refactor(shared): 抽取标题模板为公共组件，智能剪辑与AI数字人复用

- 新增公共组件 @/components/title:
  - TitleTemplateSelector: 大卡片网格选择器（我的/系统模板分组、新建/编辑/复制/导出/删除）
  - TitleTemplateEditor: 模板编辑器弹窗（预览+参数Tab）
  - TitleStyleParamsTab: 参数Tab面板（基础/描边/阴影/背景/排版）
  - TitleMiniPreview: Canvas大字预览（迁移自pages内部）
  - settings.ts: 公共 TitleStyleSettings 类型+DEFAULT（不含业务字段 aiAutoSelect/title/coverTitle）
  - utils.ts: 重构为依赖公共类型，消除 pages/generate/types 反向依赖
  - position-options.ts: 公共 POSITION_OPTIONS 常量
  - index.ts: 统一 barrel export
- 智能剪辑 TitleStylePanel（pages/generate/components/title/）：
  - 改为 import @/components/title/TitleMiniPreview，旧路径改为 re-export
  - 修复 FillPreview 类型，消除 TitleStyleSettings/TitleSettings 类型不匹配
  - generate/constants.ts 重导出 POSITION_OPTIONS 自公共层
- AI数字人 PanelTitleConfig：
  - enableTemplates 改为 true，启用大卡片网格模式（与智能剪辑 Step3 一致）
  - 共享同一份 TITLE_PRESETS 系统预设、localStorage 用户模板
- tsc/eslint/prettier/vite build 全部通过

.env.example

@@ -79,33 +79,14 @@ CELERY_BROKER_URL=redis://localhost:6379/0
 CELERY_RESULT_BACKEND=redis://localhost:6379/1
 
 
-# ==================== Worker 配置（#2073 队列分流） ====================
-#
-# 容器内跑三个独立进程：beat（只发定时任务）+ generation worker（实时高优）
-# + transcode worker（后台批量/清理）。三个进程的并发与开关独立配置。
+# ==================== Worker 配置 ====================
 
 # Worker 进程名称
 WORKER_NAME=xiaoxia-saas-worker
 
-# 总并发参考（兼容旧变量）：
-# - 若 GENERATION_CONCURRENCY 与 TRANSCODE_CONCURRENCY 都未显式设置，
-#   entrypoint 会按此总数对半分配（gen=ceil(total/2), trans=剩余，各至少 1）；
-# - 任一个 *_CONCURRENCY 显式设置后，按显式值生效，忽略此变量对应部分。
+# Worker 并发数（同时执行的任务数）
 WORKER_CONCURRENCY=4
 
-# Generation worker 并发数（用户实时任务：视频生成/TTS/音色克隆/lipsync/数字人）
-# 实时链路对延迟敏感，建议 2C 以上机器设为 2；高负载场景可加到 4。
-GENERATION_CONCURRENCY=2
-
-# Transcode worker 并发数（后台批量：素材入库转码/AI 分类打标/质量评分/查重/批量下载）
-# 后台任务可排队，独立伸缩；素材入库量大时可加到 4。
-TRANSCODE_CONCURRENCY=2
-
-# 是否在本容器启动 celery beat 进程（默认 1）。
-# 默认 beat 与 worker 同容器部署；若要独立 beat 容器部署，worker 容器设为 0、
-# beat 容器单独跑 `celery -A worker_app.celery_app beat` 并设 BEAT_ENABLED=1。
-BEAT_ENABLED=1
-
 # 每个子进程最多处理多少任务后重启（防止内存泄漏）
 WORKER_MAX_TASKS_PER_CHILD=1000
 

.gitea/workflows/ci-pipeline.yml

@@ -1302,16 +1302,8 @@ jobs:
           "${staging_user}@${staging_host}:/var/lib/xiaoxia-saas-staging/configs/douyin_cookies.txt"
         echo "✅ Douyin cookies uploaded"
 
-        # 上传 infra/docker 配置到服务器（compose 单一事实来源）
-        echo "Uploading infra/docker configs to staging server..."
-        ssh -p "$staging_port" -i "$key_path" -o StrictHostKeyChecking=no "${staging_user}@${staging_host}" \
-          "mkdir -p /var/lib/xiaoxia-saas-staging/infra/docker"
-        scp -P "$staging_port" -i "$key_path" -o StrictHostKeyChecking=no infra/docker/compose.yml \
-          "${staging_user}@${staging_host}:/var/lib/xiaoxia-saas-staging/infra/docker/compose.yml"
-        echo "✅ infra/docker/compose.yml uploaded"
-
         # 通过环境变量传递凭证，避免命令行引号转义问题
-        cat scripts/ci_staging_deploy.sh | ssh -p "$staging_port" -i "$key_path" -o StrictHostKeyChecking=no "${staging_user}@${staging_host}" "IMAGE_TAG=${GITHUB_SHA} ACR_USERNAME=${ACR_USERNAME} ACR_PASSWORD=${ACR_PASSWORD} COMPOSE_SYNC=0 sh"
+        cat scripts/ci_staging_deploy.sh | ssh -p "$staging_port" -i "$key_path" -o StrictHostKeyChecking=no "${staging_user}@${staging_host}" "IMAGE_TAG=${GITHUB_SHA} ACR_USERNAME=${ACR_USERNAME} ACR_PASSWORD=${ACR_PASSWORD} sh"
 
         # 清理 CI runner 上的渲染文件
         rm -f .env.rendered

apps/worker/scripts/backfill_asset_quality.py

@@ -1,112 +0,0 @@
-"""一次性脚本：对历史 quality_score 缺失的视频素材重新打分。
-
-背景（#2073）：镜像 97ad0ae2 时期 calculate_quality_score / classify_from_analysis
-返回 str 而非 AssetClassification 枚举，导致 calculate_asset_quality 连续报
-"'str' object has no attribute 'value'"，大量视频素材的 quality_score 卡在 NULL。
-镜像 8abdeb95 已修复枚举 bug，但历史失败记录不会自动重跑。本脚本扫描全表，
-把 quality_score IS NULL 的视频素材重新投递到 worker.calculate_asset_quality 任务。
-
-使用方式（在 worker 容器内执行）：
-    cd /app/apps/worker
-    # 干跑，只打印会重跑多少条，不发任务
-    python -m scripts.backfill_asset_quality --dry-run
-    # 正式执行
-    python -m scripts.backfill_asset_quality
-    # 只重跑最近 N 天的
-    python -m scripts.backfill_asset_quality --since-days 30
-    # 限流：每投递一批 sleep 几秒，避免瞬间打爆 transcode 队列
-    python -m scripts.backfill_asset_quality --batch-size 50 --sleep 2
-
-也可以直接在 staging 机器上 exec 进容器：
-    docker exec -e PYTHONPATH=/app:/app/apps/api:/app/packages xiaoxia-worker-staging \
-        python -m scripts.backfill_asset_quality --dry-run
-"""
-
-from __future__ import annotations
-
-import argparse
-
-# 保证可以以 python -m scripts.xxx 在容器 /app/apps/worker 下执行
-# 也兼容在 repo 根目录下执行（注入路径）
-import os
-import sys
-import time
-from datetime import UTC, datetime, timedelta
-
-_SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
-_WORKER_DIR = os.path.dirname(_SCRIPT_DIR)  # apps/worker
-_APPS_DIR = os.path.dirname(_WORKER_DIR)  # apps
-_REPO_ROOT = os.path.dirname(_APPS_DIR)  # repo root
-for p in (_REPO_ROOT, os.path.join(_REPO_ROOT, "apps", "api"), _REPO_ROOT):
-    if p not in sys.path:
-        sys.path.insert(0, p)
-
-
-def main() -> int:
-    parser = argparse.ArgumentParser(description="补打历史视频素材 quality_score")
-    parser.add_argument("--dry-run", action="store_true", help="只统计数量，不投递任务")
-    parser.add_argument("--since-days", type=int, default=0, help="只处理最近 N 天上传的素材（0=全部）")
-    parser.add_argument("--batch-size", type=int, default=50, help="每批投递数量，默认 50")
-    parser.add_argument("--sleep", type=float, default=1.0, help="批次之间 sleep 秒数，默认 1s")
-    parser.add_argument("--queue", type=str, default="transcode", help="投递队列（默认 transcode）")
-    args = parser.parse_args()
-
-    # 延迟 import，避免在 dry-run 时依赖完整 DB 环境
-    from worker_app.celery_app import celery_app
-    from worker_app.db import SessionLocal
-
-    from packages.adapters.sqlalchemy_impl.models import AssetModel
-
-    db = SessionLocal()
-    try:
-        q = db.query(AssetModel).filter(
-            AssetModel.file_type == "video",
-            AssetModel.quality_score.is_(None),
-        )
-        if args.since_days > 0:
-            cutoff = datetime.now(UTC) - timedelta(days=args.since_days)
-            q = q.filter(AssetModel.created_at >= cutoff)
-
-        # 先 count 打印
-        total = q.count()
-        print(
-            f"[backfill] 待重跑 quality_score 的视频素材: {total} 条"
-            f"{' (dry-run，不投递)' if args.dry_run else ''}"
-            f"{' (最近 ' + str(args.since_days) + ' 天)' if args.since_days > 0 else ''}",
-            flush=True,
-        )
-        if total == 0 or args.dry_run:
-            return 0
-
-        # 分批投递
-        submitted = 0
-        batch = 0
-        offset = 0
-        while True:
-            assets = q.order_by(AssetModel.created_at.desc()).offset(offset).limit(args.batch_size).all()
-            if not assets:
-                break
-            batch += 1
-            for a in assets:
-                try:
-                    celery_app.send_task(
-                        "worker.calculate_asset_quality",
-                        args=[a.id],
-                        queue=args.queue,
-                    )
-                    submitted += 1
-                except Exception as e:  # noqa: BLE001
-                    print(f"[backfill] 投递失败 asset_id={a.id}: {e}", flush=True)
-            print(f"[backfill] batch {batch}: 已累计投递 {submitted}/{total}", flush=True)
-            offset += len(assets)
-            if args.sleep > 0 and offset < total:
-                time.sleep(args.sleep)
-
-        print(f"[backfill] 完成，共投递 {submitted} 条任务到 {args.queue} 队列", flush=True)
-        return 0
-    finally:
-        db.close()
-
-
-if __name__ == "__main__":
-    sys.exit(main())

apps/worker/video_processing/oss_helpers.py

@@ -1,17 +1,7 @@
-"""OSS 工具函数 — Worker 端统一入口。
+"""OSS 工具函数 — 从 generation.py 提取的共享 OSS 操作.
 
-P1 (2026-09-28) OSS 双 endpoint 改造：默认走 packages.shared.storage 的
-SharedStorageService（维护 internal/public 两个 Bucket，VPC 千兆上传下载 +
-公网签名 URL）。同时保留旧函数签名和模块级属性，兼容历史单测的 patch 路径。
-
-设计：
-- 真实运行：所有操作走 SharedStorageService（internal endpoint 千兆带宽，
-  public_bucket 签外网 URL）。
-- 单测 patch 场景：检测到 oss_settings/oss_bucket/oss2.Bucket/requests.get 等
-  被 patch 后，回退到旧直连 oss2 逻辑，老测试的 patch 仍然生效。
-- pytest importlib 模式兼容：conftest.py 把 apps/worker 加进 pythonpath，
-  本文件可能以 video_processing.oss_helpers 和 apps.worker.video_processing.oss_helpers
-  两个名字分别加载；patch 可能打到任一份，所以检测时遍历 sys.modules 里的同名模块。
+提供 OSS 配置读取、Bucket 创建、素材上传/下载、asset_id → 本地路径解析
+等能力，供 render_edit_plan 和 generate_video 共同复用。
 """
 
 from __future__ import annotations
@@ -19,173 +9,67 @@ from __future__ import annotations
 import hashlib
 import logging
 import os
-import sys
-import time as _time
+import threading
 from pathlib import Path
+from urllib.parse import urlparse
 
-import oss2  # noqa: F401  保留模块级属性，老单测 patch(oss_helpers.oss2)
-import requests  # noqa: F401  老单测 patch(oss_helpers.requests)
+import oss2
+import requests
 
-from packages.shared.storage import OSS_CONNECT_TIMEOUT  # noqa: F401
-from packages.shared.storage import OSS_MULTIPART_NUM_THREADS  # noqa: F401
-from packages.shared.storage import OSS_MULTIPART_THRESHOLD  # noqa: F401
-from packages.shared.storage import OSS_PART_SIZE  # noqa: F401
-from packages.shared.storage import (
-    OSS_HTTP_DOWNLOAD_TIMEOUT,
-    OSS_UPLOAD_TOTAL_TIMEOUT,
-    SharedStorageService,
-    get_shared_storage_service,
-)
+from packages.shared.config import get_shared_settings
 
 logger = logging.getLogger(__name__)
 
-
-# ── 单例访问 ──────────────────────────────────────────────────────────
+# OSS 上传配置
+OSS_CONNECT_TIMEOUT = 10  # 连接超时（秒），防止 TCP 握手挂死
+OSS_UPLOAD_TOTAL_TIMEOUT = 900  # 单文件上传总超时（秒），防止网络慢时无限卡住
+OSS_MULTIPART_THRESHOLD = 100 * 1024 * 1024  # 分片上传阈值：100MB 以上走分片
+OSS_PART_SIZE = 8 * 1024 * 1024  # 分片大小：8MB
+OSS_MULTIPART_NUM_THREADS = 3  # 分片上传并发数
 
 
-def _storage() -> SharedStorageService:
-    return get_shared_storage_service()
+# ── OSS 配置 ──────────────────────────────────────────────────────────────────
 
 
-# ── 多模块实例兼容（pytest importlib 模式）────────────────────────────
+def oss_settings() -> tuple[str, str, str, str] | None:
+    """获取 OSS 配置。
 
+    统一使用 SharedSettings 读取配置，与 SharedStorageService 保持一致，
+    支持从 .env 文件加载，避免两套配置路径不一致。
 
-def _sibling_modules() -> list:
-    """返回 sys.modules 里所有指向本文件的模块实例（包含自己）。"""
-    own_file = os.path.abspath(__file__)
-    mods = []
-    for _name, mod in list(sys.modules.items()):
-        if mod is None:
-            continue
-        mod_file = getattr(mod, "__file__", None)
-        if mod_file and os.path.abspath(mod_file) == own_file:
-            mods.append(mod)
-    return mods
-
-
-def _is_mock(obj) -> bool:
-    """判断对象是否是 unittest.mock.Mock/MagicMock。"""
-    if obj is None:
-        return False
-    try:
-        from unittest.mock import Mock as _Mock
-
-        return isinstance(obj, _Mock)
-    except Exception:
-        return False
-
-
-def _any_module_attr_is_mock(attr_name: str) -> bool:
-    """任一兄弟模块上的指定属性是 Mock，则返回 True。"""
-    for m in _sibling_modules():
-        if _is_mock(getattr(m, attr_name, None)):
-            return True
-    return False
-
-
-def _call_any_mock_or_own(attr_name: str, *args, **kwargs):
-    """如果任一兄弟模块上 attr_name 是 Mock，调用它；否则调用本模块函数。"""
-    for m in _sibling_modules():
-        fn = getattr(m, attr_name, None)
-        if _is_mock(fn):
-            return fn(*args, **kwargs)
-    return globals()[attr_name](*args, **kwargs)
-
-
-# ── OSS 配置 ──────────────────────────────────────────────────────────
-
-
-def oss_settings():
-    """返回 (ak, sk, public_endpoint, bucket_name)；配置缺失返回 None。"""
-    from packages.config import get_shared_settings
-
-    s = get_shared_settings()
-    if not (s.oss_access_key_id and s.oss_access_key_secret and s.oss_endpoint and s.oss_bucket_name):
+    Returns:
+        (access_key_id, access_key_secret, endpoint, bucket_name) 元组，
+        配置缺失时返回 None。
+    """
+    settings = get_shared_settings()
+    access_key_id = settings.oss_access_key_id
+    access_key_secret = settings.oss_access_key_secret
+    endpoint = settings.oss_endpoint
+    bucket_name = settings.oss_bucket_name
+    if not all([access_key_id, access_key_secret, endpoint, bucket_name]):
         return None
-    return (
-        s.oss_access_key_id,
-        s.oss_access_key_secret,
-        s.oss_endpoint,
-        s.oss_bucket_name,
-    )
+    return access_key_id, access_key_secret, endpoint, bucket_name
 
 
-def _get_oss_settings_from_any_module():
-    """从任一兄弟模块上取 oss_settings() 的返回值（mock 场景下兄弟模块上的
-    oss_settings 可能被 patch 成返回 None 或 tuple）。返回 None 表示所有模块
-    都返回 None（无配置）；返回 tuple 表示有配置；返回 Mock 表示被 patch。"""
-    any_mock = False
-    for m in _sibling_modules():
-        fn = getattr(m, "oss_settings", None)
-        if not callable(fn):
-            continue
-        is_mock = _is_mock(fn)
-        if is_mock:
-            any_mock = True
-        try:
-            result = fn()
-        except Exception:
-            continue
-        if is_mock:
-            # 被 patch 的函数：返回值就是 mock 的 return_value
-            if result is None:
-                # patch(oss_settings, return_value=None) → 无配置场景
-                return None
-            return result  # 可能是 tuple 或 Mock
-        if isinstance(result, tuple):
-            return result
-    if any_mock:
-        return None
-    return None
+def oss_bucket() -> oss2.Bucket | None:
+    """获取 OSS Bucket 实例。
 
