ai-gateway-proxy 0.0.2 → 0.2.0

package/dist/index.d.ts

@@ -1,4 +1,4 @@
-import { LanguageModelV3CallOptions, LanguageModelV3Content, LanguageModelV3FinishReason, LanguageModelV3Usage, SharedV3Warning, SharedV3ProviderMetadata, LanguageModelV3StreamPart } from '@ai-sdk/provider';
+import { LanguageModelV4CallOptions, LanguageModelV4Content, LanguageModelV4FinishReason, LanguageModelV4Usage, SharedV4Warning, SharedV4ProviderMetadata, LanguageModelV4StreamPart } from '@ai-sdk/provider';
 
 type GatewayAuthToken = {
     token: string;
@@ -7,6 +7,7 @@ type GatewayAuthToken = {
 /**
  * Get authentication token for the AI Gateway.
  * Tries API key first (AI_GATEWAY_API_KEY env var), then falls back to Vercel OIDC.
+ * Returns `null` when neither is available, after logging why OIDC failed, and the proxy then answers 500.
  */
 declare function getGatewayAuthToken(): Promise<GatewayAuthToken | null>;
 
@@ -15,11 +16,37 @@ declare function getGatewayAuthToken(): Promise<GatewayAuthToken | null>;
  * For streaming requests, this is aggregated from the stream events.
  */
 type GatewayResponse = {
-    content: LanguageModelV3Content[];
-    finishReason: LanguageModelV3FinishReason;
-    usage: LanguageModelV3Usage;
-    warnings: SharedV3Warning[];
-    providerMetadata?: SharedV3ProviderMetadata;
+    content: LanguageModelV4Content[];
+    finishReason: LanguageModelV4FinishReason;
+    /** `undefined` when a stream ended without a `finish` event, so the usage is unknown. */
+    usage: LanguageModelV4Usage | undefined;
+    warnings: SharedV4Warning[];
+    providerMetadata?: SharedV4ProviderMetadata;
+    /**
+     * The provider's response id, model and start time, from a stream's `response-metadata` parts.
+     * `timestamp` is the string sent over the wire, not a `Date`. Neither the proxy nor the client reads it from a result.
+     */
+    response?: GatewayResponseMetadata;
+};
+/**
+ * What a stream's `response-metadata` parts say about the response, the latest value of each field winning.
+ */
+type GatewayResponseMetadata = {
+    id?: string;
+    modelId?: string;
+    timestamp?: string;
+};
+/**
+ * Why a stream had no `finish` event; the cancel reason is in `cancelReason`.
+ */
+type IncompleteStream = {
+    reason: "ended";
+} | {
+    reason: "error";
+    error: unknown;
+} | {
+    reason: "cancelled";
+    cancelReason?: unknown;
 };
 /**
  * Error response from the AI Gateway.
@@ -29,52 +56,204 @@ type GatewayError = {
         message: string;
         type: string;
         param?: unknown;
+        code?: string | number | null;
     };
+    /** The gateway's id for the failed generation, which `@ai-sdk/gateway` appends to the error message. */
+    generationId?: string | null;
+};
+/**
+ * The kind of a call that is not a language model call, named after the `@ai-sdk/gateway` model it comes from.
+ */
+type ModelCallKind = "embedding" | "image" | "video" | "speech" | "transcription" | "reranking" | "decision" | "batch";
+/**
+ * What `beforeRawRequest` receives.
+ */
+type RawRequestContext = {
+    kind: ModelCallKind;
+    /** The upstream path relative to `baseUrl`, such as `embedding-model` or `batch/start`. */
+    path: string;
+    /** The `ai-model-id` header the client sent, which names the model; a batch names it per request too. */
+    modelId: string | undefined;
+    /** The parsed JSON body, unchecked. */
+    request: unknown;
+};
+/**
+ * What `afterRawResponse` receives.
+ */
+type RawResponseContext = RawRequestContext & {
+    /**
+     * The response the proxy would send: the upstream status, `content-type` and allowlisted
+     * headers, with the upstream body on success and a `GatewayError` JSON body on failure.
+     * Its body can be read once, so read `response.clone()` to leave it for the client.
+     */
+    response: Response;
 };
 /**
  * Options for creating a gateway proxy.
  */
 type CreateGatewayProxyOptions = {
-    /** Base URL for the AI Gateway. Defaults to Vercel's gateway. */
+    /**
+     * Base URL for the AI Gateway. A trailing slash is accepted.
+     * Request paths are resolved under it and cannot leave it; one that would returns 400 Invalid path.
+     * @default "https://ai-gateway.vercel.sh/v4/ai"
+     */
     baseUrl?: string;
-    /** Additional headers to send with requests. */
+    /**
+     * Additional headers to send with requests.
+     * The proxy sets `authorization`, `ai-gateway-auth-method` and `ai-gateway-protocol-version` after these, so they win.
+     * On requests with a body it also sets `content-type`, drops entries naming the model call
+     * (`ai-language-model-*`, `ai-model-id`, `ai-*-model-specification-version`) in favour of the
+     * incoming request's, and lets the incoming `ai-o11y-*`, `idempotency-key` and `accept` override these.
+     */
     headers?: Record<string, string>;
+    /**
+     * The name of the catch-all parameter in the route.
+     * @default "segments"
+     * @example
+     * // For Next.js route `app/api/ai/[...path]/route.ts`
+     * createGatewayProxy({ segmentsParam: "path" })
+     */
+    segmentsParam?: string;
+    /**
+     * Custom function to extract the path from the request.
+     * Use this for frameworks without catch-all route support (e.g., Hono, Express).
+     *
+     * When provided, `segmentsParam` is ignored.
+     *
+     * Return a path relative to `baseUrl`, without a leading `/`.
