@shop-os/modelkit 0.2.27 → 0.3.1

package/CHANGELOG.md

@@ -9,6 +9,37 @@ header in the release PR.
 
 ## [Unreleased]
 
+## [0.3.1] - Unreleased
+
+### Added
+- `providerOptions.modelkit.inputSpec` now works on fal image rows, so a new fal image model is a database row. The spec uses the image keys (`prompt`, `references`, `aspectRatio`, `size`, `outputFormat`, `seed`) with `fields`, `values` and `static`, exactly as on Replicate image rows. The new optional `editEndpoint` (fal rows only) names the endpoint called instead of `providerModelId` when the request has references, for models that split text-to-image and edit across two endpoints. Only the mapped fields are sent: no `num_images`, so a spec row returns one image per call. A fal row without an `inputSpec` keeps the Muse-shaped request and Ideogram keeps its own path. fal video rows still reject `inputSpec`. On fal an explicit `inputSpec` wins over the built-in Ideogram path, and `editEndpoint` needs a `references` entry in `fields`.
+- `resolveRowInputSpec(req, ctx)` and `invalidInputSpecError(provider, ctx, issues)` in the shared mapper, used by the Replicate and fal adapters.
+
+### Changed
+- fal image calls decode every image in the response that has a URL and fail only when none do. The Muse path used to fail when the first image had no URL even if later ones did.
+
+### Live rows to add after publish
+Add a fal image row that carries an `inputSpec` only after every app that will call it runs 0.3.1. An older kit accepts the row but ignores the `inputSpec`, so it sends the Muse-shaped request and the call fails or loses its fields.
+
+## [0.3.0] - Unreleased
+
+### Added
+- Row type `video-tool`: video in, video out, no prompt. `constraints.tool` says what the row does (`kind: "upscale" | "interpolate" | "hdr"`) and its limits (`upscaleFactor`, `maxOutputLongEdge`, `fps`, `outputFormats`). Rows `topaz-precision` (Proteus) and `topaz-starlight` (Starlight Precise 2.6) upscale up to 4x with a 3840 px long edge, `topaz-apollo` interpolates to 16-120 fps, and `topaz-hyperion-2.5` writes an HDR master in H.265 (HDR10) or ProRes. All four run on fal, cap the source at 300 s, and are listed on `cowork` only.
+- `kit.submitVideoTool(req, { webhookUrl })` submits to fal's queue (with `fal_webhook`) and returns a JSON `VideoToolSubmission`; `videoToolStatus`, `videoToolResult` (from a verified webhook body or the stored result) and `cancelVideoTool` take it back from any process. Measured on 2026-10-02: a 30 s 1080p-to-4K Starlight job took 2,593 s and a 30 s 4K SDR to HDR job 6,695 s, so callers must not wait inside one task. Submits are one attempt: a retry could start a second paid job. Stored URLs outside `https://queue.fal.run` are refused before the key is sent.
+- `planVideoTool(row, req)` (also on `@shop-os/modelkit/catalog`): the output size, fps, resolution tier and format a request produces, or why the row refuses it. `resolutionTier(width, height)` tiers by the short edge. `UsageEvent.fps` carries the output frame rate.
+- `falWebhookBodySchema` parses fal's webhook body.
+- `capabilities.videoEdit` marks the rows offered for prompt-driven edits: `true` on `seedance-2.5`, `gemini-omni-flash`, `gemini-omni-1.1-flash`; `false` on `seedance-2.0` and `seedance-2.0-fast`, which take a source video but are not offered for edits. A row without the flag is read as `acceptsSourceVideo` by Studio.
+- `constraints.references.audioSeconds` (`{ min?, max?, total? }`), `audioFormats` (lowercase extensions) and `audioMaxBytes` describe reference audio. Set from BytePlus's docs on `seedance-2.5` (2-30 s per clip, 30 s in all) and `seedance-2.0` / `seedance-2.0-fast` (2-15 s per clip, 15 s in all), all `wav`/`mp3` up to 15 MB.
+- `constraints.extension.returns`: `"joined"` (absent, Omni) returns the whole clip so far; `"new"` returns only the added footage. `seedance-2.5` gains `videoExtension` and `extension: { segmentSeconds: 30, maxTotalSeconds: 60, returns: "new" }`; a live extend on 2026-10-02 returned only the 4 s it added, at the requested resolution.