+    P0-2 修复：endpoint 不带 scheme 时自动补 https:// 前缀，
+    确保 sign_url 等依赖 scheme 的方法返回 HTTPS URL。
 
-def _legacy_path_active() -> bool:
-    """是否走旧实现路径（兼容老单测 patch 路径，严格隔离不 fallback）。"""
-    # 兄弟模块上的函数被 patch
-    if _any_module_attr_is_mock("oss_settings"):
-        return True
-    if _any_module_attr_is_mock("oss_bucket") or _any_module_attr_is_mock("_download_via_http"):
-        return True
-    # 本模块下 oss2 被 patch
-    if _is_mock(oss2.Bucket) or _is_mock(oss2.Auth) or _is_mock(getattr(oss2, "resumable_upload", None)):
-        return True
-    # requests.get 被 patch
-    if _is_mock(requests) or _is_mock(requests.get):
-        return True
-    # 超时阈值被改成小值（老单测用 1s 做超时测试）
-    if OSS_UPLOAD_TOTAL_TIMEOUT <= 2:
-        return True
-    return False
+    P0-staging 修复：增加 connect_timeout=10s，防止网络抖动时
+    TCP 握手阶段无限挂死，导致 worker 进程卡死。
 
-
-def _ensure_scheme(endpoint: str) -> str:
-    if endpoint.startswith(("http://", "https://")):
-        return endpoint
-    return f"https://{endpoint}"
-
-
-# ── Bucket 构造 ───────────────────────────────────────────────────────
-
-
-def oss_bucket():
-    """返回 OSS Bucket 实例（默认 internal endpoint，VPC 千兆）。"""
-    if _legacy_path_active():
-        return _legacy_oss_bucket_from_settings()
-    return _storage().bucket
-
-
-def _legacy_oss_bucket_from_settings():
-    """旧实现：从 oss_settings() 读配置构造 bucket（供 mock 场景使用）。"""
-    settings = _get_oss_settings_from_any_module()
+    Returns:
+        oss2.Bucket 实例，配置缺失时返回 None。
+    """
+    settings = oss_settings()
     if settings is None:
         return None
-    try:
-        access_key_id, access_key_secret, endpoint, bucket_name = settings
-    except Exception:
-        return None
-    if not isinstance(endpoint, str):
-        endpoint = str(endpoint)
-    endpoint = _ensure_scheme(endpoint)
+    access_key_id, access_key_secret, endpoint, bucket_name = settings
+    # endpoint 无 scheme 时补 https://，与 API 端 storage.py 保持一致
+    if not endpoint.startswith(("http://", "https://")):
+        endpoint = f"https://{endpoint}"
     return oss2.Bucket(
         oss2.Auth(access_key_id, access_key_secret),
         endpoint,
@@ -194,211 +78,262 @@ def _legacy_oss_bucket_from_settings():
     )
 
 
-def public_bucket():
-    """返回公网 endpoint bucket（仅用于 sign_url）。"""
-    return _storage().public_bucket
-
-
 def normalize_storage_key(storage_key_or_url: str) -> str:
-    """标准化存储键：URL 取 path + URL decode，开头斜杠去掉。"""
-    return _storage().normalize_storage_key(storage_key_or_url)
+    """标准化存储键 — 如果是完整 URL 则提取 path 部分。
+
+    Examples:
+        "https://bucket.oss-cn-hangzhou.aliyuncs.com/path/to/file.mp4"
+        → "path/to/file.mp4"
+        "path/to/file.mp4" → "path/to/file.mp4"
+    """
+    if storage_key_or_url.startswith(("http://", "https://")):
+        return urlparse(storage_key_or_url).path.lstrip("/")
+    return storage_key_or_url.lstrip("/")
 
 
-# ── HTTP 下载（保留模块级函数方便 patch）─────────────────────────────
+# ── 上传 / 下载 ───────────────────────────────────────────────────────────────
+
+
+def download_asset(asset_storage_key: str, local_path: Path) -> bool:
+    """从 OSS 下载素材文件到本地路径。
+
+    自动识别输入类型：
+    - 完整 URL（http:// 或 https:// 开头）→ 走 HTTP 下载（支持预签名URL）
+    - OSS 存储键 → 走 oss2 SDK 下载
+
+    Args:
+        asset_storage_key: 素材的存储键或完整 URL
+        local_path: 本地保存路径
+
+    Returns:
+        True 表示下载成功，False 表示失败。
+    """
+    # 完整URL走HTTP下载（兼容预签名URL）
+    if asset_storage_key.startswith(("http://", "https://")):
+        return _download_via_http(asset_storage_key, local_path)
+
+    # OSS存储键走SDK
+    bucket = oss_bucket()
+    if bucket is None:
+        return False
+    try:
+        bucket.get_object_to_file(normalize_storage_key(asset_storage_key), str(local_path))
+        return local_path.exists() and local_path.stat().st_size > 0
+    except Exception:
+        logger.exception("下载素材失败: %s", asset_storage_key)
+        return False
 
 
 def _download_via_http(url: str, local_path: Path) -> bool:
-    """通过 HTTP 下载文件（用 oss_helpers.requests，方便单测 patch）。"""
+    """通过 HTTP 下载文件（支持预签名 URL）。
+
+    使用流式下载避免大文件内存溢出，超时 900s。
+    """
     try:
-        resp = requests.get(url, stream=True, timeout=OSS_HTTP_DOWNLOAD_TIMEOUT)
+        resp = requests.get(url, stream=True, timeout=900)
         resp.raise_for_status()
-        os.makedirs(Path(local_path).parent, exist_ok=True)
         with open(local_path, "wb") as f:
             for chunk in resp.iter_content(chunk_size=8 * 1024 * 1024):
                 if chunk:
                     f.write(chunk)
-        return Path(local_path).exists() and Path(local_path).stat().st_size > 0
+        return local_path.exists() and local_path.stat().st_size > 0
     except Exception:
-        logger.exception("HTTP下载失败: %s", url[:100])
+        logger.exception("HTTP下载素材失败: %s", url)
         return False
 
 
-# ── 下载 / 上传 ───────────────────────────────────────────────────────
+def upload_to_oss(local_path: Path | str, storage_key: str) -> str | None:
+    """上传文件到 OSS，返回公开 URL。
 
+    大文件（>100MB）自动走分片上传，降低内存峰值，减少 OOM 风险。
+    上传加总超时保护（默认 900s），防止网络异常时无限挂死。
 
-def download_asset(asset_storage_key: str, local_path: Path) -> bool:
-    """下载素材：HTTP URL 走本地 _download_via_http，OSS key 走 internal endpoint。"""
-    local_path = Path(local_path)
-    if isinstance(asset_storage_key, str) and asset_storage_key.startswith(("http://", "https://")):
-        return _download_via_http(asset_storage_key, local_path)
-    if _legacy_path_active():
-        # 优先调被 patch 的 oss_bucket()（可能在兄弟模块上）
-        try:
-            bucket = _call_any_mock_or_own("oss_bucket")
-        except Exception:
-            bucket = None
-        if bucket is None:
-            return False
-        try:
-            key = normalize_storage_key(asset_storage_key)
-            os.makedirs(local_path.parent, exist_ok=True)
-            bucket.get_object_to_file(key, str(local_path))
-            return local_path.exists() and local_path.stat().st_size > 0
-        except Exception:
-            logger.exception("下载素材失败: %s", asset_storage_key[:80])
-            return False
-    return _storage().download_asset(asset_storage_key, local_path)
+    Args:
+        local_path: 本地文件路径（Path 或 str 均可）
+        storage_key: 目标存储键
 
-
-def _legacy_upload_to_oss(local_path: Path, storage_key: str) -> str | None:
-    """旧实现：put_object_from_file / resumable_upload 二选一 + 超时保护。"""
-    bucket = _legacy_oss_bucket_from_settings()
+    Returns:
+        公开访问 URL，上传失败或 OSS 未配置时返回 None。
+    """
+    local_path = Path(local_path)  # 统一转 Path，兼容 str 调用
+    bucket = oss_bucket()
     if bucket is None:
         return None
-    settings = _get_oss_settings_from_any_module()
-    if settings is None:
-        return None
-    try:
-        _, _, endpoint, bucket_name = settings
-    except Exception:
-        return None
-    endpoint = _ensure_scheme(endpoint) if isinstance(endpoint, str) else f"https://{endpoint}"
-    public_host = endpoint.split("://", 1)[1]
-    url = f"https://{bucket_name}.{public_host}/{storage_key.lstrip('/')}"
 
-    local_path = Path(local_path)
-    try:
-        file_size = local_path.stat().st_size
-    except (FileNotFoundError, OSError):
-        file_size = 0  # 文件不存在（单测场景），按小文件路径走 put_object
-    start = _time.monotonic()
+    result: dict = {"url": None, "error": None, "file_size": 0}
+    done = threading.Event()
 
-    def _timed_out() -> bool:
-        return (_time.monotonic() - start) > OSS_UPLOAD_TOTAL_TIMEOUT
+    def _do_upload():
+        try:
+            # 尝试获取文件大小，用于分片判断和日志；stat 失败时 fallback 走普通上传
+            try:
+                file_size = local_path.stat().st_size
+                result["file_size"] = file_size
+                use_multipart = file_size >= OSS_MULTIPART_THRESHOLD
+            except OSError:
+                use_multipart = False
+                file_size = 0
 
-    try:
-        if file_size < OSS_MULTIPART_THRESHOLD:
-            if _timed_out():
-                return None
-            bucket.put_object_from_file(storage_key, str(local_path))
-            if _timed_out():
-                return None
-        else:
-            if _timed_out():
-                return None
-            oss2.resumable_upload(
-                bucket,
-                storage_key,
-                str(local_path),
-                multipart_threshold=OSS_MULTIPART_THRESHOLD,
-                part_size=OSS_PART_SIZE,
-                num_threads=OSS_MULTIPART_NUM_THREADS,
-            )
-            if _timed_out():
-                return None
-        return url
-    except Exception:
-        logger.exception("上传OSS失败: %s", storage_key[:80])
+            if use_multipart:
+                # 分片上传：降低内存峰值，每片 8MB，3 线程并发
+                logger.info(
+                    "大文件分片上传: storage_key=%s, size=%.1fMB, part_size=%dMB, threads=%d",
+                    storage_key[:80],
+                    file_size / 1024 / 1024,
+                    OSS_PART_SIZE // 1024 // 1024,
+                    OSS_MULTIPART_NUM_THREADS,
+                )
+                oss2.resumable_upload(
+                    bucket,
+                    storage_key,
+                    str(local_path),
+                    multipart_threshold=OSS_MULTIPART_THRESHOLD,
+                    part_size=OSS_PART_SIZE,
+                    num_threads=OSS_MULTIPART_NUM_THREADS,
+                )
+            else:
+                bucket.put_object_from_file(storage_key, str(local_path))
+
+            # 构造返回 URL
+            settings = oss_settings()
+            if settings:
+                _, _, endpoint, bucket_name = settings
+                endpoint_clean = endpoint.replace("https://", "").replace("http://", "")
+                result["url"] = f"https://{bucket_name}.{endpoint_clean}/{storage_key}"
+        except Exception as e:
+            result["error"] = e
+            logger.exception("上传 OSS 失败: %s", storage_key)
+        finally:
+            done.set()
+
+    upload_thread = threading.Thread(target=_do_upload, daemon=True)
+    upload_thread.start()
+    finished = done.wait(timeout=OSS_UPLOAD_TOTAL_TIMEOUT)
+
+    if not finished:
+        logger.error(
+            "OSS 上传超时（%.0fs），强制中止: storage_key=%s, size=%.1fMB",
+            OSS_UPLOAD_TOTAL_TIMEOUT,
+            storage_key[:80],
+            result["file_size"] / 1024 / 1024 if result["file_size"] else 0,
+        )
         return None
 
+    if result["error"]:
+        return None
 
-def upload_to_oss(local_path: Path | str, storage_key: str) -> str | None:
-    """上传文件到 OSS，返回公网 URL。"""
-    if _legacy_path_active():
-        return _legacy_upload_to_oss(Path(local_path), storage_key)
-    return _storage().upload_file_smart(local_path, storage_key)
+    return result["url"]
 
 
 def get_signed_download_url(storage_key_or_url: str, expires_seconds: int = 3600) -> str | None:
-    """生成预签名下载 URL（公网域名，外网可访问）。"""
-    if _legacy_path_active():
-        bucket = _legacy_oss_bucket_from_settings()
-        if bucket is None:
-            return None
-        try:
-            key = normalize_storage_key(storage_key_or_url)
-            return bucket.sign_url("GET", key, expires_seconds)
-        except Exception:
-            logger.exception("生成预签名URL失败: %s", storage_key_or_url[:80])
-            return None
-    s = _storage()
-    if s.public_bucket is None and s.bucket is None:
+    """生成预签名下载 URL（用于私有 bucket 的 URL 校验或临时下载）。
+
+    Args:
+        storage_key_or_url: 存储键或完整 URL（URL 会自动提取 path）
+        expires_seconds: 签名有效期（秒）
+
+    Returns:
+        预签名 URL，失败或 OSS 未配置时返回 None。
+    """
+    bucket = oss_bucket()
+    if bucket is None:
         return None
     try:
-        return s.get_download_url(storage_key_or_url, expires_seconds=expires_seconds)
+        storage_key = normalize_storage_key(storage_key_or_url)
+        signed = bucket.sign_url("GET", storage_key, expires_seconds)
+        logger.info("生成预签名URL: key=%s url_prefix=%s", storage_key[:80], signed[:60])
+        return signed
     except Exception:
         logger.exception("生成预签名URL失败: %s", storage_key_or_url[:80])
         return None
 
 
-# ── Asset 解析 ────────────────────────────────────────────────────────
+# ── Asset 解析 ────────────────────────────────────────────────────────────────
 
 
 def resolve_asset_path(asset_id: str, work_dir: Path) -> Path | None:
-    """从 asset_id 解析到本地路径（缓存优先，否则 OSS 下载）。
+    """从 asset_id 解析到本地文件路径。
 
-    在 wrapper 层实现缓存逻辑，方便老单测 patch(oss_helpers.download_asset)。
+    策略（按优先级）：
+    1. 如果 asset_id 是本地绝对路径（/var/storage/...）→ 安全校验后返回
+    2. 如果 work_dir 下已有缓存文件 → 返回缓存路径
+    3. 从 OSS 下载到 work_dir/{hash}.mp4 → 返回下载路径
+    4. 下载失败 → 返回 None
+
+    缓存策略：以 asset_id 的 SHA256 前 16 位为文件名，避免重复下载。
+
+    安全：
+    - 本地绝对路径必须在 ASSET_ALLOWED_DIRS 环境变量指定的目录内
+    - 文件名经过 sanitize，防止路径遍历
+    - 禁止空字节、控制字符
     """
+    from video_processing.path_security import (
+        PathSecurityError,
+        get_allowed_local_dirs,
+        is_in_allowed_dirs,
+        sanitize_filename,
+    )
+
     if not asset_id or not isinstance(asset_id, str):
         return None
+
+    # 空字节检测
     if "\x00" in asset_id:
+        logger.warning("asset_id 包含空字节，拒绝: %s", asset_id[:50])
         return None
 
-    work_dir = Path(work_dir)
-    os.makedirs(work_dir, exist_ok=True)
-
-    if asset_id.startswith("/") or ".." in Path(asset_id).parts:
-        logger.warning("非法 asset_id: %s", asset_id)
-        return None
+    # 1. 本地绝对路径 — 必须在允许的目录内
+    if asset_id.startswith("/") and os.path.exists(asset_id):
+        try:
+            resolved = Path(asset_id).resolve()
+            if is_in_allowed_dirs(resolved, get_allowed_local_dirs()):
+                return resolved
+            else:
+                logger.warning(
+                    "本地素材路径不在允许目录内，拒绝: %s (allowed=%s)",
+                    asset_id[:80],
+                    get_allowed_local_dirs(),
+                )
+                return None
+        except (OSError, PathSecurityError):
+            return None
 
+    # 2. 缓存命中（使用 hash 而非原始 ID，防止路径遍历）
     cache_hash = hashlib.sha256(asset_id.encode()).hexdigest()[:16]
-    local_path = work_dir / f"{cache_hash}.mp4"
+    safe_name = sanitize_filename(cache_hash)
+    cached_path = work_dir / f"{safe_name}.mp4"
+    if cached_path.exists() and cached_path.stat().st_size > 0:
+        return cached_path
 
