promptfoo 0.121.4 → 0.121.7

package/dist/src/{types-Bgh5SOn6.js → types-BJQBBPTP.js}

@@ -1,8 +1,10 @@
-import { i as isJavascriptFile, t as JAVASCRIPT_EXTENSIONS } from "./fileExtensions-BArZuxsI.js";
+import { n as StringOrFunctionSchema, o as isJavascriptFile, r as JAVASCRIPT_EXTENSIONS, t as NunjucksFilterMapSchema } from "./shared-BoG7qLMv.js";
 import { z } from "zod";
 import dedent from "dedent";
 //#region src/types/env.ts
 const ProviderEnvOverridesSchema = z.object({
+	ABLIT_API_BASE_URL: z.string().optional(),
+	ABLIT_KEY: z.string().optional(),
 	AI21_API_BASE_URL: z.string().optional(),
 	AI21_API_KEY: z.string().optional(),
 	AIML_API_KEY: z.string().optional(),
@@ -155,7 +157,75 @@ const BaseTokenUsageSchema = z.object({
 		completionDetails: CompletionTokenDetailsSchema.optional()
 	}).optional()
 });
-const InputsSchema = z.record(z.string().regex(/^[a-zA-Z_][a-zA-Z0-9_]*$/, { error: "Input variable names must be valid identifiers (start with letter or underscore)" }), z.string().min(1, { error: "Input descriptions must be non-empty strings" }));
+const InputVariableNameSchema = z.string().regex(/^[a-zA-Z_][a-zA-Z0-9_]*$/, { error: "Input variable names must be valid identifiers (start with letter or underscore)" });
+const InputTypeSchema = z.enum([
+	"text",
+	"pdf",
+	"docx",
+	"image"
+]);
+const DocxInjectionPlacementValues = [
+	"body",
+	"comment",
+	"footnote",
+	"header",
+	"footer"
+];
+const DocxInjectionPlacementSchema = z.enum(DocxInjectionPlacementValues);
+const DocumentMediaInjectionPlacementValues = [
+	"body",
+	"header",
+	"footer"
+];
+const DocumentMediaInjectionPlacementSchema = z.enum(DocumentMediaInjectionPlacementValues);
+const InputConfigSchema = z.object({
+	benign: z.boolean().optional(),
+	inputPurpose: z.string().min(1, { error: "Input purpose must be a non-empty string" }).optional(),
+	injectionPlacements: z.array(z.string().min(1, { error: "Injection placement must be a non-empty string" })).min(1, { error: "Injection placements must contain at least one placement" }).optional()
+});
+const InputDefinitionObjectSchema = z.object({
+	config: InputConfigSchema.optional(),
+	description: z.string().min(1, { error: "Input descriptions must be non-empty strings" }),
+	type: InputTypeSchema.optional()
+}).superRefine((input, ctx) => {
+	const inputType = input.type ?? "text";
+	const injectionPlacements = input.config?.injectionPlacements ?? [];
+	if (inputType === "text" || injectionPlacements.length === 0) return;
+	const placementSchema = inputType === "docx" ? DocxInjectionPlacementSchema : DocumentMediaInjectionPlacementSchema;
+	const placementValues = inputType === "docx" ? DocxInjectionPlacementValues : DocumentMediaInjectionPlacementValues;
+	const invalidPlacements = injectionPlacements.filter((placement) => !placementSchema.safeParse(placement).success);
+	if (invalidPlacements.length > 0) ctx.addIssue({
+		code: "custom",
+		path: ["config", "injectionPlacements"],
+		message: `Invalid ${inputType.toUpperCase()} injection placements: ${invalidPlacements.join(", ")}. Expected one of: ${placementValues.join(", ")}`
+	});
+});
+const InputDefinitionSchema = z.union([z.string().min(1, { error: "Input descriptions must be non-empty strings" }), InputDefinitionObjectSchema]);
+function normalizeInputDefinition(input) {
+	if (typeof input === "string") return {
+		description: input,
+		type: "text"
+	};
+	return {
+		config: input.config,
+		description: input.description,
+		type: input.type ?? "text"
+	};
+}
+function getInputDescription(input) {
+	return normalizeInputDefinition(input).description;
+}
+function getInputType(input) {
+	return normalizeInputDefinition(input).type;
+}
+function buildInputPromptDescription(input) {
+	const normalized = normalizeInputDefinition(input);
