修改全能创作

2026-09-15 17:45:28 +08:00
parent 243a3aa54b
commit d4f282bfa9
8 changed files with 1142 additions and 212 deletions
@@ -27,7 +27,18 @@ 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, preset_workflow_guidance
+from .creation_presets import (
+    PLOT_TWIST_PRESET,
+    PLOT_TWIST_STORY_DEPTH_OPTIONS,
+    apply_image_preset_prompt,
+    apply_plot_twist_story_contract,
+    apply_video_preset_prompt,
+    plot_twist_story_depth,
+    plot_twist_story_contract,
+    preset_guidance,
+    preset_workflow_guidance,
+    video_preset_delivery_contract,
+)
 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
@@ -41,8 +52,7 @@ MAX_TOOL_ROUNDS = 8
 MAX_BILLED_GENERATIONS = 1

 # 视频闸门阶段(落在 conversation.memory.stage;resume 靠它)
-# 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
+# clarify → strategy → plan → prompt → 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。",
@@ -51,6 +61,146 @@ _STEP_CONFIRM_LABELS = {
 }


+def is_plot_twist_conversation(conversation: CreationConversation) -> bool:
+    return conversation.mode == CreationConversation.Mode.VIDEO and conversation.preset == PLOT_TWIST_PRESET
+
+
+def active_plot_twist_story_depth(conversation: CreationConversation) -> str:
+    """时长是最终事实来源；用户中途改时长后，故事结构必须随之切换。"""
+    from_duration = plot_twist_story_depth(str((conversation.params or {}).get("duration") or ""))
+    if from_duration and from_duration["value"] != "smart":
+        return str(from_duration["value"])
+    memory = conversation.memory if isinstance(conversation.memory, dict) else {}
+    return str(memory.get("plot_twist_story_depth") or "").strip()
+
+
+def set_plot_twist_story_depth(conversation: CreationConversation, value: str) -> dict | None:
+    """卡片或自然语言选时长后，同步会话顶部参数与最终 Prompt 的故事结构。"""
+    if not is_plot_twist_conversation(conversation):
+        return None
+    depth = plot_twist_story_depth(value)
+    if depth is None:
+        return None
+    memory = dict(conversation.memory or {})
+    params = dict(conversation.params or {})
+    memory["plot_twist_story_depth"] = depth["value"]
+    if depth["duration"]:
+        params["duration"] = depth["duration"]
+    conversation.memory = memory
+    conversation.params = params
+    conversation.save(update_fields=["memory", "params", "updated_at"])
+    return depth
+
+
+def append_plot_twist_story_depth_question(conversation: CreationConversation) -> CreationMessage:
+    return append_message(
+        conversation,
+        role="assistant",
+        kind=CreationMessage.Kind.ELICIT,
+        text="你希望这支剧情带货视频做到什么程度？",
+        payload={
+            "interaction": "plot_twist_story_depth",
+            "fields": [
+                {
+                    "key": "story_depth",
+                    "label": "故事深度选择",
+                    "type": "single",
+                    "required": True,
+                    "options": [
+                        {"value": item["value"], "label": f"{item['label']}｜{item['summary']}"}
+                        for item in PLOT_TWIST_STORY_DEPTH_OPTIONS
+                    ],
+                }
+            ],
+            "submitted": False,
+            "answers": {},
+        },
+    )
+
+
+def _plot_twist_direction_fallback(conversation: CreationConversation) -> list[dict]:
+    """模型遗漏方向卡工具时的平台兜底，不能只留下「我准备了三个方向」的空话。"""
+    refs = [item for item in (conversation.pinned_refs or []) if isinstance(item, dict)]
+    product = next((str(item.get("name") or "").strip() for item in refs if item.get("type") == "product"), "这款商品")
+    return [
+        {
+            "id": "misunderstanding",
+            "title": "误会翻盘",
+            "conflict": "主角把眼前的麻烦误判成无解，情绪不断升级。",
+            "product_role": f"{product} 作为关键证据或解决工具，在最需要时自然出现。",
+            "reversal": "原来问题的答案一直在眼前，误会被当场化解。",
+            "tone": "轻喜剧、反差感强",
+        },
+        {
+            "id": "last_chance",
+            "title": "最后一次机会",
+            "conflict": "主角已经试过常规办法仍然失败，只剩最后一次选择。",
+            "product_role": f"{product} 承担最后一次可验证的尝试，完整展示正常使用过程。",
+            "reversal": "看似失败的局面被扭转，结果恰好印证核心卖点。",
+            "tone": "紧张后释然、转化更强",
+        },
+        {
+            "id": "foreshadowing",
+            "title": "伏笔回收",
+            "conflict": "开头埋下一个不起眼的细节，人物关系或目标因此受阻。",
+            "product_role": f"{product} 先作为日常细节出现，后半段成为推动结局的关键线索。",
+            "reversal": "前面的细节被回收，观众才发现结局早有铺垫。",
+            "tone": "温暖小反转、故事感更强",
+        },
+    ]
+
+
+def _coerce_plot_twist_directions(raw) -> list[dict]:
+    """把模型给的三条方向收成可直接渲染的卡片字段。"""
+    if not isinstance(raw, list):
+        return []
+    directions: list[dict] = []
+    for index, item in enumerate(raw[:3], start=1):
+        if not isinstance(item, dict):
+            continue
+        title = str(item.get("title") or item.get("name") or "").strip()
+        conflict = str(item.get("conflict") or item.get("setup") or "").strip()
+        product_role = str(item.get("product_role") or item.get("product") or "").strip()
+        reversal = str(item.get("reversal") or item.get("twist") or "").strip()
+        tone = str(item.get("tone") or item.get("feel") or "").strip()
+        if not all((title, conflict, product_role, reversal)):
+            continue
+        directions.append({
+            "id": str(item.get("id") or f"direction_{index}").strip() or f"direction_{index}",
+            "title": title,
+            "conflict": conflict,
+            "product_role": product_role,
+            "reversal": reversal,
+            "tone": tone or "剧情带货",
+        })
+    return directions if len(directions) == 3 else []
+
+
+def _append_plot_twist_direction_question(context: "AgentContext", directions: list[dict]) -> CreationMessage:
+    return append_message(
+        context.conversation,
+        role="assistant",
+        kind=CreationMessage.Kind.ELICIT,
+        text="选一个剧情方向，我会按它写成完整方案。",
+        payload={
+            "interaction": "plot_twist_directions",
+            "directions": directions,
+            "fields": [{
+                "key": "story_direction",
+                "label": "三个剧情反转方向",
+                "type": "single",
+                "required": True,
+                "options": [
+                    {"value": item["id"], "label": item["title"]}
+                    for item in directions
+                ],
+            }],
+            "submitted": False,
+            "answers": {},
+        },
+    )
+
+
 def get_video_gate_stage(conversation: CreationConversation) -> str:
    memory = conversation.memory if isinstance(conversation.memory, dict) else {}
    stage = str(memory.get("stage") or "clarify").strip()
@@ -478,6 +628,17 @@ def text_already_guides(text: str) -> bool:
    return bool(_GUIDANCE_MARKER_RE.search((text or "").strip()))


