@docyrus/animaguru 0.29.0 → 0.30.0

package/README.md

@@ -66,7 +66,7 @@ animaguru templates
 animaguru create my-video
 ```
 
-`create` takes only the project folder name. Use a new folder whose parent already exists; the command refuses existing destinations, including empty folders and symlinks. It installs PixiJS, Vite, TypeScript and FFmpeg/FFprobe tools. npm progress goes to stderr and successful results are JSON. If installation fails, the generated files remain; enter the folder and run `npm install` to resume.
+`create` takes only the project folder name. Use a new folder whose parent already exists; the command refuses existing destinations, including empty folders and symlinks. It installs PixiJS, Vite and TypeScript. The audio tools and MP4 export use FFmpeg and FFprobe 7.1 or newer from your PATH (or `FFMPEG_PATH` and `FFPROBE_PATH`); `create` checks them and, when they are missing or older, says how to install a current build (macOS: `brew install ffmpeg`) in its notes and `ffmpeg` result. CLI 0.29.1 made that change: earlier workspaces installed their own FFmpeg 4.x builds from npm. npm progress goes to stderr and successful results are JSON. If installation fails, the generated files remain; enter the folder and run `npm install` to resume.
 
 Open the generated folder in **Codex or Claude Code**, then describe the video you want. It includes:
 
@@ -311,6 +311,23 @@ animaguru assets generate PROJECT_ID --type video --model fal/flux-3-max \
 
 `--keyframes` takes 1–10 `seconds,SHA` pairs, separated by semicolons, of this project's images or sprite art. Each is reached at its second, within `--seconds`; the clip has 24 frames a second and each keyframe needs a frame of its own. `--aspect-ratio auto` (default) follows the first keyframe, or give a ratio. FLUX 3 Max always works through the prompt before filming: `--prompt-expansion quality` asks for its thorough, slower level, any other value its light one. It takes `--seed`. Its dry runs are local estimates.
 
+CLI 0.30.0 adds HeyGen Video 1's other workflows (`heygen/heygen-video-1`). Besides image-to-video, it writes videos from text (`--workflow text-to-video`, with `--aspect-ratio`) and keeps the subjects of saved files (`--workflow reference-to-video`), for 5–15 whole seconds at 480p (default), 768p, 1080p or 2K. It makes its own sound, takes no `--soundtrack`, and takes `--seed`.
+
+```sh
+# A 30-second scene from text; clips over 15 seconds need prompt rewriting.
+animaguru assets generate PROJECT_ID --type video --model heygen/heygen-video-1 \
+  --workflow text-to-video --name "Potter" \
+  --prompt "A potter shapes a clay bowl on a wheel in a sunlit studio. One continuous shot, soft workshop sounds, no music" \
+  --seconds 30 --prompt-expansion balanced --resolution 1080p --aspect-ratio 16:9 --dry-run
+# Keep a product and a place from saved images.
+animaguru assets generate PROJECT_ID --type video --model heygen/heygen-video-1 \
+  --workflow reference-to-video --name "Harness check" \
+  --prompt "A supervisor wearing the harness in Image 1 stands still beside the line in Image 2 and speaks to camera" \
+  --references HARNESS_SHA,FACTORY_SHA --seconds 10 --resolution 768p --dry-run
+```
+
+From an image or text, HeyGen's clips run to 30 seconds while it rewrites the prompt: `--prompt-expansion balanced` asks for its fastest rewriting and `quality` its fuller one. Every second of a clip over 15 seconds costs its long-clip rate. At 1080p and 2K, text-to-video and reference-to-video are 16:9 or 9:16; from an image, the clip keeps the image's shape. `reference-to-video` takes up to nine `--references` images, three `--reference-videos` (HeyGen reads the first 5 seconds of each) and three `--audio-references`, twelve files in all, named Image 1, Video 1, Audio 1 in the prompt. It takes no `--reference`, `--end-reference` or `--middle-reference`, and `--aspect-ratio auto` (default) follows the first image or clip. HeyGen takes images of up to 5 MB, and clips and audio of up to 16 MB. Up to 5 seconds of each reference clip, and each reference image after the second, add to the cost; audio references are free.
+
 ### Vectors
 
 CLI 0.27.0 adds vectors: editable SVG files drawn by Ideogram on fal.ai (`fal/ideogram-vector`), for logos, icons, mascots and flat illustrations that stay sharp at any size and can be animated shape by shape.

package/dist/animaguru.js

@@ -19761,7 +19761,7 @@ var require_ignore2 = __commonJS({
 // cli/package.json
 var package_default = {
   name: "@docyrus/animaguru",
-  version: "0.29.0",
+  version: "0.30.0",
   description: "Agent-friendly AnimaGuru CLI built with CrustJS",
   type: "module",
   bin: {
@@ -45042,6 +45042,9 @@ var MAX_CAMERA_KEYFRAMES = 12;
 var MAX_CAMERA_TURNS = 32;
 var MAX_IMAGE_KEYFRAMES = 10;
 var KEYFRAME_FPS = 24;
+var HEYGEN_RESOLUTIONS = ["480p", "768p", "1080p", "2k"];
+var HEYGEN_WIDE = { "1080p": ["16:9", "9:16"], "2k": ["16:9", "9:16"] };
+var HEYGEN_LONG = { max: 30, creditsPerSecond: { "480p": 4.5, "768p": 7.5, "1080p": 10.5, "2k": 10.5 } };
 var H3_MAX = "minimax/h3-max";
 var FLUX_3_MAX = "blackforestlabs/flux-3-max";
 var h3Endpoints = (route) => ({ "image-to-video": `${route}/image-to-video`, "text-to-video": `${route}/text-to-video`, "extend-video": `${route}/extend-video` });
@@ -45061,11 +45064,18 @@ var SPRITE_VIDEO_MODELS = [
     id: "heygen/heygen-video-1",
     label: "HeyGen Video 1",
     provider: "heygen",
-    workflows: ["image-to-video"],
+    workflows: ["image-to-video", "text-to-video", "reference-to-video"],
     resolutions: ["480p", "768p"],
+    workflowOptions: {
+      "image-to-video": { resolutions: HEYGEN_RESOLUTIONS, long: HEYGEN_LONG },
+      "text-to-video": { resolutions: HEYGEN_RESOLUTIONS, long: HEYGEN_LONG, shapes: HEYGEN_WIDE },
+      "reference-to-video": { resolutions: HEYGEN_RESOLUTIONS, shapes: HEYGEN_WIDE, referencesOnly: true, creditsPerSecond: { "480p": 7, "768p": 10, "1080p": 18, "2k": 18 } }
+    },
     modes: ["cost"],
     lastFrame: false,
-    estimatedCreditsPerSecond: { "480p": { cost: 1 }, "768p": { cost: 1.5 } },
+    seed: true,
+    promptExpansion: PROMPT_EXPANSION_MODES,
+    estimatedCreditsPerSecond: { "480p": { cost: 2 }, "768p": { cost: 3.6 }, "1080p": { cost: 10 }, "2k": { cost: 10 } },
     seconds: { min: 5, max: 15 }
   },
   {
@@ -45145,21 +45155,29 @@ function spriteVideoModel(id2) {
 function videoResolutions(model, workflow = "image-to-video") {
   return model.workflowOptions?.[workflow]?.resolutions ?? model.resolutions;
 }
-function videoSeconds(model, workflow = "image-to-video") {
-  return model.workflowOptions?.[workflow]?.seconds ?? model.seconds;
+function videoSeconds(model, workflow = "image-to-video", rewrites = false) {
+  const options = model.workflowOptions?.[workflow];
+  const range = options?.seconds ?? model.seconds;
+  return rewrites && options?.long ? { ...range, max: options.long.max } : range;
 }
+var longVideo = (model, workflow, seconds2) => Boolean(model.workflowOptions?.[workflow]?.long) && Math.round(seconds2) > videoSeconds(model, workflow).max;
 var INCLUDED_REFERENCE_TOKENS = 4096;
 var imageReferenceTokens = (size) => size && size[0] > 0 && size[1] > 0 ? 32 * Math.ceil(32 * Math.max(...size) / Math.min(...size)) : 1824;
 var clipReferenceTokens = (seconds2, resolution) => Math.ceil(Math.max(0, seconds2) * (resolution === "480p" ? 2700 : 6900));
 var audioReferenceTokens = (seconds2) => Math.ceil(Math.max(0, seconds2) * 80);
 var DEFAULT_GENERATED_REFERENCES = 2;
+var HEYGEN_REFERENCE_CLIP_SECONDS = 5;
+var HEYGEN_FREE_REFERENCE_IMAGES = 2;
 function billedVideoSeconds(model, workflow, seconds2) {
   const s = videoSeconds(model, workflow).fractional ? seconds2 : Math.round(seconds2);
   if (workflow === "3d-to-video") return Math.max(5, s);
   return workflow === "lip-sync" && s > 15 ? s * 1.2 : s;
 }
 function videoCreditsPerSecond(model, p) {
-  const override = model.workflowOptions?.[p.workflow ?? "image-to-video"]?.creditsPerSecond?.[p.resolution];
+  const workflow = p.workflow ?? "image-to-video";
+  const options = model.workflowOptions?.[workflow];
+  const long = p.seconds !== void 0 && longVideo(model, workflow, p.seconds) ? options?.long?.creditsPerSecond[p.resolution] : void 0;
+  const override = long ?? options?.creditsPerSecond?.[p.resolution];
   if (override !== void 0) return override;
   const byMode = model.estimatedCreditsPerSecond[p.resolution] ?? model.estimatedCreditsPerSecond[model.resolutions[0]];
   return byMode[p.mode] ?? byMode[model.modes[0]] ?? 0;
@@ -45167,7 +45185,7 @@ function videoCreditsPerSecond(model, p) {
 function spriteVideoEstimate(model, p, extras2 = {}) {
   const seconds2 = billedVideoSeconds(model, p.workflow ?? "image-to-video", p.seconds);
   const beyond = Math.max(0, (extras2.referenceTokens ?? 0) - INCLUDED_REFERENCE_TOKENS) * 2e-3 + (extras2.generatedReferences ?? 0) * 3;
-  return Math.max(1, Math.ceil((videoCreditsPerSecond(model, p) * seconds2 + beyond) * UNITS_PER_CREDIT));
+  return Math.max(1, Math.ceil((videoCreditsPerSecond(model, p) * (seconds2 + (extras2.referenceSeconds ?? 0)) + beyond) * UNITS_PER_CREDIT));
 }
 
 // shared/sprites/generation-models.ts
@@ -45361,7 +45379,7 @@ async function installSkills(options) {
 import { spawn } from "node:child_process";
 import { cp, mkdir as mkdir7, readFile as readFile5, readdir as readdir3, access as access2, writeFile as writeFile6, rename as rename3 } from "node:fs/promises";
 import { basename as basename2, dirname as dirname6, delimiter as delimiter2, join as join7, resolve as resolve6 } from "node:path";
-import { fileURLToPath } from "node:url";
+import { fileURLToPath, pathToFileURL } from "node:url";
 
 // shared/locales/en/animation-styles.json
 var animation_styles_default = {
@@ -46353,7 +46371,7 @@ var animation_styles_default = {
         "Rack focus between details",
         "Callouts that track parts"
       ],
