@ls-stack/agent-eval 0.53.0 → 0.54.0

package/dist/index.d.mts

@@ -144,6 +144,25 @@ declare const traceSpanSchema$2: z$1.ZodObject<{
 /** Persisted trace span shape stored for each eval case run. */
 type EvalTraceSpan$2 = z$1.infer<typeof traceSpanSchema$2>; //#endregion
 //#region ../shared/src/schemas/eval.d.ts
+/**
+ * Reducer used to collapse per-case values into a single duration or column
+ * stat.
+ * `best` selects the highest finite value and `worst` selects the lowest.
+ */
+declare const evalStatAggregateSchema$1: z$1.ZodEnum<{
+  avg: "avg";
+  min: "min";
+  max: "max";
+  sum: "sum";
+  best: "best";
+  worst: "worst";
+}>;
+/**
+ * Reducer used to collapse per-case values into a single duration or column
+ * stat.
+ * `best` selects the highest finite value and `worst` selects the lowest.
+ */
+type EvalStatAggregate$1 = z$1.infer<typeof evalStatAggregateSchema$1>;
 /** Ordered list of stats rendered in the EvalCard stats row. */
 declare const evalStatsConfigSchema$1: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
@@ -155,9 +174,25 @@ declare const evalStatsConfigSchema$1: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z
 }, z$1.core.$strip>, z$1.ZodObject<{
   hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"duration">;
+  aggregate: z$1.ZodOptional<z$1.ZodEnum<{
+    avg: "avg";
+    min: "min";
+    max: "max";
+    sum: "sum";
+    best: "best";
+    worst: "worst";
+  }>>;
 }, z$1.core.$strip>, z$1.ZodObject<{
   hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"cacheHits">;
+  aggregate: z$1.ZodOptional<z$1.ZodEnum<{
+    avg: "avg";
+    min: "min";
+    max: "max";
+    sum: "sum";
+    best: "best";
+    worst: "worst";
+  }>>;
 }, z$1.core.$strip>, z$1.ZodObject<{
   hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"column">;
@@ -168,7 +203,8 @@ declare const evalStatsConfigSchema$1: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z
     min: "min";
     max: "max";
     sum: "sum";
-    last: "last";
+    best: "best";
+    worst: "worst";
   }>;
   format: z$1.ZodOptional<z$1.ZodEnum<{
     number: "number";
@@ -1022,16 +1058,28 @@ type EvalDefinitionBase<TInput = unknown, TOutputs extends EvalOutputs = EvalOut
    * When provided, the stats render in order, left to right.
    *
    * Built-in kinds (`cases`, `passRate`, `duration`, `cacheHits`) read from
-   * the latest run summary. `cacheHits` counts Agent Eval operation-level cache
-   * hits over total cache operations, not LLM provider prompt-cache read
-   * tokens. `kind: 'column'` aggregates a score or numeric output 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, or hides
-   * when `hideIfNoValue` is true. `label`, `format`, and `numberFormat` default
-   * to the matching `ColumnDef`.
+   * the latest run. `duration` aggregates finite per-case durations using the
+   * same modes as column stats. `cacheHits` counts Agent Eval operation-level
+   * cache hits over total cache operations, not LLM provider prompt-cache read
+   * tokens. Cache-hit stats have their own aggregate mode and default to `sum`;
+   * `avg` is average per-case hit rate, and min/max/best/worst select cases by
+   * hit rate. `kind: 'column'` aggregates a score or numeric output 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, or
+   * hides when `hideIfNoValue` is true. `label`, `format`, and `numberFormat`
+   * default to the matching `ColumnDef`.
    */
   stats?: EvalStatsConfig$1;
+  /**
+   * Initial aggregate mode used for this eval's duration and column stats in
+   * the web UI.
+   *
+   * Overrides `AgentEvalsConfig.defaultStatAggregate`. Individual stat
+   * `aggregate` values still define their authored reducer and remain the
+   * fallback when neither default is configured.
+   */
+  defaultStatAggregate?: EvalStatAggregate$1;
   /**
    * Optional history chart configuration for the EvalCard in the web UI.
    *
@@ -1083,8 +1131,12 @@ type EvalRegistryEntry = {
 /** Return the in-memory registry of evals defined in the current process. */
 declare function getEvalRegistry(): Map<string, EvalRegistryEntry>;
 /**
- * Register an eval definition with the SDK so the runner can discover it
- * after importing the eval module.
+ * Execute a callback with an empty async-local eval registry.
+ *
+ * Runner internals use this when importing eval modules concurrently so
+ * `defineEval(...)` calls from one import cannot overwrite another import's
+ * registered definitions. The callback receives the scoped registry populated
+ * during its async execution.
