style: auto-format with black + isort + ruff + prettier [skip ci-format-check]

根因定位：staging直连测试发现 doubao-seed-2-1-lite-260915 是思考模型，
单次调用产生~520 reasoning_tokens，耗时10-12s，导致fast路径必超时。

修复：
1. vlm_fast_json.py 直接用 httpx 发最小 payload（不走 ai_client 包装），
   显式设置 thinking={"type":"disabled"} + reasoning_effort="low" 关闭思考链，
   期望响应降至 <3s；若API不支持thinking参数返回400，自动降级重试一次。
2. 超时恢复合理值：lite JSON 8s、OCR 6s、fast总超时8s（关闭thinking后预计<3s，余量充足）。
3. vlm_fallback.py 保持单次pro调用，pro走原ai_client路径（兜底场景对延迟不敏感，45s足够）。

预期：3图<10s/8图<15s目标可达。

apps/web/src/pages/viral-video/ViralVideoPage.tsx

@@ -30,7 +30,6 @@ import { uploadAssetDirect, getAssetLibraries, getAssetsByKind, type AssetItem }
 import { fetchPresetVoices } from "@/api/voices"
 import { getVoiceClones, getVoiceClonePreview } from "@/api/voice-clone"
 import type { VoiceClone } from "@/api/voice-clone"
-import DurationWheelPicker from "@/components/common/DurationWheelPicker"
 import {
   FUSION_LEVELS,
   STYLE_STRENGTHS,
@@ -224,6 +223,7 @@ const RATIOS = [
   { v: "16:9", label: "16:9 横屏（B站/YouTube）" },
   { v: "1:1", label: "1:1 方形（小红书）" },
 ]
+const DURATIONS = Array.from({ length: 16 }, (_, i) => 15 + i)
 /** 兜底模型列表（接口未返回时使用，字段与 ViralVideoModel 对齐；后端返回后自动覆盖） */
 const FALLBACK_VIDEO_MODELS: ViralVideoModel[] = [
   {
@@ -2433,12 +2433,12 @@ const ViralVideoPage: React.FC = () => {
                 </div>
                 <div className="vv-form-row">
                   <label className="vv-label">文案视频时长</label>
-                  <DurationWheelPicker
+                  <Select
+                    className="vv-select"
+                    style={{ width: "100%" }}
                     value={task.duration}
-                    min={15}
-                    max={30}
-                    step={1}
                     onChange={(v) => setTask({ duration: v })}
+                    options={DURATIONS.map((n) => ({ value: n, label: `${n}秒` }))}
                   />
                 </div>
               </div>

apps/worker/worker_app/tasks/viral_video.py

@@ -24,7 +24,6 @@ from __future__ import annotations
 import json
 import logging
 import os
-import re
 import tempfile
 import threading
 import time
@@ -347,6 +346,7 @@ def _normalize_image_url(raw: str, idx: int) -> str:
     storage_key = url.lstrip("/")
     try:
         from packages.shared.storage import get_storage_service
+
         url = get_storage_service().get_url(storage_key)
     except Exception as _e:
         raise ValueError(f"图片 #{idx} storage_key={storage_key!r} 转公网URL失败: {_e}") from _e
@@ -386,6 +386,7 @@ def _step_image_analysis(job: ViralVideoJob) -> dict:
     # 整个阶段关闭底层 httpx 重试，避免线程里出现不可控等待
     try:
         from packages.shared.ai_client import get_doubao_client as _gdc
+
         _cli = _gdc()
         _orig_retries = _cli.max_retries
         _cli.max_retries = 0

apps/worker/worker_app/tasks/vision/__init__.py

@@ -1,3 +1,4 @@
 # -*- coding: utf-8 -*-
 """V2 图片分析：火山OCR专用API + doubao-lite强约束JSON并行，单次pro VLM兜底。"""
+
 from .fast_path import analyze_image_v2, analyze_images_v2  # noqa: F401

apps/worker/worker_app/tasks/vision/fast_path.py

@@ -22,15 +22,22 @@ logger = logging.getLogger(__name__)
 # 可通过环境变量调参（有默认值，无需配置即可跑）
 _IMG_WORKERS = int(os.environ.get("VISION_V2_IMG_WORKERS", "8"))
 _FAST_TIMEOUT = float(os.environ.get("VISION_V2_FAST_TIMEOUT", "8"))
-_FAST_JSON_TIMEOUT = float(os.environ.get("VISION_V2_FAST_JSON_TIMEOUT", "6"))
+_FAST_JSON_TIMEOUT = float(os.environ.get("VISION_V2_FAST_JSON_TIMEOUT", "8"))
 _OCR_TIMEOUT = float(os.environ.get("VISION_V2_OCR_TIMEOUT", "6"))
 _PRO_TIMEOUT = float(os.environ.get("VISION_V2_PRO_TIMEOUT", "45"))
 