-    if local_path.exists() and local_path.stat().st_size > 0:
-        return local_path
+    # 3. 从 OSS 下载（先标准化 key，防止路径遍历注入）
+    safe_key = normalize_storage_key(asset_id)
+    # 额外校验：存储键不能包含 ../ 或绝对路径
+    if ".." in safe_key or safe_key.startswith("/"):
+        logger.warning("asset_id 包含路径遍历模式，拒绝下载: %s", asset_id[:80])
+        return None
+
+    if download_asset(safe_key, cached_path):
+        return cached_path
 
-    try:
-        ok = download_asset(asset_id, local_path)
-        if ok and local_path.exists() and local_path.stat().st_size > 0:
-            return local_path
-    except Exception:
-        logger.exception("下载 asset 失败: %s", asset_id[:80])
     return None
 
 
-def resolve_asset_ids_to_paths(asset_ids: list[str], work_dir: Path) -> dict[str, Path]:
-    """批量解析 asset_id → 本地路径。"""
+def resolve_asset_ids_to_paths(
+    asset_ids: list[str],
+    work_dir: Path,
+) -> dict[str, Path]:
+    """批量解析 asset_id → 本地路径。
+
+    Args:
+        asset_ids: 素材 ID 列表
+        work_dir: 工作目录
+
+    Returns:
+        {asset_id: local_path} 映射，仅包含成功解析的条目。
+    """
     result: dict[str, Path] = {}
     for aid in asset_ids:
-        p = resolve_asset_path(aid, work_dir)
-        if p is not None:
-            result[aid] = p
+        local_path = resolve_asset_path(aid, work_dir)
+        if local_path:
+            result[aid] = local_path
     return result
-
-
-def delete_from_oss(storage_key_or_url: str) -> bool:
-    """从 OSS 删除对象（best-effort，internal endpoint）。"""
-    s = _storage()
-    if s.bucket is None:
-        return False
-    try:
-        key = normalize_storage_key(storage_key_or_url)
-        s.delete_file(key)
-        return True
-    except Exception:
-        logger.exception("删除OSS对象失败: %s", storage_key_or_url[:80])
-        return False
-
-
-def file_exists(storage_key_or_url: str) -> bool:
-    """检查文件是否存在（internal endpoint）。"""
-    s = _storage()
-    if s.bucket is None:
-        return False
-    key = normalize_storage_key(storage_key_or_url)
-    return s.file_exists(key)
-
-
-def get_public_url(storage_key: str) -> str:
-    """返回公网 URL（不带签名）。"""
-    return _storage().get_url(storage_key)

apps/worker/video_processing/thumbnail_generator.py

@@ -221,7 +221,7 @@ def _extract_frames_via_mediakit(
     """
     import uuid
 
-    from video_processing.oss_helpers import get_signed_download_url, upload_to_oss
+    from video_processing.oss_helpers import upload_to_oss
 
     from packages.shared.mediakit_client import get_mediakit_client
 
@@ -230,17 +230,14 @@ def _extract_frames_via_mediakit(
         logger.info("[thumbnail] MediaKit 未配置，跳过智能抽帧")
         return None
 
-    video_storage_key: str = ""
-    # 1. 上传视频到 OSS，并生成预签名下载 URL（bucket 私有读，公网 URL 会 403）
+    # 1. 上传视频到 OSS 获取 URL
     try:
         video_storage_key = f"temp/{plan_id}/{uuid.uuid4().hex[:8]}_{Path(video_path).name}"
-        public_url = upload_to_oss(video_path, video_storage_key)
-        if not public_url:
+        video_url = upload_to_oss(video_path, video_storage_key)
+        if not video_url:
             logger.warning("[thumbnail] 视频上传 OSS 失败，无法使用 MediaKit")
             return None
-        # MediaKit 从公网拉取视频，必须使用预签名 URL；签名 1h 足够完成抽帧
-        video_url = get_signed_download_url(video_storage_key, expires_seconds=3600) or public_url
-        logger.info("[thumbnail] 视频已上传 OSS 并生成签名 URL: key=%s", video_storage_key[:80])
+        logger.info("[thumbnail] 视频已上传 OSS: %s", video_url[:80])
     except Exception as e:
         logger.warning("[thumbnail] 视频上传 OSS 异常: %s，降级到 ffmpeg", e)
         return None

apps/worker/worker_app/tasks/asset_quality_scoring_task.py

@@ -18,7 +18,6 @@ from worker_app.celery_app import celery_app
 from worker_app.db import SessionLocal
 
 from packages.adapters.sqlalchemy_impl.asset_repository import SQLAlchemyAssetRepository
-from packages.domain.classification import ClassificationStatus
 from packages.shared.storage import get_shared_storage_service
 
 logger = get_task_logger(__name__)
@@ -97,7 +96,7 @@ def calculate_asset_quality_task(self, asset_id: str) -> dict:
                     confidence = 1.0
                 existing_meta["classification"] = classification
                 existing_meta["classification_confidence"] = confidence
-                asset.classification_status = ClassificationStatus.COMPLETED
+                asset.classification_status = "completed"
                 asset.metadata = existing_meta
                 logger.info(
                     "[quality_score] asset=%s 自动分类完成: category=%s confidence=%.2f",
@@ -111,8 +110,6 @@ def calculate_asset_quality_task(self, asset_id: str) -> dict:
                     asset_id,
                     cls_err,
                 )
-                # 分类失败显式标记 FAILED，避免停留在 PENDING 被反复重试
-                asset.classification_status = ClassificationStatus.FAILED
 
         asset_repo.update(asset)
         db.commit()

apps/worker/worker_app/tasks/generation.py

@@ -1051,15 +1051,20 @@ def generate_video(self, task_id: str) -> dict:
                         )
                         if _meta_model:
                             meta = dict(_meta_model.extra_meta or {})
-                            # #2024/P0 finalize-400: 直接展开 _precompute_render_metadata 返回的
-                            # 完整 dict（含 file_url/fingerprint_dict/fingerprint_chunks/is_duplicate
-                            # /duplicate_of/...），避免手写字段白名单漏传字段导致 finalize 读不到数据。
                             meta["rendered_output"] = {
-                                **dict(rendered_output or {}),
-                                # file_url/duration 由外层调用方拿到的实际上传结果，优先覆盖预计算值
                                 "file_url": file_url,
                                 "file_size": file_size,
                                 "duration": duration,
+                                "width": rendered_output.get("width", 1280),
+                                "height": rendered_output.get("height", 720),
+                                "fps": rendered_output.get("fps", 25.0),
+                                "name": rendered_output.get("name", ""),
+                                "thumbnail_url": rendered_output.get("thumbnail_url", ""),
+                                "mode": rendered_output.get("mode", editing_mode.value),
+                                "fingerprint_dict": rendered_output.get("fingerprint_dict"),
+                                "batch_id": batch_id,
+                                "project_id": project_id,
+                                "user_id": user_id,
                             }
                             _meta_model.extra_meta = meta
                             _finalize_meta_session.commit()

deploy/configs/.env.staging

@@ -280,18 +280,3 @@ USE_GPU_LIPSYNC=true
 GPU_LIPSYNC_POLL_INTERVAL=5
 GPU_LIPSYNC_WAIT_TIMEOUT=1200
 GPU_WORKER_STALE_SECONDS=300
-
-# ==================== P4000 NVENC 硬件编码（GPU mezzanine relay）====================
-# 注意：这些值必须写死在模板里（不是 CI Secret），否则每次 CI 重新渲染 .env 都会被丢弃，
-# 导致 staging 发版后 GPU 编码静默降级到 CPU（P0 防复发）。
-ENABLE_GPU_ENCODE=true
-GPU_ENCODE_ENDPOINT=http://100.105.75.67:8900
-GPU_ENCODE_RELAY_BASE_URL=http://100.125.116.43:8092
-GPU_ENCODE_RELAY_INTERNAL_BASE_URL=http://xiaoxia-api-staging:8000
-GPU_ENCODE_RELAY_SECRET=0e1a8f0626438564a8b3fa92f3f2aac29e3c69bc02f2f85c
-GPU_ENCODE_VCODEC=h264_nvenc
-GPU_ENCODE_PRESET=p4
-GPU_ENCODE_CRF=23
-GPU_ENCODE_FALLBACK_CPU=true
-GPU_ENCODE_MEZZANINE_TRANSPORT=oss
-GPU_ENCODE_OSS_TMP_PREFIX=tmp/gpu-mezzanine/

infra/docker/compose.yml

@@ -10,25 +10,20 @@
 #   API_IMAGE        - API 镜像名称 (默认: xiaoxia-saas-api:dev)
 #   WORKER_IMAGE     - Worker 镜像名称 (默认: xiaoxia-saas-worker:dev)
 #   WEB_IMAGE        - Web 镜像名称 (默认: xiaoxia-saas-web:dev)
+#   WEB_DOCKERFILE   - Web Dockerfile 路径
+#   WEB_NGINX_CONF   - Nginx 配置文件路径
 #   API_PORT         - API 端口映射 (staging: 8000, production: 8001)
 #   WEB_PORT         - Web 端口映射 (staging: 3001, production: 3002)
 #   GENERATED_FILES_HOST_DIR - 生成文件的主机目录
-#   GENERATION_CONCURRENCY    - Generation worker 并发（用户实时任务，默认 2）
-#   TRANSCODE_CONCURRENCY     - Transcode worker 并发（后台/转码/AI，默认 2）
+#   WORKER_CONCURRENCY        - Worker 并发数 (默认: 4)
 #   WORKER_MAX_TASKS_PER_CHILD - Worker 每个子进程最大任务数 (默认: 100)
-#   BEAT_ENABLED              - 容器内启动 celery beat（默认 1；独立 beat 容器部署设为 0）
-#   WORKER_CONCURRENCY        - 兼容旧变量：未显式设置上面两个并发时按此总数分配
 #
-# 重要:
+# 重要: 
 #   - 生产环境不要挂载 web-dist volume，否则会导致 403
 #   - 确保环境隔离网络已创建: docker network create xiaoxia-net-${ENV}
-#   - ENV=staging  -> xiaoxia-net-staging
-#   - ENV=production -> xiaoxia-net-production
+#   - ENV=staging  → xiaoxia-net-staging
+#   - ENV=production → xiaoxia-net-production
 #
-# #2073 队列分流：worker 容器内跑三个独立进程——beat(只发定时任务)、
-# generation worker(只消费 generation 队列，实时高优)、transcode worker(消费
-# transcode + celery 队列，后台任务)。beat 不再嵌入 generation worker，
-# 不占实时任务槽位；TRANSCODE_CONCURRENCY 独立伸缩，不再依赖 WORKER_CONCURRENCY 差值。
 
 # ===========================================
 # 日志轮转配置（所有服务共享）
@@ -45,39 +40,53 @@ services:
   # =========================================
   api:
     image: ${API_IMAGE:-xiaoxia-saas-api:dev}
+    # 不在生产环境构建镜像，使用预构建的镜像
+    # build:
+    #   context: ../..
+    #   dockerfile: infra/docker/api.Dockerfile
+    
     container_name: xiaoxia-api-${ENV:-staging}
     restart: unless-stopped
     stop_grace_period: 30s
     stop_signal: SIGTERM
-
+    
+    # 环境变量文件（包含数据库密码等敏感信息）
     env_file:
       - ../../.env
-
+    
     environment:
       APP_ENV: ${APP_ENV:-staging}
       GENERATED_FILES_DIR: /app/generated
       GENERATED_FILES_URL_PREFIX: /generated-files
       PUBLIC_API_BASE_URL: ${PUBLIC_API_BASE_URL:-https://api.xiaoxiajianji.com}
-
+    
+    # 端口映射
+    # Staging: 8000 -> 8000
+    # Production: 8001 -> 8000
     ports:
       - "127.0.0.1:${API_PORT:-8000}:8000"
-
+    
+    # 共享生成文件目录 + 抖音 cookies 等运行时配置
     volumes:
       - generated-files:/app/generated
       - ../../deploy/configs:/app/configs:ro
-
+    
     networks:
       - xiaoxia-net
-
+    
+    # 健康检查配置
     healthcheck:
       test: ["CMD", "python", "-c", "import urllib.request; urllib.request.urlopen('http://localhost:8000/health', timeout=5)"]
       interval: 30s
       timeout: 10s
       retries: 3
       start_period: 40s
-
+    
     logging: *default-logging
 
+    # =========================================
+    # 资源限制建议（生产环境建议启用）
+    # =========================================
     deploy:
       resources:
         limits:
@@ -88,48 +97,39 @@ services:
           memory: 512M
 
   # =========================================
-  # Worker 服务（#2073 队列分流：beat + generation + transcode 同容器三进程）
+  # Worker 服务（Celery 任务队列）
   # =========================================
-  # 三个进程独立启动，任一退出则容器整体退出由 docker restart 拉起；
-  # 各自的并发与资源占用通过环境变量控制：
-  #   - generation：GENERATION_CONCURRENCY（默认 2），消费 generation 队列
-  #   - transcode： TRANSCODE_CONCURRENCY（默认 2），消费 transcode,celery 队列
-  #   - beat：      不消费任务，只发定时任务到 celery 默认队列
   worker:
     image: ${WORKER_IMAGE:-xiaoxia-saas-worker:dev}
-
+    
     container_name: xiaoxia-worker-${ENV:-staging}
     restart: unless-stopped
-    # 长任务（ingest HEVC 转码最长 30min、生成硬超时 11min）给足优雅关闭窗口
     stop_grace_period: 300s
     stop_signal: SIGTERM
-
+    
     env_file:
       - ../../.env
-
+    
     environment:
       APP_ENV: ${APP_ENV:-staging}
-      # 兼容旧变量：若两个 *_CONCURRENCY 均未显式设置，entrypoint 会按此总数分配
       WORKER_CONCURRENCY: ${WORKER_CONCURRENCY:-4}
       WORKER_MAX_TASKS_PER_CHILD: ${WORKER_MAX_TASKS_PER_CHILD:-100}
-      # #2073 队列独立伸缩：generation 默认 2，transcode 默认 2（不再差值计算）
+      # #1714 队列隔离：generation 队列独占 worker（默认并发 2），其余并发给转码
       GENERATION_CONCURRENCY: ${GENERATION_CONCURRENCY:-2}
-      TRANSCODE_CONCURRENCY: ${TRANSCODE_CONCURRENCY:-2}
-      # beat 默认在本容器启动；独立 beat 容器部署时设为 0
-      BEAT_ENABLED: ${BEAT_ENABLED:-1}
       GENERATED_FILES_DIR: /app/generated
       GENERATED_FILES_URL_PREFIX: /generated-files
       PUBLIC_API_BASE_URL: ${PUBLIC_API_BASE_URL:-https://api.xiaoxiajianji.com}
-
+    
     volumes:
       - generated-files:/app/generated
-
+    
     networks:
       - xiaoxia-net
 
-    # 健康检查：至少有一个 celery worker 进程在跑（beat 本身不作为存活依据）
+    # 健康检查配置
+    # 注：容器内无 pgrep/ps，扫描 /proc 所有进程的 cmdline 查找 celery 进程
     healthcheck:
-      test: ["CMD-SHELL", "grep -q 'celery.*worker' /proc/[0-9]*/cmdline 2>/dev/null || exit 1"]
+      test: ["CMD-SHELL", "grep -lq celery /proc/[0-9]*/cmdline 2>/dev/null || exit 1"]
       interval: 30s
       timeout: 10s
       retries: 3
@@ -137,8 +137,12 @@ services:
 
     logging: *default-logging
 
-    # 资源限制：容器总资源 = gen + trans + beat，按 2+2 并发场景建议 4C8G；
-    # 后续如需独立扩容/重启，可拆为 worker-generation / worker-transcode / worker-beat 三个 service。
+    # =========================================
+    # 资源限制建议（生产环境建议启用）
+    # =========================================
+    # 注意: Worker 需要处理视频，建议分配更多资源
+    # #1714 队列隔离后容器内运行 generation + transcode 两个 worker 进程，
+    # 总并发 = WORKER_CONCURRENCY（默认 4），4C8G 以上确保视频渲染不 OOM
     deploy:
       resources:
         limits:
@@ -153,21 +157,35 @@ services:
   # =========================================
   web:
     image: ${WEB_IMAGE:-xiaoxia-saas-web:dev}
-
+    # 不在生产环境构建镜像，使用 web-artifact.Dockerfile
+    # build:
+    #   context: ../..
+    #   dockerfile: ${WEB_DOCKERFILE:-infra/docker/web.Dockerfile}
+    #   args:
+    #     (NGINX_CONF no longer needed - all configs baked into image)
+    
     container_name: xiaoxia-web-${ENV:-staging}
     restart: unless-stopped
-
+    
+    # 端口映射
+    # Staging: 3001 -> 80
+    # Production: 3002 -> 80 (通过 Nginx 反向代理)
     ports:
       - "127.0.0.1:${WEB_PORT:-3001}:80"
-
+    
     networks:
       - xiaoxia-net
-
+    
+    # =========================================
+    # Nginx 配置运行时覆盖（双保险：entrypoint 也按 APP_ENV 选择配置）
+    # 确保容器使用正确环境的 nginx 配置，即使镜像构建时使用了默认配置
+    # 注意: 只覆盖 /etc/nginx/conf.d/default.conf，不挂载 /usr/share/nginx/html
+    # =========================================
     environment:
       - APP_ENV=${ENV:-staging}
     volumes:
       - ./nginx-${ENV:-staging}.conf:/etc/nginx/conf.d/default.conf:ro
-
+    
     healthcheck:
       test: ["CMD", "wget", "--spider", "-q", "http://127.0.0.1:80"]
       interval: 30s
@@ -176,6 +194,9 @@ services:
 
     logging: *default-logging
 
+    # =========================================
+    # 资源限制建议
+    # =========================================
     deploy:
       resources:
         limits:
@@ -191,6 +212,9 @@ volumes:
     driver_opts:
       type: none
       o: bind
+      # 重要: 确保主机目录存在且有正确权限
+      # Staging: /var/lib/xiaoxia-saas-staging/generated
+      # Production: /var/lib/xiaoxia-saas-production/generated
       device: ${GENERATED_FILES_HOST_DIR:?GENERATED_FILES_HOST_DIR must be set in .env}
 
