promptfoo 0.121.7 → 0.121.9

package/dist/src/{graders-BX0f2tvS.js → graders-GNRQ1357.js}

@@ -1,32 +1,32 @@
 import { t as __exportAll } from "./chunk-DEq-mXcV.js";
-import { O as isCI, S as getEnvBool, T as getEnvString, a as logger, h as extractJsonObjects, k as state, m as extractFirstJsonObject, y as safeJsonStringify } from "./logger-Ct2S6Yx-.js";
+import { O as isCI, S as getEnvBool, T as getEnvString, a as logger, h as extractJsonObjects, k as state, m as extractFirstJsonObject, y as safeJsonStringify } from "./logger-Cs0I2v_y.js";
 import { t as invariant } from "./invariant-Ddh24eXh.js";
 import { Et as CODING_AGENT_PLUGIN_DISPLAY_NAMES, L as PolicyObjectSchema, Tt as CODING_AGENT_PLUGIN_DESCRIPTIONS, W as MULTI_TURN_STRATEGIES, Wt as isApiProvider, j as PromptSchema, wt as CODING_AGENT_PLUGINS, z as isValidReusablePolicyId, zt as buildInputPromptDescription } from "./types-D6glLbdF.js";
-import { l as sleep, p as REQUEST_TIMEOUT_MS, r as fetchWithTimeout, t as fetchWithProxy } from "./fetch-It34O8Ur.js";
+import { l as sleep, m as getRequestTimeoutMs, r as fetchWithTimeout, t as fetchWithProxy } from "./fetch-CHJu2KON.js";
 import { i as isJavascriptFile } from "./fileExtensions-CXRfY3Ss.js";
-import { i as getUserEmail } from "./accounts-Ca7WIoPY.js";
-import { r as importModule } from "./esm-BTK1W7lG.js";
-import { i as extractVariablesFromTemplate, o as getNunjucksEngine } from "./render-DFfDeYUK.js";
-import { d as hasCodexDefaultCredentials, l as shouldGenerateRemote, r as getRemoteGenerationUrl, u as getCodexDefaultProviders } from "./remoteGeneration-DsaSwmG2.js";
-import { l as getPoliciesFromCloud } from "./storage-CD-GWAdx.js";
-import { r as runPython } from "./pythonUtils-D0BYebvX.js";
-import { N as parseFileUrl, O as maybeLoadToolsFromExternalFile, S as getNunjucksEngineForFilePath, T as maybeLoadFromExternalFile, k as parsePathOrGlob, w as maybeLoadConfigFromExternalFile } from "./util-Bx677_k2.js";
+import { i as getUserEmail } from "./accounts-4Xa5fcHy.js";
+import { r as importModule } from "./esm-C_yrt93A.js";
+import { i as extractVariablesFromTemplate, o as getNunjucksEngine } from "./render-CTZQAINJ.js";
+import { d as hasCodexDefaultCredentials, l as shouldGenerateRemote, r as getRemoteGenerationUrl, u as getCodexDefaultProviders } from "./remoteGeneration-DvWEm3yi.js";
+import { l as getPoliciesFromCloud } from "./storage-fbHNSuZ9.js";
+import { r as runPython } from "./pythonUtils-C4Wnz6Q7.js";
+import { N as parseFileUrl, O as maybeLoadToolsFromExternalFile, S as getNunjucksEngineForFilePath, T as maybeLoadFromExternalFile, k as parsePathOrGlob, w as maybeLoadConfigFromExternalFile } from "./util-vaCPFatZ.js";
 import { n as sha256 } from "./createHash-BtbSX3mj.js";
-import { i as getCache, o as isCacheEnabled, r as fetchWithCache } from "./cache-DIXbtkNO.js";
-import { t as OpenAiChatCompletionProvider } from "./chat-Dabu84Br.js";
-import { x as hasGoogleDefaultCredentials } from "./transform-DmvYBRll.js";
-import { C as AzureModerationProvider, S as DefaultSynthesizeProvider$1, T as AzureChatCompletionProvider, b as DefaultLlmRubricProvider, c as parseScriptParts, d as MistralEmbeddingProvider, l as OpenAiModerationProvider, m as DefaultGradingProvider$3, n as loadApiProvider, p as DefaultEmbeddingProvider$2, s as getFileHashes, u as MistralChatCompletionProvider, v as DefaultGradingJsonProvider$2, w as AzureEmbeddingProvider, x as DefaultSuggestionsProvider$2, y as DefaultGradingProvider$2 } from "./providers-DVYRZP4E.js";
-import { t as OpenAiEmbeddingProvider } from "./embedding-Bi3rxrZF.js";
+import { i as getCache, o as isCacheEnabled, r as fetchWithCache } from "./cache-CT5D7d_A.js";
+import { t as OpenAiChatCompletionProvider } from "./chat-CCtIXezK.js";
+import { x as hasGoogleDefaultCredentials } from "./transform-R1PnNdLv.js";
+import { C as AzureModerationProvider, S as DefaultSynthesizeProvider$1, T as AzureChatCompletionProvider, b as DefaultLlmRubricProvider, c as parseScriptParts, d as MistralEmbeddingProvider, l as OpenAiModerationProvider, m as DefaultGradingProvider$3, n as loadApiProvider, p as DefaultEmbeddingProvider$2, s as getFileHashes, u as MistralChatCompletionProvider, v as DefaultGradingJsonProvider$2, w as AzureEmbeddingProvider, x as DefaultSuggestionsProvider$2, y as DefaultGradingProvider$2 } from "./providers-DTPwadhk.js";
+import { t as OpenAiEmbeddingProvider } from "./embedding-cd5oPBkX.js";
 import { r as accumulateTokenUsage } from "./tokenUsageUtils-CmnQ0G2m.js";
-import { t as AnthropicMessagesProvider } from "./messages-vsE_-Lv0.js";
-import { t as OpenAiResponsesProvider } from "./responses-URRzV8qE.js";
-import { n as materializeInputVariablesWithMetadata } from "./inputVariables-DUGMb9Ka.js";
-import { a as extractPromptFromTags, c as isBasicRefusal, i as extractMaterializedVariablesFromJsonWithMetadata, l as isEmptyResponse, n as extractGoalFromPrompt, r as extractInputVarsFromPrompt, s as getShortPluginId, t as extractAllPromptsFromTags, u as removePrefix } from "./util-DNl96nNs.js";
-import { _ as isRateLimitWrapped, f as redteamProviderManager, g as createProviderRateLimitOptions } from "./shared-D6IjElRI.js";
-import { a as PROMPT_DELIMITER, n as maybeFilePath, r as normalizeInput } from "./utils-BGY69tk_.js";
-import { n as getGeneratedPromptOverLimit, r as getMaxCharsPerMessageModifierValue, t as MAX_CHARS_PER_MESSAGE_MODIFIER_KEY } from "./promptLength-B9nZEfO6.js";
+import { t as AnthropicMessagesProvider } from "./messages-Dy5vvc3H.js";
+import { t as OpenAiResponsesProvider } from "./responses-C2vhwf8F.js";
+import { n as materializeInputVariablesWithMetadata } from "./inputVariables-B7cylqzM.js";
+import { a as extractPromptFromTags, c as isBasicRefusal, i as extractMaterializedVariablesFromJsonWithMetadata, l as isEmptyResponse, n as extractGoalFromPrompt, r as extractInputVarsFromPrompt, s as getShortPluginId, t as extractAllPromptsFromTags, u as removePrefix } from "./util-D1NKMz55.js";
+import { _ as isRateLimitWrapped, f as redteamProviderManager, g as createProviderRateLimitOptions } from "./shared-BF25a7ms.js";
+import { a as PROMPT_DELIMITER, n as maybeFilePath, r as normalizeInput } from "./utils-C3qMBhJT.js";
