ai 7.0.161 → 7.0.163

package/dist/index.js

@@ -1,8 +1,30 @@
 import { t as __exportAll } from "./rolldown-runtime-D7D4PA-g.js";
 import { GatewayAuthenticationError, GatewayError, createGateway, gateway, gateway as gateway$1 } from "@ai-sdk/gateway";
-import { DEFAULT_MAX_DOWNLOAD_SIZE, DelayedPromise, DownloadError, DownloadError as DownloadError$1, EXPERIMENTAL_EMBEDDING_MODEL_MAX_INPUT_BYTES_PER_CALL, EXPERIMENTAL_EMBEDDING_MODEL_PROVIDER_OPTIONS_TRANSFORMER, asArray, asSchema, asSchema as asSchema$1, cancelResponseBody, convertAsyncIteratorToReadableStream, convertBase64ToUint8Array, convertUint8ArrayToBase64, createIdGenerator, createIdGenerator as createIdGenerator$1, delay, detectMediaType, dynamicTool, executeTool, experimental_getToolCaller, experimental_toolCaller, fetchUntrustedUrl, filterNullable, generateId, generateId as generateId$1, getErrorMessage, getRuntimeEnvironmentUserAgent, getWebSocketConstructor, isAbortError, isArrayBuffer, isBuffer, isExecutableTool, isFullMediaType, isNonNullable, isProviderReference, isProviderStreamError, isUint8Array, isUrlSupported, jsonSchema, jsonSchema as jsonSchema$1, lazySchema, normalizeHeaders, parseJsonEventStream, parseJsonEventStream as parseJsonEventStream$1, readResponseWithSizeLimit, readWebSocketMessageText, resolve, retryWithExponentialBackoff, safeParseJSON, safeValidateTypes, toWebSocketUrl, tool, tool as tool$1, validateTypes, waitForWebSocketBufferDrain, withUserAgentSuffix, zodSchema, zodSchema as zodSchema$1 } from "@ai-sdk/provider-utils";
+import { DEFAULT_MAX_DOWNLOAD_SIZE, DelayedPromise, DownloadError, DownloadError as DownloadError$1, EXPERIMENTAL_EMBEDDING_MODEL_MAX_INPUT_BYTES_PER_CALL, EXPERIMENTAL_EMBEDDING_MODEL_PROVIDER_OPTIONS_TRANSFORMER, asArray, asSchema, asSchema as asSchema$1, cancelResponseBody, convertAsyncIteratorToReadableStream, convertBase64ToUint8Array, convertUint8ArrayToBase64, createIdGenerator, createIdGenerator as createIdGenerator$1, delay, detectMediaType, dynamicTool, executeTool, experimental_getToolCaller, experimental_toolCaller, fetchUntrustedUrl, filterNullable, generateId, generateId as generateId$1, getErrorMessage, getRuntimeEnvironmentUserAgent, getTopLevelMediaType, getWebSocketConstructor, isAbortError, isArrayBuffer, isBuffer, isExecutableTool, isFullMediaType, isNonNullable, isProviderReference, isProviderStreamError, isUint8Array, isUrlSupported, jsonSchema, jsonSchema as jsonSchema$1, lazySchema, normalizeHeaders, parseJsonEventStream, parseJsonEventStream as parseJsonEventStream$1, readResponseWithSizeLimit, readWebSocketMessageText, resolve, retryWithExponentialBackoff, safeParseJSON, safeValidateTypes, toWebSocketUrl, tool, tool as tool$1, validateTypes, waitForWebSocketBufferDrain, withUserAgentSuffix, zodSchema, zodSchema as zodSchema$1 } from "@ai-sdk/provider-utils";
 import { AISDKError, AISDKError as AISDKError$1, APICallError, APICallError as APICallError$1, EmptyResponseBodyError, EmptyResponseBodyError as EmptyResponseBodyError$1, Experimental_DecisionRefusalError, Experimental_DecisionRefusalError as Experimental_DecisionRefusalError$1, Experimental_DecisionUnsupportedQuestionTypeError, Experimental_DecisionUnsupportedQuestionTypeError as Experimental_DecisionUnsupportedQuestionTypeError$1, InvalidPromptError, InvalidPromptError as InvalidPromptError$1, InvalidResponseDataError, InvalidResponseDataError as InvalidResponseDataError$1, JSONParseError, JSONParseError as JSONParseError$1, LoadAPIKeyError, LoadSettingError, NoContentGeneratedError, NoSuchModelError, NoSuchModelError as NoSuchModelError$1, NoSuchProviderReferenceError, TooManyEmbeddingValuesForCallError, TypeValidationError, TypeValidationError as TypeValidationError$1, UnsupportedFunctionalityError, UnsupportedFunctionalityError as UnsupportedFunctionalityError$1, getErrorMessage as getErrorMessage$1, isJSONArray, isJSONObject } from "@ai-sdk/provider";
 import { array, boolean, custom, discriminatedUnion, enum as enum$1, instanceof as instanceof$1, lazy, literal, looseObject, never, null as null$1, number, object, record, string, union, unknown } from "zod/v4";
+//#region src/telemetry/get-request-origin.ts
+/** Remove credentials, paths, queries, and fragments before emitting telemetry. */
+function getRequestOrigin(url) {
+	if (url == null) return void 0;
+	try {
+		const endpoint = new URL(url);
+		return endpoint.protocol === "http:" || endpoint.protocol === "https:" ? endpoint.origin : void 0;
+	} catch {
+		return;
+	}
+}
+/** Preserve request metadata while keeping unknown or sensitive URL fields out of results. */
+function sanitizeRequestMetadata(request) {
+	if (request == null || !("url" in request)) return request;
+	const { url, ...rest } = request;
+	const origin = getRequestOrigin(url);
+	return {
+		...rest,
+		...origin != null ? { url: origin } : {}
+	};
+}
+//#endregion
 //#region src/error/invalid-argument-error.ts
 const name$23 = "AI_InvalidArgumentError";
 const marker$23 = `vercel.ai.error.${name$23}`;
@@ -937,6 +959,18 @@ function asTranscriptionModelV4(model) {
 	const v3Model = model.specificationVersion === "v2" ? asTranscriptionModelV3(model) : model;
 	return new Proxy(v3Model, { get(target, prop) {
 		if (prop === "specificationVersion") return "v4";
+		if (prop === "doGenerate") return async ({ timestampGranularities, ...options }) => {
+			const result = await target.doGenerate.call(model, options);
+			if (!timestampGranularities?.length) return result;
+			return {
+				...result,
+				warnings: [...result.warnings, {
+					type: "unsupported",
+					feature: "timestampGranularities",
+					details: `Shared timestamp granularities are not supported by ${model.specificationVersion} transcription models.`
+				}]
+			};
+		};
 		return target[prop];
 	} });
 }
@@ -1207,7 +1241,7 @@ function cloneValue(value, seen) {
 }
 //#endregion
 //#region src/version.ts
-const VERSION = "7.0.161";
+const VERSION = "7.0.163";
 //#endregion
 //#region src/util/download/download.ts
 /**
@@ -1670,12 +1704,18 @@ async function downloadAssets(messages, download, supportedUrls) {
 		if (message.role === "tool") for (const part of message.content) {
 			if (part.type !== "tool-result") continue;
 			if (part.output.type !== "content") continue;
-			for (const contentPart of part.output.value) if (contentPart.type === "file") downloadableFiles.push(contentPart);
+			for (const contentPart of part.output.value) {
+				const filePart = getDownloadableFilePart(contentPart);
+				if (filePart != null) downloadableFiles.push(filePart);
+			}
 		}
 		if (message.role === "assistant" && Array.isArray(message.content)) for (const part of message.content) {
 			if (part.type !== "tool-result") continue;
 			if (part.output.type !== "content") continue;
-			for (const contentPart of part.output.value) if (contentPart.type === "file") downloadableFiles.push(contentPart);
+			for (const contentPart of part.output.value) {
+				const filePart = getDownloadableFilePart(contentPart);
+				if (filePart != null) downloadableFiles.push(filePart);
+			}
 		}
 	}
 	const plannedDownloads = downloadableFiles.map((part) => {
@@ -1700,6 +1740,57 @@ async function downloadAssets(messages, download, supportedUrls) {
 	}]).filter((file) => file != null));
 }
 /**
+* Resolves the tool result content parts that point at a remote file to the
+* tagged file shape, including the deprecated `file-url` and `image-url`
+* types, so that they are downloaded and inlined exactly like their `file`
+* replacement. Returns `undefined` for content that is already inline or
+* provider-referenced.
