promptfoo 0.120.24 → 0.120.26

package/dist/src/index.js

@@ -1,39 +1,40 @@
 import "./chunk-FhC4c-0y.js";
-import { C as isCI, S as getMaxEvalTimeMs, _ as getEnvBool, a as setLogCallback, b as getEnvString, d as getAjv, h as summarizeEvaluateResultForLogging, i as logger_default, m as safeJsonStringify, n as isDebugEnabled, o as setLogLevel, p as orderKeys, t as getLogLevel, u as extractJsonObjects, v as getEnvFloat, w as cliState_default, x as getEvalTimeoutMs, y as getEnvInt } from "./logger-q5I8CByj.js";
+import { C as isCI, S as getMaxEvalTimeMs, _ as getEnvBool, a as setLogCallback, b as getEnvString, d as getAjv, h as summarizeEvaluateResultForLogging, i as logger_default, m as safeJsonStringify, n as isDebugEnabled, o as setLogLevel, p as orderKeys, t as getLogLevel, u as extractJsonObjects, v as getEnvFloat, w as cliState_default, x as getEvalTimeoutMs, y as getEnvInt } from "./logger-w8Ozp0Td.js";
 import { t as invariant } from "./invariant-1pAf2CD1.js";
-import { r as importModule, t as getDirectory } from "./esm-D40XsWlR.js";
-import { r as runPython } from "./pythonUtils-eNq6Wsfr.js";
+import { r as importModule, t as getDirectory } from "./esm-rDtG_2rg.js";
+import { r as runPython } from "./pythonUtils-vMlk9Qp5.js";
 import { i as isJavascriptFile } from "./fileExtensions-ePDqouxn.js";
-import { n as transform, r as getProcessShim, t as TransformInputType } from "./transform-DvQWeBSR.js";
-import { $ as matchesSearchRubric, A as BeavertailsPlugin, B as getAndCheckProvider, C as HarmbenchPlugin, D as DebugAccessPlugin, E as DivergentRepetitionPlugin, F as retryWithDeduplication, G as matchesContextFaithfulness, H as matchesAnswerRelevance, I as sampleArray, J as matchesFactuality, K as matchesContextRecall, L as fetchHuggingFaceDataset, M as RedteamGraderBase, N as RedteamPluginBase, O as CrossSessionLeakPlugin, P as getCustomPolicies, Q as matchesPiScore, R as callProviderWithContext, S as ImitationPlugin, T as ExcessiveAgencyPlugin, U as matchesClassification, V as loadRubricPrompt, W as matchesClosedQa, X as matchesLlmRubric, Y as matchesGEval, Z as matchesModeration, _ as makeInlinePolicyIdSync, a as UnverifiableClaimsPlugin, at as processPrompts, b as OverreliancePlugin, c as ToolDiscoveryPlugin, ct as SUGGEST_PROMPTS_SYSTEM_MESSAGE, d as RbacPlugin, dt as loadFromJavaScriptFile, et as matchesSelectBest, f as PromptExtractionPlugin, ft as processFileReference, g as isValidPolicyObject, h as determinePolicyTypeFromId, i as VLGuardPlugin, it as DefaultSuggestionsProvider, j as AegisPlugin, k as ContractPlugin, l as SqlInjectionPlugin, lt as coerceString, m as PolicyPlugin, n as getGraderById, nt as selectMaxScore, o as UnsafeBenchPlugin, ot as readPrompts, p as PoliticsPlugin, pt as resolveContext, q as matchesContextRelevance, r as VLSUPlugin, rt as getDefaultProviders, s as ToxicChatPlugin, st as readProviderPromptMap, t as GRADERS, tt as matchesSimilarity, u as ShellInjectionPlugin, ut as getFinalTest, v as PlinyPlugin, w as HallucinationPlugin, x as IntentPlugin, y as getPiiLeakTestsForCategory, z as fail } from "./graders-BscQ6O1Q.js";
-import { $ as riskCategorySeverityMap, A as isApiProvider, C as TestGeneratorConfigSchema, Ct as CompletionTokenDetailsSchema, D as VarsSchema, E as UnifiedConfigSchema, F as ConversationMessageSchema, G as STRATEGY_COLLECTIONS, I as PartialGenerationError, K as STRATEGY_COLLECTION_MAPPINGS, L as PluginConfigSchema, M as RedteamConfigSchema, O as isGradingResult, P as ProvidersSchema, R as PolicyObjectSchema, S as TestCasesWithMetadataSchema, St as BaseTokenUsageSchema, T as TestSuiteSchema, U as DEFAULT_STRATEGIES, X as Severity, Y as isFanoutStrategy, Z as categoryAliases, _ as ScenarioSchema, _t as REMOTE_ONLY_PLUGIN_IDS, a as AtomicTestCaseSchema, at as FOUNDATION_PLUGINS, b as TestCaseWithVarsFileSchema, bt as UNALIGNED_PROVIDER_HARM_PLUGINS, c as CompletedPromptSchema, ct as LLAMA_GUARD_ENABLED_CATEGORIES, d as EvaluateOptionsSchema, dt as MULTI_INPUT_EXCLUDED_PLUGINS, et as ALIASED_PLUGIN_MAPPINGS, f as GradingConfigSchema, ft as MULTI_INPUT_VAR, g as ResultFailureReason, gt as REDTEAM_PROVIDER_HARM_PLUGINS, h as OutputFileExtension, ht as PLUGIN_CATEGORIES, i as AssertionTypeSchema, it as FINANCIAL_PLUGINS, j as isProviderOptions, k as isResultFailureReason, l as DerivedMetricSchema, lt as LLAMA_GUARD_REPLICATE_PROVIDER, m as OutputConfigSchema, mt as PII_PLUGINS, n as AssertionSchema, nt as DATASET_EXEMPT_PLUGINS, o as BaseAssertionTypesSchema, ot as HARM_PLUGINS, p as NotPrefixedAssertionTypesSchema, pt as PHARMACY_PLUGINS, q as getDefaultNFanout, r as AssertionSetSchema, rt as DEFAULT_PLUGINS, s as CommandLineOptionsSchema, st as INSURANCE_PLUGINS, t as AssertionOrSetSchema, tt as BIAS_PLUGINS, u as EvalResultsFilterMode, ut as MEDICAL_PLUGINS, v as SpecialAssertionTypesSchema, w as TestSuiteConfigSchema, wt as InputsSchema, x as TestCasesWithMetadataPromptSchema, xt as PromptSchema, y as TestCaseSchema, yt as TELECOM_PLUGINS, z as StrategyConfigSchema } from "./types-BsU_PxR3.js";
-import { A as isAnthropicProvider, C as deduplicateTestCases, D as resultIsForTestCase, E as getTestCaseDeduplicationKey, M as isOpenAiProvider, N as isProviderAllowed, O as doesProviderRefMatch, S as setupEnv, T as filterRuntimeVars, b as loadFunction, c as maybeLoadFromExternalFile, d as maybeLoadToolsFromExternalFile, h as renderEnvOnlyInObject, i as fetchCsvFromGoogleSheet, j as isGoogleProvider, k as getProviderDescription, m as readOutput, n as writeMultipleOutputs, p as readFilters, r as writeOutput, s as maybeLoadConfigFromExternalFile, t as printBorder, v as extractVariablesFromTemplates, w as extractRuntimeVars, x as parseFileUrl, y as getNunjucksEngine } from "./util-WYC3rB_p.js";
-import { F as getShareViewBaseUrl, L as VERSION, M as TERMINAL_MAX_WIDTH, N as getDefaultShareViewBaseUrl, O as DEFAULT_MAX_CONCURRENCY, P as getShareApiBaseUrl, R as FILE_METADATA_KEY, S as parseChatPrompt, c as cloudConfig, f as getCurrentTimestamp, h as REQUEST_TIMEOUT_MS, k as DEFAULT_QUERY_LIMIT, n as fetchWithRetries, p as sleep, r as fetchWithTimeout, s as CloudConfig, t as fetchWithProxy, y as isPromptfooSampleTarget, z as HUMAN_ASSERTION_TYPE } from "./fetch-BIfcW-Fd.js";
-import { a as fetchWithCache, o as getCache, r as disableCache, t as cache_exports } from "./cache-BMl3AIo5.js";
-import { A as createRateLimitRegistry, B as resolveTeamId, C as collectFileMetadata, D as loadFromPackage, E as isPackagePath, F as getCloudDatabaseId, I as getOrgContext, J as AIStudioChatProvider, L as getPluginSeverityOverridesFromCloud, M as PromptfooHarmfulCompletionProvider, O as redteamProviderManager, P as checkCloudPermissions, T as runExtensionHook, _ as extractVariablesFromJson, a as resolveProviderConfigs, b as isBasicRefusal, c as Strategies, d as pluginMatchesStrategyTargets, f as checkExfilTracking, g as extractPromptFromTags, i as resolveProvider, j as createProviderRateLimitOptions, k as TokenUsageTracker, l as loadStrategy, m as extractGoalFromPrompt, n as loadApiProvider, o as MCPProvider, q as VertexChatProvider, r as loadApiProviders, s as GoogleLiveProvider, t as getProviderIds, u as validateStrategies, v as getSessionId, w as renderPrompt, y as getShortPluginId, z as isCloudProvider } from "./providers-Co_FGgH8.js";
