@hebo-ai/gateway 0.10.4 → 0.10.6

package/README.md

@@ -218,7 +218,7 @@ Out-of-the-box model presets:
   Nova: `nova` (`v1`, `v2`, `v1.x`, `v2.x`, `latest`, `embeddings`, `all`)
 
 - **Anthropic** — `@hebo-ai/gateway/models/anthropic`  
-  Claude: `claude` (`v4.6`, `v4.5`, `v4.1`, `v4`, `v3.7`, `v3.5`, `v3`, `v4.x`, `v3.x`, `haiku`, `sonnet`, `opus`, `latest`, `all`)
+  Claude: `claude` (`v4.7`, `v4.6`, `v4.5`, `v4.1`, `v4`, `v3.7`, `v3.5`, `v3`, `v4.x`, `v3.x`, `haiku`, `sonnet`, `opus`, `latest`, `all`)
 
 - **Cohere** — `@hebo-ai/gateway/models/cohere`  
   Command: `command` (`A`, `R`, `latest`, `all`)
@@ -730,7 +730,7 @@ Normalization rules:
 
 - `enabled` -> fall-back to model default if none provided
 - `max_tokens`: fall-back to model default if model supports
-- `effort` supports: `none`, `minimal`, `low`, `medium`, `high`, `xhigh`
+- `effort` supports: `none`, `minimal`, `low`, `medium`, `high`, `xhigh`, `max`
 - Generic `effort` -> budget = percentage of `max_tokens`
   - `none`: 0%
   - `minimal`: 10%
@@ -738,6 +738,7 @@ Normalization rules:
   - `medium`: 50% (default)
   - `high`: 80%
   - `xhigh`: 95%
+  - `max`: 100%
 
 Reasoning output is surfaced as extension to the `completion` object.
 

package/dist/endpoints/chat-completions/handler.js

@@ -133,6 +133,8 @@ export const chatCompletions = (config) => {
         });
         logger.trace({ requestId: ctx.requestId, result }, "[chat] AI SDK result");
         addSpanEvent("hebo.ai-sdk.completed");
+        if (result.response.headers)
+            ctx.response = { headers: result.response.headers };
         recordTimeToFirstToken(performance.now() - start, genAiGeneralAttrs, ctx.trace);
         // Transform result.
         ctx.result = toChatCompletions(result, ctx.resolvedModelId);

package/dist/endpoints/chat-completions/schema.d.ts

@@ -737,6 +737,7 @@ declare const ChatCompletionsInputsSchema: z.ZodObject<{
         medium: "medium";
         high: "high";
         xhigh: "xhigh";
+        max: "max";
     }>>;
     service_tier: z.ZodOptional<z.ZodEnum<{
         default: "default";
@@ -768,6 +769,7 @@ declare const ChatCompletionsInputsSchema: z.ZodObject<{
             medium: "medium";
             high: "high";
             xhigh: "xhigh";
+            max: "max";
         }>>;
         max_tokens: z.ZodOptional<z.ZodNumber>;
         exclude: z.ZodOptional<z.ZodBoolean>;
@@ -997,6 +999,7 @@ export declare const ChatCompletionsBodySchema: z.ZodObject<{
         medium: "medium";
         high: "high";
         xhigh: "xhigh";
+        max: "max";
     }>>;
     service_tier: z.ZodOptional<z.ZodEnum<{
         default: "default";
@@ -1028,6 +1031,7 @@ export declare const ChatCompletionsBodySchema: z.ZodObject<{
             medium: "medium";
             high: "high";
             xhigh: "xhigh";
+            max: "max";
         }>>;
         max_tokens: z.ZodOptional<z.ZodNumber>;
         exclude: z.ZodOptional<z.ZodBoolean>;
@@ -1051,8 +1055,8 @@ export declare const ChatCompletionsBodySchema: z.ZodObject<{
 export type ChatCompletionsBody = z.infer<typeof ChatCompletionsBodySchema>;
 export declare const ChatCompletionsFinishReasonSchema: z.ZodEnum<{
     length: "length";
-    stop: "stop";
     tool_calls: "tool_calls";
+    stop: "stop";
     content_filter: "content_filter";
 }>;
 export type ChatCompletionsFinishReason = z.infer<typeof ChatCompletionsFinishReasonSchema>;
@@ -1105,8 +1109,8 @@ export declare const ChatCompletionsChoiceSchema: z.ZodObject<{
     }, z.core.$strip>;
     finish_reason: z.ZodEnum<{
         length: "length";
-        stop: "stop";
         tool_calls: "tool_calls";
+        stop: "stop";
         content_filter: "content_filter";
     }>;
     logprobs: z.ZodOptional<z.ZodUnknown>;
@@ -1179,8 +1183,8 @@ export declare const ChatCompletionsSchema: z.ZodObject<{
         }, z.core.$strip>;
         finish_reason: z.ZodEnum<{
             length: "length";
-            stop: "stop";
             tool_calls: "tool_calls";
+            stop: "stop";
             content_filter: "content_filter";
         }>;
         logprobs: z.ZodOptional<z.ZodUnknown>;
@@ -1315,8 +1319,8 @@ export declare const ChatCompletionsChoiceDeltaSchema: z.ZodObject<{
     }, z.core.$strip>;
     finish_reason: z.ZodNullable<z.ZodEnum<{
         length: "length";
-        stop: "stop";
         tool_calls: "tool_calls";
+        stop: "stop";
         content_filter: "content_filter";
     }>>;
     logprobs: z.ZodOptional<z.ZodUnknown>;
@@ -1377,8 +1381,8 @@ export declare const ChatCompletionsChunkSchema: z.ZodObject<{
         }, z.core.$strip>;
         finish_reason: z.ZodNullable<z.ZodEnum<{
             length: "length";
-            stop: "stop";
             tool_calls: "tool_calls";
+            stop: "stop";
             content_filter: "content_filter";
         }>>;
         logprobs: z.ZodOptional<z.ZodUnknown>;

package/dist/endpoints/embeddings/handler.js

@@ -81,6 +81,8 @@ export const embeddings = (config) => {
         });
         logger.trace({ requestId: ctx.requestId, result }, "[embeddings] AI SDK result");
         addSpanEvent("hebo.ai-sdk.completed");
+        if (result.responses?.[0]?.headers)
+            ctx.response = { headers: result.responses[0].headers };
         // Transform result.
         ctx.result = toEmbeddings(result, ctx.modelId);
         logger.trace({ requestId: ctx.requestId, result: ctx.result }, "[chat] Embeddings");

package/dist/endpoints/messages/converters.js

@@ -482,18 +482,6 @@ export class MessagesTransformStream extends TransformStream {
                         break;
                     }
                     case "reasoning-delta": {
-                        controller.enqueue({
-                            event: "content_block_delta",
-                            data: {
-                                type: "content_block_delta",
-                                index: blockIndex,
-                                delta: { type: "thinking_delta", thinking: part.text },
-                            },
-                        });
-                        break;
-                    }
-                    case "reasoning-end": {
-                        // Emit signature delta if available from provider metadata
                         const { signature } = extractReasoningMetadata(part.providerMetadata);
                         if (signature) {
                             controller.enqueue({
@@ -505,6 +493,19 @@ export class MessagesTransformStream extends TransformStream {
                                 },
                             });
                         }
+                        else {
+                            controller.enqueue({
+                                event: "content_block_delta",
+                                data: {
+                                    type: "content_block_delta",
+                                    index: blockIndex,
+                                    delta: { type: "thinking_delta", thinking: part.text },
+                                },
+                            });
+                        }
+                        break;
+                    }
+                    case "reasoning-end": {
                         controller.enqueue({
                             event: "content_block_stop",
                             data: { type: "content_block_stop", index: blockIndex },

