@koda-sl/baker-cli 0.123.0-dev.4a85b9f30 → 0.123.0-dev.8e4328629

package/dist/cli.js

@@ -857,7 +857,6 @@ var LINKEDIN_LIMITS = {
     choiceOptionsMax: 30,
     choiceOptionTextMax: 100,
     thankYouMessageMax: 300,
-    privacyPolicyTextMax: 2e3,
     legalDisclaimerMax: 2e3,
     consentsMax: 5,
     // Campaign Manager caps disclosure checkboxes at 5
@@ -1452,7 +1451,6 @@ var leadFormFields = {
   /** Form language, e.g. { country: "US", language: "en" }. Defaults to the account locale on LinkedIn. */
   locale: z2.object({ country: z2.string().length(2), language: z2.string().length(2) }).optional(),
   privacyPolicyUrl: httpsUrlSchema,
-  privacyPolicyText: z2.string().max(LEAD.privacyPolicyTextMax).optional(),
   questions: z2.array(leadFormQuestionSchema).min(1).max(LEAD.questionsMax),
   consents: z2.array(leadFormConsentSchema).max(LEAD.consentsMax).optional(),
   hiddenFields: z2.array(leadFormHiddenFieldSchema).max(LEAD.hiddenFieldsMax).optional(),
@@ -9392,7 +9390,7 @@ var leadFormsCreateCommand = defineCommand38({
 Required: name, headline (\u226460), privacyPolicyUrl, questions[] (\u226412; playbook: \u22644 for completion).
 Each question is a predefined profile field ({ name, predefinedField: "EMAIL" }) or a custom question ({ name, questionType: "MULTIPLE_CHOICE", options: [...] }; \u22643 custom).
 Best-practice fields the preview will nudge for if missing: 1-3 qualifying questions, consents[] (disclosure checkboxes), thankYou.message + thankYou.landingUrl|appointmentUrl.
-Also supported: locale, formImageId|formImageUrn, privacyPolicyText, hiddenFields[], legalDisclaimer, thankYou.cta. Example: baker ads linkedin lead-forms create --file form.json`
+Also supported: locale, formImageId|formImageUrn, hiddenFields[], legalDisclaimer, thankYou.cta. Example: baker ads linkedin lead-forms create --file form.json`
   },
   args: {
     ...accountArgs,
@@ -14441,23 +14439,28 @@ function isResolvableRelative(value) {
   return typeof value === "string" && value.length > 0 && !value.includes("[TODO") && !path2.isAbsolute(value);
 }
 
-// src/commands/canvas/run-record.ts
-import path3 from "path";
-var MAX_RUN_NODES = 200;
-var MAX_OUTPUTS_PER_NODE = 10;
-var MAX_FINAL_OUTPUTS = 10;
+// src/commands/canvas/node-preview.ts
 var MAX_PARAMS_PREVIEW_LENGTH = 4e3;
-function paramsPreviewFromParams(params) {
-  if (params === void 0 || params === null) return void 0;
-  const text = humanParamText(params, "prompt") ?? humanParamText(params, "source") ?? compactJson(params);
-  if (!text) return void 0;
-  return text.length > MAX_PARAMS_PREVIEW_LENGTH ? `${text.slice(0, MAX_PARAMS_PREVIEW_LENGTH - 1)}\u2026` : text;
-}
-function humanParamText(params, key) {
+var SPEC_KEYS = [
+  "model",
+  "aspect_ratio",
+  "image_size",
+  "resolution",
+  "duration",
+  "subject_type",
+  "font_size",
+  "expect",
+  "source",
+  "path",
+  "url"
+];
+var BODY_KEYS = ["prompt", "text", "subject_description"];
+function scalarField(params, key) {
   const value = params[key];
+  if (typeof value === "number") return String(value);
   if (typeof value !== "string") return void 0;
   const trimmed = value.trim();
-  return trimmed && !trimmed.startsWith("$ref:") ? trimmed : void 0;
+  return trimmed && !trimmed.startsWith(REF_PREFIX) ? trimmed : void 0;
 }
 function compactJson(params) {
   try {
@@ -14467,6 +14470,32 @@ function compactJson(params) {
     return void 0;
   }
 }
+function cap(text) {
+  return text.length > MAX_PARAMS_PREVIEW_LENGTH ? `${text.slice(0, MAX_PARAMS_PREVIEW_LENGTH - 1)}\u2026` : text;
+}
+function nodeParamsPreview(params, deps) {
+  const refsLine = deps && deps.length > 0 ? `refs: ${deps.join(", ")}` : void 0;
+  if (!params || typeof params !== "object") return refsLine;
+  const record = params;
+  const spec = [];
+  for (const key of SPEC_KEYS) {
+    const value = scalarField(record, key);
+    if (value !== void 0) spec.push(`${key}: ${value}`);
+  }
+  if (refsLine) spec.push(refsLine);
+  const body = BODY_KEYS.map((key) => scalarField(record, key)).find(Boolean);
+  const header = spec.join("\n");
+  const composed = header && body ? `${header}
+
+${body}` : header || body || compactJson(record);
+  return composed ? cap(composed) : void 0;
+}
+
+// src/commands/canvas/run-record.ts
+import path3 from "path";
+var MAX_RUN_NODES = 200;
+var MAX_OUTPUTS_PER_NODE = 10;
+var MAX_FINAL_OUTPUTS = 10;
 var MAX_CREATIVE_SLUG_LENGTH = 100;
 function creativeSlugFromCanvasPath(filePath) {
   const normalized = filePath.split(path3.sep).join("/");
@@ -14585,7 +14614,7 @@ var RunProgressTracker = class {
           outputs: [],
           deps: node.deps,
           status: "pending",
-          paramsPreview: paramsPreviewFromParams(node.params)
+          paramsPreview: nodeParamsPreview(node.params, node.deps)
         });
       }
       this.planned = true;
@@ -15014,6 +15043,12 @@ import { defineCommand as defineCommand89 } from "citty";
 import { z as z11 } from "zod";
