@gentbajko/slopify 2.4.0 → 3.0.0

package/dist/slices/rebuild/recipe-audio.js

@@ -4,25 +4,37 @@ import { defaultChunking } from "../narration/chunk.js";
 import { concatArgs } from "../narration/concat.js";
 import { normalizeNarrationText } from "../narration/plan.js";
 import { pronunciationChunks } from "../narration/pronunciation-chunks.js";
-import { narrationParts, pronunciationFutureValues, voiceValues } from "./recipe-audio-parts.js";
+import { aliasFutureValues, narrationParts, pronunciationFutureValues, voiceValues, } from "./recipe-audio-parts.js";
+import { levelRecipe } from "./recipe-loudness.js";
 import { recipe, resourceIdentity, selectedReference, } from "./recipe-model.js";
 import { narrationFileRecipe } from "./recipe-narration-text.js";
+import { pauseValues } from "./recipe-pauses.js";
 import { preparationFuture, preparationTemplate } from "./recipe-preparation.js";
+import { voiceBodyRecipes } from "./recipe-voices.js";
 export function bodyNarrationGroups(context, text) {
-    return text.articleText === null
+    return text.narrationText === null
         ? []
-        : pronunciationChunks(normalizeNarrationText(text.articleText), context.config.chunking ?? defaultChunking, usesPronunciationGlossary(context.config) && text.glossary?.ok ? text.glossary.entries : [], new Set(Object.keys(context.content.narrationOverrides)), context.content.narrationSources);
+        : pronunciationChunks(normalizeNarrationText(text.narrationText), context.config.chunking ?? defaultChunking, usesPronunciationGlossary(context.config) && text.glossary?.ok ? text.glossary.entries : [], new Set(Object.keys(context.content.narrationOverrides)), context.content.narrationSources);
 }
 export function audioRecipes(context, text) {
     const { config, content } = context;
     const recipes = [];
     if (config.sources.audio === "off")
-        return { recipes, mediaFingerprint: null, timeline: [], keys: [] };
+        return { recipes, mediaFingerprint: null, timeline: [], keys: [], levels: [] };
     const prepare = usesNarrationPreparation(config);
     const pronounce = usesPronunciationGlossary(config);
     const narrationFiles = prepare || pronounce;
     let body;
-    let bodyTranscript = text.articleText ?? text.article.fingerprint;
+    let bodyTranscript = text.narrationText ??
+        (text.descriptions.length === 0
+            ? text.article.fingerprint
+            : [
+                "described-v1",
+                text.article.fingerprint,
+                text.descriptions.map((row) => row.fingerprint),
+            ]);
+    let sections;
+    let cards = [];
     if (config.sources.audio === "provide") {
         body = recipe(context, "audio:provided", "audio", {
             kind: "provided",
@@ -32,18 +44,31 @@ export function audioRecipes(context, text) {
         }, [text.article.key], { unresolved: content.provided.audio === undefined });
         recipes.push(body);
     }
+    else if (text.script !== undefined && config.voices !== undefined) {
+        const voiced = voiceBodyRecipes(context, config.voices, text.script, text.glossary);
+        recipes.push(...voiced.preparations, ...voiced.cards, ...voiced.parts);
+        cards = voiced.cards;
+        if (narrationFiles)
+            recipes.push(narrationFileRecipe(context, "body", voiced.parts));
+        recipes.push(voiced.body);
+        body = voiced.body;
+        bodyTranscript = voiced.transcript;
+        sections = voiced.sections;
+    }
     else {
         const groups = bodyNarrationGroups(context, text);
         const parts = [];
         const transcripts = [];
         const futurePronunciation = [];
-        let pending = text.articleText === null;
+        recipes.push(...text.descriptions, ...text.cards);
+        cards = text.cards;
+        let pending = text.narrationText === null;
         for (const { key: logicalKey, text: logicalText } of groups) {
             transcripts.push({
                 original: logicalText,
                 effective: effectiveText(context, logicalKey, logicalText),
             });
-            futurePronunciation.push(...pronunciationFutureValues(context, text.glossary, text.article, logicalKey, logicalText));
+            futurePronunciation.push(...pronunciationFutureValues(context, text.glossary, text.article, logicalKey, logicalText), ...aliasFutureValues(context, logicalKey, logicalText));
             const group = narrationParts(context, logicalKey, logicalText, "body", [text.article.key], recipes, text.glossary);
             if (group.length === 0)
                 pending = true;
@@ -51,10 +76,10 @@ export function audioRecipes(context, text) {
         }
         if (transcripts.some((row) => row.effective !== row.original))
