"""全能创作 · Agent 编排循环(契约 §3/§4)。 和 script_agent.py 的根本区别:那边是「单次结构化出稿」,模型只会写脚本; 这边是**真 function calling 循环** —— 模型自己决定这一轮该反问用户、该查素材, 还是该出图/出片。 铁律(踩过就回不来的三条): 1. **视频 5–10 分钟,绝不在 SSE 里等。** 生成工具立刻返回 task_id,落一条 `generating` 消息,发 `task` 事件,收流。前端轮询完成后原地换成 `result`。 2. **`ask_user` 一旦被调用就中断循环。** 反问的意义是等人回答,继续跑下去 等于自问自答。 3. **一条用户消息最多计费生成一次。** 对话式会放大调用量,一句「多做几版」 能烧掉一堆积分。 SSE 事件见契约 §3。 """ from __future__ import annotations import json import logging import re from collections.abc import Iterator from dataclasses import dataclass from django.core.serializers.json import DjangoJSONEncoder from django.db import transaction from .creation import append_message, pin_refs from .creation_presets import preset_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 logger = logging.getLogger(__name__) # 单条用户消息的循环上限。8 轮足够「查素材 → 反问 → 写方案 → 出图」, # 再多基本是模型在原地打转。 MAX_TOOL_ROUNDS = 8 # 单条用户消息最多触发一次计费生成(契约 §4) MAX_BILLED_GENERATIONS = 1 # 记忆压缩(契约 §5):超过这么多条消息就把最老的一批压成一段摘要, # 只保留最近 KEEP_RECENT_MESSAGES 条原文。 COMPRESS_AFTER_MESSAGES = 24 KEEP_RECENT_MESSAGES = 12 # 攒够这么多条没压过的消息才重压一次。没有它的话,过了阈值以后**每一轮都要多花 # 一次模型调用**去重压那么两三句话 —— 长会话的成本会翻倍。 COMPRESS_MIN_BATCH = 8 FIELD_TYPES = ("single", "multi", "text", "asset") # 顶栏下拉里的展示名 → 火山模型名。前端给的是人看的label,submit_free_video 只认真名。 VIDEO_MODEL_BY_LABEL = { "Seedance 2.5": "doubao-seedance-2-5-260628", "Seedance 2.0": "doubao-seedance-2-0-260128", "Seedance 2.0 Fast": "doubao-seedance-2-0-fast-260128", "Seedance 2.0 Mini": "doubao-seedance-2-0-mini-260615", } DEFAULT_VIDEO_MODEL = "doubao-seedance-2-5-260628" IMAGE_MODEL_BY_LABEL = { "Seedream5.0": "volcano", "Seedream-5.0-pro": "volcano", "YQ image2": "gpt-image", "影擎-Image2": "gpt-image", } # 「智能时长」= 交给我们定,取一个口播讲得完又不烧钱的中间值 SMART_DURATION = 15 # 全能创作 video_prompt 写作规范:从专业创作 / 一键成片(口播脚本)抽硬规则, # 适配「整段出片指令」而不是 ScriptDraft JSON。只要求口径接近,不改工具形态。 _OMNI_VIDEO_PROMPT_RULES = """ 【出片脚本写法 · 对齐专业口播】 表现形式默认**口播带货**,像真人在讲刚遇到的一件事,不要详情页朗读或主播念稿。 video_prompt 必须能被出片模型直接执行,按时间轴写秒级分镜,建议结构: - 片头写清:总时长、画幅、整体光线/色调、口播语气(口语、有停顿与立场) - 然后按「0-3s / 3-8s / …」逐段写,每段同时写清这五项(不能省): 1. 景别(大特写/特写/近景/中景/全景;一条片里至少切两次景别) 2. 机位(平视/俯拍/仰拍/过肩/桌面视角) 3. 运镜(手持跟拍/推近/拉远/横摇/环绕/固定 —— 每段至少一个运镜词) 4. 动作(谁、哪只手、对什么、做什么;要连贯可拍,禁止「展示质感」等抽象词) 5. 信息变化(这几秒画面上多了/变了什么) - 人声原文单独写清,必须用「人声(仅音频,由人物口型与配音表达,不出现在画面上):…」,禁止写成「口播:…」(裸写口播会被出片模型烧成字幕);字数贴近会话时长:大约 5.0–5.7 字/秒, 15 秒约 75–85 字;太短撑不满,太长会赶。 - 钩子段画面不要「对着镜头说话」静态开场;前 15 个口播字禁止「大家好/今天分享/给你们推荐」。 - 全片只围绕一个具体情境推进一个主卖点;卖点要有看得见的证据(质地/前后变化/用法结果)。 - CTA 像跟朋友说话;禁止小黄车/立即购买/闭眼入等平台指令腔。 - **不要写字幕/花字/标题贴片/弹幕/角标/水印/购物浮层**,也不要写「无字幕」 (否定说法也容易把字画上屏)。口播只存在于声音;包装上原有印刷字除外。 - 已 @ 的角色/商品/场景参考图会自动附上,不要在 prompt 里重描长相;以图锁定性别年龄服装外形。 - 用户说「商品说话 / 商品自述 / 商品拟人」时,默认采用**无脸拟人**:商品声音是画外角色声, 商品本体不做口型、不新增卡通五官;性格靠整体倾斜、转向、弹跳、进退、镜头和音效表达。 只有用户明确要求可见的卡通五官时才例外。 - 同一场戏保持地点、光线、服装连续;要换环境就明确写下一时间段切换。 """ _PRODUCT_VOICE_HINT_RE = re.compile( r"(商品|产品|包装|瓶|机身).{0,8}(说话|开口|自述|拟人)" r"|拟人.{0,8}(商品|产品|包装|瓶|机身)" ) _VISIBLE_PRODUCT_FACE_RE = re.compile( r"(商品|产品|包装|瓶|机身).{0,12}(卡通)?(五官|眼睛|眼珠|嘴巴|嘴|口型)" r"|(卡通)?(五官|眼睛|眼珠|嘴巴|嘴|口型).{0,12}(商品|产品|包装|瓶|机身)" r"|(给|让).{0,12}(长出|加上|出现).{0,6}(卡通)?(五官|眼睛|眼珠|嘴巴|嘴|口型)" ) PRODUCT_VOICE_VISUAL_GUARD = ( "【商品角色表现·最高优先级】商品始终是参考图里的真实物件。包装正面、瓶身、机身和全部可见表面" "保持原有设计，只保留参考图本来就有的标签、图案与结构。拟人感完全通过整个商品轻微倾斜、" "转向、弹跳、进退，配合镜头、光影、环境反应和音效表达。商品台词采用画外角色声，声源不在" "画面内，商品保持完整物件形态；场景里的其他物件也保持真实原貌。整体采用精致实拍广告质感" "与克制幽默。前文若有改造商品外观的动作描述，统一改成商品整体运动的对应表达。" ) def apply_product_voice_visual_guard(conversation: CreationConversation, prompt: str) -> str: """商品拟人默认不长脸;用户明确要求可见卡通五官时尊重其创作选择。""" base = str(prompt or "").strip() recent_user_text = " ".join( conversation.messages.filter( role=CreationMessage.Role.USER, kind=CreationMessage.Kind.TEXT, ).order_by("-seq").values_list("text", flat=True)[:12] ) if _VISIBLE_PRODUCT_FACE_RE.search(recent_user_text): return base needs_guard = ( conversation.preset == "商品拟人广告" or bool(_PRODUCT_VOICE_HINT_RE.search(recent_user_text)) or bool(_PRODUCT_VOICE_HINT_RE.search(base)) ) if not needs_guard or PRODUCT_VOICE_VISUAL_GUARD in base: return base return f"{base}\n{PRODUCT_VOICE_VISUAL_GUARD}".strip() class AgentError(Exception): """Agent 循环里的业务错误,已经是可以直接给用户看的中文。""" @dataclass class AgentContext: conversation: CreationConversation user: object model_config: ModelConfig generations_used: int = 0 @property def team(self): return self.conversation.team @property def is_video(self) -> bool: return self.conversation.mode == CreationConversation.Mode.VIDEO _ASSET_PICK_LABEL = { "product": "换成哪个商品？", "character": "换成哪个角色？", "model": "换成哪个模特？", "scene": "换成哪个场景？", } _ASSET_PICK_PATTERNS = ( ("product", re.compile(r"(改|换|修改|更换|重新选|选(一个|个)?|挑).{0,8}商品")), ("character", re.compile(r"(改|换|修改|更换|重新选|选(一个|个)?).{0,8}(角色|人物)")), ("model", re.compile(r"(改|换|修改|更换|重新选|选(一个|个)?).{0,8}模特")), ("scene", re.compile(r"(改|换|修改|更换|重新选|选(一个|个)?).{0,8}场景")), ) _CHITCHAT_RE = re.compile( r"^\s*(" r"hi|hello|hey|yo|hola|" r"你好呀?|您好|嗨|哈喽|嘿|" r"在吗|在不在|有人吗|" r"早+|早安|早上好|午安|晚安|" r"嗯+|哦+|噢+|额+|呃+|" r"好的?|行|可以|ok(?:ay)?|thanks?|thank\s*you|谢谢了?|感谢|" r"收到|知道了|明白了|了解" r")[\s!！.。~～？?…]*$", re.IGNORECASE, ) _GREETING_RE = re.compile( r"^\s*(" r"hi|hello|hey|yo|hola|" r"你好呀?|您好|嗨|哈喽|嘿|" r"在吗|在不在|有人吗|" r"早+|早安|早上好|午安|晚安" r")[\s!！.。~～？?…]*$", re.IGNORECASE, ) def is_greeting(user_text: str) -> bool: """单纯打招呼要立刻回应，不能为了分析引用素材去等模型。""" return bool(_GREETING_RE.match((user_text or "").strip())) def is_pure_chitchat(user_text: str) -> bool: """纯打招呼 / 应答,没有创作意图。这类消息绝不能触发 write_strategy / write_plan。""" text = (user_text or "").strip() if not text or len(text) > 24: return False return bool(_CHITCHAT_RE.match(text)) # 明确要做内容才算出方案/出图。闲聊、吐槽、问功能都不算。 _CREATIVE_INTENT_RE = re.compile( r"(" r"帮我做|帮我拍|帮我出|帮我写|帮我改|帮我生成|" r"创作[一两]?[条个张]?|创作(?:一条|个|短)?|" r"做[一两]?[条个张](?:视频|片|图|广告)?|拍[一两]?[条个](?:视频|片|广告)?|" r"出[一两]?[条个张](?:视频|片|图|广告)?|出片|出图|出方案|写方案|改方案|重写方案|重新写|" r"生成(?:一下|一张|几张|一条)?(?:视频|图|片|广告|脚本)?|做条|做个片|短视频|带货视频|短广告|广告片|带货片|" r"换卖点|改卖点|换剧情|改剧情|重做|重新出|按这个出|确认出片|" r"分镜|脚本|口播稿|storyboard|拟人|" r"(改|换|修改|更换).{0,8}(时长|秒数|比例|尺寸|画幅|分辨率|清晰度|模型)|" r"改成\s*\d+\s*秒|改成\s*\d+\s*[:：]\s*\d+|改成.{0,8}(竖屏|横屏|比例|分辨率)" r")", re.IGNORECASE, ) # 创作 brief 常见搭配:动词 + 成片名词(「创作一条短广告」) _CREATIVE_VERB_RE = re.compile(r"创作|制作|拍摄|生成|做|拍|出|写|弄|来一条|来个") _CREATIVE_NOUN_RE = re.compile(r"视频|短片|短视频|广告|带货|出片|分镜|脚本|口播|主图|海报|成片") def has_creative_intent(user_text: str, refs: list | None = None) -> bool: """用户本轮是否明确要做片/出图/改方案。 没意图时不把 write_strategy / write_plan / generate_image 塞进 tools, 避免模型把闲聊「整理」成方案卡。仅 @ 素材不算意图。 """ text = (user_text or "").strip() if not text: return False if is_pure_chitchat(text): return False if _CREATIVE_INTENT_RE.search(text): return True # 「把商品…创作一条…短广告」这类 brief:同时有创作动词和成片名词 if len(text) >= 8 and _CREATIVE_VERB_RE.search(text) and _CREATIVE_NOUN_RE.search(text): return True return False _CONTINUE_INTENT_RE = re.compile( r"^\s*(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)[吧啊呀呢。.!！]*\s*$" ) def is_continue_intent(user_text: str) -> bool: """已有创作上下文时,这些短句是在授权继续,不是闲聊。""" return bool(_CONTINUE_INTENT_RE.match((user_text or "").strip())) def session_has_creative_context(conversation: CreationConversation) -> bool: """会话里是否已有可继续的创作进度(钉了素材 / 出过策略或方案)。 这个信号只用于判断一句短回复是否承接创作,不拿来主动追问流程确认。 """ if conversation.pinned_refs: return True if conversation.messages.filter( kind__in=( CreationMessage.Kind.STRATEGY, CreationMessage.Kind.PLAN, CreationMessage.Kind.PROMPT_FILE, CreationMessage.Kind.CONFIRM, ) ).exists(): return True # 追问发生在策略卡之前时,会话里可能还没有任何结构化产物。原始 brief 本身 # 就是创作上下文;不认它的话,刷新后一句「继续」会被当成空闲聊天。 