 var GEN_ASPECT_RATIOS = /* @__PURE__ */ new Set(["1:1", "4:5", "9:16", "16:9", "4:3", "3:4", "2:3", "3:2", "21:9"]);
 var DEFAULT_ASPECT_RATIO = "9:16";
+var SHEET_SUBJECT_TYPE = {
+  person: "person",
+  animal: "character"
+};
+var ACTOR_SHEET_MODEL = "google/gemini-3-pro-image-preview";
+var ACTOR_SHEET_IMAGE_SIZE = "4K";
 var Blueprint = z11.object({
   meta: z11.object({ estimated_aspect_ratio: z11.string().optional() }).loose().optional(),
   text_content: z11.array(z11.object({ text: z11.string().optional() }).loose()).optional()
@@ -15085,6 +15120,11 @@ function todoPath(el, label) {
   const lock = t === "person" || t === "animal" ? " \u2014 REQUIRED: this is the emotional hero; ground it in a real reference (image-library / video-library / Pinterest), do not delete this slot and free-generate" : "";
   return `[TODO: drop a real image for ${label} (${el.type})${desc}${expr}${lock}]`;
 }
+function sheetSubjectDescription(el) {
+  const desc = el.description ?? `the ${el.type}`;
+  const expr = el.expression ? `, ${el.expression} expression` : "";
+  return `${desc}${expr} \u2014 clean plate: no phone-camera UI chrome, no app interface, no watermarks, no captions or on-image text`;
+}
 function aspectRatio(blueprint, opts) {
   const requested = opts.aspectRatio ?? blueprint.meta?.estimated_aspect_ratio;
   return requested && GEN_ASPECT_RATIOS.has(requested) ? requested : DEFAULT_ASPECT_RATIO;
@@ -15108,6 +15148,7 @@ function scaffoldStaticAd(input, elementsInput, opts) {
     type: "ingest",
     params: opts.imageIsUrl ? { source: "url", url: opts.imagePath, expect: "image" } : { source: "path", path: opts.imagePath, expect: "image" }
   });
+  const includeActorSheets = opts.includeActorSheets !== false;
   const usedIds = /* @__PURE__ */ new Set(["prompt", "original", "gen", "brandfont", "type_ref"]);
   const elementSlots = [];
   assignElementLabels(elements).forEach(({ el, label }, i) => {
@@ -15119,7 +15160,25 @@ function scaffoldStaticAd(input, elementsInput, opts) {
       type: "ingest",
       params: { source: "path", path: todoPath(el, label), expect: "image" }
     });
-    elementSlots.push({ ref: `$ref:${id}.asset`, label, type: el.type });
+    let ref = `$ref:${id}.asset`;
+    const subjectType = includeActorSheets ? SHEET_SUBJECT_TYPE[el.type.toLowerCase()] : void 0;
+    if (subjectType) {
+      const sheetId = `${id}_sheet`;
+      usedIds.add(sheetId);
+      nodes.push({
+        id: sheetId,
+        type: "image_reference_sheet",
+        inputs: { references: [ref] },
+        params: {
+          model: ACTOR_SHEET_MODEL,
+          subject_description: sheetSubjectDescription(el),
+          subject_type: subjectType,
+          image_size: ACTOR_SHEET_IMAGE_SIZE
+        }
+      });
+      ref = `$ref:${sheetId}.sheet`;
+    }
+    elementSlots.push({ ref, label, type: el.type });
   });
   const hasFont = includeFont;
   if (hasFont) {
@@ -15180,6 +15239,7 @@ function scaffoldStaticAd(input, elementsInput, opts) {
 function staticAdReport(input, elementsInput, opts) {
   const blueprint = Blueprint.parse(input);
   const elements = MainElements.parse(elementsInput);
+  const includeActorSheets = opts.includeActorSheets !== false;
   return {
     element_count: elements.length,
     elements: assignElementLabels(elements).map(({ el, label }) => ({
@@ -15189,6 +15249,7 @@ function staticAdReport(input, elementsInput, opts) {
       asset_todo: todoPath(el, label)
     })),
     includes_font: opts.includeFont !== false,
+    actor_sheets: includeActorSheets ? assignElementLabels(elements).filter(({ el }) => SHEET_SUBJECT_TYPE[el.type.toLowerCase()]).map(({ label }) => label) : [],
     aspect_ratio: aspectRatio(blueprint, opts)
   };
 }
@@ -15295,7 +15356,12 @@ function canvasToDefinitionGraph(canvas) {
     };
     walkStrings(inputs, collect);
     walkStrings(params, collect);
-    return deps.size > 0 ? { id, type, deps: [...deps] } : { id, type };
+    const depList = [...deps];
+    const paramsPreview = nodeParamsPreview(params, depList);
+    const base = { id, type };
+    if (depList.length > 0) base.deps = depList;
+    if (paramsPreview) base.paramsPreview = paramsPreview;
+    return base;
   });
   const rawOutput = canvas?.output;
   const outNode = rawOutput?.node;
@@ -15366,7 +15432,7 @@ function resolveModel(kind, preferred) {
   const ids = Object.keys(MODEL_REGISTRY[kind]);
   return ids.includes(preferred) ? preferred : ids[0] ?? preferred;
 }