-import { i as generateIdFromPrompt, t as hashPrompt } from "./utils-BryuD3vq.js";
-import { a as sha256, i as randomSequence, n as isBlobStorageEnabled, t as extractAndStoreBinaryData } from "./extractor-Da8IxCeS.js";
+import { n as transform, r as getProcessShim, t as TransformInputType } from "./transform-D8dILpfZ.js";
+import { $ as matchesSearchRubric, A as BeavertailsPlugin, B as getAndCheckProvider, C as HarmbenchPlugin, D as DebugAccessPlugin, E as DivergentRepetitionPlugin, F as retryWithDeduplication, G as matchesContextFaithfulness, H as matchesAnswerRelevance, I as sampleArray, J as matchesFactuality, K as matchesContextRecall, L as fetchHuggingFaceDataset, M as RedteamGraderBase, N as RedteamPluginBase, O as CrossSessionLeakPlugin, P as getCustomPolicies, Q as matchesPiScore, R as callProviderWithContext, S as ImitationPlugin, T as ExcessiveAgencyPlugin, U as matchesClassification, V as loadRubricPrompt, W as matchesClosedQa, X as matchesLlmRubric, Y as matchesGEval, Z as matchesModeration, _ as makeInlinePolicyIdSync, a as UnverifiableClaimsPlugin, at as processPrompts, b as OverreliancePlugin, c as ToolDiscoveryPlugin, ct as SUGGEST_PROMPTS_SYSTEM_MESSAGE, d as RbacPlugin, dt as loadFromJavaScriptFile, et as matchesSelectBest, f as PromptExtractionPlugin, ft as processFileReference, g as isValidPolicyObject, h as determinePolicyTypeFromId, i as VLGuardPlugin, it as DefaultSuggestionsProvider, j as AegisPlugin, k as ContractPlugin, l as SqlInjectionPlugin, lt as coerceString, m as PolicyPlugin, n as getGraderById, nt as selectMaxScore, o as UnsafeBenchPlugin, ot as readPrompts, p as PoliticsPlugin, pt as resolveContext, q as matchesContextRelevance, r as VLSUPlugin, rt as getDefaultProviders, s as ToxicChatPlugin, st as readProviderPromptMap, t as GRADERS, tt as matchesSimilarity, u as ShellInjectionPlugin, ut as getFinalTest, v as PlinyPlugin, w as HallucinationPlugin, x as IntentPlugin, y as getPiiLeakTestsForCategory, z as fail } from "./graders-QsALpIdy.js";
+import { A as isApiProvider, C as TestGeneratorConfigSchema, Ct as BaseTokenUsageSchema, D as VarsSchema, E as UnifiedConfigSchema, F as ConversationMessageSchema, I as PartialGenerationError, J as getDefaultNFanout, K as STRATEGY_COLLECTIONS, L as PluginConfigSchema, M as RedteamConfigSchema, O as isGradingResult, P as ProvidersSchema, Q as categoryAliases, R as PolicyObjectSchema, S as TestCasesWithMetadataSchema, St as PromptSchema, T as TestSuiteSchema, Tt as InputsSchema, V as isUuid, W as DEFAULT_STRATEGIES, X as isFanoutStrategy, Z as Severity, _ as ScenarioSchema, _t as REDTEAM_PROVIDER_HARM_PLUGINS, a as AtomicTestCaseSchema, at as FINANCIAL_PLUGINS, b as TestCaseWithVarsFileSchema, bt as TELECOM_PLUGINS, c as CompletedPromptSchema, ct as INSURANCE_PLUGINS, d as EvaluateOptionsSchema, dt as MEDICAL_PLUGINS, et as riskCategorySeverityMap, f as GradingConfigSchema, ft as MULTI_INPUT_EXCLUDED_PLUGINS, g as ResultFailureReason, gt as PLUGIN_CATEGORIES, h as OutputFileExtension, ht as PII_PLUGINS, i as AssertionTypeSchema, it as DEFAULT_PLUGINS, j as isProviderOptions, k as isResultFailureReason, l as DerivedMetricSchema, lt as LLAMA_GUARD_ENABLED_CATEGORIES, m as OutputConfigSchema, mt as PHARMACY_PLUGINS, n as AssertionSchema, nt as BIAS_PLUGINS, o as BaseAssertionTypesSchema, ot as FOUNDATION_PLUGINS, p as NotPrefixedAssertionTypesSchema, pt as MULTI_INPUT_VAR, q as STRATEGY_COLLECTION_MAPPINGS, r as AssertionSetSchema, rt as DATASET_EXEMPT_PLUGINS, s as CommandLineOptionsSchema, st as HARM_PLUGINS, t as AssertionOrSetSchema, tt as ALIASED_PLUGIN_MAPPINGS, u as EvalResultsFilterMode, ut as LLAMA_GUARD_REPLICATE_PROVIDER, v as SpecialAssertionTypesSchema, vt as REMOTE_ONLY_PLUGIN_IDS, w as TestSuiteConfigSchema, wt as CompletionTokenDetailsSchema, x as TestCasesWithMetadataPromptSchema, xt as UNALIGNED_PROVIDER_HARM_PLUGINS, y as TestCaseSchema, z as StrategyConfigSchema } from "./types-DMVjYLpx.js";
+import { A as isAnthropicProvider, C as deduplicateTestCases, D as resultIsForTestCase, E as getTestCaseDeduplicationKey, M as isOpenAiProvider, N as isProviderAllowed, O as doesProviderRefMatch, S as setupEnv, T as filterRuntimeVars, b as loadFunction, c as maybeLoadFromExternalFile, d as maybeLoadToolsFromExternalFile, h as renderEnvOnlyInObject, i as fetchCsvFromGoogleSheet, j as isGoogleProvider, k as getProviderDescription, m as readOutput, n as writeMultipleOutputs, p as readFilters, r as writeOutput, s as maybeLoadConfigFromExternalFile, t as printBorder, v as extractVariablesFromTemplates, w as extractRuntimeVars, x as parseFileUrl, y as getNunjucksEngine } from "./util-DiCePfDu.js";
+import { F as getShareViewBaseUrl, L as VERSION, M as TERMINAL_MAX_WIDTH, N as getDefaultShareViewBaseUrl, O as DEFAULT_MAX_CONCURRENCY, P as getShareApiBaseUrl, R as FILE_METADATA_KEY, S as parseChatPrompt, c as cloudConfig, f as getCurrentTimestamp, h as REQUEST_TIMEOUT_MS, k as DEFAULT_QUERY_LIMIT, n as fetchWithRetries, p as sleep, r as fetchWithTimeout, s as CloudConfig, t as fetchWithProxy, y as isPromptfooSampleTarget, z as HUMAN_ASSERTION_TYPE } from "./fetch-18MuNu9i.js";
+import { a as fetchWithCache, c as NON_TRANSIENT_HTTP_STATUSES, l as isNonTransientHttpStatus, o as getCache, r as disableCache, t as cache_exports } from "./cache-BLLayYEN.js";
+import { A as createRateLimitRegistry, B as isCloudProvider, C as collectFileMetadata, D as loadFromPackage, E as isPackagePath, F as getCloudDatabaseId, I as getEvalConfigFromCloud, J as AIStudioChatProvider, L as getOrgContext, M as PromptfooHarmfulCompletionProvider, O as redteamProviderManager, P as checkCloudPermissions, R as getPluginSeverityOverridesFromCloud, T as runExtensionHook, V as resolveTeamId, _ as extractVariablesFromJson, a as resolveProviderConfigs, b as isBasicRefusal, c as Strategies, d as pluginMatchesStrategyTargets, f as checkExfilTracking, g as extractPromptFromTags, i as resolveProvider, j as createProviderRateLimitOptions, k as TokenUsageTracker, l as loadStrategy, m as extractGoalFromPrompt, n as loadApiProvider, o as MCPProvider, q as VertexChatProvider, r as loadApiProviders, s as GoogleLiveProvider, t as getProviderIds, u as validateStrategies, v as getSessionId, w as renderPrompt, y as getShortPluginId } from "./providers-BiNq_Iyc.js";
+import { i as generateIdFromPrompt, t as hashPrompt } from "./utils-CVzb4YiI.js";
+import { a as sha256, i as randomSequence, n as isBlobStorageEnabled, t as extractAndStoreBinaryData } from "./extractor-Hs7la_19.js";
 import "./genaiTracer-Ce19n68P.js";
-import { t as OpenAiChatCompletionProvider } from "./chat-DDGU7Tik.js";
