"""对话式脚本生成 agent(出稿 + 改稿一体,多模型可选,流式 SSE)。 设计: - 加载电商脚本 skill(SKILL.md + references)作为领域知识系统提示词;模型无关。 - 3 种输入模式(全自动 / 一句话 / 改稿)收敛到同一份结构化 ScriptDraft(铁律1契约)。 - 流式:边生成边吐「工具卡 + 思考」事件,给前端真 agent 体感;JSON 由后端可靠抽取,不靠模型排版。 - 计费走现有 AITask + 额度预扣(reserve→charge/release),与 generate_project_script 一致。 SSE 事件(每帧 `data: {json}\n\n`,json 带 type): tool {id,label?,status:running|done|error} —— 工具卡(加载skill/分析商品/生成分镜/提取实体/自检) reasoning {text} —— 推理模型思考流(逐字,纯展示,不进答案) delta {text} —— 模型自然语言前言(JSON 部分不外露) draft {draft} —— 规范化后的 ScriptDraft(前端结构化渲染) saved {script_version_id, version} —— 已落库的 ScriptVersion(含 segments/metadata) summary {text} —— 模型自己写的收尾交付语(当 AI 回复气泡,替代写死的「已生成」) done {} —— 结束 error {detail,error} —— 失败(已回滚额度；error 为安全业务错误对象) """ from __future__ import annotations import json import re from decimal import Decimal from functools import lru_cache from pathlib import Path from django.conf import settings from django.utils import timezone from apps.ai.generation_errors import classify_generation_error from apps.ai.models import AITask, ModelConfig from apps.billing.services.ledger import charge_reserved_credit, release_credit VALID_TONES = ["种草", "测评", "剧情", "痛点"] VALID_ROLES = ["钩子", "痛点", "卖点", "CTA"] VALID_ENTITY_TYPES = ["character", "scene", "product"] DURATION_TIERS = [15, 30, 60, 90] # --------------------------------------------------------------------------- # # skill 加载(缓存) # --------------------------------------------------------------------------- # def _skill_dir() -> Path: override = getattr(settings, "ECOMMERCE_SKILL_DIR", None) if override: return Path(override) # skills 已随后端打进镜像(core/backend/skills);优先 BASE_DIR/skills,回落仓库根(本地/旧布局)。 # 旧版只看仓库根 → 镜像里没有(构建上下文是 ./core/backend)→ 提示词为空、退化兜底。详见 services._skills_root。 base = Path(settings.BASE_DIR) for cand in (base / "skills", base.parent.parent / "skills"): if (cand / "ecommerce-video-script").is_dir(): return cand / "ecommerce-video-script" return base / "skills" / "ecommerce-video-script" @lru_cache(maxsize=1) def load_ecommerce_skill() -> str: """读取 SKILL.md + 全部 references 拼成系统提示词(领域知识)。缺文件不致命,尽量给。""" skill_dir = _skill_dir() parts: list[str] = [] main = skill_dir / "SKILL.md" if main.exists(): parts.append(main.read_text(encoding="utf-8")) ref_dir = skill_dir / "references" if ref_dir.exists(): for ref in sorted(ref_dir.glob("*.md")): parts.append(f"\n\n===== references/{ref.name} =====\n\n{ref.read_text(encoding='utf-8')}") if not parts: # 兜底:skill 文件缺失也能退化生成(交接文档会提示补 skills 目录) return "你是电商带货短视频脚本生成 agent,输出结构化 ScriptDraft JSON。" return "".join(parts) # 运行时输出协议:优先级高于 skill 里的「只输出 JSON / 不展示思考」,只为流式体感放开一句前言。 _OUTPUT_PROTOCOL = """ --- ## 运行时输出协议(AirShelf 流式展示专用,优先级高于技能正文的「只输出 JSON」) 严格按以下顺序输出,不要有别的内容: 1. 先用 **1 句中文口语**告诉用户你正在做什么(≤40 字,例:「在为这款保温杯生成 4 镜痛点脚本…」),让用户看到进展; 2. 紧接着输出**且仅输出一个** ```json 代码块,内容为符合技能契约(铁律1)的 ScriptDraft 对象; 3. json 代码块**收尾之后另起一行**,用 **1–2 句中文口语**跟用户交付这一版:做了什么、为什么这么改、可以怎么接着调(像同事汇报,**别复述 JSON 字段、别再写代码块**)。这句会作为你的回复气泡展示给用户。 ### 字段名锚定(硬性 · 下游靠它取数,跑偏即数据全空) - 分镜数组的键名**必须**叫 `segments`(禁止用 scenes / script / shots / 分镜 等同义词)。 - 每镜口播键名**必须**叫 `narration`(禁止用 voiceover / audio / line)。 - 每镜画面键名**必须**叫 `visual`,且为**一句话字符串**(禁止用 scene / screen / 画面,也禁止写成 {setting,camera,...} 对象)。 - 即使你额外附带了 scenes / shots 等创作结构,也**必须同时**给出标准 `segments` 数组,并把口播填进 `narration`、画面填进 `visual`,否则视为不合格。 """ # --------------------------------------------------------------------------- # # 提示词构建(3 模式) # --------------------------------------------------------------------------- # def _product_context(project, selling_point_ids: list[str] | None) -> str: product = project.product selling_points = product.selling_points.all() if selling_point_ids: selling_points = selling_points.filter(id__in=selling_point_ids) selling_text = "\n".join(f"- {sp.title}:{sp.detail}" for sp in selling_points) return ( f"商品标题:{product.title}\n" f"品牌:{product.brand or '未填写'}\n" f"类目:{product.category or '未填写'}\n" f"目标人群:{product.target_audience or '未填写'}\n" f"商品描述:{product.description or '未填写'}\n" f"卖点:\n{selling_text or '未勾选卖点,请根据商品信息自行提炼。'