[{"data":1,"prerenderedAt":1588},["ShallowReactive",2],{"\u002F2026-07-02":3,"\u002F2026-07-02-rel":339},{"id":4,"title":5,"body":6,"column":321,"date":322,"description":12,"extension":323,"hero_image":324,"meta":325,"navigation":326,"path":327,"seo":328,"series_id":324,"severity":324,"stem":329,"summary":330,"tags":331,"__hash__":338},"posts\u002F2026-07-02-工作流迁移与复用.md","搬一条生产线，难的不是业务逻辑",{"type":7,"value":8,"toc":307},"minimark",[9,13,16,20,24,32,35,42,45,48,55,70,73,76,95,98,119,122,125,138,141,148,174,185,188,191,194,201,204,207,210,213,220,223,253,268,271,274,301,304],[10,11,12],"p",{},"旧平台上已经跑通了漫剧和小说两条完整的创作工作流。新平台已具备图片生成、异步任务、用户鉴权和资源计费的基础设施。现在需要把这两条链路搬过来，但不是照搬代码，而是复用领域逻辑重新适配。",[10,14,15],{},"迁移看似是一个代码挪动的问题，实际上是一个架构解耦的问题。两个平台在三个维度上产生了紧密耦合，直接挪动代码会导致新平台继承旧平台的设计包袱。",[17,18,19],"h2",{"id":19},"三处不兼容的耦合",[21,22,23],"h3",{"id":23},"任务队列与调度模型",[10,25,26,27,31],{},"旧平台的工作流采用进程内任务 Map 和命令式 API 网关。漫剧工作流从剧本拆分镜头、生成资产图、生成视频、拼接整集——这一系列操作都是通过 ",[28,29,30],"code",{},"__api\u002Fcomic_*"," 这样的命令端点来驱动的，任务状态存在内存 Map 里，重启进程就丢了。",[10,33,34],{},"新平台设计了一套资源计费底座，任务需要经历预留（reserve）→ 执行 → 结算（settle）→ 失败时退款（refund）的流程。以小说工作流为例，每个阶段（设定、世界观、角色、分卷、大纲、章节）都是独立任务，生成前要冻结估算的算力点，生成后按实际输出字符数结算，多退少补。这套计费流程嵌入到新平台的 Billing 服务里，不能绕过。",[10,36,37,38,41],{},"如果我直接把旧平台的任务调度逻辑搬过来，新平台的任务会跳过 reserve 这一步，造成\"已生成但余额不足无法扣费\"的问题。同样，旧平台的视频任务如果失败只是标记 failed，没有退款逻辑，新平台则需要原子地调用 ",[28,39,40],{},"RefundCharge","。",[10,43,44],{},"两边的任务模型在原语层面就不兼容。",[21,46,47],{"id":47},"存储路径与对象约定",[10,49,50,51,54],{},"旧平台用本地 JSON 文件存储小说项目数据。小说作品、设定、世界观、角色这些结构化阶段的原始输出被序列化到磁盘上，依赖一套约定好的目录结构。漫剧项目也类似，源项目中通过 ",[28,52,53],{},"featureDir(userId, 'comic')"," 这样的函数来确定资产文件的存储根路径。",[10,56,57,58,61,62,65,66,69],{},"新平台统一使用 S3 \u002F OOS 对象存储加 Prisma 数据库。项目的元数据（作品标题、阶段状态、版本历史）存 Prisma 表，生成的内容（图片、视频、结构化文本）存对象存储并记录 key。例如小说的 ",[28,59,60],{},"NovelSection"," 表存储化后的 JSON 结构和展示文本，章节正文存在 ",[28,63,64],{},"NovelChapterVersion"," 表的 ",[28,67,68],{},"content"," 字段。",[10,71,72],{},"如果我从旧平台直接挪过来一套\"读本地 JSON\"的逻辑，新平台就要维护两套存储系统。更重要的是，新平台的计费系统依赖持久化的结构化数据——只有把规范化后的展示文本存进表里，后台才能追溯某次扣费对应的具体内容。",[21,74,75],{"id":75},"计费模型与资源定价",[10,77,78,79,82,83,86,87,90,91,94],{},"旧平台对小说的计费可能是按生成 token 次数、按模型调用、或者干脆不计费。新平台设计了一套通用资源计费框架：每个资源有一个 key（例如 ",[28,80,81],{},"image_generation","、",[28,84,85],{},"novel_text_output","），配套一个定价模型（",[28,88,89],{},"PER_CALL"," 或 ",[28,92,93],{},"PER_UNIT","），后台可以随时调价。",[10,96,97],{},"对小说工作流，新平台的选择是按可见字符数计费，不按 token。这意味着：",[99,100,101,105,112],"ul",{},[102,103,104],"li",{},"用户最终看到的文本去掉空白后的字符数才会被计入。",[102,106,107,108,111],{},"模型原始输出如果是 JSON 格式，那么 ",[28,109,110],{},"{}"," 括号、字段名、引号、逗号、缩进都不算——只计算最终展示的内容。",[102,113,114,115,118],{},"每个阶段的计费单位是 ",[28,116,117],{},"算力点 \u002F 千字","，后台可配置，用户无感知。",[10,120,121],{},"从旧平台的某种计费方式切换到这套新模型，需要重新梳理每个环节的计费触发点。如果只是把旧的生成函数搬过来调用一遍，新平台的 Billing 根本没有机会介入。",[17,123,124],{"id":124},"迁移方案的抽象层次",[10,126,127,128,82,131,82,134,137],{},"面对这三处耦合，如果采取\"完全抽象\"的思路——把工作流定义成一个通用的步骤编排引擎，每个平台只需要实现 ",[28,129,130],{},"execute_step",[28,132,133],{},"store_result",[28,135,136],{},"charge_resource"," 这样的抽象接口——理论上很优雅，但代价是引入了一层不必要的复杂性。通用引擎要支持各种平台的差异，势必要留下很多可配置项，导致理解和维护难度上升。",[10,139,140],{},"实际采取的方案是有选择的复用：",[10,142,143,147],{},[144,145,146],"strong",{},"复用稳定部分","：工作流的步骤编排和领域逻辑。小说工作流的七个阶段顺序（设定 → 宏观 → 世界观 → 角色 → 分卷 → 拆章 → 正文）是领域知识，与平台无关。这套提示词组装、JSON 解析、规范化、展示文本提取的逻辑，从旧平台完整迁移到新平台，TypeScript 化但不改核心算法。漫剧工作流的四阶段（剧本 → 资产 → 分镜 → 成片）也是如此。",[10,149,150,153,154,82,157,82,159,82,162,165,166,169,170,173],{},[144,151,152],{},"重新实现易变部分","：任务调度、存储、计费。小说模块新增 ",[28,155,156],{},"NovelProject",[28,158,60],{},[28,160,161],{},"NovelChapter",[28,163,164],{},"NovelTask"," 等 Prisma 模型，完全按新平台的设计来。生成任务在开始前调用 ",[28,167,168],{},"billing.reserveResource()","，成功后提取展示文本、调用 ",[28,171,172],{},"billing.settleResource()","，失败时退款。这套接口与 Image Generation 任务的流程完全一致，在平台已有的基础上建造。",[10,175,176,177,180,181,184],{},"具体的模型调用、LLM 的参数、生成的超时策略这些\"具体模型调用\"的细节，留在各自的适配层。