@aswless_854771076/ai_short_studio_cli 0.1.153 → 0.1.154

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (27) hide show
  1. package/README.md +1 -1
  2. package/dist/commands/canvas/node/run.js +15 -6
  3. package/dist/commands/task/wait.d.ts +2 -0
  4. package/dist/commands/task/wait.js +11 -1
  5. package/dist/domain/vvicat-api.d.ts +3 -0
  6. package/dist/domain/vvicat-api.js +40 -0
  7. package/package.json +1 -1
  8. package/skills/using-vvicat-ai-short-studio-cli/SKILL.md +1 -1
  9. package/skills/using-vvicat-ai-short-studio-cli/references/assets-and-storyboard.md +5 -5
  10. package/skills/using-vvicat-ai-short-studio-cli/references/async-pipeline.md +1 -0
  11. package/skills/using-vvicat-ai-short-studio-cli/references/commands/assets-and-tasks.md +2 -2
  12. package/skills/using-vvicat-ai-short-studio-cli/references/hypit/SKILL.md +1 -1
  13. package/skills/using-vvicat-ai-short-studio-cli/references/hypit/playbooks/formats/short-drama-workflow.md +1 -1
  14. package/skills/using-vvicat-ai-short-studio-cli/references/project-and-canvas.md +2 -2
  15. package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/design-drama-audio/prompts/voice_analysis.en.txt +1 -1
  16. package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/design-drama-audio/prompts/voice_analysis.zh.txt +1 -1
  17. package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/plan-drama-production/prompts/production_plan.en.txt +8 -5
  18. package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/assets/modules/plan-drama-production/prompts/production_plan.zh.txt +5 -2
  19. package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/editing/edit-deliver-drama.md +2 -0
  20. package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/feature-completeness.md +1 -1
  21. package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/generation/configure-generation-providers.md +1 -1
  22. package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/generation/design-drama-audio.md +2 -2
  23. package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/generation/drama-generation-service.md +1 -1
  24. package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/generation/write-drama-video-prompts.md +1 -1
  25. package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/pipeline.md +1 -1
  26. package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/storyboard/generate-blender-previz.md +1 -1
  27. package/skills/using-vvicat-ai-short-studio-cli/references/short-drama/usage-guide.md +1 -1
package/README.md CHANGED
@@ -107,7 +107,7 @@ ai-short-studio task list --project "$PROJECT_ID" --status running --json
107
107
  ai-short-studio task wait --task <task1> --task <task2> --concurrency 4 --timeout 600 --max-interval 15000 --json
108
108
  ```
109
109
 
110
- `canvas node update` 的 `--set key=<JSON>`、`--unset key`、`--title`、`--x` 和 `--y` 只修改指定字段;旧 `--config` 仍兼容。`update-many` 在一次画布版本校验中原子更新多个节点。结构化生成资源仍必须使用 `edit-resource` 或 `edit-shot`,不能通过字段更新改写 artifact。批量执行与等待默认并发为 4、上限为 16。`task list` 默认返回精简状态,`--full` 才返回完整 payload/result;精简状态包含 attempt、maxAttempts 和时间戳,`task wait` 默认 600 秒超时,并对无变化状态使用 1 秒到 15 秒指数退避;传 `--allow-timeout` 可在超时时返回当前任务快照,供调度器继续判断。Agent 应优先单进程批处理,避免为每个节点启动一个 CLI 进程。
110
+ `canvas node update` 的 `--set key=<JSON>`、`--unset key`、`--title`、`--x` 和 `--y` 只修改指定字段;旧 `--config` 仍兼容。`update-many` 在一次画布版本校验中原子更新多个节点。结构化生成资源仍必须使用 `edit-resource` 或 `edit-shot`,不能通过字段更新改写 artifact。批量执行与等待默认并发为 4、上限为 16;多节点 `--wait` 优先按项目批量轮询任务快照,任务不在快照范围时自动回退逐任务等待。`task list` 默认返回精简状态,`--full` 才返回完整 payload/result;精简状态包含 attempt、maxAttempts 和时间戳,`task wait` 默认 600 秒超时,并对无变化状态使用 1 秒到 15 秒指数退避;传 `--allow-timeout` 可在超时时返回当前任务快照,供调度器继续判断。Agent 应优先单进程批处理,避免为每个节点启动一个 CLI 进程。
111
111
 
112
112
  使用 `canvas node edit-shot <nodeId> --project <projectId> --file shot.json` 编辑镜头。命令会追加并选中不可变的人工 DOCUMENT 版本,不允许修改 `shotKey`、`shotIndex` 或来源身份,旧版本继续可选。重复提交同一节点、同一基线与同一载荷,或当前选版内容已等于目标结果时,返回已有版本并标记 `idempotent: true`。画布版本冲突会刷新画布和当前选版后重试一次;资产版本竞争因事务已回滚,会刷新画布并保留原基线重试一次。身份错误不会重试。
113
113
 
@@ -11,7 +11,7 @@ export default class CanvasNodeRun extends BaseCommand {
11
11
  project: projectFlag,
12
12
  node: Flags.string({ description: '节点 ID,可重复指定以批量运行', multiple: true }),
13
13
  concurrency: Flags.integer({ description: '批量提交并发数', default: 4, min: 1, max: 16 }),
14
- wait: Flags.boolean({ description: '兼容显式单批等待;滚动流水线应使用 task list 调度', default: false }),
14
+ wait: Flags.boolean({ description: '提交后等待任务完成;滚动流水线仍建议使用 task list 调度', default: false }),
15
15
  interval: Flags.integer({ description: '初始轮询间隔(毫秒)', default: 1_000, min: 1, max: 60_000 }),
16
16
  'max-interval': Flags.integer({ description: '无状态变化时的最大轮询间隔(毫秒)', default: 15_000, min: 1, max: 60_000 }),
17
17
  timeout: Flags.integer({ description: '等待超时(秒)', default: 600, min: 1, max: 86_400 }),
@@ -34,14 +34,23 @@ export default class CanvasNodeRun extends BaseCommand {
34
34
  const taskId = submittedTaskId(item.result);
35
35
  return taskId ? [{ nodeId: item.nodeId, taskId }] : [];
36
36
  });
37
- const completed = await mapConcurrent(tasks, flags.concurrency, async (item) => ({
38
- nodeId: item.nodeId,
39
- task: await api.waitTask(item.taskId, {
37
+ const batchTasks = tasks.length > 1 && typeof api.waitTasksInProject === 'function'
38
+ ? await api.waitTasksInProject(flags.project, tasks.map((item) => item.taskId), {
40
39
  interval: flags.interval,
41
40
  maxInterval: flags['max-interval'],
42
41
  timeoutMs: flags.timeout * 1_000,
43
- }),
44
- }));
42
+ }, 'canvas_node')
43
+ : null;
44
+ const completed = batchTasks
45
+ ? tasks.map((item, index) => ({ nodeId: item.nodeId, task: batchTasks[index] }))
46
+ : await mapConcurrent(tasks, flags.concurrency, async (item) => ({
47
+ nodeId: item.nodeId,
48
+ task: await api.waitTask(item.taskId, {
49
+ interval: flags.interval,
50
+ maxInterval: flags['max-interval'],
51
+ timeoutMs: flags.timeout * 1_000,
52
+ }),
53
+ }));
45
54
  this.print(nodeIds.length === 1 ? completed[0]?.task ?? submissions[0].result : { submissions, tasks: completed });
46
55
  if (completed.some((item) => item.task.status !== 'completed'))
47
56
  this.error('部分任务未成功完成', { exit: 8 });
@@ -4,6 +4,8 @@ export default class TaskWait extends BaseCommand {
4
4
  taskId: import("@oclif/core/interfaces").Arg<string | undefined, Record<string, unknown>>;
5
5
  };
6
6
  static flags: {
7
+ project: import("@oclif/core/interfaces").OptionFlag<string | undefined, import("@oclif/core/interfaces").CustomOptions>;
8
+ 'target-type': import("@oclif/core/interfaces").OptionFlag<string | undefined, import("@oclif/core/interfaces").CustomOptions>;
7
9
  task: import("@oclif/core/interfaces").OptionFlag<string[] | undefined, import("@oclif/core/interfaces").CustomOptions>;
8
10
  concurrency: import("@oclif/core/interfaces").OptionFlag<number, import("@oclif/core/interfaces").CustomOptions>;
9
11
  interval: import("@oclif/core/interfaces").OptionFlag<number, import("@oclif/core/interfaces").CustomOptions>;
@@ -5,6 +5,8 @@ export default class TaskWait extends BaseCommand {
5
5
  static args = { taskId: Args.string({ required: false }) };
6
6
  static flags = {
7
7
  ...BaseCommand.baseFlags,
8
+ project: Flags.string({ description: '项目 ID;与 --target-type 一起启用项目级批量轮询' }),
9
+ 'target-type': Flags.string({ description: '项目级批量轮询的任务目标类型' }),
8
10
  task: Flags.string({ description: '任务 ID,可重复指定以批量等待', multiple: true }),
9
11
  concurrency: Flags.integer({ description: '批量轮询并发数', default: 4, min: 1, max: 16 }),
10
12
  interval: Flags.integer({ description: '初始轮询间隔(毫秒)', default: 1_000, min: 1, max: 60_000 }),
@@ -18,7 +20,15 @@ export default class TaskWait extends BaseCommand {
18
20
  if (!taskIds.length)
19
21
  throw new Error('请提供 taskId 或至少一个 --task');
20
22
  const { api } = await commandApi(flags);
21
- const tasks = await mapConcurrent(taskIds, flags.concurrency, async (taskId) => await api.waitTask(taskId, {
23
+ const batchTasks = flags.project && taskIds.length > 1 && typeof api.waitTasksInProject === 'function'
24
+ ? await api.waitTasksInProject(flags.project, taskIds, {
25
+ interval: flags.interval,
26
+ maxInterval: flags['max-interval'],
27
+ timeoutMs: flags.timeout * 1_000,
28
+ allowTimeout: flags['allow-timeout'],
29
+ }, flags['target-type'])
30
+ : null;
31
+ const tasks = batchTasks ?? await mapConcurrent(taskIds, flags.concurrency, async (taskId) => await api.waitTask(taskId, {
22
32
  interval: flags.interval,
23
33
  maxInterval: flags['max-interval'],
24
34
  timeoutMs: flags.timeout * 1_000,
@@ -187,6 +187,9 @@ export declare class VvicatApi {
187
187
  waitTask(taskId: string, options?: TaskWaitOptions | number): Promise<JsonObject & {
188
188
  status: string;
189
189
  }>;
190
+ waitTasksInProject(projectId: string, taskIds: string[], options?: TaskWaitOptions, targetType?: string): Promise<(JsonObject & {
191
+ status: string;
192
+ })[] | null>;
190
193
  config(): Promise<JsonObject>;
191
194
  updateConfig(input: JsonObject): Promise<JsonObject>;
192
195
  raw(path: string, options?: ApiRequestOptions): Promise<unknown>;