+import { n as getGeneratedPromptOverLimit, r as getMaxCharsPerMessageModifierValue, t as MAX_CHARS_PER_MESSAGE_MODIFIER_KEY } from "./promptLength-ho3wbst6.js";
 import { t as REDTEAM_MEMORY_POISONING_PLUGIN_ID } from "./constants-DZGEFLsu.js";
-import { n as checkExfilTracking } from "./indirectWebPwn-CbjUG0rh.js";
+import { n as checkExfilTracking } from "./indirectWebPwn-1gHqaL91.js";
 import * as fs$2 from "fs";
 import fs from "fs";
 import path from "path";
@@ -276,13 +276,13 @@ const DefaultSuggestionsProvider$1 = new MistralChatCompletionProvider("mistral-
 const DefaultSynthesizeProvider = new MistralChatCompletionProvider("mistral-large-latest");
 //#endregion
 //#region src/providers/openai/defaults.ts
-const DEFAULT_OPENAI_GRADING_MODEL = "gpt-5.4-2026-03-05";
+const DEFAULT_OPENAI_GRADING_MODEL = "gpt-5.5-2026-04-23";
 const DefaultEmbeddingProvider = new OpenAiEmbeddingProvider("text-embedding-3-large");
 const DefaultGradingProvider = new OpenAiChatCompletionProvider(DEFAULT_OPENAI_GRADING_MODEL);
 const DefaultGradingJsonProvider = new OpenAiChatCompletionProvider(DEFAULT_OPENAI_GRADING_MODEL, { config: { response_format: { type: "json_object" } } });
 const DefaultSuggestionsProvider = new OpenAiChatCompletionProvider(DEFAULT_OPENAI_GRADING_MODEL);
 const DefaultModerationProvider = new OpenAiModerationProvider("omni-moderation-latest");
-const DefaultWebSearchProvider = new OpenAiResponsesProvider("gpt-5.4-2026-03-05", { config: { tools: [{ type: "web_search_preview" }] } });
+const DefaultWebSearchProvider = new OpenAiResponsesProvider("gpt-5.5-2026-04-23", { config: { tools: [{ type: "web_search_preview" }] } });
 async function getDefaultProviderPreferences(env) {
 	const hasAnthropicCredentials = Boolean(getEnvString("ANTHROPIC_API_KEY") || env?.ANTHROPIC_API_KEY);
 	const hasOpenAiCredentials = Boolean(getEnvString("OPENAI_API_KEY") || env?.OPENAI_API_KEY);
@@ -1396,7 +1396,7 @@ async function doRemoteGrading(payload) {
 			method: "POST",
 			headers: { "Content-Type": "application/json" },
 			body
-		}, REQUEST_TIMEOUT_MS);
+		}, getRequestTimeoutMs());
 		logger.debug(`Remote grading result: status=${status}, statusText=${statusText}, data=${JSON.stringify(data)}`);
 		if (status !== 200) throw new Error(`Remote grading failed with status ${status}: ${statusText} ${JSON.stringify(data)}`);
 		const { result } = data;
@@ -1439,7 +1439,7 @@ async function doRemoteScoringWithPi(payload, passThreshold = .5) {
 					"x-api-key": apiKey
 				},
 				body
-			}, REQUEST_TIMEOUT_MS);
+			}, getRequestTimeoutMs());
 			return convertPiResultToGradingResult(data, passThreshold);
 		} else throw new Error(`Env var WITHPI_API_KEY must be set. Visit https://docs.withpi.ai for more information.`);
 	} catch (error) {
@@ -8389,7 +8389,7 @@ function describeFilters(config) {
 }
 async function fetchDataset$1(limit, config) {
 	try {
-		const response = await fetchWithTimeout(DATASET_URL, {}, REQUEST_TIMEOUT_MS);
+		const response = await fetchWithTimeout(DATASET_URL, {}, getRequestTimeoutMs());
 		if (!response.ok) throw new Error(`HTTP status: ${response.status} ${response.statusText}`);
 		const records = parse$1(await response.text(), { columns: true });
 		logger.debug(`[harmbench] Parsed ${records.length} entries from CSV`);
@@ -15606,7 +15606,7 @@ var VLSUDatasetManager = class VLSUDatasetManager extends ImageDatasetManager {
 		}
 		logger.debug(`[vlsu] Fetching CSV from ${VLSU_CSV_URL}`);
 		try {
-			const response = await fetchWithCache(VLSU_CSV_URL, {}, REQUEST_TIMEOUT_MS, "text");
+			const response = await fetchWithCache(VLSU_CSV_URL, {}, getRequestTimeoutMs(), "text");
 			if (response.status < 200 || response.status >= 300) throw new Error(`Failed to fetch VLSU CSV: ${response.statusText}`);
 			const csvText = response.data;
 			const records = parse$1(csvText, {
@@ -16169,4 +16169,4 @@ function getGraderById(id) {
 //#endregion
 export { readProviderPromptMap as $, ExcessiveAgencyPlugin as A, withProviderCallExecutionContext as At, retryWithDeduplication as B, PlinyPlugin as C, processFileReference as Ct, ImitationPlugin as D, getAndCheckProvider as Dt, IntentPlugin as E, callProviderWithContext as Et, BeavertailsPlugin as F, matchesFactuality as G, fetchHuggingFaceDataset as H, AegisPlugin as I, matchesPiScore as J, matchesGEval as K, RedteamGraderBase as L, DebugAccessPlugin as M, CrossSessionLeakPlugin as N, HarmbenchPlugin as O, getGradingProvider as Ot, ContractPlugin as P, readPrompts as Q, RedteamPluginBase as R, makeInlinePolicyIdSync as S, loadFromJavaScriptFile as St, OverreliancePlugin as T, DEFAULT_ANTHROPIC_MODEL as Tt, isGraderFailure as U, sampleArray as V, matchesClosedQa as W, doRemoteGrading as X, matchesTrajectoryGoalSuccess as Y, processPrompts as Z, PromptExtractionPlugin as _, normalizeMatcherTokenUsage as _t, VLGuardPlugin as a, CONTEXT_FAITHFULNESS_NLI_STATEMENTS as at, determinePolicyTypeFromId as b, coerceString as bt, ToxicChatPlugin as c, CONTEXT_RECALL_NOT_ATTRIBUTED_TOKEN as ct, TeenSafetyDangerousRoleplayPlugin as d, loadRubricPrompt as dt, DEFAULT_WEB_SEARCH_PROMPT as et, TeenSafetyDangerousContentPlugin as f, renderLlmRubricPrompt as ft, RbacPlugin as g, fail as gt, ShellInjectionPlugin as h, euclideanDistance as ht, VLSUPlugin as i, CONTEXT_FAITHFULNESS_LONGFORM as it, DivergentRepetitionPlugin as j, HallucinationPlugin as k, getProviderCallExecutionContext as kt, ToolDiscoveryPlugin as l, CONTEXT_RELEVANCE as lt, SqlInjectionPlugin as m, dotProduct as mt, getGraderById as n, SUGGEST_PROMPTS_SYSTEM_MESSAGE as nt, UnverifiableClaimsPlugin as o, CONTEXT_RECALL as ot, TeenSafetyAgeRestrictedGoodsAndServicesPlugin as p, cosineSimilarity as pt, matchesLlmRubric as q, graders_exports as r, ANSWER_RELEVANCY_GENERATE as rt, UnsafeBenchPlugin as s, CONTEXT_RECALL_ATTRIBUTED_TOKEN as st, GRADERS as t, SELECT_BEST_PROMPT as tt, TeenSafetyHarmfulBodyIdealsPlugin as u, CONTEXT_RELEVANCE_BAD as ut, PoliticsPlugin as v, splitIntoSentences as vt, getPiiLeakTestsForCategory as w, getDefaultProviders as wt, isValidPolicyObject as x, getFinalTest as xt, PolicyPlugin as y, tryParse as yt, getCustomPolicies as z };
 
