image.get("bytesBase64Encoded")
+            inline_data = image.get("inlineData", {})
+            if isinstance(inline_data, dict):
+                image_ref = inline_data.get("data")
+            # 兼容旧格式
+            if not image_ref:
+                image_ref = image.get("bytesBase64Encoded")
 
         prompt = instance.get("prompt") if isinstance(instance, dict) else None
         prompt_str = str(prompt).strip() if prompt else ""
@@ -747,7 +754,7 @@ class GeminiNormalizer(FormatNormalizer):
             raise ValueError("Video prompt is required")
 
         duration_raw = params.get("durationSeconds")
-        sample_count_raw = params.get("sampleCount")
+        sample_count_raw = params.get("sampleCount") or params.get("numberOfVideos")
 
         try:
             duration_seconds = int(duration_raw) if duration_raw else 8
@@ -760,6 +767,35 @@ class GeminiNormalizer(FormatNormalizer):
         except (ValueError, TypeError):
             sample_count = 1
 
+        # 构建 extra 字段，保留所有 Veo 特有的参数
+        extra: dict[str, Any] = {
+            "personGeneration": params.get("personGeneration"),
+            "sampleCount": sample_count,
+        }
+
+        # negativePrompt - 负面提示词
+        if params.get("negativePrompt"):
+            extra["negativePrompt"] = params["negativePrompt"]
+
+        # lastFrame - 最后一帧（用于插值）
+        last_frame = params.get("lastFrame")
+        if isinstance(last_frame, dict):
+            extra["lastFrame"] = last_frame
+
+        # referenceImages - 参考图像（最多3张，仅 Veo 3.1）
+        ref_images = params.get("referenceImages")
+        if isinstance(ref_images, list) and ref_images:
+            extra["referenceImages"] = ref_images
+
+        # video - 视频扩展输入（用于视频续写）
+        video_input = instance.get("video") if isinstance(instance, dict) else None
+        if isinstance(video_input, dict):
+            extra["video"] = video_input
+
+        # seed - 种子值（Veo 3）
+        if params.get("seed") is not None:
+            extra["seed"] = params["seed"]
+
         return InternalVideoRequest(
             prompt=prompt_str,
             model=str(request.get("model") or "veo-3.1-generate-preview"),
@@ -767,10 +803,7 @@ class GeminiNormalizer(FormatNormalizer):
             aspect_ratio=str(params.get("aspectRatio") or "16:9"),
             resolution=str(params.get("resolution") or "720p"),
             reference_image_url=image_ref,
-            extra={
-                "personGeneration": params.get("personGeneration"),
-                "sampleCount": sample_count,
-            },
+            extra=extra,
         )
 
     def video_request_from_internal(self, internal: InternalVideoRequest) -> dict[str, Any]:
@@ -783,15 +816,31 @@ class GeminiNormalizer(FormatNormalizer):
             "resolution": internal.resolution,
             "durationSeconds": internal.duration_seconds,
         }
-        for key in ["personGeneration", "sampleCount"]:
+        for key in ["personGeneration", "sampleCount", "negativePrompt", "seed"]:
             if key in internal.extra:
                 parameters[key] = internal.extra[key]
 
-        return {
+        # lastFrame 和 referenceImages 需要特殊处理
+        if internal.extra.get("lastFrame"):
+            parameters["lastFrame"] = internal.extra["lastFrame"]
+        if internal.extra.get("referenceImages"):
+            parameters["referenceImages"] = internal.extra["referenceImages"]
+
+        # video 输入（视频续写）
+        if internal.extra.get("video"):
+            instance["video"] = internal.extra["video"]
+
+        result: dict[str, Any] = {
             "instances": [instance],
             "parameters": parameters,
         }
 
+        # 模型信息（用于 URL 构建）
+        if internal.model:
+            result["model"] = internal.model
+
+        return result
+
     def video_task_to_internal(self, response: dict[str, Any]) -> InternalVideoTask:
         operation_name = str(response.get("name") or "")
         done = bool(response.get("done"))
@@ -824,20 +873,30 @@ class GeminiNormalizer(FormatNormalizer):
         )
 
     def video_task_from_internal(self, internal: InternalVideoTask) -> dict[str, Any]:
-        # 优先使用 external_id（上游返回的 operation name），否则用内部 id
-        operation_name = internal.external_id or f"operations/{internal.id}"
-        if not operation_name.startswith("operations/"):
-            operation_name = f"operations/{operation_name}"
+        # 从 external_id 中提取 model 名称，用于构建 operation name
+        # external_id 格式: models/{model}/operations/{gemini_id}
+        model_name = "unknown"
+        if internal.external_id:
+            parts = internal.external_id.split("/")
+            if len(parts) >= 2 and parts[0] == "models":
+                model_name = parts[1]
+
+        # 使用我们的内部 task_id 构建 operation name，不暴露 Gemini 的 operation ID
+        # 格式: models/{model}/operations/{our_task_id}
+        operation_name = f"models/{model_name}/operations/{internal.id}"
 
         if internal.status == VideoStatus.COMPLETED:
-            urls = internal.video_urls or ([internal.video_url] if internal.video_url else [])
+            # 使用我们的内部 task_id 构建下载 URL，不暴露真实的 Gemini file_id
+            # 使用 aev_ 前缀标识这是视频任务的下载链接
+            # 格式：/v1beta/files/aev_{task_id}:download?alt=media
+            proxy_download_url = f"/v1beta/files/aev_{internal.id}:download?alt=media"
             return {
                 "name": operation_name,
                 "done": True,
                 "response": {
                     "generateVideoResponse": {
                         "generatedSamples": [
-                            {"video": {"uri": url, "mimeType": "video/mp4"}} for url in urls
+                            {"video": {"uri": proxy_download_url, "mimeType": "video/mp4"}}
                         ]
                     }
                 },

src/core/api_format/conversion/normalizers/openai.py

@@ -752,10 +752,21 @@ class OpenAINormalizer(FormatNormalizer):
                 "message": internal.error_message,
             }
 
-        for key in ["model", "size", "seconds"]:
-            if key in internal.extra:
+        # 基本字段
+        for key in ["model", "size", "prompt"]:
+            if internal.extra.get(key):
                 payload[key] = internal.extra[key]
 
+        # seconds 必须是字符串类型
+        seconds = internal.extra.get("seconds")
+        if seconds is not None:
+            payload["seconds"] = str(seconds)
+
+        # remix 相关字段
+        remixed_from = internal.extra.get("remixed_from_video_id")
+        if remixed_from:
+            payload["remixed_from_video_id"] = remixed_from
+
         return payload
 
     def video_poll_to_internal(self, response: dict[str, Any]) -> InternalVideoPollResult:
@@ -764,14 +775,18 @@ class OpenAINormalizer(FormatNormalizer):
 
         if status == "completed":
             expires_at = response.get("expires_at")
-            # 使用任务 ID 构建内容路径，由调用方拼接完整 URL
-            # 如果 task_id 不存在，说明上游响应异常
-            video_url = f"videos/{task_id}/content" if task_id else None
+            # 优先使用上游返回的直接 URL（某些代理如 API易 会返回 CDN URL）
+            # 回退到构建相对路径（标准 OpenAI API 通过 /content 端点下载）
+            direct_url = (
+                response.get("video_url") or response.get("url") or response.get("result_url")
+            )
+            # 使用直接 URL 或回退到相对路径
+            video_url = direct_url or (f"videos/{task_id}/content" if task_id else None)
             if not video_url:
                 return InternalVideoPollResult(
                     status=VideoStatus.FAILED,
-                    error_code="missing_task_id",
-                    error_message="Upstream response missing task id",
+                    error_code="missing_video_url",
+                    error_message="Upstream response missing video url",
                     raw_response=response,
                 )
             return InternalVideoPollResult(

src/core/api_format/conversion/registry.py

@@ -155,6 +155,110 @@ class FormatConversionRegistry:
             except Exception as e:
                 raise FormatConversionError(source_format, target_format, str(e)) from e
 
+    # ==================== 视频格式转换 ====================
+
+    def convert_video_request(
+        self,
+        request: dict[str, Any],
+        source_format: str,
+        target_format: str,
+    ) -> dict[str, Any]:
+        """转换视频请求格式（OpenAI <-> Gemini）
+
+        Args:
+            request: 原始视频请求
+            source_format: 源格式（如 openai:video, gemini:video）
+            target_format: 目标格式
+
+        Returns:
+            转换后的视频请求
+        """
+        # 统一使用基础格式 ID（去掉 :video 后缀）
+        src_base = self._video_format_to_base(source_format)
+        tgt_base = self._video_format_to_base(target_format)
+
+        if src_base == tgt_base:
+            return request
+
+        src = self._require_normalizer(src_base)
+        tgt = self._require_normalizer(tgt_base)
+
+        with _track_conversion_metrics(
+            "video_request", str(source_format).upper(), str(target_format).upper()
+        ):
+            try:
+                internal = src.video_request_to_internal(request)
+                return tgt.video_request_from_internal(internal)
+            except Exception as e:
+                raise FormatConversionError(source_format, target_format, str(e)) from e
+
+    def convert_video_task(
+        self,
+        task_response: dict[str, Any],
+        source_format: str,
+        target_format: str,
+    ) -> dict[str, Any]:
+        """转换视频任务响应格式（OpenAI <-> Gemini）
+
+        Args:
+            task_response: 原始任务响应
+            source_format: 源格式
+            target_format: 目标格式
+
+        Returns:
+            转换后的任务响应
+        """
+        src_base = self._video_format_to_base(source_format)
+        tgt_base = self._video_format_to_base(target_format)
+
+        if src_base == tgt_base:
+            return task_response
+
+        src = self._require_normalizer(src_base)
+        tgt = self._require_normalizer(tgt_base)
+
+        with _track_conversion_metrics(
+            "video_task", str(source_format).upper(), str(target_format).upper()
+        ):
+            try:
+                internal = src.video_task_to_internal(task_response)
+                return tgt.video_task_from_internal(internal)
+            except Exception as e:
+                raise FormatConversionError(source_format, target_format, str(e)) from e
+
+    def can_convert_video(self, source_format: str, target_format: str) -> bool:
+        """检查是否支持视频格式转换"""
+        src_base = self._video_format_to_base(source_format)
+        tgt_base = self._video_format_to_base(target_format)
+
+        if src_base == tgt_base:
+            return True
+
+        src = self.get_normalizer(src_base)
+        tgt = self.get_normalizer(tgt_base)
+
+        if src is None or tgt is None:
+            return False
+
+        # 检查是否有视频转换方法
+        return (
+            hasattr(src, "video_request_to_internal")
+            and hasattr(src, "video_task_to_internal")
+            and hasattr(tgt, "video_request_from_internal")
+            and hasattr(tgt, "video_task_from_internal")
+        )
+
+    def _video_format_to_base(self, format_id: str) -> str:
+        """将视频格式 ID 转换为基础格式 ID
+
+        例如: openai:video -> openai:chat, gemini:video -> gemini:chat
+        """
+        upper = str(format_id).upper()
+        if upper.endswith(":VIDEO"):
+            base = upper[:-6]  # 去掉 :VIDEO
+            return f"{base}:CHAT"
+        return upper
+
     # ==================== 流式转换（严格） ====================
 
     def convert_stream_chunk(

src/core/key_capabilities.py

@@ -248,10 +248,10 @@ register_capability(
 )
 
 register_capability(
-    name="gemini_files_api",
-    display_name="Gemini文件上传",
-    description="支持 Gemini Files API（上传、查询、删除），第三方 Key 通常不支持",
-    match_mode=CapabilityMatchMode.COMPATIBLE,  # 需要时选有的，不需要时都可选
-    config_mode=CapabilityConfigMode.REQUEST_PARAM,  # 从请求路径检测
-    short_name="文件上传",
+    name="gemini_files",
+    display_name="Gemini 文件 API",
+    description="支持 Gemini Files API（文件上传/管理），仅 Google 官方 API 支持",
+    match_mode=CapabilityMatchMode.COMPATIBLE,
+    config_mode=CapabilityConfigMode.USER_CONFIGURABLE,
+    short_name="文件API",
 )

src/core/modules/registry.py

@@ -236,15 +236,23 @@ class ModuleRegistry:
         # 获取启用状态
         enabled = self.is_enabled(name, db) if available else False
 
-        # 注意：配置验证失败时不自动禁用模块
-        # 自动禁用会在查询方法中产生写操作副作用，违反幂等性原则
-        # 配置验证状态通过 config_validated/config_error 字段返回，由调用方决定如何处理
+        # 配置验证失败时自动禁用模块
+        # 注意：此处故意在 get_status() 中写入，以确保模块状态与配置同步
+        # 场景：用户删除了模块所依赖的 Provider Key 后，模块应自动关闭
+        # 权衡：查询方法中的写操作副作用 vs 状态一致性保证
+        if enabled and not config_validated:
+            self.set_enabled(name, False, db)
+            enabled = False
+
+        # 计算激活状态：available && enabled && config_validated && 依赖模块都激活
+        is_active = self.is_active(name, db) if available else False
+        active = is_active and config_validated
 
         return ModuleStatus(
             name=name,
             available=available,
             enabled=enabled,
-            active=self.is_active(name, db) if available else False,
+            active=active,
             config_validated=config_validated,
             config_error=config_error,
             display_name=meta.display_name,

src/database/__init__.py

@@ -3,7 +3,7 @@
 """
 
 from ..models.database import ApiKey, Base, Usage, User, UserQuota
-from .database import create_session, get_db, get_db_url, init_db, log_pool_status
+from .database import create_session, get_db, get_db_context, get_db_url, init_db, log_pool_status
 
 __all__ = [
     "Base",
@@ -12,6 +12,7 @@ __all__ = [
     "Usage",
     "UserQuota",
     "get_db",
+    "get_db_context",
     "init_db",
     "create_session",
     "get_db_url",

src/database/database.py

@@ -4,6 +4,7 @@
 
 import time
 from collections.abc import Generator
+from contextlib import contextmanager
 from typing import Any, cast
 
 from sqlalchemy import create_engine, event
@@ -271,6 +272,31 @@ def create_session() -> Session:
     return _SessionLocal()
 
 
+@contextmanager
+def get_db_context() -> Generator[Session, None, None]:
+    """
+    获取数据库会话的上下文管理器
+
+    自动管理会话生命周期：创建、提交/回滚、关闭
+
+    示例:
+        with get_db_context() as db:
+            user = db.query(User).first()
+            # 事务在 with 块结束时自动提交或回滚
+    """
+    _ensure_engine()
+    assert _SessionLocal is not None
+    db = _SessionLocal()
+    try:
+        yield db
+        db.commit()
+    except Exception:
+        db.rollback()
+        raise
+    finally:
+        db.close()
+
+
 def get_db_url() -> str:
     """返回当前配置的数据库连接字符串（供脚本/测试使用）。"""
     return config.database_url

src/main.py

@@ -202,14 +202,14 @@ async def lifespan(app: FastAPI) -> Any:
     logger.info("启动月卡额度重置调度器...")
     from src.services.model.fetch_scheduler import get_model_fetch_scheduler
     from src.services.system.maintenance_scheduler import get_maintenance_scheduler
+    from src.services.task.task_poller import get_task_poller
     from src.services.usage.quota_scheduler import get_quota_scheduler
-    from src.services.video.task_poller import get_video_task_poller
     from src.utils.task_coordinator import StartupTaskCoordinator
 
     quota_scheduler = get_quota_scheduler()
     maintenance_scheduler = get_maintenance_scheduler()
     model_fetch_scheduler = get_model_fetch_scheduler()
-    video_task_poller = get_video_task_poller()
+    task_poller = get_task_poller()
     task_coordinator = StartupTaskCoordinator(redis_client)
 
     # 启动额度调度器
@@ -238,14 +238,14 @@ async def lifespan(app: FastAPI) -> Any:
         logger.info("检测到其他 worker 已运行模型获取调度器，本实例跳过")
         model_fetch_scheduler = None  # type: ignore[assignment]
 
-    # 启动视频任务轮询服务
-    video_poller_active = await task_coordinator.acquire("video_task_poller")
-    if video_poller_active:
-        logger.info("启动视频任务轮询服务...")
-        await video_task_poller.start()
+    # 启动异步任务轮询服务（当前仅视频）
+    task_poller_active = await task_coordinator.acquire("task_poller:video")
+    if task_poller_active:
+        logger.info("启动 TaskPoller（video）...")
+        await task_poller.start()
     else:
-        logger.info("检测到其他 worker 已运行视频任务轮询，本实例跳过")
-        video_task_poller = None  # type: ignore[assignment]
+        logger.info("检测到其他 worker 已运行 TaskPoller（video），本实例跳过")
+        task_poller = None  # type: ignore[assignment]
 
     # 启动统一的定时任务调度器
     from src.services.system.scheduler import get_scheduler
@@ -296,10 +296,10 @@ async def lifespan(app: FastAPI) -> Any:
         await model_fetch_scheduler.stop()
         await task_coordinator.release("model_fetch_scheduler")
 
-    if video_task_poller:
-        logger.info("停止视频任务轮询...")
-        await video_task_poller.stop()
-        await task_coordinator.release("video_task_poller")
+    if task_poller:
+        logger.info("停止 TaskPoller（video）...")
+        await task_poller.stop()
+        await task_coordinator.release("task_poller:video")
 
     # 停止统一的定时任务调度器
     logger.info("停止定时任务调度器...")

src/models/database.py

@@ -272,6 +272,14 @@ class Usage(Base):
     """使用记录模型"""
 
     __tablename__ = "usage"
+    __table_args__ = (
+        # Composite indexes for common query patterns (analytics / list pages)
+        Index("idx_usage_user_created", "user_id", "created_at"),
+        Index("idx_usage_apikey_created", "api_key_id", "created_at"),
+        Index("idx_usage_provider_model_created", "provider_name", "model", "created_at"),
+        Index("idx_usage_provider_created", "provider_name", "created_at"),
+        Index("idx_usage_model_created", "model", "created_at"),
+    )
 
     id = Column(String(36), primary_key=True, default=lambda: str(uuid.uuid4()), index=True)
     user_id = Column(String(36), ForeignKey("users.id", ondelete="SET NULL"), nullable=True)
@@ -347,6 +355,13 @@ class Usage(Base):
     # cancelled: 客户端主动断开连接
     status = Column(String(20), default="completed", nullable=False, index=True)
 
+    # 结算状态（与 status 解耦）
+    # - pending: 等待结算（任务未完成 / 流式未结束）
+    # - settled: 已结算（cost 已写入，可能 > 0 或 = 0）
+    # - void: 作废（不收费，如任务未开始就取消）
+    billing_status = Column(String(20), default="settled", nullable=False, index=True)
+    finalized_at = Column(DateTime(timezone=True), nullable=True)  # 结算完成时间（可选）
+
     # 完整请求和响应记录
     request_headers = Column(JSON, nullable=True)  # 客户端请求头
     request_body = Column(JSON, nullable=True)  # 请求体（7天内未压缩）
@@ -654,6 +669,14 @@ class Provider(Base):
     # 注意：如果全局配置 KEEP_PRIORITY_ON_CONVERSION=true，此字段被忽略（所有提供商都保持优先级）
     keep_priority_on_conversion = Column(Boolean, default=False, nullable=False)
 
+    # 是否允许格式转换（默认 True）
+    # - True: 该提供商可以作为格式转换的目标（如 OpenAI 客户端请求可以路由到此 Gemini 提供商）
+    # - False: 该提供商不接受需要格式转换的请求
+    # 优先级逻辑：
+    # - 全局开关 ON → 强制允许所有提供商的格式转换（忽略此字段）
+    # - 全局开关 OFF → 由此字段决定是否允许该提供商的格式转换
+    enable_format_conversion = Column(Boolean, default=False, nullable=False)
+
     # 状态
     is_active = Column(Boolean, default=True, nullable=False)
 
@@ -1350,12 +1373,26 @@ class ProviderAPIKey(Base):
     provider = relationship("Provider", back_populates="api_keys")
 
 
+def _generate_short_id(length: int = 12) -> str:
+    """生成 Gemini 风格的短 ID（小写字母+数字）"""
+    import secrets
+    import string
+
+    alphabet = string.ascii_lowercase + string.digits
+    return "".join(secrets.choice(alphabet) for _ in range(length))
+
+
 class VideoTask(Base):
     """视频生成任务"""
 
     __tablename__ = "video_tasks"
 
     id = Column(String(36), primary_key=True, default=lambda: str(uuid.uuid4()))
+    # Gemini 风格的短 ID，用于对外暴露（如 operations/xxx）
+    short_id = Column(String(16), unique=True, index=True, default=_generate_short_id)
+    request_id = Column(
+        String(100), unique=True, index=True, nullable=False
+    )  # 关联 Usage/RequestCandidate
     external_task_id = Column(String(200))
 
     # 关联
@@ -1865,6 +1902,7 @@ class RequestCandidate(Base):
         Index("idx_request_candidates_request_id", "request_id"),
         Index("idx_request_candidates_status", "status"),
         Index("idx_request_candidates_provider_id", "provider_id"),
+        Index("idx_request_candidates_created_at", "created_at"),
     )
 
     # 关系
@@ -2109,5 +2147,60 @@ class StatsUserDaily(Base):
     user = relationship("User")
 
 
+class GeminiFileMapping(Base):
+    """
+    Gemini Files API 文件与 Provider Key 的映射关系
+
+    用于持久化存储 file_id → key_id 的绑定关系，
+    确保后续 generateContent 请求使用上传时的同一 Key。
+
+    Gemini 文件有 48 小时有效期，此表中的记录也会在过期后被清理。
+    """
+
+    __tablename__ = "gemini_file_mappings"
+
+    id = Column(String(36), primary_key=True, default=lambda: str(uuid.uuid4()), index=True)
+
+    # 文件名（如 files/abc123xyz）
+    file_name = Column(String(255), nullable=False, unique=True, index=True)
+
+    # Provider Key ID（关联到 provider_api_keys 表）
+    key_id = Column(
+        String(36),
+        ForeignKey("provider_api_keys.id", ondelete="CASCADE"),
+        nullable=False,
+        index=True,
+    )
+
+    # 用户 ID（用于权限验证，可选）
+    user_id = Column(
+        String(36), ForeignKey("users.id", ondelete="CASCADE"), nullable=True, index=True
+    )
+
+    # 文件元数据（可选，用于调试）
+    display_name = Column(String(255), nullable=True)
+    mime_type = Column(String(100), nullable=True)
+
+    # 源文件哈希（用于关联相同源文件的不同上传，可选）
+    # 当同一源文件上传到多个 Key 时，可通过此字段找到所有等效文件
+    source_hash = Column(String(64), nullable=True, index=True)
+
+    # 时间戳
+    created_at = Column(
+        DateTime(timezone=True), default=lambda: datetime.now(timezone.utc), nullable=False
+    )
+    # 过期时间（Gemini 文件 48 小时后过期）
+    expires_at = Column(DateTime(timezone=True), nullable=False, index=True)
+
+    # 关系
+    key = relationship("ProviderAPIKey")
+    user = relationship("User")
+
+    __table_args__ = (
+        Index("idx_gemini_file_mappings_expires", "expires_at"),
+        Index("idx_gemini_file_mappings_source_hash", "source_hash"),
+    )
+
+
 # 导入扩展的数据库模型
 from .database_extensions import ApiKeyProviderMapping, ProviderUsageTracking

src/models/endpoint_models.py

@@ -635,6 +635,10 @@ class ProviderUpdateRequest(BaseModel):
         None,
         description="格式转换时是否保持优先级（True=保持原优先级，False=需要转换时降级）",
     )
+    enable_format_conversion: bool | None = Field(
+        None,
+        description="是否允许格式转换（提供商级别开关）",
+    )
     is_active: bool | None = None
     billing_type: str | None = Field(
         None, description="计费类型：monthly_quota/pay_as_you_go/free_tier"
@@ -667,6 +671,10 @@ class ProviderWithEndpointsSummary(BaseModel):
         default=False,
         description="格式转换时是否保持优先级（True=保持原优先级，False=需要转换时降级）",
     )
+    enable_format_conversion: bool = Field(
+        default=True,
+        description="是否允许格式转换（提供商级别开关）",
+    )
     is_active: bool
 
     # 计费相关字段

src/modules/__init__.py

@@ -7,6 +7,7 @@
 from src.core.modules.base import ModuleDefinition
 
 # 导入所有模块定义
+from src.modules.gemini_files import gemini_files_module
 from src.modules.ldap import ldap_module
 from src.modules.oauth import oauth_module
 
@@ -14,6 +15,7 @@ from src.modules.oauth import oauth_module
 ALL_MODULES: list[ModuleDefinition] = [
     ldap_module,
     oauth_module,
+    gemini_files_module,
 ]
 
 __all__ = ["ALL_MODULES"]

src/modules/gemini_files/__init__.py

@@ -0,0 +1,96 @@
+"""
+Gemini Files 文件管理模块
+
+提供 Gemini Files API 文件上传和管理功能：
+- 文件上传到 Google Gemini Files API
+- 文件映射管理（file_id → key_id）
+- 文件列表查看和删除
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any
+
+from src.core.modules.base import (
+    ModuleCategory,
+    ModuleDefinition,
+    ModuleHealth,
+    ModuleMetadata,
+)
+
+if TYPE_CHECKING:
+    from sqlalchemy.orm import Session
+
+
+def _get_router() -> Any:
+    """延迟导入路由"""
+    from src.api.admin.gemini_files import router
+
+    return router
+
+
+async def _health_check() -> ModuleHealth:
+    """健康检查"""
+    # 检查是否有可用的 gemini_files 能力的 Key
+    from src.database import create_session
+    from src.models.database import ProviderAPIKey
+
+    db = create_session()
+    try:
+        # 查找有 gemini_files 能力的 Key
+        keys = db.query(ProviderAPIKey).filter(ProviderAPIKey.is_active.is_(True)).all()
+        has_capable_key = any(
+            key.capabilities and key.capabilities.get("gemini_files", False) for key in keys
+        )
+        if has_capable_key:
+            return ModuleHealth.HEALTHY
+        return ModuleHealth.DEGRADED
+    except Exception:
+        return ModuleHealth.UNKNOWN
+    finally:
+        db.close()
+
+
+def _validate_config(db: Session) -> tuple[bool, str]:
+    """
+    验证 Gemini Files 模块配置
+
+    检查项：
+    1. 至少有一个有 gemini_files 能力的 Provider Key
+    """
+    from src.models.database import ProviderAPIKey
+
+    # 查找有 gemini_files 能力的 Key
+    keys = db.query(ProviderAPIKey).filter(ProviderAPIKey.is_active.is_(True)).all()
+    capable_keys = [
+        key for key in keys if key.capabilities and key.capabilities.get("gemini_files", False)
+    ]
+
+    if not capable_keys:
+        return (
+            False,
+            "至少启用一个具有「Gemini 文件 API」能力的 Key",
+        )
+
+    return True, ""
+
+
+gemini_files_module = ModuleDefinition(
+    metadata=ModuleMetadata(
+        name="gemini_files",
+        display_name="文件缓存",
+        description="管理 Gemini Files API 上传的文件，支持文件上传、查看和删除",
+        category=ModuleCategory.INTEGRATION,
+        env_key="GEMINI_FILES_AVAILABLE",
+        default_available=True,
+        required_packages=[],
+        api_prefix="/api/admin/gemini-files",
+        admin_route="/admin/gemini-files",
+        admin_menu_icon="FileUp",
+        admin_menu_group="system",
+        admin_menu_order=60,
+    ),
+    router_factory=_get_router,
+    health_check=_health_check,
+    validate_config=_validate_config,
+)

src/services/auth/oauth/service.py

@@ -804,8 +804,33 @@ class OAuthService:
 
         cfg = OAuthService._get_provider_config(db, provider_type)
 
