@nodaro/shared 1.7.0 → 1.10.0

package/src/model-catalog.ts

@@ -146,7 +146,7 @@ export const MODEL_RECOMMENDATIONS: readonly ModelRecommendation[] = [
   { intent: "best for typography / logos / text-heavy", modelIds: ["nano-banana-pro", "gpt-image-2"], note: "Nano Banana Pro for diagrams / complex text; GPT Image 2 for logos and short copy." },
   { intent: "cheapest realistic image", modelIds: ["z-image", "qwen", "imagen4-fast"], note: "Z-Image is the cheapest at 1 credit. Qwen / Imagen4 Fast for slightly higher quality." },
   { intent: "highest fidelity image", modelIds: ["nano-banana-pro", "imagen4-ultra", "flux-flex"], note: "Pick by family preference; all three are premium tiers." },
-  { intent: "image edit / restyle", modelIds: ["flux-kontext", "ideogram-remix", "seedream-5-lite-i2i"], note: "Flux Kontext preserves identity; Ideogram Remix is character-aware; Seedream 5 Lite for instruction-based edits." },
+  { intent: "image edit / restyle", modelIds: ["flux-kontext", "ideogram-remix", "seedream-5-pro-i2i"], note: "Flux Kontext preserves identity; Ideogram Remix is character-aware; Seedream 5 Pro for instruction-based edits (5 Lite is the budget option)." },
   { intent: "highest-resolution image (4K / 8K)", modelIds: ["topaz-image-upscale", "nano-banana-pro", "gpt-image-2"], note: "Generate at native then Topaz upscale for 8K." },
   { intent: "background removal / cutout", modelIds: ["recraft-remove-bg"], note: "1 credit, no prompt needed." },
   // video
@@ -165,6 +165,11 @@ export const MODEL_RECOMMENDATIONS: readonly ModelRecommendation[] = [
 // can consume them. Keep in sync when adding a new ratio set.)
 // =============================================================================
 const NANO_BANANA_RATIOS = ["1:1", "16:9", "9:16", "4:3", "3:4", "3:2", "2:3", "4:5", "5:4", "21:9"] as const
+// Nano Banana 2 Lite only — the model's full documented enum incl. `auto`
+// (native default) and extreme banner ratios (docs.kie.ai/market/google/
+// nano-banana-2-lite). Kept separate so the wider set can't leak to the
+// rest of the family.
+const NANO_BANANA_2_LITE_RATIOS = ["auto", "1:1", "16:9", "9:16", "4:3", "3:4", "3:2", "2:3", "4:5", "5:4", "21:9", "4:1", "1:4", "8:1", "1:8"] as const
 // Note: these arrays are synchronized with the frontend's
 // `model-options.ts` constants of the same names. The frontend now imports
 // them via the `getAspectRatioOptions(modelId)` helper; if you update one
@@ -233,6 +238,22 @@ const IMAGE_MODELS: Record<string, ModelCatalogEntry> = {
       { identifier: "nano-banana-2:4K", credits: 5, note: "4K" },
     ],
   },
+  "nano-banana-2-lite": {
+    id: "nano-banana-2-lite",
+    kind: "image",
+    // Same single-id shape as nano-banana-2: optional input images on the one
+    // endpoint (via image_urls, up to 10), so reachable for both t2i and i2i.
+    modes: ["t2i", "i2i"] as const,
+    family: "Google",
+    label: "Nano Banana 2 Lite",
+    series: "Nano Banana",
+    description: "Lightweight Nano Banana 2 (Gemini 3.1 Flash-Lite) — fast, low-cost 1K generation and editing.",
+    useCases: ["realistic", "drafts", "iteration"],
+    features: ["reference-image"],
+    aspectRatios: NANO_BANANA_2_LITE_RATIOS,
+    // 1K only — no resolution lever, flat pricing.
+    pricing: [{ identifier: "nano-banana-2-lite", credits: 2, note: "1K flat" }],
+  },
   "nano-banana-pro": {
     id: "nano-banana-pro",
     kind: "image",
@@ -691,6 +712,40 @@ const IMAGE_MODELS: Record<string, ModelCatalogEntry> = {
       { identifier: "seedream-5-lite-i2i:high", credits: 5, note: "high quality" },
     ],
   },
+  "seedream-5-pro": {
+    id: "seedream-5-pro",
+    kind: "image",
+    modes: ["t2i"] as const,
+    family: "Bytedance",
+    label: "Seedream 5 Pro",
+    series: "Seedream",
+    description: "Flagship Seedream 5 Pro — strongest instruction following and visual reasoning. Basic = 1K, high = 2K.",
+    useCases: ["realistic", "instruction"],
+    features: ["reference-image"],
+    aspectRatios: SEEDREAM_RATIOS,
+    qualities: ["basic", "high"],
+    pricing: [
+      { identifier: "seedream-5-pro", credits: 3, note: "basic / 1K default" },
+      { identifier: "seedream-5-pro:high", credits: 6, note: "high / 2K" },
+    ],
+  },
+  "seedream-5-pro-i2i": {
+    id: "seedream-5-pro-i2i",
+    kind: "image",
+    modes: ["i2i"] as const,
+    family: "Bytedance",
+    label: "Seedream 5 Pro (I2I)",
+    series: "Seedream",
+    description: "Image-to-image with Seedream 5 Pro — multi-reference instruction edits. Basic = 1K, high = 2K.",
+    useCases: ["edit", "instruction"],
+    features: ["reference-image"],
+    aspectRatios: SEEDREAM_RATIOS,
+    qualities: ["basic", "high"],
+    pricing: [
+      { identifier: "seedream-5-pro-i2i", credits: 3, note: "basic / 1K default" },
+      { identifier: "seedream-5-pro-i2i:high", credits: 6, note: "high / 2K" },
+    ],
+  },
 
