vieval 0.0.1 → 0.0.3

package/dist/{config-D2fe1SnT.mjs → config-CHN24egi.mjs}

@@ -14,4 +14,4 @@ function defineTask(definition) {
 //#endregion
 export { defineTask as n, defineEval as t };
 
-//# sourceMappingURL=config-D2fe1SnT.mjs.map
+//# sourceMappingURL=config-CHN24egi.mjs.map

package/dist/{config-D2fe1SnT.mjs.map → config-CHN24egi.mjs.map}

@@ -1 +1 @@
-{"version":3,"file":"config-D2fe1SnT.mjs","names":[],"sources":["../src/config/define.ts"],"sourcesContent":["import type { EvalDefinition, TaskDefinition } from './types'\n\n/**\n * Returns the provided vieval definition while preserving literal field types.\n */\nexport function defineEval<const TDefinition extends EvalDefinition>(definition: TDefinition): TDefinition {\n  return definition\n}\n\n/**\n * Returns the provided task definition while preserving literal field types.\n */\nexport function defineTask<const TDefinition extends TaskDefinition>(definition: TDefinition): TDefinition {\n  return definition\n}\n"],"mappings":";;;;AAKA,SAAgB,WAAqD,YAAsC;AACzG,QAAO;;;;;AAMT,SAAgB,WAAqD,YAAsC;AACzG,QAAO"}
+{"version":3,"file":"config-CHN24egi.mjs","names":[],"sources":["../src/config/define.ts"],"sourcesContent":["import type { EvalDefinition, TaskDefinition } from './types'\n\n/**\n * Returns the provided vieval definition while preserving literal field types.\n */\nexport function defineEval<const TDefinition extends EvalDefinition>(definition: TDefinition): TDefinition {\n  return definition\n}\n\n/**\n * Returns the provided task definition while preserving literal field types.\n */\nexport function defineTask<const TDefinition extends TaskDefinition>(definition: TDefinition): TDefinition {\n  return definition\n}\n"],"mappings":";;;;AAKA,SAAgB,WAAqD,YAAsC;AACzG,QAAO;;;;;AAMT,SAAgB,WAAqD,YAAsC;AACzG,QAAO"}

package/dist/config.d.mts

@@ -1,3 +1,2 @@
-import { _ as TaskDefinition, a as EvalModule, b as TaskRunOutput, c as MatrixDefinition, d as MatrixRow, f as MatrixValue, g as TaskCaseState, h as TaskCaseReporterPayload, i as EvalDefinition, l as MatrixLayer, m as TaskCaseReporterEndPayload, n as defineTask, o as EvalModuleMap, p as ScopedMatrices, r as CollectedEvalEntry, s as MatrixAxisValues, t as defineEval, u as MatrixPrimitive, v as TaskReporterHooks, y as TaskRunContext } from "./index-DP7jsORl.mjs";
-import { n as ModelDefinition, r as resolveModelByName, t as ConfigHookPlugin } from "./plugin-DVaRZY2x.mjs";
-export { CollectedEvalEntry, ConfigHookPlugin, EvalDefinition, EvalModule, EvalModuleMap, MatrixAxisValues, MatrixDefinition, MatrixLayer, MatrixPrimitive, MatrixRow, MatrixValue, ModelDefinition, ScopedMatrices, TaskCaseReporterEndPayload, TaskCaseReporterPayload, TaskCaseState, TaskDefinition, TaskReporterHooks, TaskRunContext, TaskRunOutput, defineEval, defineTask, resolveModelByName };
+import { M as ModelDefinition, N as resolveModelByName, S as TaskRunOutput, _ as TaskCaseState, a as EvalDefinition, b as TaskReporterHooks, c as MatrixAxisValues, d as MatrixPrimitive, f as MatrixRow, g as TaskCaseReporterPayload, h as TaskCaseReporterEndPayload, i as CollectedEvalEntry, l as MatrixDefinition, m as ScopedMatrices, n as defineEval, o as EvalModule, p as MatrixValue, r as defineTask, s as EvalModuleMap, t as ConfigHookPlugin, u as MatrixLayer, v as TaskDefinition, x as TaskRunContext, y as TaskReporterEventPayload } from "./index-BDMEAmf2.mjs";
+export { CollectedEvalEntry, ConfigHookPlugin, EvalDefinition, EvalModule, EvalModuleMap, MatrixAxisValues, MatrixDefinition, MatrixLayer, MatrixPrimitive, MatrixRow, MatrixValue, ModelDefinition, ScopedMatrices, TaskCaseReporterEndPayload, TaskCaseReporterPayload, TaskCaseState, TaskDefinition, TaskReporterEventPayload, TaskReporterHooks, TaskRunContext, TaskRunOutput, defineEval, defineTask, resolveModelByName };

package/dist/config.mjs

@@ -1,3 +1,3 @@
-import { n as defineTask, t as defineEval } from "./config-D2fe1SnT.mjs";
-import { t as resolveModelByName } from "./models-D_MsBtYw.mjs";
+import { n as defineTask, t as defineEval } from "./config-CHN24egi.mjs";
+import { t as resolveModelByName } from "./models-DIGdOUpJ.mjs";
 export { defineEval, defineTask, resolveModelByName };

package/dist/core/assertions/index.d.mts

@@ -1,2 +1,2 @@
-import { C as expectToolCallArgs, S as expectStructuredOutput, T as toRunScores, _ as expectMustExclude, a as CustomAssertionOptions, b as expectRegex, c as RegexAssertionOptions, d as StructuredOutputAssertionOptions, f as ToolCall, g as expectCustom, h as evaluateAssertions, i as AssertionState, l as RubricAssertionOptions, m as collectFailedAssertions, n as AssertionContext, o as MustExcludeAssertionOptions, p as ToolCallArgsAssertionOptions, r as AssertionOutcome, s as MustIncludeAssertionOptions, t as Assertion, u as RubricJudgeResult, v as expectMustInclude, w as normalizeMatchText, x as expectRubric, y as expectNot } from "../../index-oSXhM1zx.mjs";
+import { C as expectToolCallArgs, S as expectStructuredOutput, T as toRunScores, _ as expectMustExclude, a as CustomAssertionOptions, b as expectRegex, c as RegexAssertionOptions, d as StructuredOutputAssertionOptions, f as ToolCall, g as expectCustom, h as evaluateAssertions, i as AssertionState, l as RubricAssertionOptions, m as collectFailedAssertions, n as AssertionContext, o as MustExcludeAssertionOptions, p as ToolCallArgsAssertionOptions, r as AssertionOutcome, s as MustIncludeAssertionOptions, t as Assertion, u as RubricJudgeResult, v as expectMustInclude, w as normalizeMatchText, x as expectRubric, y as expectNot } from "../../index-C3gPFmcR.mjs";
 export { Assertion, AssertionContext, AssertionOutcome, AssertionState, CustomAssertionOptions, MustExcludeAssertionOptions, MustIncludeAssertionOptions, RegexAssertionOptions, RubricAssertionOptions, RubricJudgeResult, StructuredOutputAssertionOptions, ToolCall, ToolCallArgsAssertionOptions, collectFailedAssertions, evaluateAssertions, expectCustom, expectMustExclude, expectMustInclude, expectNot, expectRegex, expectRubric, expectStructuredOutput, expectToolCallArgs, normalizeMatchText, toRunScores };

package/dist/core/inference-executors/index.d.mts

@@ -1,3 +1,4 @@
+import { a as requiredEnvFrom, i as envFrom, n as EnvValueType, r as RequiredEnvFromOptions, t as EnvFromOptions } from "../../env-DtpjACOW.mjs";
 import { createOpenAI } from "@xsai-ext/providers/create";
 
 //#region src/core/inference-executors/retry-policy.d.ts
@@ -116,51 +117,6 @@ interface ProviderAdapter<TProvider> {
  */
 declare function createProviderAdapter<TProvider>(provider: TProvider, options?: RetryPolicyOptions): ProviderAdapter<TProvider>;
 //#endregion
-//#region src/core/inference-executors/env.d.ts
-/**
- * Supported env value coercion types.
- */
-type EnvValueType = 'string';
-/**
- * Common options for env readers.
- */
-interface EnvFromOptions {
-  /**
-   * Expected env value type.
-   */
-  type: EnvValueType;
-  /**
-   * Whether an empty or missing value should throw.
-   *
-   * @default false
-   */
-  required?: boolean;
-  /**
-   * Optional key name used for clearer error messages.
-   */
-  name?: string;
-}
-/**
- * Env options used by the required helper.
- *
- * `required` is intentionally omitted because this helper is always required.
- */
-type RequiredEnvFromOptions = Omit<EnvFromOptions, 'required'>;
-/**
- * Parses one env value with optional required behavior.
- *
- * Example:
- * `const apiKey = envFrom(process.env.OPENAI_API_KEY, { type: 'string', required: true, name: 'OPENAI_API_KEY' })`
- */
-declare function envFrom(value: string | undefined, options: EnvFromOptions): string | undefined;
-/**
- * Parses one required env value.
- *
- * Example:
- * `const apiKey = requiredEnvFrom(process.env.OPENAI_API_KEY, { type: 'string', name: 'OPENAI_API_KEY' })`
- */
-declare function requiredEnvFrom(value: string | undefined, options: RequiredEnvFromOptions): string;
-//#endregion
 //#region src/core/inference-executors/remote-providers/openai/index.d.ts
 /**
  * Represents the OpenAI provider instance returned by xsai.

package/dist/core/inference-executors/index.mjs

@@ -1,3 +1,4 @@
+import { n as requiredEnvFrom, t as envFrom } from "../../env-C7X81PWa.mjs";
 import process from "node:process";
 import { errorMessageFrom, errorNameFrom, sleep } from "@moeru/std";
 import { createOpenAI } from "@xsai-ext/providers/create";
@@ -125,44 +126,6 @@ function createProviderAdapter(provider, options = {}) {
 	};
 }
 //#endregion
-//#region src/core/inference-executors/env.ts
-function assertNonEmptyString(value, options) {
-	if (value == null || value.trim().length === 0) {
-		if (options.required === true) {
-			const label = options.name ?? "environment variable";
-			throw new Error(`Missing required ${label}.`);
-		}
-		return;
-	}
-	return value;
-}
-/**
-* Parses one env value with optional required behavior.
-*
-* Example:
-* `const apiKey = envFrom(process.env.OPENAI_API_KEY, { type: 'string', required: true, name: 'OPENAI_API_KEY' })`
-*/
-function envFrom(value, options) {
-	if (options.type === "string") return assertNonEmptyString(value, options);
-}
-/**
-* Parses one required env value.
-*
-* Example:
-* `const apiKey = requiredEnvFrom(process.env.OPENAI_API_KEY, { type: 'string', name: 'OPENAI_API_KEY' })`
-*/
-function requiredEnvFrom(value, options) {
-	const parsed = envFrom(value, {
-		...options,
-		required: true
-	});
-	if (parsed == null) {
-		const label = options.name ?? "environment variable";
-		throw new Error(`Missing required ${label}.`);
-	}
-	return parsed;
-}
-//#endregion
 //#region src/core/inference-executors/remote-providers/openai/index.ts
 /**
 * Normalizes provider text output to a safe string.