+        # Read all required fields first, then release DB connection before any awaits.
+        # This prevents holding a pooled connection while doing network I/O.
         auth_url = provider.get_effective_authorization_url(cfg)
         token_url = provider.get_effective_token_url(cfg)
+        redirect_uri = cfg.redirect_uri
+        client_id = cfg.client_id
+        has_secret = bool(cfg.client_secret_encrypted)
+        client_secret = cfg.get_client_secret() if has_secret else None
+
+        # Release DB connection (safe only when session has no pending changes).
+        try:
+            has_pending_changes = bool(db.new) or bool(db.dirty) or bool(db.deleted)
+        except Exception:
+            has_pending_changes = False
+        if not has_pending_changes:
+            original_expire_on_commit = getattr(db, "expire_on_commit", True)
+            db.expire_on_commit = False
+            try:
+                if db.in_transaction():
+                    db.commit()
+            except Exception:
+                try:
+                    db.rollback()
+                except Exception:
+                    pass
+            finally:
+                db.expire_on_commit = original_expire_on_commit
 
         async def _reachable(url: str) -> bool:
             try:
@@ -823,7 +848,7 @@ class OAuthService:
         secret_status = "unknown"
         details = ""
 
-        if cfg.client_secret_encrypted:
+        if has_secret and client_secret:
             # 使用无效 code 做一次 token 请求（仅做粗略判定）
             try:
                 async with httpx.AsyncClient(
@@ -834,9 +859,9 @@ class OAuthService:
                         data={
                             "grant_type": "authorization_code",
                             "code": "invalid",
-                            "redirect_uri": cfg.redirect_uri,
-                            "client_id": cfg.client_id,
-                            "client_secret": cfg.get_client_secret(),
+                            "redirect_uri": redirect_uri,
+                            "client_id": client_id,
+                            "client_secret": client_secret,
                         },
                     )
                 try:

src/services/billing/__init__.py

@@ -29,6 +29,8 @@ from src.services.billing.models import (
     CostBreakdown,
     StandardizedUsage,
 )
+from src.services.billing.schema import BillingSnapshot, CostResult
+from src.services.billing.service import BillingService
 from src.services.billing.templates import BILLING_TEMPLATE_REGISTRY, BillingTemplates
 from src.services.billing.usage_mapper import UsageMapper, map_usage, map_usage_from_response
 
@@ -44,6 +46,10 @@ __all__ = [
     # 计算器
     "BillingCalculator",
     "calculate_request_cost",
+    # 统一入口（Phase2）
+    "BillingService",
+    "BillingSnapshot",
+    "CostResult",
     # 映射器
     "UsageMapper",
     "map_usage",

src/services/billing/schema.py

@@ -0,0 +1,64 @@
+"""
+Billing schema (stable contracts)
+
+These dataclasses are meant to be stored in `Usage.request_metadata` / `Task.request_metadata`
+for auditability. They are internal-only and MUST NOT be exposed to end users without sanitizing.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from typing import Any, Literal
+
+BILLING_SNAPSHOT_SCHEMA_VERSION = "1.0"
+
+BillingSnapshotStatus = Literal["complete", "incomplete", "no_rule", "legacy"]
+
+
+@dataclass(frozen=True)
+class BillingSnapshot:
+    """Stable billing snapshot for audit."""
+
+    schema_version: str = BILLING_SNAPSHOT_SCHEMA_VERSION
+
+    # Rule info (optional for legacy/no_rule)
+    rule_id: str | None = None
+    rule_name: str | None = None
+    scope: str | None = None
+
+    # Rule expression (internal, do not expose to clients)
+    expression: str | None = None
+
+    # Dimensions
+    dimensions_used: dict[str, Any] = field(default_factory=dict)
+    missing_required: list[str] = field(default_factory=list)
+
+    # Result
+    cost: float = 0.0
+    status: BillingSnapshotStatus = "no_rule"
+
+    # Audit
+    calculated_at: str = ""  # ISO 8601
+
+    def to_dict(self) -> dict[str, Any]:
+        return {
+            "schema_version": self.schema_version,
+            "rule_id": self.rule_id,
+            "rule_name": self.rule_name,
+            "scope": self.scope,
+            "expression": self.expression,
+            "dimensions_used": self.dimensions_used,
+            "missing_required": self.missing_required,
+            "cost": self.cost,
+            "status": self.status,
+            "calculated_at": self.calculated_at,
+        }
+
+
+@dataclass(frozen=True)
+class CostResult:
+    """Billing calculation output."""
+
+    cost: float
+    status: BillingSnapshotStatus
+    snapshot: BillingSnapshot

src/services/billing/service.py

@@ -0,0 +1,145 @@
+from __future__ import annotations
+
+from datetime import datetime, timezone
+from typing import Any
+
+from sqlalchemy.orm import Session
+
+from src.config.settings import config
+from src.core.logger import logger
+from src.services.billing.dimension_collector_service import DimensionCollectorService
+from src.services.billing.formula_engine import BillingIncompleteError, FormulaEngine
+from src.services.billing.rule_service import BillingRuleService
+from src.services.model.cost import ModelCostService
+
+from .schema import BILLING_SNAPSHOT_SCHEMA_VERSION, BillingSnapshot, CostResult
+
+
+class BillingService:
+    """
+    BillingService (pure-ish application helper for billing domain).
+
+    Notes:
+    - This service **does not** write Usage rows.
+    - It may read billing rules & collectors from DB.
+    """
+
+    def __init__(self, db: Session):
+        self.db = db
+        self._formula_engine = FormulaEngine()
+        self._dimension_collector = DimensionCollectorService(db)
+
+    def collect_dimensions(
+        self,
+        *,
+        api_format: str | None,
+        task_type: str | None,
+        request: dict[str, Any] | None = None,
+        response: dict[str, Any] | None = None,
+        metadata: dict[str, Any] | None = None,
+        base_dimensions: dict[str, Any] | None = None,
+    ) -> dict[str, Any]:
+        return self._dimension_collector.collect_dimensions(
+            api_format=api_format,
+            task_type=task_type,
+            request=request,
+            response=response,
+            metadata=metadata,
+            base_dimensions=base_dimensions,
+        )
+
+    def calculate(
+        self,
+        *,
+        task_type: str,
+        model: str,
+        provider_id: str,
+        dimensions: dict[str, Any],
+        strict_mode: bool | None = None,
+    ) -> CostResult:
+        """
+        Calculate cost for a task.
+
+        Returns:
+            CostResult (includes BillingSnapshot)
+
+        Raises:
+            BillingIncompleteError: when strict_mode=True and required dims missing.
+        """
+        strict = config.billing_strict_mode if strict_mode is None else bool(strict_mode)
+
+        lookup = BillingRuleService.find_rule(
+            self.db,
+            provider_id=provider_id,
+            model_name=model,
+            task_type=task_type,
+        )
+
+        if lookup and lookup.rule and lookup.rule.expression:
+            rule = lookup.rule
+            result = self._formula_engine.evaluate(
+                expression=rule.expression,
+                variables=rule.variables or {},
+                dimensions=dimensions,
+                dimension_mappings=rule.dimension_mappings or {},
+                strict_mode=strict,
+            )
+            cost = float(result.cost) if result.status == "complete" else 0.0
+            snapshot = BillingSnapshot(
+                schema_version=BILLING_SNAPSHOT_SCHEMA_VERSION,
+                rule_id=str(rule.id),
+                rule_name=str(rule.name),
+                scope=str(getattr(lookup, "scope", None) or ""),
+                expression=str(rule.expression),
+                dimensions_used=dimensions,
+                missing_required=result.missing_required,
+                cost=cost,
+                status=result.status,
+                calculated_at=datetime.now(timezone.utc).isoformat(),
+            )
+            return CostResult(cost=cost, status=result.status, snapshot=snapshot)
+
+        # No rule fallback
+        if task_type in ("chat", "cli"):
+            input_tokens = int(dimensions.get("input_tokens") or 0)
+            output_tokens = int(dimensions.get("output_tokens") or 0)
+            cost = float(
+                ModelCostService.calculate_cost(
+                    model=model,
+                    input_tokens=input_tokens,
+                    output_tokens=output_tokens,
+                )
+            )
+            snapshot = BillingSnapshot(
+                schema_version=BILLING_SNAPSHOT_SCHEMA_VERSION,
+                rule_id=None,
+                rule_name=None,
+                scope=None,
+                expression=None,
+                dimensions_used=dimensions,
+                missing_required=[],
+                cost=cost,
+                status="legacy",
+                calculated_at=datetime.now(timezone.utc).isoformat(),
+            )
+            return CostResult(cost=cost, status="legacy", snapshot=snapshot)
+
+        logger.warning(
+            "No billing rule for task (task_type=%s, model=%s, provider_id=%s)",
+            task_type,
+            model,
+            provider_id,
+        )
+        snapshot = BillingSnapshot(
+            schema_version=BILLING_SNAPSHOT_SCHEMA_VERSION,
+            rule_id=None,
+            rule_name=None,
+            scope=None,
+            expression=None,
+            dimensions_used=dimensions,
+            missing_required=[],
+            cost=0.0,
+            status="no_rule",
+            calculated_at=datetime.now(timezone.utc).isoformat(),
+        )
+        return CostResult(cost=0.0, status="no_rule", snapshot=snapshot)

src/services/cache/aware_scheduler.py

@@ -182,6 +182,43 @@ class CacheAwareScheduler:
             "last_reservation_result": None,
         }
 
+    @staticmethod
+    def _release_db_connection_before_await(db: Session) -> None:
+        """
+        Best-effort: end a read-only transaction before awaiting async I/O.
+
+        This scheduler does a lot of async work (cache/Redis) mixed with sync SQLAlchemy reads.
+        If a SELECT has already started a transaction, the pooled connection can remain checked
+        out while we await, causing pool pressure under concurrency.
+
+        Safety:
+        - Only commits when the Session has no ORM pending changes.
+        - Temporarily disables expire_on_commit to keep already-loaded ORM objects usable.
+        """
+        try:
+            if db is None:
+                return
+            has_pending_changes = bool(db.new) or bool(db.dirty) or bool(db.deleted)
+            if has_pending_changes:
+                return
+            if not db.in_transaction():
+                return
+
+            original_expire_on_commit = getattr(db, "expire_on_commit", True)
+            db.expire_on_commit = False
+            try:
+                db.commit()
+            except Exception:
+                try:
+                    db.rollback()
+                except Exception:
+                    pass
+            finally:
+                db.expire_on_commit = original_expire_on_commit
+        except Exception:
+            # Never let this optimization break scheduling
+            return
+
     async def _ensure_initialized(self) -> None:
         """确保所有异步组件已初始化"""
         if self._affinity_manager is None:
@@ -577,6 +614,8 @@ class CacheAwareScheduler:
         Returns:
             (候选列表, global_model_id) - global_model_id 用于缓存亲和性
         """
+        # If the caller already touched the DB, release the connection before we do async work.
+        self._release_db_connection_before_await(db)
         await self._ensure_initialized()
 
         target_format = normalize_endpoint_signature(api_format)
@@ -648,6 +687,9 @@ class CacheAwareScheduler:
             provider_limit=provider_limit,
         )
 
+        # Provider query starts a transaction; release connection before entering async candidate build.
+        self._release_db_connection_before_await(db)
+
         logger.debug(
             "[Scheduler] Found %d active providers",
             len(providers),
@@ -680,8 +722,13 @@ class CacheAwareScheduler:
 
         # 2. 构建候选列表（传入 is_stream 和 capability_requirements 用于过滤）
         from src.config.settings import config
+        from src.services.system.config import SystemConfigService
 
-        global_conversion_enabled = config.format_conversion_enabled
+        # 全局格式转换开关：优先使用数据库配置，回退到环境变量
+        global_conversion_enabled = SystemConfigService.is_format_conversion_enabled(db)
+        # 如果环境变量明确禁用，则禁用（环境变量可作为强制禁用开关）
+        if not config.format_conversion_enabled:
+            global_conversion_enabled = False
         candidates = await self._build_candidates(
             db=db,
             providers=providers,
@@ -801,6 +848,9 @@ class CacheAwareScheduler:
             - supported_capabilities: 模型支持的能力列表
             - provider_model_names: Provider 侧可用的模型名称集合（主名称 + 映射名称，按 api_format 过滤）
         """
+        # Avoid holding a DB connection while awaiting cache/Redis inside ModelCacheService.
+        self._release_db_connection_before_await(db)
+
         # 使用 ModelCacheService 解析模型名称（支持映射名）
         global_model = await ModelCacheService.resolve_global_model_by_name_or_mapping(
             db, model_name
@@ -1120,18 +1170,34 @@ class CacheAwareScheduler:
                     str(getattr(endpoint, "endpoint_kind", "")).strip().lower(),
                 )
 
+                # 计算格式转换的有效开关状态（三层优先级）
+                # 全局 ON → 强制允许（跳过端点检查）
+                # 全局 OFF → 提供商 ON → 强制允许（跳过端点检查）
+                # 全局 OFF → 提供商 OFF → 看端点配置
+                provider_allows_conversion = getattr(provider, "enable_format_conversion", True)
+                effective_conversion_enabled = (
+                    global_conversion_enabled or provider_allows_conversion
+                )
+                # 如果全局或提供商开关为 ON，跳过端点配置检查
+                skip_endpoint_check = global_conversion_enabled or provider_allows_conversion
+
                 is_compatible, needs_conversion, _compat_reason = is_format_compatible(
                     client_format_str,
                     endpoint_format_str,
                     getattr(endpoint, "format_acceptance_config", None),
                     is_stream,
-                    global_conversion_enabled,
+                    effective_conversion_enabled,
+                    skip_endpoint_check=skip_endpoint_check,
                 )
                 logger.debug(
-                    "[Scheduler] Format compatibility: client=%s, endpoint=%s, compatible=%s, reason=%s",
+                    "[Scheduler] Format compatibility: client=%s, endpoint=%s, compatible=%s, "
+                    "global=%s, provider=%s, skip_endpoint=%s, reason=%s",
                     client_format_str,
                     endpoint_format_str,
                     is_compatible,
+                    global_conversion_enabled,
+                    provider_allows_conversion,
+                    skip_endpoint_check,
                     _compat_reason,
                 )
                 if not is_compatible:

src/services/candidate/__init__.py

@@ -0,0 +1,28 @@
+"""
+Candidate domain (Phase2)
+
+This package centralizes:
+- candidate resolving (Provider/Endpoint/Key combinations)
+- request_candidates recording & audit
+- failover execution policies
+"""
+
+from src.services.candidate.policy import RetryMode, RetryPolicy, SkipPolicy
+from src.services.candidate.schema import (
+    CANDIDATE_KEY_SCHEMA_VERSION,
+    CandidateKey,
+    CandidateResult,
+)
+from src.services.candidate.service import CandidateService
+
+__all__ = [
+    "CandidateService",
+    # schema
+    "CANDIDATE_KEY_SCHEMA_VERSION",
+    "CandidateKey",
+    "CandidateResult",
+    # policies
+    "RetryMode",
+    "RetryPolicy",
+    "SkipPolicy",
+]

src/services/candidate/failover.py

@@ -0,0 +1,57 @@
+from __future__ import annotations
+
+from typing import Any, Protocol
+
+from sqlalchemy.orm import Session
+
+from src.services.cache.aware_scheduler import ProviderCandidate
+
+from .policy import RetryPolicy, SkipPolicy
+from .schema import CandidateKey, CandidateResult
+
+
+class AttemptFunc(Protocol):
+    async def __call__(self, candidate: ProviderCandidate) -> Any: ...
+
+
+class FailoverEngine:
+    """
+    FailoverEngine executes candidate attempts under policies.
+
+    Phase2 scaffolding: implementation will gradually replace legacy orchestrators.
+    """
+
+    def __init__(self, db: Session):
+        self.db = db
+
+    async def execute(
+        self,
+        *,
+        candidates: list[ProviderCandidate],
+        attempt_func: AttemptFunc,
+        retry_policy: RetryPolicy,
+        skip_policy: SkipPolicy,
+        request_id: str | None = None,
+        max_candidates: int | None = None,
+    ) -> CandidateResult:
+        # NOTE: intentionally minimal for now; legacy orchestrators still in use.
+        # This will be implemented when migrating video/chat flows to CandidateService.
+        _ = (retry_policy, skip_policy, request_id, max_candidates)
+        candidate_keys: list[CandidateKey] = []
+        for idx, cand in enumerate(candidates):
+            candidate_keys.append(
+                CandidateKey(
+                    candidate_index=idx,
+                    provider_id=str(cand.provider.id),
+                    provider_name=str(cand.provider.name),
+                    endpoint_id=str(cand.endpoint.id),
+                    key_id=str(cand.key.id),
+                    key_name=str(getattr(cand.key, "name", "") or ""),
+                    auth_type=str(getattr(cand.key, "auth_type", "") or ""),
+                    priority=int(getattr(cand.key, "priority", 0) or 0),
+                    is_cached=bool(getattr(cand, "is_cached", False)),
+                    status="available",
+                )
+            )
+
+        raise NotImplementedError("FailoverEngine.execute is not implemented yet")

src/services/candidate/policy.py

@@ -0,0 +1,41 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from enum import Enum
+
+
+class RetryMode(str, Enum):
+    """Retry mode for candidate attempts."""
+
+    PRE_EXPAND = "pre_expand"  # pre-create retry slots (sync)
+    ON_DEMAND = "on_demand"  # create retry record only when retry happens
+    DISABLED = "disabled"  # no retry
+
+
+@dataclass(frozen=True)
+class RetryPolicy:
+    """Unified retry policy."""
+
+    mode: RetryMode = RetryMode.DISABLED
+    max_retries: int = 1
+    retry_on_cached_only: bool = True
+
+    @classmethod
+    def for_sync_task(cls) -> "RetryPolicy":
+        return cls(mode=RetryMode.PRE_EXPAND, max_retries=2)
+
+    @classmethod
+    def for_async_task(cls) -> "RetryPolicy":
+        return cls(mode=RetryMode.DISABLED, max_retries=1)
+
+    @classmethod
+    def for_async_submit_with_retry(cls) -> "RetryPolicy":
+        return cls(mode=RetryMode.ON_DEMAND, max_retries=2)
+
+
+@dataclass(frozen=True)
+class SkipPolicy:
+    """Rules for skipping unsupported candidates."""
+
+    allow_format_conversion: bool = True
+    supported_auth_types: set[str] | None = None

src/services/candidate/recorder.py

@@ -0,0 +1,59 @@
+from __future__ import annotations
+
+from sqlalchemy.orm import Session
+
+from src.models.database import RequestCandidate
+
+from .schema import CandidateKey
+
+
+class CandidateRecorder:
+    """Read helpers for RequestCandidate audit data."""
+
+    def __init__(self, db: Session):
+        self.db = db
+
+    def get_candidate_keys(self, request_id: str) -> list[CandidateKey]:
+        rows: list[RequestCandidate] = (
+            self.db.query(RequestCandidate)
+            .filter(RequestCandidate.request_id == request_id)
+            .order_by(RequestCandidate.candidate_index.asc(), RequestCandidate.retry_index.asc())
+            .all()
+        )
+
+        result: list[CandidateKey] = []
+        for row in rows:
+            provider_name = None
+            if getattr(row, "provider", None) is not None:
+                provider_name = getattr(row.provider, "name", None)
+
+            key_name = None
+            auth_type = None
+            priority = None
+            if getattr(row, "key", None) is not None:
+                key_name = getattr(row.key, "name", None)
+                auth_type = getattr(row.key, "auth_type", None)
+                priority = getattr(row.key, "priority", None)
+
+            result.append(
+                CandidateKey(
+                    candidate_index=int(row.candidate_index or 0),
+                    retry_index=int(row.retry_index or 0),
+                    provider_id=str(row.provider_id) if row.provider_id else None,
+                    provider_name=str(provider_name) if provider_name else None,
+                    endpoint_id=str(row.endpoint_id) if row.endpoint_id else None,
+                    key_id=str(row.key_id) if row.key_id else None,
+                    key_name=str(key_name) if key_name else None,
+                    auth_type=str(auth_type) if auth_type else None,
+                    priority=int(priority) if priority is not None else None,
+                    is_cached=bool(getattr(row, "is_cached", False)),
+                    status=str(getattr(row, "status", "") or "pending"),
+                    skip_reason=getattr(row, "skip_reason", None),
+                    error_type=getattr(row, "error_type", None),
+                    error_message=getattr(row, "error_message", None),
+                    status_code=getattr(row, "status_code", None),
+                    latency_ms=getattr(row, "latency_ms", None),
+                )
+            )
+
+        return result

src/services/candidate/resolver.py

@@ -0,0 +1,10 @@
+"""
+CandidateResolver facade import.
+
+Phase2 keeps the implementation in `services/orchestration/` for compatibility,
+and gradually migrates it into this package.
+"""
+
+from src.services.orchestration.candidate_resolver import CandidateResolver
+
+__all__ = ["CandidateResolver"]

src/services/candidate/schema.py

@@ -0,0 +1,71 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any
+
+from src.services.cache.aware_scheduler import ProviderCandidate
+
+CANDIDATE_KEY_SCHEMA_VERSION = "1.0"
+
+
+@dataclass(frozen=True)
+class CandidateKey:
+    """Stable candidate key snapshot for audit."""
+
+    schema_version: str = CANDIDATE_KEY_SCHEMA_VERSION
+
+    candidate_index: int = 0
+    retry_index: int = 0
+
+    provider_id: str | None = None
+    provider_name: str | None = None
+    endpoint_id: str | None = None
+    key_id: str | None = None
+    key_name: str | None = None
+    auth_type: str | None = None
+    priority: int | None = None
+    is_cached: bool = False
+
+    status: str = "pending"  # pending/success/failed/skipped/available/...
+    skip_reason: str | None = None
+    error_type: str | None = None
+    error_message: str | None = None
+    status_code: int | None = None
+    latency_ms: int | None = None
+
+    def to_dict(self) -> dict[str, Any]:
+        data: dict[str, Any] = {
+            "schema_version": self.schema_version,
+            "candidate_index": self.candidate_index,
+            "retry_index": self.retry_index,
+            "provider_id": self.provider_id,
+            "provider_name": self.provider_name,
+            "endpoint_id": self.endpoint_id,
+            "key_id": self.key_id,
+            "key_name": self.key_name,
+            "auth_type": self.auth_type,
+            "priority": self.priority,
+            "is_cached": self.is_cached,
+            "status": self.status,
+            "skip_reason": self.skip_reason,
+            "error_type": self.error_type,
+            "error_message": self.error_message,
+            "status_code": self.status_code,
+            "latency_ms": self.latency_ms,
+        }
+        # drop Nones for compact audit payload
+        return {k: v for k, v in data.items() if v is not None}
+
+
+@dataclass(slots=True)
+class CandidateResult:
+    """Failover execution result."""
+
+    success: bool
+    selected: ProviderCandidate | None
+    selected_index: int | None
+    candidate_keys: list[CandidateKey]
+
+    external_task_id: str | None = None
+    error: Exception | None = None
+    last_status_code: int | None = None

