ai 7.0.104 → 7.0.106

package/docs/03-ai-sdk-core/32-evaluation.mdx

@@ -11,8 +11,8 @@ is one state, not a batch of unrelated inputs.
 
 This API and the evaluation model specification are experimental and may change
 in patch releases. Pass an evaluation model instance, resolve a model through a
-provider registry, or use a string ID with an evaluation-capable default provider.
-Evaluation does not fall back to Vercel AI Gateway.
+provider registry, or use a string ID. Strings resolve through Vercel AI Gateway
+unless you configure an evaluation-capable default provider.
 
 ```ts
 import { experimental_evaluate, type Experimental_EvaluationModel } from 'ai';
@@ -125,8 +125,26 @@ evaluation models.
 
 ### Default-provider strings
 
-To use strings directly, configure the default provider once at application
-startup:
+String IDs use Vercel AI Gateway by default. Configure Gateway authentication
+with `AI_GATEWAY_API_KEY` or Vercel OIDC, then pass a Gateway model ID:
+
+```ts
+import { experimental_evaluate } from 'ai';
+
+const result = await experimental_evaluate({
+  model: 'typesafe-ai/jev-latest',
+  state: 'I was charged twice. Please refund the extra charge.',
+  questions: {
+    refund: {
+      type: 'boolean',
+      instructions: 'Is the customer asking for a refund?',
+    },
+  },
+});
+```
+
+To resolve strings through your own provider or aliases, configure the default
+provider once at application startup:
 
 ```ts
 // Using the registry above, expose an alias through a custom provider:
@@ -152,9 +170,10 @@ also resolve through this global default. Prefer model instances in aliases to
 avoid resolution cycles. Global configuration affects other AI SDK functions
 too; avoid changing it per request in a shared process.
 
-Strings require an explicitly configured provider with an `evaluationModel`
-method. There is no assumed Gateway evaluation support. Missing providers throw
-`NoSuchProviderError`; missing models or evaluation capabilities throw
+An explicitly configured default provider must expose an `evaluationModel`
+method; Gateway is used only when no default provider is configured. Missing
+registry providers throw `NoSuchProviderError`; missing models or evaluation
+capabilities throw
 `NoSuchModelError` with `modelType: 'evaluationModel'`. Unsupported model versions
 throw `UnsupportedModelVersionError`, including models returned by registries.
 The stable `ProviderV4` and `ProviderRegistryProvider` interfaces are unchanged;

package/docs/03-ai-sdk-core/45-provider-management.mdx

@@ -473,8 +473,8 @@ This simplifies provider usage and makes it easier to switch between providers w
 
 Custom providers accept `evaluationModels` aliases, and registries expose
 `evaluationModel('provider:model')`. These methods return model instances for
-`experimental_evaluate`. An explicitly configured default provider with an
-`evaluationModel` method also enables direct string IDs; evaluation does not
-implicitly use Gateway. Registry middleware for language and image models does
-not apply to evaluation. See [Evaluation](/docs/ai-sdk-core/evaluation#model-aliases-and-registries)
+`experimental_evaluate`. Direct string IDs use Gateway by default, or an
+explicitly configured default provider with an `evaluationModel` method. Registry
+middleware for language and image models does not apply to evaluation.
+See [Evaluation](/docs/ai-sdk-core/evaluation#model-aliases-and-registries)
 for aliases, default-provider configuration, and capability differences.

package/docs/07-reference/01-ai-sdk-core/14-evaluate.mdx

@@ -14,15 +14,15 @@ state. See [Evaluation](/docs/ai-sdk-core/evaluation) for examples and semantics
 
