promptfoo 0.121.7 → 0.121.8

package/dist/src/{graders-BGP99PdK.js → graders-BQt1BaQe.js}

@@ -1,32 +1,32 @@
 #!/usr/bin/env node
 import { N as state, O as getEnvString, T as getEnvBool, _ as extractJsonObjects, g as extractFirstJsonObject, j as isCI, s as logger, x as safeJsonStringify } from "./logger-BbY6ypFL.js";
-import { _ as REQUEST_TIMEOUT_MS, f as sleep, i as fetchWithTimeout, n as fetchWithProxy } from "./fetch-DXUnXkVU.js";
+import { _ as REQUEST_TIMEOUT_MS, f as sleep, i as fetchWithTimeout, n as fetchWithProxy } from "./fetch-Cpf1U1nO.js";
 import { t as invariant } from "./invariant-B2Rf6avk.js";
-import { o as getUserEmail } from "./accounts-CjFnOPmb.js";
+import { o as getUserEmail } from "./accounts-D6IBfEE0.js";
 import { C as PolicyObjectSchema, M as MULTI_TURN_STRATEGIES, St as buildInputPromptDescription, T as isValidReusablePolicyId, _ as PromptSchema, _t as CODING_AGENT_PLUGINS, p as isApiProvider, vt as CODING_AGENT_PLUGIN_DESCRIPTIONS, yt as CODING_AGENT_PLUGIN_DISPLAY_NAMES } from "./types-BFevViUY.js";
 import { i as isJavascriptFile } from "./fileExtensions-D4GCJ67J.js";
 import { r as importModule } from "./esm-Bexx2PFc.js";
 import { i as extractVariablesFromTemplate, o as getNunjucksEngine } from "./render-CSP99NLm.js";
-import { d as hasCodexDefaultCredentials, l as shouldGenerateRemote, r as getRemoteGenerationUrl, u as getCodexDefaultProviders } from "./remoteGeneration-B1_XsKXU.js";
-import { p as getPoliciesFromCloud } from "./storage-BU4qcnOb.js";
+import { d as hasCodexDefaultCredentials, l as shouldGenerateRemote, r as getRemoteGenerationUrl, u as getCodexDefaultProviders } from "./remoteGeneration-D6UjE2JT.js";
+import { p as getPoliciesFromCloud } from "./storage-EKVWZBNY.js";
 import { n as sha256 } from "./createHash-CgRvs4Fn.js";
-import { a as fetchWithCache, c as isCacheEnabled, o as getCache } from "./cache-DpPWrkTE.js";
+import { a as fetchWithCache, c as isCacheEnabled, o as getCache } from "./cache-BR77mdIR.js";
 import { r as runPython } from "./pythonUtils-CgYxeSmO.js";
-import { B as parseFileUrl, F as maybeLoadToolsFromExternalFile, I as parsePathOrGlob, M as maybeLoadFromExternalFile, j as maybeLoadConfigFromExternalFile, k as getNunjucksEngineForFilePath } from "./util-DURocbYR.js";
-import { t as OpenAiChatCompletionProvider } from "./chat-DqUFcWI0.js";
-import { x as hasGoogleDefaultCredentials } from "./transform-EtD4jAWi.js";