src/services/candidate/service.py

@@ -0,0 +1,515 @@
+from __future__ import annotations
+
+import re
+from datetime import datetime, timezone
+from typing import Any
+
+import httpx
+from sqlalchemy import update
+from sqlalchemy.orm import Session
+
+from src.config.settings import config
+from src.core.exceptions import ProviderNotAvailableException
+from src.core.logger import logger
+from src.models.database import ApiKey, RequestCandidate
+from src.services.billing.rule_service import BillingRuleLookupResult, BillingRuleService
+from src.services.cache.aware_scheduler import ProviderCandidate, get_cache_aware_scheduler
+from src.services.candidate.submit import (
+    AllCandidatesFailedError,
+    SubmitOutcome,
+    UpstreamClientRequestError,
+)
+from src.services.orchestration.error_classifier import ErrorClassifier
+from src.services.system.config import SystemConfigService
+
+from .recorder import CandidateRecorder
+from .resolver import CandidateResolver
+from .schema import CandidateKey
+
+_SENSITIVE_PATTERN = re.compile(
+    r"(api[_-]?key|token|bearer|authorization)[=:\s]+\S+",
+    re.IGNORECASE,
+)
+
+
+def _sanitize(message: str, max_length: int = 200) -> str:
+    if not message:
+        return "request_failed"
+    return _SENSITIVE_PATTERN.sub("[REDACTED]", message)[:max_length]
+
+
+class CandidateService:
+    """
+    CandidateService (Facade).
+
+    Phase2 note: this is introduced as a new domain entrypoint. Legacy orchestrators
+    still exist and will be migrated gradually to use this service.
+    """
+
+    def __init__(self, db: Session, redis_client: Any | None = None) -> None:
+        self.db = db
+        self.redis = redis_client
+        self._cache_scheduler = None
+        self._resolver: CandidateResolver | None = None
+        self._error_classifier: ErrorClassifier | None = None
+        self._recorder = CandidateRecorder(db)
+
+    async def _ensure_initialized(self) -> None:
+        if self._cache_scheduler is not None:
+            return
+
+        priority_mode = SystemConfigService.get_config(
+            self.db,
+            "provider_priority_mode",
+            "provider",
+        )
+        scheduling_mode = SystemConfigService.get_config(
+            self.db,
+            "scheduling_mode",
+            "cache_affinity",
+        )
+        self._cache_scheduler = await get_cache_aware_scheduler(
+            self.redis,
+            priority_mode=priority_mode,
+            scheduling_mode=scheduling_mode,
+        )
+        self._resolver = CandidateResolver(db=self.db, cache_scheduler=self._cache_scheduler)
+        self._error_classifier = ErrorClassifier(db=self.db, cache_scheduler=self._cache_scheduler)
+
+    async def resolve(
+        self,
+        *,
+        api_format: str,
+        model_name: str,
+        affinity_key: str,
+        user_api_key: ApiKey | None = None,
+        request_id: str | None = None,
+        is_stream: bool = False,
+        capability_requirements: dict[str, bool] | None = None,
+        preferred_key_ids: list[str] | None = None,
+    ) -> tuple[list[ProviderCandidate], str]:
+        await self._ensure_initialized()
+        assert self._resolver is not None
+        return await self._resolver.fetch_candidates(
+            api_format=api_format,
+            model_name=model_name,
+            affinity_key=affinity_key,
+            user_api_key=user_api_key,
+            request_id=request_id,
+            is_stream=is_stream,
+            capability_requirements=capability_requirements,
+            preferred_key_ids=preferred_key_ids,
+        )
+
+    def _should_stop_on_http_error(self, *, status_code: int, error_text: str) -> bool:
+        """
+        Decide whether an upstream HTTP error is a client error (no failover).
+
+        Rules:
+        - 401/403/429 are usually key/permission/ratelimit issues -> allow failover
+        - other 4xx: stop only if ErrorClassifier says it's a client error
+        """
+        if status_code in (401, 403, 429):
+            return False
+        if 400 <= status_code < 500:
+            assert self._error_classifier is not None
+            return self._error_classifier.is_client_error(error_text)
+        return False
+
+    async def submit_with_failover(
+        self,
+        *,
+        api_format: str,
+        model_name: str,
+        affinity_key: str,
+        user_api_key: ApiKey,
+        request_id: str | None,
+        task_type: str,
+        submit_func: Any,
+        extract_external_task_id: Any,
+        supported_auth_types: set[str] | None = None,
+        allow_format_conversion: bool = False,
+        capability_requirements: dict[str, bool] | None = None,
+        max_candidates: int | None = None,
+    ) -> SubmitOutcome:
+        """
+        Submit async task with failover, returning the selected candidate + external_task_id.
+
+        Phase2 submit entrypoint (replaces legacy submit orchestrator).
+        """
+        # IMPORTANT:
+        # This method awaits upstream HTTP calls. If we have an open DB transaction before awaiting,
+        # the connection can be held for a long time (pool exhaustion under concurrency).
+        #
+        # Also note SQLAlchemy's default expire_on_commit=True would expire ORM objects and may
+        # trigger unexpected lazy DB loads after we commit (potentially during the await).
+        # We disable it temporarily to keep candidate/provider/key objects in-memory.
+        original_expire_on_commit = getattr(self.db, "expire_on_commit", True)
+        self.db.expire_on_commit = False
+        await self._ensure_initialized()
+        assert self._resolver is not None
+        try:
+            candidates, _global_model_id = await self._resolver.fetch_candidates(
+                api_format=api_format,
+                model_name=model_name,
+                affinity_key=affinity_key,
+                user_api_key=user_api_key,
+                request_id=request_id,
+                is_stream=False,
+                capability_requirements=capability_requirements,
+            )
+
+            if not candidates:
+                raise ProviderNotAvailableException("No candidates available")
+
+            if max_candidates is not None and max_candidates > 0:
+                candidates = candidates[:max_candidates]
+
+            # Pre-create RequestCandidate records (no retry expand for async submit stage)
+            record_map: dict[tuple[int, int], str] = {}
+            if request_id:
+                try:
+                    record_map = self.create_candidate_records(
+                        candidates=candidates,
+                        request_id=request_id,
+                        user_api_key=user_api_key,
+                        required_capabilities=capability_requirements,
+                        expand_retries=False,
+                    )
+                except Exception as exc:
+                    logger.warning(
+                        "[CandidateService] Failed to create candidate records: %s",
+                        _sanitize(str(exc)),
+                    )
+                    record_map = {}
+
+            candidate_keys: list[dict[str, Any]] = []
+            eligible_count = 0
+            last_status_code: int | None = None
+
+            for idx, cand in enumerate(candidates):
+                now = datetime.now(timezone.utc)
+                auth_type = getattr(cand.key, "auth_type", "api_key") or "api_key"
+
+                candidate_info: dict[str, Any] = {
+                    "index": idx,
+                    "provider_id": cand.provider.id,
+                    "provider_name": cand.provider.name,
+                    "endpoint_id": cand.endpoint.id,
+                    "key_id": cand.key.id,
+                    "key_name": getattr(cand.key, "name", None),
+                    "auth_type": auth_type,
+                    "priority": getattr(cand.key, "priority", 0) or 0,
+                    "is_cached": bool(getattr(cand, "is_cached", False)),
+                }
+                candidate_keys.append(candidate_info)
+
+                record_id = record_map.get((idx, 0))
+
+                # Scheduler marked skip
+                if getattr(cand, "is_skipped", False):
+                    skip_reason = getattr(cand, "skip_reason", None) or "skipped"
+                    candidate_info.update({"skipped": True, "skip_reason": skip_reason})
+                    if record_id:
+                        # record is usually already skipped, but keep it consistent
+                        self.db.execute(
+                            update(RequestCandidate)
+                            .where(RequestCandidate.id == record_id)
+                            .values(status="skipped", skip_reason=skip_reason)
+                        )
+                    continue
+
+                # Format conversion checks
+                # 优先级：全局开关 ON 强制允许，全局开关 OFF 看提供商开关
+                needs_conversion = bool(getattr(cand, "needs_conversion", False))
+                if needs_conversion:
+                    # 1. Check handler-level switch (handler 不支持则直接跳过)
+                    if not allow_format_conversion:
+                        skip_reason = "format_conversion_not_supported"
+                        candidate_info.update({"skipped": True, "skip_reason": skip_reason})
+                        if record_id:
+                            self.db.execute(
+                                update(RequestCandidate)
+                                .where(RequestCandidate.id == record_id)
+                                .values(status="skipped", skip_reason=skip_reason)
+                            )
+                        continue
+
+                    # 2. Check global + provider switches
+                    # 全局 ON → 允许；全局 OFF → 看提供商
+                    from src.services.system.config import SystemConfigService
+
+                    global_enabled = SystemConfigService.is_format_conversion_enabled(self.db)
+                    provider_enabled = getattr(cand.provider, "enable_format_conversion", True)
+                    effective_enabled = global_enabled or provider_enabled
+
+                    if not effective_enabled:
+                        skip_reason = "format_conversion_disabled"
+                        candidate_info.update(
+                            {
+                                "skipped": True,
+                                "skip_reason": skip_reason,
+                                "global_conversion_enabled": global_enabled,
+                                "provider_conversion_enabled": provider_enabled,
+                            }
+                        )
+                        if record_id:
+                            self.db.execute(
+                                update(RequestCandidate)
+                                .where(RequestCandidate.id == record_id)
+                                .values(status="skipped", skip_reason=skip_reason)
+                            )
+                        continue
+
+                # auth_type filter
+                if supported_auth_types is not None and auth_type not in supported_auth_types:
+                    skip_reason = f"unsupported_auth_type:{auth_type}"
+                    candidate_info.update({"skipped": True, "skip_reason": skip_reason})
+                    if record_id:
+                        self.db.execute(
+                            update(RequestCandidate)
+                            .where(RequestCandidate.id == record_id)
+                            .values(status="skipped", skip_reason=skip_reason)
+                        )
+                    continue
+
+                # billing rule filter
+                rule_lookup: BillingRuleLookupResult | None = None
+                has_billing_rule = True
+                if config.billing_require_rule:
+                    rule_lookup = BillingRuleService.find_rule(
+                        self.db,
+                        provider_id=cand.provider.id,
+                        model_name=model_name,
+                        task_type=task_type,
+                    )
+                    has_billing_rule = rule_lookup is not None
+                    if not has_billing_rule:
+                        skip_reason = "billing_rule_missing"
+                        candidate_info.update(
+                            {"has_billing_rule": False, "skipped": True, "skip_reason": skip_reason}
+                        )
+                        if record_id:
+                            self.db.execute(
+                                update(RequestCandidate)
+                                .where(RequestCandidate.id == record_id)
+                                .values(status="skipped", skip_reason=skip_reason)
+                            )
+                        continue
+                candidate_info["has_billing_rule"] = has_billing_rule
+
+                eligible_count += 1
+
+                # Mark pending
+                if record_id:
+                    self.db.execute(
+                        update(RequestCandidate)
+                        .where(RequestCandidate.id == record_id)
+                        .values(status="pending", started_at=now)
+                    )
+
+                # Flush/commit BEFORE awaiting upstream submit to avoid holding DB connections
+                # during potentially slow network operations.
+                if self.db.in_transaction():
+                    try:
+                        self.db.commit()
+                    except Exception:
+                        self.db.rollback()
+                        raise
+
+                # Attempt submit (upstream HTTP)
+                try:
+                    response: httpx.Response = await submit_func(cand)
+                except Exception as exc:
+                    finished_at = datetime.now(timezone.utc)
+                    error_msg = _sanitize(str(exc))
+                    candidate_info.update(
+                        {
+                            "attempt_status": "exception",
+                            "error_type": type(exc).__name__,
+                            "error_message": error_msg,
+                        }
+                    )
+                    if record_id:
+                        self.db.execute(
+                            update(RequestCandidate)
+                            .where(RequestCandidate.id == record_id)
+                            .values(
+                                status="failed",
+                                error_type=type(exc).__name__,
+                                error_message=error_msg,
+                                finished_at=finished_at,
+                            )
+                        )
+                    continue
+
+                last_status_code = int(getattr(response, "status_code", 0) or 0)
+
+                if response.status_code >= 400:
+                    finished_at = datetime.now(timezone.utc)
+                    try:
+                        error_text = response.text or ""
+                    except Exception:
+                        error_text = ""
+                    error_msg = _sanitize(error_text)
+                    candidate_info.update(
+                        {
+                            "attempt_status": "http_error",
+                            "status_code": response.status_code,
+                            "error_message": error_msg,
+                        }
+                    )
+                    if record_id:
+                        self.db.execute(
+                            update(RequestCandidate)
+                            .where(RequestCandidate.id == record_id)
+                            .values(
+                                status="failed",
+                                status_code=response.status_code,
+                                error_type="http_error",
+                                error_message=error_msg,
+                                finished_at=finished_at,
+                            )
+                        )
+
+                    if self._should_stop_on_http_error(
+                        status_code=response.status_code, error_text=error_text
+                    ):
+                        try:
+                            self.db.commit()
+                        except Exception:
+                            self.db.rollback()
+                        raise UpstreamClientRequestError(
+                            response=response,
+                            candidate_keys=candidate_keys,
+                        )
+                    continue
+
+                # Parse JSON
+                payload: dict[str, Any] | None = None
+                try:
+                    data = response.json()
+                    if isinstance(data, dict):
+                        payload = data
+                except Exception as exc:
+                    finished_at = datetime.now(timezone.utc)
+                    error_msg = _sanitize(str(exc))
+                    candidate_info.update(
+                        {
+                            "attempt_status": "invalid_json",
+                            "error_type": type(exc).__name__,
+                            "error_message": error_msg,
+                        }
+                    )
+                    if record_id:
+                        self.db.execute(
+                            update(RequestCandidate)
+                            .where(RequestCandidate.id == record_id)
+                            .values(
+                                status="failed",
+                                status_code=response.status_code,
+                                error_type="invalid_json",
+                                error_message=error_msg,
+                                finished_at=finished_at,
+                            )
+                        )
+                    continue
+
+                external_task_id = extract_external_task_id(payload or {})
+                if not external_task_id:
+                    finished_at = datetime.now(timezone.utc)
+                    candidate_info.update(
+                        {
+                            "attempt_status": "empty_task_id",
+                            "error_message": "Upstream returned empty task id",
+                        }
+                    )
+                    if record_id:
+                        self.db.execute(
+                            update(RequestCandidate)
+                            .where(RequestCandidate.id == record_id)
+                            .values(
+                                status="failed",
+                                status_code=response.status_code,
+                                error_type="empty_task_id",
+                                error_message="Upstream returned empty task id",
+                                finished_at=finished_at,
+                            )
+                        )
+                    continue
+
+                # Success
+                finished_at = datetime.now(timezone.utc)
+                candidate_info.update({"attempt_status": "success", "selected": True})
+                if record_id:
+                    self.db.execute(
+                        update(RequestCandidate)
+                        .where(RequestCandidate.id == record_id)
+                        .values(
+                            status="success",
+                            status_code=response.status_code,
+                            finished_at=finished_at,
+                        )
+                    )
+                try:
+                    self.db.commit()
+                except Exception:
+                    self.db.rollback()
+
+                return SubmitOutcome(
+                    candidate=cand,
+                    candidate_keys=candidate_keys,
+                    external_task_id=str(external_task_id),
+                    rule_lookup=rule_lookup,
+                    upstream_payload=payload,
+                    upstream_headers=dict(response.headers),
+                    upstream_status_code=response.status_code,
+                )
+
+            # Persist candidate records before raising
+            try:
+                self.db.commit()
+            except Exception:
+                self.db.rollback()
+
+            if eligible_count == 0:
+                reason = "no_eligible_candidates"
+                if config.billing_require_rule:
+                    reason = "no_candidate_with_billing_rule"
+                raise AllCandidatesFailedError(
+                    reason=reason,
+                    candidate_keys=candidate_keys,
+                    last_status_code=last_status_code,
+                )
+
+            raise AllCandidatesFailedError(
+                reason="all_candidates_failed",
+                candidate_keys=candidate_keys,
+                last_status_code=last_status_code,
+            )
+        finally:
+            # Restore Session behavior for the rest of the request lifecycle.
+            self.db.expire_on_commit = original_expire_on_commit
+
+    def create_candidate_records(
+        self,
+        *,
+        candidates: list[ProviderCandidate],
+        request_id: str,
+        user_api_key: ApiKey,
+        required_capabilities: dict[str, bool] | None = None,
+        expand_retries: bool = True,
+    ) -> dict[tuple[int, int], str]:
+        # CandidateResolver.create_candidate_records is currently the canonical implementation
+        assert self._resolver is not None, "Call resolve() once before create_candidate_records()"
+        return self._resolver.create_candidate_records(
+            all_candidates=candidates,
+            request_id=request_id,
+            user_id=str(user_api_key.user_id),
+            user_api_key=user_api_key,
+            required_capabilities=required_capabilities,
+            expand_retries=expand_retries,
+        )
+
+    def get_candidate_keys(self, request_id: str) -> list["CandidateKey"]:
+        return self._recorder.get_candidate_keys(request_id)

src/services/candidate/submit.py

@@ -0,0 +1,66 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any, Protocol, runtime_checkable
+
+import httpx
+
+from src.services.billing.rule_service import BillingRuleLookupResult
+from src.services.cache.aware_scheduler import ProviderCandidate
+
+
+@runtime_checkable
+class SubmitFunc(Protocol):
+    async def __call__(self, candidate: ProviderCandidate) -> httpx.Response: ...
+
+
+@runtime_checkable
+class ExtractExternalTaskIdFunc(Protocol):
+    def __call__(self, payload: dict[str, Any]) -> str | None: ...
+
+
+class UpstreamClientRequestError(RuntimeError):
+    """可判定为客户端请求问题（不应 failover）的上游错误。"""
+
+    def __init__(
+        self,
+        *,
+        response: httpx.Response,
+        candidate_keys: list[dict[str, Any]],
+    ) -> None:
+        self.response = response
+        self.candidate_keys = candidate_keys
+        super().__init__(f"Upstream client error: HTTP {response.status_code}")
+
+
+class AllCandidatesFailedError(RuntimeError):
+    def __init__(
+        self,
+        *,
+        reason: str,
+        candidate_keys: list[dict[str, Any]],
+        last_status_code: int | None = None,
+    ) -> None:
+        self.reason = reason
+        self.candidate_keys = candidate_keys
+        self.last_status_code = last_status_code
+        super().__init__(f"All candidates failed: {reason}")
+
+
+class CandidateUnsupportedError(RuntimeError):
+    """候选不被当前任务支持（如 auth_type/格式转换需求不支持）。"""
+
+
+class CandidateSubmissionError(RuntimeError):
+    """候选提交异常（网络/解密/解析等）。"""
+
+
+@dataclass(slots=True)
+class SubmitOutcome:
+    candidate: ProviderCandidate
+    candidate_keys: list[dict[str, Any]]
+    external_task_id: str
+    rule_lookup: BillingRuleLookupResult | None
+    upstream_payload: dict[str, Any] | None = None
+    upstream_headers: dict[str, str] | None = None
+    upstream_status_code: int | None = None

src/services/gemini_files_mapping.py

@@ -1,21 +1,41 @@
 """
-Gemini Files API - 文件与 Key 绑定缓存
+Gemini Files API - 文件与 Key 绑定映射服务
 
 用于在上传文件后记录 file_id -> provider_key_id，
 并在后续 generateContent 请求中优先使用同一 Key。
+
+存储策略：
+- 数据库（持久化）：主存储，支持服务重启后恢复
+- Redis（缓存）：加速读取，TTL=48小时
+
+读取策略：
+1. 先查 Redis 缓存
+2. 缓存未命中时回查数据库
+3. 从数据库读取后回填缓存
+
+清理策略：
+- 数据库中 expires_at 过期的记录由定时任务清理
+- Redis 缓存由 TTL 自动过期
 """
 
 from __future__ import annotations
 
-from typing import Any, Dict, Optional, Set
+import uuid
+from datetime import datetime, timedelta, timezone
+from typing import Any
+
+from sqlalchemy import delete
+from sqlalchemy.orm import Session
 
 from src.core.cache_service import CacheService
+from src.core.logger import logger
 
 FILE_MAPPING_TTL_SECONDS = 60 * 60 * 48  # 48小时
 FILE_MAPPING_CACHE_PREFIX = "gemini_files:key"
 
 
 def _normalize_file_name(file_name: str) -> str:
+    """规范化文件名，确保以 files/ 开头"""
     name = (file_name or "").strip()
     if not name:
         return ""
@@ -23,31 +43,294 @@ def _normalize_file_name(file_name: str) -> str:
 
 
 def build_file_mapping_key(file_name: str) -> str:
+    """构建 Redis 缓存键"""
     normalized = _normalize_file_name(file_name)
     return f"{FILE_MAPPING_CACHE_PREFIX}:{normalized}" if normalized else ""
 
 
-async def store_file_key_mapping(file_name: str, key_id: str) -> None:
-    cache_key = build_file_mapping_key(file_name)
-    if not cache_key or not key_id:
+# =============================================================================
+# 异步接口（用于请求处理流程）
+# =============================================================================
+
+
+async def store_file_key_mapping(
+    file_name: str,
+    key_id: str,
+    user_id: str | None = None,
+    display_name: str | None = None,
+    mime_type: str | None = None,
+    source_hash: str | None = None,
+) -> None:
+    """
+    存储文件→Key 映射（同时写入 Redis 和数据库）
+
+    Args:
+        file_name: 文件名（如 files/abc123）
+        key_id: Provider Key ID
+        user_id: 用户 ID（可选，用于权限验证）
+        display_name: 文件显示名（可选）
+        mime_type: 文件 MIME 类型（可选）
+        source_hash: 源文件哈希（可选，用于关联相同源文件的不同上传）
+    """
+    normalized_name = _normalize_file_name(file_name)
+    if not normalized_name or not key_id:
         return
+
+    # 1. 写入 Redis 缓存
+    cache_key = build_file_mapping_key(normalized_name)
     await CacheService.set(cache_key, str(key_id), ttl_seconds=FILE_MAPPING_TTL_SECONDS)
 
+    # 2. 写入数据库（异步执行，不阻塞主流程）
+    try:
+        await _store_to_database(
+            file_name=normalized_name,
+            key_id=key_id,
+            user_id=user_id,
+            display_name=display_name,
+            mime_type=mime_type,
+            source_hash=source_hash,
+        )
+    except Exception as e:
+        # 数据库写入失败只记录警告，不影响主流程
+        logger.warning(f"Failed to persist Gemini file mapping to database: {e}")
+
+
+async def _store_to_database(
+    file_name: str,
+    key_id: str,
+    user_id: str | None = None,
+    display_name: str | None = None,
+    mime_type: str | None = None,
+    source_hash: str | None = None,
+) -> None:
+    """将映射写入数据库"""
+    from src.database import get_db_context
+    from src.models.database import GeminiFileMapping
+
+    now = datetime.now(timezone.utc)
+    expires_at = now + timedelta(hours=48)
+
+    with get_db_context() as db:
+        # 使用 upsert 逻辑：存在则更新，不存在则插入
+        existing = (
+            db.query(GeminiFileMapping).filter(GeminiFileMapping.file_name == file_name).first()
+        )
+
+        if existing:
+            # 更新现有记录
+            existing.key_id = key_id
+            existing.user_id = user_id
+            existing.display_name = display_name
+            existing.mime_type = mime_type
+            existing.source_hash = source_hash
+            existing.expires_at = expires_at
+        else:
+            # 插入新记录
+            mapping = GeminiFileMapping(
+                id=str(uuid.uuid4()),
+                file_name=file_name,
+                key_id=key_id,
+                user_id=user_id,
+                display_name=display_name,
+                mime_type=mime_type,
+                source_hash=source_hash,
+                created_at=now,
+                expires_at=expires_at,
+            )
+            db.add(mapping)
+
+        db.commit()
+
 
 async def get_file_key_mapping(file_name: str) -> str | None:
-    cache_key = build_file_mapping_key(file_name)
-    if not cache_key:
+    """
+    获取文件→Key 映射
+
+    读取策略：
+    1. 先查 Redis 缓存
+    2. 缓存未命中时回查数据库
+    3. 从数据库读取后回填缓存
+
+    Args:
+        file_name: 文件名（如 files/abc123）
+
+    Returns:
+        Provider Key ID，如果不存在或已过期则返回 None
+    """
+    normalized_name = _normalize_file_name(file_name)
+    if not normalized_name:
         return None
-    value = await CacheService.get(cache_key)
-    if value:
-        return str(value)
+
+    cache_key = build_file_mapping_key(normalized_name)
+
+    # 1. 先查 Redis 缓存
+    cached_value = await CacheService.get(cache_key)
+    if cached_value:
+        return str(cached_value)
+
+    # 2. 缓存未命中，回查数据库
+    key_id = await _get_from_database(normalized_name)
+
+    if key_id:
+        # 3. 回填缓存（使用剩余有效期或默认 TTL）
+        await CacheService.set(cache_key, key_id, ttl_seconds=FILE_MAPPING_TTL_SECONDS)
+        logger.debug(f"Gemini file mapping cache refilled from database: {normalized_name}")
+
+    return key_id
+
+
+async def _get_from_database(file_name: str) -> str | None:
+    """从数据库查询映射"""
+    from src.database import get_db_context
+    from src.models.database import GeminiFileMapping
+
+    now = datetime.now(timezone.utc)
+
+    try:
+        with get_db_context() as db:
+            mapping = (
+                db.query(GeminiFileMapping)
+                .filter(
+                    GeminiFileMapping.file_name == file_name,
+                    GeminiFileMapping.expires_at > now,  # 只返回未过期的
+                )
+                .first()
+            )
+
+            if mapping:
+                return str(mapping.key_id)
+    except Exception as e:
+        logger.warning(f"Failed to query Gemini file mapping from database: {e}")
+
     return None
 
 
+async def get_all_key_ids_for_file(file_name: str) -> list[str]:
+    """
+    获取支持指定文件的所有 Key ID 列表
+
+    当同一个源文件被上传到多个 Key 时，返回所有可用的 Key ID。
+    这允许系统在首选 Key 不可用时选择其他 Key。
+
+    Args:
+        file_name: 文件名（如 files/abc123）
+
+    Returns:
+        所有支持该文件的 Key ID 列表（包括原始映射和具有相同 source_hash 的映射）
+    """
+    from src.database import get_db_context
+    from src.models.database import GeminiFileMapping
+
+    normalized_name = _normalize_file_name(file_name)
+    if not normalized_name:
+        return []
+
+    now = datetime.now(timezone.utc)
+
+    try:
+        with get_db_context() as db:
+            # 首先获取原始映射
+            original_mapping = (
+                db.query(GeminiFileMapping)
+                .filter(
+                    GeminiFileMapping.file_name == normalized_name,
+                    GeminiFileMapping.expires_at > now,
+                )
+                .first()
+            )
+
+            if not original_mapping:
+                return []
+
+            key_ids = [str(original_mapping.key_id)]
+
+            # 如果有 source_hash，查找所有具有相同 source_hash 的映射
+            if original_mapping.source_hash:
+                related_mappings = (
+                    db.query(GeminiFileMapping)
+                    .filter(
+                        GeminiFileMapping.source_hash == original_mapping.source_hash,
+                        GeminiFileMapping.expires_at > now,
+                        GeminiFileMapping.file_name != normalized_name,  # 排除原始映射
+                    )
+                    .all()
+                )
+
+                for mapping in related_mappings:
+                    kid = str(mapping.key_id)
+                    if kid not in key_ids:
+                        key_ids.append(kid)
+
+            return key_ids
+    except Exception as e:
+        logger.warning(f"Failed to query related Gemini file mappings: {e}")
+        return []
+
+
 async def delete_file_key_mapping(file_name: str) -> None:
-    cache_key = build_file_mapping_key(file_name)
-    if cache_key:
-        await CacheService.delete(cache_key)
+    """
+    删除文件→Key 映射（同时从 Redis 和数据库删除）
+
+    Args:
+        file_name: 文件名（如 files/abc123）
+    """
+    normalized_name = _normalize_file_name(file_name)
+    if not normalized_name:
+        return
+
+    # 1. 从 Redis 删除
+    cache_key = build_file_mapping_key(normalized_name)
+    await CacheService.delete(cache_key)
+
+    # 2. 从数据库删除
+    try:
+        await _delete_from_database(normalized_name)
+    except Exception as e:
+        logger.warning(f"Failed to delete Gemini file mapping from database: {e}")
+
+
+async def _delete_from_database(file_name: str) -> None:
+    """从数据库删除映射"""
+    from src.database import get_db_context
+    from src.models.database import GeminiFileMapping
+
+    with get_db_context() as db:
+        db.execute(delete(GeminiFileMapping).where(GeminiFileMapping.file_name == file_name))
+        db.commit()
+
+
+# =============================================================================
+# 同步接口（用于定时任务等场景）
+# =============================================================================
+
+
+def cleanup_expired_mappings(db: Session) -> int:
+    """
+    清理过期的文件映射记录（同步方法，供定时任务调用）
+
+    Args:
+        db: 数据库会话
+
+    Returns:
+        删除的记录数
+    """
+    from src.models.database import GeminiFileMapping
+
+    now = datetime.now(timezone.utc)
+
+    result = db.execute(delete(GeminiFileMapping).where(GeminiFileMapping.expires_at <= now))
+    db.commit()
+
+    deleted_count = result.rowcount
+    if deleted_count > 0:
+        logger.info(f"Cleaned up {deleted_count} expired Gemini file mappings")
+
+    return deleted_count
+
+
+# =============================================================================
+# 请求解析工具函数
+# =============================================================================
 
 
 def _extract_file_name_from_uri(file_uri: str) -> str | None:

src/services/model/fetch_scheduler.py

@@ -39,6 +39,10 @@ MAX_CONCURRENT_REQUESTS = 5
 # 单个 Key 处理的超时时间（秒）
 KEY_FETCH_TIMEOUT_SECONDS = 120
 
+# 模型获取 HTTP 请求超时时间（秒）
+# 使用较短的超时（10秒），避免不支持 /models 端点的提供商长时间阻塞
+MODEL_FETCH_HTTP_TIMEOUT = 10.0
+
 # 上游模型缓存 TTL（与定时任务间隔保持一致）
 UPSTREAM_MODELS_CACHE_TTL_SECONDS = MODEL_FETCH_INTERVAL_MINUTES * 60
 