+*/
+function getDownloadableFilePart(contentPart) {
+	switch (contentPart.type) {
+		case "file": return contentPart;
+		case "file-url": return {
+			type: "file",
+			data: {
+				type: "url",
+				url: new URL(contentPart.url)
+			},
+			mediaType: contentPart.mediaType ?? getMediaTypeFromUrl(contentPart.url),
+			providerOptions: contentPart.providerOptions
+		};
+		case "image-url": return {
+			type: "file",
+			data: {
+				type: "url",
+				url: new URL(contentPart.url)
+			},
+			mediaType: "image",
+			providerOptions: contentPart.providerOptions
+		};
+		default: return;
+	}
+}
+/**
+* Resolves file data and media type for a remote tool result content part: the downloaded
+* bytes when the asset was fetched, otherwise the URL reference. `originalUrl`
+* is preserved when the URL string does not round-trip through `URL`.
+*/
+function resolveDownloadedFile({ url, originalUrl, downloadedAssets, mediaType }) {
+	const convertedPart = convertPartToLanguageModelPart({
+		type: "file",
+		data: {
+			type: "url",
+			url,
+			...url.toString() !== originalUrl ? { originalUrl } : {}
+		},
+		mediaType
+	}, downloadedAssets);
+	return {
+		data: convertedPart.data,
+		mediaType: convertedPart.mediaType
+	};
+}
+/**
 * Convert part of a user message to a LanguageModelV4Part.
 *
 * @param part - The part to convert.
@@ -1784,12 +1875,12 @@ function mapToolResultOutput({ output, provider, warnings = [], downloadedAssets
 					});
 					return {
 						type: "file",
-						data: {
-							type: "url",
+						...resolveDownloadedFile({
 							url,
-							...url.toString() !== item.url ? { originalUrl: item.url } : {}
-						},
-						mediaType,
+							originalUrl: item.url,
+							downloadedAssets,
+							mediaType
+						}),
 						providerOptions: item.providerOptions
 					};
 				}
@@ -1850,12 +1941,12 @@ function mapToolResultOutput({ output, provider, warnings = [], downloadedAssets
 					});
 					return {
 						type: "file",
-						data: {
-							type: "url",
+						...resolveDownloadedFile({
 							url,
-							...url.toString() !== item.url ? { originalUrl: item.url } : {}
-						},
-						mediaType: "image",
+							originalUrl: item.url,
+							downloadedAssets,
+							mediaType: "image"
+						}),
 						providerOptions: item.providerOptions
 					};
 				}
@@ -3401,17 +3492,7 @@ const text = () => ({
 	},
 	createElementStreamTransform() {}
 });
-/**
-* Output specification for typed object generation using schemas.
-* When the model generates a text response, it will return an object that matches the schema.
-*
-* @param schema - The schema of the object to generate.
-* @param name - Optional name of the output that should be generated. Used by some providers for additional LLM guidance, e.g. via tool or schema name.
-* @param description - Optional description of the output that should be generated. Used by some providers for additional LLM guidance, e.g. via tool or schema description.
-*
-* @returns An output specification for generating objects with the specified schema.
-*/
-const object$1 = ({ schema: inputSchema, name, description }) => {
+function object$1({ schema: inputSchema, name, description }) {
 	const schema = asSchema$1(inputSchema);
 	return {
 		name: "object",
@@ -3456,20 +3537,8 @@ const object$1 = ({ schema: inputSchema, name, description }) => {
 		},
 		createElementStreamTransform() {}
 	};
-};
-/**
-* Output specification for array generation.
-* When the model generates a text response, it will return an array of elements.
-*
-* @param element - The schema of the array elements to generate.
-* @param minItems - Optional minimum number of elements to generate.
-* @param maxItems - Optional maximum number of elements to generate.
-* @param name - Optional name of the output that should be generated. Used by some providers for additional LLM guidance, e.g. via tool or schema name.
-* @param description - Optional description of the output that should be generated. Used by some providers for additional LLM guidance, e.g. via tool or schema description.
-*
-* @returns An output specification for generating an array of elements.
-*/
-const array$1 = ({ element: inputElementSchema, minItems, maxItems, name, description }) => {
+}
+function array$1({ element: inputElementSchema, minItems, maxItems, name, description }) {
 	validateArrayBound({
 		name: "minItems",
 		value: minItems
@@ -3599,7 +3668,7 @@ const array$1 = ({ element: inputElementSchema, minItems, maxItems, name, descri
 			} });
 		}
 	};
-};
+}
 function validateArrayBound({ name, value }) {
 	if (value == null) return;
 	if (!Number.isInteger(value)) throw new InvalidArgumentError({
@@ -4064,7 +4133,7 @@ function loadBuiltinModule(id) {
 * if tracing itself throws, and prevents falling back by calling `execute` a
 * second time.
 */
-async function runWithTracingChannelSpan(message, execute) {
+async function runWithTracingChannelSpan(message, execute, recordResult = true) {
 	const tracingChannel = (await loadDiagnosticsChannel())?.tracingChannel?.(AI_SDK_TELEMETRY_TRACING_CHANNEL);
 	if (tracingChannel == null || tracingChannel.hasSubscribers === false) return await execute();
 	let executePromise;
@@ -4087,10 +4156,13 @@ async function runWithTracingChannelSpan(message, execute) {
 			hasExecutionError = true;
 			throw error;
 		});
-		return executePromise;
+		return recordResult ? executePromise : executePromise.then(() => void 0, () => {
+			throw new Error("AI SDK operation failed");
+		});
 	};
 	try {
-		return await tracingChannel.tracePromise(tracedExecute, message);
+		await tracingChannel.tracePromise(tracedExecute, message);
+		return executionResult;
 	} catch {
 		if (hasExecutionError) throw executionError;
 		if (hasExecutionResult) return executionResult;
@@ -4102,7 +4174,7 @@ async function runWithTracingChannelSpan(message, execute) {
 * Opens a long-lived tracing-channel span context and returns a runner that can
 * re-enter that context later without changing stream setup timing.
