优化视频创作

core/backend/apps/ai/creation_agent.py

@@ -27,7 +27,7 @@ from django.core.serializers.json import DjangoJSONEncoder
 from django.db import transaction
 
 from .creation import append_message, pin_refs
-from .creation_presets import apply_image_preset_prompt, preset_guidance
+from .creation_presets import apply_image_preset_prompt, preset_guidance, preset_workflow_guidance
 from .mentions import TYPE_LABELS, infer_field_types, resolve_refs, search_mentions
 from .models import CreationConversation, CreationMessage, ModelConfig
 from .services import build_provider, get_default_model, get_seed_text_model, resolve_text_model
@@ -41,12 +41,13 @@ MAX_TOOL_ROUNDS = 8
 MAX_BILLED_GENERATIONS = 1
 
 # 视频闸门阶段(落在 conversation.memory.stage;resume 靠它)
-# clarify → strategy → plan → prompt → confirm → done
+# prompt 仅兼容历史会话。新视频链路在方案确认后直接转入出片确认，Prompt 始终由平台在后台维护。
+# clarify → strategy → plan → confirm → done
 VIDEO_GATE_STAGES = ("clarify", "strategy", "plan", "prompt", "confirm", "done")
 _STEP_CONFIRM_LABELS = {
     "strategy": "创作策略已写好。确认后继续写方案；要改就点「我想改」或直接说改哪里。",
-    "plan": "视频方案已写好。确认后继续出 Prompt；要改就点「我想改」或直接说改哪里。",
+    "plan": "视频方案已写好。确认后我会整理出片细节，并带你确认生成参数；要改就点「我想改」或直接说改哪里。",
-    "prompt": "出片 Prompt 已写好。确认后选择参数并生成；要改就点「我想改」或直接说改哪里。",
+    "prompt": "出片指令已整理好。确认后核对参数并生成；要改就点「我想改」或直接说改哪里。",
 }
 
 
@@ -1552,7 +1553,7 @@ def build_system_prompt(context: AgentContext, *, allow_plan: bool = True, has_c
         "- 用户选择暂不提供某项素材时,把它当成明确授权:按已有信息和合理默认继续。除非任务客观上无法完成,否则不要再次追问同一素材。",
         "- 用户说「你来定」「你帮我选」「随便」「都行」时,就是授权你做专业判断;直接选合理方案继续,不要把选择题再抛回去。",
         "- 禁止问「要不要继续」「要不要生成」「是否开始创作」这类流程问题。缺信息用 ask_user;信息够了就写策略。"
-        "  视频每写完一步(策略/方案/Prompt)平台会出确认卡,由用户点「按这个继续」或提出修改;不要口头问流程。",
+        "  视频每写完策略或方案，平台会出确认卡；方案确认后平台会在后台整理出片指令，再让用户核对生成参数。不要口头问流程。",
         "- 用户打招呼或闲聊(hi / 你好 / 在吗 / 你在干什么 / 嗯 / 好的 / ok):"
         "  **禁止**调用 write_strategy、write_plan、generate_image,不要「整理方案」或直接开写脚本。",
         "- 会话里**还没有**商品/方向时:自然地告诉用户可以直接丢一句想法,别硬推销,也别用客服式结束语。",
@@ -1582,6 +1583,15 @@ def build_system_prompt(context: AgentContext, *, allow_plan: bool = True, has_c
     ]
     if context.is_video:
         lines.extend(["", _OMNI_VIDEO_PROMPT_RULES.strip()])
+        lines.extend([
+            "",
+            "【视频创作工作原则】",
+            "- 先从文字和已锁定素材整理商品事实、人物/服装、场景、参考视频或音频、现成脚本，以及时长、比例和语言；已经给出的信息不重复问。",
+            "- 用户已给脚本或分镜时，以它为基础补足，不强制从头重写。只改用户点名的镜头、人物、商品、台词或参数，其他内容保持。",
+            "- 可以自己决定转场、灯光、普通镜头细节；商品功能、规格、价格、活动、功效和关键使用边界不能猜，缺失时一次只问一项。",
+            "- 用户上传的人物、商品、服装和场景优先作为参考；如确实需要额外生成角色、场景或道具，先说明用途和预计积分，等用户确认。",
+            "- 出片前检查：主卖点都有画面或台词证据、口播能在时长内说完、脚本中每位人物/商品/服装/场景都有对应素材、商品正常使用、全片一致、所有 SKU 都已安排。内容超出时长时建议删减、延长或拆分，而不是硬塞。",
+        ])
         # 按会话时长给出口播字数锚点(与专业创作 narration_limit 同口径)
         try:
             dur = video_duration(params)
@@ -1591,14 +1601,13 @@ def build_system_prompt(context: AgentContext, *, allow_plan: bool = True, has_c
         hi = max(lo, min(85, int(dur * 5.7)))
         lines.append(f"- 当前按约 {dur} 秒出片,口播建议 {lo}–{hi} 字;write_plan 的 voice_chars 填这个区间。")
         lines.append(
-            "- 视频 5 步闸门(不可同轮连跳):①缺信息 ask_user 停 → ②write_strategy 停等确认 → "
+            "- 视频 4 步闸门(不可同轮连跳):①缺信息 ask_user 停 → ②write_strategy 停等确认 → "
-            "③用户确认后 write_plan 停等确认 → ④用户确认后 write_prompt(或平台代出 Prompt)停等确认 → "
+            "③用户确认后 write_plan 停等确认 → ④方案确认后平台在后台整理完整出片指令，并直接出现积分确认卡。"
-            "⑤用户确认后才出现积分确认卡出片。"
         )
         lines.append(
-            "- 只有用户明确要做片、出方案、改方案、换卖点/剧情时才调用 write_strategy / write_plan / write_prompt;"
+            "- 只有用户明确要做片、出方案、改方案、换卖点/剧情时才调用 write_strategy / write_plan;"
             "闲聊与打招呼绝对不要。用户对某一步提出修改时,只重写那一步,不要跳到后面。"
-            "video_prompt 按上面的秒级分镜规范写满,不要只给大纲,也禁止只回「好的,有需要再说」。"
+            "write_plan 里的 video_prompt 要按上面的秒级分镜规范写满，供平台后台出片使用；不要只给大纲，也禁止只回「好的,有需要再说」。"
         )
         stage = get_video_gate_stage(conversation)
         memory = conversation.memory if isinstance(conversation.memory, dict) else {}
@@ -1606,13 +1615,13 @@ def build_system_prompt(context: AgentContext, *, allow_plan: bool = True, has_c
         stage_hint = {
             "clarify": "当前阶段=澄清:缺关键信息就 ask_user;信息够了只调 write_strategy。",
             "strategy": (
-                "当前阶段=策略已确认,请调用 write_plan 写方案;不要再写策略,不要 write_prompt。"
+                "当前阶段=策略已确认,请调用 write_plan 写方案;不要再写策略。"
                 if strategy_confirmed else
-                "当前阶段=等策略确认:不要再 write_plan/write_prompt;用户确认后才会进入方案。若用户在改策略,只重调 write_strategy。"
+                "当前阶段=等策略确认:不要再写方案或出片;用户确认后才会进入方案。若用户在改策略,只重调 write_strategy。"
             ),
-            "plan": "当前阶段=等方案确认:不要 write_prompt 或出片。若用户在改方案,只重调 write_plan。",
+            "plan": "当前阶段=等方案确认:不要出片。若用户在改方案,只重调 write_plan。",
-            "prompt": "当前阶段=等 Prompt 确认:不要出积分确认/出片。若用户在改 Prompt,只重调 write_prompt。",
+            "prompt": "当前阶段=兼容历史会话的出片指令确认:不要出片，按用户反馈重写内部出片指令。",
-            "confirm": "当前阶段=等出片确认:不要再写策略/方案/Prompt;用户会在确认卡上点开始生成。",
+            "confirm": "当前阶段=等出片确认:不要再写策略/方案;用户会在确认卡上点开始生成。",
             "done": "当前阶段=已出片:等用户新的修改或新需求再行动。用户说重新来/重来/从头开始时,当作新一轮创作,从澄清或 write_strategy 重开,不要再弹旧模特/商品追问。",
         }.get(stage, "")
         if stage_hint:
@@ -1651,6 +1660,9 @@ def build_system_prompt(context: AgentContext, *, allow_plan: bool = True, has_c
         if guidance:
             lines.append(guidance)
             lines.append("用户选了这个预设,就按它的拍法来;要偏离得先问过用户。")
+        workflow_guidance = preset_workflow_guidance(conversation.preset) if context.is_video else ""
+        if workflow_guidance:
+            lines.append(f"【当前预设的工作重点】{workflow_guidance}")
 