 ## Parameters
 
-| Parameter         | Type                                              | Description                                                                                                                      |
-| ----------------- | ------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------- |
-| `model`           | `Experimental_EvaluationModel`                    | Required experimental v4 model instance or a string ID resolved by an explicitly configured evaluation-capable default provider. |
-| `state`           | `string \| object \| array`                       | Required JSON-compatible shared state.                                                                                           |
-| `questions`       | `Record<string, Experimental_EvaluationQuestion>` | Required nonempty question map.                                                                                                  |
-| `maxRetries`      | `number`                                          | Nonnegative integer; defaults to 2.                                                                                              |
-| `abortSignal`     | `AbortSignal`                                     | Cancels evaluation.                                                                                                              |
-| `headers`         | `Record<string, string>`                          | Additional HTTP headers.                                                                                                         |
-| `providerOptions` | `ProviderOptions`                                 | Provider-specific options.                                                                                                       |
+| Parameter         | Type                                              | Description                                                                                                                                 |
+| ----------------- | ------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------- |
+| `model`           | `Experimental_EvaluationModel`                    | Required experimental v4 model instance or a string ID resolved by Gateway or an explicitly configured evaluation-capable default provider. |
+| `state`           | `string \| object \| array`                       | Required JSON-compatible shared state.                                                                                                      |
+| `questions`       | `Record<string, Experimental_EvaluationQuestion>` | Required nonempty question map.                                                                                                             |
+| `maxRetries`      | `number`                                          | Nonnegative integer; defaults to 2.                                                                                                         |
+| `abortSignal`     | `AbortSignal`                                     | Cancels evaluation.                                                                                                                         |
+| `headers`         | `Record<string, string>`                          | Additional HTTP headers.                                                                                                                    |
+| `providerOptions` | `ProviderOptions`                                 | Provider-specific options.                                                                                                                  |
 
 ## Result
 

package/docs/07-reference/01-ai-sdk-core/40-provider-registry.mdx

@@ -312,6 +312,7 @@ retain experimental evaluation access.
 Unavailable evaluation capabilities or models throw `NoSuchModelError` with
 `modelType: 'evaluationModel'`; unknown registry providers throw
 `NoSuchProviderError`. These capabilities are structural extensions and are not
-added to the stable `ProviderV4` contract. Evaluation does not assume Gateway
-support. See [Evaluation](/docs/ai-sdk-core/evaluation#model-aliases-and-registries)
+added to the stable `ProviderV4` contract. Direct evaluation string IDs default
+to Gateway when no default provider is configured.
+See [Evaluation](/docs/ai-sdk-core/evaluation#model-aliases-and-registries)
 for runnable usage patterns.

package/docs/07-reference/01-ai-sdk-core/42-custom-provider.mdx

@@ -207,13 +207,15 @@ The `customProvider` function returns a `Provider` instance. It has the followin
 Pass `evaluationModels: Record<string, Experimental_EvaluationModel>` to define
 aliases for evaluation model instances or string IDs. The returned provider adds
 `evaluationModel(alias): Experimental_EvaluationModelV4`. String aliases resolve
-through an explicitly configured default provider with an `evaluationModel`
-method. Unknown aliases use the fallback provider's evaluation factory when
-available; model failures and unsupported questions do not trigger substitution.
+through Gateway by default, or an explicitly configured default provider with
+an `evaluationModel` method. Unknown aliases use the fallback provider's evaluation
+factory when available; model failures and unsupported questions do not trigger
+substitution.
 
 Unavailable evaluation capabilities or models throw `NoSuchModelError` with
 `modelType: 'evaluationModel'`; unknown registry providers throw
 `NoSuchProviderError`. These capabilities are structural extensions and are not
-added to the stable `ProviderV4` contract. Evaluation does not assume Gateway
-support. See [Evaluation](/docs/ai-sdk-core/evaluation#model-aliases-and-registries)
+added to the stable `ProviderV4` contract. Direct evaluation string IDs default
+to Gateway when no default provider is configured.
+See [Evaluation](/docs/ai-sdk-core/evaluation#model-aliases-and-registries)
 for runnable usage patterns.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ai",
-  "version": "7.0.104",
+  "version": "7.0.106",
   "type": "module",
   "description": "AI SDK by Vercel - build apps like ChatGPT, Claude, Gemini, and more with a single interface for any model using the Vercel AI Gateway or go direct to OpenAI, Anthropic, Google, or any other model provider.",
   "license": "Apache-2.0",
@@ -42,20 +42,20 @@
     }
   },
   "dependencies": {
-    "@ai-sdk/gateway": "4.0.84",
+    "@ai-sdk/gateway": "4.0.86",
     "@ai-sdk/provider": "4.0.17",
-    "@ai-sdk/provider-utils": "5.0.43"
+    "@ai-sdk/provider-utils": "5.0.44"
   },
   "devDependencies": {
-    "@ai-sdk/amazon-bedrock": "5.0.86",
-    "@ai-sdk/deepseek": "3.0.47",
-    "@ai-sdk/google": "4.0.74",
-    "@ai-sdk/groq": "4.0.44",
-    "@ai-sdk/huggingface": "2.0.51",
-    "@ai-sdk/moonshotai": "3.0.52",
-    "@ai-sdk/openai": "4.0.69",
+    "@ai-sdk/amazon-bedrock": "5.0.87",
+    "@ai-sdk/deepseek": "3.0.48",
+    "@ai-sdk/google": "4.0.75",
+    "@ai-sdk/groq": "4.0.45",
+    "@ai-sdk/huggingface": "2.0.52",
+    "@ai-sdk/moonshotai": "3.0.53",
+    "@ai-sdk/openai": "4.0.70",
     "@ai-sdk/test-server": "2.0.1",
-    "@ai-sdk/xai": "5.0.2",
+    "@ai-sdk/xai": "5.0.3",
     "@edge-runtime/vm": "^5.0.0",
     "@smithy/eventstream-codec": "^4.3.3",
     "@smithy/util-utf8": "^4.3.3",

