优化全能创作

core/backend/apps/ai/creation_agent.py

@@ -65,6 +65,8 @@ CREATION_AGENT_MODEL_TIMEOUT_SECONDS = 180
 CREATION_AGENT_TURN_TIMEOUT_SECONDS = 420
 # 输出被 max_tokens 截断后最多让模型重写几次;再失败就明确告诉用户,不再闷着重试到超时。
 MAX_TRUNCATION_RETRIES = 2
+# 已明确要求推进创作时，模型若只回说明文字而没调工具，最多强制纠正两次。
+MAX_PROGRESS_TOOL_RETRIES = 2
 LONG_VIDEO_DURATION_SLACK_SECONDS = 2
 # 单条用户消息最多触发一次计费生成(契约 §4)
 MAX_BILLED_GENERATIONS = 1
@@ -145,6 +147,16 @@ TRUNCATION_GIVE_UP_NOTICE = (
     "可以把时长改短一点，或直接点「按这个继续」让我按更紧凑的结构重写一次。"
 )
 
+PROGRESS_TOOL_RETRY_INSTRUCTION = (
+    "用户已经明确授权继续创作，本轮不能用说明文字、建议或‘继续完善方案’结束。"
+    "信息不足就调用 ask_user；信息足够就立即调用 write_strategy，并在同一轮继续调用 write_plan，"
+    "最终必须落下可见的视频架构卡。不要输出工具调用以外的解释文字。"
+)
+
+PROGRESS_TOOL_GIVE_UP_NOTICE = (
+    "这次没有成功整理出视频方案，已经停止。请把刚才的要求再发一次，我会直接生成方案卡。"
+)
+
 
 def long_video_script_covers_requested_duration(
     raw_duration: int | None,
@@ -195,6 +207,17 @@ def set_public_agent_progress(
     ):
         return
     history = list(memory.get("agent_progress_history") or [])
+    # reasoning 可能在「理解需求 / 商品 / 角色」之间来回切换。再次回到同一阶段时
+    # 把旧位置移除后放到末尾，既保留当前顺序，也不让加载卡重复刷同一句话。
+    history = [
+        item
+        for item in history
+        if not (
+            isinstance(item, dict)
+            and str(item.get("phase") or "") == phase
+            and str(item.get("detail_key") or "") == detail_key
+        )
+    ]
     history.append({"phase": phase, "detail_key": detail_key})
     memory["agent_progress_history"] = history[-12:]
     memory["agent_progress_phase"] = phase
@@ -1446,6 +1469,70 @@ def build_cast_role_briefs(count: int, appearance_prompt: str = "") -> list[str]
     return [_character_only_appearance_text(item)[:500] for item in briefs[:count]]
 