-      previewCaption: "Generated headphones turn on a stone plinth as the camera pulls back from a macro of the brushed yoke, focus racking from part to part under pinned callouts, ending on a lockup for an invented brand. Model generated with Tripo H3.1, rendered in three.js.",
+      previewCaption: "Generated headphones turn on a stone plinth as the camera pulls back from a macro of the brushed yoke, focus racking from part to part under pinned callouts, ending on a lockup for an invented brand.",
       family: "3d",
       example: true
     },
@@ -46370,7 +46388,7 @@ var animation_styles_default = {
         "Crossfaded gestures",
         "A camera that follows the character"
       ],
-      previewCaption: "A generated, auto-rigged baker walks across a painted square, turns to camera, waves and explains, his shadow falling over the painting. Character made with Tripo H3.1 and Meshy's auto-rig, rendered in three.js.",
+      previewCaption: "A generated, auto-rigged baker walks across a painted square, turns to camera, waves and explains, his shadow falling over the painting.",
       family: "3d",
       example: true
     },
@@ -46387,7 +46405,7 @@ var animation_styles_default = {
         "Little cars and figures on loops",
         "Focus pulled through the scene"
       ],
-      previewCaption: "A camera glides over a toy harbour town as little cars and boats run their loops, everything sharp only in a thin band of focus. Props generated with Tripo H3.1, rendered in three.js.",
+      previewCaption: "A camera glides over a toy harbour town as little cars and boats run their loops, everything sharp only in a thin band of focus.",
       family: "3d",
       example: true
     },
@@ -46642,6 +46660,15 @@ async function installDependencies(directory, options) {
     child.once("close", (code) => code === 0 ? resolveInstall() : reject(new Error(`npm install exited with code ${code}.`)));
   });
 }
+async function ffmpegProblem() {
+  const { findFfmpeg } = await import(pathToFileURL(join7(packageRoot, "templates/video/scripts/ffmpeg-binaries.mjs")).href);
+  try {
+    await findFfmpeg();
+    return null;
+  } catch (error62) {
+    return error62 instanceof Error ? error62.message : String(error62);
+  }
+}
 var STARTER_STORY = {
   version: 1,
   start: "opening",
@@ -46723,13 +46750,15 @@ async function createVideoProject(options, installer = installDependencies) {
   if (options.story) await writeStarterStory(directory);
   if (options.quiz) await writeStarterQuiz(directory, options.story);
   if (options.install !== false) {
-    options.log?.("Installing PixiJS, Vite, TypeScript, and the MP4 tools\u2026");
+    options.log?.("Installing PixiJS, Vite and TypeScript\u2026");
     try {
       await installer(directory, options);
     } catch (error62) {
       throw new Error(`Project files are ready at ${directory}, but dependency installation failed: ${error62 instanceof Error ? error62.message : String(error62)} Open that folder and run npm install to resume; do not rerun create into it.`);
     }
   }
+  const ffmpeg = await ffmpegProblem();
+  if (ffmpeg) options.log?.(`Note: ${ffmpeg}`);
   options.log?.(`Ready. Open ${directory} in Codex or Claude Code and describe your video.
 ${options.install === false ? "Run npm install, then " : "Run "}npm run dev inside that folder to preview.`);
   return {
@@ -46741,7 +46770,8 @@ ${options.install === false ? "Run npm install, then " : "Run "}npm run dev insi
     installed: options.install !== false,
     skills: [...skills, ...pixiSkills],
     agents: ["Codex", "Claude Code"],
-    nextSteps: ["Open this folder in Codex or Claude Code and describe your video. The agent asks for Explainer, Advertisement or General Video, then a visual style and a compact creative brief.", ...options.install === false ? ["Run npm install."] : [], "Run npm run dev to preview.", "Run npm run export:mp4 when the film is ready."]
+    ffmpeg: ffmpeg ? { ready: false, problem: ffmpeg } : { ready: true },
+    nextSteps: ["Open this folder in Codex or Claude Code and describe your video. The agent asks for Explainer, Advertisement or General Video, then a visual style and a compact creative brief.", ...options.install === false ? ["Run npm install."] : [], ...ffmpeg ? [ffmpeg] : [], "Run npm run dev to preview.", "Run npm run export:mp4 when the film is ready."]
   };
 }
 async function createWorkspace(options, installer = installDependencies) {
@@ -46794,6 +46824,8 @@ async function createWorkspace(options, installer = installDependencies) {
       throw new Error(`Workspace files are ready at ${directory}, but dependency installation failed: ${error62 instanceof Error ? error62.message : String(error62)} Run npm install there to finish.`);
     }
   }
+  const ffmpeg = result.ffmpeg;
+  if (!ffmpeg.ready) options.log?.(`Note: ${ffmpeg.problem}`);
   options.log?.(`Workspace ready: ${directory}. Open it in Codex or Claude Code with your website or brand guide, then create an animation.`);
   return {
     type: "workspace",
@@ -46804,7 +46836,8 @@ async function createWorkspace(options, installer = installDependencies) {
     installed: options.install !== false,
     branding: join7(directory, "branding"),
     animations: join7(directory, "animations"),
-    nextSteps: ["Open this workspace in Codex or Claude Code and provide your website or brand guide; the agent asks if it is missing.", "Build the shared brand kit in branding/.", "Run animaguru create animation my-video.", "Run npm run dev for the animation hub."]
+    ffmpeg,
+    nextSteps: ["Open this workspace in Codex or Claude Code and provide your website or brand guide; the agent asks if it is missing.", "Build the shared brand kit in branding/.", ...ffmpeg.ready ? [] : [ffmpeg.problem], "Run animaguru create animation my-video.", "Run npm run dev for the animation hub."]
   };
 }
 async function findWorkspace(start = process.cwd()) {
@@ -47086,14 +47119,19 @@ function videoChecks(p, ctx) {
   if (!model.workflows.includes(p.workflow)) return issue2("workflow", "Choose a way of making video supported by the video model.");
   const w = p.workflow;
   const extend2 = w === "extend-video";
-  const range = videoSeconds(model, w);
+  const rewrites = p.promptExpansion !== "disabled";
+  const range = videoSeconds(model, w, rewrites);
   const seconds2 = p.seconds ?? defaultSeconds(model, w);
   const resolution = p.resolution ?? videoResolutions(model, w)[0];
+  const options = model.workflowOptions?.[w];
   const prompt = PROMPT[w];
   if (prompt === "none" && p.prompt) issue2("prompt", `${w} takes no prompt; leave it out.`);
   if (!prompt && p.prompt.length < 3) issue2("prompt", "Describe the video in at least 3 characters.");
   if ((w === "recast" || w === "3d-to-video") && p.prompt.length > 2e3) issue2("prompt", `${w} takes a prompt of at most 2,000 characters.`);
-  if (!lengthFromMedia(w) && (seconds2 < range.min || seconds2 > range.max || !range.fractional && !Number.isInteger(seconds2))) issue2("seconds", "Choose a duration supported by the video model.");
+  if (!lengthFromMedia(w) && (seconds2 < range.min || seconds2 > range.max || !range.fractional && !Number.isInteger(seconds2))) {
+    if (!rewrites && options?.long && Number.isInteger(seconds2) && seconds2 > range.max && seconds2 <= options.long.max) issue2("seconds", `Set promptExpansion to balanced or quality for a video over ${range.max} seconds.`);
+    else issue2("seconds", "Choose a duration supported by the video model.");
+  }
   if (!videoResolutions(model, w).includes(resolution)) issue2("resolution", "Choose a resolution supported by the video model.");
   if (p.mode && !model.modes.includes(p.mode)) issue2("mode", "Choose a mode supported by the video model.");
   for (const [field, workflows] of Object.entries(TAKES)) {
@@ -47116,16 +47154,22 @@ function videoChecks(p, ctx) {
   if (p.damage && p.style !== "vhs") issue2("damage", "Only the VHS style takes tape damage.");
   if (p.soundtrack && !model.audio) issue2("soundtrack", "This video model can't use a soundtrack.");
   if (p.soundtrack && w !== "lip-sync" && seconds2 < 2) issue2("seconds", "A video with a soundtrack must be at least 2 seconds long.");
-  if (p.audioReferences.length && !model.audio) issue2("audioReferences", "Only an extension takes audio references.");
+  if (p.audioReferences.length && !model.audio && w !== "reference-to-video") issue2("audioReferences", "Only an extension takes audio references.");
   if (p.promptExpansion !== "disabled" && !model.promptExpansion?.includes(p.promptExpansion)) issue2("promptExpansion", "This video model doesn't rewrite prompts.");
   if (p.seed !== void 0 && !model.seed) issue2("seed", "This video model doesn't take a seed.");
   if (p.seed !== void 0 && p.seed >= 2 ** 31 && (w === "lip-sync" || w === "insert-video")) issue2("seed", `${w} takes a seed below 2,147,483,648.`);
   const references = p.references ?? [];
   const unique = (shas2) => new Set(shas2).size === shas2.length;
   if (!unique(references) || !unique(p.referenceVideos ?? []) || !unique(p.audioReferences)) issue2("references", "Choose each reference once.");
+  const shapes = options?.shapes?.[resolution];
+  if (shapes && !shapes.includes(p.aspectRatio ?? (w === "text-to-video" ? "16:9" : "auto"))) issue2("aspectRatio", `At ${resolution === "2k" ? "2K" : resolution}, ${w} with this video model is ${shapes.join(" or ")}; choose one as the aspect ratio.`);
   if (w === "reference-to-video") {
+    if (options?.referencesOnly) {
+      for (const field of ["reference", "endReference", "middleReference"]) if (p[field]) issue2(field, "This video model's reference-to-video takes references only, without a first, last or middle frame.");
+    }
     if (references.length + (p.referenceVideos?.length ?? 0) + p.audioReferences.length > MAX_REFERENCE_FILES) issue2("references", `reference-to-video takes at most ${MAX_REFERENCE_FILES} reference files in all.`);