-import { a as createEmptyTokenUsage, i as createEmptyAssertions, n as accumulateResponseTokenUsage, o as normalizeTokenUsage, r as accumulateTokenUsage, t as accumulateAssertionTokenUsage } from "./tokenUsageUtils-cFdLMERB.js";
-import { p as validateFunctionCall } from "./transform-BUrxadlA.js";
-import "./messages-CJgEjRQF.js";
-import "./util-CfepsNVK.js";
-import "./openai-CitF-gEN.js";
-import { l as validateFunctionCall$1 } from "./util-Bnw6EyZN.js";
-import "./completion-1RAiLcwi.js";
-import { c as setUserEmail, i as getUserEmail, l as EMAIL_OK_STATUS, r as getAuthor, s as promptForEmailUnverified, t as checkEmailStatusAndMaybeExit } from "./accounts-DxR9x_En.js";
-import { i as getRemoteGenerationUrl, l as shouldGenerateRemote, o as getRemoteHealthUrl, r as promptYesNo, s as neverGenerateRemote } from "./server-BF3HkMhe.js";
-import { t as getBlobByHash } from "./blobs-Dcm272OH.js";
-import { a as evalsTable, c as evalsToTagsTable, d as tagsTable, i as evalResultsTable, l as promptsTable, m as getDbSignalPath, o as evalsToDatasetsTable, p as getDb, r as datasetsTable, s as evalsToPromptsTable } from "./tables-Dwexr5Z6.js";
-import { n as telemetry_default } from "./telemetry-Ds5Nn81l.js";
-import { t as ellipsize } from "./text-Dx0GJOCN.js";
-import { n as getTraceStore } from "./store-CT_e5OZX.js";
-import "./base-BjsgxJp9.js";
-import "./image-B6TV9l0v.js";
-import { t as providerRegistry } from "./providerRegistry-BdKWcUa8.js";
-import { n as runRuby } from "./rubyUtils-LBsk3zIm.js";
-import { t as EvalResult } from "./evalResult-C_UM63nn.js";
+import { t as OpenAiChatCompletionProvider } from "./chat-C8Ei4f87.js";
+import { a as createEmptyTokenUsage, i as createEmptyAssertions, n as accumulateResponseTokenUsage, o as normalizeTokenUsage, r as accumulateTokenUsage, t as accumulateAssertionTokenUsage } from "./tokenUsageUtils-FZd5O_4A.js";
+import { m as validateFunctionCall } from "./transform-DMaxQwDx.js";
+import "./messages-CDZYGNlS.js";
+import "./util-mJ58qbbw.js";
+import "./responses-DLLjADw5.js";
+import "./openai-Dz3surb_.js";
+import { l as validateFunctionCall$1 } from "./util-Cl0zfT3V.js";
+import "./completion-D9_MDlnd.js";
+import { c as setUserEmail, i as getUserEmail, l as EMAIL_OK_STATUS, r as getAuthor, s as promptForEmailUnverified, t as checkEmailStatusAndMaybeExit } from "./accounts-R3gfCR_g.js";
+import { i as getRemoteGenerationUrl, l as shouldGenerateRemote, o as getRemoteHealthUrl, r as promptYesNo, s as neverGenerateRemote } from "./server-CbMTRQkg.js";
+import { t as getBlobByHash } from "./blobs-C9J2mVgo.js";
+import { a as evalsTable, c as evalsToTagsTable, d as tagsTable, i as evalResultsTable, l as promptsTable, m as getDbSignalPath, o as evalsToDatasetsTable, p as getDb, r as datasetsTable, s as evalsToPromptsTable } from "./tables-D-NSwNIb.js";
+import { n as telemetry_default } from "./telemetry-DZ_7PaVq.js";
+import { t as ellipsize } from "./text-Dm78AVGG.js";
+import { n as getTraceStore } from "./store-2ocbYY9D.js";
+import "./base-Dtp8b4_N.js";
+import "./image-BiEVdpdP.js";
+import { t as providerRegistry } from "./providerRegistry-wCWd7sKQ.js";
+import { n as runRuby } from "./rubyUtils-Cs35SDYa.js";
+import { t as EvalResult } from "./evalResult-COsVttMA.js";
 import * as fs$1 from "fs";
 import fs, { createWriteStream } from "fs";
 import * as path$2 from "path";
@@ -291,7 +292,7 @@ async function startOtlpReceiverIfNeeded(testSuite) {
 		telemetry_default.record("feature_used", { feature: "tracing" });
 		try {
 			logger_default.debug("[EvaluatorTracing] Tracing configuration detected, starting OTLP receiver");
-			const { startOTLPReceiver } = await import("./otlpReceiver-DRNetlJH.js");
+			const { startOTLPReceiver } = await import("./otlpReceiver-BO0rbDzh.js");
 			const port = testSuite.tracing.otlp.http.port || 4318;
 			const host = testSuite.tracing.otlp.http.host || "127.0.0.1";
 			logger_default.debug(`[EvaluatorTracing] Starting OTLP receiver on ${host}:${port}`);
@@ -314,7 +315,7 @@ async function startOtlpReceiverIfNeeded(testSuite) {
 async function stopOtlpReceiverIfNeeded() {
 	if (otlpReceiverStarted) try {
 		logger_default.debug("[EvaluatorTracing] Stopping OTLP receiver");
-		const { stopOTLPReceiver } = await import("./otlpReceiver-DRNetlJH.js");
+		const { stopOTLPReceiver } = await import("./otlpReceiver-BO0rbDzh.js");
 		await stopOTLPReceiver();
 		otlpReceiverStarted = false;
 		logger_default.info("[EvaluatorTracing] OTLP receiver stopped successfully");
@@ -349,7 +350,7 @@ async function generateTraceContextIfNeeded(test, evaluateOptions, testIdx, prom
 	}
 	if (!tracingEnabled) return null;
 	logger_default.debug("[EvaluatorTracing] Importing trace store");
-	const { getTraceStore } = await import("./store-KXWnHB15.js");
+	const { getTraceStore } = await import("./store-D_lq_8oQ.js");
 	const traceStore = getTraceStore();
 	const traceId = generateTraceId();
 	const spanId = generateSpanId();
@@ -2760,7 +2761,7 @@ const ASSERTION_HANDLERS = {
 	"llm-rubric": handleLlmRubric,
 	meteor: async (params) => {
 		try {
-			const { handleMeteorAssertion } = await import("./meteor-CBciquOS.js");
+			const { handleMeteorAssertion } = await import("./meteor-Dj8cTkU_.js");
 			return handleMeteorAssertion(params);
 		} catch (error) {
 			if (error instanceof Error && (error.message.includes("Cannot find module") || error.message.includes("natural\" package is required"))) return {
@@ -2870,9 +2871,9 @@ async function runAssertion({ prompt, provider, assertion, test, vars, latencyMs
 		let filePath = fileRef;
 		let functionName;
 		if (fileRef.includes(":")) {
-			const [pathPart, funcPart] = fileRef.split(":");
-			filePath = pathPart;
-			functionName = funcPart;
+			const colonIndex = fileRef.indexOf(":");
+			filePath = fileRef.slice(0, colonIndex);
+			functionName = fileRef.slice(colonIndex + 1);
 		}
 		filePath = path.resolve(basePath, filePath);
 		if (isJavascriptFile(filePath)) {
@@ -2890,7 +2891,7 @@ async function runAssertion({ prompt, provider, assertion, test, vars, latencyMs
 			};
 		}
 		else if (filePath.endsWith(".rb")) try {
-			const { runRuby } = await import("./rubyUtils-_t9Gmf7U.js");
+			const { runRuby } = await import("./rubyUtils-DRRiMFV2.js");
 			valueFromScript = await runRuby(filePath, functionName || "get_assert", [output, context]);
 			logger_default.debug(`Ruby script ${filePath} output: ${valueFromScript}`);
 		} catch (error) {
@@ -3898,7 +3899,7 @@ async function runEval({ provider, prompt, test, testSuite, delay, nunjucksFilte
 			promptIdx,
 			testIdx
 		});
-		logger_default.error("Provider call failed during eval", logContext);
+		if (!(err instanceof Error && err.name === "AbortError")) logger_default.error("Provider call failed during eval", logContext);
 		return [{
 			...setup,
 			error: errorWithStack,
@@ -4079,11 +4080,17 @@ var Evaluator = class {
 		let globalTimeout;