}" ) def build_agent_messages( *, project, mode: str, user_prompt: str, selling_point_ids: list[str] | None, base_draft: dict | None, aspect_ratio: str, total_duration: int, target_index: int | None = None, ) -> list[dict[str, str]]: system = load_ecommerce_skill() + _OUTPUT_PROTOCOL head = ( f"【画幅】{aspect_ratio}\n" f"【总时长】{total_duration} 秒(每 15 秒一镜,共 {total_duration // 15} 镜)\n" f"【商品信息】\n{_product_context(project, selling_point_ids)}" ) if mode == "revise" and base_draft and target_index is not None: # 精准改一镜:读全脚本上下文,只重写第 N 镜,强制与前后镜衔接;其余镜后端会强制保持原样。 user = ( f"【任务】只重写第 {target_index + 1} 镜(共 {len(base_draft.get('segments', []))} 镜),其余镜保持不变。\n" f"{head}\n\n" f"【现有完整脚本 JSON(读它保证与前后镜衔接)】\n{json.dumps(base_draft, ensure_ascii=False)}\n\n" f"【对第 {target_index + 1} 镜的修改意见】{user_prompt.strip() or '让这一镜更有吸引力、表达更清晰。'}\n\n" "仍输出**完整** ScriptDraft(我只会采用第 " f"{target_index + 1} 镜的改动);若意见涉及角色对白,就给这一镜填 dialogue。" ) elif mode == "revise" and base_draft: user = ( "【任务】改稿(模式③):在保留用户原意的前提下,增强钩子/节奏/卖点/CTA,并归一化到契约 JSON。\n" f"{head}\n\n" f"【现有脚本 JSON】\n{json.dumps(base_draft, ensure_ascii=False)}\n\n" f"【用户修改意见】{user_prompt.strip() or '让整体更有吸引力、转化感更强,并保持各镜衔接连贯。'}\n\n" "请输出修订后的**完整** ScriptDraft。" ) elif mode == "theme" or (user_prompt and user_prompt.strip()): user = ( "【任务】一句话主题扩写(模式②):以用户主题为脚本主轴,其余自动补全。\n" f"{head}\n\n" f"【用户主题】{user_prompt.strip()}\n\n" "请按技能流程一次性产出 ScriptDraft。" ) else: user = ( "【任务】全自动(模式①):仅凭商品与前置条件,自动定档/选 tone/造 entity/填黄金结构。\n" f"{head}\n\n" "请按技能流程一次性产出 ScriptDraft。" ) return [{"role": "system", "content": system}, {"role": "user", "content": user}] # --------------------------------------------------------------------------- # # JSON 抽取 + 契约规范化(模型无关,后端兜底) # --------------------------------------------------------------------------- # def _balanced_object(text: str) -> str | None: """从首个 '{' 起按括号深度扫描,返回第一个配平的 {...}(忽略字符串内的括号)。 避免 rfind('}') 在 JSON 后还有含花括号的散文时越界截出非法片段。""" start = text.find("{") if start == -1: return None depth = 0 in_str = False esc = False for i in range(start, len(text)): c = text[i] if in_str: if esc: esc = False elif c == "\\": esc = True elif c == '"': in_str = False continue if c == '"': in_str = True elif c == "{": depth += 1 elif c == "}": depth -= 1 if depth == 0: return text[start : i + 1] return None def _looks_like_draft(blob: str) -> bool: try: d = json.loads(blob) except (ValueError, TypeError): return False return isinstance(d, dict) and ("segments" in d or "hook" in d) def _extract_json(text: str) -> str | None: """抽取 ScriptDraft JSON。容错:模型可能先给示例 ```json 块再给正式块, 故取**最后一个**含 segments/hook 的合法围栏块;都不像草稿再退而取末个配平对象;无围栏再裸扫。""" fences = re.findall(r"```(?:json)?\s*(.+?)```", text, re.DOTALL) for block in reversed(fences): obj = _balanced_object(block) if obj and _looks_like_draft(obj): return obj for block in reversed(fences): obj = _balanced_object(block) if obj: return obj return _balanced_object(text) def _nearest_duration(value) -> int: try: value = int(value) except (TypeError, ValueError): return 60 if value in DURATION_TIERS: return value return min(DURATION_TIERS, key=lambda t: abs(t - value)) # 模型每次生成都可能换字段名(scene/screenDescription/visual…、dialogue/lines/caption…), # 与其逐一追变体,不如「优先键命中 → 否则按关键词模糊匹配」通用解析。SKIP 掉明显的非内容键, # 避免误抓(shotNo/duration/bgMusic/note 等)。 _PICK_SKIP_KEYS = { "shotno", "shotsize", "shotsizetype", "duration", "duration_seconds", "starttime", "endtime", "timerange", "time_range", "bgmusic", "bg_music", "music", "sound", "sfx", "note", "notes", "tips", "index", "role", "speaker", "entity_refs", "product_exposure", "id", "no", "transition", # 标题/编号类:含 scene 字样会误命中画面 fuzzy,显式跳过(注意:不跳裸 scene,它常=画面) "scene_id", "sceneid", "scene_no", "sceneno", "scene_number", "scenenumber", "scene_title", "scenetitle", "title", "scene_title_type", } _VISUAL_EXACT = ("visual", "visual_prompt", "visual_description", "screen_description", "screendescription", "screen", "picture", "shot_description", "scene", "画面") _VISUAL_FUZZY = ("visual", "screen", "picture", "scene", "画面", "镜头描述", "分镜画面") _NARRATION_EXACT = ("narration", "voiceover", "voice_over", "vo", "line", "caption", "subtitle", "speech", "口播", "旁白") _NARRATION_FUZZY = ("narrat", "voice", "旁白", "口播", "台词", "dialog", "caption", "subtitle", "字幕", "speech", "monolog", "audio") def _flatten_text(val) -> str: """把 字符串/字典/列表 里的文本拍平成一句。模型常把 visual 写成 {setting,camera,key_shots} 对象、把口播写成数组,这里统一抽成纯文本,避免结构化值被当空丢弃(全空根因之一)。""" if isinstance(val, str): return val.strip() if isinstance(val, dict): return " · ".join(p for p in (_flatten_text(v) for v in val.values()) if p) if isinstance(val, (list, tuple)): return " ".join(p for p in (_flatten_text(v) for v in val) if p) return "" def _pick_field(seg: dict, exact: tuple[str, ...], fuzzy: tuple[str, ...]) -> str: """从一镜里取某类文本字段:先按优先键精确命中,再按关键词在剩余键里模糊匹配(跳过非内容键)。 值允许是 字符串/对象/数组(嵌套结构拍平成一句),不再只认裸字符串。""" for key in exact: if key in seg: txt = _flatten_text(seg.get(key)) if txt: return txt for key, val in seg.items(): kl = str(key).lower() if kl in _PICK_SKIP_KEYS: continue if any(f in kl for f in fuzzy): txt = _flatten_text(val) if txt: return txt return "" # 模型给分镜数组的键名五花八门(segments/scenes/script/shots…),且常同时给一个**空的** # segments 骨架 + 真内容放在 scenes 里。所以不能「见 segments 是 list 就用」,要在所有候选里 # 挑「能解析出最多非空旁白/画面」的那个。segments(契约本名)排第一,同分时优先。 _SEGMENT_ARRAY_KEYS = ( "segments", "shots", "scenes", "script", "shot_list", "shotlist", "shotList", "scene_list", "scenes_list", "storyboard", "分镜", "镜头", ) def _resolve_segments(draft: dict) -> list: """在所有候选数组键里挑内容最丰富的分镜数组(按非空旁白/画面条数打分)。""" best: list = [] best_score = -1 for key in _SEGMENT_ARRAY_KEYS: arr = draft.get(key) if not isinstance(arr, list) or not arr: continue dict_items = [s for s in arr if isinstance(s, dict)] if not dict_items: continue score = sum( 1 for s in dict_items if _pick_field(s, _NARRATION_EXACT, _NARRATION_FUZZY) or _pick_field(s, _VISUAL_EXACT, _VISUAL_FUZZY) ) if score > best_score: # 严格大于 → 同分保留更靠前的键(segments 优先) best, best_score = arr, score return best def normalize_draft(raw_text: str, *, aspect_ratio: str, total_duration: int) -> dict: """把模型输出抽成 JSON 并按铁律1契约规范化。宽容:小问题就地修,不轻易抛错。""" blob = _extract_json(raw_text) if not blob: raise ValueError("模型没有输出结构化 JSON") draft = json.loads(blob) if not isinstance(draft, dict): raise ValueError("脚本 JSON 顶层不是对象") # 兼容模型不按契约的常见变体:① 把内容裹进 {"ScriptDraft": {...}} 外壳; # ② 用 basicInfo(camelCase)放时长/比例;③ 用 shots 代替 segments。 # 解开外壳并把别名拍平到契约字段,避免「找不到 segments → 全填空占位镜」(旁白/画面全空)。 for wrapper in ("ScriptDraft", "script_draft", "scriptDraft", "draft"): inner = draft.get(wrapper) if isinstance(inner, dict): draft = {**inner, **{k: v for k, v in draft.items() if k != wrapper}} break basic = draft.get("basicInfo") if isinstance(draft.get("basicInfo"), dict) else {} if basic: draft.setdefault("total_duration", basic.get("totalDuration") or basic.get("total_duration")) draft.setdefault("aspect_ratio", basic.get("aspectRatio") or basic.get("aspect_ratio")) draft.setdefault("theme", basic.get("theme")) # 在所有候选键里挑内容最丰富的分镜数组(空 segments 骨架会被更丰富的 scenes/shots 顶替) draft["segments"] = _resolve_segments(draft) draft["aspect_ratio"] = (draft.get("aspect_ratio") or aspect_ratio or "9:16").strip() dur = _nearest_duration(draft.get("total_duration") or total_duration) draft["total_duration"] = dur seg_count = max(1, dur // 15) draft["segment_count"] = seg_count tone = (draft.get("tone") or "").strip() draft["tone"] = tone if tone in VALID_TONES else "种草" draft["hook"] = (draft.get("hook") or "").strip() # entities 规范化:补 id / ref_index,过滤非法 type entities = draft.get("entities") if isinstance(draft.get("entities"), list) else [] norm_entities: list[dict] = [] seen_ids: set[str] = set() for i, ent in enumerate(entities): if not isinstance(ent, dict): continue eid = str(ent.get("id") or f"e{i + 1}").strip() or f"e{i + 1}" while eid in seen_ids: eid = f"{eid}_{i}" seen_ids.add(eid) etype = (ent.get("type") or "").strip() if etype not in VALID_ENTITY_TYPES: etype = "character" norm_entities.append( { "id": eid, "type": etype, "name": (ent.get("name") or eid).strip(), "visual_prompt": (ent.get("visual_prompt") or "").strip(), "ref_index": ent.get("ref_index") if isinstance(ent.get("ref_index"), int) else i + 1, "voice_ref": ent.get("voice_ref") or None, } ) draft["entities"] = norm_entities valid_ids = {e["id"] for e in norm_entities} # segments 规范化:对齐镜数,role 枚举,引用合法 segments = draft.get("segments") if isinstance(draft.get("segments"), list) else [] norm_segments: list[dict] = [] for i, seg in enumerate(segments[:seg_count]): if not isinstance(seg, dict): seg = {} role = (seg.get("role") or "").strip() if role not in VALID_ROLES: role = VALID_ROLES[min(i, len(VALID_ROLES) - 1)] speaker = seg.get("speaker") speaker = speaker if (speaker in valid_ids) else None refs = [r for r in (seg.get("entity_refs") or []) if r in valid_ids] # 对白(剧情向):[{speaker(合法 entity id 或 null=旁白), line}];默认空 = 纯口播。 # 模型变体的口播字段五花八门:dialogue(字符串/数组)/ lines(数组)/ 每项 line|text|content。 # 这里把任意数组形态归一成结构化对白,字符串形态留给下面当整句旁白。 raw_dialogue = seg.get("dialogue") dialogue = [] dialogue_array = raw_dialogue if isinstance(raw_dialogue, list) else (seg.get("lines") if isinstance(seg.get("lines"), list) else []) for d in dialogue_array: if not isinstance(d, dict): continue line = (d.get("line") or d.get("text") or d.get("content") or "").strip() if not line: continue sp = d.get("speaker") dialogue.append({"speaker": sp if sp in valid_ids else None, "line": line}) # 旁白:结构化对白/lines 优先,其次整句字符串 dialogue,再退到通用字段解析(narration/voiceover/caption/字幕…) narration = "" if isinstance(raw_dialogue, str) and raw_dialogue.strip(): narration = raw_dialogue.strip() elif dialogue: narration = " ".join(d["line"] for d in dialogue) # 扁平拼接,兼容下游字幕/配音 if not narration: narration = _pick_field(seg, _NARRATION_EXACT, _NARRATION_FUZZY) # 画面:通用解析(visual/scene/screenDescription/画面… 都能命中,跳过 shotNo/bgMusic 等) visual = _pick_field(seg, _VISUAL_EXACT, _VISUAL_FUZZY) norm_segments.append( { "index": i, "duration": 15, "role": role, "narration": narration, "speaker": speaker, "visual": visual, "product_exposure": (seg.get("product_exposure") or "").strip(), "entity_refs": refs, "dialogue": dialogue, } ) # 不足镜数则补占位镜(极少发生,避免下游镜数对不上) while len(norm_segments) < seg_count: i = len(norm_segments) norm_segments.append( { "index": i, "duration": 15, "role": VALID_ROLES[min(i, len(VALID_ROLES) - 1)], "narration": "", "speaker": None, "visual": "", "product_exposure": "", "entity_refs": [], "dialogue": [], } ) if not norm_segments: raise ValueError("脚本没有任何分镜") draft["segments"] = norm_segments return draft def _merge_single_segment(base: dict, new: dict, idx: int, aspect_ratio: str, total_duration: int) -> dict: """精准改一镜:以基准稿为底,只用新稿的第 idx 镜替换,其余镜逐字保持;合并新稿引入的新 entity(对白可能加角色)。