@ls-stack/agent-eval 0.27.0 → 0.28.0

package/dist/index.d.mts

@@ -99,6 +99,7 @@ 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>;
+  hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   align: z$1.ZodOptional<z$1.ZodEnum<{
     left: "left";
     center: "center";
@@ -381,13 +382,17 @@ type EvalStatAggregate = z$1.infer<typeof evalStatAggregateSchema>;
  * `column` aggregates a score or numeric output column across the latest run.
  */
 declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
+  hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"cases">;
 }, z$1.core.$strip>, z$1.ZodObject<{
+  hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"passRate">;
   accent: z$1.ZodOptional<z$1.ZodBoolean>;
 }, z$1.core.$strip>, z$1.ZodObject<{
+  hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"duration">;
 }, z$1.core.$strip>, z$1.ZodObject<{
+  hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"column">;
   key: z$1.ZodString;
   label: z$1.ZodOptional<z$1.ZodString>;
@@ -419,13 +424,17 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
 type EvalStatItem = z$1.infer<typeof evalStatItemSchema>;
 /** Ordered list of stats rendered in the EvalCard stats row. */
 declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
+  hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"cases">;
 }, z$1.core.$strip>, z$1.ZodObject<{
+  hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"passRate">;
   accent: z$1.ZodOptional<z$1.ZodBoolean>;
 }, z$1.core.$strip>, z$1.ZodObject<{
+  hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"duration">;
 }, z$1.core.$strip>, z$1.ZodObject<{
+  hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"column">;
   key: z$1.ZodString;
   label: z$1.ZodOptional<z$1.ZodString>;
@@ -499,6 +508,7 @@ 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>;
+    hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
     align: z$1.ZodOptional<z$1.ZodEnum<{
       left: "left";
       center: "center";
@@ -515,13 +525,17 @@ declare const evalSummarySchema: z$1.ZodObject<{
     unscored: "unscored";
   }>>;
   stats: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
+    hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
     kind: z$1.ZodLiteral<"cases">;
   }, z$1.core.$strip>, z$1.ZodObject<{
+    hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
     kind: z$1.ZodLiteral<"passRate">;
     accent: z$1.ZodOptional<z$1.ZodBoolean>;
   }, z$1.core.$strip>, z$1.ZodObject<{
+    hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
     kind: z$1.ZodLiteral<"duration">;
   }, z$1.core.$strip>, z$1.ZodObject<{
+    hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
     kind: z$1.ZodLiteral<"column">;
     key: z$1.ZodString;
     label: z$1.ZodOptional<z$1.ZodString>;
@@ -551,6 +565,7 @@ declare const evalSummarySchema: z$1.ZodObject<{
   }, z$1.core.$strip>], "kind">>>;
   charts: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
     heading: z$1.ZodOptional<z$1.ZodString>;
+    hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
     type: z$1.ZodEnum<{
       area: "area";
       line: "line";
@@ -666,6 +681,7 @@ declare const caseRowSchema: z$1.ZodObject<{
 type CaseRow = z$1.infer<typeof caseRowSchema>;
 /** Structured assertion failure metadata captured for one case run. */
 declare const assertionFailureSchema: z$1.ZodObject<{
+  name: z$1.ZodOptional<z$1.ZodString>;
   message: z$1.ZodString;
   stack: z$1.ZodOptional<z$1.ZodString>;
 }, z$1.core.$strip>;
@@ -961,10 +977,12 @@ declare const caseDetailSchema: z$1.ZodObject<{
     fileName: z$1.ZodOptional<z$1.ZodString>;
   }, z$1.core.$strip>]>]>>;
   assertionFailures: z$1.ZodArray<z$1.ZodUnion<readonly [z$1.ZodObject<{
+    name: z$1.ZodOptional<z$1.ZodString>;
     message: z$1.ZodString;
     stack: z$1.ZodOptional<z$1.ZodString>;
   }, z$1.core.$strip>, z$1.ZodPipe<z$1.ZodString, z$1.ZodTransform<{
     message: string;
+    name?: string | undefined;
     stack?: string | undefined;
   }, string>>]>>;
   logs: z$1.ZodDefault<z$1.ZodArray<z$1.ZodObject<{
@@ -1008,6 +1026,8 @@ declare const caseDetailSchema: z$1.ZodObject<{
       refresh: "refresh";
       bypass: "bypass";
     }>;
+    read: z$1.ZodOptional<z$1.ZodBoolean>;
+    stored: z$1.ZodOptional<z$1.ZodBoolean>;
     storedAt: z$1.ZodOptional<z$1.ZodString>;
     age: z$1.ZodOptional<z$1.ZodNumber>;
   }, z$1.core.$strip>>>;
@@ -1187,6 +1207,7 @@ type EvalChartTooltipExtra = z$1.infer<typeof evalChartTooltipExtraSchema>;
  */