package/dist/endpoints/messages/handler.js

@@ -124,6 +124,8 @@ export const messages = (config) => {
         });
         logger.trace({ requestId: ctx.requestId, result }, "[messages] AI SDK result");
         addSpanEvent("hebo.ai-sdk.completed");
+        if (result.response.headers)
+            ctx.response = { headers: result.response.headers };
         recordTimeToFirstToken(performance.now() - start, genAiGeneralAttrs, ctx.trace);
         ctx.result = toMessages(result, ctx.resolvedModelId);
         logger.trace({ requestId: ctx.requestId, result: ctx.result }, "[messages] Messages");

package/dist/endpoints/responses/handler.js

@@ -123,6 +123,8 @@ export const responses = (config) => {
         });
         logger.trace({ requestId: ctx.requestId, result }, "[responses] AI SDK result");
         addSpanEvent("hebo.ai-sdk.completed");
+        if (result.response.headers)
+            ctx.response = { headers: result.response.headers };
         recordTimeToFirstToken(performance.now() - start, genAiGeneralAttrs, ctx.trace);
         ctx.result = toResponses(result, ctx.resolvedModelId, ctx.body.metadata);
         logger.trace({ requestId: ctx.requestId, result: ctx.result }, "[responses] Responses");

package/dist/endpoints/responses/schema.d.ts

@@ -1489,6 +1489,7 @@ declare const ResponsesInputsSchema: z.ZodObject<{
             medium: "medium";
             high: "high";
             xhigh: "xhigh";
+            max: "max";
         }>>;
         max_tokens: z.ZodOptional<z.ZodNumber>;
         exclude: z.ZodOptional<z.ZodBoolean>;
@@ -1524,6 +1525,7 @@ declare const ResponsesInputsSchema: z.ZodObject<{
         medium: "medium";
         high: "high";
         xhigh: "xhigh";
+        max: "max";
     }>>;
     extra_body: z.ZodOptional<z.ZodType<import("@ai-sdk/provider").SharedV3ProviderMetadata, unknown, z.core.$ZodTypeInternals<import("@ai-sdk/provider").SharedV3ProviderMetadata, unknown>>>;
 }, z.core.$strip>;
@@ -1984,6 +1986,7 @@ export declare const ResponsesBodySchema: z.ZodObject<{
             medium: "medium";
             high: "high";
             xhigh: "xhigh";
+            max: "max";
         }>>;
         max_tokens: z.ZodOptional<z.ZodNumber>;
         exclude: z.ZodOptional<z.ZodBoolean>;
@@ -2019,6 +2022,7 @@ export declare const ResponsesBodySchema: z.ZodObject<{
         medium: "medium";
         high: "high";
         xhigh: "xhigh";
+        max: "max";
     }>>;
     extra_body: z.ZodOptional<z.ZodType<import("@ai-sdk/provider").SharedV3ProviderMetadata, unknown, z.core.$ZodTypeInternals<import("@ai-sdk/provider").SharedV3ProviderMetadata, unknown>>>;
     model: z.ZodString;

package/dist/endpoints/shared/schema.d.ts

@@ -24,6 +24,7 @@ export declare const ReasoningEffortSchema: z.ZodEnum<{
     medium: "medium";
     high: "high";
     xhigh: "xhigh";
+    max: "max";
 }>;
 export type ReasoningEffort = z.infer<typeof ReasoningEffortSchema>;
 export declare const ReasoningSummarySchema: z.ZodEnum<{
@@ -42,6 +43,7 @@ export declare const ReasoningConfigSchema: z.ZodObject<{
         medium: "medium";
         high: "high";
         xhigh: "xhigh";
+        max: "max";
     }>>;
     max_tokens: z.ZodOptional<z.ZodNumber>;
     exclude: z.ZodOptional<z.ZodBoolean>;

package/dist/endpoints/shared/schema.js

@@ -11,7 +11,15 @@ export const CacheControlSchema = z.object({
     ttl: z.enum(["5m", "1h", "24h"]).optional(),
 });
 export const ProviderMetadataSchema = z.record(z.string(), z.record(z.string(), z.any()));
-export const ReasoningEffortSchema = z.enum(["none", "minimal", "low", "medium", "high", "xhigh"]);
+export const ReasoningEffortSchema = z.enum([
+    "none",
+    "minimal",
+    "low",
+    "medium",
+    "high",
+    "xhigh",
+    "max",
+]);
 export const ReasoningSummarySchema = z.enum(["auto", "concise", "detailed", "none"]);
 export const ReasoningConfigSchema = z.object({
     enabled: z.optional(z.boolean()),

package/dist/errors/ai-sdk.js

@@ -1,11 +1,20 @@
 import { AISDKError, APICallError, DownloadError, EmptyResponseBodyError, InvalidArgumentError, InvalidDataContentError, InvalidMessageRoleError, InvalidPromptError, InvalidResponseDataError, InvalidStreamPartError, InvalidToolApprovalError, InvalidToolInputError, JSONParseError, LoadAPIKeyError, LoadSettingError, MessageConversionError, MissingToolResultsError, NoContentGeneratedError, NoImageGeneratedError, NoObjectGeneratedError, NoOutputGeneratedError, NoSpeechGeneratedError, NoSuchModelError, NoSuchProviderError, NoSuchToolError, NoTranscriptGeneratedError, NoVideoGeneratedError, RetryError, ToolCallNotFoundForApprovalError, ToolCallRepairError, TooManyEmbeddingValuesForCallError, TypeValidationError, UIMessageStreamError, UnsupportedModelVersionError, UnsupportedFunctionalityError, } from "ai";
 import { GatewayError } from "./gateway";
-import { STATUS_CODE } from "./utils";
+import { STATUS_TEXT } from "./utils";
+const normalizeApiCallError = (error) => {
+    const status = error.statusCode ?? (error.isRetryable ? 502 : 422);
+    const statusText = `UPSTREAM_${STATUS_TEXT(status)}`;
+    return new GatewayError(error, status, statusText, undefined, error.responseHeaders ?? undefined);
+};
 export const normalizeAiSdkError = (error) => {
     if (APICallError.isInstance(error)) {
-        const status = error.statusCode ?? (error.isRetryable ? 502 : 422);
-        const code = `UPSTREAM_${STATUS_CODE(status)}`;
-        return new GatewayError(error, status, code);
+        return normalizeApiCallError(error);
+    }
+    if (RetryError.isInstance(error)) {
+        if (APICallError.isInstance(error.lastError)) {
+            return normalizeApiCallError(error.lastError);
+        }
+        return new GatewayError(error, 502, `UPSTREAM_${STATUS_TEXT(502)}`);
     }
     if (JSONParseError.isInstance(error) ||
         InvalidResponseDataError.isInstance(error) ||
@@ -15,7 +24,6 @@ export const normalizeAiSdkError = (error) => {
         NoOutputGeneratedError.isInstance(error) ||
         InvalidStreamPartError.isInstance(error) ||
         UIMessageStreamError.isInstance(error) ||
-        RetryError.isInstance(error) ||
         DownloadError.isInstance(error) ||
         ToolCallRepairError.isInstance(error) ||
         NoImageGeneratedError.isInstance(error) ||
@@ -23,7 +31,7 @@ export const normalizeAiSdkError = (error) => {
         NoSpeechGeneratedError.isInstance(error) ||
         NoTranscriptGeneratedError.isInstance(error) ||
         NoVideoGeneratedError.isInstance(error)) {
-        return new GatewayError(error, 502, `UPSTREAM_${STATUS_CODE(502)}`);
+        return new GatewayError(error, 502, `UPSTREAM_${STATUS_TEXT(502)}`);
     }
     if (InvalidArgumentError.isInstance(error) ||
         InvalidPromptError.isInstance(error) ||
@@ -40,7 +48,7 @@ export const normalizeAiSdkError = (error) => {
         TooManyEmbeddingValuesForCallError.isInstance(error) ||
         NoSuchModelError.isInstance(error) ||
         NoSuchProviderError.isInstance(error)) {
-        return new GatewayError(error, 422, `UPSTREAM_${STATUS_CODE(422)}`);
+        return new GatewayError(error, 422, `UPSTREAM_${STATUS_TEXT(422)}`);
     }
     if (LoadSettingError.isInstance(error) || LoadAPIKeyError.isInstance(error)) {
         return new GatewayError(error, 500);

