promptfoo 0.121.5 → 0.121.8

package/dist/src/{fetch-Y5qX_kST.js → fetch-BSSAcMxf.js}

@@ -1,7 +1,8 @@
 import { t as __exportAll } from "./chunk-DEq-mXcV.js";
 import { S as getEnvBool, T as getEnvString, a as logger, i as logRequestResponse, k as state, p as sanitizeUrl, w as getEnvInt, x as getConfigDirectoryPath } from "./logger-KD8JjCRJ.js";
-import { n as VERSION } from "./version-0frU0UTr.js";
+import { n as VERSION } from "./version-CHR-EFec.js";
 import { t as invariant } from "./invariant-DIYf9sP1.js";
+import { AsyncLocalStorage } from "node:async_hooks";
 import * as fs$2 from "fs";
 import * as path$1 from "path";
 import path from "path";
@@ -61,8 +62,8 @@ function calculateCost(modelName, config, promptTokens, completionTokens, models
 	const model = models.find((m) => m.id === modelName);
 	if (!model || !model.cost) return;
 	const longContextCost = model.cost.longContext && promptTokens > model.cost.longContext.threshold ? model.cost.longContext : void 0;
-	const inputCost = config.cost ?? longContextCost?.input ?? model.cost.input;
-	const outputCost = config.cost ?? longContextCost?.output ?? model.cost.output;
+	const inputCost = config.inputCost ?? config.cost ?? longContextCost?.input ?? model.cost.input;
+	const outputCost = config.outputCost ?? config.cost ?? longContextCost?.output ?? model.cost.output;
 	return inputCost * promptTokens + outputCost * completionTokens;
 }
 /**
@@ -469,7 +470,10 @@ function writeGlobalConfigPartial(partialConfig) {
 	});
 	writeGlobalConfig(updatedConfig);
 }
-const API_HOST = getEnvString("API_HOST", "https://api.promptfoo.app");
+//#endregion
+//#region src/globalConfig/cloud.ts
+const CLOUD_API_HOST = "https://api.promptfoo.app";
+const API_HOST = getEnvString("API_HOST", CLOUD_API_HOST);
 const SHARING_CUTOFF_DATE = /* @__PURE__ */ new Date("2026-03-09T00:00:00Z");
 var CloudConfig = class {
 	config;
@@ -652,6 +656,14 @@ function isConnectionError(error) {
 /**
 * Enhanced fetch wrapper that adds logging, authentication, error handling, and optional compression
 */
+function isPromptfooCloudApiHost(url) {
+	try {
+		const targetUrl = url instanceof Request ? url.url : url.toString();
+		return new URL(targetUrl).origin === CLOUD_API_HOST;
+	} catch {
+		return false;
+	}
+}
 async function monkeyPatchFetch(url, options) {
 	const NO_LOG_URLS = [
 		R_ENDPOINT,
@@ -671,7 +683,7 @@ async function monkeyPatchFetch(url, options) {
 	} catch (e) {
 		logger.warn(`Failed to compress request body: ${e}`);
 	}
-	if (typeof url === "string" && url.startsWith("https://api.promptfoo.app") || url instanceof URL && url.host === "https://api.promptfoo.app".replace(/^https?:\/\//, "")) {
+	if (isPromptfooCloudApiHost(url)) {
 		const token = cloudConfig.getApiKey();
 		opts.headers = {
 			...opts.headers || {},
@@ -705,11 +717,31 @@ async function monkeyPatchFetch(url, options) {
 	}
 }
 //#endregion
+//#region src/util/fetch/retryContext.ts
+const fetchRetryContext = new AsyncLocalStorage();
+/**
+* Run `fn` with a fetch retry context so nested `fetchWithRetries` /
+* `fetchWithProxy` calls inherit the provider's configured `maxRetries`.
+*
+* When `maxRetries` is `undefined`, the new scope deliberately shadows any
+* outer provider context so providers without an override fall back to defaults.
+*/
+function withFetchRetryContext(maxRetries, fn) {
+	return fetchRetryContext.run({ maxRetries }, fn);
+}
+/**
+* Read the active context's `maxRetries`, or `undefined` when none is set.
+*/
+function getFetchRetryContextMaxRetries() {
+	return fetchRetryContext.getStore()?.maxRetries;
+}
+//#endregion
 //#region src/util/fetch/index.ts
 var fetch_exports = /* @__PURE__ */ __exportAll({
 	fetchWithProxy: () => fetchWithProxy,
 	fetchWithRetries: () => fetchWithRetries,
 	fetchWithTimeout: () => fetchWithTimeout,
+	getFetchWithProxyHeaders: () => getFetchWithProxyHeaders,
 	handleRateLimit: () => handleRateLimit,
 	isRateLimited: () => isRateLimited,
 	isTransientError: () => isTransientError
@@ -764,20 +796,44 @@ function getOrCreateProxyAgent(proxyUrl, tlsOptions) {
 	cachedProxyAgents.set(cacheKey, agent);
 	return agent;
 }
+/**
+* Resolve whether to disable transient-error retries. An explicit caller flag
+* always wins (a caller may opt back in even when the provider set
+* `maxRetries: 0`); otherwise we disable only when the retry context carries
+* `maxRetries === 0`.
