promptfoo 0.121.7 → 0.121.9

package/dist/src/main.js

@@ -1,45 +1,45 @@
 #!/usr/bin/env node
-import { C as getConfigDirectoryPath, D as getEnvInt, E as getEnvFloat, M as isNonInteractive, N as state, O as getEnvString, T as getEnvBool, _ as extractJsonObjects, a as isDebugEnabled, b as orderKeys, c as setLogCallback, g as extractFirstJsonObject, i as initializeRunLogging, j as isCI, l as setLogLevel, m as sanitizeObject, n as getLogLevel, s as logger, t as closeLogger, w as setConfigDirectoryPath } from "./logger-BbY6ypFL.js";
-import { F as getShareApiBaseUrl, I as getShareViewBaseUrl, M as TERMINAL_MAX_WIDTH, N as getDefaultPort, O as CLOUD_PROVIDER_PREFIX, P as getDefaultShareViewBaseUrl, R as HUMAN_ASSERTION_TYPE, _ as REQUEST_TIMEOUT_MS, f as sleep, i as fetchWithTimeout, n as fetchWithProxy, r as fetchWithRetries, t as clearAgentCache, x as isPromptfooSampleTarget } from "./fetch-DXUnXkVU.js";
-import { n as VERSION } from "./version-eRkNuGv8.js";
+import { C as getConfigDirectoryPath, D as getEnvInt, E as getEnvFloat, M as isNonInteractive, N as state, O as getEnvString, T as getEnvBool, _ as extractJsonObjects, a as isDebugEnabled, b as orderKeys, c as setLogCallback, g as extractFirstJsonObject, i as initializeRunLogging, j as isCI, l as setLogLevel, m as sanitizeObject, n as getLogLevel, s as logger, t as closeLogger, w as setConfigDirectoryPath } from "./logger-h-eiS4iv.js";
+import { F as getShareApiBaseUrl, I as getShareViewBaseUrl, M as TERMINAL_MAX_WIDTH, N as getDefaultPort, O as CLOUD_PROVIDER_PREFIX, P as getDefaultShareViewBaseUrl, R as HUMAN_ASSERTION_TYPE, f as sleep, i as fetchWithTimeout, n as fetchWithProxy, r as fetchWithRetries, t as clearAgentCache, v as getRequestTimeoutMs, x as isPromptfooSampleTarget } from "./fetch-C59QM4cf.js";
+import { n as VERSION } from "./version-CnYqmWbz.js";
 import { t as invariant } from "./invariant-B2Rf6avk.js";
-import { a as getAuthor, c as isLoggedIntoCloud, l as promptForEmailUnverified, n as checkEmailStatusAndMaybeExit, o as getUserEmail, r as clearUserEmail, s as getUserId, t as checkEmailStatus, u as setUserEmail } from "./accounts-CjFnOPmb.js";
-import { a as cloudConfig, c as writeGlobalConfigPartial, o as readGlobalConfig, r as CloudConfig } from "./cloud-D3DiFqH6.js";
-import { n as TelemetryEventSchema, r as telemetry } from "./telemetry-00ezXr_t.js";
-import { $ as HARM_PLUGINS, B as categoryAliases, D as ADDITIONAL_STRATEGIES, E as isUuid, F as getDefaultNFanout, G as ADDITIONAL_PLUGINS, H as riskCategorySeverityMap, J as CANARY_BREAKING_STRATEGY_IDS, K as ALL_PLUGINS, L as isFanoutStrategy, N as STRATEGY_COLLECTIONS, P as STRATEGY_COLLECTION_MAPPINGS, Q as FOUNDATION_PLUGINS, R as isMultiTurnStrategy, S as PluginConfigSchema, U as subCategoryDescriptions, W as ALIASED_PLUGIN_MAPPINGS, X as DEFAULT_PLUGINS, Y as DATASET_EXEMPT_PLUGINS, Z as FINANCIAL_PLUGINS, _ as PromptSchema, _t as CODING_AGENT_PLUGINS, a as EvaluateOptionsSchema, at as MULTI_INPUT_VAR, b as ConversationMessageSchema, c as TestSuiteConfigSchema, dt as REMOTE_ONLY_PLUGIN_IDS, et as INSURANCE_PLUGINS, g as RedteamGenerateOptionsSchema, gt as CODING_AGENT_CORE_PLUGINS, h as RedteamConfigSchema, ht as UNALIGNED_PROVIDER_HARM_PLUGINS, i as EvalResultsFilterMode, it as MULTI_INPUT_EXCLUDED_PLUGINS, j as DEFAULT_STRATEGIES, k as ALL_STRATEGIES, l as TestSuiteSchema, lt as REDTEAM_MODEL, mt as TELECOM_PLUGINS, n as BaseAssertionTypesSchema, nt as LLAMA_GUARD_REPLICATE_PROVIDER, o as OutputFileExtension, ot as PHARMACY_PLUGINS, p as isApiProvider, pt as TEEN_SAFETY_PLUGINS, q as BIAS_PLUGINS, r as CommandLineOptionsSchema, rt as MEDICAL_PLUGINS, s as ResultFailureReason, st as PII_PLUGINS, t as AssertionOrSetSchema, tt as LLAMA_GUARD_ENABLED_CATEGORIES, u as UnifiedConfigSchema, ut as REDTEAM_PROVIDER_HARM_PLUGINS, v as ProviderOptionsSchema, w as StrategyConfigSchema, x as PartialGenerationError, y as ProvidersSchema, z as Severity } from "./types-BFevViUY.js";
+import { c as isLoggedIntoCloud, i as getAuthor, l as promptForEmailUnverified, n as checkEmailStatusAndMaybeExit, o as getUserEmail, r as clearUserEmail, s as getUserId, t as checkEmailStatus, u as setUserEmail } from "./accounts-Bw2qPqE-.js";
+import { a as cloudConfig, c as writeGlobalConfigPartial, o as readGlobalConfig, r as CloudConfig } from "./cloud-BbAl8dyj.js";
+import { n as TelemetryEventSchema, r as telemetry } from "./telemetry-BU80NveV.js";
+import { $ as HARM_PLUGINS, B as categoryAliases, D as ADDITIONAL_STRATEGIES, E as isUuid, F as getDefaultNFanout, G as ADDITIONAL_PLUGINS, H as riskCategorySeverityMap, J as CANARY_BREAKING_STRATEGY_IDS, K as ALL_PLUGINS, L as isFanoutStrategy, N as STRATEGY_COLLECTIONS, P as STRATEGY_COLLECTION_MAPPINGS, Q as FOUNDATION_PLUGINS, R as isMultiTurnStrategy, S as PluginConfigSchema, U as subCategoryDescriptions, W as ALIASED_PLUGIN_MAPPINGS, X as DEFAULT_PLUGINS, Y as DATASET_EXEMPT_PLUGINS, Z as FINANCIAL_PLUGINS, _ as PromptSchema, _t as CODING_AGENT_PLUGINS, a as EvaluateOptionsSchema, at as MULTI_INPUT_VAR, b as ConversationMessageSchema, c as TestSuiteConfigSchema, dt as REMOTE_ONLY_PLUGIN_IDS, et as INSURANCE_PLUGINS, g as RedteamGenerateOptionsSchema, gt as CODING_AGENT_CORE_PLUGINS, h as RedteamConfigSchema, ht as UNALIGNED_PROVIDER_HARM_PLUGINS, i as EvalResultsFilterMode, it as MULTI_INPUT_EXCLUDED_PLUGINS, j as DEFAULT_STRATEGIES, k as ALL_STRATEGIES, l as TestSuiteSchema, lt as REDTEAM_MODEL, mt as TELECOM_PLUGINS, n as BaseAssertionTypesSchema, nt as LLAMA_GUARD_REPLICATE_PROVIDER, o as OutputFileExtension, ot as PHARMACY_PLUGINS, p as isApiProvider, pt as TEEN_SAFETY_PLUGINS, q as BIAS_PLUGINS, r as CommandLineOptionsSchema, rt as MEDICAL_PLUGINS, s as ResultFailureReason, st as PII_PLUGINS, t as AssertionOrSetSchema, tt as LLAMA_GUARD_ENABLED_CATEGORIES, u as UnifiedConfigSchema, ut as REDTEAM_PROVIDER_HARM_PLUGINS, v as ProviderOptionsSchema, w as StrategyConfigSchema, x as PartialGenerationError, y as ProvidersSchema, z as Severity } from "./types-C_e_SykX.js";