 
+_FEMALE_CAST_WORDS = r"女性|女生|女人|女孩|女演员|女模特|女士|女主|妈妈|母亲|女儿|姐姐|妹妹"
+_MALE_CAST_WORDS = r"男性|男生|男人|男孩|男演员|男模特|男士|男主|爸爸|父亲|儿子|哥哥|弟弟"
+_CAST_COUNT_WORDS = {"一": 1, "1": 1, "两": 2, "二": 2, "2": 2, "三": 3, "3": 3, "四": 4, "4": 4}
+
+
+def _cast_gender_from_text(text: str) -> str:
+    """只读明确的角色性别描述；不从商品或风格词推断。"""
+    female = bool(re.search(_FEMALE_CAST_WORDS, text))
+    male = bool(re.search(_MALE_CAST_WORDS, text))
+    if female == male:
+        return ""
+    return "女性" if female else "男性"
+
+
+def _shared_cast_gender_from_text(text: str, total: int) -> str:
+    """仅当文本明确说整组角色同一性别时，才把性别约束复制给每张定妆图。"""
+    matches = re.finditer(
+        rf"(?<!\d)([一1两二2三3四4])\s*(?:位|名|个|人)?\s*(?:成年|年轻|中年|年长)?\s*"
+        rf"({_FEMALE_CAST_WORDS}|{_MALE_CAST_WORDS})",
+        text,
+    )
+    genders = {
+        _cast_gender_from_text(match.group(2))
+        for match in matches
+        if _CAST_COUNT_WORDS.get(match.group(1)) == total
+    }
+    genders.discard("")
+    if len(genders) == 1:
+        return genders.pop()
+    if genders:
+        return ""
+    if total == 2:
+        if re.search(r"母女|姐妹|闺蜜|双女主", text) and not re.search(r"父子|兄弟|双男主|一男一女|男女主", text):
+            return "女性"
+        if re.search(r"父子|兄弟|双男主", text) and not re.search(r"母女|姐妹|闺蜜|双女主|一男一女|男女主", text):
+            return "男性"
+    return ""
+
+
+def _cast_gender_requirements(
+    conversation: CreationConversation,
+    role_briefs: list[str],
+    appearance_prompt: str,
+) -> list[str]:
+    """外观输入优先，其次读取已确认的视频 Prompt；只传性别，不带商品文案。"""
+    total = len(role_briefs)
+    memory = conversation.memory if isinstance(conversation.memory, dict) else {}
+    sources = [
+        appearance_prompt,
+        str(memory.get("pending_video_prompt") or ""),
+    ]
+    if getattr(conversation, "created_at", None):
+        sources.extend(
+            str(item or "")
+            for item in conversation.messages.filter(role=CreationMessage.Role.USER)
+            .order_by("-seq").values_list("text", flat=True)[:6]
+        )
+    shared = next(
+        (gender for source in sources if (gender := _shared_cast_gender_from_text(source, total))),
+        "",
+    )
+    return [_cast_gender_from_text(brief) or shared for brief in role_briefs]
+
+
 
 def _build_single_person_reference_prompt(
     *,
@@ -1454,6 +1541,7 @@ def _build_single_person_reference_prompt(
     context_brief: str,
     cast_index: int,
     cast_total: int,
+    gender_requirement: str = "",
 ) -> tuple[str, str, str]:
     """返回 (prompt, generating_label, model_name)。"""
     is_pet = is_pet_preset(conversation.preset)
@@ -1486,9 +1574,13 @@ def _build_single_person_reference_prompt(
         model_name = "平台生成宠物角色"
     else:
         role_tag = f"第{cast_index}位/共{cast_total}位" if cast_total > 1 else "唯一"
+        gender_instruction = (
+            f"这位角色必须是成年{gender_requirement}，外貌和性别不得改变。"
+            if gender_requirement else ""
+        )
         prompt = (
             f"为短视频生成一张可反复用于锁定身份的真人模特定妆参考图（{role_tag}出镜角色）。"
-            "只出现一位成年人物，正面或轻微三分之四角度，中近景，表情自然，"
+            f"只出现一位成年人物。{gender_instruction}正面或轻微三分之四角度，中近景，表情自然，"
             "五官、发型、肤色、身形和服装细节清晰，简洁中性背景，写实摄影，"
             "不要文字、水印、拼图、多人、遮挡脸部或夸张滤镜。"
             f"{no_product}"
@@ -1497,7 +1589,7 @@ def _build_single_person_reference_prompt(
             prompt += f" 用户指定的人物外观：{appearance_only}。严格保留这些外观要求。"
         if cast_total > 1:
             prompt += (
-                f" 这是多人物视频中的角色{cast_index}，必须与其他角色在性别或发型或年龄段或服装气质上"
+                f" 这是多人物视频中的角色{cast_index}，必须与其他角色在发型、年龄段或服装气质上"
                 "有清晰可辨的差异，便于整片锁脸。"
             )
             label = f"正在生成人物参考（{cast_index}/{cast_total}）"
@@ -1542,6 +1634,7 @@ def submit_generated_person_reference(
     if is_pet_preset(conversation.preset):
         total = 1
     role_briefs = build_cast_role_briefs(total, raw_appearance)
+    gender_requirements = _cast_gender_requirements(conversation, role_briefs, raw_appearance)
 