@@ -458,6 +458,46 @@ export class VvicatApi {
458
458
  await this.sleep(remaining);
459
459
  }
460
460
  }
461
+ async waitTasksInProject(projectId, taskIds, options = {}, targetType) {
462
+ const wanted = [...new Set(taskIds)];
463
+ if (!wanted.length)
464
+ return [];
465
+ const query = { projectId, ...(targetType ? { targetType } : {}), limit: 200, summary: true };
466
+ const first = await this.listTasks(query);
467
+ const snapshots = new Map(first.filter((task) => typeof task.id === 'string').map((task) => [task.id, task]));
468
+ if (wanted.some((taskId) => !snapshots.has(taskId)))
469
+ return null;
470
+ const initialInterval = Math.max(1, Math.floor(options.interval ?? 1_000));
471
+ const maxInterval = Math.max(initialInterval, Math.floor(options.maxInterval ?? 15_000));
472
+ const timeoutMs = Math.max(0, Math.floor(options.timeoutMs ?? 0));
473
+ const deadline = timeoutMs > 0 ? Date.now() + timeoutMs : null;
474
+ let delay = initialInterval;
475
+ let signature = '';
476
+ while (true) {
477
+ const current = await this.listTasks(query);
478
+ const currentIds = new Set(current.map((task) => task.id).filter((id) => typeof id === 'string'));
479
+ // 任务被分页挤出时不能沿用旧快照,否则可能把已完成任务误判为一直运行。
480
+ if (wanted.some((taskId) => !currentIds.has(taskId)))
481
+ return null;
482
+ for (const task of current) {
483
+ if (typeof task.id === 'string' && snapshots.has(task.id))
484
+ snapshots.set(task.id, task);
485
+ }
486
+ const ordered = wanted.map((taskId) => snapshots.get(taskId));
487
+ if (ordered.every((task) => ['completed', 'failed', 'canceled', 'dismissed'].includes(task.status)))
488
+ return ordered;
489
+ const nextSignature = ordered.map((task) => [task.status, task.progress, task.stage, task.flowStageIndex].join(':')).join('|');
490
+ delay = nextSignature === signature ? Math.min(maxInterval, Math.ceil(delay * 1.5)) : initialInterval;
491
+ signature = nextSignature;
492
+ if (deadline !== null && Date.now() >= deadline) {
493
+ if (options.allowTimeout)
494
+ return ordered;
495
+ throw new Error(`TASK_WAIT_TIMEOUT:${wanted.join(',')}:${ordered.map((task) => task.status).join(',')}`);
496
+ }
497
+ const remaining = deadline === null ? delay : Math.min(delay, Math.max(0, deadline - Date.now()));
498
+ await this.sleep(remaining);
499
+ }
500
+ }
461
501
  config() {
462
502
  // 读取时服务端会修复已失效的默认模型并补齐必需的能力默认值。
463
503
  return this.client.request('/api/user/api-config?redactSecrets=1');
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@aswless_854771076/ai_short_studio_cli",
3
- "version": "0.1.153",
3
+ "version": "0.1.154",
4
4
  "description": "VVICAT 无限画布项目命令行工具",
5
5
  "license": "MIT",
6
6
  "private": false,
@@ -67,7 +67,7 @@ description: Use when an agent needs to inspect or operate VVICAT infinite-canva
67
67
 
68
68
  ## 强制分流
69
69
 
70
- - H3 成片中的人物说话、旁白、环境音和 BGM 全部由 H3 原生生成;不得用 `voice-design`、`voice-clone`、`tts` 或其他外部 TTS 覆盖。独立音频交付才读取专项 TTS 流程。
70
+ - H3 成片默认使用 H3 原生对白、旁白、环境音和动作声,并默认支持当前已确认的 `voice_id`/voice reference;只有实时能力明确返回不支持时,才让用户在支持该音色的 Provider、独立后配音或放弃固定音色之间选择。不得用 `voice-design`、`voice-clone` 或外部 TTS 静默覆盖 H3 原生音轨。
71
71
  - 独立交付的背景音乐、主题曲或纯音乐使用 `music-generate`,不要用 TTS 模型模拟。默认模型键为 `starrouter::suno_music`;运行前以 `canvas settings get` 和实时节点 schema 为准。
72
72
  - 涉及剪辑、字幕、转场、音频编排或渲染时,读取 `references/hypit/` 中对应的制作规则。
73
73
  - 专业拆镜出现软提示时,可按 [增量修订流程](references/assets-and-storyboard.md#专业拆镜软提示与增量修订) 使用 `canvas node revise-storyboard` 读取诊断、预览并保存局部补丁;冲突不得盲目重试。普通编辑镜头用 `canvas node edit-shot`,编辑结构化资源用 `canvas node edit-resource`,均追加不可变版本。项目 `outputLanguage` 在创建后固定,画布只读;手工首尾帧使用 `video-frame-extract`。
@@ -28,7 +28,7 @@
28
28
 
29
29
  ## 资产生产门禁
30
30
 
31
- 专业镜头的 `assetNodeRefs` 与 `visualAssetBindings` 共用镜头选版。通过下文专业增量修订流程,为每项设定来源绑定实际 `imageNodeId`,并把这些图片接到两板的 `reference` 端口;不能按节点标题猜归属。每个绑定图片节点只能代表一个选定媒体,多候选需使用单项资源节点。服务端在付费板图前检查全部声明资产、上游真实设定边、选版与实际上送图片;提示词按实际数组编译身份和参考序号。修改绑定会使两板及下游 stale;设定换版或文字投影变化同样使旧来源失效。两板都审核通过仍要比较实际参考集合,差异不能由文字批准覆盖;引用覆盖不代表图像像素已符合设定。
31
+ 专业镜头的 `assetNodeRefs` 与 `visualAssetBindings` 共用镜头选版。通过下文专业增量修订流程,为每项设定来源绑定实际 `imageNodeId`,并把这些图片接到已选择板型的 `reference` 端口;不能按节点标题猜归属。每个绑定图片节点只能代表一个选定媒体,多候选需使用单项资源节点。服务端在付费板图前检查全部声明资产、上游真实设定边、选版与实际上送图片;提示词按实际数组编译身份和参考序号。修改绑定会使已选择板型及下游 stale;设定换版或文字投影变化同样使旧来源失效。所有已选择板型都审核通过仍要比较实际参考集合,差异不能由文字批准覆盖;引用覆盖不代表图像像素已符合设定。
32
32
 
33
33
  1. 只生产目标集实际出现的人物及形象版本、场景/子场景、关键道具及叙事版本。每项任务成功后展示真实预览,让用户确认并设置 selected version。
34
34
  2. 每个形象候选使用独立输入、独立三视图节点和独立图片资产;`count` 的多个结果也不能拼图。视觉素材未另行指定时使用 16:9,并按实时 schema 显式写入。
@@ -39,11 +39,11 @@
39
39
 
40
40
  ## 专业时间宫格与镜头故事版
41
41
 
42
- 专业拆镜每镜包含 `storyboard_strategy`:`primary_board_type` 为 `temporal` 或 `shot`,`required_board_types` 必须同时包含两类,`selection_basis` 说明选择依据,`frame_count` 使用 4/6/9/16/25,默认建议 6。时间宫格画同镜动作因果,镜头故事版画可见的空间、调度和视线关系;单幅首帧独立存在,不能替代任一板。
42
+ 专业拆镜每镜包含 `storyboard_strategy`:`primary_board_type` 必须是已选择的板型,`required_board_types` 只包含解决本镜实际制作风险所需的最小集合,`selection_basis` 说明选择依据,`frame_count` 使用 4/6/9/16/25,默认建议 6。时间宫格表达动作因果,镜头故事版表达空间、调度和视线关系;单幅首帧按需使用,不能冒充已选择的板型。
43
43
 
44
44
  物化按每镜 `required_board_types` 复用已选择的板型,并维护已有首帧,不按任务完成顺序猜身份。`canvas inspect` 的 `data.production` 提供板型和镜头归属;该字段仅由服务端维护,禁止通过 `canvas apply` 伪造。生成任务记录当前镜头选版、父拆镜版本和板型,不能从草稿文字替代来源。编辑 `storyboard_strategy.frame_count` 后同步已选择板型的格数;已有选版保留但标记 stale,分别重新生成。导演上下文不设独立的 10,000 字符门槛;最终提示词按实际供应商和模型能力校验,当前 GPT Image 2 支持 36,000 字符。不手写旁路提示词绕过来源和已选择板型合同。
45
45
 
46
- 专业提示词完成后,自动派生的视频节点继承主图、两板及对应预检,同图兼任主板时只接一次参考;回读 inspect 核对实际连线和当前版本。板节点与图片存在仅证明已派生或已生成,不证明实际审核通过或已传入视频。逐板查看真实图像并核对版本与引用;同图换标题不算另一类板,抽帧或 ASR 不能代替声画审核。
46
+ 专业提示词完成后,自动派生的视频节点继承主图、已选择的分镜板及对应预检;回读 inspect 核对实际连线和当前版本。板节点与图片存在仅证明已派生或已生成,不证明实际审核通过或已传入视频。逐板查看真实图像并核对版本与引用;同图换标题不算另一类板,抽帧或 ASR 不能代替声画审核。
47
47
 
48
48
  ## 拆镜语言匹配门禁
49
49
 
@@ -65,7 +65,7 @@
65
65
 
66
66
  逐镜读取 selected shot 完整结构,把 `characters`、`locations`、`props` 与 description、sourceText、imagePrompt、videoPrompt、photographyPlan 交叉核对。结构化 `props` 缺失时用长名和上下文扫描文字字段;简称、同名或真假版本不能唯一匹配时询问用户。
67
67
 
68
- 每镜同时核对两类边:
68
+ 每镜核对实际声明的参考边:
69
69
 
70
70
  - 已选 `image-resource → storyboard-image.reference` 的图片参考边。
71
71
  - `character/location/prop-resource → storyboard-shot/storyboard-image` 的文字设定边。
@@ -95,7 +95,7 @@
95
95
 
96
96
  每批完成后回读任务、节点、连线、资产和最终配置;先用 `canvas asset download` 或等效预览能力取得每个候选的真实图片,再在当前对话中展示候选缩略图/文件链接、资产 key、版本、用途和审核状态。用户看过并确认候选后才能写入 selected version,不能只在 Agent 内部查看后静默选版。随后实际查看所有图片,检查身份、肢体、道具、构图、画幅、裁切、文字、连续性和空结果。缺陷需再次付费生成或改选版本时,遵循已有授权;超出用户已确认的生成数量或修改范围才补充确认。
97
97
 
98
- 已选板型的 `storyboard_strategy.frame_plan` 必须登记 `beats:[{id,sourceUnitIds,description}]`,来源并集恰好等于本镜 `sourceUnitIds`;`temporal_cells` 与 `shot_cells` 各为 `frame_count` 个 `{beat_id,visible_change}`,各自覆盖全部节拍且不重复填格。时间格按节拍顺序,空间格表达几何证据,不能复制同一计划。`overflow_strategy` 固定 `return-to-breakdown`;容量不足退回拆镜,不截断、不借邻镜来源。旧策略可读,重新生成/审核前须补计划;`STORYBOARD_BOARD_FRAME_PLAN_*` 指出计划缺陷。专业镜头按下文“读取—预览—保存—复查”通过 `canvas node revise-storyboard` 保存策略补丁,改格数时同步已选择板型的计划及节点配置。结构校验不证明语义或像素符合预期。
98
+ 已选板型的 `storyboard_strategy.frame_plan` 必须登记 `beats:[{id,sourceUnitIds,description}]`,来源并集恰好等于本镜 `sourceUnitIds`;选择 `temporal` 时填写 `temporal_cells`,选择 `shot-board` 时填写 `shot_cells`;对应字段为 `frame_count` 个 `{beat_id,visible_change}`,覆盖所选板型的全部节拍且不重复填格。时间格按节拍顺序,空间格表达几何证据,不能复制同一计划。`overflow_strategy` 固定 `return-to-breakdown`;容量不足退回拆镜,不截断、不借邻镜来源。旧策略可读,重新生成/审核前须补计划;`STORYBOARD_BOARD_FRAME_PLAN_*` 指出计划缺陷。专业镜头按下文“读取—预览—保存—复查”通过 `canvas node revise-storyboard` 保存策略补丁,改格数时同步已选择板型的计划及节点配置。结构校验不证明语义或像素符合预期。
99
99
 
100
100
 
101
101
  使用前镜尾帧时,在镜头编辑器的 `continuity` JSON 中为相邻两镜填写 `camera_setup_id`、`entry_state`、`exit_state`。机位 ID 必须相同,前镜 `exit_state` 必须与本镜 `entry_state` 完全一致,本镜 `mode` 为 `previous-tail`,`sourceShotKey` 指向紧邻前镜。状态描述应包含位置、朝向、持物、姿态动势、轴线、光照和天气;换机位使用 `independent`。视频生成必须接入同一父拆镜下、已通过当前版本实际播放审核的前镜视频所提取的真实尾帧,并占用一个独立参考名额。提示词可提前规划;实际生成不可缺少尾帧。建议接线也遵守机位和状态约束。前镜换版或审核变化会阻止执行,执行中变化则仅归档历史结果。旧镜头可读,重新使用尾帧前须补齐声明并重新审核。声明一致不证明像素衔接自然,仍须检查实际视频。
@@ -24,6 +24,7 @@
24
24
  ## 并发、幂等与恢复
25
25
 
26
26
  - 每类默认并发 4,总在途任务不超过 16;用户明确要求“全部提交”时可提高提交并发,但仍服从服务端能力、费用范围和全局上限。
27
+ - 按依赖层并发推进:无依赖的镜头生成、下载、转写和 QC 应在同一并发窗口中执行;只有 previous-tail 链、版本选定、时间线和整集交付审核保持串行。不得因为单镜等待而阻塞无依赖镜头,也不得把整集流程退化为逐镜串行。
27
28
  - 同一 `nodeId + inputFingerprint` 已有 queued/running 任务时不得重复提交。任务完成后若 selected 版本、输入指纹或下游节点未变化,也不得重复触发。
28
29
  - 服务端物化下游节点可能有短暂延迟;先刷新和重试查询,不手工创建重复节点。
29
30
  - 网络、上传、`External task gone` 等可恢复错误只重试对应节点,使用有限指数退避;默认最多 3 次自动重试。内容审核、余额、权限、schema、模型能力和付费门禁错误禁止自动换模型或无限重试。
@@ -18,8 +18,8 @@ ai-short-studio canvas asset download <assetId> --project <projectId> --output <
18
18
  ai-short-studio canvas node run --project <projectId> --node <node1> --node <node2> --concurrency 4 --json
19
19
  ai-short-studio task list --project <projectId> --status running --json
20
20
  ai-short-studio task get <taskId> --json
21
- ai-short-studio task wait --task <task1> --task <task2> --concurrency 4 --timeout 600 --max-interval 15000 --json
21
+ ai-short-studio task wait --project <projectId> --target-type canvas_node --task <task1> --task <task2> --concurrency 4 --timeout 600 --max-interval 15000 --json
22
22
  ai-short-studio task cancel <taskId> --yes --json
23
23
  ```
24
24
 
