@slatesvideo/shared 0.6.9 → 0.6.11

package/dist/operations/index.js

@@ -1,3 +1,4 @@
+import { MINIMAX_MAX_REFERENCE, minimaxMaxReferenceTokens, MODEL_CAPABILITIES, GPT_QUALITY_TIERS, GPT_BACKGROUNDS } from '../prompts/model-capabilities.js';
 // Operations layer — the ONE place every Slates agent tool is defined.
 // Both the MCP server and the CLI register these as their tool / command
 // surface. Every operation:
@@ -352,11 +353,11 @@ export const VIDEO_MODELS = [
     'veo-3.1-fast',
     'veo-3.1-standard',
     'seedance-2',
-    // Seedance 2.5 is a SECOND SEAT, not a replacement: 30s takes, 30 image
-    // references, audio-only references, up to 1080p (2026-08-24) — but no 4K,
-    // and dearer than 2.0 at every shared tier, so 2.0 stays the default. Its
-    // EDIT row is not here; edit models live on slates_edit_video, same as the
-    // Kling and Omni Flash ones.
+    // Seedance 2.5 is the DEFAULT video model (Eric, 2026-09-13): 30s takes, 30
+    // image references, audio-only references, up to 1080p (2026-08-24). 2.0 stays
+    // beside it as the 4K seat, cheaper at every shared tier. 2.5's EDIT row is
+    // not here; edit models live on slates_edit_video, same as the Kling and Omni
+    // Flash ones.
     'seedance-2.5',
     'omni-flash',
     // MiniMax H3, two seats in one family (2026-08-27). Base H3 is the AUTHORED-
@@ -495,7 +496,7 @@ function editClipBounds(model) {
 const EDIT_VIDEO_RESOLUTIONS = videoResolutionUnion(['seedance-2.5-edit']);
 export const estimateGenerationCost = {
     id: 'slates_estimate_generation_cost',
-    description: 'Pre-flight cost estimate. Call before any generate_* op so the user sees "this will cost N credits" up front. Takes the SAME base model ids as the generate ops (video: "seedance-2" + duration + videoResolution; image: "nano-banana-2" + resolution) — exact registry cost keys also work. Pairs with the confirm gate.',
+    description: 'Quote credits before any generate_* op. Accepts the same base model ids and parameters as generation, or an exact registry cost key. Pairs with the confirm gate.',
     input: z.object({
         model: z.string().describe('Base model id as passed to the generate op (e.g. "seedance-2", "kling-v3.0-std", "nano-banana-2") or an exact registry cost key ("nano-banana-2-2k", "seedance-2-1080p-8s")'),
         quantity: z.number().int().min(1).max(10).optional().describe('Number of generations (default 1)'),
@@ -512,11 +513,13 @@ export const estimateGenerationCost = {
         characters: z.number().int().min(1).max(TTS_MAX_CHARACTERS).optional().describe(`${TTS_MODEL} only — the LENGTH OF THE TEXT to speak (${TTS_BUCKET_CHARS}-char buckets).`),
         videoResolution: zEnum(VIDEO_RESOLUTIONS).optional().describe('Video only. Omitted, each model quotes at its own default. Per-model ladders: see slates_generate_video\'s videoResolution.'),
         resolution: z.enum(['1k', '2k', '3k', '4k']).optional().describe('Image only (default 2k; 3k: GPT Image/seedream-5-lite).'),
-        quality: z.enum(['low', 'medium', 'high', 'xhigh', 'max']).optional().describe('GPT Image tier; default high.'),
+        quality: z.enum(GPT_QUALITY_TIERS).optional().describe('GPT Image tier; default high.'),
         aspectRatio: z.string().optional().describe('Image only. 1:1/4:3/3:4 cost more than 16:9.'),
         sound: z.boolean().optional().describe('Veo only — audio flag changes the cost key.'),
         seedanceFace: z.boolean().optional().describe('Seedance AI-face route (pricier key).'),
         seedanceRealFace: z.boolean().optional().describe('Seedance consented real-face route (premium key).'),
+        videoRefSeconds: z.number().nonnegative().optional().describe('Combined reference-video seconds, measured from the clips.'),
+        audioRefSeconds: z.number().nonnegative().optional().describe('Combined reference-audio seconds, including attached character voices.'),
