@nodaro/shared 2.14.0 → 2.16.0

package/src/model-catalog.ts

@@ -58,6 +58,10 @@ export type ModelMode =
   | "isolation"
   | "dubbing"
   | "forced-alignment"
+  // dialogue — deliberately its own mode, never "tts": the dialogue model
+  // takes a multi-speaker script shape (inputs[]), not the single-text
+  // generate_speech contract, so it must never appear in a TTS model list.
+  | "dialogue"
   // video analysis
   | "video-analysis"
   // video audit — deliberately its own mode, never "video-analysis": the
@@ -2143,18 +2147,17 @@ const AUDIO_MODELS: Record<string, ModelCatalogEntry> = {
   "elevenlabs-dialogue": {
     id: "elevenlabs-dialogue",
     kind: "audio",
-    modes: ["tts"] as const,
+    // Its own mode, never "tts": the script shape (inputs[]) doesn't fit the
+    // single-text generate_speech contract, so list_models must never offer
+    // it there. The dialogue-capable MCP verb is `generate_dialogue`.
+    modes: ["dialogue"] as const,
     family: "ElevenLabs",
     label: "ElevenLabs Dialogue v3",
     series: "ElevenLabs",
-    description: "Multi-speaker dialogue TTS — give it a script, it voices each role.",
+    description: "Multi-speaker dialogue via the direct ElevenLabs API — give it a script, it voices each role (any voice: premade, library, or cloned).",
     useCases: ["tts", "dialogue", "multi-speaker"],
+    features: ["audio-tags", "voice-cloning"],
     pricing: [{ identifier: "elevenlabs-dialogue", credits: 25, note: "per 1K chars" }],
-    // Driven only via the dialogue/character-voice path (multi-speaker script
-    // shape), NOT the single-text generate_speech verb. Hide from MCP
-    // list_models so generate_speech (TTS_PROVIDERS) can't advertise it and
-    // then 400. Re-expose if a dialogue-capable MCP verb is added.
-    mcpHidden: true,
   },
 
   // ── ElevenLabs voice utilities ──

package/src/model-constants.ts

@@ -211,16 +211,18 @@ export const TTS_TEXT_MAX = 5000
 
 /**
  * Per-model Text-to-Speech character cap (PER REQUEST), from official ElevenLabs
- * docs. turbo/multilingual accept FAR more than the old flat 5000; v3 uses the
- * conservative 3000 (official pages say 5000 but the API is widely reported to
- * hard-limit v3 at 3000 — conservative avoids prod rejections). Absent →
- * {@link TTS_TEXT_MAX}.
+ * docs. turbo/multilingual accept FAR more than the old flat 5000; v3 matches
+ * the official 5000 (probed live 2026-08-30: 4,500 AND 5,200 chars both
+ * returned 200 — the old "API hard-limits v3 at 3000" report no longer holds).
+ * Dialogue's documented 2,000 is a recommendation, not a limit (2,500 and
+ * 5,000 total chars both probed 200); we cap at 5000 like v3 — same model
+ * underneath. Absent → {@link TTS_TEXT_MAX}.
  */
 export const MAX_TTS_CHARS_BY_PROVIDER: Record<string, number> = {
   "elevenlabs-turbo": 40000,        // == eleven_flash_v2_5 (functionally equivalent)
   "elevenlabs-multilingual": 10000, // eleven_multilingual_v2
-  "elevenlabs-v3": 3000,            // conservative (official 5000 / API-reported 3000)
-  "elevenlabs-dialogue": 2000,      // text-to-dialogue recommended per-request max
+  "elevenlabs-v3": 5000,            // official cap (probed: 5,200 chars accepted; keep the clamp)
+  "elevenlabs-dialogue": 5000,      // total across lines; ≤2,000 recommended for best quality
 }
 
 /** Max TTS text length (chars) for a provider: verified override, else {@link TTS_TEXT_MAX}. */

package/src/model-tree.ts

@@ -62,6 +62,7 @@ const MODE_TO_NODE: Readonly<Partial<Record<ModelMode, string>>> = {
   "lip-sync": "lip-sync", "tts": "text-to-speech", "sfx": "text-to-audio",
   "music": "suno-generate", "voice-design": "voice-design", "voice-changer": "voice-changer",
   "isolation": "audio-isolation", "dubbing": "dubbing", "forced-alignment": "forced-alignment", "stt": "transcribe",
+  "dialogue": "text-to-dialogue",
   "video-analysis": "video-analysis",
 }
 

package/src/to-connected-references.ts

@@ -12,7 +12,7 @@ import { locationMentionSlug } from "./location-mention-slug.js"
 export interface EntityReferenceInput {
   /** The entity row id. */
   readonly id: string
-  readonly kind: "character" | "location" | "creature"
+  readonly kind: "character" | "location" | "creature" | "image"
   /** Display name — the slug is derived from this. */
   readonly name: string
   /** Resolved thumbnail/source URL; null/undefined → "" (placeholder-safe). */
@@ -30,7 +30,9 @@ export interface EntityReferenceInput {
  * `locationMentionSlug`), creature → `wired-creature` (no mention-slug machinery
  * — like `wired-object`, a bound creature AUTO-ATTACHES and gets a canonical-style
  * creature/animal-subject directive with zero typing; `{image:N:creature}` tokens
- * also resolve against it). Canonical-description fields are null here (a binding
+ * also resolve against it), image → `wired-image` (a plain media reference; it
+ * AUTO-ATTACHES, and an `@<name-slug>:<index>[:<role>]` mention re-seats it at the
+ * position it was typed). Canonical-description fields are null here (a binding
  * captures only name+variant; the picker/full-row path supplies descriptions).
  */
 export function toConnectedReference(entity: EntityReferenceInput): ConnectedReference {
@@ -47,6 +49,18 @@ export function toConnectedReference(entity: EntityReferenceInput): ConnectedRef
       variantDisplayName: entity.variant ?? "canonical",
     }
   }
+  if (entity.kind === "image") {
+    // No slug field — the resolver derives it from `defaultName` at prompt time
+    // (`knownImageSlugsFromRefs`), so a client cannot drift from the grammar and
+    // nothing changes on the wire.
+    return {
+      id: entity.id,
+      defaultName: entity.name,
+      source: "wired-image",
+      url: entity.url ?? "",
+      ...(entity.description ? { description: entity.description } : {}),
+    }
+  }
   if (entity.kind === "creature") {
     return {
       id: entity.id,

package/src/video-analysis.ts

@@ -334,6 +334,15 @@ const windowSceneBase = z.object({
   /** Effects on this shot's PICTURE. Absent ⇒ a clean image. */
   effects: z.array(z.enum(VIDEO_ANALYSIS_VISUAL_EFFECTS)).optional(),
   transitionOut: z.enum(VIDEO_ANALYSIS_TRANSITIONS).optional(),
+  // Round 4 (2026-08-30, recast shot craft rev 1.5 Appendix F.8): the edit into
+  // the next scene in the analyser's OWN WORDS — a few words of what it saw
+  // ("fast blurred pan left", "fade to black", "zoom-blur warp into the tunnel")
+  // — or absent when the take continues through the change or no edit is seen
+  // (nothing asserted: the video model decides from story and refs). The closed
+  // `transitionOut` vocabulary above is LEGACY from here on: kept so stored
+  // blueprints still parse, never emitted again. Readers take
+  // `transition ?? transitionOut`.
+  transition: z.string().trim().min(1).max(120).optional(),
   // Array of concurrent layers (music + speech + sfx together); [] = silence.
   audio: z.array(audioLayerSchema),
   /** slotId → variationId for slots wearing a NON-default look in this scene