@ls-stack/agent-eval 0.58.4 → 0.59.0

package/dist/index.d.mts

@@ -398,6 +398,18 @@ type EvalCase$1$1<TInput = unknown> = {
   input: TInput;
   tags?: string[];
 };
+/** Normalized view of one tool-call span and its common tool metadata. */
+type EvalToolCallSpan = {
+  /** Preferred tool name, using GenAI/Mastra identity metadata when present. */name: string; /** Original trace span display name. */
+  spanName: string; /** Original trace span kind. */
+  kind: string; /** Parsed tool-call arguments, or the raw value when parsing is not possible. */
+  arguments: unknown; /** Parsed tool-call result, or the raw value when parsing is not possible. */
+  result: unknown; /** Tool description from GenAI/Mastra metadata when present. */
+  description: string | undefined; /** Tool type from GenAI/Mastra metadata when present. */
+  toolType: string | undefined; /** Original span attributes. */
+  attributes: Record<string, unknown> | undefined; /** Original trace span for fields not normalized above. */
+  span: EvalTraceSpan$2;
+};
 /** Query helpers built from the flattened trace recorded for one eval case. */
 type EvalTraceTree = {
   /** Flat span list in creation order. */spans: EvalTraceSpan$2[]; /** Top-level spans whose `parentId` is `null`. */
@@ -405,10 +417,17 @@ type EvalTraceTree = {
   findSpan: (name: string) => EvalTraceSpan$2 | undefined; /** Return every span whose name exactly matches `name`. */
   findSpans: (name: string) => EvalTraceSpan$2[]; /** Return whether any span name exactly matches `name`. */
   hasSpan: (name: string) => boolean; /** Return every span whose kind exactly matches `kind`. */
-  findSpansByKind: (kind: string) => EvalTraceSpan$2[]; /** Return every span with `kind: 'tool'`. */
-  findToolCallSpans: () => EvalTraceSpan$2[]; /** Return the names of every span with `kind: 'tool'`. */
-  listToolCallSpanNames: () => string[]; /** Return whether a `kind: 'tool'` span has a name exactly matching `name`. */
-  hasToolCallSpan: (name: string) => boolean; /** Return span names in creation order, optionally filtered by kind. */
+  findSpansByKind: (kind: string) => EvalTraceSpan$2[]; /** Return every span with `kind: 'tool'` or `kind: 'tool_call'`. */
+  findToolCallSpans: () => EvalTraceSpan$2[];
+  /**
+   * Return tool-call names, preferring GenAI/Mastra tool identity attributes
+   * when available.
+   */
+  listToolCallSpanNames: () => string[]; /** Return whether a tool-call span name or tool identity matches `name`. */
+  hasToolCallSpan: (name: string) => boolean; /** Return normalized tool-call spans whose name or tool identity matches `name`. */
+  getToolCallSpans: (name: string) => EvalToolCallSpan[]; /** Return how many tool-call spans have a name or tool identity matching `toolName`. */
+  getToolCallSpanCount: (toolName: string) => number; /** Return whether a tool-call span name or tool identity appears exactly `expectedCalls` times. */
+  hasToolCallSpanCount: (toolName: string, expectedCalls: number) => boolean; /** Return span names in creation order, optionally filtered by kind. */
   listSpanNames: (kind?: string) => string[]; /** Return span names in depth-first tree order, optionally filtered by kind. */
   listSpanNamesDfs: (kind?: string) => string[]; /** Return all spans in depth-first tree order. */
   flattenDfs: () => EvalTraceSpan$2[];
@@ -2241,10 +2260,10 @@ type EvalFreshnessStatus = z$1.infer<typeof evalFreshnessStatusSchema>;
  * `best` selects the highest finite value and `worst` selects the lowest.