+
+### Live rows to add or edit after publish
+Older kits drop a `video-tool` row (unknown type) with a warning and keep loading the catalog. Pro-backend's admin API validates rows with this kit, so it must run 0.3.0 before the rows can be saved. In each environment, once pro-backend and cowork run 0.3.0 and the prices exist:
+- Add `topaz-precision`, `topaz-starlight`, `topaz-apollo` and `topaz-hyperion-2.5` as bundled.
+- `seedance-2.5`: add `videoExtension` to `constraints.supports`, set `constraints.extension` to `{ "segmentSeconds": 30, "maxTotalSeconds": 60, "returns": "new" }` and `capabilities.videoEdit` to `true`.
+- `gemini-omni-flash`, `gemini-omni-1.1-flash`: set `capabilities.videoEdit` to `true`; on `gemini-omni-1.1-flash` add `"returns": "joined"` to `constraints.extension`.
+- `seedance-2.0`, `seedance-2.0-fast`: set `capabilities.videoEdit` to `false`, and add `"audioSeconds": { "min": 2, "max": 15, "total": 15 }, "audioFormats": ["wav", "mp3"], "audioMaxBytes": 15728640` to `constraints.references`.
+- `seedance-2.5`: add `"audioSeconds": { "min": 2, "max": 30, "total": 30 }, "audioFormats": ["wav", "mp3"], "audioMaxBytes": 15728640` to `constraints.references`.
+
 ## [0.2.27] - Unreleased
 
 ### Fixed

package/README.md

@@ -51,6 +51,25 @@ await kit.generateText({
 await clip.cleanup?.();
 ```
 
+### Video tools
+
+`video-tool` rows (Topaz upscale, frame interpolation, SDR to HDR on fal) take a clip and
+return a clip with no prompt. They run as provider jobs the caller tracks, because a 30 s
+4K job can take over 40 minutes:
+
+```ts
+const plan = planVideoTool(kit.describe("topaz-precision"), { source, upscaleFactor: 2 }); // validate + price
+const sub = await kit.submitVideoTool(
+    { model: "topaz-precision", source, upscaleFactor: 2 },
+    { webhookUrl }, // fal posts here when the job ends
+);
+// store `sub` (plain JSON); later, in any process:
+const out = await kit.videoToolResult(sub, { webhook: verifiedBody }); // or poll videoToolStatus(sub)
+```
+
+`source` carries the caller's measured `width`, `height`, `durationSeconds` and `fps`; the kit
+never measures. Verifying the webhook signature is the caller's job.
+
 Keys come from `process.env` by default: bare names for pro (`OPENAI_API_KEY`, `FAL_KEY`, `ARK_API_KEY` ...), the `ENTERPRISE_` prefix for enterprise and the `SLOOSH_` prefix for sloosh. Pass `keys: { pro, enterprise, sloosh }` to override. `tier: "sloosh"` uses the sloosh keys only: a missing key is a `ModelKitError` with code `auth`, never a fallback to pro.
 
 ## Browser-safe catalog
@@ -194,8 +213,8 @@ The Bedrock model id is `<routing.bedrockProfile>.anthropic.<providerModelId>`
 breaks the pattern sets `providerOptions.modelkit.bedrockModelId` on its row.
 See the CHANGELOG entry for the caller-visible Bedrock limits.
 
-A Replicate video, image or audio row for a new model carries `providerOptions.modelkit.inputSpec`
-instead of a kit change. It is read by the Replicate adapter only. `fields` maps a canonical key to the provider field name,
+A Replicate video, image or audio row, or a fal image row, for a new model carries `providerOptions.modelkit.inputSpec`
+instead of a kit change. It is read by the Replicate and fal adapters only. `fields` maps a canonical key to the provider field name,
 `values` remaps canonical values per key, and `static` merges literal provider fields
 last. Video rows use `VIDEO_INPUT_SPEC_KEYS` (`prompt`, `durationSeconds`,
 `aspectRatio`, `resolution`, `firstFrame`, `lastFrame`, `seed`, `negativePrompt`,
@@ -204,7 +223,7 @@ last. Video rows use `VIDEO_INPUT_SPEC_KEYS` (`prompt`, `durationSeconds`,
 `seed`; map `references` to the provider's array field and remap `size` and `outputFormat`
 with `values`); audio rows use
 `AUDIO_INPUT_SPEC_KEYS` (`prompt`, `durationSeconds`, `seed`, `negativePrompt`,
-`references`, `sourceVideo`, `instrumental`). `aspectRatio: "adaptive"` is omitted.
