workers-ai-provider 3.3.0 → 4.0.0

package/dist/index.mjs

@@ -1,4 +1,4 @@
-import { C as GatewayDelegateError, S as headersToObject, _ as detectProviderByUrl, a as WORKERS_AI_ERROR_CODE_TO_STATUS, b as asText, c as parseWorkersAIErrorCode, d as isForcedToolChoice, f as normalizeMessagesForBinding$1, g as GATEWAY_PROVIDERS, h as createResumableStream, i as WorkersAIGatewayError, l as SSEDecoder, m as processText, n as createGatewayProvider, o as isAbortError, p as parseLeakedToolCalls$1, r as WorkersAIFallbackError, s as messageOf, t as createGatewayFetch, u as getToolNames, v as findProviderBySlug, w as _defineProperty, x as buildGatewayEntry, y as wireableProviders } from "./gateway-provider-BjPLZLob.mjs";
+import { C as GatewayDelegateError, S as headersToObject, _ as detectProviderByUrl, a as WORKERS_AI_ERROR_CODE_TO_STATUS, b as asText, c as parseWorkersAIErrorCode, d as isForcedToolChoice, f as normalizeMessagesForBinding$1, g as GATEWAY_PROVIDERS, h as createResumableStream, i as WorkersAIGatewayError, l as SSEDecoder, m as processText, n as createGatewayProvider, o as isAbortError, p as parseLeakedToolCalls$1, r as WorkersAIFallbackError, s as messageOf, t as createGatewayFetch, u as getToolNames, v as findProviderBySlug, w as _defineProperty, x as buildGatewayEntry, y as wireableProviders } from "./gateway-provider-BIZ2bimW.mjs";
 import { APICallError, TooManyEmbeddingValuesForCallError, UnsupportedFunctionalityError } from "@ai-sdk/provider";
 import { generateId } from "ai";
 //#region src/workersai-error.ts
@@ -294,7 +294,7 @@ function processToolCalls(output) {
 *
 * The recovery logic (which JSON shapes count as a leaked call) lives in
 * `@cloudflare/gateway-core`; this wrapper only layers the framework id on each
-* neutral result so the existing `LanguageModelV3ToolCall` shape is preserved.
+* neutral result so the existing `LanguageModelV4ToolCall` shape is preserved.
 */
 function parseLeakedToolCalls(text, knownToolNames) {
 	return parseLeakedToolCalls$1(text, knownToolNames).map((call) => ({
@@ -341,28 +341,41 @@ function salvageToolCallsFromText(output, context) {
 //#endregion
 //#region src/convert-to-workersai-chat-messages.ts
 /**
-* Normalise any LanguageModelV3DataContent value to a Uint8Array.
+* Normalise a tagged `SharedV4FileData` value to a Uint8Array.
 *
 * Handles:
-*   - Uint8Array  → returned as-is
-*   - string      → decoded from base64 (with or without data-URL prefix)
-*   - URL         → not supported (Workers AI needs raw bytes, not a reference)
+*   - { type: 'data' }      → Uint8Array returned as-is; string decoded from
+*                             base64 (with or without data-URL prefix)
+*   - { type: 'url' }       → not supported (Workers AI needs raw bytes, not a reference)
+*   - { type: 'reference' } → not supported (Workers AI has no file store to resolve it)
+*   - { type: 'text' }      → not supported for image parts
 */
-function toUint8Array$2(data) {
-	if (data instanceof Uint8Array) return data;
-	if (typeof data === "string") {
-		let base64 = data;
-		if (base64.startsWith("data:")) {
-			const commaIndex = base64.indexOf(",");
-			if (commaIndex >= 0) base64 = base64.slice(commaIndex + 1);
+function toUint8Array$2(fileData) {
+	switch (fileData.type) {
+		case "data": {
+			const data = fileData.data;
+			if (data instanceof Uint8Array) return data;
+			let base64 = data;
+			if (base64.startsWith("data:")) {
+				const commaIndex = base64.indexOf(",");
+				if (commaIndex >= 0) base64 = base64.slice(commaIndex + 1);
+			}
+			const binaryString = atob(base64);
+			const bytes = new Uint8Array(binaryString.length);
+			for (let i = 0; i < binaryString.length; i++) bytes[i] = binaryString.charCodeAt(i);
+			return bytes;
 		}
-		const binaryString = atob(base64);
-		const bytes = new Uint8Array(binaryString.length);
-		for (let i = 0; i < binaryString.length; i++) bytes[i] = binaryString.charCodeAt(i);
-		return bytes;
+		case "url": throw new Error("URL image sources are not supported by Workers AI. Provide image data as a Uint8Array or base64 string instead.");
+		case "reference": throw new UnsupportedFunctionalityError({
+			functionality: "file-part-provider-reference",
+			message: "Provider file references are not supported by Workers AI. Provide image data as a Uint8Array or base64 string instead."