例如小说模块的 ",[28,178,179],{},"novel-generation.ts"," 只负责组装提示词、调用 LLM、解析结果，不涉及数据库操作；数据库操作交给 ",[28,182,183],{},"novel-service.ts","。这样领域逻辑与平台逻辑的边界清晰，后续不同平台可以并行维护。",[17,186,187],{"id":187},"关键的取舍决定",[10,189,190],{},"完全抽象的诱惑在于\"一套代码支持多平台\"的承诺。但这需要引入足够的可配置性和接口设计，而代价是灵活性反而下降——当某个平台需要特殊处理某个步骤时，通用引擎不得不打补丁。这在两个平台差异较大的情况下尤其成立。",[10,192,193],{},"选择有针对性的复用，意味着承认重复：数据模型要各写一份，任务调度逻辑要各写一份。但这个重复是可控的，因为它们在各自平台内部是一致的。小说模块的任务预留\u002F结算逻辑完全复用 Image Generation 已有的那套，不需要新增抽象层。",[10,195,196,197,200],{},"另一个关键的取舍是用户模型的配置。旧平台可能让用户选择小说生成用哪个模型，新平台则完全隐藏模型选择，由后台配置或环境变量决定。这简化了前端和 API 的设计——用户请求体里根本不接受 ",[28,198,199],{},"model"," 字段，Billing 也不需要按模型定价。代价是失去了\"用户自行选择成本与质量的平衡\"的灵活性，但换来了计费模型的清晰和后台的可控。",[10,202,203],{},"这类决定需要在迁移前明确：哪些能力是目标平台\"必须有\"的，哪些是\"很有但可以先不做\"的。漫剧工作流迁移时明确排除了白模（3D 白模式）视频相关的所有功能，不复制 Blender 依赖、不暴露白模提示词输入、成片阶段只处理普通视频。这个决定减少了迁移的复杂性，也避免了在新平台上重新部署 Blender 和白模渲染的基础设施。",[17,205,206],{"id":206},"稳定部分与易变部分的边界",[10,208,209],{},"这个分法的关键在于，稳定部分要确实稳定。领域逻辑（各阶段的提示词、章节的上下文组装、长篇的记忆管理）在两个平台上是一样的，因为它反映的是小说创作或漫剧创作的规律。但是，一旦涉及\"这个阶段的输入从哪里读、输出存到哪里、失败后怎么处理\"，就已经是平台相关的。",[10,211,212],{},"以小说的世界观生成为例：",[10,214,215,216,219],{},"稳定部分是 ",[28,217,218],{},"worldPrompt(previousSections, projectSettings)"," 这个函数，它组装 LLM 提示词，描述\"根据设定和前序阶段的输出，生成世界观\"。这个函数与平台无关。",[10,221,222],{},"易变部分是：",[99,224,225,232,235,246],{},[102,226,227,228,231],{},"生成前如何预留算力点（新平台调 ",[28,229,230],{},"billing.reserveResource","，旧平台可能不调）",[102,233,234],{},"生成结果是一个 JSON 对象，如何从中提取展示文本（两个平台的数据模型可能不同，但提取逻辑应该是一样的——属于稳定部分）",[102,236,237,238,241,242,245],{},"提取后的展示文本存到哪里（新平台的 ",[28,239,240],{},"NovelSection.displayText"," 字段，旧平台可能是本地文件 ",[28,243,244],{},"sections\u002Fworld.json","）",[102,247,248,249,252],{},"失败时如何处理（新平台调 ",[28,250,251],{},"billing.refundResource","，旧平台可能是清除临时文件）",[10,254,255,256,82,259,261,262,82,264,267],{},"这样划分后，新平台只需要在适配层（",[28,257,258],{},"novel-routes.ts",[28,260,183],{},"）重新实现存储和计费部分，核心生成逻辑（",[28,263,179],{},[28,265,266],{},"novel-prompts.ts","）从旧平台迁移过来。",[17,269,270],{"id":270},"实际的迁移清单",[10,272,273],{},"从这个分析，迁移的具体工作清单变成：",[275,276,277,283,289,295],"ol",{},[102,278,279,282],{},[144,280,281],{},"评估可复用的代码"," — 在旧平台找出真正与平台无关的部分。对小说工作流，这包括提示词、JSON 解析、规范化逻辑。对漫剧工作流，这包括分镜拆分的逻辑、资产提取的规则。",[102,284,285,288],{},[144,286,287],{},"定义新平台的合同"," — 每个工作流步骤的输入输出是什么，资源消耗如何计量。小说阶段的输出是规范化 JSON 加展示文本，消耗的资源是可见字符数。漫剧镜头的输出是镜头配置和首帧图片，消耗的资源是视频秒数（按模型和分辨率估算）。",[102,290,291,294],{},[144,292,293],{},"实现平台适配层"," — 数据模型（Prisma 表）、任务调度（与 Billing 集成）、错误处理和重试。这部分代码是新平台特有的，不能也不应该复用。",[102,296,297,300],{},[144,298,299],{},"测试边界"," — 在适配层写测试，验证计费逻辑、任务状态机、错误恢复。在稳定部分写测试，验证生成质量、规范化正确性。",[10,302,303],{},"不采用这样的分层，而是直接把旧平台的 Next.js 路由、本地文件操作、命令式 API 端点搬到新平台，结果是新平台沦为旧平台代码的容器。后续需要升级某个依赖、调整计费模型、或对接新的生成模型时，都会发现改一个地方影响到另一个地方。",[10,305,306],{},"设置清晰的边界，允许有选择的重复，是避免这个问题的方法。",{"title":308,"searchDepth":309,"depth":309,"links":310},"",2,[311,317,318,319,320],{"id":19,"depth":309,"text":19,"children":312},[313,315,316],{"id":23,"depth":314,"text":23},3,{"id":47,"depth":314,"text":47},{"id":75,"depth":314,"text":75},{"id":124,"depth":309,"text":124},{"id":187,"depth":309,"text":187},{"id":206,"depth":309,"text":206},{"id":270,"depth":309,"text":270},"内容流水线","2026-07-02","md",null,{},true,"\u002F2026-07-02",{"title":5,"description":12},"2026-07-02-工作流迁移与复用","两条生产工作流从旧平台迁移到新平台，难点不在业务逻辑重写，而在解耦三处平台依赖——任务调度、存储约定、计费接口。",[332,333,334,335,336,337],"架构","工作流","微服务","平台化","任务队列","计费系统","zEJUEqZgsQkHJOxHWOoc4WL1iYD_NaAli4VZYCye-ZY",[340,1055,1252],{"id":341,"title":342,"body":343,"column":321,"date":1042,"description":347,"extension":323,"hero_image":324,"meta":1043,"navigation":326,"path":1044,"seo":1045,"series_id":324,"severity":324,"stem":1046,"summary":1047,"tags":1048,"__hash__":1054},"posts\u002F2026-07-08-视频包装与数字人配音.md","配音时长不可控——那就让画面跟着它走",{"type":7,"value":344,"toc":1033},[345,348,352,358,361,364,397,400,414,417,420,427,608,621,624,635,638,641,648,658,700,707,710,721,729,738,741,752,755,758,764,799,805,837,843,860,863,869,906,912,931,936,965,968,971,974,977,991,994,997,1000,1019,1022,1029],[10,346,347],{},"数字人口播的完整链路已能生成对口型视频原片，但从音频出发反向约束画面节奏、再分层合成包装成可发布的成片，这一环涉及三个工程难点，都不是生成问题，而是一致性与路径确定性问题。",[17,349,351],{"id":350},"tts-时长驱动的反向工作流","TTS 时长驱动的反向工作流",[10,353,354,355,41],{},"MiMo TTS 客户端同步调用返回 base64 音频与实际音长（单位秒），这个时长是后续所有操作的源头。