@nodaro/shared 1.18.0 → 1.20.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nodaro/shared",
-  "version": "1.18.0",
+  "version": "1.20.0",
   "description": "Shared types, model catalog, wire contracts, and structural vocabularies for the Nodaro platform and SDK.",
   "type": "module",
   "license": "Apache-2.0",

package/src/__tests__/llm-models.test.ts

@@ -20,12 +20,36 @@ import { PIPELINE_PINNABLE_SCRIPT_LLMS } from "../pipeline-types.js"
 // the NON-monetary model registry (ids, capabilities, tiers, feature
 // defaults) that stays in the published package.
 
+/** The registry's expected membership, in order — the single list both the
+ *  count guard and the id-order guard below assert against, so a model
+ *  addition can never leave one of them stale. */
+const EXPECTED_MODEL_IDS = [
+  "gemini-3-flash",
+  "gemini-3.6-flash",
+  "claude-haiku-4.5",
+  "claude-sonnet-4.6",
+  "gpt-5.2",
+  "gemini-3.1-pro",
+  "claude-opus-4.7",
+  "gpt-5.4",
+  "gpt-5.5",
+  "gpt-5.6-luna",
+  "gpt-5.6-terra",
+  "gpt-5.6-sol",
+  "claude-sonnet-5",
+  "claude-opus-4.8",
+  "claude-opus-5",
+  "claude-fable-5",
+]
+
 // ---------------------------------------------------------------------------
 // LLM_MODELS data integrity
 // ---------------------------------------------------------------------------
 describe("LLM_MODELS data integrity", () => {
-  it("should have exactly 15 models", () => {
-    expect(LLM_MODELS).toHaveLength(16)
+  it("model count matches the enumerated id list", () => {
+    // Bound to the id list asserted below rather than a hand-typed number, so
+    // the count and the title can never disagree after a model addition.
+    expect(LLM_MODELS).toHaveLength(EXPECTED_MODEL_IDS.length)
   })
 
   it("each model has all required fields", () => {
@@ -106,25 +130,7 @@ describe("LLM_MODEL_IDS", () => {
   })
 
   it("contains all expected model ids", () => {
-    const expected = [
-      "gemini-3-flash",
-      "gemini-3.6-flash",
-      "claude-haiku-4.5",
-      "claude-sonnet-4.6",
-      "gpt-5.2",
-      "gemini-3.1-pro",
-      "claude-opus-4.7",
-      "gpt-5.4",
-      "gpt-5.5",
-      "gpt-5.6-luna",
-      "gpt-5.6-terra",
-      "gpt-5.6-sol",
-      "claude-sonnet-5",
-      "claude-opus-4.8",
-      "claude-opus-5",
-      "claude-fable-5",
-    ]
-    expect(LLM_MODEL_IDS).toEqual(expected)
+    expect(LLM_MODEL_IDS).toEqual(EXPECTED_MODEL_IDS)
   })
 })
 
@@ -522,6 +528,25 @@ describe("reasoning effort registry", () => {
     expect(getLlmTier("gpt-5.5")).toBe("premium")
     expect(getLlmTier("gemini-3.6-flash")).toBe("economy")
     expect(getLlmTier("claude-fable-5")).toBe("premium")
+    expect(getLlmTier("claude-opus-5")).toBe("premium")
+  })
+
+  it("claude-opus-5 carries the full ladder, its own direct fallback, and thinkingDefaultOn", () => {
+    const m = getLlmModel("claude-opus-5")
+    expect(m?.reasoningEfforts).toEqual(["low", "medium", "high", "xhigh", "max"])
+    expect(m?.directFallbackModel).toBe("claude-opus-5")
+    // Load-bearing: consumers floor max_tokens off this flag, because Opus 5
+    // reasons even with NO thinking param sent. Dropping it silently
+    // reintroduces truncated answers at Effort=Auto on a premium model.
+    expect(m?.thinkingDefaultOn).toBe(true)
+  })
+
+  it("thinkingDefaultOn is set only where the vendor default actually reasons", () => {
+    // Opus 4.8 / 4.7 and Sonnet 5 do NOT reason when `thinking` is omitted —
+    // flagging them would inflate every call's cap for no reason.
+    for (const id of ["claude-opus-4.8", "claude-opus-4.7", "claude-sonnet-5", "claude-sonnet-4.6"]) {
+      expect(getLlmModel(id)?.thinkingDefaultOn, id).toBeUndefined()
+    }
   })
 