package/dist/errors/anthropic.d.ts

@@ -9,7 +9,8 @@ export declare const AnthropicErrorSchema: z.ZodObject<{
 export declare class AnthropicError {
     readonly type: "error";
     readonly error: z.infer<typeof AnthropicErrorSchema>["error"];
+    status: number;
     constructor(message: string, type?: string);
 }
-export declare function toAnthropicError(error: unknown): AnthropicError;
-export declare function toAnthropicErrorResponse(error: unknown, responseInit?: ResponseInit): Response;
+export declare function toAnthropicError(error: unknown, requestId?: string): AnthropicError;
+export declare function toAnthropicErrorResponse(error: unknown, init: ResponseInit): Response;

package/dist/errors/anthropic.js

@@ -14,6 +14,8 @@ export class AnthropicError {
     error;
     constructor(message, type = "api_error") {
         this.error = { type, message };
+        // internal property to derive status from error handlers without breaking official format
+        Object.defineProperty(this, "status", { value: 500, writable: true });
     }
 }
 const mapType = (status) => {
@@ -22,12 +24,12 @@ const mapType = (status) => {
             return "invalid_request_error";
         case 401:
             return "authentication_error";
+        case 402:
+            return "billing_error";
         case 403:
             return "permission_error";
         case 404:
             return "not_found_error";
-        case 402:
-            return "billing_error";
         case 413:
             return "request_too_large";
         case 429:
@@ -40,15 +42,12 @@ const mapType = (status) => {
             return status >= 500 ? "api_error" : "invalid_request_error";
     }
 };
-export function toAnthropicError(error) {
+export function toAnthropicError(error, requestId) {
     const meta = getErrorMeta(error);
-    return new AnthropicError(maybeMaskMessage(meta), mapType(meta.status));
+    const anthropicError = new AnthropicError(maybeMaskMessage(error instanceof Error ? error.message : String(error), meta.status, requestId), mapType(meta.status));
+    anthropicError.status = meta.status;
+    return anthropicError;
 }
-export function toAnthropicErrorResponse(error, responseInit) {
-    const meta = getErrorMeta(error);
-    return toResponse(new AnthropicError(maybeMaskMessage(meta, resolveRequestId(responseInit)), mapType(meta.status)), {
-        status: meta.status,
-        statusText: meta.code,
-        headers: responseInit?.headers,
-    });
+export function toAnthropicErrorResponse(error, init) {
+    return toResponse(new AnthropicError(maybeMaskMessage(error instanceof Error ? error.message : String(error), init.status ?? 500, resolveRequestId(init)), mapType(init.status ?? 500)), init);
 }

package/dist/errors/gateway.d.ts

@@ -1,5 +1,6 @@
 export declare class GatewayError extends Error {
     readonly status: number;
-    readonly code: string;
-    constructor(error: unknown, status: number, code?: string, cause?: unknown);
+    readonly statusText: string;
+    readonly headers: Record<string, string> | undefined;
+    constructor(error: unknown, status: number, statusText?: string, cause?: unknown, headers?: Record<string, string>);
 }

package/dist/errors/gateway.js

@@ -1,13 +1,19 @@
-import { STATUS_CODE } from "./utils";
+import { X_SHOULD_RETRY_HEADER } from "../utils/headers";
+import { STATUS_TEXT } from "./utils";
 export class GatewayError extends Error {
     status;
-    code;
-    constructor(error, status, code, cause) {
+    statusText;
+    headers;
+    constructor(error, status, statusText, cause, headers) {
         const isError = error instanceof Error;
         super(isError ? error.message : String(error));
         this.name = "GatewayError";
         this.cause = cause ?? (isError ? error : undefined);
         this.status = status;
-        this.code = code ?? STATUS_CODE(status);
+        this.statusText = statusText ?? STATUS_TEXT(status);
+        this.headers = headers;
+        if (!this.statusText.startsWith("UPSTREAM_")) {
+            (this.headers ??= {})[X_SHOULD_RETRY_HEADER] = "false";
+        }
     }
 }

package/dist/errors/openai.d.ts

@@ -9,7 +9,8 @@ export declare const OpenAIErrorSchema: z.ZodObject<{
 }, z.core.$strip>;
 export declare class OpenAIError {
     readonly error: z.infer<typeof OpenAIErrorSchema>["error"];
+    status: number;
     constructor(message: string, type?: string, code?: string, param?: string);
 }
-export declare function toOpenAIError(error: unknown): OpenAIError;
-export declare function toOpenAIErrorResponse(error: unknown, responseInit?: ResponseInit): Response;
+export declare function toOpenAIError(error: unknown, requestId?: string): OpenAIError;
+export declare function toOpenAIErrorResponse(error: unknown, init: ResponseInit): Response;

package/dist/errors/openai.js

@@ -14,18 +14,17 @@ export class OpenAIError {
     error;
     constructor(message, type = "server_error", code, param = "") {
         this.error = { message, type, code: code?.toLowerCase(), param };
+        // internal property to derive status from error handlers without breaking official format
+        Object.defineProperty(this, "status", { value: 500, writable: true });
     }
 }
 const mapType = (status) => (status < 500 ? "invalid_request_error" : "server_error");
-export function toOpenAIError(error) {
+export function toOpenAIError(error, requestId) {
     const meta = getErrorMeta(error);
-    return new OpenAIError(maybeMaskMessage(meta), mapType(meta.status), meta.code);
+    const openAIError = new OpenAIError(maybeMaskMessage(error instanceof Error ? error.message : String(error), meta.status, requestId), mapType(meta.status), meta.statusText);
+    openAIError.status = meta.status;
+    return openAIError;
 }
-export function toOpenAIErrorResponse(error, responseInit) {
-    const meta = getErrorMeta(error);
-    return toResponse(new OpenAIError(maybeMaskMessage(meta, resolveRequestId(responseInit)), mapType(meta.status), meta.code), {
-        ...responseInit,
-        status: meta.status,
-        statusText: meta.code,
-    });
+export function toOpenAIErrorResponse(error, init) {
+    return toResponse(new OpenAIError(maybeMaskMessage(error instanceof Error ? error.message : String(error), init.status ?? 500, resolveRequestId(init)), mapType(init.status ?? 500), init.statusText ?? "INTERNAL_SERVER_ERROR"), init);
 }

