@ls-stack/agent-eval 0.22.0 → 0.23.0

package/dist/{app-moDHbg1O.mjs → app-Cw79dJDr.mjs}

@@ -1,6 +1,6 @@
-import { E as updateManualScoreRequestSchema, T as createRunRequestSchema } from "./runOrchestration-D1edUDhp.mjs";
-import "./src-D-HuV8I-.mjs";
-import { t as getRunnerInstance } from "./runner-C9nP2VKL.mjs";
+import { E as updateManualScoreRequestSchema, T as createRunRequestSchema } from "./runOrchestration-CimthgI7.mjs";
+import "./src-CcXfWT4M.mjs";
+import { t as getRunnerInstance } from "./runner-4yNYRvmF.mjs";
 import { readFile } from "node:fs/promises";
 import { dirname, isAbsolute, join, relative, resolve, sep } from "node:path";
 import { z } from "zod/v4";

package/dist/bin.mjs

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { t as runCli } from "./cli-C0EtHhEO.mjs";
+import { t as runCli } from "./cli-D3QNOcPN.mjs";
 import { spawn } from "node:child_process";
 //#region src/bin.ts
 const moduleMocksFlag = "--experimental-test-module-mocks";

package/dist/{cli-C0EtHhEO.mjs → cli-D3QNOcPN.mjs}

@@ -1,4 +1,4 @@
-import { C as validateCharts, F as deriveScopedSummaryFromCases, Fn as getEvalRegistry, N as getEvalTitle, P as getEvalDisplayStatus, S as normalizeScoreDef, _ as loadEvalModule, a as loadPersistedRunSnapshot, b as loadConfig, c as persistCaseDetail, d as recomputePersistedCaseStatus, et as resolveApiCallsConfig, f as runTouchesEval, g as setLatestRunInfoMap, h as getTargetEvalIds, i as getLatestRunInfos, l as persistRunState, m as buildEvalSummary, n as generateRunId, o as loadPersistedRunSnapshots, p as resolveArtifactPath, r as getLastRunStatuses, s as nextShortIdFromSnapshots, tt as resolveLlmCallsConfig, u as recomputeEvalStatusesInRuns, v as parseEvalMetas, w as createFsCacheStore, x as buildDeclaredColumnDefs, y as resolveEvalDefaultConfig, z as runSummarySchema } from "./runOrchestration-D1edUDhp.mjs";
+import { C as validateCharts, F as deriveScopedSummaryFromCases, Ln as getEvalRegistry, N as getEvalTitle, P as getEvalDisplayStatus, S as normalizeScoreDef, _ as loadEvalModule, a as loadPersistedRunSnapshot, b as loadConfig, c as persistCaseDetail, d as recomputePersistedCaseStatus, et as resolveApiCallsConfig, f as runTouchesEval, g as setLatestRunInfoMap, h as getTargetEvalIds, i as getLatestRunInfos, l as persistRunState, m as buildEvalSummary, n as generateRunId, o as loadPersistedRunSnapshots, p as resolveArtifactPath, r as getLastRunStatuses, s as nextShortIdFromSnapshots, tt as resolveLlmCallsConfig, u as recomputeEvalStatusesInRuns, v as parseEvalMetas, w as createFsCacheStore, x as buildDeclaredColumnDefs, y as resolveEvalDefaultConfig, z as runSummarySchema } from "./runOrchestration-CimthgI7.mjs";
 import { createHash } from "node:crypto";
 import { mkdir, readFile, rm, writeFile } from "node:fs/promises";
 import { dirname, join, relative, resolve } from "node:path";
@@ -959,8 +959,8 @@ async function commandApp(args) {
 	const { serve } = await import("@hono/node-server");
 	const bundledWebDist = resolve(currentDir, "apps/web/dist");
 	if (existsSync(bundledWebDist)) process.env.AGENT_EVALS_WEB_DIST = bundledWebDist;
-	const appModule = await import("./app-moDHbg1O.mjs");
-	const runnerModule = await import("./runner-CyRhIzci.mjs");
+	const appModule = await import("./app-Cw79dJDr.mjs");
+	const runnerModule = await import("./runner-B-SYzW8w.mjs");
 	if (!isHonoAppModule(appModule)) throw new Error("Server app module is invalid");
 	if (!isServerRunnerModule(runnerModule)) throw new Error("Server runner module is invalid");
 	await runnerModule.initRunner();

