@ls-stack/agent-eval 0.20.0 → 0.22.0

package/dist/index.d.mts

@@ -39,8 +39,9 @@ type NumberDisplayOptions = {
   /** Number notation used when rendering the value. */notation?: 'standard' | 'compact'; /** Compact style used when `notation: 'compact'` is enabled. */
   compactDisplay?: 'short' | 'long'; /** String prepended to the rendered number, such as `$`. */
   prefix?: string; /** String appended to the rendered number, such as ` ms`. */
-  suffix?: string; /** Fixed number of decimal places to render. */
-  decimalPlaces?: number;
+  suffix?: string; /** Minimum number of decimal places to render. */
+  minDecimalPlaces?: number; /** Maximum number of decimal places to render. */
+  maxDecimalPlaces?: number;
 };
 /** Schema for numeric presentation options used by number-formatted values. */
 declare const numberDisplayOptionsSchema: z$1.ZodType<NumberDisplayOptions>;
@@ -98,7 +99,6 @@ declare const columnDefSchema: z$1.ZodObject<{
   passThreshold: z$1.ZodOptional<z$1.ZodNumber>;
   maxStars: z$1.ZodOptional<z$1.ZodNumber>;
   hideInTable: z$1.ZodOptional<z$1.ZodBoolean>;
-  sortable: z$1.ZodOptional<z$1.ZodBoolean>;
   align: z$1.ZodOptional<z$1.ZodEnum<{
     left: "left";
     center: "center";
@@ -412,6 +412,7 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
     passFail: "passFail";
     stars: "stars";
   }>>;
+  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
   accent: z$1.ZodOptional<z$1.ZodBoolean>;
 }, z$1.core.$strip>], "kind">;
 /** Single stat rendered in the EvalCard stats row. */
@@ -449,6 +450,7 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
     passFail: "passFail";
     stars: "stars";
   }>>;
+  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
   accent: z$1.ZodOptional<z$1.ZodBoolean>;
 }, z$1.core.$strip>], "kind">>;
 /** Ordered list of stats rendered in the EvalCard stats row. */
@@ -496,7 +498,6 @@ declare const evalSummarySchema: z$1.ZodObject<{
     passThreshold: z$1.ZodOptional<z$1.ZodNumber>;
     maxStars: z$1.ZodOptional<z$1.ZodNumber>;
     hideInTable: z$1.ZodOptional<z$1.ZodBoolean>;
-    sortable: z$1.ZodOptional<z$1.ZodBoolean>;
     align: z$1.ZodOptional<z$1.ZodEnum<{
       left: "left";
       center: "center";
@@ -544,6 +545,7 @@ declare const evalSummarySchema: z$1.ZodObject<{
       passFail: "passFail";
       stars: "stars";
     }>>;
+    numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
     accent: z$1.ZodOptional<z$1.ZodBoolean>;
   }, z$1.core.$strip>], "kind">>>;
   charts: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
@@ -643,7 +645,7 @@ declare const caseRowSchema: z$1.ZodObject<{
     cancelled: "cancelled";
     pending: "pending";
   }>;
-  latencyMs: z$1.ZodNullable<z$1.ZodNumber>;
+  durationMs: z$1.ZodNullable<z$1.ZodNumber>;
   costUsd: z$1.ZodOptional<z$1.ZodNullable<z$1.ZodNumber>>;
   columns: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnion<readonly [z$1.ZodType<string | number | boolean | Record<string, unknown> | unknown[] | null, unknown, z$1.core.$ZodTypeInternals<string | number | boolean | Record<string, unknown> | unknown[] | null, unknown>>, z$1.ZodUnion<readonly [z$1.ZodObject<{
     source: z$1.ZodLiteral<"repo">;
@@ -1484,7 +1486,7 @@ declare const defaultConfigKeySchema: z$1.ZodEnum<{
   cachedInputTokens: "cachedInputTokens";
   cacheCreationInputTokens: "cacheCreationInputTokens";
   reasoningTokens: "reasoningTokens";
-  llmLatencyMs: "llmLatencyMs";
+  llmDurationMs: "llmDurationMs";
 }>;
 /** Built-in eval-level output/column key. */
 type DefaultConfigKey = z$1.infer<typeof defaultConfigKeySchema>;
@@ -1499,7 +1501,7 @@ declare const removeDefaultConfigSchema: z$1.ZodUnion<readonly [z$1.ZodLiteral<t
   cachedInputTokens: "cachedInputTokens";
   cacheCreationInputTokens: "cacheCreationInputTokens";
   reasoningTokens: "reasoningTokens";
-  llmLatencyMs: "llmLatencyMs";
+  llmDurationMs: "llmDurationMs";
 }>>]>;
 /** Removal config for built-in eval-level outputs and UI metadata. */
 type RemoveDefaultConfig = z$1.infer<typeof removeDefaultConfigSchema>;
@@ -1593,7 +1595,7 @@ declare const apiCallMetricSchema: z$1.ZodObject<{
 type ApiCallMetric = z$1.infer<typeof apiCallMetricSchema>;
 /**
  * Schema for one model/provider pricing entry used to derive LLM-call costs
- * from token counts when a span does not already record explicit USD costs.