   it("gemini-3.6-flash exposes the KIE low/high thinking levels; claude-fable-5 the full Claude ladder", () => {

package/src/__tests__/smart-cut-windows.test.ts

@@ -0,0 +1,43 @@
+import { describe, it, expect } from "vitest"
+import {
+  clampSmartCutWindow,
+  SMART_CUT_WINDOW_MIN,
+  SMART_CUT_WINDOW_MAX,
+  SMART_CUT_WINDOW_DEFAULT,
+} from "../smart-cut-windows.js"
+
+describe("smart-cut search windows", () => {
+  it("keeps in-range integers untouched", () => {
+    for (const n of [1, 4, 8, 12, 24]) expect(clampSmartCutWindow(n)).toBe(n)
+  })
+
+  it("clamps to the product bounds instead of erroring", () => {
+    expect(clampSmartCutWindow(999)).toBe(SMART_CUT_WINDOW_MAX)
+    expect(clampSmartCutWindow(25)).toBe(SMART_CUT_WINDOW_MAX)
+    expect(clampSmartCutWindow(0)).toBe(SMART_CUT_WINDOW_MIN)
+    expect(clampSmartCutWindow(-5)).toBe(SMART_CUT_WINDOW_MIN)
+  })
+
+  it("rounds fractional input (a number <input> can emit one)", () => {
+    expect(clampSmartCutWindow(8.4)).toBe(8)
+    expect(clampSmartCutWindow(8.6)).toBe(9)
+  })
+
+  it('returns undefined — "use the engine default" — for anything non-numeric', () => {
+    // Load-bearing: undefined means the field is OMITTED from the request, so
+    // the engine applies its own 8/8. Coercing junk to a NUMBER instead would
+    // silently run a search width the user never chose.
+    for (const v of [undefined, null, NaN, Infinity, "8", "", {}, []]) {
+      expect(clampSmartCutWindow(v), String(v)).toBeUndefined()
+    }
+  })
+
+  it("bounds are coherent and the default sits inside them", () => {
+    expect(SMART_CUT_WINDOW_MIN).toBeLessThan(SMART_CUT_WINDOW_MAX)
+    expect(SMART_CUT_WINDOW_DEFAULT).toBeGreaterThanOrEqual(SMART_CUT_WINDOW_MIN)
+    expect(SMART_CUT_WINDOW_DEFAULT).toBeLessThanOrEqual(SMART_CUT_WINDOW_MAX)
+    // The engine route accepts up to 48; the product cap must stay within it
+    // or a clamped value would still 400 at the boundary we were protecting.
+    expect(SMART_CUT_WINDOW_MAX).toBeLessThanOrEqual(48)
+  })
+})

package/src/index.ts

@@ -889,5 +889,8 @@ export * from "./video-analysis.js"
 // --- Video-analysis pricing (duration buckets + structural credit formula) ---
 export * from "./video-analysis-pricing.js"
 
+// --- Smart-cut best-pair search windows (shared bound + clamp) ---
+export * from "./smart-cut-windows.js"
+
 export * from "./entity-asset-types.js"
 export * from "./hint-graph-types.js"

package/src/llm-models.ts

@@ -57,6 +57,20 @@ export interface LlmModelDef {
   supportsTemperature?: false
   /** Claude-only: KIE is the preferred routing, direct Anthropic the fallback. */
   preferKie?: true
+  /**
+   * true = the model reasons even when NO thinking parameter is sent, so its
+   * reasoning tokens share the `max_tokens` budget on EVERY call — not just
+   * effort-bearing ones. Claude Opus 5 flipped this default (on Opus 4.8/4.7
+   * and Sonnet 5, omitting `thinking` means no thinking at all).
+   *
+   * Consumers MUST give such a model output headroom regardless of the
+   * requested effort, or reasoning silently eats a small legacy cap and the
+   * answer truncates with `stop_reason: max_tokens` — a paid-for empty reply,
+   * not an error. `deriveParams` (llm-client.ts) and the film-pipeline's
+   * `callLLM` both floor on this flag; keep it in sync with the vendor's
+   * documented default rather than inferring it from the model name.
+   */
+  thinkingDefaultOn?: true
 }
 