 declare const evalChartConfigSchema: z$1.ZodObject<{
   heading: z$1.ZodOptional<z$1.ZodString>;
+  hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   type: z$1.ZodEnum<{
     area: "area";
     line: "line";
@@ -1275,6 +1296,7 @@ type EvalChartConfig = z$1.infer<typeof evalChartConfigSchema>;
  */
 declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
   heading: z$1.ZodOptional<z$1.ZodString>;
+  hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   type: z$1.ZodEnum<{
     area: "area";
     line: "line";
@@ -1549,6 +1571,82 @@ declare const removeDefaultConfigSchema: z$1.ZodUnion<readonly [z$1.ZodLiteral<t
 }>>]>;
 /** Removal config for built-in eval-level outputs and UI metadata. */
 type RemoveDefaultConfig = z$1.infer<typeof removeDefaultConfigSchema>;
+/** Single authored eval case with its stable identifier and input payload. */
+type EvalCase<TInput = unknown> = {
+  id: string;
+  input: TInput;
+  tags?: string[];
+};
+/** Query helpers built from the flattened trace recorded for one eval case. */
+type EvalTraceTree = {
+  spans: EvalTraceSpan[];
+  rootSpans: EvalTraceSpan[];
+  findSpan: (name: string) => EvalTraceSpan | undefined;
+  findSpansByKind: (kind: string) => EvalTraceSpan[];
+  flattenDfs: () => EvalTraceSpan[];
+  checkpoints: Map<string, unknown>;
+};
+/** Context passed to `deriveFromTracing` after execution has completed. */
+type EvalDeriveContext<TInput = unknown> = {
+  trace: EvalTraceTree;
+  input: TInput;
+  case: EvalCase<TInput>;
+};
+type MaybePromise<T> = T | Promise<T>;
+/** Function that derives one output value for a configured output key. */
+type EvalDeriveValueFn<TInput = unknown> = (ctx: EvalDeriveContext<TInput>) => MaybePromise<unknown>;
+/** Keyed `deriveFromTracing` config where each key derives one output value. */
+type EvalDeriveMap<TInput = unknown> = Record<string, EvalDeriveValueFn<TInput>>;
+/** Object-returning `deriveFromTracing` callback. */
+type EvalDeriveFn<TInput = unknown> = (ctx: EvalDeriveContext<TInput>) => Record<string, unknown> | Promise<Record<string, unknown>>;
+/** Trace-derived output config accepted globally and on eval definitions. */
+type EvalDeriveConfig<TInput = unknown> = EvalDeriveMap<TInput> | EvalDeriveFn<TInput>;
+/** Schema for keyed or object-returning trace-derived output config. */
+declare const evalDeriveConfigSchema: z$1.ZodType<EvalDeriveConfig>;
+/** UI overrides for a derived or scored column emitted by an eval. */
+type EvalColumnOverride = {
+  /** Display label shown for the column in tables and detail views. */label?: string;
+  /**
+   * Presentation preset for the value.
+   *
+   * Use this to control how the UI renders the cell and infer table behavior,
+   * for example `number`, `boolean`, `duration`, `markdown`, `json`, or
+   * file/media previews.
+   */
+  format?: ColumnFormat;
+  /**
+   * Extra options for `format: 'number'`.
+   *
+   * 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;
+  /**
+   * Hides the column from the runs table while keeping it available in detail
+   * views and raw output data.
+   */
+  hideInTable?: boolean;
+  /**
+   * Hides the column from the runs table when none of the rendered rows have a
+   * value. Missing values, `null`, and empty strings count as no value; `0` and
+   * `false` remain visible.
+   */
+  hideIfNoValue?: boolean; /** Horizontal alignment used when rendering the column cells. */
+  align?: 'left' | 'center' | 'right';
+  /**
+   * Maximum number of stars used when `format: 'stars'`.
+   *
+   * Values are still stored as normalized `0..1` numbers; the UI maps the
+   * selected star count evenly across that range.
+   */
+  maxStars?: number;
+};
+/** Column override map keyed by output or score field name. */
+type EvalColumns = Record<string, EvalColumnOverride>;
+/** Schema for UI overrides on derived or scored columns. */
+declare const evalColumnOverrideSchema: z$1.ZodType<EvalColumnOverride>;
+/** Schema for column override maps keyed by output or score field name. */
+declare const evalColumnsSchema: z$1.ZodType<EvalColumns>;
 /** Render formats supported by an LLM-call metric in the UI. */
 declare const llmCallMetricFormatSchema: z$1.ZodEnum<{
   string: "string";
@@ -1662,21 +1760,44 @@ declare const apiCallMetricSchema: z$1.ZodObject<{
 /** User-defined API-call metric authored in `agent-evals.config.ts`. */
 type ApiCallMetric = z$1.infer<typeof apiCallMetricSchema>;
 /**
- * Schema for one model/provider pricing entry used to derive LLM-call costs
- * from token counts.