package/dist/core/inference-executors/index.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"index.mjs","names":["defaultSleep"],"sources":["../../../src/core/inference-executors/retry-policy.ts","../../../src/core/inference-executors/adapters.ts","../../../src/core/inference-executors/env.ts","../../../src/core/inference-executors/remote-providers/openai/index.ts"],"sourcesContent":["import { sleep as defaultSleep, errorMessageFrom, errorNameFrom } from '@moeru/std'\n\n/**\n * Describes how provider retries should behave.\n *\n * ASCII flow:\n * attempt -> run request -> success return\n * attempt -> run request -> retriable failure -> sleep -> next attempt\n * attempt -> run request -> non-retriable failure -> throw\n */\nexport interface RetryPolicy {\n  /**\n   * Maximum number of total attempts, including the first try.\n   */\n  maxAttempts: number\n  /**\n   * Returns the wait time for a retry attempt.\n   */\n  delayMs: (attempt: number) => number\n  /**\n   * Determines whether an error can be retried safely.\n   */\n  shouldRetry: (error: unknown) => boolean\n  /**\n   * Suspends execution between retries.\n   */\n  sleep: (milliseconds: number) => Promise<void>\n}\n\n/**\n * Configures a retry policy before a provider call is executed.\n *\n * Use when:\n * - you want the default retry classifier but need to tune attempts or delay\n * - you need to replace the sleeper in tests\n *\n * Expects:\n * - `maxAttempts` to be a finite integer greater than or equal to `1`\n * - `delayMs` to return a non-negative wait time in milliseconds\n */\nexport interface RetryPolicyOptions {\n  /**\n   * Maximum total attempts, including the first request.\n   *\n   * @default 3\n   */\n  maxAttempts?: number\n  /**\n   * Computes the delay for a retry attempt.\n   *\n   * The attempt number starts at `1` for the first retry.\n   */\n  delayMs?: (attempt: number) => number\n  /**\n   * Overrides the retry classifier.\n   */\n  shouldRetry?: (error: unknown) => boolean\n  /**\n   * Overrides the sleeper used between attempts.\n   */\n  sleep?: (milliseconds: number) => Promise<void>\n}\n\nconst retryableStatusCodes = new Set([408, 425, 429, 500, 502, 503, 504])\nconst retryableErrorNames = new Set(['TimeoutError', 'FetchError'])\nconst retryableMessagePatterns = [\n  /rate limit/i,\n  /rate-limited/i,\n  /temporarily unavailable/i,\n  /service unavailable/i,\n  /server error/i,\n  /fetch failed/i,\n  /network error/i,\n  /socket hang up/i,\n  /econnreset/i,\n  /econnrefused/i,\n  /eai_again/i,\n  /etimedout/i,\n  /timed out/i,\n  /timeout/i,\n]\n\nfunction getStatusCode(error: unknown): number | undefined {\n  if (error == null || typeof error !== 'object') {\n    return undefined\n  }\n\n  const maybeStatusCode = (error as { statusCode?: unknown }).statusCode\n  if (typeof maybeStatusCode === 'number') {\n    return maybeStatusCode\n  }\n\n  const maybeStatus = (error as { status?: unknown }).status\n  if (typeof maybeStatus === 'number') {\n    return maybeStatus\n  }\n\n  const response = (error as { response?: unknown }).response\n  if (response == null || typeof response !== 'object') {\n    return undefined\n  }\n\n  const responseStatus = (response as { status?: unknown }).status\n  return typeof responseStatus === 'number' ? responseStatus : undefined\n}\n\n/**\n * Returns true when a provider failure is temporary and a retry is reasonable.\n *\n * Use when:\n * - the upstream failure is a transport problem or a 5xx/429 response\n *\n * Expects:\n * - provider errors to expose a status code, name, or message when possible\n */\nexport function isRetriableProviderError(error: unknown): boolean {\n  const statusCode = getStatusCode(error)\n\n  if (statusCode != null) {\n    return retryableStatusCodes.has(statusCode)\n  }\n\n  const errorName = errorNameFrom(error)\n  if (errorName != null && retryableErrorNames.has(errorName)) {\n    return true\n  }\n\n  const errorMessage = errorMessageFrom(error)\n  if (errorMessage == null) {\n    return false\n  }\n\n  return retryableMessagePatterns.some(pattern => pattern.test(errorMessage))\n}\n\nfunction defaultDelayMs(attempt: number): number {\n  return 500 * 2 ** (attempt - 1)\n}\n\n/**\n * Creates a retry policy for provider work.\n *\n * Use when:\n * - you need a reusable retry runner for eval-time provider calls\n * - you want to keep retry behavior deterministic in tests\n *\n * Expects:\n * - callers to treat `maxAttempts` as total attempts, not retries\n *\n * Throws:\n * - `RangeError` when `maxAttempts` is not a finite integer greater than or equal to `1`\n */\nfunction assertValidMaxAttempts(value: number): number {\n  if (!Number.isFinite(value) || !Number.isInteger(value) || value < 1) {\n    throw new RangeError('maxAttempts must be a finite integer greater than or equal to 1.')\n  }\n\n  return value\n}\n\nexport function createRetryPolicy(options: RetryPolicyOptions = {}): RetryPolicy {\n  const maxAttempts = assertValidMaxAttempts(options.maxAttempts ?? 3)\n\n  return {\n    maxAttempts,\n    delayMs: options.delayMs ?? defaultDelayMs,\n    shouldRetry: options.shouldRetry ?? isRetriableProviderError,\n    sleep: options.sleep ?? defaultSleep,\n  }\n}\n\n/**\n * Runs an operation with bounded retries.\n *\n * Use when:\n * - you are calling an LLM provider or other temporary upstream dependency\n * - non-retriable failures should bubble immediately\n *\n * Expects:\n * - the operation to be idempotent across attempts\n */\nexport async function runWithRetry<T>(operation: () => Promise<T>, policy: RetryPolicy = createRetryPolicy()): Promise<T> {\n  for (let attempt = 1; attempt <= policy.maxAttempts; attempt += 1) {\n    try {\n      return await operation()\n    }\n    catch (error) {\n      if (attempt >= policy.maxAttempts || !policy.shouldRetry(error)) {\n        throw error\n      }\n\n      const delayMilliseconds = policy.delayMs(attempt)\n      if (delayMilliseconds > 0) {\n        await policy.sleep(delayMilliseconds)\n      }\n    }\n  }\n\n  throw new Error('Retry loop exited without returning a value.')\n}\n","import type { RetryPolicy, RetryPolicyOptions } from './retry-policy'\n\nimport { createRetryPolicy, runWithRetry } from './retry-policy'\n\n/**\n * Bundles a provider with the retry policy used to call it.\n *\n * Use when:\n * - a provider instance should travel with the retry runner that governs it\n * - you want call sites to share one retry configuration object\n */\nexport interface ProviderAdapter<TProvider> {\n  /**\n   * The underlying provider instance.\n   */\n  provider: TProvider\n  /**\n   * The retry policy used for provider calls.\n   */\n  retryPolicy: RetryPolicy\n  /**\n   * Runs a provider-dependent operation with the adapter retry policy.\n   */\n  runWithRetry: <TResult>(operation: () => Promise<TResult>) => Promise<TResult>\n}\n\n/**\n * Creates a provider adapter with the default retry policy.\n *\n * Use when:\n * - you have a provider instance and want a consistent retry wrapper\n *\n * Expects:\n * - the provider to be safe to reuse across attempts\n */\nexport function createProviderAdapter<TProvider>(provider: TProvider, options: RetryPolicyOptions = {}): ProviderAdapter<TProvider> {\n  const retryPolicy = createRetryPolicy(options)\n\n  return {\n    provider,\n    retryPolicy,\n    runWithRetry: operation => runWithRetry(operation, retryPolicy),\n  }\n}\n","/**\n * Supported env value coercion types.\n */\nexport type EnvValueType = 'string'\n\n/**\n * Common options for env readers.\n */\nexport interface EnvFromOptions {\n  /**\n   * Expected env value type.\n   */\n  type: EnvValueType\n  /**\n   * Whether an empty or missing value should throw.\n   *\n   * @default false\n   */\n  required?: boolean\n  /**\n   * Optional key name used for clearer error messages.\n   */\n  name?: string\n}\n\n/**\n * Env options used by the required helper.\n *\n * `required` is intentionally omitted because this helper is always required.\n */\nexport type RequiredEnvFromOptions = Omit<EnvFromOptions, 'required'>\n\nfunction assertNonEmptyString(value: string | undefined, options: EnvFromOptions): string | undefined {\n  if (value == null || value.trim().length === 0) {\n    if (options.required === true) {\n      const label = options.name ?? 'environment variable'\n      throw new Error(`Missing required ${label}.`)\n    }\n\n    return undefined\n  }\n\n  return value\n}\n\n/**\n * Parses one env value with optional required behavior.\n *\n * Example:\n * `const apiKey = envFrom(process.env.OPENAI_API_KEY, { type: 'string', required: true, name: 'OPENAI_API_KEY' })`\n */\nexport function envFrom(\n  value: string | undefined,\n  options: EnvFromOptions,\n): string | undefined {\n  if (options.type === 'string') {\n    return assertNonEmptyString(value, options)\n  }\n\n  return undefined\n}\n\n/**\n * Parses one required env value.\n *\n * Example:\n * `const apiKey = requiredEnvFrom(process.env.OPENAI_API_KEY, { type: 'string', name: 'OPENAI_API_KEY' })`\n */\nexport function requiredEnvFrom(\n  value: string | undefined,\n  options: RequiredEnvFromOptions,\n): string {\n  const parsed = envFrom(value, {\n    ...options,\n    required: true,\n  })\n\n  if (parsed == null) {\n    const label = options.name ?? 'environment variable'\n    throw new Error(`Missing required ${label}.`)\n  }\n\n  return parsed\n}\n","import type { ProviderAdapter } from '../../adapters'\nimport type { RetryPolicyOptions } from '../../retry-policy'\n\nimport process from 'node:process'\n\nimport { createOpenAI } from '@xsai-ext/providers/create'\n\nimport { createProviderAdapter } from '../../adapters'\nimport { envFrom, requiredEnvFrom } from '../../env'\n\n/**\n * Represents the OpenAI provider instance returned by xsai.\n */\nexport type OpenAIProvider = ReturnType<typeof createOpenAI>\n\n/**\n * Represents the OpenAI adapter used by vieval.