         referenceImages: z.number().int().min(0).optional().describe(`MiniMax H3 rows only — how many reference IMAGES the generation will carry. Each image past the row's free allowance is a paid dimension of the cost key, so a quote that omits this UNDER-REPORTS a reference-heavy job. The allowances differ: minimax-h3 gives ${MINIMAX_FREE_REF_IMAGES_BY_MODEL['minimax-h3']} free, minimax-h3-max gives ${MINIMAX_FREE_REF_IMAGES_BY_MODEL['minimax-h3-max']}. Ignored by every other model. Start/end frames are free on both rows and are not reference images.`),
     }),
     async run(input, ctx) {
@@ -651,9 +654,11 @@ export const estimateGenerationCost = {
                     seedanceFace: input.seedanceFace ?? resolved.seedanceFace,
                     seedanceRealFace: input.seedanceRealFace,
                     referenceImages: input.referenceImages ?? resolved.referenceImages,
+                    videoRefSeconds: input.videoRefSeconds, audioRefSeconds: input.audioRefSeconds,
                 });
             }
         }
+        await loadDynamicPrice(ctx, byKey, key);
         const perCredits = key != null ? byKey.get(key) : undefined;
         if (key == null || perCredits == null) {
             // Every id in the error comes from the SSOT arrays. The image half was
@@ -1337,18 +1342,8 @@ async function previewAssets(ctx, refs) {
     return out;
 }
 /** The exact `model` ids `slates_generate_image` accepts. */
-/**
- * The image models whose ladder includes the 3k (1440p) class — a MIRROR of
- * `imageResolutions` in slate's MODEL_REGISTRY, which this package cannot read.
- * Exported so `pricing-consistency-check.mjs` can prove the mirror still
- * matches; without that proof a model that gains 3k in the registry just goes
- * quietly unreachable through the op.
- */
-export const THREE_K_IMAGE_MODELS = [
-    'gpt-image-2-5-flare',
-    'gpt-image-2-5-sunburst',
-    'seedream-5-lite',
-];
+/** Derived compatibility export for callers enumerating the 3k ladder. */
+export const THREE_K_IMAGE_MODELS = Object.keys(MODEL_CAPABILITIES).filter((id) => MODEL_CAPABILITIES[id].imageResolutions?.includes('3k'));
 export const IMAGE_MODELS = [
     'nano-banana-2',
     'nano-banana-2-lite',
@@ -1443,8 +1438,8 @@ export const generateImage = {
         model: zEnum(IMAGE_MODELS).optional().describe('Image model. Default nano-banana-2. Routing doctrine: slates-model-selection skill. All except nano-banana-2 require projectId.'),
         projectId: z.string().uuid().optional().describe('Save into this Slates project. Renderer refreshes live. Required for every model except nano-banana-2.'),
         resolution: z.enum(['1k', '2k', '3k', '4k']).optional().describe('1k drafts, 2k hero, 4k final. nano-banana-2-lite: 1k only. GPT Image classes 1024²/1080p/1440p/2160p. Never default this.'),
-        quality: z.enum(['low', 'medium', 'high', 'xhigh', 'max']).optional().describe('GPT Image only. UNEVEN ladder: max=4× high, xhigh~1.8×. medium drafts; default high.'),
-        backgroundMode: z.enum(['auto', 'transparent', 'opaque']).optional().describe('GPT Image only. transparent = alpha channel. Free.'),
+        quality: z.enum(GPT_QUALITY_TIERS).optional().describe('GPT Image only. UNEVEN ladder: max=4× high, xhigh~1.8×. medium drafts; default high.'),
+        backgroundMode: z.enum(GPT_BACKGROUNDS).optional().describe('GPT Image only. transparent = alpha channel. Free.'),
         aspectRatio: zEnum(IMAGE_ASPECT_RATIOS).optional().describe(`Pick from the use case: cinematic 16:9 · TikTok/Reels 9:16 · IG square 1:1 · ultra-wide 21:9. 1:1 costs most on GPT Image. Per model: ${describeAspectRatios(IMAGE_MODELS)}`),
         count: z.number().int().min(1).max(10).optional().describe('Up to 10 with projectId; headless caps at 4.'),
         referenceImageUrls: z.array(z.string().url()).max(14).optional().describe('Headless (no projectId) nano-banana-2 only. With a projectId, upload via slates_upload_reference_image. Label every image role in the prompt.'),
@@ -1481,25 +1476,10 @@ export const generateImage = {
         }
         const resolution = input.resolution;
         const imageModel = input.model ?? 'nano-banana-2';
-        // 🚨 SEEDREAM HAS A 3k CLASS TOO, AND THIS GUARD USED TO DENY IT.