  */
 //#endregion
 //#region src/evalExpect.d.ts
@@ -1729,14 +1781,14 @@ declare const columnFormatSchema: z$1.ZodEnum<{
   number: "number";
   boolean: "boolean";
   file: "file";
-  markdown: "markdown";
   json: "json";
+  duration: "duration";
+  markdown: "markdown";
   image: "image";
   html: "html";
   pdf: "pdf";
   audio: "audio";
   video: "video";
-  duration: "duration";
   percent: "percent";
   passFail: "passFail";
   stars: "stars";
@@ -1756,14 +1808,14 @@ declare const columnDefSchema: z$1.ZodObject<{
     number: "number";
     boolean: "boolean";
     file: "file";
-    markdown: "markdown";
     json: "json";
+    duration: "duration";
+    markdown: "markdown";
     image: "image";
     html: "html";
     pdf: "pdf";
     audio: "audio";
     video: "video";
-    duration: "duration";
     percent: "percent";
     passFail: "passFail";
     stars: "stars";
@@ -1777,8 +1829,8 @@ declare const columnDefSchema: z$1.ZodObject<{
   hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   align: z$1.ZodOptional<z$1.ZodEnum<{
     left: "left";
-    center: "center";
     right: "right";
+    center: "center";
   }>>;
 }, z$1.core.$strip>;
 /** Column definition exposed to the UI for eval and case tables. */
@@ -1848,9 +1900,9 @@ declare const traceAttributeDisplaySchema: z$1.ZodObject<{
     subtree: "subtree";
   }>>;
   mode: z$1.ZodOptional<z$1.ZodEnum<{
+    all: "all";
     sum: "sum";
     last: "last";
-    all: "all";
   }>>;
 }, z$1.core.$strip>;
 /**
@@ -1884,9 +1936,9 @@ declare const traceDisplayConfigSchema: z$1.ZodObject<{
       subtree: "subtree";
     }>>;
     mode: z$1.ZodOptional<z$1.ZodEnum<{
+      all: "all";
       sum: "sum";
       last: "last";
-      all: "all";
     }>>;
   }, z$1.core.$strip>>>;
 }, z$1.core.$strip>;
@@ -1924,9 +1976,9 @@ declare const traceAttributeDisplayInputSchema: z$1.ZodObject<{
     subtree: "subtree";
   }>>;
   mode: z$1.ZodOptional<z$1.ZodEnum<{
+    all: "all";
     sum: "sum";
     last: "last";
-    all: "all";
   }>>;
   transform: z$1.ZodOptional<z$1.ZodCustom<TraceAttributeTransform, TraceAttributeTransform>>;
 }, z$1.core.$strip>;
@@ -1962,9 +2014,9 @@ declare const traceDisplayInputConfigSchema: z$1.ZodObject<{
       subtree: "subtree";
     }>>;
     mode: z$1.ZodOptional<z$1.ZodEnum<{
+      all: "all";
       sum: "sum";
       last: "last";
-      all: "all";
     }>>;
     transform: z$1.ZodOptional<z$1.ZodCustom<TraceAttributeTransform, TraceAttributeTransform>>;
   }, z$1.core.$strip>>>;
@@ -1999,9 +2051,9 @@ declare const traceSpanSchema$1: z$1.ZodObject<{
   startedAt: z$1.ZodString;
   endedAt: z$1.ZodNullable<z$1.ZodString>;
   status: z$1.ZodEnum<{
-    error: "error";
     running: "running";
     cancelled: "cancelled";
+    error: "error";
     ok: "ok";
   }>;
   attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
@@ -2041,21 +2093,32 @@ declare const evalFreshnessStatusSchema: z$1.ZodEnum<{
 }>;
 /** Freshness signal derived from the latest relevant run plus git state. */
 type EvalFreshnessStatus = z$1.infer<typeof evalFreshnessStatusSchema>;
-/** Reducer used to collapse a column's per-case values into a single stat. */
+/**
+ * Reducer used to collapse per-case values into a single duration or column
+ * stat.
+ * `best` selects the highest finite value and `worst` selects the lowest.
+ */
 declare const evalStatAggregateSchema: z$1.ZodEnum<{
   avg: "avg";
+  sum: "sum";
   min: "min";
   max: "max";
-  sum: "sum";
-  last: "last";
+  best: "best";
+  worst: "worst";
 }>;
-/** Reducer used to collapse a column's per-case values into a single stat. */
+/**
+ * Reducer used to collapse per-case values into a single duration or column
+ * stat.
+ * `best` selects the highest finite value and `worst` selects the lowest.
+ */
 type EvalStatAggregate = z$1.infer<typeof evalStatAggregateSchema>;
 /**
- * One entry in the EvalCard stats row. Built-in kinds use latest run totals;
- * `cacheHits` counts Agent Eval operation-level cache hits from spans and
- * `evalTracer.cache(...)` refs, not LLM provider prompt-cache read tokens.
