fix(ui): widget key 按项目隔离，避免滚动卡死/加载串项目；fix(img): 最小180s+失败码重试3次

app.py

@@ -133,6 +133,20 @@ if "view_mode" not in st.session_state:
     st.session_state.view_mode = "workspace" # workspace, history, settings
 if "selected_img_provider" not in st.session_state:
     st.session_state.selected_img_provider = "shubiaobiao"
+if "ui_rev" not in st.session_state:
+    # used to scope Streamlit widget keys per project load to avoid cross-project state pollution
+    st.session_state.ui_rev = int(time.time() * 1000)
+
+
+def _ui_key(suffix: str) -> str:
+    """
+    Build a project-scoped widget key.
+    Streamlit widget state is keyed globally; without project scoping, switching projects can
+    reuse old widget states and even trigger frontend DOM errors (e.g. removeChild NotFoundError).
+    """
+    pid = st.session_state.get("project_id") or "NEW"
+    rev = st.session_state.get("ui_rev") or 0
+    return f"p:{pid}|r:{rev}|{suffix}"
 
 def load_project(project_id):
     """Load project state from DB"""
@@ -142,6 +156,8 @@ def load_project(project_id):
         return
     
     st.session_state.project_id = project_id
+    # bump UI revision to ensure all widget keys are isolated per project load
+    st.session_state.ui_rev = int(time.time() * 1000)
     st.session_state.script_data = data.get("script_data")
     st.session_state.view_mode = "workspace"
 
@@ -384,20 +400,21 @@ if st.session_state.view_mode == "workspace":
         col1, col2 = st.columns([1, 1])
         
         with col1:
-            product_name = st.text_input("商品标题", value=default_name)
+            product_name = st.text_input("商品标题", value=default_name, key=_ui_key("product_name"))
-            category = st.text_input("商品类目", value=default_category)
+            category = st.text_input("商品类目", value=default_category, key=_ui_key("category"))
-            price = st.text_input("价格", value=default_price)
+            price = st.text_input("价格", value=default_price, key=_ui_key("price"))
             
         with col2:
-            tags = st.text_area("评价标签 (用于提炼卖点)", value=default_tags, height=100)
+            tags = st.text_area("评价标签 (用于提炼卖点)", value=default_tags, height=100, key=_ui_key("tags"))
-            params = st.text_area("商品参数", value=default_params, height=100)
+            params = st.text_area("商品参数", value=default_params, height=100, key=_ui_key("params"))
 
         # 商家自定义风格提示
         style_hint = st.text_area(
             "商品视频重点增强提示 (可选)", 
             value=loaded_info.get("style_hint", ""),
             placeholder="例如：韩风、高级感、活力青春、简约日系...",
-            height=80
+            height=80,
+            key=_ui_key("style_hint"),
         )
 
         st.markdown("### 上传素材")
@@ -418,7 +435,7 @@ if st.session_state.view_mode == "workspace":
         
         # Model Selection (all support images; user explicitly chooses model)
         model_options = ["GPT-5.2", "Gemini 3 Pro", "Doubao Pro (Vision)"]
-        selected_model_label = st.radio("选择脚本生成模型", model_options, horizontal=True, index=0)
+        selected_model_label = st.radio("选择脚本生成模型", model_options, horizontal=True, index=0, key=_ui_key("script_model"))
         # Map label to provider key
         if selected_model_label == "GPT-5.2":
             model_provider = "shubiaobiao_gpt"
@@ -548,17 +565,17 @@ if st.session_state.view_mode == "workspace":
                     for i, item in enumerate(timeline):
                         c1, c2, c3, c4 = st.columns([3, 3, 1, 1])
                         with c1:
-                            item["text"] = st.text_input(f"旁白 #{i+1}", value=item.get("text", ""), key=f"tl_vo_{i}")
+                            item["text"] = st.text_input(f"旁白 #{i+1}", value=item.get("text", ""), key=_ui_key(f"tl_vo_{i}"))
                         with c2:
-                            item["subtitle"] = st.text_input(f"字幕 #{i+1}", value=item.get("subtitle", item.get("text", "")), key=f"tl_sub_{i}")
+                            item["subtitle"] = st.text_input(f"字幕 #{i+1}", value=item.get("subtitle", item.get("text", "")), key=_ui_key(f"tl_sub_{i}"))
                         with c3:
                             # 兼容旧格式: 如果有 start_ratio 则转换为 start_time
                             default_start = item.get("start_time", item.get("start_ratio", 0) * 12)  # 假设总时长12秒
-                            item["start_time"] = st.number_input(f"开始(秒) #{i+1}", value=float(default_start), min_value=0.0, max_value=30.0, step=0.5, key=f"tl_start_{i}")
+                            item["start_time"] = st.number_input(f"开始(秒) #{i+1}", value=float(default_start), min_value=0.0, max_value=30.0, step=0.5, key=_ui_key(f"tl_start_{i}"))
                         with c4:
                             # 兼容旧格式: 如果有 duration_ratio 则转换为 duration
                             default_dur = item.get("duration", item.get("duration_ratio", 0.25) * 12)  # 假设总时长12秒
-                            item["duration"] = st.number_input(f"时长(秒) #{i+1}", value=float(default_dur), min_value=0.5, max_value=15.0, step=0.5, key=f"tl_dur_{i}")
+                            item["duration"] = st.number_input(f"时长(秒) #{i+1}", value=float(default_dur), min_value=0.5, max_value=15.0, step=0.5, key=_ui_key(f"tl_dur_{i}"))
                         # 清理旧字段
                         item.pop("start_ratio", None)
                         item.pop("duration_ratio", None)
@@ -577,8 +594,8 @@ if st.session_state.view_mode == "workspace":
                         c_vis, c_aud = st.columns([2, 1])
                         