-        // It read `!== 'gpt-image-2'` and rejected every other model at 3k — but
-        // `seedream-5-lite` declares ['2k','3k','4k'] in the desktop registry and
-        // has a real `seedream-5-lite` cost key, so the op was refusing a request
-        // the desktop would have served. Pre-existing; found by the 2026-09-09
-        // audit, not introduced by the 2.5 swap.
-        //
-        // The ladder itself is owned by MODEL_REGISTRY in slate/src/shared/pricing.ts
-        // and is not readable from here, so THREE_K_IMAGE_MODELS is a MIRROR — declared
-        // and exported below so `pricing-consistency-check.mjs` compares it against
-        // the desktop registry's own `imageResolutions`. It used to be an inline
-        // literal with a comment admitting nothing checked it, which is how it came
-        // to deny `seedream-5-lite` a class the desktop had always served.
-        if (resolution === '3k' && !THREE_K_IMAGE_MODELS.includes(imageModel)) {
-            return ok({
-                requires_clarification: true,
-                missing: ['resolution'],
-                message: `3k (1440p) exists on ${THREE_K_IMAGE_MODELS.join(', ')} — pick 1k/2k/4k for ${imageModel}.`,
-            });
+        const resolutions = MODEL_CAPABILITIES[imageModel]?.imageResolutions ?? [];
+        if (resolution && !resolutions.includes(resolution)) {
+            return ok({ requires_clarification: true, missing: ['resolution'],
+                message: `${imageModel} accepts ${resolutions.join(', ')}.` });
         }
         // Only nano-banana-2 has a headless path — everything else routes through
         // the desktop generation pipeline, which needs a project.
@@ -1571,7 +1551,8 @@ export const generateImage = {
         const costKey = imageCostKey(imageModel, resolution, input.quality, input.aspectRatio ?? '1:1');
         const cloud = ctx.cloud();
         const registry = await cloud.get('/api/agent/models');
-        const entry = registry.models.find((m) => m.model === costKey);
+        const entry = registry.models.find((m) => m.model === costKey) ??
+            (await cloud.get(`/api/agent/models?costKey=${encodeURIComponent(costKey)}`)).models[0];
         if (!entry)
             throw new Error(`Model not in registry: ${costKey}`);
         const totalCents = creditCost(entry) * (input.count ?? 1);
@@ -1730,6 +1711,17 @@ export const generateImage = {
             params: {
                 prompt: input.prompt,
                 aspect_ratio: input.aspectRatio ?? '1:1',
+                // 🚨 THE RESOLUTION HAS TO BE ON THE WIRE, BECAUSE IT IS IN THE KEY.
+                // `imageCostKey` above bills `nano-banana-2-{resolution}`, and until
+                // 2026-09-10 this body omitted the field entirely — so fal rendered at
+                // its own default while we charged for whatever rung the caller asked
+                // for. The enum is fal's (`1K`/`2K`/`4K`), the same one the desktop's
+                // `buildFalNB2Request` sends, and the proxy now recovers the rung from it
+                // rather than trusting the key (`lib/fal-image-keys.ts`).
+                // `3k` cannot reach here — the capability guard above refuses a rung
+                // nano-banana-2 does not declare — so the map is deliberately partial
+                // rather than carrying a rung this model has no key for.
+                resolution: { '1k': '1K', '2k': '2K', '4k': '4K' }[resolution] ?? '2K',
                 // 🚨 THE HEADLESS PATH IS THE ONE PLACE WE ASK FAL FOR A BATCH, so it
                 // is the one place a provider's own `num_images` ceiling binds — and
                 // Nano Banana's is 4 (fal schema, read 2026-09-09), against the op's
@@ -1737,6 +1729,9 @@ export const generateImage = {
                 // separate single-image generations, where no batch ceiling exists.
                 // Guarded above rather than clamped here: silently making 4 when 10
                 // were asked for would bill 4 and look like a partial failure.