+ * Schema for pricing rates used to derive LLM-call costs from token counts.
+ */
+declare const llmCallPricingRateSchema: z$1.ZodObject<{
+  inputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
+  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>;
+/** Token pricing rates authored in `agent-evals.config.ts`. */
+type LlmCallPricingRate = z$1.infer<typeof llmCallPricingRateSchema>;
+/**
+ * Schema for one model's pricing config. The object key is the exact model
+ * name. Use `providers` when a model has provider-specific rates in addition
+ * to, or instead of, generic model rates.
  */
 declare const llmCallPricingSchema: z$1.ZodObject<{
-  model: z$1.ZodString;
-  provider: z$1.ZodOptional<z$1.ZodString>;
   inputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
   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>;
+  provider: z$1.ZodOptional<z$1.ZodString>;
+  providers: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodObject<{
+    inputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
+    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>>>;
 }, z$1.core.$strip>;
-/** Model/provider pricing entry authored in `agent-evals.config.ts`. */
+/** Model pricing config authored in `agent-evals.config.ts`. */
 type LlmCallPricing = z$1.infer<typeof llmCallPricingSchema>;
+/** Model-keyed pricing registry authored in `agent-evals.config.ts`. */
+type LlmCallPricingRegistry = Record<string, LlmCallPricing>;
 /** Schema for the global LLM calls config block in `agent-evals.config.ts`. */
 declare const llmCallsConfigSchema: z$1.ZodObject<{
   kinds: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
@@ -1698,15 +1819,22 @@ declare const llmCallsConfigSchema: z$1.ZodObject<{
     toolCalls: z$1.ZodOptional<z$1.ZodString>;
   }, z$1.core.$strip>>;
   derivedAttributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodCustom<CallDerivedAttribute, CallDerivedAttribute>>>;
-  pricing: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
-    model: z$1.ZodString;
-    provider: z$1.ZodOptional<z$1.ZodString>;
+  pricing: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodObject<{
     inputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
     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>;
+    provider: z$1.ZodOptional<z$1.ZodString>;
+    providers: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodObject<{
+      inputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
+      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>>>;
   }, z$1.core.$strip>>>;
   metrics: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
     label: z$1.ZodString;
@@ -1855,7 +1983,7 @@ declare const DEFAULT_API_CALLS_CONFIG: ResolvedApiCallsConfig;
  * - Missing `metrics[].format` defaults to `'string'`.
  * - Missing `metrics[].placements` defaults to `['body']`.
  * - Missing `pricing` defaults to an empty registry; built-in costs are only
- *   derived from configured pricing and token counts.
+ *   derived from configured model-keyed pricing and token counts.
  */
 declare function resolveLlmCallsConfig(input: LlmCallsConfigInput | undefined): ResolvedLlmCallsConfig;
 /**
@@ -1904,12 +2032,35 @@ type AgentEvalsConfig = {
    * definition taking precedence for matching `key` or `path` entries.
    */
   traceDisplay?: TraceDisplayInputConfig;
+  /**
+   * Workspace-wide output columns applied to every eval.
+   *
+   * Eval-level `columns` with the same key take precedence. Built-in default
+   * columns are still added first unless removed with `removeDefaultConfig`.
+   */
+  columns?: EvalColumns;
+  /**
+   * Workspace-wide trace-derived outputs applied to every eval case.
+   *
+   * Prefer the keyed map form for shared metrics:
+   * `{ toolCalls: ({ trace }) => trace.findSpansByKind('tool').length }`.
+   * The object-returning function form is also supported. Derived outputs
+   * only fill keys that were not already recorded by eval execution.
+   */
+  deriveFromTracing?: EvalDeriveConfig;
+  /**
+   * Workspace-wide stats prepended to every eval's stats row.
+   *
+   * Eval-level stats render after these, and built-in default stats are
+   * appended last unless removed with `removeDefaultConfig`.
+   */
+  stats?: EvalStatsConfig;
   /**
    * Configuration for the "LLM calls" tab in the case-run drawer.
    *
    * Determines which trace spans are treated as LLM calls (`kinds`), how
    * structured fields like `model` and `usage.inputTokens` are read from
-   * span attributes, which pricing table derives built-in costs, and which
+   * span attributes, which pricing registry 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.