@@ -260,7 +264,48 @@ class ModelFetchScheduler:
             logger.exception(f"更新 Key {key_id} 错误信息失败")
 
     async def _fetch_models_for_key_by_id(self, key_id: str) -> str:
-        """根据 Key ID 获取模型并更新，返回结果状态"""
+        """
+        根据 Key ID 获取模型并更新，返回结果状态
+
+        优化：分两个阶段处理，HTTP 请求期间不持有数据库连接，避免阻塞其他请求
+        """
+        # ========== 阶段 1：准备数据（短暂持有连接）==========
+        fetch_context = self._prepare_fetch_context(key_id)
+        if fetch_context is None:
+            return "skip"
+        if isinstance(fetch_context, str):
+            return fetch_context  # "error" or "skip"
+
+        key_id, provider_id, provider_name, api_key_value, endpoint_configs = fetch_context
+
+        # ========== 阶段 2：HTTP 请求（不持有数据库连接）==========
+        # 使用较短的超时时间（10秒），避免长时间阻塞
+        all_models, errors, has_success = await fetch_models_from_endpoints(
+            endpoint_configs, timeout=MODEL_FETCH_HTTP_TIMEOUT
+        )
+
+        # ========== 阶段 3：更新数据库（获取新连接）==========
+        return await self._update_key_after_fetch(
+            key_id=key_id,
+            provider_id=provider_id,
+            provider_name=provider_name,
+            all_models=all_models,
+            errors=errors,
+            has_success=has_success,
+        )
+
+    def _prepare_fetch_context(
+        self, key_id: str
+    ) -> tuple[str, str, str, str, list[dict]] | str | None:
+        """
+        准备获取模型所需的上下文数据
+
+        Returns:
+            - tuple: (key_id, provider_id, provider_name, api_key_value, endpoint_configs)
+            - "skip": 跳过该 Key
+            - "error": 出错
+            - None: Key 不存在
+        """
         with create_session() as db:
             key = (
                 db.query(ProviderAPIKey)
@@ -271,142 +316,154 @@ class ModelFetchScheduler:
 
             if not key:
                 logger.warning(f"Key {key_id} 不存在，跳过")
-                return "skip"
+                return None
 
             if not key.is_active or not key.auto_fetch_models:
                 logger.debug(f"Key {key_id} 已禁用或关闭自动获取，跳过")
                 return "skip"
 
-            try:
-                result = await self._fetch_models_for_key(db, key)
+            now = datetime.now(timezone.utc)
+            provider_id = key.provider_id
+
+            # 获取 Provider 和 Endpoints
+            provider = (
+                db.query(Provider)
+                .options(joinedload(Provider.endpoints))
+                .filter(Provider.id == provider_id)
+                .first()
+            )
+
+            if not provider:
+                logger.warning(f"Provider {provider_id} 不存在，跳过 Key {key.id}")
+                key.last_models_fetch_error = "Provider not found"
+                key.last_models_fetch_at = now
                 db.commit()
-                return result
+                return "error"
+
+            # Vertex AI 类型不支持自动获取模型
+            auth_type = getattr(key, "auth_type", "api_key") or "api_key"
+            if auth_type == "vertex_ai":
+                key.last_models_fetch_error = "auto_fetch_models 暂不支持 Vertex AI 类型的 Key"
+                key.last_models_fetch_at = now
+                db.commit()
+                logger.info(f"Key {key.id} 为 Vertex AI 类型，跳过自动获取模型")
+                return "skip"
+
+            # 解密 API Key
+            if not key.api_key:
+                logger.warning(f"Key {key.id} 没有 API Key，跳过")
+                key.last_models_fetch_error = "No API key configured"
+                key.last_models_fetch_at = now
+                db.commit()
+                return "error"
+
+            try:
+                api_key_value = crypto_service.decrypt(key.api_key)
             except Exception:
-                db.rollback()
-                raise
+                logger.error(f"解密 Key {key.id} 失败")
+                key.last_models_fetch_error = "Decrypt error"
+                key.last_models_fetch_at = now
+                db.commit()
+                return "error"
 
-    async def _fetch_models_for_key(
+            # 构建 api_format -> endpoint 映射
+            format_to_endpoint: dict[str, object] = {}
+            for endpoint in provider.endpoints:  # type: ignore[attr-defined]
+                if endpoint.is_active:
+                    format_to_endpoint[endpoint.api_format] = endpoint
+
+            if not format_to_endpoint:
+                logger.warning(f"Provider {provider.name} 没有活跃的端点，跳过 Key {key.id}")
+                key.last_models_fetch_error = "No active endpoints"
+                key.last_models_fetch_at = now
+                db.commit()
+                return "error"
+
+            # 使用公共函数构建所有格式的端点配置
+            endpoint_configs = build_all_format_configs(api_key_value, format_to_endpoint)  # type: ignore[arg-type]
+
+            return (key_id, provider_id, provider.name, api_key_value, endpoint_configs)
+
+    async def _update_key_after_fetch(
         self,
-        db: Session,
-        key: ProviderAPIKey,
+        key_id: str,
+        provider_id: str,
+        provider_name: str,
+        all_models: list[dict],
+        errors: list[str],
+        has_success: bool,
     ) -> str:
-        """为单个 Key 获取模型并更新 allowed_models，返回结果状态"""
+        """
+        HTTP 请求完成后更新数据库
+
+        使用新的数据库连接来更新 Key 的 allowed_models
+        """
         now = datetime.now(timezone.utc)
-        provider_id = key.provider_id
 
-        # 获取 Provider 和 Endpoints
-        provider = (
-            db.query(Provider)
-            .options(joinedload(Provider.endpoints))
-            .filter(Provider.id == provider_id)
-            .first()
-        )
+        with create_session() as db:
+            # 重新获取 Key（因为之前的连接已关闭）
+            key = db.query(ProviderAPIKey).filter(ProviderAPIKey.id == key_id).first()
+            if not key:
+                logger.warning(f"Key {key_id} 在更新时不存在")
+                return "error"
 
-        if not provider:
-            logger.warning(f"Provider {provider_id} 不存在，跳过 Key {key.id}")
-            key.last_models_fetch_error = "Provider not found"
+            # 记录获取时间
             key.last_models_fetch_at = now
-            return "error"
 
-        # Vertex AI 类型不支持自动获取模型（需要使用 Service Account 认证）
-        auth_type = getattr(key, "auth_type", "api_key") or "api_key"
-        if auth_type == "vertex_ai":
-            key.last_models_fetch_error = "auto_fetch_models 暂不支持 Vertex AI 类型的 Key"
-            key.last_models_fetch_at = now
-            logger.info(f"Key {key.id} 为 Vertex AI 类型，跳过自动获取模型")
-            return "skip"
+            # 如果没有任何成功的响应，不更新 allowed_models（保留旧数据）
+            if not has_success:
+                error_msg = "; ".join(errors) if errors else "All endpoints failed"
+                key.last_models_fetch_error = error_msg
+                logger.warning(
+                    f"Provider {provider_name} Key {key.id} 所有端点获取失败，保留现有模型列表"
+                )
+                db.commit()
+                return "error"
 
-        # 解密 API Key
-        if not key.api_key:
-            logger.warning(f"Key {key.id} 没有 API Key，跳过")
-            key.last_models_fetch_error = "No API key configured"
-            key.last_models_fetch_at = now
-            return "error"
+            # 有成功的响应，清除错误状态
+            key.last_models_fetch_error = None
 
-        try:
-            api_key_value = crypto_service.decrypt(key.api_key)
-        except Exception:
-            # 不记录异常详情，避免泄露密钥信息
-            logger.error(f"解密 Key {key.id} 失败")
-            key.last_models_fetch_error = "Decrypt error"
-            key.last_models_fetch_at = now
-            return "error"
+            # 去重获取模型 ID 列表
+            fetched_model_ids: set[str] = set()
+            for model in all_models:
+                model_id = model.get("id")
+                if model_id:
+                    fetched_model_ids.add(model_id)
 
-        # 构建 api_format -> endpoint 映射
-        format_to_endpoint: dict[str, object] = {}
-        for endpoint in provider.endpoints:  # type: ignore[attr-defined]
-            if endpoint.is_active:
-                format_to_endpoint[endpoint.api_format] = endpoint
-
-        if not format_to_endpoint:
-            logger.warning(f"Provider {provider.name} 没有活跃的端点，跳过 Key {key.id}")
-            key.last_models_fetch_error = "No active endpoints"
-            key.last_models_fetch_at = now
-            return "error"
-
-        # 使用公共函数构建所有格式的端点配置
-        endpoint_configs = build_all_format_configs(api_key_value, format_to_endpoint)  # type: ignore[arg-type]
-
-        # 并发获取模型
-        all_models, errors, has_success = await fetch_models_from_endpoints(endpoint_configs)
-
-        # 记录获取时间
-        key.last_models_fetch_at = now
-
-        # 如果没有任何成功的响应，不更新 allowed_models（保留旧数据）
-        if not has_success:
-            # 所有端点都失败时，记录错误
-            error_msg = "; ".join(errors) if errors else "All endpoints failed"
-            key.last_models_fetch_error = error_msg
-            logger.warning(
-                f"Provider {provider.name} Key {key.id} 所有端点获取失败，保留现有模型列表"
-            )
-            return "error"
-
-        # 有成功的响应，清除错误状态（部分失败不算失败）
-        key.last_models_fetch_error = None
-
-        # 去重获取模型 ID 列表
-        fetched_model_ids: set[str] = set()
-        for model in all_models:
-            model_id = model.get("id")
-            if model_id:
-                fetched_model_ids.add(model_id)
-
-        logger.info(
-            f"Provider {provider.name} Key {key.id} 获取到 {len(fetched_model_ids)} 个唯一模型"
-        )
-
-        # 写入上游模型缓存（按 model id + api_format 去重后的完整模型信息）
-        seen_keys: set[str] = set()
-        unique_models: list[dict] = []
-        for model in all_models:
-            model_id = model.get("id")
-            api_format = model.get("api_format", "")
-            unique_key = f"{model_id}:{api_format}"
-            if model_id and unique_key not in seen_keys:
-                seen_keys.add(unique_key)
-                unique_models.append(model)
-        await set_upstream_models_to_cache(
-            provider_id,  # type: ignore[arg-type]
-            key.id,  # type: ignore[arg-type]
-            unique_models,
-        )
-
-        # 更新 allowed_models（保留 locked_models）
-        has_changed = self._update_key_allowed_models(key, fetched_model_ids)
-
-        # 如果白名单有变化，触发缓存失效和自动关联检查
-        if has_changed and provider_id:
-            from src.services.model.global_model import on_key_allowed_models_changed
-
-            await on_key_allowed_models_changed(
-                db=db,
-                provider_id=provider_id,
-                allowed_models=list(key.allowed_models or []),
+            logger.info(
+                f"Provider {provider_name} Key {key.id} 获取到 {len(fetched_model_ids)} 个唯一模型"
             )
 
-        return "success"
+            # 写入上游模型缓存（按 model id + api_format 去重后的完整模型信息）
+            seen_keys: set[str] = set()
+            unique_models: list[dict] = []
+            for model in all_models:
+                model_id = model.get("id")
+                api_format = model.get("api_format", "")
+                unique_key = f"{model_id}:{api_format}"
+                if model_id and unique_key not in seen_keys:
+                    seen_keys.add(unique_key)
+                    unique_models.append(model)
+            await set_upstream_models_to_cache(provider_id, key.id, unique_models)
+
+            # 更新 allowed_models（保留 locked_models）
+            has_changed = self._update_key_allowed_models(key, fetched_model_ids)
+
+            db.commit()
+
+            # 如果白名单有变化，触发缓存失效和自动关联检查
+            if has_changed and provider_id:
+                from src.services.model.global_model import on_key_allowed_models_changed
+
+                # 使用新会话处理后续操作
+                with create_session() as db2:
+                    await on_key_allowed_models_changed(
+                        db=db2,
+                        provider_id=provider_id,
+                        allowed_models=list(key.allowed_models or []),
+                    )
+
+            return "success"
 
     def _update_key_allowed_models(self, key: ProviderAPIKey, fetched_model_ids: set[str]) -> bool:
         """

src/services/orchestration/candidate_resolver.py

@@ -156,6 +156,8 @@ class CandidateResolver:
         user_id: str,
         user_api_key: ApiKey,
         required_capabilities: dict[str, bool] | None = None,
+        *,
+        expand_retries: bool = True,
     ) -> dict[tuple[int, int], str]:
         """
         为所有候选预先创建 available 状态记录（批量插入优化）
@@ -211,9 +213,12 @@ class CandidateResolver:
                 candidate_record_map[(candidate_index, 0)] = record_id
             else:
                 # max_retries 已从 Endpoint 迁移到 Provider（Endpoint 仍可能保留旧字段用于兼容）
-                max_retries_for_candidate = (
-                    int(provider.max_retries or 2) if candidate.is_cached else 1
-                )
+                if not expand_retries:
+                    max_retries_for_candidate = 1
+                else:
+                    max_retries_for_candidate = (
+                        int(provider.max_retries or 2) if candidate.is_cached else 1
+                    )
 
                 for retry_index in range(max_retries_for_candidate):
                     record_id = str(uuid.uuid4())

src/services/provider_ops/service.py

@@ -38,6 +38,19 @@ BALANCE_CACHE_TTL = 86400
 # 认证失败缓存 TTL（60 秒，避免频繁重试但允许用户修正后快速重试）
 AUTH_FAILED_CACHE_TTL = 60
 
+# 后台余额刷新并发限制（避免启动时耗尽连接池）
+# 使用较小的值（3）确保不会对连接池造成过大压力
+_balance_refresh_semaphore: asyncio.Semaphore | None = None
+
+
+def _get_balance_refresh_semaphore() -> asyncio.Semaphore:
+    """获取余额刷新信号量（延迟初始化）"""
+    global _balance_refresh_semaphore
+    if _balance_refresh_semaphore is None:
+        # 限制为 3 个并发，确保后台任务不会占用太多连接
+        _balance_refresh_semaphore = asyncio.Semaphore(3)
+    return _balance_refresh_semaphore
+
 
 def _get_batch_balance_concurrency() -> int:
     """
@@ -98,6 +111,44 @@ class ProviderOpsService:
         # 连接器缓存 {provider_id: ProviderConnector}
         self._connectors: dict[str, ProviderConnector] = {}
 
+    def _release_db_connection_before_await(self) -> None:
+        """
+        Release pooled DB connection before long awaits (network/Redis).
+
+        SQLAlchemy Session will keep a connection checked out while a transaction is open,
+        even for read-only queries. In async code, this can exhaust the pool if we `await`
+        network I/O while holding that transaction.
+
+        Safety:
+        - Only commits when the session has no pending changes (new/dirty/deleted).
+        - Temporarily disables expire_on_commit to avoid unexpected lazy reloads.
+        """
+        try:
+            has_pending_changes = bool(self.db.new) or bool(self.db.dirty) or bool(self.db.deleted)
+        except Exception:
+            has_pending_changes = False
+
+        if has_pending_changes:
+            return
+
+        try:
+            if not self.db.in_transaction():
+                return
+        except Exception:
+            return
+
+        original_expire_on_commit = getattr(self.db, "expire_on_commit", True)
+        self.db.expire_on_commit = False
+        try:
+            self.db.commit()
+        except Exception:
+            try:
+                self.db.rollback()
+            except Exception:
+                pass
+        finally:
+            self.db.expire_on_commit = original_expire_on_commit
+
     # ==================== 配置管理 ====================
 
     def get_config(self, provider_id: str) -> ProviderOpsConfig | None:
@@ -251,6 +302,9 @@ class ProviderOpsService:
         if not actual_credentials:
             return False, "未提供凭据"
 
+        # Avoid holding a DB connection while awaiting network I/O.
+        self._release_db_connection_before_await()
+
         # 建立连接
         logger.info(
             f"尝试连接: provider_id={provider_id}, "
@@ -333,6 +387,8 @@ class ProviderOpsService:
                 )
             connector = self._connectors.get(provider_id)
 
+        # Avoid holding a DB connection while awaiting authentication checks.
+        self._release_db_connection_before_await()
         if not connector or not await connector.is_authenticated():
             return ActionResult(
                 status=ActionStatus.AUTH_EXPIRED,
@@ -373,6 +429,9 @@ class ProviderOpsService:
         # 创建操作实例
         action = architecture.get_action(action_type, merged_config)
 
+        # Avoid holding a DB connection while awaiting the upstream action.
+        self._release_db_connection_before_await()
+
         # 执行操作
         async with connector.get_client() as client:
             result = await action.execute(client)
@@ -422,6 +481,9 @@ class ProviderOpsService:
         Returns:
             操作结果（可能是缓存的）
         """
+        # Avoid holding a DB connection while awaiting Redis/cache I/O.
+        self._release_db_connection_before_await()
+
         # 尝试从缓存获取
         cached = await self._get_cached_balance(provider_id)
 
@@ -453,7 +515,20 @@ class ProviderOpsService:
 
         注意：这是一个后台任务，使用独立的短生命周期 session，
         避免长时间占用连接池资源。
+
+        使用信号量限制并发数，避免启动时多个刷新任务同时运行导致连接池耗尽。
         """
+        semaphore = _get_balance_refresh_semaphore()
+
+        # 尝试获取信号量，如果无法立即获取则跳过本次刷新
+        # 这样可以避免在连接池紧张时阻塞
+        try:
+            # 使用 wait_for 设置超时，避免无限等待
+            await asyncio.wait_for(semaphore.acquire(), timeout=5.0)
+        except asyncio.TimeoutError:
+            logger.debug(f"异步刷新余额跳过（并发限制）: provider_id={provider_id}")
+            return
+
         db = None
         try:
             # 后台任务需要创建独立的 session，因为原请求的 session 可能已关闭
@@ -469,6 +544,8 @@ class ProviderOpsService:
                     db.close()
                 except Exception:
                     pass
+            # 释放信号量
+            semaphore.release()
 
     async def _clear_balance_cache(self, provider_id: str) -> None:
         """清除余额缓存"""
@@ -762,6 +839,9 @@ class ProviderOpsService:
         if not provider_ids:
             return {}
 
+        # Release the DB connection before awaiting many async cache refreshes.
+        self._release_db_connection_before_await()
+
         # 使用信号量限制并发数，避免同时发起过多请求耗尽连接池
         concurrency = _get_batch_balance_concurrency()
         semaphore = asyncio.Semaphore(concurrency)
@@ -822,6 +902,9 @@ class ProviderOpsService:
 
         from src.utils.ssl_utils import get_ssl_context
 
+        # Avoid holding a DB connection while awaiting verify pre-processing / network.
+        self._release_db_connection_before_await()
+
         # 移除 base_url 末尾的斜杠
         base_url = base_url.rstrip("/")
 

src/services/system/config.py

@@ -136,6 +136,11 @@ class SystemConfigService:
             "value": [],
             "description": "邮箱后缀列表，配合 email_suffix_mode 使用",
         },
+        # 格式转换开关
+        "enable_format_conversion": {
+            "value": False,
+            "description": "全局格式转换开关：开启时强制允许所有提供商的格式转换；关闭时由各提供商自行决定",
+        },
         "audit_log_retention_days": {
             "value": 30,
             "description": "审计日志保留天数，超过此天数的审计日志将被自动清理",
@@ -358,6 +363,11 @@ class SystemConfigService:
         """获取敏感请求头列表"""
         return cls.get_config(db, "sensitive_headers", [])
 
+    @classmethod
+    def is_format_conversion_enabled(cls, db: Session) -> bool:
+        """检查全局格式转换是否启用"""
+        return bool(cls.get_config(db, "enable_format_conversion", True))
+
     @classmethod
     def mask_sensitive_headers(cls, db: Session, headers: dict[str, Any]) -> dict[str, Any]:
         """脱敏敏感请求头"""

src/services/system/maintenance_scheduler.py

@@ -8,6 +8,7 @@
 - 审计日志清理：定期清理过期的审计日志
 - 连接池监控：定期检查数据库连接池状态
 - Pending 状态清理：清理异常的 Pending 状态记录
+- Gemini 文件映射清理：清理过期的 Gemini 文件→Key 映射
 
 使用 APScheduler 进行任务调度，支持时区配置。
 """
@@ -103,6 +104,14 @@ class MaintenanceScheduler:
             name="审计日志清理",
         )
 
+        # Gemini 文件映射清理 - 每小时执行
+        scheduler.add_interval_job(
+            self._scheduled_gemini_file_mapping_cleanup,
+            hours=1,
+            job_id="gemini_file_mapping_cleanup",
+            name="Gemini文件映射清理",
+        )
+
         # Provider 签到任务 - 凌晨 1:05 执行
         scheduler.add_cron_job(
             self._scheduled_provider_checkin,
@@ -117,8 +126,9 @@ class MaintenanceScheduler:
 
     async def _run_startup_tasks(self) -> None:
         """启动时执行的初始化任务"""
-        # 延迟一点执行，确保系统完全启动
-        await asyncio.sleep(2)
+        # 延迟执行，等待系统完全启动（Redis 连接、其他后台任务稳定）
+        # 增加延迟时间避免与 UsageQueueConsumer 等后台任务竞争数据库连接
+        await asyncio.sleep(10)
 
         try:
             logger.info("启动时执行首次清理任务...")
@@ -170,6 +180,10 @@ class MaintenanceScheduler:
         """审计日志清理任务（定时调用）"""
         await self._perform_audit_cleanup()
 
+    async def _scheduled_gemini_file_mapping_cleanup(self) -> None:
+        """Gemini 文件映射清理任务（定时调用）"""
+        await self._perform_gemini_file_mapping_cleanup()
+
     async def _scheduled_provider_checkin(self) -> None:
         """Provider 签到任务（定时调用）"""
         await self._perform_provider_checkin()
@@ -483,6 +497,26 @@ class MaintenanceScheduler:
         finally:
             db.close()
 
+    async def _perform_gemini_file_mapping_cleanup(self) -> None:
+        """清理过期的 Gemini 文件映射记录"""
+        db = create_session()
+        try:
+            from src.services.gemini_files_mapping import cleanup_expired_mappings
+
+            deleted_count = cleanup_expired_mappings(db)
+
+            if deleted_count > 0:
+                logger.info(f"清理了 {deleted_count} 条过期的 Gemini 文件映射")
+
+        except Exception as e:
+            logger.exception(f"Gemini 文件映射清理失败: {e}")
+            try:
+                db.rollback()
+            except Exception:
+                pass
+        finally:
+            db.close()
+
     async def _perform_provider_checkin(self) -> None:
         """执行 Provider 签到任务
 
@@ -508,8 +542,21 @@ class MaintenanceScheduler:
 
             logger.info(f"开始执行 Provider 签到，共 {len(provider_ids)} 个...")
 
-            # 创建 ProviderOpsService 并执行批量余额查询（会触发签到）
-            service = ProviderOpsService(db)
+            # 释放主 session 的连接，避免在整个签到期间占用连接池
+            # （后续每个 provider 将使用独立短生命周期 session）
+            try:
+                if db.in_transaction():
+                    db.commit()
+            except Exception:
+                try:
+                    db.rollback()
+                except Exception:
+                    pass
+            try:
+                db.close()
+            except Exception:
+                pass
+            db = None
 
             # 使用信号量限制并发，避免同时发起过多请求
             concurrency = 3  # 签到任务并发数
@@ -518,7 +565,9 @@ class MaintenanceScheduler:
             async def _checkin_provider(provider_id: str) -> tuple[str, bool, str]:
                 """执行单个 Provider 的签到"""
                 async with semaphore:
+                    task_db = create_session()
                     try:
+                        service = ProviderOpsService(task_db)
                         # 触发余额查询（会先执行签到）
                         result = await service.query_balance(provider_id)
                         # 检查签到结果
@@ -537,6 +586,11 @@ class MaintenanceScheduler:
                     except Exception as e:
                         logger.warning(f"Provider {provider_id} 签到失败: {e}")
                         return provider_id, False, str(e)
+                    finally:
+                        try:
+                            task_db.close()
+                        except Exception:
+                            pass
 
             # 并行执行签到
             tasks = [_checkin_provider(pid) for pid in provider_ids]
@@ -556,7 +610,8 @@ class MaintenanceScheduler:
         except Exception as e:
             logger.exception(f"Provider 签到任务执行失败: {e}")
         finally:
-            db.close()
+            if db is not None:
+                db.close()
 
     async def _perform_cleanup(self) -> None:
         """执行清理任务"""

src/services/task/__init__.py

@@ -1,25 +1,13 @@
 """
-异步任务服务层
+任务服务层（Phase2）
 
-提供视频/图片/音频等异步任务的：
-- 提交阶段故障转移（AsyncTaskOrchestrator）
-- 终态计费与 Usage 写入（VideoTelemetry 等）
+统一任务框架相关的应用层入口：
+- 候选提交阶段：`services.candidate.CandidateService`
+- 终态结算：`services.task.application.TaskApplicationService`
 """
 
-from .orchestrator import (
-    AllCandidatesFailedError,
-    AsyncTaskOrchestrator,
-    CandidateSubmissionError,
-    CandidateUnsupportedError,
-    SubmitOutcome,
-    UpstreamClientRequestError,
-)
+from .application import TaskApplicationService
 