     resolved = resolve_refs(context.team, conversation.pinned_refs or [])
     if resolved.facts:
@@ -2344,7 +2356,7 @@ _TOOL_LABELS = {
     "generate_image": "生成图片",
     "write_strategy": "梳理创作策略",
     "write_plan": "编排视频方案",
-    "write_prompt": "编写出片 Prompt",
+    "write_prompt": "整理出片指令",
 }
 
 
@@ -2403,11 +2415,21 @@ def _elicit_payload_for_fields(
         else:
             gate_label = _GATE_LABELS.get(primary_type, _GATE_LABELS["asset"])
 
+        asset_label = {
+            "product": "商品",
+            "character": "角色",
+            "model": "模特",
+            "scene": "场景",
+        }.get(primary_type, "素材")
         gate_field = {
             "key": "_asset_gate",
             "label": gate_label,
             "type": "text",
             "required": False,
+            "options": [
+                {"value": "send", "label": f"发{asset_label}列表"},
+                {"value": "auto", "label": "你来推荐"},
+            ],
         }
         return {
             # 这是 Agent 的一句自然追问，不是让用户点选流程的卡片。
@@ -2526,7 +2548,7 @@ def _dispatch_tool(
             kind=CreationMessage.Kind.PLAN, payload=plan_payload,
         )
         events.append({"type": "message", "message": _message_payload(plan)})
-        # video_prompt 先缓存:用户确认方案后由平台展示 Prompt 卡,本步不出 Prompt/积分卡
+        # video_prompt 只存后台。用户确认方案后直接进入积分确认，不展示内部 Prompt。
         set_video_gate_stage(
             context.conversation, "plan", pending_video_prompt=video_prompt
         )
@@ -2545,12 +2567,16 @@ def _dispatch_tool(
             return {"payload": {"error": "video_prompt 不能为空,请把完整出片指令写进去"}}, False
         video_prompt = apply_product_voice_visual_guard(context.conversation, video_prompt)
         video_prompt = apply_product_reality_guard(video_prompt)
-        emitted = emit_prompt_gate(context.conversation, video_prompt)
+        set_video_gate_stage(context.conversation, "prompt", pending_video_prompt=video_prompt)
-        if not emitted:
+        confirm = emit_final_confirm_gate(context.conversation, context=context)
-            return {"payload": {"error": "无法写出 Prompt 卡"}}, False
+        if confirm is None:
-        events = [{"type": "message", "message": _message_payload(m)} for m in emitted]
+            return {"payload": {"error": "无法准备出片确认"}}, False
+        events = [
+            {"type": "message", "message": _message_payload(confirm)},
+            {"type": "credits", "estimated": int((confirm.payload or {}).get("estimated_credits") or 0)},
+        ]
         return {
-            "payload": {"awaiting_step": "prompt", "written": True},
+            "payload": {"awaiting_confirmation": True, "prepared": True},
             "_events": events,
         }, True
 

core/backend/apps/ai/creation_presets.py

@@ -1,6 +1,6 @@
 """全能创作 · 预设(契约 §6)。
 
-首页那 8+6 张预设卡不只是个名字 —— 每个预设代表一套**明确的拍法**。
+首页的预设卡不只是个名字 —— 每个预设代表一套**明确的拍法**。
 只把「达人口播种草」四个字塞进提示词,模型只能靠猜;这里给它可执行的约束。
 
 新增预设 = 在下面加一条,前端 `omni-create.tsx` 的 PRESET 列表加一张卡。两边的
@@ -9,6 +9,10 @@ key 必须是同一个中文名(会话建的时候原样存进 CreationConversat
 from __future__ import annotations
 