 		let globalAbortController;
 		const processedIndices = /* @__PURE__ */ new Set();
+		let targetUnavailable = false;
+		let targetErrorStatus;
+		const targetErrorAbortController = new AbortController();
 		let ciProgressReporter = null;
 		let progressBarManager = null;
+		let providerAbortSignal = options.abortSignal;
+		let combinedAbortSignal = options.abortSignal ? AbortSignal.any([options.abortSignal, targetErrorAbortController.signal]) : targetErrorAbortController.signal;
 		if (maxEvalTimeMs > 0) {
 			globalAbortController = new AbortController();
-			options.abortSignal = options.abortSignal ? AbortSignal.any([options.abortSignal, globalAbortController.signal]) : globalAbortController.signal;
+			providerAbortSignal = providerAbortSignal ? AbortSignal.any([providerAbortSignal, globalAbortController.signal]) : globalAbortController.signal;
+			combinedAbortSignal = AbortSignal.any([combinedAbortSignal, globalAbortController.signal]);
 			globalTimeout = setTimeout(() => {
 				evalTimedOut = true;
 				globalAbortController?.abort();
@@ -4091,7 +4098,7 @@ var Evaluator = class {
 		}
 		const vars = /* @__PURE__ */ new Set();
 		const checkAbort = () => {
-			if (options.abortSignal?.aborted) throw new Error("Operation cancelled");
+			if (combinedAbortSignal.aborted) throw new Error("Operation cancelled");
 		};
 		if (!options.silent) logger_default.info(`Starting evaluation ${this.evalRecord.id}`);
 		checkAbort();
@@ -4252,7 +4259,7 @@ var Evaluator = class {
 				const defaultProvider = testSuite.defaultTest.provider;
 				if (isApiProvider(defaultProvider)) testCase.provider = defaultProvider;
 				else if (typeof defaultProvider === "object" && defaultProvider.id) {
-					const { loadApiProvider } = await import("./providers-41mSodR_.js");
+					const { loadApiProvider } = await import("./providers-CQQrNaJk.js");
 					testCase.provider = await loadApiProvider(typeof defaultProvider.id === "function" ? defaultProvider.id() : defaultProvider.id, { options: defaultProvider });
 				} else testCase.provider = defaultProvider;
 			}
@@ -4322,7 +4329,7 @@ var Evaluator = class {
 							registers: this.registers,
 							isRedteam: testSuite.redteam != null,
 							concurrency,
-							abortSignal: options.abortSignal,
+							abortSignal: providerAbortSignal,
 							evalId: this.evalRecord.id,
 							rateLimitRegistry: this.rateLimitRegistry
 						});
@@ -4336,7 +4343,7 @@ var Evaluator = class {
 			if (evalOption.test.assert?.some((a) => a.type === "max-score")) rowsWithMaxScoreAssertion.add(evalOption.testIdx);
 		}
 		if (cliState_default.resume && this.evalRecord.persisted) try {
-			const { default: EvalResult } = await import("./evalResult-C8iVgVvr.js");
+			const { default: EvalResult } = await import("./evalResult-DyttNQ_G.js");
 			const completedPairs = await EvalResult.getCompletedIndexPairs(this.evalRecord.id, { excludeErrors: cliState_default.retryMode });
 			const originalCount = runEvalOptions.length;
 			for (let i = runEvalOptions.length - 1; i >= 0; i--) {
@@ -4389,6 +4396,14 @@ var Evaluator = class {
 					logger_default.error(`Error saving result: ${error} ${safeJsonStringify(resultSummary)}`);
 				}
 				for (const writer of this.fileWriters) await writer.write(row);
+				const httpStatus = row.response?.metadata?.http?.status;
+				if (typeof httpStatus === "number" && isNonTransientHttpStatus(httpStatus)) {
+					targetUnavailable = true;
+					targetErrorStatus = httpStatus;
+					logger_default.error(`Target returned HTTP ${httpStatus}. Aborting scan - this error will not resolve on retry.`);
+					targetErrorAbortController.abort();
+					break;
+				}
 				const { promptIdx } = row;
 				const metrics = prompts[promptIdx].metrics;
 				invariant(metrics, "Expected prompt.metrics to be set");
@@ -4552,6 +4567,7 @@ var Evaluator = class {
 		if (this.options.showProgressBar && progressBarManager) await progressBarManager.initialize(runEvalOptions, concurrency, 0);
 		try {
 			if (serialRunEvalOptions.length > 0) for (const evalStep of serialRunEvalOptions) {
+				checkAbort();
 				if (isWebUI) {
 					const provider = evalStep.provider.label || evalStep.provider.id();
 					const vars = formatVarsForDisplay(evalStep.test.vars || {}, 50);
@@ -4569,22 +4585,32 @@ var Evaluator = class {
 				await this.evalRecord.addPrompts(prompts);
 			});
 		} catch (err) {
-			if (options.abortSignal?.aborted) if (evalTimedOut) logger_default.warn(`Evaluation stopped after reaching max duration (${maxEvalTimeMs}ms)`);
-			else {
-				logger_default.info("Evaluation interrupted, saving progress...");
-				if (globalTimeout) clearTimeout(globalTimeout);
-				if (progressBarManager) progressBarManager.stop();
-				if (ciProgressReporter) ciProgressReporter.finish();
-				this.evalRecord.setVars(Array.from(vars));
-				await this.evalRecord.addPrompts(prompts);
-				updateSignalFile(this.evalRecord.id);
-				return this.evalRecord;
-			}
-			else {
+			if (combinedAbortSignal.aborted) {
+				if (evalTimedOut) logger_default.warn(`Evaluation stopped after reaching max duration (${maxEvalTimeMs}ms)`);
+				else if (!targetUnavailable) {
+					logger_default.info("Evaluation interrupted, saving progress...");
+					if (globalTimeout) clearTimeout(globalTimeout);
+					if (progressBarManager) progressBarManager.stop();
+					if (ciProgressReporter) ciProgressReporter.finish();
+					this.evalRecord.setVars(Array.from(vars));
+					await this.evalRecord.addPrompts(prompts);
+					updateSignalFile(this.evalRecord.id);
+					return this.evalRecord;
+				}
+			} else {
 				if (ciProgressReporter) ciProgressReporter.error(`Evaluation failed: ${String(err)}`);
 				throw err;
 			}
 		}
+		if (targetUnavailable) {
+			if (globalTimeout) clearTimeout(globalTimeout);
+			if (progressBarManager) progressBarManager.stop();
+			if (ciProgressReporter) ciProgressReporter.error(`Target unavailable (HTTP ${targetErrorStatus})`);
+			this.evalRecord.setVars(Array.from(vars));
+			await this.evalRecord.addPrompts(prompts);
+			updateSignalFile(this.evalRecord.id);
+			return this.evalRecord;
+		}
 		const compareRowsCount = rowsWithSelectBestAssertion.size + rowsWithMaxScoreAssertion.size;
 		if (progressBarManager) {
 			if (compareRowsCount > 0) progressBarManager.updateTotalCount(compareRowsCount);
@@ -5295,10 +5321,17 @@ function convertResultsToTable(eval_) {
 				if (keyToUpdate) result.vars[keyToUpdate] = actualPrompt;
 			}
 		}
-		if (result.metadata?.sessionId && !result.vars?.sessionId) {
-			result.vars = result.vars || {};
-			result.vars.sessionId = result.metadata.sessionId;
-			varsForHeader.add("sessionId");
+		if (!result.vars?.sessionId) {
+			const metadataSessionIds = result.metadata?.sessionIds;
+			if (Array.isArray(metadataSessionIds) && metadataSessionIds.length > 0) {
+				result.vars = result.vars || {};
+				result.vars.sessionId = metadataSessionIds.filter((id) => id != null && id !== "").map(String).join("\n");