   // ── Alibaba Qwen ──
   "qwen": {
@@ -1754,10 +1809,10 @@ const VIDEO_MODELS: Record<string, ModelCatalogEntry> = {
     id: "gemini-3-flash-video-analysis",
     kind: "video",
     modes: ["video-analysis"] as const,
-    family: "Google",
-    label: "Video Analysis (Gemini 3 Flash)",
-    series: "Gemini",
-    description: "Analyze a video into a structured shot list (scenes, camera, audio) — fast Gemini tier. Billed per duration bucket.",
+    family: "Nodaro",
+    label: "Video Analysis (Fast)",
+    series: "Video Analysis",
+    description: "Analyze a video into a structured shot list (scenes, camera, audio) — fast, economy tier. Billed per duration bucket.",
     useCases: ["video-analysis", "shot-list", "fast"],
     pricing: [
       { identifier: "video-analysis:gemini-3-flash", credits: 3, note: "10-min ceiling (no duration given)" },
@@ -1771,10 +1826,10 @@ const VIDEO_MODELS: Record<string, ModelCatalogEntry> = {
     id: "gemini-3.1-pro-video-analysis",
     kind: "video",
     modes: ["video-analysis"] as const,
-    family: "Google",
-    label: "Video Analysis (Gemini 3.1 Pro)",
-    series: "Gemini",
-    description: "Analyze a video into a structured shot list (scenes, camera, audio) — high-fidelity Gemini tier. Billed per duration bucket.",
+    family: "Nodaro",
+    label: "Video Analysis (Pro)",
+    series: "Video Analysis",
+    description: "Analyze a video into a structured shot list (scenes, camera, audio) — higher-fidelity, default tier. Billed per duration bucket.",
     useCases: ["video-analysis", "shot-list", "cinematic"],
     pricing: [
       { identifier: "video-analysis:gemini-3.1-pro", credits: 11, note: "10-min ceiling (no duration given)" },