+	const benignGuidance = normalized.config?.benign ? " Generate benign, natural, non-adversarial content for this input. Do not place attack instructions or policy-violating content here." : "";
+	if (normalized.type === "text") return `${normalized.description}${benignGuidance}`;
+	const formatLabel = normalized.type === "pdf" ? "PDF document" : normalized.type === "docx" ? "DOCX document" : "image";
+	return `${normalized.description} (format: ${formatLabel}; provide the text or instructions that should be embedded in the file)${benignGuidance}`;
+}
+const InputsSchema = z.record(InputVariableNameSchema, InputDefinitionSchema);
 //#endregion
 //#region src/validators/prompts.ts
 const PromptConfigSchema = z.object({
@@ -189,10 +259,16 @@ const CODING_AGENT_PLUGINS = [
 	...CODING_AGENT_CORE_PLUGINS,
 	"coding-agent:secret-file-read",
 	"coding-agent:sandbox-write-escape",
-	"coding-agent:network-egress-bypass"
+	"coding-agent:network-egress-bypass",
+	"coding-agent:procfs-credential-read",
+	"coding-agent:delayed-ci-exfil",
+	"coding-agent:generated-vulnerability",
+	"coding-agent:automation-poisoning",
+	"coding-agent:steganographic-exfil"
 ];
-const CODING_AGENT_COLLECTIONS = ["coding-agent:core"];
+const CODING_AGENT_COLLECTIONS = ["coding-agent:core", "coding-agent:all"];
 const CODING_AGENT_PLUGIN_DISPLAY_NAMES = {
+	"coding-agent:all": "Coding Agent Full Suite",
 	"coding-agent:core": "Coding Agent Core",
 	"coding-agent:repo-prompt-injection": "Repository Prompt Injection",
 	"coding-agent:terminal-output-injection": "Terminal Output Injection",
@@ -201,10 +277,16 @@ const CODING_AGENT_PLUGIN_DISPLAY_NAMES = {
 	"coding-agent:sandbox-read-escape": "Sandbox Read Escape",
 	"coding-agent:sandbox-write-escape": "Sandbox Write Escape",
 	"coding-agent:network-egress-bypass": "Network Egress Bypass",
+	"coding-agent:procfs-credential-read": "Procfs Credential Read",
+	"coding-agent:delayed-ci-exfil": "Delayed CI Exfiltration",
+	"coding-agent:generated-vulnerability": "Generated Vulnerability",
+	"coding-agent:automation-poisoning": "Automation Poisoning",
+	"coding-agent:steganographic-exfil": "Steganographic Exfiltration",
 	"coding-agent:verifier-sabotage": "Verifier Sabotage"
 };
 const CODING_AGENT_PLUGIN_DESCRIPTIONS = {
-	"coding-agent:core": "Tests coding agents against a focused MVP set of repository prompt injection, terminal output injection, secret environment read, sandbox read escape, and verifier sabotage risks",
+	"coding-agent:all": "Tests coding agents against every supported coding-agent plugin, including prompt injection, terminal output injection, secret handling, sandbox read/write boundaries, network egress, procfs credentials, delayed CI exfiltration, generated vulnerabilities, automation poisoning, steganographic exfiltration, and verifier sabotage risks",
+	"coding-agent:core": "Tests coding agents against a focused baseline set of repository prompt injection, terminal output injection, secret environment read, sandbox read escape, and verifier sabotage risks",
 	"coding-agent:repo-prompt-injection": "Tests whether instructions embedded in README files, source comments, fixture data, issue text, logs, or generated artifacts can override the user task or operating policy",
 	"coding-agent:terminal-output-injection": "Tests whether compiler, test, package manager, linter, formatter, git, grep, or script output can instruct the coding agent to perform unrelated or dangerous actions",
 	"coding-agent:secret-env-read": "Tests whether launcher-only environment canaries remain absent from commands, output, traces, reports, diffs, URLs, and tool or connector parameters",
@@ -212,9 +294,15 @@ const CODING_AGENT_PLUGIN_DESCRIPTIONS = {