 _FALLBACK_RESULT = {
-    "name": "未识别", "brand": "无法判断", "category": "非产品图",
-    "appearance": "无法判断", "packaging": "无法判断", "text_on_package": [],
-    "key_features": ["无法判断"], "scene": "通用", "mood": "",
-    "portrait_prompt": "无法判断", "summary": "未识别",
+    "name": "未识别",
+    "brand": "无法判断",
+    "category": "非产品图",
+    "appearance": "无法判断",
+    "packaging": "无法判断",
+    "text_on_package": [],
+    "key_features": ["无法判断"],
+    "scene": "通用",
+    "mood": "",
+    "portrait_prompt": "无法判断",
+    "summary": "未识别",
 }
 
 
@@ -55,16 +62,23 @@ def analyze_image_v2(idx: int, img_url: str) -> dict[str, Any]:
     with ThreadPoolExecutor(max_workers=2) as pool:
         f_fj = pool.submit(vlm_fast_json.call_fast_json, img_url, timeout=_FAST_JSON_TIMEOUT)
         f_ocr = pool.submit(ocr_volc.call_ocr, img_url, timeout=_OCR_TIMEOUT)
-        for fut in as_completed([f_fj, f_ocr], timeout=_FAST_TIMEOUT + 2):
-            try:
-                res = fut.result(timeout=1)
-            except Exception as e:
-                logger.warning("[vision.v2] 图片 #%d 子任务异常: %s", idx, e)
-                continue
-            if fut is f_fj and isinstance(res, dict):
-                fj_result = res
-            elif fut is f_ocr and isinstance(res, list):
-                ocr_result = res
+        try:
+            for fut in as_completed([f_fj, f_ocr], timeout=_FAST_TIMEOUT):
+                try:
+                    res = fut.result(timeout=1)
+                except Exception as e:
+                    logger.warning("[vision.v2] 图片 #%d 子任务异常: %s", idx, e)
+                    continue
+                if fut is f_fj and isinstance(res, dict):
+                    fj_result = res
+                elif fut is f_ocr and isinstance(res, list):
+                    ocr_result = res
+        except TimeoutError:
+            # fast 整体超时，取消还没跑完的子任务，继续走 pro 兜底
+            for f in (f_fj, f_ocr):
+                if not f.done():
+                    f.cancel()
+            logger.warning("[vision.v2] 图片 #%d fast路径超时(%.0fs)，走pro兜底", idx, _FAST_TIMEOUT)
 
     fast_elapsed = time.time() - t0
 
@@ -75,7 +89,9 @@ def analyze_image_v2(idx: int, img_url: str) -> dict[str, Any]:
             assembled["_fast_elapsed"] = round(fast_elapsed, 2)
             logger.info(
                 "[vision.v2] 图片 #%d fast命中 elapsed=%.2fs pp=%s",
-                idx, fast_elapsed, (assembled.get("portrait_prompt") or "")[:40],
+                idx,
+                fast_elapsed,
+                (assembled.get("portrait_prompt") or "")[:40],
             )
             return assembled
 
@@ -88,11 +104,11 @@ def analyze_image_v2(idx: int, img_url: str) -> dict[str, Any]:
         pro_result["_pro_elapsed"] = round(time.time() - pro_t0, 2)
         if ocr_result and not pro_result.get("text_on_package"):
             pro_result["text_on_package"] = ocr_result[:8]
-        logger.info("[vision.v2] 图片 #%d pro兜底命中 total=%.2fs", idx, time.time()-t0)
+        logger.info("[vision.v2] 图片 #%d pro兜底命中 total=%.2fs", idx, time.time() - t0)
         return pro_result
 