+		});
+		case "text": throw new UnsupportedFunctionalityError({
+			functionality: "file-part-inline-text",
+			message: "Inline text file parts are not supported by Workers AI chat. Pass text as a regular text part instead."
+		});
+		default: throw new Error(`Unsupported file data type: ${fileData.type}`);
 	}
-	if (data instanceof URL) throw new Error("URL image sources are not supported by Workers AI. Provide image data as a Uint8Array or base64 string instead.");
-	return null;
 }
 function assertImageMediaType(mediaType) {
 	if (!mediaType) throw new UnsupportedFunctionalityError({
@@ -445,6 +458,8 @@ function convertToWorkersAIChatMessages(prompt) {
 					reasoning += part.text;
 					break;
 				case "file": break;
+				case "reasoning-file": break;
+				case "custom": break;
 				case "tool-call":
 					toolCalls.push({
 						function: {
@@ -511,7 +526,7 @@ function convertToWorkersAIChatMessages(prompt) {
 //#endregion
 //#region src/map-workersai-usage.ts
 /**
-* Map Workers AI usage data to the AI SDK V3 usage format.
+* Map Workers AI usage data to the AI SDK v4 usage format.
 * Accepts any object that may have a `usage` property with token counts.
 *
 * Workers AI mirrors the OpenAI usage shape, including
@@ -595,7 +610,7 @@ function mapWorkersAIFinishReason(finishReasonOrResponse) {
 //#endregion
 //#region src/streaming.ts
 /**
-* Prepend a stream-start event to an existing LanguageModelV3 stream.
+* Prepend a stream-start event to an existing LanguageModelV4 stream.
 * Uses pipeThrough for proper backpressure and error propagation.
 */
 function prependStreamStart(source, warnings) {
@@ -629,7 +644,7 @@ function isNullFinalizationChunk(tc) {
 	return !(tc.id ?? null) && !name && (!args || args === "");
 }
 /**
-* Maps a Workers AI SSE stream into AI SDK V3 LanguageModelV3StreamPart events.
+* Maps a Workers AI SSE stream into AI SDK LanguageModelV4StreamPart events.
 *
 * Uses a TransformStream pipeline for proper backpressure — chunks are emitted
 * one at a time as the downstream consumer pulls, not buffered eagerly.