\n */\nexport type OpenAIProviderAdapter = ProviderAdapter<OpenAIProvider>\n\n/**\n * Configures env key names and source for OpenAI provider setup.\n */\nexport interface OpenAIEnvSourceOptions {\n  /**\n   * Environment object used for variable lookup.\n   *\n   * @default process.env\n   */\n  env?: NodeJS.ProcessEnv\n  /**\n   * Env key name for API key.\n   *\n   * @default 'OPENAI_API_KEY'\n   */\n  apiKey?: string\n  /**\n   * Env key name for base URL.\n   *\n   * @default 'OPENAI_BASE_URL'\n   */\n  baseURL?: string\n  /**\n   * Env key name for model.\n   *\n   * @default 'OPENAI_MODEL'\n   */\n  model?: string\n}\n\n/**\n * Configures fallback defaults when env values are missing.\n */\nexport interface OpenAIFromEnvDefaultOptions {\n  /**\n   * API key fallback value.\n   */\n  apiKey?: string\n  /**\n   * Base URL fallback value.\n   */\n  baseURL?: string\n  /**\n   * Model fallback value.\n   */\n  model?: string\n  /**\n   * Retry policy override passed to provider adapter.\n   */\n  retryOptions?: RetryPolicyOptions\n}\n\n/**\n * Result produced by `createOpenAIFromEnv`.\n */\nexport interface OpenAIFromEnvResult {\n  adapter: OpenAIProviderAdapter\n  apiKey: string\n  baseURL?: string\n  model: string\n}\n\n/**\n * Minimal response shape returned by text-generation calls.\n */\nexport interface OpenAITextGenerationResult {\n  /**\n   * Text output from the provider.\n   *\n   * Some OpenAI-compatible implementations may return `null`.\n   */\n  text?: string | null\n}\n\n/**\n * Normalizes provider text output to a safe string.\n *\n * Before: `{ text: null }`\n * After: `''`\n *\n * Before: `{ text: 'hello' }`\n * After: `'hello'`\n */\nexport function normalizeOpenAITextOutput(result: OpenAITextGenerationResult): string {\n  return typeof result.text === 'string' ? result.text : ''\n}\n\n/**\n * Creates an OpenAI provider adapter using environment variables with defaults.\n *\n * Example:\n * `const runtime = createOpenAIFromEnv({}, { model: 'gpt-4.1-mini' })`\n */\nexport function createOpenAIFromEnv(\n  source: OpenAIEnvSourceOptions = {},\n  defaults: OpenAIFromEnvDefaultOptions = {},\n): OpenAIFromEnvResult {\n  const env = source.env ?? process.env\n  const apiKeyEnvKey = source.apiKey ?? 'OPENAI_API_KEY'\n  const baseURLEnvKey = source.baseURL ?? 'OPENAI_BASE_URL'\n  const modelEnvKey = source.model ?? 'OPENAI_MODEL'\n\n  const apiKey = requiredEnvFrom(env[apiKeyEnvKey] ?? defaults.apiKey, {\n    name: apiKeyEnvKey,\n    type: 'string',\n  })\n  const model = requiredEnvFrom(env[modelEnvKey] ?? defaults.model, {\n    name: modelEnvKey,\n    type: 'string',\n  })\n  const baseURL = envFrom(env[baseURLEnvKey] ?? defaults.baseURL, {\n    name: baseURLEnvKey,\n    type: 'string',\n  })\n  const adapter = createOpenAIProviderAdapter(apiKey, baseURL, defaults.retryOptions)\n\n  return {\n    adapter,\n    apiKey,\n    baseURL,\n    model,\n  }\n}\n\n/**\n * Creates an OpenAI provider adapter for eval-time requests.\n *\n * Use when:\n * - an eval needs the OpenAI SDK surface plus the shared retry runner\n *\n * Expects:\n * - `apiKey` and `baseURL` to point at an OpenAI-compatible endpoint\n * - `retryOptions` to follow the same invariants as `createRetryPolicy`\n */\nexport function createOpenAIProviderAdapter(apiKey: string, baseURL?: string, retryOptions: RetryPolicyOptions = {}): OpenAIProviderAdapter {\n  return createProviderAdapter(createOpenAI(apiKey, baseURL), retryOptions)\n}\n"],"mappings":";;;;AA+DA,MAAM,uBAAuB,IAAI,IAAI;CAAC;CAAK;CAAK;CAAK;CAAK;CAAK;CAAK;CAAI,CAAC;AACzE,MAAM,sBAAsB,IAAI,IAAI,CAAC,gBAAgB,aAAa,CAAC;AACnE,MAAM,2BAA2B;CAC/B;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACD;AAED,SAAS,cAAc,OAAoC;AACzD,KAAI,SAAS,QAAQ,OAAO,UAAU,SACpC;CAGF,MAAM,kBAAmB,MAAmC;AAC5D,KAAI,OAAO,oBAAoB,SAC7B,QAAO;CAGT,MAAM,cAAe,MAA+B;AACpD,KAAI,OAAO,gBAAgB,SACzB,QAAO;CAGT,MAAM,WAAY,MAAiC;AACnD,KAAI,YAAY,QAAQ,OAAO,aAAa,SAC1C;CAGF,MAAM,iBAAkB,SAAkC;AAC1D,QAAO,OAAO,mBAAmB,WAAW,iBAAiB,KAAA;;;;;;;;;;;AAY/D,SAAgB,yBAAyB,OAAyB;CAChE,MAAM,aAAa,cAAc,MAAM;AAEvC,KAAI,cAAc,KAChB,QAAO,qBAAqB,IAAI,WAAW;CAG7C,MAAM,YAAY,cAAc,MAAM;AACtC,KAAI,aAAa,QAAQ,oBAAoB,IAAI,UAAU,CACzD,QAAO;CAGT,MAAM,eAAe,iBAAiB,MAAM;AAC5C,KAAI,gBAAgB,KAClB,QAAO;AAGT,QAAO,yBAAyB,MAAK,YAAW,QAAQ,KAAK,aAAa,CAAC;;AAG7E,SAAS,eAAe,SAAyB;AAC/C,QAAO,MAAM,MAAM,UAAU;;;;;;;;;;;;;;;AAgB/B,SAAS,uBAAuB,OAAuB;AACrD,KAAI,CAAC,OAAO,SAAS,MAAM,IAAI,CAAC,OAAO,UAAU,MAAM,IAAI,QAAQ,EACjE,OAAM,IAAI,WAAW,mEAAmE;AAG1F,QAAO;;AAGT,SAAgB,kBAAkB,UAA8B,EAAE,EAAe;AAG/E,QAAO;EACL,aAHkB,uBAAuB,QAAQ,eAAe,EAAE;EAIlE,SAAS,QAAQ,WAAW;EAC5B,aAAa,QAAQ,eAAe;EACpC,OAAO,QAAQ,SAASA;EACzB;;;;;;;;;;;;AAaH,eAAsB,aAAgB,WAA6B,SAAsB,mBAAmB,EAAc;AACxH,MAAK,IAAI,UAAU,GAAG,WAAW,OAAO,aAAa,WAAW,EAC9D,KAAI;AACF,SAAO,MAAM,WAAW;UAEnB,OAAO;AACZ,MAAI,WAAW,OAAO,eAAe,CAAC,OAAO,YAAY,MAAM,CAC7D,OAAM;EAGR,MAAM,oBAAoB,OAAO,QAAQ,QAAQ;AACjD,MAAI,oBAAoB,EACtB,OAAM,OAAO,MAAM,kBAAkB;;AAK3C,OAAM,IAAI,MAAM,+CAA+C;;;;;;;;;;;;;ACnKjE,SAAgB,sBAAiC,UAAqB,UAA8B,EAAE,EAA8B;CAClI,MAAM,cAAc,kBAAkB,QAAQ;AAE9C,QAAO;EACL;EACA;EACA,eAAc,cAAa,aAAa,WAAW,YAAY;EAChE;;;;ACVH,SAAS,qBAAqB,OAA2B,SAA6C;AACpG,KAAI,SAAS,QAAQ,MAAM,MAAM,CAAC,WAAW,GAAG;AAC9C,MAAI,QAAQ,aAAa,MAAM;GAC7B,MAAM,QAAQ,QAAQ,QAAQ;AAC9B,SAAM,IAAI,MAAM,oBAAoB,MAAM,GAAG;;AAG/C;;AAGF,QAAO;;;;;;;;AAST,SAAgB,QACd,OACA,SACoB;AACpB,KAAI,QAAQ,SAAS,SACnB,QAAO,qBAAqB,OAAO,QAAQ;;;;;;;;AAY/C,SAAgB,gBACd,OACA,SACQ;CACR,MAAM,SAAS,QAAQ,OAAO;EAC5B,GAAG;EACH,UAAU;EACX,CAAC;AAEF,KAAI,UAAU,MAAM;EAClB,MAAM,QAAQ,QAAQ,QAAQ;AAC9B,QAAM,IAAI,MAAM,oBAAoB,MAAM,GAAG;;AAG/C,QAAO;;;;;;;;;;;;;ACqBT,SAAgB,0BAA0B,QAA4C;AACpF,QAAO,OAAO,OAAO,SAAS,WAAW,OAAO,OAAO;;;;;;;;AASzD,SAAgB,oBACd,SAAiC,EAAE,EACnC,WAAwC,EAAE,EACrB;CACrB,MAAM,MAAM,OAAO,OAAO,QAAQ;CAClC,MAAM,eAAe,OAAO,UAAU;CACtC,MAAM,gBAAgB,OAAO,WAAW;CACxC,MAAM,cAAc,OAAO,SAAS;CAEpC,MAAM,SAAS,gBAAgB,IAAI,iBAAiB,SAAS,QAAQ;EACnE,MAAM;EACN,MAAM;EACP,CAAC;CACF,MAAM,QAAQ,gBAAgB,IAAI,gBAAgB,SAAS,OAAO;EAChE,MAAM;EACN,MAAM;EACP,CAAC;CACF,MAAM,UAAU,QAAQ,IAAI,kBAAkB,SAAS,SAAS;EAC9D,MAAM;EACN,MAAM;EACP,CAAC;AAGF,QAAO;EACL,SAHc,4BAA4B,QAAQ,SAAS,SAAS,aAAa;EAIjF;EACA;EACA;EACD;;;;;;;;;;;;AAaH,SAAgB,4BAA4B,QAAgB,SAAkB,eAAmC,EAAE,EAAyB;AAC1I,QAAO,sBAAsB,aAAa,QAAQ,QAAQ,EAAE,aAAa"}
+{"version":3,"file":"index.mjs","names":["defaultSleep"],"sources":["../../../src/core/inference-executors/retry-policy.ts","../../../src/core/inference-executors/adapters.ts","../../../src/core/inference-executors/remote-providers/openai/index.ts"],"sourcesContent":["import { sleep as defaultSleep, errorMessageFrom, errorNameFrom } from '@moeru/std'\n\n/**\n * Describes how provider retries should behave.\n *\n * ASCII flow:\n * attempt -> run request -> success return\n * attempt -> run request -> retriable failure -> sleep -> next attempt\n * attempt -> run request -> non-retriable failure -> throw\n */\nexport interface RetryPolicy {\n  /**\n   * Maximum number of total attempts, including the first try.\n   */\n  maxAttempts: number\n  /**\n   * Returns the wait time for a retry attempt.\n   */\n  delayMs: (attempt: number) => number\n  /**\n   * Determines whether an error can be retried safely.\n   */\n  shouldRetry: (error: unknown) => boolean\n  /**\n   * Suspends execution between retries.\n   */\n  sleep: (milliseconds: number) => Promise<void>\n}\n\n/**\n * Configures a retry policy before a provider call is executed.\n *\n * Use when:\n * - you want the default retry classifier but need to tune attempts or delay\n * - you need to replace the sleeper in tests\n *\n * Expects:\n * - `maxAttempts` to be a finite integer greater than or equal to `1`\n * - `delayMs` to return a non-negative wait time in milliseconds\n */\nexport interface RetryPolicyOptions {\n  /**\n   * Maximum total attempts, including the first request.\n   *\n   * @default 3\n   */\n  maxAttempts?: number\n  /**\n   * Computes the delay for a retry attempt.\n   *\n   * The attempt number starts at `1` for the first retry.\n   */\n  delayMs?: (attempt: number) => number\n  /**\n   * Overrides the retry classifier.\n   */\n  shouldRetry?: (error: unknown) => boolean\n  /**\n   * Overrides the sleeper used between attempts.\n   */\n  sleep?: (milliseconds: number) => Promise<void>\n}\n\nconst retryableStatusCodes = new Set([408, 425, 429, 500, 502, 503, 504])\nconst retryableErrorNames = new Set(['TimeoutError', 'FetchError'])\nconst retryableMessagePatterns = [\n  /rate limit/i,\n  /rate-limited/i,\n  /temporarily unavailable/i,\n  /service unavailable/i,\n  /server error/i,\n  /fetch failed/i,\n  /network error/i,\n  /socket hang up/i,\n  /econnreset/i,\n  /econnrefused/i,\n  /eai_again/i,\n  /etimedout/i,\n  /timed out/i,\n  /timeout/i,\n]\n\nfunction getStatusCode(error: unknown): number | undefined {\n  if (error == null || typeof error !== 'object') {\n    return undefined\n  }\n\n  const maybeStatusCode = (error as { statusCode?: unknown }).statusCode\n  if (typeof maybeStatusCode === 'number') {\n    return maybeStatusCode\n  }\n\n  const maybeStatus = (error as { status?: unknown }).status\n  if (typeof maybeStatus === 'number') {\n    return maybeStatus\n  }\n\n  const response = (error as { response?: unknown }).response\n  if (response == null || typeof response !== 'object') {\n    return undefined\n  }\n\n  const responseStatus = (response as { status?: unknown }).status\n  return typeof responseStatus === 'number' ? responseStatus : undefined\n}\n\n/**\n * Returns true when a provider failure is temporary and a retry is reasonable.