@shop-os/modelkit 0.2.9 → 0.2.11

package/CHANGELOG.md

@@ -9,6 +9,26 @@ header in the release PR.
 
 ## [Unreleased]
 
+## [0.2.11] - 2026-09-17
+
+### Added
+- Anthropic on Bedrock transport: InvokeModel via `createBedrockAnthropic`, with the same call surface as the direct route. The transport of a row is its `providerOptions.modelkit.transport` (`"bedrock"` | `"direct"`), then the app's `routing.anthropic`, then `"direct"`. Nothing changes for an app until a row or its routing says `"bedrock"`. The row value moves a model with a catalog edit and no deploy, in either direction, for example back to the Anthropic API during a Bedrock outage.
+- Credential check: when the chosen transport has no credentials in the app and the other one has, the kit uses the other one and warns once, so a catalog edit cannot take an app down. An unknown row transport is ignored with a warning. `CatalogModelSchema` (the write schema) rejects an unknown transport; `CatalogSchema` (the read schema) tolerates it, so a value from a newer kit never makes an older kit reject the catalog.
+- Bedrock model ids are derived: `<routing.bedrockProfile>.anthropic.<providerModelId>`, profile default `"us"`. The dated Haiku 4.5 and Sonnet 4.5 ids are built in. A row overrides with `providerOptions.modelkit.bedrockModelId`, and an already-prefixed `providerModelId` passes through unchanged. A new Claude row needs no kit change.
+- `keys.aws` shape: `region` plus `apiKey` (Bedrock bearer token, wins over SigV4), `accessKeyId`/`secretAccessKey`/`sessionToken`, or `credentialProvider` for task roles and other chains. `hasAwsCredentials` reports whether a tier holds usable credentials, and enterprise-to-pro fallback treats `aws` like every other key. `keysFromEnv` reads `BEDROCK_REGION` (default `us-east-1`, never `AWS_REGION`), `AWS_BEARER_TOKEN_BEDROCK`, `AWS_ACCESS_KEY_ID`/`AWS_SECRET_ACCESS_KEY` plus `AWS_SESSION_TOKEN`, under the bare and `ENTERPRISE_` prefixes. An app that has the AWS key pair for S3 uses that IAM user for Bedrock, so it needs `bedrock:InvokeModel` and `bedrock:InvokeModelWithResponseStream` before a row or the routing moves it to Bedrock.
+- Billing guard: the Bedrock model reports the canonical catalog id with provider `anthropic.messages`, so usage events match the direct route. `response.modelId` is pinned to the same id for `generate` and `stream` calls, because Bedrock echoes its own id for dated models (`anthropic.claude-haiku-4-5-20251001-v1:0`). Call-level `providerOptions.anthropic.cacheControl` is dropped with one warning per handle (Bedrock rejects it); message-level cacheControl and every other Anthropic option pass through.
+
+Known Bedrock limits for callers: no `web_search`/`web_fetch` server tools, no Files API, no `anthropic-beta` headers. Image/file URLs are downloaded by the AI SDK in the caller's process and sent as base64 (5 MB per image); the request cap is 20 MB.
+
+## [0.2.10] - 2026-09-16
+
+### Added
+- `VideoRequest.taskType` (`reference` | `edit` | `extend`) for Seedance omni-reference subtasks. `edit` sends `duration: -1` and `ratio: adaptive`; `extend` sends `ratio: adaptive`; both need a source or reference video and fail before submit otherwise. `seedance-2.5` also sends `omni_reference_task_type` so BytePlus validates at submit time instead of failing the task later with `InvalidParameter.TaskTypeConstraint`. The 2.0 rows do not list `taskType`.
+- `ReferenceLimits.imagesWithVideo`: a lower image cap when a reference video is sent. `kling-v3-omni-video` declares 7 / 4 per the Replicate schema.
+
+### Fixed
+- Omni usage bills 10 s when the caller sends no duration; the model renders 10 s by default, not 8 (live probe 2026-09-16). The `duration` response_format field is undocumented but honored, so it stays.
+
 ## [0.2.9] - 2026-09-15
 
 ### Changed

package/README.md

@@ -127,6 +127,67 @@ Opt into thinking per call, for example
 The kit keeps its `Warning` type for coercions; the SDK's warning type is `AiWarning`.
 `languageModel()` remains the low-level handle for Mastra and custom middleware; its `providerOptions` carry no kit-internal keys.
 