@@ -949,7 +964,7 @@ function getMappedStream(response, salvageContext) {
 //#region src/aisearch-chat-language-model.ts
 var AISearchChatLanguageModel = class {
 	constructor(modelId, settings, config) {
-		_defineProperty(this, "specificationVersion", "v3");
+		_defineProperty(this, "specificationVersion", "v4");
 		_defineProperty(this, "defaultObjectGenerationMode", "json");
 		_defineProperty(this, "supportedUrls", {});
 		_defineProperty(this, "modelId", void 0);
@@ -1042,7 +1057,7 @@ var WorkersAIEmbeddingModel = class {
 		return this.settings.supportsParallelCalls ?? true;
 	}
 	constructor(modelId, settings, config) {
-		_defineProperty(this, "specificationVersion", "v3");
+		_defineProperty(this, "specificationVersion", "v4");
 		_defineProperty(this, "modelId", void 0);
 		_defineProperty(this, "config", void 0);
 		_defineProperty(this, "settings", void 0);
@@ -1079,9 +1094,26 @@ var WorkersAIEmbeddingModel = class {
 };
 //#endregion
 //#region src/workersai-chat-language-model.ts
+/**
+* Map the unified `reasoning` call option (spec v4) to Workers AI's
+* `reasoning_effort`. Workers AI accepts "low" | "medium" | "high" | null
+* (null disables reasoning), so `minimal` and `xhigh` are clamped to the
+* nearest supported effort. `provider-default` (and absence) returns
+* undefined so the settings-level `reasoning_effort` still applies.
+*/
+function mapUnifiedReasoningEffort(reasoning) {
+	switch (reasoning) {
+		case void 0:
+		case "provider-default": return;
+		case "none": return null;
+		case "minimal": return "low";
+		case "xhigh": return "high";
+		default: return reasoning;
+	}
+}
 var WorkersAIChatLanguageModel = class {
 	constructor(modelId, settings, config) {
-		_defineProperty(this, "specificationVersion", "v3");
+		_defineProperty(this, "specificationVersion", "v4");
 		_defineProperty(this, "defaultObjectGenerationMode", "json");
 		_defineProperty(this, "supportedUrls", {});
 		_defineProperty(this, "modelId", void 0);
@@ -1094,7 +1126,7 @@ var WorkersAIChatLanguageModel = class {
 	get provider() {
 		return this.config.provider;
 	}
-	getArgs({ responseFormat, tools, toolChoice, maxOutputTokens, temperature, topP, frequencyPenalty, presencePenalty, seed }) {
+	getArgs({ responseFormat, tools, toolChoice, maxOutputTokens, temperature, topP, frequencyPenalty, presencePenalty, seed, reasoning }) {
 		const type = responseFormat?.type ?? "text";
 		const warnings = [];
 		if (frequencyPenalty != null) warnings.push({
@@ -1105,6 +1137,11 @@ var WorkersAIChatLanguageModel = class {
 			feature: "presencePenalty",
 			type: "unsupported"
 		});
+		if (reasoning === "minimal" || reasoning === "xhigh") warnings.push({
+			type: "compatibility",
+			feature: "reasoning",
+			details: `Workers AI supports reasoning_effort "low" | "medium" | "high"; "${reasoning}" was mapped to "${reasoning === "minimal" ? "low" : "high"}".`
+		});
 		const baseArgs = {
 			max_tokens: maxOutputTokens,
 			model: this.modelId,
@@ -1157,11 +1194,17 @@ var WorkersAIChatLanguageModel = class {
 	*
 	* `reasoning_effort: null` is a valid value ("disable reasoning"), so we
 	* check `!== undefined` rather than truthiness.
+	*
+	* The unified `reasoning` call option (spec v4) is mapped onto
+	* `reasoning_effort` between the two: an explicit per-call
+	* `providerOptions["workers-ai"].reasoning_effort` wins, then the unified
+	* option, then settings.
