@koda-sl/baker-cli 0.272.1-dev.1f1c09c80 → 0.275.0-dev.1f1c09c80

package/README.md

@@ -3648,6 +3648,9 @@ because a video model garbles a wordmark every time:
   the ad needs something other than the client's default.
 - **The spoken language is read off the script** when the spec does not declare it, so a
   Spanish ad is cast from Spanish voices without anyone having to say so.
+- **`voiceover: false`** — a music-led ad. No voice, no transcription: the `say` lines
+  become on-screen text and are captioned straight from the script, so the words are
+  exact. Give it `music` too.
 - **`--avatar <handle>`** — the same flag as on `studio generate` / `studio animate`, and the way to cast a person into an ad. Overrides `cast.avatar` in the spec.
 - **`cast.avatar`** — the handle of a cast Avatar (`baker avatars list`). The ad grounds
   every beat they appear in on that avatar's identity sheet and copies its subject

package/dist/cli.js

@@ -29566,9 +29566,32 @@ function emitBlueprintIngests(opts, nodes) {
     params: { source: "path", path: opts.blueprintStylePath ?? "./prompt.style.json", expect: "json" }
   });
 }
+function emitStaticCaptions(opts, videoRef, nodes) {
+  if (!opts.staticTranscriptPath) return { videoRef };
+  const wordsPerGroup = opts.captionWordsPerGroup ? { words_per_group: opts.captionWordsPerGroup } : {};
+  nodes.push({
+    id: "captions_transcript",
+    type: "ingest",
+    params: { source: "path", path: opts.staticTranscriptPath, expect: "json" }
+  });
+  const overlaid = nodes.find((n) => n.id === "overlaid");
+  if (overlaid) {
+    overlaid.inputs = { ...overlaid.inputs, transcript: "$ref:captions_transcript.json" };
+    overlaid.params = { ...overlaid.params, ...wordsPerGroup };
+    return { videoRef };
+  }
+  nodes.push({
+    id: "captions",
+    type: "hyperframe_render",
+    inputs: { background: videoRef, transcript: "$ref:captions_transcript.json" },
+    params: { composition: opts.captionsCompositionPath, ...wordsPerGroup }
+  });
+  return { videoRef: "$ref:captions.video", videoNode: "captions" };
+}
 function emitCaptionChain(blueprint, tracks, clock, opts, videoRef, nodes) {
   const voTrackList = tracks.filter((t) => t.kind === "vo");
-  if (!opts.captionsCompositionPath || voTrackList.length === 0) return { videoRef };
+  if (!opts.captionsCompositionPath) return { videoRef };
+  if (voTrackList.length === 0) return emitStaticCaptions(opts, videoRef, nodes);
   const voiceMixInputs = {};
   for (const t of voTrackList) voiceMixInputs[t.slot] = t.ref;
   nodes.push({
@@ -30333,7 +30356,8 @@ function buildRunRecord(result, meta, plan, finalLabels) {
       durationMs: result.stats.duration_ms
     },
     nodes,
-    finalOutputs: finalOutputs.length > 0 ? finalOutputs : void 0
+    finalOutputs: finalOutputs.length > 0 ? finalOutputs : void 0,
+    ...meta.review && meta.review.length > 0 ? { review: meta.review.slice(0, 40) } : {}
   };
 }
 function buildInitialRunRecord(runId, meta) {
@@ -31394,7 +31418,6 @@ ${describeRewrites(healed.rewrites)}
       } : void 0
     });
     await clearRunMarker(outputsDir, filePath);
-    if (poster) await poster.flush(buildRunRecord(result, recordMeta, progress?.planInfo(), finalLabels));
     const keepRuns = opts.keepRuns;
     if (keepRuns !== void 0 && Number.isFinite(keepRuns)) {
       await pruneOldRuns(outputsDir, keepRuns, result.run_id, (line) => process.stdout.write(`${line}
@@ -31409,6 +31432,11 @@ ${describeRewrites(healed.rewrites)}
         `Fix them, then re-run: edit the \`show\` of the scene files named above \u2014 a shot with a person handling equipment is where limbs and objects go wrong, so widen it, move the hands out of frame, or drop the person from that beat. Re-running re-renders ONLY the scenes you edited; everything else comes from the cache (this run: ${result.stats?.cached_nodes ?? 0} of ${result.stats?.total_nodes ?? 0} nodes were cached).`
       ] : []
     ] : [];
+    if (poster) {
+      await poster.flush(
+        buildRunRecord(result, { ...recordMeta, ...review ? { review } : {} }, progress?.planInfo(), finalLabels)
+      );
+    }
     process.stdout.write(
       `${JSON.stringify(
         {
@@ -32490,12 +32518,16 @@ import { defineCommand as defineCommand104 } from "citty";
 // src/engine/scaffold/ad-spec.ts
 import { z as z30 } from "zod";
 var oneClause = (line) => (line.match(/[.!?](\s|$)/g) ?? []).length <= 1;
+var isSpeech = (line) => new RegExp("\\p{L}\\p{L}", "u").test(line);
 var Beat = z30.object({
   /**
    * The line the voice says in this beat — ONE clause, ending in its own
    * punctuation. It becomes the caption card verbatim.
