@koda-sl/baker-cli 0.276.0-dev.1f1c09c80 → 0.277.0-dev.1f1c09c80

package/dist/cli.js

@@ -58,7 +58,7 @@ import {
   ulid,
   validateCanvasDeep,
   ytDlpBlockSignal
-} from "./chunk-4IJX4R4J.js";
+} from "./chunk-AUOSDGWY.js";
 import {
   csvOrJson,
   daysAgoIso,
@@ -3699,14 +3699,22 @@ var avatarSummarySchema = z7.object({
   subjectDescription: z7.string(),
   /** The identity sheet to wire into `--reference`. Absent until the build settles. */
   sheetUrl: z7.string().optional(),
+  /**
+   * The avatar's pinned voice.
+   *
+   * A casting essential, not a profile detail: an avatar is a face AND a voice, and
+   * casting the face without it is what makes a presenter speak in somebody else's.
+   * It lived only behind `--full`, so an ad that grounded every frame on the identity
+   * sheet still had its lines read by a voice cast from a prose description.
+   */
+  voiceId: z7.string().optional(),
+  voiceDescription: z7.string().optional(),
   coverUrl: z7.string().optional(),
   errorMessage: z7.string().optional()
 });
 var avatarDetailSchema = avatarSummarySchema.extend({
   profile: avatarProfileSchema,
   sourceUrls: z7.array(z7.string()),
-  voiceId: z7.string().optional(),
-  voiceDescription: z7.string().optional(),
   createdAt: z7.number()
 });
 var avatarsListRequestSchema = z7.object({
@@ -28705,7 +28713,7 @@ function buildPhrases(blueprint, canonical2, compositeScenes, presenterPresent,
   flush();
   return phrases;
 }
-function makeVoiceFactory(blueprint, canonical2, nodes, voiceLanguage) {
+function makeVoiceFactory(blueprint, canonical2, nodes, voiceLanguage, pinnedVoiceId) {
   const bySpeaker = /* @__PURE__ */ new Map();
   const describe = (speaker) => {
     for (const scene of blueprint.scenes)
@@ -28721,7 +28729,9 @@ function makeVoiceFactory(blueprint, canonical2, nodes, voiceLanguage) {
     const description = describe(speaker);
     const traits = parseVoiceTraits(description);
     if (voiceLanguage) traits.language = voiceLanguage;
-    nodes.push({ id, type: "voice_select", params: { description, ...traits } });
+    nodes.push(
+      pinnedVoiceId ? { id, type: "voice_select", params: { description, voice_id: pinnedVoiceId, ...traits } } : { id, type: "voice_select", params: { description, ...traits } }
+    );
     bySpeaker.set(speaker, id);
     return id;
   };
@@ -29122,7 +29132,7 @@ function buildTimeline(blueprint, slots, opts, nodes) {
     });
   }
   const canonical2 = collapseVoiceover(blueprint);
-  const ensureVoiceNode = makeVoiceFactory(blueprint, canonical2, nodes, opts.voiceLanguage);
+  const ensureVoiceNode = makeVoiceFactory(blueprint, canonical2, nodes, opts.voiceLanguage, opts.voiceId);
   const aspect = resolveAspect(blueprint.source?.aspect_ratio, opts.aspect, genAspectsFor(opts.videoModel));
   const env = {
     blueprint,
@@ -32775,7 +32785,8 @@ function adSpecCastElements(spec, avatar) {
 function motionPrompt(beat, isClosingCard, place, physics) {
   if (isClosingCard) return BRAND_PLATE;
   const shot = `${beat.show}${place}${physics}`;
-  return beat.on_camera ? shot : `${shot} Nobody in frame is speaking \u2014 mouths closed, no dialogue.`;
+  if (beat.on_camera) return shot;
+  return `${shot} Nobody in frame is speaking: mouths CLOSED and STILL throughout, no talking, no lip movement. Whoever is in shot is looking, reacting or working \u2014 never addressing the camera.`;
 }
 function staticTranscript(spec) {
   const out = [];
@@ -33219,6 +33230,10 @@ var scaffoldAdCommand = defineCommand104({
       // Told, not guessed: the voice description is written in the ad's own language
       // and the engine's trait parser reads English only.
       voiceLanguage: voiceLanguageFor(spec.data),
+      // The cast avatar's own voice. An avatar is a face AND a voice; grounding every
+      // frame on their identity sheet while a separately cast voice reads the lines is
+      // how a presenter ends up speaking in somebody else's.
+      ...avatar?.voiceId ? { voiceId: avatar.voiceId } : {},
       ...spec.data.voiceover === false ? { staticTranscriptPath: path24.relative(outDir, transcriptPath) } : {}
     };
     const canvas = scaffoldVideoCanvas(blueprint, adSpecCastElements(spec.data, avatar), opts);
@@ -33272,6 +33287,9 @@ var scaffoldAdCommand = defineCommand104({
         ] : [],
         ...avatarError ? [avatarError] : [],
         ...avatar ? [
+          ...avatar?.voiceId ? [] : [
+            `Avatar \`${handle}\` has no pinned voice, so the ad's lines are read by a voice cast from a description \u2014 it will not sound like them. Give them one (\`baker avatars\` voice) before this ships.`
+          ],
           `Every beat the customer appears in is grounded on \`${handle}\`'s identity sheet, and their subject description was copied into the frames verbatim \u2014 so it is the same face throughout and the same face as the rest of this company's work.`
         ] : [],
         ...presenterIsInEveryShot(spec.data) ? [