 */
-function openTelemetryChannelSpanContext({ message, completion }) {
+function openTelemetryChannelSpanContext({ message, completion, recordResult = true }) {
 	if (!isNodeRuntime()) {
 		Promise.resolve(completion).catch(() => {});
 		return;
@@ -4126,10 +4198,10 @@ function openTelemetryChannelSpanContext({ message, completion }) {
 		if (asyncEndPublished) return;
 		asyncEndPublished = true;
 		if (error !== void 0) {
-			context.error = error;
+			context.error = recordResult ? error : /* @__PURE__ */ new Error("AI SDK operation failed");
 			safePublish(() => tracingChannel.error.publish(context));
 		}
-		if (result !== void 0) context.result = result;
+		if (recordResult && result !== void 0) context.result = result;
 		safePublish(() => tracingChannel.asyncEnd.publish(context));
 	};
 	safePublish(() => {
@@ -4154,6 +4226,116 @@ function getGlobalTelemetryIntegrations() {
 	return globalThis.AI_SDK_TELEMETRY_INTEGRATIONS ?? [];
 }
 //#endregion
+//#region src/telemetry/filter-telemetry-event.ts
+/**
+* Filters SDK-owned content fields on a copy. Only descend into SDK event
+* containers; user context has its own inclusion policy.
+*/
+function filterTelemetryEvent(event, telemetry) {
+	const hideInputs = telemetry.recordInputs === false;
+	const hideOutputs = telemetry.recordOutputs === false;
+	if (!hideInputs && !hideOutputs || event == null || typeof event !== "object") return event;
+	const filtered = Object.assign(Object.create(Object.getPrototypeOf(event)), event);
+	const replace = (fields) => {
+		for (const [key, value] of Object.entries(fields)) if (Object.hasOwn(event, key)) filtered[key] = value;
+	};
+	const isSpeech = filtered.operationId === "ai.generateSpeech";
+	if (hideInputs) {
+		replace({
+			instructions: void 0,
+			system: void 0,
+			prompt: void 0,
+			messages: [],
+			promptMessages: [],
+			experimental_steeringMessages: [],
+			tools: void 0,
+			stepTools: void 0,
+			schema: void 0,
+			schemaName: void 0,
+			schemaDescription: void 0,
+			...typeof filtered.output === "object" ? { output: void 0 } : {},
+			providerOptions: {},
+			value: "",
+			values: [],
+			query: "",
+			documents: [],
+			state: Array.isArray(filtered.state) ? [] : typeof filtered.state === "string" ? "" : {},
+			questions: {},
+			...isSpeech ? { text: "" } : {}
+		});
+		if (filtered.toolCall != null) filtered.toolCall = {
+			...filtered.toolCall,
+			input: void 0
+		};
+		if (filtered.toolOutput != null) filtered.toolOutput = {
+			...filtered.toolOutput,
+			input: void 0
+		};
+		if (Array.isArray(filtered.ranking)) filtered.ranking = filtered.ranking.map((entry) => "document" in entry ? {
+			...entry,
+			document: typeof entry.document === "string" ? "" : {}
+		} : entry);
+	}
+	if (hideOutputs) {
+		replace({
+			content: [],
+			...!isSpeech ? { text: "" } : {},
+			reasoning: Array.isArray(filtered.reasoning) ? [] : void 0,
+			reasoningText: void 0,
+			files: [],
+			sources: [],
+			toolCalls: [],
+			staticToolCalls: [],
+			dynamicToolCalls: [],
+			toolResults: [],
+			staticToolResults: [],
+			dynamicToolResults: [],
+			responseMessages: [],
+			object: void 0,
+			objectText: "",
+			embedding: [],
+			embeddings: [],
+			ranking: [],
+			answers: {},
+			segments: []
+		});
+		if (filtered.toolOutput != null) filtered.toolOutput = {
+			...filtered.toolOutput,
+			output: void 0,
+			error: void 0
+		};
+	}
+	replace({
+		providerMetadata: void 0,
+		error: filtered.error == null ? filtered.error : /* @__PURE__ */ new Error("AI SDK operation failed")
+	});
+	for (const key of ["toolCall", "toolOutput"]) {
+		const toolEvent = filtered[key];
+		if (toolEvent != null && typeof toolEvent === "object") filtered[key] = {
+			...toolEvent,
+			providerMetadata: void 0,
+			..."error" in toolEvent ? { error: toolEvent.error == null ? toolEvent.error : /* @__PURE__ */ new Error("AI SDK operation failed") } : {}
+		};
+	}
+	for (const key of ["request", "response"]) {
+		const filterMetadata = (metadata) => {
+			if (metadata == null || typeof metadata !== "object") return metadata;
+			const filteredMetadata = { ...metadata };
+			delete filteredMetadata.body;
+			if (key === "request" ? hideInputs : hideOutputs) {
+				if ("messages" in metadata) filteredMetadata.messages = [];
+				if ("instructions" in metadata) filteredMetadata.instructions = void 0;
+			}
+			return filteredMetadata;
+		};
+		const metadata = filtered[key];
+		if (metadata != null) filtered[key] = Array.isArray(metadata) ? metadata.map(filterMetadata) : filterMetadata(metadata);
+	}
+	if (Array.isArray(filtered.steps)) filtered.steps = filtered.steps.map((step) => filterTelemetryEvent(step, telemetry));
+	if (Object.hasOwn(event, "finalStep")) filtered.finalStep = filterTelemetryEvent(filtered.finalStep, telemetry);
+	return filtered;
+}
+//#endregion
 //#region src/telemetry/create-telemetry-dispatcher.ts
 function augmentEvent(event, telemetry, filterContext = false) {
 	const augmentedEvent = Object.assign(Object.create(Object.getPrototypeOf(event)), event, {
@@ -4177,7 +4359,7 @@ function augmentEvent(event, telemetry, filterContext = false) {
 			includeToolsContext: telemetry.includeToolsContext
 		});
 	}
-	return augmentedEvent;
+	return filterTelemetryEvent(augmentedEvent, telemetry);
 }
 /**
 * Creates a telemetry dispatcher that sends telemetry events
@@ -4203,6 +4385,7 @@ function createTelemetryDispatcher({ telemetry }) {
 		includeRuntimeContext: telemetry?.includeRuntimeContext,
 		includeToolsContext: telemetry?.includeToolsContext
 	};
+	const recordResult = telemetry?.recordInputs !== false && telemetry?.recordOutputs !== false;
 	const mergeTelemetryCallback = (key, deprecatedKey) => {
 		const integrationCallbacks = integrations.map((integration) => {
 			const callback = integration[key] ?? (deprecatedKey == null ? void 0 : integration[deprecatedKey]);
@@ -4222,13 +4405,14 @@ function createTelemetryDispatcher({ telemetry }) {
 		runInTracingChannelSpan: async ({ type, event, execute }) => await runWithTracingChannelSpan({
 			type,
 			event: augmentEvent(event, telemetryMetadata, true)
-		}, execute),
+		}, execute, recordResult),
 		startTracingChannelContext: ({ type, event, completion }) => openTelemetryChannelSpanContext({
 			message: {
 				type,
 				event: augmentEvent(event, telemetryMetadata, true)
 			},
-			completion
+			completion,
+			recordResult
 		}),
 		onStart: mergeTelemetryCallback("onStart"),
 		onStepStart: mergeTelemetryCallback("onStepStart"),
@@ -4270,7 +4454,7 @@ function createTelemetryDispatcher({ telemetry }) {
 			return await runWithTracingChannelSpan({
 				type: "languageModelCall",
 				event: augmentedEvent
-			}, wrappedExecute);
+			}, wrappedExecute, recordResult);
 		},
 		/**
 		* Composes all `executeTool` wrappers around the original tool execution.