+`references`, `sourceVideo`, `instrumental`). `aspectRatio: "adaptive"` is omitted. A fal image row has no kind and builds its whole input from the spec; the optional `editEndpoint` (fal rows only) is the fal endpoint called instead of `providerModelId` when the request has references. fal rows without an `inputSpec` keep the Muse-shaped request and Ideogram keeps its own path; an `inputSpec` on an Ideogram row replaces that path. `editEndpoint` needs a `references` entry in `fields`. `inputSpec` on fal video rows is not supported.
 `validateInputSpec` is the shared contract the write schema, `dataOnlyCapable`, and
 the Replicate mapper all check. `dataOnlyCapable(row)` reports whether such a row would
 be served.
@@ -252,7 +271,7 @@ const defaultKit = createDefaultModelKit({ keys: { pro: keysFromEnv(process.env)
 
 ## Add a model
 
-1. Add a row to `src/catalog/{image,video,text,other}.ts`. Copy `constraints` from the provider docs.
+1. Add a row to `src/catalog/{image,video,video-tools,text,other}.ts`. Copy `constraints` from the provider docs.
 2. Run `pnpm test`. The catalog test checks id and alias uniqueness.
 3. Bump `version` in `package.json` in a PR that touches only `packages/modelkit/**`.
    Merging it to `dev` runs `modelkit-publish.yml`, which publishes through the npm

package/dist/catalog.cjs

@@ -36,6 +36,7 @@ __export(catalog_entry_exports, {
   SurfaceSchema: () => SurfaceSchema,
   TEXT_MODELS: () => TEXT_MODELS,
   VIDEO_MODELS: () => VIDEO_MODELS,
+  VIDEO_TOOL_MODELS: () => VIDEO_TOOL_MODELS,
   billedAudio: () => billedAudio,
   compareCatalogOrder: () => compareCatalogOrder,
   isListed: () => isListed,
@@ -46,15 +47,18 @@ __export(catalog_entry_exports, {
   nearestAspectRatio: () => nearestAspectRatio,
   normalizeExposure: () => normalizeExposure,
   outputAudio: () => outputAudio,
+  planVideoTool: () => planVideoTool,
   primeCatalog: () => primeCatalog,
   promptConventions: () => promptConventions,
   resetLiveCatalog: () => resetLiveCatalog,
+  resolutionTier: () => resolutionTier,
   resolveAudioOptions: () => resolveAudioOptions,
   resolveImageOptions: () => resolveImageOptions,
   resolveOptions: () => resolveOptions,
   resolveVideoOptions: () => resolveVideoOptions,
   sortCatalogRows: () => sortCatalogRows,
-  videoLengthViolation: () => videoLengthViolation
+  videoLengthViolation: () => videoLengthViolation,
+  videoToolConstraints: () => videoToolConstraints
 });
 module.exports = __toCommonJS(catalog_entry_exports);
 
@@ -1353,6 +1357,11 @@ var NEGATIVE_PROMPT = "blur, distort, and low quality";
 var FRAMES_EXCLUDE_REFERENCES = { framesExcludeReferences: true };
 var KLING_PROMPT = { referenceSyntax: "numbered", maxChars: 2500 };
 var SEEDANCE_PROMPT = { referenceSyntax: "numbered", maxChars: 3e4 };
+var SEEDANCE_2_AUDIO = {
+  audioSeconds: { min: 2, max: 15, total: 15 },
+  audioFormats: ["wav", "mp3"],
+  audioMaxBytes: 15 * 1024 * 1024
+};
 var H3_PROMPT = { referenceSyntax: "numbered", maxChars: 5e4 };
 var OMNI_PROMPT = { referenceSyntax: "caption", maxChars: 5e4 };
 var BYTEDANCE_LOGO = "https://d31vvfpk2ppdyq.cloudfront.net/assets/image/3bba5d81-4e57-4e71-911b-cdfa05acca2c_compressed.png";
@@ -1489,7 +1498,7 @@ var VIDEO_MODELS = [
       description: "Google's conversational video model with native audio, human support, and natural language editing. Aspect ratio only \u2014 resolution and duration are model-determined.",
       logo: GOOGLE_LOGO2
     },
-    capabilities: { acceptsSourceVideo: true }
+    capabilities: { acceptsSourceVideo: true, videoEdit: true }
   },
   {
     id: "gemini-omni-1.1-flash",
@@ -1514,7 +1523,7 @@ var VIDEO_MODELS = [
       // with a video task, so those calls carry the prompt alone).
