@shop-os/modelkit 0.7.4 → 0.7.7

package/CHANGELOG.md

@@ -9,6 +9,18 @@ header in the release PR.
 
 ## [Unreleased]
 
+## [0.7.7] - Unreleased
+
+### Fixed
+- Google `prepareMediaPart` inlines bytes only up to 10 MiB (was 20 MB) and uploads larger media through the Files API. Inline bytes are base64-encoded, so a 15 MB video made a ~20 MB request and Gemini refused it or timed out; on dev no inline video between 10.9 MB and 20 MB succeeded. Media of 10-20 MB now uploads (a few seconds slower, same tokens) and the returned `cleanup` deletes the file as before.
+
+## [0.7.6] - Unreleased
+
+ENG-3323: one xAI reference-mode rule.
+
+### Changed
+- `xaiReferenceResolution` (kit-internal) now owns "is this request in xAI reference mode" and the 720p cap for both the resolver and the `xai-video` adapter. The adapter no longer keeps its own reference-mode check or its own copy of the refuse/lower rule. Requests, refusals and warnings are unchanged.
+
 ## [0.7.4] - Unreleased
 
 ENG-3321: one source of truth for each 3D family's options.

package/README.md

@@ -309,7 +309,7 @@ const defaultKit = createDefaultModelKit({ keys: { pro: keysFromEnv(process.env)
 | Retry       | One classifier, one policy. Never on content blocks, aborts, or 4xx except 429; SDKs with their own retry (Replicate) run under a single kit attempt; a 429 on a status poll keeps polling unless `CallContext.maxRetries` is `0`. See durable media execution above.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |
 | Errors      | `ModelKitError` with `code`, `provider`, `status`, `body`. Gemini `blockReason` and fal `body.detail` are lifted into `body`. A throw with no HTTP status (fetch failed, reset, DNS) is a retryable `network` error on every adapter. An out-of-credit provider account is `provider_billing` and is never retried: HTTP 402, a 4xx whose structured error code is a billing code (`insufficient_quota`, ...), or a 403/429 whose text names the exhausted account ("Exhausted balance", "Insufficient credit"). Text calls go through `ai` and surface the SDK's own error instead.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |
 | Usage       | One event per call with the canonical model id and the row's `usage.unit` (adapters supply the quantity in that unit). Gemini image events carry the token counts too (`inputTokens`/`outputTokens`); Google's raw `usageMetadata` rides on `raw.providerMetadata.google`. Video usage is emitted on completion and carries the resolution the provider renders (the default it falls back to when the request has none), because pricing is keyed by it; Seedance keeps `usage`, `duration`, `framespersecond` and `resolution` from the poll on `raw`. Audio: Lyria 3 returns no duration, so it emits `unit: "images", quantity: 1` (token counts ride along as `inputTokens`/`outputTokens`); Replicate audio (Lyria 2, MMAudio) emits `unit: "compute_seconds"` with the prediction time.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |
-| Media parts | `prepareMediaPart` turns a `MediaRef` into an AI SDK file part. Google inlines bytes up to 20 MB and uploads larger payloads through the Files API; call the returned `cleanup` after the request. Other providers get the reference inline.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            |
+| Media parts | `prepareMediaPart` turns a `MediaRef` into an AI SDK file part. Google inlines bytes up to 10 MiB (base64 makes them ~4/3 larger inside Gemini's 20 MB request cap) and uploads larger payloads through the Files API; call the returned `cleanup` after the request. Other providers get the reference inline.                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                            |
 
 ## Add a model
 

package/dist/catalog.cjs

@@ -6401,14 +6401,14 @@ function xaiReferenceMode(model, req) {
   const takesLastFrame = (model.constraints.supports ?? []).includes("lastFrame");
   return (req.referenceImages ?? []).length > 0 || takesLastFrame && req.lastFrame !== void 0;
 }
-function applyXaiVideoRules(model, req, onUnsupported, warnings) {
-  if (model.adapter !== "xai-video" || req.resolution !== "1080p") return;
-  if (!xaiReferenceMode(model, req)) return;
+function xaiReferenceResolution(model, req, onUnsupported, warnings, provider) {
+  if (req.resolution !== "1080p" || !xaiReferenceMode(model, req)) return void 0;
   const message = xaiReferenceResolutionMessage(model.id);
   if (onUnsupported === "error") {
     throw new ModelKitError({
       code: "unsupported_option",
       message,
+      ...provider === void 0 ? {} : { provider },
       model: model.id,
       retryable: false
     });
@@ -6420,7 +6420,12 @@ function applyXaiVideoRules(model, req, onUnsupported, warnings) {
     to: XAI_REFERENCE_MAX_RESOLUTION,
     message
   });
-  req.resolution = XAI_REFERENCE_MAX_RESOLUTION;
+  return XAI_REFERENCE_MAX_RESOLUTION;
+}
+function applyXaiVideoRules(model, req, onUnsupported, warnings) {
+  if (model.adapter !== "xai-video") return;
+  const lowered = xaiReferenceResolution(model, req, onUnsupported, warnings);
+  if (lowered !== void 0) req.resolution = lowered;
 }
 function takesTargetAspectRatio(model) {
   const ratios = model.constraints.aspectRatios ?? [];