25
- 互不依赖的节点在一个 CLI 进程内批量运行;有依赖时按滚动流水线分批。`task list` 默认返回精简状态,排查时才使用 `--full`。`task wait` 默认 600 秒超时,并对无变化状态指数退避;不要用整批等待替代滚动调度。任务成功后仍需回读资产并查看真实产物,不能把成功状态当作质量验收。
25
+ 互不依赖的节点在一个 CLI 进程内批量运行;有依赖时按滚动流水线分批。`task list` 默认返回精简状态,排查时才使用 `--full`。`task wait` 携带 `--project` 时优先按项目级任务快照批量轮询,未携带时保持逐任务等待;默认 600 秒超时,并对无变化状态指数退避。任务成功后仍需回读资产并查看真实产物,不能把成功状态当作质量验收。
@@ -298,7 +298,7 @@ the material of the production, then refine how their arrangement expresses the
298
298
 
299
299
  For reference-led recreation, produce and approve a depth video from each selected source shot before
300
300
  submitting the corresponding generated shot. Without a reference video, generate and approve Blender
301
- previz only when the two still boards cannot prove a complex spatial or movement risk, then bind it directly as the motion reference; do not generate a synthetic source clip and
301
+ previz only when the selected still boards cannot prove a complex spatial or movement risk, then bind it directly as the motion reference; do not generate a synthetic source clip and
302
302
  convert it to depth. Treat character boards as identity guidance and any chosen motion reference as
303
303
  structure, motion and camera guidance. If the selected route cannot consume it, change the route.
304
304
 