 VIDEO_PRESETS: dict[str, str] = {
+    "痛点解决演示": (
+        "真实问题解决型短片。按「具体痛点 → 商品自然登场 → 正常使用步骤 → 可见结果 → 轻行动引导」推进。"
+        "痛点必须来自用户提供的商品用途；使用方法、效果与结果只写能被素材或商品事实证明的内容，禁止夸大功效或虚构测试。"
+    ),
     "真实使用演示": (
         "可信实拍型商品短片。先按商品真实用途、可见结构和已有卖点安排演示，再谈氛围和转化；"
         "每一个卖点都必须由看得见的使用步骤、材质细节或结果承接，不能凭空加性能。"
@@ -40,9 +44,14 @@ VIDEO_PRESETS: dict[str, str] = {
         "鱼眼/广角近距离透视,连续换装节奏。**人物面部和身形必须全程一致**,"
         "只有服装在变。每次换装用一个明确动作触发。"
     ),
+    "多色商品换款": (
+        "先识别并整理不同颜色、款式或 SKU，确保每款的比例、位置、外观与顺序清楚。"
+        "统一构图和机位，用点击/触碰或一个明确动作触发换款；背景、光线、机位全程稳定，最后用全家福镜头覆盖全部款式。"
+    ),
+    # 旧会话仍会保存「点触换款」，保留同一拍法约束以支持历史继续创作。
     "点触换款": (
-        "统一构图和机位,用点击/触碰动作触发商品款式切换,快速展示多个 SKU。"
+        "先识别并整理不同颜色、款式或 SKU，确保每款的比例、位置、外观与顺序清楚。"
-        "背景、光线、机位全程不变 —— 变化只发生在商品本身。"
+        "统一构图和机位，用点击/触碰或一个明确动作触发换款；背景、光线、机位全程稳定，最后用全家福镜头覆盖全部款式。"
     ),
     "探店漫游": (
         "以空间动线串联:入口 → 环境 → 关键细节 → 服务/主推项目。"
@@ -52,6 +61,37 @@ VIDEO_PRESETS: dict[str, str] = {
         "强调光影、材质和镜头节奏,建立品牌识别。慢节奏、精致构图、克制的色彩。"
         "少即是多,不要堆信息。"
     ),
+    "前后对比实测": (
+        "围绕同一对象和同一可比较条件，依次拍清使用前状态、正常使用过程、使用后状态与可见对比。"
+        "优先使用用户提供的前/中/后素材；没有事实依据时不承诺功效、不制造夸张差异。"
+    ),
+    "AI 宠物拟人": (
+        "先确定宠物的外观、性格和可执行动作，再设计一段轻巧剧情。宠物与商品互动要符合商品真实结构和正常操作方式；"
+        "趣味来自角色反应和剧情，不把商品改造成不真实的玩具或凭空增加能力。"
+    ),
+    "人物与音色替换": (
+        "先理解原视频的镜头、人物、动作、台词、画外音、字幕、场景和节奏。只替换用户指定的人物、声音或两者；"
+        "其他构图、运镜、动作、商品、场景和剪辑节奏尽量保持原样。用户未提供原视频或替换目标时，先问这一个必要信息。"
+    ),
+}
+
+# 预设除了决定最终 Prompt，也决定 Agent 在素材整理、追问与生成前检查时的工作重点。
+# 这段进入系统提示词，避免把预设退化成一句风格修饰词。
+VIDEO_PRESET_WORKFLOWS: dict[str, str] = {
+    "痛点解决演示": "先确认商品真实解决的具体问题与正常用法；生成前核对痛点、过程和结果都有可见证据。",
+    "真实使用演示": "优先核对商品真实用途、关键步骤和可证实卖点；不为氛围而增加不合理测试。",
+    "剧情反转带货": "优先确认人物关系或冲突；缺剧情时给 2–3 个可选方向，商品必须成为解决问题或反转的关键。",
+    "商品拟人广告": "先确认商品外观和性格表达方式；默认无脸拟人，商品保持真实完整，台词用画外声。",
+    "达人口播种草": "优先确认达人/人物、真实体验和主卖点；生成前核对口播字数能在时长内说完，每个卖点都有画面证明。",
+    "商品图一键成片": "优先从商品参考图锁定外观；自动补场景和动作，但不替换或改变用户商品图里的结构、颜色和包装。",
+    "鱼眼换装": "优先确认人物参考、服装套数和展示顺序；生成前核对脸、身形、场景稳定，只有服装随动作切换。",
+    "多色商品换款": "优先识别每个颜色或款式及其顺序；生成前核对所有 SKU 都已进入时间轴，比例、位置和机位保持一致。",
+    "点触换款": "优先识别每个颜色或款式及其顺序；生成前核对所有 SKU 都已进入时间轴，比例、位置和机位保持一致。",
+    "探店漫游": "优先确认门店动线与主推项目；用连续移动串联入口、环境、细节和服务，不做碎片化硬切。",
+    "品牌质感大片": "优先确认品牌气质、材质和商品主卖点；由 Agent 决定光线和镜头，不把专业选择反复抛给用户。",
+    "前后对比实测": "优先确认同一对象的前、中、后素材或可比较条件；生成前核对对比不夸大且没有缺失关键阶段。",
+    "AI 宠物拟人": "优先确认宠物角色参考、性格和商品互动方式；生成前核对宠物、商品结构和使用动作前后一致。",
+    "人物与音色替换": "先确认原视频、替换对象和替换范围；未经要求不得改变原视频的构图、动作、商品、场景或节奏。",
 }
 
 IMAGE_PRESETS: dict[str, str] = {
@@ -83,6 +123,11 @@ def preset_guidance(name: str) -> str:
     return ALL_PRESETS.get((name or "").strip(), "")
 
 
+def preset_workflow_guidance(name: str) -> str:
+    """预设名 → Agent 的素材、追问和检查重点。"""
+    return VIDEO_PRESET_WORKFLOWS.get((name or "").strip(), "")
+
+
 def apply_image_preset_prompt(name: str, prompt: str) -> str:
     """将已选图片预设的风格字段确定性并入实际出图指令，不依赖 Agent 自行复述。"""
     base = str(prompt or "").strip()

core/backend/apps/ai/test_creation_agent.py

@@ -217,6 +217,10 @@ class AskUserTests(CreationAgentBaseTests):
         self.assertEqual(payload["interaction"], "chat")
         self.assertEqual(payload["phase"], "gate")
         self.assertEqual(payload["fields"][0]["key"], "_asset_gate")
+        self.assertEqual(
+            payload["fields"][0]["options"],
+            [{"value": "send", "label": "发商品列表"}, {"value": "auto", "label": "你来推荐"}],
+        )
         self.assertEqual(payload["pending_fields"][0]["type"], "asset")
         self.assertEqual(payload["pending_fields"][0]["asset_types"], ["product"])
         self.assertIn("商品库列表", elicit[0]["message"]["text"])
@@ -903,7 +907,7 @@ class VideoPlanAndConfirmTests(CreationAgentBaseTests):
         # 方案闸门停下,不能同轮出 Prompt/积分卡
         self.assertEqual(len(fake.calls), 1)
 