package/src/model-constants.ts

@@ -37,6 +37,7 @@ export const IMAGE_PROMPT_MAX = 5000
 export const MAX_IMAGE_PROMPT_CHARS_BY_PROVIDER: Record<string, number> = {
   // ── higher than the 5000 default ──
   "nano-banana-2": 20000,      // docs.kie.ai/market/google/nano-banana-2
+  "nano-banana-2-lite": 20000, // docs.kie.ai/market/google/nano-banana-2-lite
   "nano-banana-pro": 20000,    // docs.kie.ai/market/google/pro-image-to-image
   "gpt-image-2-i2i": 20000,    // docs.kie.ai/market/gpt/gpt-image-2-image-to-image
   // ── lower than the 5000 default (over-send risk if left at default) ──
@@ -48,7 +49,8 @@ export const MAX_IMAGE_PROMPT_CHARS_BY_PROVIDER: Record<string, number> = {
   "qwen-edit": 2000,           // docs.kie.ai/market/qwen/image-edit
   // verified == 5000 default (no entry needed): imagen4(-fast/-ultra), nano-banana,
   //   nano-banana-edit, flux, flux-flex, gpt-image-2, ideogram-v3/-edit/-remix,
-  //   z-image, grok, qwen-i2i.
+  //   z-image, grok, qwen-i2i, seedream-5-pro, seedream-5-pro-i2i
+  //   (docs.kie.ai/market/seedream/5-pro-text-to-image + 5-pro-image-to-image).
   // grok-i2i: doc states 390000 (78× its t2i sibling) — treated as a KIE schema
   //   typo and left at the 5000 default per the sanity-cap decision.
   // UNVERIFIED (no limit stated in schema) → 5000 default: flux-kontext(-max)
@@ -93,6 +95,18 @@ export const SUNO_TEXT_MAX = 5000
  */
 export const PROMPT_HARD_CEILING = 20000
 
+/**
+ * Ceiling for LLM TEXT-generation node inputs (the "Generate Text"/llm-chat node
+ * + AI Writer + Generate Script `systemPrompt`/`userInput`/`prompt` fields).
+ * Distinct from PROMPT_HARD_CEILING (an image/video PROMPT budget): these inputs
+ * go straight into an LLM whose context is huge (Claude 200K / GPT 128K+ /
+ * Gemini 1M tokens), so the old flat 10000 was a false blocker on pasting a
+ * document to summarize or rewrite. 100000 chars (~25K input tokens) covers
+ * long-form inputs while still bounding abuse; the OUTPUT stays capped by each
+ * route's maxTokens.
+ */
+export const LLM_TEXT_INPUT_MAX = 100_000
+
 /**
  * Per-provider maximum VIDEO prompt length (chars), VERIFIED against each model's
  * official docs.kie.ai schema (2026-06). Absent → {@link VIDEO_PROMPT_MAX} (8000)
@@ -333,6 +347,7 @@ export const MODELS_WITH_REFERENCE_IMAGE_SUPPORT = new Set([
   "nano-banana",
   "nano-banana-pro",
   "nano-banana-2",
+  "nano-banana-2-lite",
   // T2I providers that auto-route to their i2i sibling when refs are attached
   "gpt-image",
   "gpt-image-2",
@@ -340,6 +355,7 @@ export const MODELS_WITH_REFERENCE_IMAGE_SUPPORT = new Set([
   "qwen",
   "seedream",
   "seedream-5-lite",
+  "seedream-5-pro",
   "flux",
   "flux-flex",
   // Image editing / image-to-image (reference = source image)
@@ -357,6 +373,7 @@ export const MODELS_WITH_REFERENCE_IMAGE_SUPPORT = new Set([
   "qwen-edit",
   "seedream-edit",
   "seedream-5-lite-i2i",
+  "seedream-5-pro-i2i",
   "grok-i2i",
   // Upscale / background ops (source acts as the reference)
   "recraft-remove-bg",
@@ -386,6 +403,7 @@ export const T2I_TO_I2I_VARIANT: Record<string, string> = {
   "qwen": "qwen-i2i",
   "seedream": "seedream-edit",
   "seedream-5-lite": "seedream-5-lite-i2i",
+  "seedream-5-pro": "seedream-5-pro-i2i",
   "flux": "flux-pro-i2i",
   "flux-flex": "flux-i2i",
 }
@@ -405,6 +423,7 @@ export const REF_IMAGE_MAX_LIMITS: Record<string, number> = {
   "nano-banana": 8,
   "nano-banana-pro": 8,
   "nano-banana-2": 4,
+  "nano-banana-2-lite": 10,
   "wan-2.7": 9,
   // Image-to-image (multi-source array)
   "nano-banana-edit": 8,
@@ -414,6 +433,7 @@ export const REF_IMAGE_MAX_LIMITS: Record<string, number> = {
   "flux-pro-i2i": 4,
   "seedream-edit": 16,
   "seedream-5-lite-i2i": 16,
+  "seedream-5-pro-i2i": 16,
   // Single-source i2i (one input image)
   "flux-kontext": 1,
   "flux-kontext-max": 1,
@@ -484,6 +504,8 @@ export const VARIABLE_PRICING_MODELS: Record<string, "quality" | "resolution" |
   "seedream-edit": "quality",
   "seedream-5-lite": "quality",
   "seedream-5-lite-i2i": "quality",
+  "seedream-5-pro": "quality",
+  "seedream-5-pro-i2i": "quality",
   "topaz-image-upscale": "resolution",
   "ideogram-edit": "rendering-speed",
   "ideogram-remix": "rendering-speed",
@@ -495,7 +517,7 @@ export const VARIABLE_PRICING_MODELS: Record<string, "quality" | "resolution" |
 