+     * A path that starts with `/`, contains `://`, or has a segment that fails to decode or decodes
+     * to `.` or `..` or contains `/`, `\` or a NUL byte returns 400 Invalid path.
+     *
+     * @example
+     * // Hono with a base path
+     * createGatewayProxy({
+     *   extractPath: (request) => {
+     *     const url = new URL(request.url);
+     *     return url.pathname.replace(/^\/api\/ai\//, "");
+     *   }
+     * })
+     */
+    extractPath?: (request: Request) => string | Promise<string>;
     /**
      * Called before the request is sent to the gateway.
-     * Allows modification of the request body.
+     * Allows modification of the request body. Returning `undefined` or `null` sends the
+     * parsed body unchanged and logs an error. Any other result that is not an object, such as
+     * a string or an array, is not sent: the client gets a 500 and the gateway is not called.
+     *
+     * Not called for GET, HEAD or other requests without a body, which are forwarded as is.
+     * A body that is JSON but not an object gets a 400 before this or any other language model
+     * hook runs.
+     *
+     * Not called when the path relative to `baseUrl` is exactly one of the paths `@ai-sdk/gateway`
+     * uses for other calls: `embedding-model`, `image-model`, `video-model`, `video-model/start`,
+     * `video-model/status`, `speech-model`, `transcription-model`, `reranking-model`,
+     * `decision-model`, `batch/start`, `batch/status`, `batch/results` and `batch/cancel`.
+     * Every other path runs it, whatever headers the request sends.
      */
     beforeRequest?: (context: {
-        request: LanguageModelV3CallOptions;
-    }) => LanguageModelV3CallOptions | Promise<LanguageModelV3CallOptions>;
+        request: LanguageModelV4CallOptions;
+        /** The `ai-language-model-id` header the client sent, which names the model; the body does not. */
+        modelId?: string;
+    }) => LanguageModelV4CallOptions | Promise<LanguageModelV4CallOptions>;
     /**
      * Called after a successful response is received from the gateway.
      * Allows modification of the response body.
      *
+     * For non-streaming responses, the returned object replaces the JSON body sent to the client.
+     * A result with a missing or wrong-typed field is not sent: the client gets a 500 naming the
+     * field, and on a stream the original finish event is forwarded. `null` stands for an absent
+     * field, and `warnings` may be left out. An absent `usage` keeps the gateway's, on both paths.
+     *
      * For streaming responses, this is called when the stream completes with
      * the aggregated response. The content has already been streamed, but you
      * can modify the finish event metadata (usage, finishReason, providerMetadata).
+     * If the hook throws while handling a stream, the error is logged and the
+     * original finish event is forwarded unchanged. It runs once per stream: a
+     * later finish event gets the same rewrite, or none if the hook failed.
+     *
+     * A stream that ends, fails or is cancelled before its finish event calls the hook once
+     * with the content received so far, `incomplete` set, `response.usage` undefined and
+     * `response.finishReason` `error` when the read failed or the stream sent an `error` part,
+     * `other` otherwise. Nothing can be rewritten then, so the
+     * returned value is ignored and a throw is logged.
+     *
+     * Not called for an event stream or JSON Lines body answering a request not marked
+     * `ai-language-model-streaming: true`, such as a video model call or batch results, which is
+     * forwarded as is.
+     *
+     * Not called for a 204 or 205 response, which is forwarded without a body.
+     *
+     * Not called for GET, HEAD or other requests without a body, which are forwarded as is.
+     *
+     * Not called when the path relative to `baseUrl` is exactly one of the paths `@ai-sdk/gateway`
+     * uses for other calls: `embedding-model`, `image-model`, `video-model`, `video-model/start`,
+     * `video-model/status`, `speech-model`, `transcription-model`, `reranking-model`,
+     * `decision-model`, `batch/start`, `batch/status`, `batch/results` and `batch/cancel`.
+     * Every other path runs it, whatever headers the request sends.
      */
     afterResponse?: (context: {
-        request: LanguageModelV3CallOptions;
+        request: LanguageModelV4CallOptions;
+        /** The `ai-language-model-id` header the client sent, which names the model; the body does not. */
+        modelId?: string;
         response: GatewayResponse;
+        /** Set only for a stream that did not deliver its `finish` event. */
+        incomplete?: IncompleteStream;
     }) => GatewayResponse | Promise<GatewayResponse>;
     /**
      * Called when the gateway returns an error response.
      * Allows custom error handling or modification.
+     *
+     * A returned error body is sent as JSON with the upstream status.
+     * A returned `Response` is sent as is.
+     * Returning `undefined`, `null` or a body without a string `error.message` (or with a
+     * non-string `error.type`, an `error.code` that is neither a string nor a number, or a
+     * non-string `generationId`) logs an error and sends the upstream error body instead, and so
+     * does a hook that throws.
+     *
+     * Not called for GET, HEAD or other requests without a body, which are forwarded as is.
+     *
+     * Not called when the path relative to `baseUrl` is exactly one of the paths `@ai-sdk/gateway`
+     * uses for other calls: `embedding-model`, `image-model`, `video-model`, `video-model/start`,
+     * `video-model/status`, `speech-model`, `transcription-model`, `reranking-model`,
+     * `decision-model`, `batch/start`, `batch/status`, `batch/results` and `batch/cancel`.
+     * Every other path runs it, whatever headers the request sends.
      */
     onError?: (context: {
-        request: LanguageModelV3CallOptions;
+        request: LanguageModelV4CallOptions;
+        /** The `ai-language-model-id` header the client sent, which names the model; the body does not. */
+        modelId?: string;
         error: GatewayError;
         status: number;
     }) => GatewayError | Response | Promise<GatewayError | Response>;
+    /**
+     * Called before a call that is not a language model call is sent to the gateway: a request
+     * with a body whose path relative to `baseUrl` is exactly `embedding-model`, `image-model`,
+     * `video-model`, `video-model/start`, `speech-model`, `transcription-model`,
+     * `reranking-model`, `decision-model` or `batch/start`. Job reads (`video-model/status` and
+     * `batch/status`, `batch/results`, `batch/cancel`) run no hook.