-//# sourceMappingURL=graders-BX0f2tvS.js.map
+//# sourceMappingURL=graders-GNRQ1357.js.map

package/dist/src/{graders-C0nXU_ZP.js → graders-k8qmhrPu.js}

@@ -1,36 +1,36 @@
 import { t as __exportAll } from "./chunk-DEq-mXcV.js";
-import { O as isCI, S as getEnvBool, T as getEnvString, a as logger, h as extractJsonObjects, k as state, m as extractFirstJsonObject, y as safeJsonStringify } from "./logger-KD8JjCRJ.js";
-import { m as sleep, r as fetchWithTimeout, t as fetchWithProxy, v as REQUEST_TIMEOUT_MS } from "./fetch-BufrQtvR.js";
+import { C as getEnvBool, E as getEnvString, S as state, a as logger, h as extractJsonObjects, k as isCI, m as extractFirstJsonObject, y as safeJsonStringify } from "./logger-BjYLHvGm.js";
+import { m as sleep, r as fetchWithTimeout, t as fetchWithProxy, y as getRequestTimeoutMs } from "./fetch-CBe8pU8Q.js";
 import { t as invariant } from "./invariant-DIYf9sP1.js";
-import { o as getUserEmail } from "./accounts-DanM1wq_.js";
-import { r as importModule } from "./esm-B6whoAcf.js";
-import { i as extractVariablesFromTemplate, o as getNunjucksEngine } from "./render-_6ur1fhE.js";
-import { d as hasCodexDefaultCredentials, l as shouldGenerateRemote, r as getRemoteGenerationUrl, u as getCodexDefaultProviders } from "./remoteGeneration-COpWcmWd.js";
-import { C as isValidReusablePolicyId, ft as CODING_AGENT_PLUGINS, ht as PromptSchema, k as MULTI_TURN_STRATEGIES, mt as CODING_AGENT_PLUGIN_DISPLAY_NAMES, p as isApiProvider, pt as CODING_AGENT_PLUGIN_DESCRIPTIONS, vt as buildInputPromptDescription, x as PolicyObjectSchema } from "./types-BJQBBPTP.js";
+import { o as getUserEmail } from "./accounts-BAajbKSh.js";
+import { r as importModule } from "./esm-CGER3lko.js";
+import { i as extractVariablesFromTemplate, o as getNunjucksEngine } from "./render-BoxrvB84.js";
+import { d as hasCodexDefaultCredentials, l as shouldGenerateRemote, r as getRemoteGenerationUrl, u as getCodexDefaultProviders } from "./remoteGeneration-CoxtxLvB.js";
+import { C as isValidReusablePolicyId, ft as CODING_AGENT_PLUGINS, ht as PromptSchema, k as MULTI_TURN_STRATEGIES, mt as CODING_AGENT_PLUGIN_DISPLAY_NAMES, p as isApiProvider, pt as CODING_AGENT_PLUGIN_DESCRIPTIONS, vt as buildInputPromptDescription, x as PolicyObjectSchema } from "./types-uLCKnRiK.js";
 import { o as isJavascriptFile } from "./shared-BoG7qLMv.js";
 import { n as sha256 } from "./createHash-CGVzWdjj.js";
-import { i as PROMPT_DELIMITER, n as maybeFilePath, r as normalizeInput } from "./utils-B0lzitHZ.js";
-import { i as getCache, o as isCacheEnabled, r as fetchWithCache } from "./cache-roFAE0cI.js";
+import { i as PROMPT_DELIMITER, n as maybeFilePath, r as normalizeInput } from "./utils-DMWnIu_V.js";
+import { i as getCache, o as isCacheEnabled, r as fetchWithCache } from "./cache-x9llZxbx.js";
 import { r as accumulateTokenUsage } from "./tokenUsageUtils-BjVkdk18.js";
-import { d as getPoliciesFromCloud } from "./storage-QdU-SmvD.js";
-import { r as runPython } from "./pythonUtils-CLCgQ9tt.js";
-import { A as maybeLoadConfigFromExternalFile, F as parsePathOrGlob, O as getNunjucksEngineForFilePath, P as maybeLoadToolsFromExternalFile, j as maybeLoadFromExternalFile, z as parseFileUrl } from "./util-I-Rf-KaD.js";
-import { t as OpenAiChatCompletionProvider } from "./chat-CUCorGiL.js";
-import { x as hasGoogleDefaultCredentials } from "./transform-BIMynQsA.js";
-import { t as AnthropicMessagesProvider } from "./messages-BabO-cX8.js";
-import { t as OpenAiResponsesProvider } from "./responses-B6ktc3Ra.js";
-import { C as DefaultLlmRubricProvider, D as AzureEmbeddingProvider, E as AzureModerationProvider, O as AzureChatCompletionProvider, S as DefaultGradingProvider$2, T as DefaultSynthesizeProvider$1, _ as DefaultEmbeddingProvider$2, c as parseScriptParts, g as MistralEmbeddingProvider, h as MistralChatCompletionProvider, m as OpenAiModerationProvider, n as loadApiProvider, s as getFileHashes, v as DefaultGradingProvider$3, w as DefaultSuggestionsProvider$2, x as DefaultGradingJsonProvider$2 } from "./providers-BCCz6_IX.js";
-import { t as OpenAiEmbeddingProvider } from "./embedding-C251p1-8.js";
-import { r as materializeInputVariablesWithMetadata } from "./inputVariables-B0qUChbV.js";