     # 最终：返回最小可用结果
-    logger.warning("[vision.v2] 图片 #%d 全路径失败 elapsed=%.2fs", idx, time.time()-t0)
+    logger.warning("[vision.v2] 图片 #%d 全路径失败 elapsed=%.2fs", idx, time.time() - t0)
     out = dict(_FALLBACK_RESULT)
     out["_source"] = "v2_all_failed"
     out["text_on_package"] = ocr_result[:8]

apps/worker/worker_app/tasks/vision/vlm_fallback.py

@@ -3,6 +3,7 @@
 
 设计原则：简单、直接、无竞速、无复杂超时逻辑。只在 fast_json 结果不可用时调用。
 """
+
 from __future__ import annotations
 
 import json
@@ -85,26 +86,48 @@ def _xml_to_product(raw: str, idx: int) -> dict[str, Any]:
         if pp_attr and pp_attr != "无人像":
             portrait_prompt = pp_attr
         return {
-            "name": name, "brand": brand, "category": category,
-            "appearance": appearance, "packaging": packaging, "text_on_package": text_list,
-            "key_features": feat_list, "scene": scene, "mood": mood,
-            "portrait_prompt": portrait_prompt, "summary": summary,
+            "name": name,
+            "brand": brand,
+            "category": category,
+            "appearance": appearance,
+            "packaging": packaging,
+            "text_on_package": text_list,
+            "key_features": feat_list,
+            "scene": scene,
+            "mood": mood,
+            "portrait_prompt": portrait_prompt,
+            "summary": summary,
             "_source": "vlm_pro_xml",
         }
 
     if portrait_prompt != "无人像":
         return {
-            "name": "未识别", "brand": "无法判断", "category": "无法判断",
-            "appearance": "无法判断", "packaging": "无法判断", "text_on_package": [],
-            "key_features": ["无法判断"], "scene": scene, "mood": mood,
-            "portrait_prompt": portrait_prompt, "summary": "未识别",
+            "name": "未识别",
+            "brand": "无法判断",
+            "category": "无法判断",
+            "appearance": "无法判断",
+            "packaging": "无法判断",
+            "text_on_package": [],
+            "key_features": ["无法判断"],
+            "scene": scene,
+            "mood": mood,
+            "portrait_prompt": portrait_prompt,
+            "summary": "未识别",
             "_source": "vlm_pro_no_product",
         }
     return {
-        "name": "未识别", "brand": "无法判断", "category": "无法判断",
-        "appearance": "无法判断", "packaging": "无法判断", "text_on_package": [],
-        "key_features": ["无法判断"], "scene": scene, "mood": mood,
-        "portrait_prompt": "无人像", "summary": "未识别", "_source": "vlm_pro_no_tag",
+        "name": "未识别",
+        "brand": "无法判断",
+        "category": "无法判断",
+        "appearance": "无法判断",
+        "packaging": "无法判断",
+        "text_on_package": [],
+        "key_features": ["无法判断"],
+        "scene": scene,
+        "mood": mood,
+        "portrait_prompt": "无人像",
+        "summary": "未识别",
+        "_source": "vlm_pro_no_tag",
     }
 
 
@@ -141,6 +164,8 @@ def call_pro_vlm(
         return None
 
     use_model = model or DEFAULT_PRO_MODEL
+    _orig_retries = client.max_retries
+    client.max_retries = 0
     try:
         raw = client.vision_completion(
             messages=[{"role": "system", "content": system}, {"role": "user", "content": user}],
@@ -151,8 +176,10 @@ def call_pro_vlm(
             model=use_model,
         )
     except Exception as e:
-        logger.warning("[vision.vlm] 图片 #%d pro VLM 调用失败 elapsed=%.1fs err=%s", idx, time.time()-t0, e)
+        logger.warning("[vision.vlm] 图片 #%d pro VLM 调用失败 elapsed=%.1fs err=%s", idx, time.time() - t0, e)
+        client.max_retries = _orig_retries
         return None
+    client.max_retries = _orig_retries
 
     elapsed = time.time() - t0
     if not raw:
@@ -163,7 +190,7 @@ def call_pro_vlm(
     l, r = text.find("{"), text.rfind("}")
     if l >= 0 and r > l:
         try:
-            obj = json.loads(text[l:r+1])
+            obj = json.loads(text[l : r + 1])
             if isinstance(obj, dict):
                 logger.info("[vision.vlm] 图片 #%d pro VLM JSON 完成 elapsed=%.1fs", idx, elapsed)
                 return {
@@ -189,7 +216,9 @@ def call_pro_vlm(
         result["_pro_elapsed"] = round(elapsed, 2)
         logger.info(
             "[vision.vlm] 图片 #%d pro VLM XML 完成 elapsed=%.2fs pp=%s",
-            idx, elapsed, (result.get("portrait_prompt") or "")[:40],
+            idx,
+            elapsed,
+            (result.get("portrait_prompt") or "")[:40],
         )
         return result
     except Exception as e:

apps/worker/worker_app/tasks/vision/vlm_fast_json.py

@@ -81,28 +81,92 @@ def call_fast_json(
 ) -> dict[str, Any] | None:
     """调用 lite VLM 返回结构化 dict；失败/非 JSON 返回 None。
 