  */
 declare const evalStatAggregateSchema: z$1.ZodEnum<{
-  min: "min";
-  max: "max";
   sum: "sum";
   avg: "avg";
+  min: "min";
+  max: "max";
   best: "best";
   worst: "worst";
 }>;
@@ -2273,10 +2292,10 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"duration">;
   aggregate: z$1.ZodOptional<z$1.ZodEnum<{
-    min: "min";
-    max: "max";
     sum: "sum";
     avg: "avg";
+    min: "min";
+    max: "max";
     best: "best";
     worst: "worst";
   }>>;
@@ -2284,10 +2303,10 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"cacheHits">;
   aggregate: z$1.ZodOptional<z$1.ZodEnum<{
-    min: "min";
-    max: "max";
     sum: "sum";
     avg: "avg";
+    min: "min";
+    max: "max";
     best: "best";
     worst: "worst";
   }>>;
@@ -2297,10 +2316,10 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   key: z$1.ZodString;
   label: z$1.ZodOptional<z$1.ZodString>;
   aggregate: z$1.ZodEnum<{
-    min: "min";
-    max: "max";
     sum: "sum";
     avg: "avg";
+    min: "min";
+    max: "max";
     best: "best";
     worst: "worst";
   }>;
@@ -2337,10 +2356,10 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
   hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"duration">;
   aggregate: z$1.ZodOptional<z$1.ZodEnum<{
-    min: "min";
-    max: "max";
     sum: "sum";
     avg: "avg";
+    min: "min";
+    max: "max";
     best: "best";
     worst: "worst";
   }>>;
@@ -2348,10 +2367,10 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
   hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
   kind: z$1.ZodLiteral<"cacheHits">;
   aggregate: z$1.ZodOptional<z$1.ZodEnum<{
-    min: "min";
-    max: "max";
     sum: "sum";
     avg: "avg";
+    min: "min";
+    max: "max";
     best: "best";
     worst: "worst";
   }>>;
@@ -2361,10 +2380,10 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
   key: z$1.ZodString;
   label: z$1.ZodOptional<z$1.ZodString>;
   aggregate: z$1.ZodEnum<{
-    min: "min";
-    max: "max";
     sum: "sum";
     avg: "avg";
+    min: "min";
+    max: "max";
     best: "best";
     worst: "worst";
   }>;
@@ -2464,10 +2483,10 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
     hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
     kind: z$1.ZodLiteral<"duration">;
     aggregate: z$1.ZodOptional<z$1.ZodEnum<{
-      min: "min";
-      max: "max";
       sum: "sum";
       avg: "avg";
+      min: "min";
+      max: "max";
       best: "best";
       worst: "worst";
     }>>;
@@ -2475,10 +2494,10 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
     hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
     kind: z$1.ZodLiteral<"cacheHits">;
     aggregate: z$1.ZodOptional<z$1.ZodEnum<{
-      min: "min";
-      max: "max";
       sum: "sum";
       avg: "avg";
+      min: "min";
+      max: "max";
       best: "best";
       worst: "worst";
     }>>;
@@ -2488,10 +2507,10 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
     key: z$1.ZodString;
     label: z$1.ZodOptional<z$1.ZodString>;
     aggregate: z$1.ZodEnum<{
-      min: "min";
-      max: "max";
       sum: "sum";
       avg: "avg";
+      min: "min";
+      max: "max";
       best: "best";
       worst: "worst";
     }>;
@@ -2515,10 +2534,10 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
     accent: z$1.ZodOptional<z$1.ZodBoolean>;
   }, z$1.core.$strip>], "kind">>>;
   defaultStatAggregate: z$1.ZodOptional<z$1.ZodEnum<{
-    min: "min";
-    max: "max";
     sum: "sum";
     avg: "avg";
+    min: "min";
+    max: "max";
     best: "best";
     worst: "worst";
   }>>;
@@ -2534,15 +2553,15 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
     metrics: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
       source: z$1.ZodLiteral<"builtin">;
       metric: z$1.ZodEnum<{
-        durationMs: "durationMs";
         passRate: "passRate";
+        durationMs: "durationMs";
       }>;
       label: z$1.ZodOptional<z$1.ZodString>;
       color: z$1.ZodOptional<z$1.ZodEnum<{
-        error: "error";
         success: "success";
-        accent: "accent";
+        error: "error";
         warning: "warning";
+        accent: "accent";
         accentDim: "accentDim";
         textMuted: "textMuted";
       }>>;
@@ -2554,19 +2573,19 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
       source: z$1.ZodLiteral<"column">;
       key: z$1.ZodString;
       aggregate: z$1.ZodEnum<{