 # ===========================================
@@ -199,4 +223,8 @@ volumes:
 networks:
   xiaoxia-net:
     external: true
+    # 网络名根据 ENV 变量区分，实现 staging/production 环境隔离
+    # staging:     xiaoxia-net-staging
+    # production:  xiaoxia-net-production
     name: xiaoxia-net-${ENV:-staging}
+

infra/docker/entrypoint-worker.sh

@@ -1,77 +1,48 @@
 #!/bin/bash
-# Worker 启动脚本 — #1714 + #2073 队列分流
+# Worker 启动脚本 — #1714 队列隔离
 #
-# 容器内启动三个独立进程（任一退出则整体退出由 docker restart 拉起）：
-#   1. beat：celery beat 调度器，不消费任何任务，只发定时任务到 celery 默认队列
-#   2. generation-worker：独占消费 generation 队列（用户实时任务，高优先级）
-#   3. transcode-worker：消费 transcode + celery 默认队列（后台/清理任务）
+# 部署约束：worker 容器单实例（replicas=1），容器内启动两个 celery 进程：
+#   1. generation-worker：独占消费 generation 队列（用户视频生成，高优先级），
+#      内嵌 celery beat（-B），定时清理任务只在一个进程里跑，避免重复执行；
+#   2. transcode-worker：消费 transcode + celery 默认队列（素材转码/分类/查重/
+#      配音/下载等后台任务）。
+# 转码队列积压时，generation 队列仍有独立 worker 立即领取视频生成任务。
 #
 # 环境变量：
+#   WORKER_CONCURRENCY            总并发槽参考（默认 4）；生成 worker 并发默认 2，
+#                                 可用 GENERATION_CONCURRENCY 覆盖
 #   GENERATION_CONCURRENCY        generation worker 并发（默认 2）
-#   TRANSCODE_CONCURRENCY         transcode worker 并发（默认 2）
+#   TRANSCODE_CONCURRENCY         transcode worker 并发（默认 = WORKER_CONCURRENCY - 2，最小 1）
 #   WORKER_MAX_TASKS_PER_CHILD    每个子进程最大任务数（默认 100）
-#   WORKER_CONCURRENCY            兼容旧变量：若未显式设置 GENERATION_CONCURRENCY /
-#                                 TRANSCODE_CONCURRENCY，则按比例分配（gen=ceil(total*1/2)，
-#                                 trans=剩余，各至少 1）；已显式设置时忽略此变量。
-#   BEAT_ENABLED                  是否在本容器内启动 beat 进程（默认 1）；
-#                                 若独立 beat 容器部署设为 0。
 
 set -e
 
+CONCURRENCY="${WORKER_CONCURRENCY:-4}"
 MAX_TASKS="${WORKER_MAX_TASKS_PER_CHILD:-100}"
 
-# ── 并发计算：显式 env 优先；否则从 WORKER_CONCURRENCY 按比例推导 ──
-if [ -n "$GENERATION_CONCURRENCY" ]; then
-  GEN_CONCURRENCY="$GENERATION_CONCURRENCY"
-else
-  TOTAL="${WORKER_CONCURRENCY:-4}"
-  GEN_CONCURRENCY=$(( (TOTAL + 1) / 2 ))
-  if [ "$GEN_CONCURRENCY" -lt 1 ]; then GEN_CONCURRENCY=1; fi
-fi
-
-if [ -n "$TRANSCODE_CONCURRENCY" ]; then
-  TRANS_CONCURRENCY="$TRANSCODE_CONCURRENCY"
-else
-  if [ -n "$WORKER_CONCURRENCY" ] && [ -z "$GENERATION_CONCURRENCY" ]; then
-    # 两个都没显式设置，按 WORKER_CONCURRENCY 分配剩余
-    TOTAL="$WORKER_CONCURRENCY"
-    TRANS_CONCURRENCY=$(( TOTAL - GEN_CONCURRENCY ))
-    if [ "$TRANS_CONCURRENCY" -lt 1 ]; then TRANS_CONCURRENCY=1; fi
-  else
-    # 默认 2（#2073：独立伸缩，不再依赖 WORKER_CONCURRENCY 差值）
-    TRANS_CONCURRENCY=2
+GEN_CONCURRENCY="${GENERATION_CONCURRENCY:-2}"
+if [ -z "$TRANSCODE_CONCURRENCY" ]; then
+  TRANS_CONCURRENCY=$((CONCURRENCY - GEN_CONCURRENCY))
+  if [ "$TRANS_CONCURRENCY" -lt 1 ]; then
+    TRANS_CONCURRENCY=1
   fi
+else
+  TRANS_CONCURRENCY="$TRANSCODE_CONCURRENCY"
 fi
 
-BEAT_ENABLED="${BEAT_ENABLED:-1}"
-
-PIDS=()
-
-# ── 1. Beat 调度器（独立进程，不消费任务）──
-if [ "$BEAT_ENABLED" = "1" ] || [ "$BEAT_ENABLED" = "true" ]; then
-  echo "Starting beat scheduler (schedule file=/tmp/celerybeat-schedule)"
-  celery \
-    -A worker_app.celery_app \
-    beat \
-    --loglevel=info \
-    -s /tmp/celerybeat-schedule &
-  PIDS+=($!)
-fi
-
-# ── 2. Generation worker（实时高优队列）──
-echo "Starting generation worker (queue=generation, concurrency=$GEN_CONCURRENCY)"
+echo "Starting generation worker (queue=generation, concurrency=$GEN_CONCURRENCY, beat embedded)"
 celery \
   -A worker_app.celery_app \
   worker \
   --loglevel=info \
+  "-B" \
+  -s /tmp/celerybeat-schedule \
   -Q generation \
   "--concurrency=${GEN_CONCURRENCY}" \
   "--max-tasks-per-child=${MAX_TASKS}" \
   -n generation@%h &
-PIDS+=($!)
-GEN_PID=${PIDS[1]:-${PIDS[0]}}
+GEN_PID=$!
 
-# ── 3. Transcode worker（后台 + 清理队列）──
 echo "Starting transcode worker (queues=transcode,celery, concurrency=$TRANS_CONCURRENCY)"
 celery \
   -A worker_app.celery_app \
@@ -81,22 +52,13 @@ celery \
   "--concurrency=${TRANS_CONCURRENCY}" \
   "--max-tasks-per-child=${MAX_TASKS}" \
   -n transcode@%h &
-PIDS+=($!)
-TRANS_PID=${PIDS[2]:-${PIDS[1]}}
+TRANS_PID=$!
 
-# 任一进程退出则终止其他进程，让容器整体重启
-cleanup() {
-  echo "Shutting down all celery processes..."
-  for pid in "${PIDS[@]}"; do
-    kill -TERM "$pid" 2>/dev/null || true
-  done
-}
-trap cleanup TERM INT
+# 任一进程退出则终止另一个，让容器整体重启（restart: unless-stopped）
+trap 'echo "Shutting down workers..."; kill -TERM $GEN_PID $TRANS_PID 2>/dev/null || true' TERM INT
 
-# wait -n 等待任意一个子进程退出（bash 4.3+）
-# 容器镜像基础为 python:3.11-slim，bash 版本满足
-wait -n "${PIDS[@]}"
+wait -n $GEN_PID $TRANS_PID
 EXIT_CODE=$?
-echo "One celery process exited (code=$EXIT_CODE), stopping the rest..."
-cleanup
-exit "$EXIT_CODE"
+echo "One worker exited (code=$EXIT_CODE), stopping the other..."
+kill -TERM $GEN_PID $TRANS_PID 2>/dev/null || true
+exit $EXIT_CODE

packages/adapters/sqlalchemy_impl/asset_repository.py

@@ -128,12 +128,8 @@ class SQLAlchemyAssetRepository:
             height=asset.height,
             fps=asset.fps,
             codec=asset.codec,
-            status=(asset.status.value if hasattr(asset.status, "value") else str(asset.status)),
-            classification_status=(
-                asset.classification_status.value
-                if hasattr(asset.classification_status, "value")
-                else str(asset.classification_status)
-            ),
+            status=asset.status.value,
+            classification_status=asset.classification_status.value,
             classification_result=(json.dumps(asset.metadata) if asset.metadata else None),
             quality_score=asset.quality_score,
             uploaded_by_user_id=asset.uploaded_by_user_id or "system",
@@ -146,7 +142,7 @@ class SQLAlchemyAssetRepository:
         self.session.flush()
         self._sync_asset_tags(asset.id, asset.tag_ids)
         # Issue #1776: 自动维护素材库计数（同事务内原子更新）
-        if asset.library_id and (getattr(asset.status, "value", str(asset.status)) != "deleted"):
+        if asset.library_id and asset.status.value != "deleted":
             from sqlalchemy import func
 
             self.session.query(AssetLibraryModel).filter(AssetLibraryModel.id == asset.library_id).update(
@@ -172,12 +168,8 @@ class SQLAlchemyAssetRepository:
         model.height = asset.height
         model.fps = asset.fps
         model.codec = asset.codec
-        model.status = asset.status.value if hasattr(asset.status, "value") else str(asset.status)
-        model.classification_status = (
-            asset.classification_status.value
-            if hasattr(asset.classification_status, "value")
-            else str(asset.classification_status)
-        )
+        model.status = asset.status.value
+        model.classification_status = asset.classification_status.value
         model.classification_result = json.dumps(asset.metadata) if asset.metadata else None
         model.quality_score = asset.quality_score
         model.uploaded_by_user_id = asset.uploaded_by_user_id or model.uploaded_by_user_id

packages/adapters/sqlalchemy_impl/generation_task_repository.py

@@ -43,7 +43,6 @@ def _to_domain(model: GenerationTaskModel) -> GenerationTask:
         output_height=getattr(model, "output_height", 720) or 720,
         cover_url=getattr(model, "cover_url", "") or "",
         title_config=dict(getattr(model, "title_config", {}) or {}),
-        extra_meta=dict(getattr(model, "extra_meta", {}) or {}),
         logs=model.logs or "[]",
         created_at=model.created_at,
         updated_at=model.updated_at,
@@ -89,7 +88,6 @@ class SQLAlchemyGenerationTaskRepository:
             output_height=task.output_height,
             cover_url=task.cover_url or "",
             title_config=dict(task.title_config) if task.title_config else {},
-            extra_meta=dict(task.extra_meta) if task.extra_meta else {},
             logs=task.logs,
             created_at=task.created_at,
             updated_at=task.updated_at,
@@ -324,7 +322,6 @@ class SQLAlchemyGenerationTaskRepository:
         model.output_height = task.output_height
         model.cover_url = task.cover_url or ""
         model.title_config = dict(task.title_config) if task.title_config else {}
-        model.extra_meta = dict(task.extra_meta) if task.extra_meta else {}
         model.logs = task.logs
         self.session.commit()
         return task

packages/application/generated_video_finalize.py

@@ -52,18 +52,6 @@ class RenderedOutput:
     def from_dict(cls, data: dict[str, Any]) -> "RenderedOutput":
         if not isinstance(data, dict):
             raise ValueError("rendered_output must be a dict")
-        # fingerprint_chunks 历史上有两种位置：
-        #   1) 顶层 ``fingerprint_chunks``（由 compute_render_fingerprint_and_dedup 直接返回）
-        #   2) 嵌套在 ``fingerprint_dict["chunks"]``（VideoFingerprint.to_dict() 序列化的结构）
-        # 顶层优先；顶层为空时回退到嵌套位置，兼容旧数据。
-        fp_dict = data.get("fingerprint_dict") or {}
-        chunks_raw = data.get("fingerprint_chunks")
-        if not chunks_raw and isinstance(fp_dict, dict):
-            chunks_raw = fp_dict.get("chunks")
-        # md5 同样可能在顶层或嵌套在 fingerprint_dict 内（历史数据兼容）
-        md5_value = data.get("video_fingerprint_md5")
-        if not md5_value and isinstance(fp_dict, dict):
-            md5_value = fp_dict.get("md5")
         return cls(
             file_url=str(data.get("file_url") or ""),
             file_size=int(data.get("file_size") or 0),
@@ -77,14 +65,14 @@ class RenderedOutput:
             batch_id=str(data.get("batch_id") or ""),
             project_id=str(data.get("project_id") or ""),
             user_id=str(data.get("user_id") or ""),
-            fingerprint_dict=fp_dict or None,
-            fingerprint_chunks=chunks_raw if isinstance(chunks_raw, list) else None,
+            fingerprint_dict=data.get("fingerprint_dict"),
+            fingerprint_chunks=data.get("fingerprint_chunks"),
             is_duplicate=bool(data.get("is_duplicate", False)),
             duplicate_of=data.get("duplicate_of"),
             duplicate_rate=_safe_float(data.get("duplicate_rate")),
             match_count=_safe_int(data.get("match_count")),
             visual_similarity=_safe_float(data.get("visual_similarity")),
-            video_fingerprint_md5=str(md5_value or ""),
+            video_fingerprint_md5=str(data.get("video_fingerprint_md5") or ""),
         )
 
 

packages/config/api_settings.py

@@ -51,7 +51,7 @@ class APISettings(SharedSettings):
     def validate_jwt_secret_key(cls, v):
         if v is None or v == "":
             raise ValueError(
-                "JWT_SECRET_KEY must be set via environment variable. Do not use default value in production!"
+                "JWT_SECRET_KEY must be set via environment variable. " "Do not use default value in production!"
             )
         # Block known insecure default values
         insecure_defaults = [
@@ -63,7 +63,7 @@ class APISettings(SharedSettings):
         ]
         if v.lower() in [d.lower() for d in insecure_defaults]:
             raise ValueError(
-                f"JWT_SECRET_KEY '{v}' is insecure. Please set a strong random secret via environment variable."
+                f"JWT_SECRET_KEY '{v}' is insecure. " "Please set a strong random secret via environment variable."
             )
         return v
 
@@ -248,10 +248,6 @@ class APISettings(SharedSettings):
     def OSS_ENDPOINT(self) -> str:
         return self.oss_endpoint
 
-    @property
-    def OSS_INTERNAL_ENDPOINT(self) -> str:
-        return self.effective_oss_internal_endpoint
-
     @property
     def OSS_ACCESS_KEY_ID(self) -> str:
         return self.oss_access_key_id

packages/config/base.py

@@ -47,37 +47,12 @@ class SharedSettings(BaseSettings):
 
     # ── OSS 阿里云 ──────────────────────────────────────────────────────
     oss_endpoint: str = "oss-cn-hangzhou.aliyuncs.com"
-    # 内网 endpoint：ECS VPC 内访问 OSS 用（千兆带宽、免公网流量费）。
-    # 为空时自动从 oss_endpoint 推导：若 oss_endpoint 是阿里云公网域名（形如
-    # oss-cn-<region>.aliyuncs.com），自动加 -internal 得到内网域名；其他情况
-    # （自定义域名/本地 MinIO/非阿里云）回退使用 oss_endpoint。
-    # 显式填同值可以覆盖自动推导、强制所有流量都走公网。
-    oss_internal_endpoint: str = ""
     oss_access_key_id: str = ""
     oss_access_key_secret: str = ""
     oss_bucket_name: str = "xiaoxia-autocut"
     oss_direct_upload_max_mb: int = 2000
     oss_direct_upload_expire_seconds: int = 900
 