 import { i as isJavascriptFile } from "./fileExtensions-D4GCJ67J.js";
-import { r as importModule, t as getDirectory } from "./esm-Bexx2PFc.js";
-import { a as extractVariablesFromTemplates, i as extractVariablesFromTemplate, o as getNunjucksEngine, t as renderEnvOnlyInObject } from "./render-CSP99NLm.js";
-import { a as getRemoteHealthUrl, l as shouldGenerateRemote, n as getRemoteGenerationExplicitlyDisabledError, r as getRemoteGenerationUrl, s as neverGenerateRemote } from "./remoteGeneration-B1_XsKXU.js";
-import { a as openAuthBrowser, i as checkServerRunning, n as BrowserBehaviorNames, o as openBrowser, s as promptUser, t as BrowserBehavior } from "./server-DhMHosWj.js";
-import { _ as makeRequest$1, c as getCloudDatabaseId, d as getOrgContext, f as getPluginSeverityOverridesFromCloud, g as isCloudProvider, h as getUserTeams, i as retrieveMedia, l as getConfigFromCloud, m as getProviderFromCloud, o as canCreateTargets, r as mediaExists, s as checkCloudPermissions, t as getMediaStorage, u as getEvalConfigFromCloud, v as resolveTeamFromIdentifier, y as resolveTeamId } from "./storage-BU4qcnOb.js";
+import { r as importModule, t as getDirectory } from "./esm-CYMWMaSj.js";
+import { a as extractVariablesFromTemplates, i as extractVariablesFromTemplate, o as getNunjucksEngine, t as renderEnvOnlyInObject } from "./render-jE6RKVgk.js";
+import { a as getRemoteHealthUrl, l as shouldGenerateRemote, n as getRemoteGenerationExplicitlyDisabledError, r as getRemoteGenerationUrl, s as neverGenerateRemote } from "./remoteGeneration-DW6mPZnO.js";
+import { a as openAuthBrowser, i as checkServerRunning, n as BrowserBehaviorNames, o as openBrowser, s as promptUser, t as BrowserBehavior } from "./server-4wx571OM.js";
+import { _ as makeRequest$1, c as getCloudDatabaseId, d as getOrgContext, f as getPluginSeverityOverridesFromCloud, g as isCloudProvider, h as getUserTeams, i as retrieveMedia, l as getConfigFromCloud, m as getProviderFromCloud, o as canCreateTargets, r as mediaExists, s as checkCloudPermissions, t as getMediaStorage, u as getEvalConfigFromCloud, v as resolveTeamFromIdentifier, y as resolveTeamId } from "./storage-B9lg2RCL.js";
 import { n as sha256, t as randomSequence } from "./createHash-CgRvs4Fn.js";
-import { a as fetchWithCache, n as clearCache, r as disableCache, t as cache_exports } from "./cache-DpPWrkTE.js";
-import { r as runPython } from "./pythonUtils-CgYxeSmO.js";
-import { D as normalizeProviderRef, L as readFilters, M as maybeLoadFromExternalFile, R as readOutput, _ as resultIsForTestCase, a as ComparisonEvalNotFoundError, b as getProviderDescription, c as getEvalTableOutputPromptLocationsBySize, d as fetchCsvFromGoogleSheet, f as setupEnv, g as getTestCaseDeduplicationKey, h as filterRuntimeVars, i as writeOutput, j as maybeLoadConfigFromExternalFile, l as getEvalTablePromptStrippedPayload, m as extractRuntimeVars, n as createOutputMetadata, o as evalTableToJson, p as deduplicateTestCases, r as writeMultipleOutputs, s as generateEvalCsv, t as printBorder, u as mergeComparisonTables, v as checkProviderApiKeys, y as doesProviderRefMatch } from "./util-DURocbYR.js";
-import { a as resolveProviderConfigs, d as MCPProvider, f as HttpProvider, i as resolveProvider, m as createTransformResponse, n as loadApiProvider, p as createTransformRequest, r as loadApiProviders, t as getProviderIds } from "./providers-B9KzWxAX.js";
+import { a as fetchWithCache, n as clearCache, r as disableCache, t as cache_exports } from "./cache-BuudGL4I.js";
+import { r as runPython } from "./pythonUtils-CLAoc6jg.js";
+import { D as normalizeProviderRef, L as readFilters, M as maybeLoadFromExternalFile, R as readOutput, _ as resultIsForTestCase, a as ComparisonEvalNotFoundError, b as getProviderDescription, c as getEvalTableOutputPromptLocationsBySize, d as fetchCsvFromGoogleSheet, f as setupEnv, g as getTestCaseDeduplicationKey, h as filterRuntimeVars, i as writeOutput, j as maybeLoadConfigFromExternalFile, l as getEvalTablePromptStrippedPayload, m as extractRuntimeVars, n as createOutputMetadata, o as evalTableToJson, p as deduplicateTestCases, r as writeMultipleOutputs, s as generateEvalCsv, t as printBorder, u as mergeComparisonTables, v as checkProviderApiKeys, y as doesProviderRefMatch } from "./util-CES4qWd7.js";
+import { a as resolveProviderConfigs, d as MCPProvider, f as HttpProvider, i as resolveProvider, m as createTransformResponse, n as loadApiProvider, p as createTransformRequest, r as loadApiProviders, t as getProviderIds } from "./providers-CYBHjFEg.js";
 import { a as createEmptyTokenUsage, i as createEmptyAssertions, n as accumulateResponseTokenUsage, r as accumulateTokenUsage, t as accumulateAssertionTokenUsage } from "./tokenUsageUtils-CDet74yk.js";
 import { n as escapeRegExp, t as ellipsize } from "./text-Db-Wt2u2.js";
-import { a as evalResultsTable, c as evalsToPromptsTable, d as promptsTable, g as getDb, h as closeDbIfOpen, i as datasetsTable, l as evalsToTagsTable, n as blobReferencesTable, o as evalsTable, p as tagsTable, r as configsTable, s as evalsToDatasetsTable, t as blobAssetsTable, u as modelAuditsTable } from "./tables-DafUHOeh.js";
-import { n as getBlobUrl, t as getBlobByHash } from "./blobs-DODuTK-a.js";
-import { n as createPlaceholderInputValue, r as materializeInputVariablesWithMetadata, t as buildPromptInputDescriptions } from "./inputVariables-DXFdi7AI.js";
-import { a as extractPromptFromTags, i as extractMaterializedVariablesFromJsonWithMetadata, n as extractGoalFromPrompt, r as extractInputVarsFromPrompt, s as getShortPluginId } from "./util-kDURhgJW.js";
-import { n as PromptfooHarmfulCompletionProvider } from "./promptfoo-Rjp-MeBb.js";