+*/
+function resolveTransientRetryDisabled(explicit) {
+	if (explicit !== void 0) return explicit;
+	return getFetchRetryContextMaxRetries() === 0;
+}
+function headersInitToRecord(headers) {
+	if (!headers) return {};
+	if (headers instanceof Headers) return Object.fromEntries(headers.entries());
+	if (Array.isArray(headers)) return Object.fromEntries(new Headers(headers).entries());
+	return { ...headers };
+}
+function getFetchWithProxyHeaders(url, options) {
+	const requestHeaders = url instanceof Request && options.headers === void 0 ? headersInitToRecord(url.headers) : {};
+	const optionHeaders = headersInitToRecord(options.headers);
+	return {
+		...requestHeaders,
+		...optionHeaders,
+		"x-promptfoo-version": VERSION
+	};
+}
+function getFetchUrlString(url) {
+	if (typeof url === "string") return url;
+	if (url instanceof URL) return url.toString();
+	if (url instanceof Request) return url.url;
+}
 async function fetchWithProxy(url, options = {}, abortSignal) {
 	let finalUrl = url;
-	let finalUrlString;
-	if (typeof url === "string") finalUrlString = url;
-	else if (url instanceof URL) finalUrlString = url.toString();
-	else if (url instanceof Request) finalUrlString = url.url;
+	let finalUrlString = getFetchUrlString(url);
 	if (!finalUrlString) throw new Error("Invalid URL");
 	const combinedSignal = abortSignal ? options.signal ? AbortSignal.any([options.signal, abortSignal]) : abortSignal : options.signal;
 	const finalOptions = {
 		...options,
-		headers: {
-			...options.headers,
-			"x-promptfoo-version": VERSION
-		},
+		headers: getFetchWithProxyHeaders(url, options),
 		signal: combinedSignal
 	};
 	if (typeof url === "string") try {
@@ -815,10 +871,11 @@ async function fetchWithProxy(url, options = {}, abortSignal) {
 		logger.debug(`Using proxy: ${sanitizeUrl(proxyUrl)}`);
 		finalOptions.dispatcher = getOrCreateProxyAgent(proxyUrl, tlsOptions);
 	} else finalOptions.dispatcher = getOrCreateAgent(tlsOptions);
-	const maxTransientRetries = options.disableTransientRetries ? 0 : 3;
+	const disableTransientRetries = resolveTransientRetryDisabled(options.disableTransientRetries);
+	const maxTransientRetries = disableTransientRetries ? 0 : 3;
 	for (let attempt = 0; attempt <= maxTransientRetries; attempt++) {
 		const response = await monkeyPatchFetch(finalUrl, finalOptions);
-		if (!options.disableTransientRetries && isTransientError(response) && attempt < maxTransientRetries) {
+		if (!disableTransientRetries && isTransientError(response) && attempt < maxTransientRetries) {
 			const backoffMs = Math.pow(2, attempt) * 1e3;
 			logger.debug(`Transient error (${response.status} ${response.statusText}), retry ${attempt + 1}/${maxTransientRetries} after ${backoffMs}ms`);
 			await sleep(backoffMs);
@@ -891,8 +948,17 @@ function isTransientError(response) {
 		default: return false;
 	}
 }
+function formatFetchErrorMessage(error) {
+	if (!(error instanceof Error)) return String(error);
+	const typedError = error;
+	let message = `${typedError.name}: ${typedError.message}`;
+	if (typedError.cause) message += ` (Cause: ${typedError.cause})`;
+	if (typedError.code) message += ` (Code: ${typedError.code})`;
+	return message;
+}
 async function fetchWithRetries(url, options = {}, timeout, maxRetries) {
-	maxRetries = Math.max(0, maxRetries ?? 4);
+	const contextMaxRetries = getFetchRetryContextMaxRetries();
+	maxRetries = Math.max(0, maxRetries ?? contextMaxRetries ?? 4);
 	let lastErrorMessage;
 	const backoff = getEnvInt("PROMPTFOO_REQUEST_BACKOFF_MS", 5e3);
 	for (let i = 0; i <= maxRetries; i++) {
@@ -904,21 +970,19 @@ async function fetchWithRetries(url, options = {}, timeout, maxRetries) {
 			}, timeout);
 			if (getEnvBool("PROMPTFOO_RETRY_5XX") && response.status >= 500 && response.status < 600) throw new Error(`Internal Server Error: ${response.status} ${response.statusText}`);
 			if (response && isRateLimited(response)) {
-				logger.debug(`Rate limited on URL ${url}: ${response.status} ${response.statusText}, attempt ${i + 1}/${maxRetries + 1}, waiting before retry...`);
 				lastErrorMessage = `Rate limited: ${response.status} ${response.statusText}`;
+				if (i >= maxRetries) {