关键约束在于",[144,356,357],{},"时长由上游决定，不能人为指定",[10,359,360],{},"拿一个具体例子：用户输入 60 字的口播文案，送给 MiMo preset 模式（冰糖音色、wav 格式）。MiMo 返回的不是\"这段话应该播 30 秒\"，而是\"我合成出来的音频实际是 28.5 秒\"。TTS 生成的时长由语速、音色、模型版本等因素共同决定，变量太多了。",[10,362,363],{},"在这个约束下，整个包装流程必须反向适配：",[275,365,366,376,388],{},[102,367,368,371,372,375],{},[144,369,370],{},"TTS 合成音频"," → 得到 ",[28,373,374],{},"audioDurationSec","（来自 ffprobe 或 MiMo 返回）",[102,377,378,381,382,384,385,245],{},[144,379,380],{},"飞天对口型"," → 输入 ",[28,383,374],{},"，输出\"这个长度的人物口播视频\"（飞天返回 ",[28,386,387],{},"duration",[102,389,390,393,394,396],{},[144,391,392],{},"HyperFrames 包装"," → 模板中的时间线也以这个 ",[28,395,387],{}," 为基准构建字幕、转场、片头片尾的时间点",[10,398,399],{},"如果反过来做——先定好画面框架是 60 秒，再想办法让音频往里塞——就陷入了\"剪音频、变速播放、或者音视不同步\"的泥潭。",[10,401,402,403,406,407,410,411,413],{},"实现上，MiMo 音频落地后先上传我方 S3，用 ffprobe 测长作为权威值（",[28,404,405],{},"probeVideoDurationSec"," 对音频也适用），这个测值再被传给飞天 ",[28,408,409],{},"create_by_audio"," 接口。飞天不是按输入时长生成固定时长的视频，而是真正根据音轨长度对口型——返回的 ",[28,412,387],{}," 几乎就等于输入的音频时长（考虑 AAC 编码器的 priming delay，实测偏差 0.64 帧即 21.33ms，可忽略）。",[17,415,416],{"id":416},"数字人作为独立图层而非重新生成",[10,418,419],{},"一个常见的工程误区：既然最终成片是\"数字人 + 包装\"，能不能让 HyperFrames 直接去生成数字人？不能。飞天数字人是外部供应商接口，响应慢（异步任务，等待 2-5 分钟常态），接口也不开放生成能力给外部工具（只提供对口型 API）。再者，用户可能想用同一份配音试多个数字人形象、或试多个包装模板，重新生成整条视频的成本太高。",[10,421,422,423,426],{},"设计决策是把飞天原片（已完成对口型的 MP4 文件）当作 HyperFrames composition 里的一个 ",[28,424,425],{},"\u003Cvideo>"," 图层。模板里声明这样的结构：",[428,429,433],"pre",{"className":430,"code":431,"language":432,"meta":308,"style":308},"language-html shiki shiki-themes github-light github-dark","\u003Cdiv data-track-index=\"0\">\n  \u003Cvideo data-start=\"0s\" data-duration=\"$MAIN_VIDEO_DURATION\" \n          data-has-audio=\"true\" \n          src=\"$MAIN_VIDEO_URL\">\n  \u003C\u002Fvideo>\n\u003C\u002Fdiv>\n\n\u003Cdiv data-track-index=\"1\">\n  \u003C!-- 字幕、花字、角标等包装层 -->\n\u003C\u002Fdiv>\n\n\u003Cdiv data-track-index=\"2\">\n  \u003C!-- 片头片尾、转场 -->\n\u003C\u002Fdiv>\n","html",[28,434,435,462,489,501,514,524,534,540,556,563,572,577,593,599],{"__ignoreMap":308},[436,437,440,444,448,452,455,459],"span",{"class":438,"line":439},"line",1,[436,441,443],{"class":442},"sVt8B","\u003C",[436,445,447],{"class":446},"s9eBZ","div",[436,449,451],{"class":450},"sScJk"," data-track-index",[436,453,454],{"class":442},"=",[436,456,458],{"class":457},"sZZnC","\"0\"",[436,460,461],{"class":442},">\n",[436,463,464,467,470,473,475,478,481,483,486],{"class":438,"line":309},[436,465,466],{"class":442},"  \u003C",[436,468,469],{"class":446},"video",[436,471,472],{"class":450}," data-start",[436,474,454],{"class":442},[436,476,477],{"class":457},"\"0s\"",[436,479,480],{"class":450}," data-duration",[436,482,454],{"class":442},[436,484,485],{"class":457},"\"$MAIN_VIDEO_DURATION\"",[436,487,488],{"class":442}," \n",[436,490,491,494,496,499],{"class":438,"line":314},[436,492,493],{"class":450},"          data-has-audio",[436,495,454],{"class":442},[436,497,498],{"class":457},"\"true\"",[436,500,488],{"class":442},[436,502,504,507,509,512],{"class":438,"line":503},4,[436,505,506],{"class":450},"          src",[436,508,454],{"class":442},[436,510,511],{"class":457},"\"$MAIN_VIDEO_URL\"",[436,513,461],{"class":442},[436,515,517,520,522],{"class":438,"line":516},5,[436,518,519],{"class":442},"  \u003C\u002F",[436,521,469],{"class":446},[436,523,461],{"class":442},[436,525,527,530,532],{"class":438,"line":526},6,[436,528,529],{"class":442},"\u003C\u002F",[436,531,447],{"class":446},[436,533,461],{"class":442},[436,535,537],{"class":438,"line":536},7,[436,538,539],{"emptyLinePlaceholder":326},"\n",[436,541,543,545,547,549,551,554],{"class":438,"line":542},8,[436,544,443],{"class":442},[436,546,447],{"class":446},[436,548,451],{"class":450},[436,550,454],{"class":442},[436,552,553],{"class":457},"\"1\"",[436,555,461],{"class":442},[436,557,559],{"class":438,"line":558},9,[436,560,562],{"class":561},"sJ8bj","  \u003C!