@@ -1924,10 +2075,13 @@ type AgentEvalsConfig = {
    *   metrics: [
    *     { label: 'Retries', path: 'retryCount', format: 'number' },
    *   ],
-   *   pricing: [
-   *     { model: 'gpt-4o-mini', provider: 'openai',
-   *       inputUsdPerMillion: 0.15, outputUsdPerMillion: 0.6 },
-   *   ],
+   *   pricing: {
+   *     'gpt-4o-mini': {
+   *       provider: 'openai',
+   *       inputUsdPerMillion: 0.15,
+   *       outputUsdPerMillion: 0.6,
+   *     },
+   *   },
    * }
    * ```
    */
@@ -2036,6 +2190,47 @@ declare const agentEvalsConfigSchema: z$1.ZodObject<{
       transform: z$1.ZodOptional<z$1.ZodCustom<TraceAttributeTransform, TraceAttributeTransform>>;
     }, z$1.core.$strip>>>;
   }, z$1.core.$strip>>;
+  columns: z$1.ZodOptional<z$1.ZodType<EvalColumns, unknown, z$1.core.$ZodTypeInternals<EvalColumns, unknown>>>;
+  deriveFromTracing: z$1.ZodOptional<z$1.ZodType<EvalDeriveConfig<unknown>, unknown, z$1.core.$ZodTypeInternals<EvalDeriveConfig<unknown>, unknown>>>;
+  stats: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
+    hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
+    kind: z$1.ZodLiteral<"cases">;
+  }, z$1.core.$strip>, z$1.ZodObject<{
+    hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
+    kind: z$1.ZodLiteral<"passRate">;
+    accent: z$1.ZodOptional<z$1.ZodBoolean>;
+  }, z$1.core.$strip>, z$1.ZodObject<{
+    hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
+    kind: z$1.ZodLiteral<"duration">;
+  }, z$1.core.$strip>, z$1.ZodObject<{
+    hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
+    kind: z$1.ZodLiteral<"column">;
+    key: z$1.ZodString;
+    label: z$1.ZodOptional<z$1.ZodString>;
+    aggregate: z$1.ZodEnum<{
+      last: "last";
+      sum: "sum";
+      avg: "avg";
+      min: "min";
+      max: "max";
+    }>;
+    format: z$1.ZodOptional<z$1.ZodEnum<{
+      number: "number";
+      boolean: "boolean";
+      file: "file";
+      duration: "duration";
+      json: "json";
+      markdown: "markdown";
+      image: "image";
+      audio: "audio";
+      video: "video";
+      percent: "percent";
+      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">>>;
   llmCalls: z$1.ZodOptional<z$1.ZodObject<{
     kinds: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
     attributes: z$1.ZodOptional<z$1.ZodObject<{
@@ -2056,15 +2251,22 @@ declare const agentEvalsConfigSchema: z$1.ZodObject<{
       toolCalls: z$1.ZodOptional<z$1.ZodString>;
     }, z$1.core.$strip>>;
     derivedAttributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodCustom<CallDerivedAttribute, CallDerivedAttribute>>>;
-    pricing: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
-      model: z$1.ZodString;
-      provider: z$1.ZodOptional<z$1.ZodString>;
+    pricing: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodObject<{
       inputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
       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>;
+      provider: z$1.ZodOptional<z$1.ZodString>;
+      providers: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodObject<{
+        inputUsdPerMillion: z$1.ZodOptional<z$1.ZodNumber>;
+        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>>>;
     }, z$1.core.$strip>>>;
     metrics: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
       label: z$1.ZodString;
@@ -2327,6 +2529,8 @@ declare const traceCacheRefSchema: z$1.ZodObject<{
     hit: "hit";
     miss: "miss";
   }>;
+  read: z$1.ZodOptional<z$1.ZodBoolean>;
+  stored: z$1.ZodOptional<z$1.ZodBoolean>;
   storedAt: z$1.ZodOptional<z$1.ZodString>;
   age: z$1.ZodOptional<z$1.ZodNumber>;
 }, z$1.core.$strip>;
@@ -2344,7 +2548,6 @@ declare const cacheListItemSchema: z$1.ZodObject<{
   spanName: z$1.ZodOptional<z$1.ZodString>;
   spanKind: z$1.ZodOptional<z$1.ZodString>;
   storedAt: z$1.ZodString;
-  codeFingerprint: z$1.ZodString;
   sizeBytes: z$1.ZodNumber;
 }, z$1.core.$strip>;
 /** Summary row for a single cache entry. */
@@ -2469,7 +2672,6 @@ declare const cacheEntrySchema: z$1.ZodObject<{