再整体规范化。""" merged = json.loads(json.dumps(base)) # 深拷贝 base_ids = {e.get("id") for e in merged.get("entities", []) if isinstance(e, dict)} for e in new.get("entities", []): if isinstance(e, dict) and e.get("id") and e["id"] not in base_ids: merged.setdefault("entities", []).append(e) base_ids.add(e["id"]) new_segs = new.get("segments", []) target = next((s for s in new_segs if isinstance(s, dict) and s.get("index") == idx), None) if target is None and 0 <= idx < len(new_segs): target = new_segs[idx] if not isinstance(target, dict): # 模型没产出目标镜(没按 N 镜输出)→ 抛错让上层释放额度+报错,而不是静默返回 base 空转计费 raise ValueError(f"模型未产出第 {idx + 1} 镜的改动,请重试") segs = merged.get("segments", []) if isinstance(target, dict) and 0 <= idx < len(segs): target = dict(target) target["index"] = idx segs[idx] = target merged["segments"] = segs return normalize_draft(json.dumps(merged, ensure_ascii=False), aspect_ratio=aspect_ratio, total_duration=total_duration) # --------------------------------------------------------------------------- # # 落库 # --------------------------------------------------------------------------- # def _map_entities_to_project_metadata(project, entities: list[dict]) -> None: """把结构化 entities 回填到 project.metadata,复用下游已有的 cast/scenes/*_prompts 接线 (脚本页标签 + 基础资产 seed + 故事板 @图N)。只在有内容时覆盖,空结果不清旧标签。""" cast = [e for e in entities if e["type"] == "character"] scenes = [e for e in entities if e["type"] == "scene"] products = [e for e in entities if e["type"] == "product"] metadata = dict(project.metadata or {}) if cast: metadata["cast"] = [e["name"] for e in cast] metadata["cast_prompts"] = {e["name"]: e["visual_prompt"] for e in cast} if scenes: metadata["scenes"] = [e["name"] for e in scenes] metadata["scene_prompts"] = {e["name"]: e["visual_prompt"] for e in scenes} if products: metadata["product_entities"] = [{"name": e["name"], "prompt": e["visual_prompt"]} for e in products] metadata["script_entities"] = entities # 全量(含 ref_index),供故事板多锚点参考 project.metadata = metadata project.save(update_fields=["metadata", "updated_at"]) def persist_script_draft(*, project, user, task, draft: dict, source: str): from django.db import transaction from apps.projects.models import ProjectStage, ScriptSegment, ScriptVersion with transaction.atomic(): script = ScriptVersion.objects.create( project=project, task=task, title=(draft.get("hook") or "AI 脚本")[:128], content=json.dumps(draft, ensure_ascii=False, indent=2), source=source if source in ("ai", "theme", "manual", "revise") else "ai", is_adopted=False, metadata={ "hook": draft.get("hook", ""), "tone": draft.get("tone", ""), "aspect_ratio": draft.get("aspect_ratio", "9:16"), "total_duration": draft.get("total_duration", 60), "segment_count": draft.get("segment_count", 4), "entities": draft.get("entities", []), }, ) for seg in draft["segments"]: ScriptSegment.objects.create( script_version=script, sort_order=seg["index"], duration_seconds=seg.get("duration", 15), narration=seg.get("narration", ""), visual_prompt=seg.get("visual", ""), role=seg.get("role", ""), speaker=seg.get("speaker") or "", product_exposure=seg.get("product_exposure", ""), entity_refs=seg.get("entity_refs") or [], dialogue=seg.get("dialogue") or [], product_points=[], ) _map_entities_to_project_metadata(project, draft.get("entities", [])) stage, _ = ProjectStage.objects.get_or_create(project=project, stage=ProjectStage.Stage.SCRIPT) stage.status = ProjectStage.Status.NEEDS_REVIEW stage.save(update_fields=["status", "updated_at"]) return script # --------------------------------------------------------------------------- # # 流式编排 # --------------------------------------------------------------------------- # def _sse(obj: dict) -> str: return f"data: {json.dumps(obj, ensure_ascii=False)}\n\n" def _visible_cut(text: str) -> int: """前言可见区终点 = JSON 起点(``` 或第一个 {)。之后的内容不外露,只在后端解析。""" cands = [] for marker in ("```", "{"): i = text.find(marker) if i != -1: cands.append(i) return min(cands) if cands else len(text) def _closing_summary(raw: str) -> str: """模型在 json 之后写的收尾交付语 = 给用户的回复气泡(像同事汇报「这版做了啥/怎么接着调」)。 