@@ -98,7 +98,7 @@ references, never a visible grid in the final video.
98
98
  [H3 prompt writing](../craft/h3-prompt-writing.md); Seedance uses its Endpoint's actual ordered labels.
99
99
  Do not mix dialects.
100
100
 
101
- Every final generated shot requires an independently generated temporal storyboard, an independently generated shot board, video prompt and audio reference. Do not manufacture a source clip or depth pass: add Blender previz only when the two still boards cannot prove a complex route, multi-person interaction, exact contact, axis risk, or continuous camera move, then bind that previz as the motion reference. `generate-storyboard-images` remains mandatory for every shot; optional asset boards are selected
101
+ Every final generated shot requires the independently generated storyboard boards selected by its production plan, a video prompt, and any audio reference required by its sound contract. Do not manufacture a source clip or depth pass: add Blender previz only when the selected still boards cannot prove a complex route, multi-person interaction, exact contact, axis risk, or continuous camera move, then bind that previz as the motion reference. `generate-storyboard-images` runs only for the selected board types; optional asset boards are selected
102
102
  per shot. Configure independent audio only when applicable. The
103
103
  selected production plan and video prompt documents record those references by immutable asset version;
104
104
  the vendored workflow gates verify their paths, hashes, approvals and selection state before a paid call.
@@ -15,13 +15,13 @@
15
15
 
16
16
  初始化时单独询问每个需要说话的人物的声音来源,并把选择写入 `audio-plan.voice_consistency`:
17
17
 
18
- - `first_appearance_sample`(默认):找到人物第一次出场且有清晰对白的已选视频或原声,截取实际说话区间作为声纹锚点;先展示截取时长、来源镜头、波形/试听链接和权利状态,用户确认后才能上传或绑定。
18
+ - `first_appearance_sample`(默认):只把人物第一次出场且有清晰对白的已有视频或原声截取为声纹样本,样本本身不是固定音色 ID。先展示截取时长、来源镜头、波形/试听链接和权利状态,用户确认后再由当前 Provider 创建或选择真实 `voice_id`;当前 Provider 默认支持该绑定,只有实时能力明确返回不支持时才分流。没有已有来源音频时不得用尚未生成的首镜反推音色,必须停下让用户上传样本或选择 Provider voice。
19
19
  - `user_upload`:用户上传并确认有权使用的声音样本。
20
20
  - `global_asset`:从全局声音资产库选择已审核的 voice ID 和版本。
21
21
  - `provider_voice_id`:选择实时 Provider 目录中已确认的声音 ID。
22
22
  - `none`:明确不做跨镜声音一致性绑定,后续每镜按项目确认的声音策略生成。
23
23
 
24
- 默认方案只负责建立绑定,不自动克隆或发起付费生成;绑定记录必须包含 `character_id`、`source_mode`、来源资产/镜头、起止时间、SHA-256、试听证据和用户确认记录。后续独立配音、电话音轨或视频原生声音只能引用已确认的绑定,绑定变更必须重新试听并复审。
24
+ 默认方案先建立样本证据,不自动克隆或发起付费生成;只有用户确认后才允许创建 voice ID。绑定记录必须包含 `character_id`、`source_mode`、来源资产/镜头、起止时间、SHA-256、`voice_id`/Provider、试听证据、能力校验和用户确认记录。后续独立配音、电话音轨或视频原生声音只能引用已确认的绑定;若实时能力明确返回当前视频模型不支持该 voice ID,才改用支持的 Provider voice、后配音,或由用户明确选择不固定音色。
25
25
 
26
26
  ## 查询与批处理
27
27
 
@@ -30,7 +30,7 @@ Rules:
30
30
  7. New plans must set `audio_strategy.mode=native-first`, `provider_selection=prefer-native`, and `fallback_allowed=true`. `fallback_reasons` must contain exactly: `provider-no-native-audio`, `voice-identity-drift`, `speech-intelligibility-failed`, `narration-performance-failed`, `audio-sync-failed`, and `native-ambience-failed`.
31
31
  8. Keep source and presentation orthogonal on every line: `delivery_mode=native|post_dub|external_audio`, `presentation=visible-dialogue|offscreen-dialogue|narration`, and `fallback_mode=none|post-dub|cinematic-tts|sound-design`. Never put narration or offscreen in delivery_mode.
32
32
  9. Use null for `source_audio` and `voice_binding` when no selected asset or real voice evidence exists. A native-audio fallback requires `native_audio_exception` with an allowed reason, actual evidence, an exact millisecond range, and final `mix_sources`. Offscreen dialogue or narration that was never generated is not a native-audio failure and must keep `native_audio_exception` null.
33
- 9a. `voice_consistency.source_mode` must be one of `first_appearance_sample|user_upload|global_asset|provider_voice_id|none`, taken from the user-confirmed project initialization. `first_appearance_sample` extracts a real interval from the selected video or source audio where the character first appears with clear dialogue, and records `source_asset_key`, `source_version_id`, `start_ms`, `end_ms`, `sha256`, preview evidence, and confirmation; without preview and confirmation put it in unresolved and never bind automatically. `user_upload` needs an upload receipt and rights confirmation, `global_asset` needs a reviewed global voice ID/version, and `provider_voice_id` needs live catalog evidence; `none` keeps voice_binding null.
33
+ 9a. `voice_consistency.source_mode` must be one of `first_appearance_sample|user_upload|global_asset|provider_voice_id|none`, taken from the user-confirmed project initialization. `first_appearance_sample` only extracts a real interval from an existing selected video or source audio where the character first appears with clear dialogue, and records `source_asset_key`, `source_version_id`, `start_ms`, `end_ms`, `sha256`, preview evidence, and confirmation; the sample is not a `voice_id`: after confirmation the Provider must create a real `voice_id`, pass capability and listening review, and only then bind it. Without existing source audio, preview, or confirmation put it in unresolved; never infer the voice from an ungenerated first shot. `user_upload` needs an upload receipt and rights confirmation, `global_asset` needs a reviewed global voice ID/version, and `provider_voice_id` needs live catalog evidence; `none` keeps voice_binding null.
34
34
  9b. Missing offscreen dialogue uses `delivery_mode=post_dub`, `presentation=offscreen-dialogue`, and a clone of the character's original voice made with the model required by the current audio specification. Its voice binding must include `voice_role=character`, `source=clone`, the real `target_model`, and a real `upload_receipt_id`. Narration uses `delivery_mode=post_dub`, `presentation=narration`, and an independently designed professional narrator voice. Its binding must include `voice_role=narrator`, `source=design`, `professional_narration=true`, a real `request_id`, and a `cinematic_profile` identical to performance. Both routes must bind source `speech-timing`; never estimate timing from text.
35
35
  10. Every narration line requires an executable cinematic `performance` with `tone_arc`, ordered `emotion_beats[]`, `pace`, `breath_and_pause`, and `distance_and_space`; “emotional” alone is insufficient. Non-narration performance is null. Native background sound means ambience and action sound only; BGM remains an independent licensed-music workflow.
36
36
 
@@ -30,7 +30,7 @@
30
30
  7. 新计划固定 `audio_strategy.mode=native-first`、`provider_selection=prefer-native`、`fallback_allowed=true`;`fallback_reasons` 必须完整且仅含 `provider-no-native-audio`、`voice-identity-drift`、`speech-intelligibility-failed`、`narration-performance-failed`、`audio-sync-failed`、`native-ambience-failed`。
31
31
  8. 每行分别填写声音来源 `delivery_mode=native|post_dub|external_audio`、画面呈现 `presentation=visible-dialogue|offscreen-dialogue|narration`、预案 `fallback_mode=none|post-dub|cinematic-tts|sound-design`,不得把 narration/offscreen 填进 delivery_mode。
32
32
  9. `source_audio` 和 `voice_binding` 没有已选资产或真实音色证据时为 null,不得伪造。原生失败兜底时 `native_audio_exception` 必须写受控 reason、实际 evidence、毫秒 range 和最终 `mix_sources`;原本未生成的场外音或旁白不是原生失败,`native_audio_exception` 必须为 null。
33
- 9a. `voice_consistency.source_mode` 只能是 `first_appearance_sample|user_upload|global_asset|provider_voice_id|none`,且必须来自项目初始化时用户确认的配置。`first_appearance_sample` 从人物第一次出场且有清晰对白的已选视频或原声截取真实区间,写入 `source_asset_key`、`source_version_id`、`start_ms`、`end_ms`、`sha256`、试听证据和确认记录;未展示试听或未确认时写入 unresolved,不得自动绑定。`user_upload` 必须有上传回执和权利确认,`global_asset` 必须有全局 voice ID/版本和审核证据,`provider_voice_id` 必须有实时目录证据;`none` 时 voice_binding 为 null。
33
+ 9a. `voice_consistency.source_mode` 只能是 `first_appearance_sample|user_upload|global_asset|provider_voice_id|none`,且必须来自项目初始化时用户确认的配置。`first_appearance_sample` 只从人物第一次出场且有清晰对白的已有视频或原声截取真实区间,写入 `source_asset_key`、`source_version_id`、`start_ms`、`end_ms`、`sha256`、试听证据和确认记录;样本不是 `voice_id`,必须在用户确认后经 Provider 创建真实 `voice_id` 并通过能力检查和试听审核。没有已有来源音频、未展示试听或未确认时写入 unresolved,不得自动绑定或用尚未生成的首镜反推。`user_upload` 必须有上传回执和权利确认,`global_asset` 必须有全局 voice ID/版本和审核证据,`provider_voice_id` 必须有实时目录证据;`none` 时 voice_binding 为 null。
34
34
  9b. 原本未生成的场外音设 `delivery_mode=post_dub`、`presentation=offscreen-dialogue`,绑定当前声音规范指定模型克隆的角色原声音色;voice_bindings 必须含 `voice_role=character`、`source=clone`、真实 `target_model` 和 `upload_receipt_id`。旁白设 `delivery_mode=post_dub`、`presentation=narration`,使用独立专业设计音色;voice_bindings 必须含 `voice_role=narrator`、`source=design`、`professional_narration=true`、真实 `request_id`,且 `cinematic_profile` 与 performance 逐字段一致。两者都必须先绑定原声 speech-timing,不得按文字估算时间。
35
35
  10. `presentation=narration` 必须写电影感 `performance`:`tone_arc`、按顺序的 `emotion_beats[]`、`pace`、`breath_and_pause`、`distance_and_space`,不能只写“富有情感”;非旁白为 null。背景音原生直出只包括环境声和动作声,BGM 仍由独立授权配乐流程负责。
36
36
 