   spanName: z$1.ZodOptional<z$1.ZodString>;
   spanKind: z$1.ZodOptional<z$1.ZodString>;
   storedAt: z$1.ZodString;
-  codeFingerprint: z$1.ZodString;
   recording: z$1.ZodObject<{
     returnValue: z$1.ZodUnknown;
     finalAttributes: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
@@ -2542,7 +2744,6 @@ declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
   }>;
   operationName: z$1.ZodString;
   storedAt: z$1.ZodString;
-  codeFingerprint: z$1.ZodString;
   rawKey: z$1.ZodUnknown;
 }, z$1.core.$strip>;
 /** Debug-only raw cache key entry. May contain sensitive prompt/input data. */
@@ -2560,7 +2761,6 @@ declare const cacheEntryWithDebugKeySchema: z$1.ZodObject<{
   spanName: z$1.ZodOptional<z$1.ZodString>;
   spanKind: z$1.ZodOptional<z$1.ZodString>;
   storedAt: z$1.ZodString;
-  codeFingerprint: z$1.ZodString;
   recording: z$1.ZodObject<{
     returnValue: z$1.ZodUnknown;
     finalAttributes: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
@@ -2629,7 +2829,6 @@ declare const cacheEntryWithDebugKeySchema: z$1.ZodObject<{
     }>;
     operationName: z$1.ZodString;
     storedAt: z$1.ZodString;
-    codeFingerprint: z$1.ZodString;
     rawKey: z$1.ZodUnknown;
   }, z$1.core.$strip>>;
 }, z$1.core.$strip>;
@@ -2651,7 +2850,6 @@ declare const cacheFileSchema: z$1.ZodObject<{
     spanName: z$1.ZodOptional<z$1.ZodString>;
     spanKind: z$1.ZodOptional<z$1.ZodString>;
     storedAt: z$1.ZodString;
-    codeFingerprint: z$1.ZodString;
     recording: z$1.ZodObject<{
       returnValue: z$1.ZodUnknown;
       finalAttributes: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
@@ -2728,7 +2926,6 @@ declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
     }>;
     operationName: z$1.ZodString;
     storedAt: z$1.ZodString;
-    codeFingerprint: z$1.ZodString;
     rawKey: z$1.ZodUnknown;
   }, z$1.core.$strip>>;
 }, z$1.core.$strip>;
@@ -2742,16 +2939,19 @@ type CacheDebugKeyFile = z$1.infer<typeof cacheDebugKeyFileSchema>;
  *
  * `action === 'hit'` rows reused an existing persisted cache entry.
  * `action === 'added'` rows came from a miss or refresh that wrote a persisted
- * cache entry during the run. `origin === 'caseRoot'` rows came from
- * `evalTracer.cache(...)` calls made directly from the case body (no
- * surrounding `traceSpan`), which would otherwise be invisible.
+ * cache entry during the run. `action === 'notStored'` rows executed a cached
+ * operation but did not persist it because storage was disabled for that eval
+ * scope. `origin === 'caseRoot'` rows came from `evalTracer.cache(...)` calls
+ * made directly from the case body (no surrounding `traceSpan`), which would
+ * otherwise be invisible.
  */
 type CacheActivityEntry = {
   id: string;
   source: 'span' | 'value';
   origin: 'span' | 'caseRoot';
-  action: 'hit' | 'added';
+  action: 'hit' | 'added' | 'notStored';
   status: 'hit' | 'miss' | 'refresh';
+  stored: boolean;
   name: string;
   namespace: string;
   key: string;
@@ -2840,12 +3040,6 @@ declare const updateManualScoreRequestSchema: z$1.ZodObject<{
 type UpdateManualScoreRequest = z$1.infer<typeof updateManualScoreRequestSchema>;
 //#endregion
 //#region ../sdk/src/types.d.ts
-/** Single authored eval case with its stable identifier and input payload. */
-type EvalCase<TInput> = {
-  id: string;
-  input: TInput;
-  tags?: string[];
-};
 /** Runtime output values collected from output helpers and `deriveFromTracing`. */
 type EvalOutputs = Record<string, unknown>;
 /**
@@ -2861,48 +3055,22 @@ type EvalStartTime = Date | number | string;
  * `deriveFromTracing` finish, before computed scores run.
  */
 type EvalOutputsSchema<TOutputs extends EvalOutputs> = z$1.ZodType<TOutputs>;
-/** UI overrides for a derived or scored column emitted by an eval. */
-type EvalColumnOverride = {
-  /** Display label shown for the column in tables and detail views. */label?: string;
+/** Per-eval controls for SDK operation caching. */
+type EvalCacheConfig = {
   /**
-   * Presentation preset for the value.