-        min: "min";
-        max: "max";
         sum: "sum";
         avg: "avg";
+        min: "min";
+        max: "max";
         latest: "latest";
         passThresholdRate: "passThresholdRate";
       }>;
       label: z$1.ZodOptional<z$1.ZodString>;
       color: z$1.ZodOptional<z$1.ZodEnum<{
-        error: "error";
         success: "success";
-        accent: "accent";
+        error: "error";
         warning: "warning";
+        accent: "accent";
         accentDim: "accentDim";
         textMuted: "textMuted";
       }>>;
@@ -2588,18 +2607,18 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
     tooltipExtras: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
       source: z$1.ZodLiteral<"builtin">;
       metric: z$1.ZodEnum<{
-        durationMs: "durationMs";
         passRate: "passRate";
+        durationMs: "durationMs";
       }>;
       label: z$1.ZodOptional<z$1.ZodString>;
     }, z$1.core.$strip>, z$1.ZodObject<{
       source: z$1.ZodLiteral<"column">;
       key: z$1.ZodString;
       aggregate: z$1.ZodEnum<{
-        min: "min";
-        max: "max";
         sum: "sum";
         avg: "avg";
+        min: "min";
+        max: "max";
         latest: "latest";
         passThresholdRate: "passThresholdRate";
       }>;
@@ -2698,9 +2717,9 @@ declare const caseRowSchema$1: z$1.ZodObject<{
     error: "error";
     running: "running";
     cancelled: "cancelled";
-    pending: "pending";
     pass: "pass";
     fail: "fail";
+    pending: "pending";
   }>;
   durationMs: z$1.ZodNullable<z$1.ZodNumber>;
   cacheHits: z$1.ZodOptional<z$1.ZodNumber>;
@@ -2779,9 +2798,9 @@ declare const runLogLevelSchema: z$1.ZodEnum<{
 type RunLogLevel = z$1.infer<typeof runLogLevelSchema>;
 /** Eval runner phase that emitted a captured case log. */
 declare const runLogPhaseSchema: z$1.ZodEnum<{
-  tracingAssertions: "tracingAssertions";
   eval: "eval";
   derive: "derive";
+  tracingAssertions: "tracingAssertions";
   outputsSchema: "outputsSchema";
   scorer: "scorer";
 }>;
@@ -2806,9 +2825,9 @@ declare const runLogEntrySchema: z$1.ZodObject<{
     warn: "warn";
   }>;
   phase: z$1.ZodEnum<{
-    tracingAssertions: "tracingAssertions";
     eval: "eval";
     derive: "derive";
+    tracingAssertions: "tracingAssertions";
     outputsSchema: "outputsSchema";
     scorer: "scorer";
   }>;
@@ -2901,10 +2920,10 @@ 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>;
@@ -2925,9 +2944,9 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
     error: "error";
     running: "running";
     cancelled: "cancelled";
-    pending: "pending";
     pass: "pass";
     fail: "fail";
+    pending: "pending";
   }>;
   input: z$1.ZodUnknown;
   trace: z$1.ZodArray<z$1.ZodObject<{
@@ -3073,10 +3092,10 @@ 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>;
@@ -3165,9 +3184,9 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
       warn: "warn";
     }>;
     phase: z$1.ZodEnum<{
-      tracingAssertions: "tracingAssertions";
       eval: "eval";
       derive: "derive";
+      tracingAssertions: "tracingAssertions";
       outputsSchema: "outputsSchema";
       scorer: "scorer";
     }>;
@@ -3194,10 +3213,10 @@ 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>;
@@ -3240,8 +3259,8 @@ type EvalChartType = z$1.infer<typeof evalChartTypeSchema>;
  * than from a per-case column.
  */
 declare const evalChartBuiltinMetricSchema: z$1.ZodEnum<{
-  durationMs: "durationMs";
   passRate: "passRate";
+  durationMs: "durationMs";
 }>;
 /**
  * Run-level metric sourced from the aggregated `RunSummary` for a run, rather
@@ -3250,10 +3269,10 @@ declare const evalChartBuiltinMetricSchema: z$1.ZodEnum<{
 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<{
-  min: "min";
-  max: "max";
   sum: "sum";
   avg: "avg";
+  min: "min";
+  max: "max";
   latest: "latest";
   passThresholdRate: "passThresholdRate";
 }>;
@@ -3264,10 +3283,10 @@ 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<{
-  error: "error";
   success: "success";
-  accent: "accent";
+  error: "error";
   warning: "warning";
+  accent: "accent";
   accentDim: "accentDim";
   textMuted: "textMuted";
 }>;
@@ -3288,15 +3307,15 @@ type EvalChartAxis = z$1.infer<typeof evalChartAxisSchema>;
 declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   source: z$1.ZodLiteral<"builtin">;
   metric: z$1.ZodEnum<{
-    durationMs: "durationMs";