- * `column` aggregates a score or numeric output column across the latest run.
+ * One entry in the EvalCard stats row. Built-in kinds read from the latest run;
+ * `duration` aggregates per-case durations, `cacheHits` counts Agent Eval
+ * operation-level cache hits from spans and `evalTracer.cache(...)` refs, not
+ * LLM provider prompt-cache read tokens. Cache hits use an independent
+ * aggregate mode and default to `sum`. `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>;
@@ -2067,9 +2130,25 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
 }, z$1.core.$strip>, z$1.ZodObject<{
   hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"duration">;
+  aggregate: z$1.ZodOptional<z$1.ZodEnum<{
+    avg: "avg";
+    sum: "sum";
+    min: "min";
+    max: "max";
+    best: "best";
+    worst: "worst";
+  }>>;
 }, z$1.core.$strip>, z$1.ZodObject<{
   hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"cacheHits">;
+  aggregate: z$1.ZodOptional<z$1.ZodEnum<{
+    avg: "avg";
+    sum: "sum";
+    min: "min";
+    max: "max";
+    best: "best";
+    worst: "worst";
+  }>>;
 }, z$1.core.$strip>, z$1.ZodObject<{
   hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"column">;
@@ -2077,23 +2156,24 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   label: z$1.ZodOptional<z$1.ZodString>;
   aggregate: z$1.ZodEnum<{
     avg: "avg";
+    sum: "sum";
     min: "min";
     max: "max";
-    sum: "sum";
-    last: "last";
+    best: "best";
+    worst: "worst";
   }>;
   format: z$1.ZodOptional<z$1.ZodEnum<{
     number: "number";
     boolean: "boolean";
     file: "file";
-    markdown: "markdown";
     json: "json";
+    duration: "duration";
+    markdown: "markdown";
     image: "image";
     html: "html";
     pdf: "pdf";
     audio: "audio";
     video: "video";
-    duration: "duration";
     percent: "percent";
     passFail: "passFail";
     stars: "stars";
@@ -2114,9 +2194,25 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
 }, z$1.core.$strip>, z$1.ZodObject<{
   hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"duration">;
+  aggregate: z$1.ZodOptional<z$1.ZodEnum<{
+    avg: "avg";
+    sum: "sum";
+    min: "min";
+    max: "max";
+    best: "best";
+    worst: "worst";
+  }>>;
 }, z$1.core.$strip>, z$1.ZodObject<{
   hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"cacheHits">;
+  aggregate: z$1.ZodOptional<z$1.ZodEnum<{
+    avg: "avg";
+    sum: "sum";
+    min: "min";
+    max: "max";
+    best: "best";
+    worst: "worst";
+  }>>;
 }, z$1.core.$strip>, z$1.ZodObject<{
   hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"column">;
@@ -2124,23 +2220,24 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
   label: z$1.ZodOptional<z$1.ZodString>;
   aggregate: z$1.ZodEnum<{
     avg: "avg";
+    sum: "sum";
     min: "min";
     max: "max";
-    sum: "sum";
-    last: "last";
+    best: "best";
+    worst: "worst";
   }>;
   format: z$1.ZodOptional<z$1.ZodEnum<{
     number: "number";
     boolean: "boolean";
     file: "file";
-    markdown: "markdown";
     json: "json";
+    duration: "duration";
+    markdown: "markdown";
     image: "image";
     html: "html";
     pdf: "pdf";
     audio: "audio";
     video: "video";
-    duration: "duration";
     percent: "percent";
     passFail: "passFail";
     stars: "stars";
@@ -2179,14 +2276,14 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
       number: "number";
       boolean: "boolean";
       file: "file";
-      markdown: "markdown";
       json: "json";
+      duration: "duration";
+      markdown: "markdown";
       image: "image";
       html: "html";
       pdf: "pdf";
       audio: "audio";
       video: "video";
-      duration: "duration";
       percent: "percent";
       passFail: "passFail";
       stars: "stars";
@@ -2200,18 +2297,18 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
     hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
     align: z$1.ZodOptional<z$1.ZodEnum<{
       left: "left";
-      center: "center";
       right: "right";
+      center: "center";
     }>>;
   }, z$1.core.$strip>>;
   caseCount: z$1.ZodNullable<z$1.ZodNumber>;
   caseIds: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