-import { a as extractPromptFromTags, c as isBasicRefusal, i as extractMaterializedVariablesFromJsonWithMetadata, l as isEmptyResponse, n as extractGoalFromPrompt, r as extractInputVarsFromPrompt, s as getShortPluginId, t as extractAllPromptsFromTags, u as removePrefix } from "./util-Df8YMvS1.js";
-import { _ as isRateLimitWrapped, f as redteamProviderManager, g as createProviderRateLimitOptions } from "./shared-7pmVZLNO.js";
-import { n as getGeneratedPromptOverLimit, r as getMaxCharsPerMessageModifierValue, t as MAX_CHARS_PER_MESSAGE_MODIFIER_KEY } from "./promptLength-0qIHyhA5.js";
+import { d as getPoliciesFromCloud } from "./storage-BC9TzmHh.js";
+import { r as runPython } from "./pythonUtils-CBe7VV6f.js";
+import { A as maybeLoadConfigFromExternalFile, F as parsePathOrGlob, O as getNunjucksEngineForFilePath, P as maybeLoadToolsFromExternalFile, j as maybeLoadFromExternalFile, z as parseFileUrl } from "./util-BGfk7zh7.js";
+import { t as OpenAiChatCompletionProvider } from "./chat-CvQMD-mp.js";
+import { x as hasGoogleDefaultCredentials } from "./transform-D7QlmswG.js";
+import { t as AnthropicMessagesProvider } from "./messages-C5qBfONg.js";
+import { t as OpenAiResponsesProvider } from "./responses-pvlY3oYx.js";
+import { C as DefaultLlmRubricProvider, D as AzureEmbeddingProvider, E as AzureModerationProvider, O as AzureChatCompletionProvider, S as DefaultGradingProvider$2, T as DefaultSynthesizeProvider$1, _ as DefaultEmbeddingProvider$2, c as parseScriptParts, g as MistralEmbeddingProvider, h as MistralChatCompletionProvider, m as OpenAiModerationProvider, n as loadApiProvider, s as getFileHashes, v as DefaultGradingProvider$3, w as DefaultSuggestionsProvider$2, x as DefaultGradingJsonProvider$2 } from "./providers-DDLQc0tl.js";
+import { t as OpenAiEmbeddingProvider } from "./embedding-BwTB3YDt.js";
+import { r as materializeInputVariablesWithMetadata } from "./inputVariables-DpOsOnzU.js";
+import { a as extractPromptFromTags, c as isBasicRefusal, i as extractMaterializedVariablesFromJsonWithMetadata, l as isEmptyResponse, n as extractGoalFromPrompt, r as extractInputVarsFromPrompt, s as getShortPluginId, t as extractAllPromptsFromTags, u as removePrefix } from "./util-CKZReNsz.js";
+import { _ as isRateLimitWrapped, f as redteamProviderManager, g as createProviderRateLimitOptions } from "./shared-yBQATkD5.js";
+import { n as getGeneratedPromptOverLimit, r as getMaxCharsPerMessageModifierValue, t as MAX_CHARS_PER_MESSAGE_MODIFIER_KEY } from "./promptLength-5B4_P93z.js";
 import { t as REDTEAM_MEMORY_POISONING_PLUGIN_ID } from "./constants-BjJV0cRr.js";
-import { n as checkExfilTracking } from "./indirectWebPwn-tNx9OZ35.js";
-import { AsyncLocalStorage } from "node:async_hooks";
+import { n as checkExfilTracking } from "./indirectWebPwn-DRYpRTx6.js";
 import * as fs$2 from "fs";
 import fs from "fs";
 import path from "path";
+import { AsyncLocalStorage } from "node:async_hooks";
 import yaml from "js-yaml";
 import z$1 from "zod";
 import { readFile, stat } from "fs/promises";
@@ -160,13 +160,13 @@ const DefaultSuggestionsProvider$1 = new MistralChatCompletionProvider("mistral-
 const DefaultSynthesizeProvider = new MistralChatCompletionProvider("mistral-large-latest");
 //#endregion
 //#region src/providers/openai/defaults.ts
-const DEFAULT_OPENAI_GRADING_MODEL = "gpt-5.4-2026-03-05";
+const DEFAULT_OPENAI_GRADING_MODEL = "gpt-5.5-2026-04-23";
 const DefaultEmbeddingProvider = new OpenAiEmbeddingProvider("text-embedding-3-large");
 const DefaultGradingProvider = new OpenAiChatCompletionProvider(DEFAULT_OPENAI_GRADING_MODEL);
 const DefaultGradingJsonProvider = new OpenAiChatCompletionProvider(DEFAULT_OPENAI_GRADING_MODEL, { config: { response_format: { type: "json_object" } } });
 const DefaultSuggestionsProvider = new OpenAiChatCompletionProvider(DEFAULT_OPENAI_GRADING_MODEL);
 const DefaultModerationProvider = new OpenAiModerationProvider("omni-moderation-latest");
-const DefaultWebSearchProvider = new OpenAiResponsesProvider("gpt-5.4-2026-03-05", { config: { tools: [{ type: "web_search_preview" }] } });
+const DefaultWebSearchProvider = new OpenAiResponsesProvider("gpt-5.5-2026-04-23", { config: { tools: [{ type: "web_search_preview" }] } });
 async function getDefaultProviderPreferences(env) {
 	const hasAnthropicCredentials = Boolean(getEnvString("ANTHROPIC_API_KEY") || env?.ANTHROPIC_API_KEY);