+## Anthropic on Bedrock
+
+Anthropic catalog rows can go through AWS Bedrock InvokeModel instead of the
+Anthropic API. The call surface is the same, and usage events and
+`response.modelId` still carry the catalog id.
+
+The transport of a row is decided in this order:
+
+1. The row: `providerOptions.modelkit.transport` = `"bedrock"` or `"direct"`. An
+   operator edits it in pro-admin and every app follows at its next catalog
+   refresh, with no deploy. It works in both directions, so it is also the
+   switch back during a Bedrock outage.
+2. The app: `routing.anthropic`.
+3. The default: `"direct"`, the Anthropic API.
+
+A catalog edit can never take an app down. When the chosen transport has no
+credentials in the app and the other one has, the kit uses the other one and
+logs one warning. Keep both `ANTHROPIC_API_KEY` and the AWS credentials set so
+the switch has somewhere to go.
+
+Credentials come from `keys.aws` (`region` plus an `apiKey`, a static pair, or
+a `credentialProvider`), or from the environment through `keysFromEnv`
+(`BEDROCK_REGION`, default `us-east-1`; `AWS_BEARER_TOKEN_BEDROCK`, or
+`AWS_ACCESS_KEY_ID`/`AWS_SECRET_ACCESS_KEY` plus `AWS_SESSION_TOKEN`).
+`keysFromEnv` picks up the AWS key pair an app already has for S3, so that IAM
+user needs `bedrock:InvokeModel` and `bedrock:InvokeModelWithResponseStream`
+before a row or the routing moves it to Bedrock. Bedrock rows never go through
+the Cloudflare gateway.
+
+The Bedrock model id is `<routing.bedrockProfile>.anthropic.<providerModelId>`
+(profile default `us`). The older dated models are built in. A model that
+breaks the pattern sets `providerOptions.modelkit.bedrockModelId` on its row.
+See the CHANGELOG entry for the caller-visible Bedrock limits.
+
+```ts
+import { defaultProvider } from "@aws-sdk/credential-provider-node";
+import { createDefaultModelKit, keysFromEnv } from "@shop-os/modelkit";
+
+// Task roles and other credential chains. The Anthropic key stays as the fallback.
+const kit = createDefaultModelKit({
+    keys: {
+        pro: {
+            ...keysFromEnv(process.env),
+            aws: { region: "us-east-1", credentialProvider: defaultProvider() },
+        },
+    },
+    routing: { anthropic: "bedrock" },
+});
+
+// Static keys or a bearer token from the environment:
+// BEDROCK_REGION=us-east-1 plus AWS_BEARER_TOKEN_BEDROCK, or
+// AWS_ACCESS_KEY_ID + AWS_SECRET_ACCESS_KEY (+ AWS_SESSION_TOKEN).
+const envKit = createDefaultModelKit({
+    keys: { pro: keysFromEnv(process.env) },
+    routing: { anthropic: "bedrock" },
+});
+
+// No routing: the Anthropic API, until a catalog row sets transport "bedrock".
+const defaultKit = createDefaultModelKit({ keys: { pro: keysFromEnv(process.env) } });
+```
+
 ## What the package owns
 
 | Concern     | Rule                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                    |