             bodyTranscript = ["narration-transcript-v1", transcripts.map((row) => row.effective)];
-        if (text.articleText === null) {
-            futurePronunciation.push(...pronunciationFutureValues(context, text.glossary, text.article, "audio:body:future", null));
+        if (text.narrationText === null) {
+            futurePronunciation.push(...pronunciationFutureValues(context, text.glossary, text.article, "audio:body:future", null), ...aliasFutureValues(context, "audio:body:future", null));
             if (prepare)
-                recipes.push(preparationFuture(context, "body", text.article));
+                recipes.push(preparationFuture(context, "body", text.article, text.descriptions));
         }
         if (pending && prepare)
             parts.length = 0;
@@ -69,8 +94,15 @@ export function audioRecipes(context, text) {
                     chunkingValues(context),
                     ...(prepare ? [preparationTemplate(context)] : []),
                     ...futurePronunciation,
+                    ...(text.descriptions.length === 0
+                        ? []
+                        : [["described-v1", text.descriptions.map((row) => row.fingerprint)]]),
                 ],
-            }, [text.article.key, ...preparationKeys(recipes, "body")]));
+            }, [
+                text.article.key,
+                ...text.descriptions.map((row) => row.key),
+                ...preparationKeys(recipes, "body"),
+            ]));
         recipes.push(...parts);
         if (narrationFiles)
             recipes.push(narrationFileRecipe(context, "body", parts));
@@ -81,14 +113,30 @@ export function audioRecipes(context, text) {
             values: [
                 parts.map((part) => resourceIdentity(context, part)),
                 parts.length === 1 ? "copy-single-file" : concatArgs("$parts", "$output"),
+                // Pauses between sentences; a paragraph's between pieces cut at paragraphs.
+                ...pauseValues(config, parts.map((_part, at) => at === parts.length - 1
+                    ? "end"
+                    : (config.chunking ?? defaultChunking).mode === "paragraph"
+                        ? "paragraph"
+                        : "sentence")),
             ],
-        }, parts.map((part) => part.key), { unresolved: text.articleText !== null && groups.length === 0 });
+        }, parts.map((part) => part.key), { unresolved: text.narrationText !== null && groups.length === 0 });
         recipes.push(body);
     }
     const ordered = [];
+    const levels = [];
+    const level = (join, segment) => {
+        // An uploaded narration is the user's own file: only the exports' master reaches it.
+        const levelled = config.sources.audio === "generate" ? levelRecipe(context, join, segment) : undefined;
+        if (levelled === undefined)
+            return;
+        recipes.push(levelled);
+        levels.push(levelled);
+    };
     for (const category of ["intro", "body", "outro"]) {
         if (category === "body") {
             ordered.push({ value: body, transcript: bodyTranscript });
+            level(body, "body");
             continue;
         }
         const entry = text.entries[category];
@@ -116,6 +164,7 @@ export function audioRecipes(context, text) {
                     voiceValues(context),
                     ...(prepare ? [preparationTemplate(context)] : []),
                     ...pronunciationFutureValues(context, text.glossary, text.article, logicalKey, entry.text),
+                    ...aliasFutureValues(context, logicalKey, entry.text),
                 ],
             }, [...dependencies, ...preparationKeys(recipes, category)], invalidGlossary === undefined ? {} : { unresolved: true, refusal: invalidGlossary }));
         recipes.push(...parts);
@@ -126,9 +175,11 @@ export function audioRecipes(context, text) {
             values: [
                 parts.map((part) => resourceIdentity(context, part)),
                 parts.length === 1 ? "copy-single-file" : concatArgs("$parts", "$output"),
+                ...pauseValues(config, parts.map((_part, at) => (at === parts.length - 1 ? "end" : "sentence"))),
             ],
         }, parts.map((part) => part.key));
         recipes.push(audio);
+        level(audio, category);
         if (narrationFiles)
             recipes.push(narrationFileRecipe(context, category, parts));
         ordered.push({
@@ -158,6 +209,9 @@ export function audioRecipes(context, text) {
         ]),
         timeline,
         keys: ordered.map(({ value }) => value.key),
+        levels,
+        ...(sections === undefined ? {} : { sections }),
+        ...(cards.length === 0 ? {} : { cards }),
     };
 }
 function effectiveText(context, key, original) {