-    def test_write_prompt_emits_prompt_and_step_confirm(self):
+    def test_write_prompt_keeps_internal_prompt_and_emits_confirm(self):
         fake = FakeProvider([
             _tool_chunks("write_prompt", {"video_prompt": "0-3秒 近景手持商品…"}),
             _text_chunks("不该跑到这一轮"),
@@ -912,13 +916,10 @@ class VideoPlanAndConfirmTests(CreationAgentBaseTests):
             events = _events(stream_creation_agent(conversation=self.conversation, user=self.user,
                                                    text="做条视频", model_config=self.model))
         kinds = [e["message"]["kind"] for e in events if e.get("type") == "message"]
-        self.assertIn("prompt_file", kinds)
+        self.assertNotIn("prompt_file", kinds)
-        self.assertNotIn("confirm", kinds)
+        self.assertIn("confirm", kinds)
-        elicits = [
+        self.conversation.refresh_from_db()
-            e["message"] for e in events
+        self.assertEqual((self.conversation.memory or {}).get("stage"), "confirm")
-            if e.get("type") == "message" and e["message"]["kind"] == "elicit"
-        ]
-        self.assertEqual(elicits[-1]["payload"].get("step"), "prompt")
         self.assertEqual(len(fake.calls), 1)
 
     def test_strategy_and_plan_same_round_stops_after_strategy(self):
@@ -1291,6 +1292,18 @@ class PresetGuidanceTests(CreationAgentBaseTests):
         self.assertIn("真实使用演示", system)
         self.assertIn("禁止无依据拍成破损、漏液、渗水", system)
 
+    def test_video_preset_adds_workflow_checks_not_only_a_final_prompt_style(self):
+        conversation = CreationConversation.objects.create(
+            team=self.team, created_by=self.user, mode="video", preset="多色商品换款", params={},
+        )
+        context = AgentContext(conversation=conversation, user=self.user, model_config=self.model)
+
+        system = build_system_prompt(context)
+
+        self.assertIn("当前预设的工作重点", system)
+        self.assertIn("所有 SKU 都已进入时间轴", system)
+        self.assertIn("比例、位置和机位保持一致", system)
+
     def test_product_reality_guard_is_added_to_final_prompt(self):
         from .creation_agent import apply_product_reality_guard
 
@@ -1319,10 +1332,10 @@ class PresetGuidanceTests(CreationAgentBaseTests):
         self.assertIn("避免高饱和、现代摄影感", prompt)
 
     def test_every_frontend_preset_has_guidance(self):
-        """前端 9 个视频 + 12 个图片预设都要有拍法,漏一个就等于那张卡是摆设。"""
+        """视频和图片预设都要有拍法,漏一个就等于那张卡是摆设。"""
         from .creation_presets import IMAGE_PRESETS, VIDEO_PRESETS
 
-        self.assertEqual(len(VIDEO_PRESETS), 9)
+        self.assertEqual(len(VIDEO_PRESETS), 14)
         self.assertEqual(len(IMAGE_PRESETS), 12)
         self.assertTrue(all(text.strip() for text in {**VIDEO_PRESETS, **IMAGE_PRESETS}.values()))
 

core/backend/apps/ai/views.py

@@ -40,7 +40,6 @@ from .creation_agent import (
     apply_restart_intent,
     apply_session_params,
     emit_final_confirm_gate,
-    emit_prompt_gate,
     is_greeting,
     is_restart_intent,
     restore_gated_step_after_cancel,
@@ -174,11 +173,11 @@ _STEP_CONTINUE_INSTRUCTIONS = {
         "不要再写策略,不要调用 write_prompt,不要出片。"
     ),
     "plan": (
-        "用户已确认视频方案。若尚未写出 Prompt,调用 write_prompt 展示出片 Prompt;"
+        "用户已确认视频方案。补齐完整 video_prompt 后直接调用 write_prompt 整理后台出片指令并给出积分确认卡;"
-        "不要重写策略/方案,不要出积分确认卡或直接出片。"
+        "不要重写策略/方案，也不要直接出片。"
     ),
     "prompt": (
-        "用户已确认出片 Prompt。不要再写策略/方案/Prompt;"
+        "用户已确认后台出片指令。不要再写策略/方案;"
         "平台会出积分确认卡,等用户点开始生成。"
     ),
 }
@@ -193,8 +192,8 @@ _STEP_REVISE_INSTRUCTIONS = {
         "(含完整 video_prompt 存档);写完即停,不要同轮 write_prompt 或出片。"
     ),
     "prompt": (
-        "用户要求修改出片 Prompt。根据反馈只重新调用 write_prompt 写一版修订 Prompt;"
+        "用户要求修改出片细节。根据反馈只重新调用 write_prompt 整理一版修订指令;"
-        "写完即停,不要出积分确认卡或直接出片。"
+        "完成后给出积分确认卡，不要直接出片。"
     ),
 }
 
@@ -249,17 +248,21 @@ def _handle_step_confirm_answer(
         return None, True, _STEP_CONTINUE_INSTRUCTIONS["strategy"]
 
     if step == "plan":
-        emitted = emit_prompt_gate(conversation)
+        confirm = emit_final_confirm_gate(conversation)
-        if not emitted:
+        if confirm is None:
-            # 缓存丢了:让 agent 补 write_prompt
+            # 兼容旧会话或缓存丢失:让 agent 补齐后台出片指令，再直接给出积分确认。
             return None, True, _STEP_CONTINUE_INSTRUCTIONS["plan"]
         conversation.agent_status = CreationConversation.AgentStatus.AWAITING_USER
         conversation.save(update_fields=["agent_status", "updated_at"])
-        return JsonResponse({
+        credits = int((confirm.payload or {}).get("estimated_credits") or 0)
+        body = {
             "conversation_id": str(conversation.id),
             "agent_status": conversation.agent_status,
-            "messages": [CreationMessageSerializer(m).data for m in emitted],
+            "messages": [CreationMessageSerializer(confirm).data],
-        }, status=200), False, ""
+        }
+        if credits:
+            body["estimated_credits"] = credits
+        return JsonResponse(body, status=200), False, ""
 
