@koda-sl/baker-cli 0.90.1 → 0.91.0

package/dist/cli.js

@@ -9,7 +9,7 @@ import {
   defaultRegistry,
   generateCatalog,
   validateCanvasDeep
-} from "./chunk-2E4H2GIJ.js";
+} from "./chunk-LMVDA3EZ.js";
 
 // src/cli.ts
 import { defineCommand as defineCommand141, runMain } from "citty";
@@ -8287,6 +8287,24 @@ import { promisify as promisify2 } from "util";
 var execFileAsync2 = promisify2(execFile2);
 var PYSCENEDETECT_THRESHOLD = 18;
 var PYSCENEDETECT_MIN_SCENE_LEN_S = 0.25;
+var PYSCENEDETECT_RECHECK_THRESHOLD = 27;
+var PYSCENEDETECT_RECHECK_MIN_SCENE_LEN_S = 0.6;
+function isLikelyOverSegmented(cuts, opts = {}) {
+  const minCuts = opts.minCuts ?? 6;
+  const maxMedianGap = opts.medianGapS ?? 2;
+  const sorted = [...cuts].filter((c) => Number.isFinite(c) && c > 0).sort((a, b) => a - b);
+  if (sorted.length < minCuts) return false;
+  const gaps = [];
+  let prev = 0;
+  for (const c of sorted) {
+    gaps.push(c - prev);
+    prev = c;
+  }
+  gaps.sort((a, b) => a - b);
+  const mid = Math.floor(gaps.length / 2);
+  const median = gaps.length % 2 ? gaps[mid] : (gaps[mid - 1] + gaps[mid]) / 2;
+  return median < maxMedianGap;
+}
 function timecodeToSeconds(tc) {
   const m = tc.trim().match(/^(\d+):(\d{1,2}):(\d{1,2}(?:\.\d+)?)$/);
   if (!m) return null;
@@ -8306,9 +8324,7 @@ function parsePySceneDetectCsvCuts(csv) {
   }
   return [...new Set(cuts)].sort((a, b) => a - b);
 }
-async function detectSceneCutsPySceneDetect(filePath, opts = {}) {
-  const threshold = opts.threshold ?? PYSCENEDETECT_THRESHOLD;
-  const minSceneLenS = opts.minSceneLenS ?? PYSCENEDETECT_MIN_SCENE_LEN_S;
+async function runSceneDetectOnce(filePath, threshold, minSceneLenS, timeoutMs) {
   const outDir = await mkdtemp(join2(tmpdir(), "baker-scenedetect-"));
   try {
     await execFileAsync2(
@@ -8326,7 +8342,7 @@ async function detectSceneCutsPySceneDetect(filePath, opts = {}) {
         "list-scenes",
         "--quiet"
       ],
-      { encoding: "utf-8", maxBuffer: 32 * 1024 * 1024, timeout: opts.timeout_ms ?? 12e4 }
+      { encoding: "utf-8", maxBuffer: 32 * 1024 * 1024, timeout: timeoutMs }
     );
     const csvName = (await readdir2(outDir)).find((f) => f.toLowerCase().endsWith(".csv"));
     if (!csvName) return [];
@@ -8335,6 +8351,22 @@ async function detectSceneCutsPySceneDetect(filePath, opts = {}) {
     await rm(outDir, { recursive: true, force: true });
   }
 }
+async function detectSceneCutsPySceneDetect(filePath, opts = {}) {
+  const pinned = opts.threshold !== void 0;
+  const threshold = opts.threshold ?? PYSCENEDETECT_THRESHOLD;
+  const minSceneLenS = opts.minSceneLenS ?? PYSCENEDETECT_MIN_SCENE_LEN_S;
+  const timeoutMs = opts.timeout_ms ?? 12e4;
+  const cuts = await runSceneDetectOnce(filePath, threshold, minSceneLenS, timeoutMs);
+  if (!pinned && isLikelyOverSegmented(cuts)) {
+    return await runSceneDetectOnce(
+      filePath,
+      PYSCENEDETECT_RECHECK_THRESHOLD,
+      PYSCENEDETECT_RECHECK_MIN_SCENE_LEN_S,
+      timeoutMs
+    );
+  }
+  return cuts;
+}
 
 // src/engine/scaffold/video.ts
 import { z as z3 } from "zod";
@@ -8438,7 +8470,7 @@ var FIXED_TTS_MODEL = "elevenlabs/eleven_v3";