+                //
+                // The proxy bills this COUNT (it multiplies the single-image key by it),
+                // which is what the confirm gate above has always quoted.
                 num_images: input.count ?? 1,
                 ...(hasReferenceImages
                     ? { image_urls: input.referenceImageUrls }
@@ -1815,8 +1810,8 @@ export const editImage = {
         editModel: z.enum(['nano-banana-2', 'nano-banana-2-lite', 'nano-banana-pro', 'gpt-image-2-5-flare', 'gpt-image-2-5-sunburst', 'flux-2-max', 'seedream-5-lite']).optional(),
         referenceAssetIds: z.array(z.string().uuid()).max(13).optional().describe('Nano-Banana only (NB Pro 13, NB2 Lite 3).'),
         resolution: z.enum(['1k', '2k', '3k', '4k']).optional().describe('3k = GPT Image/seedream-5-lite; nano-banana-2-lite is 1k only.'),
-        quality: z.enum(['low', 'medium', 'high', 'xhigh', 'max']).optional().describe('GPT Image tier; default high.'),
-        backgroundMode: z.enum(['auto', 'transparent', 'opaque']).optional().describe('GPT Image only. transparent = alpha channel. Free.'),
+        quality: z.enum(GPT_QUALITY_TIERS).optional().describe('GPT Image tier; default high.'),
+        backgroundMode: z.enum(GPT_BACKGROUNDS).optional().describe('GPT Image only. transparent = alpha channel. Free.'),
         aspectRatio: z.string().optional(),
         confirm: z.boolean().optional().describe('Set true to bypass the confirm gate.'),
         background: z.boolean().optional().describe(BACKGROUND_DESCRIBE),
@@ -1847,7 +1842,8 @@ export const editImage = {
             : imageCostKey(editModel, resolution, input.quality, input.aspectRatio);
         const cloud = ctx.cloud();
         const registry = await cloud.get('/api/agent/models');
-        const entry = registry.models.find((m) => m.model === costKey);
+        const entry = registry.models.find((m) => m.model === costKey) ??
+            (await cloud.get(`/api/agent/models?costKey=${encodeURIComponent(costKey)}`)).models[0];
         if (!entry)
             throw new Error(`Model not in registry: ${costKey}`);
         const totalCents = creditCost(entry);
@@ -1991,6 +1987,11 @@ export function videoCostKey(input) {
     // Mirrors minimaxCreditKey() in slate/src/shared/pricing.ts.
     if (MINIMAX_MODELS.has(input.model)) {
         const res = input.videoResolution ?? defaultVideoResolutionFor(input.model);
+        if (input.model === 'minimax-h3-max') {
+            const tokens = minimaxMaxReferenceTokens({ imagePixels: (input.referenceImages ?? 0) * MINIMAX_MAX_REFERENCE.normalizedImageEdge ** 2,
+                videoSeconds: input.videoRefSeconds ?? 0, audioSeconds: input.audioRefSeconds ?? 0, resolution: res ?? '' });
+            return `${input.model}-${res}-${input.duration}s${tokens ? `-rt${tokens}` : ''}`;
+        }
         const k = minimaxRefSurchargeCount(input.model, input.referenceImages);
         return `${input.model}-${res}-${input.duration}s${k > 0 ? `-ref${k}` : ''}`;
     }
@@ -2214,9 +2215,10 @@ function resolveVideoModel(raw) {
         'kling-v3.0-omni-pro': 'kling-v3.0-omni',
         'seedance-2.0': 'seedance-2',
         'seedance-2-0': 'seedance-2',
-        // ⚠️ The 2.5 spellings must resolve to 2.5, and the BARE `seedance` must keep
-        // resolving to 2.0 — 2.0 is the default video model and holds 1080p/4K, which
-        // 2.5 does not have at all.
+        // ⚠️ The 2.5 spellings must resolve to 2.5, and the BARE `seedance` keeps
+        // resolving to 2.0: every published CLI and MCP build that sends it expects
+        // the 4K seat, and 2.5 rejects 4K. The default video model is 2.5 (tier in
+        // MODEL_FACTS, 2026-09-13); an agent names it as `seedance-2.5`.
         'seedance-2.5': 'seedance-2.5',
         'seedance-25': 'seedance-2.5',
         'seedance-2-5': 'seedance-2.5',
@@ -2355,9 +2357,9 @@ export const generateVideo = {
         // The capacity sentences are DERIVED from MODEL_FACTS (see
         // multimodalRefSummary) rather than hand-typed, so a cap change in one
         // place cannot leave a stale number in a description an LLM reads.