package/src/batch/batch.ts

@@ -202,6 +202,7 @@ export async function startBatch<
               prompt: standardizedPrompt,
               supportedUrls,
               download: undefined,
+              abortSignal: operationAbortSignal,
               provider: batchApi.provider.split('.')[0],
             }),
             tools: preparedTools,

package/src/generate-object/generate-object.ts

@@ -379,6 +379,7 @@ export async function generateObject<
       prompt: standardizedPrompt,
       supportedUrls: await model.supportedUrls,
       download,
+      abortSignal,
       provider: model.provider.split('.')[0],
     });
 

package/src/generate-object/stream-object.ts

@@ -612,6 +612,7 @@ class DefaultStreamObjectResult<
           prompt: standardizedPrompt,
           supportedUrls: await model.supportedUrls,
           download,
+          abortSignal,
           provider: model.provider.split('.')[0],
         }),
         providerOptions,

package/src/generate-text/generate-text.ts

@@ -964,6 +964,7 @@ export async function generateText<
                 },
                 supportedUrls: await stepModel.supportedUrls,
                 download,
+                abortSignal: mergedAbortSignal,
                 provider: stepModel.provider.split('.')[0],
               });
 

package/src/generate-text/invoke-tool-callbacks-from-stream.ts

@@ -80,7 +80,7 @@ export function invokeToolCallbacksFromStream<
 
             delete ongoingToolCallToolNames[chunk.toolCallId];
 
-            if (tool?.onInputAvailable != null) {
+            if (!chunk.invalid && tool?.onInputAvailable != null) {
               await tool.onInputAvailable({
                 input: chunk.input,
                 toolCallId: chunk.toolCallId,

package/src/generate-text/prune-messages.ts

@@ -121,6 +121,7 @@ export function pruneMessages({
       }
     }
 
+    const approvalIdToToolCallId = new Map<string, string>();
     const approvalIdToToolName = new Map<string, string>();
     for (const message of messages) {
       if (
@@ -129,6 +130,8 @@ export function pruneMessages({
       ) {
         for (const part of message.content) {
           if (part.type === 'tool-approval-request') {
+            approvalIdToToolCallId.set(part.approvalId, part.toolCallId);
+
             const toolName = toolCallIdToToolName.get(part.toolCallId);
             if (toolName != null) {
               approvalIdToToolName.set(part.approvalId, toolName);
@@ -138,6 +141,16 @@ export function pruneMessages({
       }
     }
 
+    // Approval requests depend on their originating tool call. When an
+    // approval response is in the retained messages, trace through its request
+    // so pruning does not leave a pending approval without the call to execute.
+    for (const approvalId of keptApprovalIds) {
+      const toolCallId = approvalIdToToolCallId.get(approvalId);
+      if (toolCallId != null) {
+        keptToolCallIds.add(toolCallId);
+      }
+    }
+
     messages = messages.map((message, messageIndex) => {
       if (
         (message.role !== 'assistant' && message.role !== 'tool') ||

package/src/generate-text/stream-language-model-call.ts

@@ -306,6 +306,7 @@ export async function streamLanguageModelCall<
     },
     supportedUrls: await resolvedModel.supportedUrls,
     download,
+    abortSignal,
     provider: resolvedModel.provider.split('.')[0],
   });
 

package/src/generate-text/stream-text.ts

@@ -968,14 +968,16 @@ function createOutputTransformStream<
   let text = '';
   let textChunk = '';
   let textProviderMetadata: ProviderMetadata | undefined = undefined;
-  let lastPublishedValue = '';
+  let lastPublishedValue: string | undefined = undefined;
+  let hasPublishedValue = false;
 