-- 字幕、花字、角标等包装层 -->\n",[436,564,566,568,570],{"class":438,"line":565},10,[436,567,529],{"class":442},[436,569,447],{"class":446},[436,571,461],{"class":442},[436,573,575],{"class":438,"line":574},11,[436,576,539],{"emptyLinePlaceholder":326},[436,578,580,582,584,586,588,591],{"class":438,"line":579},12,[436,581,443],{"class":442},[436,583,447],{"class":446},[436,585,451],{"class":450},[436,587,454],{"class":442},[436,589,590],{"class":457},"\"2\"",[436,592,461],{"class":442},[436,594,596],{"class":438,"line":595},13,[436,597,598],{"class":561},"  \u003C!-- 片头片尾、转场 -->\n",[436,600,602,604,606],{"class":438,"line":601},14,[436,603,529],{"class":442},[436,605,447],{"class":446},[436,607,461],{"class":442},[10,609,610,613,614,616,617,620],{},[28,611,612],{},"data-has-audio=\"true\""," 这一行是命门——没有它，渲染器会给 ",[28,615,425],{}," 默认加 ",[28,618,619],{},"muted","，成片就没声音。",[10,622,623],{},"这样做的好处显而易见：",[99,625,626,629,632],{},[102,627,628],{},"数字人原片一旦生成就不动，支持单独重做配音（只需重新调用 TTS 和飞天，不影响包装渲染）",[102,630,631],{},"同一个配音可套多个模板，不需要重新对口型",[102,633,634],{},"包装层的 HTML\u002FCSS 改动不会触发数字人重生成，迭代快",[10,636,637],{},"缺点是需要镜像内内置 Chromium 和 FFmpeg（官方渲染镜像实测 3.72GB），与 API 镜像分离部署。但这换来的是确定性输出和可控的环境——生产机、开发机、CI 跑同一个 composition，出片应该帧级一致（除了字体、系统库这类版本差异导致的微调，都锁版本了）。",[17,639,640],{"id":640},"成片路径的幂等性与状态机",[10,642,643,644,647],{},"用户选定模板、确认方案后，系统提交一个 ",[28,645,646],{},"VideoRenderJob","：输入模板 ID、原片 key、字幕 cues、音频 URL 等，输出成片 objectKey。同一份输入如果重试或重新提交，必须得到同样的输出（或者快速失败）。",[10,649,650,651,90,654,657],{},"状态流转是 ",[28,652,653],{},"queued → preparing → rendering → uploading → completed",[28,655,656],{},"failed","：",[99,659,660,670,676,689,695],{},[102,661,662,665,666,669],{},[144,663,664],{},"queued","：任务入队，等待 worker 消费。这一步是防并发上限的信号量检查——飞天有并发限制（错误码 1001），Redis 信号量 ",[28,667,668],{},"yunclaude:dub:sky:sem"," 兜底（触顶不直接失败，改为入队等待）。",[102,671,672,675],{},[144,673,674],{},"preparing","：下载原片、组装模板（注入数据、渲染占位符）。这一步的幂等性来自 objectKey 的内容寻址——同一个原片 key、同一个模板版本，组装出的 composition.html 比特级相同。",[102,677,678,681,682,82,685,688],{},[144,679,680],{},"rendering","：Chromium 逐帧捕获、FFmpeg 合成。这是确定性的关键——环境锁定（Node 24、Chromium 版本、FFmpeg 版本、字体版本都在镜像里写死），同一个 composition 渲染多次输出帧级一致。M1 POC 中用 seek 点验证：",[28,683,684],{},"t=4.5s→第 135 帧",[28,686,687],{},"t=30.0s→第 900 帧","（读原片内置计数器，nb_frames=1800），长时间点零漂移。",[102,690,691,694],{},[144,692,693],{},"uploading","：上传 OSS，记录 objectKey。返回给前端时用现签（每次读时重新签，不存短时 URL）。",[102,696,697,699],{},[144,698,656],{},"：任何一步异常，立即 rollback。计费侧已扣的视频点全额退款（resource operationId 幂等）。",[10,701,702,703,706],{},"失败的兜底是 reaper（BullMQ 的死信队列处理），轮询 ",[28,704,705],{},"status=running"," 超期（>10 分钟）的任务，标记为失败并补退款。",[17,708,709],{"id":709},"音画同步与字幕对齐",[10,711,712,713,716,717,720],{},"成片里字幕何时出现、何时消失，这些时间点由分析步生成的 ",[28,714,715],{},"subtitleCues"," 定义（格式 WebVTT）。一个 cue 的结构是 ",[28,718,719],{},"start → end"," + 文本，例如：",[428,722,727],{"className":723,"code":725,"language":726},[724],"language-text","00:05.000 --> 00:08.500\n这是一段口播文案\n","text",[28,728,725],{"__ignoreMap":308},[10,730,731,732,734,735,737],{},"HyperFrames 的字幕图层根据这些 cues 生成动画：start 时刻淡入，end 时刻淡出。整个成片的时间线参考都来自主视频（",[28,733,425],{}," 元素），而主视频的时长就是飞天返回的 ",[28,736,387],{},"——由音频长度决定。",[10,739,740],{},"一个细节：AAC-LC 编码器有 priming delay（约 1024 samples@48kHz = 21.33ms），所以成片里音频起点和视频起点存在一个已知的 21ms 偏移。