 	*/
 	buildRunInputs(args, messages, options) {
 		const rawPerCall = options?.providerOptions?.["workers-ai"];
 		const perCall = rawPerCall !== null && typeof rawPerCall === "object" && !Array.isArray(rawPerCall) ? rawPerCall : {};
-		const reasoningEffort = "reasoning_effort" in perCall ? perCall.reasoning_effort : this.settings.reasoning_effort;
+		const unifiedReasoningEffort = mapUnifiedReasoningEffort(options?.reasoning);
+		const reasoningEffort = "reasoning_effort" in perCall ? perCall.reasoning_effort : unifiedReasoningEffort !== void 0 ? unifiedReasoningEffort : this.settings.reasoning_effort;
 		const chatTemplateKwargs = "chat_template_kwargs" in perCall ? perCall.chat_template_kwargs : this.settings.chat_template_kwargs;
 		return {
 			max_tokens: args.max_tokens,
@@ -1230,7 +1273,10 @@ var WorkersAIChatLanguageModel = class {
 	async doGenerate(options) {
 		const { args, warnings } = this.getArgs(options);
 		const { messages } = convertToWorkersAIChatMessages(options.prompt);
-		const inputs = this.buildRunInputs(args, messages, { providerOptions: options.providerOptions });
+		const inputs = this.buildRunInputs(args, messages, {
+			providerOptions: options.providerOptions,
+			reasoning: options.reasoning
+		});
 		const runOptions = this.getRunOptions();
 		let output;
 		try {
@@ -1269,7 +1315,8 @@ var WorkersAIChatLanguageModel = class {
 		const { messages } = convertToWorkersAIChatMessages(options.prompt);
 		const inputs = this.buildRunInputs(args, messages, {
 			stream: true,
-			providerOptions: options.providerOptions
+			providerOptions: options.providerOptions,
+			reasoning: options.reasoning
 		});
 		const runOptions = this.getRunOptions();
 		let response;
@@ -1352,7 +1399,7 @@ var WorkersAIImageModel = class {
 		this.modelId = modelId;
 		this.settings = settings;
 		this.config = config;
-		_defineProperty(this, "specificationVersion", "v3");
+		_defineProperty(this, "specificationVersion", "v4");
 	}
 	async doGenerate({ prompt, n, size, aspectRatio, seed, abortSignal }) {
 		const { width, height } = getDimensionsFromSizeString(size);
@@ -1448,7 +1495,7 @@ async function toUint8Array$1(output) {
 //#endregion
 //#region src/workersai-transcription-model.ts
 /**
-* Workers AI transcription model implementing the AI SDK's `TranscriptionModelV3` interface.
+* Workers AI transcription model implementing the AI SDK's `TranscriptionModelV4` interface.
 *
 * Supports:
 * - Whisper models (`@cf/openai/whisper`, `whisper-tiny-en`, `whisper-large-v3-turbo`)
@@ -1462,7 +1509,7 @@ var WorkersAITranscriptionModel = class {
 		this.modelId = modelId;
 		this.settings = settings;
 		this.config = config;
-		_defineProperty(this, "specificationVersion", "v3");
+		_defineProperty(this, "specificationVersion", "v4");
 	}
 	async doGenerate(options) {
 		const { audio, mediaType, abortSignal } = options;
@@ -1561,7 +1608,7 @@ function uint8ArrayToBase64(bytes) {
 //#endregion
 //#region src/workersai-speech-model.ts
 /**
-* Workers AI speech (text-to-speech) model implementing the AI SDK's `SpeechModelV3` interface.
+* Workers AI speech (text-to-speech) model implementing the AI SDK's `SpeechModelV4` interface.
 *
 * Currently supports Deepgram Aura-1 (`@cf/deepgram/aura-1`).
 * The model accepts `{ text, voice?, speed? }` and returns raw audio bytes.
@@ -1574,7 +1621,7 @@ var WorkersAISpeechModel = class {
 		this.modelId = modelId;
 		this.settings = settings;
 		this.config = config;
-		_defineProperty(this, "specificationVersion", "v3");
+		_defineProperty(this, "specificationVersion", "v4");
 	}
 	async doGenerate(options) {
 		const { text, voice, speed, abortSignal } = options;
@@ -1662,7 +1709,7 @@ async function toUint8Array(output) {
 //#endregion
 //#region src/workersai-reranking-model.ts
 /**
-* Workers AI reranking model implementing the AI SDK's `RerankingModelV3` interface.
+* Workers AI reranking model implementing the AI SDK's `RerankingModelV4` interface.
 *
 * Supports BGE reranker models (`@cf/baai/bge-reranker-base`, `bge-reranker-v2-m3`).