-var DESCRIBE_FOCUS = `the GLOBAL LAYOUT GEOMETRY above all \u2014 the column/row grid, each region's approximate bounds as a percentage of frame width and height, any panel SPLITS and their proportions (e.g. a left column split into a ~60% top photo and a ~20% bottom price box), what occupies every region, and for EVERY text block its string verbatim, relative size (x_large | large | medium | small), weight, case, color, and alignment \u2014 including small bottom-corner fine print. Also: the EXPRESSION and emotion of every person/animal (capture exaggerated or AI-edited faces), the ad_intent (what feeling it engineers and how), per-color brand_ownership (which colors are the advertiser's brand vs borrowed-functional like a red comparison column), logos by brand, and all visible text verbatim. CRITICAL \u2014 name the WINNING MECHANISM(S): the special sauce, the specific thing that makes THIS ad a candidate winner rather than a generic ad. A great ad is rarely great by accident, and the mechanism is invisible unless named, so a faithful rewrite quietly discards it. It may be VERBAL (a rhyme, pun, alliteration, rhythm/meter, repetition, antithesis, double meaning), VISUAL (an unexpected crop or scale, a visual gag or pun, a striking juxtaposition, a pattern interrupt, an exaggerated/AI-edited expression, a before/after or comparison, bold negative space, a surprising focal point), or STRUCTURAL (an unusual hook, order, or reveal). Record each as a \`winning_mechanisms\` array of { kind: verbal|visual|structural, device, why_it_works } (e.g. { kind: verbal, device: "rhyme \u2014 'today' / 'go away'", why_it_works: "memorable, playful, reframes cost as relief" }) so a rewrite can rebuild the SAUCE for our brand instead of adapting only the surface and losing what made it convert`;
+var DESCRIBE_FOCUS = "the GLOBAL LAYOUT GEOMETRY above all \u2014 the column/row grid, each region's approximate bounds as a percentage of frame width and height, any panel SPLITS and their proportions (e.g. a left column split into a ~60% top photo and a ~20% bottom price box), what occupies every region, and for EVERY text block its string verbatim, relative size (x_large | large | medium | small), weight, case, color, and alignment \u2014 including small bottom-corner fine print. Also: the EXPRESSION and emotion of every person/animal (capture exaggerated or AI-edited faces), the ad_intent (what feeling it engineers and how), per-color brand_ownership (which colors are the advertiser's brand vs borrowed-functional like a red comparison column), logos by brand, and all visible text verbatim. the TYPOGRAPHY as a first-class identity signal \u2014 for EVERY distinct typeface in the ad record a `fonts` entry: which text blocks use it, its classification (serif | slab-serif | sans | script | display | mono), a best-guess family or a close lookalike (e.g. 'a Tiempos-like high-contrast serif', 'Circular-like geometric sans'), and its weight/case/styling \u2014 because a faithful rewrite that keeps the words but loses the letterforms loses the brand; ALSO record whether the advertiser's LOGO appears in more than one lockup (a square/icon MARK and a horizontal WORDMARK are two different assets \u2014 capture BOTH). CRITICAL \u2014 name the WINNING MECHANISM(S): the special sauce, the specific thing that makes THIS ad a candidate winner rather than a generic ad. A great ad is rarely great by accident, and the mechanism is invisible unless named, so a faithful rewrite quietly discards it. It may be VERBAL (a rhyme, pun, alliteration, rhythm/meter, repetition, antithesis, double meaning), VISUAL (an unexpected crop or scale, a visual gag or pun, a striking juxtaposition, a pattern interrupt, an exaggerated/AI-edited expression, a before/after or comparison, bold negative space, a surprising focal point), or STRUCTURAL (an unusual hook, order, or reveal). Record each as a `winning_mechanisms` array of { kind: verbal|visual|structural, device, why_it_works } (e.g. { kind: verbal, device: \"rhyme \u2014 'today' / 'go away'\", why_it_works: \"memorable, playful, reframes cost as relief\" }) so a rewrite can rebuild the SAUCE for our brand instead of adapting only the surface and losing what made it convert";
 var LAYOUT_SYSTEM = "You convert an advertisement's JSON blueprint into a precise, structured LAYOUT MAP of the frame \u2014 the spatial grid a designer would rebuild it from. Be exhaustive and quantitative: every region, its bounds as a percentage of the frame, and every text block with its relative size. Output ONLY a JSON object, no prose.";
 var LAYOUT_PROMPT = `AD BLUEPRINT (from image_describe):
 {{blueprint}}
@@ -15392,7 +15458,7 @@ var SELECT_PROMPT = `AD BLUEPRINT (from image_describe):
 {{blueprint}}
 
 From this blueprint, list ONLY the elements that are prominent, important, and identity-bearing \u2014 the ones a reproduction must ground in a real asset:
-- the brand logo/wordmark (from brands_logos with function_in_image = advertiser_brand) -> type "logo"
+- the brand logo/wordmark (from brands_logos with function_in_image = advertiser_brand) -> type "logo". A brand often ships its logo in TWO lockups \u2014 a square/icon MARK and a horizontal WORDMARK \u2014 and an ad may use both (e.g. an icon in a product mock plus the wordmark in the sign-off). When more than one distinct lockup of the SAME advertiser appears, emit a SEPARATE logo element for each (label them by lockup, e.g. LOGO_MARK and LOGO_WORDMARK) so the reproduction can drop the right file in each slot instead of stretching one logo to cover both.