-    @property
-    def effective_oss_internal_endpoint(self) -> str:
-        """实际用于 SDK 内网访问的 endpoint（带 -internal 自动推导）。"""
-        if self.oss_internal_endpoint:
-            return self.oss_internal_endpoint
-        ep = self.oss_endpoint.strip()
-        scheme = ""
-        host = ep
-        if ep.startswith("https://"):
-            scheme = "https://"
-            host = ep[len("https://") :]
-        elif ep.startswith("http://"):
-            scheme = "http://"
-            host = ep[len("http://") :]
-        # 阿里云公网域名自动推导：oss-cn-<region>.aliyuncs.com → oss-cn-<region>-internal.aliyuncs.com
-        if host.endswith(".aliyuncs.com") and "-internal" not in host and host.startswith("oss-cn-"):
-            host = host[: -len(".aliyuncs.com")] + "-internal.aliyuncs.com"
-        return f"{scheme}{host}" if scheme else host
-
     # ── CosyVoice (阿里云百炼语音合成) ───────────────────────────────────
     cosyvoice_api_key: str = ""
     cosyvoice_base_url: str = "https://dashscope.aliyuncs.com/api/v1"

packages/shared/celery_queues.py

@@ -1,14 +1,14 @@
 """Celery 队列定义与路由配置（API / Worker 共享）。
 
-#1714 + #2073 队列分流：用户同步等待的实时任务路由到 `generation` 高优队列，
-由专用 generation worker 独占消费；素材入库/转码/AI 分析/查重等后台批量任务路由
-到 `transcode` 队列；beat 定时清理等轻量维护任务走默认 `celery` 队列。
-transcode / celery 队列积压时，generation 队列仍能被立即领取，不阻塞用户实时链路。
+#1714 队列隔离：用户等待的视频生成任务路由到高优先级 `generation` 队列，
+由专用 worker 进程独占消费；素材入库/转码等后台批量任务路由到 `transcode`
+队列；其余杂项任务走默认 `celery` 队列。转码队列积压时，视频生成任务
+仍能被 generation worker 立即领取执行，不会排队。
 
 队列说明：
-- generation: 用户同步等待的实时任务（视频生成、TTS、音色克隆、lipsync、AI 数字人、人声/背景提取）
-- transcode:  后台批量/异步任务（素材入库转码、AI 分类打标、质量评分、原子切片、查重、批量下载/缩略图）
-- celery:     beat 定时巡检/清理等轻量维护任务（极短、低优、不占业务槽）
+- generation: 用户提交的视频生成/预览渲染（延迟敏感，资源消耗大）
+- transcode: 素材入库（HEVC 转码）、AI 分类、素材查重（批量、可排队）
+- celery(默认): 配音、语音、下载缩略图、定时清理等杂项
 """
 
 from __future__ import annotations
@@ -20,9 +20,8 @@ QUEUE_GENERATION = "generation"
 QUEUE_TRANSCODE = "transcode"
 QUEUE_DEFAULT = "celery"
 
-# 三个消费组各自消费的队列列表（顺序即优先级：高优队列排在前面）
-WORKER_QUEUES_GENERATION = (QUEUE_GENERATION,)
-WORKER_QUEUES_TRANSCODE = (QUEUE_TRANSCODE, QUEUE_DEFAULT)
+# Worker 消费的队列列表（顺序即优先级：高优队列排在前面）
+WORKER_QUEUES = (QUEUE_GENERATION, QUEUE_TRANSCODE, QUEUE_DEFAULT)
 
 # 队列声明：持久化队列，broker 重启不丢消息
 task_queues = (
@@ -32,45 +31,15 @@ task_queues = (
 )
 
 # ── 任务路由表：task name → 队列 ──
-# 键支持 celery 标准通配符。所有生产端（API send_task / worker 内 send_task）
-# 未显式指定 queue 时按此表路由；漏配会走默认队列 celery，被 transcode worker 消费。
-# 新增实时任务务必在此表显式路由到 generation，避免落到后台队列排队。
+# 键支持 celery 标准通配符。
 task_routes = {
-    # ── 高优先级：用户同步等待的实时链路 ──
-    # 视频生成（主链路）
+    # 高优先级：用户等待的视频生成
     "worker.generate_video": {"queue": QUEUE_GENERATION},
-    # TTS 合成 / 片段合成（配音页、视频生成配乐/TTS 链路）
-    "worker.process_tts_synthesis": {"queue": QUEUE_GENERATION},
-    "worker.process_tts_segment_synthesis": {"queue": QUEUE_GENERATION},
-    # 音色克隆（用户主动上传样本等待克隆完成）
-    "worker.process_voice_clone": {"queue": QUEUE_GENERATION},
-    # 人声/背景提取（音色克隆前置步骤，用户同步等待）
-    "worker.extract_voice": {"queue": QUEUE_GENERATION},
-    "worker.extract_background": {"queue": QUEUE_GENERATION},
-    # AI 数字人渲染（用户主动触发，等待成片）
-    "ai_avatar_render.execute": {"queue": QUEUE_GENERATION},
-    # GPU MuseTalk 口型同步（用户等成片，链路子任务全部走 generation 避免跨队列阻塞）
-    "lipsync_gpu_process_async": {"queue": QUEUE_GENERATION},
-    "lipsync_tts.synthesize_and_submit": {"queue": QUEUE_GENERATION},
-    "lipsync_tts.poll_mediakit_status": {"queue": QUEUE_GENERATION},
-    "lipsync_tts.persist_output_video": {"queue": QUEUE_GENERATION},
-    # ── 后台批量：素材入库/转码 + AI 分析/打标 + 查重，积压不影响生成 ──
+    # 后台批量：素材入库/转码 + AI 分类 + 素材查重，积压不影响生成
     "worker.ingest_asset": {"queue": QUEUE_TRANSCODE},
     "worker.classify_asset": {"queue": QUEUE_TRANSCODE},
-    "worker.calculate_asset_quality": {"queue": QUEUE_TRANSCODE},
-    "worker.generate_atom_clips": {"queue": QUEUE_TRANSCODE},
-    "worker.tag_atom_clip": {"queue": QUEUE_TRANSCODE},
-    "worker.backfill_atom_clip_tags": {"queue": QUEUE_TRANSCODE},
     "worker.process_duplication_check": {"queue": QUEUE_TRANSCODE},
     "worker.check_duplicate": {"queue": QUEUE_TRANSCODE},
-    "worker.batch_download_videos": {"queue": QUEUE_TRANSCODE},
-    "worker.batch_generate_thumbnails": {"queue": QUEUE_TRANSCODE},
-    # ── beat 定时清理/巡检任务走默认 celery 队列（由 transcode worker 消费）──
-    # 未在此表显式列出的 cleanup 任务会落到默认队列 celery，不占 generation 槽位。
-    "worker.cleanup_stale_pending_tasks": {"queue": QUEUE_DEFAULT},
-    "worker.cleanup_stale_running_tasks": {"queue": QUEUE_DEFAULT},
-    "worker.cleanup_stale_ingest_jobs": {"queue": QUEUE_DEFAULT},
-    "worker.cleanup_stale_voice_clones": {"queue": QUEUE_DEFAULT},
 }
 
 # 生成任务的预取数：渲染是长任务，预取 1 避免任务被某个 worker 占住不调度
@@ -78,10 +47,11 @@ GENERATION_WORKER_PREFETCH_MULTIPLIER = 1
 
 
 def apply_queue_settings(app) -> None:
-    """把队列分流配置应用到 Celery app（API 生产端与 Worker 消费端都要调用）。
+    """把队列隔离配置应用到 Celery app（API 生产端与 Worker 消费端都要调用）。
 
     配置 task_queues / task_routes / task_default_queue。生产端靠 task_routes
-    把消息投递到对应队列；消费端靠启动参数 -Q 控制自己消费哪些队列（entrypoint）。
+    把消息投递到对应队列；消费端靠 task_queues 声明自己消费哪些队列
+    （实际消费集由启动参数 -Q 控制）。
     """
     app.conf.task_queues = task_queues
     app.conf.task_routes = task_routes

packages/shared/gpu_encoder.py

@@ -7,21 +7,13 @@
   3. 生成 relay 一次性 key，构造两个带 token 的 URL：
      - put_url：给 P4000 回传结果，走 relay_base_url（Tailscale host:8092）
      - get/del_url：worker 自己下载+清理用，走 relay_internal_base_url（Docker DNS 直连 API）
-  4. 【冷启动防护】距上次成功通信 >60s 时，先 GET /health 预热 Tailscale 链路（短超时快速失败）
-  5. POST P4000 /api/render/sync：inputs={"in.mp4": "<mezzanine-get-url>"}, output_url="<put_url>"
+  4. POST P4000 /api/render/sync：inputs={"in.mp4": "<mezzanine-get-url>"}, output_url="<put_url>"
      ffmpeg_args: -i in.mp4 [-vf <vf>] -c:v h264_nvenc ... -an/-c:a aac -f mp4 pipe:1
-     - 首字节用短超时(默认20s)，避免链路卡死空等上百秒；首字节到达后放宽到 ffmpeg_timeout+60s
-  6. P4000 从 relay GET mezzanine → h264_nvenc 编码 → PUT 最终 mp4 到 put_url
-  7. 本客户端通过 get_url（Docker 内网）下载最终文件到 output_path，然后 DELETE 清理
-  8. 删除 relay 上的 mezzanine 临时文件（以及 OSS fallback 的 key）
+  5. P4000 从 relay GET mezzanine → h264_nvenc 编码 → PUT 最终 mp4 到 put_url
+  6. 本客户端通过 get_url（Docker 内网）下载最终文件到 output_path，然后 DELETE 清理
+  7. 删除 relay 上的 mezzanine 临时文件（以及 OSS fallback 的 key）
 
 任何环节失败抛 GpuEncodeError，调用方应 fallback 到 CPU libx264。
-
-冷启动/链路卡顿背景（2026-09-27 实测）：P4000 与 staging 之间走 Tailscale，长时间空闲
-（>7h）后首次请求曾出现 150s 延迟才真正开始下载 mezzanine，期间 ffmpeg 尚未启动、GPU 空闲。
-根因在服务端/网络层（可能是 Tailscale DERP 打洞或 httpx 连接池重建），本客户端通过
-pre_warm + 首字节短超时做兜底：预热打通链路 + 20s 内收不到首字节就快速失败让 CPU fallback，
-不再让用户等满 150s+。
 """
 
 from __future__ import annotations
@@ -71,13 +63,6 @@ class GpuEncoderClient:
         mezzanine_transport: str = "relay",
         sync_timeout: int = 300,
         health_timeout: float = 3.0,
-        # 提交编码任务前先发一次 /health 预热 Tailscale 链路，避免长时间空闲后首次请求
-        # 因 DERP 打洞/NAT 映射过期/Tailscale 连接重建而阻塞上百秒。
-        pre_warm: bool = True,
-        # POST 首次响应超时：P4000 已收到请求后应该在数秒内开始下载 inputs；
-        # 如果超过这个值还没收到任何响应字节，说明链路/服务卡住，快速失败让调用方 fallback CPU。
-        # 注意：ffmpeg 编码本身靠 body.timeout 控制（300s），不应该被这个超时影响。
-        post_first_byte_timeout: float = 20.0,
         vcodec: str = "h264_nvenc",
         preset: str = "p4",
         crf: int = 23,
@@ -95,16 +80,12 @@ class GpuEncoderClient:
         self.mezzanine_transport = mezzanine_transport.lower()  # "relay" | "oss"
         self.sync_timeout = sync_timeout
         self.health_timeout = health_timeout
-        self.pre_warm = pre_warm
-        self.post_first_byte_timeout = post_first_byte_timeout
         self.vcodec = vcodec
         self.preset = preset
         self.crf = crf
         self.bitrate = bitrate
         self._relay_secret = relay_secret
         self.oss_tmp_prefix = oss_tmp_prefix.rstrip("/") + "/" if oss_tmp_prefix else "tmp/gpu-mezzanine/"
-        # 上次与 P4000 成功通信的时间戳（用于判断是否需要 pre_warm 预热）
-        self._last_ok_ts: float = 0.0
 
     RELAY_PATH_PREFIX = "/api/v1/internal/gpu-relay"
 
@@ -147,16 +128,13 @@ class GpuEncoderClient:
         except (urllib.error.URLError, socket.timeout, TimeoutError, json.JSONDecodeError, ConnectionError) as e:
             return GpuHealth(healthy=False, error=f"health probe failed: {e}")
         try:
-            h = GpuHealth(
+            return GpuHealth(
                 healthy=data.get("status") == "healthy",
                 worker=str(data.get("worker", "")),
                 gpu_name=(data.get("gpu") or {}).get("name", ""),
                 nvenc_h264=bool((data.get("nvenc") or {}).get("h264_nvenc")),
                 nvenc_hevc=bool((data.get("nvenc") or {}).get("hevc_nvenc")),
             )
-            if h.healthy:
-                self._last_ok_ts = time.time()
-            return h
         except Exception as e:  # noqa: BLE001
             return GpuHealth(healthy=False, error=f"malformed health response: {e}")
 
@@ -249,8 +227,7 @@ class GpuEncoderClient:
                 ffmpeg_args.append("-an")
             ffmpeg_args.extend(["-f", "mp4", "pipe:1"])
 
-            # 4. pre-warm then call P4000 sync render
-            self._warm_up_if_needed()
+            # 4. call P4000 sync render
             body = {
                 "inputs": {"in.mp4": input_url},
                 "ffmpeg_args": ffmpeg_args,
@@ -258,7 +235,6 @@ class GpuEncoderClient:
                 "timeout": int(timeout),
             }
             job = self._post_sync(body)
-            self._last_ok_ts = time.time()
             logger.info(
                 "[gpu-encoder] P4000 done: job_id=%s rc=%s size=%s dur=%ss transport=%s",
                 job.get("job_id"),
@@ -323,38 +299,9 @@ class GpuEncoderClient:
             raise GpuEncodeError("GPU_ENCODE_RELAY_SECRET not set")
         return secret
 
-    def _warm_up_if_needed(self) -> None:
-        """POST 前预热：如果距上次成功通信超过 idle 阈值，先打 /health 打通 Tailscale 链路。
-
-        背景：Tailscale 在长时间空闲（几小时）后，到对端的直连 NAT 映射可能过期，
-        首次请求会走 DERP 中继打洞；极少数情况下打洞/重连会卡住上百秒（曾观测到 150s 延迟）。
-        预热请求本身走短超时快速失败，不会阻塞主流程；预热成功后再发 POST。
-        """
-        if not self.pre_warm:
-            return
-        idle = time.time() - self._last_ok_ts
-        # 空闲超过 60s 才预热（正常流水线里相邻任务间隔通常 <10s，没必要每次都打）
-        if idle < 60:
-            return
-        url = f"{self.endpoint}/health"
-        t0 = time.time()
-        try:
-            with urllib.request.urlopen(url, timeout=min(self.health_timeout, 3.0)) as resp:
-                resp.read()
-            self._last_ok_ts = time.time()
-            logger.debug("[gpu-encoder] pre-warm ok: took=%.2fs idle=%.0fs", time.time() - t0, idle)
-        except (urllib.error.URLError, socket.timeout, TimeoutError, ConnectionError, OSError) as e:
-            # 预热失败不致命——主 POST 会带自己的超时，再失败就抛 GpuEncodeError 让调用方 fallback
-            logger.warning("[gpu-encoder] pre-warm probe failed (will try POST anyway): %s", e)
-
     def _post_sync(self, body: dict[str, Any]) -> dict[str, Any]:
         url = f"{self.endpoint}/api/render/sync"
-        ffmpeg_timeout = body.get("timeout", self.sync_timeout)
-        # 连接 + 首字节用短超时（防链路卡死数百秒）；首字节到达后给 ffmpeg 留足编码+上传时间
-        # Python urllib 的 timeout 是整个请求总超时，所以用"两段式"：
-        #   阶段1：先 read(1) 拿首字节，用短超时；
-        #   阶段2：再 read() 读完整 body，用 ffmpeg_timeout+60。
-        connect_timeout = min(max(self.post_first_byte_timeout, 5.0), 30.0)
+        req_timeout = body.get("timeout", self.sync_timeout) + 60
         payload = json.dumps(body).encode("utf-8")
         req = urllib.request.Request(
             url,
@@ -363,45 +310,14 @@ class GpuEncoderClient:
             method="POST",
         )
         t0 = time.time()
-        first_byte_ok = False
-        resp = None
         try:
-            resp = urllib.request.urlopen(req, timeout=connect_timeout)
-            # 读首字节 —— 如果 P4000/链路卡死，这里会在 connect_timeout 内抛超时
-            first_chunk = resp.read(1)
-            first_byte_ok = True
-            logger.debug(
-                "[gpu-encoder] P4000 first byte in %.2fs (connect_timeout=%.1fs)",
-                time.time() - t0,
-                connect_timeout,
-            )
-            # 剩余用长超时（给底层socket放宽时限；如果是mock/不支持，则跳过）
-            try:
-                resp.fp._sock.settimeout(ffmpeg_timeout + 60)
-            except (AttributeError, OSError):
-                pass
-            rest = resp.read()
-            raw = (first_chunk + rest).decode("utf-8")
-            resp.close()
-            resp = None
+            with urllib.request.urlopen(req, timeout=req_timeout) as resp:
+                raw = resp.read().decode("utf-8")
         except urllib.error.HTTPError as e:
             detail = e.read().decode("utf-8", errors="replace")[:1000]
             raise GpuEncodeError(f"P4000 HTTP {e.code}: {detail}") from e
-        except (urllib.error.URLError, socket.timeout, TimeoutError, ConnectionError, OSError) as e:
-            waited = time.time() - t0
-            hint = "first-byte" if not first_byte_ok else "ffmpeg/upload"
-            # 统一以 "connection error" 开头，便于上层 fallback 逻辑用关键词识别；
-            # 末尾再附带具体错误（timed out / refused ...）供排障
-            raise GpuEncodeError(
-                f"P4000 {hint} connection error after {waited:.1f}s "
-                f"(connect_timeout={connect_timeout:.0f}s, ffmpeg_timeout={ffmpeg_timeout}s): {e}"
-            ) from e
-        finally:
-            if resp is not None:
-                try:
-                    resp.close()
-                except Exception:
-                    pass
+        except (urllib.error.URLError, socket.timeout, TimeoutError, ConnectionError) as e:
+            raise GpuEncodeError(f"P4000 connection error: {e}") from e
         try:
             result = json.loads(raw)
         except json.JSONDecodeError as e:
@@ -535,8 +451,6 @@ def _build_client_from_settings() -> Optional[GpuEncoderClient]:
         mezzanine_transport=getattr(settings, "gpu_encode_mezzanine_transport", "relay") or "relay",
         sync_timeout=getattr(settings, "gpu_encode_sync_timeout", 300),
         health_timeout=getattr(settings, "gpu_encode_health_timeout", 3.0),
-        pre_warm=getattr(settings, "gpu_encode_pre_warm", True),
-        post_first_byte_timeout=getattr(settings, "gpu_encode_post_first_byte_timeout", 20.0),
         vcodec=getattr(settings, "gpu_encode_vcodec", "h264_nvenc"),
         preset=getattr(settings, "gpu_encode_preset", "p4"),
         crf=getattr(settings, "gpu_encode_crf", 23),

packages/shared/storage.py

@@ -5,13 +5,6 @@
 - Worker端 oss_helpers 的高级能力（分片上传/超时保护/HTTP下载/Asset路径解析）
 
 所有服务都通过这个统一入口与存储交互，消除重复实现。
-
-P1 (2026-09-28) OSS 双 endpoint 分离：
-- 内部 bucket（self.bucket）：使用 internal endpoint（VPC 千兆带宽），
-  用于所有 SDK 上传/下载/删除/object_exists 操作；
-- 公网 bucket（self.public_bucket）：使用公网 endpoint，仅用于 sign_url
-  生成给前端/P4000/MediaKit 等外网访问方用的预签名 URL；
-- public_url 永远拼公网域名，不随 internal endpoint 变化。
 """
 
 from __future__ import annotations
@@ -49,44 +42,19 @@ OSS_MULTIPART_NUM_THREADS = 3  # 分片上传并发数
 OSS_HTTP_DOWNLOAD_TIMEOUT = 300  # HTTP下载超时（秒）
 