@@ -24,7 +24,7 @@ Return strict JSON only: {"episode_key":"...","source_versions":{},"shots":[...]
24
24
 
25
25
  Preflight selection gate: before choosing image_strategy, storyboard_strategy, a Provider, or any paid task, read the director book, selected storyboard, and user limits shot by shot. Classify the main uncertainty as one of: single-frame composition/first-last state, temporal action and exact state, narrative causality/reveal/emotional turn, blocking and spatial route, action breakdown for fights/chases/impacts, choreography for dance/martial arts/sports, scene-overview, scene-plan, scene-turnaround, director-track delivery, or comprehensive hero-sequence control. Select the smallest primary type that resolves the actual risk. In `selection_basis`, cite the source evidence and explain why competing types were not selected; never label every shot `shot-board` because it is the default or because the model prefers it.
26
26
 
27
- Complete board contract: a standard video shot's `required_board_types` contains at least `temporal` and `shot-board`; when the primary type is `narrative`, `blocking`, `action`, `choreography`, `comprehensive`, or `director-track`, retain the two required video boards and add the primary type. Add `previz-3d` only when static boards cannot prove the spatial risk; it never replaces temporal or shot-board. Non-video assets such as single-frame, scene-overview, scene-plan, and scene-turnaround do not automatically receive the two video boards. `image_strategy.board_type` remains `shot-board` for compatibility with the existing image executor and does not represent the preflight primary story-board type.
27
+ Complete board contract: a standard video shot selects the smallest `required_board_types` set that resolves its actual production risk. `temporal`, `shot-board`, `narrative`, `blocking`, `action`, `choreography`, `comprehensive`, `director-track`, and `previz-3d` are selectable types; add multiple types only when each has a separate observable purpose. `previz-3d` is an additional spatial reference when selected and never replaces another selected type. Non-video assets such as single-frame, scene-overview, scene-plan, and scene-turnaround do not receive video boards automatically. `image_strategy.board_type` remains `shot-board` for compatibility with the existing image executor and does not represent the preflight primary story-board type.
28
28
 
29
29
  Rules:
30
30
  1. Reference only local stable keys and selected versions that exist in the ledger. Put missing or wrong versions in unresolved.
@@ -33,13 +33,16 @@ Rules:
33
33
  4. totals summarizes task count by modality, material count, shot count, and total duration. Never calculate, guess, or display a price.
34
34
  5. approved must be false when unresolved is non-empty, reference use is unconfirmed, the production plan is unconfirmed, or a required generation parameter is missing.
35
35
  6. Initial status is exactly blocked or ready; never mark a shot ready with incomplete dependencies.
36
- 7. prompt_profile is exactly seedance2 or h3. Final generation uses character images, a temporal storyboard, a shot board, an audio reference, and the video prompt. Add Blender previz only when the two still boards cannot prove a complex route, multi-person interaction, exact contact, axis risk, or continuous camera move. Seedance 2.0 uses `full-reference`, and H3 uses `Ref2VA`. If the Provider cannot carry the inputs this shot actually requires, mark unresolved; never silently omit one.
37
- 8. When the Provider supports MiniMax H3 native audio (video.native-audio), default audio_strategy.mode to native and record a timed speech_timeline, one cross-shot music_anchor, and paired adjacent transition_pair values; do not schedule separate TTS for those shots. Use post-dub/independent only when the user explicitly requests dubbing or the Provider/model is confirmed to lack native audio; do not silently fall back to dubbing just because capability is unknown. Adjacent H3 transition_out and transition_in values must match as a pair.
38
- 9. Every shot uses image_strategy.mode=generate, board_type=shot-board, and panel_grid_size=2..16 (default 6, adjusted to actual beats), producing independent temporal-storyboard and shot-board assets. Standard projects default to storyboard_strategy.mode=image and previz_strategy={"mode":"none"}. Use storyboard_strategy.mode=blender with previz_strategy={"mode":"blender","purpose":"motion-reference","fps":12} only when the two boards still cannot unambiguously express a complex route, multi-person interaction, exact contact, axis risk, or continuous camera move; name that spatial risk in review_checks. Only StarRouter MiniMax-H3 currently accepts the Blender motion reference through TempFile.org hosting; otherwise mark unresolved. Do not bind future asset versions. Composition elements require source, selected-asset, or confirmed-blocking evidence; block unsupported elements in unresolved. overflow_strategy is exactly compose-assets or reject.
36
+ 6a. Each shot's executable action must be split into singular observable beats. When one shot asks for multiple independent state changes, complex contact, prop release, reaction, and dialogue at once, split it into sequential beats or shots and explain the split in `selection_basis` and `review_checks`; never stack actions that cannot be verified together into one prompt.
37
+ 7. prompt_profile is exactly seedance2 or h3. Final generation uses the selected character images, selected storyboard boards, any required audio reference, and the video prompt. Add Blender previz only when the selected still boards cannot prove a complex route, multi-person interaction, exact contact, axis risk, or continuous camera move. Seedance 2.0 uses `full-reference`, and H3 uses `Ref2VA`. If the Provider cannot carry the inputs this shot actually requires, mark unresolved; never silently omit one.
38
+ 8. When the Provider supports MiniMax H3 native audio (video.native-audio), default audio_strategy.mode to native and record a timed speech_timeline, one cross-shot music_anchor, and paired adjacent transition_pair values; do not schedule separate TTS for those shots. When the project has confirmed `voice_consistency` with a mode other than `none`, every visible speaker must reference a confirmed voice binding; missing a real `voice_id`, source evidence, or capability check goes into `unresolved`, and prompt adjectives such as “low” or “bright” never substitute for character binding. Use post-dub/independent only when the user explicitly requests dubbing or the Provider/model is confirmed to lack native audio; do not silently fall back to dubbing just because capability is unknown. Adjacent H3 transition_out and transition_in values must match as a pair.
39
+ 9. Every shot uses image_strategy.mode=generate and panel_grid_size=2..16 (default 6, adjusted to actual beats), producing only the independent storyboard assets listed in required_board_types. Standard projects default to storyboard_strategy.mode=image and previz_strategy={"mode":"none"}. Use storyboard_strategy.mode=blender with previz_strategy={"mode":"blender","purpose":"motion-reference","fps":12} only when the two boards still cannot unambiguously express a complex route, multi-person interaction, exact contact, axis risk, or continuous camera move; name that spatial risk in review_checks. Only StarRouter MiniMax-H3 currently accepts the Blender motion reference through TempFile.org hosting; otherwise mark unresolved. Do not bind future asset versions. Composition elements require source, selected-asset, or confirmed-blocking evidence; block unsupported elements in unresolved. overflow_strategy is exactly compose-assets or reject.
39
40
  9a. Before generation, realize `required_board_types`: a video shot must generate only the board types selected in its `storyboard_strategy.required_board_types`; narrative, blocking, action, choreography, comprehensive, director-track, temporal, shot-board, and previz-3d are selectable types. Add multiple types only when the production goal needs each one; `previz-3d` is an additional spatial reference when selected. Do not auto-create an unselected board type. `image_strategy.board_type=shot-board` is only a compatibility field for the executor and must not override `primary_board_type`.
40
41
  10. Tailor review_checks to actual shot risk; every item must be a concrete, named, observable phrase (e.g. "mirror reflection consistency", "contact point readable through rain") derived from the contract's ai_risks and evidence elements: dialogue checks speaker identity and every visible mouth state; group shots check headcount, order, facing, and limb ownership; prop interaction checks before/contact/after states and adjacent objects; walking or fights check foot contact, weight, direction, speed, and pose across cuts; in-frame screens check whether precise post-production graphics are required. When composition risks apply, turn each into an explicit check: mirrors/glossy water verify the reflection matches the real person and actual dressing, with no extra fingers, camera, or reflection anomaly; rain, fog, neon, and smoke/dust verify the weather and colored light are evidenced, directions are consistent, and neither buries the contact point nor identity; crowd shots verify headcount, identity, costume-color anchors, and no extra bystanders; weapons and multi-hand interaction verify prop form, holding relationship, hand count, and before/contact/after states; a narrow door gap or foreground occlusion verifies visible width, both eyes intact, and the key action uncovered; high-speed trails verify both eyes, the face, and the contact point stay sharp and the subject is not a blob; front-and-back pincer/encirclement verifies both threats' identity, distance, and facing stay readable in foreground and background; vertical shots verify faces, key props, and contact points sit in the upper-middle safe area, the bottom subtitle safe area is clear, and motion edges keep a 10% margin. Keep only the risks this shot actually has. Never copy one generic checklist onto every shot.
41
42
 