-import { C as DefaultSuggestionsProvider$2, D as AzureChatCompletionProvider, E as AzureEmbeddingProvider, S as DefaultLlmRubricProvider, T as AzureModerationProvider, b as DefaultGradingJsonProvider$2, c as OpenAiModerationProvider, g as DefaultGradingProvider$3, h as DefaultEmbeddingProvider$2, l as MistralChatCompletionProvider, n as loadApiProvider, o as getFileHashes, s as parseScriptParts, u as MistralEmbeddingProvider, w as DefaultSynthesizeProvider$1, x as DefaultGradingProvider$2 } from "./providers-B9KzWxAX.js";
-import { t as OpenAiEmbeddingProvider } from "./embedding--UZVe4_7.js";
+import { B as parseFileUrl, F as maybeLoadToolsFromExternalFile, I as parsePathOrGlob, M as maybeLoadFromExternalFile, j as maybeLoadConfigFromExternalFile, k as getNunjucksEngineForFilePath } from "./util-jZRrXe1P.js";
+import { t as OpenAiChatCompletionProvider } from "./chat-DTdf-J5Q.js";
+import { x as hasGoogleDefaultCredentials } from "./transform-DtTfiGoh.js";
+import { C as DefaultSuggestionsProvider$2, D as AzureChatCompletionProvider, E as AzureEmbeddingProvider, S as DefaultLlmRubricProvider, T as AzureModerationProvider, b as DefaultGradingJsonProvider$2, c as OpenAiModerationProvider, g as DefaultGradingProvider$3, h as DefaultEmbeddingProvider$2, l as MistralChatCompletionProvider, n as loadApiProvider, o as getFileHashes, s as parseScriptParts, u as MistralEmbeddingProvider, w as DefaultSynthesizeProvider$1, x as DefaultGradingProvider$2 } from "./providers-B7TyByfj.js";
+import { t as OpenAiEmbeddingProvider } from "./embedding-CU78FMnw.js";
 import { r as accumulateTokenUsage } from "./tokenUsageUtils-CDet74yk.js";
-import { t as AnthropicMessagesProvider } from "./messages-DGUlSNU7.js";
-import { t as OpenAiResponsesProvider } from "./responses-B3W2JvOQ.js";
+import { t as AnthropicMessagesProvider } from "./messages-K9A8RxBM.js";
+import { t as OpenAiResponsesProvider } from "./responses-CrmWv6iz.js";
 import { r as materializeInputVariablesWithMetadata } from "./inputVariables-DXFdi7AI.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-kDURhgJW.js";
-import { _ as isRateLimitWrapped, f as redteamProviderManager, g as createProviderRateLimitOptions } from "./shared-9WHQ1oNE.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-ETfU_sS9.js";
+import { _ as isRateLimitWrapped, f as redteamProviderManager, g as createProviderRateLimitOptions } from "./shared-DNvim54U.js";
 import { a as PROMPT_DELIMITER, n as maybeFilePath, r as normalizeInput } from "./utils-BFOh20Gb.js";
 import { n as getGeneratedPromptOverLimit, r as getMaxCharsPerMessageModifierValue, t as MAX_CHARS_PER_MESSAGE_MODIFIER_KEY } from "./promptLength-4X-Wd8PG.js";
 import { t as REDTEAM_MEMORY_POISONING_PLUGIN_ID } from "./constants-DH5XYLKZ.js";
