@nodaro/shared 1.9.0 → 1.11.0

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):
@@ -1147,10 +1198,11 @@ export const SEEDANCE_2_EXTEND_STITCH = {
   trimHeadFrames: 3,
   /** Boundary audio fade length (seconds), timeline-preserving. */
   audioFadeSec: 0.15,
-  /** Seconds of the source's TAIL passed as the @video_1 reference —
-   *  spike-validated: a short tail keeps the model focused on continuing
-   *  the boundary motion instead of re-staging the whole clip. */
-  referenceTailSeconds: 1,
+  /** 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
 
 /**
@@ -1341,8 +1393,13 @@ export const AUDIO_ADDON_PROVIDERS = new Set([
  *                     lip-synced spoken dialogue. Toggle is offered; the
  *                     pipeline still uses TTS + lip-sync for dialogue.
  *   "native_speech" — bakes spoken dialogue + lip movement from the prompt
- *                     (VEO 3.x). The pipeline injects the dialogue line, enables
- *                     audio, and revoices the clip to the character's saved voice.
+ *                     (VEO 3.x always-on; Kling 2.6/3.0 behind the `sound`
+ *                     toggle — probe-verified 2026-07-16: scripted lines come
+ *                     back word-for-word with articulated lips on the KIE
+ *                     path, matching the official Kling 2.6 audio guide and
+ *                     Kling 3.0 prompting docs). The pipeline injects the
+ *                     dialogue line, enables audio, and revoices the clip to
+ *                     the character's saved voice.
  *   "audio_driven"  — lip-syncs to a supplied reference-audio track (Seedance
  *                     2.0 multimodal). The pipeline synthesises the character's
  *                     voice first, feeds it as reference audio, and skips the
@@ -1358,6 +1415,14 @@ export interface VideoAudioCapability {
   alwaysOn?: boolean
   /** Enabling audio raises the credit cost (Kling — see AUDIO_ADDON_PROVIDERS). */
   affectsCost?: boolean
+  /**
+   * The model generates audio when the caller expresses NO intent (its own
+   * config default is on — kling-3.0's `sound: true`, kling-3-omni's
+   * `generate_audio: true`). The `:audio` credit suffix mirrors this default
+   * so an intent-less request is billed for the audio it actually produces
+   * (see buildVideoCreditModelIdentifier). Absent ⇒ default off.
+   */
+  defaultOn?: boolean
 }
 
 /**
@@ -1374,9 +1439,22 @@ export const VIDEO_AUDIO_CAPABILITY: Record<string, VideoAudioCapability> = {
   veo3: { mode: "native_speech", alwaysOn: true },
   "veo3.1": { mode: "native_speech", alwaysOn: true },
   veo3_lite: { mode: "native_speech", alwaysOn: true },
-  // Kling 2.6 / 3.0 — ambient sound/SFX toggle; not lip-synced speech. Cost-affecting.
-  kling: { mode: "ambient", field: "sound", affectsCost: true },
-  "kling-3.0": { mode: "ambient", field: "sound", affectsCost: true },
+  // Kling 2.6 / 3.0 — native spoken dialogue + lip sync behind the `sound`
+  // toggle (probe-verified on the KIE path 2026-07-16: scripted quoted lines
+  // are spoken verbatim with articulated lips; the official Kling 2.6 audio
+  // guide documents speech/dialogue/narration/singing with [Character@Voice]
+  // binding, zh+en voices). Cost-affecting: `:audio` credit suffix.
+  // kling-3.0's model default is sound ON (kie models.ts extraParams +
+  // kling3-client `?? true`) → defaultOn keeps billing aligned with what an
+  // intent-less request actually generates; kling 2.6 defaults OFF.
+  kling: { mode: "native_speech", field: "sound", affectsCost: true },
+  "kling-3.0": { mode: "native_speech", field: "sound", affectsCost: true, defaultOn: true },
+  // Kling 3.0 Omni (Replicate kwaivgi/kling-v3-omni-video) — the Omni tier's
+  // headline feature IS native dialogue (per-character voices, unified audio
+  // timeline). Lever is Replicate's `generate_audio` (default true in our
+  // provider config → defaultOn); audio is priced into the flat per-duration
+  // rate, so NOT cost-affecting (no :audio composite).
+  "kling-3-omni": { mode: "native_speech", field: "generateAudio", defaultOn: true },
   // Seedance 1.x — optional ambient audio (generate_audio); not dialogue.
   seedance: { mode: "ambient", field: "generateAudio" },
   // Seedance 2.0 — multimodal; lip-syncs to a supplied reference-audio track.
@@ -1405,11 +1483,11 @@ export function videoModelSupportsAudio(model: string | undefined): boolean {
 
 /**
  * True when the model can produce lip-synced spoken DIALOGUE — either natively
- * (VEO) or driven by a supplied audio track (Seedance 2.0). Drives the Story→Video
- * dialogue auto-pick: in-model speech + character revoice (VEO) / character-voiced
- * reference audio (Seedance 2.0) vs. the TTS + separate-lip-sync fallback.
- * Ambient-only models (Kling, Seedance 1.x) return `false` — their audio is SFX,
- * not speech.
+ * (VEO 3.x, Kling 2.6/3.0/Omni) or driven by a supplied audio track (Seedance
+ * 2.0). Drives the Story→Video dialogue auto-pick: in-model speech + character
+ * revoice (VEO, Kling) / character-voiced reference audio (Seedance 2.0) vs.
+ * the TTS + separate-lip-sync fallback. Ambient-only models (Seedance 1.x)
+ * return `false` — their audio is SFX, not speech.
  */
 export function videoModelCanSpeakDialogue(model: string | undefined): boolean {
   const mode = getVideoAudioCapability(model).mode

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/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" })