 export const LLM_MODELS: readonly LlmModelDef[] = [
@@ -143,7 +157,11 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
   {
     id: "claude-opus-4.7",
     displayName: "Claude Opus 4.7",
-    desc: "Highest quality, complex tasks",
+    // Superlatives belong to the CURRENT top model only — `desc` renders in every
+    // model picker, so leaving "highest quality" on an older Opus steers
+    // quality-critical work backwards (all three Opus entries bill premium, so
+    // the mis-steer is invisible on the invoice).
+    desc: "Older Opus, complex tasks",
     tier: "premium",
     kieFormat: "messages",
     kieSlugOrModel: "claude-opus-4-7",
@@ -276,6 +294,9 @@ export const LLM_MODELS: readonly LlmModelDef[] = [
     reasoningEfforts: ["low", "medium", "high", "xhigh", "max"],
     supportsTemperature: false,
     preferKie: true,
+    // Opus 5 reasons with NO thinking param sent (vendor default flipped from
+    // Opus 4.8/4.7) — so every call needs output headroom, not just xhigh/max.
+    thinkingDefaultOn: true,
   },
   {
     id: "claude-fable-5",

package/src/model-catalog.ts

@@ -1812,14 +1812,14 @@ const VIDEO_MODELS: Record<string, ModelCatalogEntry> = {
     family: "Nodaro",
     label: "Video Analysis (Fast — legacy)",
     series: "Video Analysis",
-    description: "Legacy fast-tier analysis model (pre-2026-07). Kept so stored raw-model configs keep running and pricing unchanged; new fast-tier runs use the current fast model.",
+    description: "Legacy fast-tier analysis model (pre-2026-07). Kept so stored raw-model configs keep running and keep pricing under their own identifier; new fast-tier runs use the current fast model.",
     useCases: ["video-analysis", "shot-list", "fast"],
     pricing: [
-      { identifier: "video-analysis:gemini-3-flash", credits: 3, note: "10-min ceiling (no duration given)" },
-      { identifier: "video-analysis:gemini-3-flash:60s", credits: 1 },
-      { identifier: "video-analysis:gemini-3-flash:180s", credits: 1 },
-      { identifier: "video-analysis:gemini-3-flash:360s", credits: 2 },
-      { identifier: "video-analysis:gemini-3-flash:600s", credits: 3, note: "10-min ceiling" },
+      { identifier: "video-analysis:gemini-3-flash", credits: 9, note: "10-min ceiling (no duration given)" },
+      { identifier: "video-analysis:gemini-3-flash:60s", credits: 2 },
+      { identifier: "video-analysis:gemini-3-flash:180s", credits: 3 },
+      { identifier: "video-analysis:gemini-3-flash:360s", credits: 6 },
+      { identifier: "video-analysis:gemini-3-flash:600s", credits: 9, note: "10-min ceiling" },
     ],
   },
   "gemini-3.6-flash-video-analysis": {
@@ -1832,11 +1832,11 @@ const VIDEO_MODELS: Record<string, ModelCatalogEntry> = {
     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-3.6-flash", credits: 9, note: "10-min ceiling (no duration given)" },
-      { identifier: "video-analysis:gemini-3.6-flash:60s", credits: 2 },
-      { identifier: "video-analysis:gemini-3.6-flash:180s", credits: 2 },
-      { identifier: "video-analysis:gemini-3.6-flash:360s", credits: 5 },
-      { identifier: "video-analysis:gemini-3.6-flash:600s", credits: 9, note: "10-min ceiling" },
+      { identifier: "video-analysis:gemini-3.6-flash", credits: 25, note: "10-min ceiling (no duration given)" },
+      { identifier: "video-analysis:gemini-3.6-flash:60s", credits: 5 },
+      { identifier: "video-analysis:gemini-3.6-flash:180s", credits: 6 },
+      { identifier: "video-analysis:gemini-3.6-flash:360s", credits: 15 },
+      { identifier: "video-analysis:gemini-3.6-flash:600s", credits: 25, note: "10-min ceiling" },
     ],
   },
   "gemini-3.1-pro-video-analysis": {
@@ -1849,11 +1849,11 @@ const VIDEO_MODELS: Record<string, ModelCatalogEntry> = {
     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)" },
-      { identifier: "video-analysis:gemini-3.1-pro:60s", credits: 2 },
-      { identifier: "video-analysis:gemini-3.1-pro:180s", credits: 3 },
-      { identifier: "video-analysis:gemini-3.1-pro:360s", credits: 7 },
-      { identifier: "video-analysis:gemini-3.1-pro:600s", credits: 11, note: "10-min ceiling" },
+      { identifier: "video-analysis:gemini-3.1-pro", credits: 33, note: "10-min ceiling (no duration given)" },