-import { n as checkExfilTracking } from "./indirectWebPwn-uyWdHx04.js";
+import { n as checkExfilTracking } from "./indirectWebPwn-BMTXXznx.js";
 import { AsyncLocalStorage } from "node:async_hooks";
 import z$1 from "zod";
 import * as fs$2 from "fs";
@@ -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);
@@ -16165,4 +16165,4 @@ function getGraderById(id) {
 //#endregion
 export { DEFAULT_WEB_SEARCH_PROMPT as $, DivergentRepetitionPlugin as A, sampleArray as B, getPiiLeakTestsForCategory as C, getDefaultProviders as Ct, HarmbenchPlugin as D, getGradingProvider as Dt, ImitationPlugin as E, getAndCheckProvider as Et, AegisPlugin as F, matchesGEval as G, isGraderFailure as H, RedteamGraderBase as I, matchesTrajectoryGoalSuccess as J, matchesLlmRubric as K, RedteamPluginBase as L, CrossSessionLeakPlugin as M, ContractPlugin as N, HallucinationPlugin as O, getProviderCallExecutionContext as Ot, BeavertailsPlugin as P, readProviderPromptMap as Q, getCustomPolicies as R, PlinyPlugin as S, processFileReference as St, IntentPlugin as T, callProviderWithContext as Tt, matchesClosedQa as U, fetchHuggingFaceDataset as V, matchesFactuality as W, processPrompts as X, doRemoteGrading as Y, readPrompts as Z, PoliticsPlugin as _, splitIntoSentences as _t, UnverifiableClaimsPlugin as a, CONTEXT_RECALL as at, isValidPolicyObject as b, getFinalTest as bt, ToolDiscoveryPlugin as c, CONTEXT_RELEVANCE as ct, TeenSafetyDangerousContentPlugin as d, renderLlmRubricPrompt as dt, SELECT_BEST_PROMPT as et, TeenSafetyAgeRestrictedGoodsAndServicesPlugin as f, cosineSimilarity as ft, PromptExtractionPlugin as g, normalizeMatcherTokenUsage as gt, RbacPlugin as h, fail as ht, VLGuardPlugin as i, CONTEXT_FAITHFULNESS_NLI_STATEMENTS as it, DebugAccessPlugin as j, ExcessiveAgencyPlugin as k, withProviderCallExecutionContext as kt, TeenSafetyHarmfulBodyIdealsPlugin as l, CONTEXT_RELEVANCE_BAD as lt, ShellInjectionPlugin as m, euclideanDistance as mt, getGraderById as n, ANSWER_RELEVANCY_GENERATE as nt, UnsafeBenchPlugin as o, CONTEXT_RECALL_ATTRIBUTED_TOKEN as ot, SqlInjectionPlugin as p, dotProduct as pt, matchesPiScore as q, VLSUPlugin as r, CONTEXT_FAITHFULNESS_LONGFORM as rt, ToxicChatPlugin as s, CONTEXT_RECALL_NOT_ATTRIBUTED_TOKEN as st, GRADERS as t, SUGGEST_PROMPTS_SYSTEM_MESSAGE as tt, TeenSafetyDangerousRoleplayPlugin as u, loadRubricPrompt as ut, PolicyPlugin as v, tryParse as vt, OverreliancePlugin as w, DEFAULT_ANTHROPIC_MODEL as wt, makeInlinePolicyIdSync as x, loadFromJavaScriptFile as xt, determinePolicyTypeFromId as y, coerceString as yt, retryWithDeduplication as z };
 