-import { A as DivergentRepetitionPlugin, B as sampleArray, C as getPiiLeakTestsForCategory, Ct as getDefaultProviders, D as HarmbenchPlugin, E as ImitationPlugin, F as AegisPlugin, I as RedteamGraderBase, L as RedteamPluginBase, M as CrossSessionLeakPlugin, N as ContractPlugin, O as HallucinationPlugin, P as BeavertailsPlugin, Q as readProviderPromptMap, R as getCustomPolicies, S as PlinyPlugin, T as IntentPlugin, V as fetchHuggingFaceDataset, X as processPrompts, Y as doRemoteGrading, Z as readPrompts, _ as PoliticsPlugin, a as UnverifiableClaimsPlugin, b as isValidPolicyObject, c as ToolDiscoveryPlugin, d as TeenSafetyDangerousContentPlugin, f as TeenSafetyAgeRestrictedGoodsAndServicesPlugin, g as PromptExtractionPlugin, h as RbacPlugin, i as VLGuardPlugin, j as DebugAccessPlugin, k as ExcessiveAgencyPlugin, l as TeenSafetyHarmfulBodyIdealsPlugin, m as ShellInjectionPlugin, o as UnsafeBenchPlugin, p as SqlInjectionPlugin, r as VLSUPlugin, s as ToxicChatPlugin, t as GRADERS, u as TeenSafetyDangerousRoleplayPlugin, v as PolicyPlugin, w as OverreliancePlugin, x as makeInlinePolicyIdSync, y as determinePolicyTypeFromId, z as retryWithDeduplication } from "./graders-BGP99PdK.js";
-import { f as redteamProviderManager, m as TokenUsageTracker } from "./shared-9WHQ1oNE.js";
-import { c as assertions_default, l as readAssertions, n as evaluate$1, o as accumulateNamedMetric, s as doesPromptRefMatch, u as runAssertions } from "./evaluator-DRoiYB2q.js";
-import { i as generateIdFromPrompt } from "./utils-BFOh20Gb.js";
-import { n as TRANSFORM_KEYS, t as INLINE_FUNCTION_LABEL } from "./transform-BnSTnFlp.js";
-import { n as getTraceStore } from "./store-DKd5592Q.js";
-import { n as setupSignalWatcher, t as readSignalEvalId } from "./signal-CSurUUyV.js";
-import { n as isBlobStorageEnabled } from "./extractor-CIW3iN-b.js";
-import { n as getGeneratedPromptOverLimit, r as getMaxCharsPerMessageModifierValue, t as MAX_CHARS_PER_MESSAGE_MODIFIER_KEY } from "./promptLength-4X-Wd8PG.js";
-import { a as createEvalId, n as EvalQueries, s as getEvalSummaries, t as Eval } from "./eval-0VRANImH.js";
-import { n as sanitizeProvider, t as EvalResult } from "./evalResult-eUkJv9Ko.js";
-import { n as renderPrompt } from "./evaluatorHelpers-BsYP_muT.js";
+import { a as evalResultsTable, c as evalsToPromptsTable, d as promptsTable, g as getDb, h as closeDbIfOpen, i as datasetsTable, l as evalsToTagsTable, n as blobReferencesTable, o as evalsTable, p as tagsTable, r as configsTable, s as evalsToDatasetsTable, t as blobAssetsTable, u as modelAuditsTable } from "./tables-D_ZGKiGc.js";
+import { n as getBlobUrl, t as getBlobByHash } from "./blobs-CA4gXarv.js";
+import { n as createPlaceholderInputValue, t as buildPromptInputDescriptions } from "./inputVariables-qJmuSN5i.js";
+import { a as extractPromptFromTags, i as extractMaterializedVariablesFromJsonWithMetadata, n as extractGoalFromPrompt, s as getShortPluginId } from "./util-D9grslJP.js";
+import { a as assertRemoteMaterializationHandled, l as requiresRemoteMaterialization, n as PromptfooHarmfulCompletionProvider } from "./promptfoo-wxe28-q9.js";
+import { A as DivergentRepetitionPlugin, B as sampleArray, C as getPiiLeakTestsForCategory, Ct as getDefaultProviders, D as HarmbenchPlugin, E as ImitationPlugin, F as AegisPlugin, I as RedteamGraderBase, L as RedteamPluginBase, M as CrossSessionLeakPlugin, N as ContractPlugin, O as HallucinationPlugin, P as BeavertailsPlugin, Q as readProviderPromptMap, R as getCustomPolicies, S as PlinyPlugin, T as IntentPlugin, V as fetchHuggingFaceDataset, X as processPrompts, Y as doRemoteGrading, Z as readPrompts, _ as PoliticsPlugin, a as UnverifiableClaimsPlugin, b as isValidPolicyObject, c as ToolDiscoveryPlugin, d as TeenSafetyDangerousContentPlugin, f as TeenSafetyAgeRestrictedGoodsAndServicesPlugin, g as PromptExtractionPlugin, h as RbacPlugin, i as VLGuardPlugin, j as DebugAccessPlugin, k as ExcessiveAgencyPlugin, l as TeenSafetyHarmfulBodyIdealsPlugin, m as ShellInjectionPlugin, o as UnsafeBenchPlugin, p as SqlInjectionPlugin, r as VLSUPlugin, s as ToxicChatPlugin, t as GRADERS, u as TeenSafetyDangerousRoleplayPlugin, v as PolicyPlugin, w as OverreliancePlugin, x as makeInlinePolicyIdSync, y as determinePolicyTypeFromId, z as retryWithDeduplication } from "./graders-BnO-xoa-.js";
+import { f as redteamProviderManager, m as TokenUsageTracker } from "./shared-DsbObxvy.js";
+import { c as assertions_default, l as readAssertions, n as evaluate$1, o as accumulateNamedMetric, s as doesPromptRefMatch, u as runAssertions } from "./evaluator-BpaUYhLY.js";
+import { i as generateIdFromPrompt } from "./utils-Cdx3ZJr7.js";
+import { n as TRANSFORM_KEYS, t as INLINE_FUNCTION_LABEL } from "./transform-C5pxvKUV.js";
+import { n as getTraceStore } from "./store-u1XNK3e8.js";
+import { n as setupSignalWatcher, t as readSignalEvalId } from "./signal-CRrWj2xI.js";
+import { n as isBlobStorageEnabled } from "./extractor-CtadTcJm.js";
+import { n as getGeneratedPromptOverLimit, r as getMaxCharsPerMessageModifierValue, t as MAX_CHARS_PER_MESSAGE_MODIFIER_KEY } from "./promptLength-PkjlYFx0.js";
+import { a as createEvalId, n as EvalQueries, s as getEvalSummaries, t as Eval } from "./eval-DmGWplIN.js";
+import { n as sanitizeProvider, t as EvalResult } from "./evalResult-ByGP9zuP.js";
+import { n as renderPrompt } from "./evaluatorHelpers-DDG44JeI.js";
 import { t as formatDuration } from "./formatDuration-CMVNrYvE.js";
-import { i as pluginMatchesStrategyTargets, n as loadStrategy, r as validateStrategies, t as Strategies } from "./strategies-DRJjGTIY.js";
-import { n as shouldUseInkUI } from "./interactiveCheck-C4QlIuoR.js";
+import { i as pluginMatchesStrategyTargets, n as loadStrategy, r as validateStrategies, t as Strategies } from "./strategies-Ptz_P4P4.js";
+import { n as shouldUseInkUI } from "./interactiveCheck-BLnDrfjg.js";
 import { fileURLToPath } from "node:url";
 import { Command, InvalidArgumentError } from "commander";
 import { z } from "zod";
@@ -111,7 +111,7 @@ function runCommand(program) {
 			hasGithubPr: !!cmdObj.githubPr,
 			hasGuidance: !!(cmdObj.guidance || cmdObj.guidanceFile)
 		});
-		const { executeScan } = await import("./scanner-DS0109SS.js");