 var FIXED_SFX_MODEL = "elevenlabs/eleven_text_to_sound_v2";
 var FIXED_MUSIC_MODEL = "elevenlabs/music-v1";
 var FIXED_VOICE_CONVERT_MODEL = "elevenlabs/eleven_multilingual_sts_v2";
-var MUSIC_BED_GAIN_DB = -12;
+var MUSIC_BED_GAIN_DB = -20;
 var AMBIENT_BED_GAIN_DB = -20;
 var TRANSITION_DEFAULT_S = 0.4;
 var XFADE_BY_TYPE = {
@@ -8821,7 +8853,16 @@ function todoPath2(el, label) {
   return `[TODO: drop one real source image for ${label} (${el.type})${desc}${expr} \u2014 reused across every frame it appears in${fresh}${same}]`;
 }
 function buildElementSlots(elements) {
-  const usedIds = /* @__PURE__ */ new Set(["prompt", "spine", "overlaid", "audio_mix", "final", "music_bed"]);
+  const usedIds = /* @__PURE__ */ new Set([
+    "prompt",
+    "spine",
+    "overlaid",
+    "captions",
+    "captions_transcript",
+    "audio_mix",
+    "final",
+    "music_bed"
+  ]);
   const slots = [];
   assignElementLabels2(elements).forEach(({ el, label }, i) => {
     let id = sanitizeId2(`el_${label}`, `el_${i}`);
@@ -8892,6 +8933,9 @@ function buildFramePrompt(edge, sceneIndex, framePrompt, present, hasAnchor, mod
     "a studio) contains any text or graphics, DO NOT reproduce them \u2014 render the subject/scene",
     "only, leaving the regions where overlays will sit clean. Imperfect/garbled letterforms or",
     "stray icons are the worst outcome; leave those areas blank.",
+    "A SCREEN/UI surface \u2014 an app, website, chat, dashboard, or phone display \u2014 is NEVER",
+    "rendered here: leave any phone/screen OFF or blank-screened. The real interface is",
+    "composited later as a screenshot or a brand HTML block, never AI-generated.",
     "",
     "FRAMING \u2014 ONE UNCUT FRAME:",
     "Render ONE single uncut photographic frame: NO split screen, NO panels, NO dividing line,",
@@ -8959,8 +9003,13 @@ function ingestFrameRef(url, edge, ctx, nodes) {
 function buildFrameRef(edge, url, framePrompt, present, ctx, nodes) {
   const tag = ctx.tag ?? "";
   if (ctx.reuse && url) return ingestFrameRef(url, edge, ctx, nodes);
-  const hasOriginal = Boolean(url);
-  const originalRef = hasOriginal && url ? ingestFrameRef(url, edge, ctx, nodes) : void 0;
+  const hasPersonOrAnimal = present.some((s) => {
+    const t = s.type.toLowerCase();
+    return t === "person" || t === "animal";
+  });
+  const useOriginalAnchor = Boolean(url) && !hasPersonOrAnimal;
+  const hasOriginal = useOriginalAnchor;
+  const originalRef = useOriginalAnchor && url ? ingestFrameRef(url, edge, ctx, nodes) : void 0;
   const reference = [...present.map((s) => s.ref), ...originalRef ? [originalRef] : []];
   const genParams = {
     model: ctx.imageModel,
@@ -9151,12 +9200,21 @@ function emitSceneClip(i, scene, present, mode, nativeTurn, ambientBroll, frames
   return { ref: `$ref:s${i}${tag}_clip_trim.video`, scene_s: lengths.dur, out };
 }
 var COMPOSITE_LAYOUTS = /* @__PURE__ */ new Set(["split_screen", "pip", "keyed_overlay"]);
+var UI_SURFACE_RE = /\b(?:app|ui|web ?site|web ?page|website|browser|chat|interface|mock-?up|in[- ]?app|dashboard|app screen|phone screen|screen[- ]?(?:recording|capture|grab|share))\b/i;