package/dist/index.d.mts

@@ -2758,6 +2758,12 @@ type EvalCase<TInput> = {
 };
 /** Runtime output values collected from output helpers and `deriveFromTracing`. */
 type EvalOutputs = Record<string, unknown>;
+/**
+ * Initial wall-clock time used by an eval's shifted Date clock.
+ *
+ * Pass `'now'` to opt one eval back into the real current clock.
+ */
+type EvalStartTime = Date | number | string;
 /**
  * Schema used to validate and type an eval's collected runtime outputs.
  *
@@ -2925,6 +2931,23 @@ type EvalDefinitionBase<TInput = unknown, TOutputs extends EvalOutputs = EvalOut
    * delay case finalization; late mutations are not guaranteed to persist.
    */
   waitForBackgroundJobs?: boolean;
+  /**
+   * Optional initial wall-clock time for this eval's runtime.
+   *
+   * When set, `new Date()` and `Date.now()` inside case generation, execution,
+   * tracing, derived outputs, and scorers start from this wall-clock value and
+   * then continue advancing with real elapsed time. The default is
+   * `2026-04-10T00:00:00.000Z`. Pass `'now'` to use the real current clock for
+   * this eval. Timers are not faked, so `setTimeout` and other asynchronous
+   * work still run normally.
+   */
+  startTime?: EvalStartTime;
+  /**
+   * Freeze the eval Date clock at `startTime` until `advanceEvalTime(...)`
+   * moves it manually. Defaults to `false`, so eval time advances with real
+   * elapsed time from the configured `startTime`.
+   */
+  freezeTime?: boolean;
   execute: (ctx: EvalExecuteContext<TInput, TOutputs>) => Promise<void> | void;
   deriveFromTracing?: (ctx: EvalDeriveContext<TInput>) => Partial<TOutputs> | Promise<Partial<TOutputs>>;
   scores?: Record<string, EvalScoreDef<TInput, TOutputs>>;
@@ -3018,6 +3041,9 @@ declare function defineEval<TInput = unknown, TOutputs extends EvalOutputs = Eva
 declare function repoFile(path: string, mimeType?: string): RepoFileRef;
 //#endregion
 //#region ../sdk/src/runtime.d.ts
+declare global {
+  var __agentEvalsRealDate: DateConstructor | undefined;
+}
 /**
  * Raw-key debug payload passed alongside cache writes.
  *
@@ -3063,7 +3089,15 @@ type CacheRecordingFrame = {
 };
 /** Mutable per-case runtime state stored in async local storage. */
 type EvalCaseScope = {
-  caseId: string; /** Stable prefix used by `nextEvalId()` for this eval case scope. */
+  caseId: string; /** Initial wall-clock time used by Date APIs inside this eval case. */
+  startTime: EvalStartTime | undefined; /** Mutable shifted wall-clock state shared across this eval case's phases. */
+  evalClockState: {
+    startMs: number;
+    realStartMs: number;
+    offsetMs: number;
+    frozen: boolean;
+    shifted: boolean;
+  }; /** Stable prefix used by `nextEvalId()` for this eval case scope. */
   idPrefix: string | undefined; /** Monotonic per-scope counter used by `nextEvalId()`. */
   nextEvalIdCounter: number; /** Authored input for the current case, when provided by the runner. */
   input?: unknown;
@@ -3101,11 +3135,28 @@ type EvalCaseScope = {
  * modules imported while a run is being prepared.
  */
 type EvalRuntimeScope = 'env' | 'cases' | 'eval' | 'derive' | 'outputsSchema' | 'scorer';
+/** Time unit accepted by `advanceEvalTime(unit, amount)`. */
+type EvalTimeUnit = 'millisecond' | 'milliseconds' | 'second' | 'seconds' | 'minute' | 'minutes' | 'hour' | 'hours' | 'day' | 'days';
 type EvalLogLevelInput = RunLogLevel | 'warning';
 /** Error thrown when an eval assertion fails during case execution. */
 declare class EvalAssertionError extends Error {
   constructor(message: string);
 }
+/**
+ * Return the wall-clock start time captured for the active eval.
+ *
+ * For `startTime: 'now'`, this is the real time captured when the eval clock
+ * context was created.
+ */
+declare function getEvalStartTime(): Date;
+/**
+ * Advance the active eval's shifted Date clock and return the new time.
+ *
+ * Throws outside an active shifted eval clock. Evals that set
+ * `startTime: 'now'` use the real current clock unless `freezeTime: true` is
+ * also set.
+ */
+declare function advanceEvalTime(unit: EvalTimeUnit, amount: number): Date;
 /** Return the current eval scope for the active async context, if any. */
 declare function getCurrentScope(): EvalCaseScope | undefined;
 /**
@@ -3155,7 +3206,9 @@ type RunInEvalScopeOptions = {
   idPrefix?: string; /** Cache adapter + mode attached to the scope before `fn` runs. */
   cacheContext?: CacheScopeContext; /** Whether registered background jobs should settle before scope finalizes. */
   waitForBackgroundJobs?: boolean; /** Eval runner phase exposed through `isInEvalScope()`. Defaults to `eval`. */
-  runtimeScope?: EvalRuntimeScope;
+  runtimeScope?: EvalRuntimeScope; /** Initial wall-clock time used by `new Date()` and `Date.now()` in this eval. */
+  startTime?: EvalStartTime; /** Whether Date APIs stay frozen until advanced manually. */
+  freezeTime?: boolean;
 };
 /** Execute a callback while `isInEvalScope()` reports a runner phase. */
 declare function runInEvalRuntimeScope<T>(runtimeScope: EvalRuntimeScope, fn: () => Promise<T> | T): Promise<T>;
@@ -3570,4 +3623,4 @@ declare function createRunner({
  */
 declare function runCli(argv: string[]): Promise<void>;
 //#endregion
-export { type AgentEvalsConfig, type ApiCallEntry, type ApiCallMetric, type ApiCallMetricFormat, type ApiCallMetricPlacement, type ApiCallMetricValue, type ApiCallsConfigInput, type AssertionFailure, type CacheActivityEntry, type CacheAdapter, type CacheDebugKeyEntry, type CacheDebugKeyFile, type CacheDebugKeyWrite, type CacheEntry, type CacheEntryWithDebugKey, type CacheFile, type CacheHitEntry, type CacheKeyHashInput, type CacheKeyHashOptions, type CacheListItem, type CacheMode, type CacheOperationType, type CacheRecording, type CacheRecordingFrame, type CacheRecordingOp, type CacheScopeContext, type CacheStatus, type CaptureEvalSpanErrorLevel, type CaptureEvalSpanErrorOptions, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type CreateRunRequest, DEFAULT_API_CALLS_CONFIG, DEFAULT_LLM_CALLS_CONFIG, type DefaultConfigKey, type DerivedStatus, EvalAssertionError, type EvalCase, type EvalCaseScope, type EvalChartAggregate, type EvalChartAxis, type EvalChartBuiltinMetric, type EvalChartColor, type EvalChartConfig, type EvalChartMetric, type EvalChartTooltipExtra, type EvalChartType, type EvalChartsConfig, type EvalColumnOverride, type EvalColumns, type EvalDefinition, type EvalDeriveContext, type EvalDisplayStatus, type EvalExecuteContext, type EvalFreshnessStatus, type EvalManualScoreDef, type EvalOutputs, type EvalOutputsSchema, type EvalRunner, type EvalRuntimeScope, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalSetOutput, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, type EvalTraceSpan, type EvalTraceSpanError, type EvalTraceSpanWarning, type EvalTraceTree, type FileRef, type JsonCell, type LlmCallEntry, type LlmCallMetric, type LlmCallMetricFormat, type LlmCallMetricPlacement, type LlmCallMetricValue, type LlmCallPricing, type LlmCallsConfigInput, type NumberDisplayOptions, type RemoveDefaultConfig, type RepoFileRef, type ResolvedApiCallMetric, type ResolvedApiCallsConfig, type ResolvedLlmCallMetric, type