+     *
+     * A returned value replaces the body sent upstream; `undefined` or `null` sends the body as is.
+     */
+    beforeRawRequest?: (context: RawRequestContext) => unknown;
+    /**
+     * Called with the response to a call `beforeRawRequest` would run for, on success and on an
+     * upstream error status, but not on a redirect, which is answered with a 502, nor when the
+     * upstream request fails.
+     *
+     * A returned `Response` is sent instead. Returning nothing sends `response`, which must then
+     * be unread; a hook that read it without returning a `Response`, or returned anything else,
+     * is a 500. When the hook returns another `Response`, throws or returns anything else, an
+     * unread `response.body` is cancelled unless the returned `Response` is built on it.
+     */
+    afterRawResponse?: (context: RawResponseContext) => Response | undefined | Promise<Response | undefined>;
 };
 /**
- * Next.js route handler function signature for the gateway proxy.
+ * Route handler function signature for the gateway proxy.
+ * Compatible with Next.js App Router and other frameworks.
+ * Takes a Fetch API `Request`, not a Node or Express request object.
  */
-type CreateGatewayProxyFn = (request: Request, context: {
-    params: Promise<{
-        segments?: string[];
-    }>;
+type CreateGatewayProxyFn = (request: Request, context?: {
+    params: Promise<Record<string, string | string[] | undefined>>;
 }) => Promise<Response>;
 /**
  * Result of createGatewayProxy - a handler function with HTTP method exports.
@@ -90,17 +269,17 @@ type CreateGatewayProxyResult = CreateGatewayProxyFn & {
 /**
  * Creates a proxy handler for the Vercel AI Gateway.
  *
- * Uses AI SDK V3 Language Model.
+ * Uses AI SDK V4 Language Model.
  * @see https://github.com/vercel/ai/blob/main/packages/gateway/src/gateway-language-model.ts
  */
-declare const createGatewayProxy: ({ baseUrl, headers, beforeRequest, afterResponse, onError, }?: CreateGatewayProxyOptions) => CreateGatewayProxyResult;
+declare const createGatewayProxy: ({ baseUrl, headers, segmentsParam, extractPath, beforeRequest, afterResponse, onError, beforeRawRequest, afterRawResponse, }?: CreateGatewayProxyOptions) => CreateGatewayProxyResult;
 
 /**
  * Aggregator for collecting stream content.
  * Tracks text deltas, reasoning, tool calls, and other content to build the final GatewayResponse.
  *
- * Uses AI SDK V3 protocol types.
- * @see https://github.com/vercel/ai/blob/main/packages/provider/src/language-model/v3/language-model-v3-stream-part.ts
+ * Uses AI SDK V4 protocol types.
+ * @see https://github.com/vercel/ai/blob/main/packages/provider/src/language-model/v4/language-model-v4-stream-part.ts
  */
 declare class StreamContentAggregator {
     private content;
@@ -110,15 +289,20 @@ declare class StreamContentAggregator {
     private finishReason;
     private usage;
     private providerMetadata;
+    private response;
     /**
      * Process a stream part and update aggregated state.
-     * Uses LanguageModelV3StreamPart type from @ai-sdk/provider.
+     * Uses LanguageModelV4StreamPart type from @ai-sdk/provider.
      *
-     * @see https://github.com/vercel/ai/blob/main/packages/provider/src/language-model/v3/language-model-v3-stream-part.ts
+     * @see https://github.com/vercel/ai/blob/main/packages/provider/src/language-model/v4/language-model-v4-stream-part.ts
      */
-    process(part: LanguageModelV3StreamPart): void;
+    process(part: LanguageModelV4StreamPart): void;
+    private recordResponseMetadata;
+    private openText;
+    private openReasoning;
     /**
-     * Get the aggregated response.
+     * Get the aggregated response. Before a `finish` part, `usage` is undefined and
+     * `finishReason` is `error` after an `error` part and `other` otherwise.
      */
     getResponse(): GatewayResponse;
 }
@@ -126,16 +310,20 @@ declare class StreamContentAggregator {
  * Creates a transform stream that:
  * 1. Passes through all SSE events immediately
  * 2. Aggregates content using StreamContentAggregator
- * 3. Calls afterResponse when finish event is received
- * 4. Modifies the finish event based on afterResponse result
+ * 3. Calls afterResponse once: on the first finish event, or with partial content if the
+ *    stream ends, fails or is cancelled first
+ * 4. Modifies every finish event based on that afterResponse result
+ *
+ * Pass the client request's `signal` when it is also the upstream fetch's signal: a read that
+ * fails because it aborted is then reported as a cancel rather than an upstream failure.
  *
  * Uses EventSourceParserStream from @ai-sdk/provider-utils for SSE parsing.
  * EventSourceParserStream is re-exported from the eventsource-parser package.