+      { identifier: "video-analysis:gemini-3.1-pro:60s", credits: 6 },
+      { identifier: "video-analysis:gemini-3.1-pro:180s", credits: 8 },
+      { identifier: "video-analysis:gemini-3.1-pro:360s", credits: 20 },
+      { identifier: "video-analysis:gemini-3.1-pro:600s", credits: 33, note: "10-min ceiling" },
     ],
   },
   // Both mixed tiers are variants of the same advanced multi-engine analysis
@@ -1869,11 +1869,11 @@ const VIDEO_MODELS: Record<string, ModelCatalogEntry> = {
     description: "Our most advanced analysis tier — multiple analysis engines combined into one result for maximum completeness and accuracy. Billed per duration bucket.",
     useCases: ["video-analysis", "shot-list", "premium", "most-complete"],
     pricing: [
-      { identifier: "video-analysis:mixed", credits: 14, note: "10-min ceiling (no duration given)" },
-      { identifier: "video-analysis:mixed:60s", credits: 3 },
-      { identifier: "video-analysis:mixed:180s", credits: 4 },
-      { identifier: "video-analysis:mixed:360s", credits: 9 },
-      { identifier: "video-analysis:mixed:600s", credits: 14, note: "10-min ceiling" },
+      { identifier: "video-analysis:mixed", credits: 57, note: "10-min ceiling (no duration given)" },
+      { identifier: "video-analysis:mixed:60s", credits: 10 },
+      { identifier: "video-analysis:mixed:180s", credits: 13 },
+      { identifier: "video-analysis:mixed:360s", credits: 35 },
+      { identifier: "video-analysis:mixed:600s", credits: 57, note: "10-min ceiling" },
     ],
   },
 }

package/src/smart-cut-windows.ts

@@ -0,0 +1,50 @@
+/**
+ * Smart-cut best-pair SEARCH WINDOWS — the shared bound + clamp for
+ * generate-video-pro's `smartCutFramesPrev` / `smartCutFramesNext`.
+ *
+ * What they do: the best-pair matcher (the mode the node calls "legacy-8x8")
+ * PSNR-compares the last N frames of a segment against the first M of the
+ * next, ends the previous clip ON the best match and starts the next right
+ * AFTER its twin — so the duplicated frame plays once and motion stays
+ * continuous. N and M are those windows. Absent → the engine's own 8/8
+ * default, which is byte-identical to the behavior before they were
+ * exposed.
+ *
+ * Why wider helps: a continuation can re-enact a longer stretch of the
+ * previous tail than 8 frames covers, and a match outside the window is
+ * simply never found — the boundary silently falls back to the fixed
+ * freeze-trims. recast pins 24/24 for exactly this reason.
+ *
+ * Why a shared clamp: the canvas node (single-node Run) and the orchestrator
+ * (workflow Run) are two independent send paths into the same engine route,
+ * whose Zod schema rejects out-of-range values. Deriving both from this one
+ * function means a stale or hand-edited node value degrades to a legal
+ * request instead of 400-ing an entire multi-segment run at finalize time —
+ * and the two paths cannot drift apart.
+ */
+
+/** Widest window the UI offers. The engine route itself accepts up to 48;
+ *  24 is the product cap — it already covers a full second of re-enactment
+ *  at 24fps, and every frame added past the real overlap only costs match
+ *  time and invites a spurious pairing. */
+export const SMART_CUT_WINDOW_MAX = 24
+/** Narrowest meaningful window — one frame each side. */
+export const SMART_CUT_WINDOW_MIN = 1
+/** The engine's default when the field is absent. Documented here so the UI
+ *  can show it as the placeholder without hardcoding a second copy. */
+export const SMART_CUT_WINDOW_DEFAULT = 8
+
+/**
+ * Narrow an arbitrary node value to a legal window, or `undefined` to mean
+ * "let the engine use its default". Non-numbers, NaN, and non-integers all
+ * collapse to `undefined` rather than to a guessed number: a malformed value
+ * should fall back to the proven default, not silently pick a different
+ * search width than the user asked for.
+ */
+export function clampSmartCutWindow(value: unknown): number | undefined {
+  if (typeof value !== "number" || !Number.isFinite(value)) return undefined
+  const n = Math.round(value)
+  if (n < SMART_CUT_WINDOW_MIN) return SMART_CUT_WINDOW_MIN
+  if (n > SMART_CUT_WINDOW_MAX) return SMART_CUT_WINDOW_MAX
+  return n
+}