-        videoReferenceAssetIds: z.array(z.string()).optional().describe(`Reference VIDEOS, cited in the prompt as "video 1", "video 2"… in the order given. ${multimodalRefModels().join(' / ')} only. ${multimodalRefSummary('seedance-2')} ${multimodalRefSummary('seedance-2.5')} Billing switches to the vref key (input+output seconds) — pass videoReferenceSecondsEach. Over the cap is REFUSED, never trimmed.`),
+        videoReferenceAssetIds: z.array(z.string()).optional().describe(`Reference VIDEOS, cited in the prompt as "video 1", "video 2"… in the order given. ${multimodalRefModels().join(' / ')} only; per-model caps are on audioReferenceAssetIds. Billing switches to the vref key (input+output seconds) — pass videoReferenceSecondsEach. Over the cap is REFUSED, never trimmed.`),
         videoReferenceSecondsEach: z.array(z.number()).optional().describe('REQUIRED with videoReferenceAssetIds, same order and length: each clip\'s duration in seconds. Feeds the vref cost key; the server re-probes and corrects an understated value upward.'),
-        audioReferenceAssetIds: z.array(z.string()).optional().describe(`Reference AUDIO, cited as "audio 1", "audio 2"… in the order given. No billing surcharge. ${multimodalRefSummary('seedance-2')} ${multimodalRefSummary('seedance-2.5')}`),
+        audioReferenceAssetIds: z.array(z.string()).optional().describe(`Reference AUDIO, cited as "audio 1", "audio 2"… in the order given. ${multimodalRefModels().join(' / ')} only. No billing surcharge. ${multimodalRefModels().map(multimodalRefSummary).join(' ')}`),
         audioReferenceSpokenText: z.array(z.string()).optional().describe('The exact words in each reference clip — same order and length as audioReferenceAssetIds, "" for a clip with no speech. The model RE-TRANSCRIBES a take rather than using it verbatim, so audio decides voice/accent/timing and only this decides the WORDS. Omit it and the words are a guess.'),
         sound: z.boolean().optional().describe('Kling Omni / Veo / Seedance: enable audio generation. Default true.'),
         audioLanguage: z.enum(['EN', 'ZH', 'JA', 'KO', 'ES']).optional().describe('Kling Omni only — language for dialogue.'),
@@ -2681,25 +2683,27 @@ export const generateVideo = {
         }
         const cloud = ctx.cloud();
         const registry = await cloud.get('/api/agent/models');
-        const costKey = videoCostKey({
-            model: input.model,
-            duration: input.duration,
-            videoResolution: input.videoResolution,
-            sound: input.sound,
-            seedanceFace: input.seedanceFace,
-            seedanceRealFace: input.seedanceRealFace,
-            // MiniMax H3 base: reference images past the fifth are a PAID key
-            // dimension. Counted from the same four arrays the request sends, so the
-            // quote and the server's own re-derivation see the same number.
-            referenceImages: minimaxRefImageCount(input),
-            // Σ ceil(d - 0.05) over every reference clip, both shapes — the same
-            // expression the desktop's estimateCost and the handler's key builder
-            // use. Quoting only the singular would understate a multi-clip call.
-            videoRefSeconds: (input.videoReferenceAssetId && input.videoReferenceSeconds
-                ? Math.ceil(input.videoReferenceSeconds - 0.05)
-                : 0) +
-                (input.videoReferenceSecondsEach ?? []).reduce((n, d) => n + (d > 0 ? Math.ceil(d - 0.05) : 0), 0),
-        });
+        const costKey = input.model === 'minimax-h3-max'
+            ? (await ctx.desktop().get('/agent/generation/video-quote', { request: JSON.stringify(input) })).costKey
+            : videoCostKey({
+                model: input.model,
+                duration: input.duration,
+                videoResolution: input.videoResolution,
+                sound: input.sound,
+                seedanceFace: input.seedanceFace,
+                seedanceRealFace: input.seedanceRealFace,
+                // MiniMax H3 base: reference images past the fifth are a PAID key
+                // dimension. Counted from the same four arrays the request sends, so the
+                // quote and the server's own re-derivation see the same number.
+                referenceImages: minimaxRefImageCount(input),
+                // Σ ceil(d - 0.05) over every reference clip, both shapes — the same
+                // expression the desktop's estimateCost and the handler's key builder
+                // use. Quoting only the singular would understate a multi-clip call.