但这是编码器层的常数，不是渲染问题，接受即可。",[10,742,743,744,747,748,751],{},"关键词高亮（比如把卖点词着色为黄色）需要在分析时做标记，例如 HTML 标签：",[28,745,746],{},"\u003Cspan class=\"highlight\">关键词\u003C\u002Fspan>","，然后 CSS 定义颜色。模板编写规约里明确禁止动画 ",[28,749,750],{},"letterSpacing"," 等布局属性（会在逐帧捕获时 snap 到整数像素产生抖动），只允许 transform（x\u002Fy\u002Fscale\u002Fopacity）。",[17,753,754],{"id":754},"成片路径的端到端烟测",[10,756,757],{},"发版前的验收分六个阶段：",[10,759,760,763],{},[144,761,762],{},"A. 发版前置","（缺一项线上就炸）",[99,765,766,777,783,793,796],{},[102,767,768,769,772,773,776],{},"后台已配渲染单价（resourceKey ",[28,770,771],{},"video_render_sec","），且 ",[28,774,775],{},"enabled"," 勾选",[102,778,779,780,245],{},"灰度名单已配置（环境变量 ",[28,781,782],{},"VIDEO_RENDER_HTML_USER_IDS",[102,784,785,786,789,790,245],{},"发版机 ",[28,787,788],{},"release.sh"," 已改（支持第 6 个镜像 ",[28,791,792],{},"yc-video-render",[102,794,795],{},"K8s 集群配额已提升（requestQuota 新增 2C\u002F2Gi、limitsQuota 新增 4C\u002F4Gi）",[102,797,798],{},"构建机磁盘充足（≥10GB 空闲）",[10,800,801,804],{},[144,802,803],{},"B. 发版后基础设施","（不通过立即 rollout undo）",[99,806,807,813,820,827,834],{},[102,808,809,810,812],{},"迁移已执行（",[28,811,646],{}," 表已建）",[102,814,815,816,819],{},"渲染 worker 已起（",[28,817,818],{},"READY 1\u002F1","，镜像 tag 与本次发版一致）",[102,821,822,823,826],{},"容器内 hyperframes CLI 可用（",[28,824,825],{},"hyperframes --version"," 返回 0.7.70）",[102,828,829,830,833],{},"容器内中文字体已装（",[28,831,832],{},"fc-list :lang=zh"," 非空）",[102,835,836],{},"worker 连上 Redis 队列（日志无 crash loop）",[10,838,839,842],{},[144,840,841],{},"C. 回归防线：未放量用户零感知","（最高优先级）",[99,844,845,848,851,854],{},[102,846,847],{},"不在灰度名单的用户进增强步看不到\"包装模板\"卡",[102,849,850],{},"旧链路（ffmpeg）完整出片，字幕\u002F转场\u002FBGM 都在",[102,852,853],{},"旧链路仍生成 AI 插片，计费时间线出现 seedance 扣费",[102,855,856,857,859],{},"未放量时 ",[28,858,646],{}," 表没有新行",[10,861,862],{},"这段最重要是因为旧链路服务着所有现存数字人用户。一个新功能开关不应该波及灰度外的用户。",[10,864,865,868],{},[144,866,867],{},"D. 灰度用户正向流程","（核心价值）",[99,870,871,874,881,888,891,894,897,900,903],{},[102,872,873],{},"模板列表可见：增强步看到\"不加包装\"+\"美食探店\"两张卡",[102,875,876,877,880],{},"选模板后出片最终 ",[28,878,879],{},"completed","，可播放",[102,882,883,884,887],{},"成片有口播声音（这是 ",[28,885,886],{},"data-has-audio"," 命门检验）",[102,889,890],{},"片头\u002F片尾\u002F角标都在：0-3s 品牌片头、右上角全程角标、最后 3s CTA",[102,892,893],{},"中文不乱码：片头标题与字幕汉字正常",[102,895,896],{},"字幕关键词高亮：关键词黄色，其余白色，标签没被打碎",[102,898,899],{},"进度实时可见：出片过程中进度条推进，不死在一个数字",[102,901,902],{},"成片链接可下载且长期有效：15 分钟后刷新页面仍能播放",[102,904,905],{},"渲染时长符合预期：60s 成片≤5 分钟（M1 POC 实测 39.5s）",[10,907,908,911],{},[144,909,910],{},"E. 资金正确性","（看后台账单，不能只看页面）",[99,913,914,919,922,925,928],{},[102,915,916,917,245],{},"扣的是视频点不是算力点（计费时间线条目 title 首段是 ",[28,918,469],{},[102,920,921],{},"不再扣 seedance 插片钱（灰度用户的时间线里没有插片扣费）",[102,923,924],{},"结算按实际秒数（units ≈ 成片时长整秒向上取整）",[102,926,927],{},"余额不足时不产生任务、不扣费",[102,929,930],{},"重复提交不重复扣费（operationId 幂等）",[10,932,933],{},[144,934,935],{},"F. 异常路径",[99,937,938,944,950,953,956,959],{},[102,939,940,941,943],{},"渲染失败会退款：任务置 ",[28,942,656],{},"，计费时间线出现退款条目",[102,945,946,947,949],{},"失败同步反映到增强任务：增强任务也变 ",[28,948,656],{},"，前端不会永远停在\"渲染中\"",[102,951,952],{},"排队中可取消并全额退：返回成功，退款到账",[102,954,955],{},"已在渲染中不可取消：返回 409，不退款（CPU 已烧）",[102,957,958],{},"worker 重启不丢任务：任务被重新捞起或置失败退款",[102,960,961,962,964],{},"超时判失败：>10 分钟置 ",[28,963,656],{}," 并退款",[10,966,967],{},"每一条都写了具体的检查命令和判据。比如验证成片有声音，就是直接播放、耳朵听；验证字幕关键词高亮，就是肉眼看颜色；验证退款，就是对比出片前后的计费时间线。",[17,969,970],{"id":970},"设计的权衡",[10,972,973],{},"这套方案的成本是什么？",[10,975,976],{},"首先是镜像大小：官方渲染镜像装了 280 多个 apt 包、node_modules、Chromium 和 FFmpeg，构建出的镜像实测 3.72GB，单独占用一个 K8s node pool，构建时间约 10 分钟，推拉镜像耗时显著上升。但这是\"要么装进 API 镜像肥到 6GB+、拖累三个部署，要么单独镜像\"的取舍——选了单独。",[10,978,979,980,982,983,986,987,990],{},"其次是模板编写规约的学习成本。一个看起来\"普通的网页动画\"可能在逐帧 seek 渲染时炸：GSAP 退场动画必须挂在 clip 内层并补 hard kill，禁止 ",[28,981,750],{}," 等布局属性，中文字体必须显式 ",[28,984,985],{},"@font-face","（容器内用 Noto Sans CJK）。",[28,988,989],{},"hyperframes check"," 作为模板上架的强制 gate，能一次性拦截这类问题。",[10,992,993],{},"获得的是什么？確定性输出。同一个模板、同一份配音，渲染 100 次得到 100 个帧级一致的成片（前提是输入稳定，不变数字人形象、不变字体库版本）。