    */
-  say: z30.string().min(1).refine(oneClause, "a beat is one clause: split a second sentence into its own beat"),
+  say: z30.string().min(1).refine(
+    isSpeech,
+    "a beat's line has to be words someone says \u2014 a dot or a dash is not a line, and an ad with no voiceover is not supported yet"
+  ).refine(oneClause, "a beat is one clause: split a second sentence into its own beat"),
   /** What is on screen while it is said, as a shot brief for the video model. */
   show: z30.string().min(1),
   /**
@@ -32597,6 +32629,24 @@ var AdSpec = z30.object({
       cta: z30.string().min(1).optional()
     })
   ]).optional(),
+  /**
+   * Whether a voice reads the lines.
+   *
+   * On by default. Turned off, the ad carries its music and its pictures and the `say`
+   * lines become ON-SCREEN text instead of spoken ones — which is how most music-led
+   * social ads are actually built.
+   *
+   * It exists because an agent wanted exactly this and had no way to ask: it wrote "·"
+   * as every line with a voice description reading "No voiceover; music only". The dot
+   * was sent to the voice model, the near-silence that came back was transcribed, and
+   * the transcriber hallucinated "Thank you." — which became the ad's only caption card,
+   * held over all sixteen seconds, in English, in a Spanish ad.
+   *
+   * The captions are BETTER on this path, not worse: they come from the script itself
+   * rather than from transcribing a rendered read, so the text is exact and no speech
+   * model is between the copy and the screen.
+   */
+  voiceover: z30.boolean().default(true),
   beats: z30.array(Beat).min(2)
 });
 function presenterIsInEveryShot(spec) {
@@ -32687,6 +32737,10 @@ function beatSeconds(line) {
   return allowed.find((d) => d >= spoken) ?? allowed[allowed.length - 1];
 }
 var BRAND_PLATE = "A solid flat brand colour plate \u2014 an end card. No people, no objects, no scenery, no text of any kind.";
+var ASKS_FOR_LEGIBLE = /\b(focus on the (number|amount|total|price|figure|text)|readable|legible|clearly (shows?|visible)|highlighted (number|amount|total)|showing the (number|amount|total|price))\b/i;
+function briefWantsLegibleText(show) {
+  return ASKS_FOR_LEGIBLE.test(show);
+}
 function endCardWanted(spec) {
   if (spec.end_card === false) return false;
   return Boolean(spec.brand?.name || spec.brand?.logo);
@@ -32718,6 +32772,30 @@ 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 : `${shot} Nobody in frame is speaking \u2014 mouths closed, no dialogue.`;
+}
+function staticTranscript(spec) {
+  const out = [];
+  let clock = 0;
+  for (const beat of spec.beats) {
+    const duration = beatSeconds(beat.say);
+    const words2 = beat.say.trim().split(/\s+/).filter(Boolean);
+    const each = words2.length > 0 ? duration / words2.length : duration;
+    words2.forEach((text2, i) => {
+      const start = clock + i * each;
+      out.push({
+        text: text2,
+        start: Math.round(start * 100) / 100,
+        end: Math.round((start + each) * 100) / 100
+      });
+    });
+    clock += duration;
+  }
+  return out;
+}
 function adSpecToBlueprint(spec) {
   const total = spec.beats.length;
   const market = marketFor(spec);
@@ -32742,7 +32820,10 @@ function adSpecToBlueprint(spec) {
       // `on_camera` is the engine's own switch for "this line is narration" —
       // absent, it guesses from who is in frame, which is how a still photograph
       // once got lip-synced.