       // First-to-last-frame interpolation: `<LAST_FRAME>` goes with `<FIRST_FRAME>`.
       supports: ["videoExtension", "lastFrame"],
-      extension: { segmentSeconds: 10, maxTotalSeconds: 40 },
+      extension: { segmentSeconds: 10, maxTotalSeconds: 40, returns: "joined" },
       prompt: OMNI_PROMPT,
       audio: "always"
     },
@@ -1526,7 +1535,11 @@ var VIDEO_MODELS = [
       logo: GOOGLE_LOGO2,
       description: "Google Gemini Omni Flash 1.1 conversational video generation, editing and extension; text, image, and up to 10-second video input, 3-10 second output at up to 4K and 24 FPS, extendable to 40 seconds."
     },
-    capabilities: { acceptsSourceVideo: true, lastFrameRequiresFirstFrame: true }
+    capabilities: {
+      acceptsSourceVideo: true,
+      videoEdit: true,
+      lastFrameRequiresFirstFrame: true
+    }
   },
   {
     // vibe-triggers and Space Builder drop the ratio when a frame is wired; Cowork always sends it.
@@ -1548,7 +1561,8 @@ var VIDEO_MODELS = [
         videos: 3,
         audios: 3,
         audioOnly: false,
-        videoSeconds: { min: 2, max: 15, total: 15 }
+        videoSeconds: { min: 2, max: 15, total: 15 },
+        ...SEEDANCE_2_AUDIO
       },
       sourceVideoSeconds: { min: 2, max: 15, inReferenceTotal: true },
       ratioIgnoredWith: ["firstFrame", "sourceVideo"],
@@ -1570,7 +1584,7 @@ var VIDEO_MODELS = [
       description: "High-quality AI video generation with reference image conditioning. Supports 4-15 second clips with human preprocessing.",
       logo: BYTEDANCE_LOGO
     },
-    capabilities: { acceptsSourceVideo: true, ...FRAMES_EXCLUDE_REFERENCES }
+    capabilities: { acceptsSourceVideo: true, videoEdit: false, ...FRAMES_EXCLUDE_REFERENCES }
   },
   {
     id: "seedance-2.0-fast",
@@ -1590,7 +1604,8 @@ var VIDEO_MODELS = [
         videos: 3,
         audios: 3,
         audioOnly: false,
-        videoSeconds: { min: 2, max: 15, total: 15 }
+        videoSeconds: { min: 2, max: 15, total: 15 },
+        ...SEEDANCE_2_AUDIO
       },
       sourceVideoSeconds: { min: 2, max: 15, inReferenceTotal: true },
       ratioIgnoredWith: ["firstFrame", "sourceVideo"],
@@ -1612,7 +1627,7 @@ var VIDEO_MODELS = [
       logo: BYTEDANCE_LOGO,
       description: "Speed- and cost-optimized ByteDance Seedance 2.0 Fast; text plus image, video, and audio references to generate or edit 4-15 second clips."
     },
-    capabilities: { acceptsSourceVideo: true, ...FRAMES_EXCLUDE_REFERENCES }
+    capabilities: { acceptsSourceVideo: true, videoEdit: false, ...FRAMES_EXCLUDE_REFERENCES }
   },
   {
     // omni_reference_task_type edit/extend: adaptive ratio, edit sends duration -1.
@@ -1632,12 +1647,20 @@ var VIDEO_MODELS = [
         images: 30,
         videos: 10,
         audios: 10,
-        videoSeconds: { min: 2, max: 30, total: 30, editMin: 4 }
+        videoSeconds: { min: 2, max: 30, total: 30, editMin: 4 },
+        // BytePlus: 2-30 s per clip, 30 s in all, wav/mp3, 15 MB each.
+        audioSeconds: { min: 2, max: 30, total: 30 },
+        audioFormats: ["wav", "mp3"],
+        audioMaxBytes: 15 * 1024 * 1024
       },
       sourceVideoSeconds: { min: 2, max: 30, editMin: 4, inReferenceTotal: true },
       ratioIgnoredWith: ["firstFrame", "sourceVideo"],
       // No seed: 2.x echoes it but does not honour it (two seeded runs differed, 2026-10-01).
-      supports: ["generateAudio", "lastFrame", "watermark", "taskType"],
+      supports: ["generateAudio", "lastFrame", "watermark", "taskType", "videoExtension"],
+      // Extend (probe 2026-10-02, 480p): returns ONLY the added footage, continuing from
+      // the source's last frame at the source's framing and fps but at the requested
+      // resolution, so the caller joins the clips. Forward only in practice.