+_NUMERIC_REPLY_INSTRUCTION_RE = re.compile(
+    r"(?:你|请)?\s*(?:直接)?(?:回复|选择|选)\s*(?:数字|编号)?\s*"
+    r"1\s*(?:[/、，,]\s*2)(?:\s*(?:[/、，,]\s*3))?[^。！？!\n]*(?:[。！？!]|$)"
+)
+
+
+def strip_numeric_reply_instruction(text: str) -> str:
+    """选择器会把 1/2/3 直接做成按钮，正文不再要求用户手输数字。"""
+    return _NUMERIC_REPLY_INSTRUCTION_RE.sub("", text or "").strip()
+
+
 def default_reply_hint(
    conversation: CreationConversation | None = None,
    *,
@@ -776,7 +937,7 @@ VIDEO_MODELS = ["Seedance 2.5", "Seedance 2.0", "Seedance 2.0 Fast", "Seedance 2
 IMAGE_MODELS = ["Seedream5.0", "YQ image2"]
 RATIOS = ["16:9", "9:16", "4:3", "3:4", "1:1"]
 RESOLUTIONS = ["480p", "720p", "1080p"]
-VIDEO_DURATIONS = ["智能时长", "4 秒", "5 秒", "6 秒", "8 秒", "10 秒", "12 秒", "15 秒", "30 秒"]
+VIDEO_DURATIONS = ["智能时长", "4 秒", "5 秒", "6 秒", "8 秒", "10 秒", "12 秒", "15 秒", "30 秒", "60 秒"]
 IMAGE_COUNTS = ["1 张", "2 张", "4 张", "8 张"]
 SESSION_PARAM_KEYS = ("duration", "ratio", "resolution", "video_model", "count")
 _PARAM_TO_STORED = {"video_model": "model"}
@@ -865,6 +1026,9 @@ def apply_session_params(conversation, fields, answers: dict) -> bool:
    if changed:
        conversation.params = current
        conversation.save(update_fields=["params", "updated_at"])
+        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 ""))
    return changed