-    if (!references.length && !p.referenceVideos?.length && !p.audioReferences.length && !p.reference && !p.endReference) issue2("references", "reference-to-video needs reference images, videos or audio, or a first or last frame.");
+    const frames = !options?.referencesOnly && (p.reference || p.endReference);
+    if (!references.length && !p.referenceVideos?.length && !p.audioReferences.length && !frames) issue2("references", `reference-to-video needs reference images, videos or audio${options?.referencesOnly ? "" : ", or a first or last frame"}.`);
     if (Boolean(p.middleReference) !== (p.middleSeconds !== void 0)) issue2("middleSeconds", "A middle image needs middleSeconds, and middleSeconds a middle image.");
     if (p.middleReference && (!p.reference || !p.endReference)) issue2("middleReference", "A middle image needs the starting and ending images too.");
     if (p.middleReference && !["480p", "768p"].includes(resolution)) issue2("middleReference", "A middle image needs 480p or 768p.");
@@ -47276,6 +47320,12 @@ var imageEstimate = (p) => Math.ceil((130 + 40 * p.references.length) * spriteIm
 var vectorEstimate = (p) => p.workflow === "image-to-vector" ? 100 : p.quality === "low" ? 130 : p.quality === "high" ? p.references.length ? 320 : 200 : 160;
 function videoExtras(p, facts = {}) {
   const w = p.workflow;
+  if (spriteVideoModel(p.model).provider === "heygen") {
+    if (w !== "reference-to-video") return {};
+    const clips3 = (p.referenceVideos ?? []).reduce((sum, sha2) => sum + Math.min(HEYGEN_REFERENCE_CLIP_SECONDS, facts.seconds?.(sha2) ?? HEYGEN_REFERENCE_CLIP_SECONDS), 0);
+    const seconds2 = clips3 + Math.max(0, (p.references?.length ?? 0) - HEYGEN_FREE_REFERENCE_IMAGES);
+    return seconds2 ? { referenceSeconds: seconds2 } : {};
+  }
   if (w !== "reference-to-video" && w !== "insert-video" && w !== "3d-to-video") return {};
   const images = (p.references ?? []).reduce((sum, sha2) => sum + imageReferenceTokens(facts.size?.(sha2)), 0);
   if (w === "3d-to-video") return { referenceTokens: images + clipReferenceTokens(p.seconds, p.resolution), generatedReferences: p.references?.length ? 0 : p.generatedReferences ?? DEFAULT_GENERATED_REFERENCES };
@@ -50931,15 +50981,15 @@ Confirm code: ${device.user_code}`);
     { name: "transparent", type: "boolean", description: "Image: transparent background" },
     { name: "references", type: "string", description: "Comma-separated blob SHAs of this project's images or sprite art, named Image 1, Image 2\u2026 in the prompt. Image: up to 4 to follow, such as a character sheet. Vector from text: up to 5, the first leading. Video: subjects or looks to keep (reference-to-video and insert-video up to 9, 3d-to-video up to 8), or recast's new people, one photo each (up to 4)" },
     { name: "mask", type: "string", description: "Image: edit only these rectangles of Image 1 (the first --references image), as x,y,width,height fractions of its size from the top left, up to 8 separated by semicolons; the rest is kept" },
-    { name: "workflow", type: "string", choices: [...VIDEO_WORKFLOWS, ...VECTOR_WORKFLOWS], description: "Video: image-to-video (default), text-to-video or extend-video; H3 Max also reference-to-video, lip-sync, camera-controls, insert-video, recast, relight, 3d-to-video and style, FLUX 3 Max keyframes-to-video, as the model's videoProfiles allow. Vector: text-to-vector (default) or image-to-vector, which traces --source" },
+    { name: "workflow", type: "string", choices: [...VIDEO_WORKFLOWS, ...VECTOR_WORKFLOWS], description: "Video: image-to-video (default), text-to-video or extend-video; H3 Max also reference-to-video, lip-sync, camera-controls, insert-video, recast, relight, 3d-to-video and style, FLUX 3 Max keyframes-to-video, HeyGen Video 1 reference-to-video, as the model's videoProfiles allow. Vector: text-to-vector (default) or image-to-vector, which traces --source" },
     { name: "reference", type: "string", description: "Video: starting image blob SHA from this project's assets or sprites; the speaking face for lip-sync, the image the camera moves around for camera-controls. Model: the image or sprite art to build it from (image-to-3D); without one the prompt describes it" },
     { name: "end-reference", type: "string", description: "Video: ending image blob SHA; models with endFrameOnly accept it without --reference" },
     { name: "middle-reference", type: "string", description: "Reference-to-video: image blob SHA reached at --middle-seconds; needs --reference and --end-reference, 480p or 768p" },
     { name: "middle-seconds", type: "number", description: "Reference-to-video: when the middle image is reached, in seconds" },
     { name: "source", type: "string", description: "Blob SHA of a video asset in this project: the clip to extend, insert a scene into, recast or relight, or the 3D render 3d-to-video turns into footage. Vector image-to-vector: the image or sprite art to trace" },
     { name: "soundtrack", type: "string", description: "Video from an image or text: blob SHA of a music, speech or SFX asset pinned as the soundtrack (2 s or longer). Lip-sync: the speech to say (5 s\u201315 min), which sets the length" },
-    { name: "audio-references", type: "string", description: "Extension or reference-to-video: up to 3 comma-separated audio asset blob SHAs (2\u201315 s each, 15 s total) guiding the sound; refer to them as Audio 1, Audio 2\u2026 in the prompt" },
-    { name: "reference-videos", type: "string", description: "Reference-to-video or insert-video: up to 3 comma-separated video asset blob SHAs (2\u201315 s each, 15 s total) guiding motion, named Video 1, Video 2\u2026 in the prompt" },
+    { name: "audio-references", type: "string", description: "Extension or reference-to-video: up to 3 comma-separated audio asset blob SHAs (on fal 2\u201315 s each, 15 s total) guiding the sound; refer to them as Audio 1, Audio 2\u2026 in the prompt" },
+    { name: "reference-videos", type: "string", description: "Reference-to-video or insert-video: up to 3 comma-separated video asset blob SHAs (on fal 2\u201315 s each, 15 s total; HeyGen reads the first 5 s of each) guiding motion, named Video 1, Video 2\u2026 in the prompt" },
     { name: "camera", type: "string", description: 'Camera-controls: 2\u201312 keyframes time,azimuth,elevation,distance (time 0\u20131, angles in degrees, elevation \u221290\u201390) separated by semicolons, e.g. "0,0,0,1;1,90,15,1.2"' },
     { name: "keyframes", type: "string", description: 'Keyframes-to-video (FLUX 3 Max): 1\u201310 seconds,image-SHA pairs separated by semicolons, e.g. "0,SHA;2.5,SHA;5,SHA"; each image is reached at its second, on a frame of its own of the 24 fps clip, within --seconds' },
     { name: "insert-at", type: "number", description: "Insert-video: seconds into --source where the new scene starts (1.625\u201360)" },
@@ -50952,10 +51002,10 @@ Confirm code: ${device.user_code}`);
     { name: "generated-references", type: "number", description: "3d-to-video without --references: how many reference images it may draw, 1\u20138 (default 2)" },
     { name: "style", type: "string", choices: VIDEO_STYLES, description: "Style workflow: the preset look, at 768p for 5\u201315 whole seconds" },
     { name: "damage", type: "string", choices: VHS_DAMAGE_LEVELS, description: "VHS style: tape damage (fal's default medium)" },
-    { name: "seconds", type: "number", description: "Video or SFX duration in seconds (default 5); H3 takes fractions (6.5), FLUX 3 Max 5\u201320 whole seconds; extensions add this many seconds, insert-video's scene is 5\u201313; lip-sync, recast, relight and 3d-to-video take their input's length; SFX supports 0.5\u201330" },
+    { name: "seconds", type: "number", description: "Video or SFX duration in seconds (default 5); H3 takes fractions (6.5), FLUX 3 Max 5\u201320 whole seconds, HeyGen Video 1 5\u201315 (up to 30 from an image or text with --prompt-expansion balanced or quality); extensions add this many seconds, insert-video's scene is 5\u201313; lip-sync, recast, relight and 3d-to-video take their input's length; SFX supports 0.5\u201330" },
     { name: "resolution", type: "string", choices: VIDEO_RESOLUTIONS, description: "Video resolution supported by the selected model and workflow (default its first)" },
     { name: "mode", type: "string", choices: ["cost", "speed", "quality"], description: "Video generation mode supported by the selected model (default its first)" },
-    { name: "prompt-expansion", type: "string", choices: PROMPT_EXPANSION_MODES, description: "Video: how much the model may rewrite the prompt (default disabled); extensions treat any other value as on. FLUX 3 Max always reasons about the prompt: quality thoroughly, any other value lightly" },
+    { name: "prompt-expansion", type: "string", choices: PROMPT_EXPANSION_MODES, description: "Video: how much the model may rewrite the prompt (default disabled); extensions treat any other value as on. FLUX 3 Max always reasons about the prompt: quality thoroughly, any other value lightly. HeyGen Video 1 needs balanced or quality for clips over 15 s" },
     { name: "seed", type: "number", description: "Video or vector from text: seed for repeatable results, where the model takes one" },
     { name: "extend-output", type: "string", choices: EXTEND_OUTPUTS, description: "Extension: extended (default, source plus new footage) or continuation (a 1.625 s overlap, then the new footage)" },
     { name: "voice", type: "string", description: "Gemini voice name or ElevenLabs voice ID (defaults: Kore / George)" },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@docyrus/animaguru",
-  "version": "0.29.0",
+  "version": "0.30.0",
   "description": "Agent-friendly AnimaGuru CLI built with CrustJS",
   "type": "module",
   "bin": {

package/skills/animaguru/SKILL.md

@@ -62,7 +62,7 @@ animaguru assets generate PROJECT_ID --type speech --name "Welcome" \
   --prompt "Welcome, traveler." --voice Kore --delivery "Warm and welcoming" --yes
 ```
 