 
 // Models where quality=high triggers composite credit identifier
-export const HIGH_QUALITY_PROVIDERS = new Set(["gpt-image", "gpt-image-i2i", "seedream", "seedream-edit", "seedream-5-lite", "seedream-5-lite-i2i"])
+export const HIGH_QUALITY_PROVIDERS = new Set(["gpt-image", "gpt-image-i2i", "seedream", "seedream-edit", "seedream-5-lite", "seedream-5-lite-i2i", "seedream-5-pro", "seedream-5-pro-i2i"])
 
 // Models where resolution=2K triggers composite credit identifier
 export const TWO_K_RESOLUTION_PROVIDERS = new Set(["flux", "flux-pro-i2i", "flux-flex", "flux-i2i"])
@@ -522,6 +544,7 @@ export const IMAGE_GEN_PROVIDERS = [
   "flux",
   "nano-banana-pro",
   "nano-banana-2",
+  "nano-banana-2-lite",
   "grok",
   "gpt-image",
   "gpt-image-2",
@@ -532,6 +555,7 @@ export const IMAGE_GEN_PROVIDERS = [
   "qwen",
   "seedream",
   "seedream-5-lite",
+  "seedream-5-pro",
   "flux-flex",
   "flux-kontext",
   "flux-kontext-max",
@@ -548,6 +572,7 @@ export const IMAGE_GEN_PROVIDERS = [
 export const IMAGE_I2I_PROVIDERS = [
   "nano-banana",
   "nano-banana-2",
+  "nano-banana-2-lite",
   "nano-banana-pro",
   "grok-i2i",
   "flux-i2i",
@@ -561,6 +586,7 @@ export const IMAGE_I2I_PROVIDERS = [
   "qwen-edit",
   "seedream-edit",
   "seedream-5-lite-i2i",
+  "seedream-5-pro-i2i",
   "flux-kontext",
   "flux-kontext-max",
   // Replicate Open (uncensored) — multi-image Kontext via Replicate
@@ -1003,10 +1029,12 @@ export const IMAGE_MASK_MODE: Record<ImageGenProvider, ImageMaskMode> = {
   "nano-banana": "prompt",
   "nano-banana-pro": "prompt",
   "nano-banana-2": "prompt",
+  "nano-banana-2-lite": "prompt",
   "gpt-image": "prompt",
   "gpt-image-2": "prompt",
   "seedream": "prompt",
   "seedream-5-lite": "prompt",
+  "seedream-5-pro": "prompt",
   "qwen": "prompt",
   "flux-kontext": "prompt",
   "flux-kontext-max": "prompt",
@@ -1129,6 +1157,29 @@ export const SEEDANCE_2_REF_LIMITS = {
   audio: 3,
 } as const
 