package/dist/errors/utils.d.ts

@@ -16,11 +16,11 @@ export declare const STATUS_CODES: {
     readonly 503: "SERVICE_UNAVAILABLE";
     readonly 504: "GATEWAY_TIMEOUT";
 };
-export declare const STATUS_CODE: (status: number) => "BAD_REQUEST" | "UNAUTHORIZED" | "PAYMENT_REQUIRED" | "FORBIDDEN" | "NOT_FOUND" | "METHOD_NOT_ALLOWED" | "CONFLICT" | "PAYLOAD_TOO_LARGE" | "UNSUPPORTED_MEDIA_TYPE" | "UNPROCESSABLE_ENTITY" | "TOO_MANY_REQUESTS" | "CLIENT_CLOSED_REQUEST" | "INTERNAL_SERVER_ERROR" | "BAD_GATEWAY" | "SERVICE_UNAVAILABLE" | "GATEWAY_TIMEOUT";
+export declare const STATUS_TEXT: (status: number) => "BAD_REQUEST" | "UNAUTHORIZED" | "PAYMENT_REQUIRED" | "FORBIDDEN" | "NOT_FOUND" | "METHOD_NOT_ALLOWED" | "CONFLICT" | "PAYLOAD_TOO_LARGE" | "UNSUPPORTED_MEDIA_TYPE" | "UNPROCESSABLE_ENTITY" | "TOO_MANY_REQUESTS" | "CLIENT_CLOSED_REQUEST" | "INTERNAL_SERVER_ERROR" | "BAD_GATEWAY" | "SERVICE_UNAVAILABLE" | "GATEWAY_TIMEOUT";
 export type ErrorMeta = {
     status: number;
-    code: string;
-    message: string;
+    statusText: string;
+    headers: Record<string, string>;
 };
 export declare function getErrorMeta(error: unknown): ErrorMeta;
-export declare function maybeMaskMessage(meta: ErrorMeta, requestId?: string): string;
+export declare function maybeMaskMessage(message: string, status: number, requestId?: string): string;

package/dist/errors/utils.js

@@ -19,37 +19,37 @@ export const STATUS_CODES = {
     503: "SERVICE_UNAVAILABLE",
     504: "GATEWAY_TIMEOUT",
 };
-export const STATUS_CODE = (status) => {
+export const STATUS_TEXT = (status) => {
     const label = STATUS_CODES[status];
     if (label)
         return label;
     return status >= 400 && status < 500 ? STATUS_CODES[400] : STATUS_CODES[500];
 };
-// FUTURE: always return a wrapped GatewayError?
 export function getErrorMeta(error) {
-    const message = error instanceof Error ? error.message : String(error);
     let status;
-    let code;
+    let statusText;
+    let headers;
     if (error instanceof GatewayError) {
-        ({ status, code } = error);
+        ({ status, statusText, headers } = error);
     }
     else {
         const normalized = normalizeAiSdkError(error);
         if (normalized) {
-            ({ status, code } = normalized);
+            ({ status, statusText, headers } = normalized);
         }
         else {
             status = 500;
-            code = STATUS_CODE(status);
+            statusText = STATUS_TEXT(status);
+            headers = {};
         }
     }
-    return { status, code, message };
+    return { status, statusText, headers: headers ?? {} };
 }
-export function maybeMaskMessage(meta, requestId) {
+export function maybeMaskMessage(message, status, requestId) {
     // FUTURE: consider masking all upstream errors, also 4xx
-    if (!(isProduction() && meta.status >= 500)) {
-        return meta.message;
+    if (!(isProduction() && status >= 500)) {
+        return message;
     }
     // FUTURE: always attach requestId to errors (masked and unmasked)
-    return `${STATUS_CODE(meta.status)} (${requestId ?? "see requestId in response headers"})`;
+    return `${STATUS_TEXT(status)} (${requestId ?? "see requestId in response headers"})`;
 }

package/dist/lifecycle.js

@@ -2,6 +2,7 @@ import { parseConfig } from "./config";
 import { toAnthropicError, toAnthropicErrorResponse } from "./errors/anthropic";
 import { GatewayError } from "./errors/gateway";
 import { toOpenAIError, toOpenAIErrorResponse } from "./errors/openai";
+import { getErrorMeta } from "./errors/utils";
 import { logger } from "./logger";
 import { getBaggageAttributes } from "./telemetry/baggage";
 import { instrumentFetch } from "./telemetry/fetch";
@@ -53,7 +54,7 @@ export const winterCgHandler = (run, config) => {
                     requestId: ctx.requestId,
                     err: reason ?? ctx.request.signal.reason,
                 });
-                const isUpstreamError = reason instanceof GatewayError && reason.code.startsWith("UPSTREAM_");
+                const isUpstreamError = reason instanceof GatewayError && reason.statusText.startsWith("UPSTREAM_");
                 span.recordError(reason, realStatus >= 500 || isUpstreamError);
             }
             span.setAttributes({ "http.response.status_code_effective": realStatus });
@@ -76,10 +77,10 @@ export const winterCgHandler = (run, config) => {
                 }
                 if (!ctx.response) {
                     ctx.result = (await run(ctx, parsedConfig));
-                    const formatError = ctx.operation === "messages" ? toAnthropicError : toOpenAIError;
-                    ctx.response = toResponse(ctx.result, prepareResponseInit(ctx.requestId), {
+                    const toError = ctx.operation === "messages" ? toAnthropicError : toOpenAIError;
+                    ctx.response = toResponse(ctx.result, prepareResponseInit(ctx.requestId, ctx.response), {
                         onDone: finalize,
-                        formatError,
+                        toError: (error) => toError(error, ctx.requestId),
                     });
                 }
                 if (parsedConfig.hooks?.onResponse) {
@@ -111,11 +112,10 @@ export const winterCgHandler = (run, config) => {
                 const errorPayload = ctx.request.signal.aborted
                     ? new GatewayError(error ?? ctx.request.signal.reason, 499)
                     : error;
-                const errorResponseInit = prepareResponseInit(ctx.requestId);
-                ctx.response ??=
-                    ctx.operation === "messages"
-                        ? toAnthropicErrorResponse(errorPayload, errorResponseInit)
-                        : toOpenAIErrorResponse(errorPayload, errorResponseInit);
+                if (!(ctx.response instanceof Response)) {
+                    const toErrorResponse = ctx.operation === "messages" ? toAnthropicErrorResponse : toOpenAIErrorResponse;
+                    ctx.response = toErrorResponse(errorPayload, prepareResponseInit(ctx.requestId, getErrorMeta(errorPayload)));
+                }
                 finalize(ctx.response.status, error);
             }
         });

package/dist/middleware/utils.js

@@ -19,6 +19,9 @@ export function calculateReasoningBudgetFromEffort(effort, maxTokens, minTokens
         case "xhigh":
             percentage = 0.95;
             break;
+        case "max":
+            percentage = 1.0;
+            break;
     }
     return Math.max(minTokens, Math.floor(maxTokens * percentage));
 }

package/dist/models/amazon/middleware.js

@@ -28,6 +28,7 @@ function mapNovaEffort(effort) {
             return "medium";
         case "high":
         case "xhigh":
+        case "max":
             return "high";
     }
     return undefined;

package/dist/models/anthropic/middleware.d.ts

@@ -1,5 +1,5 @@
 import type { LanguageModelMiddleware } from "ai";
 import type { ChatCompletionsReasoningEffort } from "../../endpoints/chat-completions/schema";
-export declare function mapClaudeReasoningEffort(effort: ChatCompletionsReasoningEffort, modelId: string): "low" | "medium" | "high" | "max" | undefined;
+export declare function mapClaudeReasoningEffort(effort: ChatCompletionsReasoningEffort, modelId: string): "low" | "medium" | "high" | "xhigh" | "max" | undefined;
 export declare const claudeReasoningMiddleware: LanguageModelMiddleware;
 export declare const claudePromptCachingMiddleware: LanguageModelMiddleware;