package/dist/slices/rebuild/recipe-build.js

@@ -1,7 +1,12 @@
+import { referenceForThumbnail } from "../admission/rules.js";
 import { audioRecipes } from "./recipe-audio.js";
 import { documentRecipes } from "./recipe-document.js";
 import { editPlan } from "./recipe-edit.js";
 import { exportRecipes } from "./recipe-exports.js";
+import { masterPlan } from "./recipe-loudness.js";
+import { resourceIdentity } from "./recipe-model.js";
+import { imageReference, referenceRecipe } from "./recipe-reference.js";
+import { withReviews } from "./recipe-reviews.js";
 import { shortsRecipes } from "./recipe-shorts.js";
 import { textRecipes } from "./recipe-text.js";
 import { thumbnailRecipes, visualAssets, visualRecipes } from "./recipe-visual.js";
@@ -11,16 +16,21 @@ export function buildRecipes(context) {
     const audio = audioRecipes(context, text);
     const exports = exportRecipes(context, audio);
     const captions = exports.find((value) => value.key === "subtitles:files");
-    const thumbnail = thumbnailRecipes(context, text.recipes);
+    const timing = exports.find((value) => value.key === "subtitles:timing");
+    // The establishing image, when it is on: every other image is drawn from it.
+    const reference = referenceRecipe(context);
+    const drawnFrom = imageReference(context, reference);
+    const thumbnail = thumbnailRecipes(context, text.recipes, referenceForThumbnail(context.config) ? drawnFrom : undefined);
     const youtube = youtubeRecipes(context, exports);
-    return [
+    return withReviews(context, [
         ...text.recipes,
         ...audio.recipes,
         ...exports,
         ...youtube,
-        ...shortsRecipes(context, exports),
+        ...(reference === undefined ? [] : [reference]),
+        ...shortsRecipes(context, exports, drawnFrom, audio.cards ?? [], masterPlan(context, audio.levels, "video")),
         ...thumbnail,
         ...documentRecipes(context, text, thumbnail),
-        ...visualAssets(context, visualRecipes(context.config, context.content, audio.mediaFingerprint, captions?.fingerprint ?? null, (images) => editPlan(context, exports, youtube, images))),
-    ];
+        ...visualAssets(context, visualRecipes(context.config, context.content, audio.mediaFingerprint, captions?.fingerprint ?? null, (images) => editPlan(context, exports, youtube, images, audio.cards ?? []), drawnFrom, timing === undefined ? null : resourceIdentity(context, timing), masterPlan(context, audio.levels, "video"))),
+    ]);
 }

package/dist/slices/rebuild/recipe-cast.js

@@ -0,0 +1,18 @@
+import { castMentions } from "../channels/cast-match.js";
+// The cast members an image request is drawn with: the ones its texts mention, at most
+// `castMembersPerImage` of them in the order they are first mentioned. Undefined when none is,
+// so the request - and its fingerprint - is exactly what it was before the channel had a cast.
+// A member's name and description go with its pictures, since the note that introduces each
+// picture to the provider names them.
+export const castMembersPerImage = 4;
+export function castFor(config, ...texts) {
+    const text = texts.filter((value) => typeof value === "string" && value !== "").join("\n");
+    const mentioned = castMentions(text, config.cast).slice(0, castMembersPerImage);
+    return mentioned.length === 0
+        ? undefined
+        : mentioned.map((member) => ({
+            name: member.name,
+            description: member.description,
+            images: member.images,
+        }));
+}

