@lingjingai/scriptctl 0.74.2 → 0.75.0

package/dist/cli.d.ts

@@ -57,7 +57,7 @@ type OperationalFailure = FailureClassification & FailureContext & {
 };
 
 type LlmProtocol = "anthropic" | "openai-compatible" | "google-gemini" | "google-vertex";
-type LlmTaskId = "direct.episode-titles" | "direct.batch-extract" | "direct.episode-synopsis" | "direct.synopsis-reduction" | "direct.script-synopsis" | "direct.asset-grouping" | "direct.asset-curation" | "direct.asset-curation-audit" | "direct.location-canonicalization" | "direct.state-curation" | "direct.state-binding" | "metadata.global" | "metadata.descriptions" | "synopsis.chunk" | "synopsis.episode" | "synopsis.arc" | "synopsis.script-overview" | "ingest.body-start" | "ingest.arc-summary" | "ingest.text-normalize" | "ingest.episode-assets" | "ingest.name-fix" | "ingest.mention-resolve" | "ingest.asset-extract" | "ingest.asset-curate" | "ingest.asset-recall" | "ingest.asset-refine" | "ingest.state-reconcile" | "ingest.asset-audit" | "ingest.speaker-audit" | "ingest.video-serial" | "ingest.comic-serial";
+type LlmTaskId = "direct.episode-titles" | "direct.batch-extract" | "direct.episode-synopsis" | "direct.synopsis-reduction" | "direct.script-synopsis" | "direct.asset-grouping" | "direct.asset-curation" | "direct.asset-curation-audit" | "direct.location-canonicalization" | "direct.state-curation" | "direct.state-binding" | "metadata.global" | "metadata.descriptions" | "synopsis.chunk" | "synopsis.episode" | "synopsis.arc" | "synopsis.script-overview" | "ingest.body-start" | "ingest.arc-summary" | "ingest.text-normalize" | "ingest.episode-assets" | "ingest.name-fix" | "ingest.mention-resolve" | "ingest.asset-extract" | "ingest.asset-curate" | "ingest.asset-recall" | "ingest.asset-refine" | "ingest.state-reconcile" | "ingest.asset-audit" | "ingest.speaker-audit" | "ingest.video-serial" | "ingest.comic-serial" | "soundtrack.analyze";
 type JsonSchema = Record<string, unknown>;
 interface TextGenerationRequest {
     system: string;
@@ -141,6 +141,8 @@ interface DocumentGenerationRequest {
     user: string;
     responseJsonSchema?: JsonSchema;
     acceptsText?: (text: string) => boolean;
+    temperature?: number;
+    seed?: number;
     signal?: AbortSignal;
 }
 interface DocumentUnderstandingPort {
@@ -154,6 +156,8 @@ interface VideoGenerationRequest {
     user: string;
     fps: number;
     maxOutputTokens: number;
+    temperature?: number;
+    seed?: number;
     responseJsonSchema?: JsonSchema;
     acceptsText?: (text: string) => boolean;
     localPath?: string;

package/dist/cli.js

@@ -1 +1 @@
-import{a,b}from"./chunk-GIHAHL2F.js";export{a as buildProgram,b as main};
+import{a,b}from"./chunk-7AXMHZL5.js";export{a as buildProgram,b as main};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lingjingai/scriptctl",
-  "version": "0.74.2",
+  "version": "0.75.0",
   "description": "剧本阶段统一 CLI：素材转 Script v4，并兼容 Script v3 的本地写作、精修、校验和发布。",
   "type": "module",
   "bin": {

package/skills/scriptctl/SKILL.md

@@ -214,11 +214,15 @@ scriptctl review -p "$S" -s <视频目录> --converge-only --resume-log <已有
 scriptctl review -p "$S" -s <漫画ingest工作区>  # 对照 PDF 逐话审核并修正
 scriptctl create --title "…" -p "$S"          # 从零建空剧本
 scriptctl view -w <workspace>                 # 渲染或起审阅页
+scriptctl soundtrack -s <视频或目录> --episodes 1,3-5 -w <workspace>  # 听音标 BGM、音效的时间和情绪
+scriptctl soundtrack-view -w <workspace> --host <地址>  # 起网页服务，看各集汇总，点进去细看
 scriptctl checkout / status / publish         # 远端工作副本
 ```
 
 `ingest` 见 [references/ingest-workflow.md](references/ingest-workflow.md)。
 
+`soundtrack` 把结果写到 `<workspace>/soundtrack/ep_NNN.json` 和 `ep_NNN.html`，并更新汇总各集的 `index.html`，不碰 `script.json`。`soundtrack-view` 用这个目录起网页服务并一直运行：默认只有本机能打开，默认端口 8766，要在别的电脑上看就用 `--host` 传这台机器的局域网或 Tailscale 地址，这个地址上没有登录保护。单个文件视为一集，`--episodes` 指定它的集号；目录按 `--episodes` 过滤。同一来源重跑沿用已有结果，`--force` 才重新调模型。每集调用 2 次模型：第一轮只听音频定声音和起止，第二轮带清单看视频，给每段写 `emotion`、`on_screen`、`why`、`improve`，一段音乐跨几场戏时按剧情切开。`audio_emotion` 是纯听音频的情绪，`source: video` 的声音是第二轮才补的，可能按画面猜出来，要重点听。有没有和起止以人耳为准。这个任务不走豆包后备。需要 PATH 里有 ffmpeg 和 ffprobe。
+
 单个 PDF 或分话 PDF 目录自动进入漫画 ingest；`--step` 完成下一话，重跑续接已接受进度。漫画正文通过 `extend.scriptctl.source` 保存原 PDF 页码及画格/气泡/文字区域，`view` 可点击正文定位并高亮原页区域。文档模型通过 `ingest.comic-serial` 配置，内置路由为 NNYYAI `gemini-3.8-flash`，使用 `NNYYAI_API_KEY`、base64 PDF 与 `responseSchema`。
 
 `ingest` 的 Gemini 文本、结构化工具、视频和 PDF 文档请求先用配置的主模型，再按 3.8 Flash →