     passRate: "passRate";
+    durationMs: "durationMs";
   }>;
   label: z$1.ZodOptional<z$1.ZodString>;
   color: z$1.ZodOptional<z$1.ZodEnum<{
-    error: "error";
     success: "success";
-    accent: "accent";
+    error: "error";
     warning: "warning";
+    accent: "accent";
     accentDim: "accentDim";
     textMuted: "textMuted";
   }>>;
@@ -3308,19 +3327,19 @@ declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   source: z$1.ZodLiteral<"column">;
   key: z$1.ZodString;
   aggregate: z$1.ZodEnum<{
-    min: "min";
-    max: "max";
     sum: "sum";
     avg: "avg";
+    min: "min";
+    max: "max";
     latest: "latest";
     passThresholdRate: "passThresholdRate";
   }>;
   label: z$1.ZodOptional<z$1.ZodString>;
   color: z$1.ZodOptional<z$1.ZodEnum<{
-    error: "error";
     success: "success";
-    accent: "accent";
+    error: "error";
     warning: "warning";
+    accent: "accent";
     accentDim: "accentDim";
     textMuted: "textMuted";
   }>>;
@@ -3335,18 +3354,18 @@ type EvalChartMetric = z$1.infer<typeof evalChartMetricSchema>;
 declare const evalChartTooltipExtraSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   source: z$1.ZodLiteral<"builtin">;
   metric: z$1.ZodEnum<{
-    durationMs: "durationMs";
     passRate: "passRate";
+    durationMs: "durationMs";
   }>;
   label: z$1.ZodOptional<z$1.ZodString>;
 }, z$1.core.$strip>, z$1.ZodObject<{
   source: z$1.ZodLiteral<"column">;
   key: z$1.ZodString;
   aggregate: z$1.ZodEnum<{
-    min: "min";
-    max: "max";
     sum: "sum";
     avg: "avg";
+    min: "min";
+    max: "max";
     latest: "latest";
     passThresholdRate: "passThresholdRate";
   }>;
@@ -3371,15 +3390,15 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
   metrics: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
     source: z$1.ZodLiteral<"builtin">;
     metric: z$1.ZodEnum<{
-      durationMs: "durationMs";
       passRate: "passRate";
+      durationMs: "durationMs";
     }>;
     label: z$1.ZodOptional<z$1.ZodString>;
     color: z$1.ZodOptional<z$1.ZodEnum<{
-      error: "error";
       success: "success";
-      accent: "accent";
+      error: "error";
       warning: "warning";
+      accent: "accent";
       accentDim: "accentDim";
       textMuted: "textMuted";
     }>>;
@@ -3391,19 +3410,19 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
     source: z$1.ZodLiteral<"column">;
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
-      min: "min";
-      max: "max";
       sum: "sum";
       avg: "avg";
+      min: "min";
+      max: "max";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
     label: z$1.ZodOptional<z$1.ZodString>;
     color: z$1.ZodOptional<z$1.ZodEnum<{
-      error: "error";
       success: "success";
-      accent: "accent";
+      error: "error";
       warning: "warning";
+      accent: "accent";
       accentDim: "accentDim";
       textMuted: "textMuted";
     }>>;
@@ -3425,18 +3444,18 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
   tooltipExtras: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
     source: z$1.ZodLiteral<"builtin">;
     metric: z$1.ZodEnum<{
-      durationMs: "durationMs";
       passRate: "passRate";
+      durationMs: "durationMs";
     }>;
     label: z$1.ZodOptional<z$1.ZodString>;
   }, z$1.core.$strip>, z$1.ZodObject<{
     source: z$1.ZodLiteral<"column">;
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
-      min: "min";
-      max: "max";
       sum: "sum";
       avg: "avg";
+      min: "min";
+      max: "max";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
@@ -3461,15 +3480,15 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
   metrics: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
     source: z$1.ZodLiteral<"builtin">;
     metric: z$1.ZodEnum<{
-      durationMs: "durationMs";
       passRate: "passRate";
+      durationMs: "durationMs";
     }>;
     label: z$1.ZodOptional<z$1.ZodString>;
     color: z$1.ZodOptional<z$1.ZodEnum<{
-      error: "error";
       success: "success";
-      accent: "accent";
+      error: "error";
       warning: "warning";
+      accent: "accent";
       accentDim: "accentDim";
       textMuted: "textMuted";
     }>>;
@@ -3481,19 +3500,19 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
     source: z$1.ZodLiteral<"column">;
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
-      min: "min";