 - trust/rating/certification/app-store/review badges (brands_logos with function_in_image = trust_badge | review_platform | certification_or_seal | app_store_badge | payment_method) -> type "badge"
 - a showcased/hero product or package (a foreground entry in subjects that the ad is selling) -> type "product"
 - a foreground person (from people) \u2014 keep it when its identity matters OR it is the emotional/hero focal point of the ad, even a generic one with no brand identity -> type "person"
@@ -15536,7 +15602,11 @@ var scaffoldStaticAdCommand = defineCommand89({
     "layout-model": { type: "string", description: "Override the text_generate model id for the layout pass" },
     "gen-model": { type: "string", description: "Override the image_generate model id" },
     aspect: { type: "string", description: "Force the output aspect ratio (else inferred from the image)" },
-    "skip-font": { type: "boolean", description: "Skip the brand-font \u2192 type-specimen slot" }
+    "skip-font": { type: "boolean", description: "Skip the brand-font \u2192 type-specimen slot" },
+    "skip-actor-sheets": {
+      type: "boolean",
+      description: "Ground each person/animal on its lone dropped photo instead of a generated multi-view sheet"
+    }
   },
   async run({ args }) {
     const slug = args.slug ? String(args.slug) : void 0;
@@ -15586,7 +15656,8 @@ var scaffoldStaticAdCommand = defineCommand89({
       imageIsUrl: canvasImageIsUrl,
       blueprintPath: canvasBlueprintPath,
       aspectRatio: args.aspect ? String(args.aspect) : void 0,
-      includeFont: !args["skip-font"]
+      includeFont: !args["skip-font"],
+      includeActorSheets: !args["skip-actor-sheets"]
     };
     let canvas;
     let report;
@@ -15653,7 +15724,8 @@ var scaffoldStaticAdCommand = defineCommand89({
           checklist: {
             edit_prompt: `Edit ${path9.basename(blueprintPath)} \u2014 it is the blueprint generated from your image; rewrite it into the ad you want (palette, copy, claims, subjects). It feeds the generator directly.`,
             assets_to_supply: report.elements,
-            font_slot: report.includes_font ? "Drop a brand font at the [TODO] brandfont path, or delete the brandfont + type_ref nodes to skip it." : "skipped (--skip-font)",
+            font_slot: report.includes_font ? "Drop a brand font at the [TODO] brandfont path (the describe pass recorded the ad's typefaces under `fonts` in prompt.json \u2014 match those). The font is wired into the render as a TYPE SPECIMEN reference so generated text takes the brand letterforms. Delete the brandfont + type_ref nodes to skip it." : "skipped (--skip-font)",
+            actor_sheets: report.actor_sheets.length > 0 ? `Each living hero (${report.actor_sheets.join(", ")}) is fused into a generated multi-view reference sheet (image_reference_sheet) that the render grounds on \u2014 so drop ONE clean photo at that hero's ingest and the sheet builds the consistent turnaround. Pass --skip-actor-sheets to ground on the lone photo instead.` : "none (no person/animal heroes detected, or --skip-actor-sheets)",
             note: "Populate as you go: for each [TODO] ingest slot, source its real asset and wire it into the slot right away \u2014 one at a time, not all sourced first then reconciled at the end. When every slot is filled, `baker canvas validate` then `baker canvas run`. Running generates a billed image \u2014 it is not free."
           }
         },
@@ -16475,8 +16547,8 @@ function extendPresenceByPromptMentions(slots, blueprint) {
     }
   });
 }
-var ACTOR_SHEET_MODEL = "google/gemini-3-pro-image-preview";
-var SHEET_SUBJECT_TYPE = {
+var ACTOR_SHEET_MODEL2 = "google/gemini-3-pro-image-preview";
+var SHEET_SUBJECT_TYPE2 = {
   person: "person",
   animal: "character",
   product: "product",
@@ -16484,7 +16556,7 @@ var SHEET_SUBJECT_TYPE = {
 };
 function buildElementSheets(slots, nodes) {
   for (const slot of slots) {
-    const subjectType = SHEET_SUBJECT_TYPE[slot.type.toLowerCase()];
+    const subjectType = SHEET_SUBJECT_TYPE2[slot.type.toLowerCase()];
     if (!subjectType) continue;
     if (slot.sameAs) continue;
     if (slot.presence.size < 1) continue;
@@ -16497,7 +16569,7 @@ function buildElementSheets(slots, nodes) {
       // The lone dropped ingest is the source; the sheet fans it into a turnaround.
       inputs: { references: [slot.ref] },
       params: {
-        model: ACTOR_SHEET_MODEL,
+        model: ACTOR_SHEET_MODEL2,
         // The clean-plate clause mirrors the frame prompts' CLEAN PLATE block: a sheet
         // that comes back with a fake camera app baked in (P2-22) poisons EVERY frame
         // grounded on it, so the suppression must live on the sheet too.
@@ -17426,10 +17498,16 @@ function makePresenterPresent(slots, canonical, opts = {}) {
     return presence.has(sceneIndex);
   };
 }
-var PAUSE_GAP_S = 0.6;
 var SEEDANCE_SAFE_MAX_S = SEEDANCE_DURATIONS.find((d) => d >= 10) ?? 10;
 var PHRASE_MAX_S = SEEDANCE_SAFE_MAX_S;
-var SEEDANCE_MAX_WORDS_PER_TAKE = 16;
+var PAUSE_GAP_S = 0.6;
+function isAdjacentShownCut(ln, lastShownScene, scenes) {
+  if (!ln.shown || lastShownScene === null) return false;
+  return ln.sceneIndex === lastShownScene + 1 && scenes[ln.sceneIndex]?.continues_previous !== true;
+}
+function breaksPhrase(cur, ln, lineCover, lineClipStart, scenes) {