 *
@@ -1678,7 +1725,7 @@ var WorkersAIRerankingModel = class {
 		this.modelId = modelId;
 		this.settings = settings;
 		this.config = config;
-		_defineProperty(this, "specificationVersion", "v3");
+		_defineProperty(this, "specificationVersion", "v4");
 	}
 	async doRerank(options) {
 		const { documents, query, topN, abortSignal } = options;
@@ -1773,7 +1820,7 @@ function createClientFallbackModel(legs) {
 		throw new WorkersAIFallbackError(attempts);
 	}
 	return {
-		specificationVersion: "v3",
+		specificationVersion: "v4",
 		provider: primary.provider,
 		modelId: primary.modelId,
 		supportedUrls: primary.supportedUrls,
@@ -1825,7 +1872,7 @@ function selectTransport(opts, resumeExplicitlyTrue, runCatalog = true, gatewayA
 	const wantsCaching = opts.cacheTtl !== void 0 || opts.skipCache === true;
 	const gatewayOnly = wantsServerFallback || wantsCaching;
 	const feature = wantsServerFallback ? "fallback.mode:\"server\"" : "caching (cacheTtl/skipCache)";
-	if (runCatalog && !gatewayAvailable && (opts.transport === "gateway" || gatewayOnly)) throw new GatewayDelegateError("config", `${opts.transport === "gateway" ? "transport:\"gateway\"" : feature} is unavailable: this provider is on the unified run catalog but is not a native gateway provider, so it has no gateway path (no caching, server-side fallback, or transport:"gateway"). Use the default run path, or fallback.mode:"client".`);
+	if (runCatalog && !gatewayAvailable && (opts.transport === "gateway" || opts.byok || gatewayOnly)) throw new GatewayDelegateError("config", `${opts.transport === "gateway" ? "transport:\"gateway\"" : opts.byok ? "byok" : feature} is unavailable: this provider is on the unified run catalog but is not a native gateway provider, so it has no gateway path (no caching, server-side fallback, BYOK, or transport:"gateway"). Use the default run path, or fallback.mode:"client".`);
 	if (!runCatalog) {
 		if (opts.transport === "run") throw new GatewayDelegateError("config", "transport:\"run\" is unavailable: this provider is not on the unified-billing run catalog, so it can only be reached through the gateway path (BYOK).");
 		if (resumeExplicitlyTrue) throw new GatewayDelegateError("config", "resume:true is unavailable: this provider is not on the resumable run catalog (cf-aig-run-id requires the unified-billing run path).");
@@ -1835,6 +1882,15 @@ function selectTransport(opts, resumeExplicitlyTrue, runCatalog = true, gatewayA
 			warnings
 		};
 	}
+	if (opts.byok) {
+		if (opts.transport === "run") throw new GatewayDelegateError("config", "transport:\"run\" cannot forward a BYOK key — BYOK is a gateway-path feature. Drop transport:\"run\" (or set transport:\"gateway\").");
+		if (resumeExplicitlyTrue) throw new GatewayDelegateError("config", "byok cannot provide resume — cf-aig-run-id is only on the unified-billing run path.");
+		return {
+			transport: "gateway",
+			resumeEnabled: false,
+			warnings
+		};
+	}
 	if (opts.transport === "run" && gatewayOnly) throw new GatewayDelegateError("config", `transport:"run" cannot satisfy ${feature}: those features are only available on the gateway path. Use the gateway transport, or fallback.mode:"client".`);
 	if (opts.transport === "gateway" && resumeExplicitlyTrue) throw new GatewayDelegateError("config", "transport:\"gateway\" cannot provide resume — cf-aig-run-id is only on the run path.");
 	if (gatewayOnly) {
@@ -2146,7 +2202,7 @@ function makeServerFallbackModel(params) {
 		})[method](options);
 	};
 	return {
-		specificationVersion: "v3",
+		specificationVersion: "v4",
 		provider: refModel.provider,
 		modelId: refModel.modelId,