+/**
+ * KIE r2v REFERENCE-VIDEO MINIMUM (seconds) for the Seedance 2.0 family — the
+ * provider hard-rejects shorter reference clips with a 400 BEFORE generation:
+ * "the parameter video duration (seconds) specified in the request must be
+ * greater than or equal to 1.8 for model dreamina-seedance-2-0-fast in r2v"
+ * (2026-07-13, job dbf95612 — every 1.0s continuation tail failed
+ * deterministically). Mirror of the reference-AUDIO maximum below.
+ */
+export const SEEDANCE_2_R2V_MIN_REF_VIDEO_SEC = 1.8
+
+/**
+ * The continuation-reference length (seconds) every Seedance-2 chaining
+ * feature actually cuts — extend-node tails, generate-video-pro segment
+ * tails, edit-video-pro refOut/refIn brackets — AND the length the gvp/evp
+ * credit formulas bill per continuation join. Clears the provider floor
+ * above with margin while staying short enough to keep the model focused on
+ * continuing the boundary motion instead of re-staging the whole clip.
+ * Guarded ≥ floor by model-constants tests; the private-plugin twin
+ * (nodaro-cloud-plugins chain.ts TAIL_SEC / bridge-math.ts MIN_REF) is
+ * guarded by that repo's r2v-ref-floor.test.ts — keep the two in sync.
+ */
+export const SEEDANCE_2_CONTINUATION_REF_SEC = 2
+
 /**
  * Trim-stitch parameters for the seedance-2-extend provider (spike-validated
  * 2026-06-11):
@@ -1139,12 +1190,19 @@ export const SEEDANCE_2_REF_LIMITS = {
  * boundary clicks without shifting sync (combineVideos cut+crossfade path).
  */
 export const SEEDANCE_2_EXTEND_STITCH = {
-  /** Frames dropped from the END of the source clip. */
+  /** Frames dropped from the END of the source clip (smart-cut FALLBACK —
+   *  used only when the PSNR boundary matcher finds no genuine match). */
   trimTailFrames: 4,
-  /** Frames dropped from the START of the generated extension. */
+  /** Frames dropped from the START of the generated extension (smart-cut
+   *  fallback, see above). */
   trimHeadFrames: 3,
   /** Boundary audio fade length (seconds), timeline-preserving. */
   audioFadeSec: 0.15,
+  /** Seconds of the source's TAIL passed as the @video_1 reference — short
+   *  keeps the model focused on continuing the boundary motion; the value is
+   *  floored by SEEDANCE_2_R2V_MIN_REF_VIDEO_SEC (the original spike's 1s
+   *  tail is now provider-rejected). */
+  referenceTailSeconds: SEEDANCE_2_CONTINUATION_REF_SEC,
 } as const
 
 /**

package/src/node-default-mappings.ts

@@ -174,6 +174,8 @@ const QUALITY_MAP: Record<string, QualityMapping> = {
   "seedream-edit":       { field: "quality",    values: { low: "basic",  mid: "basic",  high: "high" } },
   "seedream-5-lite":     { field: "quality",    values: { low: "basic",  mid: "basic",  high: "high" } },
   "seedream-5-lite-i2i": { field: "quality",    values: { low: "basic",  mid: "basic",  high: "high" } },
+  "seedream-5-pro":      { field: "quality",    values: { low: "basic",  mid: "basic",  high: "high" } },
+  "seedream-5-pro-i2i":  { field: "quality",    values: { low: "basic",  mid: "basic",  high: "high" } },
   // Video gen — resolution-style (720p/1080p)
   "veo3":                { field: "resolution", values: { low: "720p", mid: "1080p", high: "1080p" } },
   "veo3.1":              { field: "resolution", values: { low: "720p", mid: "1080p", high: "1080p" } },
@@ -248,6 +250,7 @@ const GENERATE_IMAGE_PROVIDER_TO_MODEL: Record<string, string> = {
   "nano-banana": "gemini-2.5-flash-image",
   "nano-banana-pro": "gemini-2.5-flash-image",
   "nano-banana-2": "gemini-2.5-flash-image",
+  "nano-banana-2-lite": "gemini-3.1-flash-lite-image",
 }
 