-  function resetAttemptState() {
+  function resetOutputState() {
     firstTextChunkId = undefined;
     text = '';
     textChunk = '';
     textProviderMetadata = undefined;
-    lastPublishedValue = '';
+    lastPublishedValue = undefined;
+    hasPublishedValue = false;
   }
 
   function enqueueChunk({
@@ -1020,11 +1022,15 @@ function createOutputTransformStream<
   >({
     async transform(chunk, controller) {
       if (isStreamRetryBoundaryPart(chunk)) {
-        resetAttemptState();
+        resetOutputState();
         controller.enqueue(chunk);
         return;
       }
 
+      if (chunk.type === 'start-step') {
+        resetOutputState();
+      }
+
       // ensure that we publish the last text chunk before the step finish:
       if (chunk.type === 'finish-step' && textChunk.length > 0) {
         publishTextChunk({ controller });
@@ -1096,9 +1102,10 @@ function createOutputTransformStream<
           typeof result.partial === 'string'
             ? result.partial
             : JSON.stringify(result.partial);
-        if (currentValue !== lastPublishedValue) {
+        if (!hasPublishedValue || currentValue !== lastPublishedValue) {
           publishTextChunk({ controller, partialOutput: result.partial });
           lastPublishedValue = currentValue;
+          hasPublishedValue = true;
         }
       }
     },
@@ -1414,6 +1421,8 @@ class DefaultStreamTextResult<
     const initialResponseMessages: Array<ResponseMessage> = [];
     let stepMessagesForNextStep: Array<ModelMessage> | undefined;
     let currentStepMessages: Array<ModelMessage> = [];
+    let isAborted = false;
+    let currentStepModel = model;
 
     // provider-assigned text/reasoning part IDs are only unique within a
     // single model call (e.g. Anthropic uses the content block index, which
@@ -1661,8 +1670,8 @@ class DefaultStreamTextResult<
             new DefaultStepResult({
               callId,
               stepNumber: recordedSteps.length,
-              provider: model.provider,
-              modelId: model.modelId,
+              provider: currentStepModel.provider,
+              modelId: currentStepModel.modelId,
               runtimeContext,
               toolsContext,
               content: recordedContent,
@@ -1691,8 +1700,8 @@ class DefaultStreamTextResult<
 
           logWarnings({
             warnings: recordedWarnings,
-            provider: model.provider,
-            model: model.modelId,
+            provider: currentStepModel.provider,
+            model: currentStepModel.modelId,
           });
 
           recordedSteps.push(currentStepResult);
@@ -1743,6 +1752,10 @@ class DefaultStreamTextResult<
           // aggregate results:
           self._steps.resolve(recordedSteps);
 
+          if (isAborted) {
+            return;
+          }
+
           // call onEnd callback:
           const finalStep = recordedSteps[recordedSteps.length - 1];
           const content = recordedSteps.flatMap(step => step.content);
@@ -1845,6 +1858,8 @@ class DefaultStreamTextResult<
       async pull(controller) {
         // abort handling:
         async function abort() {
+          isAborted = true;
+
           await notify({
             event: {
               callId,
@@ -1885,6 +1900,7 @@ class DefaultStreamTextResult<
           if (isAbortError(error) && abortSignal?.aborted) {
             await abort();
           } else {
+            await telemetryDispatcher.onError?.({ callId, error });
             controller.error(error);
           }
         }
@@ -2328,6 +2344,7 @@ class DefaultStreamTextResult<
           const stepModel = resolveLanguageModel(
             prepareStepResult?.model ?? model,
           );
+          currentStepModel = stepModel;
 
           const stepActiveTools = filterActiveTools({
             tools,
@@ -2752,7 +2769,7 @@ class DefaultStreamTextResult<
           const createStepResponse = () => ({
             id: generateId(),
             timestamp: new Date(),
-            modelId: model.modelId,
+            modelId: stepModel.modelId,
           });
           let stepResponse: {
             id: string;
@@ -3375,7 +3392,7 @@ class DefaultStreamTextResult<
           InferPartialOutput<OUTPUT>
         >({
           transform({ partialOutput }, controller) {
-            if (partialOutput != null) {
+            if (partialOutput !== undefined) {
               controller.enqueue(partialOutput);
             }
           },

package/src/middleware/extract-json-middleware.ts

@@ -19,6 +19,28 @@ function stripMarkdownCodeFenceSuffix(text: string): string {
   return text.replace(/\n?```\s*$/, '').trimEnd();
 }
 