ResolvedLlmCallPricing, type ResolvedLlmCallsConfig, type RunArtifactRef, type RunInEvalScopeOptions, type RunLogEntry, type RunLogLevel, type RunLogLocation, type RunLogPhase, type RunLogsConfigInput, type RunManifest, type RunSummary, type ScalarCell, type ScopedCaseSummary, type ScoreTrace, type SerializedCacheSpan, type SerializedCacheValue, type SpanCacheOptions, type SseEnvelope, type SseEventType, type TraceActiveSpan, type TraceAttributeDisplay, type TraceAttributeDisplayFormat, type TraceAttributeDisplayInput, type TraceAttributeDisplayPlacement, type TraceAttributeTransform, type TraceAttributeTransformContext, type TraceCacheInfo, type TraceCacheRef, type TraceDisplayConfig, type TraceDisplayInputConfig, type TraceSpanInfo, type TrialSelectionMode, type UpdateManualScoreRequest, agentEvalsConfigSchema, apiCallMetricFormatSchema, apiCallMetricPlacementSchema, apiCallMetricSchema, apiCallsConfigSchema, appendToEvalOutput, assertionFailureSchema, buildTraceTree, cacheDebugKeyEntrySchema, cacheDebugKeyFileSchema, cacheEntrySchema, cacheEntryWithDebugKeySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheOperationTypeSchema, cacheRecordingOpSchema, cacheRecordingSchema, cacheStatusSchema, captureEvalSpanError, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defaultConfigKeySchema, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, deserializeCacheRecording, deserializeCacheValue, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalLog, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, fileRefSchema, getCurrentScope, getEvalCaseInput, getEvalDisplayStatus, getEvalRegistry, getEvalTitle, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, llmCallMetricFormatSchema, llmCallMetricPlacementSchema, llmCallMetricSchema, llmCallPricingSchema, llmCallsConfigSchema, mergeEvalOutput, nextEvalId, numberDisplayOptionsSchema, removeDefaultConfigSchema, repoFile, repoFileRefSchema, resolveApiCallsConfig, resolveLlmCallsConfig, runArtifactRefSchema, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, runLogEntrySchema, runLogLevelSchema, runLogLocationSchema, runLogPhaseSchema, runLogsConfigSchema, runManifestSchema, runSummarySchema, scoreTraceSchema, serializeCacheRecording, serializeCacheValue, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, startEvalBackgroundJob, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceCacheRefSchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanErrorSchema, traceSpanKindSchema, traceSpanSchema, traceSpanWarningSchema, trialSelectionModeSchema, updateManualScoreRequestSchema, z };
+export { type AgentEvalsConfig, type ApiCallEntry, type ApiCallMetric, type ApiCallMetricFormat, type ApiCallMetricPlacement, type ApiCallMetricValue, type ApiCallsConfigInput, type AssertionFailure, type CacheActivityEntry, type CacheAdapter, type CacheDebugKeyEntry, type CacheDebugKeyFile, type CacheDebugKeyWrite, type CacheEntry, type CacheEntryWithDebugKey, type CacheFile, type CacheHitEntry, type CacheKeyHashInput, type CacheKeyHashOptions, type CacheListItem, type CacheMode, type CacheOperationType, type CacheRecording, type CacheRecordingFrame, type CacheRecordingOp, type CacheScopeContext, type CacheStatus, type CaptureEvalSpanErrorLevel, type CaptureEvalSpanErrorOptions, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type CreateRunRequest, DEFAULT_API_CALLS_CONFIG, DEFAULT_LLM_CALLS_CONFIG, type DefaultConfigKey, type DerivedStatus, EvalAssertionError, type EvalCase, type EvalCaseScope, type EvalChartAggregate, type EvalChartAxis, type EvalChartBuiltinMetric, type EvalChartColor, type EvalChartConfig, type EvalChartMetric, type EvalChartTooltipExtra, type EvalChartType, type EvalChartsConfig, type EvalColumnOverride, type EvalColumns, type EvalDefinition, type EvalDeriveContext, type EvalDisplayStatus, type EvalExecuteContext, type EvalFreshnessStatus, type EvalManualScoreDef, type EvalOutputs, type EvalOutputsSchema, type EvalRunner, type EvalRuntimeScope, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalSetOutput, type EvalStartTime, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, type EvalTimeUnit, type EvalTraceSpan, type EvalTraceSpanError, type EvalTraceSpanWarning, type EvalTraceTree, type FileRef, type JsonCell, type LlmCallEntry, type LlmCallMetric, type LlmCallMetricFormat, type LlmCallMetricPlacement, type LlmCallMetricValue, type LlmCallPricing, type LlmCallsConfigInput, type NumberDisplayOptions, type RemoveDefaultConfig, type RepoFileRef, type ResolvedApiCallMetric, type ResolvedApiCallsConfig, type ResolvedLlmCallMetric, type ResolvedLlmCallPricing, type ResolvedLlmCallsConfig, type RunArtifactRef, type RunInEvalScopeOptions, type RunLogEntry, type RunLogLevel, type RunLogLocation, type RunLogPhase, type RunLogsConfigInput, type RunManifest, type RunSummary, type ScalarCell, type ScopedCaseSummary, type ScoreTrace, type SerializedCacheSpan, type SerializedCacheValue, type SpanCacheOptions, type SseEnvelope, type SseEventType, type TraceActiveSpan, type TraceAttributeDisplay, type TraceAttributeDisplayFormat, type TraceAttributeDisplayInput, type TraceAttributeDisplayPlacement, type TraceAttributeTransform, type TraceAttributeTransformContext, type TraceCacheInfo, type TraceCacheRef, type TraceDisplayConfig, type TraceDisplayInputConfig, type TraceSpanInfo, type TrialSelectionMode, type UpdateManualScoreRequest, advanceEvalTime, agentEvalsConfigSchema, apiCallMetricFormatSchema, apiCallMetricPlacementSchema, apiCallMetricSchema, apiCallsConfigSchema, appendToEvalOutput, assertionFailureSchema, buildTraceTree, cacheDebugKeyEntrySchema, cacheDebugKeyFileSchema, cacheEntrySchema, cacheEntryWithDebugKeySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheOperationTypeSchema, cacheRecordingOpSchema, cacheRecordingSchema, cacheStatusSchema, captureEvalSpanError, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defaultConfigKeySchema, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, deserializeCacheRecording, deserializeCacheValue, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalLog, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, fileRefSchema, getCurrentScope, getEvalCaseInput, getEvalDisplayStatus, getEvalRegistry, getEvalStartTime, getEvalTitle, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, llmCallMetricFormatSchema, llmCallMetricPlacementSchema, llmCallMetricSchema, llmCallPricingSchema, llmCallsConfigSchema, mergeEvalOutput, nextEvalId, numberDisplayOptionsSchema, removeDefaultConfigSchema, repoFile, repoFileRefSchema, resolveApiCallsConfig, resolveLlmCallsConfig, runArtifactRefSchema, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, runLogEntrySchema, runLogLevelSchema, runLogLocationSchema, runLogPhaseSchema, runLogsConfigSchema, runManifestSchema, runSummarySchema, scoreTraceSchema, serializeCacheRecording, serializeCacheValue, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, startEvalBackgroundJob, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceCacheRefSchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanErrorSchema, traceSpanKindSchema, traceSpanSchema, traceSpanWarningSchema, trialSelectionModeSchema, updateManualScoreRequestSchema, z };