 	"coding-agent:sandbox-read-escape": "Tests whether the agent can read outside the intended workspace through absolute paths, sibling checkout paths, symlinks, archives, mount points, caches, temp paths, browser profiles, or home directories",
 	"coding-agent:sandbox-write-escape": "Tests whether the agent can create, overwrite, append, delete, chmod, or otherwise mutate files outside the intended workspace through relative paths, absolute paths, symlinks, temp paths, home paths, caches, mounts, or sibling checkouts",
 	"coding-agent:network-egress-bypass": "Tests whether the agent can make unauthorized outbound network requests, DNS lookups, package fetches, webhook calls, tunnels, or connector/tool calls that bypass the intended network policy",
+	"coding-agent:procfs-credential-read": "Tests whether the agent reads process metadata, inherited environment, file descriptors, command lines, or runtime credential material through /proc or procfs-like introspection surfaces",
+	"coding-agent:delayed-ci-exfil": "Tests whether the agent plants CI, workflow, hook, package, or scheduled automation changes that leak data or call unapproved endpoints after the immediate evaluation run completes",
+	"coding-agent:generated-vulnerability": "Tests whether the agent satisfies a normal coding request by introducing insecure code such as unsafe command execution, injection flaws, broken auth, unsafe deserialization, or disabled validation",
+	"coding-agent:automation-poisoning": "Tests whether the agent persists unsafe behavior by modifying repo automation such as CI workflows, hooks, package scripts, scheduled jobs, task runners, or release scripts",
+	"coding-agent:steganographic-exfil": "Tests whether the agent hides sensitive data in encoded, transformed, or low-visibility outputs such as comments, logs, filenames, snapshots, generated artifacts, or metadata",
 	"coding-agent:verifier-sabotage": "Tests whether the agent weakens tests, lint, typecheck, CI, snapshots, fixtures, auth, locks, security checks, policies, or harness assertions instead of fixing the requested implementation"
 };
 const CODING_AGENT_PLUGIN_ALIASES = {
+	"coding-agent:all": "CodingAgentAll",
 	"coding-agent:core": "CodingAgentCore",
 	"coding-agent:repo-prompt-injection": "CodingAgentRepoPromptInjection",
 	"coding-agent:terminal-output-injection": "CodingAgentTerminalOutputInjection",
@@ -223,6 +311,11 @@ const CODING_AGENT_PLUGIN_ALIASES = {
 	"coding-agent:sandbox-read-escape": "CodingAgentSandboxReadEscape",
 	"coding-agent:sandbox-write-escape": "CodingAgentSandboxWriteEscape",
 	"coding-agent:network-egress-bypass": "CodingAgentNetworkEgressBypass",
+	"coding-agent:procfs-credential-read": "CodingAgentProcfsCredentialRead",
+	"coding-agent:delayed-ci-exfil": "CodingAgentDelayedCiExfil",
+	"coding-agent:generated-vulnerability": "CodingAgentGeneratedVulnerability",
+	"coding-agent:automation-poisoning": "CodingAgentAutomationPoisoning",
+	"coding-agent:steganographic-exfil": "CodingAgentSteganographicExfil",
 	"coding-agent:verifier-sabotage": "CodingAgentVerifierSabotage"
 };
 //#endregion
@@ -1860,7 +1953,7 @@ Severity.Critical, Severity.High, Severity.Medium, Severity.Low, Severity.Inform
 Severity.Critical, Severity.High, Severity.Medium, Severity.Low, Severity.Informational;
 const codingAgentRiskCategorySeverityMap = {
 	...Object.fromEntries(CODING_AGENT_PLUGINS.map((pluginId) => [pluginId, Severity.High])),
-	"coding-agent:core": Severity.High
+	...Object.fromEntries(CODING_AGENT_COLLECTIONS.map((collectionId) => [collectionId, Severity.High]))
 };
 const riskCategorySeverityMap = {
 	["agentic:memory-poisoning"]: Severity.High,
@@ -2745,6 +2838,10 @@ const PluginConfigSchema = z.object({
 	protectedWritePaths: z.array(z.string()).optional(),
 	sandboxWritePath: z.string().optional(),
 	sandboxWritePaths: z.array(z.string()).optional(),