     if step == "prompt":
         confirm = emit_final_confirm_gate(conversation)

core/frontend/src/design-restraint.css

@@ -274,14 +274,14 @@ aside.sidebar {
 .brand { display: flex; align-items: center; justify-content: center; min-width: 0; color: var(--accent-black); }
 .brand-clip {
   display: block;
-  width: 112px;
+  width: 132px;
   height: 90px;
   overflow: hidden;
   transition: width var(--t-base), height var(--t-base);
 }
 .brand-logo {
   display: block;
-  width: 112px;
+  width: 132px;
   height: 90px;
   max-width: none;
   margin: 0;

core/frontend/src/omni-create-page.css

@@ -790,47 +790,9 @@
   transform: translateY(7px);
 }
 
-.omni-case-head {
-  display: flex;
-  align-items: flex-end;
-  justify-content: flex-start;
-  gap: 20px;
-  margin-bottom: 15px;
-}
-
-.omni-case-head-label {
-  flex: 0 0 auto;
-  color: #252b35;
-  font-size: 17px;
-  font-weight: 750;
-  line-height: 30px;
-}
-
-.omni-case-filters {
-  display: flex;
-  align-items: center;
-  gap: 4px;
-}
-
-.omni-case-filters button {
-  height: 30px;
-  padding: 0 12px;
-  border-radius: 8px;
-  color: #707681;
-  background: transparent;
-  font-size: 13px;
-  font-weight: 600;
-  cursor: pointer;
-}
-
-.omni-case-filters button.active {
-  color: #fff;
-  background: var(--black);
-}
-
 .omni-case-grid {
   display: grid;
-  grid-template-columns: repeat(3, minmax(0, 1fr));
+  grid-template-columns: repeat(4, minmax(0, 1fr));
   gap: 14px;
 }
 
@@ -840,24 +802,24 @@
   gap: 14px;
 }
 
-.omni-image-preset-intro {
+.omni-preset-intro {
   margin: 0 0 14px;
   text-align: center;
 }
 
-.omni-image-preset-intro strong,
+.omni-preset-intro strong,
-.omni-image-preset-intro span {
+.omni-preset-intro span {
   display: block;
 }
 
-.omni-image-preset-intro strong {
+.omni-preset-intro strong {
   color: var(--accent-black);
   font-size: 16px;
   font-weight: 500;
   line-height: 1.4;
 }
 
-.omni-image-preset-intro span {
+.omni-preset-intro span {
   margin-top: 4px;
   color: var(--black-alpha-56);
   font-size: 13px;
@@ -899,6 +861,12 @@
   background: #edf0f4;
 }
 
+/* 视频样片本身是竖版，按竖构图展示，避免横向裁切掉人物和商品。 */
+.omni-case-card[data-mode="video"] .omni-case-visual {
+  height: auto;
+  aspect-ratio: 9 / 15;
+}
+
 .omni-case-visual img,
 .omni-case-visual video {
   width: 100%;

core/frontend/src/omni-session-page.css

@@ -2250,23 +2250,25 @@
 .omni-gate-actions {
   display: flex;
   flex-wrap: wrap;
-  gap: 8px;
+  gap: 6px;
 }
 
 .omni-gate-actions button {
-  padding: 8px 14px;
+  min-height: 30px;
-  border: 1px solid rgba(34, 42, 54, .12);
+  padding: 0 10px;
-  border-radius: 999px;
+  border: 1px solid var(--border-faint);
-  background: #fff;
+  border-radius: var(--r-md);
-  color: #575d66;
+  background: var(--surface);
-  font-size: 13px;
+  color: var(--accent-black);
+  font-size: 12px;
+  font-weight: 500;
   cursor: pointer;
 }
 
 .omni-gate-actions button.primary {
-  border-color: var(--klein);
+  border-color: var(--heat-20);
-  background: var(--klein);
+  background: var(--heat-12);
-  color: #fff;
+  color: var(--heat);
 }
 
 .omni-gate-actions button:disabled {

core/frontend/src/routes/omni-create.tsx

@@ -35,18 +35,16 @@ type PresetItem = {
   desc?: string;
   starter?: string;
   cover: string;
+  previewVideo?: string;
 };
 