+				varsForHeader.add("sessionId");
+			} else if (result.metadata?.sessionId) {
+				result.vars = result.vars || {};
+				result.vars.sessionId = result.metadata.sessionId;
+				varsForHeader.add("sessionId");
+			}
 		}
 		const transformDisplayVars = result.response?.metadata?.transformDisplayVars;
 		if (transformDisplayVars) {
@@ -5354,7 +5387,12 @@ function convertResultsToTable(eval_) {
 				model: result.response.video.model,
 				aspectRatio: result.response.video.aspectRatio,
 				resolution: result.response.video.resolution
-			} : void 0
+			} : void 0,
+			images: result.response?.images?.map((img) => ({
+				data: img.data,
+				blobRef: img.blobRef,
+				mimeType: img.mimeType
+			}))
 		};
 		invariant(result.promptId, "Prompt ID is required");
 		row.testIdx = result.testIdx;
@@ -5420,7 +5458,12 @@ function convertEvalResultToTableCell(result) {
 			model: result.response.video.model,
 			aspectRatio: result.response.video.aspectRatio,
 			resolution: result.response.video.resolution
-		} : void 0
+		} : void 0,
+		images: result.response?.images?.map((img) => ({
+			data: img.data,
+			blobRef: img.blobRef,
+			mimeType: img.mimeType
+		}))
 	};
 }
 function convertTestResultsToTableRow(results, varsForHeader) {
@@ -5429,10 +5472,13 @@ function convertTestResultsToTableRow(results, varsForHeader) {
 		outputs: [],
 		vars: Object.values(varsForHeader).map((varName) => {
 			if (varName === "sessionId") {
-				const sessionId = results[0].testCase.vars?.sessionId;
-				const varValue = sessionId == null || sessionId === "" ? results[0].metadata?.sessionId ?? "" : sessionId;
-				if (typeof varValue === "string") return varValue;
-				return JSON.stringify(varValue);
+				const sessionIdFromVars = results[0].testCase.vars?.sessionId;
+				if (sessionIdFromVars != null && sessionIdFromVars !== "") return typeof sessionIdFromVars === "string" ? sessionIdFromVars : JSON.stringify(sessionIdFromVars);
+				const metadataSessionIds = results[0].metadata?.sessionIds;
+				if (Array.isArray(metadataSessionIds) && metadataSessionIds.length > 0) return metadataSessionIds.filter((id) => id != null && id !== "").map(String).join("\n");
+				const metadataSessionId = results[0].metadata?.sessionId;
+				if (metadataSessionId != null) return typeof metadataSessionId === "string" ? metadataSessionId : JSON.stringify(metadataSessionId);
+				return "";
 			}
 			const varValue = results[0].testCase.vars?.[varName] ?? "";
 			if (typeof varValue === "string") return varValue;
@@ -5712,7 +5758,13 @@ var Eval = class Eval {
 	_resultsLoaded = false;
 	runtimeOptions;
 	_shared = false;
+	/** Total wall-clock duration. For redteam evals: generationDurationMs + evaluationDurationMs.
+	*  For non-redteam evals: equals evaluationDurationMs (generation phase is N/A). */
 	durationMs;
+	/** Time spent generating adversarial test cases (redteam only). */
+	generationDurationMs;
+	/** Time spent running the evaluation phase. */
+	evaluationDurationMs;
 	/**
 	* The shareable URL for this evaluation, if it has been shared.
 	* Set by the evaluate() function when sharing is enabled.
@@ -5731,8 +5783,11 @@ var Eval = class Eval {
 		const eval_ = evalData[0];
 		const datasetId = datasetResults[0]?.datasetId;
 		const resultsObj = eval_.results;
-		const rawDurationMs = resultsObj && "durationMs" in resultsObj ? resultsObj.durationMs : void 0;
-		const durationMs = typeof rawDurationMs === "number" && Number.isFinite(rawDurationMs) && rawDurationMs >= 0 ? rawDurationMs : void 0;
+		const validateDuration = (raw) => typeof raw === "number" && Number.isFinite(raw) && raw >= 0 ? raw : void 0;
+		const rawDurationMs = validateDuration(resultsObj?.["durationMs"]);
+		const generationDurationMs = validateDuration(resultsObj?.["generationDurationMs"]);
+		const evaluationDurationMs = validateDuration(resultsObj?.["evaluationDurationMs"]);
+		const durationMs = rawDurationMs ?? (generationDurationMs != null || evaluationDurationMs != null ? (generationDurationMs ?? 0) + (evaluationDurationMs ?? 0) : void 0);
 		const evalInstance = new Eval(eval_.config, {
 			id: eval_.id,
 			createdAt: new Date(eval_.createdAt),
@@ -5743,7 +5798,9 @@ var Eval = class Eval {
 			persisted: true,
 			vars: eval_.vars || [],
 			runtimeOptions: eval_.runtimeOptions ?? void 0,
-			durationMs
+			durationMs,
+			generationDurationMs,
+			evaluationDurationMs
 		});
 		if (eval_.results && "table" in eval_.results) evalInstance.oldResults = eval_.results;
 		if (!eval_.vars || eval_.vars.length === 0) {
@@ -5868,6 +5925,8 @@ var Eval = class Eval {
 		this.vars = opts?.vars || [];
 		this.runtimeOptions = opts?.runtimeOptions;
 		this.durationMs = opts?.durationMs;
+		this.generationDurationMs = opts?.generationDurationMs;
+		this.evaluationDurationMs = opts?.evaluationDurationMs;
 	}
 	version() {
 		/**
@@ -5898,7 +5957,13 @@ var Eval = class Eval {
 		if (this.useOldResults()) {
 			invariant(this.oldResults, "Old results not found");
 			updateObj.results = this.oldResults;
-		} else if (this.durationMs !== void 0) updateObj.results = { durationMs: this.durationMs };
+		} else if (this.durationMs !== void 0 || this.generationDurationMs !== void 0 || this.evaluationDurationMs !== void 0) {
+			let expr = sql`CASE WHEN json_valid(${evalsTable.results}) AND json_type(${evalsTable.results}) = 'object' THEN ${evalsTable.results} ELSE '{}' END`;
+			if (this.durationMs !== void 0) expr = sql`json_set(${expr}, '$.durationMs', ${this.durationMs})`;
+			if (this.generationDurationMs !== void 0) expr = sql`json_set(${expr}, '$.generationDurationMs', ${this.generationDurationMs})`;
+			if (this.evaluationDurationMs !== void 0) expr = sql`json_set(${expr}, '$.evaluationDurationMs', ${this.evaluationDurationMs})`;
+			updateObj.results = expr;
+		}
 		db.update(evalsTable).set(updateObj).where(eq(evalsTable.id, this.id)).run();
 		this.persisted = true;
 	}
@@ -5908,8 +5973,17 @@ var Eval = class Eval {
 	addVar(varName) {
 		this.vars.push(varName);
 	}
+	/** Sets the evaluation phase duration and recomputes the total. Called by the evaluator. */
 	setDurationMs(durationMs) {
-		this.durationMs = durationMs;
+		if (!Number.isFinite(durationMs) || durationMs < 0) return;
+		this.evaluationDurationMs = durationMs;
+		this.durationMs = (this.generationDurationMs ?? 0) + durationMs;
+	}
+	/** Sets the generation phase duration and recomputes the total. Called by doRedteamRun. */
+	setGenerationDurationMs(durationMs) {
+		if (!Number.isFinite(durationMs) || durationMs < 0) return;
+		this.generationDurationMs = durationMs;
+		this.durationMs = durationMs + (this.evaluationDurationMs ?? 0);
 	}
 	getPrompts() {
 		if (this.useOldResults()) {
@@ -5947,6 +6021,27 @@ var Eval = class Eval {
 	async getTotalResultRowCount() {
 		return getTotalResultRowCount(this.id);
 	}
+	/**
+	* Find a non-transient HTTP error status from evaluation results.