\n *\n * Use when:\n * - the upstream failure is a transport problem or a 5xx/429 response\n *\n * Expects:\n * - provider errors to expose a status code, name, or message when possible\n */\nexport function isRetriableProviderError(error: unknown): boolean {\n  const statusCode = getStatusCode(error)\n\n  if (statusCode != null) {\n    return retryableStatusCodes.has(statusCode)\n  }\n\n  const errorName = errorNameFrom(error)\n  if (errorName != null && retryableErrorNames.has(errorName)) {\n    return true\n  }\n\n  const errorMessage = errorMessageFrom(error)\n  if (errorMessage == null) {\n    return false\n  }\n\n  return retryableMessagePatterns.some(pattern => pattern.test(errorMessage))\n}\n\nfunction defaultDelayMs(attempt: number): number {\n  return 500 * 2 ** (attempt - 1)\n}\n\n/**\n * Creates a retry policy for provider work.\n *\n * Use when:\n * - you need a reusable retry runner for eval-time provider calls\n * - you want to keep retry behavior deterministic in tests\n *\n * Expects:\n * - callers to treat `maxAttempts` as total attempts, not retries\n *\n * Throws:\n * - `RangeError` when `maxAttempts` is not a finite integer greater than or equal to `1`\n */\nfunction assertValidMaxAttempts(value: number): number {\n  if (!Number.isFinite(value) || !Number.isInteger(value) || value < 1) {\n    throw new RangeError('maxAttempts must be a finite integer greater than or equal to 1.')\n  }\n\n  return value\n}\n\nexport function createRetryPolicy(options: RetryPolicyOptions = {}): RetryPolicy {\n  const maxAttempts = assertValidMaxAttempts(options.maxAttempts ?? 3)\n\n  return {\n    maxAttempts,\n    delayMs: options.delayMs ?? defaultDelayMs,\n    shouldRetry: options.shouldRetry ?? isRetriableProviderError,\n    sleep: options.sleep ?? defaultSleep,\n  }\n}\n\n/**\n * Runs an operation with bounded retries.\n *\n * Use when:\n * - you are calling an LLM provider or other temporary upstream dependency\n * - non-retriable failures should bubble immediately\n *\n * Expects:\n * - the operation to be idempotent across attempts\n */\nexport async function runWithRetry<T>(operation: () => Promise<T>, policy: RetryPolicy = createRetryPolicy()): Promise<T> {\n  for (let attempt = 1; attempt <= policy.maxAttempts; attempt += 1) {\n    try {\n      return await operation()\n    }\n    catch (error) {\n      if (attempt >= policy.maxAttempts || !policy.shouldRetry(error)) {\n        throw error\n      }\n\n      const delayMilliseconds = policy.delayMs(attempt)\n      if (delayMilliseconds > 0) {\n        await policy.sleep(delayMilliseconds)\n      }\n    }\n  }\n\n  throw new Error('Retry loop exited without returning a value.')\n}\n","import type { RetryPolicy, RetryPolicyOptions } from './retry-policy'\n\nimport { createRetryPolicy, runWithRetry } from './retry-policy'\n\n/**\n * Bundles a provider with the retry policy used to call it.\n *\n * Use when:\n * - a provider instance should travel with the retry runner that governs it\n * - you want call sites to share one retry configuration object\n */\nexport interface ProviderAdapter<TProvider> {\n  /**\n   * The underlying provider instance.\n   */\n  provider: TProvider\n  /**\n   * The retry policy used for provider calls.\n   */\n  retryPolicy: RetryPolicy\n  /**\n   * Runs a provider-dependent operation with the adapter retry policy.\n   */\n  runWithRetry: <TResult>(operation: () => Promise<TResult>) => Promise<TResult>\n}\n\n/**\n * Creates a provider adapter with the default retry policy.\n *\n * Use when:\n * - you have a provider instance and want a consistent retry wrapper\n *\n * Expects:\n * - the provider to be safe to reuse across attempts\n */\nexport function createProviderAdapter<TProvider>(provider: TProvider, options: RetryPolicyOptions = {}): ProviderAdapter<TProvider> {\n  const retryPolicy = createRetryPolicy(options)\n\n  return {\n    provider,\n    retryPolicy,\n    runWithRetry: operation => runWithRetry(operation, retryPolicy),\n  }\n}\n","import type { ProviderAdapter } from '../../adapters'\nimport type { RetryPolicyOptions } from '../../retry-policy'\n\nimport process from 'node:process'\n\nimport { createOpenAI } from '@xsai-ext/providers/create'\n\nimport { createProviderAdapter } from '../../adapters'\nimport { envFrom, requiredEnvFrom } from '../../env'\n\n/**\n * Represents the OpenAI provider instance returned by xsai.\n */\nexport type OpenAIProvider = ReturnType<typeof createOpenAI>\n\n/**\n * Represents the OpenAI adapter used by vieval.\n */\nexport type OpenAIProviderAdapter = ProviderAdapter<OpenAIProvider>\n\n/**\n * Configures env key names and source for OpenAI provider setup.\n */\nexport interface OpenAIEnvSourceOptions {\n  /**\n   * Environment object used for variable lookup.\n   *\n   * @default process.env\n   */\n  env?: NodeJS.ProcessEnv\n  /**\n   * Env key name for API key.\n   *\n   * @default 'OPENAI_API_KEY'\n   */\n  apiKey?: string\n  /**\n   * Env key name for base URL.\n   *\n   * @default 'OPENAI_BASE_URL'\n   */\n  baseURL?: string\n  /**\n   * Env key name for model.\n   *\n   * @default 'OPENAI_MODEL'\n   */\n  model?: string\n}\n\n/**\n * Configures fallback defaults when env values are missing.\n */\nexport interface OpenAIFromEnvDefaultOptions {\n  /**\n   * API key fallback value.\n   */\n  apiKey?: string\n  /**\n   * Base URL fallback value.\n   */\n  baseURL?: string\n  /**\n   * Model fallback value.\n   */\n  model?: string\n  /**\n   * Retry policy override passed to provider adapter.\n   */\n  retryOptions?: RetryPolicyOptions\n}\n\n/**\n * Result produced by `createOpenAIFromEnv`.\n */\nexport interface OpenAIFromEnvResult {\n  adapter: OpenAIProviderAdapter\n  apiKey: string\n  baseURL?: string\n  model: string\n}\n\n/**\n * Minimal response shape returned by text-generation calls.\n */\nexport interface OpenAITextGenerationResult {\n  /**\n   * Text output from the provider.\n   *\n   * Some OpenAI-compatible implementations may return `null`.\n   */\n  text?: string | null\n}\n\n/**\n * Normalizes provider text output to a safe string.\n *\n * Before: `{ text: null }`\n * After: `''`\n *\n * Before: `{ text: 'hello' }`\n * After: `'hello'`\n */\nexport function normalizeOpenAITextOutput(result: OpenAITextGenerationResult): string {\n  return typeof result.text === 'string' ? result.text : ''\n}\n\n/**\n * Creates an OpenAI provider adapter using environment variables with defaults.\n *\n * Example:\n * `const runtime = createOpenAIFromEnv({}, { model: 'gpt-4.1-mini' })`\n */\nexport function createOpenAIFromEnv(\n  source: OpenAIEnvSourceOptions = {},\n  defaults: OpenAIFromEnvDefaultOptions = {},\n): OpenAIFromEnvResult {\n  const env = source.env ?? process.env\n  const apiKeyEnvKey = source.apiKey ?? 'OPENAI_API_KEY'\n  const baseURLEnvKey = source.baseURL ?? 'OPENAI_BASE_URL'\n  const modelEnvKey = source.model ?? 'OPENAI_MODEL'\n\n  const apiKey = requiredEnvFrom(env[apiKeyEnvKey] ?? defaults.apiKey, {\n    name: apiKeyEnvKey,\n    type: 'string',\n  })\n  const model = requiredEnvFrom(env[modelEnvKey] ?? defaults.model, {\n    name: modelEnvKey,\n    type: 'string',\n  })\n  const baseURL = envFrom(env[baseURLEnvKey] ?? defaults.baseURL, {\n    name: baseURLEnvKey,\n    type: 'string',\n  })\n  const adapter = createOpenAIProviderAdapter(apiKey, baseURL, defaults.retryOptions)\n\n  return {\n    adapter,\n    apiKey,\n    baseURL,\n    model,\n  }\n}\n\n/**\n * Creates an OpenAI provider adapter for eval-time requests.\n *\n * Use when:\n * - an eval needs the OpenAI SDK surface plus the shared retry runner\n *\n * Expects:\n * - `apiKey` and `baseURL` to point at an OpenAI-compatible endpoint\n * - `retryOptions` to follow the same invariants as `createRetryPolicy`\n */\nexport function createOpenAIProviderAdapter(apiKey: string, baseURL?: string, retryOptions: RetryPolicyOptions = {}): OpenAIProviderAdapter {\n  return createProviderAdapter(createOpenAI(apiKey, baseURL), retryOptions)\n}\n"],"mappings":";;;;;AA+DA,MAAM,uBAAuB,IAAI,IAAI;CAAC;CAAK;CAAK;CAAK;CAAK;CAAK;CAAK;CAAI,CAAC;AACzE,MAAM,sBAAsB,IAAI,IAAI,CAAC,gBAAgB,aAAa,CAAC;AACnE,MAAM,2BAA2B;CAC/B;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACA;CACD;AAED,SAAS,cAAc,OAAoC;AACzD,KAAI,SAAS,QAAQ,OAAO,UAAU,SACpC;CAGF,MAAM,kBAAmB,MAAmC;AAC5D,KAAI,OAAO,oBAAoB,SAC7B,QAAO;CAGT,MAAM,cAAe,MAA+B;AACpD,KAAI,OAAO,gBAAgB,SACzB,QAAO;CAGT,MAAM,WAAY,MAAiC;AACnD,KAAI,YAAY,QAAQ,OAAO,aAAa,SAC1C;CAGF,MAAM,iBAAkB,SAAkC;AAC1D,QAAO,OAAO,mBAAmB,WAAW,iBAAiB,KAAA;;;;;;;;;;;AAY/D,SAAgB,yBAAyB,OAAyB;CAChE,MAAM,aAAa,cAAc,MAAM;AAEvC,KAAI,cAAc,KAChB,QAAO,qBAAqB,IAAI,WAAW;CAG7C,MAAM,YAAY,cAAc,MAAM;AACtC,KAAI,aAAa,QAAQ,oBAAoB,IAAI,UAAU,CACzD,QAAO;CAGT,MAAM,eAAe,iBAAiB,MAAM;AAC5C,KAAI,gBAAgB,KAClB,QAAO;AAGT,QAAO,yBAAyB,MAAK,YAAW,QAAQ,KAAK,aAAa,CAAC;;AAG7E,SAAS,eAAe,SAAyB;AAC/C,QAAO,MAAM,MAAM,UAAU;;;;;;;;;;;;;;;AAgB/B,SAAS,uBAAuB,OAAuB;AACrD,KAAI,CAAC,OAAO,SAAS,MAAM,IAAI,CAAC,OAAO,UAAU,MAAM,IAAI,QAAQ,EACjE,OAAM,IAAI,WAAW,mEAAmE;AAG1F,QAAO;;AAGT,SAAgB,kBAAkB,UAA8B,EAAE,EAAe;AAG/E,QAAO;EACL,aAHkB,uBAAuB,QAAQ,eAAe,EAAE;EAIlE,SAAS,QAAQ,WAAW;EAC5B,aAAa,QAAQ,eAAe;EACpC,OAAO,QAAQ,SAASA;EACzB;;;;;;;;;;;;AAaH,eAAsB,aAAgB,WAA6B,SAAsB,mBAAmB,EAAc;AACxH,MAAK,IAAI,UAAU,GAAG,WAAW,OAAO,aAAa,WAAW,EAC9D,KAAI;AACF,SAAO,MAAM,WAAW;UAEnB,OAAO;AACZ,MAAI,WAAW,OAAO,eAAe,CAAC,OAAO,YAAY,MAAM,CAC7D,OAAM;EAGR,MAAM,oBAAoB,OAAO,QAAQ,QAAQ;AACjD,MAAI,oBAAoB,EACtB,OAAM,OAAO,MAAM,kBAAkB;;AAK3C,OAAM,IAAI,MAAM,+CAA+C;;;;;;;;;;;;;ACnKjE,SAAgB,sBAAiC,UAAqB,UAA8B,EAAE,EAA8B;CAClI,MAAM,cAAc,kBAAkB,QAAQ;AAE9C,QAAO;EACL;EACA;EACA,eAAc,cAAa,aAAa,WAAW,YAAY;EAChE;;;;;;;;;;;;;AC6DH,SAAgB,0BAA0B,QAA4C;AACpF,QAAO,OAAO,OAAO,SAAS,WAAW,OAAO,OAAO;;;;;;;;AASzD,SAAgB,oBACd,SAAiC,EAAE,EACnC,WAAwC,EAAE,EACrB;CACrB,MAAM,MAAM,OAAO,OAAO,QAAQ;CAClC,MAAM,eAAe,OAAO,UAAU;CACtC,MAAM,gBAAgB,OAAO,WAAW;CACxC,MAAM,cAAc,OAAO,SAAS;CAEpC,MAAM,SAAS,gBAAgB,IAAI,iBAAiB,SAAS,QAAQ;EACnE,MAAM;EACN,MAAM;EACP,CAAC;CACF,MAAM,QAAQ,gBAAgB,IAAI,gBAAgB,SAAS,OAAO;EAChE,MAAM;EACN,MAAM;EACP,CAAC;CACF,MAAM,UAAU,QAAQ,IAAI,kBAAkB,SAAS,SAAS;EAC9D,MAAM;EACN,MAAM;EACP,CAAC;AAGF,QAAO;EACL,SAHc,4BAA4B,QAAQ,SAAS,SAAS,aAAa;EAIjF;EACA;EACA;EACD;;;;;;;;;;;;AAaH,SAAgB,4BAA4B,QAAgB,SAAkB,eAAmC,EAAE,EAAyB;AAC1I,QAAO,sBAAsB,aAAa,QAAQ,QAAQ,EAAE,aAAa"}