package/dist/index.mjs

@@ -1,4 +1,4 @@
-import { $ as removeDefaultConfigSchema, $t as columnKindSchema, A as extractApiCalls, An as setEvalOutput, At as cacheFileSchema, B as DEFAULT_API_CALLS_CONFIG, Bt as traceAttributeDisplayFormatSchema, Cn as incrementEvalOutput, Ct as evalChartTooltipExtraSchema, D as sseEnvelopeSchema, Dn as runInEvalRuntimeScope, Dt as cacheDebugKeyFileSchema, E as updateManualScoreRequestSchema, En as nextEvalId, Et as cacheDebugKeyEntrySchema, F as deriveScopedSummaryFromCases, Fn as getEvalRegistry, Ft as cacheRecordingSchema, G as apiCallMetricSchema, Gt as traceDisplayInputConfigSchema, H as agentEvalsConfigSchema, Ht as traceAttributeDisplayPlacementSchema, I as deriveStatusFromCaseRows, It as cacheStatusSchema, J as llmCallMetricFormatSchema, Jt as traceSpanSchema, K as apiCallsConfigSchema, Kt as traceSpanErrorSchema, L as deriveStatusFromChildStatuses, Lt as serializedCacheSpanSchema, M as getNestedAttribute, Mn as startEvalBackgroundJob, Mt as cacheModeSchema, N as getEvalTitle, Nn as repoFile, Nt as cacheOperationTypeSchema, O as extractCacheEntries, On as runInEvalScope, Ot as cacheEntrySchema, P as getEvalDisplayStatus, Pn as defineEval, Pt as cacheRecordingOpSchema, Q as llmCallsConfigSchema, Qt as columnFormatSchema, R as runManifestSchema, Rt as spanCacheOptionsSchema, Sn as getEvalCaseInput, St as evalChartMetricSchema, T as createRunRequestSchema, Tn as mergeEvalOutput, Tt as evalChartsConfigSchema, U as apiCallMetricFormatSchema, Ut as traceAttributeDisplaySchema, V as DEFAULT_LLM_CALLS_CONFIG, Vt as traceAttributeDisplayInputSchema, W as apiCallMetricPlacementSchema, Wt as traceDisplayConfigSchema, X as llmCallMetricSchema, Xt as cellValueSchema, Y as llmCallMetricPlacementSchema, Yt as traceSpanWarningSchema, Z as llmCallPricingSchema, Zt as columnDefSchema, _n as appendToEvalOutput, _t as evalChartAggregateSchema, an as z, at as caseDetailSchema, bn as evalLog, bt as evalChartColorSchema, cn as evalSpan, ct as evalStatAggregateSchema, dn as hashCacheKeySync, dt as evalSummarySchema, en as fileRefSchema, et as resolveApiCallsConfig, fn as deserializeCacheRecording, ft as runLogEntrySchema, gn as EvalAssertionError, gt as scoreTraceSchema, hn as serializeCacheValue, ht as runLogPhaseSchema, in as runArtifactRefSchema, it as assertionFailureSchema, j as extractLlmCalls, jn as setScopeCacheContext, jt as cacheListItemSchema, k as extractCacheHits, kn as runInExistingEvalScope, kt as cacheEntryWithDebugKeySchema, ln as evalTracer, lt as evalStatItemSchema, mn as serializeCacheRecording, mt as runLogLocationSchema, nn as numberDisplayOptionsSchema, nt as runLogsConfigSchema, on as buildTraceTree, ot as caseRowSchema, pn as deserializeCacheValue, pt as runLogLevelSchema, q as defaultConfigKeySchema, qt as traceSpanKindSchema, rn as repoFileRefSchema, rt as trialSelectionModeSchema, sn as captureEvalSpanError, st as evalFreshnessStatusSchema, tn as jsonCellSchema, tt as resolveLlmCallsConfig, un as hashCacheKey, ut as evalStatsConfigSchema, vt as evalChartAxisSchema, wn as isInEvalScope, wt as evalChartTypeSchema, xn as getCurrentScope, xt as evalChartConfigSchema, yn as evalAssert, yt as evalChartBuiltinMetricSchema, z as runSummarySchema, zt as traceCacheRefSchema } from "./runOrchestration-D1edUDhp.mjs";
-import { n as createRunner, t as runCli } from "./cli-C0EtHhEO.mjs";
-import "./src-D-HuV8I-.mjs";
-export { DEFAULT_API_CALLS_CONFIG, DEFAULT_LLM_CALLS_CONFIG, EvalAssertionError, agentEvalsConfigSchema, apiCallMetricFormatSchema, apiCallMetricPlacementSchema, apiCallMetricSchema, apiCallsConfigSchema, appendToEvalOutput, assertionFailureSchema, buildTraceTree, cacheDebugKeyEntrySchema, cacheDebugKeyFileSchema, cacheEntrySchema, cacheEntryWithDebugKeySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheOperationTypeSchema, cacheRecordingOpSchema, cacheRecordingSchema, cacheStatusSchema, captureEvalSpanError, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defaultConfigKeySchema, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, deserializeCacheRecording, deserializeCacheValue, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalLog, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, fileRefSchema, getCurrentScope, getEvalCaseInput, getEvalDisplayStatus, getEvalRegistry, getEvalTitle, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, llmCallMetricFormatSchema, llmCallMetricPlacementSchema, llmCallMetricSchema, llmCallPricingSchema, llmCallsConfigSchema, mergeEvalOutput, nextEvalId, numberDisplayOptionsSchema, removeDefaultConfigSchema, repoFile, repoFileRefSchema, resolveApiCallsConfig, resolveLlmCallsConfig, runArtifactRefSchema, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, runLogEntrySchema, runLogLevelSchema, runLogLocationSchema, runLogPhaseSchema, runLogsConfigSchema, runManifestSchema, runSummarySchema, scoreTraceSchema, serializeCacheRecording, serializeCacheValue, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, startEvalBackgroundJob, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceCacheRefSchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanErrorSchema, traceSpanKindSchema, traceSpanSchema, traceSpanWarningSchema, trialSelectionModeSchema, updateManualScoreRequestSchema, z };
+import { $ as removeDefaultConfigSchema, $t as columnKindSchema, A as extractApiCalls, An as runInEvalScope, At as cacheFileSchema, B as DEFAULT_API_CALLS_CONFIG, Bt as traceAttributeDisplayFormatSchema, Cn as getEvalCaseInput, Ct as evalChartTooltipExtraSchema, D as sseEnvelopeSchema, Dn as mergeEvalOutput, Dt as cacheDebugKeyFileSchema, E as updateManualScoreRequestSchema, En as isInEvalScope, Et as cacheDebugKeyEntrySchema, F as deriveScopedSummaryFromCases, Fn as repoFile, Ft as cacheRecordingSchema, G as apiCallMetricSchema, Gt as traceDisplayInputConfigSchema, H as agentEvalsConfigSchema, Ht as traceAttributeDisplayPlacementSchema, I as deriveStatusFromCaseRows, In as defineEval, It as cacheStatusSchema, J as llmCallMetricFormatSchema, Jt as traceSpanSchema, K as apiCallsConfigSchema, Kt as traceSpanErrorSchema, L as deriveStatusFromChildStatuses, Ln as getEvalRegistry, Lt as serializedCacheSpanSchema, M as getNestedAttribute, Mn as setEvalOutput, Mt as cacheModeSchema, N as getEvalTitle, Nn as setScopeCacheContext, Nt as cacheOperationTypeSchema, O as extractCacheEntries, On as nextEvalId, Ot as cacheEntrySchema, P as getEvalDisplayStatus, Pn as startEvalBackgroundJob, Pt as cacheRecordingOpSchema, Q as llmCallsConfigSchema, Qt as columnFormatSchema, R as runManifestSchema, Rt as spanCacheOptionsSchema, Sn as getCurrentScope, St as evalChartMetricSchema, T as createRunRequestSchema, Tn as incrementEvalOutput, Tt as evalChartsConfigSchema, U as apiCallMetricFormatSchema, Ut as traceAttributeDisplaySchema, V as DEFAULT_LLM_CALLS_CONFIG, Vt as traceAttributeDisplayInputSchema, W as apiCallMetricPlacementSchema, Wt as traceDisplayConfigSchema, X as llmCallMetricSchema, Xt as cellValueSchema, Y as llmCallMetricPlacementSchema, Yt as traceSpanWarningSchema, Z as llmCallPricingSchema, Zt as columnDefSchema, _n as advanceEvalTime, _t as evalChartAggregateSchema, an as z, at as caseDetailSchema, bn as evalAssert, bt as evalChartColorSchema, cn as evalSpan, ct as evalStatAggregateSchema, dn as hashCacheKeySync, dt as evalSummarySchema, en as fileRefSchema, et as resolveApiCallsConfig, fn as deserializeCacheRecording, ft as runLogEntrySchema, gn as EvalAssertionError, gt as scoreTraceSchema, hn as serializeCacheValue, ht as runLogPhaseSchema, in as runArtifactRefSchema, it as assertionFailureSchema, j as extractLlmCalls, jn as runInExistingEvalScope, jt as cacheListItemSchema, k as extractCacheHits, kn as runInEvalRuntimeScope, kt as cacheEntryWithDebugKeySchema, ln as evalTracer, lt as evalStatItemSchema, mn as serializeCacheRecording, mt as runLogLocationSchema, nn as numberDisplayOptionsSchema, nt as runLogsConfigSchema, on as buildTraceTree, ot as caseRowSchema, pn as deserializeCacheValue, pt as runLogLevelSchema, q as defaultConfigKeySchema, qt as traceSpanKindSchema, rn as repoFileRefSchema, rt as trialSelectionModeSchema, sn as captureEvalSpanError, st as evalFreshnessStatusSchema, tn as jsonCellSchema, tt as resolveLlmCallsConfig, un as hashCacheKey, ut as evalStatsConfigSchema, vn as appendToEvalOutput, vt as evalChartAxisSchema, wn as getEvalStartTime, wt as evalChartTypeSchema, xn as evalLog, xt as evalChartConfigSchema, yt as evalChartBuiltinMetricSchema, z as runSummarySchema, zt as traceCacheRefSchema } from "./runOrchestration-CimthgI7.mjs";
+import { n as createRunner, t as runCli } from "./cli-D3QNOcPN.mjs";
+import "./src-CcXfWT4M.mjs";
+export { DEFAULT_API_CALLS_CONFIG, DEFAULT_LLM_CALLS_CONFIG, EvalAssertionError, advanceEvalTime, agentEvalsConfigSchema, apiCallMetricFormatSchema, apiCallMetricPlacementSchema, apiCallMetricSchema, apiCallsConfigSchema, appendToEvalOutput, assertionFailureSchema, buildTraceTree, cacheDebugKeyEntrySchema, cacheDebugKeyFileSchema, cacheEntrySchema, cacheEntryWithDebugKeySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheOperationTypeSchema, cacheRecordingOpSchema, cacheRecordingSchema, cacheStatusSchema, captureEvalSpanError, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defaultConfigKeySchema, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, deserializeCacheRecording, deserializeCacheValue, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalLog, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, fileRefSchema, getCurrentScope, getEvalCaseInput, getEvalDisplayStatus, getEvalRegistry, getEvalStartTime, getEvalTitle, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, llmCallMetricFormatSchema, llmCallMetricPlacementSchema, llmCallMetricSchema, llmCallPricingSchema, llmCallsConfigSchema, mergeEvalOutput, nextEvalId, numberDisplayOptionsSchema, removeDefaultConfigSchema, repoFile, repoFileRefSchema, resolveApiCallsConfig, resolveLlmCallsConfig, runArtifactRefSchema, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, runLogEntrySchema, runLogLevelSchema, runLogLocationSchema, runLogPhaseSchema, runLogsConfigSchema, runManifestSchema, runSummarySchema, scoreTraceSchema, serializeCacheRecording, serializeCacheValue, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, startEvalBackgroundJob, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceCacheRefSchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanErrorSchema, traceSpanKindSchema, traceSpanSchema, traceSpanWarningSchema, trialSelectionModeSchema, updateManualScoreRequestSchema, z };