-      max: "max";
       sum: "sum";
       avg: "avg";
+      min: "min";
+      max: "max";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
     label: z$1.ZodOptional<z$1.ZodString>;
     color: z$1.ZodOptional<z$1.ZodEnum<{
-      error: "error";
       success: "success";
-      accent: "accent";
+      error: "error";
       warning: "warning";
+      accent: "accent";
       accentDim: "accentDim";
       textMuted: "textMuted";
     }>>;
@@ -3515,18 +3534,18 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
   tooltipExtras: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
     source: z$1.ZodLiteral<"builtin">;
     metric: z$1.ZodEnum<{
-      durationMs: "durationMs";
       passRate: "passRate";
+      durationMs: "durationMs";
     }>;
     label: z$1.ZodOptional<z$1.ZodString>;
   }, z$1.core.$strip>, z$1.ZodObject<{
     source: z$1.ZodLiteral<"column">;
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
-      min: "min";
-      max: "max";
       sum: "sum";
       avg: "avg";
+      min: "min";
+      max: "max";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
@@ -3555,8 +3574,8 @@ declare const runManifestSchema$1: z$1.ZodObject<{
   target: z$1.ZodObject<{
     mode: z$1.ZodEnum<{
       all: "all";
-      evalIds: "evalIds";
       caseIds: "caseIds";
+      evalIds: "evalIds";
     }>;
     evalKeys: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
     files: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
@@ -3570,9 +3589,9 @@ declare const runManifestSchema$1: z$1.ZodObject<{
     median: "median";
   }>>>;
   cacheMode: z$1.ZodOptional<z$1.ZodEnum<{
-    use: "use";
-    bypass: "bypass";
     refresh: "refresh";
+    bypass: "bypass";
+    use: "use";
   }>>;
 }, z$1.core.$strip>;
 /** Persisted lifecycle metadata for a single eval run. */
@@ -3668,8 +3687,8 @@ 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<{
-  apiCalls: "apiCalls";
   costUsd: "costUsd";
+  apiCalls: "apiCalls";
   llmTurns: "llmTurns";
   inputTokens: "inputTokens";
   outputTokens: "outputTokens";
@@ -3687,6 +3706,18 @@ type EvalCase$1<TInput = unknown> = {
   input: TInput;
   tags?: string[];
 };
+/** Normalized view of one tool-call span and its common tool metadata. */
+type EvalToolCallSpan$1 = {
+  /** Preferred tool name, using GenAI/Mastra identity metadata when present. */name: string; /** Original trace span display name. */
+  spanName: string; /** Original trace span kind. */
+  kind: string; /** Parsed tool-call arguments, or the raw value when parsing is not possible. */
+  arguments: unknown; /** Parsed tool-call result, or the raw value when parsing is not possible. */
+  result: unknown; /** Tool description from GenAI/Mastra metadata when present. */
+  description: string | undefined; /** Tool type from GenAI/Mastra metadata when present. */
+  toolType: string | undefined; /** Original span attributes. */
+  attributes: Record<string, unknown> | undefined; /** Original trace span for fields not normalized above. */
+  span: EvalTraceSpan$1;
+};
 /** Query helpers built from the flattened trace recorded for one eval case. */
 type EvalTraceTree$1 = {
   /** Flat span list in creation order. */spans: EvalTraceSpan$1[]; /** Top-level spans whose `parentId` is `null`. */
@@ -3694,10 +3725,17 @@ type EvalTraceTree$1 = {
   findSpan: (name: string) => EvalTraceSpan$1 | undefined; /** Return every span whose name exactly matches `name`. */
   findSpans: (name: string) => EvalTraceSpan$1[]; /** Return whether any span name exactly matches `name`. */
   hasSpan: (name: string) => boolean; /** Return every span whose kind exactly matches `kind`. */
-  findSpansByKind: (kind: string) => EvalTraceSpan$1[]; /** Return every span with `kind: 'tool'`. */
-  findToolCallSpans: () => EvalTraceSpan$1[]; /** Return the names of every span with `kind: 'tool'`. */
-  listToolCallSpanNames: () => string[]; /** Return whether a `kind: 'tool'` span has a name exactly matching `name`. */
-  hasToolCallSpan: (name: string) => boolean; /** Return span names in creation order, optionally filtered by kind. */
+  findSpansByKind: (kind: string) => EvalTraceSpan$1[]; /** Return every span with `kind: 'tool'` or `kind: 'tool_call'`. */
+  findToolCallSpans: () => EvalTraceSpan$1[];
+  /**
+   * Return tool-call names, preferring GenAI/Mastra tool identity attributes
+   * when available.