recent_user_texts = conversation.messages.filter( role=CreationMessage.Role.USER, kind=CreationMessage.Kind.TEXT, ).order_by("-seq").values_list("text", flat=True)[:8] return any(has_creative_intent(item) for item in recent_user_texts) def wanted_asset_pick(user_text: str, refs: list | None) -> str | None: """用户说「改商品」却没点名时,启动对应素材的自然确认。""" if refs: return None text = user_text or "" for type_, pattern in _ASSET_PICK_PATTERNS: if pattern.search(text): return type_ return None def requested_asset_card_from_context( conversation: CreationConversation, user_text: str, ) -> str | None: """「发一下卡片我选」没有说类型时,沿用最近一次素材追问的类型。""" text = str(user_text or "") wants_card = re.search( r"(发|打开|展示|看看|看下|给我).{0,10}(卡片|列表|商品库|素材库)" r"|(卡片|列表).{0,10}(选|选择|看看|看下)", text, ) if not wants_card: return None recent = conversation.messages.filter( kind=CreationMessage.Kind.ELICIT ).order_by("-seq")[:8] for message in recent: payload = message.payload or {} fields = payload.get("pending_fields") or payload.get("fields") or [] for field in fields: if not isinstance(field, dict): continue inferred = infer_field_types(field) if len(inferred) == 1 and inferred[0] in TYPE_LABELS: return inferred[0] return None VIDEO_MODELS = ["Seedance 2.5", "Seedance 2.0", "Seedance 2.0 Fast", "Seedance 2.0 Mini"] 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"] 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"} _PARAM_PICK_LABEL = { "duration": "想改成多少秒？", "ratio": "想换成什么画幅？比如 9:16 竖屏或 16:9 横屏。", "resolution": "想换成什么清晰度？直接说 480p、720p 或 1080p 就行。", "video_model": "想换哪个模型？直接告诉我模型名就行。", "count": "这次想出几张？", } def _param_options(key: str, is_video: bool) -> list[str]: if key == "duration": return VIDEO_DURATIONS if key == "ratio": return RATIOS if key == "resolution": return RESOLUTIONS if key == "count": return IMAGE_COUNTS return VIDEO_MODELS if is_video else IMAGE_MODELS def session_param_fields(keys: list[str], is_video: bool) -> list[dict]: fields = [] # 对话式追问一次只问一件事，避免又退化成参数问卷。 for key in keys[:1]: if key not in _PARAM_PICK_LABEL: continue options = [{"value": item, "label": item} for item in _param_options(key, is_video)] fields.append({ "key": key, "label": _PARAM_PICK_LABEL[key], "type": "single", "required": True, "options": options, }) return fields def wanted_param_keys(user_text: str, *, is_video: bool) -> list[str]: """用户说「改时长」「换模型」时弹出参数卡。不和「改模特」抢。""" text = user_text or "" keys: list[str] = [] if re.search(r"(改|换|修改|更换).{0,8}(时长|秒数)|改成\s*\d+\s*秒", text): keys.append("duration" if is_video else "count") if re.search(r"(改|换|修改|更换).{0,8}(比例|尺寸|画幅)", text): keys.append("ratio") if re.search(r"(改|换|修改|更换).{0,8}(分辨率|清晰度)", text): keys.append("resolution") if re.search(r"(改|换|修改|更换).{0,8}模型", text) and "模特" not in text: keys.append("video_model") if re.search(r"(改|换|修改|更换).{0,8}张数", text): keys.append("count") if not keys and re.search(r"(改|换|修改).{0,6}(参数|设置|规格)", text): keys = ["duration", "video_model", "ratio"] if is_video else ["count", "video_model", "ratio"] seen = set() out = [] for key in keys: if key in seen: continue seen.add(key) out.append(key) return out def apply_session_params(conversation, fields, answers: dict) -> bool: """追问卡里选出的时长/模型等写回会话参数。返回是否有改动。""" 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 for key, raw in (answers or {}).items(): field = field_by_key.get(str(key)) or {} if field.get("type") == "asset": continue stored = _PARAM_TO_STORED.get(str(key), str(key)) if stored not in {"model", "ratio", "resolution", "duration", "count"}: continue value = "、".join(raw) if isinstance(raw, list) else str(raw or "").strip() if not value or current.get(stored) == value: continue current[stored] = value changed = True if changed: conversation.params = current conversation.save(update_fields=["params", "updated_at"]) return changed def snapshot_session_params(conversation) -> dict: params = conversation.params or {} return { "model": str(params.get("model") or ""), "resolution": str(params.get("resolution") or ""), "ratio": str(params.get("ratio") or ""), "duration": str(params.get("duration") or ""), "count": str(params.get("count") or params.get("duration") or ""), } def confirm_param_options(is_video: bool) -> dict: return { "model": VIDEO_MODELS if is_video else IMAGE_MODELS, "resolution": RESOLUTIONS if is_video else [], "ratio": RATIOS, "duration": VIDEO_DURATIONS if is_video else [], "count": IMAGE_COUNTS if not is_video else [], } def apply_confirm_params(conversation, incoming: dict | None) -> tuple[dict, bool]: """确认卡上改的参数写回会话。返回 (最新 params, 视频时长是否变了)。""" current = dict(conversation.params or {}) old_duration = str(current.get("duration") or "") changed = False for key, raw in (incoming or {}).items(): if key not in {"model", "ratio", "resolution", "duration", "count"}: continue value = str(raw or "").strip() if not value or current.get(key) == value: continue current[key] = value changed = True duration_changed = ( conversation.mode == CreationConversation.Mode.VIDEO and str(current.get("duration") or "") != old_duration and bool(str(current.get("duration") or "")) and bool(old_duration) ) if changed: conversation.params = current conversation.save(update_fields=["params", "updated_at"]) return snapshot_session_params(conversation), duration_changed # ---------------------------------------------------------------- 工具 schema def tool_schemas(context: AgentContext, *, allow_plan: bool = True) -> list[dict]: """给模型看的工具清单。图片会话不暴露 generate_video,反之亦然 —— 会话 mode 是定死的(契约 §0),把不该用的工具摆出来只会诱导模型走错路。 allow_plan=False 时隐藏 write_strategy / write_plan / generate_image,闲聊用不出来。""" tools = [ { "type": "function", "function": { "name": "ask_user", "description": ( "缺少必要信息时向用户追问。" "只在信息**确实缺失且无法合理推断**时用;能自己定的就自己定,别把用户当填表机器。" "用户要选/改/换商品、角色、模特、场景时**必须**调这个工具," "type 用 asset 并填 asset_types。系统会以类Agent自然对话轻量询问是否需要发送商品列表(同时支持直接输入名字或由你推荐),避免一上来粗暴弹出大卡片打断交流。" "一次只问 1 项,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。" ), "parameters": { "type": "object", "properties": { "fields": { "type": "array", "maxItems": 1, "items": { "type": "object", "properties": { "key": {"type": "string", "description": "英文标识,如 product / duration"}, "label": {"type": "string", "description": "问题原文,中文"}, "type": {"type": "string", "enum": list(FIELD_TYPES)}, "required": {"type": "boolean"}, "options": { "type": "array", "items": { "type": "object", "properties": { "value": {"type": "string"}, "label": {"type": "string"}, }, "required": ["value", "label"], }, }, "asset_types": { "type": "array", "items": {"type": "string", "enum": list(TYPE_LABELS)}, }, "placeholder": {"type": "string"}, }, "required": ["key", "label", "type"], }, } }, "required": ["fields"], }, }, }, { "type": "function", "function": { "name": "search_library", "description": "在团队的商品库/模特库/角色/场景/资产库里找素材。用户说了名字但没 @ 时用它找回来。", "parameters": { "type": "object", "properties": { "query": {"type": "string"}, "types": {"type": "array", "items": {"type": "string", "enum": list(TYPE_LABELS)}}, }, "required": ["query"], }, }, }, ] if not allow_plan: # 闲聊轮次不给 ask_user,避免模型追问「要不要出片」; # 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: tools.append({ "type": "function", "function": { "name": "write_strategy", "description": ( "写「创作策略理解」卡:说清这条片给谁看、他为什么会信、你想让他信什么、整体创作方向。" "四个字段都必须写具体非空文案,禁止空字符串。" "仅当用户明确要做片/出方案时,在 write_plan 之前调一次;打招呼或闲聊不要调。" ), "parameters": { "type": "object", "properties": { "target": {"type": "string", "description": "这条视频给谁看,要具体到人群特征"}, "trust": {"type": "string", "description": "用户为什么相信,靠什么建立可信度"}, "belief": {"type": "string", "description": "希望用户看完相信什么"}, "direction": {"type": "string", "description": "创作方向一句话,说清是什么类型的片"}, }, "required": ["target", "trust", "belief", "direction"], }, }, }) tools.append({ "type": "function", "function": { "name": "write_plan", "description": ( "写「视频最终方案」卡并请用户确认。