package/dist/runChild.mjs

@@ -1,4 +1,4 @@
-import { R as runManifestSchema, T as createRunRequestSchema, Tt as evalChartsConfigSchema, Zt as columnDefSchema, b as loadConfig, t as executeRun, ut as evalStatsConfigSchema, v as parseEvalMetas, vn as configureEvalRunLogs, w as createFsCacheStore, z as runSummarySchema } from "./runOrchestration-D1edUDhp.mjs";
+import { R as runManifestSchema, T as createRunRequestSchema, Tt as evalChartsConfigSchema, Zt as columnDefSchema, b as loadConfig, t as executeRun, ut as evalStatsConfigSchema, v as parseEvalMetas, w as createFsCacheStore, yn as configureEvalRunLogs, z as runSummarySchema } from "./runOrchestration-CimthgI7.mjs";
 import { createHash } from "node:crypto";
 import { readFile } from "node:fs/promises";
 import { relative } from "node:path";

package/dist/{runOrchestration-D1edUDhp.mjs → runOrchestration-CimthgI7.mjs}

@@ -49,9 +49,46 @@ function repoFile(path, mimeType) {
 //#region ../sdk/src/runtime.ts
 const scopeStorage = new AsyncLocalStorage();
 const runtimeScopeStorage = new AsyncLocalStorage();
+const evalClockStorage = new AsyncLocalStorage();
 let activeEvalScopeCount = 0;
 let activeEvalRuntimeScopeCount = 0;
 let consoleCaptureEnabled = true;
+const defaultEvalStartTimeMs = Date.parse("2026-04-10T00:00:00.000Z");
+const realDate = globalThis.__agentEvalsRealDate ?? Date;
+globalThis.__agentEvalsRealDate = realDate;
+function toDateConstructorArg(value) {
+	if (typeof value === "string" || typeof value === "number" || value instanceof realDate) return value;
+	return Number(value);
+}
+function toDateNumberArg(value) {
+	return typeof value === "number" ? value : Number(value);
+}
+function constructDateFromArgs(args) {
+	if (args.length === 0) return new realDate();
+	if (args.length === 1) return new realDate(toDateConstructorArg(args[0]));
+	return new realDate(toDateNumberArg(args[0]), toDateNumberArg(args[1]), args[2] === void 0 ? 1 : toDateNumberArg(args[2]), args[3] === void 0 ? 0 : toDateNumberArg(args[3]), args[4] === void 0 ? 0 : toDateNumberArg(args[4]), args[5] === void 0 ? 0 : toDateNumberArg(args[5]), args[6] === void 0 ? 0 : toDateNumberArg(args[6]));
+}
+const evalDate = new Proxy(realDate, {
+	apply(target, thisArg, argArray_) {
+		const nowMs = getEvalClockNowMs();
+		if (nowMs !== null) return new target(nowMs).toString();
+		return target.call(thisArg);
+	},
+	construct(target, argArray, newTarget_) {
+		const nowMs = getEvalClockNowMs();
+		if (argArray.length === 0 && nowMs !== null) return new target(nowMs);
+		return constructDateFromArgs(Array.from(argArray));
+	},
+	get(target, property) {
+		if (property === "now") return getEvalDateNow;
+		if (property === "parse") return target.parse;
+		if (property === "UTC") return target.UTC;
+		if (property === "prototype") return target.prototype;
+		if (property === "name") return target.name;
+		if (property === "length") return target.length;
+	}
+});
+globalThis.Date = evalDate;
 const maxLogMessageLength = 2e4;
 const maxLogStringLength = 1e4;
 const maxLogArrayLength = 100;
@@ -79,6 +116,82 @@ var EvalAssertionError = class extends Error {
 		this.name = "EvalAssertionError";
 	}
 };
+function getEvalClockStateNowMs(state) {
+	const elapsedMs = state.frozen ? 0 : realDate.now() - state.realStartMs;
+	return state.startMs + elapsedMs + state.offsetMs;
+}
+function getEvalClockNowMs() {
+	const state = evalClockStorage.getStore();
+	if (state?.shifted !== true) return null;
+	return getEvalClockStateNowMs(state);
+}
+function getEvalDateNow() {
+	return getEvalClockNowMs() ?? realDate.now();
+}
+/** Return the host process clock, bypassing the eval Date shim. */
+function getRealDateNowMs() {
+	return realDate.now();
+}
+/** Return the shifted wall-clock time for a stored eval clock state. */
+function getEvalClockStateTimeMs(state) {
+	if (!state.shifted) return null;
+	return getEvalClockStateNowMs(state);
+}
+/**
+* Return the wall-clock start time captured for the active eval.
+*
+* For `startTime: 'now'`, this is the real time captured when the eval clock
+* context was created.
+*/
+function getEvalStartTime() {
+	const state = evalClockStorage.getStore();
+	if (state === void 0) throw new Error("getEvalStartTime() must be called inside an active eval");
+	return new realDate(state.startMs);
+}
+function resolveEvalStartTimeMs(startTime) {
+	if (startTime === void 0) return defaultEvalStartTimeMs;
+	if (startTime === "now") return realDate.now();
+	const ms = startTime instanceof realDate ? startTime.getTime() : typeof startTime === "number" ? startTime : realDate.parse(startTime);
+	if (Number.isFinite(ms)) return ms;
+	throw new Error(`Invalid eval startTime "${String(startTime)}". Use a Date, timestamp, ISO date string, or "now".`);
+}
+function createEvalClockState(startTime, freezeTime) {
+	const nowMs = realDate.now();
+	return {
+		startMs: startTime === "now" ? nowMs : resolveEvalStartTimeMs(startTime),
+		realStartMs: nowMs,
+		offsetMs: 0,
+		frozen: freezeTime,
+		shifted: startTime !== "now" || freezeTime
+	};
+}
+/** Execute a callback with the eval Date clock shifted from `startTime`. */
+async function runWithEvalClock(startTime, fn, options = {}) {
+	return await evalClockStorage.run(createEvalClockState(startTime, options.freezeTime === true), fn);
+}
+function getEvalTimeUnitMs(unit) {
+	if (unit === "millisecond" || unit === "milliseconds") return 1;
+	if (unit === "second" || unit === "seconds") return 1e3;
+	if (unit === "minute" || unit === "minutes") return 6e4;
+	if (unit === "hour" || unit === "hours") return 36e5;
+	if (unit === "day" || unit === "days") return 864e5;
+	throw new Error(`Unsupported eval time unit "${unit}"`);
+}
+/**
+* Advance the active eval's shifted Date clock and return the new time.
+*
+* Throws outside an active shifted eval clock. Evals that set
+* `startTime: 'now'` use the real current clock unless `freezeTime: true` is
+* also set.
+*/
+function advanceEvalTime(unit, amount) {
+	const state = evalClockStorage.getStore();
+	if (state === void 0) throw new Error("advanceEvalTime() must be called inside an active eval");
+	if (!state.shifted) throw new Error("advanceEvalTime() requires a shifted eval clock. Remove startTime: \"now\" or set freezeTime: true to use it.");
+	if (!Number.isFinite(amount)) throw new Error("advanceEvalTime() amount must be a finite number");
+	state.offsetMs += getEvalTimeUnitMs(unit) * amount;
+	return new realDate(getEvalClockStateNowMs(state));
+}
 /** Return the current eval scope for the active async context, if any. */
 function getCurrentScope() {
 	if (activeEvalScopeCount === 0) return void 0;