+   */
+  listToolCallSpanNames: () => string[]; /** Return whether a tool-call span name or tool identity matches `name`. */
+  hasToolCallSpan: (name: string) => boolean; /** Return normalized tool-call spans whose name or tool identity matches `name`. */
+  getToolCallSpans: (name: string) => EvalToolCallSpan$1[]; /** Return how many tool-call spans have a name or tool identity matching `toolName`. */
+  getToolCallSpanCount: (toolName: string) => number; /** Return whether a tool-call span name or tool identity appears exactly `expectedCalls` times. */
+  hasToolCallSpanCount: (toolName: string, expectedCalls: number) => boolean; /** Return span names in creation order, optionally filtered by kind. */
   listSpanNames: (kind?: string) => string[]; /** Return span names in depth-first tree order, optionally filtered by kind. */
   listSpanNamesDfs: (kind?: string) => string[]; /** Return all spans in depth-first tree order. */
   flattenDfs: () => EvalTraceSpan$1[];
@@ -4529,9 +4567,9 @@ declare function extractApiCalls(spans: EvalTraceSpan$1[], config: ResolvedApiCa
  * - `refresh`: never read, always write (forces re-execution and overwrites).
  */
 declare const cacheModeSchema: z$1.ZodEnum<{
-  use: "use";
-  bypass: "bypass";
   refresh: "refresh";
+  bypass: "bypass";
+  use: "use";
 }>;
 /** Mode controlling how cached spans behave during a run. */
 type CacheMode = z$1.infer<typeof cacheModeSchema>;
@@ -4545,17 +4583,17 @@ declare const spanCacheOptionsSchema: z$1.ZodObject<{
 type SpanCacheOptions = z$1.infer<typeof spanCacheOptionsSchema>;
 /** Category of operation stored in the eval cache. */
 declare const cacheOperationTypeSchema: z$1.ZodEnum<{
-  span: "span";
   value: "value";
+  span: "span";
 }>;
 /** Category of operation stored in the eval cache. */
 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>;
@@ -4572,10 +4610,10 @@ 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>;
@@ -4771,8 +4809,8 @@ declare const cacheEntrySchema: z$1.ZodObject<{
   key: z$1.ZodString;
   namespace: z$1.ZodString;
   operationType: z$1.ZodOptional<z$1.ZodEnum<{
-    span: "span";
     value: "value";
+    span: "span";
   }>>;
   operationName: z$1.ZodOptional<z$1.ZodString>;
   spanName: z$1.ZodOptional<z$1.ZodString>;
@@ -4878,8 +4916,8 @@ declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
   key: z$1.ZodString;
   namespace: z$1.ZodString;
   operationType: z$1.ZodEnum<{
-    span: "span";
     value: "value";
+    span: "span";
   }>;
   operationName: z$1.ZodString;
   storedAt: z$1.ZodString;
@@ -4889,8 +4927,8 @@ declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
     key: z$1.ZodString;
     namespace: z$1.ZodString;
     operationType: z$1.ZodOptional<z$1.ZodEnum<{
-      span: "span";
       value: "value";
+      span: "span";
     }>>;
     operationName: z$1.ZodOptional<z$1.ZodString>;
     spanName: z$1.ZodOptional<z$1.ZodString>;
@@ -4996,8 +5034,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
   key: z$1.ZodString;
   namespace: z$1.ZodString;
   operationType: z$1.ZodOptional<z$1.ZodEnum<{
-    span: "span";
     value: "value";
+    span: "span";
   }>>;
   operationName: z$1.ZodOptional<z$1.ZodString>;
   spanName: z$1.ZodOptional<z$1.ZodString>;
@@ -5094,8 +5132,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
     key: z$1.ZodString;
     namespace: z$1.ZodString;
     operationType: z$1.ZodEnum<{
-      span: "span";
       value: "value";
+      span: "span";
     }>;
     operationName: z$1.ZodString;
     storedAt: z$1.ZodString;
@@ -5105,8 +5143,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
       key: z$1.ZodString;
       namespace: z$1.ZodString;
       operationType: z$1.ZodOptional<z$1.ZodEnum<{
-        span: "span";
         value: "value";
+        span: "span";
       }>>;
       operationName: z$1.ZodOptional<z$1.ZodString>;
       spanName: z$1.ZodOptional<z$1.ZodString>;
@@ -5212,8 +5250,8 @@ declare const cacheFileSchema: z$1.ZodObject<{
     key: z$1.ZodString;
     namespace: z$1.ZodString;
     operationType: z$1.ZodOptional<z$1.ZodEnum<{
-      span: "span";
       value: "value";
+      span: "span";
     }>>;
     operationName: z$1.ZodOptional<z$1.ZodString>;
     spanName: z$1.ZodOptional<z$1.ZodString>;