   lastRunStatus: z$1.ZodNullable<z$1.ZodEnum<{
+    running: "running";
+    cancelled: "cancelled";
     error: "error";
     pass: "pass";
     fail: "fail";
-    running: "running";
-    cancelled: "cancelled";
     unscored: "unscored";
   }>>;
   stats: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
@@ -2224,9 +2321,25 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
   }, z$1.core.$strip>, z$1.ZodObject<{
     hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
     kind: z$1.ZodLiteral<"duration">;
+    aggregate: z$1.ZodOptional<z$1.ZodEnum<{
+      avg: "avg";
+      sum: "sum";
+      min: "min";
+      max: "max";
+      best: "best";
+      worst: "worst";
+    }>>;
   }, z$1.core.$strip>, z$1.ZodObject<{
     hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
     kind: z$1.ZodLiteral<"cacheHits">;
+    aggregate: z$1.ZodOptional<z$1.ZodEnum<{
+      avg: "avg";
+      sum: "sum";
+      min: "min";
+      max: "max";
+      best: "best";
+      worst: "worst";
+    }>>;
   }, z$1.core.$strip>, z$1.ZodObject<{
     hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
     kind: z$1.ZodLiteral<"column">;
@@ -2234,23 +2347,24 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
     label: z$1.ZodOptional<z$1.ZodString>;
     aggregate: z$1.ZodEnum<{
       avg: "avg";
+      sum: "sum";
       min: "min";
       max: "max";
-      sum: "sum";
-      last: "last";
+      best: "best";
+      worst: "worst";
     }>;
     format: z$1.ZodOptional<z$1.ZodEnum<{
       number: "number";
       boolean: "boolean";
       file: "file";
-      markdown: "markdown";
       json: "json";
+      duration: "duration";
+      markdown: "markdown";
       image: "image";
       html: "html";
       pdf: "pdf";
       audio: "audio";
       video: "video";
-      duration: "duration";
       percent: "percent";
       passFail: "passFail";
       stars: "stars";
@@ -2258,6 +2372,14 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
     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">>>;
+  defaultStatAggregate: z$1.ZodOptional<z$1.ZodEnum<{
+    avg: "avg";
+    sum: "sum";
+    min: "min";
+    max: "max";
+    best: "best";
+    worst: "worst";
+  }>>;
   charts: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
     heading: z$1.ZodOptional<z$1.ZodString>;
     hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
@@ -2275,8 +2397,8 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
       }>;
       label: z$1.ZodOptional<z$1.ZodString>;
       color: z$1.ZodOptional<z$1.ZodEnum<{
-        success: "success";
         error: "error";
+        success: "success";
         accent: "accent";
         accentDim: "accentDim";
         warning: "warning";
@@ -2291,16 +2413,16 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
       key: z$1.ZodString;
       aggregate: z$1.ZodEnum<{
         avg: "avg";
+        sum: "sum";
         min: "min";
         max: "max";
-        sum: "sum";
         latest: "latest";
         passThresholdRate: "passThresholdRate";
       }>;
       label: z$1.ZodOptional<z$1.ZodString>;
       color: z$1.ZodOptional<z$1.ZodEnum<{
-        success: "success";
         error: "error";
+        success: "success";
         accent: "accent";
         accentDim: "accentDim";
         warning: "warning";
@@ -2333,9 +2455,9 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
       key: z$1.ZodString;
       aggregate: z$1.ZodEnum<{
         avg: "avg";
+        sum: "sum";
         min: "min";
         max: "max";
-        sum: "sum";
         latest: "latest";
         passThresholdRate: "passThresholdRate";
       }>;
@@ -2431,12 +2553,12 @@ declare const caseRowSchema$1: z$1.ZodObject<{
   evalId: z$1.ZodString;
   tags: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
   status: z$1.ZodEnum<{
+    pending: "pending";
+    running: "running";
+    cancelled: "cancelled";
     error: "error";
     pass: "pass";
     fail: "fail";
-    running: "running";
-    cancelled: "cancelled";
-    pending: "pending";
   }>;
   durationMs: z$1.ZodNullable<z$1.ZodNumber>;
   cacheHits: z$1.ZodOptional<z$1.ZodNumber>;
@@ -2532,9 +2654,9 @@ declare const scoreTraceSchema: z$1.ZodObject<{