+      extension: { segmentSeconds: 30, maxTotalSeconds: 60, returns: "new" },
       prompt: SEEDANCE_PROMPT,
       audio: "optional"
     },
@@ -1654,7 +1677,7 @@ var VIDEO_MODELS = [
       logo: BYTEDANCE_LOGO,
       description: "ByteDance Seedance 2.5 for long-form storytelling and omni-reference generation/editing; up to 50 reference assets and 4-30 second clips at up to 1080p."
     },
-    capabilities: { acceptsSourceVideo: true, ...FRAMES_EXCLUDE_REFERENCES }
+    capabilities: { acceptsSourceVideo: true, videoEdit: true, ...FRAMES_EXCLUDE_REFERENCES }
   },
   {
     id: "grok-imagine",
@@ -2018,11 +2041,116 @@ var VIDEO_MODELS = [
   }
 ];
 
+// src/catalog/video-tools.ts
+var TOPAZ_LOGO = "https://cdn.jsdelivr.net/npm/@lobehub/icons-static-png@1.97.0/light/topazlabs.png";
+var TOPAZ_SOURCE_SECONDS = { max: 300 };
+var TOPAZ_USAGE = { unit: "seconds" };
+var VIDEO_TOOL_MODELS = [
+  {
+    id: "topaz-precision",
+    aliases: [],
+    label: "Topaz Precision",
+    type: "video-tool",
+    adapter: "fal",
+    providerModelId: "topaz/upscale/video/precision",
+    lifecycle: "active",
+    exposure: ["cowork"],
+    constraints: {
+      supports: [],
+      sourceVideoSeconds: TOPAZ_SOURCE_SECONDS,
+      tool: { kind: "upscale", upscaleFactor: { min: 1, max: 4 }, maxOutputLongEdge: 3840 }
+    },
+    defaults: {},
+    providerOptions: { fal: { model: "Proteus" } },
+    usage: TOPAZ_USAGE,
+    pricingKind: "video_per_second",
+    display: {
+      provider: "topaz",
+      logo: TOPAZ_LOGO,
+      description: "Topaz Proteus upscaling: sharper, cleaner footage up to 4K while staying faithful to the source.",
+      sortOrder: 20
+    }
+  },
+  {
+    id: "topaz-starlight",
+    aliases: [],
+    label: "Topaz Starlight",
+    type: "video-tool",
+    adapter: "fal",
+    providerModelId: "topaz/upscale/video/generative",
+    lifecycle: "active",
+    exposure: ["cowork"],
+    constraints: {
+      supports: [],
+      sourceVideoSeconds: TOPAZ_SOURCE_SECONDS,
+      tool: { kind: "upscale", upscaleFactor: { min: 1, max: 4 }, maxOutputLongEdge: 3840 }
+    },
+    defaults: {},
+    providerOptions: { fal: { model: "Starlight Precise 2.6" } },
+    usage: TOPAZ_USAGE,
+    pricingKind: "video_per_second",
+    display: {
+      provider: "topaz",
+      logo: TOPAZ_LOGO,
+      description: "Topaz Starlight Precise 2.6: generative upscaling that rebuilds fine detail, up to 4K. Slower and costlier than Precision.",
+      sortOrder: 10
+    }
+  },
+  {
+    id: "topaz-apollo",
+    aliases: [],
+    label: "Topaz Apollo",
+    type: "video-tool",
+    adapter: "fal",
+    providerModelId: "topaz/interpolate/video",
+    lifecycle: "active",
+    exposure: ["cowork"],
+    constraints: {
+      supports: [],
+      sourceVideoSeconds: TOPAZ_SOURCE_SECONDS,
+      tool: { kind: "interpolate", fps: { min: 16, max: 120 } }
+    },
+    defaults: {},
+    providerOptions: { fal: { model: "Apollo" } },
+    usage: TOPAZ_USAGE,
+    pricingKind: "video_per_second",
+    display: {
+      provider: "topaz",
+      logo: TOPAZ_LOGO,
+      description: "Topaz Apollo frame interpolation: smoother motion at up to 120 fps, same length as the source."