-//# sourceMappingURL=graders-BGP99PdK.js.map
+//# sourceMappingURL=graders-BQt1BaQe.js.map

package/dist/src/{graders-ClrU2fnd.cjs → graders-CBQ2s6gz.cjs}

@@ -2,31 +2,31 @@ const require_rolldown_runtime = require("./rolldown-runtime-D_mwlA32.cjs");
 const require_logger = require("./logger-cfNpzI4o.cjs");
 const require_invariant = require("./invariant-QtnLD03y.cjs");
 const require_types = require("./types-CxJvaY2S.cjs");
-const require_fetch = require("./fetch-Dw4XZHjj.cjs");
+const require_fetch = require("./fetch-Cfm4EuOB.cjs");
 const require_fileExtensions = require("./fileExtensions-BhdwzYaD.cjs");
-const require_accounts = require("./accounts-CmWzeD2d.cjs");
+const require_accounts = require("./accounts-BVz5gHLK.cjs");
 const require_esm = require("./esm-BIKakvNa.cjs");
 const require_render = require("./render-BNTrbmBw.cjs");
-const require_remoteGeneration = require("./remoteGeneration-DS9N3pgB.cjs");
-const require_storage = require("./storage-CA-v9V2v.cjs");
+const require_remoteGeneration = require("./remoteGeneration-BpIYlb_O.cjs");
+const require_storage = require("./storage-A4WnAeN3.cjs");
 const require_pythonUtils = require("./pythonUtils-Cokhluq3.cjs");
-const require_util = require("./util-SPsvFONY.cjs");
+const require_util = require("./util-COnzevXE.cjs");
 const require_createHash = require("./createHash-CSiqnK5P.cjs");
-const require_cache = require("./cache-CPGUA4Yl.cjs");
-const require_chat = require("./chat-DxTDQ83C.cjs");
-const require_transform = require("./transform-DhNkAUs8.cjs");
-const require_providers = require("./providers-BDVVIQM6.cjs");
-const require_embedding = require("./embedding-BbrwopfX.cjs");
+const require_cache = require("./cache-h5MWOBZI.cjs");
+const require_chat = require("./chat-DSyYuTYT.cjs");
+const require_transform = require("./transform-DJkt81VY.cjs");
+const require_providers = require("./providers-V6RBuieY.cjs");
+const require_embedding = require("./embedding-BQIApR18.cjs");
 const require_tokenUsageUtils = require("./tokenUsageUtils-_B-P8IAi.cjs");
-const require_messages = require("./messages-DBPir0TQ.cjs");
-const require_responses = require("./responses-1UFFF9N_.cjs");
+const require_messages = require("./messages-BrZEnHsV.cjs");
+const require_responses = require("./responses-CUARGrhY.cjs");
 const require_inputVariables = require("./inputVariables-Dq9W-Z3a.cjs");
-const require_util$1 = require("./util-CN8om2rz.cjs");
-const require_shared = require("./shared-WkgnDkcg.cjs");
+const require_util$1 = require("./util-DNtUsus_.cjs");
+const require_shared = require("./shared-aHWko3P1.cjs");
 const require_utils = require("./utils-Ve6kuJsa.cjs");
 const require_promptLength = require("./promptLength-BbBbDHNj.cjs");
 const require_constants = require("./constants-a2kYssQk.cjs");
-const require_indirectWebPwn = require("./indirectWebPwn-BJ22AbQa.cjs");
+const require_indirectWebPwn = require("./indirectWebPwn-D7NA9Nsv.cjs");
 let fs = require("fs");
 fs = require_rolldown_runtime.__toESM(fs, 1);
 let path = require("path");
@@ -285,13 +285,13 @@ const DefaultSuggestionsProvider$1 = new require_providers.MistralChatCompletion
 const DefaultSynthesizeProvider = new require_providers.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 require_embedding.OpenAiEmbeddingProvider("text-embedding-3-large");
 const DefaultGradingProvider = new require_chat.OpenAiChatCompletionProvider(DEFAULT_OPENAI_GRADING_MODEL);
 const DefaultGradingJsonProvider = new require_chat.OpenAiChatCompletionProvider(DEFAULT_OPENAI_GRADING_MODEL, { config: { response_format: { type: "json_object" } } });
 const DefaultSuggestionsProvider = new require_chat.OpenAiChatCompletionProvider(DEFAULT_OPENAI_GRADING_MODEL);
 const DefaultModerationProvider = new require_providers.OpenAiModerationProvider("omni-moderation-latest");
-const DefaultWebSearchProvider = new require_responses.OpenAiResponsesProvider("gpt-5.4-2026-03-05", { config: { tools: [{ type: "web_search_preview" }] } });
+const DefaultWebSearchProvider = new require_responses.OpenAiResponsesProvider("gpt-5.5-2026-04-23", { config: { tools: [{ type: "web_search_preview" }] } });
 async function getDefaultProviderPreferences(env) {
 	const hasAnthropicCredentials = Boolean(require_logger.getEnvString("ANTHROPIC_API_KEY") || env?.ANTHROPIC_API_KEY);
 	const hasOpenAiCredentials = Boolean(require_logger.getEnvString("OPENAI_API_KEY") || env?.OPENAI_API_KEY);
@@ -16681,4 +16681,4 @@ Object.defineProperty(exports, "withProviderCallExecutionContext", {
 	}
 });
 
-//# sourceMappingURL=graders-ClrU2fnd.cjs.map
+//# sourceMappingURL=graders-CBQ2s6gz.cjs.map