+  return cur.speaker !== ln.speaker || ln.start - cur.end > PAUSE_GAP_S || isAdjacentShownCut(ln, cur.lastShownScene, scenes) || Math.max(cur.coverEnd, lineCover) - Math.min(cur.clipStart, lineClipStart) > PHRASE_MAX_S;
+}
 var JOIN_DEDUP_MAX_WORDS = 4;
 function joinKey(word) {
   return word.toLowerCase().replace(/[^\p{L}\p{N}]+/gu, "");
@@ -17472,10 +17550,7 @@ function collapseVoiceover(blueprint) {
   const presenter = [...presenters][0];
   return (speaker) => NARRATOR_SPEAKERS.has(speaker.toLowerCase()) ? presenter : speaker;
 }
-function buildPhrases(blueprint, canonical, compositeScenes, presenterPresent, presentStrict) {
-  const casts = castIdSet(blueprint);
-  const cameraOn = onCameraDialogue(blueprint);
-  const sceneEndS = (i) => blueprint.scenes[i]?.end_s ?? blueprint.scenes[i]?.start_s ?? 0;
+function multiSpeakerScenes(blueprint, casts, cameraOn, canonical, presentStrict) {
   const multiSpeaker = /* @__PURE__ */ new Set();
   blueprint.scenes.forEach((scene, i) => {
     const onCamAll = new Set(
@@ -17485,32 +17560,45 @@ function buildPhrases(blueprint, canonical, compositeScenes, presenterPresent, p
     const effective = onCamPresent.length > 0 ? new Set(onCamPresent) : onCamAll;
     if (effective.size >= 2) multiSpeaker.add(i);
   });
-  const lines = blueprint.scenes.flatMap(
-    (scene, sceneIndex) => compositeScenes.has(sceneIndex) ? [] : (scene.dialogue ?? []).filter((l) => Boolean(l.line?.trim())).map((l) => {
+  return multiSpeaker;
+}
+function lineClipWindow(ln, scenes) {
+  if (!ln.shown) return { cover: ln.end, clipStart: ln.start };
+  const sc = scenes[ln.sceneIndex];
+  const sceneEnd = sc?.end_s ?? sc?.start_s ?? 0;
+  return { cover: Math.max(ln.end, sceneEnd), clipStart: Math.min(ln.start, sc?.start_s ?? ln.start) };
+}
+function dialogueLines(blueprint, ctx) {
+  return blueprint.scenes.flatMap((scene, sceneIndex) => {
+    if (ctx.compositeScenes.has(sceneIndex)) return [];
+    return (scene.dialogue ?? []).filter((l) => Boolean(l.line?.trim())).map((l) => {
       const raw = l.speaker ?? "voiceover";
-      const sp = canonical(raw);
       const text = l.line.trim();
       const start = l.start_s ?? scene.start_s ?? 0;
+      const shown = l.on_camera !== false && !sceneIsAllGraphic(scene) && isOnCameraSpeaker(raw, ctx.casts, ctx.cameraOn) && !ctx.multiSpeaker.has(sceneIndex) && ctx.presenterPresent(ctx.canonical(raw), sceneIndex);
       return {
         sceneIndex,
-        speaker: sp,
-        // Shown = a cast member speaking AND their element is actually on screen
-        // here (not a cutaway). A b-roll cutaway mid-phrase fails this and gets
-        // its own clip while the phrase voice plays under it. An explicit
-        // deconstruct voiceover stamp (`on_camera: false`) wins over element
-        // presence — a speaker pictured in a photo is "present" but not talking.
-        // An all-graphic composition (no camera region) is voiceover by
-        // definition: nobody is on screen to lip-sync.
-        shown: l.on_camera !== false && !sceneIsAllGraphic(scene) && isOnCameraSpeaker(raw, casts, cameraOn) && !multiSpeaker.has(sceneIndex) && presenterPresent(sp, sceneIndex),
+        speaker: ctx.canonical(raw),
+        shown,
         start,
-        // Real speech end. When the deconstruct gives no end_s, estimate it from
-        // the words — NOT the scene end (which would fabricate continuity across
-        // a long silent b-roll gap and wrongly merge two separate phrases).
         end: l.end_s ?? start + estSpeechS(text),
         text
       };
-    })
-  ).sort((a, b) => a.start - b.start);
+    });
+  }).sort((a, b) => a.start - b.start);
+}
+function buildPhrases(blueprint, canonical, compositeScenes, presenterPresent, presentStrict) {
+  const casts = castIdSet(blueprint);
+  const cameraOn = onCameraDialogue(blueprint);
+  const multiSpeaker = multiSpeakerScenes(blueprint, casts, cameraOn, canonical, presentStrict);
+  const lines = dialogueLines(blueprint, {
+    compositeScenes,
+    multiSpeaker,
+    canonical,
+    casts,
+    cameraOn,
+    presenterPresent
+  });
   const phrases = [];
   let cur = null;
   const flush = () => {
@@ -17528,16 +17616,8 @@ function buildPhrases(blueprint, canonical, compositeScenes, presenterPresent, p
     cur = null;
   };
   for (const ln of lines) {
-    const lineCover = ln.shown ? Math.max(ln.end, sceneEndS(ln.sceneIndex)) : ln.end;
-    const lineClipStart = ln.shown ? Math.min(ln.start, blueprint.scenes[ln.sceneIndex]?.start_s ?? ln.start) : ln.start;
-    const breakRun = !cur || cur.speaker !== ln.speaker || ln.start - cur.end > PAUSE_GAP_S || // Cap by SCENE COVERAGE span, not line end — a presenter run whose sliced scenes span
-    // more than one Seedance clip splits into the next take here (at this scene's
-    // boundary, never mid-scene), so no segment ever reads past the generated clip.
-    Math.max(cur.coverEnd, lineCover) - Math.min(cur.clipStart, lineClipStart) > PHRASE_MAX_S || // Cap by SPOKEN WORDS too: a dense read that fits the time ceiling can still cram a
-    // monologue Seedance can't lip-sync cleanly. Break before this line pushes the run
-    // past the word cap (a fresh run always accepts its own first line, so one long line
-    // is never blocked — it just can't be sub-split without a mid-word cut).