这对 CI 自动回归、成片质量审核都很有意义。还有灵活性：配音、数字人、包装模板可以独立迭代，不互相阻塞。",[17,995,996],{"id":996},"线上部署的最后两步",[10,998,999],{},"发版后的两个关键项，缺一项都会导致灰度用户无法下单：",[275,1001,1002,1011],{},[102,1003,1004,657,1007,1010],{},[144,1005,1006],{},"后台配价",[28,1008,1009],{},"resourceKey=video_render_sec"," 的单价必须填（单位元\u002F秒），且勾选 enabled。没配的话 chargeResource 会返回 404，用户点开始出片就直接报错。",[102,1012,1013,657,1016,1018],{},[144,1014,1015],{},"灰度名单",[28,1017,782],{}," 环境变量决定了谁能看到模板卡。留空 = 全员走旧 ffmpeg 链路（可用于紧急回滚），指定用户 ID = 该用户进入新链路。发版初期应只填 1-2 个测试账号。",[10,1020,1021],{},"这两个配置是代码之外的硬依赖，容易遗漏。烟测清单里放在最前面，作为\"不做后面全走不通\"的前置项。",[10,1023,1024,1025,1028],{},"整个方案的核心原则是",[144,1026,1027],{},"分层隔离与路径确定性","：TTS 决定时长、飞天决定人像、HyperFrames 决定包装，各层独立演进，成片路径从输入到输出一条流水线，无分支、无条件、无随机。这对一个产生可发布物料的流水线来说，是底线。",[1030,1031,1032],"style",{},"html pre.shiki code .sVt8B, html code.shiki .sVt8B{--shiki-default:#24292E;--shiki-dark:#E1E4E8}html pre.shiki code .s9eBZ, html code.shiki .s9eBZ{--shiki-default:#22863A;--shiki-dark:#85E89D}html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html pre.shiki code .sJ8bj, html code.shiki .sJ8bj{--shiki-default:#6A737D;--shiki-dark:#6A737D}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":308,"searchDepth":309,"depth":309,"links":1034},[1035,1036,1037,1038,1039,1040,1041],{"id":350,"depth":309,"text":351},{"id":416,"depth":309,"text":416},{"id":640,"depth":309,"text":640},{"id":709,"depth":309,"text":709},{"id":754,"depth":309,"text":754},{"id":970,"depth":309,"text":970},{"id":996,"depth":309,"text":996},"2026-07-08",{},"\u002F2026-07-08",{"title":342,"description":347},"2026-07-08-视频包装与数字人配音","数字人口播成片的最后一环：配音时长不可控，如何反向驱动画面？原片如何作为独立图层无损合成？成片路径如何保证确定性？",[1049,1050,1051,1052,1053],"TTS","数字人","HyperFrames","音视同步","流水线设计","ISg-n2hvjyCKUW_WMTk0Ais8ji7YhOK5rXPmNyUBWQk",{"id":4,"title":5,"body":1056,"column":321,"date":322,"description":12,"extension":323,"hero_image":324,"meta":1249,"navigation":326,"path":327,"seo":1250,"series_id":324,"severity":324,"stem":329,"summary":330,"tags":1251,"__hash__":338},{"type":7,"value":1057,"toc":1238},[1058,1060,1062,1064,1066,1070,1072,1076,1078,1080,1084,1092,1094,1096,1106,1108,1120,1122,1124,1132,1134,1138,1154,1160,1162,1164,1166,1170,1172,1174,1176,1178,1182,1184,1202,1212,1214,1216,1234,1236],[10,1059,12],{},[10,1061,15],{},[17,1063,19],{"id":19},[21,1065,23],{"id":23},[10,1067,26,1068,31],{},[28,1069,30],{},[10,1071,34],{},[10,1073,37,1074,41],{},[28,1075,40],{},[10,1077,44],{},[21,1079,47],{"id":47},[10,1081,50,1082,54],{},[28,1083,53],{},[10,1085,57,1086,61,1088,65,1090,69],{},[28,1087,60],{},[28,1089,64],{},[28,1091,68],{},[10,1093,72],{},[21,1095,75],{"id":75},[10,1097,78,1098,82,1100,86,1102,90,1104,94],{},[28,1099,81],{},[28,1101,85],{},[28,1103,89],{},[28,1105,93],{},[10,1107,97],{},[99,1109,1110,1112,1116],{},[102,1111,104],{},[102,1113,107,1114,111],{},[28,1115,110],{},[102,1117,114,1118,118],{},[28,1119,117],{},[10,1121,121],{},[17,1123,124],{"id":124},[10,1125,127,1126,82,1128,82,1130,137],{},[28,1127,130],{},[28,1129,133],{},[28,1131,136],{},[10,1133,140],{},[10,1135,1136,147],{},[144,1137,146],{},[10,1139,1140,153,1142,82,1144,82,1146,82,1148,165,1150,169,1152,173],{},[144,1141,152],{},[28,1143,156],{},[28,1145,60],{},[28,1147,161],{},[28,1149,164],{},[28,1151,168],{},[28,1153,172],{},[10,1155,176,1156,180,1158,184],{},[28,1157,179],{},[28,1159,183],{},[17,1161,187],{"id":187},[10,1163,190],{},[10,1165,193],{},[10,1167,196,1168,200],{},[28,1169,199],{},[10,1171,203],{},[17,1173,206],{"id":206},[10,1175,209],{},[10,1177,212],{},[10,1179,215,1180,219],{},[28,1181,218],{},[10,1183,222],{},[99,1185,1186,1190,1192,1198],{},[102,1187,227,1188,231],{},[28,1189,230],{},[102,1191,234],{},[102,1193,237,1194,241,1196,245],{},[28,1195,240],{},[28,1197,244],{},[102,1199,248,1200,252],{},[28,1201,251],