 	const hasOpenAiCredentials = Boolean(getEnvString("OPENAI_API_KEY") || env?.OPENAI_API_KEY);
@@ -1456,7 +1456,7 @@ async function doRemoteGrading(payload) {
 			method: "POST",
 			headers: { "Content-Type": "application/json" },
 			body
-		}, REQUEST_TIMEOUT_MS);
+		}, getRequestTimeoutMs());
 		logger.debug(`Remote grading result: status=${status}, statusText=${statusText}, data=${JSON.stringify(data)}`);
 		if (status !== 200) throw new Error(`Remote grading failed with status ${status}: ${statusText} ${JSON.stringify(data)}`);
 		const { result } = data;
@@ -1499,7 +1499,7 @@ async function doRemoteScoringWithPi(payload, passThreshold = .5) {
 					"x-api-key": apiKey
 				},
 				body
-			}, REQUEST_TIMEOUT_MS);
+			}, getRequestTimeoutMs());
 			return convertPiResultToGradingResult(data, passThreshold);
 		} else throw new Error(`Env var WITHPI_API_KEY must be set. Visit https://docs.withpi.ai for more information.`);
 	} catch (error) {
@@ -8389,7 +8389,7 @@ function describeFilters(config) {
 }
 async function fetchDataset$1(limit, config) {
 	try {
-		const response = await fetchWithTimeout(DATASET_URL, {}, REQUEST_TIMEOUT_MS);
+		const response = await fetchWithTimeout(DATASET_URL, {}, getRequestTimeoutMs());
 		if (!response.ok) throw new Error(`HTTP status: ${response.status} ${response.statusText}`);
 		const records = parse$1(await response.text(), { columns: true });
 		logger.debug(`[harmbench] Parsed ${records.length} entries from CSV`);
@@ -15606,7 +15606,7 @@ var VLSUDatasetManager = class VLSUDatasetManager extends ImageDatasetManager {
 		}
 		logger.debug(`[vlsu] Fetching CSV from ${VLSU_CSV_URL}`);
 		try {
-			const response = await fetchWithCache(VLSU_CSV_URL, {}, REQUEST_TIMEOUT_MS, "text");
+			const response = await fetchWithCache(VLSU_CSV_URL, {}, getRequestTimeoutMs(), "text");
 			if (response.status < 200 || response.status >= 300) throw new Error(`Failed to fetch VLSU CSV: ${response.statusText}`);
 			const csvText = response.data;
 			const records = parse$1(csvText, {
@@ -16169,4 +16169,4 @@ function getGraderById(id) {
 //#endregion
 export { SUGGEST_PROMPTS_SYSTEM_MESSAGE as $, ExcessiveAgencyPlugin as A, DEFAULT_ANTHROPIC_MODEL as At, isGraderFailure as B, PlinyPlugin as C, getGradingProvider as Ct, ImitationPlugin as D, retryWithDeduplication as Dt, IntentPlugin as E, getCustomPolicies as Et, BeavertailsPlugin as F, matchesPiScore as G, matchesFactuality as H, AegisPlugin as I, processPrompts as J, matchesTrajectoryGoalSuccess as K, RedteamGraderBase as L, DebugAccessPlugin as M, CrossSessionLeakPlugin as N, HarmbenchPlugin as O, sampleArray as Ot, ContractPlugin as P, SELECT_BEST_PROMPT as Q, RedteamPluginBase as R, makeInlinePolicyIdSync as S, getAndCheckProvider as St, OverreliancePlugin as T, withProviderCallExecutionContext as Tt, matchesGEval as U, matchesClosedQa as V, matchesLlmRubric as W, readProviderPromptMap as X, readPrompts as Y, DEFAULT_WEB_SEARCH_PROMPT as Z, PromptExtractionPlugin as _, coerceString as _t, VLGuardPlugin as a, CONTEXT_RECALL_NOT_ATTRIBUTED_TOKEN as at, determinePolicyTypeFromId as b, processFileReference as bt, ToxicChatPlugin as c, loadRubricPrompt as ct, TeenSafetyDangerousRoleplayPlugin as d, dotProduct as dt, ANSWER_RELEVANCY_GENERATE as et, TeenSafetyDangerousContentPlugin as f, euclideanDistance as ft, RbacPlugin as g, tryParse as gt, ShellInjectionPlugin as h, splitIntoSentences as ht, VLSUPlugin as i, CONTEXT_RECALL_ATTRIBUTED_TOKEN as it, DivergentRepetitionPlugin as j, HallucinationPlugin as k, getDefaultProviders as kt, ToolDiscoveryPlugin as l, renderLlmRubricPrompt as lt, SqlInjectionPlugin as m, normalizeMatcherTokenUsage as mt, getGraderById as n, CONTEXT_FAITHFULNESS_NLI_STATEMENTS as nt, UnverifiableClaimsPlugin as o, CONTEXT_RELEVANCE as ot, TeenSafetyAgeRestrictedGoodsAndServicesPlugin as p, fail as pt, doRemoteGrading as q, graders_exports as r, CONTEXT_RECALL as rt, UnsafeBenchPlugin as s, CONTEXT_RELEVANCE_BAD as st, GRADERS as t, CONTEXT_FAITHFULNESS_LONGFORM as tt, TeenSafetyHarmfulBodyIdealsPlugin as u, cosineSimilarity as ut, PoliticsPlugin as v, getFinalTest as vt, getPiiLeakTestsForCategory as w, getProviderCallExecutionContext as wt, isValidPolicyObject as x, callProviderWithContext as xt, PolicyPlugin as y, loadFromJavaScriptFile as yt, fetchHuggingFaceDataset as z };
 