package/dist/core/processors/results/index.d.mts

@@ -1,4 +1,4 @@
-import { I as AggregatedRunResults } from "../../../index-DP7jsORl.mjs";
+import { B as AggregatedRunResults } from "../../../index-BDMEAmf2.mjs";
 
 //#region src/core/processors/results/policies/hybrid-threshold.d.ts
 /**

package/dist/core/runner/index.d.mts

@@ -1,2 +1,2 @@
-import { A as asProjectRelativePath, B as RunScoreKind, C as RunnerTaskState, D as TaskExecutionContext, E as CreateTaskExecutionContextOptions, F as AggregatedProviderSummary, G as RunnerMatrixInput, H as CreateRunnerScheduleOptions, I as AggregatedRunResults, J as ScheduledTaskMatrix, K as RunnerMatrixSelection, L as AggregatedRunSummary, M as CreateVievalRunnerRuntimeContextOptions, N as RunnerRuntimeContext, O as TaskModelSelectionOptions, P as createRunnerRuntimeContext, R as RunResult, S as RunnerExecutionError, T as runScheduledTasks, U as InferenceExecutor, V as aggregateRunResults, W as RunnerMatrixDefinition, X as createRunnerSchedule, Y as ScheduledTaskMatrixMeta, j as collectEvalEntries, k as createTaskExecutionContext, q as ScheduledTask, w as ScheduledTaskExecutor, x as RunScheduledTasksOptions, z as RunScore } from "../../index-DP7jsORl.mjs";
-export { AggregatedProviderSummary, AggregatedRunResults, AggregatedRunSummary, CreateRunnerScheduleOptions, CreateTaskExecutionContextOptions, CreateVievalRunnerRuntimeContextOptions, InferenceExecutor, RunResult, RunScheduledTasksOptions, RunScore, RunScoreKind, RunnerExecutionError, RunnerMatrixDefinition, RunnerMatrixInput, RunnerMatrixSelection, RunnerRuntimeContext, RunnerTaskState, ScheduledTask, ScheduledTaskExecutor, ScheduledTaskMatrix, ScheduledTaskMatrixMeta, TaskExecutionContext, TaskModelSelectionOptions, aggregateRunResults, asProjectRelativePath, collectEvalEntries, createRunnerRuntimeContext, createRunnerSchedule, createTaskExecutionContext, runScheduledTasks };
+import { $ as ScheduledTaskMatrixMeta, A as TaskModelSelectionOptions, B as AggregatedRunResults, C as RunScheduledTasksOptions, D as runScheduledTasks, E as ScheduledTaskExecutor, F as collectEvalEntries, G as aggregateRunResults, H as RunResult, I as CreateVievalRunnerRuntimeContextOptions, J as RunnerMatrixDefinition, K as CreateRunnerScheduleOptions, L as RunnerRuntimeContext, O as CreateTaskExecutionContextOptions, P as asProjectRelativePath, Q as ScheduledTaskMatrix, R as createRunnerRuntimeContext, T as RunnerTaskState, U as RunScore, V as AggregatedRunSummary, W as RunScoreKind, X as RunnerMatrixSelection, Y as RunnerMatrixInput, Z as ScheduledTask, at as CacheFileOptions, et as createRunnerSchedule, it as CacheFileHandle, j as createTaskExecutionContext, k as TaskExecutionContext, nt as createFilesystemTaskCacheRuntime, ot as CacheNamespace, q as InferenceExecutor, rt as normalizeCacheFilePathSegments, st as TaskCacheRuntime, tt as CreateFilesystemTaskCacheRuntimeOptions, w as RunnerExecutionError, z as AggregatedProviderSummary } from "../../index-BDMEAmf2.mjs";
+export { AggregatedProviderSummary, AggregatedRunResults, AggregatedRunSummary, CacheFileHandle, CacheFileOptions, CacheNamespace, CreateFilesystemTaskCacheRuntimeOptions, CreateRunnerScheduleOptions, CreateTaskExecutionContextOptions, CreateVievalRunnerRuntimeContextOptions, InferenceExecutor, RunResult, RunScheduledTasksOptions, RunScore, RunScoreKind, RunnerExecutionError, RunnerMatrixDefinition, RunnerMatrixInput, RunnerMatrixSelection, RunnerRuntimeContext, RunnerTaskState, ScheduledTask, ScheduledTaskExecutor, ScheduledTaskMatrix, ScheduledTaskMatrixMeta, TaskCacheRuntime, TaskExecutionContext, TaskModelSelectionOptions, aggregateRunResults, asProjectRelativePath, collectEvalEntries, createFilesystemTaskCacheRuntime, createRunnerRuntimeContext, createRunnerSchedule, createTaskExecutionContext, normalizeCacheFilePathSegments, runScheduledTasks };

package/dist/core/runner/index.mjs

@@ -1,2 +1,2 @@
-import { a as runScheduledTasks, c as aggregateRunResults, i as RunnerExecutionError, n as createRunnerSchedule, o as asProjectRelativePath, r as createRunnerRuntimeContext, s as collectEvalEntries, t as createTaskExecutionContext } from "../../runner-4ZsOveoY.mjs";
-export { RunnerExecutionError, aggregateRunResults, asProjectRelativePath, collectEvalEntries, createRunnerRuntimeContext, createRunnerSchedule, createTaskExecutionContext, runScheduledTasks };
+import { a as runScheduledTasks, c as aggregateRunResults, i as RunnerExecutionError, l as createFilesystemTaskCacheRuntime, n as createRunnerSchedule, o as asProjectRelativePath, r as createRunnerRuntimeContext, s as collectEvalEntries, t as createTaskExecutionContext, u as normalizeCacheFilePathSegments } from "../../runner-Dpy-eivM.mjs";
+export { RunnerExecutionError, aggregateRunResults, asProjectRelativePath, collectEvalEntries, createFilesystemTaskCacheRuntime, createRunnerRuntimeContext, createRunnerSchedule, createTaskExecutionContext, normalizeCacheFilePathSegments, runScheduledTasks };

package/dist/env-C7X81PWa.mjs

@@ -0,0 +1,41 @@
+//#region src/core/inference-executors/env.ts
+function assertNonEmptyString(value, options) {
+	if (value == null || value.trim().length === 0) {
+		if (options.required === true) {
+			const label = options.name ?? "environment variable";
+			throw new Error(`Missing required ${label}.`);
+		}
+		return;
+	}
+	return value;
+}
+/**
+* Parses one env value with optional required behavior.
+*
+* Example:
+* `const apiKey = envFrom(process.env.OPENAI_API_KEY, { type: 'string', required: true, name: 'OPENAI_API_KEY' })`
+*/
+function envFrom(value, options) {
+	if (options.type === "string") return assertNonEmptyString(value, options);
+}
+/**
+* Parses one required env value.
+*
+* Example:
+* `const apiKey = requiredEnvFrom(process.env.OPENAI_API_KEY, { type: 'string', name: 'OPENAI_API_KEY' })`
+*/
+function requiredEnvFrom(value, options) {
+	const parsed = envFrom(value, {
+		...options,
+		required: true
+	});
+	if (parsed == null) {
+		const label = options.name ?? "environment variable";
+		throw new Error(`Missing required ${label}.`);
+	}
+	return parsed;
+}
+//#endregion
+export { requiredEnvFrom as n, envFrom as t };
+
+//# sourceMappingURL=env-C7X81PWa.mjs.map