@@ -4559,7 +4743,10 @@ var DefaultSteering = class {
 					message,
 					sentMessages
 				})) {
-					delivered.push(message);
+					delivered.push({
+						id,
+						message
+					});
 					resolve({
 						type: "delivered",
 						id,
@@ -4722,7 +4909,7 @@ async function toResponseMessages({ content: inputContent, tools }) {
 	for (const part of inputContent) {
 		if (part.type === "source") continue;
 		if ((part.type === "tool-result" || part.type === "tool-error") && !part.providerExecuted) continue;
-		if (part.type === "text" && part.text.length === 0) continue;
+		if (part.type === "text" && part.text.length === 0 && part.providerMetadata == null) continue;
 		switch (part.type) {
 			case "text":
 				content.push({
@@ -4973,9 +5160,19 @@ async function validateApprovedToolApprovals({ approvedToolApprovals, tools, too
 	const approved = [];
 	const denied = [];
 	const invalid = [];
-	for (const approval of approvedToolApprovals) {
+	for (const collectedApproval of approvedToolApprovals) {
+		const tool = getOwn(tools, collectedApproval.toolCall.toolName);
+		const { toolMetadata: _toolMetadata, dynamic: _dynamic, ...toolCallFromHistory } = collectedApproval.toolCall;
+		const approval = {
+			...collectedApproval,
+			toolCall: {
+				...toolCallFromHistory,
+				...tool?.metadata != null ? { toolMetadata: tool.metadata } : {},
+				...tool?.type === "dynamic" ? { dynamic: true } : {},
+				title: tool?.title
+			}
+		};
 		const { approvalRequest, toolCall } = approval;
-		const tool = getOwn(tools, toolCall.toolName);
 		if (toolApprovalSecret != null) {
 			if (approvalRequest.signature == null) throw new InvalidToolApprovalSignatureError({
 				approvalId: approvalRequest.approvalId,
@@ -5462,10 +5659,10 @@ async function generateTextInternal({ model: modelArg, tools, toolChoice, instru
 								messages: prepareStepResult?.messages ?? stepInputMessages,
 								toolCallerMessages
 							});
-							const deliveredSteeringMessages = steeringRun?.resolveTakenMessages({
+							const deliveredSteeringMessages = (steeringRun?.resolveTakenMessages({
 								stepNumber,
 								sentMessages: stepMessages
-							}) ?? [];
+							}) ?? []).map(({ message }) => message);
 							const promptMessages = await convertToLanguageModelPrompt({
 								prompt: {
 									instructions: stepInstructions,
@@ -5547,6 +5744,7 @@ async function generateTextInternal({ model: modelArg, tools, toolChoice, instru
 								};
 								return {
 									...result,
+									...result.request != null ? { request: sanitizeRequestMetadata(result.request) } : {},
 									response: responseData
 								};
 							});
@@ -5584,6 +5782,7 @@ async function generateTextInternal({ model: modelArg, tools, toolChoice, instru
 									usage: stepUsage,
 									content: modelCallContent,
 									responseId: currentModelResponse.response.id,
+									...currentModelResponse.request?.url != null ? { request: { url: getRequestOrigin(currentModelResponse.request.url) } } : {},
 									...currentModelResponse.providerMetadata != null ? { providerMetadata: currentModelResponse.providerMetadata } : {},
 									performance: {
 										responseTimeMs,
@@ -6425,6 +6624,22 @@ const uiMessageChunkSchema = lazySchema(() => zodSchema$1(z.union([
 		mediaType: z.string(),
 		providerMetadata: providerMetadataSchema.optional()
 	}),
+	z.looseObject({
+		type: z.literal("steering-message"),
+		id: z.string(),
+		parts: z.array(z.union([z.looseObject({
+			type: z.literal("text"),
+			text: z.string(),
+			providerMetadata: providerMetadataSchema.optional()
+		}), z.looseObject({
+			type: z.literal("file"),
+			mediaType: z.string(),
+			url: z.string(),
+			filename: z.string().optional(),
+			providerReference: z.record(z.string(), z.string()).optional(),
+			providerMetadata: providerMetadataSchema.optional()
+		})]))
+	}),
 	z.looseObject({
 		type: z.custom((value) => typeof value === "string" && value.startsWith("data-"), { message: "Type must start with \"data-\"" }),
 		id: z.string().optional(),
@@ -7068,6 +7283,14 @@ function processUIMessageStream({ stream, messageMetadataSchema, dataPartSchemas
 					write();
 					break;
 				}
+				case "steering-message":
+					state.message.parts.push({
+						type: "steering-message",
+						id: chunk.id,
+						parts: chunk.parts
+					});
+					write();
+					break;
 				case "start-step":
 					state.message.parts.push({ type: "step-start" });
 					break;
@@ -7415,6 +7638,11 @@ function toUIMessageChunk(part, { tools, sendReasoning = true, sendSources = fal
 			type: "error",
 			errorText: onError(part.error)
 		};
+		case "steering-message": return {
+			type: "steering-message",
+			id: part.id,
+			parts: convertUserMessageToUIParts(part.message)
+		};
 		case "start-step": return { type: "start-step" };
 		case "finish-step": return { type: "finish-step" };
 		case "start":
@@ -7433,10 +7661,53 @@ function toUIMessageChunk(part, { tools, sendReasoning = true, sendSources = fal
 			};
 		case "abort": return part;
 		case "tool-input-end":
+		case "ping":
 		case "raw": return;
 		default: throw new Error(`Unknown chunk type: ${partType}`);
 	}
 }
+function convertUserMessageToUIParts(message) {
+	if (typeof message.content === "string") return [{
+		type: "text",
+		text: message.content
+	}];
+	return message.content.map((part) => {
+		const providerMetadata = part.providerOptions != null ? { providerMetadata: part.providerOptions } : {};
+		if (part.type === "text") return {
+			type: "text",
+			text: part.text,
+			...providerMetadata
+		};
+		const { data, mediaType: dataUrlMediaType } = convertToLanguageModelV4FilePart(part.type === "image" ? part.image : part.data);
+		const mediaType = dataUrlMediaType ?? part.mediaType ?? "image";
+		const filename = part.type === "file" ? part.filename : void 0;
+		const file = {
+			type: "file",
+			mediaType,
+			...filename != null ? { filename } : {},
+			...providerMetadata
+		};
+		switch (data.type) {
+			case "url": return {
+				...file,
+				url: data.url.toString()
+			};
+			case "data": return {
+				...file,
+				url: `data:${mediaType};base64,${convertDataContentToBase64String(data.data)}`
+			};
+			case "text": return {
+				...file,
+				url: `data:${mediaType};base64,${convertUint8ArrayToBase64(new TextEncoder().encode(data.text))}`
+			};
+			case "reference": return {
+				...file,
+				url: "",
+				providerReference: data.reference
+			};
+		}
+	});
+}
 //#endregion
 //#region src/ui-message-stream/to-ui-message-stream.ts
 /**
@@ -7969,6 +8240,7 @@ const isOutputChunkType = {
 	"model-call-response-metadata": false,
 	"model-call-end": false,
 	error: false,
+	ping: false,
 	raw: false
 };
 function isOutputChunk$1(chunk) {
@@ -8274,7 +8546,7 @@ async function streamLanguageModelCall({ model, tools, toolOrder, output, toolCh
 		callbacks: onLanguageModelCallStart
 	});
 	const callStartTimestampMs = now$3();
-	const { stream: languageModelStream, response, request } = await executeLanguageModelCallInTelemetryContext({
+	const { stream: languageModelStream, response, request: rawRequest } = await executeLanguageModelCallInTelemetryContext({