42
- 11. Only a shot whose previz_strategy.mode is blender runs direct-blender-previz before generate-blender-previz renders and scores it. Previz never replaces the temporal storyboard, shot board, character, scene, prop, wardrobe, material, or final-look references.
43
+ 11. Only a shot whose previz_strategy.mode is blender runs direct-blender-previz before generate-blender-previz renders and scores it. Previz never replaces the selected storyboard boards, character, scene, prop, wardrobe, material, or final-look references.
44
+
45
+ 12. `scope_this_round` must state the exact shots, assets, and stop stage allowed in this run. Expanding shot scope, adding paid tasks, or changing the Provider requires a new production-plan version and renewed user confirmation; execution logs must not silently expand the scope.
43
46
 
44
47
 
45
48
  13. `video_strategy.visible_character_keys` lists every `char-*` that is actually visible in the shot and exactly matches the asset keys in storyboard `characters[].appearance`; an off-screen speaker is included only when visible. Every listed key binds one selected image in reference_assets with `role=character_identity`; never bind only the protagonist or speaker. Use an empty array for a shot with no visible character. `video_strategy.visible_event_keys` lists every stable visible event key that may produce action sound, or an empty array when none exists. Final video-prompts bind one temporal storyboard, one shot board, and one audio reference. Bind exactly one Blender reference only when the plan enables it. An asset board is optional and never replaces either a character image or either storyboard.
@@ -33,11 +33,14 @@
33
33
  4. totals 只汇总各模态任务数、素材数量、镜头数和总时长,不包含价格或费用估算。
34
34
  5. unresolved 非空、参考素材用途未确认或任何制作参数缺失时,approved 必须为 false。
35
35
  6. status 初始只能是 blocked 或 ready;依赖不齐不得写 ready。
36
- 8. Provider 支持 MiniMax H3 原生音频(video.native-audio)时,audio_strategy.mode 默认写 native,并记录逐秒 speech_timeline、跨镜统一 music_anchor 和相邻镜 transition_pair;这类镜头不得另排 TTS。仅当用户明确要求后配、或 Provider/模型明确声明不支持原生音频时才写 post-dub/independent 并排入独立音频流程;能力未明确时默认按原生处理,不得因信息缺失而静默转后配。相邻 H3 镜头的 transition_out 与 transition_in 必须成对一致。
36
+ 7. 每个镜头的可执行动作必须按单一可观察节拍拆分:同一镜同时要求多个独立状态变化、复杂接触、道具脱手、反应和对白时,必须拆成连续 beat 或多个镜头,并在 `selection_basis` 与 `review_checks` 说明拆分依据;不得把不可同时验证的动作堆进一个提示词。
37
+ 8. Provider 支持 MiniMax H3 原生音频(video.native-audio)时,audio_strategy.mode 默认写 native,并记录逐秒 speech_timeline、跨镜统一 music_anchor 和相邻镜 transition_pair;这类镜头不得另排 TTS。项目已确认 `voice_consistency` 且模式不是 `none` 时,每个可见说话人都必须引用已确认的 voice binding;缺少真实 `voice_id`、来源证据或能力校验时写入 unresolved,不得只靠提示词中的“低沉/高亢/语速”冒充人物绑定。仅当用户明确要求后配、或 Provider/模型明确声明不支持原生音频时才写 post-dub/independent 并排入独立音频流程;能力未明确时默认按原生处理,不得因信息缺失而静默转后配。相邻 H3 镜头的 transition_out 与 transition_in 必须成对一致。
37
38
  9a. 生成前必须先落实 `required_board_types`:只生成其中声明的板型;主类型必须在集合中,每个板型都必须有对应的可观察风险或用途,未选择的板型不得自动生成。若选择 `previz-3d`,它仅为空间参考,并按其自身合同审核。`image_strategy.board_type=shot-board` 只是执行器兼容字段,不得覆盖 `primary_board_type`。
38
39
  10. review_checks 按逐镜真实风险填写,每个检查项必须是具体、具名、可观察的短语(如「镜面反射一致性」「雨幕接触点可读」),并从合同 ai_risks 与证据元素反推:对白检查说话人与每张可见嘴部状态;多人镜检查人数、顺序、朝向和肢体归属;道具交互检查接触前/中/后状态及邻接物;步行/打斗检查落脚、重心、方向、速度和跨镜姿态;画内屏幕检查是否需要后期精确图形。命中构图风险时必须逐项落为检查:镜面/光滑水面核对映像与真人及现场陈设一致、无多指、无多余摄像机或反射穿帮;雨、雾、霓虹、烟尘核对天气光色有据、雨丝/烟向统一且不淹没接触点与人物身份;群体镜核对人数、身份、服色锚点且无多余路人;武器与多手交互核对形制、持有关系、手数与接触前/中/后;窄门缝或前景遮挡核对可见宽度、双眼完整与关键动作不被盖;高速拖影核对双眼、面部与接触点清晰可辨、主体不糊;前后夹击/包围核对前后景两方身份、距离与朝向可辨;竖屏镜核对脸、关键道具与接触点在中上安全区、底部字幕安全区不被占、动势边缘留 10% 余量。只保留本镜命中的风险项,不得复制同一组空泛检查覆盖所有镜头。
39
40
 
40
- 11. 仅当 previz_strategy.mode=blender 时,白模才由 direct-blender-previz 先生成导演合同,再由 generate-blender-previz 生成和评分;白模不能替代时间故事版、镜头分镜板、人物、场景、道具、服装、材质或最终画质参考。
41
+ 11. 仅当 previz_strategy.mode=blender 时,白模才由 direct-blender-previz 先生成导演合同,再由 generate-blender-previz 生成和评分;白模不能替代已选择的分镜板、人物、场景、道具、服装、材质或最终画质参考。
42
+
43
+ 12. `scope_this_round` 必须明确本轮实际允许生成的镜头、资产和停止阶段;任何扩大镜头范围、追加付费任务或改变 Provider 的行为都必须生成新的 production-plan 版本并重新取得用户确认,不能在执行日志中静默扩大范围。
41
44
 
42
45
 
43
46
 
@@ -3,3 +3,5 @@
3
3
  只接收 `edit-drama-timeline` 按 [本地 Remotion 剪辑方案](../editing-workflow.md) 完成并通过审片的渲染版本。用 `media-tools.mjs qc` 核对分辨率、画幅、帧率、实际时长、编码、采样率、声道、响度和黑帧/冻结/静音事件;目标时长不要求逐秒命中,默认约 ±15% 的自然偏差不作为失败,超出时核对叙事理由和用户确认。从头到尾观看并听完,复核叙事、画面、声音、全部切点、字幕和片尾。交付前必须特别确认字幕不是默认黑底贴纸、说话人层级清楚、文字处于安全区;逐切点检查动作方向、视线、环境声和空间轴是否连续,并抽检淡化精确起始帧,避免仅靠长黑帧检测漏掉单帧黑闪。任何关键动作仅靠生成模型“应该会发生”而未在画面中实际成立,均不得标记为通过。
4
4
 
5
5
  批准版视频、UTF-8 SRT 和 ASS 放入 `delivery/<episode-key>/` 后,运行 `node "${CODEX_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-.}}/scripts/editing-store.mjs" manifest <项目> <episode-key> <成片> <SRT> <ASS>`;脚本会重新校验该集时间线与完整审片记录,计算文件哈希,并以独占方式建立 `delivery/<episode-key>/manifest.json`。所有选中剧集均登记本 Skill 的 manifest 证据并通过 delivery 门禁后,执行 `node "${CODEX_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-.}}/scripts/workflow.mjs" complete <项目>` 固化项目终态。交付不得夹带临时缓存、未批准候选或覆盖既有 manifest。
6
+
7
+ 如果平台、对象存储或分享流程另行上传交付文件,上传完成后必须把实际远端文件重新下载或读取服务端元数据,核对文件大小、SHA-256、时长、音视频流、字幕版本和分辨率;远端副本与本地 manifest 不一致时不得标记交付完成,也不能用压缩副本替代母版而不记录新的交付版本。
@@ -52,7 +52,7 @@
52
52
  ## 当前边界
53
53
 