package/dist/env-C7X81PWa.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"env-C7X81PWa.mjs","names":[],"sources":["../src/core/inference-executors/env.ts"],"sourcesContent":["/**\n * Supported env value coercion types.\n */\nexport type EnvValueType = 'string'\n\n/**\n * Common options for env readers.\n */\nexport interface EnvFromOptions {\n  /**\n   * Expected env value type.\n   */\n  type: EnvValueType\n  /**\n   * Whether an empty or missing value should throw.\n   *\n   * @default false\n   */\n  required?: boolean\n  /**\n   * Optional key name used for clearer error messages.\n   */\n  name?: string\n}\n\n/**\n * Env options used by the required helper.\n *\n * `required` is intentionally omitted because this helper is always required.\n */\nexport type RequiredEnvFromOptions = Omit<EnvFromOptions, 'required'>\n\nfunction assertNonEmptyString(value: string | undefined, options: EnvFromOptions): string | undefined {\n  if (value == null || value.trim().length === 0) {\n    if (options.required === true) {\n      const label = options.name ?? 'environment variable'\n      throw new Error(`Missing required ${label}.`)\n    }\n\n    return undefined\n  }\n\n  return value\n}\n\n/**\n * Parses one env value with optional required behavior.\n *\n * Example:\n * `const apiKey = envFrom(process.env.OPENAI_API_KEY, { type: 'string', required: true, name: 'OPENAI_API_KEY' })`\n */\nexport function envFrom(\n  value: string | undefined,\n  options: EnvFromOptions,\n): string | undefined {\n  if (options.type === 'string') {\n    return assertNonEmptyString(value, options)\n  }\n\n  return undefined\n}\n\n/**\n * Parses one required env value.\n *\n * Example:\n * `const apiKey = requiredEnvFrom(process.env.OPENAI_API_KEY, { type: 'string', name: 'OPENAI_API_KEY' })`\n */\nexport function requiredEnvFrom(\n  value: string | undefined,\n  options: RequiredEnvFromOptions,\n): string {\n  const parsed = envFrom(value, {\n    ...options,\n    required: true,\n  })\n\n  if (parsed == null) {\n    const label = options.name ?? 'environment variable'\n    throw new Error(`Missing required ${label}.`)\n  }\n\n  return parsed\n}\n"],"mappings":";AAgCA,SAAS,qBAAqB,OAA2B,SAA6C;AACpG,KAAI,SAAS,QAAQ,MAAM,MAAM,CAAC,WAAW,GAAG;AAC9C,MAAI,QAAQ,aAAa,MAAM;GAC7B,MAAM,QAAQ,QAAQ,QAAQ;AAC9B,SAAM,IAAI,MAAM,oBAAoB,MAAM,GAAG;;AAG/C;;AAGF,QAAO;;;;;;;;AAST,SAAgB,QACd,OACA,SACoB;AACpB,KAAI,QAAQ,SAAS,SACnB,QAAO,qBAAqB,OAAO,QAAQ;;;;;;;;AAY/C,SAAgB,gBACd,OACA,SACQ;CACR,MAAM,SAAS,QAAQ,OAAO;EAC5B,GAAG;EACH,UAAU;EACX,CAAC;AAEF,KAAI,UAAU,MAAM;EAClB,MAAM,QAAQ,QAAQ,QAAQ;AAC9B,QAAM,IAAI,MAAM,oBAAoB,MAAM,GAAG;;AAG/C,QAAO"}

package/dist/env-DtpjACOW.d.mts

@@ -0,0 +1,47 @@
+//#region src/core/inference-executors/env.d.ts
+/**
+ * Supported env value coercion types.
+ */
+type EnvValueType = 'string';
+/**
+ * Common options for env readers.
+ */
+interface EnvFromOptions {
+  /**
+   * Expected env value type.
+   */
+  type: EnvValueType;
+  /**
+   * Whether an empty or missing value should throw.
+   *
+   * @default false
+   */
+  required?: boolean;
+  /**
+   * Optional key name used for clearer error messages.
+   */
+  name?: string;
+}
+/**
+ * Env options used by the required helper.
+ *
+ * `required` is intentionally omitted because this helper is always required.
+ */
+type RequiredEnvFromOptions = Omit<EnvFromOptions, 'required'>;
+/**
+ * Parses one env value with optional required behavior.
+ *
+ * Example:
+ * `const apiKey = envFrom(process.env.OPENAI_API_KEY, { type: 'string', required: true, name: 'OPENAI_API_KEY' })`
+ */
+declare function envFrom(value: string | undefined, options: EnvFromOptions): string | undefined;
+/**
+ * Parses one required env value.
+ *
+ * Example:
+ * `const apiKey = requiredEnvFrom(process.env.OPENAI_API_KEY, { type: 'string', name: 'OPENAI_API_KEY' })`
+ */
+declare function requiredEnvFrom(value: string | undefined, options: RequiredEnvFromOptions): string;
+//#endregion
+export { requiredEnvFrom as a, envFrom as i, EnvValueType as n, RequiredEnvFromOptions as r, EnvFromOptions as t };
+//# sourceMappingURL=env-DtpjACOW.d.mts.map

package/dist/expect-B2vaoRVZ.d.mts

@@ -0,0 +1,10 @@
+import * as _$vitest from "vitest";
+
+//#region src/expect.d.ts
+/**
+ * Re-exported expect with vieval custom matchers pre-installed.
+ */
+declare const expect: _$vitest.ExpectStatic;
+//#endregion
+export { expect as t };
+//# sourceMappingURL=expect-B2vaoRVZ.d.mts.map

package/dist/{expect-i9WZWGrA.mjs → expect-CaXiUkwY.mjs}

@@ -1,4 +1,4 @@
-import { n as globalExpect, t as installVievalExpectMatchers } from "./expect-extensions-CwPtgTz8.mjs";
+import { n as getRuntimeExpect, t as installVievalExpectMatchers } from "./expect-extensions-BOzwV5EJ.mjs";
 //#region src/expect.ts
 let isInstalled = false;
 function ensureExpectMatchersInstalled() {
@@ -10,8 +10,8 @@ ensureExpectMatchersInstalled();
 /**
 * Re-exported expect with vieval custom matchers pre-installed.
 */
-const expect = globalExpect;
+const expect = getRuntimeExpect();
 //#endregion
 export { expect as t };
 
-//# sourceMappingURL=expect-i9WZWGrA.mjs.map
+//# sourceMappingURL=expect-CaXiUkwY.mjs.map

package/dist/expect-CaXiUkwY.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"expect-CaXiUkwY.mjs","names":[],"sources":["../src/expect.ts"],"sourcesContent":["import { installVievalExpectMatchers } from './testing/expect-extensions'\nimport { getRuntimeExpect } from './testing/runtime-expect'\n\nlet isInstalled = false\n\nfunction ensureExpectMatchersInstalled(): void {\n  if (isInstalled) {\n    return\n  }\n\n  installVievalExpectMatchers()\n  isInstalled = true\n}\n\nensureExpectMatchersInstalled()\n\n/**\n * Re-exported expect with vieval custom matchers pre-installed.\n */\nexport const expect = getRuntimeExpect()\n"],"mappings":";;AAGA,IAAI,cAAc;AAElB,SAAS,gCAAsC;AAC7C,KAAI,YACF;AAGF,8BAA6B;AAC7B,eAAc;;AAGhB,+BAA+B;;;;AAK/B,MAAa,SAAS,kBAAkB"}

package/dist/expect-extensions-BOzwV5EJ.mjs

@@ -0,0 +1,197 @@
+import { d as normalizeMatchText } from "./assertions-DcAjfVDA.mjs";
+import { ASYMMETRIC_MATCHERS_OBJECT, ChaiStyleAssertions, GLOBAL_EXPECT, JestAsymmetricMatchers, JestChaiExpect, JestExtend, addCustomEqualityTesters, chai, customMatchers, getState, setState } from "@vitest/expect";
+//#region src/testing/runtime-expect.ts
+let isPluginInstalled = false;
+let runtimeExpectInstance;
+/**
+* Installs Vitest expect plugins once for process-local runtime assertions.
+*
+* Use when:
+* - running eval tasks outside Vitest worker runtime
+* - building an `expect` instance that does not rely on Vitest internal state
+*
+* Expects:
+* - `@vitest/expect` is available in runtime dependencies
+*
+* Returns:
+* - nothing; side-effects are applied to `chai`
+*/
+function ensureRuntimeExpectPluginsInstalled() {
+	if (isPluginInstalled) return;
+	chai.use(JestExtend);
+	chai.use(JestChaiExpect);
+	chai.use(ChaiStyleAssertions);
+	chai.use(JestAsymmetricMatchers);
+	isPluginInstalled = true;
+}
+/**
+* Creates a Vitest-compatible `expect` instance without worker-state coupling.
+*
+* Use when:
+* - CLI runtime needs assertion helpers from `vieval/expect`
+* - code is executed outside `vitest run`
+*
+* Expects:
+* - plugins from {@link ensureRuntimeExpectPluginsInstalled} are installed
+* - callers do not depend on Vitest worker-only features (snapshot/poll internals)
+*
+* Returns:
+* - standalone expect instance with core matcher APIs and `extend`
+*/
+function createRuntimeExpect() {
+	ensureRuntimeExpectPluginsInstalled();
+	const runtimeExpect = ((value, message) => {
+		setState({ assertionCalls: getState(runtimeExpect).assertionCalls + 1 }, runtimeExpect);
+		return chai.expect(value, message);
+	});
+	Object.assign(runtimeExpect, chai.expect);
+	Object.assign(runtimeExpect, globalThis[ASYMMETRIC_MATCHERS_OBJECT]);
+	runtimeExpect.getState = () => getState(runtimeExpect);
+	runtimeExpect.setState = (state) => setState(state, runtimeExpect);
+	runtimeExpect.assert = chai.assert;
+	const chaiExpectWithExtend = chai.expect;
+	runtimeExpect.extend = (matchers) => chaiExpectWithExtend.extend(runtimeExpect, matchers);
+	runtimeExpect.addEqualityTesters = (customTesters) => addCustomEqualityTesters(customTesters);
+	runtimeExpect.unreachable = (message) => {
+		chai.assert.fail(`expected${message ? ` "${message}" ` : " "}not to be reached`);
+	};
+	runtimeExpect.setState({
+		assertionCalls: 0,
+		currentTestName: "",
+		expectedAssertionsNumber: null,
+		expectedAssertionsNumberErrorGen: null,
+		isExpectingAssertions: false,
+		isExpectingAssertionsError: null
+	});
+	runtimeExpect.extend(customMatchers);
+	return runtimeExpect;
+}
+/**
+* Returns process-local runtime `expect` instance used by Vieval.
+*
+* Use when:
+* - you need matcher assertions in eval files and CLI runtime
+* - importing from `vitest` would crash outside Vitest worker contexts
+*
+* Expects:
+* - single-process usage (instance is memoized per process)
+*
+* Returns:
+* - memoized runtime `expect` instance
+*/
+function getRuntimeExpect() {
+	if (runtimeExpectInstance != null) return runtimeExpectInstance;
+	runtimeExpectInstance = createRuntimeExpect();
+	Object.defineProperty(globalThis, GLOBAL_EXPECT, {
+		configurable: true,
+		value: runtimeExpectInstance,
+		writable: true
+	});
+	return runtimeExpectInstance;
+}
+//#endregion
+//#region src/testing/expect-extensions.ts
+function toKeywordArray(keywords) {
+	if (typeof keywords === "string") return [keywords];
+	return keywords;
+}
+/**
+* Registers vieval custom matchers on Vitest `expect`.
+*
+* Call stack:
+*
+* {@link installVievalExpectMatchers}
+*   -> `expect.extend(...)`
+*     -> `expect(received).toMustInclude(...)`
+*     -> `expect(received).toScoreRubricGreaterThan(...)`
+*
+* Use when:
+* - eval suites need domain assertions while preserving native Vitest ergonomics
+* - callers want native `.not` chaining with the same matchers
+*/
+function installVievalExpectMatchers() {
+	getRuntimeExpect().extend({
+		toMustExclude(received, keywords, options = {}) {
+			const keywordList = toKeywordArray(keywords);
+			if (typeof received !== "string") return {
+				message: () => "Expected received value to be a string.",
+				pass: false
+			};
+			const normalizedText = normalizeMatchText(received, options.caseSensitive ?? false);
+			const forbiddenMatches = keywordList.filter((keyword) => {
+				return normalizedText.includes(normalizeMatchText(keyword, options.caseSensitive ?? false));
+			});
+			const pass = forbiddenMatches.length === 0;
+			return {
+				message: () => {
+					if (pass) return `Expected text to include forbidden keywords: ${keywordList.join(", ")}`;
+					return `Expected text not to include forbidden keywords, but matched: ${forbiddenMatches.join(", ")}`;
+				},
+				pass
+			};
+		},
+		toMustInclude(received, keywords, options = {}) {
+			const keywordList = toKeywordArray(keywords);
+			if (typeof received !== "string") return {
+				message: () => "Expected received value to be a string.",
+				pass: false
+			};
+			const normalizedText = normalizeMatchText(received, options.caseSensitive ?? false);
+			const matches = keywordList.filter((keyword) => {
+				return normalizedText.includes(normalizeMatchText(keyword, options.caseSensitive ?? false));
+			});
+			const mode = options.mode ?? "all";
+			const pass = mode === "all" ? matches.length === keywordList.length : matches.length > 0;
+			return {
+				message: () => {
+					if (pass) return `Expected text not to match required keywords, but matched: ${matches.join(", ")}`;
+					return `Expected text to match required keywords (${mode}), but matched ${matches.length}/${keywordList.length}.`;
+				},
+				pass
+			};
+		},
+		toScoreRubricGreaterThan(received, threshold) {
+			const score = typeof received === "number" ? received : received?.score;
+			if (typeof score !== "number") return {
+				message: () => "Expected received value to be a number or RubricJudgeResult.",
+				pass: false
+			};
+			const pass = score > threshold;
+			return {
+				message: () => {
+					if (pass) return `Expected rubric score ${score} to be less than or equal to ${threshold}.`;
+					return `Expected rubric score ${score} to be greater than ${threshold}.`;
+				},
+				pass
+			};
+		},
+		toSatisfyStructuredOutput(received, validator) {
+			const pass = validator(received);
+			return {
+				message: () => pass ? "Expected structured output validator to fail." : "Expected structured output validator to pass.",
+				pass
+			};
+		},
+		toSatisfyToolCallArgs(received, toolName, validator) {
+			const toolCalls = received?.toolCalls;
+			if (toolCalls == null) return {
+				message: () => "Expected received value to provide toolCalls array.",
+				pass: false
+			};
+			const targetCall = toolCalls.find((call) => call.name === toolName);
+			if (targetCall == null) return {
+				message: () => `Expected tool call ${toolName} to exist.`,
+				pass: false
+			};
+			const pass = validator(targetCall.args);
+			return {
+				message: () => pass ? `Expected tool call args for ${toolName} to fail validation.` : `Expected tool call args for ${toolName} to pass validation.`,
+				pass
+			};
+		}
+	});
+}
+//#endregion
+export { getRuntimeExpect as n, installVievalExpectMatchers as t };
+
+//# sourceMappingURL=expect-extensions-BOzwV5EJ.mjs.map