+ * from token counts.
  */
 declare const llmCallPricingSchema: z$1.ZodObject<{
   model: z$1.ZodString;
@@ -1602,6 +1604,7 @@ declare const llmCallPricingSchema: z$1.ZodObject<{
   outputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
   cachedInputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
   cacheCreationInputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
+  cacheCreationInput1hUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
   reasoningUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
 }, z$1.core.$strip>;
 /** Model/provider pricing entry authored in `agent-evals.config.ts`. */
@@ -1616,15 +1619,9 @@ declare const llmCallsConfigSchema: z$1.ZodObject<{
     outputTokens: z$1.ZodOptional<z$1.ZodString>;
     cachedInputTokens: z$1.ZodOptional<z$1.ZodString>;
     cacheCreationInputTokens: z$1.ZodOptional<z$1.ZodString>;
+    cacheCreationInput1hTokens: z$1.ZodOptional<z$1.ZodString>;
     reasoningTokens: z$1.ZodOptional<z$1.ZodString>;
-    totalTokens: z$1.ZodOptional<z$1.ZodString>;
-    tokensPerSecond: z$1.ZodOptional<z$1.ZodString>;
-    cost: z$1.ZodOptional<z$1.ZodString>;
-    inputCost: z$1.ZodOptional<z$1.ZodString>;
-    outputCost: z$1.ZodOptional<z$1.ZodString>;
-    cachedInputCost: z$1.ZodOptional<z$1.ZodString>;
-    cacheCreationInputCost: z$1.ZodOptional<z$1.ZodString>;
-    reasoningCost: z$1.ZodOptional<z$1.ZodString>;
+    latencyMs: z$1.ZodOptional<z$1.ZodString>;
     steps: z$1.ZodOptional<z$1.ZodString>;
     finishReason: z$1.ZodOptional<z$1.ZodString>;
     input: z$1.ZodOptional<z$1.ZodString>;
@@ -1639,6 +1636,7 @@ declare const llmCallsConfigSchema: z$1.ZodObject<{
     outputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
     cachedInputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
     cacheCreationInputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
+    cacheCreationInput1hUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
     reasoningUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
   }, z$1.core.$strip>>>;
   metrics: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
@@ -1712,15 +1710,9 @@ type ResolvedLlmCallsConfig = {
     outputTokens: string;
     cachedInputTokens: string;
     cacheCreationInputTokens: string;
+    cacheCreationInput1hTokens: string;
     reasoningTokens: string;
-    totalTokens: string;
-    tokensPerSecond: string;
-    cost: string;
-    inputCost: string;
-    outputCost: string;
-    cachedInputCost: string;
-    cacheCreationInputCost: string;
-    reasoningCost: string;
+    latencyMs: string;
     steps: string;
     finishReason: string;
     input: string;
@@ -1774,6 +1766,7 @@ type ResolvedLlmCallPricing = {
   outputUsdPerMillion?: number;
   cachedInputUsdPerMillion?: number;
   cacheCreationInputUsdPerMillion?: number;
+  cacheCreationInput1hUsdPerMillion?: number;
   reasoningUsdPerMillion?: number;
 };
 /** Default LLM-calls config the UI uses before the workspace fetch resolves. */
@@ -1789,8 +1782,8 @@ declare const DEFAULT_API_CALLS_CONFIG: ResolvedApiCallsConfig;
  *   attribute path.
  * - Missing `metrics[].format` defaults to `'string'`.
  * - Missing `metrics[].placements` defaults to `['body']`.
- * - Missing `pricing` defaults to an empty registry; explicit span costs still
- *   take precedence over derived costs.
+ * - Missing `pricing` defaults to an empty registry; built-in costs are only
+ *   derived from configured pricing and token counts.
  */
 declare function resolveLlmCallsConfig(input: LlmCallsConfigInput | undefined): ResolvedLlmCallsConfig;
 /**
@@ -1844,10 +1837,10 @@ type AgentEvalsConfig = {
    *
    * Determines which trace spans are treated as LLM calls (`kinds`), how
    * structured fields like `model` and `usage.inputTokens` are read from
-   * span attributes, and which custom user-defined metrics are surfaced on
-   * each call. All fields are optional and fall back to the documented
-   * defaults; the LLM calls tab is shown automatically when at least one
-   * matching span exists in a case run.