 const VIDEO_PRESETS: PresetItem[] = [
-  { name: "真实使用演示", category: "commerce", mode: "video", title: "真实使用演示", desc: "从商品真实用途出发，拍清使用过程与卖点证据，避免不合理的夸张演示。", starter: "基于商品的真实用途创作一条可信的使用演示：拍清使用步骤、产品状态和核心卖点，不夸张、不出现不合理的使用场景。", cover: "/assets/prototype/video-oneclick-film-v3.png" },
+  { name: "痛点解决演示", category: "commerce", mode: "video", title: "痛点解决演示", desc: "从真实痛点到正常使用，再用看得见的过程证明商品价值。", starter: "基于商品的真实用途创作一条痛点解决演示：先让人看见困扰，再拍清使用步骤和真实结果，不夸张、不出现不合理的测试。", cover: "/assets/video-presets/pain-point-demo.jpg", previewVideo: "/assets/video-presets/pain-point-demo.mp4" },
-  { name: "剧情反转带货", category: "story", mode: "video", title: "剧情反转带货", desc: "用人物冲突与意外反转建立记忆点，商品承担剧情中的关键作用。", starter: "创作一条有前后反转的剧情带货视频，让商品自然成为解决问题的关键。", cover: "/assets/prototype/photo-1529139574466-a303027c1d8b.jpg" },
+  { name: "剧情反转带货", category: "story", mode: "video", title: "剧情反转带货", desc: "用人物冲突与意外反转建立记忆点，商品承担剧情中的关键作用。", starter: "创作一条有前后反转的剧情带货视频，让商品自然成为解决问题的关键。", cover: "/assets/video-presets/plot-twist-commerce.jpg", previewVideo: "/assets/video-presets/plot-twist-commerce.mp4" },
-  { name: "商品拟人广告", category: "commerce", mode: "video", title: "商品拟人广告", desc: "让商品成为故事主角，通过个性、动作和情绪完成趣味表达。", starter: "把商品塑造成有性格的拟人角色，创作一条轻快、有趣的短广告。", cover: "/assets/prototype/photo-1608248543803-ba4f8c70ae0b.jpg" },
+  { name: "达人口播种草", category: "speaker", mode: "video", title: "达人口播种草", desc: "用真实体验与生活化表达建立信任，适合电商和本地生活内容。", starter: "创作一条真实自然的达人口播种草视频，重点讲清使用场景和核心卖点。", cover: "/assets/video-presets/creator-recommendation.jpg", previewVideo: "/assets/video-presets/creator-recommendation.mp4" },
-  { name: "达人口播种草", category: "speaker", mode: "video", title: "达人口播种草", desc: "用真实体验与生活化表达建立信任，适合电商和本地生活内容。", starter: "创作一条真实自然的达人口播种草视频，重点讲清使用场景和核心卖点。", cover: "/assets/prototype/photo-1494790108377-be9c29b29330.jpg" },
+  { name: "鱼眼换装", category: "visual", mode: "video", title: "鱼眼换装", desc: "强调近距离透视与连续变装节奏，适合服饰和人物视觉内容。", starter: "创作一条鱼眼镜头风格的连续换装视频，人物和服装需要保持稳定。", cover: "/assets/video-presets/rhythm-outfit-change.jpg", previewVideo: "/assets/video-presets/rhythm-outfit-change.mp4" },
-  { name: "商品图一键成片", category: "commerce", mode: "video", title: "商品图一键成片", desc: "从商品参考图建立镜头语言，自动补齐场景、动作与转场。", starter: "根据我上传的商品图片生成一条节奏清晰的电商短视频。", cover: "/assets/prototype/photo-1556228578-8c89e6adf883.jpg" },
+  { name: "多色商品换款", category: "visual", mode: "video", title: "多色商品换款", desc: "统一商品比例与机位，用动作连续展示不同颜色和款式。", starter: "创作一条多色商品换款视频：统一构图展示不同颜色或款式，用明确动作触发切换，最后给出全款总览。", cover: "/assets/video-presets/multi-sku-switch.jpg", previewVideo: "/assets/video-presets/multi-sku-switch.mp4" },
-  { name: "鱼眼换装", category: "visual", mode: "video", title: "鱼眼换装", desc: "强调近距离透视与连续变装节奏，适合服饰和人物视觉内容。", starter: "创作一条鱼眼镜头风格的连续换装视频，人物和服装需要保持稳定。", cover: "/assets/prototype/photo-1483985988355-763728e1935b.jpg" },
+  { name: "AI 宠物拟人", category: "story", mode: "video", title: "AI 宠物拟人", desc: "让宠物角色参与有趣小剧情，同时保留商品真实结构与用途。", starter: "创作一条 AI 宠物拟人短片：宠物有明确性格和动作，商品以真实结构和正常用法自然参与剧情。", cover: "/assets/video-presets/ai-pet-personification.jpg", previewVideo: "/assets/video-presets/ai-pet-personification.mp4" },
-  { name: "点触换款", category: "visual", mode: "video", title: "点触换款", desc: "用统一构图和动作触发商品变化，快速展示多个款式或 SKU。", starter: "创作一条通过点击动作连续切换不同商品款式的短视频。", cover: "/assets/prototype/unbranded-running-shoe-remix.png" },
-  { name: "探店漫游", category: "story", mode: "video", title: "探店漫游", desc: "以空间动线串联门店环境、服务细节与主推项目。", starter: "根据门店素材创作一条有路线感和空间氛围的探店漫游视频。", cover: "/assets/prototype/photo-1497366754035-f200968a6e72.jpg" },
-  { name: "品牌质感大片", category: "visual", mode: "video", title: "品牌质感大片", desc: "通过统一的光影、材质与镜头节奏建立更强的品牌识别。", starter: "创作一条强调光影、材质和品牌气质的高级感商品短片。", cover: "/assets/prototype/video-free-cinematic.png" },
 ];
 
 const IMAGE_PRESETS: PresetItem[] = [
@@ -64,19 +62,6 @@ const IMAGE_PRESETS: PresetItem[] = [
   { name: "新中式", category: "style", mode: "image", title: "新中式", cover: "/assets/image-presets/neo-chinese.png" },
 ];
 
-const VIDEO_FILTERS = [
-  { key: "all", label: "全部" },
-  { key: "story", label: "剧情" },
-  { key: "commerce", label: "电商" },
-  { key: "speaker", label: "口播" },
-  { key: "visual", label: "视觉" },
-];
-
-const IMAGE_FILTERS = [
-  { key: "all", label: "全部" },
-  { key: "style", label: "风格" },
-];
-
 const MENTION_REF_LIMIT = 5;
 
 const MENTION_TABS: Array<{ type: CreationRef["type"]; label: string; Icon: typeof ImageIcon }> = [
@@ -134,7 +119,6 @@ export function OmniCreatePage({
   const [typeLabels, setTypeLabels] = useState<Record<string, string>>({});
   const [pendingRefs, setPendingRefs] = useState<CreationRef[]>([]);
   const [previewCase, setPreviewCase] = useState<PresetItem | null>(null);
-  const [activeCategory, setActiveCategory] = useState("all");
   const fileInputRef = useRef<HTMLInputElement>(null);
   const toolsRef = useRef<HTMLDivElement>(null);
   const [starting, setStarting] = useState(false);
@@ -155,7 +139,6 @@ export function OmniCreatePage({
       setRatio("9:16");
       setDuration("智能时长");
     }
-    setActiveCategory("all");
     setSelectedCase((prev) => (prev && prev.mode !== outputMode ? null : prev));
   }, [outputMode, modelConfigs]);
 
@@ -170,17 +153,13 @@ export function OmniCreatePage({
     return () => document.removeEventListener("mousedown", onDown);
   }, []);
 
-  const filters = outputMode === "video" ? VIDEO_FILTERS : IMAGE_FILTERS;
   const presets = useMemo(() => {
-    const list = outputMode === "video" ? VIDEO_PRESETS : IMAGE_PRESETS;
+    return outputMode === "video" ? VIDEO_PRESETS : IMAGE_PRESETS;
-    if (activeCategory === "all") return list;
+  }, [outputMode]);
-    return list.filter((item) => item.category === activeCategory);
-  }, [outputMode, activeCategory]);
 
   const applyPreset = (preset: PresetItem) => {
     setSelectedCase(preset);
-    // 图片预设只作为隐藏的生成风格，用户输入区保留自己的创作描述，不把内部 Prompt 展示出来。
+    // 预设只作为后台的创作方向。无论图片还是视频，都不把内部生成指令塞进用户输入区。
-    if (preset.mode !== "image" && preset.starter) setPrompt(preset.starter);
     setPreviewCase(null);
     onNotify?.("info", `已选择预设：${preset.name}`);
   };
@@ -460,32 +439,12 @@ export function OmniCreatePage({
 