-    cur.words + wordCount(ln.text) > SEEDANCE_MAX_WORDS_PER_TAKE;
+    const { cover: lineCover, clipStart: lineClipStart } = lineClipWindow(ln, blueprint.scenes);
+    const breakRun = !cur || breaksPhrase(cur, ln, lineCover, lineClipStart, blueprint.scenes);
     if (breakRun || !cur) {
       flush();
       cur = {
@@ -17548,7 +17628,7 @@ function buildPhrases(blueprint, canonical, compositeScenes, presenterPresent, p
         coverEnd: lineCover,
         clipStart: lineClipStart,
         texts: [ln.text],
-        words: wordCount(ln.text),
+        lastShownScene: ln.shown ? ln.sceneIndex : null,
         shown: /* @__PURE__ */ new Set()
       };
     } else {
@@ -17556,7 +17636,7 @@ function buildPhrases(blueprint, canonical, compositeScenes, presenterPresent, p
       cur.end = Math.max(cur.end, ln.end);
       cur.coverEnd = Math.max(cur.coverEnd, lineCover);
       cur.clipStart = Math.min(cur.clipStart, lineClipStart);
-      cur.words += wordCount(ln.text);
+      if (ln.shown) cur.lastShownScene = ln.sceneIndex;
     }
     if (ln.shown) cur.shown.add(ln.sceneIndex);
   }
@@ -17685,11 +17765,12 @@ function emitPhraseClip(phrase, voiceNode, env, nodes, out) {
     est_speech_s: Math.round(estSpeechWindowS(phrase.text, phrase.start_s, phrase.end_s) * 100) / 100,
     speech_words: wordCount(phrase.text)
   });
-  registerMergedPhraseSlices(phrase, clipRef, clipStart, genDur, env, out);
+  registerMergedPhraseSlices(phrase, clipRef, clipStart, genDur, Boolean(chained), env, out);
 }
-function registerMergedPhraseSlices(phrase, clipRef, clipStart, genDur, env, out) {
+function registerMergedPhraseSlices(phrase, clipRef, clipStart, genDur, chained, env, out) {
   const shown = [...phrase.shownScenes].sort((a, b) => a - b);
   let r = 0;
+  let firstRegistered = true;
   while (r < shown.length) {
     const first = shown[r];
     let last = first;
@@ -17706,8 +17787,10 @@ function registerMergedPhraseSlices(phrase, clipRef, clipStart, genDur, env, out
       // the clip hits the whole-clip fast path instead of a needless re-encode + tiny shift.
       offset: rawOffset < 0.05 ? 0 : rawOffset,
       len: Math.max(0.5, runEnd - firstStart),
-      clipDur: genDur
+      clipDur: genDur,
+      ...firstRegistered && chained ? { continuesFrame: true } : {}
     });
+    firstRegistered = false;
     for (let s = first + 1; s <= last; s++) {
       out.sceneSlice.set(s, { clipRef, offset: 0, len: 0, clipDur: genDur, skip: true });
     }
@@ -17865,7 +17948,8 @@ function emitBrollScene(scene, i, isLast, env, nodes, out, prevEndFrame) {
     emitFlashHold(i, scene, env.slots, ctx, lengths, lengths.out, env.outAr, nodes, out.clips);
     return void 0;
   }
-  const first = scene.continues_previous && prevEndFrame ? prevEndFrame : buildFrameRef(
+  const sharesPrevFrame = Boolean(scene.continues_previous && prevEndFrame);
+  const first = sharesPrevFrame && prevEndFrame ? prevEndFrame : buildFrameRef(
     "start",
     scene.start_frame_asset?.url,
     scene.start_frame_prompt,
@@ -17913,15 +17997,16 @@ function emitBrollScene(scene, i, isLast, env, nodes, out, prevEndFrame) {
       out.nativeSegments
     );
   }
-  out.clips.push(clip);
+  out.clips.push(sharesPrevFrame ? { ...clip, continuesFrame: true } : clip);
   return last;
 }
 function emitPresenterSliceClip(i, slice, env, nodes, out) {
   if (slice.skip) return;
+  const cont = slice.continuesFrame ? { continuesFrame: true } : {};
   const normDims = env.genAr !== env.outAr ? canvasDims(env.outAr) : void 0;
   const whole = slice.offset === 0 && Math.abs(slice.len - slice.clipDur) <= 0.05 && !normDims;
   if (whole) {
-    out.clips.push({ ref: slice.clipRef, scene_s: slice.len, out: null });
+    out.clips.push({ ref: slice.clipRef, scene_s: slice.len, out: null, ...cont });
     return;
   }
   nodes.push({
@@ -17930,7 +18015,7 @@ function emitPresenterSliceClip(i, slice, env, nodes, out) {
     inputs: { clip: slice.clipRef },
     params: { args: trimArgs(slice.len, slice.offset, normDims), outputs: { video: { kind: "video", ext: "mp4" } } }
   });
-  out.clips.push({ ref: `$ref:s${i}_seg.video`, scene_s: slice.len, out: null });
+  out.clips.push({ ref: `$ref:s${i}_seg.video`, scene_s: slice.len, out: null, ...cont });
 }
 function buildTimeline(blueprint, slots, opts, nodes) {
   const reuse = opts.frames === "reuse";
@@ -18290,25 +18375,40 @@ function lastSceneEnd(blueprint) {
   for (const s of blueprint.scenes) end = Math.max(end, s.end_s ?? 0);
   return end > 0 ? end : 8;
 }
-function concatArgs(count) {
+function seamDropOps(clips, i, seam) {
+  if (seam === "off") return null;
+  if (seam === "head" && i > 0 && clips[i]?.continuesFrame) return "trim=start_frame=1,setpts=PTS-STARTPTS";
+  if (seam === "tail" && clips[i + 1]?.continuesFrame) return "reverse,trim=start_frame=1,setpts=PTS-STARTPTS,reverse";
+  return null;
+}
+function concatArgs(clips, seam) {
   const inputs = [];
-  let labels = "";
-  for (let i = 0; i < count; i++) {
+  const pre = [];
+  const labels = [];
+  clips.forEach((_, i) => {
     inputs.push("-i", `{{in.c${i}}}`);
-    labels += `[${i}:v]`;
-  }
-  return [...inputs, "-filter_complex", `${labels}concat=n=${count}:v=1:a=0[v]`, "-map", "[v]", "{{out.video}}"];
+    const ops = seamDropOps(clips, i, seam);
+    if (ops) {
+      pre.push(`[${i}:v]${ops}[c${i}]`);
+      labels.push(`[c${i}]`);
+    } else {
+      labels.push(`[${i}:v]`);
+    }
+  });