@@ -911,6 +1075,9 @@ 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):
+            # 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 ""))
    return snapshot_session_params(conversation), duration_changed


@@ -993,6 +1160,41 @@ def tool_schemas(context: AgentContext, *, allow_plan: bool = True) -> list[dict
        # 换商品/改参数仍由 wanted_asset_pick / wanted_param_keys 兜底追问。
        return [t for t in tools if t.get("function", {}).get("name") == "search_library"]
    if context.is_video:
+        if is_plot_twist_conversation(context.conversation) and active_plot_twist_story_depth(context.conversation):
+            tools.append({
+                "type": "function",
+                "function": {
+                    "name": "present_story_directions",
+                    "description": (
+                        "剧情反转带货预设在用户选定故事深度后，先调用此工具展示 3 个可点击的剧情方向。"
+                        "每条都必须有不同的冲突、商品承担的实际作用、反转和情绪；不能只说‘我准备了三个方向’。"
+                        "用户点击其一后才可 write_strategy；本工具调用后必须停下来等待选择。"
+                    ),
+                    "parameters": {
+                        "type": "object",
+                        "properties": {
+                            "directions": {
+                                "type": "array",
+                                "minItems": 3,
+                                "maxItems": 3,
+                                "items": {
+                                    "type": "object",
+                                    "properties": {
+                                        "id": {"type": "string"},
+                                        "title": {"type": "string", "description": "短标题，不超过 10 个字"},
+                                        "conflict": {"type": "string", "description": "人物处境与开场冲突"},
+                                        "product_role": {"type": "string", "description": "商品如何自然推进剧情"},
+                                        "reversal": {"type": "string", "description": "最终反转如何回收"},
+                                        "tone": {"type": "string", "description": "情绪与转化倾向"},
+                                    },
+                                    "required": ["title", "conflict", "product_role", "reversal", "tone"],
+                                },
+                            },
+                        },
+                        "required": ["directions"],
+                    },
+                },
+            })
        tools.append({
            "type": "function",
            "function": {
@@ -1233,7 +1435,7 @@ def video_duration(params: dict) -> int:
    digits = "".join(ch for ch in raw if ch.isdigit())
    if not digits:
        return SMART_DURATION
-    return max(4, min(int(digits), 30))
+    return max(4, min(int(digits), 60))