 
-def _make_bucket(
-    auth,
-    endpoint: str,
-    bucket_name: str,
-    *,
-    connect_timeout: int = OSS_CONNECT_TIMEOUT,
-    app_name: str = "",
-):
-    """构造 oss2.Bucket，自动补 https:// 前缀。"""
-    if not endpoint.startswith(("http://", "https://")):
-        endpoint = f"https://{endpoint}"
-    kwargs: dict = {"connect_timeout": connect_timeout}
-    if app_name:
-        kwargs["app_name"] = app_name
-    return oss2.Bucket(auth, endpoint, bucket_name, **kwargs)
-
-
 class SharedStorageService(StoragePort):
-    """统一存储服务 — 实现 StoragePort，API 和 Worker 共用。"""
+    """统一存储服务 — 实现 StoragePort，API 和 Worker 共用。
 
-    # 类级默认值，方便单测 mock __init__ 后实例仍有这些属性
-    bucket: Optional[object] = None
-    public_bucket: Optional[object] = None
-    public_endpoint: str = ""
-    internal_endpoint: str = ""
-    public_url: str = ""
-    local_url_prefix: str = "/generated-files"
-    bucket_name: str = ""
+    整合了原 SharedStorageService + oss_helpers 的全部能力。
+    """
 
     def __init__(self):
         settings = get_shared_settings()
         self.bucket_name = settings.oss_bucket_name
-        self.public_endpoint = settings.oss_endpoint  # 公网 endpoint，用于签名 URL
-        self.internal_endpoint = settings.effective_oss_internal_endpoint  # 内网 endpoint，SDK 用
-        self.public_url = f"https://{settings.oss_bucket_name}.{self._public_host()}"
+        self.endpoint = settings.oss_endpoint
+        self.public_url = f"https://{settings.oss_bucket_name}.{settings.oss_endpoint}"
         self.local_url_prefix = os.getenv("GENERATED_FILES_URL_PREFIX", "/generated-files")
-        self.bucket: Optional[object] = None  # internal: SDK 上传/下载/删除
-        self.public_bucket: Optional[object] = None  # public: sign_url 给外网
+        self.bucket = None
 
         self.access_key_id = settings.oss_access_key_id
         self.access_key_secret = settings.oss_access_key_secret
@@ -97,26 +65,21 @@ class SharedStorageService(StoragePort):
         if has_key_id and has_key_secret:
             if oss2 is not None:
                 try:
+                    # endpoint 不带 scheme 时补 https:// 前缀
+                    bucket_endpoint = self.endpoint
+                    if not bucket_endpoint.startswith(("http://", "https://")):
+                        bucket_endpoint = f"https://{bucket_endpoint}"
                     auth = oss2.Auth(self.access_key_id, self.access_key_secret)
-                    self.bucket = _make_bucket(
+                    self.bucket = oss2.Bucket(
                         auth,
-                        self.internal_endpoint,
+                        bucket_endpoint,
                         self.bucket_name,
-                        app_name="xiaoxia-internal",
+                        connect_timeout=OSS_CONNECT_TIMEOUT,
                     )
-                    self.public_bucket = _make_bucket(
-                        auth,
-                        self.public_endpoint,
-                        self.bucket_name,
-                        app_name="xiaoxia-public",
-                    )
-                    same_ep = self.internal_endpoint == self.public_endpoint
                     logger.info(
-                        "OSS initialized: public_ep=%s internal_ep=%s bucket=%s dual=%s",
-                        self.public_endpoint,
-                        self.internal_endpoint,
+                        "OSS initialized: endpoint=%s bucket=%s",
+                        self.endpoint,
                         self.bucket_name,
-                        "no" if same_ep else "yes",
                     )
                 except Exception as error:
                     logger.error("Failed to initialize OSS bucket client: %s", error)
@@ -130,35 +93,26 @@ class SharedStorageService(StoragePort):
                 missing.append("OSS_ACCESS_KEY_SECRET")
             logger.error("OSS credentials not configured — missing: %s", ", ".join(missing))
 
-    def _public_host(self) -> str:
-        ep = self.public_endpoint
-        if ep.startswith("https://"):
-            return ep[len("https://") :]
-        if ep.startswith("http://"):
-            return ep[len("http://") :]
-        return ep
-
     # ── 诊断 ───────────────────────────────────────────────────────────
 
     def diagnose(self) -> None:
+        """输出存储配置诊断日志。"""
         key_id_display = (
             f"{self.access_key_id[:4]}...{self.access_key_id[-4:]}" if len(self.access_key_id) > 8 else "(empty)"
         )
         logger.info(
-            "[OSS诊断] public_ep=%s internal_ep=%s bucket=%s ak=%s",
-            self.public_endpoint,
-            self.internal_endpoint,
+            "[OSS诊断] endpoint=%s bucket_name=%s access_key_id=%s",
+            self.endpoint,
             self.bucket_name,
             key_id_display,
         )
         if self.bucket is None:
-            logger.error("[OSS诊断] ❌ bucket(internal)=None")
+            logger.error(
+                "[OSS诊断] ❌ bucket=None — 预签名URL不可用！"
+                "原因: OSS_ACCESS_KEY_ID/OSS_ACCESS_KEY_SECRET 未配置或 oss2 未安装。"
+            )
         else:
-            logger.info("[OSS诊断] ✅ bucket(internal) 就绪")
-        if self.public_bucket is None:
-            logger.error("[OSS诊断] ❌ public_bucket=None")
-        else:
-            logger.info("[OSS诊断] ✅ public_bucket 就绪，公网签名URL可用")
+            logger.info("[OSS诊断] ✅ bucket 已配置，预签名URL可用")
 
     # ── 工具方法 ───────────────────────────────────────────────────────
 
@@ -168,15 +122,20 @@ class SharedStorageService(StoragePort):
         return path.startswith(f"{self.local_url_prefix}/")
 
     def _normalize_storage_key(self, storage_key_or_url: str) -> str:
+        """从 URL 提取存储键，并做 URL 解码。
+
+        防止 URL 编码的字符（空格=%20、中文=%XX）导致签名不匹配。
+        """
         if storage_key_or_url.startswith("http://") or storage_key_or_url.startswith("https://"):
             parsed = urlparse(storage_key_or_url)
             return unquote(parsed.path.lstrip("/"))
         return storage_key_or_url.lstrip("/")
 
     def normalize_storage_key(self, storage_key_or_url: str) -> str:
+        """从 URL 提取存储键（公开方法）。"""
         return self._normalize_storage_key(storage_key_or_url)
 
-    # ── 上传（SDK 走 internal endpoint）───────────────────────────────
+    # ── 上传 ───────────────────────────────────────────────────────────
 
     def upload_file(
         self,
@@ -184,22 +143,21 @@ class SharedStorageService(StoragePort):
         storage_key: str,
         content_type: str = "application/octet-stream",
     ) -> str:
+        """上传文件到存储，返回公开 URL（简单上传，API端原有行为）。
+
+        - 路径字符串 → bucket.put_object_from_file
+        - 类文件对象 → bucket.put_object
+        - bucket未配置 → 抛 RuntimeError
+        """
         if self.bucket is None:
             raise RuntimeError("OSS storage is not configured")
+
         try:
             if isinstance(file_or_path, (str, Path)):
-                self.bucket.put_object_from_file(
-                    storage_key,
-                    str(file_or_path),
-                    headers={"Content-Type": content_type},
-                )
+                self.bucket.put_object_from_file(storage_key, str(file_or_path), headers={"Content-Type": content_type})
             else:
-                file_or_path.seek(0)
-                self.bucket.put_object(
-                    storage_key,
-                    file_or_path,
-                    headers={"Content-Type": content_type},
-                )
+                file_or_path.seek(0)  # type: ignore[attr-defined]
+                self.bucket.put_object(storage_key, file_or_path, headers={"Content-Type": content_type})
             return f"{self.public_url}/{storage_key}"
         except Exception as e:
             raise Exception(f"Failed to upload file to OSS: {e}") from e
@@ -209,6 +167,14 @@ class SharedStorageService(StoragePort):
         local_path: str | Path,
         storage_key: str,
     ) -> Optional[str]:
+        """智能上传：大文件自动分片+超时保护（从 oss_helpers 合并）。
+
+        - 大文件（>100MB）走分片上传，3 线程并发
+        - 总超时 300s，防止网络异常时挂死
+        - 成功返回 URL，失败返回 None（不抛异常）
+
+        Worker端 oss_helpers.upload_to_oss 的统一入口。
+        """
         local_path = Path(local_path)
         if not local_path.exists():
             logger.error("上传文件不存在: %s", local_path)
@@ -232,10 +198,10 @@ class SharedStorageService(StoragePort):
 
                 if use_multipart:
                     logger.info(
-                        "大文件分片上传(internal): key=%s size=%.1fMB part=%dMB threads=%d",
+                        "大文件分片上传: storage_key=%s, size=%.1fMB, part_size=%dMB, threads=%d",
                         storage_key[:80],
-                        file_size / 1048576,
-                        OSS_PART_SIZE // 1048576,
+                        file_size / 1024 / 1024,
+                        OSS_PART_SIZE // 1024 // 1024,
                         OSS_MULTIPART_NUM_THREADS,
                     )
                     oss2.resumable_upload(
@@ -248,6 +214,7 @@ class SharedStorageService(StoragePort):
                     )
                 else:
                     self.bucket.put_object_from_file(storage_key, str(local_path))
+
                 result["url"] = f"{self.public_url}/{storage_key}"
             except Exception as e:
                 result["error"] = e
@@ -255,54 +222,73 @@ class SharedStorageService(StoragePort):
             finally:
                 done.set()
 
-        t = threading.Thread(target=_do_upload, daemon=True)
-        t.start()
+        upload_thread = threading.Thread(target=_do_upload, daemon=True)
+        upload_thread.start()
         finished = done.wait(timeout=OSS_UPLOAD_TOTAL_TIMEOUT)
 
         if not finished:
             logger.error(
-                "OSS 上传超时(%ds): key=%s size=%.1fMB",
+                "OSS 上传超时（%.0fs），强制中止: storage_key=%s, size=%.1fMB",
                 OSS_UPLOAD_TOTAL_TIMEOUT,
                 storage_key[:80],
-                result["file_size"] / 1048576 if result["file_size"] else 0,
+                result["file_size"] / 1024 / 1024 if result["file_size"] else 0,
             )
             return None
-        return None if result["error"] else result["url"]
 
-    # ── 下载（SDK 走 internal endpoint）───────────────────────────────
+        if result["error"]:
+            return None
+
+        return result["url"]
+
+    # ── 下载 ───────────────────────────────────────────────────────────
 
     def download_file(self, storage_key: str, local_path: str | Path) -> None:
+        """从 OSS 下载文件（简单下载，API端原有行为）。
+
+        bucket未配置 → 抛 RuntimeError
+        """
         if self.bucket is None:
             raise RuntimeError("OSS storage is not configured")
+
         local_path = Path(local_path)
         os.makedirs(local_path.parent, exist_ok=True)
         try:
-            self.bucket.get_object_to_file(
-                self._normalize_storage_key(storage_key),
-                str(local_path),
-            )
+            self.bucket.get_object_to_file(self._normalize_storage_key(storage_key), str(local_path))
         except Exception as e:
             raise Exception(f"Failed to download file from OSS: {e}") from e
 
     def download_asset(self, asset_storage_key: str, local_path: str | Path) -> bool:
+        """下载素材（从 oss_helpers 合并）。
+
+        自动识别输入类型：
+        - 完整 URL → 走 HTTP 下载（支持预签名URL）
+        - 存储键 → 走 oss2 SDK 下载
+
+        成功返回 True，失败返回 False（不抛异常）。
+        """
         local_path = Path(local_path)
         os.makedirs(local_path.parent, exist_ok=True)
+
+        # 完整URL走HTTP下载（兼容预签名URL）
         if asset_storage_key.startswith(("http://", "https://")):
             return self._download_via_http(asset_storage_key, local_path)
+
+        # OSS存储键走SDK
         if self.bucket is None:
-            logger.error("OSS not configured: %s", asset_storage_key[:80])
+            logger.error("OSS not configured, cannot download: %s", asset_storage_key[:80])
             return False
         try:
-            self.bucket.get_object_to_file(
-                self._normalize_storage_key(asset_storage_key),
-                str(local_path),
-            )
+            self.bucket.get_object_to_file(self._normalize_storage_key(asset_storage_key), str(local_path))
             return local_path.exists() and local_path.stat().st_size > 0
         except Exception:
             logger.exception("下载素材失败: %s", asset_storage_key)
             return False
 
     def _download_via_http(self, url: str, local_path: Path) -> bool:
+        """通过 HTTP 下载文件（支持预签名 URL）。
+
+        流式下载避免大文件内存溢出。
+        """
         try:
             resp = requests.get(url, stream=True, timeout=OSS_HTTP_DOWNLOAD_TIMEOUT)
             resp.raise_for_status()
@@ -312,64 +298,83 @@ class SharedStorageService(StoragePort):
                         f.write(chunk)
             return local_path.exists() and local_path.stat().st_size > 0
         except Exception:
-            logger.exception("HTTP下载失败: %s", url[:100])
+            logger.exception("HTTP下载素材失败: %s", url[:100])
             return False
 
-    # ── URL 生成（sign_url 用 public_bucket 签公网域名）───────────────
+    # ── URL 生成 ──────────────────────────────────────────────────────
 
     def get_url(self, storage_key: str) -> str:
+        """获取公开 URL。"""
         return f"{self.public_url}/{storage_key}"
 
-    def _sign_bucket(self):
-        """签名优先用 public_bucket，回退到 bucket。"""
-        return self.public_bucket or self.bucket
-
     def get_download_url(self, storage_key_or_url: str, expires_seconds: int = 3600) -> str:
-        sign_bucket = self._sign_bucket()
-        if sign_bucket is None:
+        """获取预签名下载 URL。
+
+        bucket未配置时降级为公开URL；本地产物URL直接返回。
+        """
+        if self.bucket is None:
             if self._is_local_generated_url(storage_key_or_url):
                 return storage_key_or_url
-            logger.warning("OSS bucket not configured, returning raw URL: %s", storage_key_or_url[:200])
+            logger.warning(
+                "get_download_url: OSS bucket not configured, returning raw URL. key=%s",
+                storage_key_or_url[:200],
+            )
             return self.get_url(self.normalize_storage_key(storage_key_or_url))
 
         storage_key = self.normalize_storage_key(storage_key_or_url)
         try:
-            signed = sign_bucket.sign_url("GET", storage_key, expires_seconds)
+            signed = self.bucket.sign_url("GET", storage_key, expires_seconds)
             logger.info(
-                "signed URL generated for key=%s prefix=%s",
+                "get_download_url: signed URL generated. key=%s url_prefix=%s",
                 storage_key[:80],
                 signed[:60],
             )
             return signed
         except Exception:
-            logger.exception("get_download_url: sign_url 失败，返回 raw URL: %s", storage_key[:200])
+            logger.exception(
+                "get_download_url: sign_url failed, falling back to raw URL. key=%s",
+                storage_key[:200],
+            )
             return self.get_url(storage_key)
 