+	* Returns the first non-transient status (401, 403, 404, 500, 501) found, or undefined.
+	*
+	* For persisted evals: Uses efficient O(1) database query with LIMIT 1.
+	* For non-persisted evals: Falls back to scanning in-memory results.
+	*/
+	async findTargetErrorStatus() {
+		const scanInMemory = () => {
+			for (const result of this.results) {
+				const status = result.response?.metadata?.http?.status;
+				if (typeof status === "number" && isNonTransientHttpStatus(status)) return status;
+			}
+		};
+		if (!this.persisted) return scanInMemory();
+		try {
+			return getDb().select({ httpStatus: sql`CAST(json_extract(${evalResultsTable.response}, '$.metadata.http.status') AS INTEGER)` }).from(evalResultsTable).where(and(eq(evalResultsTable.evalId, this.id), sql`json_extract(${evalResultsTable.response}, '$.metadata.http.status') IN (${sql.join(NON_TRANSIENT_HTTP_STATUSES.map((s) => sql`${s}`), sql`, `)})`)).limit(1).get()?.httpStatus ?? void 0;
+		} catch {
+			return scanInMemory();
+		}
+	}
 	async fetchResultsByTestIdx(testIdx) {
 		return await EvalResult.findManyByEvalId(this.id, { testIdx });
 	}
@@ -6169,7 +6264,12 @@ var Eval = class Eval {
 			};
 		}
 		const allResults = await EvalResult.findManyByEvalIdAndTestIndices(this.id, testIndices);
-		if (allResults.some((result) => result.metadata?.sessionId && !result.testCase?.vars?.sessionId) && !vars.includes("sessionId")) {
+		if (allResults.some((result) => {
+			const hasSessionIds = Array.isArray(result.metadata?.sessionIds) && result.metadata.sessionIds.length > 0;
+			const hasSessionId = Boolean(result.metadata?.sessionId);
+			const notInVars = !result.testCase?.vars?.sessionId;
+			return (hasSessionIds || hasSessionId) && notInVars;
+		}) && !vars.includes("sessionId")) {
 			vars.push("sessionId");
 			vars.sort();
 		}
@@ -6230,7 +6330,9 @@ var Eval = class Eval {
 			failures: 0,
 			errors: 0,
 			tokenUsage: createEmptyTokenUsage(),
-			durationMs: this.durationMs
+			durationMs: this.durationMs,
+			generationDurationMs: this.generationDurationMs,
+			evaluationDurationMs: this.evaluationDurationMs
 		};
 		for (const prompt of this.prompts) {
 			stats.successes += prompt.metrics?.testPassCount ?? 0;
@@ -7528,6 +7630,24 @@ function validateTestProviderReferences(tests, providers, defaultTest, scenarios
 	});
 }
 
+//#endregion
+//#region src/util/config/extensions.ts
+/**
+* Supported config file extensions, sorted by frequency of use.
+* Order matters: loaders try each in sequence and stop at the first match.
+*/
+const DEFAULT_CONFIG_EXTENSIONS = [
+	"yaml",
+	"yml",
+	"json",
+	"cjs",
+	"cts",
+	"js",
+	"mjs",
+	"mts",
+	"ts"
+];
+
 //#endregion
 //#region src/util/config/load.ts
 /**
@@ -7536,6 +7656,34 @@ function validateTestProviderReferences(tests, providers, defaultTest, scenarios
 function isTestCaseWithVars(test) {
 	return typeof test === "object" && test !== null && "vars" in test;
 }
+/**
+* When --providers is used alongside a config file that has providers defined,
+* maps each CLI provider token to a matching config provider (preserving its config
+* options like num_ctx, temperature). Unmatched tokens are kept as bare strings.
+*
+* Matching priority per token:
+* 1. Exact match on provider id
+* 2. Exact match on provider label
+* 3. Provider-prefix match: config id ends with `:cliProvider` (e.g. CLI `llama3.1:8b`
+*    matches config `ollama:llama3.1:8b`). First match wins if multiple configs share a suffix.