**仅当用户明确要做片/出方案/改方案时调用**;" "打招呼或闲聊不要调。调完会等用户点确认,确认后平台直接按 video_prompt 出片。" "usp / points / timeline 是给用户看的卡片正文,必须写满具体文案,禁止空着只交 video_prompt。" "video_prompt 按系统里的「出片脚本写法」写成口播秒级分镜(专业创作同口径),不要只写大纲。" "先调 write_strategy 再调它。" ), "parameters": { "type": "object", "properties": { "usp": {"type": "string", "description": "主打卖点,全片只讲这一个核心价值"}, "points": { "type": "array", "maxItems": 3, "items": {"type": "string"}, "description": "核心支撑卖点,最多 3 条", }, "timeline": { "type": "array", "items": { "type": "object", "properties": { "start": {"type": "number"}, "end": {"type": "number"}, "stage": {"type": "string", "description": "Hook / 过桥 / 正文 / CTA"}, "desc": {"type": "string"}, }, "required": ["start", "end", "stage"], }, }, "voice_chars": { "type": "array", "items": {"type": "integer"}, "description": "口播字数区间 [下限, 上限]", }, "video_prompt": { "type": "string", "description": ( "交给出片模型的完整口播带货指令(对齐专业创作口径)。" "必须含:总时长与画幅、整体光线色调、按 0-Ns 分段的秒级分镜" "(每段写清景别/机位/运镜/具体动作/信息变化)、人声(仅音频抬头)原文、一个主卖点与可见证据、口语 CTA。禁止「口播:」字样。" "禁止字幕/花字/贴片及「无字幕」字样;禁止详情页腔与「大家好」开场。" "已 @ 素材会自动作参考图,勿重描长相。" ), }, }, "required": ["usp", "points", "video_prompt"], }, }, }) else: tools.append({ "type": "function", "function": { "name": "generate_image", "description": ( "生成图片。prompt 必须是完整、可独立执行的画面描述(主体/动作/环境/光线/构图/风格)," "不要写成对用户说的话。已 @ 引用的素材会自动作为参考图带上,不用在 prompt 里重复描述它们的外观。" ), "parameters": { "type": "object", "properties": { "prompt": {"type": "string"}, "count": {"type": "integer", "minimum": 1, "maximum": 4}, }, "required": ["prompt"], }, }, }) return tools # ---------------------------------------------------------------- 工具执行 def _coerce_fields(raw) -> list[dict]: """把模型给的 fields 规整成契约 §2 的 Field。脏数据丢弃而不是抛 —— 模型偶尔漏个 type 不该让整条对话崩掉。""" fields: list[dict] = [] for item in (raw or [])[:1]: if not isinstance(item, dict): continue key = str(item.get("key") or "").strip() label = str(item.get("label") or "").strip() type_ = str(item.get("type") or "").strip() if not key or not label or type_ not in FIELD_TYPES: continue field = { "key": key, "label": label, "type": type_, "required": bool(item.get("required", True)), } options = [ {"value": str(o.get("value")), "label": str(o.get("label"))} for o in (item.get("options") or []) if isinstance(o, dict) and o.get("value") and o.get("label") ] if type_ in ("single", "multi"): if not options: continue # 单选/多选没选项 = 废卡,丢掉 field["options"] = options if type_ == "asset": asset_types = [t for t in (item.get("asset_types") or []) if t in TYPE_LABELS] field["asset_types"] = asset_types or list(TYPE_LABELS) if type_ == "text": field["placeholder"] = str(item.get("placeholder") or "") inferred = infer_field_types(field) # 商品/角色这类必须出素材卡,文字单选钉不上参考图 if key in SESSION_PARAM_KEYS: fields.append(field) continue if len(inferred) == 1 and inferred[0] in TYPE_LABELS and type_ != "asset": field["type"] = "asset" field["asset_types"] = inferred field.pop("options", None) field.pop("placeholder", None) fields.append(field) return fields def _normalize_model_label(value: str) -> str: return "".join(ch for ch in str(value or "").lower() if ch.isalnum()) def image_model_name(params: dict) -> str | None: """出图模型 label → 供应商模型名;目录优先,认不出返回 None 让下游用默认。""" from django.db.models import Q from .models import ModelConfig label = str(params.get("model") or "").strip() if not label: return None mapped = IMAGE_MODEL_BY_LABEL.get(label) if mapped: return mapped hit = ( ModelConfig.objects.filter(capability=ModelConfig.Capability.IMAGE, status=ModelConfig.Status.ACTIVE) .filter(Q(display_name=label) | Q(name=label)) .order_by("created_at") .first() ) return hit.name if hit else None def video_model_name(params: dict) -> str: """会话参数里的模型 label → 供应商模型名。 先认历史写死映射,再按 ModelConfig.display_name / name 查目录 —— 后台新加模型不用改代码。 """ from django.db.models import Q from .models import ModelConfig label = str(params.get("model") or "").strip() if not label: return DEFAULT_VIDEO_MODEL mapped = VIDEO_MODEL_BY_LABEL.get(label) if not mapped: norm = _normalize_model_label(label) for k, v in VIDEO_MODEL_BY_LABEL.items(): if _normalize_model_label(k) == norm: mapped = v break if mapped: return mapped hit = ( ModelConfig.objects.filter(capability=ModelConfig.Capability.VIDEO, status=ModelConfig.Status.ACTIVE) .filter(Q(display_name=label) | Q(name=label)) .order_by("created_at") .first() ) if hit is None: hit = ( ModelConfig.objects.filter(capability=ModelConfig.Capability.VIDEO, status=ModelConfig.Status.ACTIVE) .filter(Q(display_name__icontains=label) | Q(name__icontains=label)) .order_by("created_at") .first() ) return hit.name if hit else DEFAULT_VIDEO_MODEL def video_duration(params: dict) -> int: """「15 秒」→ 15;「智能时长」/ 解析不出 → SMART_DURATION。""" raw = str(params.get("duration") or "") digits = "".join(ch for ch in raw if ch.isdigit()) if not digits: return SMART_DURATION return max(4, min(int(digits), 30)) def _image_count(params: dict, raw) -> int: """出图张数:首页选过「N 张」就用它,否则用模型传的 count,默认 1,上限 8。""" label = str((params or {}).get("count") or (params or {}).get("duration") or "") if "张" in label: digits = "".join(ch for ch in label if ch.isdigit()) if digits: raw = digits try: count = int(raw or 1) except (TypeError, ValueError): count = 1 return max(1, min(count, 8)) def _run_search_library(context: AgentContext, args: dict) -> dict: results = search_mentions( context.team, q=str(args.get("query") or "").strip(), types=[t for t in (args.get("types") or []) if t in TYPE_LABELS] or None, limit=5, ) return { "results": [ {"type": r["type"], "id": r["id"], "name": r["name"], "kind": TYPE_LABELS[r["type"]]} for r in results ] } def _run_generate_image(context: AgentContext, args: dict) -> tuple[dict, list]: """提交出图。返回 (给模型看的结果, AITask 列表)。 出图也是异步的(worker 出图 ~30s),所以这里同样只提交不等待 —— 和视频一条路子, 前端拿 task_id 轮询 GET /api/ai/generate-image/?ids=… """ from .services import enqueue_standalone_images prompt = str(args.get("prompt") or "").strip() if not prompt: raise AgentError("生成失败:模型没有给出画面描述") params = context.conversation.params or {} resolved = resolve_refs(context.team, context.conversation.pinned_refs or []) reference_image_ids = [r["asset_id"] for r in resolved.references if r.get("asset_id")] count = _image_count(params, args.get("count")) tasks = enqueue_standalone_images( team=context.team, user=context.user, prompt=prompt, mode="image", count=count, ratio=params.get("ratio") or None, image_model=image_model_name(params) or params.get("model") or None, reference_image_ids=reference_image_ids or None, feature="omni_create", ) context.generations_used += 1 return ( {"submitted": True, "count": len(tasks), "note": "已提交生成,结果稍后回填,不要重复提交"}, list(tasks), ) def _video_submit_params(context: AgentContext, prompt: str) -> tuple[dict, list]: """拼 submit_free_video 的入参。references 直接用 resolve_refs 的产物 —— 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_product_voice_visual_guard(context.conversation, prompt) submit = { "prompt": prompt, "feature": "omni_create", "mode": "universal", "model": video_model_name(params), "aspect_ratio": params.get("ratio") or "9:16", "resolution": params.get("resolution") or "720p", "duration": video_duration(params), "generate_audio": True, "references": resolved.references, } return submit, resolved.references def estimate_video_credits(context: AgentContext) -> int: """确认按钮旁的预计积分。