     startedAt: z$1.ZodString;
     endedAt: z$1.ZodNullable<z$1.ZodString>;
     status: z$1.ZodEnum<{
-      error: "error";
       running: "running";
       cancelled: "cancelled";
+      error: "error";
       ok: "ok";
     }>;
     attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
@@ -2585,9 +2707,9 @@ declare const scoreTraceSchema: z$1.ZodObject<{
         subtree: "subtree";
       }>>;
       mode: z$1.ZodOptional<z$1.ZodEnum<{
+        all: "all";
         sum: "sum";
         last: "last";
-        all: "all";
       }>>;
     }, z$1.core.$strip>>>;
   }, z$1.core.$strip>;
@@ -2597,10 +2719,10 @@ declare const scoreTraceSchema: z$1.ZodObject<{
     namespace: z$1.ZodString;
     key: z$1.ZodString;
     status: z$1.ZodEnum<{
+      bypass: "bypass";
+      refresh: "refresh";
       hit: "hit";
       miss: "miss";
-      refresh: "refresh";
-      bypass: "bypass";
     }>;
     read: z$1.ZodOptional<z$1.ZodBoolean>;
     stored: z$1.ZodOptional<z$1.ZodBoolean>;
@@ -2618,12 +2740,12 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
   evalId: z$1.ZodString;
   tags: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
   status: z$1.ZodEnum<{
+    pending: "pending";
+    running: "running";
+    cancelled: "cancelled";
     error: "error";
     pass: "pass";
     fail: "fail";
-    running: "running";
-    cancelled: "cancelled";
-    pending: "pending";
   }>;
   input: z$1.ZodUnknown;
   trace: z$1.ZodArray<z$1.ZodObject<{
@@ -2635,9 +2757,9 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
     startedAt: z$1.ZodString;
     endedAt: z$1.ZodNullable<z$1.ZodString>;
     status: z$1.ZodEnum<{
-      error: "error";
       running: "running";
       cancelled: "cancelled";
+      error: "error";
       ok: "ok";
     }>;
     attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
@@ -2688,9 +2810,9 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
         subtree: "subtree";
       }>>;
       mode: z$1.ZodOptional<z$1.ZodEnum<{
+        all: "all";
         sum: "sum";
         last: "last";
-        all: "all";
       }>>;
     }, z$1.core.$strip>>>;
   }, z$1.core.$strip>;
@@ -2704,9 +2826,9 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
       startedAt: z$1.ZodString;
       endedAt: z$1.ZodNullable<z$1.ZodString>;
       status: z$1.ZodEnum<{
-        error: "error";
         running: "running";
         cancelled: "cancelled";
+        error: "error";
         ok: "ok";
       }>;
       attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
@@ -2757,9 +2879,9 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
           subtree: "subtree";
         }>>;
         mode: z$1.ZodOptional<z$1.ZodEnum<{
+          all: "all";
           sum: "sum";
           last: "last";
-          all: "all";
         }>>;
       }, z$1.core.$strip>>>;
     }, z$1.core.$strip>;
@@ -2769,10 +2891,10 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
       namespace: z$1.ZodString;
       key: z$1.ZodString;
       status: z$1.ZodEnum<{
+        bypass: "bypass";
+        refresh: "refresh";
         hit: "hit";
         miss: "miss";
-        refresh: "refresh";
-        bypass: "bypass";
       }>;
       read: z$1.ZodOptional<z$1.ZodBoolean>;
       stored: z$1.ZodOptional<z$1.ZodBoolean>;
@@ -2838,10 +2960,10 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
     namespace: z$1.ZodString;
     key: z$1.ZodString;
     status: z$1.ZodEnum<{
+      bypass: "bypass";
+      refresh: "refresh";
       hit: "hit";
       miss: "miss";
-      refresh: "refresh";
-      bypass: "bypass";
     }>;
     read: z$1.ZodOptional<z$1.ZodBoolean>;
     stored: z$1.ZodOptional<z$1.ZodBoolean>;
@@ -2895,9 +3017,9 @@ type EvalChartBuiltinMetric = z$1.infer<typeof evalChartBuiltinMetricSchema>;
 /** Reducer applied to a numeric column across all cases of a single run. */
 declare const evalChartAggregateSchema: z$1.ZodEnum<{
   avg: "avg";
+  sum: "sum";
   min: "min";
   max: "max";
-  sum: "sum";
   latest: "latest";
   passThresholdRate: "passThresholdRate";
 }>;
@@ -2908,8 +3030,8 @@ type EvalChartAggregate = z$1.infer<typeof evalChartAggregateSchema>;
  * not emit raw hex so authored evals stay decoupled from the web theme.