@@ -349,7 +462,9 @@ async function runInExistingEvalScope(scope, runtimeScope, fn) {
 	activeEvalScopeCount++;
 	try {
 		return await scopeStorage.run(scope, async () => {
-			return await runInEvalRuntimeScope(runtimeScope, fn);
+			return await evalClockStorage.run(scope.evalClockState, async () => {
+				return await runInEvalRuntimeScope(runtimeScope, fn);
+			});
 		});
 	} finally {
 		activeEvalScopeCount--;
@@ -362,6 +477,8 @@ async function runInExistingEvalScope(scope, runtimeScope, fn) {
 async function runInEvalScope(caseId, fn, options = {}) {
 	const scope = {
 		caseId,
+		startTime: options.startTime,
+		evalClockState: createEvalClockState(options.startTime, options.freezeTime === true),
 		idPrefix: options.idPrefix,
 		nextEvalIdCounter: 0,
 		input: options.input,
@@ -1493,9 +1610,12 @@ function mergeSpanAttribute(span, key, patch) {
 		...patch
 	} });
 }
-function finishSpanWithoutThrownError(span) {
+function addElapsedMsToTimestamp(isoTimestamp, elapsedMs) {
+	return new Date(new Date(isoTimestamp).getTime() + elapsedMs).toISOString();
+}
+function finishSpanWithoutThrownError(span, realStartedAt) {
 	span.status = hasSpanError(span) ? "error" : "ok";
-	span.endedAt = (/* @__PURE__ */ new Date()).toISOString();
+	span.endedAt = addElapsedMsToTimestamp(span.startedAt, getRealDateNowMs() - realStartedAt);
 }
 function createSpanHandle(span) {
 	return {
@@ -1737,9 +1857,11 @@ async function traceSpanInternal(info, fn) {
 	const scope = getCurrentScope();
 	if (!scope) return await fn(noopActiveSpan());
 	const id = generateSpanId();
+	const parentId = scope.activeSpanStack.at(-1)?.id ?? null;
+	const realStartedAt = getRealDateNowMs();
 	const spanRecord = {
 		id,
-		parentId: scope.activeSpanStack.at(-1)?.id ?? null,
+		parentId,
 		caseId: scope.caseId,
 		kind: info.kind,
 		name: info.name,
@@ -1779,7 +1901,7 @@ async function traceSpanInternal(info, fn) {
 					const recording = deserializeCacheRecording(hit.recording);
 					replayRecording(scope, spanRecord, recording, { generateSpanId });
 					spanRecord.status = recording.finalStatus ?? (hasSpanError(spanRecord) ? "error" : "ok");
-					spanRecord.endedAt = (/* @__PURE__ */ new Date()).toISOString();
+					spanRecord.endedAt = addElapsedMsToTimestamp(spanRecord.startedAt, getRealDateNowMs() - realStartedAt);
 					return recording.returnValue;
 				}
 				mergeSpanAttributes(spanRecord, { "cache.status": "miss" });
@@ -1798,7 +1920,7 @@ async function traceSpanInternal(info, fn) {
 				scope.recordingStack.pop();
 			}
 			appendSubSpanOps(scope, frame);
-			finishSpanWithoutThrownError(spanRecord);
+			finishSpanWithoutThrownError(spanRecord, realStartedAt);
 			if (ctx.mode !== "bypass") {
 				const recording = {
 					returnValue: bodyResult,
@@ -1832,11 +1954,11 @@ async function traceSpanInternal(info, fn) {
 			return bodyResult;
 		}
 		const result = await fn(activeSpan);
-		finishSpanWithoutThrownError(spanRecord);
+		finishSpanWithoutThrownError(spanRecord, realStartedAt);
 		return result;
 	} catch (error) {
 		spanRecord.status = "error";
-		spanRecord.endedAt = (/* @__PURE__ */ new Date()).toISOString();
+		spanRecord.endedAt = addElapsedMsToTimestamp(spanRecord.startedAt, getRealDateNowMs() - realStartedAt);
 		spanRecord.error = normalizeTraceError(error);
 		throw error;
 	} finally {
@@ -5235,7 +5357,9 @@ async function runCase(params) {
 			mode: cacheMode,
 			evalId,
 			codeFingerprint
-		} : void 0
+		} : void 0,
+		startTime: evalDef.startTime,
+		freezeTime: evalDef.freezeTime
 	});
 	const traceTree = buildTraceTree(scope.spans, scope.checkpoints);
 	const nonAssertError = executeError && !(executeError instanceof EvalAssertionError) ? executeError : null;
@@ -5276,6 +5400,7 @@ async function runCase(params) {
 	}
 	const scoreResults = /* @__PURE__ */ new Map();
 	const scoringTraces = {};
+	const scoreStartTime = getEvalClockStateTimeMs(scope.evalClockState) ?? evalDef.startTime;
 	if (!nonAssertError && scope.assertionFailures.length === 0 && evalDef.scores) for (const [key, def] of Object.entries(evalDef.scores)) {
 		const { compute, passThreshold, label } = normalizeScoreDef(def);
 		const scoreRun = await runInEvalScope(evalCase.id, async () => {
@@ -5295,7 +5420,9 @@ async function runCase(params) {
 				mode: cacheMode,
 				evalId: `${evalId}__score__${key}`,
 				codeFingerprint
-			} : void 0
+			} : void 0,
+			startTime: scoreStartTime,
+			freezeTime: evalDef.freezeTime
 		});
 		const { trace, traceDisplay } = resolveTracePresentation(scoreRun.scope.spans, globalTraceDisplay, evalDef.traceDisplay);
 		scope.logs.push(...scoreRun.scope.logs.map((entry) => ({
@@ -5608,7 +5735,7 @@ async function executeRun({ runState, request, runDir, config, evals, cacheStore
 					await runInEvalRuntimeScope("cases", async () => {
 						await entry.use(async (evalDef) => {
 							const cases = filterEvalCases(resolveRunnableEvalCases({
-								cases: typeof evalDef.cases === "function" ? await evalDef.cases() : evalDef.cases ?? [],
+								cases: await runWithEvalClock(evalDef.startTime, async () => typeof evalDef.cases === "function" ? await evalDef.cases() : evalDef.cases ?? [], { freezeTime: evalDef.freezeTime }),
 								evalId: evalMeta.id
 							}), request.target.evalIds, request.target.caseIds, evalMeta.id);
 							runState.summary.totalCases += cases.length;
@@ -5811,4 +5938,4 @@ function toLastRunStatus(status) {
 	return status === "pending" ? null : status;
 }
 //#endregion
-export { removeDefaultConfigSchema as $, columnKindSchema as $t, extractApiCalls as A, setEvalOutput as An, cacheFileSchema as At, DEFAULT_API_CALLS_CONFIG as B, traceAttributeDisplayFormatSchema as Bt, validateCharts as C, incrementEvalOutput as Cn, evalChartTooltipExtraSchema as Ct, sseEnvelopeSchema as D, runInEvalRuntimeScope as Dn, cacheDebugKeyFileSchema as Dt, updateManualScoreRequestSchema as E, nextEvalId as En, cacheDebugKeyEntrySchema as Et, deriveScopedSummaryFromCases as F, getEvalRegistry as Fn, cacheRecordingSchema as Ft, apiCallMetricSchema as G, traceDisplayInputConfigSchema as Gt, agentEvalsConfigSchema as H, traceAttributeDisplayPlacementSchema as Ht, deriveStatusFromCaseRows as I, cacheStatusSchema as It, llmCallMetricFormatSchema as J, traceSpanSchema as Jt, apiCallsConfigSchema as K, traceSpanErrorSchema as Kt, deriveStatusFromChildStatuses as L, serializedCacheSpanSchema as Lt, getNestedAttribute as M, startEvalBackgroundJob as Mn, cacheModeSchema as Mt, getEvalTitle as N, repoFile as Nn, cacheOperationTypeSchema as Nt, extractCacheEntries as O, runInEvalScope as On, cacheEntrySchema as