算不出来返回 0,前端就不显示 —— 估价失败绝不能挡住出片(用户仍会在扣费环节看到真实数字)。""" from apps.billing.pricing import quote_video_estimate params = context.conversation.params or {} submit, references = _video_submit_params(context, "") model_config = ModelConfig.objects.filter( name=submit["model"], capability=ModelConfig.Capability.VIDEO ).first() if model_config is None: return 0 try: _tokens, quote = quote_video_estimate( model_config, aspect_ratio=submit["aspect_ratio"], resolution=submit["resolution"], duration=submit["duration"], references=references, team=context.team, ) return int(quote.points) except Exception: # noqa: BLE001 — 估价挂了不该挡住出片 logger.warning("omni create: video estimate failed", exc_info=True) return 0 def estimate_image_credits(context: AgentContext) -> int: """出图确认卡预计积分:挂牌单价(含团队系数)逐张取整后再 × 张数,与 enqueue 逐任务预留同口径。""" from apps.billing.pricing import quote_flat from apps.ai.services import resolve_image_model, get_default_model params = context.conversation.params or {} model_name = image_model_name(params) or str(params.get("model") or "").strip() or None model_config = resolve_image_model(model_name) if model_name else None if model_config is None: model_config = get_default_model(ModelConfig.Capability.IMAGE) if model_config is None: return 0 count = _image_count(params, None) try: per = quote_flat(model_config, units=1, team=context.team) return int(per.points) * count except Exception: # noqa: BLE001 logger.warning("omni create: image estimate failed", exc_info=True) return 0 def submit_confirmed_video(*, conversation: CreationConversation, user, confirm_message: CreationMessage): """用户点了确认 → 直接按方案卡里存好的 video_prompt 出片。 **这里不再跑一轮模型**:方案已经确认过了,再让模型决定一次既费钱又可能它不调工具。 返回 (生成中消息, 错误文案),两者必有其一。 """ from .free_video import submit_free_video payload = confirm_message.payload or {} prompt = str(payload.get("video_prompt") or "").strip() if not prompt: return None, "这条方案没有存下出片指令，请让我重新写一次方案。" context = AgentContext(conversation=conversation, user=user, model_config=None) submit, _references = _video_submit_params(context, prompt) try: task = submit_free_video(team=conversation.team, user=user, params=submit) except ValueError as exc: # 校验类错误(时长/比例/额度),给用户看原文 return None, str(exc) message = append_message( conversation, role="assistant", kind=CreationMessage.Kind.GENERATING, payload={"task_id": str(task.id), "kind": "video", "prompt": prompt}, task=task, ) _remember_artifact(conversation, prompt, "video") return message, "" def submit_confirmed_image(*, conversation: CreationConversation, user, confirm_message: CreationMessage): """用户点了确认 → 按确认卡里存的画面描述出图。同样不跑一轮模型。""" payload = confirm_message.payload or {} prompt = str(payload.get("prompt") or payload.get("image_prompt") or "").strip() if not prompt: return None, "这条方案没有存下出图指令，请让我重新写一次。" context = AgentContext(conversation=conversation, user=user, model_config=None) try: _result, tasks = _run_generate_image(context, {"prompt": prompt}) except AgentError as exc: return None, str(exc) except ValueError as exc: return None, str(exc) message = None for task in tasks: message = append_message( conversation, role="assistant", kind=CreationMessage.Kind.GENERATING, payload={"task_id": str(task.id), "kind": "image", "prompt": prompt}, task=task, ) if message is None: return None, "出图没有提交成功，请再试一次。" _remember_artifact(conversation, prompt, "image") return message, "" # ---------------------------------------------------------------- 提示词 def get_creation_chat_model(requested: ModelConfig | None = None) -> ModelConfig | None: """全能创作编排固定优先 Seed 2.1 Pro;显式传入的可用模型仍尊重用户选择。""" if requested is not None: return resolve_text_model(requested) return get_seed_text_model() or resolve_text_model(None) def _creation_model_sees_images(model_config: ModelConfig | None) -> bool: """对话模型能不能收图。参考图只在能看图时才塞进 chat messages,避免纯文本模型整轮失败。""" if model_config is None: return False if getattr(model_config, "capability", "") == ModelConfig.Capability.VISION: return True name = str(getattr(model_config, "name", "") or "").lower() # 豆包 Seed 2.x / 1.6 文本档都支持图文;vl / vision 后缀同理。 if name.startswith("doubao-seed-") or "vision" in name or name.endswith("-vl") or "-vl-" in name: return True metadata = model_config.metadata if isinstance(getattr(model_config, "metadata", None), dict) else {} capabilities = metadata.get("capabilities") if isinstance(metadata.get("capabilities"), dict) else {} features = {str(item) for item in capabilities.get("features") or []} return bool({"vision", "image_input", "multimodal"} & features) def _prefer_vision_text_model(current: ModelConfig | None, team, refs: list | None) -> ModelConfig | None: """有参考图时,尽量换成能看图的文本模型(豆包 Seed 等),否则聊天侧完全看不见男女。""" if current is not None and _creation_model_sees_images(current): return current if not _ref_image_urls(team, refs): return current qs = ( ModelConfig.objects.select_related("provider") .filter( capability=ModelConfig.Capability.TEXT, status=ModelConfig.Status.ACTIVE, provider__status="active", ) .order_by("created_at") ) for candidate in qs: if _creation_model_sees_images(candidate): return candidate return get_default_model(ModelConfig.Capability.VISION) or current def _ref_image_urls(team, refs: list | None) -> list[str]: resolved = resolve_refs(team, refs or []) urls: list[str] = [] seen: set[str] = set() for item in resolved.references: url = str(item.get("url") or "").strip() if not url or url in seen: continue seen.add(url) urls.append(url) return urls[:6] def _attach_ref_images(messages: list[dict], image_urls: list[str]) -> list[dict]: """把锁定素材图挂到最近一条 user 消息上(OpenAI image_url 格式)。""" if not image_urls or not messages: return messages note = ( f"【参考图·请亲眼看】下面 {len(image_urls)} 张是用户锁定的素材。" "人物的性别、年龄段、发型、服装必须以图为准;看不清再问用户,禁止凭文件名猜测性别。" ) out = [dict(message) for message in messages] index = next((i for i in range(len(out) - 1, -1, -1) if out[i].get("role") == "user"), None) if index is None: content = [{"type": "text", "text": note}] content.extend({"type": "image_url", "image_url": {"url": url}} for url in image_urls) out.append({"role": "user", "content": content}) return out last = dict(out[index]) raw = last.get("content") if isinstance(raw, list): content = list(raw) text_bits = [str(item.get("text") or "") for item in content if isinstance(item, dict) and item.get("type") == "text"] if not any(note[:8] in bit for bit in text_bits): content.append({"type": "text", "text": note}) existing = { (item.get("image_url") or {}).get("url") for item in content if isinstance(item, dict) and item.get("type") == "image_url" } content.extend( {"type": "image_url", "image_url": {"url": url}} for url in image_urls if url not in existing ) else: content = [{"type": "text", "text": f"{raw or ''}\n\n{note}".strip()}] content.extend({"type": "image_url", "image_url": {"url": url}} for url in image_urls) last["content"] = content out[index] = last return out def build_system_prompt(context: AgentContext, *, allow_plan: bool = True, has_context: bool = False) -> str: conversation = context.conversation params = conversation.params or {} kind = "视频" if context.is_video else "图片" lines = [ "你是影擎「全能创作」的创作 agent,帮电商商家做短视频和商品图。", f"本次会话产出的是**{kind}**,这一点在整个会话里不会改变 —— 用户要另一种就请他新开一个创作。", "", "【怎么说话】", "- 说人话,像个懂创作、会一起把事做完的同事;自然、简短、有判断,不要写成客服话术或需求确认清单。", "- 禁止用「好的」「收到」「明白了」「有需要再说」「我将为你」起手;这些句子没有信息量,也很像机器人。", "- 不要逐字复述用户刚说过的话。需要承接时,只说你的判断或下一步,例如「这个方向能做,我先把反转落在商品登场上。」", "- 一次只推进一步,但不要把能直接做的事停在寒暄、确认或客套话上。", "- 缺信息时一次只问一个真正影响结果的问题,不要把对话做成问卷。", "- 缺少商品、模特、角色、场景等素材时：**绝不要直接弹出大块素材选择卡**打断对话。