+  const graph = [...pre, `${labels.join("")}concat=n=${clips.length}:v=1:a=0[v]`].join(";");
+  return [...inputs, "-filter_complex", graph, "-map", "[v]", "{{out.video}}"];
 }
 function clipInputLen(c) {
   return c.scene_s + (c.out?.dur ?? 0);
 }
-function xfadeSpineArgs(clips) {
+function xfadeSpineArgs(clips, seam) {
   const n = clips.length;
   const inputs = [];
   const filt = [];
   for (let i = 0; i < n; i++) {
     inputs.push("-i", `{{in.c${i}}}`);
-    filt.push(`[${i}:v]format=yuv420p,fps=30,setsar=1,settb=AVTB[c${i}]`);
+    const ops = seamDropOps(clips, i, seam);
+    filt.push(`[${i}:v]format=yuv420p,fps=30,setsar=1,settb=AVTB${ops ? `,${ops}` : ""}[c${i}]`);
   }
   let cur = "c0";
   let accLen = clipInputLen(clips[0]);
@@ -18330,13 +18430,13 @@ function xfadeSpineArgs(clips) {
   }
   return [...inputs, "-filter_complex", filt.join(";"), "-map", "[v]", "{{out.video}}"];
 }
-function buildSpine(clips, nodes) {
+function buildSpine(clips, seam, nodes) {
   const inputs = {};
   clips.forEach((c, i) => {
     inputs[`c${i}`] = c.ref;
   });
   const hasTransition = clips.length > 1 && clips.some((c) => c.out);
-  const args = hasTransition ? xfadeSpineArgs(clips) : concatArgs(clips.length);
+  const args = hasTransition ? xfadeSpineArgs(clips, seam) : concatArgs(clips, seam);
   nodes.push({
     id: "spine",
     type: "ffmpeg",
@@ -18374,7 +18474,7 @@ function scaffoldVideoCanvas(input, elementsInput, opts) {
   });
   buildElementSheets(slots, nodes);
   const { clips, voTracks, vo_segments, talking_scenes } = buildTimeline(blueprint, slots, opts, nodes);
-  let videoRef = buildSpine(clips, nodes);
+  let videoRef = buildSpine(clips, opts.seamDedup ?? "head", nodes);
   let videoNode = "spine";
   const overlays = blueprint.scenes.flatMap((s) => s.overlays ?? []);
   const floating = blueprint.scenes.flatMap((s) => s.floating_elements ?? []);
@@ -18616,7 +18716,7 @@ function buildMotionBoard(blueprint) {
   });
 }
 var VIDEO_GUIDE = [
-  "Scaffolded by `baker canvas scaffold-video` \u2014 a runnable reproduction of your reference video, built like an editing timeline. The VOICE is cut at PAUSES, not at visual cuts: each continuous-speech PHRASE is ONE Seedance clip (native lip-sync + audio) re-voiced to one brand voice, so a sentence never breaks mid-word across a cut. Each scene's PICTURE is independent: a scene that SHOWS the speaker slices its window out of the phrase clip; a b-roll cutaway gets its own silent clip (or a still hold for a sub-2s flash) laid over the continuing voice; a pure-voiceover stretch is one ElevenLabs tts read. Every clip gets a CLEAN-PLATE start AND end keyframe (no baked text), RECAST to your dropped reference assets \u2014 Seedance interpolates real in-shot motion between them. Each frame grounds ONLY on its own extracted frame + el_* slots (never another generated frame), so all frames render in PARALLEL (no cross-frame cascade). A SPLIT-SCREEN / PICTURE-IN-PICTURE / KEYED-PRESENTER scene is reproduced as one clip PER REGION, stacked or overlaid (see `metadata.todo.composition`). On-screen text/graphics are a separate HTML overlay layer you paint; audio is the voice + SFX + a ducked music bed, normalized stereo. It is a STARTING POINT, not a locked render: add, delete, reorder, split, merge, or re-time scenes freely (a b-roll cutaway INSIDE a phrase lands at an approximate beat \u2014 nudge it) \u2014 see `metadata.todo.full_flexibility`.",
+  "Scaffolded by `baker canvas scaffold-video` \u2014 a runnable reproduction of your reference video, built like an editing timeline. It is a sequence of clear SHOTS separated at COMPLETE BREAKS (hard cuts): two adjacent presenter shots at a cut are TWO clips, never glued into one take. What stays continuous is the VOICE \u2014 a voiceover narration is ONE read across the b-roll it plays over, and a b-roll CUTAWAY between two on-camera moments leaves the presenter shot continuous with the insert sliced in \u2014 and each person keeps ONE brand voice (all their clips' native audio re-voiced in a single per-speaker pass), so timbre holds across the cuts. A presenter shot is ONE Seedance clip (native lip-sync + audio); a pure-voiceover stretch is one ElevenLabs tts read; a sub-2s flash is a still hold. A single shot too long for one clip splits into takes that share a boundary frame \u2014 the spine drops the duplicated frame (`--seam-dedup head|tail|off`). Every clip gets a CLEAN-PLATE start AND end keyframe (no baked text), RECAST to your dropped reference assets \u2014 Seedance interpolates real in-shot motion between them. Each frame grounds ONLY on its own extracted frame + el_* slots (never another generated frame), so all frames render in PARALLEL (no cross-frame cascade). A SPLIT-SCREEN / PICTURE-IN-PICTURE / KEYED-PRESENTER scene is reproduced as one clip PER REGION, stacked or overlaid (see `metadata.todo.composition`). On-screen text/graphics are a separate HTML overlay layer you paint; audio is the voice + SFX + a ducked music bed, normalized stereo. It is a STARTING POINT, not a locked render: add, delete, reorder, split, merge, or re-time scenes freely \u2014 see `metadata.todo.full_flexibility`.",
   "",
   "WHAT TO DO NEXT:",
   "0. RE-CRAFT THE SCRIPT FIRST (don't clone). This reference already won in-market, but copying a video is much harder than a static: the hook is targeting and may not transfer, and the message must become TRUE for our brand. Work the `metadata.todo.script_recraft` checklist \u2014 for each scene judge its role (hook/body/CTA), decide keep/cut/reorder/replace, and re-author every line for OUR customer's pain + OUR offer. See `references/script-craft.md` (hook/body/CTA framework) and the `meta-ads-playbook` skill. Most of the work lives here.",