package/dist/expect-extensions-BOzwV5EJ.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"expect-extensions-BOzwV5EJ.mjs","names":[],"sources":["../src/testing/runtime-expect.ts","../src/testing/expect-extensions.ts"],"sourcesContent":["import type { ExpectStatic, MatchersObject, MatcherState, Tester } from '@vitest/expect'\n\nimport {\n  addCustomEqualityTesters,\n  ASYMMETRIC_MATCHERS_OBJECT,\n  chai,\n  ChaiStyleAssertions,\n  customMatchers,\n  getState,\n  GLOBAL_EXPECT,\n  JestAsymmetricMatchers,\n  JestChaiExpect,\n  JestExtend,\n  setState,\n} from '@vitest/expect'\n\nlet isPluginInstalled = false\nlet runtimeExpectInstance: ExpectStatic | undefined\n\n/**\n * Installs Vitest expect plugins once for process-local runtime assertions.\n *\n * Use when:\n * - running eval tasks outside Vitest worker runtime\n * - building an `expect` instance that does not rely on Vitest internal state\n *\n * Expects:\n * - `@vitest/expect` is available in runtime dependencies\n *\n * Returns:\n * - nothing; side-effects are applied to `chai`\n */\nfunction ensureRuntimeExpectPluginsInstalled(): void {\n  if (isPluginInstalled) {\n    return\n  }\n\n  chai.use(JestExtend)\n  chai.use(JestChaiExpect)\n  chai.use(ChaiStyleAssertions)\n  chai.use(JestAsymmetricMatchers)\n  isPluginInstalled = true\n}\n\n/**\n * Creates a Vitest-compatible `expect` instance without worker-state coupling.\n *\n * Use when:\n * - CLI runtime needs assertion helpers from `vieval/expect`\n * - code is executed outside `vitest run`\n *\n * Expects:\n * - plugins from {@link ensureRuntimeExpectPluginsInstalled} are installed\n * - callers do not depend on Vitest worker-only features (snapshot/poll internals)\n *\n * Returns:\n * - standalone expect instance with core matcher APIs and `extend`\n */\nfunction createRuntimeExpect(): ExpectStatic {\n  ensureRuntimeExpectPluginsInstalled()\n\n  const runtimeExpect = ((value: unknown, message?: string) => {\n    const currentState = getState(runtimeExpect)\n    setState({ assertionCalls: currentState.assertionCalls + 1 }, runtimeExpect)\n    return chai.expect(value, message)\n  }) as unknown as ExpectStatic\n\n  Object.assign(runtimeExpect, chai.expect)\n  Object.assign(runtimeExpect, (globalThis as Record<PropertyKey, unknown>)[ASYMMETRIC_MATCHERS_OBJECT] as object)\n\n  runtimeExpect.getState = () => getState(runtimeExpect)\n  runtimeExpect.setState = (state: Partial<MatcherState>) => setState(state, runtimeExpect)\n  runtimeExpect.assert = chai.assert\n  // NOTICE:\n  // Chai's public `ExpectStatic` type does not expose Vitest's plugin-added `extend`.\n  // Runtime `chai.expect.extend` exists after `JestExtend` plugin installation.\n  // Source/context: `@vitest/expect` plugin pipeline in `dist/index.js`.\n  // Removal condition: remove this cast if upstream exposes `extend` on Chai expect types.\n  const chaiExpectWithExtend = chai.expect as unknown as {\n    extend: (expect: ExpectStatic, matchers: MatchersObject) => void\n  }\n  runtimeExpect.extend = (matchers: MatchersObject) => chaiExpectWithExtend.extend(runtimeExpect, matchers)\n  runtimeExpect.addEqualityTesters = (customTesters: Tester[]) => addCustomEqualityTesters(customTesters)\n  runtimeExpect.unreachable = (message?: string) => {\n    chai.assert.fail(`expected${message ? ` \"${message}\" ` : ' '}not to be reached`)\n  }\n\n  runtimeExpect.setState({\n    assertionCalls: 0,\n    currentTestName: '',\n    expectedAssertionsNumber: null,\n    expectedAssertionsNumberErrorGen: null,\n    isExpectingAssertions: false,\n    isExpectingAssertionsError: null,\n  })\n\n  runtimeExpect.extend(customMatchers)\n\n  return runtimeExpect\n}\n\n/**\n * Returns process-local runtime `expect` instance used by Vieval.\n *\n * Use when:\n * - you need matcher assertions in eval files and CLI runtime\n * - importing from `vitest` would crash outside Vitest worker contexts\n *\n * Expects:\n * - single-process usage (instance is memoized per process)\n *\n * Returns:\n * - memoized runtime `expect` instance\n */\nexport function getRuntimeExpect(): ExpectStatic {\n  if (runtimeExpectInstance != null) {\n    return runtimeExpectInstance\n  }\n\n  runtimeExpectInstance = createRuntimeExpect()\n  Object.defineProperty(globalThis, GLOBAL_EXPECT, {\n    configurable: true,\n    value: runtimeExpectInstance,\n    writable: true,\n  })\n\n  return runtimeExpectInstance\n}\n","import type { RubricJudgeResult, ToolCall } from '../core/assertions'\n\nimport { normalizeMatchText } from '../core/assertions'\nimport { getRuntimeExpect } from './runtime-expect'\n\n/**\n * Options for keyword-based matcher behavior.\n */\nexport interface KeywordMatcherOptions {\n  /**\n   * Case-sensitive matching toggle.\n   *\n   * @default false\n   */\n  caseSensitive?: boolean\n  /**\n   * Match mode.\n   *\n   * @default 'all'\n   */\n  mode?: 'all' | 'any'\n}\n\n/**\n * Shape used by tool-call matchers.\n */\nexport interface ToolCallContainer {\n  /**\n   * Tool calls to inspect.\n   */\n  toolCalls?: readonly ToolCall[]\n}\n\nfunction toKeywordArray(keywords: string | readonly string[]): readonly string[] {\n  if (typeof keywords === 'string') {\n    return [keywords]\n  }\n\n  return keywords\n}\n\n/**\n * Registers vieval custom matchers on Vitest `expect`.\n *\n * Call stack:\n *\n * {@link installVievalExpectMatchers}\n *   -> `expect.extend(...)`\n *     -> `expect(received).toMustInclude(...)`\n *     -> `expect(received).toScoreRubricGreaterThan(...)`\n *\n * Use when:\n * - eval suites need domain assertions while preserving native Vitest ergonomics\n * - callers want native `.not` chaining with the same matchers\n */\nexport function installVievalExpectMatchers(): void {\n  const expect = getRuntimeExpect()\n\n  expect.extend({\n    toMustExclude(received: unknown, keywords: string | readonly string[], options: KeywordMatcherOptions = {}) {\n      const keywordList = toKeywordArray(keywords)\n\n      if (typeof received !== 'string') {\n        return {\n          message: () => 'Expected received value to be a string.',\n          pass: false,\n        }\n      }\n\n      const normalizedText = normalizeMatchText(received, options.caseSensitive ?? false)\n      const forbiddenMatches = keywordList.filter((keyword) => {\n        return normalizedText.includes(normalizeMatchText(keyword, options.caseSensitive ?? false))\n      })\n\n      const pass = forbiddenMatches.length === 0\n\n      return {\n        message: () => {\n          if (pass) {\n            return `Expected text to include forbidden keywords: ${keywordList.join(', ')}`\n          }\n\n          return `Expected text not to include forbidden keywords, but matched: ${forbiddenMatches.join(', ')}`\n        },\n        pass,\n      }\n    },\n\n    toMustInclude(received: unknown, keywords: string | readonly string[], options: KeywordMatcherOptions = {}) {\n      const keywordList = toKeywordArray(keywords)\n\n      if (typeof received !== 'string') {\n        return {\n          message: () => 'Expected received value to be a string.',\n          pass: false,\n        }\n      }\n\n      const normalizedText = normalizeMatchText(received, options.caseSensitive ?? false)\n      const matches = keywordList.filter((keyword) => {\n        return normalizedText.includes(normalizeMatchText(keyword, options.caseSensitive ?? false))\n      })\n\n      const mode = options.mode ?? 