package/dist/slices/rebuild/recipe-checkpoints.js

@@ -1,5 +1,6 @@
 import { fingerprint } from "../../kernel/runner/work.js";
 import { checkpointClosure, checkpointFingerprint } from "../checkpoints/fingerprint.js";
+import { imageCountsOf } from "../images/scale.js";
 import { buildRecipes } from "./recipe-build.js";
 export function reviewCheckpointSet(runs, catalogue, attachments, fontHash) {
     return runs.flatMap((run, runIndex) => {
@@ -11,15 +12,16 @@ export function reviewCheckpointSet(runs, catalogue, attachments, fontHash) {
                 id,
                 { source: "provide", assetId: id, prompt: null, templateKey: null },
             ])
-            : config.imagePrompts.flatMap((prompt, index) => Array.from({ length: prompt.number }, (_, number) => [
-                `${index}:${number}`,
-                {
-                    source: "generate",
-                    assetId: null,
-                    prompt: run.rendered[`imagePrompts.${index}`] ?? null,
-                    templateKey: `imagePrompts.${index}`,
-                },
-            ]));
+            : // The count the run's first revision plans (`revisions/adopt-content.ts`).
+                imageCountsOf(config).flatMap((count, index) => Array.from({ length: count }, (_, number) => [
+                    `${index}:${number}`,
+                    {
+                        source: "generate",
+                        assetId: null,
+                        prompt: run.rendered[`imagePrompts.${index}`] ?? null,
+                        templateKey: `imagePrompts.${index}`,
+                    },
+                ]));
         const content = {
             articleMarkdown: config.sources.article === "provide" ? config.provided.article : undefined,
             provided: { audio: config.provided.audio, thumbnail: config.provided.thumbnail },

package/dist/slices/rebuild/recipe-describe.js

@@ -0,0 +1,93 @@
+import { usesDescribedNarration, usesFigureCards, usesShorts } from "../admission/rules.js";
+import { describedBlocks, narrationBlocks, spokenNarration, } from "../narration/blocks.js";
+import { describeMessages, spokenPassage } from "../narration/describe.js";
+import { recipe } from "./recipe-model.js";
+import { llmInput, matchingText, renderedPrompt } from "./recipe-text.js";
+// "Describe tables and figures in the narration": one text-model step per table, figure,
+// diagram, equation or code block (`narration:describe:<n>`, in reading order), whose answer
+// is said in the block's place. The narration is chunked, prepared and spoken from that
+// spoken text, so the captions and the word timing follow what is said; the article, the
+// PDF and the reading view keep the block itself.
+//
+// Off (or absent) plans nothing here and the narration is the flattened article, so a project
+// saved before the setting keeps every fingerprint.
+export const describeFutureKey = "narration:describe:future";
+export function describing(context) {
+    return usesDescribedNarration(context.config);
+}
+// What a description request is made of before its block is known: the instruction family,
+// the model and the style guidance. The future step's template, so a change to any of them
+// reaches the narration still to be written.
+export function describeTemplate(context) {
+    const input = llmInput(context, describeMessages({ kind: "table", index: 1, source: "", section: null }, renderedPrompt(context, "narration"), context.config.language));
+    return [
+        "narration-description-v1",
+        JSON.parse(JSON.stringify({ ...input, thinkingConfig: null })),
+        context.config.audio?.skipCode === true,
+    ];
+}
+// The step that stands in for the descriptions while the text they describe is unwritten.
+export function describeFuture(context, dependency) {
+    return recipe(context, describeFutureKey, "audio", {
+        kind: "deferred",
+        version: 1,
+        operation: "narration-description",
+        template: [dependency.fingerprint, describeTemplate(context)],
+    }, [dependency.key]);
+}
+// The Markdown's described blocks as steps, numbered from `offset + 1`, and the spoken text
+// once every one of them has answered.
+export function describedText(context, markdown, dependsOn, offset = 0) {
+    const { config } = context;
+    const blocks = narrationBlocks(markdown, {
+        code: config.audio?.skipCode === true ? "skip" : "describe",
+        language: config.language,
+    });
+    const style = renderedPrompt(context, "narration");
+    const passages = new Map();
+    const cards = [];
+    const recipes = describedBlocks(blocks).map((block) => {
+        const index = offset + block.index;
+        const value = recipe(context, `narration:describe:${String(index)}`, "audio", {
+            ...llmInput(context, describeMessages(block, style, config.language)),
+            describe: {
+                kind: block.kind,
+                index,
+                ...(block.image === undefined ? {} : { image: block.image }),
+            },
+        }, dependsOn);
+        const answer = matchingText(context, value, "text");
+        passages.set(block.index, answer === null ? null : spokenPassage(answer));
+        if (usesFigureCards(config))
+            cards.push(cardRecipe(context, block, index, dependsOn));
+        return value;
+    });
+    return {
+        recipes,
+        cards,
+        spoken: spokenNarration(blocks, (block) => passages.get(block.index) ?? null),
+    };
+}
+// The card the video shows while block `index` is described: drawn locally from the block in
+// the brand kit's title font and colour (the 3.0 look otherwise), in the video's format, and
+// upright too when a 16:9 project makes Shorts. Regenerated with its own token.
+function cardRecipe(context, block, index, dependsOn) {
+    const { config } = context;
+    const formats = config.format === "16:9" && usesShorts(config) ? ["16:9", "9:16"] : [config.format];
+    return recipe(context, `figure:card:${String(index)}`, "images", {
+        kind: "local",
+        version: 1,
+        operation: "figure-card-v1",
+        values: {
+            kind: block.kind,
+            index,
+            source: block.source,
+            section: block.section,
+            image: block.image ?? null,
+            formats,
+            fontId: config.titleStyle?.fontId ?? config.subtitles?.fontId ?? "default",
+            color: config.titleStyle?.color ?? null,
+            language: config.language ?? "en",
+        },
+    }, dependsOn);
+}