+* 4. No match: keep raw CLI string for fresh provider creation
+*/
+function resolveCliProvidersWithConfig(cliProviders, configProviders) {
+	if (!configProviders || !Array.isArray(configProviders)) return cliProviders;
+	const indexed = configProviders.map((cp, i) => ({
+		provider: cp,
+		...getProviderIdAndLabel(cp, i)
+	}));
+	return cliProviders.map((cliProvider) => {
+		const exactId = indexed.find((entry) => entry.id === cliProvider);
+		if (exactId) return exactId.provider;
+		const exactLabel = indexed.find((entry) => entry.label === cliProvider);
+		if (exactLabel) return exactLabel.provider;
+		const prefixMatch = indexed.find((entry) => entry.id.endsWith(":" + cliProvider));
+		if (prefixMatch) return prefixMatch.provider;
+		return cliProvider;
+	});
+}
 async function dereferenceConfig(rawConfig) {
 	if (getEnvBool("PROMPTFOO_DISABLE_REF_PARSER")) return rawConfig;
 	const extractFunctionParameters = (functions) => {
@@ -7672,7 +7820,7 @@ async function combineConfigs(configPaths) {
 	const configs = [];
 	for (const configPath of configPaths) {
 		const globPaths = globSync(path$2.resolve(process$1.cwd(), configPath), { windowsPathsNoEscape: true });
-		if (globPaths.length === 0) throw new Error(`No configuration file found at ${configPath}. Run "promptfoo init" to create one or pass --config path/to/promptfooconfig.yaml.`);
+		if (globPaths.length === 0) throw new Error(`No configuration file found at ${configPath}. Run "${promptfooCommand("init")}" to create one or pass --config path/to/promptfooconfig.yaml.`);
 		for (const globPath of globPaths) {
 			const config = await readConfig(globPath);
 			configs.push(config);
@@ -7885,7 +8033,7 @@ async function resolveConfigs(cmdObj, _defaultConfig, type) {
 		tags: fileConfig.tags || defaultConfig.tags,
 		description: cmdObj.description || fileConfig.description || defaultConfig.description,
 		prompts: cmdObj.prompts || fileConfig.prompts || defaultConfig.prompts || [],
-		providers: cmdObj.providers || fileConfig.providers || defaultConfig.providers || [],
+		providers: fileConfig.providers || defaultConfig.providers || [],
 		tests: cmdObj.tests || cmdObj.vars || fileConfig.tests || defaultConfig.tests || [],
 		scenarios: fileConfig.scenarios || defaultConfig.scenarios,
 		env: fileConfig.env || defaultConfig.env,
@@ -7900,11 +8048,14 @@ async function resolveConfigs(cmdObj, _defaultConfig, type) {
 		evaluateOptions: fileConfig.evaluateOptions || defaultConfig.evaluateOptions
 	};
 	const hasPrompts = [config.prompts].flat().filter(Boolean).length > 0;
-	const hasProviders = [config.providers].flat().filter(Boolean).length > 0;
+	const hasProviders = cmdObj.providers && cmdObj.providers.length > 0 || [config.providers].flat().filter(Boolean).length > 0;
 	if (!Boolean(configPaths) && !hasPrompts && !hasProviders && !isCI()) {
+		const extList = DEFAULT_CONFIG_EXTENSIONS.join(", ");
 		logger_default.warn(dedent`
       ${chalk.yellow.bold("⚠️  No promptfooconfig found")}
 
+      ${chalk.white(`Searched in ${chalk.bold(process$1.cwd())} for promptfooconfig.{${extList}}`)}
+
       ${chalk.white("Try running with:")}
 
       ${chalk.cyan(`${promptfooCommand("")} eval -c ${chalk.bold("path/to/promptfooconfig.yaml")}`)}
@@ -7925,8 +8076,9 @@ async function resolveConfigs(cmdObj, _defaultConfig, type) {
 	}
 	invariant(Array.isArray(config.providers), "providers must be an array");
 	const resolvedProviderConfigs = resolveProviderConfigs(config.providers, { basePath });
+	const cliFilteredProviderConfigs = (cmdObj.providers ? resolveCliProvidersWithConfig(cmdObj.providers, resolvedProviderConfigs) : resolvedProviderConfigs) ?? [];
 	const filterOption = cmdObj.filterProviders || cmdObj.filterTargets;
-	const filteredProviderConfigs = filterProviderConfigs(resolvedProviderConfigs, filterOption);
+	const filteredProviderConfigs = filterProviderConfigs(cliFilteredProviderConfigs, filterOption);
 	if (filterOption && Array.isArray(filteredProviderConfigs) && filteredProviderConfigs.length === 0) logger_default.warn(`No providers matched the filter "${filterOption}". Check your --filter-providers/--filter-targets value.`);
 	let parsedPrompts = await readPrompts(config.prompts, cmdObj.prompts ? void 0 : basePath);
 	if (cmdObj.filterPrompts) {
@@ -7962,7 +8114,7 @@ async function resolveConfigs(cmdObj, _defaultConfig, type) {
 	}
 	const parsedProviderPromptMap = readProviderPromptMap({ providers: filteredProviderConfigs }, parsedPrompts);
 	if (parsedPrompts.length === 0) {
-		logger_default.error("No prompts found");
+		logger_default.error("No prompts found. Add a `prompts:` entry to your config or pass --prompts path/to/prompt.txt.");
 		process$1.exit(1);
 	}
 	const defaultTest = {
@@ -9570,6 +9722,7 @@ async function synthesize({ abortSignal, delay, entities: entitiesOverride, inje
 		seen.add(key);
 		return true;
 	});
+	const needsGoalExtraction = strategies.some((s) => Strategies.find((def) => def.id === s.id)?.requiresGoalExtraction);
 	await validateStrategies(strategies);
 	await validateSharpDependency(strategies, plugins);
 	const redteamProvider = await redteamProviderManager.getProvider({ provider });
@@ -9761,13 +9914,15 @@ async function synthesize({ abortSignal, delay, entities: entitiesOverride, inje
 			if (!Array.isArray(allPluginTests) || allPluginTests.length === 0) logger_default.warn(`Failed to generate tests for ${plugin.id}`);
 			else {
 				const testCasesWithMetadata = allPluginTests;
-				logger_default.debug(`Extracting goal for ${testCasesWithMetadata.length} tests from ${plugin.id}...`);
-				for (const testCase of testCasesWithMetadata) {
-					const promptVar = testCase.vars?.[injectVar];
-					const prompt = Array.isArray(promptVar) ? promptVar[0] : String(promptVar);
-					const policy = getPolicyText(testCase.metadata);
-					const extractedGoal = await extractGoalFromPrompt(prompt, purpose, plugin.id, policy);
-					testCase.metadata.goal = extractedGoal;
+				if (needsGoalExtraction) {
+					logger_default.debug(`Extracting goal for ${testCasesWithMetadata.length} tests from ${plugin.id}...`);
+					for (const testCase of testCasesWithMetadata) {
+						const promptVar = testCase.vars?.[injectVar];
+						const prompt = Array.isArray(promptVar) ? promptVar[0] : String(promptVar);
+						const policy = getPolicyText(testCase.metadata);
+						const extractedGoal = await extractGoalFromPrompt(prompt, purpose, plugin.id, policy);
+						testCase.metadata.goal = extractedGoal;
+					}
 				}
 				testCases.push(...testCasesWithMetadata);
 			}
@@ -9802,13 +9957,15 @@ async function synthesize({ abortSignal, delay, entities: entitiesOverride, inje
 					...t.metadata || {}
 				}
 			}));
-			logger_default.debug(`Extracting goal for ${testCasesWithMetadata.length} custom tests from ${plugin.id}...`);
-			for (const testCase of testCasesWithMetadata) {
-				const promptVar = testCase.vars?.[injectVar];
-				const prompt = Array.isArray(promptVar) ? promptVar[0] : String(promptVar);
-				const policy = getPolicyText(testCase.metadata);
-				const extractedGoal = await extractGoalFromPrompt(prompt, purpose, plugin.id, policy);
-				testCase.metadata.goal = extractedGoal;
+			if (needsGoalExtraction) {
+				logger_default.debug(`Extracting goal for ${testCasesWithMetadata.length} custom tests from ${plugin.id}...`);
+				for (const testCase of testCasesWithMetadata) {
+					const promptVar = testCase.vars?.[injectVar];
+					const prompt = Array.isArray(promptVar) ? promptVar[0] : String(promptVar);
+					const policy = getPolicyText(testCase.metadata);
+					const extractedGoal = await extractGoalFromPrompt(prompt, purpose, plugin.id, policy);
+					testCase.metadata.goal = extractedGoal;
+				}
 			}
 			testCases.push(...testCasesWithMetadata);
 			logger_default.debug(`Added ${customTests.length} custom test cases from ${plugin.id}`);
@@ -10831,17 +10988,7 @@ async function loadDefaultConfig(dir, configName = "promptfooconfig") {
 	if (configCache.has(cacheKey)) return configCache.get(cacheKey);