+					logger.debug(`Rate limited on URL ${url}: ${response.status} ${response.statusText}, attempt ${i + 1}/${maxRetries + 1}, no retries remain.`);
+					break;
+				}
+				logger.debug(`Rate limited on URL ${url}: ${response.status} ${response.statusText}, attempt ${i + 1}/${maxRetries + 1}, waiting before retry...`);
 				await handleRateLimit(response);
 				continue;
 			}
 			return response;
 		} catch (error) {
 			if (error instanceof Error && error.name === "AbortError") throw error;
-			let errorMessage;
-			if (error instanceof Error) {
-				const typedError = error;
-				errorMessage = `${typedError.name}: ${typedError.message}`;
-				if (typedError.cause) errorMessage += ` (Cause: ${typedError.cause})`;
-				if (typedError.code) errorMessage += ` (Code: ${typedError.code})`;
-			} else errorMessage = String(error);
+			const errorMessage = formatFetchErrorMessage(error);
 			logger.debug(`Request to ${url} failed (attempt #${i + 1}), retrying: ${errorMessage}`);
 			if (i < maxRetries) await sleep(Math.pow(2, i) * (backoff + 1e3 * Math.random()));
 			lastErrorMessage = errorMessage;
@@ -927,6 +991,6 @@ async function fetchWithRetries(url, options = {}, timeout, maxRetries) {
 	throw new Error(`Request failed after ${maxRetries} retries: ${lastErrorMessage}`);
 }
 //#endregion
-export { TERMINAL_MAX_WIDTH as A, toTitleCase as C, CONSENT_ENDPOINT as D, CLOUD_PROVIDER_PREFIX as E, FILE_METADATA_KEY as F, HUMAN_ASSERTION_TYPE as I, getDefaultShareViewBaseUrl as M, getShareApiBaseUrl as N, EVENTS_ENDPOINT as O, getShareViewBaseUrl as P, parseChatPrompt as S, transformTools as T, isOpenAIToolArray as _, CloudConfig as a, openaiToolChoiceToBedrock as b, writeGlobalConfig as c, sleep as d, parseRateLimitHeaders as f, calculateCost as g, REQUEST_TIMEOUT_MS as h, fetch_exports as i, getDefaultPort as j, R_ENDPOINT as k, writeGlobalConfigPartial as l, LONG_RUNNING_MODEL_TIMEOUT_MS as m, fetchWithRetries as n, cloudConfig as o, parseRetryAfter as p, fetchWithTimeout as r, readGlobalConfig as s, fetchWithProxy as t, getCurrentTimestamp as u, isOpenAIToolChoice as v, transformToolChoice as w, openaiToolsToBedrock as x, isPromptfooSampleTarget as y };
+export { CONSENT_ENDPOINT as A, openaiToolChoiceToBedrock as C, transformToolChoice as D, toTitleCase as E, getDefaultShareViewBaseUrl as F, getShareApiBaseUrl as I, getShareViewBaseUrl as L, R_ENDPOINT as M, TERMINAL_MAX_WIDTH as N, transformTools as O, getDefaultPort as P, FILE_METADATA_KEY as R, isPromptfooSampleTarget as S, parseChatPrompt as T, LONG_RUNNING_MODEL_TIMEOUT_MS as _, getFetchWithProxyHeaders as a, isOpenAIToolArray as b, CloudConfig as c, writeGlobalConfig as d, writeGlobalConfigPartial as f, parseRetryAfter as g, parseRateLimitHeaders as h, fetch_exports as i, EVENTS_ENDPOINT as j, CLOUD_PROVIDER_PREFIX as k, cloudConfig as l, sleep as m, fetchWithRetries as n, withFetchRetryContext as o, getCurrentTimestamp as p, fetchWithTimeout as r, isPromptfooCloudApiHost as s, fetchWithProxy as t, readGlobalConfig as u, REQUEST_TIMEOUT_MS as v, openaiToolsToBedrock as w, isOpenAIToolChoice as x, calculateCost as y, HUMAN_ASSERTION_TYPE as z };
 
-//# sourceMappingURL=fetch-Y5qX_kST.js.map
+//# sourceMappingURL=fetch-BSSAcMxf.js.map

package/dist/src/fetch-BodQTrMU.js

@@ -0,0 +1,3 @@
+#!/usr/bin/env node
+import { n as fetchWithProxy } from "./fetch-Cpf1U1nO.js";
+export { fetchWithProxy };

package/dist/src/{fetch-NuqXW1Xb.cjs → fetch-Cfm4EuOB.cjs}

@@ -1,14 +1,16 @@
-const require_logger = require("./logger-COuQb2xB.cjs");
-const require_invariant = require("./invariant-kfQ8Bu82.cjs");
-const require_version = require("./version-D9zu9FWB.cjs");
+const require_rolldown_runtime = require("./rolldown-runtime-D_mwlA32.cjs");
+const require_logger = require("./logger-cfNpzI4o.cjs");
+const require_invariant = require("./invariant-QtnLD03y.cjs");
+const require_version = require("./version-VzUqOBZk.cjs");
 let fs = require("fs");
-fs = require_logger.__toESM(fs);
+fs = require_rolldown_runtime.__toESM(fs, 1);
 let path = require("path");
-path = require_logger.__toESM(path);
+path = require_rolldown_runtime.__toESM(path, 1);
 let js_yaml = require("js-yaml");
-js_yaml = require_logger.__toESM(js_yaml);
+js_yaml = require_rolldown_runtime.__toESM(js_yaml, 1);