  */
 declare const evalChartColorSchema: z$1.ZodEnum<{
-  success: "success";
   error: "error";
+  success: "success";
   accent: "accent";
   accentDim: "accentDim";
   warning: "warning";
@@ -2937,8 +3059,8 @@ declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   }>;
   label: z$1.ZodOptional<z$1.ZodString>;
   color: z$1.ZodOptional<z$1.ZodEnum<{
-    success: "success";
     error: "error";
+    success: "success";
     accent: "accent";
     accentDim: "accentDim";
     warning: "warning";
@@ -2953,16 +3075,16 @@ declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   key: z$1.ZodString;
   aggregate: z$1.ZodEnum<{
     avg: "avg";
+    sum: "sum";
     min: "min";
     max: "max";
-    sum: "sum";
     latest: "latest";
     passThresholdRate: "passThresholdRate";
   }>;
   label: z$1.ZodOptional<z$1.ZodString>;
   color: z$1.ZodOptional<z$1.ZodEnum<{
-    success: "success";
     error: "error";
+    success: "success";
     accent: "accent";
     accentDim: "accentDim";
     warning: "warning";
@@ -2988,9 +3110,9 @@ declare const evalChartTooltipExtraSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObj
   key: z$1.ZodString;
   aggregate: z$1.ZodEnum<{
     avg: "avg";
+    sum: "sum";
     min: "min";
     max: "max";
-    sum: "sum";
     latest: "latest";
     passThresholdRate: "passThresholdRate";
   }>;
@@ -3020,8 +3142,8 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
     }>;
     label: z$1.ZodOptional<z$1.ZodString>;
     color: z$1.ZodOptional<z$1.ZodEnum<{
-      success: "success";
       error: "error";
+      success: "success";
       accent: "accent";
       accentDim: "accentDim";
       warning: "warning";
@@ -3036,16 +3158,16 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
       avg: "avg";
+      sum: "sum";
       min: "min";
       max: "max";
-      sum: "sum";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
     label: z$1.ZodOptional<z$1.ZodString>;
     color: z$1.ZodOptional<z$1.ZodEnum<{
-      success: "success";
       error: "error";
+      success: "success";
       accent: "accent";
       accentDim: "accentDim";
       warning: "warning";
@@ -3078,9 +3200,9 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
       avg: "avg";
+      sum: "sum";
       min: "min";
       max: "max";
-      sum: "sum";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
@@ -3110,8 +3232,8 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
     }>;
     label: z$1.ZodOptional<z$1.ZodString>;
     color: z$1.ZodOptional<z$1.ZodEnum<{
-      success: "success";
       error: "error";
+      success: "success";
       accent: "accent";
       accentDim: "accentDim";
       warning: "warning";
@@ -3126,16 +3248,16 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
       avg: "avg";
+      sum: "sum";
       min: "min";
       max: "max";
-      sum: "sum";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
     label: z$1.ZodOptional<z$1.ZodString>;
     color: z$1.ZodOptional<z$1.ZodEnum<{
-      success: "success";
       error: "error";
+      success: "success";
       accent: "accent";
       accentDim: "accentDim";
       warning: "warning";
@@ -3168,9 +3290,9 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
       avg: "avg";
+      sum: "sum";
       min: "min";
       max: "max";
-      sum: "sum";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
@@ -3185,11 +3307,11 @@ declare const runManifestSchema$1: z$1.ZodObject<{
   id: z$1.ZodString;
   shortId: z$1.ZodString;
   status: z$1.ZodEnum<{
-    error: "error";
-    running: "running";
-    cancelled: "cancelled";
     pending: "pending";
+    running: "running";
     completed: "completed";
+    cancelled: "cancelled";
+    error: "error";
   }>;
   temporary: z$1.ZodDefault<z$1.ZodOptional<z$1.ZodBoolean>>;
   startedAt: z$1.ZodString;
@@ -3198,9 +3320,9 @@ declare const runManifestSchema$1: z$1.ZodObject<{
   evalSourceFingerprints: z$1.ZodDefault<z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodString>>>;
   target: z$1.ZodObject<{
     mode: z$1.ZodEnum<{
-      caseIds: "caseIds";
       all: "all";
       evalIds: "evalIds";
+      caseIds: "caseIds";
     }>;
     evalKeys: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
     files: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
@@ -3214,9 +3336,9 @@ declare const runManifestSchema$1: z$1.ZodObject<{
     median: "median";
   }>>>;
   cacheMode: z$1.ZodOptional<z$1.ZodEnum<{
-    refresh: "refresh";
-    bypass: "bypass";
     use: "use";
+    bypass: "bypass";
+    refresh: "refresh";
   }>>;
 }, z$1.core.$strip>;
 /** Persisted lifecycle metadata for a single eval run. */
@@ -3225,11 +3347,11 @@ type RunManifest = z$1.infer<typeof runManifestSchema$1>;
 declare const runSummarySchema$1: z$1.ZodObject<{
   runId: z$1.ZodString;
   status: z$1.ZodEnum<{
-    error: "error";
-    running: "running";
-    cancelled: "cancelled";
     pending: "pending";
+    running: "running";
     completed: "completed";
+    cancelled: "cancelled";
+    error: "error";
   }>;
   totalCases: z$1.ZodNumber;
   passedCases: z$1.ZodNumber;
@@ -3312,8 +3434,8 @@ type TrialSelectionMode = z$1.infer<typeof trialSelectionModeSchema>;
 /** Built-in eval-level output/column keys. */
 /** Removal config for built-in eval-level outputs and UI metadata. */
 declare const removeDefaultConfigSchema: z$1.ZodUnion<readonly [z$1.ZodLiteral<true>, z$1.ZodArray<z$1.ZodEnum<{
-  costUsd: "costUsd";
   apiCalls: "apiCalls";
+  costUsd: "costUsd";
   llmTurns: "llmTurns";
   inputTokens: "inputTokens";
   outputTokens: "outputTokens";
@@ -3818,6 +3940,15 @@ type AgentEvalsConfig$1 = {
    * appended last unless removed with `removeDefaultConfig`.