+    # ── 浏览器直传 POST ────────────────────────────────────────────────
+
     def get_upload_url(
         self,
         storage_key_or_url: str,
         expires_seconds: int = 3600,
         content_type: str = "video/mp4",
     ) -> str:
-        sign_bucket = self._sign_bucket()
-        if sign_bucket is None:
+        """获取预签名 PUT 上传 URL（供外部 Worker 上传结果文件）。
+
+        bucket未配置时降级为 public_url（本地/开发环境）；
+        本地产物 key 原样返回。
+        """
+        if self.bucket is None:
             if self._is_local_generated_url(storage_key_or_url):
                 return storage_key_or_url
-            logger.warning("get_upload_url: OSS 未配置，返回 raw URL: %s", storage_key_or_url[:200])
+            logger.warning(
+                "get_upload_url: OSS bucket not configured, returning raw URL. key=%s",
+                storage_key_or_url[:200],
+            )
             return self.get_url(self.normalize_storage_key(storage_key_or_url))
 
         storage_key = self.normalize_storage_key(storage_key_or_url)
         try:
+            # oss2 sign_url 支持 'PUT'，需指定 headers 才能限定 Content-Type
             headers = {"Content-Type": content_type} if content_type else None
-            signed = sign_bucket.sign_url("PUT", storage_key, expires_seconds, headers=headers)
+            signed = self.bucket.sign_url("PUT", storage_key, expires_seconds, headers=headers)
             logger.info(
-                "get_upload_url: 公网签名PUT URL已生成 key=%s prefix=%s",
+                "get_upload_url: signed PUT URL generated. key=%s url_prefix=%s",
                 storage_key[:80],
                 signed[:60],
             )
             return signed
         except Exception:
-            logger.exception("get_upload_url: sign_url 失败，返回 raw URL: %s", storage_key[:200])
+            logger.exception(
+                "get_upload_url: sign_url failed, falling back to raw URL. key=%s",
+                storage_key[:200],
+            )
             return self.get_url(storage_key)
 
     def create_direct_upload_post(
@@ -379,6 +384,7 @@ class SharedStorageService(StoragePort):
         max_size_bytes: int,
         expires_seconds: int,
     ) -> dict[str, object]:
+        """创建浏览器直传 POST 表单。"""
         if not self.access_key_id or not self.access_key_secret:
             raise RuntimeError("OSS storage is not configured")
         normalized_key = self.normalize_storage_key(storage_key)
@@ -425,17 +431,19 @@ class SharedStorageService(StoragePort):
             },
         }
 
-    # ── 文件操作（internal endpoint）──────────────────────────────────
+    # ── 文件操作 ───────────────────────────────────────────────────────
 
     def delete_file(self, storage_key: str) -> None:
+        """删除文件（不抛异常）。"""
         if self.bucket is None:
             return
         try:
             self.bucket.delete_object(storage_key)
         except Exception as error:
-            logger.warning("OSS delete 失败", extra={"storage_key": storage_key, "error": str(error)})
+            logger.warning("Failed to delete file from OSS", extra={"storage_key": storage_key, "error": str(error)})
 
     def file_exists(self, storage_key: str) -> bool:
+        """检查文件是否存在。"""
         if self.bucket is None:
             return False
         return self.bucket.object_exists(storage_key)
@@ -443,6 +451,17 @@ class SharedStorageService(StoragePort):
     # ── Asset 路径解析（Worker 用）────────────────────────────────────
 
     def resolve_asset_path(self, asset_id: str, work_dir: str | Path) -> Optional[Path]:
+        """从 asset_id 解析到本地文件路径。
+
+        策略（按优先级）：
+        1. 本地绝对路径（在允许目录内）→ 直接返回
+        2. work_dir 缓存命中 → 返回缓存路径
+        3. 从OSS下载到缓存 → 返回下载路径
+        4. 全部失败 → None
+
+        从 oss_helpers.resolve_asset_path 合并而来。
+        """
+        # 延迟导入，避免循环依赖
         from video_processing.path_security import (  # type: ignore[import-not-found]
             PathSecurityError,
             get_allowed_local_dirs,
@@ -452,35 +471,47 @@ class SharedStorageService(StoragePort):
 
         if not asset_id or not isinstance(asset_id, str):
             return None
+
         work_dir = Path(work_dir)
         os.makedirs(work_dir, exist_ok=True)
+
+        # 空字节检测
         if "\x00" in asset_id:
-            logger.warning("asset_id 含空字节，拒绝: %s", asset_id[:50])
+            logger.warning("asset_id 包含空字节，拒绝: %s", asset_id[:50])
             return None
+
+        # 1. 本地绝对路径 — 必须在允许的目录内
         if asset_id.startswith("/") and os.path.exists(asset_id):
             try:
                 resolved = Path(asset_id).resolve()
                 if is_in_allowed_dirs(resolved, get_allowed_local_dirs()):
                     return resolved
-                logger.warning(
-                    "本地素材路径不在允许目录: %s allowed=%s",
-                    asset_id[:80],
-                    get_allowed_local_dirs(),
-                )
-                return None
+                else:
+                    logger.warning(
+                        "本地素材路径不在允许目录内，拒绝: %s (allowed=%s)",
+                        asset_id[:80],
+                        get_allowed_local_dirs(),
+                    )
+                    return None
             except (OSError, PathSecurityError):
                 return None
+
+        # 2. 缓存命中（SHA256 hash 防路径遍历）
         cache_hash = hashlib.sha256(asset_id.encode()).hexdigest()[:16]
         safe_name = sanitize_filename(cache_hash)
         cached_path = work_dir / f"{safe_name}.mp4"
         if cached_path.exists() and cached_path.stat().st_size > 0:
             return cached_path
+
+        # 3. 从 OSS 下载（先标准化 key，防路径遍历注入）
         safe_key = self.normalize_storage_key(asset_id)
         if ".." in safe_key or safe_key.startswith("/"):
-            logger.warning("asset_id 含路径遍历: %s", asset_id[:80])
+            logger.warning("asset_id 包含路径遍历模式，拒绝下载: %s", asset_id[:80])
             return None
+
         if self.download_asset(safe_key, cached_path):
             return cached_path
+
         return None
 
     def resolve_asset_ids_to_paths(
@@ -488,20 +519,22 @@ class SharedStorageService(StoragePort):
         asset_ids: list[str],
         work_dir: str | Path,
     ) -> dict[str, Path]:
+        """批量解析 asset_id → 本地路径。"""
         result: dict[str, Path] = {}
         for aid in asset_ids:
-            p = self.resolve_asset_path(aid, work_dir)
-            if p:
-                result[aid] = p
+            local_path = self.resolve_asset_path(aid, work_dir)
+            if local_path:
+                result[aid] = local_path
         return result
 
 
-# ── 单例 ────────────────────────────────────────────────────────────────
+# ── 单例管理 ────────────────────────────────────────────────────────────
 
 _storage_service: Optional[SharedStorageService] = None
 
 
 def get_shared_storage_service() -> SharedStorageService:
+    """获取统一存储服务单例。"""
     global _storage_service
     if _storage_service is None:
         _storage_service = SharedStorageService()
@@ -509,6 +542,7 @@ def get_shared_storage_service() -> SharedStorageService:
     return _storage_service
 
 
+# 向后兼容别名
 def get_storage_service() -> SharedStorageService:
-    """向后兼容别名。"""
+    """向后兼容：返回统一存储服务。"""
     return get_shared_storage_service()

scripts/ci_staging_deploy.sh

@@ -1,7 +1,6 @@
 #!/bin/sh
 # ===========================================
 # Staging 部署脚本（SSH 模式，并行优化版）
-# worker 已收敛到 infra/docker/compose.yml 单一事实来源；api/web 暂保留 docker run。
 # ===========================================
 set -eu
 
@@ -49,11 +48,6 @@ GENERATED_DIR="${GENERATED_DIR:-/var/lib/xiaoxia-saas-staging/generated}"
 LEGACY_ASSETS_DIR="${LEGACY_ASSETS_DIR:-/var/lib/xiaoxia-saas-staging/legacy-assets}"
 NGINX_CONF_FILE="${NGINX_CONF_FILE:-/var/lib/xiaoxia-saas-staging/nginx-staging.conf}"
 COOKIES_FILE="${COOKIES_FILE:-/var/lib/xiaoxia-saas-staging/configs/douyin_cookies.txt}"
-INFRA_DOCKER_DIR="${INFRA_DOCKER_DIR:-/var/lib/xiaoxia-saas-staging/infra/docker}"
-COMPOSE_PROJECT="${COMPOSE_PROJECT:-xiaoxia-staging}"
-COMPOSE_ENV_VALUE="${COMPOSE_ENV_VALUE:-staging}"
-# COMPOSE_SYNC: CI workflow 已通过 scp 把 infra/docker/compose.yml 上传到服务器时设为 0 跳过同步
-COMPOSE_SYNC="${COMPOSE_SYNC:-1}"
 
 SKIP_MIGRATION="${SKIP_MIGRATION:-false}"
 SKIP_ROLLBACK="${SKIP_ROLLBACK:-false}"
@@ -63,6 +57,8 @@ if [ -z "$IMAGE_TAG" ]; then
   exit 1
 fi
 
+# .env 文件由 CI 从模板 + Secrets 渲染后通过 SCP 上传到服务器
+# 如果文件不存在，说明 CI 渲染步骤失败或未执行
 if [ ! -f "$ENV_FILE" ]; then
   echo "ERROR: $ENV_FILE 不存在。CI 应先在 render_env 步骤渲染并上传此文件"
   exit 1
@@ -71,7 +67,7 @@ echo "✅ .env file found: $ENV_FILE ($(wc -l < "$ENV_FILE") lines)"
 mkdir -p "$GENERATED_DIR"
 mkdir -p "$LEGACY_ASSETS_DIR"
 mkdir -p "$(dirname "$COOKIES_FILE")"
-mkdir -p "$INFRA_DOCKER_DIR"
+# 抖音 cookies 文件：CI workflow 已通过 scp 上传；如果不存在（非 CI 环境）则创建占位
 if [ ! -f "$COOKIES_FILE" ] || [ "$(wc -c < "$COOKIES_FILE" 2>/dev/null || echo 0)" -lt 200 ]; then
   printf '# Netscape HTTP Cookie File\n# 抖音 cookies 占位（CI 应通过 scp 上传真实 cookies）\n' > "$COOKIES_FILE"
   echo "WARNING: Douyin cookies not found or too small at $COOKIES_FILE (extraction will 503)"
@@ -80,6 +76,7 @@ else
 fi
 
 # ── 写入 Staging Nginx 配置 ──
+# 运行时覆盖 nginx 配置，确保 upstream 指向正确的 staging 网络
 echo "Writing staging nginx config..."
 cat > "$NGINX_CONF_FILE" << 'NGINX_EOF'
 server {
@@ -125,26 +122,8 @@ server {
 NGINX_EOF
 echo "✅ Nginx config written: $NGINX_CONF_FILE"
 
-# ── 确认 infra/docker/compose.yml 存在 ──
-# CI workflow 在执行本脚本前已通过 scp 把 infra/docker/compose.yml 上传到 $INFRA_DOCKER_DIR
-# （workflow 里做：scp infra/docker/compose.yml <host>:$INFRA_DOCKER_DIR/compose.yml）。
-# 这里只做存在性检查 + nginx 软链；不再 curl 私有仓库（SSH 环境无 Gitea token）。
-COMPOSE_FILE_PATH="$INFRA_DOCKER_DIR/compose.yml"
-if [ ! -f "$COMPOSE_FILE_PATH" ]; then
-  echo "ERROR: $COMPOSE_FILE_PATH 不存在。CI workflow 应先 scp infra/docker/compose.yml 到服务器"
-  exit 1
-fi
-echo "✅ compose.yml ready: $COMPOSE_FILE_PATH ($(wc -l < "$COMPOSE_FILE_PATH") lines)"
-ln -sf "$NGINX_CONF_FILE" "$INFRA_DOCKER_DIR/nginx-${COMPOSE_ENV_VALUE}.conf" 2>/dev/null || true
-
-# 封装 docker compose 调用：统一 --env-file（compose 默认只读取 project 目录下的 .env，
-# 我们的 .env 在 $INFRA_DOCKER_DIR/../../.env，必须显式传入才能读到 GENERATED_FILES_HOST_DIR 等变量）
-compose() {
-  (cd "$INFRA_DOCKER_DIR" && docker compose --env-file "$ENV_FILE" -p "$COMPOSE_PROJECT" "$@")
-}
-
 echo "==========================================="
-echo "  Staging 部署 - $IMAGE_TAG"
+echo "  Staging 部署 - $IMAGE_TAG (并行优化版)"
 echo "==========================================="
 
 echo "Recording current image versions for rollback..."
@@ -165,7 +144,6 @@ for c in xiaoxia-api-staging xiaoxia-worker-staging xiaoxia-web-staging; do
   fi
 done
 
-# ── 回滚函数 ──
 rollback() {
   echo ""
   echo "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!"
@@ -179,37 +157,9 @@ rollback() {
   fi
 
   echo "Stopping new containers..."
-  docker rm -f xiaoxia-api-staging xiaoxia-web-staging 2>/dev/null || true
-  if [ -n "$PREV_WORKER_IMAGE" ]; then
-    echo "Rolling back Worker to: $PREV_WORKER_IMAGE (via compose)"
-    compose up -d --no-deps worker 2>&1 || echo "WARN: compose rollback failed, fallback to docker run"
-    # 镜像通过 env 注入：compose 默认读 WORKER_IMAGE（未设则用 :dev），这里用临时 env 覆盖
-    if ! docker inspect xiaoxia-worker-staging >/dev/null 2>&1; then
-      echo "Fallback: docker run previous worker image"
-      docker run -d \
-        --name xiaoxia-worker-staging \
-        --env-file "$ENV_FILE" \
-        --network "xiaoxia-net-${COMPOSE_ENV_VALUE}" \
-        -e APP_ENV="$COMPOSE_ENV_VALUE" \
-        -e APP_VERSION="$(echo "$PREV_WORKER_IMAGE" | grep -oE '[^:]+$')" \
-        -e WORKER_MAX_TASKS_PER_CHILD=100 \
-        -e GENERATION_CONCURRENCY="${GENERATION_CONCURRENCY:-2}" \
-        -e TRANSCODE_CONCURRENCY="${TRANSCODE_CONCURRENCY:-2}" \
-        -e BEAT_ENABLED=1 \
-        -e GENERATED_FILES_DIR=/app/generated \
-        -e GENERATED_FILES_URL_PREFIX=/generated-files \
-        -e PUBLIC_API_BASE_URL=https://staging-api.xiaoxiajianji.com \
-        -v "$GENERATED_DIR:/app/generated" \
-        --restart unless-stopped \
-        --health-cmd "grep -q 'celery.*worker' /proc/[0-9]*/cmdline 2>/dev/null || exit 1" \
-        --health-interval 30s \
-        --health-timeout 10s \
-        --health-retries 3 \
-        --health-start-period 40s \
-        --log-driver json-file --log-opt max-size=50m --log-opt max-file=3 \
-        "$PREV_WORKER_IMAGE" || true
-    fi
-  fi
+  docker rm -f xiaoxia-api-staging 2>/dev/null || true
+  docker rm -f xiaoxia-worker-staging 2>/dev/null || true
+  docker rm -f xiaoxia-web-staging 2>/dev/null || true
 
   LOG_OPTS="--log-driver json-file --log-opt max-size=50m --log-opt max-file=3"
 