+let node_async_hooks = require("node:async_hooks");
 let fs_promises = require("fs/promises");
-fs_promises = require_logger.__toESM(fs_promises);
+fs_promises = require_rolldown_runtime.__toESM(fs_promises, 1);
 let proxy_from_env = require("proxy-from-env");
 let undici = require("undici");
 let util = require("util");
@@ -74,7 +76,10 @@ function writeGlobalConfigPartial(partialConfig) {
 	});
 	writeGlobalConfig(updatedConfig);
 }
-const API_HOST = require_logger.getEnvString("API_HOST", "https://api.promptfoo.app");
+//#endregion
+//#region src/globalConfig/cloud.ts
+const CLOUD_API_HOST = "https://api.promptfoo.app";
+const API_HOST = require_logger.getEnvString("API_HOST", CLOUD_API_HOST);
 const SHARING_CUTOFF_DATE = /* @__PURE__ */ new Date("2026-03-09T00:00:00Z");
 var CloudConfig = class {
 	config;
@@ -274,8 +279,8 @@ function calculateCost(modelName, config, promptTokens, completionTokens, models
 	const model = models.find((m) => m.id === modelName);
 	if (!model || !model.cost) return;
 	const longContextCost = model.cost.longContext && promptTokens > model.cost.longContext.threshold ? model.cost.longContext : void 0;
-	const inputCost = config.cost ?? longContextCost?.input ?? model.cost.input;
-	const outputCost = config.cost ?? longContextCost?.output ?? model.cost.output;
+	const inputCost = config.inputCost ?? config.cost ?? longContextCost?.input ?? model.cost.input;
+	const outputCost = config.outputCost ?? config.cost ?? longContextCost?.output ?? model.cost.output;
 	return inputCost * promptTokens + outputCost * completionTokens;
 }
 /**
@@ -651,6 +656,14 @@ function isConnectionError(error) {
 /**
 * Enhanced fetch wrapper that adds logging, authentication, error handling, and optional compression
 */
+function isPromptfooCloudApiHost(url) {
+	try {
+		const targetUrl = url instanceof Request ? url.url : url.toString();
+		return new URL(targetUrl).origin === CLOUD_API_HOST;
+	} catch {
+		return false;
+	}
+}
 async function monkeyPatchFetch(url, options) {
 	const NO_LOG_URLS = [
 		R_ENDPOINT,
@@ -670,7 +683,7 @@ async function monkeyPatchFetch(url, options) {
 	} catch (e) {
 		require_logger.logger.warn(`Failed to compress request body: ${e}`);
 	}
-	if (typeof url === "string" && url.startsWith("https://api.promptfoo.app") || url instanceof URL && url.host === "https://api.promptfoo.app".replace(/^https?:\/\//, "")) {
+	if (isPromptfooCloudApiHost(url)) {
 		const token = cloudConfig.getApiKey();
 		opts.headers = {
 			...opts.headers || {},
@@ -704,11 +717,31 @@ async function monkeyPatchFetch(url, options) {
 	}
 }
 //#endregion
+//#region src/util/fetch/retryContext.ts
+const fetchRetryContext = new node_async_hooks.AsyncLocalStorage();
+/**
+* Run `fn` with a fetch retry context so nested `fetchWithRetries` /
+* `fetchWithProxy` calls inherit the provider's configured `maxRetries`.
+*
+* When `maxRetries` is `undefined`, the new scope deliberately shadows any
+* outer provider context so providers without an override fall back to defaults.
+*/
+function withFetchRetryContext(maxRetries, fn) {
+	return fetchRetryContext.run({ maxRetries }, fn);
+}
+/**
+* Read the active context's `maxRetries`, or `undefined` when none is set.
+*/
+function getFetchRetryContextMaxRetries() {
+	return fetchRetryContext.getStore()?.maxRetries;
+}
+//#endregion
 //#region src/util/fetch/index.ts
-var fetch_exports = /* @__PURE__ */ require_logger.__exportAll({
+var fetch_exports = /* @__PURE__ */ require_rolldown_runtime.__exportAll({
 	fetchWithProxy: () => fetchWithProxy,
 	fetchWithRetries: () => fetchWithRetries,
 	fetchWithTimeout: () => fetchWithTimeout,
+	getFetchWithProxyHeaders: () => getFetchWithProxyHeaders,
 	handleRateLimit: () => handleRateLimit,
 	isRateLimited: () => isRateLimited,
 	isTransientError: () => isTransientError
@@ -763,20 +796,44 @@ function getOrCreateProxyAgent(proxyUrl, tlsOptions) {
 	cachedProxyAgents.set(cacheKey, agent);
 	return agent;
 }
+/**
+* Resolve whether to disable transient-error retries. An explicit caller flag
+* always wins (a caller may opt back in even when the provider set
+* `maxRetries: 0`); otherwise we disable only when the retry context carries
+* `maxRetries === 0`.