像懂创作、懂电商的专业伙伴一样先自然询问用户想推什么商品/用哪个角色，询问是否需要把商品列表发给他选，同时说明也可以直接输入商品名或由你推荐。", "- 只有当用户明确说要发列表（如「发给我」「发列表」「给我看看」「我来选」）时，才展示可视化卡片。", "- 用户直接打字输入商品名时，直接采纳该商品并继续推进创作方案，不要强迫用户去卡片里点选。", "- 调性、受众、文案、时长等其他信息才用聊天追问,问题末尾顺手告诉用户下一句怎么回。", "- 用户刚回答过你的问题时,直接沿着答案继续;不要复述答案,也不要额外回一句「收到」。", "- 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。用户已经提出创作需求且信息够了,就直接写策略/方案;真正扣积分前另有确认卡。", "- 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。", "- 会话里**已有**商品或方向时:针对用户这句话本身自然回应;别用「要现在生成还是先调细节」把工作又抛回给用户。", "- 没有明确创作指令时不要自己写策略卡/方案卡,也不要主动追问流程确认;等用户给出具体想法或修改。", "", "【什么时候反问】", "- 只有信息**确实缺失且无法合理推断**时才调用 ask_user;能自己定的就自己定。", "- ask_user 一次只问一件事。type=asset 会以类Agent自然追问轻量询问是否发送列表或由你推荐；其他类型显示普通聊天问题，让用户直接输入。", "- 商品是谁、给谁看、什么调性 —— 这些缺了会直接影响成片,值得问。", "- 让用户选商品/角色/模特/场景时,ask_user 必须用 type=asset 并填 asset_types。", "- 用户说「改商品」「换角色」却没点名是哪个:立刻 ask_user 启动素材确认,不要强行出大卡片。", "- 用户说改时长/模型/比例/分辨率但没给新值:立刻 ask_user,type=single 提供可识别的候选值;聊天里只显示问题,用户直接输入。回答后旧方案作废,必须按新参数重新 write_plan。", "- 光线、构图、镜头这些专业判断是你的活,不要反过来问用户。", ] if context.is_video: lines.extend(["", _OMNI_VIDEO_PROMPT_RULES.strip()]) # 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) except Exception: # noqa: BLE001 dur = SMART_DURATION lo = max(1, int(dur * 5.0)) 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("- 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。真要写方案时必须先 write_strategy 再 write_plan,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」。") else: lines.extend([ "", "【出图】", "- 决定出图时必须调用 generate_image,不要只口头说「我这就出图」。", "- 一次用户消息只出一轮;张数用会话已定参数,不要自己加张。", ]) if not allow_plan: lines.extend([ "", "【本轮闸门】", "- 用户本轮没有明确说「去做/出方案」。没有 write_strategy / write_plan / generate_image。", "- **禁止**本轮直接写出策略卡或方案卡。", ]) if has_context: lines.extend([ "- 会话已有素材或方向:针对用户本轮实际说的话自然回应;不要问要不要继续、要不要生成。", "- 禁止只回「在呢」「有需要随时招呼」「收到」这种空话。", ]) else: lines.extend([ "- 会话还没有创作进度:短回一句,告诉用户直接说想做什么即可;不要用客服式结束语,也不要硬推销出片。", ]) if params: meta = "、".join(f"{k}:{v}" for k, v in params.items() if v) if meta: lines.append(f"\n【会话已定参数】{meta}(出片按这套;用户改动后必须按新值重写方案)") if conversation.preset: # 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 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。") resolved = resolve_refs(context.team, conversation.pinned_refs or []) if resolved.facts: lines.append("\n【本次会话已锁定的素材事实】") lines.append(resolved.facts_text) lines.append( "以上素材的参考图会附给你看,出片时也会自动锁人锁物。" "人物性别、年龄段、发型、服装、商品颜色外形必须以图为准;" "图上看不清或没附图时,必须问用户,禁止凭文件名猜测男女。" ) memory = conversation.memory or {} if memory.get("summary"): lines.append(f"\n【前情提要】{memory['summary']}") artifacts = memory.get("artifacts") or [] if artifacts: recent = artifacts[-3:] lines.append("\n【本会话已生成过】") for index, item in enumerate(recent, 1): lines.append(f"{index}. {item.get('prompt', '')[:120]}") lines.append( "用户说「改成…」「换成…」时,是要在**最后一次生成**的基础上重新生成一版," "把改动合进完整 prompt 再调生成工具 —— 不要只写改动部分。" ) return "\n".join(lines) def _answer_label(field: dict, raw) -> str: """追问卡的内部值翻成人能读的标签,避免把 UUID / gate key 喂回模型。""" values = raw if isinstance(raw, list) else [raw] options = { str(item.get("value")): str(item.get("label") or item.get("value") or "") for item in (field.get("options") or []) if isinstance(item, dict) } return "、".join(options.get(str(value), str(value)) for value in values if value not in (None, "")) def build_messages(context: AgentContext, *, allow_plan: bool = True, has_context: bool = False) -> list[dict]: """会话历史 → 模型消息。只喂对模型有意义的:文字、追问和它的答案、生成过什么。 策略卡/方案卡这类结构化产物压成一句话,原样塞 JSON 只会挤爆上下文。 长会话只喂最近 KEEP_RECENT_MESSAGES 条原文,更早的靠 system 里的【前情提要】 ——摘要在 compress_memory() 里生成,不在这里现算。 """ messages = [{"role": "system", "content": build_system_prompt(context, allow_plan=allow_plan, has_context=has_context)}] history = list(context.conversation.messages.all()) if len(history) > COMPRESS_AFTER_MESSAGES: history = history[-KEEP_RECENT_MESSAGES:] for message in history: if message.kind == CreationMessage.Kind.TEXT: if message.text.strip(): messages.append({"role": message.role, "content": message.text}) elif message.kind == CreationMessage.Kind.ELICIT: payload = message.payload or {} answers = payload.get("answers") or {} fields = [item for item in (payload.get("fields") or []) if isinstance(item, dict)] field_by_key = {str(item.get("key") or ""): item for item in fields} if payload.get("interaction") == "chat" or ( payload.get("interaction") == "asset_picker" and payload.get("answered_via") == "chat" ): question = message.text.strip() or str((fields[0] if fields else {}).get("label") or "").strip() if question: messages.append({"role": "assistant", "content": question}) # 聊天式回答本身就是下一条 TEXT user 消息,这里不再合成一条伪造答案。 continue if answers: if payload.get("phase") == "gate": choice = str(answers.get("_asset_gate") or "") if choice == "skip": answer_text = "(用户选择暂不添加这项素材,要求按现有信息继续原任务)" elif choice == "auto": answer_text = "(用户希望你帮忙挑选/决定素材,按合理推荐继续推进)" elif choice == "send": answer_text = "(用户希望打开素材列表继续选择)" else: answer_text = f"(用户指定了素材：{choice})" else: joined = "；".join( f"{field_by_key.get(str(key), {}).get('label') or key}：" f"{_answer_label(field_by_key.get(str(key), {}), value)}" for key, value in answers.items() ) answer_text = f"(用户完成了刚才的选择：{joined})" messages.append({"role": "assistant", "content": "(我请用户补充了一项会影响创作的信息)"}) messages.append({"role": "user", "content": answer_text}) else: messages.append({"role": "assistant", "content": "(我正在等用户回答刚才的问题)"}) elif message.kind in (CreationMessage.Kind.GENERATING, CreationMessage.Kind.RESULT): prompt = (message.payload or {}).get("prompt") or "" messages.append({"role": "assistant", "content": f"(我生成了一版,prompt:{prompt[:200]})"}) elif message.kind == CreationMessage.Kind.ERROR: messages.append({"role": "assistant", "content": f"(上一次生成失败:{message.text})"}) if _creation_model_sees_images(context.model_config): messages = _attach_ref_images( messages, _ref_image_urls(context.team, context.conversation.pinned_refs or []), ) return messages # ---------------------------------------------------------------- 流式循环 def _sse(event: dict) -> str: """一帧 SSE。**必须用 DjangoJSONEncoder** —— 消息里带 UUID(task 外键)和 datetime(created_at),标准 json.dumps 直接抛 TypeError,整条流当场断掉。""" return f"data: {json.dumps(event, ensure_ascii=False, cls=DjangoJSONEncoder)}\n\n" def _merge_tool_call_deltas(buffer: dict, deltas: list) -> None: """OpenAI 流式把一次 tool_call 的 arguments 拆成很多片,按 index 拼回来。 name 只在第一片出现,arguments 要逐片累加 —— 直接覆盖会只剩最后一个字符。""" for delta in deltas or []: if not isinstance(delta, dict): continue index = delta.get("index", 0) slot = buffer.setdefault(index, {"name": "", "arguments": ""}) function = delta.get("function") or {} if function.get("name"): slot["name"] = function["name"] if function.get("arguments"): slot["arguments"] += function["arguments"] def _parse_arguments(raw: str) -> dict: """解析工具 arguments。模型偶发包 markdown 围栏或夹杂前后缀,尽量救回 JSON。""" text = (raw or "").strip() if not text: return {} if text.startswith("```"): text = text.strip("`") if text.lower().startswith("json"): text = text[4:].lstrip() text = text.strip() try: parsed = json.loads(text) except ValueError: start, end = text.find("{"), text.rfind("}") if start < 0 or end <= start: return {} try: parsed = json.loads(text[start : end + 1]) except ValueError: return {} return parsed if isinstance(parsed, dict) else {} def _pick_str(args: dict, *keys: str) -> str: """按候选键取非空字符串;兼容中文别名与嵌套一层 dict。""" for key in keys: value = args.get(key) if isinstance(value, dict): # 偶发 {"text": "..."} / {"value": "..."} for nested in ("text", "value", "content", "desc", "description"): inner = value.get(nested) if isinstance(inner, str) and inner.strip(): return inner.strip() continue if value is None: continue text = str(value).strip() if text: return text return "" def _coerce_points(raw) -> list[str]: """points 规整成最多 3 条非空文案。