 		...languageModelCallStartEvent,
 		execute: async () => await resolvedModel.doStream({
 			...callSettings,
@@ -8288,6 +8560,7 @@ async function streamLanguageModelCall({ model, tools, toolOrder, output, toolCh
 			includeRawChunks
 		})
 	});
+	const request = sanitizeRequestMetadata(rawRequest);
 	return {
 		stream: createAsyncIterableStream(languageModelStream.pipeThrough(createLanguageModelV4StreamPartToLanguageModelStreamPartTransform({
 			tools,
@@ -8303,6 +8576,7 @@ async function streamLanguageModelCall({ model, tools, toolOrder, output, toolCh
 			generateId,
 			now: now$3,
 			callStartTimestampMs,
+			request,
 			onLanguageModelCallEnd,
 			collectContent
 		}))),
@@ -8310,7 +8584,7 @@ async function streamLanguageModelCall({ model, tools, toolOrder, output, toolCh
 		request
 	};
 }
-function createLanguageModelV4StreamPartToLanguageModelStreamPartTransform({ tools, instructions, messages, repairToolCall, refineToolInput, abortSignal, callId, provider, modelId, toolChoice, generateId, now, callStartTimestampMs, onLanguageModelCallEnd, collectContent }) {
+function createLanguageModelV4StreamPartToLanguageModelStreamPartTransform({ tools, instructions, messages, repairToolCall, refineToolInput, abortSignal, callId, provider, modelId, toolChoice, generateId, now, callStartTimestampMs, request, onLanguageModelCallEnd, collectContent }) {
 	const toolCallsByToolCallId = /* @__PURE__ */ new Map();
 	const modelCallContent = [];
 	const rawModelCallContent = [];
@@ -8488,6 +8762,7 @@ function createLanguageModelV4StreamPartToLanguageModelStreamPartTransform({ too
 						usage,
 						content: modelCallContent,
 						responseId,
+						...request?.url != null ? { request: { url: getRequestOrigin(request.url) } } : {},
 						...chunk.providerMetadata != null ? { providerMetadata: chunk.providerMetadata } : {},
 						performance
 					},
@@ -9782,10 +10057,20 @@ var DefaultStreamTextResult = class {
 						toolCallerMessages
 					});
 					currentStepMessages = stepMessages;
-					currentStepSteeringMessages = steeringRun?.resolveTakenMessages({
+					const deliveredSteeringMessages = steeringRun?.resolveTakenMessages({
 						stepNumber: currentStep,
 						sentMessages: stepMessages
 					}) ?? [];
+					currentStepSteeringMessages = deliveredSteeringMessages.map(({ message }) => message);
+					if (deliveredSteeringMessages.length > 0) self.addStream(new ReadableStream({ start(controller) {
+						for (const { id, message } of deliveredSteeringMessages) controller.enqueue({
+							type: "steering-message",
+							id,
+							message,
+							stepNumber: currentStep
+						});
+						controller.close();
+					} }));
 					const stepInstructions = prepareStepResult?.instructions ?? prepareStepResult?.system ?? instructionsForNextStep;
 					instructionsForNextStep = stepInstructions;
 					const stepProviderOptions = mergeObjects(providerOptions, prepareStepResult?.providerOptions);
@@ -9930,7 +10215,7 @@ var DefaultStreamTextResult = class {
 									clearFirstChunkTimeout();
 									resetChunkTimeout();
 								}
-								if (shouldBufferToolParts && value.type !== "error" && (isToolPart || bufferedAttemptParts.length > 0)) {
+								if (shouldBufferToolParts && value.type !== "error" && value.type !== "ping" && (isToolPart || bufferedAttemptParts.length > 0)) {
 									bufferedAttemptParts.push(value);
 									continue;
 								}
@@ -10095,6 +10380,7 @@ var DefaultStreamTextResult = class {
 								case "tool-input-end":
 								case "tool-input-delta":
 								case "tool-approval-request":
+								case "ping":
 								case "tool-output-denied":
 									enqueueStepPart(controller, chunk);
 									break;
@@ -11033,24 +11319,7 @@ async function convertToModelMessages(messages, options) {
 			modelMessages.push({
 				role: "user",
 				content: message.parts.map((part) => {
-					if (isTextUIPart(part)) return {
-						type: "text",
-						text: part.text,
-						...part.providerMetadata != null ? { providerOptions: part.providerMetadata } : {}
-					};
-					if (isFileUIPart(part)) return {
-						type: "file",
-						mediaType: part.mediaType,
-						filename: part.filename,
-						data: part.providerReference != null ? {
-							type: "reference",
-							reference: part.providerReference
-						} : {
-							type: "url",
-							url: new URL(part.url)
-						},
-						...part.providerMetadata != null ? { providerOptions: part.providerMetadata } : {}
-					};
+					if (isTextUIPart(part) || isFileUIPart(part)) return convertUserContentUIPart(part);
 					if (isDataUIPart(part)) return options?.convertDataPart?.(part);
 				}).filter(isNonNullable)
 			});
@@ -11206,6 +11475,13 @@ async function convertToModelMessages(messages, options) {
 				}
 				for (const part of message.parts) if (isCustomContentUIPart(part) || isTextUIPart(part) || isReasoningUIPart(part) || isReasoningFileUIPart(part) || isFileUIPart(part) || isToolUIPart(part) || isDataUIPart(part)) block.push(part);
 				else if (part.type === "step-start") await processBlock();
+				else if (part.type === "steering-message") {
+					await processBlock();
+					modelMessages.push({
+						role: "user",
+						content: part.parts.map(convertUserContentUIPart)
+					});
+				}
 				await processBlock();
 				break;
 			}
@@ -11220,6 +11496,26 @@ async function convertToModelMessages(messages, options) {
 	}
 	return modelMessages;
 }
+function convertUserContentUIPart(part) {
+	if (part.type === "text") return {
+		type: "text",
+		text: part.text,
+		...part.providerMetadata != null ? { providerOptions: part.providerMetadata } : {}
+	};
+	return {
+		type: "file",
+		mediaType: part.mediaType,
+		filename: part.filename,
+		data: part.providerReference != null ? {
+			type: "reference",
+			reference: part.providerReference
+		} : {
+			type: "url",
+			url: new URL(part.url)
+		},
+		...part.providerMetadata != null ? { providerOptions: part.providerMetadata } : {}
+	};
+}
 //#endregion
 //#region src/ui/validate-ui-messages.ts
 const toolMetadataSchema = z.record(z.string(), jsonValueSchema.optional());
@@ -11312,6 +11608,22 @@ const uiMessagesSchema = lazySchema(() => {
 				url: z.string(),
 				providerMetadata: providerMetadataSchema.optional()
 			}),
+			z.object({
+				type: z.literal("steering-message"),
+				id: z.string(),
+				parts: z.array(z.union([z.object({
+					type: z.literal("text"),
+					text: z.string(),
+					providerMetadata: providerMetadataSchema.optional()
+				}), z.object({
+					type: z.literal("file"),
+					mediaType: z.string(),
+					filename: z.string().optional(),
+					url: z.string(),
+					providerReference: providerReferenceSchema$1.optional(),
+					providerMetadata: providerMetadataSchema.optional()
+				})]))
+			}),
 			z.object({ type: z.literal("step-start") }),
 			z.object({
 				type: z.string().startsWith("data-"),
@@ -13854,7 +14166,7 @@ async function generateObject(options) {
 		const usage = asLanguageModelUsage(generateResult.usage);
 		const warnings = generateResult.warnings;
 		const resultProviderMetadata = generateResult.providerMetadata;