    */
   stats?: EvalStatsConfig;
+  /**
+   * Initial aggregate mode used for duration and column stats on every eval
+   * card.
+   *
+   * Per-eval `defaultStatAggregate` overrides this value. Individual stat
+   * `aggregate` values still define their authored reducer and remain the
+   * fallback when no default aggregate is configured.
+   */
+  defaultStatAggregate?: EvalStatAggregate;
   /**
    * Configuration for the "LLM calls" tab in the case-run drawer.
    *
@@ -4118,9 +4249,9 @@ declare function extractApiCalls(spans: EvalTraceSpan$1[], config: ResolvedApiCa
  * - `refresh`: never read, always write (forces re-execution and overwrites).
  */
 declare const cacheModeSchema: z$1.ZodEnum<{
-  refresh: "refresh";
-  bypass: "bypass";
   use: "use";
+  bypass: "bypass";
+  refresh: "refresh";
 }>;
 /** Mode controlling how cached spans behave during a run. */
 type CacheMode = z$1.infer<typeof cacheModeSchema>;
@@ -4141,10 +4272,10 @@ declare const cacheOperationTypeSchema: z$1.ZodEnum<{
 type CacheOperationType = z$1.infer<typeof cacheOperationTypeSchema>;
 /** Status of a cache lookup recorded on a span or case scope. */
 declare const cacheStatusSchema: z$1.ZodEnum<{
+  bypass: "bypass";
+  refresh: "refresh";
   hit: "hit";
   miss: "miss";
-  refresh: "refresh";
-  bypass: "bypass";
 }>;
 /** Status of a cache lookup recorded on a span or case scope. */
 type CacheStatus = z$1.infer<typeof cacheStatusSchema>;
@@ -4161,10 +4292,10 @@ declare const traceCacheRefSchema: z$1.ZodObject<{
   namespace: z$1.ZodString;
   key: z$1.ZodString;
   status: z$1.ZodEnum<{
+    bypass: "bypass";
+    refresh: "refresh";
     hit: "hit";
     miss: "miss";
-    refresh: "refresh";
-    bypass: "bypass";
   }>;
   read: z$1.ZodOptional<z$1.ZodBoolean>;
   stored: z$1.ZodOptional<z$1.ZodBoolean>;
@@ -4239,9 +4370,9 @@ declare const cacheRecordingSchema: z$1.ZodObject<{
   returnValue: z$1.ZodUnknown;
   finalAttributes: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
   finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
-    error: "error";
     running: "running";
     cancelled: "cancelled";
+    error: "error";
     ok: "ok";
   }>>;
   finalError: z$1.ZodOptional<z$1.ZodObject<{
@@ -4312,9 +4443,9 @@ declare const cacheEntrySchema: z$1.ZodObject<{
     returnValue: z$1.ZodUnknown;
     finalAttributes: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
     finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
-      error: "error";
       running: "running";
       cancelled: "cancelled";
+      error: "error";
       ok: "ok";
     }>>;
     finalError: z$1.ZodOptional<z$1.ZodObject<{
@@ -4402,9 +4533,9 @@ declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
       returnValue: z$1.ZodUnknown;
       finalAttributes: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
       finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
-        error: "error";
         running: "running";
         cancelled: "cancelled";
+        error: "error";
         ok: "ok";
       }>>;
       finalError: z$1.ZodOptional<z$1.ZodObject<{
@@ -4481,9 +4612,9 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
     returnValue: z$1.ZodUnknown;
     finalAttributes: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
     finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
-      error: "error";
       running: "running";
       cancelled: "cancelled";
+      error: "error";
       ok: "ok";
     }>>;
     finalError: z$1.ZodOptional<z$1.ZodObject<{
@@ -4562,9 +4693,9 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
         returnValue: z$1.ZodUnknown;
         finalAttributes: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
         finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
-          error: "error";
           running: "running";
           cancelled: "cancelled";
+          error: "error";
           ok: "ok";
         }>>;
         finalError: z$1.ZodOptional<z$1.ZodObject<{
@@ -4641,9 +4772,9 @@ declare const cacheFileSchema: z$1.ZodObject<{
       returnValue: z$1.ZodUnknown;
       finalAttributes: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
       finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
-        error: "error";
         running: "running";
         cancelled: "cancelled";
+        error: "error";
         ok: "ok";
       }>>;
       finalError: z$1.ZodOptional<z$1.ZodObject<{