         <section
           className={`omni-case-library${outputMode === "image" ? " is-image-presets" : ""}`}
-          aria-labelledby={outputMode === "video" ? "omniCaseTitle" : undefined}
+          aria-label={outputMode === "image" ? "图片风格预设" : "视频创作预设"}
-          aria-label={outputMode === "image" ? "图片风格预设" : undefined}
         >
-          {outputMode === "video" ? (
+          <div className="omni-preset-intro">
-            <div className="omni-case-head">
+            <strong>{outputMode === "video" ? "从一个视频方向开始" : "从一个风格开始"}</strong>
-              <strong className="omni-case-head-label" id="omniCaseTitle">预设分类</strong>
+            <span>{outputMode === "video" ? "可选，也可以直接描述你想拍的视频。" : "可选，也可以直接描述你想要的画面。"}</span>
-              <div className="omni-case-filters" role="tablist" aria-label="预设筛选">
+          </div>
-                {filters.map((filter) => (
-                  <button
-                    type="button"
-                    key={filter.key}
-                    className={activeCategory === filter.key ? "active" : ""}
-                    onClick={() => setActiveCategory(filter.key)}
-                  >
-                    {filter.label}
-                  </button>
-                ))}
-              </div>
-            </div>
-          ) : null}
-          {outputMode === "image" ? (
-            <div className="omni-image-preset-intro">
-              <strong>从一个风格开始</strong>
-              <span>可选，也可以直接描述你想要的画面。</span>
-            </div>
-          ) : null}
           <div className="omni-case-grid">
             {presets.map((card) => (
               <button
@@ -493,16 +452,39 @@ export function OmniCreatePage({
                 className={`omni-case-card${selectedCase?.name === card.name ? " active" : ""}`}
                 data-mode={card.mode}
                 key={card.name}
+                onMouseEnter={(event) => {
+                  const video = event.currentTarget.querySelector<HTMLVideoElement>("video");
+                  void video?.play().catch(() => undefined);
+                }}
+                onMouseLeave={(event) => {
+                  const video = event.currentTarget.querySelector<HTMLVideoElement>("video");
+                  if (!video) return;
+                  video.pause();
+                  video.currentTime = 0;
+                }}
+                onFocus={(event) => {
+                  const video = event.currentTarget.querySelector<HTMLVideoElement>("video");
+                  void video?.play().catch(() => undefined);
+                }}
+                onBlur={(event) => {
+                  const video = event.currentTarget.querySelector<HTMLVideoElement>("video");
+                  if (!video) return;
+                  video.pause();
+                  video.currentTime = 0;
+                }}
                 onClick={() => {
                   if (card.mode === "image") applyPreset(card);
                   else setPreviewCase(card);
                 }}
               >
                 <span className="omni-case-visual">
-                  <img src={card.cover} alt="" />
+                  {card.previewVideo ? (
-                  <span className="omni-case-play">
+                    <video muted loop playsInline preload="metadata" poster={card.cover} aria-label={`${card.title} 样片预览`}>
-                    {card.mode === "video" ? <Play className="lucide-play" /> : <ImageIcon />}
+                      <source src={card.previewVideo} type="video/mp4" />
-                  </span>
+                    </video>
+                  ) : (
+                    <img src={card.cover} alt="" />
+                  )}
                 </span>
                 <span className="omni-case-copy">
                   <strong>{card.title}</strong>
@@ -530,8 +512,15 @@ export function OmniCreatePage({
                 <X />
               </button>
               <div className="omni-preset-media-frame">
-                <img src={previewCase.cover} alt="预设示例" />
+                {previewCase.previewVideo ? (
-                {previewCase.mode === "video" && (
+                  <video controls playsInline preload="metadata" poster={previewCase.cover}>
+                    <source src={previewCase.previewVideo} type="video/mp4" />
+                    当前浏览器不支持播放该示例视频。
+                  </video>
+                ) : (
+                  <img src={previewCase.cover} alt="预设示例" />
+                )}
+                {previewCase.mode === "video" && !previewCase.previewVideo && (
                   <span className="omni-preset-video-mark">
                     <Play />
                   </span>

core/frontend/src/routes/omni-session.tsx

@@ -399,17 +399,6 @@ function collectSessionAssets(
       });
     }
 
-    if (message.kind === "prompt_file") {
-      const title = String(message.payload?.title || "视频生成Prompt.md");
-      const body = String(message.payload?.body || "");
-      push({
-        key: `prompt:${message.id || title}`,
-        kind: "document",
-        title,
-        promptTitle: title,
-        promptBody: body || "暂无内容",
-      });
-    }
   }
 
   return out;
@@ -512,7 +501,7 @@ function PlanCard({ payload }: { payload: Record<string, unknown> }) {
     <section className="omni-video-plan-card">
       <header className="omni-video-plan-head">
         <strong>视频最终方案</strong>
-        <span>请确认后再出 Prompt</span>
+        <span>确认后进入出片参数核对</span>
       </header>
       <section className="omni-plan-section">
         <strong>卖点做减法</strong>
@@ -555,9 +544,9 @@ function PlanCard({ payload }: { payload: Record<string, unknown> }) {
           </span>
         ) : null}
         <span className="omni-plan-final-input">
-          最终输入：
+          出片准备：
           <b>
-            Prompt.md + <em className="omni-plan-point-body">{String(payload.ref_count ?? 0)}</em> 组参考素材
+            已整合生成参数与 <em className="omni-plan-point-body">{String(payload.ref_count ?? 0)}</em> 组参考素材
           </b>
         </span>
       </div>
@@ -565,30 +554,21 @@ function PlanCard({ payload }: { payload: Record<string, unknown> }) {
   );
 }
 