+function getPotentialSuffixStart(text: string): number {
+  let index = text.length;
+
+  while (index > 0 && /\s/.test(text[index - 1])) {
+    index--;
+  }
+
+  let backtickCount = 0;
+  while (index > 0 && backtickCount < 3 && text[index - 1] === '`') {
+    index--;
+    backtickCount++;
+  }
+
+  if (backtickCount > 0) {
+    while (index > 0 && /\s/.test(text[index - 1])) {
+      index--;
+    }
+  }
+
+  return index;
+}
+
 /**
  * Middleware that extracts JSON from text content by stripping
  * markdown code fences and other formatting.
@@ -75,8 +97,6 @@ export function extractJsonMiddleware(options?: {
         }
       > = createIdMap();
 
-      const SUFFIX_BUFFER_SIZE = 12;
-
       return {
         stream: stream.pipeThrough(
           new TransformStream<
@@ -145,13 +165,19 @@ export function extractJsonMiddleware(options?: {
                   }
                 }
 
-                // Stream content
-                if (
-                  block.phase === 'streaming' &&
-                  block.buffer.length > SUFFIX_BUFFER_SIZE
-                ) {
-                  const toStream = block.buffer.slice(0, -SUFFIX_BUFFER_SIZE);
-                  block.buffer = block.buffer.slice(-SUFFIX_BUFFER_SIZE);
+                // Stream content while retaining anything that could still
+                // become trailing whitespace or a markdown fence suffix.
+                if (block.phase === 'streaming') {
+                  const potentialSuffixStart = getPotentialSuffixStart(
+                    block.buffer,
+                  );
+                  const toStream = block.buffer.slice(0, potentialSuffixStart);
+                  block.buffer = block.buffer.slice(potentialSuffixStart);
+
+                  if (toStream.length === 0) {
+                    return;
+                  }
+
                   controller.enqueue({
                     type: 'text-delta',
                     id: chunk.id,

package/src/middleware/simulate-streaming-middleware.ts

@@ -25,7 +25,13 @@ export function simulateStreamingMiddleware(): LanguageModelMiddleware {
             switch (part.type) {
               case 'text': {
                 if (part.text.length > 0) {
-                  controller.enqueue({ type: 'text-start', id: String(id) });
+                  controller.enqueue({
+                    type: 'text-start',
+                    id: String(id),
+                    ...(part.providerMetadata != null
+                      ? { providerMetadata: part.providerMetadata }
+                      : {}),
+                  });
                   controller.enqueue({
                     type: 'text-delta',
                     id: String(id),

package/src/model/resolve-model.ts

@@ -222,10 +222,8 @@ export function resolveEvaluationModel(
 ): EvaluationModelV4 {
   if (typeof model === 'string') {
     // Use the original provider so experimental methods and their receiver survive.
-    // Evaluation does not default to Gateway.
-    const provider = globalThis.AI_SDK_DEFAULT_PROVIDER as
-      | EvaluationProvider
-      | undefined;
+    const provider = (globalThis.AI_SDK_DEFAULT_PROVIDER ??
+      gateway) as EvaluationProvider;
 
     if (typeof provider?.evaluationModel !== 'function') {
       throw new NoSuchModelError({

package/src/prompt/convert-to-language-model-prompt.ts

@@ -36,7 +36,8 @@ import { MissingToolResultsError } from '../error/missing-tool-result-error';
 export async function convertToLanguageModelPrompt({
   prompt,
   supportedUrls,
-  download = createDefaultDownloadFunction(),
+  download,
+  abortSignal,
   // `provider` is only needed here to convert legacy tool output types via `mapToolResultOutput`.
   // TODO: remove in v8 when "file-id" and "image-file-id" types are removed
   provider,
@@ -44,11 +45,12 @@ export async function convertToLanguageModelPrompt({
   prompt: StandardizedPrompt;
   supportedUrls: Record<string, RegExp[]>;
   download: DownloadFunction | undefined;
+  abortSignal?: AbortSignal;
   provider?: string;
 }): Promise<LanguageModelV4Prompt> {
   const downloadedAssets = await downloadAssets(
     prompt.messages,
-    download,
+    download ?? createDefaultDownloadFunction(undefined, abortSignal),
     supportedUrls,
   );
 