  * @see https://github.com/vercel/ai/blob/main/packages/provider-utils/src/parse-json-event-stream.ts
  * @see https://github.com/rexxars/eventsource-parser
  */
-declare function createStreamTransformer(body: ReadableStream<Uint8Array>, requestBody: LanguageModelV3CallOptions, afterResponse: NonNullable<CreateGatewayProxyOptions["afterResponse"]>): ReadableStream<Uint8Array>;
+declare function createStreamTransformer(body: ReadableStream<Uint8Array>, requestBody: LanguageModelV4CallOptions, afterResponse: NonNullable<CreateGatewayProxyOptions["afterResponse"]>, signal?: AbortSignal): ReadableStream<Uint8Array>;
 
 export { StreamContentAggregator, createGatewayProxy, createStreamTransformer, getGatewayAuthToken };
-export type { CreateGatewayProxyFn, CreateGatewayProxyOptions, CreateGatewayProxyResult, GatewayAuthToken, GatewayError, GatewayResponse };
+export type { CreateGatewayProxyFn, CreateGatewayProxyOptions, CreateGatewayProxyResult, GatewayAuthToken, GatewayError, GatewayResponse, GatewayResponseMetadata, IncompleteStream, ModelCallKind, RawRequestContext, RawResponseContext };
 //# sourceMappingURL=index.d.ts.map

package/dist/index.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"index.d.ts","sources":["../src/auth.ts","../src/types.ts","../src/proxy.ts","../src/stream.ts"],"sourcesContent":["import { getVercelOidcToken } from \"@vercel/oidc\";\n\nexport type GatewayAuthToken = {\n\ttoken: string;\n\tauthMethod: \"api-key\" | \"oidc\";\n};\n\n/**\n * Get authentication token for the AI Gateway.\n * Tries API key first (AI_GATEWAY_API_KEY env var), then falls back to Vercel OIDC.\n */\nexport async function getGatewayAuthToken(): Promise<GatewayAuthToken | null> {\n\tconst apiKey = process.env.AI_GATEWAY_API_KEY;\n\tif (apiKey) {\n\t\treturn { token: apiKey, authMethod: \"api-key\" };\n\t}\n\n\ttry {\n\t\tconst oidcToken = await getVercelOidcToken();\n\t\treturn { token: oidcToken, authMethod: \"oidc\" };\n\t} catch {\n\t\treturn null;\n\t}\n}\n","import type {\n\tLanguageModelV3CallOptions,\n\tLanguageModelV3Content,\n\tLanguageModelV3FinishReason,\n\tLanguageModelV3Usage,\n\tSharedV3ProviderMetadata,\n\tSharedV3Warning,\n} from \"@ai-sdk/provider\";\n\n/**\n * The response body from the AI Gateway for non-streaming requests.\n * For streaming requests, this is aggregated from the stream events.\n */\nexport type GatewayResponse = {\n\tcontent: LanguageModelV3Content[];\n\tfinishReason: LanguageModelV3FinishReason;\n\tusage: LanguageModelV3Usage;\n\twarnings: SharedV3Warning[];\n\tproviderMetadata?: SharedV3ProviderMetadata;\n};\n\n/**\n * Error response from the AI Gateway.\n */\nexport type GatewayError = {\n\terror: {\n\t\tmessage: string;\n\t\ttype: string;\n\t\tparam?: unknown;\n\t};\n};\n\n/**\n * Options for creating a gateway proxy.\n */\nexport type CreateGatewayProxyOptions = {\n\t/** Base URL for the AI Gateway. Defaults to Vercel's gateway. */\n\tbaseUrl?: string;\n\t/** Additional headers to send with requests. */\n\theaders?: Record<string, string>;\n\t/**\n\t * Called before the request is sent to the gateway.\n\t * Allows modification of the request body.\n\t */\n\tbeforeRequest?: (context: {\n\t\trequest: LanguageModelV3CallOptions;\n\t}) => LanguageModelV3CallOptions | Promise<LanguageModelV3CallOptions>;\n\t/**\n\t * Called after a successful response is received from the gateway.\n\t * Allows modification of the response body.\n\t *\n\t * For streaming responses, this is called when the stream completes with\n\t * the aggregated response. The content has already been streamed, but you\n\t * can modify the finish event metadata (usage, finishReason, providerMetadata).\n\t */\n\tafterResponse?: (context: {\n\t\trequest: LanguageModelV3CallOptions;\n\t\tresponse: GatewayResponse;\n\t}) => GatewayResponse | Promise<GatewayResponse>;\n\t/**\n\t * Called when the gateway returns an error response.\n\t * Allows custom error handling or modification.\n\t */\n\tonError?: (context: {\n\t\trequest: LanguageModelV3CallOptions;\n\t\terror: GatewayError;\n\t\tstatus: number;\n\t}) => GatewayError | Response | Promise<GatewayError | Response>;\n};\n\n/**\n * Next.js route handler function signature for the gateway proxy.