+   * Whether cached spans and value caches may read existing persisted entries.
    *
-   * Use this to control how the UI renders the cell and infer table behavior,
-   * for example `number`, `boolean`, `duration`, `markdown`, `json`, or
-   * file/media previews.
+   * Defaults to `true`. Set to `false` when this eval should always execute
+   * cached operations instead of replaying previous results.
    */
-  format?: ColumnFormat;
+  read?: boolean;
   /**
-   * Extra options for `format: 'number'`.
+   * Whether cached spans and value caches may persist entries after execution.
    *
-   * Use this to add a prefix or suffix, control minimum and maximum decimal
-   * places, or switch to compact notation such as `1.2K`.
+   * Defaults to `true`. Set to `false` when this eval may reuse existing cache
+   * entries but must not create or refresh stored cache files.
    */
-  numberFormat?: NumberDisplayOptions;
-  /**
-   * Hides the column from the runs table while keeping it available in detail
-   * views and raw output data.
-   */
-  hideInTable?: boolean; /** Horizontal alignment used when rendering the column cells. */
-  align?: 'left' | 'center' | 'right';
-  /**
-   * Maximum number of stars used when `format: 'stars'`.
-   *
-   * Values are still stored as normalized `0..1` numbers; the UI maps the
-   * selected star count evenly across that range.
-   */
-  maxStars?: number;
-};
-/** Column override map keyed by output or score field name. */
-type EvalColumns = Record<string, EvalColumnOverride>;
-/** Query helpers built from the flattened trace recorded for one eval case. */
-type EvalTraceTree = {
-  spans: EvalTraceSpan[];
-  rootSpans: EvalTraceSpan[];
-  findSpan: (name: string) => EvalTraceSpan | undefined;
-  findSpansByKind: (kind: string) => EvalTraceSpan[];
-  flattenDfs: () => EvalTraceSpan[];
-  checkpoints: Map<string, unknown>;
+  store?: boolean;
 };
 /** Type-safe output writer passed to an eval's `execute` function. */
 type EvalSetOutput<TOutputs extends EvalOutputs = EvalOutputs> = <TKey extends Extract<keyof TOutputs, string>>(
@@ -2930,12 +3098,6 @@ type EvalExecuteContext<TInput, TOutputs extends EvalOutputs = EvalOutputs> = {
    */
   setOutput: EvalSetOutput<TOutputs>;
 };
-/** Context passed to `deriveFromTracing` after execution has completed. */
-type EvalDeriveContext<TInput> = {
-  trace: EvalTraceTree;
-  input: TInput;
-  case: EvalCase<TInput>;
-};
 /** Context passed to score functions after outputs have been collected. */
 type EvalScoreContext<TInput, TOutputs extends EvalOutputs = EvalOutputs> = {
   input: TInput;
@@ -2995,8 +3157,27 @@ type EvalDefinitionOutputSchemaConfig<TOutputs extends EvalOutputs> = [EvalOutpu
   outputsSchema: EvalOutputsSchema<TOutputs>;
 };
 type EvalDefinitionBase<TInput = unknown, TOutputs extends EvalOutputs = EvalOutputs> = {
+  /**
+   * Stable eval identifier within the authored eval file.
+   *
+   * The runner combines this value with the workspace-relative file path to
+   * form the eval key used for targeting, persisted runs, and UI navigation.
+   */
   id: string;
+  /**
+   * Human-readable eval name shown in the CLI and web UI.
+   *
+   * When omitted, consumers fall back to `id`.
+   */
   title?: string;
+  /**
+   * Per-eval cache controls. Both `read` and `store` default to `true`.
+   *
+   * `read: false` skips cache lookups for this eval. `store: false` prevents
+   * new or refreshed entries from being written while still allowing reads
+   * unless `read` is also disabled.
+   */
+  cache?: EvalCacheConfig;
   /**
    * Authored cases for this eval.
    *
@@ -3004,6 +3185,13 @@ type EvalDefinitionBase<TInput = unknown, TOutputs extends EvalOutputs = EvalOut
    * eval once using a synthetic case with empty object input.
    */
   cases?: EvalCase<TInput>[] | (() => Promise<EvalCase<TInput>[]>);
+  /**
+   * Output and score column display overrides for this eval.
+   *
+   * Use this to label, format, group, hide, or otherwise customize columns
+   * produced by default config, output helpers, `deriveFromTracing`, scores,
+   * or manual scores.
+   */
   columns?: EvalColumns;
   /**
    * Per-eval trace attribute display rules for the UI.
@@ -3038,8 +3226,30 @@ type EvalDefinitionBase<TInput = unknown, TOutputs extends EvalOutputs = EvalOut
    * elapsed time from the configured `startTime`.