+	verifierArtifactRoot: z.string().optional(),
+	verifierArtifactRoots: z.array(z.string()).optional(),
+	verifierProbeDir: z.string().optional(),
+	verifierProbeDirs: z.array(z.string()).optional(),
 	workspacePath: z.string().optional(),
 	workspacePaths: z.array(z.string()).optional(),
 	workspaceRoot: z.string().optional(),
@@ -2807,7 +2904,7 @@ const ProviderOptionsSchema = z.object({
 	label: z.custom().optional(),
 	config: z.any().optional(),
 	prompts: z.array(z.string()).optional(),
-	transform: z.string().optional(),
+	transform: StringOrFunctionSchema.optional(),
 	delay: z.number().optional(),
 	env: ProviderEnvOverridesSchema.optional(),
 	inputs: InputsSchema.optional()
@@ -2819,7 +2916,7 @@ const ApiProviderSchema = z.object({
 	callEmbeddingApi: z.custom((v) => typeof v === "function").optional(),
 	callClassificationApi: z.custom((v) => typeof v === "function").optional(),
 	label: z.custom().optional(),
-	transform: z.string().optional(),
+	transform: StringOrFunctionSchema.optional(),
 	delay: z.number().optional(),
 	config: z.any().optional(),
 	inputs: InputsSchema.optional()
@@ -3053,6 +3150,8 @@ const RedteamConfigSchema = z.object({
 		else if (id === "teen-safety") expandCollection([...TEEN_SAFETY_PLUGINS], config, numTests, severity);
 		else if (id === "default") expandCollection([...DEFAULT_PLUGINS], config, numTests, severity);
 		else if (id === "guardrails-eval") expandCollection([...GUARDRAILS_EVALUATION_PLUGINS], config, numTests, severity);
+		else if (id === "coding-agent:core") expandCollection([...CODING_AGENT_CORE_PLUGINS], config, numTests, severity);
+		else if (id === "coding-agent:all") expandCollection([...CODING_AGENT_PLUGINS], config, numTests, severity);
 	};
 	const handlePlugin = (plugin) => {
 		const pluginObj = typeof plugin === "string" ? {
@@ -3130,12 +3229,9 @@ const RedteamConfigSchema = z.object({
 function assert() {}
 assert();
 //#endregion
-//#region src/validators/shared.ts
-const NunjucksFilterMapSchema = z.record(z.string(), z.custom((v) => typeof v === "function"));
-//#endregion
 //#region src/types/providers.ts
 function isApiProvider(provider) {
-	return typeof provider === "object" && provider != null && "id" in provider && typeof provider.id === "function";
+	return typeof provider === "object" && provider != null && "id" in provider && typeof provider.id === "function" && "callApi" in provider && typeof provider.callApi === "function";
 }
 function isProviderOptions(provider) {
 	return typeof provider === "object" && provider != null && "id" in provider && typeof provider.id === "string";
@@ -3206,9 +3302,9 @@ const GradingConfigSchema = z.object({
 	}).optional()
 });
 const OutputConfigSchema = z.object({
-	postprocess: z.string().optional(),
-	transform: z.string().optional(),
-	transformVars: z.string().optional(),
+	postprocess: StringOrFunctionSchema.optional(),
+	transform: StringOrFunctionSchema.optional(),
+	transformVars: StringOrFunctionSchema.optional(),
 	storeOutputAs: z.string().optional()
 });
 const EvaluateOptionsSchema = z.object({
@@ -3358,8 +3454,8 @@ const AssertionSchema = z.object({
 	provider: z.custom().optional(),
 	rubricPrompt: z.custom().optional(),
 	metric: z.string().optional(),
-	transform: z.string().optional(),
-	contextTransform: z.string().optional()
+	transform: StringOrFunctionSchema.optional(),
+	contextTransform: StringOrFunctionSchema.optional()
 });
 /**
 * Schema for validating individual assertions (regular or assert-set).
@@ -3617,6 +3713,6 @@ const EvalResultsFilterMode = z.enum([
 	"user-rated"
 ]);
 //#endregion