+                videoRefSeconds: (input.videoReferenceAssetId && input.videoReferenceSeconds
+                    ? Math.ceil(input.videoReferenceSeconds - 0.05)
+                    : 0) +
+                    (input.videoReferenceSecondsEach ?? []).reduce((n, d) => n + (d > 0 ? Math.ceil(d - 0.05) : 0), 0),
+            });
         // Hard consent gate, checked before any spend: the real-face route is
         // consent-attested by design (the desktop enforces it too).
         if (input.seedanceRealFace && !input.realFaceConsent) {
@@ -2709,7 +2713,8 @@ export const generateVideo = {
                 message: 'Real-person generation needs consent: confirm with the user that they hold the rights/consent to this likeness, then retry with realFaceConsent=true.',
             });
         }
-        const entry = registry.models.find((m) => m.model === costKey);
+        const entry = registry.models.find((m) => m.model === costKey) ??
+            (await cloud.get(`/api/agent/models?costKey=${encodeURIComponent(costKey)}`)).models[0];
         if (!entry) {
             throw new Error(`Model variant not in registry: ${costKey}. ` +
                 // Prefixes DERIVED from the model list, so a new family cannot be
@@ -3067,7 +3072,8 @@ export const generateAudio = {
             durationSeconds: seconds,
             characters: input.model === TTS_MODEL ? input.prompt.length : undefined,
         });
-        const entry = registry.models.find((m) => m.model === costKey);
+        const entry = registry.models.find((m) => m.model === costKey) ??
+            (await cloud.get(`/api/agent/models?costKey=${encodeURIComponent(costKey)}`)).models[0];
         if (!entry) {
             throw new Error(`Audio variant not in registry: ${costKey}. Available audio models: ${AUDIO_MODELS.join(' | ')}.`);
         }
@@ -3179,7 +3185,8 @@ export const generateLipSync = {
         }
         const cloud = ctx.cloud();
         const registry = await cloud.get('/api/agent/models');
-        const entry = registry.models.find((m) => m.model === costKey);
+        const entry = registry.models.find((m) => m.model === costKey) ??
+            (await cloud.get(`/api/agent/models?costKey=${encodeURIComponent(costKey)}`)).models[0];
         if (!entry)
             throw new Error(`Model variant not in registry: ${costKey}`);
         const totalCents = creditCost(entry);
@@ -3275,7 +3282,8 @@ export const generateMotionTransfer = {
         const costKey = motionModel === 'kling-mc-std' ? 'kling-mc-std-5s' : 'kling-mc-pro-5s';
         const cloud = ctx.cloud();
         const registry = await cloud.get('/api/agent/models');
-        const entry = registry.models.find((m) => m.model === costKey);
+        const entry = registry.models.find((m) => m.model === costKey) ??
+            (await cloud.get(`/api/agent/models?costKey=${encodeURIComponent(costKey)}`)).models[0];
         if (!entry)
             throw new Error(`Model variant not in registry: ${costKey}`);
         const totalCents = creditCost(entry);
@@ -3454,7 +3462,8 @@ export const editVideo = {
                 : klingEditCostKey(model, billedSeconds);
         const cloud = ctx.cloud();
         const registry = await cloud.get('/api/agent/models');
-        const entry = registry.models.find((m) => m.model === costKey);
+        const entry = registry.models.find((m) => m.model === costKey) ??