取最后一个 json 对象之后的文字,去掉收尾围栏与空白;没写或残留花括号就返回空(前端兜底)。""" js = _extract_json(raw) tail = "" if js: idx = raw.rfind(js) if idx != -1: tail = raw[idx + len(js):] tail = re.sub(r"`+", " ", tail).strip() # 去掉 json 收尾的 ``` 围栏 if "{" in tail or len(tail) < 4: return "" return tail[:160] def stream_script_agent( *, project, user, model_config: ModelConfig, mode: str = "auto", user_prompt: str = "", selling_point_ids: list[str] | None = None, base_version_id: str | None = None, aspect_ratio: str = "9:16", total_duration: int = 60, target_index: int | None = None, ): """生成 SSE 帧字符串的同步生成器,供 StreamingHttpResponse 包裹。 target_index 非空 = 精准只改第 N 镜(读全脚本上下文,后端强制保留其余镜原样)。""" from apps.ai.services import create_ai_task, stream_routed_text_request yield _sse({"type": "tool", "id": "skill", "label": "加载电商脚本技能", "status": "running"}) skill_loaded = bool(load_ecommerce_skill()) yield _sse({"type": "tool", "id": "skill", "status": "done" if skill_loaded else "error"}) yield _sse({"type": "tool", "id": "analyze", "label": f"分析商品:{project.product.title}", "status": "running"}) base_draft = None if mode == "revise" and base_version_id: base_draft = _load_base_draft(project, base_version_id) if base_draft is None: target_index = None # 没有基准稿就退回整版生成,单镜改无从谈起 # 改稿以基准稿的时长/镜数为准,避免请求侧默认值(前端可能硬编码 60)把 90s/6镜稿的尾镜挤掉 effective_duration = (base_draft.get("total_duration") if base_draft else None) or total_duration # 精准改一镜:镜号越界直接报错返回,绝不建任务/扣费(避免计费空转的静默 no-op) if target_index is not None and base_draft is not None: seg_n = len(base_draft.get("segments", [])) if not (0 <= target_index < seg_n): yield _sse({"type": "error", "detail": f"镜号越界:第 {target_index + 1} 镜(共 {seg_n} 镜)"}) return messages = build_agent_messages( project=project, mode=mode, user_prompt=user_prompt, selling_point_ids=selling_point_ids, base_draft=base_draft, aspect_ratio=aspect_ratio, total_duration=effective_duration, # 改稿用基准稿时长,prompt head 才不会误导模型镜数(否则模型按60s只出4镜) target_index=target_index, ) yield _sse({"type": "tool", "id": "analyze", "status": "done"}) task_type = AITask.Type.SCRIPT_OPTIMIZATION if mode == "revise" else AITask.Type.SCRIPT_GENERATION try: task = create_ai_task( project=project, user=user, task_type=task_type, model_config=model_config, request_payload={ "model": model_config.name, "endpoint": model_config.endpoint, "mode": mode, "aspect_ratio": aspect_ratio, "total_duration": total_duration, "base_version_id": str(base_version_id or ""), "target_index": target_index, "model_routing_v1": True, }, ) except Exception as exc: # noqa: BLE001 — 多为额度不足 internal_kind = "user_credit_insufficient" if str(exc).strip().lower() == "insufficient credit" else "" yield _sse(_script_error_event(exc, internal_kind=internal_kind)) return reservation = task.credit_reservation # 额度是否已结算(charge 成功 / release 失败)。客户端中途断连时,生成器被 .close() 抛 # GeneratorExit —— 它是 BaseException 不是 Exception,普通 except 抓不到,会让预扣额度冻结。 # 故用 try/finally 兜底:任何未结算路径(含断连)都释放预扣。 settled = False try: yield _sse({"type": "tool", "id": "generate", "label": "按黄金结构生成分镜", "status": "running"}) full: list[str] = [] shown = 0 forwarding = True try: task.status = AITask.Status.SUBMITTED task.submitted_at = timezone.now() task.save(update_fields=["status", "submitted_at", "updated_at"]) def validate_script_text(raw_text: str) -> dict: candidate = normalize_draft( raw_text, aspect_ratio=aspect_ratio, total_duration=effective_duration, ) if target_index is not None and base_draft: return _merge_single_segment( base_draft, candidate, target_index, aspect_ratio, effective_duration, ) return candidate routed_stream = stream_routed_text_request( task=task, primary_model=model_config, messages=messages, streaming=True, structured_output=True, business_operation="script_generate", temperature=0.85, validate_text=validate_script_text, request_summary={ "mode": mode, "target_index": target_index, "base_version_id": str(base_version_id or ""), "aspect_ratio": aspect_ratio, "total_duration": effective_duration, }, ) while True: try: ev = next(routed_stream) except StopIteration as completed: routed = completed.value break et = ev.get("type") if et == "reasoning": # 思考流:推理模型在出 JSON 前会先想很久,把思考逐字下发给前端(像对话一样可见), # 不进 full(不是答案正文,纯展示)。