+*/
+function resolveTransientRetryDisabled(explicit) {
+	if (explicit !== void 0) return explicit;
+	return getFetchRetryContextMaxRetries() === 0;
+}
+function headersInitToRecord(headers) {
+	if (!headers) return {};
+	if (headers instanceof Headers) return Object.fromEntries(headers.entries());
+	if (Array.isArray(headers)) return Object.fromEntries(new Headers(headers).entries());
+	return { ...headers };
+}
+function getFetchWithProxyHeaders(url, options) {
+	const requestHeaders = url instanceof Request && options.headers === void 0 ? headersInitToRecord(url.headers) : {};
+	const optionHeaders = headersInitToRecord(options.headers);
+	return {
+		...requestHeaders,
+		...optionHeaders,
+		"x-promptfoo-version": require_version.VERSION
+	};
+}
+function getFetchUrlString(url) {
+	if (typeof url === "string") return url;
+	if (url instanceof URL) return url.toString();
+	if (url instanceof Request) return url.url;
+}
 async function fetchWithProxy(url, options = {}, abortSignal) {
 	let finalUrl = url;
-	let finalUrlString;
-	if (typeof url === "string") finalUrlString = url;
-	else if (url instanceof URL) finalUrlString = url.toString();
-	else if (url instanceof Request) finalUrlString = url.url;
+	let finalUrlString = getFetchUrlString(url);
 	if (!finalUrlString) throw new Error("Invalid URL");
 	const combinedSignal = abortSignal ? options.signal ? AbortSignal.any([options.signal, abortSignal]) : abortSignal : options.signal;
 	const finalOptions = {
 		...options,
-		headers: {
-			...options.headers,
-			"x-promptfoo-version": require_version.VERSION
-		},
+		headers: getFetchWithProxyHeaders(url, options),
 		signal: combinedSignal
 	};
 	if (typeof url === "string") try {
@@ -814,10 +871,11 @@ async function fetchWithProxy(url, options = {}, abortSignal) {
 		require_logger.logger.debug(`Using proxy: ${require_logger.sanitizeUrl(proxyUrl)}`);
 		finalOptions.dispatcher = getOrCreateProxyAgent(proxyUrl, tlsOptions);
 	} else finalOptions.dispatcher = getOrCreateAgent(tlsOptions);
-	const maxTransientRetries = options.disableTransientRetries ? 0 : 3;
+	const disableTransientRetries = resolveTransientRetryDisabled(options.disableTransientRetries);
+	const maxTransientRetries = disableTransientRetries ? 0 : 3;
 	for (let attempt = 0; attempt <= maxTransientRetries; attempt++) {
 		const response = await monkeyPatchFetch(finalUrl, finalOptions);
-		if (!options.disableTransientRetries && isTransientError(response) && attempt < maxTransientRetries) {
+		if (!disableTransientRetries && isTransientError(response) && attempt < maxTransientRetries) {
 			const backoffMs = Math.pow(2, attempt) * 1e3;
 			require_logger.logger.debug(`Transient error (${response.status} ${response.statusText}), retry ${attempt + 1}/${maxTransientRetries} after ${backoffMs}ms`);
 			await sleep(backoffMs);
@@ -890,8 +948,17 @@ function isTransientError(response) {
 		default: return false;
 	}
 }
+function formatFetchErrorMessage(error) {
+	if (!(error instanceof Error)) return String(error);
+	const typedError = error;
+	let message = `${typedError.name}: ${typedError.message}`;
+	if (typedError.cause) message += ` (Cause: ${typedError.cause})`;
+	if (typedError.code) message += ` (Code: ${typedError.code})`;
+	return message;
+}
 async function fetchWithRetries(url, options = {}, timeout, maxRetries) {
-	maxRetries = Math.max(0, maxRetries ?? 4);
+	const contextMaxRetries = getFetchRetryContextMaxRetries();
+	maxRetries = Math.max(0, maxRetries ?? contextMaxRetries ?? 4);
 	let lastErrorMessage;
 	const backoff = require_logger.getEnvInt("PROMPTFOO_REQUEST_BACKOFF_MS", 5e3);
 	for (let i = 0; i <= maxRetries; i++) {
@@ -903,21 +970,19 @@ async function fetchWithRetries(url, options = {}, timeout, maxRetries) {
 			}, timeout);
 			if (require_logger.getEnvBool("PROMPTFOO_RETRY_5XX") && response.status >= 500 && response.status < 600) throw new Error(`Internal Server Error: ${response.status} ${response.statusText}`);
 			if (response && isRateLimited(response)) {
-				require_logger.logger.debug(`Rate limited on URL ${url}: ${response.status} ${response.statusText}, attempt ${i + 1}/${maxRetries + 1}, waiting before retry...`);
 				lastErrorMessage = `Rate limited: ${response.status} ${response.statusText}`;
+				if (i >= maxRetries) {
+					require_logger.logger.debug(`Rate limited on URL ${url}: ${response.status} ${response.statusText}, attempt ${i + 1}/${maxRetries + 1}, no retries remain.`);
+					break;
+				}
+				require_logger.logger.debug(`Rate limited on URL ${url}: ${response.status} ${response.statusText}, attempt ${i + 1}/${maxRetries + 1}, waiting before retry...`);
 				await handleRateLimit(response);
 				continue;
 			}
 			return response;