+   * span attributes, which pricing table derives built-in costs, and which
+   * custom user-defined metrics are surfaced on each call. All fields are
+   * optional and fall back to the documented defaults; the LLM calls tab is
+   * shown automatically when at least one matching span exists in a case run.
    *
    * @example
    * ```ts
@@ -1980,15 +1973,9 @@ declare const agentEvalsConfigSchema: z$1.ZodObject<{
       outputTokens: z$1.ZodOptional<z$1.ZodString>;
       cachedInputTokens: z$1.ZodOptional<z$1.ZodString>;
       cacheCreationInputTokens: z$1.ZodOptional<z$1.ZodString>;
+      cacheCreationInput1hTokens: z$1.ZodOptional<z$1.ZodString>;
       reasoningTokens: z$1.ZodOptional<z$1.ZodString>;
-      totalTokens: z$1.ZodOptional<z$1.ZodString>;
-      tokensPerSecond: z$1.ZodOptional<z$1.ZodString>;
-      cost: z$1.ZodOptional<z$1.ZodString>;
-      inputCost: z$1.ZodOptional<z$1.ZodString>;
-      outputCost: z$1.ZodOptional<z$1.ZodString>;
-      cachedInputCost: z$1.ZodOptional<z$1.ZodString>;
-      cacheCreationInputCost: z$1.ZodOptional<z$1.ZodString>;
-      reasoningCost: z$1.ZodOptional<z$1.ZodString>;
+      latencyMs: z$1.ZodOptional<z$1.ZodString>;
       steps: z$1.ZodOptional<z$1.ZodString>;
       finishReason: z$1.ZodOptional<z$1.ZodString>;
       input: z$1.ZodOptional<z$1.ZodString>;
@@ -2003,6 +1990,7 @@ declare const agentEvalsConfigSchema: z$1.ZodObject<{
       outputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
       cachedInputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
       cacheCreationInputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
+      cacheCreationInput1hUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
       reasoningUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
     }, z$1.core.$strip>>>;
     metrics: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
@@ -2033,7 +2021,7 @@ declare const agentEvalsConfigSchema: z$1.ZodObject<{
     cachedInputTokens: "cachedInputTokens";
     cacheCreationInputTokens: "cacheCreationInputTokens";
     reasoningTokens: "reasoningTokens";
-    llmLatencyMs: "llmLatencyMs";
+    llmDurationMs: "llmDurationMs";
   }>>]>>;
   apiCalls: z$1.ZodOptional<z$1.ZodObject<{
     kinds: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
@@ -2102,7 +2090,8 @@ type LlmCallEntry = {
   cachedInputTokens: number | null;
   cacheCreationInputTokens: number | null;
   reasoningTokens: number | null;
-  totalTokens: number | null;
+  totalTokens: number | null; /** Time to first token for the LLM call in milliseconds, when reported by the span. */
+  latencyMs: number | null;
   tokensPerSecond: number | null;
   costUsd: number | null;
   inputCostUsd: number | null;
@@ -2112,8 +2101,8 @@ type LlmCallEntry = {
   reasoningCostUsd: number | null; /** Number of inference rounds. Derived from the array length when `stepDetails` is set. */
   stepCount: number | null; /** Per-step breakdown when the configured `steps` attribute resolves to an array. */
   stepDetails: unknown[] | null;
-  finishReason: string | null;
-  latencyMs: number | null;
+  finishReason: string | null; /** Elapsed LLM call span duration in milliseconds. */
+  durationMs: number | null;
   input: unknown;
   output: unknown;
   reasoning: unknown;
@@ -2127,16 +2116,22 @@ type LlmCallEntry = {
  * shape consumed by the LLM calls tab.
  *
  * Spans whose `kind` is not in `config.kinds` are dropped. Structured fields
- * (`model`, token counts, explicit cost, etc.) are read via
+ * (`model`, token counts, latency, etc.) are read via
  * `getNestedAttribute` from the configured paths, with safe coercion to
- * `string | null` / `number | null`. When explicit USD costs are absent,
- * configured model pricing derives per-token-type costs from token counts.