-Images use the staff-managed route automatically. Use `--model` to select a model for video, music, speech or sound effects; discover the choices with `catalog` and the server's configured subset with `assets list`. Asset defaults are per type, independently of sprite project defaults. Image backgrounds are retained unless `--transparent` is supplied. CLI 0.17.0 adds `--references`, `--mask` and the 16:9 and 9:16 image shapes. `--references SHA,SHA` (up to four) makes an image follow this project's saved images, uploads or sprite art. To keep a recurring character consistent, generate a character sheet once and reference its `file` SHA in every later image; refer to references in the prompt as Image 1, Image 2 and so on, and say what to take from each. `--aspect-ratio` is 1:1, 3:2, 2:3, 16:9 or 9:16; use 16:9 or 9:16 for video scenes. `--mask x,y,width,height[;…]` (up to eight rectangles, in fractions of Image 1 from the top left) edits only those parts of Image 1, the first reference, and keeps the rest and Image 1's shape. Look at the image before choosing the rectangles. Videos require an image asset's `file` SHA from `assets list`, or a character's `art` SHA from `projects get`, in the same project. Validate video options against the selected model's `videoProfiles` in `catalog`; omitted video settings use five seconds and that provider's first resolution/mode. MiniMax H3 Max and H3 Max Turbo (`fal/minimax-h3-max`, `fal/minimax-h3-max-turbo`) also take `--workflow text-to-video` (with `--aspect-ratio`) and `--workflow extend-video` (with `--source VIDEO_SHA`, 2K resolution, `--extend-output`, and up to three comma-separated `--audio-references`). From images, they accept `--end-reference` with or without `--reference` and a `--soundtrack AUDIO_SHA`. All workflows take `--prompt-expansion` and `--seed`. These need CLI 0.8.0 or newer. CLI 0.18.0 adds H3 Max's (not Turbo's) `--workflow reference-to-video` (keep a character across clips from up to nine `--references`, three `--reference-videos` and three `--audio-references`, named Image 1, Video 1, Audio 1 in the prompt; optional `--reference`, `--end-reference`, `--middle-reference` with `--middle-seconds`), `lip-sync` (the face in `--reference` says `--soundtrack` speech of 5 s–15 min; no `--prompt`; `--transcribe`), `camera-controls` (`--camera "time,azimuth,elevation,distance;…"`, 2–12 keyframes around `--reference`; without a prompt only the camera moves), `insert-video` (a 5–13 s scene inside `--source` from `--insert-at` to `--resume-at`, guided by `--references`/`--reference-videos`), `recast` (people in `--references` replace those in `--source`; pass `--subject-consent` only after each person agreed), `relight` (`--source` relit from a `--lighting` sphere image; no prompt), `3d-to-video` (a 3D render `--source` into footage, from `--references` or `--generated-references`) and `style` (`--style vhs|retro-toon-70s|low-poly|hand-drawn|16bit-pixel`, `--damage` for VHS, 768p, 5–15 whole seconds). Lip-sync, recast, relight and 3d-to-video take their input's length; their dry runs, like reference-to-video's and insert-video's, are server estimates because the price depends on the files. H3 takes fractional `--seconds` (6.5). CLI 0.21.0 adds FLUX 3 Max (`fal/flux-3-max`): image-to-video (with an optional `--end-reference`) and text-to-video for 5–20 whole seconds at 720p or 1080p with its own sound and no `--soundtrack`, and `--workflow keyframes-to-video`, which passes through 1–10 `--keyframes "SECONDS,SHA;…"` images, each reached at its second within `--seconds`. It always reasons about the prompt: `--prompt-expansion quality` thoroughly, other values lightly. For Gemini speech, `--prompt` is spoken verbatim and `--delivery` is a separate instruction. Music uses Lyria; the Gemini TTS models are for speech.
+Images use the staff-managed route automatically. Use `--model` to select a model for video, music, speech or sound effects; discover the choices with `catalog` and the server's configured subset with `assets list`. Asset defaults are per type, independently of sprite project defaults. Image backgrounds are retained unless `--transparent` is supplied. CLI 0.17.0 adds `--references`, `--mask` and the 16:9 and 9:16 image shapes. `--references SHA,SHA` (up to four) makes an image follow this project's saved images, uploads or sprite art. To keep a recurring character consistent, generate a character sheet once and reference its `file` SHA in every later image; refer to references in the prompt as Image 1, Image 2 and so on, and say what to take from each. `--aspect-ratio` is 1:1, 3:2, 2:3, 16:9 or 9:16; use 16:9 or 9:16 for video scenes. `--mask x,y,width,height[;…]` (up to eight rectangles, in fractions of Image 1 from the top left) edits only those parts of Image 1, the first reference, and keeps the rest and Image 1's shape. Look at the image before choosing the rectangles. Videos require an image asset's `file` SHA from `assets list`, or a character's `art` SHA from `projects get`, in the same project. Validate video options against the selected model's `videoProfiles` in `catalog`; omitted video settings use five seconds and that provider's first resolution/mode. MiniMax H3 Max and H3 Max Turbo (`fal/minimax-h3-max`, `fal/minimax-h3-max-turbo`) also take `--workflow text-to-video` (with `--aspect-ratio`) and `--workflow extend-video` (with `--source VIDEO_SHA`, 2K resolution, `--extend-output`, and up to three comma-separated `--audio-references`). From images, they accept `--end-reference` with or without `--reference` and a `--soundtrack AUDIO_SHA`. All workflows take `--prompt-expansion` and `--seed`. These need CLI 0.8.0 or newer. CLI 0.18.0 adds H3 Max's (not Turbo's) `--workflow reference-to-video` (keep a character across clips from up to nine `--references`, three `--reference-videos` and three `--audio-references`, named Image 1, Video 1, Audio 1 in the prompt; optional `--reference`, `--end-reference`, `--middle-reference` with `--middle-seconds`), `lip-sync` (the face in `--reference` says `--soundtrack` speech of 5 s–15 min; no `--prompt`; `--transcribe`), `camera-controls` (`--camera "time,azimuth,elevation,distance;…"`, 2–12 keyframes around `--reference`; without a prompt only the camera moves), `insert-video` (a 5–13 s scene inside `--source` from `--insert-at` to `--resume-at`, guided by `--references`/`--reference-videos`), `recast` (people in `--references` replace those in `--source`; pass `--subject-consent` only after each person agreed), `relight` (`--source` relit from a `--lighting` sphere image; no prompt), `3d-to-video` (a 3D render `--source` into footage, from `--references` or `--generated-references`) and `style` (`--style vhs|retro-toon-70s|low-poly|hand-drawn|16bit-pixel`, `--damage` for VHS, 768p, 5–15 whole seconds). Lip-sync, recast, relight and 3d-to-video take their input's length; their dry runs, like reference-to-video's and insert-video's, are server estimates because the price depends on the files. H3 takes fractional `--seconds` (6.5). CLI 0.21.0 adds FLUX 3 Max (`fal/flux-3-max`): image-to-video (with an optional `--end-reference`) and text-to-video for 5–20 whole seconds at 720p or 1080p with its own sound and no `--soundtrack`, and `--workflow keyframes-to-video`, which passes through 1–10 `--keyframes "SECONDS,SHA;…"` images, each reached at its second within `--seconds`. It always reasons about the prompt: `--prompt-expansion quality` thoroughly, other values lightly. CLI 0.30.0 adds HeyGen Video 1's (`heygen/heygen-video-1`) `--workflow text-to-video` (with `--aspect-ratio`) and `reference-to-video` (up to nine `--references`, three `--reference-videos`, of which it reads the first 5 s, and three `--audio-references`, named Image 1, Video 1, Audio 1 in the prompt; no `--reference`, `--end-reference` or `--middle-reference`), 1080p and 2K (16:9 or 9:16 from text or references), and `--seed`. From an image or text it runs to 30 whole seconds with `--prompt-expansion balanced` or `quality`, and every second past 15 costs more. For Gemini speech, `--prompt` is spoken verbatim and `--delivery` is a separate instruction. Music uses Lyria; the Gemini TTS models are for speech.
 