package/dist/slices/rebuild/recipe-edit.js

@@ -15,7 +15,7 @@ import { recipe, resourceIdentity, selectedReference, } from "./recipe-model.js"
 // that step runs) has landed, so until then one deferred `animate:future` stands in for the
 // requests and carries their cost, the way `shorts:future` does.
 export const animateFutureKey = "animate:future";
-export function editPlan(context, exports, youtube, images) {
+export function editPlan(context, exports, youtube, images, cards = []) {
     const { config } = context;
     const edit = videoEditOf(config);
     if (config.sources.video !== "generate")
@@ -38,6 +38,13 @@ export function editPlan(context, exports, youtube, images) {
         values.push(["look-v1", edit.vignette, edit.grain, edit.grade, edit.atmosphere]);
     if (usesChapterCards(config))
         values.push(["cards-v1", config.subtitles?.fontId ?? "default"]);
+    // The brand kit's end screen and title style; nothing when unset, so a project without them
+    // keeps its export fingerprint.
+    const endScreen = config.endScreen?.text.trim() ?? "";
+    if (endScreen !== "")
+        values.push(["end-screen-v1", endScreen]);
+    if (config.titleStyle !== undefined && (usesChapterCards(config) || endScreen !== ""))
+        values.push(["title-style", config.titleStyle.fontId ?? null, config.titleStyle.color ?? null]);
     if (timing !== undefined) {
         values.push(["timing", resourceIdentity(context, timing)]);
         dependsOn.push(timing.key);
@@ -48,6 +55,11 @@ export function editPlan(context, exports, youtube, images) {
         else
             values.push(["chapters", "headings", headingsIdentity(context)]);
     }
+    // "Show tables and figures on screen": each card, placed where its description is spoken.
+    if (cards.length > 0) {
+        values.push(["figure-cards-v1", cards.map((value) => resourceIdentity(context, value))]);
+        dependsOn.push(...cards.map((value) => value.key));
+    }
     const recipes = usesAnimation(config)
         ? animateRecipes(context, images, timing, fromDescription ? description : undefined)
         : [];