\n */\nexport type CreateGatewayProxyFn = (\n\trequest: Request,\n\tcontext: { params: Promise<{ segments?: string[] }> },\n) => Promise<Response>;\n\n/**\n * Result of createGatewayProxy - a handler function with HTTP method exports.\n */\nexport type CreateGatewayProxyResult = CreateGatewayProxyFn & {\n\tGET: CreateGatewayProxyFn;\n\tPOST: CreateGatewayProxyFn;\n\tPUT: CreateGatewayProxyFn;\n\tDELETE: CreateGatewayProxyFn;\n\tPATCH: CreateGatewayProxyFn;\n};\n","import type { LanguageModelV3CallOptions } from \"@ai-sdk/provider\";\nimport { getGatewayAuthToken } from \"./auth\";\nimport { createStreamTransformer } from \"./stream\";\nimport type {\n\tCreateGatewayProxyOptions,\n\tCreateGatewayProxyResult,\n\tGatewayError,\n\tGatewayResponse,\n} from \"./types\";\n\n/**\n * Creates a proxy handler for the Vercel AI Gateway.\n *\n * Uses AI SDK V3 Language Model.\n * @see https://github.com/vercel/ai/blob/main/packages/gateway/src/gateway-language-model.ts\n */\nexport const createGatewayProxy = ({\n\tbaseUrl = \"https://ai-gateway.vercel.sh/v1/ai\",\n\theaders,\n\tbeforeRequest,\n\tafterResponse,\n\tonError,\n}: CreateGatewayProxyOptions = {}): CreateGatewayProxyResult => {\n\tconst handler = async (\n\t\trequest: Request,\n\t\t{ params }: { params: Promise<{ segments?: string[] }> },\n\t) => {\n\t\tconst auth = await getGatewayAuthToken();\n\t\tif (!auth) {\n\t\t\treturn new Response(\"AI Gateway authentication not configured\", {\n\t\t\t\tstatus: 500,\n\t\t\t});\n\t\t}\n\n\t\tconst { segments } = await params;\n\t\tconst path = segments?.join(\"/\") ?? \"\";\n\t\tconst searchParams = new URL(request.url).searchParams;\n\t\tconst upstream = `${baseUrl}/${path}${searchParams.toString() ? `?${searchParams.toString()}` : \"\"}`;\n\n\t\tlet requestBody: LanguageModelV3CallOptions;\n\t\ttry {\n\t\t\trequestBody = await request.clone().json();\n\t\t} catch {\n\t\t\treturn new Response(\"Invalid JSON body\", { status: 400 });\n\t\t}\n\n\t\t// Apply beforeRequest hook\n\t\tif (beforeRequest) {\n\t\t\trequestBody = await beforeRequest({ request: requestBody });\n\t\t}\n\n\t\tconst isStreaming = request.headers.get(\"ai-language-model-streaming\") === \"true\";\n\n\t\ttry {\n\t\t\tconst res = await fetch(upstream, {\n\t\t\t\tmethod: request.method,\n\t\t\t\tbody: JSON.stringify(requestBody),\n\t\t\t\tduplex: \"half\",\n\t\t\t\theaders: {\n\t\t\t\t\t\"content-type\": \"application/json\",\n\t\t\t\t\t...headers,\n\t\t\t\t\tauthorization: `Bearer ${auth.token}`,\n\t\t\t\t\t\"ai-gateway-auth-method\": auth.authMethod,\n\t\t\t\t\t\"ai-gateway-protocol-version\": \"0.0.1\",\n\t\t\t\t\t\"ai-language-model-id\": request.headers.get(\"ai-language-model-id\") ?? \"\",\n\t\t\t\t\t\"ai-language-model-streaming\": isStreaming ? \"true\" : \"false\",\n\t\t\t\t},\n\t\t\t} as RequestInit);\n\n\t\t\t// Handle error responses\n\t\t\tif (!res.ok) {\n\t\t\t\tlet errorBody: GatewayError;\n\t\t\t\ttry {\n\t\t\t\t\terrorBody = await res.json();\n\t\t\t\t} catch {\n\t\t\t\t\terrorBody = {\n\t\t\t\t\t\terror: {\n\t\t\t\t\t\t\tmessage: res.statusText || \"Unknown error\",\n\t\t\t\t\t\t\ttype: \"gateway_error\",\n\t\t\t\t\t\t},\n\t\t\t\t\t};\n\t\t\t\t}\n\n\t\t\t\tif (onError) {\n\t\t\t\t\tconst result = await onError({\n\t\t\t\t\t\trequest: requestBody,\n\t\t\t\t\t\terror: errorBody,\n\t\t\t\t\t\tstatus: res.status,\n\t\t\t\t\t});\n\n\t\t\t\t\t// If onError returns a Response, use it directly\n\t\t\t\t\tif (result instanceof Response) {\n\t\t\t\t\t\treturn result;\n\t\t\t\t\t}\n\n\t\t\t\t\t// Otherwise, return the modified error\n\t\t\t\t\treturn new Response(JSON.stringify(result), {\n\t\t\t\t\t\tstatus: res.status,\n\t\t\t\t\t\tstatusText: res.statusText,\n\t\t\t\t\t\theaders: { \"content-type\": \"application/json\" },\n\t\t\t\t\t});\n\t\t\t\t}\n\n\t\t\t\treturn new Response(JSON.stringify(errorBody), {\n\t\t\t\t\tstatus: res.status,\n\t\t\t\t\tstatusText: res.statusText,\n\t\t\t\t\theaders: { \"content-type\": \"application/json\" },\n\t\t\t\t});\n\t\t\t}\n\n\t\t\t// For streaming responses\n\t\t\tif (isStreaming) {\n\t\t\t\tif (!res.body) {\n\t\t\t\t\treturn new Response(\"No response body\", { status: 500 });\n\t\t\t\t}\n\n\t\t\t\t// If no afterResponse hook, pass through directly\n\t\t\t\tif (!afterResponse) {\n\t\t\t\t\treturn new Response(res.body, {\n\t\t\t\t\t\tstatus: res.status,\n\t\t\t\t\t\tstatusText: res.statusText,\n\t\t\t\t\t\theaders: {\n\t\t\t\t\t\t\t\"content-type\": res.headers.get(\"content-type\") || \"text/event-stream\",\n\t\t\t\t\t\t},\n\t\t\t\t\t});\n\t\t\t\t}\n\n\t\t\t\t// Transform the stream to call afterResponse on finish\n\t\t\t\tconst transformedStream = createStreamTransformer(res.body, requestBody, afterResponse);\n\n\t\t\t\treturn new Response(transformedStream, {\n\t\t\t\t\tstatus: res.status,\n\t\t\t\t\tstatusText: res.statusText,\n\t\t\t\t\theaders: {\n\t\t\t\t\t\t\"content-type\": res.headers.get(\"content-type\") || \"text/event-stream\",\n\t\t\t\t\t},\n\t\t\t\t});\n\t\t\t}\n\n\t\t\t// For non-streaming responses, parse and optionally transform\n\t\t\tlet responseBody: GatewayResponse = await res.json();\n\n\t\t\tif (afterResponse) {\n\t\t\t\tresponseBody = await afterResponse({\n\t\t\t\t\trequest: requestBody,\n\t\t\t\t\tresponse: responseBody,\n\t\t\t\t});\n\t\t\t}\n\n\t\t\treturn new Response(JSON.stringify(responseBody), {\n\t\t\t\tstatus: res.status,\n\t\t\t\tstatusText: res.statusText,\n\t\t\t\theaders: { \"content-type\": \"application/json\" },\n\t\t\t});\n\t\t} catch {\n\t\t\treturn new Response(\"Error proxying request to AI Gateway\", {\n\t\t\t\tstatus: 500,\n\t\t\t});\n\t\t}\n\t};\n\n\thandler.GET = handler;\n\thandler.POST = handler;\n\thandler.PUT = handler;\n\thandler.DELETE = handler;\n\thandler.PATCH = handler;\n\n\treturn handler;\n};\n","import type {\n\tLanguageModelV3CallOptions,\n\tLanguageModelV3Content,\n\tLanguageModelV3FinishReason,\n\tLanguageModelV3StreamPart,\n\tLanguageModelV3Usage,\n\tSharedV3ProviderMetadata,\n\tSharedV3Warning,\n} from \"@ai-sdk/provider\";\nimport { EventSourceParserStream } from \"@ai-sdk/provider-utils\";\nimport type { CreateGatewayProxyOptions, GatewayResponse } from \"./types\";\n\n/**\n * Aggregator for collecting stream content.\n * Tracks text deltas, reasoning, tool calls, and other content to build the final GatewayResponse.\n *\n * Uses AI SDK V3 protocol types.\n * @see https://github.com/vercel/ai/blob/main/packages/provider/src/language-model/v3/language-model-v3-stream-part.ts\n */\nexport class StreamContentAggregator {\n\tprivate content: LanguageModelV3Content[] = [];\n\tprivate warnings: SharedV3Warning[] = [];\n\n\t// Text content being built (keyed by id)\n\tprivate textContent = new Map<string, { type: \"text\"; text: string }>();\n\n\t// Reasoning content being built (keyed by id)\n\tprivate reasoningContent = new Map<string, { type: \"reasoning\"; text: string }>();\n\n\tprivate finishReason: LanguageModelV3FinishReason = {\n\t\tunified: \"stop\",\n\t\traw: undefined,\n\t};\n\n\tprivate usage: LanguageModelV3Usage = {\n\t\tinputTokens: {\n\t\t\ttotal: undefined,\n\t\t\tnoCache: undefined,\n\t\t\tcacheRead: undefined,\n\t\t\tcacheWrite: undefined,\n\t\t},\n\t\toutputTokens: {\n\t\t\ttotal: undefined,\n\t\t\ttext: undefined,\n\t\t\treasoning: undefined,\n\t\t},\n\t};\n\n\tprivate providerMetadata: SharedV3ProviderMetadata | undefined;\n\n\t/**\n\t * Process a stream part and update aggregated state.\n\t * Uses LanguageModelV3StreamPart type from @ai-sdk/provider.\n\t *\n\t * @see https://github.com/vercel/ai/blob/main/packages/provider/src/language-model/v3/language-model-v3-stream-part.ts\n\t */\n\tprocess(part: LanguageModelV3StreamPart): void {\n\t\tswitch (part.type) {\n\t\t\tcase \"stream-start\":\n\t\t\t\tthis.warnings.push(...part.warnings);\n\t\t\t\tbreak;\n\n\t\t\tcase \"text-start\":\n\t\t\t\tthis.textContent.set(part.id, { type: \"text\", text: \"\" });\n\t\t\t\tbreak;\n\n\t\t\tcase \"text-delta\":\n\t\t\t\t{\n\t\t\t\t\tconst text = this.textContent.get(part.id);\n\t\t\t\t\tif (text) {\n\t\t\t\t\t\ttext.text += part.delta;\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t\tbreak;\n\n\t\t\tcase \"text-end\":\n\t\t\t\t{\n\t\t\t\t\tconst text = this.textContent.get(part.id);\n\t\t\t\t\tif (text) {\n\t\t\t\t\t\tthis.content.push(text);\n\t\t\t\t\t\tthis.textContent.delete(part.id);\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t\tbreak;\n\n\t\t\tcase \"reasoning-start\":\n\t\t\t\tthis.reasoningContent.set(part.id, { type: \"reasoning\", text: \"\" });\n\t\t\t\tbreak;\n\n\t\t\tcase \"reasoning-delta\":\n\t\t\t\t{\n\t\t\t\t\tconst reasoning = this.reasoningContent.get(part.id);\n\t\t\t\t\tif (reasoning) {\n\t\t\t\t\t\treasoning.text += part.delta;\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t\tbreak;\n\n\t\t\tcase \"reasoning-end\":\n\t\t\t\t{\n\t\t\t\t\tconst reasoning = this.reasoningContent.get(part.id);\n\t\t\t\t\tif (reasoning) {\n\t\t\t\t\t\tthis.content.push(reasoning);\n\t\t\t\t\t\tthis.reasoningContent.delete(part.id);\n\t\t\t\t\t}\n\t\t\t\t}\n\t\t\t\tbreak;\n\n\t\t\tcase \"tool-input-start\":\n\t\t\tcase \"tool-input-delta\":\n\t\t\tcase \"tool-input-end\":\n\t\t\t\t// Tool input is streamed incrementally but we wait for the complete tool-call event\n\t\t\t\tbreak;\n\n\t\t\tcase \"tool-call\":\n\t\t\t\tthis.content.push(part);\n\t\t\t\tbreak;\n\n\t\t\tcase \"tool-result\":\n\t\t\t\tthis.content.push(part);\n\t\t\t\tbreak;\n\n\t\t\tcase \"tool-approval-request\":\n\t\t\t\tthis.content.push(part);\n\t\t\t\tbreak;\n\n\t\t\tcase \"file\":\n\t\t\t\tthis.content.push(part);\n\t\t\t\tbreak;\n\n\t\t\tcase \"source\":\n\t\t\t\tthis.content.push(part);\n\t\t\t\tbreak;\n\n\t\t\tcase \"finish\":\n\t\t\t\t// Finalize any remaining content\n\t\t\t\tfor (const [, text] of this.textContent) {\n\t\t\t\t\tthis.content.push(text);\n\t\t\t\t}\n\t\t\t\tthis.textContent.clear();\n\n\t\t\t\tfor (const [, reasoning] of this.reasoningContent) {\n\t\t\t\t\tthis.content.push(reasoning);\n\t\t\t\t}\n\t\t\t\tthis.reasoningContent.clear();\n\n\t\t\t\tthis.finishReason = part.finishReason;\n\t\t\t\tthis.usage = part.usage;\n\t\t\t\tthis.providerMetadata = part.providerMetadata;\n\t\t\t\tbreak;\n\t\t}\n\t}\n\n\t/**\n\t * Get the aggregated response.\n\t */\n\tgetResponse(): GatewayResponse {\n\t\treturn {\n\t\t\tcontent: this.content,\n\t\t\tfinishReason: this.finishReason,\n\t\t\tusage: this.usage,\n\t\t\twarnings: this.warnings,\n\t\t\tproviderMetadata: this.providerMetadata,\n\t\t};\n\t}\n}\n\n/**\n * Creates a transform stream that:\n * 1. Passes through all SSE events immediately\n * 2. Aggregates content using StreamContentAggregator\n * 3. Calls afterResponse when finish event is received\n * 4. Modifies the finish event based on afterResponse result\n *\n * Uses EventSourceParserStream from @ai-sdk/provider-utils for SSE parsing.\n * EventSourceParserStream is re-exported from the eventsource-parser package.\n * @see https://github.com/vercel/ai/blob/main/packages/provider-utils/src/parse-json-event-stream.ts\n * @see https://github.com/rexxars/eventsource-parser\n */\nexport function createStreamTransformer(\n\tbody: ReadableStream<Uint8Array>,\n\trequestBody: LanguageModelV3CallOptions,\n\tafterResponse: NonNullable<CreateGatewayProxyOptions[\"afterResponse\"]>,\n): ReadableStream<Uint8Array> {\n\tconst encoder = new TextEncoder();\n\tconst aggregator = new StreamContentAggregator();\n\n\t// Use EventSourceParserStream for SSE parsing (re-exported from eventsource-parser via @ai-sdk/provider-utils)\n\tconst eventStream = body\n\t\t.pipeThrough(new TextDecoderStream() as TransformStream<Uint8Array, string>)\n\t\t.pipeThrough(new EventSourceParserStream());\n\n\treturn new ReadableStream({\n\t\tasync start(controller) {\n\t\t\tconst reader = eventStream.getReader();\n\n\t\t\ttry {\n\t\t\t\twhile (true) {\n\t\t\t\t\tconst { done, value } = await reader.read();\n\t\t\t\t\tif (done) break;\n\n\t\t\t\t\tconst { data } = value;\n\n\t\t\t\t\t// Handle DONE marker (end of stream)\n\t\t\t\t\tif (data === \"DONE\" || data === \"[DONE]\") {\n\t\t\t\t\t\tcontroller.enqueue(encoder.encode(`data: ${data}\\n\\n`));\n\t\t\t\t\t\tcontinue;\n\t\t\t\t\t}\n\n\t\t\t\t\t// Parse the JSON event data\n\t\t\t\t\tlet parsed: LanguageModelV3StreamPart;\n\t\t\t\t\ttry {\n\t\t\t\t\t\tparsed = JSON.parse(data);\n\t\t\t\t\t} catch {\n\t\t\t\t\t\t// Not valid JSON, pass through as-is\n\t\t\t\t\t\tcontroller.enqueue(encoder.encode(`data: ${data}\\n\\n`));\n\t\t\t\t\t\tcontinue;\n\t\t\t\t\t}\n\n\t\t\t\t\t// Track content for aggregation\n\t\t\t\t\taggregator.process(parsed);\n\n\t\t\t\t\t// Handle finish event specially - call afterResponse hook\n\t\t\t\t\tif (parsed.type === \"finish\") {\n\t\t\t\t\t\tconst modifiedResponse = await afterResponse({\n\t\t\t\t\t\t\trequest: requestBody,\n\t\t\t\t\t\t\tresponse: aggregator.getResponse(),\n\t\t\t\t\t\t});\n\n\t\t\t\t\t\t// Emit modified finish event\n\t\t\t\t\t\tconst modifiedFinish: LanguageModelV3StreamPart = {\n\t\t\t\t\t\t\ttype: \"finish\",\n\t\t\t\t\t\t\tfinishReason: modifiedResponse.finishReason,\n\t\t\t\t\t\t\tusage: modifiedResponse.usage,\n\t\t\t\t\t\t\tproviderMetadata: modifiedResponse.providerMetadata,\n\t\t\t\t\t\t};\n\n\t\t\t\t\t\tcontroller.enqueue(encoder.encode(`data: ${JSON.stringify(modifiedFinish)}\\n\\n`));\n\t\t\t\t\t\tcontinue;\n\t\t\t\t\t}\n\n\t\t\t\t\t// Pass through event unchanged\n\t\t\t\t\tcontroller.enqueue(encoder.encode(`data: ${data}\\n\\n`));\n\t\t\t\t}\n\n\t\t\t\tcontroller.close();\n\t\t\t} catch (error) {\n\t\t\t\tcontroller.error(error);\n\t\t\t} finally {\n\t\t\t\treader.releaseLock();\n\t\t\t}\n\t\t},\n\t});\n}\n"],"names":[],"mappings":";;AAAO,KAAA,gBAAA;AACP;AACA;AACA;AACA;AACA;AACA;AACA;AACO,iBAAA,mBAAA,IAAA,OAAA,CAAA,gBAAA;;ACPP;AACA;AACA;AACA;AACO,KAAA,eAAA;AACP,aAAA,sBAAA;AACA,kBAAA,2BAAA;AACA,WAAA,oBAAA;AACA,cAAA,eAAA;AACA,uBAAA,wBAAA;AACA;AACA;AACA;AACA;AACO,KAAA,YAAA;AACP;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACO,KAAA,yBAAA;AACP;AACA;AACA;AACA,cAAA,MAAA;AACA;AACA;AACA;AACA;AACA;AACA,iBAAA,0BAAA;AACA,UAAA,0BAAA,GAAA,OAAA,CAAA,0BAAA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,iBAAA,0BAAA;AACA,kBAAA,eAAA;AACA,UAAA,eAAA,GAAA,OAAA,CAAA,eAAA;AACA;AACA;AACA;AACA;AACA;AACA,iBAAA,0BAAA;AACA,eAAA,YAAA;AACA;AACA,UAAA,YAAA,GAAA,QAAA,GAAA,OAAA,CAAA,YAAA,GAAA,QAAA;AACA;AACA;AACA;AACA;AACO,KAAA,oBAAA,aAAA,OAAA;AACP,YAAA,OAAA;AACA;AACA;AACA,MAAA,OAAA,CAAA,QAAA;AACA;AACA;AACA;AACO,KAAA,wBAAA,GAAA,oBAAA;AACP,SAAA,oBAAA;AACA,UAAA,oBAAA;AACA,SAAA,oBAAA;AACA,YAAA,oBAAA;AACA,WAAA,oBAAA;AACA;;AC3EA;AACA;AACA;AACA;AACA;AACA;AACO,cAAA,kBAAA,kEAAA,yBAAA,KAAA,wBAAA;;ACLP;AACA;AACA;AACA;AACA;AACA;AACA;AACO,cAAA,uBAAA;AACP;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,kBAAA,yBAAA;AACA;AACA;AACA;AACA,mBAAA,eAAA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACO,iBAAA,uBAAA,OAAA,cAAA,CAAA,UAAA,gBAAA,0BAAA,iBAAA,WAAA,CAAA,yBAAA,qBAAA,cAAA,CAAA,UAAA;;;;"}
+{"version":3,"file":"index.d.ts","sources":["../src/auth.ts","../src/types.ts","../src/proxy.ts","../src/stream.ts"],"mappings":";;AAAA,KAAK,gBAAgB;AACrB;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,iBAAiB,mBAAmB,IAAI,OAAO,CAAC,gBAAgB;;ACRhE;AACA;AACA;AACA;AACA,KAAK,eAAe;AACpB,aAAa,sBAAsB;AACnC,kBAAkB,2BAA2B;AAC7C;AACA,WAAW,oBAAoB;AAC/B,cAAc,eAAe;AAC7B,uBAAuB,wBAAwB;AAC/C;AACA;AACA;AACA;AACA,eAAe,uBAAuB;AACtC;AACA;AACA;AACA;AACA,KAAK,uBAAuB;AAC5B;AACA;AACA;AACA;AACA;AACA;AACA;AACA,KAAK,gBAAgB;AACrB;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,KAAK,YAAY;AACjB;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,KAAK,aAAa;AAClB;AACA;AACA;AACA,KAAK,iBAAiB;AACtB,UAAU,aAAa;AACvB;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,KAAK,kBAAkB,GAAG,iBAAiB;AAC3C;AACA;AACA;AACA;AACA;AACA,cAAc,QAAQ;AACtB;AACA;AACA;AACA;AACA,KAAK,yBAAyB;AAC9B;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,cAAc,MAAM;AACpB;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,4BAA4B,OAAO,cAAc,OAAO;AACxD;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,iBAAiB,0BAA0B;AAC3C;AACA;AACA,UAAU,0BAA0B,GAAG,OAAO,CAAC,0BAA0B;AACzE;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,iBAAiB,0BAA0B;AAC3C;AACA;AACA,kBAAkB,eAAe;AACjC;AACA,qBAAqB,gBAAgB;AACrC,UAAU,eAAe,GAAG,OAAO,CAAC,eAAe;AACnD;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,iBAAiB,0BAA0B;AAC3C;AACA;AACA,eAAe,YAAY;AAC3B;AACA,UAAU,YAAY,GAAG,QAAQ,GAAG,OAAO,CAAC,YAAY,GAAG,QAAQ;AACnE;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,iCAAiC,iBAAiB;AAClD;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,iCAAiC,kBAAkB,KAAK,QAAQ,eAAe,OAAO,CAAC,QAAQ;AAC/F;AACA;AACA;AACA;AACA;AACA;AACA,KAAK,oBAAoB,aAAa,OAAO;AAC7C,YAAY,OAAO,CAAC,MAAM;AAC1B,MAAM,OAAO,CAAC,QAAQ;AACtB;AACA;AACA;AACA,KAAK,wBAAwB,GAAG,oBAAoB;AACpD,SAAS,oBAAoB;AAC7B,UAAU,oBAAoB;AAC9B,SAAS,oBAAoB;AAC7B,YAAY,oBAAoB;AAChC,WAAW,oBAAoB;AAC/B;;AC7PA;AACA;AACA;AACA;AACA;AACA;AACA,cAAc,kBAAkB,kIAAkI,yBAAyB,KAAK,wBAAwB;;ACLxN;AACA;AACA;AACA;AACA;AACA;AACA;AACA,cAAc,uBAAuB;AACrC;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,kBAAkB,yBAAyB;AAC3C;AACA;AACA;AACA;AACA;AACA;AACA;AACA,mBAAmB,eAAe;AAClC;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA;AACA,iBAAiB,uBAAuB,OAAO,cAAc,CAAC,UAAU,gBAAgB,0BAA0B,iBAAiB,WAAW,CAAC,yBAAyB,6BAA6B,WAAW,GAAG,cAAc,CAAC,UAAU;;","names":[]}