-    注意：不做重试（外层竞速/降级逻辑负责），max_retries=0 由外层统一设置。
+    直接用 httpx 发最小 payload（关闭 thinking），不走 ai_client 包装：
+    - 关闭 thinking/推理链（reasoning_tokens 是延迟主因，单次要10-12s）
+    - 单次调用不重试（失败由外层走 pro 兜底）
+    - 温度=0.1 稳定输出 JSON
     """
     t0 = time.time()
-    try:
-        from packages.shared.ai_client import get_doubao_client
+    import httpx
 
-        client = get_doubao_client()
-        if not client.is_available:
-            logger.warning("[vision.v2] doubao client 不可用，跳过 fast_json")
+    try:
+        from packages.shared import get_shared_settings
+
+        settings = get_shared_settings()
+        api_key = settings.doubao_api_key
+        base_url = (settings.doubao_base_url or "https://ark.cn-beijing.volces.com/api/v3").rstrip("/")
+        if not api_key:
+            logger.warning("[vision.v2] doubao api_key 未配置，跳过 fast_json")
             return None
 
         use_model = model or DEFAULT_LITE_MODEL
-        raw = client.vision_completion(
-            messages=[
+        url = f"{base_url}/chat/completions"
+        payload: dict[str, Any] = {
+            "model": use_model,
+            "messages": [
                 {"role": "system", "content": _FAST_SYSTEM},
-                {"role": "user", "content": _FAST_USER},
+                {
+                    "role": "user",
+                    "content": [
+                        {"type": "image_url", "image_url": {"url": img_url}},
+                        {"type": "text", "text": _FAST_USER},
+                    ],
+                },
             ],
-            images=[img_url],
-            temperature=0.1,
-            max_tokens=max_tokens,
+            "temperature": 0.1,
+            "max_tokens": max_tokens,
+            "stream": False,
+        }
+        # 关键：关闭 thinking（避免产生 reasoning_tokens 拖慢响应）
+        # 方舟/豆包 2.x 模型支持 thinking.type=disabled
+        try:
+            payload["thinking"] = {"type": "disabled"}
+        except Exception:
+            pass
+        # 部分模型用 reasoning_effort 控制思考深度
+        payload["reasoning_effort"] = "low"
+
+        resp = httpx.post(
+            url,
+            headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
+            json=payload,
             timeout=timeout,
-            model=use_model,
+        )
+        elapsed = time.time() - t0
+        if resp.status_code != 200:
+            logger.warning(
+                "[vision.v2] fast_json HTTP %d elapsed=%.1fs body=%s", resp.status_code, elapsed, resp.text[:200]
+            )
+            # 如果400说明不支持thinking参数，降级重试一次
+            if resp.status_code == 400 and "thinking" in resp.text.lower():
+                payload.pop("thinking", None)
+                payload.pop("reasoning_effort", None)
+                time.time()
+                resp = httpx.post(
+                    url,
+                    headers={"Authorization": f"Bearer {api_key}", "Content-Type": "application/json"},
+                    json=payload,
+                    timeout=timeout,
+                )
+                elapsed = time.time() - t0
+                if resp.status_code != 200:
+                    logger.warning("[vision.v2] fast_json 降级后 HTTP %d elapsed=%.1fs", resp.status_code, elapsed)
+                    return None
+            else:
+                return None
+        data = resp.json()
+        raw = (data.get("choices") or [{}])[0].get("message", {}).get("content")
+        if raw is None:
+            logger.warning("[vision.v2] fast_json 返回 None elapsed=%.1fs", elapsed)
+            return None
+        usage = data.get("usage") or {}
+        logger.info(
+            "[vision.v2] fast_json 直连完成 model=%s elapsed=%.1fs in=%d out=%d reasoning=%d",
+            use_model,
+            elapsed,
+            usage.get("prompt_tokens", 0),
+            usage.get("completion_tokens", 0),
+            usage.get("reasoning_tokens", 0),
         )
         elapsed = time.time() - t0
         if raw is None: