promptfoo 0.121.8 → 0.121.9

package/dist/src/{util-NYQvo1C7.cjs → util-Cs372i1_.cjs}

@@ -1,12 +1,12 @@
 const require_rolldown_runtime = require("./rolldown-runtime-D_mwlA32.cjs");
-const require_logger = require("./logger-cfNpzI4o.cjs");
-const require_fetch = require("./fetch-Cfm4EuOB.cjs");
-const require_util = require("./util-COnzevXE.cjs");
+const require_logger = require("./logger-Cm0Nb8YX.cjs");
+const require_fetch = require("./fetch-B2HvaVUc.cjs");
+const require_util = require("./util-BuYzkbEz.cjs");
 let openai = require("openai");
 openai = require_rolldown_runtime.__toESM(openai, 1);
 //#region src/providers/openai/util.ts
 const ajv = require_logger.getAjv();
-const GPT_5_4_LONG_CONTEXT_THRESHOLD = 272e3;
+const GPT_5_LONG_CONTEXT_THRESHOLD = 272e3;
 const OPENAI_CHAT_MODELS = [
 	...["gpt-4o-mini-tts", "gpt-4o-mini-tts-2025-12-15"].map((model) => ({
 		id: model,
@@ -358,13 +358,25 @@ const OPENAI_CHAT_MODELS = [
 			output: 4 / 1e6
 		}
 	})),
+	...["gpt-5.5", "gpt-5.5-2026-04-23"].map((model) => ({
+		id: model,
+		cost: {
+			input: 5 / 1e6,
+			output: 30 / 1e6,
+			longContext: {
+				threshold: GPT_5_LONG_CONTEXT_THRESHOLD,
+				input: 10 / 1e6,
+				output: 45 / 1e6
+			}
+		}
+	})),
 	...["gpt-5.4", "gpt-5.4-2026-03-05"].map((model) => ({
 		id: model,
 		cost: {
 			input: 2.5 / 1e6,
 			output: 15 / 1e6,
 			longContext: {
-				threshold: GPT_5_4_LONG_CONTEXT_THRESHOLD,
+				threshold: GPT_5_LONG_CONTEXT_THRESHOLD,
 				input: 5 / 1e6,
 				output: 22.5 / 1e6
 			}
@@ -384,13 +396,6 @@ const OPENAI_CHAT_MODELS = [
 			output: 1.25 / 1e6
 		}
 	})),
-	...["gpt-5.5", "gpt-5.5-2026-04-23"].map((model) => ({
-		id: model,
-		cost: {
-			input: 5 / 1e6,
-			output: 30 / 1e6
-		}
-	})),
 	...["gpt-audio", "gpt-audio-2025-08-28"].map((model) => ({
 		id: model,
 		cost: {
@@ -429,7 +434,7 @@ const OPENAI_RESPONSES_ONLY_MODELS = [...["gpt-5.4-pro", "gpt-5.4-pro-2026-03-05
 		input: 30 / 1e6,
 		output: 180 / 1e6,
 		longContext: {
-			threshold: GPT_5_4_LONG_CONTEXT_THRESHOLD,
+			threshold: GPT_5_LONG_CONTEXT_THRESHOLD,
 			input: 60 / 1e6,
 			output: 270 / 1e6
 		}
@@ -702,4 +707,4 @@ Object.defineProperty(exports, "validateFunctionCall", {
 	}
 });
 
-//# sourceMappingURL=util-NYQvo1C7.cjs.map
+//# sourceMappingURL=util-Cs372i1_.cjs.map

package/dist/src/{util-X4KQgyVD.js → util-Cw9nJJWm.js}

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
-import { v as calculateCost } from "./fetch-Cpf1U1nO.js";
-import { w as parseDataUrl } from "./transform-DtTfiGoh.js";
+import { _ as calculateCost } from "./fetch-C59QM4cf.js";
+import { w as parseDataUrl } from "./transform-AfC8gqCB.js";
 //#region src/providers/anthropic/util.ts
 const ANTHROPIC_MODELS = [
 	...["claude-mythos-preview"].map((model) => ({
@@ -333,6 +333,7 @@ function processAnthropicTools(tools = []) {
 		else if (tool.type === "web_fetch_20260309") processedTools.push(transformWebFetchToolV2(tool));
 		else if (tool.type === "web_search_20250305") processedTools.push(transformWebSearchTool(tool));
 		else if (tool.type === "web_search_20260209") processedTools.push(transformWebSearchTool20260209(tool));
+		else if (tool.type === "memory_20250818") processedTools.push(tool);
 		else processedTools.push(tool);
 		else processedTools.push(tool);
 		if ("strict" in tool && tool.strict === true) addRequiredBetaFeature("structured-outputs-2025-11-13");
@@ -413,4 +414,4 @@ function transformWebSearchTool20260209(config) {
 //#endregion
 export { isClaudeOpus47Model as a, processAnthropicTools as c, getTokenUsage as i, calculateAnthropicCost as n, outputFromMessage as o, getRefusalDetails as r, parseMessages as s, ANTHROPIC_MODELS as t };
 
-//# sourceMappingURL=util-X4KQgyVD.js.map
+//# sourceMappingURL=util-Cw9nJJWm.js.map

package/dist/src/{util-DpV6KT5i.js → util-D1NKMz55.js}

@@ -1,10 +1,10 @@
-import { a as logger, y as safeJsonStringify } from "./logger-Ct2S6Yx-.js";
+import { a as logger, y as safeJsonStringify } from "./logger-Cs0I2v_y.js";
 import { H as DATASET_PLUGINS, Q as pluginDescriptions } from "./types-D6glLbdF.js";
-import { p as REQUEST_TIMEOUT_MS } from "./fetch-Doks14zQ.js";
-import { r as getRemoteGenerationUrl, s as neverGenerateRemote } from "./remoteGeneration-CJC3E0aW.js";
-import { r as fetchWithCache } from "./cache-DdriHsNX.js";
+import { m as getRequestTimeoutMs } from "./fetch-CHJu2KON.js";
+import { r as getRemoteGenerationUrl, s as neverGenerateRemote } from "./remoteGeneration-DvWEm3yi.js";
+import { r as fetchWithCache } from "./cache-CT5D7d_A.js";
 import { n as escapeRegExp } from "./text-DwYK5EBn.js";
-import { n as materializeInputVariablesWithMetadata } from "./inputVariables-DUGMb9Ka.js";
+import { n as materializeInputVariablesWithMetadata } from "./inputVariables-B7cylqzM.js";
 //#region src/redteam/util.ts
 /**
 * Regex pattern for matching <Prompt> tags in multi-input redteam generation output.
@@ -292,7 +292,7 @@ async function extractGoalFromPrompt(prompt, purpose, pluginId, policy) {
 			method: "POST",
 			headers: { "Content-Type": "application/json" },
 			body: JSON.stringify(requestBody)
-		}, REQUEST_TIMEOUT_MS);
+		}, getRequestTimeoutMs());
 		logger.debug(`Goal extraction response - Status: ${status} ${statusText || ""}, Data: ${JSON.stringify(data)}`);
 		if (status !== 200) {
 			logger.warn(`Failed to extract goal from prompt: HTTP ${status} ${statusText || ""}, Response Data: ${JSON.stringify(data)}`);
@@ -324,4 +324,4 @@ function getSessionId(response, context) {
 //#endregion
 export { extractPromptFromTags as a, isBasicRefusal as c, extractMaterializedVariablesFromJsonWithMetadata as i, isEmptyResponse as l, extractGoalFromPrompt as n, getSessionId as o, extractInputVarsFromPrompt as r, getShortPluginId as s, extractAllPromptsFromTags as t, removePrefix as u };
 
-//# sourceMappingURL=util-DpV6KT5i.js.map
+//# sourceMappingURL=util-D1NKMz55.js.map

package/dist/src/{util-ETfU_sS9.js → util-D9grslJP.js}

@@ -1,11 +1,11 @@
 #!/usr/bin/env node
-import { s as logger, x as safeJsonStringify } from "./logger-BbY6ypFL.js";
-import { _ as REQUEST_TIMEOUT_MS } from "./fetch-Cpf1U1nO.js";
-import { A as DATASET_PLUGINS, V as pluginDescriptions } from "./types-BFevViUY.js";
-import { r as getRemoteGenerationUrl, s as neverGenerateRemote } from "./remoteGeneration-D6UjE2JT.js";
-import { a as fetchWithCache } from "./cache-BR77mdIR.js";
+import { s as logger, x as safeJsonStringify } from "./logger-h-eiS4iv.js";
+import { v as getRequestTimeoutMs } from "./fetch-C59QM4cf.js";
+import { A as DATASET_PLUGINS, V as pluginDescriptions } from "./types-C_e_SykX.js";
+import { r as getRemoteGenerationUrl, s as neverGenerateRemote } from "./remoteGeneration-DW6mPZnO.js";
+import { a as fetchWithCache } from "./cache-BuudGL4I.js";
 import { n as escapeRegExp } from "./text-Db-Wt2u2.js";
-import { r as materializeInputVariablesWithMetadata } from "./inputVariables-DXFdi7AI.js";
+import { r as materializeInputVariablesWithMetadata } from "./inputVariables-qJmuSN5i.js";
 //#region src/redteam/util.ts
 /**
 * Regex pattern for matching <Prompt> tags in multi-input redteam generation output.
@@ -293,7 +293,7 @@ async function extractGoalFromPrompt(prompt, purpose, pluginId, policy) {
 			method: "POST",
 			headers: { "Content-Type": "application/json" },
 			body: JSON.stringify(requestBody)
-		}, REQUEST_TIMEOUT_MS);
+		}, getRequestTimeoutMs());
 		logger.debug(`Goal extraction response - Status: ${status} ${statusText || ""}, Data: ${JSON.stringify(data)}`);
 		if (status !== 200) {
 			logger.warn(`Failed to extract goal from prompt: HTTP ${status} ${statusText || ""}, Response Data: ${JSON.stringify(data)}`);
@@ -325,4 +325,4 @@ function getSessionId(response, context) {
 //#endregion
 export { extractPromptFromTags as a, isBasicRefusal as c, extractMaterializedVariablesFromJsonWithMetadata as i, isEmptyResponse as l, extractGoalFromPrompt as n, getSessionId as o, extractInputVarsFromPrompt as r, getShortPluginId as s, extractAllPromptsFromTags as t, removePrefix as u };
 
-//# sourceMappingURL=util-ETfU_sS9.js.map
+//# sourceMappingURL=util-D9grslJP.js.map

package/dist/src/{util-CMrHV35u.js → util-SNUDkWJ8.js}

@@ -1,5 +1,5 @@
-import { m as calculateCost } from "./fetch-Doks14zQ.js";
-import { w as parseDataUrl } from "./transform-Wp6s_5QE.js";
+import { p as calculateCost } from "./fetch-CHJu2KON.js";
+import { w as parseDataUrl } from "./transform-R1PnNdLv.js";
 //#region src/providers/anthropic/util.ts
 const ANTHROPIC_MODELS = [
 	...["claude-mythos-preview"].map((model) => ({
@@ -332,6 +332,7 @@ function processAnthropicTools(tools = []) {
 		else if (tool.type === "web_fetch_20260309") processedTools.push(transformWebFetchToolV2(tool));
 		else if (tool.type === "web_search_20250305") processedTools.push(transformWebSearchTool(tool));
 		else if (tool.type === "web_search_20260209") processedTools.push(transformWebSearchTool20260209(tool));
+		else if (tool.type === "memory_20250818") processedTools.push(tool);
 		else processedTools.push(tool);
 		else processedTools.push(tool);
 		if ("strict" in tool && tool.strict === true) addRequiredBetaFeature("structured-outputs-2025-11-13");
@@ -412,4 +413,4 @@ function transformWebSearchTool20260209(config) {
 //#endregion
 export { isClaudeOpus47Model as a, processAnthropicTools as c, getTokenUsage as i, calculateAnthropicCost as n, outputFromMessage as o, getRefusalDetails as r, parseMessages as s, ANTHROPIC_MODELS as t };
 
-//# sourceMappingURL=util-CMrHV35u.js.map
+//# sourceMappingURL=util-SNUDkWJ8.js.map

package/dist/src/{util-DRfqa4xz.js → util-vaCPFatZ.js}

@@ -1,12 +1,12 @@
-import { S as getEnvBool, a as logger, f as sanitizeObject, k as state } from "./logger-Ct2S6Yx-.js";
+import { S as getEnvBool, a as logger, f as sanitizeObject, k as state } from "./logger-Cs0I2v_y.js";
 import { t as invariant } from "./invariant-Ddh24eXh.js";
 import { Gt as isProviderOptions, Wt as isApiProvider, g as ResultFailureReason, h as OutputFileExtension } from "./types-D6glLbdF.js";
-import { N as TERMINAL_MAX_WIDTH, t as fetchWithProxy } from "./fetch-Doks14zQ.js";
-import { n as VERSION } from "./version-qVEN5qCm.js";
+import { N as TERMINAL_MAX_WIDTH, t as fetchWithProxy } from "./fetch-CHJu2KON.js";
+import { n as VERSION } from "./version-BNhpIb2U.js";
 import { i as isJavascriptFile, t as JAVASCRIPT_EXTENSIONS } from "./fileExtensions-CXRfY3Ss.js";
-import { o as safeResolve, r as importModule, t as getDirectory } from "./esm-BTK1W7lG.js";
-import { n as renderVarsInObject, o as getNunjucksEngine } from "./render-DFfDeYUK.js";
-import { r as runPython } from "./pythonUtils-D0BYebvX.js";
+import { o as safeResolve, r as importModule, t as getDirectory } from "./esm-C_yrt93A.js";
+import { n as renderVarsInObject, o as getNunjucksEngine } from "./render-CTZQAINJ.js";
+import { r as runPython } from "./pythonUtils-C4Wnz6Q7.js";
 import * as fs$2 from "fs";
 import fs from "fs";
 import * as path$1 from "path";
@@ -1307,4 +1307,4 @@ function printBorder() {
 //#endregion
 export { readFilters as A, getResolvedRelativePath as C, maybeLoadResponseFormatFromExternalFile as D, maybeLoadFromExternalFileWithVars as E, loadFunction as M, parseFileUrl as N, maybeLoadToolsFromExternalFile as O, getNunjucksEngineForFilePath as S, maybeLoadFromExternalFile as T, isProviderAllowed as _, setupEnv as a, normalizeProviderRef as b, filterRuntimeVars as c, checkProviderApiKeys as d, doesProviderRefMatch as f, isOpenAiProvider as g, isGoogleProvider as h, fetchCsvFromGoogleSheet as i, readOutput as j, parsePathOrGlob as k, getTestCaseDeduplicationKey as l, isAnthropicProvider as m, writeMultipleOutputs as n, deduplicateTestCases as o, getProviderDescription as p, writeOutput as r, extractRuntimeVars as s, printBorder as t, resultIsForTestCase as u, isProviderConfigFileReference as v, maybeLoadConfigFromExternalFile as w, readProviderConfigFile as x, loadProviderConfigsFromFile as y };
 
-//# sourceMappingURL=util-DRfqa4xz.js.map
+//# sourceMappingURL=util-vaCPFatZ.js.map

package/dist/src/{util-DTq3jq2z.js → util-weQgMdSy.js}

@@ -1,5 +1,5 @@
-import { y as calculateCost } from "./fetch-BSSAcMxf.js";
-import { w as parseDataUrl } from "./transform-BufxPIQL.js";
+import { v as calculateCost } from "./fetch-CBe8pU8Q.js";
+import { w as parseDataUrl } from "./transform-D7QlmswG.js";
 //#region src/providers/anthropic/util.ts
 const ANTHROPIC_MODELS = [
 	...["claude-mythos-preview"].map((model) => ({
@@ -332,6 +332,7 @@ function processAnthropicTools(tools = []) {
 		else if (tool.type === "web_fetch_20260309") processedTools.push(transformWebFetchToolV2(tool));
 		else if (tool.type === "web_search_20250305") processedTools.push(transformWebSearchTool(tool));
 		else if (tool.type === "web_search_20260209") processedTools.push(transformWebSearchTool20260209(tool));
+		else if (tool.type === "memory_20250818") processedTools.push(tool);
 		else processedTools.push(tool);
 		else processedTools.push(tool);
 		if ("strict" in tool && tool.strict === true) addRequiredBetaFeature("structured-outputs-2025-11-13");
@@ -412,4 +413,4 @@ function transformWebSearchTool20260209(config) {
 //#endregion
 export { isClaudeOpus47Model as a, processAnthropicTools as c, getTokenUsage as i, calculateAnthropicCost as n, outputFromMessage as o, getRefusalDetails as r, parseMessages as s, ANTHROPIC_MODELS as t };
 
-//# sourceMappingURL=util-DTq3jq2z.js.map
+//# sourceMappingURL=util-weQgMdSy.js.map

package/dist/src/{utils-Ve6kuJsa.cjs → utils-B_HCobst.cjs}

@@ -1,4 +1,4 @@
-const require_logger = require("./logger-cfNpzI4o.cjs");
+const require_logger = require("./logger-Cm0Nb8YX.cjs");
 const require_createHash = require("./createHash-CSiqnK5P.cjs");
 //#region src/prompts/constants.ts
 const PROMPT_DELIMITER = require_logger.getEnvString("PROMPTFOO_PROMPT_SEPARATOR") || "---";
@@ -122,4 +122,4 @@ Object.defineProperty(exports, "normalizeInput", {
 	}
 });
 
-//# sourceMappingURL=utils-Ve6kuJsa.cjs.map
+//# sourceMappingURL=utils-B_HCobst.cjs.map

package/dist/src/{utils-BGY69tk_.js → utils-C3qMBhJT.js}

@@ -1,4 +1,4 @@
-import { T as getEnvString } from "./logger-Ct2S6Yx-.js";
+import { T as getEnvString } from "./logger-Cs0I2v_y.js";
 import { n as sha256 } from "./createHash-BtbSX3mj.js";
 //#region src/prompts/constants.ts
 const PROMPT_DELIMITER = getEnvString("PROMPTFOO_PROMPT_SEPARATOR") || "---";
@@ -93,4 +93,4 @@ function hashPrompt(prompt) {
 //#endregion
 export { PROMPT_DELIMITER as a, generateIdFromPrompt as i, maybeFilePath as n, normalizeInput as r, hashPrompt as t };
 
-//# sourceMappingURL=utils-BGY69tk_.js.map
+//# sourceMappingURL=utils-C3qMBhJT.js.map

package/dist/src/{utils-BFOh20Gb.js → utils-Cdx3ZJr7.js}

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { O as getEnvString } from "./logger-BbY6ypFL.js";
+import { O as getEnvString } from "./logger-h-eiS4iv.js";
 import { n as sha256 } from "./createHash-CgRvs4Fn.js";
 //#region src/prompts/constants.ts
 const PROMPT_DELIMITER = getEnvString("PROMPTFOO_PROMPT_SEPARATOR") || "---";
@@ -94,4 +94,4 @@ function hashPrompt(prompt) {
 //#endregion
 export { PROMPT_DELIMITER as a, generateIdFromPrompt as i, maybeFilePath as n, normalizeInput as r, hashPrompt as t };
 
-//# sourceMappingURL=utils-BFOh20Gb.js.map
+//# sourceMappingURL=utils-Cdx3ZJr7.js.map

package/dist/src/{utils-B0lzitHZ.js → utils-DMWnIu_V.js}

@@ -1,4 +1,4 @@
-import { T as getEnvString } from "./logger-KD8JjCRJ.js";
+import { E as getEnvString } from "./logger-BjYLHvGm.js";
 import { n as sha256 } from "./createHash-CGVzWdjj.js";
 //#region src/models/prompt.ts
 /**
@@ -93,4 +93,4 @@ function hashPrompt(prompt) {
 //#endregion
 export { generateIdFromPrompt as a, PROMPT_DELIMITER as i, maybeFilePath as n, normalizeInput as r, hashPrompt as t };
 
-//# sourceMappingURL=utils-B0lzitHZ.js.map
+//# sourceMappingURL=utils-DMWnIu_V.js.map

package/dist/src/{version-CHR-EFec.js → version-BGiEZvX0.js}

@@ -3,7 +3,7 @@
 * Application version from package.json.
 * Injected at build time, or read from npm environment in development.
 */
-const VERSION = "0.121.8";
+const VERSION = "0.121.9";
 /**
 * PostHog analytics key.
 * Only populated during production builds via PROMPTFOO_POSTHOG_KEY env var.
@@ -13,4 +13,4 @@ const POSTHOG_KEY = "phc_E5n5uHnDo2eREJL1uqX1cIlbkoRby4yFWt3V94HqRRg";
 //#endregion
 export { VERSION as n, POSTHOG_KEY as t };
 
-//# sourceMappingURL=version-CHR-EFec.js.map
+//# sourceMappingURL=version-BGiEZvX0.js.map

package/dist/src/{version-qVEN5qCm.js → version-BNhpIb2U.js}

@@ -3,7 +3,7 @@
 * Application version from package.json.
 * Injected at build time, or read from npm environment in development.
 */
-const VERSION = "0.121.8";
+const VERSION = "0.121.9";
 /**
 * PostHog analytics key.
 * Only populated during production builds via PROMPTFOO_POSTHOG_KEY env var.
@@ -13,4 +13,4 @@ const POSTHOG_KEY = "phc_E5n5uHnDo2eREJL1uqX1cIlbkoRby4yFWt3V94HqRRg";
 //#endregion
 export { VERSION as n, POSTHOG_KEY as t };
 
-//# sourceMappingURL=version-qVEN5qCm.js.map
+//# sourceMappingURL=version-BNhpIb2U.js.map

package/dist/src/{version-F0YDgb7J.js → version-CnYqmWbz.js}

@@ -4,7 +4,7 @@
 * Application version from package.json.
 * Injected at build time, or read from npm environment in development.
 */
-const VERSION = "0.121.8";
+const VERSION = "0.121.9";
 /**
 * PostHog analytics key.
 * Only populated during production builds via PROMPTFOO_POSTHOG_KEY env var.
@@ -14,4 +14,4 @@ const POSTHOG_KEY = "phc_E5n5uHnDo2eREJL1uqX1cIlbkoRby4yFWt3V94HqRRg";
 //#endregion
 export { VERSION as n, POSTHOG_KEY as t };
 
-//# sourceMappingURL=version-F0YDgb7J.js.map
+//# sourceMappingURL=version-CnYqmWbz.js.map

package/dist/src/{version-VzUqOBZk.cjs → version-XjGPawtU.cjs}

@@ -3,7 +3,7 @@
 * Application version from package.json.
 * Injected at build time, or read from npm environment in development.
 */
-const VERSION = "0.121.8";
+const VERSION = "0.121.9";
 /**
 * PostHog analytics key.
 * Only populated during production builds via PROMPTFOO_POSTHOG_KEY env var.
@@ -24,4 +24,4 @@ Object.defineProperty(exports, "VERSION", {
 	}
 });
 
-//# sourceMappingURL=version-VzUqOBZk.cjs.map
+//# sourceMappingURL=version-XjGPawtU.cjs.map