package/dist/models/anthropic/middleware.js

@@ -6,11 +6,28 @@ const isClaude = (family, version) => {
         modelId.includes(`claude-${family}-${dashed}`);
 };
 const isClaude4 = (modelId) => modelId.includes("claude-") && modelId.includes("-4");
+const isOpus47 = isClaude("opus", "4.7");
 const isOpus46 = isClaude("opus", "4.6");
 const isOpus45 = isClaude("opus", "4.5");
 const isOpus4 = isClaude("opus", "4");
 const isSonnet46 = isClaude("sonnet", "4.6");
 export function mapClaudeReasoningEffort(effort, modelId) {
+    if (isOpus47(modelId)) {
+        switch (effort) {
+            case "none":
+            case "minimal":
+            case "low":
+                return "low";
+            case "medium":
+                return "medium";
+            case "high":
+                return "high";
+            case "xhigh":
+                return "xhigh";
+            case "max":
+                return "max";
+        }
+    }
     if (isOpus46(modelId)) {
         switch (effort) {
             case "none":
@@ -22,6 +39,7 @@ export function mapClaudeReasoningEffort(effort, modelId) {
             case "high":
                 return "high";
             case "xhigh":
+            case "max":
                 return "max";
         }
     }
@@ -34,11 +52,14 @@ export function mapClaudeReasoningEffort(effort, modelId) {
             return "medium";
         case "high":
         case "xhigh":
+        case "max":
             return "high";
     }
     return undefined;
 }
 function getMaxOutputTokens(modelId) {
+    if (isOpus47(modelId))
+        return 128_000;
     if (isOpus46(modelId))
         return 128_000;
     if (isOpus45(modelId))
@@ -69,9 +90,13 @@ export const claudeReasoningMiddleware = {
         }
         else if (reasoning.effort) {
             if (isClaude4(modelId)) {
+                // @ts-expect-error AI SDK type missing "xhigh" effort level (native on Opus 4.7+)
                 target.effort = mapClaudeReasoningEffort(reasoning.effort, modelId);
             }
-            if (isOpus46(modelId)) {
+            if (isOpus47(modelId)) {
+                target.thinking = { type: "adaptive" };
+            }
+            else if (isOpus46(modelId)) {
                 target.thinking = clampedMaxTokens
                     ? // @ts-expect-error AI SDK type missing type:adaptive with budgetToken
                         { type: "adaptive", budgetTokens: clampedMaxTokens }

package/dist/models/anthropic/presets.d.ts

@@ -107,6 +107,18 @@ export declare const claudeOpus45: import("../../utils/preset").Preset<"anthropi
     context: number;
     providers: readonly ["anthropic", "bedrock", "vertex", "azure"];
 }>;
+export declare const claudeOpus47: import("../../utils/preset").Preset<"anthropic/claude-opus-4.7", CatalogModel, {
+    name: string;
+    capabilities: ("reasoning" | "temperature" | "attachments" | "tool_call" | "structured_output")[];
+    context: number;
+    created: string;
+    knowledge: string;
+    modalities: {
+        input: readonly ["text", "image", "pdf", "file"];
+        output: readonly ["text"];
+    };
+    providers: readonly ["anthropic", "bedrock", "vertex", "azure"];
+}>;
 export declare const claudeOpus46: import("../../utils/preset").Preset<"anthropic/claude-opus-4.6", CatalogModel, {
     name: string;
     capabilities: ("reasoning" | "temperature" | "attachments" | "tool_call" | "structured_output")[];
@@ -144,7 +156,18 @@ export declare const claudeOpus4: import("../../utils/preset").Preset<"anthropic
     providers: readonly ["anthropic", "bedrock", "vertex", "azure"];
 }>;
 export declare const claude: {
-    readonly latest: readonly [import("../../utils/preset").Preset<"anthropic/claude-sonnet-4.6", CatalogModel, {
+    readonly latest: readonly [import("../../utils/preset").Preset<"anthropic/claude-opus-4.7", CatalogModel, {
+        name: string;
+        capabilities: ("reasoning" | "temperature" | "attachments" | "tool_call" | "structured_output")[];
+        context: number;
+        created: string;
+        knowledge: string;
+        modalities: {
+            input: readonly ["text", "image", "pdf", "file"];
+            output: readonly ["text"];
+        };
+        providers: readonly ["anthropic", "bedrock", "vertex", "azure"];
+    }>, import("../../utils/preset").Preset<"anthropic/claude-sonnet-4.6", CatalogModel, {
         name: string;
         capabilities: ("reasoning" | "temperature" | "attachments" | "tool_call" | "structured_output")[];
         created: string;
@@ -266,6 +289,17 @@ export declare const claude: {
         };
         context: number;
         providers: readonly ["anthropic", "bedrock", "vertex", "azure"];
+    }> | import("../../utils/preset").Preset<"anthropic/claude-opus-4.7", CatalogModel, {
+        name: string;
+        capabilities: ("reasoning" | "temperature" | "attachments" | "tool_call" | "structured_output")[];
+        context: number;
+        created: string;
+        knowledge: string;
+        modalities: {
+            input: readonly ["text", "image", "pdf", "file"];
+            output: readonly ["text"];
+        };
+        providers: readonly ["anthropic", "bedrock", "vertex", "azure"];
     }> | import("../../utils/preset").Preset<"anthropic/claude-opus-4.6", CatalogModel, {
         name: string;
         capabilities: ("reasoning" | "temperature" | "attachments" | "tool_call" | "structured_output")[];
@@ -300,7 +334,18 @@ export declare const claude: {
         context: number;
         providers: readonly ["anthropic", "bedrock", "vertex", "azure"];
     }>)[];
-    readonly "v4.x": readonly [import("../../utils/preset").Preset<"anthropic/claude-sonnet-4.6", CatalogModel, {
+    readonly "v4.x": readonly [import("../../utils/preset").Preset<"anthropic/claude-opus-4.7", CatalogModel, {
+        name: string;
+        capabilities: ("reasoning" | "temperature" | "attachments" | "tool_call" | "structured_output")[];
+        context: number;
+        created: string;
+        knowledge: string;
+        modalities: {
+            input: readonly ["text", "image", "pdf", "file"];
+            output: readonly ["text"];
+        };
+        providers: readonly ["anthropic", "bedrock", "vertex", "azure"];
+    }>, import("../../utils/preset").Preset<"anthropic/claude-sonnet-4.6", CatalogModel, {
         name: string;
         capabilities: ("reasoning" | "temperature" | "attachments" | "tool_call" | "structured_output")[];
         created: string;
@@ -434,6 +479,18 @@ export declare const claude: {
         context: number;
         providers: readonly ["anthropic", "bedrock", "vertex", "azure"];
     }>];
+    readonly "v4.7": readonly [import("../../utils/preset").Preset<"anthropic/claude-opus-4.7", CatalogModel, {
+        name: string;
+        capabilities: ("reasoning" | "temperature" | "attachments" | "tool_call" | "structured_output")[];
+        context: number;
+        created: string;
+        knowledge: string;
+        modalities: {
+            input: readonly ["text", "image", "pdf", "file"];
+            output: readonly ["text"];
+        };
+        providers: readonly ["anthropic", "bedrock", "vertex", "azure"];
+    }>];
     readonly "v4.6": readonly [import("../../utils/preset").Preset<"anthropic/claude-sonnet-4.6", CatalogModel, {
         name: string;
         capabilities: ("reasoning" | "temperature" | "attachments" | "tool_call" | "structured_output")[];
@@ -663,7 +720,18 @@ export declare const claude: {
         context: number;
         providers: readonly ["anthropic", "bedrock", "vertex", "azure"];
     }>];
-    readonly opus: readonly [import("../../utils/preset").Preset<"anthropic/claude-opus-4.6", CatalogModel, {
+    readonly opus: readonly [import("../../utils/preset").Preset<"anthropic/claude-opus-4.7", CatalogModel, {
+        name: string;
+        capabilities: ("reasoning" | "temperature" | "attachments" | "tool_call" | "structured_output")[];
+        context: number;
+        created: string;
+        knowledge: string;
+        modalities: {
+            input: readonly ["text", "image", "pdf", "file"];
+            output: readonly ["text"];
+        };
+        providers: readonly ["anthropic", "bedrock", "vertex", "azure"];
+    }>, import("../../utils/preset").Preset<"anthropic/claude-opus-4.6", CatalogModel, {
         name: string;
         capabilities: ("reasoning" | "temperature" | "attachments" | "tool_call" | "structured_output")[];
         created: string;