+function regionIsUiSurface(r) {
+  return UI_SURFACE_RE.test(`${r.panel ?? ""} ${r.summary ?? ""} ${r.frame_prompt ?? ""}`);
+}
+function isUiOnlyComposite(regions) {
+  const ui = regions.filter(regionIsUiSurface).length;
+  return ui >= 1 && regions.length - ui <= 1;
+}
 function layeredComposition(scene) {
   const comp = scene.composition;
   const layout = (comp?.layout ?? "").toLowerCase();
   if (!COMPOSITE_LAYOUTS.has(layout)) return null;
   const regions = (comp?.regions ?? []).filter((r) => Boolean(r) && typeof r === "object");
   if (regions.length < 2) return null;
+  if (isUiOnlyComposite(regions)) return null;
   return { layout, regions, comp: comp ?? {} };
 }
 function splitAxisOf(comp, regions) {
@@ -9320,24 +9378,15 @@ function emitFlashHold(i, scene, slots, ctx, lengths, out, ar, nodes, clips) {
   });
   clips.push({ ref: `$ref:s${i}_clip.video`, scene_s: lengths.dur, out });
 }
-function musicScriptDigest(blueprint) {
-  const lines = blueprint.scenes.flatMap((s) => (s.dialogue ?? []).map((d) => d.line?.trim())).filter((l) => Boolean(l));
-  const script = lines.join(" ").slice(0, 500);
+function musicArcDigest(blueprint) {
   const roles = blueprint.scenes.map((s) => s.narrative_role).filter((r) => Boolean(r));
   const arc = roles.length > 0 ? roles.join(" \u2192 ") : "";
-  const parts = [];
-  if (script) {
-    parts.push(
-      `Ad script (the bed must SUPPORT these words \u2014 leave room for the voice, swell on the payoff): "${script}"`
-    );
-  }
-  if (arc) parts.push(`Emotional arc across scenes: ${arc}. Shape the bed's energy to this arc.`);
-  return parts.length > 0 ? `
+  return arc ? `
 
-${parts.join("\n")}` : "";
+Emotional arc across scenes: ${arc}. Shape the bed's energy to this arc, swelling on the payoff. Purely instrumental \u2014 no vocals, no singing, no spoken words.` : "";
 }
 function musicBedPrompt(blueprint, musicPrompt) {
-  const digest = musicScriptDigest(blueprint);
+  const digest = musicArcDigest(blueprint);
   const track2 = blueprint.global?.music?.identified_track;
   const title = track2?.title?.trim();
   const vibe = title ? `
@@ -9881,13 +9930,21 @@ function buildSfxMusic(blueprint, nodes) {
   const musicPrompt = blueprint.global?.music?.music_prompt;
   if (musicPrompt) {
     const total = blueprint.source?.duration_s ?? lastSceneEnd(blueprint);
-    const startAt = Math.min(Math.max(blueprint.global?.music?.starts_at_s ?? 0, 0), Math.max(total - 0.5, 0));
+    const hookEnd = blueprint.scenes[0]?.end_s ?? 0;
+    const startAt = Math.min(Math.max(blueprint.global?.music?.starts_at_s ?? 0, hookEnd), Math.max(total - 0.5, 0));
     const totalMs = Math.round((total - startAt) * 1e3);
     const musicMs = Math.min(Math.max(totalMs, 3e3), ELEVENLABS_MAX_MUSIC_LENGTH_MS);
     nodes.push({
       id: "music_bed",
       type: "music",
-      params: { model: FIXED_MUSIC_MODEL, prompt: musicBedPrompt(blueprint, musicPrompt), music_length_ms: musicMs }
+      // force_instrumental: the model is vocal-capable; without this it can SING the
+      // mood (and feeding it the script made it sing the ad). The voice owns the words.