'all'\n      const pass = mode === 'all' ? matches.length === keywordList.length : matches.length > 0\n\n      return {\n        message: () => {\n          if (pass) {\n            return `Expected text not to match required keywords, but matched: ${matches.join(', ')}`\n          }\n\n          return `Expected text to match required keywords (${mode}), but matched ${matches.length}/${keywordList.length}.`\n        },\n        pass,\n      }\n    },\n\n    toScoreRubricGreaterThan(received: unknown, threshold: number) {\n      const score = typeof received === 'number'\n        ? received\n        : (received as RubricJudgeResult | null)?.score\n\n      if (typeof score !== 'number') {\n        return {\n          message: () => 'Expected received value to be a number or RubricJudgeResult.',\n          pass: false,\n        }\n      }\n\n      const pass = score > threshold\n\n      return {\n        message: () => {\n          if (pass) {\n            return `Expected rubric score ${score} to be less than or equal to ${threshold}.`\n          }\n\n          return `Expected rubric score ${score} to be greater than ${threshold}.`\n        },\n        pass,\n      }\n    },\n\n    toSatisfyStructuredOutput<T>(received: unknown, validator: (value: unknown) => value is T) {\n      const pass = validator(received)\n\n      return {\n        message: () => pass\n          ? 'Expected structured output validator to fail.'\n          : 'Expected structured output validator to pass.',\n        pass,\n      }\n    },\n\n    toSatisfyToolCallArgs(\n      received: unknown,\n      toolName: string,\n      validator: (args: unknown) => boolean,\n    ) {\n      const toolCalls = (received as ToolCallContainer | null)?.toolCalls\n\n      if (toolCalls == null) {\n        return {\n          message: () => 'Expected received value to provide toolCalls array.',\n          pass: false,\n        }\n      }\n\n      const targetCall = toolCalls.find(call => call.name === toolName)\n      if (targetCall == null) {\n        return {\n          message: () => `Expected tool call ${toolName} to exist.`,\n          pass: false,\n        }\n      }\n\n      const pass = validator(targetCall.args)\n\n      return {\n        message: () => pass\n          ? `Expected tool call args for ${toolName} to fail validation.`\n          : `Expected tool call args for ${toolName} to pass validation.`,\n        pass,\n      }\n    },\n  })\n}\n\ninterface VievalCustomMatchers {\n  /**\n   * Asserts that text includes required keywords.\n   *\n   * Example:\n   * `expect('calm answer').toMustInclude(['calm'])`\n   */\n  toMustInclude: (keywords: string | readonly string[], options?: KeywordMatcherOptions) => void\n  /**\n   * Asserts that text excludes forbidden keywords.\n   *\n   * Example:\n   * `expect('calm answer').toMustExclude(['bestmove'])`\n   */\n  toMustExclude: (keywords: string | readonly string[], options?: KeywordMatcherOptions) => void\n  /**\n   * Asserts rubric score is greater than a threshold.\n   *\n   * Example:\n   * `expect({ score: 0.91 }).toScoreRubricGreaterThan(0.8)`\n   */\n  toScoreRubricGreaterThan: (threshold: number) => void\n  /**\n   * Asserts structured output satisfies a validator.\n   *\n   * Example:\n   * `expect(value).toSatisfyStructuredOutput(isMyShape)`\n   */\n  toSatisfyStructuredOutput: <TValue>(validator: (value: unknown) => value is TValue) => void\n  /**\n   * Asserts selected tool-call args satisfy validator.\n   *\n   * Example:\n   * `expect({ toolCalls }).toSatisfyToolCallArgs('builtIn_sparkCommand', isSparkArgs)`\n   */\n  toSatisfyToolCallArgs: (toolName: string, validator: (args: unknown) => boolean) => void\n}\n\n/* eslint-disable unused-imports/no-unused-vars */\ndeclare module '@vitest/expect' {\n  interface Matchers<T = any> extends VievalCustomMatchers {}\n  interface Assertion<T = any> extends VievalCustomMatchers {}\n}\n\ndeclare module 'vitest' {\n  interface Assertion extends VievalCustomMatchers {}\n  interface Matchers<T = any> extends VievalCustomMatchers {}\n}\n/* eslint-enable unused-imports/no-unused-vars */\n"],"mappings":";;;AAgBA,IAAI,oBAAoB;AACxB,IAAI;;;;;;;;;;;;;;AAeJ,SAAS,sCAA4C;AACnD,KAAI,kBACF;AAGF,MAAK,IAAI,WAAW;AACpB,MAAK,IAAI,eAAe;AACxB,MAAK,IAAI,oBAAoB;AAC7B,MAAK,IAAI,uBAAuB;AAChC,qBAAoB;;;;;;;;;;;;;;;;AAiBtB,SAAS,sBAAoC;AAC3C,sCAAqC;CAErC,MAAM,kBAAkB,OAAgB,YAAqB;AAE3D,WAAS,EAAE,gBADU,SAAS,cAAc,CACJ,iBAAiB,GAAG,EAAE,cAAc;AAC5E,SAAO,KAAK,OAAO,OAAO,QAAQ;;AAGpC,QAAO,OAAO,eAAe,KAAK,OAAO;AACzC,QAAO,OAAO,eAAgB,WAA4C,4BAAsC;AAEhH,eAAc,iBAAiB,SAAS,cAAc;AACtD,eAAc,YAAY,UAAiC,SAAS,OAAO,cAAc;AACzF,eAAc,SAAS,KAAK;CAM5B,MAAM,uBAAuB,KAAK;AAGlC,eAAc,UAAU,aAA6B,qBAAqB,OAAO,eAAe,SAAS;AACzG,eAAc,sBAAsB,kBAA4B,yBAAyB,cAAc;AACvG,eAAc,eAAe,YAAqB;AAChD,OAAK,OAAO,KAAK,WAAW,UAAU,KAAK,QAAQ,MAAM,IAAI,mBAAmB;;AAGlF,eAAc,SAAS;EACrB,gBAAgB;EAChB,iBAAiB;EACjB,0BAA0B;EAC1B,kCAAkC;EAClC,uBAAuB;EACvB,4BAA4B;EAC7B,CAAC;AAEF,eAAc,OAAO,eAAe;AAEpC,QAAO;;;;;;;;;;;;;;;AAgBT,SAAgB,mBAAiC;AAC/C,KAAI,yBAAyB,KAC3B,QAAO;AAGT,yBAAwB,qBAAqB;AAC7C,QAAO,eAAe,YAAY,eAAe;EAC/C,cAAc;EACd,OAAO;EACP,UAAU;EACX,CAAC;AAEF,QAAO;;;;AC7FT,SAAS,eAAe,UAAyD;AAC/E,KAAI,OAAO,aAAa,SACtB,QAAO,CAAC,SAAS;AAGnB,QAAO;;;;;;;;;;;;;;;;AAiBT,SAAgB,8BAAoC;AACnC,mBAAkB,CAE1B,OAAO;EACZ,cAAc,UAAmB,UAAsC,UAAiC,EAAE,EAAE;GAC1G,MAAM,cAAc,eAAe,SAAS;AAE5C,OAAI,OAAO,aAAa,SACtB,QAAO;IACL,eAAe;IACf,MAAM;IACP;GAGH,MAAM,iBAAiB,mBAAmB,UAAU,QAAQ,iBAAiB,MAAM;GACnF,MAAM,mBAAmB,YAAY,QAAQ,YAAY;AACvD,WAAO,eAAe,SAAS,mBAAmB,SAAS,QAAQ,iBAAiB,MAAM,CAAC;KAC3F;GAEF,MAAM,OAAO,iBAAiB,WAAW;AAEzC,UAAO;IACL,eAAe;AACb,SAAI,KACF,QAAO,gDAAgD,YAAY,KAAK,KAAK;AAG/E,YAAO,iEAAiE,iBAAiB,KAAK,KAAK;;IAErG;IACD;;EAGH,cAAc,UAAmB,UAAsC,UAAiC,EAAE,EAAE;GAC1G,MAAM,cAAc,eAAe,SAAS;AAE5C,OAAI,OAAO,aAAa,SACtB,QAAO;IACL,eAAe;IACf,MAAM;IACP;GAGH,MAAM,iBAAiB,mBAAmB,UAAU,QAAQ,iBAAiB,MAAM;GACnF,MAAM,UAAU,YAAY,QAAQ,YAAY;AAC9C,WAAO,eAAe,SAAS,mBAAmB,SAAS,QAAQ,iBAAiB,MAAM,CAAC;KAC3F;GAEF,MAAM,OAAO,QAAQ,QAAQ;GAC7B,MAAM,OAAO,SAAS,QAAQ,QAAQ,WAAW,YAAY,SAAS,QAAQ,SAAS;AAEvF,UAAO;IACL,eAAe;AACb,SAAI,KACF,QAAO,8DAA8D,QAAQ,KAAK,KAAK;AAGzF,YAAO,6CAA6C,KAAK,iBAAiB,QAAQ,OAAO,GAAG,YAAY,OAAO;;IAEjH;IACD;;EAGH,yBAAyB,UAAmB,WAAmB;GAC7D,MAAM,QAAQ,OAAO,aAAa,WAC9B,WACC,UAAuC;AAE5C,OAAI,OAAO,UAAU,SACnB,QAAO;IACL,eAAe;IACf,MAAM;IACP;GAGH,MAAM,OAAO,QAAQ;AAErB,UAAO;IACL,eAAe;AACb,SAAI,KACF,QAAO,yBAAyB,MAAM,+BAA+B,UAAU;AAGjF,YAAO,yBAAyB,MAAM,sBAAsB,UAAU;;IAExE;IACD;;EAGH,0BAA6B,UAAmB,WAA2C;GACzF,MAAM,OAAO,UAAU,SAAS;AAEhC,UAAO;IACL,eAAe,OACX,kDACA;IACJ;IACD;;EAGH,sBACE,UACA,UACA,WACA;GACA,MAAM,YAAa,UAAuC;AAE1D,OAAI,aAAa,KACf,QAAO;IACL,eAAe;IACf,MAAM;IACP;GAGH,MAAM,aAAa,UAAU,MAAK,SAAQ,KAAK,SAAS,SAAS;AACjE,OAAI,cAAc,KAChB,QAAO;IACL,eAAe,sBAAsB,SAAS;IAC9C,MAAM;IACP;GAGH,MAAM,OAAO,UAAU,WAAW,KAAK;AAEvC,UAAO;IACL,eAAe,OACX,+BAA+B,SAAS,wBACxC,+BAA+B,SAAS;IAC5C;IACD;;EAEJ,CAAC"}

package/dist/expect.d.mts

@@ -1,2 +1,2 @@
-import { t as expect } from "./expect-0jPJ7Zio.mjs";
+import { t as expect } from "./expect-B2vaoRVZ.mjs";
 export { expect };

package/dist/expect.mjs

@@ -1,2 +1,2 @@
-import { t as expect } from "./expect-i9WZWGrA.mjs";
+import { t as expect } from "./expect-CaXiUkwY.mjs";
 export { expect };