 		} catch (error) {
 			if (error instanceof Error && error.name === "AbortError") throw error;
-			let errorMessage;
-			if (error instanceof Error) {
-				const typedError = error;
-				errorMessage = `${typedError.name}: ${typedError.message}`;
-				if (typedError.cause) errorMessage += ` (Cause: ${typedError.cause})`;
-				if (typedError.code) errorMessage += ` (Code: ${typedError.code})`;
-			} else errorMessage = String(error);
+			const errorMessage = formatFetchErrorMessage(error);
 			require_logger.logger.debug(`Request to ${url} failed (attempt #${i + 1}), retrying: ${errorMessage}`);
 			if (i < maxRetries) await sleep(Math.pow(2, i) * (backoff + 1e3 * Math.random()));
 			lastErrorMessage = errorMessage;
@@ -1034,6 +1099,12 @@ Object.defineProperty(exports, "getDefaultShareViewBaseUrl", {
 		return getDefaultShareViewBaseUrl;
 	}
 });
+Object.defineProperty(exports, "getFetchWithProxyHeaders", {
+	enumerable: true,
+	get: function() {
+		return getFetchWithProxyHeaders;
+	}
+});
 Object.defineProperty(exports, "getShareApiBaseUrl", {
 	enumerable: true,
 	get: function() {
@@ -1058,6 +1129,12 @@ Object.defineProperty(exports, "isOpenAIToolChoice", {
 		return isOpenAIToolChoice;
 	}
 });
+Object.defineProperty(exports, "isPromptfooCloudApiHost", {
+	enumerable: true,
+	get: function() {
+		return isPromptfooCloudApiHost;
+	}
+});
 Object.defineProperty(exports, "isPromptfooSampleTarget", {
 	enumerable: true,
 	get: function() {
@@ -1124,6 +1201,12 @@ Object.defineProperty(exports, "transformTools", {
 		return transformTools;
 	}
 });
+Object.defineProperty(exports, "withFetchRetryContext", {
+	enumerable: true,
+	get: function() {
+		return withFetchRetryContext;
+	}
+});
 Object.defineProperty(exports, "writeGlobalConfig", {
 	enumerable: true,
 	get: function() {
@@ -1137,4 +1220,4 @@ Object.defineProperty(exports, "writeGlobalConfigPartial", {
 	}
 });
 
-//# sourceMappingURL=fetch-NuqXW1Xb.cjs.map
+//# sourceMappingURL=fetch-Cfm4EuOB.cjs.map

package/dist/src/{fetch-B6ch2nU2.js → fetch-Cpf1U1nO.js}

@@ -1,8 +1,9 @@
 #!/usr/bin/env node
 import { D as getEnvInt, N as state, O as getEnvString, T as getEnvBool, h as sanitizeUrl, o as logRequestResponse, s as logger } from "./logger-BbY6ypFL.js";
-import { n as VERSION } from "./version-CbpiUINz.js";
+import { n as VERSION } from "./version-F0YDgb7J.js";
 import { t as invariant } from "./invariant-B2Rf6avk.js";
-import { a as cloudConfig } from "./cloud-Da0bofJd.js";
+import { a as cloudConfig, n as CLOUD_API_HOST } from "./cloud-DqF5N1aJ.js";
+import { AsyncLocalStorage } from "node:async_hooks";
 import path from "path";
 import yaml from "js-yaml";
 import * as fsPromises from "fs/promises";
@@ -60,8 +61,8 @@ function calculateCost(modelName, config, promptTokens, completionTokens, models
 	const model = models.find((m) => m.id === modelName);
 	if (!model || !model.cost) return;
 	const longContextCost = model.cost.longContext && promptTokens > model.cost.longContext.threshold ? model.cost.longContext : void 0;
-	const inputCost = config.cost ?? longContextCost?.input ?? model.cost.input;
-	const outputCost = config.cost ?? longContextCost?.output ?? model.cost.output;
+	const inputCost = config.inputCost ?? config.cost ?? longContextCost?.input ?? model.cost.input;
+	const outputCost = config.outputCost ?? config.cost ?? longContextCost?.output ?? model.cost.output;
 	return inputCost * promptTokens + outputCost * completionTokens;
 }
 /**
@@ -455,6 +456,14 @@ function isConnectionError(error) {
 /**
 * Enhanced fetch wrapper that adds logging, authentication, error handling, and optional compression
 */
+function isPromptfooCloudApiHost(url) {
+	try {
+		const targetUrl = url instanceof Request ? url.url : url.toString();
+		return new URL(targetUrl).origin === CLOUD_API_HOST;
+	} catch {
+		return false;
+	}
+}
 async function monkeyPatchFetch(url, options) {
 	const NO_LOG_URLS = [
 		R_ENDPOINT,
@@ -474,7 +483,7 @@ async function monkeyPatchFetch(url, options) {
 	} catch (e) {
 		logger.warn(`Failed to compress request body: ${e}`);
 	}
-	if (typeof url === "string" && url.startsWith("https://api.promptfoo.app") || url instanceof URL && url.host === "https://api.promptfoo.app".replace(/^https?:\/\//, "")) {
+	if (isPromptfooCloudApiHost(url)) {
 		const token = cloudConfig.getApiKey();
 		opts.headers = {
 			...opts.headers || {},
@@ -508,6 +517,25 @@ async function monkeyPatchFetch(url, options) {
 	}
 }
 //#endregion
+//#region src/util/fetch/retryContext.ts
+const fetchRetryContext = new AsyncLocalStorage();
+/**
+* Run `fn` with a fetch retry context so nested `fetchWithRetries` /
+* `fetchWithProxy` calls inherit the provider's configured `maxRetries`.