package/dist/slices/rebuild/recipe-exports.js

@@ -1,12 +1,21 @@
+import { projectLanguage } from "../../kernel/ports/languages.js";
 import { usesShorts, usesYoutubeDescription } from "../admission/rules.js";
+import { usesShortMode } from "../admission/short-mode.js";
+import { reviewsNarration } from "../reviews/rules.js";
+import { timingOperation } from "../subtitles/model.js";
 import { audioExportArgs } from "../video/audio-export-args.js";
 import { editNeedsTiming } from "../video/edit-settings.js";
+import { usesVoices } from "../voices/model.js";
+import { portraitValues } from "../voices/portraits.js";
+import { masterPlan } from "./recipe-loudness.js";
 import { recipe, resourceIdentity, } from "./recipe-model.js";
 export function exportRecipes(context, audio) {
     const { config, content } = context;
     if (audio.mediaFingerprint === null)
         return [];
     const recipes = [];
+    // Level the volume: the listening files are mastered to the audio files' target.
+    const master = masterPlan(context, audio.levels, "audioFiles");
     if (config.sources.video === "off")
         recipes.push(recipe(context, "export:wav", "video", {
             kind: "local",
@@ -15,30 +24,51 @@ export function exportRecipes(context, audio) {
             values: [
                 audio.mediaFingerprint,
                 audioExportArgs([{ kind: "body", path: "$body", seconds: 0 }], "$output"),
+                ...master.values,
             ],
-        }, audio.keys));
+        }, [...audio.keys, ...master.keys]));
     const captions = config.subtitles !== undefined && config.subtitles.mode !== "off";
-    // The YouTube description's chapters, the shorts' clips and captions, and the video's cuts,
-    // chapter cards and chapter openers use the same word timing, so it runs for them even with
-    // captions off; only the caption files below wait for captions.
+    // The YouTube description's chapters, the shorts' clips and captions, a short's own
+    // captions, and the video's cuts, chapter cards and chapter openers use the same word timing,
+    // so it runs for them even with captions off; only the caption files below wait for captions.
+    const voices = usesVoices(config) ? config.voices : undefined;
     if (!captions &&
         !usesYoutubeDescription(config) &&
         !usesShorts(config) &&
-        !editNeedsTiming(config))
+        !usesShortMode(config) &&
+        !editNeedsTiming(config) &&
+        !reviewsNarration(config) &&
+        voices?.audioFiles !== true)
         return recipes;
     const timing = recipe(context, "subtitles:timing", "video", {
         kind: "local",
         version: 1,
-        operation: "wav2vec2-en-a19f851-v2-omissions",
+        operation: timingOperation(projectLanguage(config)),
         // The lead-in moves every word, so the edge silence is part of the timing.
         values: [
             audio.timeline,
             config.silenceGapSeconds,
-            config.subtitles?.language ?? "en",
+            // The project language; an English project reads "en" here as it always did.
+            config.language ?? config.subtitles?.language ?? "en",
             config.edgeSilenceSeconds,
+            // Each word learns its speaker and turn on a multi-voice run.
+            ...(voices === undefined ? [] : ["voice-words-v1"]),
         ],
     }, audio.keys);
     recipes.push(timing);
+    if (voices?.audioFiles === true)
+        recipes.push(recipe(context, "voices:files", "video", {
+            kind: "local",
+            version: 1,
+            operation: "audio-files-v1",
+            values: [
+                audio.mediaFingerprint,
+                resourceIdentity(context, timing),
+                (audio.sections ?? []).map((section) => [section.title, section.firstTurn]),
+                config.title,
+                ...master.values,
+            ],
+        }, [...audio.keys, timing.key, ...master.keys], { unresolved: audio.sections === undefined }));
     if (config.subtitles === undefined || config.subtitles.mode === "off")
         return recipes;
     const cues = content.subtitleCues;
@@ -53,7 +83,12 @@ export function exportRecipes(context, audio) {
             kind: "local",
             version: 1,
             operation: "manual-cues-v1",
-            values: [cues.audioFingerprint, cues.cues.map((cue) => ({ ...cue }))],
+            // A cue's speaker only when it has one, so captions edited before speakers were
+            // kept (and every one-voice run's) keep their fingerprint.
+            values: [
+                cues.audioFingerprint,
+                cues.cues.map(({ speaker, ...cue }) => speaker === undefined ? { ...cue } : { ...cue, speaker }),
+            ],
         }, audio.keys);
     recipes.push(cueRecipe);
     recipes.push(recipe(context, "subtitles:files", "video", {
@@ -66,6 +101,11 @@ export function exportRecipes(context, audio) {
             config.subtitles.fontId,
             config.subtitles.fontSize,
             config.subtitles.position,
+            // Only when the brand kit set them, so captions made before keep their fingerprint.
+            ...(config.subtitles.color === undefined && config.subtitles.outlineColor === undefined
+                ? []
+                : [["colours", config.subtitles.color ?? null, config.subtitles.outlineColor ?? null]]),
+            ...(voices === undefined ? [] : [captionSpeakerValues(voices, audio.mediaFingerprint)]),
         ],
     }, [cueRecipe.key]));
     return recipes;
@@ -80,3 +120,17 @@ export function manualCuesNeedReview(context, recipes) {
         cues.audioFingerprint !==
             context.manifest.outputs.find((value) => value.workKey === timing.key && value.state === "ready")?.fingerprint);
 }
+// How the captions show the speakers: names, colours by place, name tags and, for a podcast or
+// interview, the speaker panel, which runs to the end of the narration. The panel's portraits
+// only when a speaker has one; the render reads the caption file's fingerprint, so it follows.
+function captionSpeakerValues(voices, media) {
+    const portraits = portraitValues(voices);
+    return [
+        "voice-captions-v1",
+        voices.format,
+        voices.nameTags,
+        voices.speakers.map((speaker) => [speaker.id, speaker.name.trim()]),
+        media,
+        ...(portraits === undefined ? [] : [["portraits", [...portraits]]]),
+    ];
+}