+            (await cloud.get(`/api/agent/models?costKey=${encodeURIComponent(costKey)}`)).models[0];
         if (!entry)
             throw new Error(`Model variant not in registry: ${costKey}`);
         const totalCents = creditCost(entry);
@@ -4229,8 +4238,9 @@ function shotParamsShape(described) {
         duration: d(z.number().int().min(1).max(360).optional(), 'Seconds for video or duration-based audio; TTS uses text length.'),
         videoResolution: d(zEnum(VIDEO_RESOLUTIONS).optional(), 'Validated against the model when the Shot is saved.'),
         imageResolution: d(z.enum(['1k', '2k', '3k', '4k']).optional(), 'Image models only.'),
-        gptQuality: d(z.enum(['low', 'medium', 'high', 'xhigh', 'max']).optional(), 'GPT Image 2.5 only.'),
-        gptBackground: d(z.enum(['auto', 'transparent', 'opaque']).optional(), 'GPT Image only.'),
+        gptQuality: d(z.enum(GPT_QUALITY_TIERS).optional(), 'GPT Image 2.5 only.'),
+        gptBackground: d(z.enum(GPT_BACKGROUNDS).optional(), 'GPT Image only.'),
+        detachedVoiceCharacterIds: z.array(z.string()).optional().describe('Character voices removed from this recipe.'),
         imageQuantity: d(z.number().int().min(1).max(10).optional(), 'Image models only.'),
         negativePrompt: z.string().optional(),
         sound: d(z.boolean().optional(), 'Video models that co-generate audio.'),
@@ -4467,7 +4477,7 @@ function shotCostKey(detail) {
         const duration = fires?.duration ?? p.duration;
         if (!duration)
             return null;
-        const billed = (d) => (d > 0 ? Math.ceil(d - 0.05) : 0);
+        const billed = (d) => model === 'minimax-h3-max' ? Math.max(0, d) : (d > 0 ? Math.ceil(d - 0.05) : 0);
         return videoCostKey({
             model: model,
             duration,
@@ -4480,6 +4490,7 @@ function shotCostKey(detail) {
             // of these read 0 there. That is why a listing quote is announced as a
             // floor and `slates_get_shot` is the exact one.
             referenceImages: (detail.references ?? []).filter((r) => r.kind === 'image').length,
+            audioRefSeconds: (detail.references ?? []).filter((r) => r.kind === 'audio').reduce((n, r) => n + (r.durationSeconds ?? 0), 0),
             videoRefSeconds: (detail.references ?? [])
                 .filter((r) => r.kind === 'video')
                 .reduce((n, r) => n + billed(r.durationSeconds ?? 0), 0),
@@ -4499,6 +4510,14 @@ function shotCostKey(detail) {
     }
     return null;
 }
+/** Dynamic reference keys are priced by the same server calculator that debits them. */
+async function loadDynamicPrice(ctx, byKey, key) {
+    if (!key || byKey.has(key))
+        return;
+    const response = await ctx.cloud().get(`/api/agent/models?costKey=${encodeURIComponent(key)}`);
+    for (const row of response.models)
+        byKey.set(row.model, creditCost(row));
+}
 /** Credits for one Shot, and how many generations it fires.
  *
  *  `imageQuantity` multiplies IMAGE models only — the same condition the
@@ -4760,6 +4779,8 @@ export const listShots = {
         // auditing.
         const registry = await ctx.cloud().get('/api/agent/models');
         const byKey = new Map(registry.models.map((m) => [m.model, creditCost(m)]));
+        for (const row of rows)
+            await loadDynamicPrice(ctx, byKey, shotCostKey(row));
         let total = 0;
         let unpriced = 0;
         const shots = rows.map((s) => {
@@ -4811,6 +4832,7 @@ export const getShot = {
         const r = await desktop.get('/agent/shots/get', { id: input.shotId });
         const registry = await ctx.cloud().get('/api/agent/models');
         const byKey = new Map(registry.models.map((m) => [m.model, creditCost(m)]));
+        await loadDynamicPrice(ctx, byKey, shotCostKey(r.shot));
         const q = shotQuote(r.shot, byKey);
         return ok({ ...r.shot, cost_key: q.key, credits: q.credits }, `"${r.shot.name || 'Untitled'}" — ${r.shot.model ?? 'no model set'}, ` +
             (q.key && !r.shot.blocked
@@ -4894,6 +4916,7 @@ export const generateFromShots = {
         }
         const registry = await ctx.cloud().get('/api/agent/models');
         const byKey = new Map(registry.models.map((m) => [m.model, creditCost(m)]));
+        await Promise.all(details.map((d) => loadDynamicPrice(ctx, byKey, shotCostKey(d))));
         const quotes = details.map((d) => ({ detail: d, ...shotQuote(d, byKey) }));
         const total = quotes.reduce((n, q) => n + q.credits, 0);
         const largest = quotes.reduce((m, q) => (q.credits > m ? q.credits : m), 0);

package/dist/prompts/model-capabilities.d.ts

@@ -84,7 +84,70 @@ export interface VoiceCloneCapability {
         max: number;
     };
 }
+export declare const GPT_QUALITY_TIERS: readonly ["low", "medium", "high", "xhigh", "max"];
+export type GptQuality = (typeof GPT_QUALITY_TIERS)[number];
+export declare const GPT_BACKGROUNDS: readonly ["auto", "transparent", "opaque"];
+export type GptBackground = (typeof GPT_BACKGROUNDS)[number];
+export type ImageResolution = '1k' | '2k' | '3k' | '4k';
+export declare const GPT_IMAGE_25_SIZES: Record<string, Record<string, {
+    width: number;
+    height: number;
+}>>;
+/**
+ * fal's named ~1MP presets per aspect ratio, with custom dims where fal has no
+ * preset. The `1k` rung of every non-GPT image model resolves through this.