兼容纯字符串、对象数组、dict。""" if raw is None: return [] items: list = [] if isinstance(raw, str): text = raw.strip() if not text: return [] # 中文分号/换行拆条 for part in text.replace("\r", "\n").replace("；", "\n").split("\n"): part = part.strip(" ·•-、，,") if part: items.append(part) elif isinstance(raw, dict): # {"P0": "...", "P1": "..."} 或 {"0": "..."} for key in sorted(raw.keys(), key=lambda k: str(k)): val = raw[key] if isinstance(val, dict): text = _pick_str(val, "text", "value", "content", "desc", "point", "label") else: text = str(val or "").strip() if text: items.append(text) elif isinstance(raw, (list, tuple)): for item in raw: if isinstance(item, dict): text = _pick_str(item, "text", "value", "content", "desc", "point", "label") else: text = str(item or "").strip() if text: items.append(text) else: text = str(raw).strip() if text: items.append(text) # 过滤单字符噪声(字符串被误当成 list 迭代时的残留) cleaned = [p for p in items if len(p) > 1] return cleaned[:3] def _coerce_voice_chars(raw, fallback: list[int] | None = None) -> list[int]: """voice_chars 必须是 [下限, 上限];模型常误塞语气文案或单个整数。""" if isinstance(raw, (list, tuple)) and len(raw) >= 2: try: lo, hi = int(raw[0]), int(raw[1]) if lo > 0 and hi >= lo: return [lo, hi] except (TypeError, ValueError): pass if isinstance(raw, (int, float)) and int(raw) > 0: n = int(raw) return [max(1, n - 5), n + 5] return list(fallback or []) def _coerce_timeline(raw) -> list[dict]: items: list[dict] = [] for item in (raw or []) if isinstance(raw, list) else []: if not isinstance(item, dict): continue try: start = float(item.get("start")) end = float(item.get("end")) except (TypeError, ValueError): continue stage = str(item.get("stage") or "").strip() if not stage: continue entry = {"start": start, "end": end, "stage": stage} desc = str(item.get("desc") or item.get("description") or "").strip() if desc: entry["desc"] = desc items.append(entry) return items def _default_plan_matrix(usp: str, points: list[str]) -> dict: """设计稿里的「卖点覆盖矩阵」:有 USP/支撑点就自动铺一版,避免方案卡干瘪。""" rows = [{"point": "主打卖点 USP", "hits": [1, 3]}] labels = ["体验卖点 P0", "视觉卖点 P0", "转化卖点 P0"] for index, point in enumerate(points[:3]): label = labels[index] if index < len(labels) else f"支撑卖点 P{index}" # 点名用文案前缀,hits 错落分布到 4 镜 hit = (index % 4) + 1 rows.append({"point": label if not point else f"{label}", "hits": [hit]}) return {"shots": 4, "rows": rows} def _coerce_strategy_args(args: dict) -> dict: return { "target": _pick_str( args, "target", "audience", "who", "给谁看", "目标人群", "人群", ), "trust": _pick_str( args, "trust", "credibility", "为什么相信", "信任", "可信度", ), "belief": _pick_str( args, "belief", "希望相信", "想让他信什么", "认知", "takeaway", ), "direction": _pick_str( args, "direction", "创作方向", "方向", "style", "路线", ), } def _coerce_plan_card_args(args: dict) -> dict: usp = _pick_str(args, "usp", "主打卖点", "卖点", "core_usp", "main_point") points = _coerce_points( args.get("points") if args.get("points") is not None else args.get("支撑点") or args.get("supports") or args.get("selling_points") ) # 兼容 point1/point2/point3 展开写法(设计稿 blueprint) if not points: for key in ("point1", "point2", "point3", "P0", "P1", "P2"): text = _pick_str(args, key) if text: points.append(text) points = points[:3] timeline = _coerce_timeline(args.get("timeline") or args.get("时间轴")) matrix = args.get("matrix") if not isinstance(matrix, dict) or not matrix.get("rows"): matrix = _default_plan_matrix(usp, points) if (usp or points) else {} return { "usp": usp, "points": points, "timeline": timeline, "matrix": matrix, "voice_chars": args.get("voice_chars"), } def iter_creation_agent_events( *, conversation: CreationConversation, user, text: str, refs: list[dict] | None = None, model_config: ModelConfig | None = None, record_user_message: bool = True, force_creative_turn: bool = False, continuation_instruction: str = "", ) -> Iterator[dict]: """一条用户消息 → 事件 dict 流(message/delta/tool/done/error)。消息已落库。""" refs = refs or [] fast_greeting = record_user_message and is_greeting(text) # 打招呼不依赖模型，模型未配置也能即时回应；其他消息保持原有的先校验模型行为。 if not fast_greeting: model_config = get_creation_chat_model(model_config) if model_config is None: yield {"type": "error", "detail": "没有可用的文本模型,请先在模型库配置"} return context = AgentContext(conversation=conversation, user=user, model_config=model_config) try: user_message = None with transaction.atomic(): pin_refs(conversation, refs) if record_user_message: user_message = append_message(conversation, role="user", text=text, refs=refs) if user_message is not None: yield {"type": "message", "message": _message_payload(user_message)} # 单纯打招呼无论有没有参考素材、有没有既有上下文，都不值得等模型。 # 素材先收下，等用户说清用途再分析，避免「你好」卡住还冒出一大段建议。 if fast_greeting: reply = append_message( conversation, role="assistant", text=( "你好，素材我收到了。想拿它做什么，直接说就行。" if refs else "你好。想做什么，直接说就行。" ), ) yield {"type": "message", "message": _message_payload(reply)} yield {"type": "done"} return # 空会话里的简短应答：短回一句就够。已有商品/方向时走模型， # 针对用户实际说的话自然回应，但不给方案工具。 has_context = session_has_creative_context(conversation) if record_user_message and is_pure_chitchat(text) and not refs and not has_context: reply = append_message( conversation, role="assistant", text="我在。想做什么，直接丢一句想法给我就行。", ) yield {"type": "message", "message": _message_payload(reply)} yield {"type": "done"} return model_config = _prefer_vision_text_model(model_config, conversation.team, conversation.pinned_refs or []) context.model_config = model_config resolved = resolve_refs(context.team, refs) if resolved.missing: names = "、".join(r.get("name") or "某个素材" for r in resolved.missing) note = append_message( conversation, role="assistant", text=f"有几个引用的素材已经找不到了({names}),我先按其余信息继续。", ) yield {"type": "message", "message": _message_payload(note)} # 压缩放在**建消息之前**:摘要要进这一轮的 system 提示词才有意义。 # 用户消息已经先回显了,所以这一小段等待不会看起来像卡住。 compress_memory(context) allow_plan = ( force_creative_turn or has_creative_intent(text, refs) or (has_context and is_continue_intent(text)) ) provider = build_provider(model_config) messages = build_messages(context, allow_plan=allow_plan, has_context=has_context) if continuation_instruction.strip(): # 卡片答案已经在历史里,这里仅给本轮一个不落库的执行指令。这样既不会多出 # 一条伪造的用户气泡,也不会让模型把「跳过素材」误判成闲聊后停住。 messages.append({"role": "user", "content": continuation_instruction.strip()}) tools = tool_schemas(context, allow_plan=allow_plan) from .creation import is_agent_cancel_requested 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 yield {"type": "cancelled"} return text_buffer: list[str] = [] tool_buffer: dict = {} for chunk in provider.chat_completion_stream( model=model_config.name, messages=messages, endpoint=model_config.endpoint or "chat/completions", extra_body={"tools": tools}, ): kind = chunk.get("type") if kind == "reasoning": yield {"type": "reasoning", "text": chunk.get("text", "")} elif kind == "delta": piece = chunk.get("text", "") text_buffer.append(piece) yield {"type": "delta", "text": piece} elif kind == "tool_call": _merge_tool_call_deltas(tool_buffer, chunk.get("tool_calls")) said = "".join(text_buffer).strip() calls = [tool_buffer[i] for i in sorted(tool_buffer) if tool_buffer[i].get("name")] fallback_fields = None allow_pick = False if not calls: requested_card = requested_asset_card_from_context(conversation, text) wanted_pick = wanted_asset_pick(text, refs) pick = requested_card or wanted_pick param_keys = wanted_param_keys(text, is_video=context.is_video) if pick: fallback_fields = [{ "key": pick, "label": _ASSET_PICK_LABEL.get(pick, _ASSET_CARD_LABELS.get(pick, "请选择素材")), "type": "asset", "required": True, "asset_types": [pick], }] allow_pick = bool(requested_card) elif param_keys: fallback_fields = session_param_fields(param_keys, context.is_video) elif allow_plan and not said: # 部分推理模型会只给 reasoning_content 后结束本轮。