@@ -5318,8 +5356,8 @@ declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
     key: z$1.ZodString;
     namespace: z$1.ZodString;
     operationType: z$1.ZodEnum<{
-      span: "span";
       value: "value";
+      span: "span";
     }>;
     operationName: z$1.ZodString;
     storedAt: z$1.ZodString;
@@ -5329,8 +5367,8 @@ declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
       key: z$1.ZodString;
       namespace: z$1.ZodString;
       operationType: z$1.ZodOptional<z$1.ZodEnum<{
-        span: "span";
         value: "value";
+        span: "span";
       }>>;
       operationName: z$1.ZodOptional<z$1.ZodString>;
       spanName: z$1.ZodOptional<z$1.ZodString>;
@@ -5504,8 +5542,8 @@ type SseEnvelope = z$1.infer<typeof sseEnvelopeSchema$1>; //#endregion
 //#region src/schemas/api.d.ts
 /** Lifecycle state for an app config reload triggered by `agent-evals.config.ts`. */
 declare const configReloadStatusSchema: z$1.ZodEnum<{
-  idle: "idle";
   pending: "pending";
+  idle: "idle";
   reloading: "reloading";
 }>;
 /** Status for config reloads in the long-running app server. */
@@ -5513,8 +5551,8 @@ type ConfigReloadStatus = z$1.infer<typeof configReloadStatusSchema>;
 /** UI/API-visible state for config reloads in `agent-evals app`. */
 declare const configReloadStateSchema$1: z$1.ZodObject<{
   status: z$1.ZodEnum<{
-    idle: "idle";
     pending: "pending";
+    idle: "idle";
     reloading: "reloading";
   }>;
   activeRunCount: z$1.ZodNumber;
@@ -5528,8 +5566,8 @@ declare const createRunRequestSchema$1: z$1.ZodObject<{
   target: z$1.ZodObject<{
     mode: z$1.ZodEnum<{
       all: "all";
-      evalIds: "evalIds";
       caseIds: "caseIds";
+      evalIds: "evalIds";
     }>;
     evalKeys: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
     files: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
@@ -5541,9 +5579,9 @@ 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<{
-      use: "use";
-      bypass: "bypass";
       refresh: "refresh";
+      bypass: "bypass";
+      use: "use";
     }>>;
   }, z$1.core.$strip>>;
   manualInputs: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
@@ -7292,4 +7330,4 @@ declare function defineEval<TInput = unknown, TOutputs extends EvalOutputs = Eva
 /** Return whether the active eval case has tags matching the typed input. */
 declare function matchesEvalTags(input: EvalTagMatchInput): boolean;
 //#endregion
-export { AgentEvalTagRegistry, 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 CacheRepairSummary, type CacheScopeContext, type CacheSerializationOptions, type CacheStatus, type CallDerivedAttribute, type CallDerivedAttributeContext, type CallDerivedAttributesConfig, type CallDerivedAttributesFn, type CaptureEvalSpanErrorLevel, type CaptureEvalSpanErrorOptions, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type ConfigReloadState, type ConfigReloadStatus, type CreateRunRequest, type DefaultConfigKey, type DerivedStatus, type DiscoveryIssue, EvalAssertionError, type EvalCacheConfig, 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, EvalDefinition, type EvalDeriveConfig, type EvalDeriveContext, type EvalDeriveFn, type EvalDeriveMap, type EvalDeriveValueFn, type EvalDisplayStatus, type EvalExecuteContext, type EvalExpectation, type EvalFreshnessStatus, type EvalManualInputConfig, type EvalManualScoreDef, type EvalOutputs, type EvalOutputsSchema, type EvalRunner, type EvalRuntimeScope, EvalRuntimeUsageError, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalSetOutput, type EvalStartTime, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, EvalTag, EvalTagMatchInput, type EvalTraceTree, type EvalTracingAssertionsConfig, type EvalTracingAssertionsFn, type JsonCell, type LlmCallCostBreakdown, type LlmCallCostCurrency, type LlmCallEntry, type LlmCallMetric, type LlmCallMetricFormat, type LlmCallMetricPlacement, type LlmCallMetricValue, type LlmCallPricing, type LlmCallPricingRate, type LlmCallPricingRegistry, type LlmCallSimulatedTokens, type LlmCallsConfigInput, type LlmCostScenario, type ManualInputDescriptor, type ManualInputFieldDescriptor, type ManualInputFieldKind, type ManualInputFieldOverride, type ManualInputFieldsConfig, type ManualInputFileValue, type