package/dist/catalog.cjs

@@ -1520,6 +1520,7 @@ var VIDEO_MODELS = [
   },
   {
     // generateAudio defaults to false here (trigger-workflows); Space Builder defaults it to true.
+    // omni_reference_task_type edit/extend: adaptive ratio, edit sends duration -1.
     id: "seedance-2.5",
     aliases: ["seedance-2-5", "dreamina-seedance-2-5-260628"],
     label: "Seedance 2.5",
@@ -1534,7 +1535,7 @@ var VIDEO_MODELS = [
       durations: { min: 4, max: 30 },
       references: { images: 30, videos: 10, audios: 10 },
       ratioIgnoredWith: ["firstFrame", "sourceVideo"],
-      supports: ["generateAudio", "seed", "lastFrame", "watermark"]
+      supports: ["generateAudio", "seed", "lastFrame", "watermark", "taskType"]
     },
     defaults: {
       durationSeconds: 5,
@@ -1643,6 +1644,9 @@ var VIDEO_MODELS = [
   },
   {
     // mode standard/pro selects 720p/1080p; the endpoint takes no separate resolution.
+    // Replicate: "Max 7 [images] without video, 4 with video"; 4K does not support
+    // reference_video. No `videos` limit yet: the adapter sends no reference video,
+    // so `imagesWithVideo` stays dormant until one is wired through.
     id: "kling-v3-omni-video",
     aliases: ["kwaivgi/kling-v3-omni-video", "kling-v3-omni"],
     label: "Kling V3 Omni",
@@ -1655,7 +1659,7 @@ var VIDEO_MODELS = [
       aspectRatios: [...KLING_RATIOS],
       resolutions: ["720p", "1080p"],
       durations: { min: 3, max: 15 },
-      references: { images: 7 },
+      references: { images: 7, imagesWithVideo: 4 },
       supports: ["generateAudio", "lastFrame"]
     },
     defaults: { durationSeconds: 5, aspectRatio: "16:9", generateAudio: false, mode: "pro" },
@@ -1931,6 +1935,7 @@ var SupportedOptionSchema = import_zod.z.enum([
   "upscale",
   "watermark",
   "safetyChecker",
+  "taskType",
   "instrumental"
 ]);
 var UsageUnitSchema = import_zod.z.enum([
@@ -1953,6 +1958,7 @@ var ProviderOptionValueSchema = import_zod.z.lazy(
 var ProviderOptionsSchema = import_zod.z.record(import_zod.z.string(), import_zod.z.record(import_zod.z.string(), ProviderOptionValueSchema));
 var ReferenceLimitsSchema = import_zod.z.looseObject({
   images: import_zod.z.number().optional(),
+  imagesWithVideo: import_zod.z.number().optional(),
   videos: import_zod.z.number().optional(),
   audios: import_zod.z.number().optional(),
   total: import_zod.z.number().optional(),
@@ -2014,6 +2020,7 @@ var DefaultsSchema = import_zod.z.looseObject({
   promptExpansion: import_zod.z.enum(["fast", "balanced", "quality"]).optional(),
   watermark: import_zod.z.boolean().optional(),
   safetyChecker: import_zod.z.boolean().optional(),
+  taskType: import_zod.z.enum(["reference", "edit", "extend"]).optional(),
   mode: import_zod.z.enum(["standard", "pro"]).optional(),
   instrumental: import_zod.z.boolean().optional()
 });
@@ -2080,7 +2087,18 @@ var CatalogModelSchemaBase = import_zod.z.looseObject({
     });
   }
 });
-var CatalogModelSchema = CatalogModelSchemaBase;
+var CatalogModelSchema = CatalogModelSchemaBase.superRefine(
+  (row, ctx) => {
+    const transport = row.providerOptions?.["modelkit"]?.["transport"];
+    if (transport !== void 0 && transport !== "direct" && transport !== "bedrock") {
+      ctx.addIssue({
+        code: "custom",
+        path: ["providerOptions", "modelkit", "transport"],
+        message: 'transport must be "direct" or "bedrock"'
+      });
+    }
+  }
+);
 var CatalogSchema = import_zod.z.array(CatalogModelSchemaBase).superRefine((rows, ctx) => {
   const seenIds = /* @__PURE__ */ new Map();
   rows.forEach((row, index) => {
@@ -2550,6 +2568,9 @@ function resolveVideoOptions(model, request, onUnsupported = "error") {
     model,
     warnings
   );
+  if (limits.imagesWithVideo !== void 0 && (limits.videos ?? 0) > 0 && videos.length > 0 && images.length > limits.imagesWithVideo) {
+    images = trim("referenceImages", images, limits.imagesWithVideo, model, warnings);
+  }
   if (limits.total !== void 0) {
     let over = images.length + videos.length + audios.length - limits.total;
     if (over > 0) {
@@ -2583,7 +2604,8 @@ function resolveVideoOptions(model, request, onUnsupported = "error") {
     "promptExpansion",
     "lastFrame",
     "watermark",
-    "safetyChecker"
+    "safetyChecker",
+    "taskType"
   ]);
   return { request: merged, warnings };
 }