+    }
+  },
+  {
+    id: "topaz-hyperion-2.5",
+    aliases: [],
+    label: "Topaz Hyperion 2.5",
+    type: "video-tool",
+    adapter: "fal",
+    providerModelId: "topaz/sdr-to-hdr/video",
+    lifecycle: "active",
+    exposure: ["cowork"],
+    constraints: {
+      supports: [],
+      sourceVideoSeconds: TOPAZ_SOURCE_SECONDS,
+      tool: { kind: "hdr", outputFormats: ["h265", "prores"] }
+    },
+    defaults: {},
+    usage: TOPAZ_USAGE,
+    pricingKind: "video_per_second",
+    display: {
+      provider: "topaz",
+      logo: TOPAZ_LOGO,
+      description: "Topaz Hyperion 2.5 SDR to HDR: a 10-bit HDR master in H.265 (HDR10) or ProRes 422 HQ, at the source's size and frame rate."
+    }
+  }
+];
+
 // src/catalog/models.ts
 var MODELS = [
   ...TEXT_MODELS,
   ...IMAGE_MODELS,
   ...VIDEO_MODELS,
+  ...VIDEO_TOOL_MODELS,
   ...OTHER_MODELS
 ];
 
@@ -2189,6 +2317,10 @@ var ALL_KEY_SET = /* @__PURE__ */ new Set([
   ...IMAGE_INPUT_SPEC_KEYS,
   ...AUDIO_INPUT_SPEC_KEYS
 ]);
+var TYPES_BY_ADAPTER = /* @__PURE__ */ new Map([
+  ["replicate", ["video", "image", "audio"]],
+  ["fal", ["image"]]
+]);
 var ALLOWED_KEYS_BY_TYPE = /* @__PURE__ */ new Map([
   ["video", VIDEO_KEY_SET],
   ["image", IMAGE_KEY_SET],
@@ -2227,7 +2359,8 @@ var StaticSchema = import_zod2.z.record(import_zod2.z.string(), ProviderScalarSc
 var InputSpecSchema = import_zod2.z.strictObject({
   fields: FieldsSchema,
   values: ValuesSchema.optional(),
-  static: StaticSchema.optional()
+  static: StaticSchema.optional(),
+  editEndpoint: import_zod2.z.string().min(1).optional()
 }).superRefine((spec, ctx) => {
   const mappedNames = new Set(Object.values(spec.fields));
   for (const key of Object.keys(spec.static ?? {})) {
@@ -2252,18 +2385,21 @@ function validateInputSpec(raw, row) {
     };
   }
   const spec = parsed.data;
-  if (row.adapter !== "replicate") {
+  const supportedTypes = TYPES_BY_ADAPTER.get(row.adapter);
+  if (supportedTypes === void 0) {
     return {
       ok: false,
-      issues: [{ path: [], message: "inputSpec is only supported on replicate rows" }]
+      issues: [
+        { path: [], message: "inputSpec is only supported on replicate and fal rows" }
+      ]
     };
   }
   const issues = [];
   const allowed = ALLOWED_KEYS_BY_TYPE.get(row.type);
-  if (allowed === void 0) {
+  if (allowed === void 0 || !supportedTypes.includes(row.type)) {
     issues.push({
       path: [],
-      message: `inputSpec is not supported for "${row.type}" rows`
+      message: row.adapter === "fal" ? `inputSpec on fal rows is only supported for image rows, not "${row.type}"` : `inputSpec is not supported for "${row.type}" rows`
     });
   } else {
     for (const key of Object.keys(spec.fields)) {
@@ -2283,6 +2419,19 @@ function validateInputSpec(raw, row) {
       }
     }
   }
+  if (spec.editEndpoint !== void 0) {
+    if (row.adapter !== "fal") {
+      issues.push({
+        path: ["editEndpoint"],
+        message: "inputSpec.editEndpoint is only supported on fal rows"
+      });
+    } else if (!("references" in spec.fields)) {
+      issues.push({
+        path: ["editEndpoint"],
+        message: 'inputSpec.editEndpoint needs a "references" entry in inputSpec.fields'
+      });
+    }
+  }
   const seenNames = /* @__PURE__ */ new Map();
   for (const [key, name] of Object.entries(spec.fields)) {
     const first = seenNames.get(name);
@@ -2328,7 +2477,15 @@ var ADAPTER_IDS = [
   "higgsfield",
   "xai-video"
 ];
-var ModelTypeSchema = import_zod3.z.enum(["text", "image", "video", "audio", "upscale", "bg-remove"]);
+var ModelTypeSchema = import_zod3.z.enum([
+  "text",
+  "image",
+  "video",
+  "audio",
+  "upscale",
+  "bg-remove",
+  "video-tool"
+]);
 var LifecycleSchema = import_zod3.z.enum(["active", "deprecated", "retired"]);