 export function deriveLinkedFields(

package/src/node-mappable-fields.ts

@@ -15,6 +15,12 @@ export const NODE_MAPPABLE_FIELDS: Readonly<Record<string, readonly string[]>> =
   // for back-compat with un-migrated workflow JSON). Mirrors text-to-video so
   // fieldMappings/{} injection AND missing-ref detection work on the live node.
   "generate-video":      ["prompt", "negativePrompt"],
+  // Trimmed multi-segment stitch variant of generate-video — prompt only, no
+  // negativePrompt field on the node.
+  "generate-video-pro":  ["prompt"],
+  // Span-replace sibling of generate-video-pro — prompt only, no
+  // negativePrompt field on the node.
+  "edit-video-pro":      ["prompt"],
   "video-analysis":      ["analysisFocus", "youtubeUrl"],
   "video-to-video":      ["prompt"],
   "text-to-speech":      ["directText"],

package/src/producer-types.ts

@@ -30,6 +30,16 @@ export const VIDEO_PRODUCER_TYPES: ReadonlySet<string> = new Set([
   // through to the imageUrl/videoUrl/audioUrl/text default and downstream consumers
   // could silently misroute the output.
   "generate-video",
+  // Generate Video Pro — Seedance-2-family multi-segment stitch variant of
+  // generate-video (same "emits videoUrl" contract; a trimmed provider +
+  // handle set). Must mirror generate-video here or its output can't connect
+  // downstream (the recurring "cannot connect the outputs" bug class).
+  "generate-video-pro",
+  // Edit Video Pro — Seedance-2-family span-replace sibling of generate-
+  // video-pro (same "emits videoUrl" contract; source video + prompt in,
+  // ONE video out). Must mirror generate-video-pro here or its output can't
+  // connect downstream (the recurring "cannot connect the outputs" bug class).
+  "edit-video-pro",
   "upload-video",
   "youtube-video",
   "combine-videos",

package/src/selector.ts

@@ -706,7 +706,16 @@ export function applyRangeIndices(
   return result
 }
 
-/** mulberry32 PRNG — deterministic given a seed (any 32-bit signed integer). */
+/** mulberry32 PRNG — deterministic given a seed (any 32-bit signed integer).
+ *
+ * TWIN WARNING: backend/src/providers/video/audio-fx.ts carries its own
+ * private mulberry32 (cosmetically different — `>>> 0` seed masking — but the
+ * same sequence). The two are deliberately INDEPENDENT and both frozen: the
+ * audio-fx copy seeds every reverb impulse response that the committed
+ * characterization goldens (and every customer's rendered reverb waveform)
+ * are pinned to. Do NOT "deduplicate" them into one shared export or "sync"
+ * one to match the other — a change to either sequence is a silent behavior
+ * change in its domain. */
 function mulberry32(seed: number): () => number {
   let a = seed | 0
   return () => {