@@ -4730,9 +4861,9 @@ declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
         returnValue: z$1.ZodUnknown;
         finalAttributes: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
         finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
-          error: "error";
           running: "running";
           cancelled: "cancelled";
+          error: "error";
           ok: "ok";
         }>>;
         finalError: z$1.ZodOptional<z$1.ZodObject<{
@@ -4889,9 +5020,9 @@ type ConfigReloadState = z$1.infer<typeof configReloadStateSchema$1>;
 declare const createRunRequestSchema$1: z$1.ZodObject<{
   target: z$1.ZodObject<{
     mode: z$1.ZodEnum<{
-      caseIds: "caseIds";
       all: "all";
       evalIds: "evalIds";
+      caseIds: "caseIds";
     }>;
     evalKeys: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
     files: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
@@ -4903,9 +5034,9 @@ declare const createRunRequestSchema$1: z$1.ZodObject<{
   temporary: z$1.ZodOptional<z$1.ZodBoolean>;
   cache: z$1.ZodOptional<z$1.ZodObject<{
     mode: z$1.ZodDefault<z$1.ZodEnum<{
-      refresh: "refresh";
-      bypass: "bypass";
       use: "use";
+      bypass: "bypass";
+      refresh: "refresh";
     }>>;
   }, z$1.core.$strip>>;
   manualInputs: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
@@ -5242,9 +5373,25 @@ declare const evalSummarySchema: z$1.ZodObject<{
   }, z$1.core.$strip>, z$1.ZodObject<{
     hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
     kind: z$1.ZodLiteral<"duration">;
+    aggregate: z$1.ZodOptional<z$1.ZodEnum<{
+      avg: "avg";
+      min: "min";
+      max: "max";
+      sum: "sum";
+      best: "best";
+      worst: "worst";
+    }>>;
   }, z$1.core.$strip>, z$1.ZodObject<{
     hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
     kind: z$1.ZodLiteral<"cacheHits">;
+    aggregate: z$1.ZodOptional<z$1.ZodEnum<{
+      avg: "avg";
+      min: "min";
+      max: "max";
+      sum: "sum";
+      best: "best";
+      worst: "worst";
+    }>>;
   }, z$1.core.$strip>, z$1.ZodObject<{
     hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
     kind: z$1.ZodLiteral<"column">;
@@ -5255,7 +5402,8 @@ declare const evalSummarySchema: z$1.ZodObject<{
       min: "min";
       max: "max";
       sum: "sum";
-      last: "last";
+      best: "best";
+      worst: "worst";
     }>;
     format: z$1.ZodOptional<z$1.ZodEnum<{
       number: "number";
@@ -5276,6 +5424,14 @@ declare const evalSummarySchema: z$1.ZodObject<{
     numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions$1, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions$1, unknown>>>;
     accent: z$1.ZodOptional<z$1.ZodBoolean>;
   }, z$1.core.$strip>], "kind">>>;
+  defaultStatAggregate: z$1.ZodOptional<z$1.ZodEnum<{
+    avg: "avg";
+    min: "min";
+    max: "max";
+    sum: "sum";
+    best: "best";
+    worst: "worst";
+  }>>;
   charts: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
     heading: z$1.ZodOptional<z$1.ZodString>;
     hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
@@ -5555,8 +5711,8 @@ declare const caseDetailSchema: z$1.ZodObject<{
       }>>;
       mode: z$1.ZodOptional<z$1.ZodEnum<{
         sum: "sum";
-        last: "last";
         all: "all";
+        last: "last";
       }>>;
     }, z$1.core.$strip>>>;
   }, z$1.core.$strip>;
@@ -5624,8 +5780,8 @@ declare const caseDetailSchema: z$1.ZodObject<{
         }>>;
         mode: z$1.ZodOptional<z$1.ZodEnum<{
           sum: "sum";
-          last: "last";
           all: "all";
+          last: "last";
         }>>;
       }, z$1.core.$strip>>>;
     }, z$1.core.$strip>;
@@ -6343,6 +6499,22 @@ type EvalRunner = {
   deleteRun(runId: string): Promise<{
     deleted: boolean;
   }>;
+  /**
+   * Convert a temporary persisted run into durable run history.
+   *
+   * Returns the updated run when found. Already-durable runs are treated as a
+   * no-op success so UI callers can refresh their cached copy idempotently.
+   */
+  promoteRun(runId: string): Promise<{
+    promoted: boolean;
+    run: {
+      manifest: RunManifest$1;
+      summary: RunSummary$1;
+      cases: CaseRow$1[];
+    };
+  } | {
+    promoted: false;
+  }>;
   /**
    * Validate a `CreateRunRequest`'s `manualInputs` map against each targeted
    * eval's authored `manualInput.schema`. Returns `ok: true` with the parsed