     memory = dict(conversation.memory or {})
     memory["person_source"] = "platform_generate"
@@ -1566,6 +1659,7 @@ def submit_generated_person_reference(
             context_brief=context_brief,
             cast_index=index,
             cast_total=total,
+            gender_requirement=gender_requirements[index - 1],
         )
         tasks = enqueue_standalone_images(
             team=conversation.team,
@@ -1876,6 +1970,8 @@ VIDEO_MODEL_BY_LABEL = {
     "Seedance 2.0 Fast": "doubao-seedance-2-0-fast-260128",
     "Seedance 2.0 Mini": "doubao-seedance-2-0-mini-260615",
 }
+FAST_VIDEO_MODEL = "doubao-seedance-2-0-fast-260128"
+FAST_VIDEO_MODEL_LABEL = "Seedance 2.0 Fast"
 DEFAULT_VIDEO_MODEL = "doubao-seedance-2-5-260628"
 IMAGE_MODEL_BY_LABEL = {
     "Seedream5.0": "volcano",
@@ -2588,6 +2684,7 @@ _CONTINUE_INTENT_RE = re.compile(
     r"^\s*("
     r"继续|继续做|接着|接着做|往下做|开始吧|开做吧|就这样|就按这个|按这个来|照这个做|直接做|直接来|"
     r"继续完善方案|完善方案|继续创作|继续推进|往下推进|"
+    r"推荐|推荐一下|你来推荐|帮我推荐|按你推荐(?:的)?|你来定|你决定|按你来|你看着办|随便|都行|"
     r"按当前描述继续|按这个继续|没有其他调整[，,]?按当前描述继续|"
     r"使用这[个位只]角色继续创作|使用这[个位只]宠物角色继续创作"
     r")[吧啊呀呢。.！!]*\s*$"
@@ -2873,6 +2970,7 @@ def apply_session_params(conversation, fields, answers: dict) -> bool:
     current = dict(conversation.params or {})
     field_by_key = {str(item.get("key") or ""): item for item in (fields or []) if isinstance(item, dict)}
     changed = False
+    model_user_selected = False
     for key, raw in (answers or {}).items():
         field = field_by_key.get(str(key)) or {}
         if field.get("type") == "asset":
@@ -2889,9 +2987,17 @@ def apply_session_params(conversation, fields, answers: dict) -> bool:
                 value = f"{MAX_VIDEO_DURATION} 秒"
         current[stored] = value
         changed = True
+        if stored == "model":
+            model_user_selected = True
     if changed:
         conversation.params = current
-        conversation.save(update_fields=["params", "updated_at"])
+        update_fields = ["params", "updated_at"]
+        if model_user_selected and conversation.mode == CreationConversation.Mode.VIDEO:
+            memory = dict(conversation.memory or {})
+            memory["video_model_user_selected"] = True
+            conversation.memory = memory
+            update_fields.append("memory")
+        conversation.save(update_fields=update_fields)
         if is_plot_twist_conversation(conversation):
             # 用户在参数卡里改了时长，剧情结构也必须立即跟着切换。
             set_plot_twist_story_depth(conversation, str(current.get("duration") or ""))
@@ -2958,11 +3064,17 @@ def apply_confirm_params(conversation, incoming: dict | None) -> tuple[dict, boo
     )
     if changed:
         conversation.params = current
-        conversation.save(update_fields=["params", "updated_at"])
-        if is_plot_twist_conversation(conversation):
+        update_fields = ["params", "updated_at"]
+        if model_changed:
+            memory = dict(conversation.memory or {})
+            memory["video_model_user_selected"] = True
+            conversation.memory = memory
+            update_fields.append("memory")
+        conversation.save(update_fields=update_fields)
+        if duration_changed and is_plot_twist_conversation(conversation):
             # 在确认卡改时长也要切换故事契约；随后视图会要求重写旧方案。
             set_plot_twist_story_depth(conversation, str(current.get("duration") or ""))
-    needs_rebuild = duration_changed or model_changed
+    needs_rebuild = duration_changed
     if needs_rebuild:
         # 角色和商品素材继续保留；只让 GPT 基于新参数重写受影响的架构与 Prompt。
         set_video_gate_stage(conversation, "strategy", clear_pending_prompt=True)
@@ -3616,17 +3728,20 @@ def resolve_smart_video_duration(
         params["duration"] = f"{duration} 秒"
     selected_model = video_model_name(params)
     switched = False
-    # Seedance 2.5 是当前唯一可稳定承载 16–30 秒单段、以及 31–180 秒分段的模型。
-    # 总时长不在单段能力表里时，后续会拆成多条 <=30 秒的 2.5 任务。
-    if duration > 15 and selected_model != DEFAULT_VIDEO_MODEL:
-        params["model"] = "Seedance 2.5"
+    memory = dict(conversation.memory or {})
+    model_user_selected = bool(memory.get("video_model_user_selected"))
+    # 默认按目标时长选模型：15 秒及以下优先 Fast，超过 15 秒使用 2.5。
+    # 用户在最终确认卡手动换过模型后尊重其选择；只有能力不支持时才自动回退。
+    preferred_model = FAST_VIDEO_MODEL if duration <= 15 else DEFAULT_VIDEO_MODEL
+    preferred_label = FAST_VIDEO_MODEL_LABEL if duration <= 15 else "Seedance 2.5"
+    if not model_user_selected and selected_model != preferred_model:
+        params["model"] = preferred_label
         switched = True
     elif not _model_supports_duration(selected_model, duration):
-        # 智能模式可自动选能完成完整脚本的模型；确认卡会清楚展示变更，用户仍能手动调整。
-        if _model_supports_duration(DEFAULT_VIDEO_MODEL, duration):
-            params["model"] = "Seedance 2.5"
+        if _model_supports_duration(preferred_model, duration):
+            params["model"] = preferred_label
             switched = True
-    memory = dict(conversation.memory or {})
+            memory.pop("video_model_user_selected", None)
     memory["smart_duration_resolved"] = duration
     if switched:
         memory["smart_duration_model_switched"] = True
@@ -4980,6 +5095,7 @@ def iter_creation_agent_events(
         turn_deadline = turn_started + CREATION_AGENT_TURN_TIMEOUT_SECONDS
         extra_body = creation_model_extra_body(model_config, tools)
         truncation_retries = 0
+        progress_tool_retries = 0
         for _round in range(MAX_TOOL_ROUNDS):
             if is_agent_cancel_requested(conversation.id):
                 # 用户终止:干净收束,不落 ERROR,已落库消息保留
@@ -5199,6 +5315,27 @@ def iter_creation_agent_events(
                             turn_has_gate = True
                 break
 