 __all__ = [
-    "AsyncTaskOrchestrator",
-    "SubmitOutcome",
-    "AllCandidatesFailedError",
-    "UpstreamClientRequestError",
-    "CandidateUnsupportedError",
-    "CandidateSubmissionError",
+    "TaskApplicationService",
 ]

src/services/task/application.py

@@ -0,0 +1,312 @@
+from __future__ import annotations
+
+from datetime import datetime, timezone
+from typing import Any
+
+from sqlalchemy.orm import Session
+
+from src.config.settings import config
+from src.core.logger import logger
+from src.models.database import ApiKey, Provider, Usage, User, VideoTask
+from src.services.billing.dimension_collector_service import DimensionCollectorService
+from src.services.billing.formula_engine import BillingIncompleteError, FormulaEngine
+from src.services.billing.rule_service import BillingRuleService
+from src.services.usage.service import UsageService
+
+
+class TaskApplicationService:
+    """
+    TaskApplicationService (Phase2)
+
+    当前仅先收敛"终态结算"入口，用于替代旧版 VideoTelemetry 直写 Usage 的流程。
+    后续将扩展 submit/cancel 并迁移候选编排逻辑。
+    """
+
+    def __init__(self, db: Session, *, redis_client: Any | None = None) -> None:
+        self.db = db
+        self.redis = redis_client
+
+    async def finalize_video_task(self, task: VideoTask) -> bool:
+        """
+        更新视频任务的计费信息（轮询完成后调用）。
+
+        异步任务的计费流程：
+        1. 提交成功时：Usage 已结算（billing_status='settled'，费用=0）
+        2. 轮询完成时：更新实际费用（成功则计费，失败则保持0）
+
+        返回 True 表示成功更新，False 表示无需更新（如已是最终状态）
+        """
+        request_id = getattr(task, "request_id", None) or task.id
+
+        existing = self.db.query(Usage).filter(Usage.request_id == request_id).first()
+        if not existing:
+            # Usage 不存在，尝试创建并结算（兜底逻辑）
+            logger.warning(
+                "Usage not found for video task, creating fallback: task_id=%s request_id=%s",
+                task.id,
+                request_id,
+            )
+            return await self._create_fallback_usage(task, request_id)
+
+        # 检查是否已有计费更新标记（避免重复计费）
+        metadata = existing.request_metadata or {}
+        if metadata.get("billing_updated_at"):
+            logger.debug(
+                "Video task billing already updated: task_id=%s request_id=%s",
+                task.id,
+                request_id,
+            )
+            return False
+
+        # 计算异步任务总耗时（ms）
+        response_time_ms: int | None = None
+        if task.submitted_at and task.completed_at:
+            delta = task.completed_at - task.submitted_at
+            response_time_ms = int(delta.total_seconds() * 1000)
+
+        # === 收集计费维度 ===
+        base_dimensions: dict[str, Any] = {
+            "duration_seconds": task.duration_seconds,
+            "resolution": task.resolution,
+            "aspect_ratio": task.aspect_ratio,
+            "size": task.size or "",
+            "retry_count": task.retry_count,
+        }
+
+        collector_metadata: dict[str, Any] = {
+            "task": {
+                "id": task.id,
+                "external_task_id": task.external_task_id,
+                "model": task.model,
+                "duration_seconds": task.duration_seconds,
+                "resolution": task.resolution,
+                "aspect_ratio": task.aspect_ratio,
+                "size": task.size,
+                "retry_count": task.retry_count,
+                "video_size_bytes": task.video_size_bytes,
+            },
+            "result": {
+                "video_url": task.video_url,
+                "video_urls": task.video_urls or [],
+            },
+        }
+
+        dims = DimensionCollectorService(self.db).collect_dimensions(
+            api_format=task.provider_api_format,
+            task_type="video",
+            request=task.original_request_body or {},
+            response=(
+                (task.request_metadata or {}).get("poll_raw_response")
+                if isinstance(task.request_metadata, dict)
+                else None
+            ),
+            metadata=collector_metadata,
+            base_dimensions=base_dimensions,
+        )
+
+        # === 计算成本（优先使用冻结的 billing_rule_snapshot）===
+        rule_snapshot = None
+        if isinstance(task.request_metadata, dict):
+            rule_snapshot = task.request_metadata.get("billing_rule_snapshot")
+
+        expression = None
+        variables: dict[str, Any] | None = None
+        dimension_mappings: dict[str, dict[str, Any]] | None = None
+        rule_id = None
+        rule_name = None
+        rule_scope = None
+
+        if isinstance(rule_snapshot, dict) and rule_snapshot.get("status") == "ok":
+            rule_id = rule_snapshot.get("rule_id")
+            rule_name = rule_snapshot.get("rule_name")
+            rule_scope = rule_snapshot.get("scope")
+            expression = rule_snapshot.get("expression")
+            variables = rule_snapshot.get("variables") or {}
+            dimension_mappings = rule_snapshot.get("dimension_mappings") or {}
+        else:
+            lookup = BillingRuleService.find_rule(
+                self.db,
+                provider_id=task.provider_id,
+                model_name=task.model,
+                task_type="video",
+            )
+            if lookup:
+                rule = lookup.rule
+                rule_id = rule.id
+                rule_name = rule.name
+                rule_scope = getattr(lookup, "scope", None)
+                expression = rule.expression
+                variables = rule.variables or {}
+                dimension_mappings = rule.dimension_mappings or {}
+
+        billing_snapshot: dict[str, Any] = {
+            "schema_version": "1.0",
+            "rule_id": str(rule_id) if rule_id else None,
+            "rule_name": str(rule_name) if rule_name else None,
+            "scope": str(rule_scope) if rule_scope else None,
+            "expression": str(expression) if expression else None,
+            "dimensions_used": dims,
+            "missing_required": [],
+            "cost": 0.0,
+            "status": "no_rule",
+            "calculated_at": datetime.now(timezone.utc).isoformat(),
+        }
+
+        cost = 0.0
+        # 只有任务成功时才计费
+        if task.status == "completed" and expression:
+            engine = FormulaEngine()
+            try:
+                result = engine.evaluate(
+                    expression=str(expression),
+                    variables=variables,
+                    dimensions=dims,
+                    dimension_mappings=dimension_mappings,
+                    strict_mode=config.billing_strict_mode,
+                )
+                billing_snapshot["status"] = result.status
+                billing_snapshot["missing_required"] = result.missing_required
+                if result.status == "complete":
+                    cost = float(result.cost)
+                billing_snapshot["cost"] = cost
+            except BillingIncompleteError as exc:
+                # strict_mode=true：标记任务失败并隐藏产物，避免"免费放行"
+                task.status = "failed"
+                task.error_code = "billing_incomplete"
+                task.error_message = f"Missing required dimensions: {exc.missing_required}"
+                task.video_url = None
+                task.video_urls = None
+                billing_snapshot["status"] = "incomplete"
+                billing_snapshot["missing_required"] = exc.missing_required
+                billing_snapshot["cost"] = 0.0
+            except Exception as exc:
+                billing_snapshot["status"] = "incomplete"
+                billing_snapshot["error"] = str(exc)
+                billing_snapshot["cost"] = 0.0
+
+        # 回写到 task.request_metadata 便于审计/重算
+        # 重新赋值整个字典，确保 SQLAlchemy 检测到变更
+        metadata = dict(task.request_metadata) if task.request_metadata else {}
+        metadata["billing_snapshot"] = billing_snapshot
+        task.request_metadata = metadata
+
+        # === 更新已结算的 Usage 计费信息 ===
+        updated = UsageService.update_settled_billing(
+            self.db,
+            request_id=request_id,
+            total_cost_usd=cost,
+            request_cost_usd=cost,
+            status="completed" if task.status == "completed" else "failed",
+            status_code=200 if task.status == "completed" else 500,
+            error_message=(
+                None
+                if task.status == "completed"
+                else (task.error_message or task.error_code or "video_task_failed")
+            ),
+            response_time_ms=response_time_ms,
+            billing_snapshot=billing_snapshot,
+            extra_metadata={
+                "dimensions": dims,
+                "raw_response_ref": {
+                    "video_task_id": task.id,
+                    "field": "video_tasks.request_metadata.poll_raw_response",
+                },
+            },
+        )
+
+        if updated:
+            logger.debug(
+                "Updated video task billing: task_id=%s request_id=%s cost=%.6f",
+                task.id,
+                request_id,
+                cost,
+            )
+        else:
+            logger.warning(
+                "Failed to update video task billing (may already be updated): "
+                "task_id=%s request_id=%s",
+                task.id,
+                request_id,
+            )
+
+        return updated
+
+    async def _create_fallback_usage(self, task: VideoTask, request_id: str) -> bool:
+        """
+        兜底逻辑：当 Usage 不存在时创建完整记录。
+        这种情况理论上不应发生（submit 阶段已创建），但保留以防万一。
+        """
+        user_obj = self.db.query(User).filter(User.id == task.user_id).first()
+        api_key_obj = (
+            self.db.query(ApiKey).filter(ApiKey.id == task.api_key_id).first()
+            if task.api_key_id
+            else None
+        )
+        provider_obj = (
+            self.db.query(Provider).filter(Provider.id == task.provider_id).first()
+            if task.provider_id
+            else None
+        )
+        provider_name = provider_obj.name if provider_obj else "unknown"
+
+        # 计算响应时间
+        response_time_ms: int | None = None
+        if task.submitted_at and task.completed_at:
+            delta = task.completed_at - task.submitted_at
+            response_time_ms = int(delta.total_seconds() * 1000)
+
+        try:
+            await UsageService.record_usage_with_custom_cost(
+                db=self.db,
+                user=user_obj,
+                api_key=api_key_obj,
+                provider=provider_name,
+                model=task.model,
+                request_type="video",
+                total_cost_usd=0.0,  # 兜底记录不计费
+                request_cost_usd=0.0,
+                input_tokens=0,
+                output_tokens=0,
+                cache_creation_input_tokens=0,
+                cache_read_input_tokens=0,
+                api_format=task.client_api_format,
+                endpoint_api_format=task.provider_api_format,
+                has_format_conversion=bool(task.format_converted),
+                is_stream=False,
+                response_time_ms=response_time_ms,
+                first_byte_time_ms=None,
+                status_code=200 if task.status == "completed" else 500,
+                error_message=(
+                    None
+                    if task.status == "completed"
+                    else (task.error_message or task.error_code or "video_task_failed")
+                ),
+                metadata={
+                    "fallback_created": True,
+                    "video_task_id": task.id,
+                },
+                request_headers=(
+                    (task.request_metadata or {}).get("request_headers")
+                    if isinstance(task.request_metadata, dict)
+                    else None
+                ),
+                request_body=task.original_request_body,
+                provider_request_headers=None,
+                response_headers=None,
+                client_response_headers=None,
+                response_body=None,
+                request_id=request_id,
+                provider_id=task.provider_id,
+                provider_endpoint_id=task.endpoint_id,
+                provider_api_key_id=task.key_id,
+                status="completed" if task.status == "completed" else "failed",
+                target_model=None,
+            )
+            return True
+        except Exception as exc:
+            logger.exception(
+                "Failed to create fallback usage for video task=%s: %s",
+                task.id,
+                str(exc),
+            )
+            return False

src/services/task/context.py

@@ -0,0 +1,36 @@
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from enum import Enum
+
+
+class TaskMode(str, Enum):
+    SYNC = "sync"
+    ASYNC = "async"
+
+
+@dataclass(slots=True)
+class TaskContext:
+    """
+    TaskContext (pure DTO)
+
+    - Only primitive types / IDs
+    - Serializable & safe to pass across processes
+    """
+
+    request_id: str
+    task_type: str  # chat/cli/video/image/audio
+    task_mode: TaskMode
+
+    user_id: str
+    api_key_id: str
+
+    client_ip: str = ""
+    user_agent: str = ""
+    start_time: float = 0.0
+
+    api_format: str | None = None
+    model: str | None = None
+    mapped_model: str | None = None
+
+    capability_requirements: dict[str, bool] = field(default_factory=dict)

src/services/task/impl/__init__.py

@@ -1,3 +1,7 @@
-"""Task telemetry implementations for concrete task types (video/image/audio)."""
+"""Per-task-type implementations (Phase2).
+
+Currently includes:
+- video: polling adapter
+"""
 
 __all__ = []

src/services/task/impl/video_poller.py

@@ -0,0 +1,626 @@
+"""
+Video task poller adapter.
+
+Implements the video-specific poll/normalize/update logic used by TaskPollerService.
+
+优化：HTTP 请求期间不持有数据库连接，避免阻塞其他请求。
+采用三阶段处理：准备数据 -> HTTP 请求 -> 更新数据库。
+"""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass
+from datetime import datetime, timedelta, timezone
+from typing import Any
+
+from sqlalchemy.orm import Session
+
+from src.api.handlers.base.request_builder import ProviderAuthInfo, get_provider_auth
+from src.api.handlers.base.video_handler_base import (
+    normalize_gemini_operation_id,
+    sanitize_error_message,
+)
+from src.clients.http_client import HTTPClientPool
+from src.config.settings import config
+from src.core.api_format import (
+    build_upstream_headers_for_endpoint,
+    get_extra_headers_from_endpoint,
+    make_signature_key,
+)
+from src.core.api_format.conversion.internal_video import InternalVideoPollResult, VideoStatus
+from src.core.api_format.conversion.normalizers.gemini import GeminiNormalizer
+from src.core.api_format.conversion.normalizers.openai import OpenAINormalizer
+from src.core.crypto import crypto_service
+from src.core.logger import logger
+from src.database import create_session
+from src.models.database import ProviderAPIKey, ProviderEndpoint, VideoTask
+from src.services.task.application import TaskApplicationService
+
+
+@dataclass(slots=True)
+class VideoPollContext:
+    """视频轮询上下文，保存 HTTP 请求所需的数据（不依赖数据库会话）"""
+
+    task_id: str
+    external_task_id: str
+    provider_api_format: str
+    base_url: str
+    upstream_key: str
+    headers: dict[str, str]
+    # 用于更新任务的原始数据
+    poll_count: int
+    retry_count: int
+    poll_interval_seconds: int
+    max_poll_count: int
+    current_status: str
+
+
+# 永久性错误指示词（用于降级判断，不应重试）
+_PERMANENT_ERROR_INDICATORS = frozenset(
+    {
+        "not found",
+        "404",
+        "unauthorized",
+        "401",
+        "forbidden",
+        "403",
+        "invalid request",
+        "invalid api key",
+        "does not exist",
+    }
+)
+
+
+class PollHTTPError(RuntimeError):
+    """HTTP 轮询错误，携带状态码便于区分临时/永久错误"""
+
+    def __init__(self, status_code: int, message: str):
+        # 确保错误信息包含状态码
+        full_message = f"HTTP {status_code}: {message}" if message else f"HTTP {status_code}"
+        super().__init__(full_message)
+        self.status_code = status_code
+        self.original_message = message
+
+
+class VideoTaskPollerAdapter:
+    task_type = "video"
+
+    # scheduler
+    job_id = "task_poller:video"
+    job_name = "视频任务轮询"
+    interval_seconds = config.video_poll_interval_seconds
+
+    # distributed lock
+    lock_key = "task_poller:video:lock"
+    lock_ttl = 60
+
+    # execution
+    batch_size = config.video_poll_batch_size
+    concurrency = config.video_poll_concurrency
+    consecutive_failure_alert_threshold = 5
+    max_backoff_seconds = 300
+
+    def __init__(self) -> None:
+        self._openai_normalizer = OpenAINormalizer()
+        self._gemini_normalizer = GeminiNormalizer()
+
+    def sanitize_error_message(self, message: str) -> str:
+        return sanitize_error_message(message)
+
+    def list_due_task_ids(self, db: Session, *, now: datetime, limit: int) -> list[str]:
+        tasks = (
+            db.query(VideoTask)
+            .filter(
+                VideoTask.status.in_(
+                    [
+                        VideoStatus.SUBMITTED.value,
+                        VideoStatus.QUEUED.value,
+                        VideoStatus.PROCESSING.value,
+                    ]
+                ),
+                VideoTask.next_poll_at <= now,
+                VideoTask.poll_count < VideoTask.max_poll_count,
+            )
+            .order_by(VideoTask.next_poll_at.asc())
+            .limit(limit)
+            .all()
+        )
+        return [t.id for t in tasks]
+
+    def get_task(self, db: Session, task_id: str) -> VideoTask | None:
+        # SQLAlchemy 1.4+ API
+        return db.get(VideoTask, task_id)
+
+    # ==================== 分阶段处理方法（优化数据库连接占用）====================
+
+    async def prepare_poll_context(
+        self, db: Session, task: VideoTask
+    ) -> VideoPollContext | InternalVideoPollResult:
+        """
+        阶段 1：准备轮询上下文（短暂持有数据库连接）
+
+        Returns:
+            VideoPollContext: 成功时返回上下文
+            InternalVideoPollResult: 失败时返回错误结果
+        """
+        if not task.endpoint_id or not task.key_id:
+            return InternalVideoPollResult(
+                status=VideoStatus.FAILED,
+                error_code="missing_provider_info",
+                error_message="Task missing endpoint_id or key_id",
+            )
+
+        endpoint = self._get_endpoint(db, task.endpoint_id)
+        key = self._get_key(db, task.key_id)
+
+        if not key.api_key:
+            return InternalVideoPollResult(
+                status=VideoStatus.FAILED,
+                error_code="provider_config_error",
+                error_message="Provider key not properly configured",
+            )
+
+        try:
+            upstream_key = crypto_service.decrypt(key.api_key)
+        except Exception:
+            logger.warning("Failed to decrypt provider key for task %s", task.id)
+            return InternalVideoPollResult(
+                status=VideoStatus.FAILED,
+                error_code="decryption_error",
+                error_message="Failed to decrypt provider key",
+            )
+
+        provider_format = (task.provider_api_format or "").strip().lower()
+        if not provider_format:
+            provider_format = make_signature_key(
+                str(getattr(endpoint, "api_family", "")).strip().lower(),
+                str(getattr(endpoint, "endpoint_kind", "")).strip().lower(),
+            )
+
+        # 构建请求头
+        if provider_format.startswith("gemini:"):
+            auth_info = await get_provider_auth(endpoint, key)
+        else:
+            auth_info = None
+        headers = self._build_headers(provider_format, upstream_key, endpoint, auth_info)
+
+        return VideoPollContext(
+            task_id=task.id,
+            external_task_id=task.external_task_id or "",
+            provider_api_format=provider_format,
+            base_url=endpoint.base_url or "",
+            upstream_key=upstream_key,
+            headers=headers,
+            poll_count=task.poll_count,
+            retry_count=task.retry_count,
+            poll_interval_seconds=task.poll_interval_seconds,
+            max_poll_count=task.max_poll_count,
+            current_status=task.status,
+        )
+
+    async def poll_task_http(self, ctx: VideoPollContext) -> InternalVideoPollResult:
+        """
+        阶段 2：执行 HTTP 请求（不持有数据库连接）
+
+        Args:
+            ctx: 轮询上下文
+
+        Returns:
+            InternalVideoPollResult: 轮询结果
+        """
+        if not ctx.external_task_id:
+            return InternalVideoPollResult(
+                status=VideoStatus.FAILED,
+                error_code="missing_external_task_id",
+                error_message="Task missing external_task_id",
+            )
+
+        if ctx.provider_api_format.startswith("gemini:"):
+            return await self._poll_gemini_with_context(ctx)
+        return await self._poll_openai_with_context(ctx)
+
+    async def update_task_after_poll(
+        self,
+        task_id: str,
+        result: InternalVideoPollResult,
+        ctx: VideoPollContext | None,
+        redis_client: Any | None,
+        error_exception: Exception | None = None,
+    ) -> None:
+        """
+        阶段 3：更新数据库（获取新的数据库连接）
+
+        Args:
+            task_id: 任务 ID
+            result: 轮询结果
+            ctx: 轮询上下文（准备阶段就失败时为 None）
+            redis_client: Redis 客户端
+            error_exception: 如果 HTTP 请求失败，传入异常对象
+        """
+        with create_session() as db:
+            task = db.get(VideoTask, task_id)
+            if not task:
+                logger.warning("Task %s disappeared during poll update", task_id)
+                return
+
+            if error_exception is not None and ctx is not None:
+                # HTTP 请求失败（需要 ctx 来计算 backoff）
+                self._handle_poll_error(task, error_exception, ctx)
+            elif result.status == VideoStatus.COMPLETED:
+                task.status = VideoStatus.COMPLETED.value
+                task.video_url = result.video_url
+                task.video_expires_at = result.expires_at
+                task.completed_at = datetime.now(timezone.utc)
+                task.progress_percent = 100
+                if result.video_urls:
+                    task.video_urls = result.video_urls
+                self._attach_poll_raw_response(task, result)
+            elif result.status == VideoStatus.FAILED:
+                task.status = VideoStatus.FAILED.value
+                task.error_code = result.error_code
+                task.error_message = result.error_message
+                task.completed_at = datetime.now(timezone.utc)
+                self._attach_poll_raw_response(task, result)
+            else:
+                task.poll_count += 1
+                task.progress_percent = result.progress_percent
+                task.next_poll_at = datetime.now(timezone.utc) + timedelta(
+                    seconds=task.poll_interval_seconds
+                )
+
+            # 超时检查
+            task.updated_at = datetime.now(timezone.utc)
+            if task.poll_count >= task.max_poll_count and task.status not in [
+                VideoStatus.COMPLETED.value,
+                VideoStatus.FAILED.value,
+                VideoStatus.CANCELLED.value,
+            ]:
+                task.status = VideoStatus.FAILED.value
+                task.error_code = "poll_timeout"
+                task.error_message = f"Task timed out after {task.poll_count} polls"
+                task.completed_at = datetime.now(timezone.utc)
+
+            # 终态结算
+            if task.status in (VideoStatus.COMPLETED.value, VideoStatus.FAILED.value):
+                try:
+                    await TaskApplicationService(db, redis_client=redis_client).finalize_video_task(
+                        task
+                    )
+                except Exception as exc:
+                    logger.exception(
+                        "Failed to record video usage for task=%s: %s",
+                        task.id,
+                        sanitize_error_message(str(exc)),
+                    )
+
+            db.commit()
+
+    def _handle_poll_error(self, task: VideoTask, exc: Exception, ctx: VideoPollContext) -> None:
+        """处理轮询错误"""
+        task.poll_count += 1
+        error_msg = sanitize_error_message(str(exc))
+        logger.warning("Poll error for task %s: %s", task.id, error_msg)
+        task.progress_message = f"Poll error: {error_msg}"
+
+        status_code = exc.status_code if isinstance(exc, PollHTTPError) else None
+        is_permanent = self._is_permanent_error(exc, status_code=status_code)
+        if is_permanent:
+            task.status = VideoStatus.FAILED.value
+            task.error_code = "poll_permanent_error"
+            task.error_message = error_msg
+            task.completed_at = datetime.now(timezone.utc)
+        else:
+            backoff = min(
+                ctx.poll_interval_seconds * (2 ** min(ctx.retry_count, 5)),
+                self.max_backoff_seconds,
+            )
+            task.retry_count += 1
+            task.next_poll_at = datetime.now(timezone.utc) + timedelta(seconds=backoff)
+
+    async def _poll_openai_with_context(self, ctx: VideoPollContext) -> InternalVideoPollResult:
+        """使用上下文进行 OpenAI 轮询（不需要数据库）"""
+        url = self._build_openai_url(ctx.base_url, ctx.external_task_id)
+
+        client = await HTTPClientPool.get_default_client_async()
+        response = await client.get(url, headers=ctx.headers)
+        if response.status_code >= 400:
+            error_message = self._extract_error_message(response.text, response.status_code)
+            raise PollHTTPError(response.status_code, error_message)
+
+        payload = response.json()
+        return self._openai_normalizer.video_poll_to_internal(payload)
+
+    async def _poll_gemini_with_context(self, ctx: VideoPollContext) -> InternalVideoPollResult:
+        """使用上下文进行 Gemini 轮询（不需要数据库）"""
+        operation_name = normalize_gemini_operation_id(ctx.external_task_id)
+        url = self._build_gemini_url(ctx.base_url, operation_name)
+
+        logger.debug(
+            "[VideoPoller] Gemini poll: task=%s external_id=%s url=%s",
+            ctx.task_id,
+            ctx.external_task_id,
+            url,
+        )
+
+        client = await HTTPClientPool.get_default_client_async()
+        response = await client.get(url, headers=ctx.headers)
+        if response.status_code >= 400:
+            logger.warning(
+                "[VideoPoller] Gemini poll failed: task=%s status=%s response=%s",
+                ctx.task_id,
+                response.status_code,
+                response.text[:500] if response.text else "(empty)",
+            )
+            error_message = self._extract_error_message(response.text, response.status_code)
+            raise PollHTTPError(response.status_code, error_message)
+
+        payload = response.json()
+        return self._gemini_normalizer.video_poll_to_internal(payload)
+
+    # ==================== 旧版方法（保留兼容性）====================
+
+    async def poll_single_task(
+        self, db: Session, task: VideoTask, *, redis_client: Any | None
+    ) -> None:
+        """
+        旧版单任务轮询方法（保留向后兼容）
+
+        注意：此方法在 HTTP 请求期间持有数据库连接，建议使用分阶段方法。
+        """
+        try:
+            result = await self._poll_task_status(db, task)
+            if result.status == VideoStatus.COMPLETED:
+                task.status = VideoStatus.COMPLETED.value
+                task.video_url = result.video_url
+                task.video_expires_at = result.expires_at
+                task.completed_at = datetime.now(timezone.utc)
+                task.progress_percent = 100
+                if result.video_urls:
+                    task.video_urls = result.video_urls
+                self._attach_poll_raw_response(task, result)
+            elif result.status == VideoStatus.FAILED:
+                task.status = VideoStatus.FAILED.value
+                task.error_code = result.error_code
+                task.error_message = result.error_message
+                task.completed_at = datetime.now(timezone.utc)
+                self._attach_poll_raw_response(task, result)
+            else:
+                task.poll_count += 1
+                task.progress_percent = result.progress_percent
+                task.next_poll_at = datetime.now(timezone.utc) + timedelta(
+                    seconds=task.poll_interval_seconds
+                )
+        except Exception as exc:
+            task.poll_count += 1
+            error_msg = sanitize_error_message(str(exc))
+            logger.warning("Poll error for task %s: %s", task.id, error_msg)
+            task.progress_message = f"Poll error: {error_msg}"
+
+            status_code = exc.status_code if isinstance(exc, PollHTTPError) else None
+            is_permanent = self._is_permanent_error(exc, status_code=status_code)
+            if is_permanent:
+                task.status = VideoStatus.FAILED.value
+                task.error_code = "poll_permanent_error"
+                task.error_message = error_msg
+                task.completed_at = datetime.now(timezone.utc)
+            else:
+                backoff = min(
+                    task.poll_interval_seconds * (2 ** min(task.retry_count, 5)),
+                    self.max_backoff_seconds,
+                )
+                task.retry_count += 1
+                task.next_poll_at = datetime.now(timezone.utc) + timedelta(seconds=backoff)
+
+        # 超时：超过最大轮询次数且未进入终态
+        task.updated_at = datetime.now(timezone.utc)
+        if task.poll_count >= task.max_poll_count and task.status not in [
+            VideoStatus.COMPLETED.value,
+            VideoStatus.FAILED.value,
+            VideoStatus.CANCELLED.value,
+        ]:
+            task.status = VideoStatus.FAILED.value
+            task.error_code = "poll_timeout"
+            task.error_message = f"Task timed out after {task.poll_count} polls"
+            task.completed_at = datetime.now(timezone.utc)
+
+        # 终态结算
+        if task.status in (VideoStatus.COMPLETED.value, VideoStatus.FAILED.value):
+            try:
+                await TaskApplicationService(db, redis_client=redis_client).finalize_video_task(
+                    task
+                )
+            except Exception as exc:
+                logger.exception(
+                    "Failed to record video usage for task=%s: %s",
+                    task.id,
+                    sanitize_error_message(str(exc)),
+                )
+
+    def _attach_poll_raw_response(self, task: VideoTask, result: InternalVideoPollResult) -> None:
+        if not result.raw_response:
+            return
+        # 重新赋值整个字典，确保 SQLAlchemy 检测到变更
+        # （直接修改 JSON 字段内部不会自动标记为 dirty）
+        metadata = dict(task.request_metadata) if task.request_metadata else {}
+        metadata["poll_raw_response"] = result.raw_response
+        task.request_metadata = metadata
+
+    def _is_permanent_error(self, exc: Exception, status_code: int | None = None) -> bool:
+        if status_code is not None:
+            return 400 <= status_code < 500 and status_code != 429
+        error_msg = str(exc).lower()
+        return any(indicator in error_msg for indicator in _PERMANENT_ERROR_INDICATORS)
+
+    async def _poll_task_status(self, db: Session, task: VideoTask) -> InternalVideoPollResult:
+        if not task.endpoint_id or not task.key_id:
+            return InternalVideoPollResult(
+                status=VideoStatus.FAILED,
+                error_code="missing_provider_info",
+                error_message="Task missing endpoint_id or key_id",
+            )
+        endpoint = self._get_endpoint(db, task.endpoint_id)
+        key = self._get_key(db, task.key_id)
+        if not key.api_key:
+            return InternalVideoPollResult(
+                status=VideoStatus.FAILED,
+                error_code="provider_config_error",
+                error_message="Provider key not properly configured",
+            )
+        try:
+            upstream_key = crypto_service.decrypt(key.api_key)
+        except Exception:
+            logger.warning("Failed to decrypt provider key for task %s", task.id)
+            return InternalVideoPollResult(
+                status=VideoStatus.FAILED,
+                error_code="decryption_error",
+                error_message="Failed to decrypt provider key",
+            )
+
+        provider_format = (task.provider_api_format or "").strip().lower()
+        if not provider_format:
+            provider_format = make_signature_key(
+                str(getattr(endpoint, "api_family", "")).strip().lower(),
+                str(getattr(endpoint, "endpoint_kind", "")).strip().lower(),
+            )
+
+        if provider_format.startswith("gemini:"):
+            auth_info = await get_provider_auth(endpoint, key)
+            return await self._poll_gemini(task, endpoint, upstream_key, auth_info)
+        return await self._poll_openai(task, endpoint, upstream_key)
+
+    async def _poll_openai(
+        self,
+        task: VideoTask,
+        endpoint: ProviderEndpoint,
+        upstream_key: str,
+    ) -> InternalVideoPollResult:
+        if not task.external_task_id:
+            return InternalVideoPollResult(
+                status=VideoStatus.FAILED,
+                error_code="missing_external_task_id",
+                error_message="Task missing external_task_id",
+            )
+        url = self._build_openai_url(endpoint.base_url, task.external_task_id)
+        endpoint_sig = (task.provider_api_format or "").strip().lower() or make_signature_key(
+            str(getattr(endpoint, "api_family", "")).strip().lower(),
+            str(getattr(endpoint, "endpoint_kind", "")).strip().lower(),
+        )
+        headers = self._build_headers(endpoint_sig, upstream_key, endpoint)
+
+        client = await HTTPClientPool.get_default_client_async()
+        response = await client.get(url, headers=headers)
+        if response.status_code >= 400:
+            error_message = self._extract_error_message(response.text, response.status_code)
+            raise PollHTTPError(response.status_code, error_message)
+
+        payload = response.json()
+        return self._openai_normalizer.video_poll_to_internal(payload)
+
+    async def _poll_gemini(
+        self,
+        task: VideoTask,
+        endpoint: ProviderEndpoint,
+        upstream_key: str,
+        auth_info: ProviderAuthInfo | None,
+    ) -> InternalVideoPollResult:
+        if not task.external_task_id:
+            return InternalVideoPollResult(
+                status=VideoStatus.FAILED,
+                error_code="missing_external_task_id",
+                error_message="Task missing external_task_id",
+            )
+        operation_name = normalize_gemini_operation_id(task.external_task_id)
+        url = self._build_gemini_url(endpoint.base_url, operation_name)
+        endpoint_sig = (task.provider_api_format or "").strip().lower() or make_signature_key(
+            str(getattr(endpoint, "api_family", "")).strip().lower(),
+            str(getattr(endpoint, "endpoint_kind", "")).strip().lower(),
+        )
+        headers = self._build_headers(endpoint_sig, upstream_key, endpoint, auth_info)
+
+        logger.debug(
+            "[VideoPoller] Gemini poll: task=%s external_id=%s url=%s",
+            task.id,
+            task.external_task_id,
+            url,
+        )
+
+        client = await HTTPClientPool.get_default_client_async()
+        response = await client.get(url, headers=headers)
+        if response.status_code >= 400:
+            logger.warning(
+                "[VideoPoller] Gemini poll failed: task=%s status=%s response=%s",
+                task.id,
+                response.status_code,
+                response.text[:500] if response.text else "(empty)",
+            )
+            error_message = self._extract_error_message(response.text, response.status_code)
+            raise PollHTTPError(response.status_code, error_message)
+
+        payload = response.json()
+        return self._gemini_normalizer.video_poll_to_internal(payload)
+
+    def _build_openai_url(self, base_url: str | None, task_id: str) -> str:
+        base = (base_url or "https://api.openai.com").rstrip("/")
+        if base.endswith("/v1"):
+            return f"{base}/videos/{task_id}"
+        return f"{base}/v1/videos/{task_id}"
+
+    def _build_gemini_url(self, base_url: str | None, operation_name: str) -> str:
+        base = (base_url or "https://generativelanguage.googleapis.com").rstrip("/")
+        if base.endswith("/v1beta"):
+            base = base[: -len("/v1beta")]
+        return f"{base}/v1beta/{operation_name}"
+
+    def _build_headers(
+        self,
+        endpoint_sig: str,
+        upstream_key: str,
+        endpoint: ProviderEndpoint,
+        auth_info: ProviderAuthInfo | None = None,
+    ) -> dict[str, str]:
+        extra_headers = get_extra_headers_from_endpoint(endpoint)
+        headers = build_upstream_headers_for_endpoint(
+            {},
+            endpoint_sig,
+            upstream_key,
+            endpoint_headers=extra_headers,
+        )
+        if auth_info:
+            headers.pop("x-goog-api-key", None)
+            headers[auth_info.auth_header] = auth_info.auth_value
+        return headers
+
+    def _get_endpoint(self, db: Session, endpoint_id: str) -> ProviderEndpoint:
+        endpoint = db.query(ProviderEndpoint).filter(ProviderEndpoint.id == endpoint_id).first()
+        if not endpoint:
+            raise RuntimeError("Provider endpoint not found")
+        return endpoint
+
+    def _get_key(self, db: Session, key_id: str) -> ProviderAPIKey:
+        key = db.query(ProviderAPIKey).filter(ProviderAPIKey.id == key_id).first()
+        if not key:
+            raise RuntimeError("Provider key not found")
+        return key
+
+    def _extract_error_message(self, response_text: str | None, status_code: int) -> str:
+        """从响应中提取有意义的错误信息"""
+        if not response_text:
+            return f"Request failed with status {status_code}"
+
+        # 尝试解析 JSON 格式的错误
+        try:
+            data = json.loads(response_text)
+            # OpenAI 格式: {"error": {"message": "..."}}
+            if isinstance(data.get("error"), dict):
+                error_obj = data["error"]
+                message = error_obj.get("message") or error_obj.get("detail") or str(error_obj)
+                return sanitize_error_message(message)
+            # Gemini 格式: {"error": {"message": "...", "code": 404}}
+            if "message" in data:
+                return sanitize_error_message(data["message"])
+        except (json.JSONDecodeError, TypeError, KeyError):
+            pass
+
+        # 回退到原始文本（截断）
+        return sanitize_error_message(response_text[:500])