@@ -18700,9 +18800,9 @@ function buildVideoTodo(report, overlayCount, floatingCount, opts, blueprint) {
       voice_description: d.voice_description,
       line: d.line
     })),
-    talking_head_note: "PHRASE-NATIVE: a continuous-speech phrase where the speaker is shown is ONE Seedance clip (the full phrase quoted in s<anchor>_clip's prompt + generate_audio) so lips+voice are generated together \u2014 no tts, no veed-lipsync. Scenes that show the speaker slice their window out of that clip (s<i>_seg); edit the phrase line in the s<anchor>_clip prompt to re-author it. A pure-voiceover phrase (speaker never shown) is one ElevenLabs tts read instead.",
-    voice_note: "ONE voice per person: a single voice_select is reused across all that person's phrases (on-camera AND off \u2014 the deconstruct's `voiceover` label folds into the sole presenter). Each presenter phrase's native audio is re-voiced to that brand voice via audio_voice_convert (eleven_multilingual_sts_v2, one convert per phrase, timing preserved so lips stay matched). Set voice_select.voice_id's gender/language to match the creator.",
-    native_timing: "The voice is cut at PAUSES, not at visual cuts, so a sentence spanning a cut stays one continuous read (no mid-word break). The clip is generated long enough for the estimated speech; if a line runs longer than its phrase window the voice continues a beat into the following pause (natural VO continuity). `metadata.video.talking_scenes` carries each phrase's scene_s vs est_speech_s. CAVEAT: a b-roll cutaway INSIDE a phrase lands at an approximate (proportional) time \u2014 Seedance exposes no word timing \u2014 so if a cutaway is off its beat, nudge the scene boundary (it's a starting point).",
+    talking_head_note: "SHOT-NATIVE: a presenter shot is ONE Seedance clip (its line quoted in s<anchor>_clip's prompt + generate_audio) so lips+voice are generated together \u2014 no tts, no veed-lipsync. Two adjacent presenter shots at a hard cut are SEPARATE clips; a cutaway phrase (the presenter on camera, cut to b-roll, back on camera) stays one clip and slices its on-camera windows (s<i>_seg). Edit the line in the s<anchor>_clip prompt to re-author it. A pure-voiceover phrase (speaker never shown) is one ElevenLabs tts read instead.",
+    voice_note: "ONE voice per person: a single voice_select is reused across all that person's shots (on-camera AND off \u2014 the deconstruct's `voiceover` label folds into the sole presenter). Every presenter clip's native audio is extracted and re-voiced to that brand voice through a SINGLE merged audio_voice_convert per speaker (<voice>_conv, eleven_multilingual_sts_v2, timing preserved so lips stay matched) \u2014 so timbre stays consistent across the separate shot clips. Set voice_select.voice_id's gender/language to match the creator.",
+    native_timing: "Clips separate at COMPLETE BREAKS between shots, but the VOICE stays continuous where it should: a voiceover narration is ONE read across the b-roll it plays over, and a cutaway leaves the presenter's read continuous under the insert. Each clip is generated long enough for its estimated speech. `metadata.video.talking_scenes` carries each shot's scene_s vs est_speech_s. CAVEAT: a b-roll cutaway INSIDE a phrase lands at an approximate (proportional) time \u2014 Seedance exposes no word timing \u2014 so if a cutaway is off its beat, nudge the scene boundary (it's a starting point).",
     craft: {
       note: "Production-craft principles that raise every clip's realism. Full rationale: references/video-craft.md (production craft); references/script-craft.md + meta-ads-playbook for the hook/message layer.",
       principles: [
@@ -19011,6 +19111,12 @@ async function materializeReferenceVideo(fileArg2) {
   await writeFile3(dest, bytes);
   return dest;
 }
+function resolveSeamDedup(raw) {
+  if (raw === void 0) return "head";
+  const v = String(raw);
+  if (v === "head" || v === "tail" || v === "off") return v;
+  throw new Error(`--seam-dedup must be "head", "tail", or "off" (got "${v}")`);
+}
 function resolveModels2(args) {
   const pick = (flag, kind, fallback) => args[flag] ? String(args[flag]) : resolveModel2(kind, fallback);
   return {
@@ -19122,6 +19228,10 @@ var scaffoldVideoCommand = defineCommand90({
       type: "boolean",
       description: "Give silent b-roll scenes native diegetic ambient mixed deep under the music bed (off by default)"
     },
+    "seam-dedup": {
+      type: "string",
+      description: `How to dedup the frame two clips SHARE when a long shot is split for length: "head" (default, drop the second clip's first frame), "tail" (drop the first clip's last frame), or "off" (keep both).`
+    },
     "max-scenes": { type: "string", description: "Cap the number of scenes the deconstruct emits" },
     "shot-threshold": {
       type: "string",
@@ -19248,6 +19358,7 @@ var scaffoldVideoCommand = defineCommand90({
       blueprintStylePath: path12.relative(outDir, blueprintStylePath),
       frames,
       ambient: Boolean(args.ambient),
+      seamDedup: resolveSeamDedup(args["seam-dedup"]),
       ...args.aspect ? { aspect: String(args.aspect) } : {},
       ...args.resolution ? { resolution: String(args.resolution) } : {}
     };