                         with c_vis:
-                            new_visual = st.text_area(f"Visual Prompt (Scene {scene['id']})", value=scene.get("visual_prompt", ""), height=80, key=f"vp_{i}")
+                            new_visual = st.text_area(f"Visual Prompt (Scene {scene['id']})", value=scene.get("visual_prompt", ""), height=80, key=_ui_key(f"vp_{i}"))
-                            new_video = st.text_area(f"Video Prompt (Scene {scene['id']})", value=scene.get("video_prompt", ""), height=80, key=f"vidp_{i}")
+                            new_video = st.text_area(f"Video Prompt (Scene {scene['id']})", value=scene.get("video_prompt", ""), height=80, key=_ui_key(f"vidp_{i}"))
                             scene["visual_prompt"] = new_visual
                             scene["video_prompt"] = new_video
                             
@@ -589,7 +606,7 @@ if st.session_state.view_mode == "workspace":
                                 new_ft_text = st.text_input(
                                     f"Fancy Text (Scene {scene['id']})",
                                     value=ft.get("text", ""),
-                                    key=f"ft_{i}",
+                                    key=_ui_key(f"ft_{i}"),
                                 )
                                 # 兼容：旧数据可能没有 fancy_text 字段
                                 if not isinstance(scene.get("fancy_text"), dict):
@@ -652,7 +669,7 @@ if st.session_state.view_mode == "workspace":
             if not st.session_state.scene_images:
                 # Model Selection for Image Gen
                 img_model_options = ["Shubiaobiao (Gemini)", "Doubao (Volcengine)", "Gemini (Direct)", "Doubao (Group Image)"]
-                selected_img_model = st.radio("选择生图模型", img_model_options, horizontal=True)
+                selected_img_model = st.radio("选择生图模型", img_model_options, horizontal=True, key=_ui_key("img_model"))
                 