src/services/task/impl/video_telemetry.py

@@ -1,320 +0,0 @@
-"""
-VideoTelemetry（Phase3）
-
-将 Video 异步任务的“终态计费 + Usage 写入 + required 缺失告警”从 poller 中抽离出来，
-便于未来 Image/Audio 复用相同框架。
-"""
-
-from __future__ import annotations
-
-from datetime import datetime, timezone
-from typing import Any
-
-from sqlalchemy.orm import Session
-
-from src.api.handlers.base.video_handler_base import sanitize_error_message
-from src.config.settings import config
-from src.core.api_format.conversion.internal_video import VideoStatus
-from src.core.logger import logger
-from src.models.database import ApiKey, Provider, User, VideoTask
-from src.services.billing.dimension_collector_service import DimensionCollectorService
-from src.services.billing.formula_engine import BillingIncompleteError, FormulaEngine
-from src.services.billing.rule_service import BillingRuleService
-from src.services.usage.service import UsageService
-
-
-class VideoTelemetry:
-    def __init__(self, db: Session, *, redis_client: Any | None = None) -> None:
-        self.db = db
-        self.redis = redis_client
-        self._formula_engine = FormulaEngine()
-
-    async def record_terminal_usage(self, task: VideoTask) -> None:
-        """
-        为视频任务终态写入 Usage：
-        - COMPLETED: 使用 FormulaEngine 计算 cost（或 no_rule / incomplete -> cost=0）
-        - FAILED: cost=0
-
-        该方法可能会在 strict_mode 缺失 required 维度时将任务降级为 FAILED 并隐藏产物。
-        """
-        request_id = None
-        if isinstance(task.request_metadata, dict):
-            request_id = task.request_metadata.get("request_id")
-        request_id = request_id or task.id
-
-        # 计算异步任务总耗时（ms）
-        response_time_ms = None
-        if task.submitted_at and task.completed_at:
-            delta = task.completed_at - task.submitted_at
-            response_time_ms = int(delta.total_seconds() * 1000)
-
-        # 基础维度（无需 collectors 也可计费）
-        base_dimensions: dict[str, Any] = {
-            "duration_seconds": task.duration_seconds,
-            "resolution": task.resolution,
-            "aspect_ratio": task.aspect_ratio,
-            "size": task.size or "",
-            "retry_count": task.retry_count,
-        }
-
-        # collectors 可用的 metadata（结构稳定，便于配置 path）
-        collector_metadata: dict[str, Any] = {
-            "task": {
-                "id": task.id,
-                "external_task_id": task.external_task_id,
-                "model": task.model,
-                "duration_seconds": task.duration_seconds,
-                "resolution": task.resolution,
-                "aspect_ratio": task.aspect_ratio,
-                "size": task.size,
-                "retry_count": task.retry_count,
-                "video_size_bytes": task.video_size_bytes,
-            },
-            "result": {
-                "video_url": task.video_url,
-                "video_urls": task.video_urls or [],
-            },
-        }
-
-        # 维度采集：base + collectors 覆盖/补全
-        dims = DimensionCollectorService(self.db).collect_dimensions(
-            api_format=task.provider_api_format,
-            task_type="video",
-            request=task.original_request_body or {},
-            response=(
-                (task.request_metadata or {}).get("poll_raw_response")
-                if isinstance(task.request_metadata, dict)
-                else None
-            ),
-            metadata=collector_metadata,
-            base_dimensions=base_dimensions,
-        )
-
-        # 取冻结的 rule_snapshot；若缺失则回退 DB 查找（兼容旧任务）
-        rule_snapshot = None
-        if isinstance(task.request_metadata, dict):
-            rule_snapshot = task.request_metadata.get("billing_rule_snapshot")
-
-        billing_snapshot: dict[str, Any] = {
-            "status": "complete",
-            "missing_required": [],
-            "strict_mode": config.billing_strict_mode,
-        }
-        cost = 0.0
-
-        if task.status == VideoStatus.FAILED.value:
-            billing_snapshot["billed_reason"] = "task_failed"
-        else:
-            # COMPLETED：计算成本
-            expression = None
-            variables = None
-            dimension_mappings = None
-            rule_id = None
-            rule_name = None
-            rule_scope = None
-
-            if isinstance(rule_snapshot, dict) and rule_snapshot.get("status") == "ok":
-                rule_id = rule_snapshot.get("rule_id")
-                rule_name = rule_snapshot.get("rule_name")
-                rule_scope = rule_snapshot.get("scope")
-                expression = rule_snapshot.get("expression")
-                variables = rule_snapshot.get("variables")
-                dimension_mappings = rule_snapshot.get("dimension_mappings")
-            else:
-                lookup = BillingRuleService.find_rule(
-                    self.db,
-                    provider_id=task.provider_id,
-                    model_name=task.model,
-                    task_type="video",
-                )
-                if lookup:
-                    rule = lookup.rule
-                    rule_id = rule.id
-                    rule_name = rule.name
-                    rule_scope = lookup.scope
-                    expression = rule.expression
-                    variables = rule.variables
-                    dimension_mappings = rule.dimension_mappings
-
-            if not expression:
-                billing_snapshot["status"] = "no_rule"
-                billing_snapshot["cost_breakdown"] = {"total": 0.0}
-                logger.warning(
-                    "No billing rule for video task (request_id=%s, model=%s, provider_id=%s)",
-                    request_id,
-                    task.model,
-                    task.provider_id,
-                )
-            else:
-                billing_snapshot.update(
-                    {
-                        "rule_id": rule_id,
-                        "rule_name": rule_name,
-                        "rule_scope": rule_scope,
-                        "expression": expression,
-                        "variables": variables or {},
-                    }
-                )
-                try:
-                    result = self._formula_engine.evaluate(
-                        expression=expression,
-                        variables=variables or {},
-                        dimensions=dims,
-                        dimension_mappings=dimension_mappings or {},
-                        strict_mode=config.billing_strict_mode,
-                    )
-                    billing_snapshot["status"] = result.status
-                    billing_snapshot["missing_required"] = result.missing_required
-                    billing_snapshot["resolved_values"] = result.resolved_values
-                    if result.status == "complete":
-                        cost = result.cost
-                    else:
-                        logger.error(
-                            "Billing incomplete due to missing required dimensions "
-                            "(request_id=%s, model=%s, missing=%s)",
-                            request_id,
-                            task.model,
-                            result.missing_required,
-                        )
-                        cost = 0.0
-                        await self._maybe_alert_missing_required(
-                            model=task.model,
-                            missing_required=result.missing_required,
-                        )
-                    if result.error:
-                        billing_snapshot["error"] = result.error
-                except BillingIncompleteError as exc:
-                    logger.error(
-                        "Billing strict mode triggered (request_id=%s, model=%s, missing=%s)",
-                        request_id,
-                        task.model,
-                        exc.missing_required,
-                    )
-                    billing_snapshot["status"] = "incomplete"
-                    billing_snapshot["missing_required"] = exc.missing_required
-                    billing_snapshot["resolved_values"] = {}
-                    billing_snapshot["error"] = "strict_mode_missing_required"
-                    cost = 0.0
-
-                    # strict_mode=true：标记任务失败并隐藏产物，避免"免费放行"
-                    task.status = VideoStatus.FAILED.value
-                    task.error_code = "billing_incomplete"
-                    task.error_message = f"Missing required dimensions: {exc.missing_required}"
-                    task.video_url = None
-                    task.video_urls = None
-
-                    await self._maybe_alert_missing_required(
-                        model=task.model,
-                        missing_required=exc.missing_required,
-                    )
-
-                billing_snapshot["cost_breakdown"] = {"total": cost}
-
-        # 将 billing_snapshot 回写到 task.request_metadata 便于对账（不会影响 usage 的单独存档）
-        if task.request_metadata is None:
-            task.request_metadata = {}
-        if isinstance(task.request_metadata, dict):
-            task.request_metadata["billing_snapshot"] = billing_snapshot
-
-        usage_metadata: dict[str, Any] = {
-            "billing_snapshot": billing_snapshot,
-            "dimensions": dims,
-            "raw_response_ref": {
-                "video_task_id": task.id,
-                "field": "video_tasks.request_metadata.poll_raw_response",
-            },
-        }
-
-        # 查询关联对象（用于写入 usage.user_id/api_key_id 等）
-        user_obj = self.db.query(User).filter(User.id == task.user_id).first()
-        api_key_obj = (
-            self.db.query(ApiKey).filter(ApiKey.id == task.api_key_id).first()
-            if task.api_key_id
-            else None
-        )
-        provider_obj = (
-            self.db.query(Provider).filter(Provider.id == task.provider_id).first()
-            if task.provider_id
-            else None
-        )
-        provider_name = provider_obj.name if provider_obj else "unknown"
-
-        await UsageService.record_usage_with_custom_cost(
-            db=self.db,
-            user=user_obj,
-            api_key=api_key_obj,
-            provider=provider_name,
-            model=task.model,
-            request_type="video",
-            total_cost_usd=cost,
-            request_cost_usd=cost,
-            input_tokens=0,
-            output_tokens=0,
-            cache_creation_input_tokens=0,
-            cache_read_input_tokens=0,
-            api_format=task.client_api_format,
-            endpoint_api_format=task.provider_api_format,
-            has_format_conversion=bool(task.format_converted),
-            is_stream=False,
-            response_time_ms=response_time_ms,
-            first_byte_time_ms=None,
-            status_code=200 if task.status == VideoStatus.COMPLETED.value else 500,
-            error_message=(
-                None
-                if task.status == VideoStatus.COMPLETED.value
-                else (task.error_message or task.error_code or "video_task_failed")
-            ),
-            metadata=usage_metadata,
-            request_headers=(
-                (task.request_metadata or {}).get("request_headers")
-                if isinstance(task.request_metadata, dict)
-                else None
-            ),
-            request_body=task.original_request_body,
-            provider_request_headers=None,
-            response_headers=None,
-            client_response_headers=None,
-            response_body=None,
-            request_id=request_id,
-            provider_id=task.provider_id,
-            provider_endpoint_id=task.endpoint_id,
-            provider_api_key_id=task.key_id,
-            status="completed" if task.status == VideoStatus.COMPLETED.value else "failed",
-            target_model=None,
-        )
-
-    async def _maybe_alert_missing_required(
-        self, *, model: str, missing_required: list[str]
-    ) -> None:
-        """required 维度缺失告警：同一 (model, dimension) 1 小时内 >= 10 次触发升级告警。"""
-        if not missing_required:
-            return
-        if not self.redis:
-            logger.error(
-                "Missing required billing dimensions (model=%s): %s", model, missing_required
-            )
-            return
-
-        # 按小时 bucket 聚合
-        now = datetime.now(timezone.utc)
-        hour_bucket = now.strftime("%Y%m%d%H")
-        for dim in missing_required:
-            key = f"billing:missing_required:{model}:{dim}:{hour_bucket}"
-            try:
-                count = await self.redis.incr(key)
-                if count == 1:
-                    await self.redis.expire(key, 3700)
-                if count >= 10:
-                    logger.warning(
-                        "Billing required dimension missing frequently (model=%s, dim=%s, count=%s/hour)",
-                        model,
-                        dim,
-                        count,
-                    )
-            except Exception as exc:
-                logger.warning(
-                    "Failed to record billing alert counter: %s", sanitize_error_message(str(exc))
-                )
-
-
-__all__ = ["VideoTelemetry"]

src/services/task/lifecycle.py

@@ -0,0 +1,27 @@
+from __future__ import annotations
+
+from enum import Enum
+
+
+class TaskStatus(str, Enum):
+    """Generic task status (progress)."""
+
+    PENDING = "pending"
+    STREAMING = "streaming"
+
+    SUBMITTED = "submitted"
+    QUEUED = "queued"
+    PROCESSING = "processing"
+
+    COMPLETED = "completed"
+    FAILED = "failed"
+    CANCELLED = "cancelled"
+    EXPIRED = "expired"
+
+
+class BillingStatus(str, Enum):
+    """Billing settlement status (Usage.billing_status)."""
+
+    PENDING = "pending"
+    SETTLED = "settled"
+    VOID = "void"