+*
+* When `maxRetries` is `undefined`, the new scope deliberately shadows any
+* outer provider context so providers without an override fall back to defaults.
+*/
+function withFetchRetryContext(maxRetries, fn) {
+	return fetchRetryContext.run({ maxRetries }, fn);
+}
+/**
+* Read the active context's `maxRetries`, or `undefined` when none is set.
+*/
+function getFetchRetryContextMaxRetries() {
+	return fetchRetryContext.getStore()?.maxRetries;
+}
+//#endregion
 //#region src/util/fetch/index.ts
 const cachedAgents = /* @__PURE__ */ new Map();
 const cachedProxyAgents = /* @__PURE__ */ new Map();
@@ -569,20 +597,44 @@ function getOrCreateProxyAgent(proxyUrl, tlsOptions) {
 	cachedProxyAgents.set(cacheKey, agent);
 	return agent;
 }
+/**
+* Resolve whether to disable transient-error retries. An explicit caller flag
+* always wins (a caller may opt back in even when the provider set
+* `maxRetries: 0`); otherwise we disable only when the retry context carries
+* `maxRetries === 0`.
+*/
+function resolveTransientRetryDisabled(explicit) {
+	if (explicit !== void 0) return explicit;
+	return getFetchRetryContextMaxRetries() === 0;
+}
+function headersInitToRecord(headers) {
+	if (!headers) return {};
+	if (headers instanceof Headers) return Object.fromEntries(headers.entries());
+	if (Array.isArray(headers)) return Object.fromEntries(new Headers(headers).entries());
+	return { ...headers };
+}
+function getFetchWithProxyHeaders(url, options) {
+	const requestHeaders = url instanceof Request && options.headers === void 0 ? headersInitToRecord(url.headers) : {};
+	const optionHeaders = headersInitToRecord(options.headers);
+	return {
+		...requestHeaders,
+		...optionHeaders,
+		"x-promptfoo-version": VERSION
+	};
+}
+function getFetchUrlString(url) {
+	if (typeof url === "string") return url;
+	if (url instanceof URL) return url.toString();
+	if (url instanceof Request) return url.url;
+}
 async function fetchWithProxy(url, options = {}, abortSignal) {
 	let finalUrl = url;
-	let finalUrlString;
-	if (typeof url === "string") finalUrlString = url;
-	else if (url instanceof URL) finalUrlString = url.toString();
-	else if (url instanceof Request) finalUrlString = url.url;
+	let finalUrlString = getFetchUrlString(url);
 	if (!finalUrlString) throw new Error("Invalid URL");
 	const combinedSignal = abortSignal ? options.signal ? AbortSignal.any([options.signal, abortSignal]) : abortSignal : options.signal;
 	const finalOptions = {
 		...options,
-		headers: {
-			...options.headers,
-			"x-promptfoo-version": VERSION
-		},
+		headers: getFetchWithProxyHeaders(url, options),
 		signal: combinedSignal
 	};
 	if (typeof url === "string") try {
@@ -620,10 +672,11 @@ async function fetchWithProxy(url, options = {}, abortSignal) {
 		logger.debug(`Using proxy: ${sanitizeUrl(proxyUrl)}`);
 		finalOptions.dispatcher = getOrCreateProxyAgent(proxyUrl, tlsOptions);
 	} else finalOptions.dispatcher = getOrCreateAgent(tlsOptions);
-	const maxTransientRetries = options.disableTransientRetries ? 0 : 3;
+	const disableTransientRetries = resolveTransientRetryDisabled(options.disableTransientRetries);
+	const maxTransientRetries = disableTransientRetries ? 0 : 3;
 	for (let attempt = 0; attempt <= maxTransientRetries; attempt++) {
 		const response = await monkeyPatchFetch(finalUrl, finalOptions);
-		if (!options.disableTransientRetries && isTransientError(response) && attempt < maxTransientRetries) {
+		if (!disableTransientRetries && isTransientError(response) && attempt < maxTransientRetries) {
 			const backoffMs = Math.pow(2, attempt) * 1e3;
 			logger.debug(`Transient error (${response.status} ${response.statusText}), retry ${attempt + 1}/${maxTransientRetries} after ${backoffMs}ms`);
 			await sleep(backoffMs);
@@ -696,8 +749,17 @@ function isTransientError(response) {
 		default: return false;
 	}
 }
+function formatFetchErrorMessage(error) {
+	if (!(error instanceof Error)) return String(error);
+	const typedError = error;
+	let message = `${typedError.name}: ${typedError.message}`;
+	if (typedError.cause) message += ` (Cause: ${typedError.cause})`;
+	if (typedError.code) message += ` (Code: ${typedError.code})`;
+	return message;