package/dist/src/{graders-C0nXU_ZP.js → graders-DaRU98zs.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-KD8JjCRJ.js";
-import { m as sleep, r as fetchWithTimeout, t as fetchWithProxy, v as REQUEST_TIMEOUT_MS } from "./fetch-BufrQtvR.js";
+import { m as sleep, r as fetchWithTimeout, t as fetchWithProxy, v as REQUEST_TIMEOUT_MS } from "./fetch-BSSAcMxf.js";
 import { t as invariant } from "./invariant-DIYf9sP1.js";
-import { o as getUserEmail } from "./accounts-DanM1wq_.js";
+import { o as getUserEmail } from "./accounts-BWjqwsrf.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 { d as hasCodexDefaultCredentials, l as shouldGenerateRemote, r as getRemoteGenerationUrl, u as getCodexDefaultProviders } from "./remoteGeneration--D6WjzUm.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 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 getCache, o as isCacheEnabled, r as fetchWithCache } from "./cache-CrioYnaa.js";
 import { r as accumulateTokenUsage } from "./tokenUsageUtils-BjVkdk18.js";
-import { d as getPoliciesFromCloud } from "./storage-QdU-SmvD.js";
+import { d as getPoliciesFromCloud } from "./storage-95htjpLs.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 { A as maybeLoadConfigFromExternalFile, F as parsePathOrGlob, O as getNunjucksEngineForFilePath, P as maybeLoadToolsFromExternalFile, j as maybeLoadFromExternalFile, z as parseFileUrl } from "./util-DwNJzqOV.js";
+import { t as OpenAiChatCompletionProvider } from "./chat-BSos6PvZ.js";
+import { x as hasGoogleDefaultCredentials } from "./transform-BufxPIQL.js";
+import { t as AnthropicMessagesProvider } from "./messages-Bg29Nbit.js";
+import { t as OpenAiResponsesProvider } from "./responses-5Gf5HNOi.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-iH3Sw1yo.js";
+import { t as OpenAiEmbeddingProvider } from "./embedding-CV8lmCnU.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 { 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-DGqkTb2-.js";
+import { _ as isRateLimitWrapped, f as redteamProviderManager, g as createProviderRateLimitOptions } from "./shared-CRq0gGVf.js";
 import { n as getGeneratedPromptOverLimit, r as getMaxCharsPerMessageModifierValue, t as MAX_CHARS_PER_MESSAGE_MODIFIER_KEY } from "./promptLength-0qIHyhA5.js";
 import { t as REDTEAM_MEMORY_POISONING_PLUGIN_ID } from "./constants-BjJV0cRr.js";
-import { n as checkExfilTracking } from "./indirectWebPwn-tNx9OZ35.js";
+import { n as checkExfilTracking } from "./indirectWebPwn-BZFPV7Q9.js";
 import { AsyncLocalStorage } from "node:async_hooks";
 import * as fs$2 from "fs";
 import fs from "fs";
@@ -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);
@@ -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-DaRU98zs.js.map

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

@@ -2,31 +2,31 @@ 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 { 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, p as REQUEST_TIMEOUT_MS, r as fetchWithTimeout, t as fetchWithProxy } from "./fetch-Doks14zQ.js";
 import { i as isJavascriptFile } from "./fileExtensions-CXRfY3Ss.js";
-import { i as getUserEmail } from "./accounts-Ca7WIoPY.js";
+import { i as getUserEmail } from "./accounts-DAv_0iE7.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 { d as hasCodexDefaultCredentials, l as shouldGenerateRemote, r as getRemoteGenerationUrl, u as getCodexDefaultProviders } from "./remoteGeneration-CJC3E0aW.js";
+import { l as getPoliciesFromCloud } from "./storage-B2Ql_oq4.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 { N as parseFileUrl, O as maybeLoadToolsFromExternalFile, S as getNunjucksEngineForFilePath, T as maybeLoadFromExternalFile, k as parsePathOrGlob, w as maybeLoadConfigFromExternalFile } from "./util-DRfqa4xz.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-DdriHsNX.js";
+import { t as OpenAiChatCompletionProvider } from "./chat-BBEnnpQk.js";
+import { x as hasGoogleDefaultCredentials } from "./transform-Wp6s_5QE.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-BGc7tDtQ.js";
+import { t as OpenAiEmbeddingProvider } from "./embedding-CStK0TV6.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 { t as AnthropicMessagesProvider } from "./messages-CBulRaud.js";
+import { t as OpenAiResponsesProvider } from "./responses-BVi7xIdv.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 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-DpV6KT5i.js";
+import { _ as isRateLimitWrapped, f as redteamProviderManager, g as createProviderRateLimitOptions } from "./shared-CSIGeGLl.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 REDTEAM_MEMORY_POISONING_PLUGIN_ID } from "./constants-DZGEFLsu.js";
-import { n as checkExfilTracking } from "./indirectWebPwn-CbjUG0rh.js";
+import { n as checkExfilTracking } from "./indirectWebPwn-CiWB-vVH.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);
@@ -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-KXzjnIim.js.map