- * `totalTokens` falls back to a sum of input + output + cached when no
- * explicit total attribute is present. The `steps` attribute path may resolve
- * to either a number (rendered as the inference-round count) or an array of
- * per-step detail objects (rendered as a Steps section in the body, with
- * `stepCount` derived from the array length). `latencyMs` is `null` while the
- * span is still running. User-defined `metrics` whose path resolves to
+ * `string | null` / `number | null`. `latencyMs` is an explicit
+ * time-to-first-token attribute; full span elapsed time is reported separately
+ * as `durationMs`. Built-in USD costs are derived only from configured model
+ * pricing and token counts. `totalTokens` is always derived from input +
+ * output tokens. Cached input and cache creation tokens are reported
+ * separately because they are subsets of input/output usage. The main cache
+ * creation token field is treated as the total write count; optional one-hour
+ * cache creation tokens only split that total for cost calculation. Base input
+ * cost uses input minus cache read/write tokens so cached tokens are not
+ * charged twice. Cache read/write costs still contribute to the total USD cost
+ * at their configured rates. The `steps` attribute path may resolve to an array
+ * of per-step detail objects, with `stepCount` derived from the array length.
+ * `durationMs` and `tokensPerSecond` are `null` while the span is still
+ * running. User-defined `metrics` whose path resolves to
  * `undefined` are dropped, but `null`, `0`, and `false` are preserved as
  * legitimate values worth displaying. Original span order is preserved so the
  * LLM calls tab matches the ordering in the Trace tab.
@@ -2161,8 +2156,8 @@ type ApiCallEntry = {
   status: EvalTraceSpan['status'];
   method: string | null;
   url: string | null;
-  statusCode: number | null;
-  latencyMs: number | null;
+  statusCode: number | null; /** Elapsed API call duration in milliseconds. */
+  durationMs: number | null;
   request: unknown;
   response: unknown;
   requestBody: unknown;
@@ -2179,10 +2174,10 @@ type ApiCallEntry = {
  *
  * Spans whose `kind` is not in `config.kinds` are dropped. Structured fields
  * (`method`, `url`, `statusCode`, etc.) are read via `getNestedAttribute` from
- * the configured paths. `durationMs` takes precedence for latency, with a
- * fallback to the span start/end timestamps. User-defined `metrics` whose path
- * resolves to `undefined` are dropped, but `null`, `0`, and `false` are
- * preserved as legitimate values worth displaying. Original span order is
+ * the configured paths. An explicit `durationMs` attribute takes precedence,
+ * with a fallback to the span start/end timestamps. User-defined `metrics`
+ * whose path resolves to `undefined` are dropped, but `null`, `0`, and `false`
+ * are preserved as legitimate values worth displaying. Original span order is
  * preserved so the API calls tab matches the ordering in the Trace tab.
  */
 declare function extractApiCalls(spans: EvalTraceSpan[], config: ResolvedApiCallsConfig): ApiCallEntry[];
@@ -2784,7 +2779,7 @@ type EvalColumnOverride = {
   /**
    * Extra options for `format: 'number'`.
    *
-   * Use this to add a prefix or suffix, force a fixed number of decimal
+   * Use this to add a prefix or suffix, control minimum and maximum decimal
    * places, or switch to compact notation such as `1.2K`.
    */
   numberFormat?: NumberDisplayOptions;
@@ -2792,8 +2787,7 @@ type EvalColumnOverride = {
    * Hides the column from the runs table while keeping it available in detail
    * views and raw output data.
    */
-  hideInTable?: boolean; /** Whether the UI should allow sorting rows by this column. */
-  sortable?: boolean; /** Horizontal alignment used when rendering the column cells. */
+  hideInTable?: boolean; /** Horizontal alignment used when rendering the column cells. */
   align?: 'left' | 'center' | 'right';
   /**
    * Maximum number of stars used when `format: 'stars'`.
@@ -2954,7 +2948,8 @@ type EvalDefinitionBase<TInput = unknown, TOutputs extends EvalOutputs = EvalOut
    * column across the latest run's cases — `key` must match one of the eval's
    * score or column keys, and only finite numeric values participate in the
    * reduction. When no case has a numeric value for the key the stat renders
-   * an em dash. `label` and `format` default to the matching `ColumnDef`.
+   * an em dash. `label`, `format`, and `numberFormat` default to the matching
+   * `ColumnDef`.
    */
   stats?: EvalStatsConfig;
   /**

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-Cv1kiOAG.mjs";
-import { n as createRunner, t as runCli } from "./cli-weogme5U.mjs";
-import "./src-B879LZfo.mjs";
+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 };

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-Cv1kiOAG.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, vn as configureEvalRunLogs, w as createFsCacheStore, z as runSummarySchema } from "./runOrchestration-D1edUDhp.mjs";
 import { createHash } from "node:crypto";
 import { readFile } from "node:fs/promises";
 import { relative } from "node:path";