+      params: {
+        model: FIXED_MUSIC_MODEL,
+        prompt: musicBedPrompt(blueprint, musicPrompt),
+        music_length_ms: musicMs,
+        force_instrumental: true
+      }
     });
     tracks.push({
       slot: "music",
@@ -9931,9 +9988,20 @@ function normalizeAnim(animation) {
   const mapped = animation === "slide" ? "slide_up" : animation;
   return SUPPORTED_ANIMS.has(mapped) ? mapped : void 0;
 }
+var FACE_ZONE_POSITIONS = /* @__PURE__ */ new Set([
+  "center",
+  "centre",
+  "mid-center",
+  "mid-centre",
+  "middle-center",
+  "center-center",
+  "mid",
+  "middle"
+]);
 function positionClass(position) {
   const p = (position ?? "bottom_center").toLowerCase().replace(/[^a-z]+/g, "-");
-  return `pos-${p}`;
+  const safe = FACE_ZONE_POSITIONS.has(p) ? "bottom-center" : p;
+  return `pos-${safe}`;
 }
 function collectCaptions(blueprint) {
   return blueprint.scenes.flatMap((scene) => {
@@ -10004,6 +10072,26 @@ function floatingStub(fe, sceneStart) {
     `<img class="ov ${positionClass(fe.position)}" src="your-${slug}.png" data-start="${at}" data-dur="${dur}" alt="" /> -->`
   ].join("\n");
 }
+function uiPipStub(scene) {
+  const comp = scene.composition;
+  const layout = (comp?.layout ?? "").toLowerCase();
+  if (!COMPOSITE_LAYOUTS.has(layout)) return "";
+  const regions = (comp?.regions ?? []).filter((r) => Boolean(r) && typeof r === "object");
+  if (regions.length < 2 || !isUiOnlyComposite(regions)) return "";
+  const ui = regions.find(regionIsUiSurface);
+  const at = scene.start_s ?? 0;
+  const dur = Math.max(0.5, Math.round(((scene.end_s ?? at + 2.5) - at) * 100) / 100);
+  const label = commentSafe(ui?.summary || ui?.frame_prompt || ui?.panel || "the app screen");
+  return [
+    `<!-- PHONE UI @ ${at}s for ${dur}s \u2014 the app/site screen this scene shows: ${label}.`,
+    "     Build it as a REAL surface, NEVER AI: capture the live page \u2014",
+    "       baker images screenshot https://<brand-domain>/<path>  (image-library skill)",
+    "     \u2014 OR hand-build a brand-accurate HTML screen; then frame it in a phone mockup:",
+    "       npx hyperframes add phone-scroll   (writes compositions/phone-scroll.html)",
+    "     drop the screenshot as screenshot.png in this dir and nest it as a PIP clip:",
+    `     <div data-composition-src="compositions/phone-scroll.html" data-start="${at}" data-duration="${dur}" data-track-index="2" data-width="1080" data-height="1920"></div> -->`
+  ].join("\n");
+}
 function buildOverlayHtml(input) {
   const blueprint = VideoBlueprint.parse(input);
   const blocks = [
@@ -10029,6 +10117,8 @@ function buildOverlayHtml(input) {
     const sceneStart = scene.start_s ?? 0;
     const floats = z3.array(FloatingElement).safeParse(scene.floating_elements ?? []);
     const parts = (floats.success ? floats.data.map((fe) => floatingStub(fe, sceneStart)) : []).filter(Boolean);
+    const pip = uiPipStub(scene);
+    if (pip) parts.push(pip);
     if (parts.length > 0) blocks.push(parts.join("\n"));
   }
   return blocks.join("\n\n");
@@ -10110,13 +10200,14 @@ function scaffoldVideoCanvas(input, elementsInput, opts) {
       params: { source: "path", path: todoPath2(elements[i], slot.label), expect: "image" }
     });
   });
-  if (opts.actorSheets) applyActorSheets(slots, nodes);
+  applyActorSheets(slots, nodes);
   const { clips, voTracks, vo_segments, talking_scenes } = buildTimeline(blueprint, slots, opts, nodes);
   let videoRef = buildSpine(clips, nodes);
   let videoNode = "spine";