package/dist/models/anthropic/presets.js

@@ -87,6 +87,15 @@ export const claudeOpus45 = presetFor()("anthropic/claude-opus-4.5", {
     created: "2025-11-01",
     knowledge: "2025-05",
 });
+export const claudeOpus47 = presetFor()("anthropic/claude-opus-4.7", {
+    ...CLAUDE_BASE,
+    ...CLAUDE_PDF_MODALITIES,
+    name: "Claude Opus 4.7",
+    capabilities: [...CLAUDE_BASE.capabilities, "reasoning"],
+    context: 1_000_000,
+    created: "2026-04-16",
+    knowledge: "2026-01",
+});
 export const claudeOpus46 = presetFor()("anthropic/claude-opus-4.6", {
     ...CLAUDE_BASE,
     ...CLAUDE_PDF_MODALITIES,
@@ -112,6 +121,7 @@ export const claudeOpus4 = presetFor()("anthropic/claude-opus-4", {
     knowledge: "2025-03",
 });
 const claudeAtomic = {
+    "v4.7": [claudeOpus47],
     "v4.6": [claudeSonnet46, claudeOpus46],
     "v4.5": [claudeHaiku45, claudeSonnet45, claudeOpus45],
     "v4.1": [claudeOpus41],
@@ -121,10 +131,11 @@ const claudeAtomic = {
     v3: [claudeHaiku3],
     haiku: [claudeHaiku45, claudeHaiku35, claudeHaiku3],
     sonnet: [claudeSonnet46, claudeSonnet45, claudeSonnet4, claudeSonnet37, claudeSonnet35],
-    opus: [claudeOpus46, claudeOpus45, claudeOpus41, claudeOpus4],
+    opus: [claudeOpus47, claudeOpus46, claudeOpus45, claudeOpus41, claudeOpus4],
 };
 const claudeGroups = {
     "v4.x": [
+        ...claudeAtomic["v4.7"],
         ...claudeAtomic["v4.6"],
         ...claudeAtomic["v4.5"],
         ...claudeAtomic["v4.1"],
@@ -135,6 +146,6 @@ const claudeGroups = {
 export const claude = {
     ...claudeAtomic,
     ...claudeGroups,
-    latest: [...claudeAtomic["v4.6"]],
+    latest: [...claudeAtomic["v4.7"], ...claudeAtomic["v4.6"]],
     all: Object.values(claudeAtomic).flat(),
 };

package/dist/models/google/middleware.js

@@ -29,6 +29,7 @@ export function mapGeminiReasoningEffort(effort, modelId) {
                 return "medium";
             case "high":
             case "xhigh":
+            case "max":
                 return "high";
         }
     }
@@ -43,6 +44,7 @@ export function mapGeminiReasoningEffort(effort, modelId) {
             return "medium";
         case "high":
         case "xhigh":
+        case "max":
             return "high";
     }
     return undefined;

package/dist/models/openai/middleware.js

@@ -28,6 +28,7 @@ function mapGptOssReasoningEffort(effort) {
             return "medium";
         case "high":
         case "xhigh":
+        case "max":
             return "high";
     }
     return undefined;
@@ -52,7 +53,7 @@ export const openAIReasoningMiddleware = {
             target.reasoningEffort = "none";
         }
         else if (reasoning.effort) {
-            target.reasoningEffort = reasoning.effort;
+            target.reasoningEffort = reasoning.effort === "max" ? "xhigh" : reasoning.effort;
         }
         // FUTURE: warn that reasoning.max_tokens (not supported) was ignored
         delete unknown["reasoning"];

package/dist/models/types.d.ts

@@ -1,5 +1,5 @@
 import type { ProviderId } from "../providers/types";
-export declare const CANONICAL_MODEL_IDS: readonly ["anthropic/claude-opus-4.6", "anthropic/claude-sonnet-4.6", "anthropic/claude-haiku-4.5", "anthropic/claude-sonnet-4.5", "anthropic/claude-opus-4.5", "anthropic/claude-opus-4.1", "anthropic/claude-opus-4", "anthropic/claude-sonnet-4", "anthropic/claude-sonnet-3.7", "anthropic/claude-sonnet-3.5", "anthropic/claude-haiku-3.5", "anthropic/claude-haiku-3", "openai/gpt-oss-20b", "openai/gpt-oss-120b", "openai/gpt-5", "openai/gpt-5-pro", "openai/gpt-5.2", "openai/gpt-5.2-chat", "openai/gpt-5.2-pro", "openai/gpt-5.2-codex", "openai/gpt-5.3-codex", "openai/gpt-5.3-codex-spark", "openai/gpt-5.3-chat", "openai/gpt-5.4", "openai/gpt-5.4-mini", "openai/gpt-5.4-nano", "openai/gpt-5.4-pro", "openai/gpt-5-mini", "openai/gpt-5-nano", "openai/gpt-5-codex", "openai/gpt-5.1-codex", "openai/gpt-5.1-codex-max", "openai/gpt-5.1-codex-mini", "openai/gpt-5.1-chat", "openai/gpt-5.1", "openai/text-embedding-3-small", "openai/text-embedding-3-large", "amazon/nova-micro", "amazon/nova-lite", "amazon/nova-pro", "amazon/nova-premier", "amazon/nova-2-lite", "amazon/nova-2-multimodal-embeddings", "google/gemini-2.5-flash-lite", "google/gemini-2.5-flash", "google/gemini-2.5-pro", "google/gemini-3-flash-preview", "google/gemini-3.1-flash-lite-preview", "google/gemini-3.1-pro-preview", "google/gemini-embedding-2-preview", "google/embedding-001", "google/gemma-3-1b", "google/gemma-3-4b", "google/gemma-3-12b", "google/gemma-3-27b", "google/gemma-4-e2b", "google/gemma-4-e4b", "google/gemma-4-26b-a4b", "google/gemma-4-31b", "meta/llama-3.1-8b", "meta/llama-3.1-70b", "meta/llama-3.1-405b", "meta/llama-3.2-1b", "meta/llama-3.2-3b", "meta/llama-3.2-11b", "meta/llama-3.2-90b", "meta/llama-3.3-70b", "meta/llama-4-scout", "meta/llama-4-maverick", "cohere/embed-v4.0", "cohere/embed-english-v3.0", "cohere/embed-english-light-v3.0", "cohere/embed-multilingual-v3.0", "cohere/embed-multilingual-light-v3.0", "cohere/command-a", "cohere/command-r7b", "cohere/command-a-translate", "cohere/command-a-reasoning", "cohere/command-a-vision", "cohere/command-r", "cohere/command-r-plus", "voyage/voyage-2-code", "voyage/voyage-2-law", "voyage/voyage-2-finance", "voyage/voyage-3-code", "voyage/voyage-3-large", "voyage/voyage-3.5-lite", "voyage/voyage-3.5", "voyage/voyage-4-lite", "voyage/voyage-4", "voyage/voyage-4-large"];
+export declare const CANONICAL_MODEL_IDS: readonly ["anthropic/claude-opus-4.7", "anthropic/claude-opus-4.6", "anthropic/claude-sonnet-4.6", "anthropic/claude-haiku-4.5", "anthropic/claude-sonnet-4.5", "anthropic/claude-opus-4.5", "anthropic/claude-opus-4.1", "anthropic/claude-opus-4", "anthropic/claude-sonnet-4", "anthropic/claude-sonnet-3.7", "anthropic/claude-sonnet-3.5", "anthropic/claude-haiku-3.5", "anthropic/claude-haiku-3", "openai/gpt-oss-20b", "openai/gpt-oss-120b", "openai/gpt-5", "openai/gpt-5-pro", "openai/gpt-5.2", "openai/gpt-5.2-chat", "openai/gpt-5.2-pro", "openai/gpt-5.2-codex", "openai/gpt-5.3-codex", "openai/gpt-5.3-codex-spark", "openai/gpt-5.3-chat", "openai/gpt-5.4", "openai/gpt-5.4-mini", "openai/gpt-5.4-nano", "openai/gpt-5.4-pro", "openai/gpt-5-mini", "openai/gpt-5-nano", "openai/gpt-5-codex", "openai/gpt-5.1-codex", "openai/gpt-5.1-codex-max", "openai/gpt-5.1-codex-mini", "openai/gpt-5.1-chat", "openai/gpt-5.1", "openai/text-embedding-3-small", "openai/text-embedding-3-large", "amazon/nova-micro", "amazon/nova-lite", "amazon/nova-pro", "amazon/nova-premier", "amazon/nova-2-lite", "amazon/nova-2-multimodal-embeddings", "google/gemini-2.5-flash-lite", "google/gemini-2.5-flash", "google/gemini-2.5-pro", "google/gemini-3-flash-preview", "google/gemini-3.1-flash-lite-preview", "google/gemini-3.1-pro-preview", "google/gemini-embedding-2-preview", "google/embedding-001", "google/gemma-3-1b", "google/gemma-3-4b", "google/gemma-3-12b", "google/gemma-3-27b", "google/gemma-4-e2b", "google/gemma-4-e4b", "google/gemma-4-26b-a4b", "google/gemma-4-31b", "meta/llama-3.1-8b", "meta/llama-3.1-70b", "meta/llama-3.1-405b", "meta/llama-3.2-1b", "meta/llama-3.2-3b", "meta/llama-3.2-11b", "meta/llama-3.2-90b", "meta/llama-3.3-70b", "meta/llama-4-scout", "meta/llama-4-maverick", "cohere/embed-v4.0", "cohere/embed-english-v3.0", "cohere/embed-english-light-v3.0", "cohere/embed-multilingual-v3.0", "cohere/embed-multilingual-light-v3.0", "cohere/command-a", "cohere/command-r7b", "cohere/command-a-translate", "cohere/command-a-reasoning", "cohere/command-a-vision", "cohere/command-r", "cohere/command-r-plus", "voyage/voyage-2-code", "voyage/voyage-2-law", "voyage/voyage-2-finance", "voyage/voyage-3-code", "voyage/voyage-3-large", "voyage/voyage-3.5-lite", "voyage/voyage-3.5", "voyage/voyage-4-lite", "voyage/voyage-4", "voyage/voyage-4-large"];
 export type CanonicalModelId = (typeof CANONICAL_MODEL_IDS)[number];
 export type ModelId = CanonicalModelId | (string & {});
 export type CatalogModel = {