{},[10,1203,255,1204,82,1206,261,1208,82,1210,267],{},[28,1205,258],{},[28,1207,183],{},[28,1209,179],{},[28,1211,266],{},[17,1213,270],{"id":270},[10,1215,273],{},[275,1217,1218,1222,1226,1230],{},[102,1219,1220,282],{},[144,1221,281],{},[102,1223,1224,288],{},[144,1225,287],{},[102,1227,1228,294],{},[144,1229,293],{},[102,1231,1232,300],{},[144,1233,299],{},[10,1235,303],{},[10,1237,306],{},{"title":308,"searchDepth":309,"depth":309,"links":1239},[1240,1245,1246,1247,1248],{"id":19,"depth":309,"text":19,"children":1241},[1242,1243,1244],{"id":23,"depth":314,"text":23},{"id":47,"depth":314,"text":47},{"id":75,"depth":314,"text":75},{"id":124,"depth":309,"text":124},{"id":187,"depth":309,"text":187},{"id":206,"depth":309,"text":206},{"id":270,"depth":309,"text":270},{},{"title":5,"description":12},[332,333,334,335,336,337],{"id":1253,"title":1254,"body":1255,"column":321,"date":1574,"description":1259,"extension":323,"hero_image":324,"meta":1575,"navigation":326,"path":1576,"seo":1577,"series_id":324,"severity":324,"stem":1578,"summary":1579,"tags":1580,"__hash__":1587},"posts\u002F2026-06-25-长文写作的记忆分层.md","三十万字之后，AI 该记住什么？",{"type":7,"value":1256,"toc":1565},[1257,1260,1263,1266,1269,1272,1275,1278,1282,1285,1291,1299,1305,1310,1316,1321,1327,1332,1343,1346,1350,1353,1359,1364,1370,1375,1380,1385,1391,1396,1402,1407,1410,1413,1417,1420,1423,1426,1429,1432,1439,1477,1488,1499,1502,1505,1514,1523,1536,1549,1552,1559,1562],[10,1258,1259],{},"长篇小说写到三十万字后，\"记住前面写了什么\"变成最大的瓶颈。角色在第五章确立的设定，到第二十章突然对不上；伏笔埋了二十章没回收；世界观里的时间线出现逻辑洞；前面说过的势力关系后面又改了。全量注入上一章或全书到上下文里根本不可能，纯向量检索又容易因为相似度不够而漏掉关键的硬约束。",[10,1261,1262],{},"这个问题需要分层的记忆系统。根据信息的稳定性和用途把记忆分成三层：设定层、事实层、文本层。每层的存储方式和检索策略完全不同。",[17,1264,1265],{"id":1265},"为什么不能统一用向量库",[10,1267,1268],{},"一个很自然的想法是\"把所有信息存到向量库里，需要时向量检索\"。但设定信息不行。",[10,1270,1271],{},"世界规则、人物档案这类信息是硬约束，一旦因为向量相似度不够高而没被召回，就会产生直接的设定冲突。模型可能生成\"主角这章掌握了某个禁忌知识\"，但系统没召回\"世界规则里明确说这个知识是自杀性的\"，结果后续剧情完全跑偏。向量检索是概率性的，概率性检索不能承载硬约束。",[10,1273,1274],{},"事实信息（已发生的事件、角色状态变化、伏笔）可以用向量检索，因为\"不完美的匹配\"还能通过更多文本从模型推理出来。如果系统检索到\"角色在某章失去了一条胳膊\"这个事实后又检索到\"这个事件在剧情中的含义是……\"，模型有足够的上下文修复漏掉的细节。",[10,1276,1277],{},"原文（正文片段）是纯量级的数据，全量注入的成本太高，用摘要替代是更经济的做法。",[17,1279,1281],{"id":1280},"设定层结构化存储全量或定向注入","设定层：结构化存储、全量或定向注入",[10,1283,1284],{},"设定层装的是世界级的硬规则和静态档案，包括：",[10,1286,1287,1290],{},[144,1288,1289],{},"世界规则","（约 600 字压缩）",[99,1292,1293,1296],{},[102,1294,1295],{},"现实是否稳定、超常能力是否公开、死亡是否可逆、信息获取是否受限",[102,1297,1298],{},"这个世界\"允许什么、不允许什么\"",[10,1300,1301,1304],{},[144,1302,1303],{},"系统性规则","（通常 4-6 条）",[99,1306,1307],{},[102,1308,1309],{},"力量体系的上限、交易与代价的原理、禁忌知识的危害方式",[10,1311,1312,1315],{},[144,1313,1314],{},"主要势力","（通常 3-5 个）",[99,1317,1318],{},[102,1319,1320],{},"每个势力的名称、目标、常用手段、与其他势力的关系",[10,1322,1323,1326],{},[144,1324,1325],{},"核心人物档案","（通常 4-8 个）",[99,1328,1329],{},[102,1330,1331],{},"每个人物的身份、核心目标、掌握的信息、心理底线",[10,1333,1334,1335,1338,1339,1342],{},"这些东西在整部小说中是不变的（或变化极缓），写作的任何环节都需要遵循它们。设定层必须在调用 AI 生成章节时",[144,1336,1337],{},"全量注入","或",[144,1340,1341],{},"按需定向注入","。全量注入适合小说世界相对简洁的情况；如果世界复杂设定众多，可以做定向注入——比如这一章涉及势力 A，就只注入 A 的档案和相关规则。",[10,1344,1345],{},"关键是：设定层的记忆片段永远不能因为上下文窗口压力而被省略。这是整个系统的天花板。",[17,1347,1349],{"id":1348},"事实层向量检索-时间过滤","事实层：向量检索 + 时间过滤",[10,1351,1352],{},"事实层装的是已经发生的、会影响后续剧情的信息。",[10,1354,1355,1358],{},[144,1356,1357],{},"章节摘要","（每章一句话）",[99,1360,1361],{},[102,1362,1363],{},"\"主角从势力 B 得到了关键物品 X，但同时被势力 A 注意到了\"",[10,1365,1366,1369],{},[144,1367,1368],{},"角色动态状态","（追加式、非覆盖）",[99,1371,1372],{},[102,1373,1374],{},"某角色掌握了什么新信息、失去了什么能力、和某人的关系发生了什么变化",[10,1376,1377],{},[144,1378,1379],{},"伏笔记录",[99,1381,1382],{},[102,1383,1384],{},"埋下的伏笔及其状态（待回收 \u002F 已回收）、涉及的章号",[10,1386,1387,1390],{},[144,1388,1389],{},"时间线","（关键事件及其时间距离）",[99,1392,1393],{},[102,1394,1395],{},"\"第五章后的第三天，X 事件发生\"",[10,1397,1398,1401],{},[144,1399,1400],{},"世界新设定","（剧情中确立、原设定没有的）",[99,1403,1404],{},[102,1405,1406],{},"\"这个世界原本不知道 X，但第十二章中 Y 揭露了 X\"",[10,1408,1409],{},"这一层用向量检索的原因是数据量大（几十万字的小说可能有几百条事实）且不需要完美精确。