+		const { executeScan } = await import("./scanner-B3DPvw5u.js");
 		await executeScan(repoPath, cmdObj);
 	});
 }
@@ -4389,7 +4389,7 @@ async function doEval(cmdObj, defaultConfig, defaultConfigPath, evaluateOptions)
 		const passRate = successes / totalTests * 100;
 		if (cmdObj.table && getLogLevel() !== "debug" && totalTests < 500) {
 			const table = await evalRecord.getTable();
-			const outputTable = generateTable(table);
+			const outputTable = generateTable(table, cmdObj.tableCellMaxLength ?? commandLineOptions?.tableCellMaxLength);
 			logger.info("\n" + outputTable.toString());
 			if (table.body.length > 25) {
 				const rowsLeft = table.body.length - 25;
@@ -4546,7 +4546,7 @@ function evalCommand(program, defaultConfig, defaultConfigPath) {
 		};
 	}, {}).option("-j, --max-concurrency <number>", `Maximum number of concurrent API calls (default: 4)`).option("--repeat <number>", "Number of times to run each test (default: 1)").option("--delay <number>", "Delay between each test (in milliseconds) (default: 0)").option("--no-cache", "Do not read or write results to disk cache", defaultConfig?.commandLineOptions?.cache ?? defaultConfig?.evaluateOptions?.cache).option("--remote", "Force remote inference wherever possible (used for red teams)", false).option("-n, --filter-first-n <number>", "Only run the first N tests").option("--filter-pattern <pattern>", "Only run tests whose description matches the regular expression pattern").option("--filter-prompts <pattern>", "Only run tests with prompts whose id or label matches the regex pattern").option("--filter-providers, --filter-targets <providers>", "Only run tests with these providers (regex match)").option("--filter-sample <number>", "Only run a random sample of N tests").option("--filter-failing <path or id>", "Path to json output file or eval ID to filter non-passing tests from (failures + errors)").option("--filter-failing-only <path or id>", "Path to json output file or eval ID to filter assertion failures from (excludes errors)").option("--filter-errors-only <path or id>", "Path to json output file or eval ID to filter error tests from").option("--filter-metadata <key=value>", "Only run tests whose metadata matches the key=value pair. Can be specified multiple times for AND logic (e.g. --filter-metadata type=unit --filter-metadata env=prod)", (value, previous) => {
 		return previous ? [...previous, value] : [value];
-	}).option("-o, --output <paths...>", "Path to output file (csv, txt, json, yaml, yml, html), default is no output file").option("--table", "Output table in CLI", defaultConfig?.commandLineOptions?.table ?? true).option("--no-table", "Do not output table in CLI", defaultConfig?.commandLineOptions?.table).option("--table-cell-max-length <number>", "Truncate console table cells to this length", "250").option("--share", "Create a shareable URL", defaultConfig?.commandLineOptions?.share).option("--no-share", "Do not share, this overrides the config file").option("--resume [evalId]", "Resume a paused/incomplete evaluation. Defaults to latest when omitted").option("--retry-errors", "Retry all ERROR results from the latest evaluation").option("--no-write", "Do not write results to promptfoo directory", defaultConfig?.commandLineOptions?.write).option("--grader <provider>", "Model that will grade outputs", defaultConfig?.commandLineOptions?.grader).option("--suggest-prompts <number>", "Generate N new prompts and append them to the prompt list").option("-w, --watch", "Watch for changes in config and re-run").option("-x, --extension <paths...>", "Extension hooks to run (e.g., file://handler.js:afterAll)").option("--description <description>", "Description of the eval run").option("--no-progress-bar", "Do not show progress bar").action(async (opts, command) => {
+	}).option("-o, --output <paths...>", "Path to output file (csv, txt, json, yaml, yml, html), default is no output file").option("--table", "Output table in CLI", defaultConfig?.commandLineOptions?.table ?? true).option("--no-table", "Do not output table in CLI", defaultConfig?.commandLineOptions?.table).option("--table-cell-max-length <number>", "Truncate console table cells to this length").option("--share", "Create a shareable URL", defaultConfig?.commandLineOptions?.share).option("--no-share", "Do not share, this overrides the config file").option("--resume [evalId]", "Resume a paused/incomplete evaluation. Defaults to latest when omitted").option("--retry-errors", "Retry all ERROR results from the latest evaluation").option("--no-write", "Do not write results to promptfoo directory", defaultConfig?.commandLineOptions?.write).option("--grader <provider>", "Model that will grade outputs", defaultConfig?.commandLineOptions?.grader).option("--suggest-prompts <number>", "Generate N new prompts and append them to the prompt list").option("-w, --watch", "Watch for changes in config and re-run").option("-x, --extension <paths...>", "Extension hooks to run (e.g., file://handler.js:afterAll)").option("--description <description>", "Description of the eval run").option("--no-progress-bar", "Do not show progress bar").action(async (opts, command) => {
 		let validatedOpts;
 		try {
 			validatedOpts = EvalCommandSchema.parse(opts);
@@ -5590,7 +5590,7 @@ async function fetchRemoteGeneration(task, prompts) {
 			method: "POST",
 			headers: { "Content-Type": "application/json" },
 			body: JSON.stringify(body)
-		}, REQUEST_TIMEOUT_MS, "json");
+		}, getRequestTimeoutMs(), "json");
 		return RedTeamGenerationResponse.parse(response.data).result;
 	} catch (error) {
 		logger.warn(`Error using remote generation for task '${task}': ${error}`);
@@ -5763,7 +5763,7 @@ const DATASET_URL$2 = `https://raw.githubusercontent.com/promptfoo/promptfoo/ref
 const DATASET_URL_MULTILINGUAL = `https://raw.githubusercontent.com/promptfoo/promptfoo/refs/tags/${DATASET_VERSION}/examples/cyberseceval/prompt_injection_multilingual.json`;
 async function fetchDataset$2(limit, isMultilingual) {
 	try {
-		const response = await fetchWithTimeout(isMultilingual ? DATASET_URL_MULTILINGUAL : DATASET_URL$2, {}, REQUEST_TIMEOUT_MS);
+		const response = await fetchWithTimeout(isMultilingual ? DATASET_URL_MULTILINGUAL : DATASET_URL$2, {}, getRequestTimeoutMs());
 		if (!response.ok) throw new Error(`[CyberSecEval] HTTP status: ${response.status} ${response.statusText}`);
 		const data = await response.json();
 		if (!data || !Array.isArray(data)) throw new Error(`[CyberSecEval] Invalid response from ${DATASET_URL$2}`);
@@ -5831,7 +5831,7 @@ async function fetchDataset$1(limit) {
 		logger.debug(`[DoNotAnswer] Fetching dataset from ${DATASET_URL$1}`);
 		let csvData;
 		if (DATASET_URL$1.startsWith("http")) {
-			const response = await fetchWithTimeout(DATASET_URL$1, {}, REQUEST_TIMEOUT_MS);
+			const response = await fetchWithTimeout(DATASET_URL$1, {}, getRequestTimeoutMs());
 			if (!response.ok) throw new Error(`[DoNotAnswer] HTTP status: ${response.status} ${response.statusText}`);
 			csvData = await response.text();
 		} else try {
@@ -6366,7 +6366,7 @@ async function fetchDataset(limit) {
 		logger.debug(`[XSTest] Fetching dataset from ${DATASET_URL}`);
 		let csvData;
 		if (DATASET_URL.startsWith("http")) {
-			const response = await fetchWithTimeout(DATASET_URL, {}, REQUEST_TIMEOUT_MS);
+			const response = await fetchWithTimeout(DATASET_URL, {}, getRequestTimeoutMs());
 			if (!response.ok) throw new Error(`[XSTest] HTTP status: ${response.status} ${response.statusText}`);
 			csvData = await response.text();
 		} else try {
@@ -6644,11 +6644,12 @@ async function fetchRemoteTestCases(key, purpose, injectVar, n, config) {
 			method: "POST",
 			headers: { "Content-Type": "application/json" },
 			body
-		}, REQUEST_TIMEOUT_MS);
+		}, getRequestTimeoutMs());
 		if (status !== 200 || !data || !data.result || !Array.isArray(data.result)) {
 			logger.error(`Error generating test cases for ${key}: ${statusText} ${JSON.stringify(data)}`);
 			return [];
 		}
+		if (requiresRemoteMaterialization(config?.inputs)) assertRemoteMaterializationHandled(data, `Remote plugin generation for ${key}`);
 		const ret = data.result;
 		logger.debug(`Received remote generation for ${key}:\n${JSON.stringify(ret)}`);
 		return ret;
@@ -6657,31 +6658,6 @@ async function fetchRemoteTestCases(key, purpose, injectVar, n, config) {
 		return [];
 	}
 }
-async function materializeRemoteTestCaseInputs({ config, injectVar, pluginId, provider, purpose, testCases }) {
-	const inputs = config.inputs;
-	if (!inputs || Object.keys(inputs).length === 0) return testCases;
-	return Promise.all(testCases.map(async (testCase, materializationIndex) => {
-		const inputVars = extractInputVarsFromPrompt(String(testCase.vars?.[injectVar] ?? ""), inputs);
-		if (!inputVars) return testCase;
-		const materializedVars = await materializeInputVariablesWithMetadata(inputVars, inputs, {
-			materializationIndex,
-			pluginId,
-			provider,
-			purpose
-		});
-		return {
-			...testCase,
-			vars: {
-				...testCase.vars || {},
-				...materializedVars.vars
-			},
-			metadata: {
-				...testCase.metadata || {},
-				...materializedVars.metadata ? { inputMaterialization: materializedVars.metadata } : {}
-			}
-		};
-	}));
-}
 function createPluginFactory(PluginClass, key, validate) {
 	return {
 		key,
@@ -6693,14 +6669,7 @@ function createPluginFactory(PluginClass, key, validate) {
 				return new PluginClass(provider, purpose, injectVar, configWithDefaults).generateTests(n, delayMs);
 			}
 			const pluginId = getShortPluginId(key);
-			const testCases = await materializeRemoteTestCaseInputs({
-				config: configWithDefaults ?? {},
-				injectVar,
-				pluginId,
-				provider,
-				purpose,
-				testCases: await fetchRemoteTestCases(key, purpose, injectVar, n, configWithDefaults ?? {})
-			});
+			const testCases = await fetchRemoteTestCases(key, purpose, injectVar, n, configWithDefaults ?? {});
 			const computedModifiers = computeModifiersFromConfig(configWithDefaults);
 			return testCases.map((testCase) => ({
 				...testCase,
@@ -6787,14 +6756,7 @@ const piiPlugins = PII_PLUGINS.map((category) => ({
 	action: async (params) => {
 		if (shouldGenerateRemote()) {
 			const pluginId = getShortPluginId(category);
-			const testCases = await materializeRemoteTestCaseInputs({
-				config: params.config ?? {},
-				injectVar: params.injectVar,
-				pluginId,
-				provider: params.provider,
-				purpose: params.purpose,
-				testCases: await fetchRemoteTestCases(category, params.purpose, params.injectVar, params.n, params.config ?? {})
-			});
+			const testCases = await fetchRemoteTestCases(category, params.purpose, params.injectVar, params.n, params.config ?? {});
 			const computedModifiers = computeModifiersFromConfig(params.config);
 			return testCases.map((testCase) => ({
 				...testCase,
@@ -6826,14 +6788,7 @@ const biasPlugins = BIAS_PLUGINS.map((category) => ({
 			return [];
 		}
 		const pluginId = getShortPluginId(category);
-		const testCases = await materializeRemoteTestCaseInputs({
-			config: params.config ?? {},
-			injectVar: params.injectVar,
-			pluginId,
-			provider: params.provider,
-			purpose: params.purpose,
-			testCases: await fetchRemoteTestCases(category, params.purpose, params.injectVar, params.n, params.config ?? {})
-		});
+		const testCases = await fetchRemoteTestCases(category, params.purpose, params.injectVar, params.n, params.config ?? {});
 		const computedModifiers = computeModifiersFromConfig(params.config);
 		return testCases.map((testCase) => ({
 			...testCase,
@@ -6852,21 +6807,14 @@ function createRemotePlugin(key, validate) {
 	return {
 		key,
 		validate,
-		action: async ({ provider, purpose, injectVar, n, config }) => {
+		action: async ({ purpose, injectVar, n, config }) => {
 			const configWithDefaults = applyDefaultRemotePluginConfig(key, config);
 			if (neverGenerateRemote()) {
 				logger.error(getRemoteGenerationExplicitlyDisabledError(`${key} plugin`));
 				return [];
 			}
 			const pluginId = getShortPluginId(key);
-			const testCases = await materializeRemoteTestCaseInputs({
-				config: configWithDefaults ?? {},
-				injectVar,
-				pluginId,
-				provider,
-				purpose,
-				testCases: await fetchRemoteTestCases(key, purpose, injectVar, n, configWithDefaults ?? {})
-			});
+			const testCases = await fetchRemoteTestCases(key, purpose, injectVar, n, configWithDefaults ?? {});
 			const computedModifiers = computeModifiersFromConfig(configWithDefaults);
 			const testsWithMetadata = testCases.map((testCase) => ({
 				...testCase,
@@ -6933,6 +6881,15 @@ async function validateSharpDependency(strategies, plugins, checkSharp = isSharp
 }
 //#endregion
 //#region src/redteam/index.ts
+const MATERIALIZED_MULTI_INPUT_PROMPT_METADATA_KEY = "__promptfooMaterializedMultiInputPrompt";
+function getMaterializedMultiInputPromptSnapshot(metadata) {
+	const snapshot = metadata?.[MATERIALIZED_MULTI_INPUT_PROMPT_METADATA_KEY];
+	return typeof snapshot === "string" ? snapshot : void 0;
+}
+function getMaterializedMultiInputPromptMetadata(vars) {
+	const prompt = vars?.[MULTI_INPUT_VAR];
+	return typeof prompt === "string" ? { [MATERIALIZED_MULTI_INPUT_PROMPT_METADATA_KEY]: prompt } : void 0;
+}
 function getPolicyText(metadata) {
 	if (!metadata || metadata.policy === void 0 || metadata.policy === null) return;
 	const policyValue = metadata.policy;
@@ -6945,12 +6902,19 @@ function getPolicyText(metadata) {
 async function rematerializeStrategyInputVars(testCase, injectVar, provider, purpose, materializationIndex) {
 	const inputs = testCase.metadata?.pluginConfig?.inputs;
 	const inputMaterialization = testCase.metadata?.inputMaterialization;
-	if (!inputs || Object.keys(inputs).length === 0 || !testCase.vars?.[injectVar]) return {
+	const materializedPromptSnapshot = getMaterializedMultiInputPromptSnapshot(testCase.metadata);
+	const currentInjectVar = testCase.vars?.[injectVar];
+	if (!inputs || Object.keys(inputs).length === 0 || !currentInjectVar) return {
+		inputMaterialization,
+		vars: testCase.vars
+	};
+	const promptChangedSinceMaterialization = typeof currentInjectVar === "string" && typeof materializedPromptSnapshot === "string" && currentInjectVar !== materializedPromptSnapshot;
+	if (Boolean(inputMaterialization) && Object.keys(inputs).every((key) => Object.prototype.hasOwnProperty.call(testCase.vars ?? {}, key)) && !promptChangedSinceMaterialization) return {
 		inputMaterialization,
 		vars: testCase.vars
 	};
 	try {
-		const materializedVars = await extractMaterializedVariablesFromJsonWithMetadata(JSON.parse(String(testCase.vars[injectVar])), inputs, {
+		const materializedVars = await extractMaterializedVariablesFromJsonWithMetadata(JSON.parse(String(currentInjectVar)), inputs, {
 			materializationIndex,
 			pluginId: String(testCase.metadata?.pluginId || "unknown-plugin"),
 			provider,
@@ -7182,7 +7146,8 @@ function addLanguageToPluginMetadata(test, lang, plugin, maxCharsPerMessage, tes
 				...test.metadata?.modifiers,
 				...languageToAdd
 			},
-			...languageToAdd
+			...languageToAdd,
+			...getMaterializedMultiInputPromptMetadata(test.vars)
 		}
 	};
 }
@@ -7281,7 +7246,8 @@ async function applyStrategies(testCases, strategies, injectVar, provider, purpo
 					...t?.metadata?.pluginId && { pluginId: t.metadata.pluginId },
 					...t?.metadata?.pluginConfig && { pluginConfig: t.metadata.pluginConfig },
 					...inputMaterialization && { inputMaterialization },
-					...Object.keys(strategyConfig).length > 0 && { strategyConfig }
+					...Object.keys(strategyConfig).length > 0 && { strategyConfig },
+					...getMaterializedMultiInputPromptMetadata(vars)
 				}
 			};
 		})));
@@ -7627,14 +7593,21 @@ async function synthesize$1({ abortSignal, delay, entities: entitiesOverride, in
 				}
 			});
 			const languageResults = await Promise.allSettled(languagePromises);
-			for (const result of languageResults) if (result.status === "fulfilled") {
+			for (const [index, result] of languageResults.entries()) if (result.status === "fulfilled") {
 				const { lang, tests, requested, generated } = result.value;
 				allPluginTests.push(...tests);
 				resultsPerLanguage[lang || "default"] = {
 					requested,
 					generated
 				};
-			} else logger.warn(`[Language Processing] Error generating tests for ${plugin.id}: ${result.reason}`);
+			} else {
+				const lang = languages[index];
+				logger.warn(`[Language Processing] Error generating tests for ${plugin.id}: ${result.reason}`);
+				resultsPerLanguage[lang || "default"] = {
+					requested: plugin.numTests,
+					generated: 0
+				};
+			}
 			logger.debug(`[Language Processing] Total tests generated for ${plugin.id}: ${allPluginTests.length} (across ${languages.length} language(s))`);
 			if (!Array.isArray(allPluginTests) || allPluginTests.length === 0) logger.warn(`Failed to generate tests for ${plugin.id}`);
 			else {
@@ -7668,35 +7641,51 @@ async function synthesize$1({ abortSignal, delay, entities: entitiesOverride, in
 				generated: allPluginTests.length
 			};
 		} else if (plugin.id.startsWith("file://")) try {
-			const customTests = await new CustomPlugin(redteamProvider, purpose, injectVar, plugin.id, resolvePluginConfigWithMaxChars(plugin.config, maxCharsPerMessage)).generateTests(plugin.numTests, delay);
-			const testCasesWithMetadata = filterOversizedTestCases(customTests.map((t) => {
-				const includePluginConfig = !(t.metadata && Object.hasOwn(t.metadata, "pluginConfig") && t.metadata.pluginConfig === void 0);
-				const pluginConfigWithMaxChars = {
+			const languageConfig = plugin.config?.language ?? language;
+			const languages = Array.isArray(languageConfig) ? languageConfig : languageConfig ? [languageConfig] : [void 0];
+			const allCustomTests = [];
+			const resultsPerLanguage = {};
+			const languagePromises = languages.map(async (lang) => {
+				const resolvedConfig = {