 def _image_count(params: dict, raw) -> int:
@@ -1306,8 +1508,14 @@ def _video_submit_params(context: AgentContext, prompt: str) -> tuple[dict, list
    它已经排好 角色 → 场景 → 商品 的顺序,那正是出片模型 @图N 的语义依据。"""
    params = context.conversation.params or {}
    resolved = resolve_refs(context.team, context.conversation.pinned_refs or [])
+    prompt = apply_video_preset_prompt(context.conversation.preset, prompt)
    prompt = apply_product_voice_visual_guard(context.conversation, prompt)
    prompt = apply_product_reality_guard(prompt)
+    prompt = apply_plot_twist_story_contract(
+        context.conversation.preset,
+        active_plot_twist_story_depth(context.conversation),
+        prompt,
+    )
    submit = {
        "prompt": prompt,
        "feature": "omni_create",
@@ -1562,7 +1770,9 @@ def build_system_prompt(context: AgentContext, *, allow_plan: bool = True, has_c
        "",
        "【每轮必须给引导】",
        "- 每一轮回复结束时,用户必须知道下一步怎么做:要么调用 ask_user 弹出可选项,"
-        "要么在文字里明确告诉用户该回复什么(例如「请回复：1… 2…」或「直接说想改的卖点」)。",
+        "要么在文字里明确告诉用户该回复什么(例如「直接说想改的卖点」)。",
+        "- 当你给出 2–3 个方向时,必须调用 ask_user 生成可点击选项；正文只解释各方向，"
+        "绝不要求用户回复数字、编号或 1/2/3。",
        "- 禁止只丢一段解释/分析就结束、让用户不知道该回什么。",
        "- 用户已经给出可直接执行的图片需求(主体、场景或氛围已足够)时,直接调用 generate_image 进入确认卡;"
        "不要只描述你准备怎么拍,再让用户继续补一句。",
@@ -1601,8 +1811,8 @@ 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 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。"
+            "- 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 → "
+            "③用户确认后 write_plan 停等确认 → ④方案确认后展示完整出片指令并停等确认 → ⑤再显示积分确认卡。"
        )
        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;"
@@ -1657,12 +1867,36 @@ def build_system_prompt(context: AgentContext, *, allow_plan: bool = True, has_c
        # 只给名字模型只能靠猜;把这个预设的拍法约束一起给它
        guidance = preset_guidance(conversation.preset)
        lines.append(f"\n【创作预设】{conversation.preset}")
+        if context.is_video:
+            lines.append(
+                "预设必须贯穿本次对话：据此判断该补问哪些必要事实和素材、给什么创意选择、"
+                "如何写脚本/分镜、商品如何出现、生成前检查什么；不能把预设当成最后才附加的一行风格词。"
+            )
        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。")
+        if is_plot_twist_conversation(conversation):
+            depth = active_plot_twist_story_depth(conversation)
+            if not depth:
+                lines.append("剧情反转带货尚未选择故事深度：必须先调用 ask_user 让用户选 15 秒、30 秒、60 秒或智能推荐；此时禁止给剧情方向、策略或方案。")
+            else:
+                lines.append(plot_twist_story_contract(depth))
+                memory = conversation.memory if isinstance(conversation.memory, dict) else {}
+                selected_direction = str(memory.get("plot_twist_story_direction") or "").strip()
+                if selected_direction:
+                    lines.append(f"用户已选剧情方向：{selected_direction}。直接围绕此方向写策略，不要重发方向卡。")
+                else:
+                    lines.append(
+                        "【强制下一步】现在必须调用 present_story_directions，直接展示 3 张完整剧情方向卡。"
+                        "每张卡写清冲突、商品如何推进剧情、反转和情绪；不得只在文字中说‘我准备了三个方向’，"
+                        "不得先 write_strategy / write_plan，也不要让用户输入编号。"
+                    )
        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}")
+        delivery_contract = video_preset_delivery_contract(conversation.preset) if context.is_video else ""
+        if delivery_contract:
+            lines.append(f"【当前预设必须贯穿脚本与出片】{delivery_contract}")

    resolved = resolve_refs(context.team, conversation.pinned_refs or [])
    if resolved.facts:
@@ -2039,6 +2273,20 @@ def iter_creation_agent_events(
            yield {"type": "done"}
            return

+        # 剧情反转预设必须先选故事深度。平台直接落选择卡，不能把这一步交给模型猜，
+        # 否则默认 15 秒会吞掉 30/60 秒该有的人物关系和冲突发展。
+        if is_plot_twist_conversation(conversation) and not active_plot_twist_story_depth(conversation):
+            explicit_depth = plot_twist_story_depth(text)
+            if explicit_depth is not None and explicit_depth["value"] != "smart":
+                set_plot_twist_story_depth(conversation, explicit_depth["value"])
+            else:
+                question = append_plot_twist_story_depth_question(conversation)
+                conversation.agent_status = CreationConversation.AgentStatus.AWAITING_USER
+                conversation.save(update_fields=["agent_status", "updated_at"])
+                yield {"type": "message", "message": _message_payload(question)}
+                yield {"type": "done"}
+                return
+
        model_config = _prefer_vision_text_model(model_config, conversation.team, conversation.pinned_refs or [])
        context.model_config = model_config

@@ -2095,7 +2343,7 @@ def iter_creation_agent_events(
                elif kind == "tool_call":
                    _merge_tool_call_deltas(tool_buffer, chunk.get("tool_calls"))

-            said = "".join(text_buffer).strip()
+            said = strip_numeric_reply_instruction("".join(text_buffer))
            calls = [tool_buffer[i] for i in sorted(tool_buffer) if tool_buffer[i].get("name")]
            fallback_fields = None
            allow_pick = False
@@ -2131,6 +2379,28 @@ def iter_creation_agent_events(
                            "asset_types": ["product"],
                        }]

+            # 方向卡是剧情反转预设的固定入口。模型偶尔只会说「我准备了三个方向」而忘了调工具，
+            # 此处直接补上可点击卡，不能让用户面对一段空话再自己追问。
+            memory = conversation.memory if isinstance(conversation.memory, dict) else {}
+            needs_plot_twist_directions = (
+                not calls
+                and allow_plan
+                and is_plot_twist_conversation(conversation)
+                and bool(active_plot_twist_story_depth(conversation))
+                and not str(memory.get("plot_twist_story_direction") or "").strip()
+            )
+            if needs_plot_twist_directions:
+                result, _stop = _dispatch_tool(
+                    context,
+                    "present_story_directions",
+                    {"directions": _plot_twist_direction_fallback(conversation)},
+                )
+                for event in result.get("_events", []):
+                    yield event
+                    if event.get("type") == "message":
+                        turn_has_gate = True
+                break
+
            # 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)
@@ -2352,6 +2622,7 @@ def run_creation_agent_turn(