Ot, getEvalDisplayStatus as P, defineEval as Pn, cacheRecordingOpSchema as Pt, llmCallsConfigSchema as Q, columnFormatSchema as Qt, runManifestSchema as R, spanCacheOptionsSchema as Rt, normalizeScoreDef as S, getEvalCaseInput as Sn, evalChartMetricSchema as St, createRunRequestSchema as T, mergeEvalOutput as Tn, evalChartsConfigSchema as Tt, apiCallMetricFormatSchema as U, traceAttributeDisplaySchema as Ut, DEFAULT_LLM_CALLS_CONFIG as V, traceAttributeDisplayInputSchema as Vt, apiCallMetricPlacementSchema as W, traceDisplayConfigSchema as Wt, llmCallMetricSchema as X, cellValueSchema as Xt, llmCallMetricPlacementSchema as Y, traceSpanWarningSchema as Yt, llmCallPricingSchema as Z, columnDefSchema as Zt, loadEvalModule as _, appendToEvalOutput as _n, evalChartAggregateSchema as _t, loadPersistedRunSnapshot as a, z$1 as an, caseDetailSchema as at, loadConfig as b, evalLog as bn, evalChartColorSchema as bt, persistCaseDetail as c, evalSpan as cn, evalStatAggregateSchema as ct, recomputePersistedCaseStatus as d, hashCacheKeySync as dn, evalSummarySchema as dt, fileRefSchema as en, resolveApiCallsConfig as et, runTouchesEval as f, deserializeCacheRecording as fn, runLogEntrySchema as ft, setLatestRunInfoMap as g, EvalAssertionError as gn, scoreTraceSchema as gt, getTargetEvalIds as h, serializeCacheValue as hn, runLogPhaseSchema as ht, getLatestRunInfos as i, runArtifactRefSchema as in, assertionFailureSchema as it, extractLlmCalls as j, setScopeCacheContext as jn, cacheListItemSchema as jt, extractCacheHits as k, runInExistingEvalScope as kn, cacheEntryWithDebugKeySchema as kt, persistRunState as l, evalTracer as ln, evalStatItemSchema as lt, buildEvalSummary as m, serializeCacheRecording as mn, runLogLocationSchema as mt, generateRunId as n, numberDisplayOptionsSchema as nn, runLogsConfigSchema as nt, loadPersistedRunSnapshots as o, buildTraceTree as on, caseRowSchema as ot, resolveArtifactPath as p, deserializeCacheValue as pn, runLogLevelSchema as pt, defaultConfigKeySchema as q, traceSpanKindSchema as qt, getLastRunStatuses as r, repoFileRefSchema as rn, trialSelectionModeSchema as rt, nextShortIdFromSnapshots as s, captureEvalSpanError as sn, evalFreshnessStatusSchema as st, executeRun as t, jsonCellSchema as tn, resolveLlmCallsConfig as tt, recomputeEvalStatusesInRuns as u, hashCacheKey as un, evalStatsConfigSchema as ut, parseEvalMetas as v, configureEvalRunLogs as vn, evalChartAxisSchema as vt, createFsCacheStore as w, isInEvalScope as wn, evalChartTypeSchema as wt, buildDeclaredColumnDefs as x, getCurrentScope as xn, evalChartConfigSchema as xt, resolveEvalDefaultConfig as y, evalAssert as yn, evalChartBuiltinMetricSchema as yt, runSummarySchema as z, traceCacheRefSchema as zt };
+export { removeDefaultConfigSchema as $, columnKindSchema as $t, extractApiCalls as A, runInEvalScope as An, cacheFileSchema as At, DEFAULT_API_CALLS_CONFIG as B, traceAttributeDisplayFormatSchema as Bt, validateCharts as C, getEvalCaseInput as Cn, evalChartTooltipExtraSchema as Ct, sseEnvelopeSchema as D, mergeEvalOutput as Dn, cacheDebugKeyFileSchema as Dt, updateManualScoreRequestSchema as E, isInEvalScope as En, cacheDebugKeyEntrySchema as Et, deriveScopedSummaryFromCases as F, repoFile as Fn, cacheRecordingSchema as Ft, apiCallMetricSchema as G, traceDisplayInputConfigSchema as Gt, agentEvalsConfigSchema as H, traceAttributeDisplayPlacementSchema as Ht, deriveStatusFromCaseRows as I, defineEval as In, cacheStatusSchema as It, llmCallMetricFormatSchema as J, traceSpanSchema as Jt, apiCallsConfigSchema as K, traceSpanErrorSchema as Kt, deriveStatusFromChildStatuses as L, getEvalRegistry as Ln, serializedCacheSpanSchema as Lt, getNestedAttribute as M, setEvalOutput as Mn, cacheModeSchema as Mt, getEvalTitle as N, setScopeCacheContext as Nn, cacheOperationTypeSchema as Nt, extractCacheEntries as O, nextEvalId as On, cacheEntrySchema as Ot, getEvalDisplayStatus as P, startEvalBackgroundJob as Pn, cacheRecordingOpSchema as Pt, llmCallsConfigSchema as Q, columnFormatSchema as Qt, runManifestSchema as R, spanCacheOptionsSchema as Rt, normalizeScoreDef as S, getCurrentScope as Sn, evalChartMetricSchema as St, createRunRequestSchema as T, incrementEvalOutput as Tn, evalChartsConfigSchema as Tt, apiCallMetricFormatSchema as U, traceAttributeDisplaySchema as Ut, DEFAULT_LLM_CALLS_CONFIG as V, traceAttributeDisplayInputSchema as Vt, apiCallMetricPlacementSchema as W, traceDisplayConfigSchema as Wt, llmCallMetricSchema as X, cellValueSchema as Xt, llmCallMetricPlacementSchema as Y, traceSpanWarningSchema as Yt, llmCallPricingSchema as Z, columnDefSchema as Zt, loadEvalModule as _, advanceEvalTime as _n, evalChartAggregateSchema as _t, loadPersistedRunSnapshot as a, z$1 as an, caseDetailSchema as at, loadConfig as b, evalAssert as bn, evalChartColorSchema as bt, persistCaseDetail as c, evalSpan as cn, evalStatAggregateSchema as ct, recomputePersistedCaseStatus as d, hashCacheKeySync as dn, evalSummarySchema as dt, fileRefSchema as en, resolveApiCallsConfig as et, runTouchesEval as f, deserializeCacheRecording as fn, runLogEntrySchema as ft, setLatestRunInfoMap as g, EvalAssertionError as gn, scoreTraceSchema as gt, getTargetEvalIds as h, serializeCacheValue as hn, runLogPhaseSchema as ht, getLatestRunInfos as i, runArtifactRefSchema as in, assertionFailureSchema as it, extractLlmCalls as j, runInExistingEvalScope as jn, cacheListItemSchema as jt, extractCacheHits as k, runInEvalRuntimeScope as kn, cacheEntryWithDebugKeySchema as kt, persistRunState as l, evalTracer as ln, evalStatItemSchema as lt, buildEvalSummary as m, serializeCacheRecording as mn, runLogLocationSchema as mt, generateRunId as n, numberDisplayOptionsSchema as nn, runLogsConfigSchema as nt, loadPersistedRunSnapshots as o, buildTraceTree as on, caseRowSchema as ot, resolveArtifactPath as p, deserializeCacheValue as pn, runLogLevelSchema as pt, defaultConfigKeySchema as q, traceSpanKindSchema as qt, getLastRunStatuses as r, repoFileRefSchema as rn, trialSelectionModeSchema as rt, nextShortIdFromSnapshots as s, captureEvalSpanError as sn, evalFreshnessStatusSchema as st, executeRun as t, jsonCellSchema as tn, resolveLlmCallsConfig as tt, recomputeEvalStatusesInRuns as u, hashCacheKey as un, evalStatsConfigSchema as ut, parseEvalMetas as v, appendToEvalOutput as vn, evalChartAxisSchema as vt, createFsCacheStore as w, getEvalStartTime as wn, evalChartTypeSchema as wt, buildDeclaredColumnDefs as x, evalLog as xn, evalChartConfigSchema as xt, resolveEvalDefaultConfig as y, configureEvalRunLogs as yn, evalChartBuiltinMetricSchema as yt, runSummarySchema as z, traceCacheRefSchema as zt };