package/dist/slices/rebuild/recipe-input-schema.js

@@ -1,6 +1,7 @@
 import { z } from "zod";
 import { messageRoles, thinkingModes } from "../../kernel/ports/llm.js";
 import { llmDocumentsSchema } from "../../kernel/ports/llm-documents.js";
+import { describedKinds } from "../narration/blocks.js";
 import { deferredOperations, localOperations } from "./recipe-model.js";
 const value = z.lazy(() => z.union([
     z.null(),
@@ -23,7 +24,9 @@ export const recipeInputSchema = z.discriminatedUnion("kind", [
             .object({
             budget: z.number().finite().optional(),
             level: z.enum(["minimal", "low", "medium", "high"]).optional(),
-            effort: z.enum(["none", "minimal", "low", "medium", "high", "xhigh"]).optional(),
+            effort: z
+                .enum(["none", "minimal", "low", "medium", "high", "xhigh", "max", "ultra"])
+                .optional(),
         })
             .strict()
             .nullable(),
@@ -40,6 +43,22 @@ export const recipeInputSchema = z.discriminatedUnion("kind", [
         })
             .strict()
             .optional(),
+        // A multi-voice run's script or speaker split: the answer is checked as a script.
+        script: z
+            .object({
+            speakers: z.array(z.object({ id: z.string(), name: z.string() }).strict()),
+            attribute: z.boolean(),
+        })
+            .strict()
+            .optional(),
+        describe: z
+            .object({
+            kind: z.enum(describedKinds),
+            index: z.number().int().positive(),
+            image: z.string().optional(),
+        })
+            .strict()
+            .optional(),
     })
         .strict(),
     z
@@ -56,6 +75,19 @@ export const recipeInputSchema = z.discriminatedUnion("kind", [
         segment: z.enum(["body", "intro", "outro"]),
         pronunciation: z.null(),
         wholeRequest: z.boolean().optional(),
+        // A multi-voice run's turn requests; without them every such piece failed to store.
+        speaker: z.string().optional(),
+        turn: z.number().int().optional(),
+        dialogue: z
+            .array(z
+            .object({
+            speaker: z.string(),
+            turn: z.number().int(),
+            voice: z.string(),
+            text: z.string(),
+        })
+            .strict())
+            .optional(),
     })
         .strict(),
     z
@@ -67,6 +99,16 @@ export const recipeInputSchema = z.discriminatedUnion("kind", [
         prompt: z.string(),
         aspect: z.enum(["16:9", "9:16"]),
         animate: z.object({ image: z.string(), seconds: z.number() }).strict().optional(),
+        thinking: z.enum(thinkingModes).optional(),
+        reference: z
+            .object({ fingerprint: z.string(), assetId: z.string().nullable() })
+            .strict()
+            .optional(),
+        cast: z
+            .array(z
+            .object({ name: z.string(), description: z.string(), images: z.array(z.string()) })
+            .strict())
+            .optional(),
     })
         .strict(),
     z

package/dist/slices/rebuild/recipe-legacy.js

@@ -1,4 +1,5 @@
 import { z } from "zod";
+import { referenceKey } from "../admission/model.js";
 import { buildRecipes } from "./recipe-build.js";
 export function baselineFingerprints(project, outputs, pieces, content, stages = []) {
     const recipes = buildRecipes({
@@ -34,6 +35,15 @@ export function legacyOutputWorkKey(output, config) {
             return `image:${output.id}`;
         case "thumbnail":
             return "thumbnail:image";
+        case "script_md":
+            return "script:attribute";
+        case "audio_mp3":
+        case "audio_m4b":
+            return "voices:files";
+        case "audio_levelled":
+            return `level:${output.meta.segment ?? "body"}`;
+        case "reference":
+            return referenceKey;
         case "notes":
             return "research:notes";
         case "article_md":
@@ -68,6 +78,8 @@ export function legacyOutputWorkKey(output, config) {
         // Only ever made by a revision, never by the legacy stages.
         case "animated_image":
             return `animate:${String(output.meta.index ?? 1)}`;
+        case "figure_card":
+            return `figure:card:${String(output.meta.index ?? 1)}`;
         case "subtitles_srt":
         case "subtitles_vtt":
         case "subtitle_ass":