-function PromptFileCard({
+function PromptFileCard({ payload }: { payload: Record<string, unknown> }) {
-  payload,
-  onView,
-}: {
-  payload: Record<string, unknown>;
-  onView: () => void;
-}) {
   return (
     <section className="omni-prompt-file-card">
       <span className="omni-prompt-file-icon">
         <FileText />
       </span>
       <div className="omni-prompt-file-copy">
-        <strong>{String(payload.title || "视频生成Prompt.md")}</strong>
+        <strong>出片指令已准备好</strong>
-        <small>已整合策略、镜头、口播、节奏与 {String(payload.ref_count ?? 0)} 组参考素材</small>
+        <small>已根据方案、参数和 {String(payload.ref_count ?? 0)} 组参考素材在后台整理完成</small>
       </div>
-      <button type="button" onClick={onView}>
-        查看 Prompt
-      </button>
     </section>
   );
 }
 
-/** 视频 5 步闸门:策略/方案/Prompt 下方的「按这个继续 / 我想改」。
+/** 视频阶段闸门:策略/方案下方的「按这个继续 / 我想改」。
  * 「我想改」在卡片内展开反馈输入，不再依赖底部 composer。 */
 function StepConfirmCard({
   message,
@@ -606,14 +586,14 @@ function StepConfirmCard({
   const answers = (message.payload.answers as Record<string, string> | undefined) || {};
   const action = String(answers.step_action || "");
   const stepLabel =
-    step === "strategy" ? "创作策略" : step === "plan" ? "视频方案" : step === "prompt" ? "出片 Prompt" : "这一步";
+    step === "strategy" ? "创作策略" : step === "plan" ? "视频方案" : step === "prompt" ? "出片指令" : "这一步";
   const placeholder =
     step === "strategy"
       ? "说说你想怎么改策略…"
       : step === "plan"
         ? "说说你想怎么改方案…"
         : step === "prompt"
-          ? "说说你想怎么改 Prompt…"
+          ? "说说你想怎么改出片细节…"
           : "说说你想怎么改…";
   const [revising, setRevising] = useState(false);
   const [feedback, setFeedback] = useState("");
@@ -781,21 +761,6 @@ function ElicitCard({
     );
   }
 
-  if (interaction === "chat") {
-    return (
-      <div className="omni-chat-row agent">
-        <span className="omni-chat-avatar">
-          <Sparkles />
-        </span>
-        <div className="omni-chat-bubble">
-          <p className="omni-chat-text">
-            {message.text || fields[0]?.label || "这项你想怎么定？"}
-          </p>
-        </div>
-      </div>
-    );
-  }
-
   const toggleMulti = (key: string, value: string) => {
     setAnswers((prev) => {
       const current = Array.isArray(prev[key]) ? (prev[key] as string[]) : [];
@@ -828,12 +793,12 @@ function ElicitCard({
         <div className="omni-chat-bubble omni-gate-bubble">
           <p className="omni-chat-text">{message.text || gateLabel}</p>
           {!submitted ? (
-            <div className="omni-gate-actions">
+            <div className="omni-gate-actions" aria-label="商品选择操作">
-              {(gateField?.options || []).map((option) => (
+              {(gateField?.options || []).map((option, index) => (
                 <button
                   type="button"
                   key={option.value}
-                  className={option.value === "send" ? "primary" : ""}
+                  className={index === 0 ? "primary" : ""}
                   disabled={disabled}
                   onClick={() => {
                     if (!gateField?.key) return;
@@ -854,6 +819,32 @@ function ElicitCard({
     );
   }
 
+  if (interaction === "chat") {
+    const question = message.text || fields[0]?.label || "这项你想怎么定？";
+    return (
+      <div className="omni-chat-row agent">
+        <span className="omni-chat-avatar">
+          <Sparkles />
+        </span>
+        <div className="omni-chat-stack">
+          <div className="omni-chat-bubble">
+            <p className="omni-chat-text">{question}</p>
+          </div>
+          {!submitted ? (
+            <div className="omni-reply-guide" aria-label="下一步引导">
+              <p><span>下一步</span>直接在下方输入你的想法，回答这一项就行。</p>
+              <div className="omni-reply-guide-options">
+                <button type="button" className="primary" onClick={() => document.getElementById("omniSessionPrompt")?.focus()}>
+                  现在填写
+                </button>
+              </div>
+            </div>
+          ) : null}
+        </div>
+      </div>
+    );
+  }
+
   return (
     <div className="omni-elicit-slot">
     <section className={`omni-elicit-card${submitted ? " is-submitted" : ""}`}>
@@ -1051,11 +1042,15 @@ function ConfirmCard({
   );
 }
 
-function ResultCard({ payload }: { payload: Record<string, unknown> }) {
+function ResultCard({
+  payload,
+}: {
+  payload: Record<string, unknown>;
+}) {
   const assets = (payload.assets as Array<Record<string, string>> | undefined) || [];
   const first = assets[0] || {};
   const meta = [payload.model, payload.resolution, payload.ratio].filter(Boolean).join(" · ");
-  const isVideo = first.type === "video";
+  const isGeneratedVideo = first.type === "video";
   const [preview, setPreview] = useState<{ src: string; kind: "image" | "video"; name: string } | null>(null);
 
   return (
@@ -1100,7 +1095,7 @@ function ResultCard({ payload }: { payload: Record<string, unknown> }) {
       </div>
       <div className="omni-result-info">
         <div>
-          <strong>{isVideo ? "视频已生成" : assets.length > 1 ? `已生成 ${assets.length} 张图` : "图片已生成"}</strong>
+          <strong>{isGeneratedVideo ? "视频已生成" : assets.length > 1 ? `已生成 ${assets.length} 张图` : "图片已生成"}</strong>
           <small>{meta}</small>
         </div>
       </div>
@@ -2178,10 +2173,6 @@ export function OmniSessionPage({
           <PromptFileCard
             key={message.clientKey || message.id}
             payload={message.payload}
-            onView={() => setPromptView({
-              title: String(message.payload.title || "视频生成Prompt.md"),
-              body: String(message.payload.body || "暂无内容"),
-            })}
           />
         );
       case "confirm": {
@@ -2213,7 +2204,12 @@ export function OmniSessionPage({
       case "generating":
         return <ProcessCard key={message.clientKey || message.id} payload={message.payload} />;
       case "result":
-        return <ResultCard key={message.clientKey || message.id} payload={message.payload} />;
+        return (
+          <ResultCard
+            key={message.clientKey || message.id}
+            payload={message.payload}
+          />
+        );
       default: {
         const refs = message.refs || [];
         const body = stripMentionText(message.text, refs);