   const overlays = blueprint.scenes.flatMap((s) => s.overlays ?? []);
   const floating = blueprint.scenes.flatMap((s) => s.floating_elements ?? []);
-  if (overlays.length > 0 || floating.length > 0) {
+  const hasUiPip = blueprint.scenes.some((s) => uiPipStub(s) !== "");
+  if (overlays.length > 0 || floating.length > 0 || hasUiPip) {
     nodes.push({
       id: "overlaid",
       type: "hyperframe_render",
@@ -10126,10 +10217,28 @@ function scaffoldVideoCanvas(input, elementsInput, opts) {
     videoRef = "$ref:overlaid.video";
     videoNode = "overlaid";
   }
+  if (opts.captionsCompositionPath && opts.transcriptPath) {
+    nodes.push({
+      id: "captions_transcript",
+      type: "ingest",
+      params: { source: "path", path: opts.transcriptPath, expect: "json" }
+    });
+    nodes.push({
+      id: "captions",
+      type: "hyperframe_render",
+      inputs: { background: videoRef, transcript: "$ref:captions_transcript.asset" },
+      params: { composition: opts.captionsCompositionPath }
+    });
+    videoRef = "$ref:captions.video";
+    videoNode = "captions";
+  }
   const tracks = [...voTracks, ...buildSfxMusic(blueprint, nodes)];
   if (tracks.length > 0) {
     const mixInputs = {};
     for (const t of tracks) mixInputs[t.slot] = t.ref;
+    const musicTrack = tracks.find((t) => t.kind === "music");
+    const voiceSlots = tracks.filter((t) => t.kind === "vo").map((t) => t.slot);
+    const duck = musicTrack && voiceSlots.length > 0 ? { duck: { track: musicTrack.slot, against: voiceSlots } } : {};
     nodes.push({
       id: "audio_mix",
       type: "audio_timeline",
@@ -10140,7 +10249,8 @@ function scaffoldVideoCanvas(input, elementsInput, opts) {
           start_s: t.start_s,
           ...t.gain_db !== void 0 ? { gain_db: t.gain_db } : {}
         })),
-        total_ms: Math.round((blueprint.source?.duration_s ?? lastSceneEnd(blueprint)) * 1e3)
+        total_ms: Math.round((blueprint.source?.duration_s ?? lastSceneEnd(blueprint)) * 1e3),
+        ...duck
       }
     });
     nodes.push({
@@ -10432,6 +10542,7 @@ function resolveShippedCanvasDir(name, startDir, exists = existsSync3, maxDepth
 
 // src/commands/canvas/scaffold-video.ts
 var SHIPPED_COMPOSITION_DIR = resolveShippedCanvasDir("video-overlay-composition", import.meta.dirname);
+var SHIPPED_CAPTIONS_DIR = resolveShippedCanvasDir("tiktok-captions-composition", import.meta.dirname);
 function resolveModel2(kind, preferred) {
   const ids = Object.keys(MODEL_REGISTRY[kind]);
   return ids.includes(preferred) ? preferred : ids[0] ?? preferred;
@@ -10463,6 +10574,24 @@ async function loadAssetText2(ref, label) {
   }
   throw new Error(`${label}: output had no readable path or url`);
 }
+async function loadTranscriptBestEffort(ref) {
+  if (!ref) return void 0;
+  try {
+    return await loadAssetText2(ref, "deconstruct transcript");
+  } catch {
+    return void 0;
+  }
+}
+async function stageCaptions(outDir, transcript) {
+  const text = transcript?.trim();
+  if (!text || text === "[]") return {};
+  const transcriptPath = path5.join(outDir, "transcript.json");
+  await writeFile2(transcriptPath, `${text}
+`, "utf8");
+  const compositionPath = path5.join(outDir, "tiktok-captions-composition");
+  await cp(SHIPPED_CAPTIONS_DIR, compositionPath, { recursive: true });
+  return { compositionPath, transcriptPath };
+}
 function parseElements2(raw) {
   const parsed = JSON.parse(raw);
   if (Array.isArray(parsed)) return parsed;
@@ -10568,10 +10697,12 @@ async function runAnalysisPasses(deconstructCanvas, selectModel) {