 var SurfaceSchema = import_zod3.z.enum(["cowork", "space_builder"]);
 var SURFACES2 = SurfaceSchema.options;
@@ -2408,7 +2565,22 @@ var ReferenceLimitsSchema = import_zod3.z.looseObject({
   audios: import_zod3.z.number().optional(),
   total: import_zod3.z.number().optional(),
   audioOnly: import_zod3.z.boolean().optional(),
-  videoSeconds: VideoSecondsLimitsSchema.optional()
+  videoSeconds: VideoSecondsLimitsSchema.optional(),
+  audioSeconds: import_zod3.z.looseObject({
+    min: import_zod3.z.number().optional(),
+    max: import_zod3.z.number().optional(),
+    total: import_zod3.z.number().optional()
+  }).optional(),
+  audioFormats: import_zod3.z.array(import_zod3.z.string()).optional(),
+  audioMaxBytes: import_zod3.z.number().positive().optional()
+});
+var RangeSchema = import_zod3.z.looseObject({ min: import_zod3.z.number(), max: import_zod3.z.number() });
+var VideoToolConstraintsSchema = import_zod3.z.looseObject({
+  kind: import_zod3.z.enum(["upscale", "interpolate", "hdr"]),
+  upscaleFactor: RangeSchema.optional(),
+  maxOutputLongEdge: import_zod3.z.number().positive().optional(),
+  fps: RangeSchema.optional(),
+  outputFormats: import_zod3.z.array(import_zod3.z.enum(["h265", "prores"])).min(1).optional()
 });
 var ModelConstraintsSchema = import_zod3.z.looseObject({
   aspectRatios: import_zod3.z.array(AspectRatioSchema).optional(),
@@ -2423,7 +2595,14 @@ var ModelConstraintsSchema = import_zod3.z.looseObject({
   supports: import_zod3.z.array(SupportedOptionSchema).optional(),
   ratioIgnoredWith: import_zod3.z.array(import_zod3.z.enum(["firstFrame", "sourceVideo"])).optional(),
   sourceVideoSeconds: VideoSecondsLimitsSchema.omit({ total: true }).extend({ inReferenceTotal: import_zod3.z.boolean().optional() }).optional(),
-  extension: import_zod3.z.looseObject({ segmentSeconds: import_zod3.z.number(), maxTotalSeconds: import_zod3.z.number() }).optional(),
+  // segmentSeconds/maxTotalSeconds stay required: a kit before 0.3.0 drops a row whose
+  // extension lacks them, so every extension block keeps them.
+  extension: import_zod3.z.looseObject({
+    segmentSeconds: import_zod3.z.number(),
+    maxTotalSeconds: import_zod3.z.number(),
+    returns: import_zod3.z.enum(["new", "joined"]).optional()
+  }).optional(),
+  tool: VideoToolConstraintsSchema.optional(),
   // Loose on purpose: the read path keeps a row whose value is invalid (the accessors in
   // conventions.ts read it as "not set"); CatalogModelSchema below rejects it on write.
   audio: import_zod3.z.unknown().optional(),
@@ -2495,6 +2674,7 @@ var CatalogDisplaySchema = import_zod3.z.looseObject({
 });
 var CatalogCapabilitiesSchema = import_zod3.z.looseObject({
   acceptsSourceVideo: import_zod3.z.boolean().optional(),
+  videoEdit: import_zod3.z.boolean().optional(),
   framesExcludeReferences: import_zod3.z.boolean().optional(),
   lastFrameRequiresFirstFrame: import_zod3.z.boolean().optional(),
   reasoning: import_zod3.z.boolean().optional(),
@@ -2537,6 +2717,13 @@ var CatalogModelSchemaBase = import_zod3.z.looseObject({
       message: "active text rows need contextWindow"
     });
   }
+  if (row.type === "video-tool" && row.constraints.tool === void 0) {
+    ctx.addIssue({
+      code: "custom",
+      path: ["constraints", "tool"],
+      message: "video-tool rows need constraints.tool"
+    });
+  }
   if (row.replacedBy !== void 0 && row.replacedBy === row.id) {
     ctx.addIssue({
       code: "custom",
@@ -2842,6 +3029,114 @@ function resetLiveCatalog() {
   liveCatalog.replace([...MODELS]);
 }
 
+// src/catalog/video-tools-spec.ts
+var import_zod6 = require("zod");
+var RangeSchema2 = import_zod6.z.object({ min: import_zod6.z.number(), max: import_zod6.z.number() });
+var ToolSchema = import_zod6.z.object({
+  kind: import_zod6.z.enum(["upscale", "interpolate", "hdr"]),