-		const request = generateResult.request ?? {};
+		const request = sanitizeRequestMetadata(generateResult.request) ?? {};
 		const response = responseData;
 		logWarnings({
 			warnings,
@@ -14313,7 +14625,8 @@ var DefaultStreamObjectResult = class {
 				}
 			} };
 			const startTimestampMs = now();
-			const { stream, response, request } = await retry(() => model.doStream(callOptions));
+			const { stream, response, request: rawRequest } = await retry(() => model.doStream(callOptions));
+			const request = sanitizeRequestMetadata(rawRequest);
 			self._request.resolve(request ?? {});
 			let warnings;
 			let usage = createNullLanguageModelUsage();
@@ -15244,6 +15557,10 @@ function smoothStream({ delayInMs = 10, chunking = "word", _internal: { delay: d
 			}
 		}
 		return new TransformStream({ async transform(chunk, controller) {
+			if (chunk.type === "ping") {
+				controller.enqueue(chunk);
+				return;
+			}
 			if (chunk.type !== "text-delta" && chunk.type !== "reasoning-delta") {
 				flushBuffer(controller);
 				controller.enqueue(chunk);
@@ -15414,10 +15731,14 @@ async function experimental_generateVideo({ model: modelArg, prompt: promptArg,
 		return retry(() => model.doGenerate(callOptions));
 	}));
 	const videos = [];
+	let uploadedVideos = 0;
 	const responses = [];
 	const providerMetadata = {};
 	for (const result of results) {
 		for (const videoData of result.videos) switch (videoData.type) {
+			case "uploaded":
+				uploadedVideos++;
+				break;
 			case "url": {
 				const { data, mediaType: downloadedMediaType } = await downloadFn({
 					url: new URL(videoData.url),
@@ -15461,19 +15782,30 @@ async function experimental_generateVideo({ model: modelArg, prompt: promptArg,
 		});
 		if (result.providerMetadata != null) mergeProviderMetadata(providerMetadata, result.providerMetadata);
 	}
-	if (videos.length === 0) throw new NoVideoGeneratedError({ responses });
+	if (videos.length === 0 && uploadedVideos === 0) throw new NoVideoGeneratedError({ responses });
 	if (warnings.length > 0) logWarnings({
 		warnings,
 		provider: model.provider,
 		model: model.modelId
 	});
-	return {
+	const result = {
 		video: videos[0],
 		videos,
+		uploadedVideos,
 		warnings,
 		responses,
 		providerMetadata
 	};
+	if (videos.length === 0) Object.defineProperty(result, "video", {
+		enumerable: false,
+		get() {
+			throw new NoVideoGeneratedError({
+				message: "Videos were delivered to the upload destination. Use uploadedVideos instead of video.",
+				responses
+			});
+		}
+	});
+	return result;
 }
 async function executeStartStatusFlow({ model, callOptions, poll: pollConfig, webhook: webhookFactory, maxRetries, retry }) {
 	const earlyWarnings = [];
@@ -18029,6 +18361,7 @@ var BrowserRealtimeWebRTC = class {
 var RealtimeAttempt = class {
 	constructor() {
 		this.abort = new AbortController();
+		this.reportedTransportErrors = /* @__PURE__ */ new WeakSet();
 		this.timers = /* @__PURE__ */ new Map();
 		this.active = true;
 		this.ready = false;
@@ -18534,9 +18867,13 @@ var AbstractRealtimeSession = class {
 	}
 	validateConnection() {
 		const { model, api } = this.options;
-		if ((!this.continuous || api.session != null) && this.options.maxPlaybackBufferSeconds != null) throw new Error("maxPlaybackBufferSeconds is supported only for continuous PCM sessions");
-		const connection = api.session != null ? "webrtc" : api.token != null ? "client-secret-websocket" : "server-websocket";
-		if (!(model.capabilities?.connections ?? ["client-secret-websocket"]).includes(connection) || model.capabilities?.transports != null && !model.capabilities.transports.includes(api.session != null ? "webrtc" : "websocket")) throw new Error(`Realtime model does not support ${connection}`);
+		if ((!this.continuous || api.session != null || api.provider != null) && this.options.maxPlaybackBufferSeconds != null) throw new Error("maxPlaybackBufferSeconds is supported only for continuous PCM sessions");
+		const connection = api.provider != null ? "provider" : api.session != null ? "webrtc" : api.token != null ? "client-secret-websocket" : "server-websocket";
+		if (!(model.capabilities?.connections ?? ["client-secret-websocket"]).includes(connection) || api.provider == null && model.capabilities?.transports != null && !model.capabilities.transports.includes(api.session != null ? "webrtc" : "websocket")) throw new Error(`Realtime model does not support ${connection}`);
+		if (api.provider != null) {
+			if (model.createTransport == null) throw new Error("Realtime model does not support a provider transport");
+			return;
+		}
 		if (api.session != null) {
 			if (model.getWebRTCConfig == null) throw new Error("Realtime model does not support WebRTC configuration");
 			return;
@@ -18560,7 +18897,7 @@ var AbstractRealtimeSession = class {
 			if (connectOptions?.capture === false) this.stopAudioCapture();
 			if (!current()) return;
 			if (connectOptions?.stream != null && connectOptions.stream !== this.suppliedStream) {
-				if (this.captureRequested && !this.sessionLifecycle && !this.continuous && api.session == null) this.startAudioCapture(connectOptions.stream);
+				if (this.captureRequested && !this.sessionLifecycle && !this.continuous && api.session == null && api.provider == null) this.startAudioCapture(connectOptions.stream);
 				else this.suppliedStream = connectOptions.stream;
 			}
 			if (!current()) return;
@@ -18582,7 +18919,7 @@ var AbstractRealtimeSession = class {
 			attempt.timer("startup", this.options.startupTimeoutMs ?? 3e4, () => {
 				if (current()) this.fail(/* @__PURE__ */ new Error("Realtime session startup timed out"));
 			});
-			const reportedTransportErrors = /* @__PURE__ */ new WeakSet();
+			const reportedTransportErrors = attempt.reportedTransportErrors;
 			const callbacks = {
 				model,
 				onEvent: async (event) => {
@@ -18590,7 +18927,7 @@ var AbstractRealtimeSession = class {
 					try {
 						await this.handleServerEvent(event, attempt);
 					} catch (error) {
-						if (current()) this.fail(error, (this.rtc ?? this.pcm ?? this.transport)?.finish());
+						if (current()) this.fail(error, (this.providerTransport ?? this.rtc ?? this.pcm ?? this.transport)?.finish());
 					}
 				},
 				onError: (error) => {
@@ -18618,6 +18955,13 @@ var AbstractRealtimeSession = class {
 				onClose: (error) => {
 					if (!current()) return;
 					const finalizationConfirmed = model.capabilities?.finalization === "session-close" && this.state.session?.finalization === "confirmed";
+					if (api.provider != null && attempt.closing) {
+						if (error != null && !finalizationConfirmed && !reportedTransportErrors.has(error)) {
+							reportedTransportErrors.add(error);
+							this.reportError(error, attempt);
+						}
+						return;
+					}
 					if (error != null && !finalizationConfirmed) this.fail(error, void 0, !reportedTransportErrors.has(error));
 					else if (!attempt.ready && !finalizationConfirmed) this.fail(/* @__PURE__ */ new Error("Realtime connection closed before becoming ready"));
 					else this.disconnect();
@@ -18635,7 +18979,21 @@ var AbstractRealtimeSession = class {