package/dist/src/{hydra-DE4xWwyc.js → hydra-BU6GjYoQ.js}

@@ -1,18 +1,18 @@
 import { _ as isValidJson, a as logger } from "./logger-Ct2S6Yx-.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-Doks14zQ.js";
+import { o as isLoggedIntoCloud } from "./accounts-DAv_0iE7.js";
+import { l as shouldGenerateRemote, n as getRemoteGenerationExplicitlyDisabledError, s as neverGenerateRemote, t as getRemoteGenerationDisabledError } from "./remoteGeneration-CJC3E0aW.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 { a as extractPromptFromTags, c as isBasicRefusal, o as getSessionId, r as extractInputVarsFromPrompt } from "./util-DpV6KT5i.js";
+import { a as assertRemoteMaterializationHandled, s as buildRemoteMaterializedInputVariables, t as PromptfooChatCompletionProvider } from "./promptfoo-CqEpj6Sr.js";
+import { c as getTargetResponse, i as externalizeResponseForRedteamHistory, l as isConversationEndedResponse, o as getGraderAssertionValue, t as buildGraderResultAssertion, u as isValidChatMessageArray } from "./shared-CSIGeGLl.js";
+import { n as isBlobStorageEnabled, r as shouldAttemptRemoteBlobUpload } from "./extractor-D3Fv_Tdh.js";
+import { n as checkExfilTracking } from "./indirectWebPwn-CiWB-vVH.js";
+import { n as renderPrompt } from "./evaluatorHelpers-CrRObe2z.js";
+import { t as Strategies } from "./strategies-CD1gHeeQ.js";
+import { a as fetchTraceContext, i as extractTraceIdFromTraceparent, n as formatTraceForMetadata, r as formatTraceSummary, t as resolveTracingOptions } from "./tracingOptions-BOP1FsRt.js";
 import { t as applyRuntimeTransforms } from "./runtimeTransform-DigbjU1r.js";
 //#region src/redteam/providers/hydra/index.ts
 const DEFAULT_MAX_TURNS = 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-KXzjnIim.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-BU6GjYoQ.js.map

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

@@ -1,17 +1,17 @@
 import { _ as isValidJson, a as logger } from "./logger-KD8JjCRJ.js";
-import { m as sleep } from "./fetch-BufrQtvR.js";
+import { m as sleep } from "./fetch-BSSAcMxf.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-BWjqwsrf.js";
+import { l as shouldGenerateRemote, n as getRemoteGenerationExplicitlyDisabledError, s as neverGenerateRemote, t as getRemoteGenerationDisabledError } from "./remoteGeneration--D6WjzUm.js";
 import { a as createEmptyTokenUsage, n as accumulateResponseTokenUsage } from "./tokenUsageUtils-BjVkdk18.js";
-import { n as isBlobStorageEnabled, r as shouldAttemptRemoteBlobUpload } from "./extractor-BR7XAzAL.js";
+import { n as isBlobStorageEnabled, r as shouldAttemptRemoteBlobUpload } from "./extractor-Jp53vs-6.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 { a as extractPromptFromTags, c as isBasicRefusal, o as getSessionId, r as extractInputVarsFromPrompt } from "./util-DGqkTb2-.js";
+import { a as assertRemoteMaterializationHandled, s as buildRemoteMaterializedInputVariables, t as PromptfooChatCompletionProvider } from "./promptfoo-CReYAtfb.js";
+import { c as getTargetResponse, i as externalizeResponseForRedteamHistory, l as isConversationEndedResponse, o as getGraderAssertionValue, t as buildGraderResultAssertion, u as isValidChatMessageArray } from "./shared-CRq0gGVf.js";
+import { n as checkExfilTracking } from "./indirectWebPwn-BZFPV7Q9.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 { a as fetchTraceContext, i as extractTraceIdFromTraceparent, n as formatTraceForMetadata, r as formatTraceSummary, t as resolveTracingOptions } from "./tracingOptions-BvmDa_jH.js";
 import { t as applyRuntimeTransforms } from "./runtimeTransform-BJOpL9Yc.js";
 //#region src/redteam/providers/hydra/index.ts
 const DEFAULT_MAX_TURNS = 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-DaRU98zs.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-BmXHxlyl.js.map