 					...resolvePluginConfigWithMaxChars(plugin.config, maxCharsPerMessage),
-					...t.metadata?.pluginConfig ?? {}
+					...lang ? { language: lang } : {},
+					...hasMultipleInputs ? { inputs } : {}
 				};
-				const modifiers = {
-					...buildRedteamModifiers({
+				const customPluginConfig = {
+					...resolvedConfig,
+					modifiers: buildRedteamModifiers({
 						maxCharsPerMessage,
-						pluginConfig: pluginConfigWithMaxChars,
+						pluginConfig: resolvedConfig,
 						testGenerationInstructions
-					}),
-					...t.metadata?.modifiers
+					})
 				};
+				const testCasesWithMetadata = filterOversizedTestCases((await new CustomPlugin(redteamProvider, purpose, injectVar, plugin.id, customPluginConfig).generateTests(plugin.numTests, delay)).map((t) => addLanguageToPluginMetadata(t, lang, plugin, maxCharsPerMessage, testGenerationInstructions)), injectVar, `Custom plugin ${plugin.id}`, maxCharsPerMessage);
 				return {
-					...t,
-					metadata: {
-						...t.metadata || {},
-						pluginId: plugin.id,
-						...includePluginConfig && { pluginConfig: pluginConfigWithMaxChars },
-						severity: plugin.severity || getPluginSeverity(plugin.id, resolvePluginConfig(plugin.config)),
-						modifiers
-					}
+					lang,
+					tests: testCasesWithMetadata,
+					requested: plugin.numTests,
+					generated: testCasesWithMetadata.length
+				};
+			});
+			const languageResults = await Promise.allSettled(languagePromises);
+			for (const [index, result] of languageResults.entries()) if (result.status === "fulfilled") {
+				const { lang, tests, requested, generated } = result.value;
+				allCustomTests.push(...tests);
+				resultsPerLanguage[lang || "default"] = {
+					requested,
+					generated
+				};
+			} else {
+				const lang = languages[index];
+				logger.warn(`[Language Processing] Error generating tests for custom plugin ${plugin.id}: ${result.reason}`);
+				resultsPerLanguage[lang || "default"] = {
+					requested: plugin.numTests,
+					generated: 0
 				};
-			}), injectVar, `Custom plugin ${plugin.id}`, maxCharsPerMessage);
+			}
 			if (needsGoalExtraction) {
-				logger.debug(`Extracting goal for ${testCasesWithMetadata.length} custom tests from ${plugin.id}...`);
-				for (const testCase of testCasesWithMetadata) {
+				logger.debug(`Extracting goal for ${allCustomTests.length} custom tests from ${plugin.id}...`);
+				for (const testCase of allCustomTests) {
 					const promptVar = testCase.vars?.[injectVar];
 					const prompt = Array.isArray(promptVar) ? promptVar[0] : String(promptVar);
 					const policy = getPolicyText(testCase.metadata);
@@ -7704,13 +7693,21 @@ async function synthesize$1({ abortSignal, delay, entities: entitiesOverride, in
 					testCase.metadata.goal = extractedGoal;
 				}
 			}
-			testCases.push(...testCasesWithMetadata);
-			logger.debug(`Added ${customTests.length} custom test cases from ${plugin.id}`);
-			const displayId = getPluginDisplayId(plugin);
-			pluginResults[displayId] = {
-				requested: plugin.numTests,
-				generated: testCasesWithMetadata.length
+			testCases.push(...allCustomTests);
+			logger.debug(`Added ${allCustomTests.length} custom test cases from ${plugin.id}`);
+			const baseDisplayId = getPluginDisplayId(plugin);
+			if (languages.filter((lang) => lang !== void 0).length > 1) for (const [langKey, result] of Object.entries(resultsPerLanguage)) {
+				const displayId = langKey === "en" ? baseDisplayId : `(${langKey}) ${baseDisplayId}`;
+				pluginResults[displayId] = {
+					requested: result.requested,
+					generated: result.generated
+				};
+			}
+			else pluginResults[baseDisplayId] = {
+				requested: plugin.numTests * languages.length,
+				generated: allCustomTests.length
 			};
+			progressBar?.increment(plugin.numTests * languages.length);
 		} catch (e) {
 			logger.error(`Error generating tests for custom plugin ${plugin.id}: ${e}`);
 			const displayId = getPluginDisplayId(plugin);
@@ -9671,7 +9668,7 @@ async function checkForUpdates() {
 	} catch {
 		return false;
 	}
-	if (semverGt(latestVersion, "0.121.7")) {
+	if (semverGt(latestVersion, "0.121.9")) {
 		const border = "=".repeat(TERMINAL_MAX_WIDTH);
 		logger.info(`\n${border}
 ${chalk.yellow("⚠️")} The current version of promptfoo ${chalk.yellow(VERSION)} is lower than the latest available version ${chalk.green(latestVersion)}.
@@ -12487,7 +12484,7 @@ async function handleGoatStrategy(ctx) {
 		method: "POST",
 		headers: { "Content-Type": "application/json" },
 		body: JSON.stringify(goatBody)
-	}, REQUEST_TIMEOUT_MS);
+	}, getRequestTimeoutMs());
 	if (!response.ok) throw new Error(`GOAT task failed with status ${response.status}: ${await response.text()}`);
 	const data = await response.json();
 	const attackerMessage = data?.message;
@@ -12518,7 +12515,7 @@ async function handleMischievousUserStrategy(ctx) {
 		method: "POST",
 		headers: { "Content-Type": "application/json" },
 		body: JSON.stringify(mischievousBody)
-	}, REQUEST_TIMEOUT_MS);
+	}, getRequestTimeoutMs());
 	if (!response.ok) throw new Error(`Mischievous User task failed with status ${response.status}: ${await response.text()}`);
 	const data = await response.json();
 	const result = data?.result;
@@ -12575,7 +12572,7 @@ async function handleHydraStrategy(ctx) {
 		method: "POST",
 		headers: { "Content-Type": "application/json" },
 		body: JSON.stringify(hydraBody)
-	}, REQUEST_TIMEOUT_MS);
+	}, getRequestTimeoutMs());
 	if (!response.ok) throw new Error(`Hydra task failed with status ${response.status}: ${await response.text()}`);
 	const data = await response.json();
 	const rawResult = data?.result;
@@ -12626,7 +12623,7 @@ async function handleCrescendoLikeStrategy(ctx) {
 		method: "POST",
 		headers: { "Content-Type": "application/json" },
 		body: JSON.stringify(providerRequest)
-	}, REQUEST_TIMEOUT_MS);
+	}, getRequestTimeoutMs());
 	if (!response.ok) throw new Error(`Crescendo task failed with status ${response.status}: ${await response.text()}`);
 	const data = await response.json();