 For ElevenLabs speech, choose `eleven_v4` or `eleven_v4_turbo`, pass an ElevenLabs voice ID with `--voice` (George is the default), and put audio tags in `--prompt` rather than using `--delivery`. For sound effects, use `--type sfx`; the only model is `eleven_text_to_sound_v2`. SFX supports `--seconds` from 0.5 to 30 (default 5), `--loop`, and `--prompt-influence` from 0 to 1 (default 0.3). These additions require CLI 0.5.0 or newer; check `catalog` before using them with an installed CLI.
 

package/templates/video/README.md

@@ -10,7 +10,7 @@ Example prompt:
 
 ## Preview
 
-Requires Node.js 24+. Dependencies are installed by `animaguru create NAME`. If installation was interrupted, run `npm install` inside this folder to finish it.
+Requires Node.js 24+, and FFmpeg 7.1 or newer with FFprobe for the audio tools and MP4 export: they run the `ffmpeg` and `ffprobe` on PATH, or `FFMPEG_PATH` and `FFPROBE_PATH` (macOS: `brew install ffmpeg`; Windows: `winget install Gyan.FFmpeg`; Debian 13: `sudo apt install ffmpeg`). Dependencies are installed by `animaguru create NAME`. If installation was interrupted, run `npm install` inside this folder to finish it.
 
 ```sh
 npm run dev
@@ -71,9 +71,9 @@ npm run export:mp4 -- --burn-captions
 npm run export:mp4 -- --watermark
 ```
 
-Open the local capture URL printed by the command and keep it open until complete. The script runs a separate loopback Vite server (default port 5196), streams high-quality JPEG frames from the Pixi canvas to the bundled FFmpeg binary, and mixes the configured soundtrack. The initial starter exports with a silent AAC track. Exporting local media does not call AnimaGuru or spend credits. FFmpeg/FFprobe binaries are supplied through platform-specific npm dependencies; installation needs a supported platform and network access.
+Open the local capture URL printed by the command and keep it open until complete. The script runs a separate loopback Vite server (default port 5196), streams high-quality JPEG frames from the Pixi canvas to FFmpeg, and mixes the configured soundtrack. An AAC soundtrack at 48 kHz that lasts the whole film, such as `npm run audio:mix` writes, is copied into the MP4 as it is; any other soundtrack is encoded to AAC and padded to length. The initial starter exports with a silent AAC track. Exporting local media does not call AnimaGuru or spend credits. The export stops before capturing when FFmpeg or FFprobe is missing or older than 7.1.
 
-Completed H.264/AAC MP4s and JSON verification reports appear in `production/exports/`. Captions are not drawn into the MP4 picture: the film's `captions` resource track is saved beside it as `SLUG-LANGUAGE.srt` (for YouTube and desktop players, which load a same-named file automatically) and `SLUG-LANGUAGE.vtt` (for an HTML5 `<track>`). `--burn-captions` instead draws them into the picture and writes no subtitle files, so players cannot show them twice. `"poster"` in `film.config.json` (seconds) picks the frame saved as `SLUG-LANGUAGE.jpg`, a caption-free thumbnail 1280 pixels on its longer side and under YouTube's 2 MB limit; without it, no thumbnail is saved. The live player and HTML5 export keep their CC toggle. `--watermark` also saves `SLUG-LANGUAGE-watermarked.mp4`, the same film with AnimaGuru's "Made with AnimaGuru" badge in its bottom-right corner, drawn with the font, logo and layout AnimaGuru's web app uses; it is the preview AnimaGuru's lead pipeline plays before a lead claims the clean film. The picture is encoded again with the badge and verified like the film; the sound is copied. Existing files are preserved. Frame retries are deduplicated. Ctrl+C removes only this session's unfinished output. The verification checks duration, exact frame count, dimensions, pixel format, codecs and audio language tag. The exporter supports even dimensions up to 4096 pixels, 24/25/30/60 fps and films up to ten minutes. Set duration to an exact number of frames.
+Completed H.264/AAC MP4s and JSON verification reports appear in `production/exports/`. Captions are not drawn into the MP4 picture: the film's `captions` resource track is saved beside it as `SLUG-LANGUAGE.srt` (for YouTube and desktop players, which load a same-named file automatically) and `SLUG-LANGUAGE.vtt` (for an HTML5 `<track>`). `--burn-captions` instead draws them into the picture and writes no subtitle files, so players cannot show them twice. `"poster"` in `film.config.json` (seconds) picks the frame saved as `SLUG-LANGUAGE.jpg`, a caption-free thumbnail 1280 pixels on its longer side and under YouTube's 2 MB limit; without it, no thumbnail is saved. The live player and HTML5 export keep their CC toggle. `--watermark` also saves `SLUG-LANGUAGE-watermarked.mp4`, the same film with AnimaGuru's "Made with AnimaGuru" badge in its bottom-right corner, drawn with the font, logo and layout AnimaGuru's web app uses; it is the preview AnimaGuru's lead pipeline plays before a lead claims the clean film. The picture is encoded again with the badge and verified like the film; the sound is copied. Existing files are preserved. Frame retries are deduplicated. Ctrl+C removes only this session's unfinished output. The verification checks duration, exact frame count, dimensions, pixel format, codecs and audio language tag. It also measures the MP4's own sound (integrated loudness and true peak, saved as `soundtrack` in the JSON report) and warns when its true peak is above -1 dBTP. The exporter supports even dimensions up to 4096 pixels, 24/25/30/60 fps and films up to ten minutes. Set duration to an exact number of frames.
 
 ## Capture API
 

package/templates/video/package.json

@@ -25,8 +25,6 @@
     "three": "0.186.1"
   },
   "devDependencies": {
-    "@ffmpeg-installer/ffmpeg": "1.1.0",
-    "@ffprobe-installer/ffprobe": "2.1.2",
     "@types/three": "0.186.0",
     "fflate": "0.8.3",
     "playwright-core": "1.64.0",

package/templates/video/scripts/audio-mix.mjs

@@ -7,13 +7,35 @@ import { mkdir, readFile, writeFile, access } from "node:fs/promises";
 import path from "node:path";
 import { fileURLToPath } from "node:url";
 import { projectArguments, projectContext } from "./project-context.mjs";
-import { ffmpeg, measureLoudness, parseFlags, projectPath, readJson, round } from "./audio-tools.mjs";
+import { ffmpeg, loudnormReport, measureLoudness, parseFlags, projectPath, readJson, round } from "./audio-tools.mjs";
 
 const RATE = 48000;
 export const MIX_DEFAULTS = { musicDb: -9, duck: { threshold: 0.03, ratio: 6, attack: 20, release: 350 }, loudness: { i: -14, tp: -2, lra: 9 } };
+// The largest LRA target loudnorm accepts (FFmpeg 7.1 and newer).
+const LRA_LIMIT = 50;
+
+/**
+ * The second loudnorm pass: one linear gain from the first pass's measurements. loudnorm stays linear only while the
+ * measured loudness range fits the LRA target and the gain keeps the true peak under TP; otherwise it switches to
+ * dynamic mode without an error and rides the gain, which can pump. The LRA target does nothing else on a linear
+ * pass, so it is widened to the mix's own range unless mix.json sets loudness.lra. `blockers` say what keeps the pass
+ * from staying linear, each with the change that would let it.
+ */
+export function loudnormPass(first, loudness, { keepRange = false } = {}) {
+  const lra = keepRange ? loudness.lra : Math.min(LRA_LIMIT, Math.max(loudness.lra, Math.ceil(first.range * 10) / 10));
+  const gain = loudness.i - first.integrated, peak = first.truePeak + gain, blockers = [];
+  if (first.range > lra) blockers.push(keepRange
+    ? { reason: `its loudness range of ${round(first.range, 1)} LU is wider than loudness.lra (${lra} LU)`, fix: "raise or remove loudness.lra" }
+    : { reason: `its loudness range of ${round(first.range, 1)} LU is wider than the ${LRA_LIMIT} LU loudnorm accepts`, fix: "bring the quietest and loudest passages closer together" });
+  // The loudest target one gain reaches without passing the ceiling, rounded down so it stays under it.
+  const loudest = Math.floor((loudness.tp - first.truePeak + first.integrated) * 10) / 10;
+  if (peak > loudness.tp) blockers.push({ reason: `a ${gain >= 0 ? "+" : ""}${round(gain, 1)} dB gain would lift its true peak to ${round(peak, 1)} dBTP, above the ${loudness.tp} dBTP ceiling`, fix: `lower loudness.i to ${loudest} or below` });
+  const filter = `loudnorm=I=${loudness.i}:TP=${loudness.tp}:LRA=${lra}:measured_I=${first.integrated}:measured_TP=${first.truePeak}:measured_LRA=${first.range}:measured_thresh=${first.threshold}:offset=${first.offset}:linear=true:print_format=json`;
+  return { filter, lra, gain, blockers };
+}
 