54
54
  - StarRouter 当前声明图片、视频、同步语音、音乐、语音转写和翻译能力;长文本异步语音返回 TAR,尚未开放为可直接剪辑的音频资产。
55
- - CosyVoice 自定义音色设计/克隆只用于独立 TTS,必须使用当前目标模型生成的真实音色 ID,并通过试听和审核;不得覆盖 H3 原生对白。七维后配兜底、SeedVR2.5 超分及 MuseTalk 口型驱动仍缺少已验证的统一生产入口,当前明确禁用。
55
+ - CosyVoice 自定义音色设计/克隆只在当前目标模型明确支持 voice reference 时使用,必须使用真实音色 ID 并通过试听和审核;不支持时不得声称 H3 原生对白已固定。七维后配兜底、SeedVR2.5 超分及 MuseTalk 口型驱动仍缺少已验证的统一生产入口,当前明确禁用。
56
56
  - 音乐目录不提供抓站下载器。在线结果只打开官方搜索入口,登录、试听、下载和当次许可证判断由用户在站点完成;插件只接收本地文件与证据,并在 audio-plan 和时间线两次核对用途。
57
57
  - 插件不内置大型模型价格表或替用户估价;生成前读取实时 Provider 目录并要求用户确认完整制作计划。
58
58
  - TempFile.org 临时参考地址对持有链接者公开,有效期按上传收据核对;不保证长期可用,也不替代本地资产。
@@ -16,4 +16,4 @@ StarRouter 已列出的主要模型包括图片 `gpt-image-2`,视频 `MiniMax-
16
16
 
17
17
  无限画布使用音乐时,模型键为 `starrouter::suno_music`,节点 kind 为 `music-generate`,输出 AUDIO。节点 config 使用 `prompt`、`title`、`tags`、`lyrics` 和 `makeInstrumental`;`prompt` 必填,`lyrics` 与 `makeInstrumental: true` 互斥。提交使用 `/suno/submit/MUSIC`,结果通过 `/suno/fetch/{task_id}` 轮询,成功结果只登记实际音频而非封面或视频。
18
18
 
19
- StarRouter 当前未验证 CosyVoice 自定义音色设计/克隆、SeedVR2.5 超分和 MuseTalk 口型驱动的等价能力;相关远端操作暂时禁用并提示用户。短剧最终视频统一通过 `hypit short-drama production plan/build` 执行现有审核与付费门禁。
19
+ StarRouter 当前未验证 CosyVoice 自定义音色设计/克隆、SeedVR2.5 超分和 MuseTalk 口型驱动的等价能力;相关操作须先核对实时能力,未验证时停下并提示用户选择其他已支持方案。短剧最终视频统一通过 `hypit short-drama production plan/build` 执行现有审核与付费门禁。
@@ -2,9 +2,9 @@
2
2
 
3
3
  优先在视频模型支持时保留原生对白、环境声和动作声。逐镜审核语言、说话者、文本、可懂度、时间同步、环境动作声和未声明音乐。失败时记录实际区间与原因;合格部分不重复生成。
4
4
 
5
- 独立语音先用 `voice_analysis` 分离说话人与台词,再按项目初始化时确认的 `audio-plan.voice_consistency` 绑定声音。默认 `first_appearance_sample` 从人物第一次出场且有清晰对白的已选视频或原声截取实际说话区间,保存来源镜头、起止时间、SHA-256 并先展示试听;用户确认后才可作为 voice reference。`user_upload` 使用用户上传且已确认有权使用的样本,`global_asset` 使用全局资产库已审核 voice ID,`provider_voice_id` 使用实时 Provider 目录声音,`none` 不建立跨镜绑定。需要复用原声时,才使用 `voice-clone` 生成真实音色 ID;需要新声音时使用 `voice-design`。生成前核对逐字文本和使用权,输出作为未选候选逐条试听、审核后再选版。普通语音模型可用于已确认参数的配音;模型可用性由实时目录和运行时探测。已授权的外部音频可以导入项目并按同样流程审核。
5
+ 独立语音先用 `voice_analysis` 分离说话人与台词,再按项目初始化时确认的 `audio-plan.voice_consistency` 绑定声音。默认 `first_appearance_sample` 只从人物第一次出场且有清晰对白的已有视频或原声截取实际说话区间,保存来源镜头、起止时间、SHA-256 并先展示试听;样本不是 `voice_id`,用户确认后才调用当前 Provider 的 `voice-clone` 创建真实音色 ID,再试听、审核并绑定。没有已有来源音频时不得用尚未生成的首镜反推音色,必须让用户上传样本或选择 Provider voice。`user_upload` 使用用户上传且已确认有权使用的样本,`global_asset` 使用全局资产库已审核 voice ID,`provider_voice_id` 使用实时 Provider 目录声音,`none` 不建立跨镜绑定。当前 Provider 默认支持已确认 voice ID,只有实时能力明确返回不支持时,才切换支持的 Provider voice、后配音或明确放弃固定音色;不能把未收到不支持结果的 H3 原生音频判为未固定。生成前核对逐字文本和使用权,输出作为未选候选逐条试听、审核后再选版。
6
6
 
7
- CosyVoice 自定义音色设计/克隆在独立 TTS 流程中可用,但必须使用与目标模型匹配的真实音色 ID,并经过试听和审核;它不能覆盖 H3 原生对白。MuseTalk 与旧 RunningHub 口型驱动仍暂时禁用。旧 audio-plan 中的失效配置不得继续提交付费任务;向用户说明缺失能力,并先重新确认当前可用模型和制作方案。不得把临时音色或未通过审核的配音默认为正式结果。
7
+ CosyVoice 自定义音色设计/克隆只有在实时能力明确支持且目标模型匹配时,才可在独立 TTS 流程中使用真实音色 ID,并经过试听和审核;它不能静默覆盖 H3 原生对白。MuseTalk 与旧 RunningHub 口型驱动仍暂时禁用。旧 audio-plan 中的失效配置不得继续提交付费任务;向用户说明缺失能力,并先重新确认当前可用模型和制作方案。不得把临时音色或未通过审核的配音默认为正式结果。
8
8
 
9
9
  现有 `cinematic-tts` 和 `generate_audio_fallback` 合同暂存以读取旧项目,但因自定义音色能力未验证,调用会明确拒绝。重新设计时先从当前 selected 原声资产建立对齐证据;最多三轮,每轮必须经八维审核并绑定最终 selected 音频,禁止自动追加第四次。
10
10
 
@@ -8,6 +8,6 @@
8
8
 
9
9
  同一依赖层的图片目标可批量提交;派生资产等母资产审核选版后进入下一层。并发由 Runtime Endpoint 和现有任务门禁控制,不通过切换厂商绕开限流。模型目录以项目配置探测为准,当前主要包括 `gpt-image-2`、`MiniMax-H3`、Seedance、`speech-2.8-hd`、`speech-2.8-turbo`、`suno_music` 和默认转写 `whisper-1`。
10
10
 
11
- StarRouter 尚无已验证的 CosyVoice 自定义音色设计/克隆、SeedVR2.5 超分和 MuseTalk 口型驱动等价能力。这些操作暂时禁用,调用时明确报错。旧项目中保存的 RunningHub、Comfly 直连、百炼或 MuseTalk 配置不能继续提交付费任务,需先改为 StarRouter 并重新确认计划。
11
+ StarRouter 尚无已验证的 CosyVoice 自定义音色设计/克隆、SeedVR2.5 超分和 MuseTalk 口型驱动等价能力。调用前必须核对实时能力,未验证时明确停下并报告缺口。旧项目中保存的 RunningHub、Comfly 直连、百炼或 MuseTalk 配置不能继续提交付费任务,需先改为 StarRouter 并重新确认计划。
12
12
 
13
13
  如果 MCP 工具列表出现缺失或占位,停止提交并修复配置;不得绕过 Hypit Runtime 直接调用旧插件适配器。
@@ -22,7 +22,7 @@ H3 使用 Ref2VA 六段英文正文,保留原文对白,最终发给 Provider
22
22
  每个 `char-*` 引用必须带 `identity_binding:{profile_name,profile_sha256,appearance_id}` 和 `identity_constraints:{age_class,grooming_and_makeup,costume_signature,memory_anchors}`。提交前执行 `assertCharacterReadyForVisuals` 并精确核对人物档案 SHA、规范名、appearance 与全部可见约束,不得从 asset key 猜人物。模板只把妆发、服装标识和记忆锚点编译进可见描述,不暴露绑定元数据;连续性 `visual_identity` 有冲突或换装缺少剧情证据时写入 errors。
23
23
 
24
24
 
25
- 每镜引用清单还必须含至少一个 `image/character_identity` 人物参考,以及各一个独立的 `image/temporal_storyboard` 故事版、`image/shot_board` 分镜板和 `audio/audio_reference` 音频参考;音频参考必须是当前镜头时间范围对应的已选片段,不得用整集音轨或其他镜头音频代替。两类分镜不得指向同一版本。人物、场景、道具设定板按风险以 `image/asset_board` 追加。任一必选项缺失都不得批准提示词或提交视频。
25
+ 每镜引用清单必须包含实际 `storyboard_strategy.required_board_types` 对应的已审核分镜板;只有选择了 `temporal` 或 `shot-board` 才分别加入 `image/temporal_storyboard` 或 `image/shot_board`,不能因为执行器兼容字段自动补板。需要对白、旁白或声音连续性证据时,才加入当前镜头时间范围对应的 `audio/audio_reference`,不得用整集音轨或其他镜头音频代替。人物参考按可见人物和当前 Provider 能力加入;人物、场景、道具设定板按风险以 `image/asset_board` 追加。任一已选择板型或合同要求的参考缺失都不得批准提示词或提交视频。
26
26
 