package/dist/{runner-C9nP2VKL.mjs → runner-4yNYRvmF.mjs}

@@ -1,5 +1,5 @@
-import { n as createRunner } from "./cli-C0EtHhEO.mjs";
-import "./src-D-HuV8I-.mjs";
+import { n as createRunner } from "./cli-D3QNOcPN.mjs";
+import "./src-CcXfWT4M.mjs";
 //#region ../../apps/server/src/runner.ts
 let runnerInstance = null;
 function getRunnerInstance() {

package/dist/{runner-CyRhIzci.mjs → runner-B-SYzW8w.mjs}

@@ -1,2 +1,2 @@
-import { n as initRunner, t as getRunnerInstance } from "./runner-C9nP2VKL.mjs";
+import { n as initRunner, t as getRunnerInstance } from "./runner-4yNYRvmF.mjs";
 export { getRunnerInstance, initRunner };

package/dist/src-CcXfWT4M.mjs

@@ -0,0 +1,3 @@
+import "./runOrchestration-CimthgI7.mjs";
+import "./cli-D3QNOcPN.mjs";
+export {};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ls-stack/agent-eval",
-  "version": "0.22.0",
+  "version": "0.23.0",
   "type": "module",
   "bin": {
     "agent-evals": "./dist/bin.mjs"

package/skills/agent-eval/SKILL.md

@@ -156,6 +156,19 @@ for settlement; promise and span errors keep their normal behavior. Use
 `waitForBackgroundJob: false` on a span, or `waitForBackgroundJobs: false` on an
 eval definition, when background work should not delay finalization.
 
+Eval Date APIs use a shifted wall clock by default: `new Date()` and
+`Date.now()` start at `2026-04-10T00:00:00.000Z` during case generation,
+execution, tracing, derived outputs, and scorers, then continue advancing with
+real elapsed time. Set `startTime` on a specific `defineEval(...)` to use
+another initial clock value, or set `startTime: 'now'` for that eval to use the
+real current clock. Timers are not faked, so async waits still run normally.
+Set `freezeTime: true` to keep Date APIs frozen until they are moved manually.
+Use `getEvalStartTime()` to read the captured wall-clock start as a `Date`.
+Use `advanceEvalTime(unit, amount)` inside an eval to move the shifted clock
+forward; supported units are `millisecond(s)`, `second(s)`, `minute(s)`,
+`hour(s)`, and `day(s)`. It throws for evals with `startTime: 'now'`, unless
+`freezeTime: true` is also set.
+
 For libraries or observability exporters that already emit span lifecycle
 events, use `evalTracer.startSpan(...)`, `evalTracer.updateSpan(...)`,
 `evalTracer.endSpan(...)`, or `evalTracer.recordSpan(...)` to translate those

package/dist/src-D-HuV8I-.mjs

@@ -1,3 +0,0 @@
-import "./runOrchestration-D1edUDhp.mjs";
-import "./cli-C0EtHhEO.mjs";
-export {};