-//# sourceMappingURL=graders-C0nXU_ZP.js.map
+//# sourceMappingURL=graders-k8qmhrPu.js.map

package/dist/src/{hydra-BSNZZm2M.js → hydra-BrgGALmB.js}

@@ -1,20 +1,20 @@
 #!/usr/bin/env node
-import { s as logger, y as isValidJson } from "./logger-BbY6ypFL.js";
-import { f as sleep } from "./fetch-DXUnXkVU.js";
+import { s as logger, y as isValidJson } from "./logger-h-eiS4iv.js";
+import { f as sleep } from "./fetch-C59QM4cf.js";
 import { t as invariant } from "./invariant-B2Rf6avk.js";
-import { c as isLoggedIntoCloud } from "./accounts-CjFnOPmb.js";
-import { l as shouldGenerateRemote, n as getRemoteGenerationExplicitlyDisabledError, s as neverGenerateRemote, t as getRemoteGenerationDisabledError } from "./remoteGeneration-B1_XsKXU.js";
+import { c as isLoggedIntoCloud } from "./accounts-Bw2qPqE-.js";
+import { l as shouldGenerateRemote, n as getRemoteGenerationExplicitlyDisabledError, s as neverGenerateRemote, t as getRemoteGenerationDisabledError } from "./remoteGeneration-DW6mPZnO.js";
 import { a as createEmptyTokenUsage, n as accumulateResponseTokenUsage } from "./tokenUsageUtils-CDet74yk.js";
-import { r as materializeInputVariablesWithMetadata } from "./inputVariables-DXFdi7AI.js";
-import { a as extractPromptFromTags, c as isBasicRefusal, o as getSessionId, r as extractInputVarsFromPrompt } from "./util-kDURhgJW.js";
-import { t as PromptfooChatCompletionProvider } from "./promptfoo-Rjp-MeBb.js";
-import { c as getTargetResponse, i as externalizeResponseForRedteamHistory, l as isConversationEndedResponse, o as getGraderAssertionValue, t as buildGraderResultAssertion, u as isValidChatMessageArray } from "./shared-9WHQ1oNE.js";
-import { n as isBlobStorageEnabled, r as shouldAttemptRemoteBlobUpload } from "./extractor-CIW3iN-b.js";
-import { n as checkExfilTracking } from "./indirectWebPwn-uyWdHx04.js";
-import { n as renderPrompt } from "./evaluatorHelpers-BsYP_muT.js";
-import { t as Strategies } from "./strategies-DRJjGTIY.js";
-import { a as fetchTraceContext, i as extractTraceIdFromTraceparent, n as formatTraceForMetadata, r as formatTraceSummary, t as resolveTracingOptions } from "./tracingOptions-BnwKCkSB.js";
-import { t as applyRuntimeTransforms } from "./runtimeTransform-Dgh_D7DU.js";
+import { r as materializeInputVariablesWithMetadata } from "./inputVariables-qJmuSN5i.js";
+import { a as extractPromptFromTags, c as isBasicRefusal, o as getSessionId, r as extractInputVarsFromPrompt } from "./util-D9grslJP.js";
+import { a as assertRemoteMaterializationHandled, s as buildRemoteMaterializedInputVariables, t as PromptfooChatCompletionProvider } from "./promptfoo-wxe28-q9.js";
+import { c as getTargetResponse, i as externalizeResponseForRedteamHistory, l as isConversationEndedResponse, o as getGraderAssertionValue, t as buildGraderResultAssertion, u as isValidChatMessageArray } from "./shared-DsbObxvy.js";
+import { n as isBlobStorageEnabled, r as shouldAttemptRemoteBlobUpload } from "./extractor-CtadTcJm.js";
+import { n as checkExfilTracking } from "./indirectWebPwn-BXUoJl4I.js";
+import { n as renderPrompt } from "./evaluatorHelpers-DDG44JeI.js";
+import { t as Strategies } from "./strategies-Ptz_P4P4.js";
+import { a as fetchTraceContext, i as extractTraceIdFromTraceparent, n as formatTraceForMetadata, r as formatTraceSummary, t as resolveTracingOptions } from "./tracingOptions-D2F0D7es.js";
+import { t as applyRuntimeTransforms } from "./runtimeTransform-s14J7yxm.js";
 //#region src/redteam/providers/hydra/index.ts
 const DEFAULT_MAX_TURNS = 10;
 const DEFAULT_MAX_BACKTRACKS = 10;
@@ -112,11 +112,12 @@ var HydraProvider = class {
 		let storedGraderResult = void 0;
 		let lastTargetResponse = void 0;
 		let backtrackCount = 0;
+		let agentFailureError;
 		const redteamHistory = [];
 		let lastTransformResult;
 		let lastTransformDisplayVars;
 		let lastFinalAttackPrompt;
-		const { getGraderById } = await import("./graders-BDT7dif6.js");
+		const { getGraderById } = await import("./graders-D4vc1PFi.js");
 		let assertToUse = test?.assert?.find((a) => a.type && a.type.includes(test.metadata?.pluginId));
 		if (!assertToUse) assertToUse = test?.assert?.find((a) => a.type);
 		let previousTraceSummary;
@@ -160,6 +161,7 @@ var HydraProvider = class {
 					testRunId,
 					error: agentResp.error
 				});
+				agentFailureError = agentResp.error;
 				continue;
 			}
 			let nextMessage;
@@ -170,18 +172,31 @@ var HydraProvider = class {
 			}
 			if (!nextMessage) {
 				logger.info("[Hydra] Missing message from agent", { turn });
+				agentFailureError = "Hydra agent did not return an attack message";
 				continue;
 			}
 			let processedMessage = nextMessage;
 			const extractedPrompt = extractPromptFromTags(nextMessage);
 			if (extractedPrompt) processedMessage = extractedPrompt;
+			if (this.config.inputs && shouldGenerateRemote()) assertRemoteMaterializationHandled(agentResp, "Hydra multi-input generation");
 			const currentInputVars = extractInputVarsFromPrompt(processedMessage, this.config.inputs);
-			const currentRenderInputVars = (currentInputVars && this.config.inputs ? await materializeInputVariablesWithMetadata(currentInputVars, this.config.inputs, {
+			let materializedInputVars;
+			if (this.config.inputs && shouldGenerateRemote() && !currentInputVars && !agentResp.materializedVars) {
+				logger.warn("[Hydra] Remote multi-input generation returned an invalid prompt format", {
+					turn,
+					messagePreview: processedMessage.slice(0, 200)
+				});
+				agentFailureError = "Hydra remote multi-input generation returned an invalid prompt format";
+				continue;
+			}
+			if ((currentInputVars || agentResp.materializedVars) && this.config.inputs) if (shouldGenerateRemote()) materializedInputVars = buildRemoteMaterializedInputVariables(agentResp, currentInputVars ?? {}, this.config.inputs);
+			else materializedInputVars = await materializeInputVariablesWithMetadata(currentInputVars, this.config.inputs, {
 				materializationIndex: turn,
 				pluginId: "hydra",
 				provider: this.agentProvider,
 				purpose: test?.metadata?.purpose
-			}) : void 0)?.vars ?? currentInputVars;
+			});
+			const currentRenderInputVars = materializedInputVars?.vars ?? currentInputVars;
 			this.conversationHistory.push({
 				role: "user",
 				content: processedMessage
@@ -513,13 +528,16 @@ var HydraProvider = class {
 			role: msg.role,
 			content: msg.content
 		}));
+		const targetProbeCount = totalTokenUsage.numRequests ?? 0;
+		const hydraRoundsCompleted = this.conversationHistory.filter((m) => m.role === "user").length;
+		const failClosedError = targetProbeCount === 0 ? agentFailureError || "Hydra did not execute any target probes" : void 0;
 		return {
 			output: lastTargetResponse?.output || "",
-			...lastTargetResponse?.error ? { error: lastTargetResponse.error } : {},
+			...failClosedError ? { error: failClosedError } : lastTargetResponse?.error ? { error: lastTargetResponse.error } : {},
 			metadata: {
 				sessionId: this.sessionId || getSessionId(lastTargetResponse, context),
 				messages,
-				hydraRoundsCompleted: this.conversationHistory.filter((m) => m.role === "user").length,
+				hydraRoundsCompleted,
 				hydraBacktrackCount: backtrackCount,
 				hydraResult: vulnerabilityAchieved,
 				stopReason,
@@ -540,4 +558,4 @@ var HydraProvider = class {
 //#endregion
 export { HydraProvider, HydraProvider as default };
 
-//# sourceMappingURL=hydra-BSNZZm2M.js.map
+//# sourceMappingURL=hydra-BrgGALmB.js.map