@@ -218,10 +168,10 @@ rollback() {
     docker run -d \
       --name xiaoxia-api-staging \
       --env-file "$ENV_FILE" \
-      --network "xiaoxia-net-${COMPOSE_ENV_VALUE}" \
+      --network xiaoxia-net-staging \
       -p 127.0.0.1:8000:8000 \
-      -e APP_ENV="$COMPOSE_ENV_VALUE" \
-      -e APP_VERSION="$(echo "$PREV_API_IMAGE" | grep -oE '[^:]+$')" \
+      -e APP_ENV=staging \
+      -e APP_VERSION="$(echo $PREV_API_IMAGE | grep -oE '[^:]+$')" \
       -e GENERATED_FILES_DIR=/app/generated \
       -e GENERATED_FILES_URL_PREFIX=/generated-files \
       -e PUBLIC_API_BASE_URL=https://staging-api.xiaoxiajianji.com \
@@ -233,7 +183,31 @@ rollback() {
       --health-retries 3 \
       --health-start-period 40s \
       $LOG_OPTS \
-      "$PREV_API_IMAGE" || true
+      "$PREV_API_IMAGE" &
+  fi
+
+  if [ -n "$PREV_WORKER_IMAGE" ]; then
+    echo "Rolling back Worker to: $PREV_WORKER_IMAGE"
+    docker run -d \
+      --name xiaoxia-worker-staging \
+      --env-file "$ENV_FILE" \
+      --network xiaoxia-net-staging \
+      -e APP_ENV=staging \
+      -e APP_VERSION="$(echo $PREV_WORKER_IMAGE | grep -oE '[^:]+$')" \
+      -e WORKER_CONCURRENCY=1 \
+      -e WORKER_MAX_TASKS_PER_CHILD=100 \
+      -e GENERATED_FILES_DIR=/app/generated \
+      -e GENERATED_FILES_URL_PREFIX=/generated-files \
+      -e PUBLIC_API_BASE_URL=https://staging-api.xiaoxiajianji.com \
+      -v "$GENERATED_DIR:/app/generated" \
+      --restart unless-stopped \
+      --health-cmd "grep -lq celery /proc/[0-9]*/cmdline 2>/dev/null || exit 1" \
+      --health-interval 30s \
+      --health-timeout 10s \
+      --health-retries 3 \
+      --health-start-period 30s \
+      $LOG_OPTS \
+      "$PREV_WORKER_IMAGE" &
   fi
 
   if [ -n "$PREV_WEB_IMAGE" ]; then
@@ -244,7 +218,7 @@ rollback() {
     fi
     docker run -d \
       --name xiaoxia-web-staging \
-      --network "xiaoxia-net-${COMPOSE_ENV_VALUE}" \
+      --network xiaoxia-net-staging \
       -p 127.0.0.1:3001:80 \
       --restart unless-stopped \
       $LEGACY_VOLUME \
@@ -254,25 +228,27 @@ rollback() {
       --health-timeout 5s \
       --health-retries 3 \
       $LOG_OPTS \
-      "$PREV_WEB_IMAGE" || true
+      "$PREV_WEB_IMAGE" &
   fi
 
-  sleep 3
+  wait
 
-  echo "Waiting for rolled-back API to become healthy..."
-  i=0
-  while [ "$i" -lt 40 ]; do
-    if curl -sf --max-time 5 http://127.0.0.1:8000/health >/dev/null 2>&1; then
-      echo "Rolled-back API is healthy!"
-      break
+  if [ -n "$PREV_API_IMAGE" ]; then
+    echo "Waiting for rolled-back API to become healthy..."
+    i=0
+    while [ "$i" -lt 40 ]; do
+      if curl -sf --max-time 5 http://127.0.0.1:8000/health >/dev/null 2>&1; then
+        echo "Rolled-back API is healthy!"
+        break
+      fi
+      i=$((i + 1))
+      echo "  Waiting... ($i/40)"
+      sleep 3
+    done
+    if [ "$i" -ge 40 ]; then
+      echo "WARN: Rolled-back API did not become healthy within 120s"
+      docker logs --tail 30 xiaoxia-api-staging
     fi
-    i=$((i + 1))
-    echo "  Waiting... ($i/40)"
-    sleep 3
-  done
-  if [ "$i" -ge 40 ]; then
-    echo "WARN: Rolled-back API did not become healthy within 120s"
-    docker logs --tail 30 xiaoxia-api-staging 2>/dev/null || true
   fi
 
   echo ""
@@ -284,7 +260,7 @@ rollback() {
   echo "Previous Web:    ${PREV_WEB_IMAGE:-none}"
   echo ""
   echo "部署失败，已自动回滚到上一版本"
-  docker ps --format "table {{.Names}}\t{{.Status}}\t{{.Image}}" | grep staging || true
+  docker ps --format "table {{.Names}}\t{{.Status}}\t{{.Image}}" | grep staging
   exit 1
 }
 
@@ -296,6 +272,10 @@ if [ -n "$REGISTRY_TOKEN" ]; then
   retry_docker_login
 fi
 
+# ---- 并行 Pull 三个镜像 ----
+# 注意：这里必须使用 IMAGE_TAG（commit SHA）做确定性部署，不要改成 :dev。
+# :dev 是 floating tag，可能被并发构建覆盖，导致部署版本不可重现、回滚混乱。
+# Watchtower 可监听 :dev 做非关键路径的自动同步；正式部署/回滚一律锚定 SHA。
 REGISTRY_API="${REGISTRY}/xiaoxia-saas-api:${IMAGE_TAG}"
 REGISTRY_WORKER="${REGISTRY}/xiaoxia-saas-worker:${IMAGE_TAG}"
 REGISTRY_WEB="${REGISTRY}/xiaoxia-saas-web:${IMAGE_TAG}"
@@ -324,6 +304,7 @@ for svc in api worker web; do
   elif grep -qE "Digest:|Status: Downloaded" "$PULL_LOG_DIR/$svc.log" 2>/dev/null; then
     echo "  OK $svc"
   else
+    # 检查docker pull返回值不直接，用镜像是否存在来判断
     img_var="REGISTRY_$(echo $svc | tr '[:lower:]' '[:upper:]')"
     img_val=$(eval echo "\$$img_var")
     if docker image inspect "$img_val" >/dev/null 2>&1; then
@@ -346,7 +327,7 @@ fi
 
 echo "All images pulled."
 
-# ====== 镜像内容校验 ======
+# ====== 镜像内容校验（CI 加固 - 防止静默部署损坏/过期镜像） ======
 echo ""
 echo "=========================================="
 echo "  镜像内容校验"
@@ -355,6 +336,7 @@ echo "=========================================="
 VERIFY_FAILED=0
 DEPLOY_MANIFEST="${GENERATED_DIR}/deploy-manifest.json"
 
+# 读取上次部署的 manifest（用于对比）
 PREV_MANIFEST=""
 if [ -f "$DEPLOY_MANIFEST" ]; then
   PREV_MANIFEST=$(cat "$DEPLOY_MANIFEST")
@@ -366,12 +348,14 @@ for svc in api worker web; do
   img_var="REGISTRY_$(echo $svc | tr '[:lower:]' '[:upper:]')"
   img_val=$(eval echo "\$$img_var")
 
+  # 1. 检查镜像是否存在
   if ! docker image inspect "$img_val" >/dev/null 2>&1; then
     echo "  ❌ $svc: 镜像不存在 ($img_val)"
     VERIFY_FAILED=$((VERIFY_FAILED + 1))
     continue
   fi
 
+  # 2. 检查 layers 有效性
   LAYER_COUNT=$(docker inspect --format='{{len .RootFS.Layers}}' "$img_val" 2>/dev/null || echo "0")
   if [ "$LAYER_COUNT" -eq 0 ]; then
     echo "  ❌ $svc: 镜像无有效 layers ($img_val)"
@@ -379,12 +363,14 @@ for svc in api worker web; do
     continue
   fi
 
+  # 3. 获取 digest 和创建时间
   IMG_ID=$(docker inspect --format='{{.Id}}' "$img_val")
   IMG_CREATED=$(docker inspect --format='{{.Created}}' "$img_val")
   IMG_SIZE=$(docker inspect --format='{{.Size}}' "$img_val")
   echo "  ✅ $svc: ${LAYER_COUNT} layers, size=${IMG_SIZE}, created=${IMG_CREATED}"
   echo "     id: $IMG_ID"
 
+  # 4. 对比上次部署
   CHANGED="unchanged"
   if [ -n "$PREV_MANIFEST" ]; then
     PREV_ID=$(echo "$PREV_MANIFEST" | grep "\"${svc}_id\"" | sed 's/.*: *"\(.*\)".*/\1/' 2>/dev/null || echo "")
@@ -412,6 +398,7 @@ if [ "$VERIFY_FAILED" -gt 0 ]; then
   exit 1
 fi
 
+# 写入新 manifest
 cat > "$DEPLOY_MANIFEST" <<MANIFEST_EOF
 {
   "deployed_at": "$(date -u +%Y-%m-%dT%H:%M:%SZ)",
@@ -457,14 +444,14 @@ for c in xiaoxia-postgres-staging xiaoxia-redis-staging; do
   fi
 done
 
-docker network create "xiaoxia-net-${COMPOSE_ENV_VALUE}" 2>/dev/null || true
+docker network create xiaoxia-net-staging 2>/dev/null || true
 
 if [ "$SKIP_MIGRATION" != "true" ]; then
   echo "Running database migrations..."
   docker run --rm \
     --env-file "$ENV_FILE" \
-    --network "xiaoxia-net-${COMPOSE_ENV_VALUE}" \
-    -e APP_ENV="$COMPOSE_ENV_VALUE" \
+    --network xiaoxia-net-staging \
+    -e APP_ENV=staging \
     "$REGISTRY_API" sh -c "cd /app && alembic upgrade head" || {
     echo "ERROR: Database migration failed"
     exit 1
@@ -475,28 +462,29 @@ else
 fi
 
 echo "Stopping old containers..."
-docker stop -t 10 xiaoxia-web-staging 2>/dev/null || true
-docker stop -t 30 xiaoxia-api-staging 2>/dev/null || true
+# 优雅关闭：先 stop（发 SIGTERM，等待），再 rm
+# Worker 需要更长时间（视频任务最长可能5分钟）
 docker stop -t 120 xiaoxia-worker-staging 2>/dev/null || true
-docker rm -f xiaoxia-api-staging xiaoxia-web-staging 2>/dev/null || true
-docker rm -f xiaoxia-worker-staging 2>/dev/null || true
+docker stop -t 30 xiaoxia-api-staging 2>/dev/null || true
+docker stop -t 10 xiaoxia-web-staging 2>/dev/null || true
+docker rm xiaoxia-worker-staging xiaoxia-api-staging xiaoxia-web-staging 2>/dev/null || true
 
 LOG_OPTS="--log-driver json-file --log-opt max-size=50m --log-opt max-file=3"
 
-echo "Starting all containers..."
+# ---- 并行启动三个容器 ----
+echo "Starting all containers (parallel)..."
 
 LEGACY_VOLUME=""
 if [ -d "$LEGACY_ASSETS_DIR" ] && [ "$(ls -A "$LEGACY_ASSETS_DIR" 2>/dev/null)" ]; then
   LEGACY_VOLUME="-v ${LEGACY_ASSETS_DIR}:/usr/share/nginx/html/assets-legacy/assets:ro"
 fi
 
-# ── API: 暂保留 docker run（TODO: 后续收敛到 compose）──
 docker run -d \
   --name xiaoxia-api-staging \
   --env-file "$ENV_FILE" \
-  --network "xiaoxia-net-${COMPOSE_ENV_VALUE}" \
+  --network xiaoxia-net-staging \
   -p 127.0.0.1:8000:8000 \
-  -e APP_ENV="$COMPOSE_ENV_VALUE" \
+  -e APP_ENV=staging \
   -e APP_VERSION="$IMAGE_TAG" \
   -e GENERATED_FILES_DIR=/app/generated \
   -e GENERATED_FILES_URL_PREFIX=/generated-files \
@@ -514,18 +502,31 @@ docker run -d \
   "$REGISTRY_API" &
 PID_API_START=$!
 
-# ── Worker: 通过 compose 启动（单一事实来源）──
-# compose.yml 定义：三进程(beat+generation+transcode)、独立并发、BEAT_ENABLED、
-# healthcheck 匹配 'celery.*worker'（不把 beat 算活）、资源限制 4C/8G。
-# WORKER_IMAGE 通过环境变量覆盖镜像 tag（compose.yml 默认 :dev）。
-echo "Starting worker via docker compose (from $INFRA_DOCKER_DIR)..."
-WORKER_IMAGE="$REGISTRY_WORKER" APP_VERSION="$IMAGE_TAG" compose up -d --no-deps worker &
+docker run -d \
+  --name xiaoxia-worker-staging \
+  --env-file "$ENV_FILE" \
+  --network xiaoxia-net-staging \
+  -e APP_ENV=staging \
+  -e APP_VERSION="$IMAGE_TAG" \
+  -e WORKER_CONCURRENCY=1 \
+  -e WORKER_MAX_TASKS_PER_CHILD=100 \
+  -e GENERATED_FILES_DIR=/app/generated \
+  -e GENERATED_FILES_URL_PREFIX=/generated-files \
+  -e PUBLIC_API_BASE_URL=https://staging-api.xiaoxiajianji.com \
+  -v "$GENERATED_DIR:/app/generated" \
+  --restart unless-stopped \
+  --health-cmd "grep -lq celery /proc/[0-9]*/cmdline 2>/dev/null || exit 1" \
+  --health-interval 30s \
+  --health-timeout 10s \
+  --health-retries 3 \
+  --health-start-period 30s \
+  $LOG_OPTS \
+  "$REGISTRY_WORKER" &
 PID_WORKER_START=$!
 
-# ── Web: 暂保留 docker run（TODO: 后续收敛到 compose）──
 docker run -d \
   --name xiaoxia-web-staging \
-  --network "xiaoxia-net-${COMPOSE_ENV_VALUE}" \
+  --network xiaoxia-net-staging \
   -p 127.0.0.1:3001:80 \
   --restart unless-stopped \
   $LEGACY_VOLUME \
@@ -562,8 +563,9 @@ if [ "$START_FAILED" -gt 0 ]; then
   rollback
 fi
 
+# ---- 并行等待 API 和 Web 健康 ----
 echo ""
-echo "Waiting for all services health (parallel)..."
+echo "Waiting for API + Web health (parallel)..."
 
 HEALTH_LOG_DIR="/tmp/staging-health-$$"
 mkdir -p "$HEALTH_LOG_DIR"
@@ -598,60 +600,36 @@ PID_API_HEALTH=$!
 ) > "$HEALTH_LOG_DIR/web.log" 2>&1 &
 PID_WEB_HEALTH=$!
 
-(
-  i=0
-  while [ "$i" -lt 20 ]; do
-    hc=$(docker inspect -f '{{if .State.Health}}{{.State.Health.Status}}{{else}}{{.State.Status}}{{end}}' xiaoxia-worker-staging 2>/dev/null || echo "missing")
-    if [ "$hc" = "healthy" ]; then
-      echo "Worker healthy after $((i * 3))s"
-      exit 0
-    fi
-    if [ "$hc" = "unhealthy" ]; then
-      echo "Worker UNHEALTHY after $((i * 3))s"
-      docker logs --tail 30 xiaoxia-worker-staging 2>/dev/null || true
-      exit 1
-    fi
-    i=$((i + 1))
-    sleep 3
-  done
-  echo "Worker health unknown after 60s (last: $hc)"
-  exit 1
-) > "$HEALTH_LOG_DIR/worker.log" 2>&1 &
-PID_WORKER_HEALTH=$!
-
 set +e
 wait $PID_API_HEALTH
 API_EXIT=$?
 wait $PID_WEB_HEALTH
 WEB_EXIT=$?
-wait $PID_WORKER_HEALTH
-WORKER_EXIT=$?
 set -e
 
 echo ""
 echo "健康检查结果:"
+API_OK=0
+WEB_OK=0
 if [ "$API_EXIT" -eq 0 ]; then
-  echo "  OK API:    $(cat "$HEALTH_LOG_DIR/api.log")"
+  echo "  OK API: $(cat "$HEALTH_LOG_DIR/api.log")"
+  API_OK=1
 else
-  echo "  FAIL API:  120s未就绪"
-  docker logs --tail 50 xiaoxia-api-staging 2>/dev/null || true
+  echo "  FAIL API: 120s未就绪"
+  docker logs --tail 50 xiaoxia-api-staging
 fi
+
 if [ "$WEB_EXIT" -eq 0 ]; then
-  echo "  OK Web:    $(cat "$HEALTH_LOG_DIR/web.log")"
+  echo "  OK Web: $(cat "$HEALTH_LOG_DIR/web.log")"
+  WEB_OK=1
 else
-  echo "  FAIL Web:  30s未就绪"
-  docker logs --tail 30 xiaoxia-web-staging 2>/dev/null || true
-fi
-if [ "$WORKER_EXIT" -eq 0 ]; then
-  echo "  OK Worker: $(cat "$HEALTH_LOG_DIR/worker.log")"
-else
-  echo "  FAIL Worker: $(cat "$HEALTH_LOG_DIR/worker.log")"
-  docker logs --tail 50 xiaoxia-worker-staging 2>/dev/null || true
+  echo "  FAIL Web: 30s未就绪"
+  docker logs --tail 30 xiaoxia-web-staging
 fi
 
 rm -rf "$HEALTH_LOG_DIR"
 
-if [ "$API_EXIT" -ne 0 ] || [ "$WEB_EXIT" -ne 0 ] || [ "$WORKER_EXIT" -ne 0 ]; then
+if [ "$API_OK" -eq 0 ] || [ "$WEB_OK" -eq 0 ]; then
   echo ""
   echo "ERROR: 健康检查失败"
   rollback
@@ -662,9 +640,8 @@ docker image prune -af --filter "until=168h" 2>/dev/null || true
 docker builder prune -af --filter "until=168h" 2>/dev/null || true
 
 echo ""
-echo "=== Staging deployment complete ==="
+echo "=== Staging deployment complete (并行优化版) ==="
 echo "API:      http://127.0.0.1:8000"
 echo "Web:      http://127.0.0.1:3001"
-echo "Worker:   managed by docker compose (project=$COMPOSE_PROJECT)"
 echo "Version:  $IMAGE_TAG"
 docker ps --format "table {{.Names}}\t{{.Status}}\t{{.Image}}" | grep staging