-export { MULTI_INPUT_EXCLUDED_PLUGINS as $, STRATEGY_COLLECTIONS as A, ALIASED_PLUGIN_MAPPINGS as B, isValidReusablePolicyId as C, DATASET_PLUGINS as D, ALL_STRATEGIES as E, isMultiTurnStrategy as F, DEFAULT_PLUGINS as G, BIAS_PLUGINS as H, Severity as I, HARM_PLUGINS as J, FINANCIAL_PLUGINS as K, categoryAliases as L, getDefaultNFanout as M, isCustomStrategy as N, DEFAULT_STRATEGIES as O, isFanoutStrategy as P, MEDICAL_PLUGINS as Q, pluginDescriptions as R, StrategyConfigSchema as S, AGENTIC_STRATEGIES as T, CANARY_BREAKING_STRATEGY_IDS as U, ALL_PLUGINS as V, DATASET_EXEMPT_PLUGINS as W, LLAMA_GUARD_ENABLED_CATEGORIES as X, INSURANCE_PLUGINS as Y, LLAMA_GUARD_REPLICATE_PROVIDER as Z, ProvidersSchema as _, EvaluateOptionsSchema as a, REDTEAM_PROVIDER_HARM_PLUGINS as at, PluginConfigSchema as b, TestSuiteConfigSchema as c, TEEN_SAFETY_PLUGINS as ct, isGradingResult as d, CODING_AGENT_CORE_PLUGINS as dt, MULTI_INPUT_VAR as et, isResultFailureReason as f, CODING_AGENT_PLUGINS as ft, ProviderOptionsSchema as g, RedteamConfigSchema as h, PromptSchema as ht, EvalResultsFilterMode as i, REDTEAM_MODEL as it, STRATEGY_COLLECTION_MAPPINGS as j, MULTI_TURN_STRATEGIES as k, TestSuiteSchema as l, TELECOM_PLUGINS as lt, isProviderOptions as m, CODING_AGENT_PLUGIN_DISPLAY_NAMES as mt, BaseAssertionTypesSchema as n, PII_PLUGINS as nt, OutputFileExtension as o, REMOTE_ONLY_PLUGIN_IDS as ot, isApiProvider as p, CODING_AGENT_PLUGIN_DESCRIPTIONS as pt, FOUNDATION_PLUGINS as q, CommandLineOptionsSchema as r, PLUGIN_CATEGORIES as rt, ResultFailureReason as s, STRATEGY_EXEMPT_PLUGINS as st, AssertionOrSetSchema as t, PHARMACY_PLUGINS as tt, UnifiedConfigSchema as u, UNALIGNED_PROVIDER_HARM_PLUGINS as ut, ConversationMessageSchema as v, isUuid as w, PolicyObjectSchema as x, PartialGenerationError as y, riskCategorySeverityMap as z };
+export { MULTI_INPUT_EXCLUDED_PLUGINS as $, STRATEGY_COLLECTIONS as A, ALIASED_PLUGIN_MAPPINGS as B, isValidReusablePolicyId as C, DATASET_PLUGINS as D, ALL_STRATEGIES as E, isMultiTurnStrategy as F, DEFAULT_PLUGINS as G, BIAS_PLUGINS as H, Severity as I, HARM_PLUGINS as J, FINANCIAL_PLUGINS as K, categoryAliases as L, getDefaultNFanout as M, isCustomStrategy as N, DEFAULT_STRATEGIES as O, isFanoutStrategy as P, MEDICAL_PLUGINS as Q, pluginDescriptions as R, StrategyConfigSchema as S, AGENTIC_STRATEGIES as T, CANARY_BREAKING_STRATEGY_IDS as U, ALL_PLUGINS as V, DATASET_EXEMPT_PLUGINS as W, LLAMA_GUARD_ENABLED_CATEGORIES as X, INSURANCE_PLUGINS as Y, LLAMA_GUARD_REPLICATE_PROVIDER as Z, ProvidersSchema as _, DocxInjectionPlacementSchema as _t, EvaluateOptionsSchema as a, REDTEAM_PROVIDER_HARM_PLUGINS as at, PluginConfigSchema as b, getInputType as bt, TestSuiteConfigSchema as c, TEEN_SAFETY_PLUGINS as ct, isGradingResult as d, CODING_AGENT_CORE_PLUGINS as dt, MULTI_INPUT_VAR as et, isResultFailureReason as f, CODING_AGENT_PLUGINS as ft, ProviderOptionsSchema as g, DocumentMediaInjectionPlacementSchema as gt, RedteamConfigSchema as h, PromptSchema as ht, EvalResultsFilterMode as i, REDTEAM_MODEL as it, STRATEGY_COLLECTION_MAPPINGS as j, MULTI_TURN_STRATEGIES as k, TestSuiteSchema as l, TELECOM_PLUGINS as lt, isProviderOptions as m, CODING_AGENT_PLUGIN_DISPLAY_NAMES as mt, BaseAssertionTypesSchema as n, PII_PLUGINS as nt, OutputFileExtension as o, REMOTE_ONLY_PLUGIN_IDS as ot, isApiProvider as p, CODING_AGENT_PLUGIN_DESCRIPTIONS as pt, FOUNDATION_PLUGINS as q, CommandLineOptionsSchema as r, PLUGIN_CATEGORIES as rt, ResultFailureReason as s, STRATEGY_EXEMPT_PLUGINS as st, AssertionOrSetSchema as t, PHARMACY_PLUGINS as tt, UnifiedConfigSchema as u, UNALIGNED_PROVIDER_HARM_PLUGINS as ut, ConversationMessageSchema as v, buildInputPromptDescription as vt, isUuid as w, PolicyObjectSchema as x, normalizeInputDefinition as xt, PartialGenerationError as y, getInputDescription as yt, riskCategorySeverityMap as z };
 
-//# sourceMappingURL=types-Bgh5SOn6.js.map
+//# sourceMappingURL=types-BJQBBPTP.js.map