package/dist/src/{hydra-BxdG4nkg.js → hydra-C38LOovb.js}

@@ -1,18 +1,18 @@
-import { _ as isValidJson, a as logger } from "./logger-KD8JjCRJ.js";
-import { m as sleep } from "./fetch-BufrQtvR.js";
+import { _ as isValidJson, a as logger } from "./logger-BjYLHvGm.js";
+import { m as sleep } from "./fetch-CBe8pU8Q.js";
 import { t as invariant } from "./invariant-DIYf9sP1.js";
-import { c as isLoggedIntoCloud } from "./accounts-DanM1wq_.js";
-import { l as shouldGenerateRemote, n as getRemoteGenerationExplicitlyDisabledError, s as neverGenerateRemote, t as getRemoteGenerationDisabledError } from "./remoteGeneration-COpWcmWd.js";
+import { c as isLoggedIntoCloud } from "./accounts-BAajbKSh.js";
+import { l as shouldGenerateRemote, n as getRemoteGenerationExplicitlyDisabledError, s as neverGenerateRemote, t as getRemoteGenerationDisabledError } from "./remoteGeneration-CoxtxLvB.js";
 import { a as createEmptyTokenUsage, n as accumulateResponseTokenUsage } from "./tokenUsageUtils-BjVkdk18.js";
-import { n as isBlobStorageEnabled, r as shouldAttemptRemoteBlobUpload } from "./extractor-BR7XAzAL.js";
-import { r as materializeInputVariablesWithMetadata } from "./inputVariables-B0qUChbV.js";
-import { a as extractPromptFromTags, c as isBasicRefusal, o as getSessionId, r as extractInputVarsFromPrompt } from "./util-Df8YMvS1.js";
-import { t as PromptfooChatCompletionProvider } from "./promptfoo-BDrfT30-.js";
-import { c as getTargetResponse, i as externalizeResponseForRedteamHistory, l as isConversationEndedResponse, o as getGraderAssertionValue, t as buildGraderResultAssertion, u as isValidChatMessageArray } from "./shared-7pmVZLNO.js";
-import { n as checkExfilTracking } from "./indirectWebPwn-tNx9OZ35.js";
+import { n as isBlobStorageEnabled, r as shouldAttemptRemoteBlobUpload } from "./extractor-DAnauX2V.js";
+import { r as materializeInputVariablesWithMetadata } from "./inputVariables-DpOsOnzU.js";
+import { a as extractPromptFromTags, c as isBasicRefusal, o as getSessionId, r as extractInputVarsFromPrompt } from "./util-CKZReNsz.js";
+import { a as assertRemoteMaterializationHandled, s as buildRemoteMaterializedInputVariables, t as PromptfooChatCompletionProvider } from "./promptfoo-DVKqidd8.js";
+import { c as getTargetResponse, i as externalizeResponseForRedteamHistory, l as isConversationEndedResponse, o as getGraderAssertionValue, t as buildGraderResultAssertion, u as isValidChatMessageArray } from "./shared-yBQATkD5.js";
+import { n as checkExfilTracking } from "./indirectWebPwn-DRYpRTx6.js";
 import { n as renderPrompt, t as Strategies } from "./server/index.js";
-import { a as fetchTraceContext, i as extractTraceIdFromTraceparent, n as formatTraceForMetadata, r as formatTraceSummary, t as resolveTracingOptions } from "./tracingOptions-Chi74lOD.js";
-import { t as applyRuntimeTransforms } from "./runtimeTransform-BJOpL9Yc.js";
+import { a as fetchTraceContext, i as extractTraceIdFromTraceparent, n as formatTraceForMetadata, r as formatTraceSummary, t as resolveTracingOptions } from "./tracingOptions-PLePpCmo.js";
+import { t as applyRuntimeTransforms } from "./runtimeTransform-BmhEztlx.js";
 //#region src/redteam/providers/hydra/index.ts
 const DEFAULT_MAX_TURNS = 10;
 const DEFAULT_MAX_BACKTRACKS = 10;
@@ -110,11 +110,12 @@ var HydraProvider = class {
 		let storedGraderResult = void 0;
 		let lastTargetResponse = void 0;
 		let backtrackCount = 0;
+		let agentFailureError;
 		const redteamHistory = [];
 		let lastTransformResult;
 		let lastTransformDisplayVars;
 		let lastFinalAttackPrompt;
-		const { getGraderById } = await import("./graders-C0nXU_ZP.js").then((n) => n.r);
+		const { getGraderById } = await import("./graders-k8qmhrPu.js").then((n) => n.r);
 		let assertToUse = test?.assert?.find((a) => a.type && a.type.includes(test.metadata?.pluginId));
 		if (!assertToUse) assertToUse = test?.assert?.find((a) => a.type);
 		let previousTraceSummary;
@@ -158,6 +159,7 @@ var HydraProvider = class {
 					testRunId,
 					error: agentResp.error
 				});
+				agentFailureError = agentResp.error;
 				continue;
 			}
 			let nextMessage;
@@ -168,18 +170,31 @@ var HydraProvider = class {
 			}
 			if (!nextMessage) {
 				logger.info("[Hydra] Missing message from agent", { turn });
+				agentFailureError = "Hydra agent did not return an attack message";
 				continue;
 			}
 			let processedMessage = nextMessage;
 			const extractedPrompt = extractPromptFromTags(nextMessage);
 			if (extractedPrompt) processedMessage = extractedPrompt;
+			if (this.config.inputs && shouldGenerateRemote()) assertRemoteMaterializationHandled(agentResp, "Hydra multi-input generation");
 			const currentInputVars = extractInputVarsFromPrompt(processedMessage, this.config.inputs);
-			const currentRenderInputVars = (currentInputVars && this.config.inputs ? await materializeInputVariablesWithMetadata(currentInputVars, this.config.inputs, {
+			let materializedInputVars;
+			if (this.config.inputs && shouldGenerateRemote() && !currentInputVars && !agentResp.materializedVars) {
+				logger.warn("[Hydra] Remote multi-input generation returned an invalid prompt format", {
+					turn,
+					messagePreview: processedMessage.slice(0, 200)
+				});
+				agentFailureError = "Hydra remote multi-input generation returned an invalid prompt format";
+				continue;
+			}
+			if ((currentInputVars || agentResp.materializedVars) && this.config.inputs) if (shouldGenerateRemote()) materializedInputVars = buildRemoteMaterializedInputVariables(agentResp, currentInputVars ?? {}, this.config.inputs);
+			else materializedInputVars = await materializeInputVariablesWithMetadata(currentInputVars, this.config.inputs, {
 				materializationIndex: turn,
 				pluginId: "hydra",
 				provider: this.agentProvider,
 				purpose: test?.metadata?.purpose
-			}) : void 0)?.vars ?? currentInputVars;
+			});
+			const currentRenderInputVars = materializedInputVars?.vars ?? currentInputVars;
 			this.conversationHistory.push({
 				role: "user",
 				content: processedMessage
@@ -511,13 +526,16 @@ var HydraProvider = class {
 			role: msg.role,
 			content: msg.content
 		}));
+		const targetProbeCount = totalTokenUsage.numRequests ?? 0;
+		const hydraRoundsCompleted = this.conversationHistory.filter((m) => m.role === "user").length;
+		const failClosedError = targetProbeCount === 0 ? agentFailureError || "Hydra did not execute any target probes" : void 0;
 		return {
 			output: lastTargetResponse?.output || "",
-			...lastTargetResponse?.error ? { error: lastTargetResponse.error } : {},
+			...failClosedError ? { error: failClosedError } : lastTargetResponse?.error ? { error: lastTargetResponse.error } : {},
 			metadata: {
 				sessionId: this.sessionId || getSessionId(lastTargetResponse, context),
 				messages,
-				hydraRoundsCompleted: this.conversationHistory.filter((m) => m.role === "user").length,
+				hydraRoundsCompleted,
 				hydraBacktrackCount: backtrackCount,
 				hydraResult: vulnerabilityAchieved,
 				stopReason,
@@ -538,4 +556,4 @@ var HydraProvider = class {
 //#endregion
 export { HydraProvider, HydraProvider as default };
 
-//# sourceMappingURL=hydra-BxdG4nkg.js.map
+//# sourceMappingURL=hydra-C38LOovb.js.map