模型在知道\"五章前主角失去了左腿\"和\"十章前主角加入了某组织\"的基础上，能够合理推理出后续事件。即使系统漏掉了某条非关键事实，模型也不太会产生硬冲突。",[10,1411,1412],{},"时间过滤很重要。检索结果应该默认优先最近的事件，因为离当前章节越近的事件通常影响力越大。\"三章前角色的转折\"比\"五十章前的背景\"更应该被注入。",[17,1414,1416],{"id":1415},"文本层只取最近段落","文本层：只取最近段落",[10,1418,1419],{},"文本层就是原始的正文片段，用于衔接语气和细节。一个几十万字的小说，全部原文根本放不进上下文。",[10,1421,1422],{},"解决方案是简单的：只注入前一章的结尾（约 800 字），这足以让模型维持住文笔连贯性和情感线的延续。对于需要回顾很久之前的情节细节的场景，用事实层的摘要替代——\"第五章中，X 因为 Y 而死亡\"这一条摘要比翻出整个第五章的原文高效得多。",[10,1424,1425],{},"当然也存在\"某章需要直接引用或高度呼应某个很久以前的场景细节\"的情况。这时候文本层可以临时扩大范围，但这是特例，不是常态。",[17,1427,1428],{"id":1428},"具体的实现策略",[10,1430,1431],{},"设定层在每次生成章节前直接注入——要么全部、要么按这章涉及的范围选择。不需要任何检索逻辑，就是结构化的数据块。",[10,1433,1434,1435,1438],{},"事实层维护一个 ",[28,1436,1437],{},"memory.json"," 文件，存储：",[99,1440,1441,1447,1453,1459,1465,1471],{},[102,1442,1443,1446],{},[28,1444,1445],{},"synopsis","：全书概要，每章更新后重写，封顶 1000 字",[102,1448,1449,1452],{},[28,1450,1451],{},"chapterDigests","：逐章一句话摘要",[102,1454,1455,1458],{},[28,1456,1457],{},"characters","：角色的动态状态（与静态档案分开）",[102,1460,1461,1464],{},[28,1462,1463],{},"foreshadow","：伏笔列表，标记待回收或已回收",[102,1466,1467,1470],{},[28,1468,1469],{},"timeline","：关键事件及章号和相对时间",[102,1472,1473,1476],{},[28,1474,1475],{},"worldFacts","：剧情中新确立的设定事实",[10,1478,1479,1480,1483,1484,1487],{},"每章生成完成后，系统调用一次 AI，输入\"旧记忆 + 本章正文\"，要求输出",[144,1481,1482],{},"增量","——这一章新增了什么伏笔、角色状态如何变化、有没有新设定。然后用纯函数 ",[28,1485,1486],{},"normalizeMemory()"," 合并增量到旧记忆里：新伏笔追加、已回收伏笔置状态、角色按 name upsert（同 ID 的覆盖）、synopsis 重写、timeline 追加。",[10,1489,1490,1491,1494,1495,1498],{},"生成下一章时，",[28,1492,1493],{},"buildChapterMessages()"," 的流程变成：注入设定块 → 注入记忆块（由 ",[28,1496,1497],{},"buildMemoryContext()"," 生成） → 注入上一章结尾 → 生成新章。记忆块里包含：最近 6 章摘要、全部角色现状、最后 8 条未回收伏笔、最后 5 条时间线、最后 10 条世界新设定。这些都是硬上限，防止记忆块因为章数增多而无限膨胀。",[17,1500,1501],{"id":1501},"一致性与失败处理",[10,1503,1504],{},"这套流水线的核心难点不是单点生成质量，而是一致性和失败恢复。",[10,1506,1507,1510,1511,1513],{},[144,1508,1509],{},"一致性来自增量而不是合并","。每章记忆更新要求 AI 只输出增量，而不是整个记忆的重写。这样设计有两个好处：一是防止 AI 每次都改掉前面的内容（一种无意义的膨胀），二是增量小得多，解析 JSON 时出错的概率降低。合并逻辑交给纯函数 ",[28,1512,1486],{},"，这个函数可以单测，保证逻辑稳定。",[10,1515,1516,1519,1520,1522],{},[144,1517,1518],{},"同一章重建幂等","。由于 ",[28,1521,1451],{}," 按章号存储，相同章号的摘要会覆盖而不是追加，所以即使记忆重建时对同一章调用多次，也不会出现重复记录。角色状态的 upsert 机制也是同理。",[10,1524,1525,1528,1529,1531,1532,1535],{},[144,1526,1527],{},"记忆更新失败不影响正文","。每章正文先落盘、记忆更新在后。如果记忆更新失败（比如 AI 返回格式错误、JSON 解析异常），降级处理：用这章的 title 或 summary 作为摘要追加到 ",[28,1530,1451],{},"，",[28,1533,1534],{},"lastChapterNo"," 照常推进，其余记忆保持不变。这样下一章的生成可以继续进行，用户不会察觉到记忆层的故障。",[10,1537,1538,1541,1542,1545,1546,1548],{},[144,1539,1540],{},"重建任务的断点续跑","。存量书如果想补全记忆，后台可以跑 ",[28,1543,1544],{},"runMemoryRebuild()"," 任务，按章序遍历，跳过 ",[28,1547,1534],{}," 以内的章（已纳入过的）。这个任务可停、可继续，防止重复扣费。",[17,1550,1551],{"id":1551},"记忆分层的边界",[10,1553,1554,1555,1558],{},"这个设计对应的问题空间是：",[144,1556,1557],{},"长篇写作中保持内容一致性和连贯性","。它解决的是系统层面的记忆管理，不覆盖人工的内容审校和设定调整。",[10,1560,1561],{},"如果作者在某个时刻决定\"我要改前面某角色的设定\"，那是人为修改设定层、然后手动落盘的过程。系统的记忆不会自动同步这种改动，因为改动涉及主观判断。类似的，如果某章生成出来的内容和记忆产生冲突（比如模型莫名其妙加了个新势力），那也需要人工介入编辑，而不是靠记忆系统自动修正。",[10,1563,1564],{},"记忆系统的职责是：提供足够的上下文约束，让模型生成时减少无意义的冲突；一旦冲突出现，系统快速降级，不中断工作流。",{"title":308,"searchDepth":309,"depth":309,"links":1566},[1567,1568,1569,1570,1571,1572,1573],{"id":1265,"depth":309,"text":1265},{"id":1280,"depth":309,"text":1281},{"id":1348,"depth":309,"text":1349},{"id":1415,"depth":309,"text":1416},{"id":1428,"depth":309,"text":1428},{"id":1501,"depth":309,"text":1501},{"id":1551,"depth":309,"text":1551},"2026-06-25",{},"\u002F2026-06-25",{"title":1254,"description":1259},"2026-06-25-长文写作的记忆分层","长篇到几十万字后，用分层记忆解决剧情断裂和设定崩坏——设定层硬约束、事实层概率检索、文本层按需取用，层级的存储与检索策略完全不同。",[1581,1582,1583,1584,1585,1586],"长篇写作","记忆管理","内容生成","AI辅助创作","向量检索","一致性维护","IK_-YIn1bPi-9X-l1KlQn3a-D0cTe45drq3MhS7m-AA",1785406912221]