package/dist/models/types.js

@@ -1,5 +1,6 @@
 export const CANONICAL_MODEL_IDS = [
     // Anthropic
+    "anthropic/claude-opus-4.7",
     "anthropic/claude-opus-4.6",
     "anthropic/claude-sonnet-4.6",
     "anthropic/claude-haiku-4.5",

package/dist/providers/bedrock/canonical.js

@@ -9,6 +9,7 @@ import { withCanonicalIds } from "../registry";
 const MAPPING = {
     // Require Inference Profiles and can't be resolved from standard name mapping
     "anthropic/claude-haiku-4.5": "{ip}anthropic.claude-haiku-4-5-20251001-v1:0",
+    "anthropic/claude-opus-4.7": "{ip}anthropic.claude-opus-4-7",
     "anthropic/claude-sonnet-4.6": "{ip}anthropic.claude-sonnet-4-6",
     "anthropic/claude-sonnet-4.5": "{ip}anthropic.claude-sonnet-4-5-20250929-v1:0",
     "anthropic/claude-opus-4.6": "{ip}anthropic.claude-opus-4-6-v1",

package/dist/providers/bedrock/middleware.js

@@ -1,6 +1,7 @@
 import { modelMiddlewareMatcher } from "../../middleware/matcher";
 import { calculateReasoningBudgetFromEffort } from "../../middleware/utils";
-const isClaude46 = (modelId) => modelId.includes("-4-6");
+const BEDROCK_EFFORT_CAPABLE = ["-4-6", "-4-7"];
+const isBedrockEffortCapable = (modelId) => BEDROCK_EFFORT_CAPABLE.some((tag) => modelId.includes(tag));
 // https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html
 export const bedrockServiceTierMiddleware = {
     specificationVersion: "v3",
@@ -83,7 +84,7 @@ export const bedrockClaudeReasoningMiddleware = {
             }
         }
         // FUTURE: bedrock currently does not support "effort" for other 4.x models
-        if (effort !== undefined && isClaude46(model.modelId)) {
+        if (effort !== undefined && isBedrockEffortCapable(model.modelId)) {
             target.maxReasoningEffort = effort;
         }
         delete bedrock.thinking;

package/dist/telemetry/gen-ai.js

@@ -1,5 +1,5 @@
 import { metrics } from "@opentelemetry/api";
-import { STATUS_CODE } from "../errors/utils";
+import { STATUS_TEXT } from "../errors/utils";
 const getMeter = () => metrics.getMeter("@hebo/gateway");
 let requestDurationHistogram;
 let timePerOutputTokenHistogram;
@@ -75,7 +75,7 @@ export const recordRequestDuration = (duration, status, ctx, signalLevel) => {
         return;
     const attrs = getGenAiGeneralAttributes(ctx, signalLevel);
     if (status !== 200) {
-        attrs["error.type"] = `${status} ${STATUS_CODE(status).toLowerCase()}`;
+        attrs["error.type"] = `${status} ${STATUS_TEXT(status).toLowerCase()}`;
     }
     getRequestDurationHistogram().record(duration / 1000, attrs);
 };

package/dist/types.d.ts

@@ -69,8 +69,10 @@ export type GatewayContext = {
     result?: ChatCompletions | ChatCompletionsStream | Embeddings | Messages | MessagesStream | Model | ModelList | Responses | ResponsesStream;
     /**
      * Response object returned by the handler.
+     * Handlers may set this to a `ResponseInit` containing upstream response
+     * headers; the lifecycle merges allowlisted headers into the final `Response`.
      */
-    response?: Response;
+    response?: Response | ResponseInit;
     /**
      * Per-request telemetry signal level override.
      * When set (via body parameter or hook), overrides `cfg.telemetry.signals.gen_ai`

package/dist/utils/headers.d.ts

@@ -1,4 +1,9 @@
 export declare const REQUEST_ID_HEADER = "x-request-id";
+export declare const RETRY_AFTER_HEADER = "retry-after";
+export declare const RETRY_AFTER_MS_HEADER = "retry-after-ms";
+export declare const X_SHOULD_RETRY_HEADER = "x-should-retry";
 type HeaderSource = Request | ResponseInit | undefined;
 export declare const resolveRequestId: (source: HeaderSource) => string | undefined;
+export declare const filterResponseHeaders: (upstream?: HeadersInit) => Record<string, string>;
+export declare const buildRetryHeaders: (status: number, upstream?: Record<string, string>) => Record<string, string>;
 export {};