+ */
+export declare const FAL_1MP_SIZES: Record<string, string | {
+    width: number;
+    height: number;
+}>;
+/** Pixel dims for a megapixel target at an aspect ratio, rounded to multiples of 8. */
+export declare function computeFalDimensions(aspectRatio: string, targetMP: number): {
+    width: number;
+    height: number;
+};
+/**
+ * The `image_size` a non-GPT fal image request carries, for one model × aspect ×
+ * resolution rung.
+ *
+ * 🚨 THIS IS A BILLING INPUT, WHICH IS WHY IT LIVES HERE (moved out of
+ * slate/src/main/api/fal.ts, 2026-09-10). The resolution rung is a segment of
+ * every image cost key, and nothing in the REQUEST names it — fal is told pixel
+ * dimensions, not "2k". The proxy therefore recovers the rung by running this
+ * function over the model's declared `imageResolutions` × `aspectRatios` and
+ * matching the body's `image_size`, exactly as it recovers a GPT Image rung from
+ * `GPT_IMAGE_25_SIZES`. A second copy of this arithmetic would mean the desktop
+ * and the server could disagree about what a request is worth, silently.
+ *
+ * GPT Image does NOT come through here — that family carries explicit pixel
+ * classes in `GPT_IMAGE_25_SIZES` and an explicit `quality` rung.
+ */
+export declare function falImageSize(model: string, aspectRatio?: string, resolution?: string): string | {
+    width: number;
+    height: number;
+};
 export interface ModelCapability {
+    imageResolutions?: ImageResolution[];
+    /**
+     * Images ONE request may ask the provider for in a single batch.
+     *
+     * 🚨 ABSENT MEANS ONE, AND THAT IS A BILLING BOUND, NOT A HINT. A cost key
+     * prices a single image, so a request that returns N images has to be charged
+     * N times — the proxy multiplies the debit by the batch size and refuses a
+     * batch larger than this (`slates-api/src/lib/fal-image-keys.ts`). Only a model
+     * with a READ provider ceiling gets a number here; everything else fans out as
+     * separate generations, which is what the desktop does for every model.
+     */
+    maxBatchImages?: number;
+    /** Per-file reference-audio bounds, independently of the combined cap. */
+    referenceVideoDuration?: {
+        min: number;
+        max: number;
+    };
+    referenceAudioDuration?: {
+        min: number;
+        max: number;
+    };
     aspectRatios: AspectRatio[];
     /** Provider-keyed overrides. `fal` is the one that matters — see AGENT_ROUTE_PROVIDER. */
     providerAspectRatios?: Record<string, AspectRatio[]>;
@@ -185,4 +248,21 @@ export declare function describeVideoResolutions(models: readonly string[]): str
 export declare function describeDurations(models: readonly string[]): string;
 /** e.g. "seedance-2: 9 · seedance-2.5: 30 · omni-flash: 7 · seedance-2.5-edit: 0 (prompt + source clip only)" */
 export declare function describeReferenceImageCaps(models: readonly string[]): string;
+/** H3 Max reference accounting, fal's worked tables read 2026-09-09.
+ * https://fal.ai/models/minimax/h3-max/reference-to-video
+ * 1080p video-reference pricing is unpublished; never infer it from output rates.
+ */
+export declare const MINIMAX_MAX_REFERENCE: {
+    readonly freeTokens: 4096;
+    readonly imagePixelsPerToken: 1024;
+    readonly normalizedImageEdge: 1024;
+    readonly audioTokensPerSecond: 80;
+    readonly videoTokensPerSecond: Partial<Record<VideoResolution, number>>;
+};
+export declare function minimaxMaxReferenceTokens(input: {
+    imagePixels: number;
+    videoSeconds: number;
+    audioSeconds: number;
+    resolution: string;
+}): number;
 //# sourceMappingURL=model-capabilities.d.ts.map