 					});
 				}
 			};
-			if (api.session != null) {
+			if (api.provider != null) {
+				const transport = model.createTransport(callbacks);
+				if (!current()) {
+					transport.dispose();
+					return;
+				}
+				this.providerTransport = transport;
+				await transport.connect({
+					api: api.provider,
+					sessionConfig,
+					stream: connectOptions?.stream,
+					capture: connectOptions?.capture,
+					abortSignal: attempt.abort.signal
+				});
+			} else if (api.session != null) {
 				this.rtc = new BrowserRealtimeWebRTC({
 					...callbacks,
 					disconnectTimeoutMs: this.options.rtcDisconnectTimeoutMs
@@ -18751,8 +19109,12 @@ var AbstractRealtimeSession = class {
 	closeFailed(attempt, error) {
 		if (this.attempt !== attempt || !attempt.active || attempt.transportClosing) return;
 		attempt.clearTimer("close");
-		const transport = this.rtc ?? this.pcm ?? this.transport;
+		const transport = this.providerTransport ?? this.rtc ?? this.pcm ?? this.transport;
 		this.drainAttempt(attempt, transport?.finish());
+		if (error instanceof Error) {
+			if (attempt.reportedTransportErrors.has(error)) return;
+			attempt.reportedTransportErrors.add(error);
+		}
 		this.reportError(error, attempt);
 	}
 	disconnect() {
@@ -18764,16 +19126,19 @@ var AbstractRealtimeSession = class {
 		const audio = this.audio;
 		const pcm = this.pcm;
 		const rtc = this.rtc;
+		const providerTransport = this.providerTransport;
 		this.transport = void 0;
 		this.audio = void 0;
 		this.pcm = void 0;
 		this.rtc = void 0;
+		this.providerTransport = void 0;
 		this.commands = void 0;
 		attempt?.retire();
 		transport?.dispose();
 		audio?.dispose();
 		pcm?.dispose();
 		rtc?.dispose();
+		providerTransport?.dispose();
 		if (this.attempt !== attempt) return;
 		const session = this.state.session;
 		this.applyState({
@@ -18787,10 +19152,27 @@ var AbstractRealtimeSession = class {
 			} } : {}
 		});
 	}
-	/** Wait for final usage when the model supports a session-close acknowledgement. */
+	/** Wait for provider cleanup or supported session-close finalization. */
 	close(options) {
 		const attempt = this.attempt;
 		if (attempt?.closePromise != null) return attempt.closePromise;
+		const providerTransport = this.providerTransport;
+		if (providerTransport != null && attempt?.active) {
+			const promise = attempt.beginClose();
+			attempt.clearTimer("startup");
+			this.applyState({
+				...this.state,
+				status: "closing"
+			});
+			if (this.attempt !== attempt || !attempt.active) return promise;
+			this.stopAudioCapture();
+			if (this.attempt !== attempt || !attempt.active) return promise;
+			attempt.timer("close", this.options.closeTimeoutMs ?? 15e3, () => {
+				if (this.attempt === attempt && attempt.active) this.fail(/* @__PURE__ */ new Error("Realtime provider session cleanup timed out"));
+			});
+			providerTransport.close().then(() => this.finishAttempt(attempt), (error) => this.closeFailed(attempt, error));
+			return promise;
+		}
 		if (this.state.status !== "connected" || this.options.model.capabilities?.finalization !== "session-close" || attempt == null) {
 			this.disconnect();
 			return Promise.resolve();
@@ -18817,8 +19199,16 @@ var AbstractRealtimeSession = class {
 		return promise;
 	}
 	sendTransport(event, guard, automaticAudio = false) {
-		const transport = this.rtc ?? this.pcm ?? this.transport;
+		const transport = this.providerTransport ?? this.rtc ?? this.pcm ?? this.transport;
 		if (transport == null) throw new Error("Realtime connection is not open");
+		if (transport === this.providerTransport) {
+			const attempt = this.attempt;
+			const sent = transport.sendEvent(event, guard);
+			sent.catch((error) => {
+				if (attempt?.active && this.attempt === attempt) this.reportError(error, attempt);
+			});
+			return sent;
+		}
 		if (transport === this.transport) return this.transport.sendEvent(event, guard, automaticAudio);
 		return transport.sendEvent(event, guard);
 	}
@@ -18904,14 +19294,14 @@ var AbstractRealtimeSession = class {
 		this.maybeRequestToolResponse();
 	}
 	maybeRequestToolResponse() {
-		if (this.state.status !== "connected" || this.attempt?.closing || !this.responseToolCallsClosed || this.toolCallsInResponse.size === 0 || [...this.toolCallsInResponse].some((id) => !this.submittedToolOutputs.has(id))) return;
+		if (this.providerTransport != null || this.state.status !== "connected" || this.attempt?.closing || !this.responseToolCallsClosed || this.toolCallsInResponse.size === 0 || [...this.toolCallsInResponse].some((id) => !this.submittedToolOutputs.has(id))) return;
 		this.sendEvent({ type: "response-create" });
 		this.toolCallsInResponse.clear();
 		this.submittedToolOutputs.clear();
 		this.responseToolCallsClosed = false;
 	}
 	startAudioCapture(stream) {
-		const legacy = !this.sessionLifecycle && !this.continuous && this.options.api.session == null;
+		const legacy = !this.sessionLifecycle && !this.continuous && this.options.api.session == null && this.options.api.provider == null;
 		if (!legacy && this.state.status !== "connected") throw new Error("Realtime session is not accepting capture");
 		if (this.attempt?.active && (this.attempt.closing || this.attempt.cause != null)) throw new Error("Realtime session is not accepting capture");
 		this.suppliedStream = stream;
@@ -18930,6 +19320,7 @@ var AbstractRealtimeSession = class {
 		const accepting = () => !this.attempt?.closing && this.attempt?.cause == null && (this.state.status === "connected" || !this.sessionLifecycle && this.state.status === "connecting");
 		if (!accepting()) throw new Error("Realtime session is not accepting capture");
 		this.captureRequested = true;
+		if (this.providerTransport != null) return this.providerTransport.startCapture(this.suppliedStream);
 		if (this.rtc != null) return this.rtc.startCapture(this.suppliedStream);
 		if (this.pcm != null) return this.pcm.resumeCapture(this.suppliedStream);
 		const audio = this.audio;
@@ -18951,7 +19342,11 @@ var AbstractRealtimeSession = class {
 		const pcm = this.pcm;
 		const audio = this.audio;
 		const rtc = this.rtc;
+		const providerTransport = this.providerTransport;
 		const attempt = this.attempt;
+		if (providerTransport != null) providerTransport.stopCapture().catch((error) => {
+			if (attempt?.active && this.attempt === attempt) this.reportError(error, attempt);
+		});
 		if (rtc != null) rtc.stopCapture().catch((error) => {
 			if (attempt?.active && this.attempt === attempt) this.reportError(error, attempt);
 		});
@@ -18959,10 +19354,10 @@ var AbstractRealtimeSession = class {
 		audio?.stopCapture();
 	}
 	stopPlayback() {
-		(this.rtc ?? this.pcm ?? this.audio)?.stopPlayback();
+		(this.providerTransport ?? this.rtc ?? this.pcm ?? this.audio)?.stopPlayback();
 	}
 	async resumePlayback() {
-		await (this.rtc ?? this.pcm ?? this.audio)?.resumePlayback();
+		await (this.providerTransport ?? this.rtc ?? this.pcm ?? this.audio)?.resumePlayback();
 	}
 	dispose() {
 		this.disconnect();
@@ -19695,7 +20090,7 @@ const originalGenerateCallId$1 = createIdGenerator$1({
 * @returns A result object that contains the generated transcript.