27
27
 
28
28
  每次模板编译统一输出严格 JSON `{prompt_profile,input_mode,prompt,duration,references,continuity,audio_policy,errors}`。Seedance 2.0 固定 `full-reference`,`prompt` 必须为中文自然语言,并让 `@图片N/@视频N/@音频N` 与引用清单及提交数组一一对应;总素材不超过 12 个,图片≤9、视频≤3且合计2–15秒、音频≤3且合计≤15秒,时长4–15秒。H3 固定 `Ref2VA`,英文六段必须严格按 `subject_definitions → summary → retention_analysis → detailed_description → overall_soundscape → non_diegetic_music` 排列,仅对白、歌词和画内文字保留原语言。不得混用两套语法。
@@ -54,7 +54,7 @@
54
54
 
55
55
  图片、音频和视频生成成功后,必须立即用 `node scripts/asset-ledger.mjs <import|fetch|decode> ...` 把结果存入项目的 `assets/`,登记 SHA-256、文件大小、本地相对路径和 provenance,再允许选版或进入下游。独立配音先保存并选中 `audio-plan/vNNN.json`。资产生成与媒体制作阶段结束时用 `snapshot-stage-evidence.mjs` 生成不可变证据,避免后续合法修改总账导致早期凭证失效。远程 URL 只是临时传输结果,不得作为最终资产。
56
56
 
57
- 声音采用 `native-first`:先选同时满足画面引用和 `video.native-audio` 的 StarRouter 模型,原生生成对白、电影感旁白、环境声和动作声;BGM 走独立配乐与许可证流程。每个原生声候选必须完成 `speech_intelligibility`、`speaker_identity`、`narration_performance`、`ambience_action_sync`、`lip_sync`、`technical_audio`、`undeclared_music` 七维审核。失败时把受控 reason 和证据写入 `native_audio_exception`;当前缺少已验证的自定义音色、七维后配和口型驱动能力,不得用旧厂商自动补做这些步骤。
57
+ 声音采用 `native-first`:先选同时满足画面引用和 `video.native-audio` 的 StarRouter 模型,原生生成对白、电影感旁白、环境声和动作声;BGM 走独立配乐与许可证流程。项目已确认固定 `voice_id` 时默认走原生绑定,只有实时能力明确返回不支持该 voice reference 时才停在声音方案确认并分流,不静默改用其他音色。每个原生声候选必须完成 `speech_intelligibility`、`speaker_identity`、`narration_performance`、`ambience_action_sync`、`lip_sync`、`technical_audio`、`undeclared_music` 七维审核。失败时把受控 reason 和证据写入 `native_audio_exception`;能力返回不支持的自定义音色、七维后配和口型驱动能力不得伪装成可用,也不得用旧厂商自动补做。
58
58
 
59
59
  媒体能力按 `character-appeal → continuity-plan → previous-tail → native-audio → native-audio-review → licensed-music → local-media-editing → delivery` 推进。previous-tail 是有前置证据的可选节点;历史状态机中的 audio-fallback、lip-sync 和 video-upscale 节点目前禁用,不能为了填满流程伪造任务或切到旧厂商。
60
60
 
@@ -15,6 +15,6 @@
15
15
 
16
16
  没有 Blender 可执行文件但能安装官方 `bpy` Python 包时,可用 `uv run --python 3.11 --with bpy python <插件目录>/scripts/blender-previz.py ...` 验证无界面批处理。实际项目输出登记为 `other-previz-epNNN-NNN` 的 `other` 资产,版本文件落在 `assets/other/<key>/vNNN.mp4`;provenance 使用 `origin=generated`、`created_by=codex`、`model_or_workflow=blender-eevee-previz`,`prompt_document` 指向当前 storyboard 镜头,`parameters` 必须同时写入计划时长 `duration`、项目内导演合同路径 `direction_contract` 和该合同文件的 `direction_contract_sha256`。缺少任一绑定都不得登记候选。
17
17
 
18
- MP4 生成成功不等于白模完成。登记和审核前必须用 `ffprobe` 核对 MP4 实际时长、平均帧率、帧数和分辨率与合同、制作计划一致,并重新核对 MP4 与合同 SHA-256;不能只相信 provenance 声明。随后完整观看,至少每 0.5–1 秒抽查一次,并在动作入点、接触点、完成点、每个多机位切点和镜尾逐帧核对;检查剧情因果与节拍、人物调度与表演、相机动机与运动曲线、构图与视觉层级、空间/轴线/连续性、物理与接触、技术交付七项。把媒体 `asset_sha256`、各项得分、P0/P1/P2 问题以及与导演合同同序的 `hard_gates[]` 写入 `shot-reviews.json`;每项硬门禁必须包含 `gate/status/frame/observation` 的实际逐帧证据,任一失败即不得选版。审核入口会在自动选版前重新验证合同路径与哈希、实际媒体参数和当前制作计划版本。正式镜头还必须给出真实来源剧本、导演本、结构化分镜和制作计划版本;没有真实来源的临时演示只能标记 `test_only`,不得判断“符合剧情”、不得选版或进入正式视频。
18
+ MP4 生成成功不等于白模完成。登记和审核前必须用 `ffprobe` 核对 MP4 实际时长、平均帧率、帧数和分辨率与合同、制作计划一致,并重新核对 MP4 与合同 SHA-256;不能只相信 provenance 声明。随后完整观看,结合首/中/尾和自动检测出的异常区间记录实际观察;仅在动作入点、接触点、完成点、每个多机位切点和镜尾逐帧核对。检查剧情因果与节拍、人物调度与表演、相机动机与运动曲线、构图与视觉层级、空间/轴线/连续性、物理与接触、技术交付七项。把媒体 `asset_sha256`、各项得分、P0/P1/P2 问题以及与导演合同同序的 `hard_gates[]` 写入 `shot-reviews.json`;每项硬门禁必须包含 `gate/status/frame/observation` 的实际逐帧证据,任一失败即不得选版。审核入口会在自动选版前重新验证合同路径与哈希、实际媒体参数和当前制作计划版本。正式镜头还必须给出真实来源剧本、导演本、结构化分镜和制作计划版本;没有真实来源的临时演示只能标记 `test_only`,不得判断“符合剧情”、不得选版或进入正式视频。
19
19
 
20
20
  人数、身份、剧情因果、对白对象关系、轴线、物理接触、关键物可见性或剧本要求的镜尾反应任一未在实际画面中成立,都不得用合同字段或总分抵消。存在 P0/P1、任一硬门禁失败、总分低于 85/100 或任一项低于该项满分 70% 时,必须明确标记“不通过”,不得以“完成”“可用”或“已通过”交付;需要展示失败样片时必须同时展示审计结论。登记的 provenance `parameters.duration`、导演合同 `duration_seconds` 与制作计划镜头时长必须完全一致。通过后才可选中预演资产。`purpose=motion-reference` 当前只允许 StarRouter `MiniMax-H3`,把 selected 预演版本作为 `type=video, role=reference_video` 的执行期派生引用写入新 video-prompts 版本;公网临时发布使用 Runtime 的 TempFile.org 配置。
@@ -63,7 +63,7 @@
63
63
  | `generate-storyboard-images` | 按每镜 required_board_types 独立生成已选择的分镜板 | 制作计划批准后、视频提示词编译前 | Provider、格数、尺寸、候选数、参考图、制作计划、选版 |
64
64
  | `direct-blender-previz` / `generate-blender-previz` | standard 仅为已选静态板无法证明的复杂调度生成可选白模;原创项目仅为复杂调度追加 review 白模 | 已选分镜板之后、视频提示词编译前 | 剧情节拍、调度、轴线、运镜、评分与选版 |
65
65
  | `generate-drama-videos` | 生成并下载逐镜视频 | 已选择的分镜板、按需审核的白模主运动参考、音频参考与视频提示词确认后 | Provider、模型/工作流、时长、分辨率、参考素材、声音、制作计划、重试 |
66
- | `design-drama-audio` | 原生音频优先;做七维复听,并允许导入已授权外部音频;自定义音色克隆和口型驱动暂时禁用 | 制作计划或逐镜视频阶段 | StarRouter 语音模型、声音权利、文本、失败区间、制作计划、选版 |
66
+ | `design-drama-audio` | 原生音频优先;固定音色只在实时 Provider 能力明确支持且用户确认后使用,否则停下让用户选择支持的 Provider、后配音或不固定;做七维复听,并允许导入已授权外部音频 | 制作计划或逐镜视频阶段 | StarRouter 语音模型、声音权利、文本、失败区间、制作计划、选版 |
67
67
  | `transform-drama-media` | 本地裁剪、抽帧、剪接等版本化媒体操作;远端口型和超分暂时禁用 | 媒体生产中按需 | 来源资产、操作范围、结果选版与专项审核 |
68
68
  | `monitor-drama-tasks` | 本地登记、去重、查询异步任务并校验完成文件 | 媒体生产贯穿执行 | 取消任务、扩大重试或重新付费 |
69
69
  | `recover-drama-pipeline` | 恢复中断任务、缺失下载和 stale 下游 | 失败或续作时 | 重提、换 Provider/模型、重新付费、回退阶段 |