src/services/task/orchestrator.py

@@ -1,635 +0,0 @@
-"""
-AsyncTaskOrchestrator
-
-提交阶段故障转移（多候选尝试）：
-- 目标：拿到 external_task_id 后锁定 provider/endpoint/key，后续轮询不再切换。
-- 仅覆盖“提交阶段”；轮询阶段由各 task poller 使用已锁定的信息执行。
-"""
-
-from __future__ import annotations
-
-import re
-import uuid
-from dataclasses import dataclass
-from datetime import datetime, timezone
-from typing import Any, Protocol, runtime_checkable
-
-import httpx
-from redis.asyncio import Redis
-from sqlalchemy.orm import Session
-
-from src.config.settings import config
-from src.core.exceptions import ProviderNotAvailableException
-from src.core.logger import logger
-from src.models.database import ApiKey, RequestCandidate
-from src.services.billing.rule_service import BillingRuleLookupResult, BillingRuleService
-from src.services.cache.aware_scheduler import ProviderCandidate, get_cache_aware_scheduler
-from src.services.orchestration.candidate_resolver import CandidateResolver
-from src.services.orchestration.error_classifier import ErrorClassifier
-from src.services.system.config import SystemConfigService
-
-_SENSITIVE_PATTERN = re.compile(
-    r"(api[_-]?key|token|bearer|authorization)[=:\s]+\S+",
-    re.IGNORECASE,
-)
-
-
-def _sanitize(message: str, max_length: int = 200) -> str:
-    if not message:
-        return "request_failed"
-    return _SENSITIVE_PATTERN.sub("[REDACTED]", message)[:max_length]
-
-
-@runtime_checkable
-class SubmitFunc(Protocol):
-    async def __call__(self, candidate: ProviderCandidate) -> httpx.Response: ...
-
-
-@runtime_checkable
-class ExtractExternalTaskIdFunc(Protocol):
-    def __call__(self, payload: dict[str, Any]) -> str | None: ...
-
-
-class UpstreamClientRequestError(RuntimeError):
-    """可判定为客户端请求问题（不应 failover）的上游错误。"""
-
-    def __init__(
-        self,
-        *,
-        response: httpx.Response,
-        candidate_keys: list[dict[str, Any]],
-    ) -> None:
-        self.response = response
-        self.candidate_keys = candidate_keys
-        super().__init__(f"Upstream client error: HTTP {response.status_code}")
-
-
-class AllCandidatesFailedError(RuntimeError):
-    def __init__(
-        self,
-        *,
-        reason: str,
-        candidate_keys: list[dict[str, Any]],
-        last_status_code: int | None = None,
-    ) -> None:
-        self.reason = reason
-        self.candidate_keys = candidate_keys
-        self.last_status_code = last_status_code
-        super().__init__(f"All candidates failed: {reason}")
-
-
-class CandidateUnsupportedError(RuntimeError):
-    """候选不被当前任务支持（如 auth_type/格式转换需求不支持）。"""
-
-
-class CandidateSubmissionError(RuntimeError):
-    """候选提交异常（网络/解密/解析等）。"""
-
-
-@dataclass(slots=True)
-class SubmitOutcome:
-    candidate: ProviderCandidate
-    candidate_keys: list[dict[str, Any]]
-    external_task_id: str
-    rule_lookup: BillingRuleLookupResult | None
-    upstream_payload: dict[str, Any] | None = None
-
-
-class AsyncTaskOrchestrator:
-    """
-    异步任务编排器：只负责提交阶段的候选遍历与错误处理策略。
-    """
-
-    def __init__(self, db: Session, *, redis_client: Redis | None = None) -> None:
-        self.db = db
-        self.redis = redis_client
-        self._candidate_resolver: CandidateResolver | None = None
-        self._error_classifier: ErrorClassifier | None = None
-
-        self._cache_scheduler = None
-        # 候选记录映射：{candidate_index: RequestCandidate}
-        self._candidate_records: dict[int, RequestCandidate] = {}
-
-    def _create_candidate_records(
-        self,
-        candidates: list[ProviderCandidate],
-        request_id: str | None,
-        user_api_key: ApiKey,
-    ) -> dict[int, RequestCandidate]:
-        """
-        为所有候选预创建 RequestCandidate 记录。
-
-        Args:
-            candidates: 候选列表
-            request_id: 请求 ID
-            user_api_key: 用户 API Key
-
-        Returns:
-            {candidate_index: RequestCandidate} 映射
-        """
-        if not request_id:
-            return {}
-
-        now = datetime.now(timezone.utc)
-        records: dict[int, RequestCandidate] = {}
-
-        for idx, cand in enumerate(candidates):
-            record = RequestCandidate(
-                id=str(uuid.uuid4()),
-                request_id=request_id,
-                candidate_index=idx,
-                retry_index=0,
-                user_id=user_api_key.user_id if user_api_key else None,
-                api_key_id=user_api_key.id if user_api_key else None,
-                provider_id=cand.provider.id,
-                endpoint_id=cand.endpoint.id,
-                key_id=cand.key.id,
-                status="available",
-                is_cached=bool(getattr(cand, "is_cached", False)),
-                created_at=now,
-            )
-            self.db.add(record)
-            records[idx] = record
-
-        try:
-            self.db.flush()
-        except Exception as exc:
-            logger.warning(
-                "[AsyncTaskOrchestrator] Failed to create candidate records: %s",
-                str(exc),
-            )
-            self.db.rollback()
-            return {}
-
-        return records
-
-    def _update_candidate_record(
-        self,
-        idx: int,
-        *,
-        status: str,
-        skip_reason: str | None = None,
-        status_code: int | None = None,
-        error_type: str | None = None,
-        error_message: str | None = None,
-        started_at: datetime | None = None,
-        finished_at: datetime | None = None,
-    ) -> None:
-        """更新候选记录状态。"""
-        record = self._candidate_records.get(idx)
-        if not record:
-            return
-
-        record.status = status
-        if skip_reason is not None:
-            record.skip_reason = skip_reason
-        if status_code is not None:
-            record.status_code = status_code
-        if error_type is not None:
-            record.error_type = error_type
-        if error_message is not None:
-            record.error_message = error_message
-        if started_at is not None:
-            record.started_at = started_at
-        if finished_at is not None:
-            record.finished_at = finished_at
-
-        try:
-            self.db.flush()
-        except Exception as exc:
-            logger.warning(
-                "[AsyncTaskOrchestrator] Failed to update candidate record %d: %s",
-                idx,
-                str(exc),
-            )
-
-    def _commit_candidate_records(self) -> None:
-        """提交候选记录到数据库。"""
-        if not self._candidate_records:
-            return
-        try:
-            self.db.commit()
-        except Exception as exc:
-            logger.warning(
-                "[AsyncTaskOrchestrator] Failed to commit candidate records: %s",
-                str(exc),
-            )
-            self.db.rollback()
-
-    async def _ensure_initialized(self) -> None:
-        if self._cache_scheduler is not None:
-            return
-
-        # 使用 SystemConfigService 读取运行时调度策略（与 Chat/CLI 一致）
-        priority_mode = SystemConfigService.get_config(
-            self.db,
-            "provider_priority_mode",
-            "provider",
-        )
-        scheduling_mode = SystemConfigService.get_config(
-            self.db,
-            "scheduling_mode",
-            "cache_affinity",
-        )
-        self._cache_scheduler = await get_cache_aware_scheduler(
-            self.redis,
-            priority_mode=priority_mode,
-            scheduling_mode=scheduling_mode,
-        )
-
-        self._candidate_resolver = CandidateResolver(
-            db=self.db,
-            cache_scheduler=self._cache_scheduler,
-        )
-        self._error_classifier = ErrorClassifier(db=self.db, cache_scheduler=self._cache_scheduler)
-
-    def _should_stop_on_http_error(self, *, status_code: int, error_text: str) -> bool:
-        """
-        判断某个上游 HTTP 错误是否为“客户端错误”（不应 failover）。
-
-        规则：
-        - 401/403/429：一般是 key/权限/限流问题，优先 failover
-        - 其他 4xx：若 ErrorClassifier 判断为客户端请求错误，则停止
-        """
-        if status_code in (401, 403, 429):
-            return False
-        if 400 <= status_code < 500:
-            assert self._error_classifier is not None
-            return self._error_classifier.is_client_error(error_text)
-        return False
-
-    async def submit_with_failover(
-        self,
-        *,
-        api_format: str,
-        model_name: str,
-        affinity_key: str,
-        user_api_key: ApiKey,
-        request_id: str | None,
-        task_type: str,
-        submit_func: SubmitFunc,
-        extract_external_task_id: ExtractExternalTaskIdFunc,
-        supported_auth_types: set[str] | None = None,
-        allow_format_conversion: bool = False,
-        capability_requirements: dict[str, bool] | None = None,
-        max_candidates: int | None = None,
-    ) -> SubmitOutcome:
-        """
-        提交异步任务并在失败时自动尝试下一个候选，直到拿到 external_task_id。
-
-        Returns:
-            SubmitOutcome（包含选中的候选 + external_task_id + candidate_keys + billing rule lookup）
-
-        Raises:
-            UpstreamClientRequestError: 判定为客户端请求错误（不应 failover）
-            ProviderNotAvailableException: 没有可用候选（调度器层面）
-            AllCandidatesFailedError: 有候选但全部提交失败
-        """
-        await self._ensure_initialized()
-        assert self._candidate_resolver is not None
-
-        logger.info(
-            "[AsyncTaskOrchestrator] submit_with_failover: "
-            "api_format=%s, model=%s, task_type=%s, request_id=%s",
-            api_format,
-            model_name,
-            task_type,
-            request_id,
-        )
-
-        candidates, _global_model_id = await self._candidate_resolver.fetch_candidates(
-            api_format=api_format,
-            model_name=model_name,
-            affinity_key=affinity_key,
-            user_api_key=user_api_key,
-            request_id=request_id,
-            is_stream=False,
-            capability_requirements=capability_requirements,
-        )
-
-        logger.info(
-            "[AsyncTaskOrchestrator] fetch_candidates returned %d candidates for model=%s",
-            len(candidates),
-            model_name,
-        )
-
-        # 如果没有候选，直接抛出异常
-        if not candidates:
-            logger.error(
-                "[AsyncTaskOrchestrator] No candidates returned from fetch_candidates for model=%s",
-                model_name,
-            )
-            raise ProviderNotAvailableException("No candidates available")
-
-        if max_candidates is not None and max_candidates > 0:
-            candidates = candidates[:max_candidates]
-
-        # 创建候选记录（用于链路追踪）
-        self._candidate_records = self._create_candidate_records(
-            candidates=candidates,
-            request_id=request_id,
-            user_api_key=user_api_key,
-        )
-
-        candidate_keys: list[dict[str, Any]] = []
-        eligible_count = 0
-        last_status_code: int | None = None
-
-        for idx, cand in enumerate(candidates):
-            submit_started_at = datetime.now(timezone.utc)
-            auth_type = getattr(cand.key, "auth_type", "api_key") or "api_key"
-            candidate_info: dict[str, Any] = {
-                "index": idx,
-                "provider_id": cand.provider.id,
-                "provider_name": cand.provider.name,
-                "endpoint_id": cand.endpoint.id,
-                "key_id": cand.key.id,
-                "key_name": cand.key.name,
-                "auth_type": auth_type,
-                "priority": getattr(cand.key, "priority", 0) or 0,
-                "is_cached": bool(getattr(cand, "is_cached", False)),
-            }
-            candidate_keys.append(candidate_info)
-
-            logger.info(
-                "[AsyncTaskOrchestrator] Checking candidate %d: provider=%s, is_skipped=%s, skip_reason=%s, needs_conversion=%s, auth_type=%s",
-                idx,
-                cand.provider.name,
-                getattr(cand, "is_skipped", False),
-                getattr(cand, "skip_reason", None),
-                getattr(cand, "needs_conversion", False),
-                auth_type,
-            )
-
-            # 调度器层面标记为跳过（健康/熔断/并发等）
-            if getattr(cand, "is_skipped", False):
-                skip_reason = getattr(cand, "skip_reason", None) or "skipped"
-                candidate_info.update(
-                    {
-                        "skipped": True,
-                        "skip_reason": skip_reason,
-                    }
-                )
-                self._update_candidate_record(idx, status="skipped", skip_reason=skip_reason)
-                logger.info(
-                    "[AsyncTaskOrchestrator] Candidate %d skipped: is_skipped=True, reason=%s",
-                    idx,
-                    cand.skip_reason,
-                )
-                continue
-
-            # 视频/图片等直连 upstream 的 handler 目前不支持跨格式转换
-            if not allow_format_conversion and bool(getattr(cand, "needs_conversion", False)):
-                candidate_info.update(
-                    {"skipped": True, "skip_reason": "format_conversion_not_supported"}
-                )
-                self._update_candidate_record(
-                    idx, status="skipped", skip_reason="format_conversion_not_supported"
-                )
-                logger.info("[AsyncTaskOrchestrator] Candidate %d skipped: needs_conversion", idx)
-                continue
-
-            # auth_type 过滤
-            if supported_auth_types is not None and auth_type not in supported_auth_types:
-                skip_reason = f"unsupported_auth_type:{auth_type}"
-                candidate_info.update(
-                    {
-                        "skipped": True,
-                        "skip_reason": skip_reason,
-                    }
-                )
-                self._update_candidate_record(idx, status="skipped", skip_reason=skip_reason)
-                logger.info(
-                    "[AsyncTaskOrchestrator] Candidate %d skipped: unsupported_auth_type=%s",
-                    idx,
-                    auth_type,
-                )
-                continue
-
-            # billing rule 过滤（可选）
-            rule_lookup: BillingRuleLookupResult | None = None
-            has_billing_rule = True
-            if config.billing_require_rule:
-                logger.info(
-                    "[AsyncTaskOrchestrator] Checking billing rule for candidate %d (billing_require_rule=True)",
-                    idx,
-                )
-                rule_lookup = BillingRuleService.find_rule(
-                    self.db,
-                    provider_id=cand.provider.id,
-                    model_name=model_name,
-                    task_type=task_type,
-                )
-                has_billing_rule = rule_lookup is not None
-                logger.info(
-                    "[AsyncTaskOrchestrator] Billing rule lookup result: has_rule=%s",
-                    has_billing_rule,
-                )
-                if not has_billing_rule:
-                    candidate_info.update(
-                        {
-                            "has_billing_rule": False,
-                            "skipped": True,
-                            "skip_reason": "billing_rule_missing",
-                        }
-                    )
-                    self._update_candidate_record(
-                        idx, status="skipped", skip_reason="billing_rule_missing"
-                    )
-                    logger.info(
-                        "[AsyncTaskOrchestrator] Candidate %d skipped: billing_rule_missing", idx
-                    )
-                    continue
-            candidate_info["has_billing_rule"] = has_billing_rule
-
-            logger.info("[AsyncTaskOrchestrator] Candidate %d eligible, attempting submit", idx)
-            eligible_count += 1
-
-            # 更新记录为 pending 状态（开始尝试）
-            self._update_candidate_record(idx, status="pending", started_at=submit_started_at)
-
-            # 尝试提交
-            try:
-                response = await submit_func(cand)
-            except Exception as exc:
-                finished_at = datetime.now(timezone.utc)
-                logger.error(
-                    "[AsyncTaskOrchestrator] Candidate %d submit exception: %s: %s",
-                    idx,
-                    type(exc).__name__,
-                    str(exc),
-                )
-                error_msg = _sanitize(str(exc))
-                candidate_info.update(
-                    {
-                        "attempt_status": "exception",
-                        "error_type": type(exc).__name__,
-                        "error_message": error_msg,
-                    }
-                )
-                self._update_candidate_record(
-                    idx,
-                    status="failed",
-                    error_type=type(exc).__name__,
-                    error_message=error_msg,
-                    finished_at=finished_at,
-                )
-                continue
-
-            logger.info(
-                "[AsyncTaskOrchestrator] Candidate %d submit response: status_code=%d",
-                idx,
-                response.status_code,
-            )
-
-            last_status_code = int(getattr(response, "status_code", 0) or 0)
-
-            # 上游错误：决定是否停止
-            if response.status_code >= 400:
-                finished_at = datetime.now(timezone.utc)
-                error_text = ""
-                try:
-                    error_text = response.text or ""
-                except Exception:
-                    error_text = ""
-                error_msg = _sanitize(error_text)
-                candidate_info.update(
-                    {
-                        "attempt_status": "http_error",
-                        "status_code": response.status_code,
-                        "error_message": error_msg,
-                    }
-                )
-                self._update_candidate_record(
-                    idx,
-                    status="failed",
-                    status_code=response.status_code,
-                    error_type="http_error",
-                    error_message=error_msg,
-                    finished_at=finished_at,
-                )
-                if self._should_stop_on_http_error(
-                    status_code=response.status_code, error_text=error_text
-                ):
-                    self._commit_candidate_records()
-                    raise UpstreamClientRequestError(
-                        response=response,
-                        candidate_keys=candidate_keys,
-                    )
-                continue
-
-            # 解析任务 ID（200 但缺字段也视为失败并 failover）
-            payload: dict[str, Any] | None = None
-            try:
-                data = response.json()
-                if isinstance(data, dict):
-                    payload = data
-                logger.info(
-                    "[AsyncTaskOrchestrator] Candidate %d response payload: %s",
-                    idx,
-                    str(payload)[:500] if payload else "None",
-                )
-            except Exception as exc:
-                finished_at = datetime.now(timezone.utc)
-                logger.error(
-                    "[AsyncTaskOrchestrator] Candidate %d invalid JSON: %s",
-                    idx,
-                    str(exc),
-                )
-                error_msg = _sanitize(str(exc))
-                candidate_info.update(
-                    {
-                        "attempt_status": "invalid_json",
-                        "error_type": type(exc).__name__,
-                        "error_message": error_msg,
-                    }
-                )
-                self._update_candidate_record(
-                    idx,
-                    status="failed",
-                    status_code=response.status_code,
-                    error_type="invalid_json",
-                    error_message=error_msg,
-                    finished_at=finished_at,
-                )
-                continue
-
-            external_task_id = extract_external_task_id(payload or {})
-            logger.info(
-                "[AsyncTaskOrchestrator] Candidate %d extracted task_id: %s",
-                idx,
-                external_task_id,
-            )
-            if not external_task_id:
-                finished_at = datetime.now(timezone.utc)
-                candidate_info.update(
-                    {
-                        "attempt_status": "empty_task_id",
-                        "error_message": "Upstream returned empty task id",
-                    }
-                )
-                self._update_candidate_record(
-                    idx,
-                    status="failed",
-                    status_code=response.status_code,
-                    error_type="empty_task_id",
-                    error_message="Upstream returned empty task id",
-                    finished_at=finished_at,
-                )
-                logger.warning(
-                    "[AsyncTaskOrchestrator] Candidate %d: empty task_id, payload keys: %s",
-                    idx,
-                    list(payload.keys()) if payload else [],
-                )
-                continue
-
-            # 成功
-            finished_at = datetime.now(timezone.utc)
-            candidate_info.update({"attempt_status": "success", "selected": True})
-            self._update_candidate_record(
-                idx,
-                status="success",
-                status_code=response.status_code,
-                finished_at=finished_at,
-            )
-            self._commit_candidate_records()
-            return SubmitOutcome(
-                candidate=cand,
-                candidate_keys=candidate_keys,
-                external_task_id=str(external_task_id),
-                rule_lookup=rule_lookup,
-                upstream_payload=payload,
-            )
-
-        # 没有任何候选可尝试
-        if not candidates:
-            raise ProviderNotAvailableException("No candidates available")
-
-        # 提交所有候选记录
-        self._commit_candidate_records()
-
-        if eligible_count == 0:
-            reason = "no_eligible_candidates"
-            if config.billing_require_rule:
-                reason = "no_candidate_with_billing_rule"
-            raise AllCandidatesFailedError(
-                reason=reason,
-                candidate_keys=candidate_keys,
-                last_status_code=last_status_code,
-            )
-
-        raise AllCandidatesFailedError(
-            reason="all_candidates_failed",
-            candidate_keys=candidate_keys,
-            last_status_code=last_status_code,
-        )
-
-
-__all__ = [
-    "AsyncTaskOrchestrator",
-    "SubmitOutcome",
-    "AllCandidatesFailedError",
-    "UpstreamClientRequestError",
-    "CandidateUnsupportedError",
-    "CandidateSubmissionError",
-]

src/services/task/task_poller.py

@@ -0,0 +1,240 @@
+"""
+Task poller (Phase2)
+
+Provides a generic polling skeleton for async tasks.
+Currently wired with a video poller adapter.
+
+优化：HTTP 请求期间不持有数据库连接，避免阻塞其他请求。
+采用三阶段处理：准备数据 -> HTTP 请求 -> 更新数据库。
+"""
+
+from __future__ import annotations
+
+import asyncio
+from datetime import datetime, timezone
+from typing import Any, Protocol, runtime_checkable
+from uuid import uuid4
+
+from sqlalchemy.orm import Session
+
+from src.core.api_format.conversion.internal_video import InternalVideoPollResult
+from src.core.logger import logger
+from src.database import create_session
+from src.services.system.scheduler import get_scheduler
+from src.services.task.impl.video_poller import VideoPollContext, VideoTaskPollerAdapter
+
+
+@runtime_checkable
+class TaskPollerAdapter(Protocol):
+    task_type: str
+
+    # scheduler
+    job_id: str
+    job_name: str
+    interval_seconds: int
+
+    # distributed lock (optional, best-effort)
+    lock_key: str
+    lock_ttl: int
+
+    # execution
+    batch_size: int
+    concurrency: int
+    consecutive_failure_alert_threshold: int
+
+    def list_due_task_ids(self, db: Session, *, now: datetime, limit: int) -> list[str]: ...
+
+    def get_task(self, db: Session, task_id: str) -> Any | None: ...
+
+    # 分阶段处理方法（推荐使用）
+    async def prepare_poll_context(
+        self, db: Session, task: Any
+    ) -> Any: ...  # Returns context or error result
+
+    async def poll_task_http(self, ctx: Any) -> Any: ...  # Returns poll result
+
+    async def update_task_after_poll(
+        self,
+        task_id: str,
+        result: Any,
+        ctx: Any,
+        redis_client: Any | None,
+        error_exception: Exception | None = None,
+    ) -> None: ...
+
+    # 旧版方法（保留兼容性）
+    async def poll_single_task(
+        self, db: Session, task: Any, *, redis_client: Any | None
+    ) -> None: ...
+
+    def sanitize_error_message(self, message: str) -> str: ...
+
+
+class TaskPollerService:
+    """Generic background poller for async tasks."""
+
+    def __init__(self, adapter: TaskPollerAdapter) -> None:
+        self.adapter = adapter
+        self._lock = asyncio.Lock()
+        self.redis: Any | None = None
+        self._semaphore: asyncio.Semaphore | None = None
+        self._consecutive_failures = 0
+
+    async def start(self) -> None:
+        if self._semaphore is None:
+            self._semaphore = asyncio.Semaphore(self.adapter.concurrency)
+
+        # lazy import to avoid redis hard dependency in local runs
+        from src.clients.redis_client import get_redis_client
+
+        if self.redis is None:
+            self.redis = await get_redis_client(require_redis=False)
+
+        scheduler = get_scheduler()
+        scheduler.add_interval_job(
+            self.poll_pending_tasks,
+            seconds=self.adapter.interval_seconds,
+            job_id=self.adapter.job_id,
+            name=self.adapter.job_name,
+        )
+
+    async def stop(self) -> None:
+        scheduler = get_scheduler()
+        scheduler.remove_job(self.adapter.job_id)
+
+    async def poll_pending_tasks(self) -> None:
+        async with self._lock:
+            token = await self._acquire_redis_lock()
+            if token is None:
+                return
+
+            try:
+                with create_session() as db:
+                    now = datetime.now(timezone.utc)
+                    task_ids = self.adapter.list_due_task_ids(
+                        db, now=now, limit=self.adapter.batch_size
+                    )
+
+                if not task_ids:
+                    self._consecutive_failures = 0
+                    return
+
+                poll_results: list[bool] = []
+
+                if self._semaphore is None:
+                    self._semaphore = asyncio.Semaphore(self.adapter.concurrency)
+                semaphore = self._semaphore
+
+                async def poll_with_semaphore(task_id: str) -> None:
+                    async with semaphore:
+                        try:
+                            # ========== 阶段 1：准备数据（短暂持有连接）==========
+                            with create_session() as task_db:
+                                task_obj = self.adapter.get_task(task_db, task_id)
+                                if not task_obj:
+                                    logger.warning(
+                                        "[%s] Task %s disappeared during poll",
+                                        self.adapter.task_type,
+                                        task_id,
+                                    )
+                                    poll_results.append(True)
+                                    return
+
+                                ctx_or_result = await self.adapter.prepare_poll_context(
+                                    task_db, task_obj
+                                )
+
+                            # 检查是否是错误结果（而非上下文）
+                            if isinstance(ctx_or_result, InternalVideoPollResult):
+                                # 准备阶段就失败了，直接更新任务状态
+                                await self.adapter.update_task_after_poll(
+                                    task_id=task_id,
+                                    result=ctx_or_result,
+                                    ctx=None,  # type: ignore[arg-type]
+                                    redis_client=self.redis,
+                                )
+                                poll_results.append(True)
+                                return
+
+                            ctx: VideoPollContext = ctx_or_result
+
+                            # ========== 阶段 2：HTTP 请求（不持有数据库连接）==========
+                            error_exception: Exception | None = None
+                            try:
+                                result = await self.adapter.poll_task_http(ctx)
+                            except Exception as http_exc:
+                                # HTTP 请求失败，记录异常以便后续处理
+                                error_exception = http_exc
+                                result = InternalVideoPollResult(
+                                    status=None,  # type: ignore[arg-type]
+                                    error_message=str(http_exc),
+                                )
+
+                            # ========== 阶段 3：更新数据库（获取新连接）==========
+                            await self.adapter.update_task_after_poll(
+                                task_id=task_id,
+                                result=result,
+                                ctx=ctx,
+                                redis_client=self.redis,
+                                error_exception=error_exception,
+                            )
+                            poll_results.append(True)
+                        except Exception as exc:
+                            logger.exception(
+                                "[%s] Unexpected error polling task %s: %s",
+                                self.adapter.task_type,
+                                task_id,
+                                self.adapter.sanitize_error_message(str(exc)),
+                            )
+                            poll_results.append(False)
+
+                async with asyncio.TaskGroup() as tg:
+                    for tid in task_ids:
+                        tg.create_task(poll_with_semaphore(tid))
+
+                batch_failures = sum(1 for r in poll_results if r is False)
+                if batch_failures == len(task_ids):
+                    self._consecutive_failures += 1
+                    if (
+                        self._consecutive_failures
+                        >= self.adapter.consecutive_failure_alert_threshold
+                    ):
+                        logger.error(
+                            "[ALERT] %s poller: %d consecutive batches failed.",
+                            self.adapter.task_type,
+                            self._consecutive_failures,
+                        )
+                else:
+                    self._consecutive_failures = 0
+            finally:
+                await self._release_redis_lock(token)
+
+    async def _acquire_redis_lock(self) -> str | None:
+        if not self.redis:
+            return "no_redis"
+        token = str(uuid4())
+        acquired = await self.redis.set(
+            self.adapter.lock_key, token, nx=True, ex=self.adapter.lock_ttl
+        )
+        return token if acquired else None
+
+    async def _release_redis_lock(self, token: str) -> None:
+        if not self.redis or token == "no_redis":
+            return
+        script = """
+        if redis.call('GET', KEYS[1]) == ARGV[1] then
+            return redis.call('DEL', KEYS[1])
+        end
+        return 0
+        """
+        await self.redis.eval(script, 1, self.adapter.lock_key, token)
+
+
+_task_poller: TaskPollerService | None = None
+
+
+def get_task_poller() -> TaskPollerService:
+    global _task_poller
+    if _task_poller is None:
+        _task_poller = TaskPollerService(VideoTaskPollerAdapter())
+    return _task_poller

src/services/usage/service.py

@@ -576,6 +576,12 @@ class UsageService:
         """更新已存在的 Usage 记录（内部方法）"""
         # 更新关键字段
         existing_usage.provider_name = usage_params["provider_name"]
+        existing_usage.model = usage_params["model"]
+        existing_usage.request_type = usage_params["request_type"]
+        existing_usage.api_format = usage_params["api_format"]
+        existing_usage.endpoint_api_format = usage_params["endpoint_api_format"]
+        existing_usage.has_format_conversion = usage_params["has_format_conversion"]
+        existing_usage.is_stream = usage_params["is_stream"]
         existing_usage.status = usage_params["status"]
         existing_usage.status_code = usage_params["status_code"]
         existing_usage.error_message = usage_params["error_message"]
@@ -621,6 +627,10 @@ class UsageService:
         existing_usage.provider_endpoint_id = usage_params["provider_endpoint_id"]
         existing_usage.provider_api_key_id = usage_params["provider_api_key_id"]
 
+        # 更新元数据（如 billing_snapshot/dimensions 等）
+        if usage_params.get("request_metadata") is not None:
+            existing_usage.request_metadata = usage_params["request_metadata"]
+
         # 更新模型映射信息
         if target_model is not None:
             existing_usage.target_model = target_model
@@ -1000,6 +1010,11 @@ class UsageService:
                 .values(monthly_used_usd=Provider.monthly_used_usd + actual_total_cost)
             )
 
+        # 结算标记：record_usage_async 写入的 Usage 通常为终态记录
+        if status not in ("pending", "streaming"):
+            usage.billing_status = "settled"
+            usage.finalized_at = datetime.now(timezone.utc)
+
         db.commit()  # 立即提交事务，释放数据库锁
         return usage
 
@@ -1172,6 +1187,11 @@ class UsageService:
                 .values(monthly_used_usd=Provider.monthly_used_usd + actual_total_cost)
             )
 
+        # 结算标记：终态请求写入 settled + finalized_at
+        if status not in ("pending", "streaming"):
+            usage.billing_status = "settled"
+            usage.finalized_at = datetime.now(timezone.utc)
+
         # 提交事务
         try:
             db.commit()
@@ -1297,17 +1317,39 @@ class UsageService:
             is_free_tier=is_free_tier,
         )
 
-        # Upsert（与 record_usage 保持一致）
+        # Upsert（并发幂等：优先用 billing_status 作为结算闸门）
+        from sqlalchemy import update
+
         existing_usage = db.query(Usage).filter(Usage.request_id == request_id).first()
         if existing_usage:
-            # 避免重复记账：若已是终态记录，直接返回（批量接口也采用该策略）
-            if existing_usage.status not in ("pending", "streaming"):
+            # 避免重复记账：若已结算/作废，直接返回（防止并发重复加计数）
+            if getattr(existing_usage, "billing_status", None) in ("settled", "void"):
                 logger.debug(
-                    "record_usage_with_custom_cost: request_id=%s already finalized (status=%s), skip",
+                    "record_usage_with_custom_cost: request_id=%s already finalized (billing_status=%s), skip",
                     request_id,
-                    existing_usage.status,
+                    getattr(existing_usage, "billing_status", None),
                 )
                 return existing_usage
+
+            # 并发闸门：只有 billing_status='pending' 的那一次调用可以继续
+            now = datetime.now(timezone.utc)
+            claim = db.execute(
+                update(Usage)
+                .where(
+                    Usage.request_id == request_id,
+                    Usage.billing_status == "pending",
+                )
+                .values(billing_status="settled", finalized_at=now)
+            )
+            if claim.rowcount != 1:
+                # 已被其他 worker 抢先处理（或被 VOID）
+                latest = db.query(Usage).filter(Usage.request_id == request_id).first()
+                return latest or existing_usage
+
+            # 同步 ORM 对象（避免后续代码读到旧值）
+            existing_usage.billing_status = "settled"
+            existing_usage.finalized_at = now
+
             cls._update_existing_usage(existing_usage, usage_params, target_model)
             usage = existing_usage
         else:
@@ -1382,6 +1424,11 @@ class UsageService:
                 .values(monthly_used_usd=Provider.monthly_used_usd + actual_total_cost)
             )
 
+        # 结算标记：record_usage_with_custom_cost 写入/更新的 Usage 通常为终态记录
+        if status not in ("pending", "streaming"):
+            usage.billing_status = "settled"
+            usage.finalized_at = datetime.now(timezone.utc)
+
         try:
             db.commit()
         except Exception as e:
@@ -2147,35 +2194,31 @@ class UsageService:
     # ========== 请求状态追踪方法 ==========
 
     @classmethod
-    def create_pending_usage(
+    def begin_pending_usage(
         cls,
         db: Session,
         request_id: str,
         user: User | None,
         api_key: ApiKey | None,
         model: str,
+        *,
         is_stream: bool = False,
+        request_type: str = "chat",
         api_format: str | None = None,
         request_headers: dict[str, Any] | None = None,
         request_body: Any | None = None,
     ) -> Usage:
         """
-        创建 pending 状态的使用记录（在请求开始时调用）
+        创建（或返回已有）pending Usage 记录，但**不提交事务**。
 