 */
 const defaultDownload = createDownload();
-async function transcribe({ model, audio, providerOptions = {}, maxRetries: maxRetriesArg, abortSignal, headers, telemetry, download: downloadFn = defaultDownload, _internal: { generateCallId = originalGenerateCallId$1 } = {} }) {
+async function transcribe({ model, audio, providerOptions = {}, timestampGranularities, maxRetries: maxRetriesArg, abortSignal, headers, telemetry, download: downloadFn = defaultDownload, _internal: { generateCallId = originalGenerateCallId$1 } = {} }) {
 	const resolvedModel = resolveTranscriptionModel(model);
 	if (!resolvedModel) throw new Error("Model could not be resolved");
 	const { maxRetries, retry } = prepareRetries({
@@ -19718,7 +20113,8 @@ async function transcribe({ model, audio, providerOptions = {}, maxRetries: maxR
 		inputAudioFormat: void 0,
 		maxRetries,
 		headers,
-		providerOptions
+		providerOptions,
+		timestampGranularities
 	});
 	const executeWithStartEvent = async ({ startEvent, execute }) => await runInTracingChannelSpan({
 		type: "transcribe",
@@ -19737,12 +20133,15 @@ async function transcribe({ model, audio, providerOptions = {}, maxRetries: maxR
 	let downloadedMediaType;
 	try {
 		if (audio instanceof URL) {
-			const downloadResult = await downloadFn({
-				url: audio,
-				abortSignal
-			});
-			audioData = downloadResult.data;
-			downloadedMediaType = downloadResult.mediaType;
+			if (resolvedModel.doGenerateFromUrl != null && downloadFn === defaultDownload && (audio.protocol === "https:" || audio.protocol === "http:")) audioData = audio;
+			else {
+				const downloadResult = await downloadFn({
+					url: audio,
+					abortSignal
+				});
+				audioData = downloadResult.data;
+				downloadedMediaType = downloadResult.mediaType;
+			}
 		} else audioData = convertDataContentToUint8Array(audio);
 	} catch (error) {
 		preparationFailed = true;
@@ -19769,25 +20168,49 @@ async function transcribe({ model, audio, providerOptions = {}, maxRetries: maxR
 			throw error;
 		}
 	}
-	const mediaType = downloadedMediaType ?? detectMediaType({
+	let mediaType = audioData instanceof URL ? void 0 : getAudioMediaType(downloadedMediaType) ?? detectMediaType({
 		data: audioData,
 		topLevelType: "audio"
 	}) ?? "audio/wav";
+	let byteLength = audioData instanceof URL ? void 0 : audioData.byteLength;
 	const startEvent = createStartEvent({
-		byteLength: audioData.byteLength,
+		byteLength,
 		mediaType
 	});
 	try {
 		return await executeWithStartEvent({
 			startEvent,
 			execute: async () => {
-				const result = await retry(() => resolvedModel.doGenerate({
-					audio: audioData,
-					abortSignal,
-					headers: headersWithUserAgent,
-					providerOptions,
-					mediaType
-				}));
+				const result = await retry(async () => {
+					const callOptions = {
+						abortSignal,
+						headers: headersWithUserAgent,
+						providerOptions,
+						...timestampGranularities != null ? { timestampGranularities } : {}
+					};
+					if (audioData instanceof URL) {
+						const nativeResult = await resolvedModel.doGenerateFromUrl?.({
+							...callOptions,
+							audio: audioData
+						});
+						if (nativeResult != null) return nativeResult;
+						const downloaded = await downloadFn({
+							url: audioData,
+							abortSignal
+						});
+						audioData = downloaded.data;
+						mediaType = getAudioMediaType(downloaded.mediaType) ?? detectMediaType({
+							data: audioData,
+							topLevelType: "audio"
+						}) ?? "audio/wav";
+						byteLength = audioData.byteLength;
+					}
+					return resolvedModel.doGenerate({
+						...callOptions,
+						audio: audioData,
+						mediaType: mediaType ?? "audio/wav"
+					});
+				});
 				logWarnings({
 					warnings: result.warnings,
 					provider: resolvedModel.provider,
@@ -19801,11 +20224,12 @@ async function transcribe({ model, audio, providerOptions = {}, maxRetries: maxR
 						provider: resolvedModel.provider,
 						modelId: resolvedModel.modelId,
 						audio: {
-							byteLength: audioData.byteLength,
+							byteLength,
 							mediaType
 						},
 						text: result.text,
 						segments: result.segments,
+						words: result.words,
 						language: result.language,
 						durationInSeconds: result.durationInSeconds,
 						usage: result.usage,
@@ -19818,6 +20242,7 @@ async function transcribe({ model, audio, providerOptions = {}, maxRetries: maxR
 				return new DefaultTranscriptionResult({
 					text: result.text,
 					segments: result.segments,
+					words: result.words,
 					language: result.language,
 					durationInSeconds: result.durationInSeconds,
 					warnings: result.warnings,
@@ -19834,10 +20259,22 @@ async function transcribe({ model, audio, providerOptions = {}, maxRetries: maxR
 		throw error;
 	}
 }
+/**
+* Normalizes a downloaded `Content-Type` header value and returns it only
+* when it is a full `audio/<subtype>` media type. Generic or mislabeled types
+* (e.g. `application/octet-stream`, `binary/octet-stream`, `text/plain`)
+* return `undefined` so that the media type is detected from the audio bytes
+* instead.
+*/
+function getAudioMediaType(contentType) {
+	const mediaType = contentType?.split(";")[0].trim().toLowerCase();
+	return mediaType != null && getTopLevelMediaType(mediaType) === "audio" && isFullMediaType(mediaType) ? mediaType : void 0;
+}
 var DefaultTranscriptionResult = class {
 	constructor(options) {
 		this.text = options.text;
 		this.segments = options.segments;
+		this.words = options.words;
 		this.language = options.language;
 		this.durationInSeconds = options.durationInSeconds;
 		this.warnings = options.warnings;