@koda-sl/baker-cli 0.295.0-dev.f941e60fb → 0.297.0-dev.306c0952c

package/dist/cli.js

@@ -54,6 +54,7 @@ import {
   sha256Hex,
   spineInputFlags,
   spineInputOps,
+  splitScriptFor,
   spokenWordsIn,
   supportsLastFrame,
   supportsParam,
@@ -62,7 +63,7 @@ import {
   ulid,
   validateCanvasDeep,
   ytDlpBlockSignal
-} from "./chunk-ER3KCAW3.js";
+} from "./chunk-YMHERSTK.js";
 import {
   csvOrJson,
   daysAgoIso,
@@ -38222,6 +38223,15 @@ var scaffoldStaticAdCommand = defineCommand110({
   }
 });
 
+// src/commands/studio/avatarModel.ts
+function pinnedModelHint(input) {
+  if (!input.avatar || !input.model) return null;
+  if (input.model === DEFAULT_VIDEO_GENERATE_MODEL) return null;
+  const speaks = input.generateAudio ?? true;
+  if (!speaks) return null;
+  return `You pinned --model ${input.model} on a speaking avatar. Without it this clip routes to ${DEFAULT_VIDEO_GENERATE_MODEL}, which was chosen for exactly this case: the better picture, both orientations, and audio always on. Drop --model unless the user asked for a specific one \u2014 and note ${input.model} may not take the length or the reference you are counting on.`;
+}
+
 // src/commands/canvas/scaffold-ad.ts
 import { copyFile, cp, mkdir as mkdir7, readFile as readFile16, stat as stat4, writeFile as writeFile9 } from "fs/promises";
 import path24 from "path";
@@ -38953,6 +38963,10 @@ function resolveShippedCanvasDir(name, startDir, exists2 = existsSync5, maxDepth
 }
 
 // src/commands/canvas/scaffold-ad.ts
+function pinnedOnAvatar(handle, model) {
+  const hint = pinnedModelHint({ avatar: handle ?? void 0, model, generateAudio: true });
+  return hint ? [hint] : [];
+}
 var AD_IMAGE_MODEL = "openai/gpt-image-2.5-sunburst";
 var SHIPPED_COMPOSITION_DIR = resolveShippedCanvasDir("video-overlay-composition", import.meta.dirname);
 var SHIPPED_CAPTIONS_DIR = resolveShippedCanvasDir("tiktok-captions-composition", import.meta.dirname);
@@ -39275,6 +39289,10 @@ var scaffoldAdCommand = defineCommand111({
       ok: true,
       data: { canvas: outPath, beats: spec.data.beats.length, slug },
       hints: [
+        // Same reasoning as `studio animate`: a cast avatar that speaks routes to the
+        // default model on purpose, and pinning one discards that silently. An ad's
+        // beats always speak, so there is no silent case to exclude here.
+        ...pinnedOnAvatar(handle, requestedModel),
         // `--until frames` on anything carrying a face or a mark, because that is where
         // getting it wrong is expensive: the pictures cost a fraction of the clips, and
         // every defect anyone has caught in a finished ad was visible in them. The run
@@ -57381,15 +57399,6 @@ import { defineCommand as defineCommand208 } from "citty";
 // src/commands/studio/animate.ts
 import { defineCommand as defineCommand200 } from "citty";
 
-// src/commands/studio/avatarModel.ts
-function pinnedModelHint(input) {
-  if (!input.avatar || !input.model) return null;
-  if (input.model === DEFAULT_VIDEO_GENERATE_MODEL) return null;
-  const speaks = input.generateAudio ?? true;
-  if (!speaks) return null;
-  return `You pinned --model ${input.model} on a speaking avatar. Without it this clip routes to ${DEFAULT_VIDEO_GENERATE_MODEL}, which was chosen for exactly this case: the better picture, both orientations, and audio always on. Drop --model unless the user asked for a specific one \u2014 and note ${input.model} may not take the length or the reference you are counting on.`;
-}
-
 // src/commands/studio/batch.ts
 function projectBatch(generation, full) {
   const compact3 = {
@@ -57610,13 +57619,18 @@ var spokenWords = spokenWordsIn;
 function wordsThatFit(seconds) {
   return Math.floor(seconds * WORDS_PER_SECOND2);
 }
-function spokenPaceHint(prompt, seconds) {
-  if (!seconds) return null;
+function spokenPaceHint(prompt, seconds, model) {
   const words2 = spokenWords(prompt);
+  if (words2 === 0) return null;
+  const split = model ? splitScriptFor(words2, model) : null;
+  if (split) {
+    return `The line is ${words2} words \u2014 about ${Math.ceil(words2 / WORDS_PER_SECOND2)}s of speech \u2014 and ${model} renders at most ${split.maxSeconds}s (~${Math.floor(split.maxSeconds * WORDS_PER_SECOND2)} words). It will not refuse and it will not cut: it will RUSH, swallowing syllables. Split the script into ${split.clips} takes of roughly ${split.wordsPerClip} words each, rendered separately and cut together \u2014 break it where the thought breaks, not mid-sentence. Or move to a model that renders longer (bytedance/seedance-2.5 goes to 30s).`;
+  }
+  if (!seconds) return null;
   const fits = wordsThatFit(seconds);
-  if (words2 === 0 || words2 <= fits) return null;
+  if (words2 <= fits) return null;
   const rate = (words2 / seconds).toFixed(1);
-  return `The line is ${words2} words and this clip is ${seconds}s \u2014 that is ${rate} words a second against a natural 2.5. The model will not run over, it will RUSH: syllables swallowed, endings clipped, and the take unusable however good the picture. Cut the line to about ${fits} words, or give it the length it needs (${Math.ceil(words2 / 2.5)}s). Do not re-render the same line at the same length and hope.`;
+  return `The line is ${words2} words and this clip is ${seconds}s \u2014 that is ${rate} words a second against a natural ${WORDS_PER_SECOND2}. The model will not run over, it will RUSH: syllables swallowed, endings clipped, and the take unusable however good the picture. Cut the line to about ${fits} words, or give it the length it needs (${Math.ceil(words2 / WORDS_PER_SECOND2)}s). Do not re-render the same line at the same length and hope.`;
 }
 
 // src/commands/studio/animate.ts
@@ -57849,7 +57863,7 @@ function costHintsFor(body) {
   ];
   const pinned = pinnedModelHint({ avatar: body.avatar, model: body.model, generateAudio: body.generateAudio });
   if (pinned) hints2.push(pinned);
-  const pace = spokenPaceHint(body.prompt, body.duration);
+  const pace = spokenPaceHint(body.prompt, body.duration, model);
   if (pace) hints2.push(pace);
   if (count > 1) {
     hints2.push(