package/dist/src/{hydra-DE4xWwyc.js → hydra-CXc0p63W.js}

@@ -1,19 +1,19 @@
-import { _ as isValidJson, a as logger } from "./logger-Ct2S6Yx-.js";
+import { _ as isValidJson, a as logger } from "./logger-Cs0I2v_y.js";
 import { t as invariant } from "./invariant-Ddh24eXh.js";
-import { l as sleep } from "./fetch-It34O8Ur.js";
-import { o as isLoggedIntoCloud } from "./accounts-Ca7WIoPY.js";
-import { l as shouldGenerateRemote, n as getRemoteGenerationExplicitlyDisabledError, s as neverGenerateRemote, t as getRemoteGenerationDisabledError } from "./remoteGeneration-DsaSwmG2.js";
+import { l as sleep } from "./fetch-CHJu2KON.js";
+import { o as isLoggedIntoCloud } from "./accounts-4Xa5fcHy.js";
+import { l as shouldGenerateRemote, n as getRemoteGenerationExplicitlyDisabledError, s as neverGenerateRemote, t as getRemoteGenerationDisabledError } from "./remoteGeneration-DvWEm3yi.js";
 import { a as createEmptyTokenUsage, n as accumulateResponseTokenUsage } from "./tokenUsageUtils-CmnQ0G2m.js";
-import { n as materializeInputVariablesWithMetadata } from "./inputVariables-DUGMb9Ka.js";
-import { a as extractPromptFromTags, c as isBasicRefusal, o as getSessionId, r as extractInputVarsFromPrompt } from "./util-DNl96nNs.js";
-import { t as PromptfooChatCompletionProvider } from "./promptfoo-Cm4hiy1Y.js";
-import { c as getTargetResponse, i as externalizeResponseForRedteamHistory, l as isConversationEndedResponse, o as getGraderAssertionValue, t as buildGraderResultAssertion, u as isValidChatMessageArray } from "./shared-D6IjElRI.js";
-import { n as isBlobStorageEnabled, r as shouldAttemptRemoteBlobUpload } from "./extractor-CxRtnaHl.js";
-import { n as checkExfilTracking } from "./indirectWebPwn-CbjUG0rh.js";
-import { n as renderPrompt } from "./evaluatorHelpers-DuqFFfq7.js";
-import { t as Strategies } from "./strategies-CiSeroPH.js";
-import { a as fetchTraceContext, i as extractTraceIdFromTraceparent, n as formatTraceForMetadata, r as formatTraceSummary, t as resolveTracingOptions } from "./tracingOptions-ji2OuXbT.js";
-import { t as applyRuntimeTransforms } from "./runtimeTransform-DigbjU1r.js";
+import { n as materializeInputVariablesWithMetadata } from "./inputVariables-B7cylqzM.js";
+import { a as extractPromptFromTags, c as isBasicRefusal, o as getSessionId, r as extractInputVarsFromPrompt } from "./util-D1NKMz55.js";
+import { a as assertRemoteMaterializationHandled, s as buildRemoteMaterializedInputVariables, t as PromptfooChatCompletionProvider } from "./promptfoo-CEsoPlBa.js";
+import { c as getTargetResponse, i as externalizeResponseForRedteamHistory, l as isConversationEndedResponse, o as getGraderAssertionValue, t as buildGraderResultAssertion, u as isValidChatMessageArray } from "./shared-BF25a7ms.js";
+import { n as isBlobStorageEnabled, r as shouldAttemptRemoteBlobUpload } from "./extractor-Bjz9sdTQ.js";
+import { n as checkExfilTracking } from "./indirectWebPwn-1gHqaL91.js";
+import { n as renderPrompt } from "./evaluatorHelpers-C8C8WQTv.js";
+import { t as Strategies } from "./strategies-4Nnr_9xt.js";
+import { a as fetchTraceContext, i as extractTraceIdFromTraceparent, n as formatTraceForMetadata, r as formatTraceSummary, t as resolveTracingOptions } from "./tracingOptions-WpE9IqC8.js";
+import { t as applyRuntimeTransforms } from "./runtimeTransform-Bw9Tttpp.js";
 //#region src/redteam/providers/hydra/index.ts
 const DEFAULT_MAX_TURNS = 10;
 const DEFAULT_MAX_BACKTRACKS = 10;
@@ -111,11 +111,12 @@ var HydraProvider = class {
 		let storedGraderResult = void 0;
 		let lastTargetResponse = void 0;
 		let backtrackCount = 0;
+		let agentFailureError;
 		const redteamHistory = [];
 		let lastTransformResult;
 		let lastTransformDisplayVars;
 		let lastFinalAttackPrompt;
-		const { getGraderById } = await import("./graders-BX0f2tvS.js").then((n) => n.r);
+		const { getGraderById } = await import("./graders-GNRQ1357.js").then((n) => n.r);
 		let assertToUse = test?.assert?.find((a) => a.type && a.type.includes(test.metadata?.pluginId));
 		if (!assertToUse) assertToUse = test?.assert?.find((a) => a.type);
 		let previousTraceSummary;
@@ -159,6 +160,7 @@ var HydraProvider = class {
 					testRunId,
 					error: agentResp.error
 				});
+				agentFailureError = agentResp.error;
 				continue;
 			}
 			let nextMessage;
@@ -169,18 +171,31 @@ var HydraProvider = class {
 			}
 			if (!nextMessage) {
 				logger.info("[Hydra] Missing message from agent", { turn });
+				agentFailureError = "Hydra agent did not return an attack message";
 				continue;
 			}
 			let processedMessage = nextMessage;
 			const extractedPrompt = extractPromptFromTags(nextMessage);
 			if (extractedPrompt) processedMessage = extractedPrompt;
+			if (this.config.inputs && shouldGenerateRemote()) assertRemoteMaterializationHandled(agentResp, "Hydra multi-input generation");
 			const currentInputVars = extractInputVarsFromPrompt(processedMessage, this.config.inputs);
-			const currentRenderInputVars = (currentInputVars && this.config.inputs ? await materializeInputVariablesWithMetadata(currentInputVars, this.config.inputs, {
+			let materializedInputVars;
+			if (this.config.inputs && shouldGenerateRemote() && !currentInputVars && !agentResp.materializedVars) {
+				logger.warn("[Hydra] Remote multi-input generation returned an invalid prompt format", {
+					turn,
+					messagePreview: processedMessage.slice(0, 200)
+				});
+				agentFailureError = "Hydra remote multi-input generation returned an invalid prompt format";
+				continue;
+			}
+			if ((currentInputVars || agentResp.materializedVars) && this.config.inputs) if (shouldGenerateRemote()) materializedInputVars = buildRemoteMaterializedInputVariables(agentResp, currentInputVars ?? {}, this.config.inputs);
+			else materializedInputVars = await materializeInputVariablesWithMetadata(currentInputVars, this.config.inputs, {
 				materializationIndex: turn,
 				pluginId: "hydra",
 				provider: this.agentProvider,
 				purpose: test?.metadata?.purpose
-			}) : void 0)?.vars ?? currentInputVars;
+			});
+			const currentRenderInputVars = materializedInputVars?.vars ?? currentInputVars;
 			this.conversationHistory.push({
 				role: "user",
 				content: processedMessage
@@ -512,13 +527,16 @@ var HydraProvider = class {
 			role: msg.role,
 			content: msg.content
 		}));
+		const targetProbeCount = totalTokenUsage.numRequests ?? 0;
+		const hydraRoundsCompleted = this.conversationHistory.filter((m) => m.role === "user").length;
+		const failClosedError = targetProbeCount === 0 ? agentFailureError || "Hydra did not execute any target probes" : void 0;
 		return {
 			output: lastTargetResponse?.output || "",
-			...lastTargetResponse?.error ? { error: lastTargetResponse.error } : {},
+			...failClosedError ? { error: failClosedError } : lastTargetResponse?.error ? { error: lastTargetResponse.error } : {},
 			metadata: {
 				sessionId: this.sessionId || getSessionId(lastTargetResponse, context),
 				messages,
-				hydraRoundsCompleted: this.conversationHistory.filter((m) => m.role === "user").length,
+				hydraRoundsCompleted,
 				hydraBacktrackCount: backtrackCount,
 				hydraResult: vulnerabilityAchieved,
 				stopReason,
@@ -539,4 +557,4 @@ var HydraProvider = class {
 //#endregion
 export { HydraProvider, HydraProvider as default };
 
-//# sourceMappingURL=hydra-DE4xWwyc.js.map
+//# sourceMappingURL=hydra-CXc0p63W.js.map