ManualInputSelectOption, type MaterializeManualInputFilesResult, type NumberDisplayOptions, type ReadManualInputFileResult, type RemoveDefaultConfig, type ResolvedApiCallMetric, type ResolvedApiCallsConfig, type ResolvedCallDerivedAttribute, type ResolvedLlmCallCostCurrency, type ResolvedLlmCallMetric, type ResolvedLlmCallPricing, type ResolvedLlmCallsConfig, 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, appendToEvalOutput, buildTraceTree, captureEvalSpanError, cleanupStagedManualInputFiles, createRunner, defineEval, deserializeCacheRecording, deserializeCacheValue, evalAssert, evalExpect, evalLog, evalSpan, evalTime, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, getCurrentScope, getEvalCaseInput, getEvalRegistry, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, isManualInputFileValue, manualInputFileValueSchema, matchesEvalTags, materializeManualInputFiles, mergeEvalOutput, nextEvalId, readManualInputFile, repoFile, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, serializeCacheRecording, serializeCacheValue, setEvalOutput, setScopeCacheContext, simulateLlmCallCost, simulateTokenAllocation, stageManualInputFile, stageManualInputFileFromPath, startEvalBackgroundJob, z };
+export { AgentEvalTagRegistry, 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 CacheRepairSummary, type CacheScopeContext, type CacheSerializationOptions, type CacheStatus, type CallDerivedAttribute, type CallDerivedAttributeContext, type CallDerivedAttributesConfig, type CallDerivedAttributesFn, type CaptureEvalSpanErrorLevel, type CaptureEvalSpanErrorOptions, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type ConfigReloadState, type ConfigReloadStatus, type CreateRunRequest, type DefaultConfigKey, type DerivedStatus, type DiscoveryIssue, EvalAssertionError, type EvalCacheConfig, 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, EvalDefinition, type EvalDeriveConfig, type EvalDeriveContext, type EvalDeriveFn, type EvalDeriveMap, type EvalDeriveValueFn, type EvalDisplayStatus, type EvalExecuteContext, type EvalExpectation, type EvalFreshnessStatus, type EvalManualInputConfig, type EvalManualScoreDef, type EvalOutputs, type EvalOutputsSchema, type EvalRunner, type EvalRuntimeScope, EvalRuntimeUsageError, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalSetOutput, type EvalStartTime, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, EvalTag, EvalTagMatchInput, type EvalToolCallSpan, type EvalTraceTree, type EvalTracingAssertionsConfig, type EvalTracingAssertionsFn, type JsonCell, type LlmCallCostBreakdown, type LlmCallCostCurrency, type LlmCallEntry, type LlmCallMetric, type LlmCallMetricFormat, type LlmCallMetricPlacement, type LlmCallMetricValue, type LlmCallPricing, type LlmCallPricingRate, type LlmCallPricingRegistry, type LlmCallSimulatedTokens, type LlmCallsConfigInput, type LlmCostScenario, type ManualInputDescriptor, type ManualInputFieldDescriptor, type ManualInputFieldKind, type ManualInputFieldOverride, type ManualInputFieldsConfig, type ManualInputFileValue, type ManualInputSelectOption, type MaterializeManualInputFilesResult, type NumberDisplayOptions, type ReadManualInputFileResult, type RemoveDefaultConfig, type ResolvedApiCallMetric, type ResolvedApiCallsConfig, type ResolvedCallDerivedAttribute, type ResolvedLlmCallCostCurrency, type ResolvedLlmCallMetric, type ResolvedLlmCallPricing, type ResolvedLlmCallsConfig, 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, appendToEvalOutput, buildTraceTree, captureEvalSpanError, cleanupStagedManualInputFiles, createRunner, defineEval, deserializeCacheRecording, deserializeCacheValue, evalAssert, evalExpect, evalLog, evalSpan, evalTime, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, getCurrentScope, getEvalCaseInput, getEvalRegistry, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, isManualInputFileValue, manualInputFileValueSchema, matchesEvalTags, materializeManualInputFiles, mergeEvalOutput, nextEvalId, readManualInputFile, repoFile, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, serializeCacheRecording, serializeCacheValue, setEvalOutput, setScopeCacheContext, simulateLlmCallCost, simulateTokenAllocation, stageManualInputFile, stageManualInputFileFromPath, startEvalBackgroundJob, z };