package/dist/utils/headers.js

@@ -1,22 +1,69 @@
 export const REQUEST_ID_HEADER = "x-request-id";
+export const RETRY_AFTER_HEADER = "retry-after";
+export const RETRY_AFTER_MS_HEADER = "retry-after-ms";
+export const X_SHOULD_RETRY_HEADER = "x-should-retry";
+const RESPONSE_HEADER_ALLOWLIST = [
+    RETRY_AFTER_HEADER,
+    RETRY_AFTER_MS_HEADER,
+    X_SHOULD_RETRY_HEADER,
+];
+const RETRYABLE_STATUS_CODES = new Set([408, 409, 429, 500, 502, 503, 504]);
+const DEFAULT_RETRY_AFTER_MS = 1000;
 export const resolveRequestId = (source) => {
     if (!source)
         return undefined;
     if (source instanceof Request) {
         return source.headers.get(REQUEST_ID_HEADER) ?? undefined;
     }
-    const headers = source.headers;
-    if (!headers)
+    if (!source.headers)
         return undefined;
+    return getHeader(source.headers, REQUEST_ID_HEADER);
+};
+function getHeader(headers, key) {
     if (headers instanceof Headers) {
-        return headers.get(REQUEST_ID_HEADER) ?? undefined;
+        return headers.get(key) ?? undefined;
     }
     if (Array.isArray(headers)) {
-        for (const [key, value] of headers) {
-            if (key.toLowerCase() === REQUEST_ID_HEADER)
-                return value;
+        for (const [k, v] of headers) {
+            if (k.toLowerCase() === key.toLowerCase()) {
+                return v;
+            }
+        }
+        return undefined;
+    }
+    return headers[key] ?? headers[key.toLowerCase()];
+}
+export const filterResponseHeaders = (upstream) => {
+    if (!upstream)
+        return {};
+    const filtered = {};
+    for (const key of RESPONSE_HEADER_ALLOWLIST) {
+        const value = getHeader(upstream, key);
+        if (value !== undefined) {
+            filtered[key] = value;
         }
+    }
+    return filtered;
+};
+function deriveRetryAfterMs(retryAfter) {
+    if (retryAfter === undefined)
+        return undefined;
+    const num = Number(retryAfter);
+    if (Number.isFinite(num) && num > 0)
+        return num * 1000;
+    const dateMs = Date.parse(retryAfter);
+    if (!Number.isFinite(dateMs))
         return undefined;
+    const deltaMs = dateMs - Date.now();
+    return deltaMs > 0 ? deltaMs : undefined;
+}
+export const buildRetryHeaders = (status, upstream = {}) => {
+    if (!RETRYABLE_STATUS_CODES.has(status)) {
+        upstream[X_SHOULD_RETRY_HEADER] = "false";
+        return upstream;
     }
-    return headers[REQUEST_ID_HEADER];
+    upstream[RETRY_AFTER_MS_HEADER] ??= String(deriveRetryAfterMs(upstream[RETRY_AFTER_HEADER]) ?? DEFAULT_RETRY_AFTER_MS);
+    upstream[RETRY_AFTER_HEADER] = String(Math.ceil((Number(upstream[RETRY_AFTER_MS_HEADER]) || DEFAULT_RETRY_AFTER_MS) / 1000));
+    upstream[X_SHOULD_RETRY_HEADER] ??= "true";
+    return upstream;
 };

package/dist/utils/response.d.ts

@@ -1,7 +1,7 @@
 import type { SseFrame } from "./stream";
-export declare const prepareResponseInit: (requestId: string) => ResponseInit;
-export declare const mergeResponseInit: (defaultHeaders: HeadersInit, responseInit?: ResponseInit) => ResponseInit;
+export declare const prepareResponseInit: (requestId: string, upstream?: ResponseInit) => ResponseInit;
+export declare const mergeResponseInit: (headers: Record<string, string>, responseInit?: ResponseInit) => ResponseInit;
 export declare const toResponse: (result: ReadableStream<SseFrame> | Uint8Array<ArrayBuffer> | object | string, responseInit?: ResponseInit, streamOptions?: {
     onDone?: (status: number, reason?: unknown) => void;
-    formatError?: (error: unknown) => unknown;
+    toError?: (error: unknown) => unknown;
 }) => Response;

package/dist/utils/response.js

@@ -1,19 +1,23 @@
-import { REQUEST_ID_HEADER } from "./headers";
+import { buildRetryHeaders, filterResponseHeaders, REQUEST_ID_HEADER } from "./headers";
 import { toSseStream } from "./stream";
 const TEXT_ENCODER = new TextEncoder();
-export const prepareResponseInit = (requestId) => ({
-    headers: { [REQUEST_ID_HEADER]: requestId },
-});
-export const mergeResponseInit = (defaultHeaders, responseInit) => {
-    const headers = new Headers(defaultHeaders);
+export const prepareResponseInit = (requestId, upstream) => {
+    const init = upstream ?? {};
+    init.headers = filterResponseHeaders(upstream?.headers);
+    if (init.status && init.status >= 400)
+        init.headers = buildRetryHeaders(init.status, init.headers);
+    init.headers[REQUEST_ID_HEADER] = requestId;
+    return init;
+};
+export const mergeResponseInit = (headers, responseInit) => {
+    if (!responseInit)
+        return { headers };
     const override = responseInit?.headers;
     if (override) {
         new Headers(override).forEach((value, key) => {
-            headers.set(key, value);
+            headers[key] = value;
         });
     }
-    if (!responseInit)
-        return { headers };
     return {
         status: responseInit.status,
         statusText: responseInit.statusText,

package/dist/utils/stream.d.ts

@@ -5,6 +5,6 @@ export type SseFrame<T = unknown, E extends string | undefined = string | undefi
 export type SseErrorFrame = SseFrame<Error, "error" | undefined>;
 export declare function toSseStream(src: ReadableStream<SseFrame>, options?: {
     onDone?: (status: number, reason?: unknown) => void;
+    toError?: (error: unknown) => unknown;
     keepAliveMs?: number;
-    formatError?: (error: unknown) => unknown;
 }): ReadableStream<Uint8Array>;

package/dist/utils/stream.js

@@ -1,4 +1,3 @@
-import { toOpenAIError } from "../errors/openai";
 const TEXT_ENCODER = new TextEncoder();
 const SSE_DONE_CHUNK = TEXT_ENCODER.encode("data: [DONE]\n\n");
 const SSE_KEEP_ALIVE_CHUNK = TEXT_ENCODER.encode(": keep-alive\n\n");
@@ -59,13 +58,9 @@ export function toSseStream(src, options = {}) {
                 }
                 const value = result.value;
                 if (value.event === "error" || value.data instanceof Error) {
-                    const error = options.formatError
-                        ? options.formatError(value.data)
-                        : toOpenAIError(value.data);
+                    const error = options.toError?.(value.data) ?? value.data;
                     controller.enqueue(TEXT_ENCODER.encode(serializeSseFrame({ event: value.event, data: error })));
-                    const openAiError = toOpenAIError(value.data);
-                    const errorStatus = openAiError?.error.type === "invalid_request_error" ? 422 : 502;
-                    done(controller, errorStatus, value.data);
+                    done(controller, error["status"] ?? 502, value.data);
                     reader.cancel(value.data).catch(() => { });
                     return;
                 }
@@ -74,12 +69,9 @@ export function toSseStream(src, options = {}) {
             }
             catch (error) {
                 try {
-                    const errorPayload = options.formatError
-                        ? options.formatError(error)
-                        : toOpenAIError(error);
                     controller.enqueue(TEXT_ENCODER.encode(serializeSseFrame({
                         event: "error",
-                        data: errorPayload,
+                        data: options.toError?.(error) ?? error,
                     })));
                 }
                 catch { }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hebo-ai/gateway",
-  "version": "0.10.4",
+  "version": "0.10.6",
   "description": "AI gateway as a framework. For full control over models, routing & lifecycle. OpenAI /chat/completions, OpenResponses /responses & Anthropic /messages.",
   "keywords": [
     "ai",