-      dialogue: [
+      // No dialogue at all when nobody speaks: an empty array is what stops the engine
+      // wiring a voice track, and no voice track is what makes this a music-led ad
+      // rather than one with a silent narrator.
+      dialogue: spec.voiceover === false ? [] : [
         {
           line: beat.say,
           start_s: Math.round(start * 100) / 100,
@@ -32754,11 +32835,7 @@ function adSpecToBlueprint(spec) {
           ...spec.voice?.description ? { voice_description: spec.voice.description } : {}
         }
       ],
-      motion_prompt: isClosingCard ? BRAND_PLATE : beat.on_camera ? `${beat.show}${place}${physics}` : (
-        // The subject must not be forming words: the voice belongs to the
-        // narration, and a mouth moving under it reads as dubbing.
-        `${beat.show}${place}${physics} Nobody in frame is speaking \u2014 mouths closed, no dialogue.`
-      ),
+      motion_prompt: motionPrompt(beat, isClosingCard, place, physics),
       start_frame_prompt: isClosingCard ? BRAND_PLATE : `${beat.show}${place}${physics}`
     };
   });
@@ -32768,8 +32845,9 @@ function adSpecToBlueprint(spec) {
       transcript: spec.beats.map((b) => b.say).join(" "),
       voiceover: {
         // Narration over the picture. Never `on_camera`: that is what tells the
-        // engine a mouth is a lip-sync candidate.
-        mode: "voiceover",
+        // engine a mouth is a lip-sync candidate. A music-led ad says `none`, which is
+        // what keeps a voice out of it entirely.
+        mode: spec.voiceover === false ? "none" : "voiceover",
         ...spec.voice?.description ? { voice_description: spec.voice.description } : {}
       },
       ...spec.brand?.name || spec.brand?.logo ? {
@@ -32983,7 +33061,7 @@ registerSchema({
     spec: {
       type: "string",
       required: true,
-      description: 'Path to the ad spec JSON. Shape: { format?, market?, brand?, cast?, end_card?, voice?, music?, beats: [{ say, show, on_camera?, cast? }] }. Fill `brand` from src/brand/BRAND.md \u2014 `palette` (its hex tokens, most important first) and `logo` (the repo path to the mark) are what make the ad look like the client rather than like stock. `market` is where the ad is SET; omitted, it is inferred from the voice language, and getting it wrong is what fills a Spanish ad with British houses. `say` is ONE clause ending in its own punctuation \u2014 it becomes a caption card verbatim, so two sentences in one beat produce a card holding both. `show` is the shot brief for that line. Set `on_camera` ONLY when that beat\'s subject talks to camera. `cast` is WHO the ad is about, and an ad with people in it needs one: `{ "avatar": "marta" }` names a cast avatar (`baker avatars list`) and every beat is grounded on that avatar\'s identity sheet with its subject description copied verbatim \u2014 the same face here as in the rest of the company\'s work. Without it each beat invents its own stranger, which is how one 28-second ad came back with five different men playing one customer. `{ "description": "..." }` is the fallback when there is no avatar, and it scaffolds a canvas that asks you to drop a photo before it can run \u2014 so prefer the avatar. A beat sets `cast: false` for a shot they are not in. `end_card` is on by default whenever `brand` is set: the last beat becomes a flat brand plate with the mark and a call to action drawn over it. `{ "cta": "..." }` sets the button copy, `false` keeps the footage.'
+      description: 'Path to the ad spec JSON. Shape: { format?, market?, brand?, cast?, end_card?, voice?, music?, beats: [{ say, show, on_camera?, cast? }] }. Fill `brand` from src/brand/BRAND.md \u2014 `palette` (its hex tokens, most important first) and `logo` (the repo path to the mark) are what make the ad look like the client rather than like stock. `market` is where the ad is SET; omitted, it is inferred from the voice language, and getting it wrong is what fills a Spanish ad with British houses. `say` is ONE clause ending in its own punctuation \u2014 it becomes a caption card verbatim, so two sentences in one beat produce a card holding both. `show` is the shot brief for that line. Set `on_camera` ONLY when that beat\'s subject talks to camera. `voiceover: false` makes it a MUSIC-LED ad: nobody speaks, the `say` lines become on-screen text captioned straight from the script, and it needs `music`. `cast` is WHO the ad is about, and an ad with people in it needs one: `{ "avatar": "marta" }` names a cast avatar (`baker avatars list`) and every beat is grounded on that avatar\'s identity sheet with its subject description copied verbatim \u2014 the same face here as in the rest of the company\'s work. Without it each beat invents its own stranger, which is how one 28-second ad came back with five different men playing one customer. `{ "description": "..." }` is the fallback when there is no avatar, and it scaffolds a canvas that asks you to drop a photo before it can run \u2014 so prefer the avatar. A beat sets `cast: false` for a shot they are not in. `end_card` is on by default whenever `brand` is set: the last beat becomes a flat brand plate with the mark and a call to action drawn over it. `{ "cta": "..." }` sets the button copy, `false` keeps the footage.'