 _TOOL_LABELS = {
    "ask_user": "向你确认",
+    "present_story_directions": "整理剧情方向",
    "search_library": "查找素材",
    "generate_image": "生成图片",
    "write_strategy": "梳理创作策略",
@@ -2476,6 +2747,23 @@ def _dispatch_tool(
            "_events": [{"type": "message", "message": _message_payload(message)}],
        }, True

+    if name == "present_story_directions":
+        if not is_plot_twist_conversation(context.conversation):
+            return {"payload": {"error": "当前会话不是剧情反转带货预设"}}, False
+        directions = _coerce_plot_twist_directions(args.get("directions"))
+        if not directions:
+            return {
+                "payload": {
+                    "error": "必须提供恰好 3 个完整剧情方向，每个都要有标题、冲突、商品作用、反转和情绪。"
+                }
+            }, False
+        message = _append_plot_twist_direction_question(context, directions)
+        set_video_gate_stage(context.conversation, "clarify")
+        return {
+            "payload": {"presented": True, "count": 3},
+            "_events": [{"type": "message", "message": _message_payload(message)}],
+        }, True
+
    if name == "search_library":
        return {"payload": _run_search_library(context, args)}, False

@@ -2515,8 +2803,14 @@ def _dispatch_tool(
        video_prompt = str(args.get("video_prompt") or "").strip()
        if not video_prompt:
            return {"payload": {"error": "video_prompt 不能为空,请把完整出片指令写进去"}}, False
+        video_prompt = apply_video_preset_prompt(context.conversation.preset, video_prompt)
        video_prompt = apply_product_voice_visual_guard(context.conversation, video_prompt)
        video_prompt = apply_product_reality_guard(video_prompt)
+        video_prompt = apply_plot_twist_story_contract(
+            context.conversation.preset,
+            active_plot_twist_story_depth(context.conversation),
+            video_prompt,
+        )
        card = _coerce_plan_card_args(args if isinstance(args, dict) else {})
        if not card["usp"] or not card["points"]:
            return {
@@ -2548,7 +2842,7 @@ def _dispatch_tool(
            kind=CreationMessage.Kind.PLAN, payload=plan_payload,
        )
        events.append({"type": "message", "message": _message_payload(plan)})
-        # 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。
+        # video_prompt 先作为方案产物存档；用户确认方案后会展示指令文件供确认。
        set_video_gate_stage(
            context.conversation, "plan", pending_video_prompt=video_prompt
        )
@@ -2565,18 +2859,23 @@ def _dispatch_tool(
            video_prompt = get_pending_video_prompt(context.conversation)
        if not video_prompt:
            return {"payload": {"error": "video_prompt 不能为空,请把完整出片指令写进去"}}, False
+        video_prompt = apply_video_preset_prompt(context.conversation.preset, video_prompt)
        video_prompt = apply_product_voice_visual_guard(context.conversation, video_prompt)
        video_prompt = apply_product_reality_guard(video_prompt)
-        set_video_gate_stage(context.conversation, "prompt", pending_video_prompt=video_prompt)
-        confirm = emit_final_confirm_gate(context.conversation, context=context)
-        if confirm is None:
-            return {"payload": {"error": "无法准备出片确认"}}, False
+        video_prompt = apply_plot_twist_story_contract(
+            context.conversation.preset,
+            active_plot_twist_story_depth(context.conversation),
+            video_prompt,
+        )
+        prompt_messages = emit_prompt_gate(context.conversation, video_prompt)
+        if not prompt_messages:
+            return {"payload": {"error": "无法准备出片指令"}}, False
        events = [
-            {"type": "message", "message": _message_payload(confirm)},
-            {"type": "credits", "estimated": int((confirm.payload or {}).get("estimated_credits") or 0)},
+            {"type": "message", "message": _message_payload(message)}
+            for message in prompt_messages
        ]
        return {
-            "payload": {"awaiting_confirmation": True, "prepared": True},
+            "payload": {"awaiting_step": "prompt", "prepared": True},
            "_events": events,
        }, True