这是「卡在生成分镜」假死的根因修复。 rpiece = ev.get("text") or "" if rpiece: yield _sse({"type": "reasoning", "text": rpiece}) continue if et == "delta": full.append(ev["text"]) if forwarding: text = "".join(full) cut = _visible_cut(text) if cut < len(text): forwarding = False visible = text[:cut] if len(visible) > shown: piece = visible[shown:] shown = len(visible) if piece.strip(): yield _sse({"type": "delta", "text": piece}) elif et == "done": continue raw, _provider_response, draft = routed.value except Exception as exc: # noqa: BLE001 _fail_task(task, reservation, str(exc)) settled = True yield _sse({"type": "tool", "id": "generate", "status": "error"}) yield _sse(_script_error_event(exc, reference_id=str(task.id))) return yield _sse({"type": "tool", "id": "generate", "status": "done"}) yield _sse( { "type": "tool", "id": "extract", "label": f"提取实体 {len(draft['entities'])} 个 · {len(draft['segments'])} 镜", "status": "done", } ) yield _sse({"type": "tool", "id": "check", "label": "自检:镜数 / ≤55字 / 违规词", "status": "done"}) yield _sse({"type": "draft", "draft": draft}) from django.db import transaction try: with transaction.atomic(): task.status = AITask.Status.SUCCEEDED task.response_payload = {"raw": raw[:8000]} task.actual_cost = task.estimated_cost task.completed_at = timezone.now() task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"]) charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost) source = "revise" if mode == "revise" else ("theme" if mode == "theme" else "ai") script = persist_script_draft(project=project, user=user, task=task, draft=draft, source=source) settled = True # charge 已提交 except Exception as exc: # noqa: BLE001 — 落库失败:atomic 已回滚 charge,补释放预留 _fail_task(task, reservation, f"保存脚本失败:{exc}") settled = True yield _sse(_script_error_event(exc, reference_id=str(task.id), internal_kind="processing_failed")) return from apps.projects.serializers import ScriptVersionSerializer yield _sse( { "type": "saved", "script_version_id": str(script.id), "version": ScriptVersionSerializer(script).data, } ) # 模型自己写的收尾交付语 → AI 回复气泡(没写则前端兜底默认句) summary = _closing_summary(raw) if summary: yield _sse({"type": "summary", "text": summary}) yield _sse({"type": "done"}) finally: # 断连(GeneratorExit)或任何 settled=False 的退出路径:释放预扣,避免额度冻结 if not settled: _fail_task(task, reservation, "stream aborted (client disconnected)") def _fail_task(task, reservation, message: str) -> None: try: task.status = AITask.Status.FAILED task.error_message = message[:2000] task.completed_at = timezone.now() task.save(update_fields=["status", "error_message", "completed_at", "updated_at"]) finally: try: release_credit(reservation=reservation, reason=message[:200]) except Exception: # noqa: BLE001 pass def _script_error_event( exc: Exception, *, reference_id: str | None = None, internal_kind: str = "", ) -> dict: """脚本 SSE 的安全失败帧；原始异常只落任务记录/日志，不回传普通用户。""" public_error = classify_generation_error( exc, operation="script_generate", internal_kind=internal_kind, reference_id=reference_id, ) return { "type": "error", "detail": public_error.fallback_message, "error": public_error.as_dict(), } def _load_base_draft(project, base_version_id: str) -> dict | None: from apps.projects.models import ScriptVersion try: version = ScriptVersion.objects.get(project=project, id=base_version_id) except (ScriptVersion.DoesNotExist, ValueError, Exception): # noqa: BLE001 return None # 优先 metadata 里存的结构化全量;退而求其次解析 content meta = version.metadata or {} if meta.get("entities") is not None or meta.get("hook"): try: return json.loads(version.content) except (ValueError, TypeError): pass try: return json.loads(version.content) except (ValueError, TypeError): return None def _draft_from_version(version) -> dict: """从 ScriptVersion 的 DB 行(segments + metadata)重建 ScriptDraft —— 比解析可能已 stale 的 content 可靠 (用户增删/改镜后 content 不一定同步)。