 /** The music part of the graph for input `input`, as edited by audio:music (a trim or a bar-aligned cut). */
-// Padding counts samples at RATE: the bundled Linux ffmpeg (@ffmpeg-installer) has no apad whole_dur option.
+// Padding counts samples at RATE, so the padded length is exact.
 export function musicGraph(edit, { input = 0, duration, musicDb = MIX_DEFAULTS.musicDb, label = "music" }) {
   const headIn = Math.max(0, Number(edit?.headIn ?? 0)), fadeIn = Number(edit?.fadeIn ?? (headIn > 0.05 ? 0.25 : 0.02)), fadeOut = Number(edit?.fadeOut ?? 0.6);
   const ending = `atrim=0:${duration},apad=whole_len=${Math.round(duration * RATE)},afade=t=out:st=${round(Math.max(0, duration - fadeOut))}:d=${fadeOut},volume=${musicDb}dB,aformat=sample_fmts=fltp:sample_rates=${RATE}:channel_layouts=stereo[${label}]`;
@@ -81,8 +103,10 @@ async function main() {
   await ffmpeg(["-y", "-loglevel", "error", ...(music ? ["-i", music] : []), ...(vo ? ["-i", vo] : []), "-filter_complex", graph, "-map", "[mix]", "-ar", String(RATE), "-ac", "2", "-c:a", "pcm_s24le", "-t", String(duration), premaster]);
   // Two-pass loudnorm: measure, then normalise linearly with the measured values.
   const first = await measureLoudness(premaster, loudness);
-  const second = `loudnorm=I=${loudness.i}:TP=${loudness.tp}:LRA=${loudness.lra}:measured_I=${first.integrated}:measured_TP=${first.truePeak}:measured_LRA=${first.range}:measured_thresh=${first.threshold}:offset=${first.offset}:linear=true`;
-  await ffmpeg(["-y", "-loglevel", "error", "-i", premaster, "-af", second, "-ar", String(RATE), "-ac", "2", "-c:a", "pcm_s24le", "-t", String(duration), master]);
+  const pass = loudnormPass(first, loudness, { keepRange: mix.loudness?.lra !== undefined });
+  // No -loglevel error here: loudnorm reports which mode it used at the info level.
+  const { stderr } = await ffmpeg(["-y", "-nostats", "-i", premaster, "-af", pass.filter, "-ar", String(RATE), "-ac", "2", "-c:a", "pcm_s24le", "-t", String(duration), master]);
+  const mode = loudnormReport(stderr)?.normalization_type;
   await ffmpeg(["-y", "-loglevel", "error", "-i", master, "-c:a", "aac", "-b:a", "192k", "-ar", String(RATE), "-ac", "2", "-movflags", "+faststart", output]);
   const result = await measureLoudness(output, loudness);
 
@@ -93,7 +117,9 @@ async function main() {
 
   const parts = [vo ? `voiceover ${path.relative(base, vo)}` : null, music ? `music ${path.relative(base, music)}${edit?.type === "cut" ? ` (${edit.headIn}–${edit.cutA} s → ${edit.cutB} s)` : edit ? ` (from ${edit.headIn} s)` : ""}` : null].filter(Boolean);
   console.log(`Mixed ${parts.join(" + ")}${music && vo ? `, music at ${mix.musicDb ?? MIX_DEFAULTS.musicDb} dB, ducked under the voice` : ""}.`);
-  console.log(`Before normalising: ${round(first.integrated, 1)} LUFS, ${round(first.truePeak, 1)} dBTP. Encoded AAC: ${round(result.integrated, 1)} LUFS, ${round(result.truePeak, 1)} dBTP (target ${loudness.i} LUFS, ${loudness.tp} dBTP).`);
+  const how = mode === "linear" ? ` Normalised with one ${pass.gain >= 0 ? "+" : ""}${round(pass.gain, 1)} dB gain.` : mode === "dynamic" ? " Normalised with dynamic gain." : "";
+  console.log(`Before normalising: ${round(first.integrated, 1)} LUFS, ${round(first.truePeak, 1)} dBTP, ${round(first.range, 1)} LU range.${how} Encoded AAC: ${round(result.integrated, 1)} LUFS, ${round(result.truePeak, 1)} dBTP (target ${loudness.i} LUFS, ${loudness.tp} dBTP).`);
+  if (mode === "dynamic") console.log(`Warning: loudnorm rode the gain instead of applying one gain, which can pump${pass.blockers.length ? `, because ${pass.blockers.map((item) => item.reason).join(" and ")}. To keep one gain, ${pass.blockers.map((item) => item.fix).join(" and ")} in production/audio/mix.json and rerun` : ""}. Listen to the mix before exporting.`);
   if (result.truePeak > -1) console.log("Warning: the encoded true peak is above -1 dBTP; lower loudness.tp in production/audio/mix.json and rerun.");
   if (Math.abs(result.integrated - loudness.i) > 1) console.log("Warning: the encoded loudness is more than 1 LU off target; check for long silences or a very dynamic bed.");
   console.log(`Wrote ${path.relative(base, output)}; film.config.json audio is "${audio}". Listen once with npm run dev before exporting.`);

package/templates/video/scripts/audio-tools.mjs

@@ -1,29 +1,20 @@
-/** Shared helpers for audio:vo, audio:music and audio:mix: the bundled ffmpeg, PCM decoding and small DSP routines. */
+/** Shared helpers for audio:vo, audio:music and audio:mix: the installed ffmpeg, PCM decoding and small DSP routines. */
 import { execFile, spawn } from "node:child_process";
 import { readFile, writeFile } from "node:fs/promises";
 import path from "node:path";
 import { promisify } from "node:util";
+import { ffmpegBinaries } from "./ffmpeg-binaries.mjs";
 
 const run = promisify(execFile);
-let binaries;
-
-/** The ffmpeg/ffprobe builds installed with the workspace; loaded lazily so the pure helpers import anywhere. */
-export async function audioBinaries() {
-  if (!binaries) {
-    const [{ default: ffmpeg }, { default: ffprobe }] = await Promise.all([import("@ffmpeg-installer/ffmpeg"), import("@ffprobe-installer/ffprobe")]);
-    binaries = { ffmpeg: ffmpeg.path, ffprobe: ffprobe.path };
-  }
-  return binaries;
-}
 
 export async function ffmpeg(args) {
-  const { ffmpeg: binary } = await audioBinaries();
+  const { ffmpeg: binary } = await ffmpegBinaries();
   try { return await run(binary, ["-hide_banner", "-nostdin", ...args], { maxBuffer: 64 * 1024 * 1024 }); }
   catch (error) { throw new Error(`ffmpeg failed: ${String(error.stderr ?? error.message).trim().split("\n").slice(-3).join(" ")}`); }
 }
 
 export async function mediaDuration(file) {
-  const { ffprobe } = await audioBinaries();
+  const { ffprobe } = await ffmpegBinaries();
   const { stdout } = await run(ffprobe, ["-v", "error", "-show_entries", "format=duration", "-of", "csv=p=0", file]);
   const duration = Number(stdout);
   if (!Number.isFinite(duration)) throw new Error(`Could not read the duration of ${file}.`);
@@ -32,7 +23,7 @@ export async function mediaDuration(file) {
 
 /** Decodes any audio file to mono Float32 samples at the given rate. */
 export async function decodePcm(file, sampleRate = 22050) {
-  const { ffmpeg: binary } = await audioBinaries();
+  const { ffmpeg: binary } = await ffmpegBinaries();
   return new Promise((resolve, reject) => {
     const child = spawn(binary, ["-v", "error", "-nostdin", "-i", file, "-ac", "1", "-ar", String(sampleRate), "-f", "f32le", "pipe:1"], { stdio: ["ignore", "pipe", "pipe"] });
     const chunks = []; let errors = "";
@@ -48,12 +39,17 @@ export async function decodePcm(file, sampleRate = 22050) {
   });
 }
 
-/** EBU R128 measurement with ffmpeg's loudnorm; returns the measured values its second pass needs. */
-export async function measureLoudness(file, { i = -14, tp = -2, lra = 9, filter = "" } = {}) {
-  const { stderr } = await ffmpeg(["-nostats", "-i", file, "-af", `${filter ? `${filter},` : ""}loudnorm=I=${i}:TP=${tp}:LRA=${lra}:print_format=json`, "-f", "null", "-"]);
+/** The report loudnorm prints at the info log level when it finishes with print_format=json, or null. */
+export function loudnormReport(stderr) {
   const json = stderr.match(/\{\s*"input_i"[\s\S]*?\}/)?.[0];
-  if (!json) throw new Error(`Loudness could not be measured for ${file}.`);
-  const values = JSON.parse(json);
+  return json ? JSON.parse(json) : null;
+}
+
+/** EBU R128 measurement with ffmpeg's loudnorm; returns the measured values its second pass needs. A video's picture is skipped. */
+export async function measureLoudness(file, { i = -14, tp = -2, lra = 9, filter = "" } = {}) {
+  const { stderr } = await ffmpeg(["-nostats", "-i", file, "-vn", "-af", `${filter ? `${filter},` : ""}loudnorm=I=${i}:TP=${tp}:LRA=${lra}:print_format=json`, "-f", "null", "-"]);
+  const values = loudnormReport(stderr);
+  if (!values) throw new Error(`Loudness could not be measured for ${file}.`);
   return { integrated: Number(values.input_i), truePeak: Number(values.input_tp), range: Number(values.input_lra), threshold: Number(values.input_thresh), offset: Number(values.target_offset) };
 }
 

package/templates/video/scripts/export-mp4.mjs

@@ -5,12 +5,13 @@ import { mkdir, access, rename, rm, stat, writeFile, readFile, realpath } from "
 import { createHash, randomUUID } from "node:crypto";
 import { fileURLToPath } from "node:url";
 import path from "node:path";
-import ffmpeg from "@ffmpeg-installer/ffmpeg";
-import ffprobe from "@ffprobe-installer/ffprobe";
+import { ffmpegBinaries } from "./ffmpeg-binaries.mjs";
 import { projectArguments, projectContext, publicFile } from "./project-context.mjs";
 import { workspacePlugin } from "./workspace-plugin.mjs";
 import { subtitleCues, toSrt, toVtt } from "./subtitles.mjs";
 import { captureHeadless, takeHeadlessFlag } from "./headless-browser.mjs";
+import { BT709_FRAMES, canCopySoundtrack, isBt709, mp4Language } from "./mp4-tools.mjs";
+import { measureLoudness } from "./audio-tools.mjs";
 