    */
   freezeTime?: boolean;
+  /**
+   * Run one eval case.
+   *
+   * The callback receives the authored case input and a typed `setOutput`
+   * helper. It may record outputs, run assertions, start traced work, and
+   * return either synchronously or asynchronously. Thrown errors fail the
+   * active case and skip later computed scores for that case.
+   */
   execute: (ctx: EvalExecuteContext<TInput, TOutputs>) => Promise<void> | void;
-  deriveFromTracing?: (ctx: EvalDeriveContext<TInput>) => Partial<TOutputs> | Promise<Partial<TOutputs>>;
+  /**
+   * Derive additional output fields from the case trace after `execute`.
+   *
+   * Prefer the keyed map form when each key has one derivation. The
+   * object-returning callback form is also supported. Derived values only fill
+   * keys not already recorded during execution.
+   */
+  deriveFromTracing?: EvalDeriveConfig<TInput>;
+  /**
+   * Computed score columns for each case.
+   *
+   * Each key becomes a persisted score column. A score can be a bare callback
+   * or an object with UI metadata and an optional `passThreshold`; thresholds
+   * fail a case only when the computed value is strictly below the threshold.
+   */
   scores?: Record<string, EvalScoreDef<TInput, TOutputs>>;
   /**
    * Score columns whose values are entered in the web UI after a run.
@@ -3061,8 +3271,8 @@ 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`, `format`, and `numberFormat` default to the matching
-   * `ColumnDef`.
+   * an em dash, or hides when `hideIfNoValue` is true. `label`, `format`, and
+   * `numberFormat` default to the matching `ColumnDef`.
    */
   stats?: EvalStatsConfig;
   /**
@@ -3077,7 +3287,8 @@ type EvalDefinitionBase<TInput = unknown, TOutputs extends EvalOutputs = EvalOut
    * the run summary. Column metrics aggregate a score or numeric output column
    * across the run using an `aggregate` reducer (`avg`, `sum`, `min`, `max`,
    * `latest`, `passThresholdRate`). `passThresholdRate` requires a score column
-   * with `passThreshold`.
+   * with `passThreshold`. Set `hideIfNoValue` to hide a chart until at least
+   * one metric has a numeric value in the rendered history window.
    */
   charts?: EvalChartsConfig;
   /**
@@ -3176,7 +3387,6 @@ type CacheDebugKeyWrite = {
   rawKey: unknown;
   operationType: CacheOperationType;
   operationName: string;
-  codeFingerprint: string;
 };
 /**
  * Adapter used by the SDK to read and write cache entries.
@@ -3199,8 +3409,21 @@ type CacheAdapter = {
 type CacheScopeContext = {
   adapter: CacheAdapter;
   mode: CacheMode;
-  evalId: string; /** Hash of the eval source file, stored as cache metadata for inspection. */
-  codeFingerprint: string;
+  evalId: string;
+  /**
+   * Whether cache lookups are allowed for this eval scope. Defaults to `true`.
+   *
+   * Run-level `bypass` and `refresh` modes still take precedence and skip
+   * reads even when this is enabled.
+   */
+  read?: boolean;
+  /**
+   * Whether cache writes are allowed for this eval scope. Defaults to `true`.
+   *
+   * Run-level `bypass` still takes precedence and skips writes even when this
+   * is enabled.
+   */
+  store?: boolean;
 };
 /** Active recording frame captured while a cached operation body executes. */
 type CacheRecordingFrame = {
@@ -3397,7 +3620,7 @@ declare function incrementEvalOutput(key: string, delta: number): void;
 declare function evalAssert(condition: unknown, message: string): asserts condition;
 //#endregion
 //#region ../sdk/src/cacheSerialization.d.ts
-declare const serializedCacheValueMarker = "__agentEvalsCacheSerialization";
+declare const serializedCacheValueMarker = "__aecs";
 declare const jsonSafeCacheValueVersion = "json-safe-v1";
 type JsonSafeCacheValueType = 'ArrayBuffer' | 'BigInt' | 'Blob' | 'CompressedJson' | 'CompressedString' | 'Date' | 'Error' | 'File' | 'Float64Array' | 'Headers' | 'Map' | 'Number' | 'Object' | 'RegExp' | 'Set' | 'URL' | 'URLSearchParams' | 'Undefined';
 type JsonSafeSerializedCacheValue = {
@@ -3409,16 +3632,32 @@ type JsonSafeSerializedCacheValue = {
 };
 /** JSON-safe persisted representation for one rich cached value. */
 type SerializedCacheValue = JsonSafeSerializedCacheValue;
+/** Options controlling how rich cache values are persisted as JSON-safe data. */
+type CacheSerializationOptions = {
+  /**
+   * Preserve JavaScript `undefined` values with explicit tagged wrappers.