package/src/video-analysis-pricing.ts

@@ -10,13 +10,16 @@
  * these.
  *
  * The measured-rate constants and the $-derived `videoAnalysisBucketCredits`
- * formula live in `backend/src/lib/pricing/video-analysis-cost.ts` (core,
- * not ee/ — the MCP tool description builder needs them regardless of
- * edition). They were moved out of this package (published Apache-2.0 on
- * npm — an irrevocable grant) per the 2026-07-06 public-flip IP audit, S5.
+ * formula that GENERATE these numbers live PRIVATELY in the
+ * `@nodaroai/cloud-plugins` package (`src/plugins/video-analysis/cost.ts`) —
+ * never in this public repo. They were first moved out of this package
+ * (published Apache-2.0 on npm) per the 2026-07-06 public-flip IP audit S5,
+ * then out of the app repo entirely alongside the rest of the video-analysis
+ * node. A cross-check test in that private package guards this table so the
+ * public numbers can't silently drift from the formula.
  *
  * `VIDEO_ANALYSIS_BUCKET_CREDITS` below is the precomputed OUTPUT of that
- * backend formula for every (model × bucket) combination — a plain credit
+ * private formula for every (model × bucket) combination — a plain credit
  * lookup table, not a formula, mirroring the same wire-contract pattern
  * `VIDEO_CLIP_CREDITS` uses in `film-pricing.ts`. It is what the frontend's
  * client-side cost preview (`estimateNodeCredits` in
@@ -34,11 +37,12 @@ const WINDOW_LEN = 150, WINDOW_STRIDE = 145, WINDOW_OVERLAP = 5
 export const VIDEO_ANALYSIS_WINDOW = { LEN: WINDOW_LEN, STRIDE: WINDOW_STRIDE, OVERLAP: WINDOW_OVERLAP, SINGLE_MAX: 180 } as const
 
 /**
- * Precomputed credit cost per (model, bucket) — the OUTPUT of the backend's
- * `videoAnalysisBucketCredits` formula, not a formula itself. Regenerate by
- * running that function for every `VIDEO_ANALYSIS_LLM_MODELS` × duration
- * bucket combination whenever the underlying rate/token constants change
- * (backend test guards drift). Keep in sync with
+ * Precomputed credit cost per (model, bucket) — the OUTPUT of the private
+ * `videoAnalysisBucketCredits` formula (in `@nodaroai/cloud-plugins`), not a
+ * formula itself. Regenerate by running that function for every
+ * `VIDEO_ANALYSIS_LLM_MODELS` × duration bucket combination whenever the
+ * underlying rate/token constants change (the plugin's cost test guards drift).
+ * Keep in sync with
  * `docs/nodes/processing-video/video-analysis.md`.
  */
 export const VIDEO_ANALYSIS_BUCKET_CREDITS: Record<string, number> = {

package/src/video-analysis.ts

@@ -34,11 +34,18 @@ export const entitySlotSchema = z.object({
 })
 export type EntitySlot = z.infer<typeof entitySlotSchema>
 
-const audioSchema = z.object({
-  mode: z.enum(["speech", "music", "sfx", "silence"]),
-  content: z.string(), // speech: verbatim quote; music/sfx: gen-ready description; silence: "" allowed
+/**
+ * One concurrent sound layer in a scene. Real footage stacks sound (music bed
+ * under dialogue over ambient sfx), so a scene carries an ARRAY of these — an
+ * empty array means genuine silence. `content`: speech = verbatim words;
+ * music/sfx = gen-ready description. `voice` is speech-only voice-casting.
+ */
+const audioLayerSchema = z.object({
+  mode: z.enum(["speech", "music", "sfx"]),
+  content: z.string().min(1),
   voice: z.string().optional(),
 })
+export type AudioLayer = z.infer<typeof audioLayerSchema>
 
 const windowSceneBase = z.object({
   startSec: z.number().min(0),
@@ -48,7 +55,8 @@ const windowSceneBase = z.object({
   camera: z.string(),
   visual: z.string().min(1),
   transitionOut: z.enum(["cut", "fade", "wipe", "whip"]).optional(),
-  audio: audioSchema,
+  // Array of concurrent layers (music + speech + sfx together); [] = silence.
+  audio: z.array(audioLayerSchema),
 })
 // .strip() (default) drops model-emitted oversized/slotRefs — validator-computed only.
 const windowSceneSchema = windowSceneBase.refine((s) => s.endSec > s.startSec, { message: "endSec must be > startSec" })