 // The cloud Animator makes no MP4s: the user exports the film in their own browser, from the draft page.
 if (process.env.ANIMAGURU_CLOUD_WORKSPACE) throw new Error("MP4 export is not available in the cloud Animator. The user exports the MP4 in their browser from the draft page.");
@@ -40,7 +41,9 @@ const { width, height, fps, duration, language } = config;
 if (![width, height].every((n) => Number.isInteger(n) && n >= 2 && n <= 4096 && n % 2 === 0) || ![24, 25, 30, 60].includes(fps) || !Number.isFinite(duration) || duration <= 0 || duration > 600 || Math.abs(duration * fps - Math.round(duration * fps)) > 0.000001) throw new Error("Use even dimensions up to 4096px, 24/25/30/60 fps, and a frame-aligned duration up to 600s in film.config.json.");
 if (typeof language !== "string" || !/^[a-z]{2,3}(?:-[A-Za-z]{2,4})?$/.test(language)) throw new Error("Set a language code in film.config.json.");
 if (typeof config.slug !== "string" || !/^[a-z0-9][a-z0-9-]{0,79}$/.test(config.slug)) throw new Error("Use a lowercase, hyphenated film slug (1–80 characters).");
-const languageTag = ({ en: "eng", tr: "tur", de: "deu", fr: "fra", es: "spa" })[language.split("-")[0]] ?? "und";
+const languageTag = mp4Language(language);
+// Before any capture starts: a missing or pre-7.1 FFmpeg stops here with how to install a current one.
+const binaries = await ffmpegBinaries();
 let audioSource;
 if (config.audio !== null) {
   if (typeof config.audio !== "string" || !config.audio) throw new Error("audio must be null or a local path relative to public/.");
@@ -89,6 +92,8 @@ let ownsPartial = false;
 let server;
 let lastActivity = Date.now();
 const run = promisify(execFile);
+// An AAC soundtrack that already fits the film, such as audio:mix's, is copied rather than encoded a second time.
+const copySoundtrack = audioSource ? canCopySoundtrack(JSON.parse((await run(binaries.ffprobe, ["-v", "error", "-select_streams", "a:0", "-show_entries", "stream=codec_name,profile,sample_rate,start_time,duration:format=duration", "-of", "json", audioSource])).stdout), duration) : false;
 
 async function readBody(request, limit = 24 * 1024 * 1024) {
   let bytes = 0;
@@ -121,13 +126,15 @@ function pngDimensions(bytes) {
   return { width: bytes.readUInt32BE(16), height: bytes.readUInt32BE(20) };
 }
 
-/** Reads an encoded MP4 back and checks that it is the whole film: frames, duration, dimensions, codecs and language. */
+/** Reads an encoded MP4 back and checks that it is the whole film: frames, duration, dimensions, codecs, colour tags and language. */
 async function verify(file, what) {
-  const { stdout } = await run(ffprobe.path, ["-v", "error", "-count_frames", "-show_streams", "-show_format", "-of", "json", file], { maxBuffer: 1024 * 1024 });
+  const { stdout } = await run(binaries.ffprobe, ["-v", "error", "-count_frames", "-show_streams", "-show_format", "-of", "json", file], { maxBuffer: 1024 * 1024 });
   const probe = JSON.parse(stdout);
   const video = probe.streams.find((stream) => stream.codec_type === "video");
   const audio = probe.streams.find((stream) => stream.codec_type === "audio");
   if (!video || video.codec_name !== "h264" || video.width !== width || video.height !== height || Number(video.nb_read_frames) !== settings.frames || video.pix_fmt !== "yuv420p" || !audio || audio.codec_name !== "aac" || audio.tags?.language !== languageTag || Math.abs(Number(probe.format.duration) - settings.duration) > 0.15) throw new Error(`The encoded ${what} did not pass duration, frame count, dimensions, language, or codec verification.`);
+  // Players guess the colours of a video without these tags, and some guess differently.
+  if (!isBt709(video)) throw new Error(`The encoded ${what} is not tagged BT.709 limited range (${["color_space", "color_primaries", "color_transfer", "color_range"].map((key) => video[key] ?? "unknown").join(", ")}).`);
   return probe;
 }
 
@@ -140,7 +147,7 @@ async function watermarkCopy() {
   await writeFile(watermarkedPartial, "", { flag: "wx" });
   ownsWatermarkedPartial = true;
   // The badge arrives on stdin as a single PNG frame, which the overlay keeps over the whole film.
-  const child = spawn(ffmpeg.path, [
+  const child = spawn(binaries.ffmpeg, [
     "-hide_banner", "-loglevel", "warning", "-y", "-i", partial, "-f", "png_pipe", "-i", "pipe:0",
     "-filter_complex", `[1:v]scale=out_color_matrix=bt709:out_range=tv,format=yuva420p[mark];[0:v][mark]overlay=x=${badge.x}:y=${badge.y}:format=yuv420,format=yuv420p[film]`,
     "-map", "[film]", "-map", "0:a:0",
@@ -201,15 +208,16 @@ async function route(request, response, url) {
     settings.captions = !subtitles.length ? "none" : burnCaptions ? "burned" : "subtitle-files";
     await writeFile(partial, "", { flag: "wx" });
     ownsPartial = true;
-    encoder = spawn(ffmpeg.path, [
+    encoder = spawn(binaries.ffmpeg, [
       "-hide_banner", "-loglevel", "warning", "-y",
       "-f", "image2pipe", "-framerate", String(fps), "-vcodec", "mjpeg", "-i", "pipe:0",
       ...(audioSource ? ["-i", audioSource] : ["-f", "lavfi", "-i", "anullsrc=r=48000:cl=stereo"]),
       "-map", "0:v:0", "-map", "1:a:0",
       "-c:v", "libx264", "-preset", "medium", "-crf", "18",
-      "-vf", `scale=${width}:${height}:in_color_matrix=bt601:out_color_matrix=bt709:in_range=full:out_range=tv,format=yuv420p`,
+      // setparams labels the frames: from FFmpeg 7.1 on, the -color_* options below no longer tag frames that do not say so.
+      "-vf", `scale=${width}:${height}:in_color_matrix=bt601:out_color_matrix=bt709:in_range=full:out_range=tv,format=yuv420p,${BT709_FRAMES}`,
       "-color_primaries", "bt709", "-color_trc", "bt709", "-colorspace", "bt709",
-      "-af", "apad", "-c:a", "aac", "-b:a", "192k", "-ar", "48000",
+      ...(copySoundtrack ? ["-c:a", "copy"] : ["-af", "apad", "-c:a", "aac", "-b:a", "192k", "-ar", "48000"]),
       "-metadata:s:a:0", `language=${languageTag}`, "-metadata:s:v:0", `language=${languageTag}`,
       "-t", String(settings.duration), "-movflags", "+faststart", partial,
     ], { stdio: ["pipe", "ignore", "pipe"] });
@@ -275,6 +283,10 @@ async function route(request, response, url) {
     await encoderDone;
     console.log("Checking the encoded video and soundtrack…");
     const probe = await verify(partial, "file");
+    // Players hear the MP4's own sound, so its level is measured from the MP4, not assumed from the soundtrack file.
+    const sound = audioSource ? await measureLoudness(partial).then(
+      (levels) => ({ file: config.audio, copied: copySoundtrack, integratedLufs: levels.integrated, truePeakDb: levels.truePeak, rangeLu: levels.range }),
+      (error) => ({ file: config.audio, copied: copySoundtrack, error: error.message })) : null;
     if (watermark) {
       console.log("Making the watermarked copy…");
       await watermarkCopy();
@@ -287,11 +299,13 @@ async function route(request, response, url) {
     for (const file of subtitleFiles) await writeFile(file.path, file.text);
     const thumbnailFile = thumbnailImage ? destination.replace(/\.mp4$/, ".jpg") : null;
     if (thumbnailFile) await writeFile(thumbnailFile, thumbnailImage);
-    await writeFile(destination.replace(/\.mp4$/, ".json"), JSON.stringify({ file: path.basename(destination), title: config.title, audio: config.audio, subtitles: subtitleFiles.map((file) => path.basename(file.path)), thumbnail: thumbnailFile && { file: path.basename(thumbnailFile), time: thumbnail.time }, ...(watermark ? { watermarked: { file: path.basename(watermarked), badge: { x: badge.x, y: badge.y, width: badge.width, height: badge.height } } } : {}), settings, verification: probe }, null, 2) + "\n");
+    await writeFile(destination.replace(/\.mp4$/, ".json"), JSON.stringify({ file: path.basename(destination), title: config.title, audio: config.audio, subtitles: subtitleFiles.map((file) => path.basename(file.path)), thumbnail: thumbnailFile && { file: path.basename(thumbnailFile), time: thumbnail.time }, soundtrack: sound, ...(watermark ? { watermarked: { file: path.basename(watermarked), badge: { x: badge.x, y: badge.y, width: badge.width, height: badge.height } } } : {}), settings, verification: probe }, null, 2) + "\n");
     const bytes = (await stat(destination)).size;
     finished = true;
     clearInterval(idleCheck);
     console.log(`Saved: ${destination}\nSize: ${(bytes / 1024 / 1024).toFixed(1)} MiB\nVerified: ${settings.frames} frames, ${settings.duration}s, H.264 video and AAC audio.`);
+    if (sound) console.log(`Sound: ${sound.file}, ${sound.copied ? "copied without a second AAC encode" : "encoded to AAC"}; ${sound.error ? `not measured (${sound.error})` : Number.isFinite(sound.integratedLufs) ? `the MP4 measures ${sound.integratedLufs.toFixed(1)} LUFS, ${sound.truePeakDb.toFixed(1)} dBTP` : "the MP4's sound is silent"}.`);
+    if (sound?.truePeakDb > -1) console.log("Warning: the MP4's sound peaks above -1 dBTP, where players and platforms can clip it. Lower the soundtrack's true peak (for npm run audio:mix, loudness.tp in production/audio/mix.json) and export again.");
     if (subtitleFiles.length) console.log(`Subtitles (${subtitles.length} captions, ${language}): ${subtitleFiles.map((file) => file.path).join(", ")}`);
     console.log(thumbnailFile ? `Thumbnail (${thumbnail.time}s): ${thumbnailFile}` : 'No thumbnail: set "poster" in film.config.json to the seconds of the frame to save as one.');
     if (watermark) console.log(`Watermarked copy ("Made with AnimaGuru"): ${watermarked}`);