+}
 async function fetchWithRetries(url, options = {}, timeout, maxRetries) {
-	maxRetries = Math.max(0, maxRetries ?? 4);
+	const contextMaxRetries = getFetchRetryContextMaxRetries();
+	maxRetries = Math.max(0, maxRetries ?? contextMaxRetries ?? 4);
 	let lastErrorMessage;
 	const backoff = getEnvInt("PROMPTFOO_REQUEST_BACKOFF_MS", 5e3);
 	for (let i = 0; i <= maxRetries; i++) {
@@ -709,21 +771,19 @@ async function fetchWithRetries(url, options = {}, timeout, maxRetries) {
 			}, timeout);
 			if (getEnvBool("PROMPTFOO_RETRY_5XX") && response.status >= 500 && response.status < 600) throw new Error(`Internal Server Error: ${response.status} ${response.statusText}`);
 			if (response && isRateLimited(response)) {
-				logger.debug(`Rate limited on URL ${url}: ${response.status} ${response.statusText}, attempt ${i + 1}/${maxRetries + 1}, waiting before retry...`);
 				lastErrorMessage = `Rate limited: ${response.status} ${response.statusText}`;
+				if (i >= maxRetries) {
+					logger.debug(`Rate limited on URL ${url}: ${response.status} ${response.statusText}, attempt ${i + 1}/${maxRetries + 1}, no retries remain.`);
+					break;
+				}
+				logger.debug(`Rate limited on URL ${url}: ${response.status} ${response.statusText}, attempt ${i + 1}/${maxRetries + 1}, waiting before retry...`);
 				await handleRateLimit(response);
 				continue;
 			}
 			return response;
 		} catch (error) {
 			if (error instanceof Error && error.name === "AbortError") throw error;
-			let errorMessage;
-			if (error instanceof Error) {
-				const typedError = error;
-				errorMessage = `${typedError.name}: ${typedError.message}`;
-				if (typedError.cause) errorMessage += ` (Cause: ${typedError.cause})`;
-				if (typedError.code) errorMessage += ` (Code: ${typedError.code})`;
-			} else errorMessage = String(error);
+			const errorMessage = formatFetchErrorMessage(error);
 			logger.debug(`Request to ${url} failed (attempt #${i + 1}), retrying: ${errorMessage}`);
 			if (i < maxRetries) await sleep(Math.pow(2, i) * (backoff + 1e3 * Math.random()));
 			lastErrorMessage = errorMessage;
@@ -732,6 +792,6 @@ async function fetchWithRetries(url, options = {}, timeout, maxRetries) {
 	throw new Error(`Request failed after ${maxRetries} retries: ${lastErrorMessage}`);
 }
 //#endregion
-export { getDefaultPort as A, transformToolChoice as C, EVENTS_ENDPOINT as D, CONSENT_ENDPOINT as E, HUMAN_ASSERTION_TYPE as F, getShareApiBaseUrl as M, getShareViewBaseUrl as N, R_ENDPOINT as O, FILE_METADATA_KEY as P, toTitleCase as S, CLOUD_PROVIDER_PREFIX as T, isOpenAIToolChoice as _, handleRateLimit as a, openaiToolsToBedrock as b, getCurrentTimestamp as c, parseRateLimitHeaders as d, parseRetryAfter as f, isOpenAIToolArray as g, calculateCost as h, fetchWithTimeout as i, getDefaultShareViewBaseUrl as j, TERMINAL_MAX_WIDTH as k, sleep as l, REQUEST_TIMEOUT_MS as m, fetchWithProxy as n, isRateLimited as o, LONG_RUNNING_MODEL_TIMEOUT_MS as p, fetchWithRetries as r, isTransientError as s, clearAgentCache as t, sleepWithAbort as u, isPromptfooSampleTarget as v, transformTools as w, parseChatPrompt as x, openaiToolChoiceToBedrock as y };
+export { EVENTS_ENDPOINT as A, openaiToolsToBedrock as C, transformTools as D, transformToolChoice as E, getShareApiBaseUrl as F, getShareViewBaseUrl as I, FILE_METADATA_KEY as L, TERMINAL_MAX_WIDTH as M, getDefaultPort as N, CLOUD_PROVIDER_PREFIX as O, getDefaultShareViewBaseUrl as P, HUMAN_ASSERTION_TYPE as R, openaiToolChoiceToBedrock as S, toTitleCase as T, REQUEST_TIMEOUT_MS as _, getFetchWithProxyHeaders as a, isOpenAIToolChoice as b, isTransientError as c, getCurrentTimestamp as d, sleep as f, LONG_RUNNING_MODEL_TIMEOUT_MS as g, parseRetryAfter as h, fetchWithTimeout as i, R_ENDPOINT as j, CONSENT_ENDPOINT as k, withFetchRetryContext as l, parseRateLimitHeaders as m, fetchWithProxy as n, handleRateLimit as o, sleepWithAbort as p, fetchWithRetries as r, isRateLimited as s, clearAgentCache as t, isPromptfooCloudApiHost as u, calculateCost as v, parseChatPrompt as w, isPromptfooSampleTarget as x, isOpenAIToolArray as y };
 
-//# sourceMappingURL=fetch-B6ch2nU2.js.map
+//# sourceMappingURL=fetch-Cpf1U1nO.js.map