不能让用户 # 只看到自己那条消息；没有商品上下文时，收敛成一条自然追问。 has_product = any( isinstance(ref, dict) and ref.get("type") == "product" for ref in (conversation.pinned_refs or []) ) if not has_product: fallback_fields = [{ "key": "product", "label": "这条想推哪款商品？", "type": "asset", "required": True, "asset_types": ["product"], }] # 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) if said and not asks_user: bubble = append_message(conversation, role="assistant", text=said) yield {"type": "message", "message": _message_payload(bubble)} if not calls: if fallback_fields: result, _stop = _dispatch_tool( context, "ask_user", {"fields": fallback_fields}, allow_pick=allow_pick ) for event in result.get("_events", []): yield event break messages.append({ "role": "assistant", "content": said or None, "tool_calls": [ {"id": f"call_{i}", "type": "function", "function": {"name": c["name"], "arguments": c["arguments"]}} for i, c in enumerate(calls) ], }) stop = False for index, call in enumerate(calls): if is_agent_cancel_requested(conversation.id): yield {"type": "cancelled"} return name = call["name"] args = _parse_arguments(call["arguments"]) yield {"type": "tool", "id": name, "label": _TOOL_LABELS.get(name, name), "status": "running"} try: result, stop_after = _dispatch_tool(context, name, args) except AgentError as exc: yield {"type": "tool", "id": name, "label": _TOOL_LABELS.get(name, name), "status": "error"} failure = append_message( conversation, role="assistant", kind=CreationMessage.Kind.ERROR, text=str(exc), ) yield {"type": "message", "message": _message_payload(failure)} yield {"type": "done"} return yield {"type": "tool", "id": name, "label": _TOOL_LABELS.get(name, name), "status": "done"} for event in result.get("_events", []): yield event messages.append({ "role": "tool", "tool_call_id": f"call_{index}", "content": json.dumps(result.get("payload", {}), ensure_ascii=False), }) stop = stop or stop_after if stop: break yield {"type": "done"} except Exception as exc: # noqa: BLE001 — SSE 里任何未捕获异常都会变成前端「白屏卡死」 logger.exception("creation agent stream failed: %s", exc) yield {"type": "error", "detail": "生成过程出错了,请再试一次"} def stream_creation_agent( *, conversation: CreationConversation, user, text: str, refs: list[dict] | None = None, model_config: ModelConfig | None = None, record_user_message: bool = True, force_creative_turn: bool = False, continuation_instruction: str = "", ) -> Iterator[str]: """兼容旧 SSE 消费方(单测 / 调试)。生产路径走 Celery + poll。""" for event in iter_creation_agent_events( conversation=conversation, user=user, text=text, refs=refs, model_config=model_config, record_user_message=record_user_message, force_creative_turn=force_creative_turn, continuation_instruction=continuation_instruction, ): yield _sse(event) def run_creation_agent_turn( *, conversation_id: str, user_id: str, text: str = "", refs: list | None = None, model_config_id: str | None = None, record_user_message: bool = True, force_creative_turn: bool = False, continuation_instruction: str = "", ) -> str: """Celery worker 入口:跑完一轮 tool loop,消息落库;更新 agent_status;释放团队锁。 不复用 CreationMessage.kind=generating / AITask —— 那些是确认后出片/出图用的。 """ from apps.accounts.models import User from .creation import finish_agent_planning conversation = ( CreationConversation.objects.select_related("team", "created_by") .filter(id=conversation_id) .first() ) if conversation is None: return conversation_id user = User.objects.filter(id=user_id).first() or conversation.created_by model_config = None if model_config_id: model_config = ( ModelConfig.objects.select_related("provider") .filter(id=model_config_id, capability=ModelConfig.Capability.TEXT, status=ModelConfig.Status.ACTIVE) .first() ) awaiting_user = False user_cancelled = False try: for event in iter_creation_agent_events( conversation=conversation, user=user, text=text or "", refs=refs or [], model_config=model_config, record_user_message=record_user_message, force_creative_turn=force_creative_turn, continuation_instruction=continuation_instruction or "", ): if not isinstance(event, dict): continue if event.get("type") == "cancelled": user_cancelled = True awaiting_user = False break if event.get("type") == "message": message = event.get("message") or {} kind = message.get("kind") if isinstance(message, dict) else None if kind in (CreationMessage.Kind.ELICIT, CreationMessage.Kind.CONFIRM): awaiting_user = True elif event.get("type") == "error": detail = str(event.get("detail") or "生成过程出错了,请再试一次") # 循环内多数错误已落 ERROR 消息;这里兜底一条,避免前端只看到卡在 planning last = conversation.messages.order_by("-seq").first() if last is None or last.kind != CreationMessage.Kind.ERROR: append_message( conversation, role="assistant", kind=CreationMessage.Kind.ERROR, text=detail, ) except Exception as exc: # noqa: BLE001 — worker 不能把异常冒成无限 planning logger.exception("creation agent turn failed: %s", exc) try: append_message( conversation, role="assistant", kind=CreationMessage.Kind.ERROR, text="生成过程出错了,请再试一次", ) except Exception: # noqa: BLE001 logger.exception("creation agent turn: failed to append error message") awaiting_user = False finally: try: conversation.refresh_from_db(fields=["agent_status", "team_id"]) except Exception: # noqa: BLE001 pass # 用户已终止时 API 侧多半已 idle+放锁;这里再收一次保证幂等,且不落入 awaiting_user finish_agent_planning( conversation, awaiting_user=(awaiting_user and not user_cancelled), ) return conversation_id _TOOL_LABELS = { "ask_user": "向你确认", "search_library": "查找素材", "generate_image": "生成图片", "write_strategy": "梳理创作策略", "write_plan": "编排视频方案", } _ASSET_CARD_LABELS = { "product": "这条要展示哪款商品？", "character": "这条想用哪个角色？", "model": "这条想用哪位模特？", "scene": "这条想放在哪个场景里？", "asset": "这一步要使用哪项素材？", } _GATE_LABELS = { "product": "这条想推哪款商品？可以直接告诉我商品名，或者需要我把商品库列表发给你选吗？", "character": "这条想用哪个角色？可以直接告诉我角色名，或者需要我发角色列表给你选吗？", "model": "这条想用哪位模特？可以直接告诉我，或者需要我把模特库发给你选吗？", "scene": "这条想放在哪个场景里？可以直接告诉我，或者需要我把场景库发给你选吗？", "asset": "这一步想用哪项素材？可以直接告诉我，或者需要我把素材列表发给你选吗？", } def _guided_elicit_text(field: dict) -> str: """追问既要问清一件事,也要让用户知道下一句怎么回。""" label = str(field.get("label") or "这项你想怎么定？").strip() if field.get("type") == "asset" or field.get("key") == "_asset_gate": return label if str(field.get("key") or "") in SESSION_PARAM_KEYS: return label return f"{label} 直接用一句话告诉我就行，不用整理成完整需求。" def _elicit_payload_for_fields( fields: list[dict], *, allow_pick: bool = False, is_video: bool = True ) -> dict: """素材问题默认出轻量对话闸门；用户明确要求发列表时才出 pick 选择卡。""" field = dict(fields[0]) if field.get("type") == "asset": asset_types = [item for item in (field.get("asset_types") or []) if item in _GATE_LABELS] primary_type = asset_types[0] if asset_types else "product" if allow_pick: field["label"] = _ASSET_CARD_LABELS.get(primary_type, _ASSET_CARD_LABELS["asset"]) return { "interaction": "asset_picker", "phase": "pick", "fields": [field], "submitted": False, "answers": {}, } if primary_type == "product": gate_label = ( "这条视频想推哪款商品？可以直接告诉我商品名，或者需要我把商品库列表发给你选吗？" if is_video else "这次想做哪款商品？可以直接告诉我商品名，或者需要我把商品库列表发给你选吗？" ) else: gate_label = _GATE_LABELS.get(primary_type, _GATE_LABELS["asset"]) gate_field = { "key": "_asset_gate", "label": gate_label, "type": "text", "required": False, } return { # 这是 Agent 的一句自然追问，不是让用户点选流程的卡片。 # 只有用户明确要求发列表后，才会生成上面 allow_pick 分支的商品卡。 "interaction": "chat", "phase": "gate", "fields": [gate_field], "pending_fields": [field], "submitted": False, "answers": {}, } return { "interaction": "chat", "fields": [field], "submitted": False, "answers": {}, } def _dispatch_tool( context: AgentContext, name: str, args: dict, *, allow_pick: bool = False ) -> tuple[dict, bool]: """执行一个工具。返回 (结果, 是否中断循环)。 结果里的 `_events` 会原样转发给前端,`payload` 回喂给模型。 """ if name == "ask_user": fields = _coerce_fields(args.get("fields")) if not fields: return {"payload": {"error": "fields 不合法,请重新组织问题"}}, False payload = _elicit_payload_for_fields(fields, allow_pick=allow_pick, is_video=context.is_video) display_field = (payload.get("fields") or fields)[0] message = append_message( context.conversation, role="assistant", kind=CreationMessage.Kind.ELICIT, text=_guided_elicit_text(display_field), payload=payload, ) # 反问一旦发出就必须停,等人回答。继续跑等于自问自答。 return { "payload": {"asked": True}, "_events": [{"type": "message", "message": _message_payload(message)}], }, True if name == "search_library": return {"payload": _run_search_library(context, args)}, False if name == "write_strategy": strategy_payload = _coerce_strategy_args(args if isinstance(args, dict) else {}) # 空卡会落成「只有标签没有正文」——拒绝,让模型把四字段写满再调 if not all(strategy_payload.values()): return { "payload": { "error": ( "创作策略卡四字段都不能为空:请填写具体的 target / trust / belief / direction" "(给谁看、为什么信、希望他信什么、创作方向),不要留空。" ) } }, False message = append_message( context.conversation, role="assistant", kind=CreationMessage.Kind.STRATEGY, payload=strategy_payload, ) # 策略卡只是「我理解对了吗」,不打断 —— 模型接着就该写方案 return { "payload": {"written": True}, "_events": [{"type": "message", "message": _message_payload(message)}], }, False if name == "write_plan": 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 card = _coerce_plan_card_args(args if isinstance(args, dict) else {}) if not card["usp"] or not card["points"]: return { "payload": { "error": ( "方案卡缺正文:请填写非空的 usp(主打卖点)和 points(1–3 条核心支撑)," "不要只写 video_prompt。卡片上要让用户看见卖点文案。" ) } }, False duration = 15 try: duration = int(float(str((context.conversation.params or {}).get("duration") or "15").replace("秒", "").strip() or "15")) except (TypeError, ValueError): duration = 15 lo = max(20, round(duration * 3.4)) hi = max(lo + 1, round(duration * 4)) plan_payload = { "usp": card["usp"], "points": card["points"], "timeline": card["timeline"], "matrix": card["matrix"], "voice_chars": _coerce_voice_chars(card["voice_chars"], [lo, hi]), "ref_count": len(context.conversation.pinned_refs or []), } events = [] plan = append_message( context.conversation, role="assistant", kind=CreationMessage.Kind.PLAN, payload=plan_payload, ) events.append({"type": "message", "message": _message_payload(plan)}) prompt_file = append_message( context.conversation, role="assistant", kind=CreationMessage.Kind.PROMPT_FILE, payload={"title": "视频生成Prompt.md", "body": video_prompt, "ref_count": plan_payload["ref_count"]}, ) events.append({"type": "message", "message": _message_payload(prompt_file)}) credits = estimate_video_credits(context) # video_prompt 存在确认卡里:用户点确认后直接照它出片,不再跑一轮模型 confirm = append_message( context.conversation, role="assistant", kind=CreationMessage.Kind.CONFIRM, payload={ "kind": "video", "label": "开始生成", "estimated_credits": credits, "video_prompt": video_prompt, "submitted": False, "params": snapshot_session_params(context.conversation), "param_options": confirm_param_options(True), }, ) events.append({"type": "message", "message": _message_payload(confirm)}) events.append({"type": "credits", "estimated": credits}) # 方案卡写着「仅需确认一次」—— 停在这里等人点,别自己往下出片 return {"payload": {"awaiting_confirmation": True}, "_events": events}, True if name == "generate_image": if context.generations_used >= MAX_BILLED_GENERATIONS: # 一条用户消息只计费一次。模型想连出好几版时在这里挡住。 return {"payload": {"error": "本轮已经生成过一次了,请让用户看过再决定要不要改"}}, True prompt = str(args.get("prompt") or "").strip() if not prompt: return {"payload": {"error": "生成失败:模型没有给出画面描述"}}, False # 出图也走确认卡:用户先看当前模型/比例/张数,点了才提交。 credits = estimate_image_credits(context) confirm = append_message( context.conversation, role="assistant", kind=CreationMessage.Kind.CONFIRM, payload={ "kind": "image", "label": "开始生成", "estimated_credits": credits, "prompt": prompt, "submitted": False, "params": snapshot_session_params(context.conversation), "param_options": confirm_param_options(False), }, ) events = [ {"type": "message", "message": _message_payload(confirm)}, {"type": "credits", "estimated": credits}, ] return {"payload": {"awaiting_confirmation": True}, "_events": events}, True return {"payload": {"error": f"未知工具 {name}"}}, False def _summarize_source(messages: list[CreationMessage]) -> str: """要压缩的那批消息 → 喂给模型的纯文本。只取有信息量的部分。""" lines = [] for message in messages: if message.kind == CreationMessage.Kind.TEXT and message.text.strip(): who = "用户" if message.role == "user" else "我" lines.append(f"{who}:{message.text.strip()}") elif message.kind == CreationMessage.Kind.ELICIT: answers = (message.payload or {}).get("answers") or {} if answers: lines.append("用户确认:" + "；".join(f"{k}={v}" for k, v in answers.items())) elif message.kind in (CreationMessage.Kind.GENERATING, CreationMessage.Kind.RESULT): prompt = (message.payload or {}).get("prompt") or "" if prompt: lines.append(f"我生成了一版:{prompt[:120]}") return "\n".join(lines) def compress_memory(context: AgentContext) -> None: """把早期消息压成一段摘要存进 conversation.memory.summary(契约 §5)。 只在消息数超过阈值时做;压缩失败**静默跳过** —— 摘要是锦上添花, 为它把整条对话打断不值得。压缩额外调一次模型,所以按 summarized_upto 记进度,同一批消息不重复压。 """ conversation = context.conversation history = list(conversation.messages.all()) if len(history) <= COMPRESS_AFTER_MESSAGES: return memory = dict(conversation.memory or {}) cutoff = len(history) - KEEP_RECENT_MESSAGES if cutoff - int(memory.get("summarized_upto") or 0) < COMPRESS_MIN_BATCH: return # 没压过的还不够一批,攒着 —— 每轮重压一次太贵 source = _summarize_source(history[:cutoff]) if not source.strip(): memory["summarized_upto"] = cutoff conversation.memory = memory conversation.save(update_fields=["memory", "updated_at"]) return previous = memory.get("summary") or "" instruction = ( "把下面这段创作对话压成一段中文摘要,150 字以内。" "保留:用户明确提过的要求和否决过的方向、已确认的设定、生成过什么。" "丢掉:寒暄、过程性的话。直接输出摘要正文,不要前言。\n\n" + (f"【已有摘要】{previous}\n\n" if previous else "") + f"【新增对话】\n{source}" ) try: provider = build_provider(context.model_config) pieces = [] for chunk in provider.chat_completion_stream( model=context.model_config.name, messages=[{"role": "user", "content": instruction}], endpoint=context.model_config.endpoint or "chat/completions", ): if chunk.get("type") == "delta": pieces.append(chunk.get("text", "")) summary = "".join(pieces).strip() except Exception: # noqa: BLE001 — 摘要失败不该打断对话 logger.warning("omni create: memory compression failed", exc_info=True) return if not summary: return memory["summary"] = summary[:400] memory["summarized_upto"] = cutoff conversation.memory = memory conversation.save(update_fields=["memory", "updated_at"]) def _remember_artifact(conversation: CreationConversation, prompt: str, kind: str) -> None: """记进产物索引,让下一轮「把背景换成夜景」能定位到这一版(契约 §5)。""" memory = dict(conversation.memory or {}) artifacts = list(memory.get("artifacts") or []) artifacts.append({"prompt": prompt, "kind": kind}) memory["artifacts"] = artifacts[-10:] conversation.memory = memory conversation.save(update_fields=["memory", "updated_at"]) def _message_payload(message: CreationMessage) -> dict: from .serializers import CreationMessageSerializer return CreationMessageSerializer(message).data