+   *
+   * Disabled by default so undefined object fields, array items, map entries,
+   * and set items are omitted instead of being written to cache files.
+   */
+  preserveUndefined?: boolean;
+};
 /**
  * Serialize one cached value while keeping plain JSON as plain JSON.
  *
- * Rich runtime values use small tagged wrappers.
+ * Rich runtime values use small tagged wrappers. Undefined values are omitted
+ * by default; pass `preserveUndefined: true` to round-trip them explicitly.
  */
-declare function serializeCacheValue(value: unknown): Promise<unknown>;
+declare function serializeCacheValue(value: unknown, options?: CacheSerializationOptions | undefined): Promise<unknown>;
 /** Revive one cached value, while preserving legacy JSON-round-tripped data. */
 declare function deserializeCacheValue(value: unknown): unknown;
-/** Serialize all rich values captured in a cache recording before persistence. */
-declare function serializeCacheRecording(recording: CacheRecording): Promise<CacheRecording>;
+/**
+ * Serialize all rich values captured in a cache recording before persistence.
+ *
+ * Undefined values are omitted by default; pass `preserveUndefined: true` to
+ * retain the legacy explicit undefined wrappers in the recording payload.
+ */
+declare function serializeCacheRecording(recording: CacheRecording, options?: CacheSerializationOptions | undefined): Promise<CacheRecording>;
 /** Revive all rich values captured in a cache recording after lookup. */
 declare function deserializeCacheRecording(recording: CacheRecording): CacheRecording;
 //#endregion
@@ -3637,6 +3876,15 @@ type CacheClearFilter = {
   key?: string;
 };
 //#endregion
+//#region ../runner/src/recalculateDerivedAttributes.d.ts
+type RecalculateDerivedAttributesResult = {
+  updated: true;
+  caseDetail: CaseDetail;
+} | {
+  updated: false;
+  reason: string;
+};
+//#endregion
 //#region ../runner/src/runner.d.ts
 /** Imperative runner interface used by the server and CLI. */
 type EvalRunner = {
@@ -3703,7 +3951,11 @@ type EvalRunner = {
    */
   recomputeStatusesForEval(evalKey: string): Promise<{
     updatedRuns: number;
-  }>;
+  }>; /** Recalculate configured LLM/API derived attributes for one persisted case trace. */
+  recalculateDerivedAttributesForCase(params: {
+    runId: string;
+    caseId: string;
+  }): Promise<RecalculateDerivedAttributesResult>;
   /**
    * Delete terminal persisted runs that touch one eval from memory and disk.
    * Accepts the exact eval key, with a legacy fallback for unique eval ids.
@@ -3754,4 +4006,4 @@ 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 CallDerivedAttribute, type CallDerivedAttributeContext, 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, type DiscoveryIssue, 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 EvalExpectation, 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 ResolvedCallDerivedAttribute, 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, applyDerivedCallAttributes, assertionFailureSchema, buildCaseKey, buildEvalKey, 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, discoveryIssueSchema, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalExpect, evalFreshnessStatusSchema, evalLog, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, fileRefSchema, getCaseRowCaseKey, getCaseRowEvalKey, 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 };
+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 CacheSerializationOptions, type CacheStatus, type CallDerivedAttribute, type CallDerivedAttributeContext, 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, type DiscoveryIssue, EvalAssertionError, type EvalCacheConfig, 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 EvalDeriveConfig, type EvalDeriveContext, type EvalDeriveFn, type EvalDeriveMap, type EvalDeriveValueFn, type EvalDisplayStatus, type EvalExecuteContext, type EvalExpectation, 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 LlmCallPricingRate, type LlmCallPricingRegistry, type LlmCallsConfigInput, type NumberDisplayOptions, type RemoveDefaultConfig, type RepoFileRef, type ResolvedApiCallMetric, type ResolvedApiCallsConfig, type ResolvedCallDerivedAttribute, 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, applyDerivedCallAttributes, assertionFailureSchema, buildCaseKey, buildEvalKey, 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, discoveryIssueSchema, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalColumnOverrideSchema, evalColumnsSchema, evalDeriveConfigSchema, evalExpect, evalFreshnessStatusSchema, evalLog, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, fileRefSchema, getCaseRowCaseKey, getCaseRowEvalKey, getCurrentScope, getEvalCaseInput, getEvalDisplayStatus, getEvalRegistry, getEvalStartTime, getEvalTitle, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, llmCallMetricFormatSchema, llmCallMetricPlacementSchema, llmCallMetricSchema, llmCallPricingRateSchema, 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 };