package/src/video-analysis-pricing.ts

@@ -47,27 +47,27 @@ export const VIDEO_ANALYSIS_WINDOW = { LEN: WINDOW_LEN, STRIDE: WINDOW_STRIDE, O
  */
 export const VIDEO_ANALYSIS_BUCKET_CREDITS: Record<string, number> = {
   // Legacy fast-tier model (pre-2026-07) — kept for stored raw-id configs.
-  "video-analysis:gemini-3-flash:60s": 1,
-  "video-analysis:gemini-3-flash:180s": 1,
-  "video-analysis:gemini-3-flash:360s": 2,
-  "video-analysis:gemini-3-flash:600s": 3,
+  "video-analysis:gemini-3-flash:60s": 2,
+  "video-analysis:gemini-3-flash:180s": 3,
+  "video-analysis:gemini-3-flash:360s": 6,
+  "video-analysis:gemini-3-flash:600s": 9,
   // Current fast tier — regenerated from the private formula for its backing
   // model; higher than the legacy fast schedule but still ≤ pro per bucket.
-  "video-analysis:gemini-3.6-flash:60s": 2,
-  "video-analysis:gemini-3.6-flash:180s": 2,
-  "video-analysis:gemini-3.6-flash:360s": 5,
-  "video-analysis:gemini-3.6-flash:600s": 9,
-  "video-analysis:gemini-3.1-pro:60s": 2,
-  "video-analysis:gemini-3.1-pro:180s": 3,
-  "video-analysis:gemini-3.1-pro:360s": 7,
-  "video-analysis:gemini-3.1-pro:600s": 11,
+  "video-analysis:gemini-3.6-flash:60s": 5,
+  "video-analysis:gemini-3.6-flash:180s": 6,
+  "video-analysis:gemini-3.6-flash:360s": 15,
+  "video-analysis:gemini-3.6-flash:600s": 25,
+  "video-analysis:gemini-3.1-pro:60s": 6,
+  "video-analysis:gemini-3.1-pro:180s": 8,
+  "video-analysis:gemini-3.1-pro:360s": 20,
+  "video-analysis:gemini-3.1-pro:600s": 33,
   // Mixed tiers (`mixed` + `mixed-fast`) share ONE credit family — they are
   // variants of the same engine plan (plan internals live in the private
   // analysis plugin). Admin-tunable via model_pricing like every other row.
-  "video-analysis:mixed:60s": 3,
-  "video-analysis:mixed:180s": 4,
-  "video-analysis:mixed:360s": 9,
-  "video-analysis:mixed:600s": 14,
+  "video-analysis:mixed:60s": 10,
+  "video-analysis:mixed:180s": 13,
+  "video-analysis:mixed:360s": 35,
+  "video-analysis:mixed:600s": 57,
 }
 
 /**