package/templates/video/scripts/ffmpeg-binaries.mjs

@@ -0,0 +1,54 @@
+/**
+ * The FFmpeg and FFprobe the audio tools and MP4 export run: the ones on PATH, or FFMPEG_PATH and FFPROBE_PATH.
+ * The workspace installs none of its own. Both must be FFmpeg 7.1 or newer, which every script here is tested with.
+ */
+import { execFile } from "node:child_process";
+import { promisify } from "node:util";
+
+const run = promisify(execFile);
+export const MINIMUM_FFMPEG = [7, 1];
+
+/** The release in the first line of `ffmpeg -version` (9.0.2, 7.1.5-0+deb13u1, n8.0.1, 8.0-essentials_build…), or null for a git snapshot. */
+export function ffmpegRelease(text) {
+  const match = /^ff(?:mpeg|probe) version n?(\d+)\.(\d+)(?:\.(\d+))?/.exec(String(text).trim());
+  return match ? match.slice(1).map((part) => Number(part ?? 0)) : null;
+}
+
+/** How to install a current FFmpeg on this platform. */
+export function ffmpegInstallHint(platform = process.platform) {
+  const install = platform === "darwin" ? "Install it with `brew install ffmpeg`"
+    : platform === "win32" ? "Install it with `winget install Gyan.FFmpeg`, then open a new terminal"
+    : "Install your distribution's ffmpeg package if it is 7.1 or newer (Debian 13: `sudo apt install ffmpeg`), or a static build";
+  return `${install}, or point FFMPEG_PATH and FFPROBE_PATH at an ffmpeg and ffprobe of ${MINIMUM_FFMPEG.join(".")} or newer.`;
+}
+
+const older = (release, minimum) => release[0] !== minimum[0] ? release[0] < minimum[0] : release[1] < minimum[1];
+
+/** Runs `NAME -version` and throws, with how to install a current build, when it is missing or older than 7.1. */
+async function check(binary, name) {
+  let stdout;
+  try { ({ stdout } = await run(binary, ["-version"])); }
+  catch (error) {
+    if (error.code === "ENOENT") throw new Error(`${name} was not found ${binary === name ? "on PATH" : `at ${binary}`}. The audio tools and MP4 export need FFmpeg ${MINIMUM_FFMPEG.join(".")} or newer. ${ffmpegInstallHint()}`);
+    throw new Error(`${binary} -version failed: ${String(error.stderr || error.message).trim().split("\n").at(-1)}. ${ffmpegInstallHint()}`);
+  }
+  const release = ffmpegRelease(stdout);
+  if (release && older(release, MINIMUM_FFMPEG)) {
+    const version = String(stdout).trim().split("\n")[0].split(/\s+/)[2];
+    throw new Error(`${name} ${version} at ${binary} is too old: the audio tools and MP4 export need FFmpeg ${MINIMUM_FFMPEG.join(".")} or newer. ${ffmpegInstallHint()}`);
+  }
+}
+
+/** Finds and checks ffmpeg and ffprobe for the given environment; resolves to the commands to run. */
+export async function findFfmpeg(env = process.env) {
+  const ffmpeg = env.FFMPEG_PATH || "ffmpeg", ffprobe = env.FFPROBE_PATH || "ffprobe";
+  await Promise.all([check(ffmpeg, "ffmpeg"), check(ffprobe, "ffprobe")]);
+  return { ffmpeg, ffprobe };
+}
+
+let binaries;
+/** findFfmpeg() for this process, checked once. */
+export function ffmpegBinaries() {
+  binaries ??= findFfmpeg();
+  return binaries;
+}

package/templates/video/scripts/mp4-tools.mjs

@@ -0,0 +1,33 @@
+/** Decisions npm run export:mp4 makes about the MP4's streams, apart from its capture server so they can be tested. */
+
+// MP4 language tags are ISO 639-2 codes: FFmpeg writes a two-letter code such as "sv" as no tag at all.
+const LANGUAGE_TAGS = { en: "eng", tr: "tur", sv: "swe", de: "deu", fr: "fra", es: "spa" };
+
+/** The MP4 stream language tag for a film.config.json language such as "sv" or "en-GB"; "und" for one not listed. */
+export function mp4Language(language) {
+  return LANGUAGE_TAGS[language.split("-")[0]] ?? "und";
+}
+
+/**
+ * Labels the film's frames BT.709 limited range at the end of its filter chain. FFmpeg 7.1 and newer no longer write
+ * -color_primaries and -color_trc as plain tags: frames decoded from the capture's JPEGs carry no primaries or transfer,
+ * and without this label the MP4 says "unknown" for both (FFmpeg 9.0.2), which players fill in by guessing.
+ */
+export const BT709_FRAMES = "setparams=color_primaries=bt709:color_trc=bt709:colorspace=bt709:range=tv";
+
+/** Whether ffprobe's video stream is tagged BT.709 limited range, as the film is encoded. */
+export function isBt709(stream) {
+  return stream?.color_space === "bt709" && stream.color_primaries === "bt709" && stream.color_transfer === "bt709" && stream.color_range === "tv";
+}
+
+/**
+ * Whether the soundtrack goes into the MP4 as it is, from ffprobe's JSON for its first audio stream: AAC-LC at 48 kHz
+ * that starts at 0 and lasts the whole film, as npm run audio:mix writes it. Copying keeps the loudness and true peak
+ * that audio:mix measured; a second AAC encode adds its own overshoot. Anything else is encoded and padded to length.
+ */
+export function canCopySoundtrack(probe, duration) {
+  const stream = probe?.streams?.[0];
+  const length = Number(stream?.duration ?? probe?.format?.duration);
+  return stream?.codec_name === "aac" && stream.profile === "LC" && Number(stream.sample_rate) === 48000
+    && Math.abs(Number(stream.start_time ?? Number.NaN)) < 0.001 && length >= duration - 0.05;
+}

package/templates/video/scripts/review-audio.mjs

@@ -1,19 +1,19 @@
 import path from "node:path";
 import { execFile } from "node:child_process";
 import { promisify } from "node:util";
-import ffmpeg from "@ffmpeg-installer/ffmpeg";
-import ffprobe from "@ffprobe-installer/ffprobe";
+import { ffmpegBinaries } from "./ffmpeg-binaries.mjs";
 
 const run=promisify(execFile);
 export async function inspectAudio(root,config) {
   if(!config.audio)return {status:"silent",message:"No soundtrack configured; audio checks are not applicable."};
   const publicRoot=path.resolve(root,"public"),file=path.resolve(publicRoot,config.audio);
   if(!file.startsWith(publicRoot+path.sep))throw new Error("Soundtrack must be inside public/.");
-  const {stdout}=await run(ffprobe.path,["-v","error","-show_streams","-show_format","-of","json",file]);
+  const {ffmpeg,ffprobe}=await ffmpegBinaries();
+  const {stdout}=await run(ffprobe,["-v","error","-show_streams","-show_format","-of","json",file]);
   const probe=JSON.parse(stdout),stream=probe.streams.find((s)=>s.codec_type==="audio");
   if(!stream)throw new Error("Configured soundtrack has no audio stream.");
   // Same targets as npm run audio:mix (−14 LUFS, −2 dBTP); the measured input values do not depend on them.
-  const {stderr}=await run(ffmpeg.path,["-hide_banner","-nostats","-i",file,"-af","loudnorm=I=-14:TP=-2:LRA=9:print_format=json","-f","null","-"],{maxBuffer:1024*1024});
+  const {stderr}=await run(ffmpeg,["-hide_banner","-nostats","-i",file,"-af","loudnorm=I=-14:TP=-2:LRA=9:print_format=json","-f","null","-"],{maxBuffer:1024*1024});
   const json=stderr.match(/\{\s*"input_i"[\s\S]*?\}/)?.[0];
   if(!json)throw new Error("Audio loudness could not be measured.");
   const levels=JSON.parse(json),duration=Number(probe.format.duration);