+  upscaleFactor: RangeSchema2.optional(),
+  maxOutputLongEdge: import_zod6.z.number().positive().optional(),
+  fps: RangeSchema2.optional(),
+  outputFormats: import_zod6.z.array(import_zod6.z.enum(["h265", "prores"])).min(1).optional()
+});
+function videoToolConstraints(row) {
+  if (row.type !== "video-tool") return void 0;
+  const parsed = ToolSchema.safeParse(row.constraints.tool);
+  if (!parsed.success) return void 0;
+  const { kind, upscaleFactor, maxOutputLongEdge, fps, outputFormats } = parsed.data;
+  return {
+    kind,
+    ...upscaleFactor !== void 0 && { upscaleFactor },
+    ...maxOutputLongEdge !== void 0 && { maxOutputLongEdge },
+    ...fps !== void 0 && { fps },
+    ...outputFormats !== void 0 && { outputFormats }
+  };
+}
+function resolutionTier(width, height) {
+  const shortEdge = Math.min(width, height);
+  if (shortEdge <= 360) return "360p";
+  if (shortEdge <= 480) return "480p";
+  if (shortEdge <= 720) return "720p";
+  if (shortEdge <= 1080) return "1080p";
+  if (shortEdge <= 1440) return "2k";
+  return "4k";
+}
+function fail(error) {
+  return { ok: false, error };
+}
+function planVideoTool(row, req) {
+  const tool = videoToolConstraints(row);
+  if (!tool) return fail(`${row.id} is not a video tool`);
+  const { width, height, durationSeconds, fps } = req.source;
+  if (!(width > 0 && height > 0 && durationSeconds > 0)) {
+    return fail("The source video's width, height and duration are required");
+  }
+  const maxSeconds = row.constraints.sourceVideoSeconds?.max;
+  if (maxSeconds !== void 0 && durationSeconds > maxSeconds) {
+    return fail(`${row.label} takes source videos up to ${maxSeconds} seconds.`);
+  }
+  const base = {
+    durationSeconds,
+    ...fps !== void 0 && { fps }
+  };
+  if (tool.kind === "upscale") {
+    const factor = req.upscaleFactor;
+    if (factor === void 0) return fail("An upscale factor is required");
+    const range = tool.upscaleFactor;
+    if (range && (factor < range.min || factor > range.max)) {
+      return fail(`${row.label} upscales by ${range.min}x to ${range.max}x.`);
+    }
+    const outWidth = Math.round(width * factor);
+    const outHeight = Math.round(height * factor);
+    const limit = tool.maxOutputLongEdge;
+    if (limit !== void 0 && Math.max(outWidth, outHeight) > limit) {
+      return fail(
+        `A ${factor}x upscale of a ${width}x${height} video passes the ${limit}px limit.`
+      );
+    }
+    return {
+      ok: true,
+      output: {
+        ...base,
+        width: outWidth,
+        height: outHeight,
+        resolution: resolutionTier(outWidth, outHeight),
+        format: "h264"
+      }
+    };
+  }
+  if (tool.kind === "interpolate") {
+    const target = req.targetFps;
+    if (target === void 0 || !Number.isInteger(target)) {
+      return fail("A whole-number target frame rate is required");
+    }
+    const range = tool.fps;
+    if (range && (target < range.min || target > range.max)) {
+      return fail(`${row.label} outputs ${range.min} to ${range.max} fps.`);
+    }
+    return {
+      ok: true,
+      output: {
+        durationSeconds,
+        fps: target,
+        width,
+        height,
+        resolution: resolutionTier(width, height),
+        format: "h264"
+      }
+    };
+  }
+  const formats = tool.outputFormats ?? ["h265"];
+  const format = req.outputFormat ?? formats[0] ?? "h265";
+  if (!formats.includes(format)) {
+    return fail(`${row.label} writes ${formats.join(" or ")}.`);
+  }
+  return {
+    ok: true,
+    output: { ...base, width, height, resolution: resolutionTier(width, height), format }
+  };
+}
+
 // src/resolve-options.ts
 var SIZE_ORDER = ["512", "1K", "2K", "4K"];
 var RESOLUTION_ORDER = ["360p", "480p", "720p", "768p", "1080p", "2k", "4k"];