     },
     avatar: {
       type: "string",
@@ -33051,6 +33129,19 @@ var scaffoldAdCommand = defineCommand104({
     if (flagAvatar) {
       spec.data.cast = { ...spec.data.cast, avatar: flagAvatar };
     }
+    const contradicts = spec.data.beats.map((beat, i) => ({ i, show: beat.show })).filter(({ show }) => briefWantsLegibleText(show));
+    if (contradicts.length > 0) {
+      writeJson({
+        ok: false,
+        error: {
+          code: "VALIDATION_ERROR",
+          message: `Beat ${contradicts.map(({ i }) => i + 1).join(", ")} asks the picture for a readable figure (e.g. "${contradicts[0]?.show.trim().slice(0, 90)}"). Every frame also carries "no readable text or numbers", so the two fight and the picture wins \u2014 one such brief rendered a bill reading $58,085.00 over a total of $139.00.`,
+          fix: "Say what the shot LOOKS like, not what it reads: 'a bill in her hands, her face falling'. The ad's numbers belong in the `say` line, where the caption track renders them correctly."
+        }
+      });
+      process.exit(1);
+      return;
+    }
     const handle = flagAvatar || spec.data.cast?.avatar?.trim();
     let avatar = null;
     let avatarError = null;
@@ -33095,6 +33186,11 @@ var scaffoldAdCommand = defineCommand104({
     await writeSceneFiles(outDir, blueprint);
     await writeFile9(blueprintStylePath, renderStyleProjectionFromValue(blueprint), "utf8");
     const logoPath = spec.data.brand?.logo?.trim();
+    const transcriptPath = path24.join(outDir, "transcript.json");
+    if (spec.data.voiceover === false) {
+      await writeFile9(transcriptPath, `${JSON.stringify(staticTranscript(spec.data), null, 2)}
+`, "utf-8");
+    }
     const opts = {
       imageModel: AD_IMAGE_MODEL,
       // The sheet on the SAME model as the frames it feeds. A sheet from another
@@ -33114,7 +33210,8 @@ var scaffoldAdCommand = defineCommand104({
       captionWordsPerGroup: 8,
       // 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)
+      voiceLanguage: voiceLanguageFor(spec.data),
+      ...spec.data.voiceover === false ? { staticTranscriptPath: path24.relative(outDir, transcriptPath) } : {}
     };
     const canvas = scaffoldVideoCanvas(blueprint, adSpecCastElements(spec.data, avatar), opts);
     const renderNode3 = canvas.nodes.find((n) => n.type === "hyperframe_render");
@@ -33178,6 +33275,9 @@ var scaffoldAdCommand = defineCommand104({
         ...spec.data.cast && !avatar ? [
           "This cast has a description but no avatar, so the canvas asks you to drop a real source photo before it can run. `baker avatars list` \u2014 casting one and passing `cast.avatar` fills that slot for you and keeps the face consistent across every piece of work."
         ] : [],
+        ...spec.data.voiceover === false ? [
+          "Nobody speaks in this ad, so the `say` lines are ON-SCREEN text: they are captioned straight from the script, exactly as written, with no voice and no transcription in between. Give it `music` \u2014 a music-led ad with no bed is a silent one."
+        ] : [],
         "The caption cards are your `say` lines verbatim \u2014 read them back before running; that is the copy the viewer sees.",
         "Every clip renders with nobody speaking unless a beat set `on_camera`. If a beat needs a presenter talking, set it there rather than describing it in `show`.",
         ...spec.data.brand?.palette?.length && spec.data.brand?.logo ? [] : [