     }
   };
   let blueprint;
+  let transcript;
   try {
     const r1 = await engine.run(deconstructCanvas, {});
     addCredits(r1.stats);
     blueprint = JSON.parse(await loadAssetText2(r1.outputs_by_node.deconstruct?.analysis, "deconstruct output"));
+    transcript = await loadTranscriptBestEffort(r1.outputs_by_node.deconstruct?.transcript);
   } catch (e) {
     if (e instanceof ValidationError) return fail2("validation", JSON.stringify(e.issues));
     if (e instanceof SyntaxError) return fail2("read_outputs", e.message);
@@ -10582,7 +10713,7 @@ async function runAnalysisPasses(deconstructCanvas, selectModel) {
     const r2 = await engine.run(buildSelectCanvas(selectModel, slimJson), {});
     addCredits(r2.stats);
     const elements = parseElements2(await loadAssetText2(r2.outputs_by_node.select?.text, "selection output"));
-    return { blueprint, elements, creditsSpent: sawCredits ? credits : void 0 };
+    return { blueprint, elements, transcript, creditsSpent: sawCredits ? credits : void 0 };
   } catch (e) {
     if (e instanceof ValidationError) return fail2("validation", JSON.stringify(e.issues));
     if (e instanceof SyntaxError) return fail2("read_outputs", e.message);
@@ -10602,14 +10733,10 @@ var scaffoldVideoCommand = defineCommand76({
       type: "boolean",
       description: "Give silent b-roll scenes native diegetic ambient mixed deep under the music bed (off by default)"
     },
-    "actor-sheets": {
-      type: "boolean",
-      description: "Lock a recast person/animal that recurs across \u22652 scenes to ONE turnaround sheet grounding every frame"
-    },
     "max-scenes": { type: "string", description: "Cap the number of scenes the deconstruct emits" },
     "shot-threshold": {
       type: "string",
-      description: "PySceneDetect content threshold (default 18; lower = more/softer cuts, higher = fewer)"
+      description: "PySceneDetect content threshold. Default is adaptive (18, auto re-checked at 27 when a continuous shot looks over-segmented); pinning a value disables the re-check. Lower = more/softer cuts, higher = fewer."
     },
     language: { type: "string", description: "Transcript/dialogue language hint (e.g. fr, en)" },
     focus: { type: "string", description: "Known provenance/emphasis to ground the deconstruct" },
@@ -10641,7 +10768,7 @@ var scaffoldVideoCommand = defineCommand76({
       focus: args.focus ? String(args.focus) : void 0,
       shotCuts
     });
-    const { blueprint, elements, creditsSpent } = await runAnalysisPasses(deconstructCanvas, selectModel);
+    const { blueprint, elements, transcript, creditsSpent } = await runAnalysisPasses(deconstructCanvas, selectModel);
     await mkdir(outDir, { recursive: true });
     const annotated = annotateBlueprintWithElements(blueprint, elements);
     await writeFile2(blueprintPath, `${JSON.stringify(annotated, null, 2)}
@@ -10659,14 +10786,16 @@ var scaffoldVideoCommand = defineCommand76({
       );
     }
     await writeFile2(indexPath, injected, "utf8");
+    const captions = await stageCaptions(outDir, transcript);
     const opts = {
       imageModel,
       videoModel,
       overlayCompositionPath: compositionDest,
+      captionsCompositionPath: captions.compositionPath,
+      transcriptPath: captions.transcriptPath,
       blueprintPath,
       frames,
-      ambient: Boolean(args.ambient),
-      actorSheets: Boolean(args["actor-sheets"])
+      ambient: Boolean(args.ambient)
     };
     let canvas;
     let report;