 	const rawResult = data?.result;
@@ -13274,7 +13271,7 @@ router.get("/", async (_req, res) => {
 				};
 			} catch (error) {
 				logger.debug(`Failed to fetch latest version: ${error}`);
-				latestVersion = versionCache.latestVersion ?? "0.121.7";
+				latestVersion = versionCache.latestVersion ?? "0.121.9";
 			}
 		}
 		const selfHosted = getEnvBool("PROMPTFOO_SELF_HOSTED");
@@ -13283,7 +13280,7 @@ router.get("/", async (_req, res) => {
 			selfHosted,
 			isNpx
 		});
-		const resolvedLatestVersion = latestVersion ?? "0.121.7";
+		const resolvedLatestVersion = latestVersion ?? "0.121.9";
 		const response = {
 			currentVersion: VERSION,
 			latestVersion: resolvedLatestVersion,
@@ -16134,7 +16131,7 @@ function shouldUseInkList() {
 async function runInkList(options) {
 	const [React, { renderInteractive }, { ListApp }] = await Promise.all([
 		import("react"),
-		import("./render-DznWrxGO.js"),
+		import("./render-MiyTvDY4.js"),
 		import("./ListApp-DLmM02JS.js")
 	]);
 	let result = { cancelled: false };
@@ -17848,7 +17845,7 @@ function registerRedteamGenerateTool(server) {
 					maxConcurrency,
 					delay,
 					language,
-					provider: provider || "openai:chat:gpt-5.4-2026-03-05",
+					provider: provider || "openai:chat:gpt-5.5-2026-04-23",
 					force,
 					write,
 					remote
@@ -17874,7 +17871,7 @@ function registerRedteamGenerateTool(server) {
 					entities: result.defaultTest && typeof result.defaultTest === "object" && "metadata" in result.defaultTest ? result.defaultTest.metadata?.entities || [] : [],
 					generatedAt: (/* @__PURE__ */ new Date()).toISOString(),
 					language,
-					provider: provider || "openai:chat:gpt-5.4-2026-03-05"
+					provider: provider || "openai:chat:gpt-5.5-2026-04-23"
 				},
 				nextSteps: {
 					runEvaluation: write ? "Run \"redteam_run\" to execute the generated tests" : `Run "redteam_run" with output: "${output || "redteam.yaml"}" to execute the tests`,
@@ -18443,8 +18440,8 @@ function registerRunEvaluationTool(server) {
 					}
 					filteredTestSuite.tests = filteredTests;
 				}
-				const { evaluate } = await import("./evaluator-DNdJF1Gv.js");
-				const evalRecord = await (await import("./eval-DscR5iOM.js")).default.create(config, filteredTestSuite.prompts, { id: `mcp-eval-${Date.now()}` });
+				const { evaluate } = await import("./evaluator-B01pCdUY.js");
+				const evalRecord = await (await import("./eval-N_VV5NZg.js")).default.create(config, filteredTestSuite.prompts, { id: `mcp-eval-${Date.now()}` });
 				logger.debug(`Running filtered eval with ${filteredTestSuite.tests?.length || 0} test cases, ${filteredTestSuite.prompts.length} prompts, ${filteredTestSuite.providers.length} providers`);
 				const startTime = Date.now();
 				const result = await evaluate(filteredTestSuite, evalRecord, {

package/dist/src/{memoryPoisoning-Dp-btinn.cjs → memoryPoisoning-BXDi4XCK.cjs}

@@ -1,12 +1,12 @@
-const require_logger = require("./logger-cfNpzI4o.cjs");
+const require_logger = require("./logger-Cm0Nb8YX.cjs");
 const require_invariant = require("./invariant-QtnLD03y.cjs");
-const require_fetch = require("./fetch-Dw4XZHjj.cjs");
-const require_version = require("./version-BWCSaByA.cjs");
-const require_accounts = require("./accounts-CmWzeD2d.cjs");
-const require_remoteGeneration = require("./remoteGeneration-DS9N3pgB.cjs");
+const require_fetch = require("./fetch-B2HvaVUc.cjs");
+const require_version = require("./version-XjGPawtU.cjs");
+const require_accounts = require("./accounts-DAcQXPwE.cjs");
+const require_remoteGeneration = require("./remoteGeneration-CD3c_rzc.cjs");
 const require_tokenUsageUtils = require("./tokenUsageUtils-_B-P8IAi.cjs");
-const require_shared = require("./shared-WkgnDkcg.cjs");
-const require_promptLength = require("./promptLength-BbBbDHNj.cjs");
+const require_shared = require("./shared-DlmtkptR.cjs");
+const require_promptLength = require("./promptLength-CAcg7BmO.cjs");
 const require_constants = require("./constants-a2kYssQk.cjs");
 //#region src/redteam/providers/agentic/memoryPoisoning.ts
 var MemoryPoisoningProvider = class {
@@ -103,4 +103,4 @@ var MemoryPoisoningProvider = class {
 //#endregion
 exports.MemoryPoisoningProvider = MemoryPoisoningProvider;
 
-//# sourceMappingURL=memoryPoisoning-Dp-btinn.cjs.map
+//# sourceMappingURL=memoryPoisoning-BXDi4XCK.cjs.map

package/dist/src/{memoryPoisoning-CM83NWYl.js → memoryPoisoning-D-V4yXjj.js}

@@ -1,13 +1,13 @@
 #!/usr/bin/env node
-import { s as logger } from "./logger-BbY6ypFL.js";
-import { n as fetchWithProxy } from "./fetch-DXUnXkVU.js";
-import { n as VERSION } from "./version-eRkNuGv8.js";
+import { s as logger } from "./logger-h-eiS4iv.js";
+import { n as fetchWithProxy } from "./fetch-C59QM4cf.js";
+import { n as VERSION } from "./version-CnYqmWbz.js";
 import { t as invariant } from "./invariant-B2Rf6avk.js";
-import { o as getUserEmail } from "./accounts-CjFnOPmb.js";
-import { r as getRemoteGenerationUrl } from "./remoteGeneration-B1_XsKXU.js";
+import { o as getUserEmail } from "./accounts-Bw2qPqE-.js";
+import { r as getRemoteGenerationUrl } from "./remoteGeneration-DW6mPZnO.js";
 import { a as createEmptyTokenUsage, n as accumulateResponseTokenUsage } from "./tokenUsageUtils-CDet74yk.js";
-import { d as messagesToRedteamHistory } from "./shared-9WHQ1oNE.js";
-import { i as throwIfTargetPromptExceedsMaxChars } from "./promptLength-4X-Wd8PG.js";
+import { d as messagesToRedteamHistory } from "./shared-DsbObxvy.js";
+import { i as throwIfTargetPromptExceedsMaxChars } from "./promptLength-PkjlYFx0.js";
 import { t as REDTEAM_MEMORY_POISONING_PLUGIN_ID } from "./constants-DH5XYLKZ.js";
 //#region src/redteam/providers/agentic/memoryPoisoning.ts
 var MemoryPoisoningProvider = class {
@@ -104,4 +104,4 @@ var MemoryPoisoningProvider = class {
 //#endregion
 export { MemoryPoisoningProvider };
 
-//# sourceMappingURL=memoryPoisoning-CM83NWYl.js.map
+//# sourceMappingURL=memoryPoisoning-D-V4yXjj.js.map