total_duration 按真实镜数算,避免 normalize 按 stale 值截/补镜。""" meta = version.metadata or {} segs = list(version.segments.order_by("sort_order")) return { "hook": meta.get("hook", ""), "tone": meta.get("tone", ""), "aspect_ratio": meta.get("aspect_ratio", "9:16"), "total_duration": max(int(meta.get("total_duration") or 0), 15 * len(segs)) or 60, "segment_count": len(segs), "entities": meta.get("entities", []), "segments": [ { "index": s.sort_order, "duration": s.duration_seconds or 15, "role": s.role or "", "narration": s.narration or "", "speaker": s.speaker or None, "visual": s.visual_prompt or "", "product_exposure": s.product_exposure or "", "entity_refs": s.entity_refs or [], "dialogue": s.dialogue or [], } for s in segs ], } def regenerate_segment_via_agent(*, project, user, model_config: ModelConfig, segment, instruction: str = ""): """非流式·精准改一镜(「场次刷新」按钮复用 agent 单镜逻辑):读全脚本上下文,只重写该镜,落新 ScriptVersion。 与 stream_script_agent 的 target_index 分支同源,但同步返回(不走 SSE)。计费 reserve→charge/release 闭环。""" from django.db import transaction from apps.ai.services import create_ai_task, execute_routed_text_request from apps.billing.services.ledger import charge_reserved_credit base_draft = _draft_from_version(segment.script_version) # 用 DB 行重建基准,别用可能 stale 的 content target_index = segment.sort_order aspect_ratio = (base_draft.get("aspect_ratio") or "9:16").strip() seg_n = len(base_draft.get("segments") or []) if not (0 <= target_index < seg_n): raise ValueError(f"镜号越界:第 {target_index + 1} 镜(共 {seg_n} 镜)") total_duration = base_draft["total_duration"] # 已按真实镜数算,normalize 不会截掉用户增删后的镜 messages = build_agent_messages( project=project, mode="revise", user_prompt=instruction, selling_point_ids=None, base_draft=base_draft, aspect_ratio=aspect_ratio, total_duration=total_duration, target_index=target_index, ) task = create_ai_task( project=project, user=user, task_type=AITask.Type.SCRIPT_OPTIMIZATION, model_config=model_config, request_payload={ "model": model_config.name, "endpoint": model_config.endpoint, "mode": "revise", "target_index": target_index, "model_routing_v1": True, }, ) reservation = task.credit_reservation # 每条真实尝试的平台成本由统一执行器累计；用户积分仍只结算这一条脚本任务。 task.base_cost = Decimal("0") task.save(update_fields=["base_cost", "updated_at"]) # 实际平台成本由每条 AIModelAttempt 累加；用户积分仍只结算这一条逻辑任务。 task.base_cost = Decimal("0") task.save(update_fields=["base_cost", "updated_at"]) try: task.status = AITask.Status.SUBMITTED task.submitted_at = timezone.now() task.save(update_fields=["status", "submitted_at", "updated_at"]) def validate_segment_text(raw_text: str) -> dict: candidate = normalize_draft(raw_text, aspect_ratio=aspect_ratio, total_duration=total_duration) return _merge_single_segment(base_draft, candidate, target_index, aspect_ratio, total_duration) routed = execute_routed_text_request( task=task, primary_model=model_config, messages=messages, streaming=False, structured_output=True, business_operation="script_generate", temperature=0.3, validate_text=validate_segment_text, request_summary={ "mode": "revise", "target_index": target_index, "base_version_id": str(segment.script_version_id), }, ) raw, _response, draft = routed.value with transaction.atomic(): task.status = AITask.Status.SUCCEEDED task.response_payload = {"raw": raw[:8000]} task.actual_cost = task.estimated_cost task.completed_at = timezone.now() task.save(update_fields=["status", "response_payload", "actual_cost", "completed_at", "updated_at"]) charge_reserved_credit(reservation=reservation, actual_amount=task.actual_cost) script = persist_script_draft(project=project, user=user, task=task, draft=draft, source="revise") return script except Exception as exc: _fail_task(task, reservation, str(exc) or "单镜重跑失败") raise