package/src/prompt/create-tool-model-output.ts

@@ -1,9 +1,5 @@
 import { getErrorMessage, type JSONValue } from '@ai-sdk/provider';
-import {
-  parseJSON,
-  type Tool,
-  type ToolResultOutput,
-} from '@ai-sdk/provider-utils';
+import type { Tool, ToolResultOutput } from '@ai-sdk/provider-utils';
 
 export async function createToolModelOutput({
   toolCallId,
@@ -21,7 +17,7 @@ export async function createToolModelOutput({
   if (errorMode === 'text') {
     return { type: 'error-text', value: getErrorMessage(output) };
   } else if (errorMode === 'json') {
-    return { type: 'error-json', value: await toJSONValue(output) };
+    return { type: 'error-json', value: toJSONValue(output) };
   }
 
   if (tool?.toModelOutput) {
@@ -30,16 +26,29 @@ export async function createToolModelOutput({
 
   return typeof output === 'string'
     ? { type: 'text', value: output }
-    : { type: 'json', value: await toJSONValue(output) };
+    : { type: 'json', value: toJSONValue(output) };
 }
 
-async function toJSONValue(value: unknown): Promise<JSONValue> {
+/**
+ * Normalizes an in-process tool output to a plain JSON value
+ * (applies `toJSON`, converts `Date`, drops `undefined`, etc.)
+ * by round-tripping it through `JSON.stringify`.
+ *
+ * The parsed text is produced by `JSON.stringify` from a value that is
+ * already materialized in this process, so it is not untrusted input.
+ * `JSON.parse` is used deliberately instead of the secure parser from
+ * `@ai-sdk/provider-utils`: the secure parser rejects own `__proto__` and
+ * `constructor.prototype` keys, which are valid data in tool outputs
+ * (e.g. rows from an external API). `JSON.parse` defines `__proto__` as an
+ * own data property and never modifies the prototype chain, so preserving
+ * these keys here is safe. Do not reuse this pattern for text that comes
+ * from outside the process; use `parseJSON` / `safeParseJSON` instead.
+ */
+function toJSONValue(value: unknown): JSONValue {
   if (value === undefined) {
     return null;
   }
 
   const serialized = JSON.stringify(value);
-  return serialized === undefined
-    ? null
-    : await parseJSON({ text: serialized });
+  return serialized === undefined ? null : JSON.parse(serialized);
 }

package/src/ui-message-stream/read-ui-message-stream.ts

@@ -65,11 +65,17 @@ export function readUIMessageStream<UI_MESSAGE extends UIMessage>({
 }): AsyncIterableStream<UI_MESSAGE> {
   let controller: ReadableStreamDefaultController<UI_MESSAGE> | undefined;
   let hasErrored = false;
+  let isCancelled = false;
+  const abortController = new AbortController();
 
   const outputStream = new ReadableStream<UI_MESSAGE>({
     start(controllerParam) {
       controller = controllerParam;
     },
+    cancel() {
+      isCancelled = true;
+      abortController.abort();
+    },
   });
 
   const state = createStreamingUIMessageState<UI_MESSAGE>({
@@ -98,17 +104,20 @@ export function readUIMessageStream<UI_MESSAGE extends UIMessage>({
         return job({
           state,
           write: () => {
-            controller?.enqueue(createUIMessageSnapshot(state.message));
+            if (!isCancelled) {
+              controller?.enqueue(createUIMessageSnapshot(state.message));
+            }
           },
         });
       },
       onError: handleError,
     }),
     onError: handleError,
+    abortSignal: abortController.signal,
   }).finally(() => {
-    // Only close if no error occurred. Calling close() on an errored controller
-    // throws "Invalid state: Controller is already closed" TypeError.
-    if (!hasErrored) {
+    // Only close if no error or cancellation occurred. Both erroring and
+    // cancelling a stream close its controller.
+    if (!hasErrored && !isCancelled) {
       controller?.close();
     }
   });

package/src/util/download/download-function.ts

@@ -34,12 +34,15 @@ export type DownloadFunction = (
  * Downloads the file if it is not supported by the model.
  */
 export const createDefaultDownloadFunction =
-  (download: typeof originalDownload = originalDownload): DownloadFunction =>
+  (
+    download: typeof originalDownload = originalDownload,
+    abortSignal?: AbortSignal,
+  ): DownloadFunction =>
   requestedDownloads =>
     Promise.all(
       requestedDownloads.map(async requestedDownload =>
         requestedDownload.isUrlSupportedByModel
           ? null
-          : await download(requestedDownload),
+          : await download({ ...requestedDownload, abortSignal }),
       ),
     );