+            # 用户已经明确说「推荐 / 继续 / 做方案」，模型却只回一段说明文字时，
+            # 不能把加载状态正常收起后只留“继续完善方案”。把这一轮文本作废并强制重试工具调用，
+            # 直到真正落下追问卡或视频架构卡。
+            if not calls and allow_plan and context.is_video and not fallback_fields:
+                progress_tool_retries += 1
+                if progress_tool_retries <= MAX_PROGRESS_TOOL_RETRIES:
+                    if said:
+                        messages.append({"role": "assistant", "content": said})
+                    messages.append({"role": "user", "content": PROGRESS_TOOL_RETRY_INSTRUCTION})
+                    continue
+                clear_public_agent_progress(conversation)
+                failure = append_message(
+                    conversation,
+                    role="assistant",
+                    kind=CreationMessage.Kind.ERROR,
+                    text=PROGRESS_TOOL_GIVE_UP_NOTICE,
+                )
+                yield {"type": "message", "message": _message_payload(failure)}
+                yield {"type": "done"}
+                return
+
             # ask_user 自己会落一条可追踪的聊天问题。模型同时吐出的过渡文案不再
             # 另存一条,否则界面会连续出现两遍几乎相同的问题。
             asks_user = bool(fallback_fields) or any(call.get("name") == "ask_user" for call in calls)