-        Args:
-            db: 数据库会话
-            request_id: 请求ID
-            user: 用户对象
-            api_key: API Key 对象
-            model: 模型名称
-            is_stream: 是否流式请求
-            api_format: API 格式
-            request_headers: 请求头
-            request_body: 请求体
-
-        Returns:
-            创建的 Usage 记录
+        适用场景：
+        - ApplicationService 在同一事务内创建 pending usage + task + candidates
+        - submit 幂等：重复调用同一 request_id 时返回已有记录
         """
+        existing = db.query(Usage).filter(Usage.request_id == request_id).first()
+        if existing:
+            return existing
+
         # 根据配置决定是否记录请求详情
         should_log_headers = SystemConfigService.should_log_headers(db)
         should_log_body = SystemConfigService.should_log_body(db)
@@ -2204,21 +2247,359 @@ class UsageService:
             output_tokens=0,
             total_tokens=0,
             total_cost_usd=0.0,
-            request_type="chat",
+            request_type=request_type,
             api_format=api_format,
             is_stream=is_stream,
             status="pending",
+            billing_status="pending",
             request_headers=processed_request_headers,
             request_body=processed_request_body,
         )
 
         db.add(usage)
+        db.flush()
+        return usage
+
+    @classmethod
+    def create_pending_usage(
+        cls,
+        db: Session,
+        request_id: str,
+        user: User | None,
+        api_key: ApiKey | None,
+        model: str,
+        is_stream: bool = False,
+        request_type: str = "chat",
+        api_format: str | None = None,
+        request_headers: dict[str, Any] | None = None,
+        request_body: Any | None = None,
+    ) -> Usage:
+        """
+        创建 pending 状态的使用记录（在请求开始时调用）
+
+        Args:
+            db: 数据库会话
+            request_id: 请求ID
+            user: 用户对象
+            api_key: API Key 对象
+            model: 模型名称
+            is_stream: 是否流式请求
+            api_format: API 格式
+            request_headers: 请求头
+            request_body: 请求体
+
+        Returns:
+            创建的 Usage 记录
+        """
+        usage = cls.begin_pending_usage(
+            db,
+            request_id=request_id,
+            user=user,
+            api_key=api_key,
+            model=model,
+            is_stream=is_stream,
+            request_type=request_type,
+            api_format=api_format,
+            request_headers=request_headers,
+            request_body=request_body,
+        )
         db.commit()
 
         logger.debug(f"创建 pending 使用记录: request_id={request_id}, model={model}")
 
         return usage
 
+    # ========== billing_status 并发幂等 finalize ==========
+
+    @classmethod
+    def finalize_settled(
+        cls,
+        db: Session,
+        request_id: str,
+        *,
+        total_cost_usd: float,
+        request_cost_usd: float | None = None,
+        status: str = "completed",
+        status_code: int = 200,
+        error_message: str | None = None,
+        response_time_ms: int | None = None,
+        billing_snapshot: dict[str, Any] | None = None,
+        extra_metadata: dict[str, Any] | None = None,
+    ) -> bool:
+        """
+        并发安全的幂等 finalize（settled）。
+
+        约定：
+        - 仅当 billing_status='pending' 时才会生效（rowcount==1）
+        - 不在本方法内 commit，由调用方决定事务提交时机
+        """
+        from sqlalchemy import update
+
+        now = datetime.now(timezone.utc)
+        cost = float(total_cost_usd)
+        request_cost = float(request_cost_usd) if request_cost_usd is not None else cost
+
+        result = db.execute(
+            update(Usage)
+            .where(
+                Usage.request_id == request_id,
+                Usage.billing_status == "pending",
+            )
+            .values(
+                billing_status="settled",
+                finalized_at=now,
+                total_cost_usd=cost,
+                request_cost_usd=request_cost,
+                status=status,
+                status_code=status_code,
+                error_message=error_message,
+                response_time_ms=response_time_ms,
+            )
+        )
+        if result.rowcount != 1:
+            return False
+
+        # 写入审计快照（只在本次 finalize 生效时执行）
+        usage = db.query(Usage).filter(Usage.request_id == request_id).first()
+        if usage:
+            metadata = usage.request_metadata or {}
+            if billing_snapshot is not None:
+                metadata["billing_snapshot"] = billing_snapshot
+            if extra_metadata:
+                metadata.update(extra_metadata)
+            usage.request_metadata = metadata
+
+        return True
+
+    @classmethod
+    def finalize_void(
+        cls,
+        db: Session,
+        request_id: str,
+        *,
+        reason: str | None = None,
+        status_code: int = 499,
+    ) -> bool:
+        """
+        并发安全的幂等 finalize（void，不收费）。
+
+        约定：
+        - 仅当 billing_status='pending' 时才会生效（rowcount==1）
+        - 不在本方法内 commit，由调用方决定事务提交时机
+        """
+        from sqlalchemy import update
+
+        now = datetime.now(timezone.utc)
+        result = db.execute(
+            update(Usage)
+            .where(
+                Usage.request_id == request_id,
+                Usage.billing_status == "pending",
+            )
+            .values(
+                billing_status="void",
+                finalized_at=now,
+                total_cost_usd=0.0,
+                request_cost_usd=0.0,
+                status="cancelled",
+                status_code=status_code,
+                error_message=reason,
+                response_time_ms=None,
+            )
+        )
+        return result.rowcount == 1
+
+    @classmethod
+    def finalize_submitted(
+        cls,
+        db: Session,
+        request_id: str,
+        *,
+        provider_name: str,
+        provider_id: str | None = None,
+        provider_endpoint_id: str | None = None,
+        provider_api_key_id: str | None = None,
+        response_time_ms: int | None = None,
+        status_code: int = 200,
+        endpoint_api_format: str | None = None,
+        provider_request_headers: dict[str, Any] | None = None,
+        response_headers: dict[str, Any] | None = None,
+        response_body: Any | None = None,
+    ) -> bool:
+        """
+        异步任务提交成功时的幂等结算。
+
+        将 pending 使用记录标记为 settled，费用暂时为 0。
+        后续轮询完成后通过 update_settled_billing 更新实际费用。
+
+        约定：
+        - 仅当 billing_status='pending' 时才会生效（rowcount==1）
+        - 不在本方法内 commit，由调用方决定事务提交时机
+        """
+        from sqlalchemy import update
+
+        now = datetime.now(timezone.utc)
+
+        # 处理响应头和响应体
+        should_log_headers = SystemConfigService.should_log_headers(db)
+        should_log_body = SystemConfigService.should_log_body(db)
+
+        processed_provider_headers = None
+        if should_log_headers and provider_request_headers:
+            processed_provider_headers = SystemConfigService.mask_sensitive_headers(
+                db, provider_request_headers
+            )
+
+        processed_response_headers = None
+        if should_log_headers and response_headers:
+            processed_response_headers = dict(response_headers)
+
+        processed_response_body = None
+        if should_log_body and response_body:
+            processed_response_body = SystemConfigService.truncate_body(
+                db, response_body, is_request=False
+            )
+
+        values: dict[str, Any] = {
+            "billing_status": "settled",
+            "finalized_at": now,
+            "total_cost_usd": 0.0,
+            "request_cost_usd": 0.0,
+            "status": "completed",
+            "status_code": status_code,
+            "response_time_ms": response_time_ms,
+            "provider_name": provider_name,
+            "provider_id": provider_id,
+            "provider_endpoint_id": provider_endpoint_id,
+            "provider_api_key_id": provider_api_key_id,
+            "endpoint_api_format": endpoint_api_format,
+        }
+
+        if processed_provider_headers is not None:
+            values["provider_request_headers"] = processed_provider_headers
+        if processed_response_headers is not None:
+            values["response_headers"] = processed_response_headers
+        if processed_response_body is not None:
+            values["response_body"] = processed_response_body
+
+        result = db.execute(
+            update(Usage)
+            .where(
+                Usage.request_id == request_id,
+                Usage.billing_status == "pending",
+            )
+            .values(**values)
+        )
+        return result.rowcount == 1
+
+    @classmethod
+    def update_settled_billing(
+        cls,
+        db: Session,
+        request_id: str,
+        *,
+        total_cost_usd: float,
+        request_cost_usd: float | None = None,
+        status: str = "completed",
+        status_code: int = 200,
+        error_message: str | None = None,
+        response_time_ms: int | None = None,
+        billing_snapshot: dict[str, Any] | None = None,
+        extra_metadata: dict[str, Any] | None = None,
+    ) -> bool:
+        """
+        更新已结算记录的计费信息（用于异步任务轮询完成后）。
+
+        与 finalize_settled 不同：
+        - finalize_settled: pending -> settled（首次结算）
+        - update_settled_billing: settled -> settled（更新费用）
+
+        约定：
+        - 仅当 billing_status='settled' 时才会生效
+        - 不在本方法内 commit，由调用方决定事务提交时机
+        """
+        from sqlalchemy import update
+
+        now = datetime.now(timezone.utc)
+        cost = float(total_cost_usd)
+        request_cost = float(request_cost_usd) if request_cost_usd is not None else cost
+
+        values: dict[str, Any] = {
+            "total_cost_usd": cost,
+            "request_cost_usd": request_cost,
+            "status": status,
+            "status_code": status_code,
+        }
+        if error_message is not None:
+            values["error_message"] = error_message
+        if response_time_ms is not None:
+            values["response_time_ms"] = response_time_ms
+
+        result = db.execute(
+            update(Usage)
+            .where(
+                Usage.request_id == request_id,
+                Usage.billing_status == "settled",
+            )
+            .values(**values)
+        )
+        if result.rowcount != 1:
+            return False
+
+        # 写入审计快照
+        usage = db.query(Usage).filter(Usage.request_id == request_id).first()
+        if usage:
+            metadata = usage.request_metadata or {}
+            if billing_snapshot is not None:
+                metadata["billing_snapshot"] = billing_snapshot
+            if extra_metadata:
+                metadata.update(extra_metadata)
+            metadata["billing_updated_at"] = now.isoformat()
+            usage.request_metadata = metadata
+
+        return True
+
+    @classmethod
+    def void_settled(
+        cls,
+        db: Session,
+        request_id: str,
+        *,
+        reason: str | None = None,
+        status_code: int = 499,
+    ) -> bool:
+        """
+        将已结算的记录作废（用于异步任务取消）。
+
+        与 finalize_void 不同：
+        - finalize_void: pending -> void（未结算时作废）
+        - void_settled: settled -> void（已结算后取消，费用归零）
+
+        约定：
+        - 仅当 billing_status='settled' 时才会生效
+        - 不在本方法内 commit，由调用方决定事务提交时机
+        """
+        from sqlalchemy import update
+
+        now = datetime.now(timezone.utc)
+        result = db.execute(
+            update(Usage)
+            .where(
+                Usage.request_id == request_id,
+                Usage.billing_status == "settled",
+            )
+            .values(
+                billing_status="void",
+                finalized_at=now,
+                total_cost_usd=0.0,
+                request_cost_usd=0.0,
+                status="cancelled",
+                status_code=status_code,
+                error_message=reason,
+            )
+        )
+        return result.rowcount == 1
+
     @classmethod
     def update_usage_status(
         cls,
@@ -2235,6 +2616,7 @@ class UsageService:
         api_format: str | None = None,
         endpoint_api_format: str | None = None,
         has_format_conversion: bool | None = None,
+        status_code: int | None = None,
     ) -> Usage | None:
         """
         快速更新使用记录状态
@@ -2253,6 +2635,7 @@ class UsageService:
             api_format: API 格式（可选，用于获取按格式配置的倍率）
             endpoint_api_format: 端点原生 API 格式（可选）
             has_format_conversion: 是否发生了格式转换（可选）
+            status_code: HTTP 状态码（可选）
 
         Returns:
             更新后的 Usage 记录，如果未找到则返回 None
@@ -2295,6 +2678,16 @@ class UsageService:
             usage.endpoint_api_format = endpoint_api_format
         if has_format_conversion is not None:
             usage.has_format_conversion = has_format_conversion
+        if status_code is not None:
+            usage.status_code = status_code
+
+        # 结算状态：当请求进入终态时，将 billing_status 标记为 settled
+        # 注意：取消是否应 VOID/部分结算由更高层策略决定；这里默认终态均视为已结算。
+        if status in ("completed", "failed", "cancelled"):
+            if getattr(usage, "billing_status", None) == "pending":
+                usage.billing_status = "settled"
+            if getattr(usage, "finalized_at", None) is None:
+                usage.finalized_at = datetime.now(timezone.utc)
 
         db.commit()
 

src/services/video/__init__.py

@@ -1,10 +0,0 @@
-"""
-视频相关服务
-"""
-
-from src.services.video.task_poller import VideoTaskPollerService, get_video_task_poller
-
-__all__ = [
-    "VideoTaskPollerService",
-    "get_video_task_poller",
-]

src/services/video/task_poller.py

@@ -1,446 +0,0 @@
-"""
-视频任务后台轮询服务
-"""
-
-from __future__ import annotations
-
-import asyncio
-from datetime import datetime, timedelta, timezone
-from typing import Any
-from uuid import uuid4
-
-from sqlalchemy.orm import Session
-
-from src.api.handlers.base.request_builder import ProviderAuthInfo, get_provider_auth
-from src.api.handlers.base.video_handler_base import (
-    normalize_gemini_operation_id,
-    sanitize_error_message,
-)
-from src.clients.http_client import HTTPClientPool
-from src.clients.redis_client import get_redis_client
-from src.config.settings import config
-from src.core.api_format import (
-    build_upstream_headers_for_endpoint,
-    get_extra_headers_from_endpoint,
-    make_signature_key,
-)
-from src.core.api_format.conversion.internal_video import InternalVideoPollResult, VideoStatus
-from src.core.api_format.conversion.normalizers.gemini import GeminiNormalizer
-from src.core.api_format.conversion.normalizers.openai import OpenAINormalizer
-from src.core.crypto import crypto_service
-from src.core.logger import logger
-from src.database import create_session
-from src.models.database import ProviderAPIKey, ProviderEndpoint, VideoTask
-from src.services.system.scheduler import get_scheduler
-from src.services.task.impl.video_telemetry import VideoTelemetry
-
-# 永久性错误指示词（用于降级判断，不应重试）
-_PERMANENT_ERROR_INDICATORS = frozenset(
-    {
-        "not found",
-        "404",
-        "unauthorized",
-        "401",
-        "forbidden",
-        "403",
-        "invalid request",
-        "invalid api key",
-        "does not exist",
-    }
-)
-
-
-class PollHTTPError(RuntimeError):
-    """HTTP 轮询错误，携带状态码便于区分临时/永久错误"""
-
-    def __init__(self, status_code: int, message: str):
-        super().__init__(message)
-        self.status_code = status_code
-
-
-class VideoTaskPollerService:
-    """后台轮询视频生成任务状态"""
-
-    LOCK_KEY = "video_task_poller:lock"
-    LOCK_TTL = 60
-    MAX_BACKOFF_SECONDS = 300
-    # 连续失败告警阈值
-    CONSECUTIVE_FAILURE_ALERT_THRESHOLD = 5
-
-    def __init__(self) -> None:
-        self._lock = asyncio.Lock()
-        self.redis = None
-        self._openai_normalizer = OpenAINormalizer()
-        self._gemini_normalizer = GeminiNormalizer()
-        # 追踪连续失败次数（用于告警）
-        self._consecutive_failures = 0
-        # 从配置读取参数
-        self._batch_size = config.video_poll_batch_size
-        self._concurrency = config.video_poll_concurrency
-        # Semaphore 延迟初始化，避免在事件循环外创建
-        self._semaphore: asyncio.Semaphore | None = None
-
-    async def start(self) -> None:
-        # 在事件循环内初始化 Semaphore
-        if self._semaphore is None:
-            self._semaphore = asyncio.Semaphore(self._concurrency)
-        if self.redis is None:
-            self.redis = await get_redis_client(require_redis=False)
-
-        scheduler = get_scheduler()
-        scheduler.add_interval_job(
-            self.poll_pending_tasks,
-            seconds=config.video_poll_interval_seconds,
-            job_id="video_task_poller",
-            name="视频任务轮询",
-        )
-
-    async def stop(self) -> None:
-        """停止轮询服务"""
-        scheduler = get_scheduler()
-        scheduler.remove_job("video_task_poller")
-
-    async def poll_pending_tasks(self) -> None:
-        async with self._lock:
-            token = await self._acquire_redis_lock()
-            if token is None:
-                return
-
-            try:
-                with create_session() as db:
-                    now = datetime.now(timezone.utc)
-                    tasks = (
-                        db.query(VideoTask)
-                        .filter(
-                            VideoTask.status.in_(
-                                [
-                                    VideoStatus.SUBMITTED.value,
-                                    VideoStatus.QUEUED.value,
-                                    VideoStatus.PROCESSING.value,
-                                ]
-                            ),
-                            VideoTask.next_poll_at <= now,
-                            VideoTask.poll_count < VideoTask.max_poll_count,
-                        )
-                        .order_by(VideoTask.next_poll_at.asc())
-                        .limit(self._batch_size)
-                        .all()
-                    )
-
-                    if not tasks:
-                        # 无任务时重置连续失败计数
-                        self._consecutive_failures = 0
-                        return
-
-                    # 提取任务 ID 列表，释放查询 session 后逐个轮询
-                    task_ids = [t.id for t in tasks]
-
-                # 并发轮询：每个任务使用独立 session，避免共享 session 的并发风险
-                poll_results: list[bool] = []
-
-                # 确保 semaphore 已初始化（在 start 中初始化，此处防御性检查）
-                if self._semaphore is None:
-                    self._semaphore = asyncio.Semaphore(self._concurrency)
-                semaphore = self._semaphore
-
-                async def poll_with_semaphore(task_id: str) -> None:
-                    """带信号量的轮询，结果写入 poll_results"""
-                    async with semaphore:
-                        try:
-                            with create_session() as task_db:
-                                task_obj = task_db.query(VideoTask).get(task_id)
-                                if not task_obj:
-                                    logger.warning("Task %s disappeared during poll", task_id)
-                                    poll_results.append(True)
-                                    return
-                                await self._poll_single_task(task_db, task_obj)
-                                task_db.commit()
-                            poll_results.append(True)
-                        except Exception as exc:
-                            logger.exception(
-                                "Unexpected error polling task %s: %s",
-                                task_id,
-                                sanitize_error_message(str(exc)),
-                            )
-                            poll_results.append(False)
-
-                async with asyncio.TaskGroup() as tg:
-                    for tid in task_ids:
-                        tg.create_task(poll_with_semaphore(tid))
-
-                batch_failures = sum(1 for r in poll_results if r is False)
-
-                # 更新连续失败计数并检查告警阈值
-                if batch_failures == len(task_ids):
-                    self._consecutive_failures += 1
-                    if self._consecutive_failures >= self.CONSECUTIVE_FAILURE_ALERT_THRESHOLD:
-                        logger.error(
-                            "[ALERT] Video task poller: %d consecutive batches failed. "
-                            "Provider connectivity or configuration issue suspected.",
-                            self._consecutive_failures,
-                        )
-                else:
-                    self._consecutive_failures = 0
-            finally:
-                await self._release_redis_lock(token)
-
-    async def _poll_single_task(self, db: Session, task: VideoTask) -> None:
-        try:
-            result = await self._poll_task_status(db, task)
-            if result.status == VideoStatus.COMPLETED:
-                task.status = VideoStatus.COMPLETED.value
-                task.video_url = result.video_url
-                task.video_expires_at = result.expires_at
-                task.completed_at = datetime.now(timezone.utc)
-                task.progress_percent = 100
-                # 存储多视频 URL（Gemini sampleCount > 1 时）
-                if result.video_urls:
-                    task.video_urls = result.video_urls
-                # 保存上游原始响应（用于审计/重算）
-                self._attach_poll_raw_response(task, result)
-            elif result.status == VideoStatus.FAILED:
-                task.status = VideoStatus.FAILED.value
-                task.error_code = result.error_code
-                task.error_message = result.error_message
-                task.completed_at = datetime.now(timezone.utc)
-                self._attach_poll_raw_response(task, result)
-            else:
-                task.poll_count += 1
-                task.progress_percent = result.progress_percent
-                task.next_poll_at = datetime.now(timezone.utc) + timedelta(
-                    seconds=task.poll_interval_seconds
-                )
-        except Exception as exc:
-            task.poll_count += 1
-            error_msg = sanitize_error_message(str(exc))
-            logger.warning("Poll error for task %s: %s", task.id, error_msg)
-            task.progress_message = f"Poll error: {error_msg}"
-
-            # 区分临时性错误和永久性错误
-            status_code = exc.status_code if isinstance(exc, PollHTTPError) else None
-            is_permanent = self._is_permanent_error(exc, status_code=status_code)
-            if is_permanent:
-                task.status = VideoStatus.FAILED.value
-                task.error_code = "poll_permanent_error"
-                task.error_message = error_msg
-                task.completed_at = datetime.now(timezone.utc)
-            else:
-                # 临时性错误：指数退避重试
-                backoff = min(
-                    task.poll_interval_seconds * (2 ** min(task.retry_count, 5)),
-                    self.MAX_BACKOFF_SECONDS,
-                )
-                task.retry_count += 1
-                task.next_poll_at = datetime.now(timezone.utc) + timedelta(seconds=backoff)
-
-        # 检查是否超过最大轮询次数（超时）
-        task.updated_at = datetime.now(timezone.utc)
-        if task.poll_count >= task.max_poll_count and task.status not in [
-            VideoStatus.COMPLETED.value,
-            VideoStatus.FAILED.value,
-            VideoStatus.CANCELLED.value,
-        ]:
-            task.status = VideoStatus.FAILED.value
-            task.error_code = "poll_timeout"
-            task.error_message = f"Task timed out after {task.poll_count} polls"
-            task.completed_at = datetime.now(timezone.utc)
-
-        # 终态写入 Usage（复用外层 per-task session）
-        if task.status in (VideoStatus.COMPLETED.value, VideoStatus.FAILED.value):
-            try:
-                await VideoTelemetry(db, redis_client=self.redis).record_terminal_usage(task)
-            except Exception as exc:
-                logger.exception(
-                    "Failed to record video usage for task=%s: %s",
-                    task.id,
-                    sanitize_error_message(str(exc)),
-                )
-
-    def _attach_poll_raw_response(self, task: VideoTask, result: InternalVideoPollResult) -> None:
-        if not result.raw_response:
-            return
-        if task.request_metadata is None:
-            task.request_metadata = {}
-        # 仅在终态写一次，避免污染 request_metadata
-        task.request_metadata["poll_raw_response"] = result.raw_response
-
-    def _is_permanent_error(self, exc: Exception, status_code: int | None = None) -> bool:
-        """判断是否为永久性错误（不应重试）"""
-        # 优先使用 HTTP 状态码判断
-        if status_code is not None:
-            # 4xx 客户端错误（除 429 限流）通常是永久性错误
-            return 400 <= status_code < 500 and status_code != 429
-
-        # 降级到字符串匹配
-        error_msg = str(exc).lower()
-        return any(indicator in error_msg for indicator in _PERMANENT_ERROR_INDICATORS)
-
-    async def _poll_task_status(self, db: Session, task: VideoTask) -> InternalVideoPollResult:
-        if not task.endpoint_id or not task.key_id:
-            return InternalVideoPollResult(
-                status=VideoStatus.FAILED,
-                error_code="missing_provider_info",
-                error_message="Task missing endpoint_id or key_id",
-            )
-        endpoint = self._get_endpoint(db, task.endpoint_id)
-        key = self._get_key(db, task.key_id)
-        if not key.api_key:
-            return InternalVideoPollResult(
-                status=VideoStatus.FAILED,
-                error_code="provider_config_error",
-                error_message="Provider key not properly configured",
-            )
-        try:
-            upstream_key = crypto_service.decrypt(key.api_key)
-        except Exception:
-            logger.warning("Failed to decrypt provider key for task %s", task.id)
-            return InternalVideoPollResult(
-                status=VideoStatus.FAILED,
-                error_code="decryption_error",
-                error_message="Failed to decrypt provider key",
-            )
-
-        provider_format = (task.provider_api_format or "").strip().lower()
-        if not provider_format:
-            provider_format = make_signature_key(
-                str(getattr(endpoint, "api_family", "")).strip().lower(),
-                str(getattr(endpoint, "endpoint_kind", "")).strip().lower(),
-            )
-
-        if provider_format.startswith("gemini:"):
-            auth_info = await get_provider_auth(endpoint, key)
-            return await self._poll_gemini(task, endpoint, upstream_key, auth_info)
-        return await self._poll_openai(task, endpoint, upstream_key)
-
-    async def _poll_openai(
-        self,
-        task: VideoTask,
-        endpoint: ProviderEndpoint,
-        upstream_key: str,
-    ) -> InternalVideoPollResult:
-        if not task.external_task_id:
-            return InternalVideoPollResult(
-                status=VideoStatus.FAILED,
-                error_code="missing_external_task_id",
-                error_message="Task missing external_task_id",
-            )
-        url = self._build_openai_url(endpoint.base_url, task.external_task_id)
-        endpoint_sig = (task.provider_api_format or "").strip().lower() or make_signature_key(
-            str(getattr(endpoint, "api_family", "")).strip().lower(),
-            str(getattr(endpoint, "endpoint_kind", "")).strip().lower(),
-        )
-        headers = self._build_headers(endpoint_sig, upstream_key, endpoint)
-
-        client = await HTTPClientPool.get_default_client_async()
-        response = await client.get(url, headers=headers)
-        if response.status_code >= 400:
-            raise PollHTTPError(
-                response.status_code,
-                sanitize_error_message(response.text or "Poll error"),
-            )
-
-        payload = response.json()
-        return self._openai_normalizer.video_poll_to_internal(payload)
-
-    async def _poll_gemini(
-        self,
-        task: VideoTask,
-        endpoint: ProviderEndpoint,
-        upstream_key: str,
-        auth_info: ProviderAuthInfo | None,
-    ) -> InternalVideoPollResult:
-        if not task.external_task_id:
-            return InternalVideoPollResult(
-                status=VideoStatus.FAILED,
-                error_code="missing_external_task_id",
-                error_message="Task missing external_task_id",
-            )
-        operation_name = normalize_gemini_operation_id(task.external_task_id)
-        url = self._build_gemini_url(endpoint.base_url, operation_name)
-        endpoint_sig = (task.provider_api_format or "").strip().lower() or make_signature_key(
-            str(getattr(endpoint, "api_family", "")).strip().lower(),
-            str(getattr(endpoint, "endpoint_kind", "")).strip().lower(),
-        )
-        headers = self._build_headers(endpoint_sig, upstream_key, endpoint, auth_info)
-
-        client = await HTTPClientPool.get_default_client_async()
-        response = await client.get(url, headers=headers)
-        if response.status_code >= 400:
-            raise PollHTTPError(
-                response.status_code,
-                sanitize_error_message(response.text or "Poll error"),
-            )
-
-        payload = response.json()
-        return self._gemini_normalizer.video_poll_to_internal(payload)
-
-    def _build_openai_url(self, base_url: str | None, task_id: str) -> str:
-        base = (base_url or "https://api.openai.com").rstrip("/")
-        if base.endswith("/v1"):
-            return f"{base}/videos/{task_id}"
-        return f"{base}/v1/videos/{task_id}"
-
-    def _build_gemini_url(self, base_url: str | None, operation_name: str) -> str:
-        base = (base_url or "https://generativelanguage.googleapis.com").rstrip("/")
-        if base.endswith("/v1beta"):
-            base = base[: -len("/v1beta")]
-        return f"{base}/v1beta/{operation_name}"
-
-    def _build_headers(
-        self,
-        endpoint_sig: str,
-        upstream_key: str,
-        endpoint: ProviderEndpoint,
-        auth_info: ProviderAuthInfo | None = None,
-    ) -> dict[str, str]:
-        extra_headers = get_extra_headers_from_endpoint(endpoint)
-        headers = build_upstream_headers_for_endpoint(
-            {},
-            endpoint_sig,
-            upstream_key,
-            endpoint_headers=extra_headers,
-        )
-        if auth_info:
-            headers.pop("x-goog-api-key", None)
-            headers[auth_info.auth_header] = auth_info.auth_value
-        return headers
-
-    def _get_endpoint(self, db: Session, endpoint_id: str) -> ProviderEndpoint:
-        endpoint = db.query(ProviderEndpoint).filter(ProviderEndpoint.id == endpoint_id).first()
-        if not endpoint:
-            raise RuntimeError("Provider endpoint not found")
-        return endpoint
-
-    def _get_key(self, db: Session, key_id: str) -> ProviderAPIKey:
-        key = db.query(ProviderAPIKey).filter(ProviderAPIKey.id == key_id).first()
-        if not key:
-            raise RuntimeError("Provider key not found")
-        return key
-
-    async def _acquire_redis_lock(self) -> str | None:
-        if not self.redis:
-            return "no_redis"
-        token = str(uuid4())
-        acquired = await self.redis.set(self.LOCK_KEY, token, nx=True, ex=self.LOCK_TTL)
-        return token if acquired else None
-
-    async def _release_redis_lock(self, token: str) -> None:
-        if not self.redis or token == "no_redis":
-            return
-        script = """
-        if redis.call('GET', KEYS[1]) == ARGV[1] then
-            return redis.call('DEL', KEYS[1])
-        end
-        return 0
-        """
-        await self.redis.eval(script, 1, self.LOCK_KEY, token)
-
-
-_video_task_poller: VideoTaskPollerService | None = None
-
-
-def get_video_task_poller() -> VideoTaskPollerService:
-    global _video_task_poller
-    if _video_task_poller is None:
-        _video_task_poller = VideoTaskPollerService()
-    return _video_task_poller