 		supportedUrls: refModel.supportedUrls,
@@ -2189,8 +2245,11 @@ function createWorkersAI(options) {
 		isBinding
 	});
 	const toGatewayOptions = (gateway) => typeof gateway === "string" ? { id: gateway } : gateway;
-	const createDynamicRouteModel = (modelId, settings = {}) => {
-		if (settings.fallback || settings.transport === "gateway" || settings.resume === true || settings.onProgress || settings.onResumeExpired || settings.byok) throw new Error(`"${modelId}" is an AI Gateway dynamic route. Dynamic routes use AI.run with OpenAI-compatible chat-completions wire format; fallback, gateway transport, resume, BYOK, and resume callbacks must be configured on the dynamic route or gateway instead of per call.`);
+	const createRunPathModel = (modelId, settings = {}, kind) => {
+		if (settings.fallback || settings.transport === "gateway" || settings.resume === true || settings.onProgress || settings.onResumeExpired || settings.byok) {
+			const subject = kind === "dynamic" ? `"${modelId}" is an AI Gateway dynamic route` : `"${modelId}" routes through the bare unified-billing run path because no \`providers\` are configured`;
+			throw new Error(`${subject}. It uses AI.run with OpenAI-compatible chat-completions wire format; fallback, gateway transport, resume, BYOK, and resume callbacks are gateway-delegate features — configure provider plugins to use them (createWorkersAI({ binding: env.AI, providers: [openai] })), and for dynamic routes set caching/fallback on the route or gateway instead of per call.`);
+		}
 		const gateway = { ...toGatewayOptions(settings.gateway) ?? options.gateway ?? { id: DEFAULT_GATEWAY_ID } };
 		if (settings.metadata) gateway.metadata = {
 			...gateway.metadata ?? {},
@@ -2216,7 +2275,7 @@ function createWorkersAI(options) {
 		delete chatSettings.byok;
 		const plugin = options.providers?.find((p) => p.wireFormat === DYNAMIC_ROUTE_WIRE_FORMAT);
 		if (!plugin) {
-			if (options.providers?.length) throw new Error(`"${modelId}" is an AI Gateway dynamic route. Dynamic routes return OpenAI-compatible chat-completions wire format on the AI.run path, so configure the OpenAI provider plugin: import { openai } from "workers-ai-provider/openai"; createWorkersAI({ binding: env.AI, providers: [openai] }).`);
+			if (kind === "dynamic" && options.providers?.length) throw new Error(`"${modelId}" is an AI Gateway dynamic route. Dynamic routes return OpenAI-compatible chat-completions wire format on the AI.run path, so configure the OpenAI provider plugin: import { openai } from "workers-ai-provider/openai"; createWorkersAI({ binding: env.AI, providers: [openai] }).`);
 			return createChatModel(modelId, chatSettings);
 		}
 		const fetchImpl = (async (_input, init) => {
@@ -2263,8 +2322,11 @@ A gateway defaults to "default" but can be set via \`gateway\`. Otherwise use a
 	};
 	const isGatewaySlug = (id) => typeof id === "string" && !id.startsWith("@") && !id.startsWith("dynamic/") && id.includes("/");
 	const buildChat = (modelId, settings) => {
-		if (typeof modelId === "string" && modelId.startsWith("dynamic/")) return createDynamicRouteModel(modelId, settings);
-		if (isGatewaySlug(modelId)) return getDelegate(modelId)(modelId, settings);
+		if (typeof modelId === "string" && modelId.startsWith("dynamic/")) return createRunPathModel(modelId, settings, "dynamic");
+		if (isGatewaySlug(modelId)) {
+			if (!options.providers?.length) return createRunPathModel(modelId, settings, "catalog");
+			return getDelegate(modelId)(modelId, settings);
+		}
 		return createChatModel(modelId, settings);
 	};
 	const createImageModel = (modelId, settings = {}) => new WorkersAIImageModel(modelId, settings, {