 	let defaultConfig = {};
 	let defaultConfigPath;
-	for (const ext of [
-		"yaml",
-		"yml",
-		"json",
-		"cjs",
-		"cts",
-		"js",
-		"mjs",
-		"mts",
-		"ts"
-	]) {
+	for (const ext of DEFAULT_CONFIG_EXTENSIONS) {
 		const configPath = path.join(dir, `${configName}.${ext}`);
 		const maybeConfig = await maybeReadConfig(configPath);
 		if (maybeConfig) {
@@ -10947,11 +11094,31 @@ function formatDuration(seconds) {
 * ```
 */
 function generateEvalSummary(params) {
-	const { evalId, isRedteam, writeToDatabase, shareableUrl, wantsToShare, hasExplicitDisable, cloudEnabled, activelySharing = false, tokenUsage, successes, failures, errors, duration, maxConcurrency, tracker } = params;
+	const { evalId, isRedteam, writeToDatabase, shareableUrl, wantsToShare, hasExplicitDisable, cloudEnabled, activelySharing = false, tokenUsage, successes, failures, errors, duration, maxConcurrency, tracker, targetErrorStatus } = params;
 	const lines = [];
 	const completionType = isRedteam ? "Red team" : "Eval";
-	const completionMessage = writeToDatabase && shareableUrl ? `${chalk.green("✓")} ${completionType} complete: ${shareableUrl}` : writeToDatabase && activelySharing ? `${chalk.green("✓")} ${completionType} complete` : writeToDatabase ? `${chalk.green("✓")} ${completionType} complete (ID: ${chalk.cyan(evalId)})` : `${chalk.green("✓")} ${completionType} complete`;
+	const wasAborted = targetErrorStatus != null;
+	let completionMessage;
+	if (wasAborted) {
+		completionMessage = `${chalk.red("✗")} ${completionType} aborted`;
+		if (writeToDatabase) completionMessage += ` (ID: ${chalk.cyan(evalId)})`;
+	} else if (writeToDatabase && shareableUrl) completionMessage = `${chalk.green("✓")} ${completionType} complete: ${shareableUrl}`;
+	else if (writeToDatabase && activelySharing) completionMessage = `${chalk.green("✓")} ${completionType} complete`;
+	else if (writeToDatabase) completionMessage = `${chalk.green("✓")} ${completionType} complete (ID: ${chalk.cyan(evalId)})`;
+	else completionMessage = `${chalk.green("✓")} ${completionType} complete`;
 	lines.push(completionMessage);
+	if (wasAborted && targetErrorStatus != null) {
+		lines.push("");
+		lines.push(chalk.red.bold("Scan stopped: Target is unavailable and will not recover on retry."));
+		lines.push(chalk.red(`  Target returned HTTP ${targetErrorStatus}`));
+		lines.push("");
+		lines.push(chalk.yellow("Possible causes:"));
+		lines.push(chalk.yellow("  • Invalid API key or authentication (401/403)"));
+		lines.push(chalk.yellow("  • Target endpoint does not exist (404)"));
+		lines.push(chalk.yellow("  • Server does not support the request (501)"));
+		lines.push("");
+		lines.push(chalk.cyan("To fix: Check your target configuration and credentials."));
+	}
 	if (writeToDatabase && !shareableUrl && !wantsToShare && !activelySharing) {
 		lines.push("");
 		lines.push(`» View results: ${chalk.green.bold("promptfoo view")}`);
@@ -11194,6 +11361,21 @@ async function doEval(cmdObj, defaultConfig, defaultConfigPath, evaluateOptions)
 	let testSuite = void 0;
 	let _basePath = void 0;
 	let commandLineOptions = void 0;
+	const configArgs = Array.isArray(cmdObj.config) ? cmdObj.config : typeof cmdObj.config === "string" ? [cmdObj.config] : [];
+	const uuidConfigArgs = configArgs.filter((configArg) => isUuid(configArg));
+	if (configArgs.length > 1 && uuidConfigArgs.length > 0) throw new Error("Cloud config UUID mode supports exactly one -c value. Use: promptfoo eval -c <cloud-config-uuid>");
+	if (configArgs.length === 1 && uuidConfigArgs.length === 1) {
+		const cloudConfigId = uuidConfigArgs[0];
+		if (cmdObj.watch) throw new Error("--watch is not supported when using a cloud config UUID with -c. Use a local config file path for watch mode.");
+		try {
+			defaultConfig = await getEvalConfigFromCloud(cloudConfigId);
+		} catch (error) {
+			const reason = error instanceof Error ? error.message : String(error);
+			throw new Error(`Failed to load cloud eval config "${cloudConfigId}". ${reason}. Cloud UUID inputs do not fall back to local file paths. Check authentication and that the UUID exists.`);
+		}
+		cmdObj.config = void 0;
+		defaultConfigPath = void 0;
+	}
 	const runEvaluation = async (initialization) => {
 		const startTime = Date.now();
 		telemetry_default.record("command_used", {
@@ -11217,7 +11399,7 @@ async function doEval(cmdObj, defaultConfig, defaultConfigPath, evaluateOptions)
 						...defaultConfig,
 						...dirConfig
 					};
-				} else logger_default.warn(`No configuration file found in directory: ${configPath}`);
+				} else logger_default.warn(`No configuration file found in directory: ${configPath}. Looked for promptfooconfig.{${DEFAULT_CONFIG_EXTENSIONS.join(",")}}. Run "${promptfooCommand("init")}" or pass --config path/to/promptfooconfig.yaml.`);
 			}
 		}
 		const resumeRaw = cmdObj.resume;
@@ -11497,6 +11679,7 @@ async function doEval(cmdObj, defaultConfig, defaultConfigPath, evaluateOptions)
 		const isRedteam = Boolean(config.redteam);
 		const duration = Math.round((Date.now() - startTime) / 1e3);
 		const tracker = TokenUsageTracker.getInstance();
+		const targetErrorStatus = await evalRecord.findTargetErrorStatus();
 		const summaryLines = generateEvalSummary({
 			evalId: evalRecord.id,
 			isRedteam,
@@ -11512,7 +11695,8 @@ async function doEval(cmdObj, defaultConfig, defaultConfigPath, evaluateOptions)
 			errors,
 			duration,
 			maxConcurrency,
-			tracker
+			tracker,
+			targetErrorStatus
 		});
 		if (cmdObj.write && wantsToShare && !canShareEval) {
 			logger_default.info(summaryLines[0]);
@@ -11567,7 +11751,7 @@ async function doEval(cmdObj, defaultConfig, defaultConfigPath, evaluateOptions)
 			if (initialization) {
 				const configPaths = (cmdObj.config || [defaultConfigPath]).filter(Boolean);
 				if (!configPaths.length) {
-					logger_default.error("Could not locate config file(s) to watch");
+					logger_default.error(`Could not locate config file(s) to watch. Pass --config path/to/promptfooconfig.yaml or run from a directory containing promptfooconfig.{${DEFAULT_CONFIG_EXTENSIONS.join(",")}}.`);
 					process.exitCode = 1;
 					return ret;
 				}
@@ -11725,6 +11909,7 @@ async function doRedteamRun(options) {
 	logger_default.info("Generating test cases...");
 	const { maxConcurrency, ...passThroughOptions } = options;
 	let redteamConfig;
+	const generationStartTime = Date.now();
 	try {
 		redteamConfig = await doGenerateRedteam({
 			...passThroughOptions,
@@ -11748,6 +11933,7 @@ async function doRedteamRun(options) {
 		}
 		throw error;
 	}
+	const generationDurationMs = Date.now() - generationStartTime;
 	if (!redteamConfig || !fs$1.existsSync(redteamPath)) {
 		logger_default.info("No test cases generated. Skipping scan.");
 		if (verboseToggleCleanup) verboseToggleCleanup();
@@ -11770,7 +11956,14 @@ async function doRedteamRun(options) {
 		abortSignal: options.abortSignal,
 		progressCallback: options.progressCallback
 	});
-	logger_default.info(chalk.green("\nRed team scan complete!"));
+	if (evalResult && generationDurationMs >= 0) {
+		evalResult.setGenerationDurationMs(generationDurationMs);
+		if (evalResult.persisted) await evalResult.save();
+		const totalMs = evalResult.durationMs ?? 0;
+		const evalMs = evalResult.evaluationDurationMs ?? 0;
+		logger_default.info(chalk.gray(`Total scan time: ${formatDuration(totalMs / 1e3)} (generation: ${formatDuration(generationDurationMs / 1e3)}, evaluation: ${formatDuration(evalMs / 1e3)})`));
+	}
+	if (evalResult ? await evalResult.findTargetErrorStatus() != null : false) {} else logger_default.info(chalk.green("\nRed team scan complete!"));
 	if (!evalResult?.shared) if (options.liveRedteamConfig) logger_default.info(chalk.blue(`To view the results, click the ${chalk.bold("View Report")} button or run ${chalk.bold(promptfooCommand("redteam report"))} on the command line.`));
 	else logger_default.info(chalk.blue(`To view the results, run ${chalk.bold(promptfooCommand("redteam report"))}`));
 	setLogCallback(null);