                 # Map to provider
                 if "Group Image" in selected_img_model:
@@ -1116,17 +1133,17 @@ if st.session_state.view_mode == "workspace":
                 for i, item in enumerate(timeline):
                     c1, c2, c3, c4 = st.columns([3, 3, 1, 1])
                     with c1:
-                        item["text"] = st.text_input(f"旁白 #{i+1}", value=item.get("text", ""), key=f"tune_vo_{i}")
+                        item["text"] = st.text_input(f"旁白 #{i+1}", value=item.get("text", ""), key=_ui_key(f"tune_vo_{i}"))
                     with c2:
-                        item["subtitle"] = st.text_input(f"字幕 #{i+1}", value=item.get("subtitle", item.get("text", "")), key=f"tune_sub_{i}")
+                        item["subtitle"] = st.text_input(f"字幕 #{i+1}", value=item.get("subtitle", item.get("text", "")), key=_ui_key(f"tune_sub_{i}"))
                     with c3:
                         # 兼容旧格式: 如果有 start_ratio 则转换为 start_time
                         default_start = item.get("start_time", item.get("start_ratio", 0) * 12)
-                        item["start_time"] = st.number_input(f"开始(秒) #{i+1}", value=float(default_start), min_value=0.0, max_value=30.0, step=0.5, key=f"tune_start_{i}")
+                        item["start_time"] = st.number_input(f"开始(秒) #{i+1}", value=float(default_start), min_value=0.0, max_value=30.0, step=0.5, key=_ui_key(f"tune_start_{i}"))
                     with c4:
                         # 兼容旧格式: 如果有 duration_ratio 则转换为 duration
                         default_dur = item.get("duration", item.get("duration_ratio", 0.25) * 12)
-                        item["duration"] = st.number_input(f"时长(秒) #{i+1}", value=float(default_dur), min_value=0.5, max_value=15.0, step=0.5, key=f"tune_dur_{i}")
+                        item["duration"] = st.number_input(f"时长(秒) #{i+1}", value=float(default_dur), min_value=0.5, max_value=15.0, step=0.5, key=_ui_key(f"tune_dur_{i}"))
                     # 清理旧字段
                     item.pop("start_ratio", None)
                     item.pop("duration_ratio", None)
@@ -1145,7 +1162,7 @@ if st.session_state.view_mode == "workspace":
                         st.markdown(f"**Scene {scene['id']}**")
                         ft = scene.get("fancy_text", {})
                         ft_text = ft.get("text", "") if isinstance(ft, dict) else ""
-                        new_ft = st.text_input(f"花字", value=ft_text, key=f"tune_ft_{i}")
+                        new_ft = st.text_input(f"花字", value=ft_text, key=_ui_key(f"tune_ft_{i}"))
                         # 兼容：旧数据可能没有 fancy_text 字段
                         if not isinstance(scene.get("fancy_text"), dict):
                             scene["fancy_text"] = {}

modules/image_gen.py

@@ -28,11 +28,12 @@ def _env_int(name: str, default: int) -> int:
 
 
 # Tunables: slow channels can be hot; default conservative but adjustable.
-IMG_SUBMIT_TIMEOUT_S = _env_int("IMG_SUBMIT_TIMEOUT_S", 180)
+# IMPORTANT: we enforce minimums to avoid accidental misconfig (e.g. 120s) causing flaky UX.
-IMG_POLL_TIMEOUT_S = _env_int("IMG_POLL_TIMEOUT_S", 30)
+IMG_SUBMIT_TIMEOUT_S = max(_env_int("IMG_SUBMIT_TIMEOUT_S", 180), 180)
-IMG_MAX_RETRIES = _env_int("IMG_MAX_RETRIES", 3)
+IMG_POLL_TIMEOUT_S = max(_env_int("IMG_POLL_TIMEOUT_S", 30), 10)
-IMG_POLL_INTERVAL_S = _env_int("IMG_POLL_INTERVAL_S", 2)
+IMG_MAX_RETRIES = max(_env_int("IMG_MAX_RETRIES", 3), 3)
-IMG_POLL_MAX_RETRIES = _env_int("IMG_POLL_MAX_RETRIES", 90)  # 90*2s ~= 180s
+IMG_POLL_INTERVAL_S = max(_env_int("IMG_POLL_INTERVAL_S", 2), 1)
+IMG_POLL_MAX_RETRIES = max(_env_int("IMG_POLL_MAX_RETRIES", 90), 90)  # 90*2s ~= 180s
 
 
 def _is_retryable_exception(e: Exception) -> bool:
@@ -43,6 +44,22 @@ def _is_retryable_exception(e: Exception) -> bool:
     # Transient provider errors often contain these keywords
     if any(k in msg for k in ["timeout", "temporarily", "temporarily unavailable", "gateway", "rate", "try again"]):
         return True
+    # Treat common HTTP transient status codes as retryable when they bubble up as RuntimeError text
+    # Examples from our code: "Shubiaobiao 提交失败 (429): ..." / "Doubao Image Failed (502): ..."
+    try:
+        import re
+        m = re.search(r"\((\d{3})\)", msg)
+        if not m:
+            # sometimes formatted like "429:" without parentheses
+            m = re.search(r"\b(\d{3})\b", msg)
+        if m:
+            code = int(m.group(1))
+            if code in (408, 409, 425, 429, 500, 502, 503, 504):
+                return True
+            if 500 <= code <= 599:
+                return True
+    except Exception:
+        pass
     return False