@ls-stack/agent-eval 0.58.3 → 0.58.5

package/dist/index.d.mts

@@ -405,10 +405,11 @@ 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 the names of every span with `kind: 'tool'` or `kind: 'tool_call'`. */
+  listToolCallSpanNames: () => string[]; /** Return whether a tool-call span has a name exactly matching `name`. */
+  hasToolCallSpan: (name: string) => boolean; /** Return whether a tool-call span name appears exactly `expectedCalls` times. */
+  hasNToolCallSpans: (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[];
@@ -2242,9 +2243,9 @@ type EvalFreshnessStatus = z$1.infer<typeof evalFreshnessStatusSchema>;
  */
 declare const evalStatAggregateSchema: z$1.ZodEnum<{
   sum: "sum";
+  avg: "avg";
   min: "min";
   max: "max";
-  avg: "avg";
   best: "best";
   worst: "worst";
 }>;
@@ -2274,9 +2275,9 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   kind: z$1.ZodLiteral<"duration">;
   aggregate: z$1.ZodOptional<z$1.ZodEnum<{
     sum: "sum";
+    avg: "avg";
     min: "min";
     max: "max";
-    avg: "avg";
     best: "best";
     worst: "worst";
   }>>;
@@ -2285,9 +2286,9 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   kind: z$1.ZodLiteral<"cacheHits">;
   aggregate: z$1.ZodOptional<z$1.ZodEnum<{
     sum: "sum";
+    avg: "avg";
     min: "min";
     max: "max";
-    avg: "avg";
     best: "best";
     worst: "worst";
   }>>;
@@ -2298,9 +2299,9 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   label: z$1.ZodOptional<z$1.ZodString>;
   aggregate: z$1.ZodEnum<{
     sum: "sum";
+    avg: "avg";
     min: "min";
     max: "max";
-    avg: "avg";
     best: "best";
     worst: "worst";
   }>;
@@ -2338,9 +2339,9 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
   kind: z$1.ZodLiteral<"duration">;
   aggregate: z$1.ZodOptional<z$1.ZodEnum<{
     sum: "sum";
+    avg: "avg";
     min: "min";
     max: "max";
-    avg: "avg";
     best: "best";
     worst: "worst";
   }>>;
@@ -2349,9 +2350,9 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
   kind: z$1.ZodLiteral<"cacheHits">;
   aggregate: z$1.ZodOptional<z$1.ZodEnum<{
     sum: "sum";
+    avg: "avg";
     min: "min";
     max: "max";
-    avg: "avg";
     best: "best";
     worst: "worst";
   }>>;
@@ -2362,9 +2363,9 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
   label: z$1.ZodOptional<z$1.ZodString>;
   aggregate: z$1.ZodEnum<{
     sum: "sum";
+    avg: "avg";
     min: "min";
     max: "max";
-    avg: "avg";
     best: "best";
     worst: "worst";
   }>;
@@ -2465,9 +2466,9 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
     kind: z$1.ZodLiteral<"duration">;
     aggregate: z$1.ZodOptional<z$1.ZodEnum<{
       sum: "sum";
+      avg: "avg";
       min: "min";
       max: "max";
-      avg: "avg";
       best: "best";
       worst: "worst";
     }>>;
@@ -2476,9 +2477,9 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
     kind: z$1.ZodLiteral<"cacheHits">;
     aggregate: z$1.ZodOptional<z$1.ZodEnum<{
       sum: "sum";
+      avg: "avg";
       min: "min";
       max: "max";
-      avg: "avg";
       best: "best";
       worst: "worst";
     }>>;
@@ -2489,9 +2490,9 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
     label: z$1.ZodOptional<z$1.ZodString>;
     aggregate: z$1.ZodEnum<{
       sum: "sum";
+      avg: "avg";
       min: "min";
       max: "max";
-      avg: "avg";
       best: "best";
       worst: "worst";
     }>;
@@ -2516,9 +2517,9 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
   }, z$1.core.$strip>], "kind">>>;
   defaultStatAggregate: z$1.ZodOptional<z$1.ZodEnum<{
     sum: "sum";
+    avg: "avg";
     min: "min";
     max: "max";
-    avg: "avg";
     best: "best";
     worst: "worst";
   }>>;
@@ -2539,8 +2540,8 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
       }>;
       label: z$1.ZodOptional<z$1.ZodString>;
       color: z$1.ZodOptional<z$1.ZodEnum<{
-        error: "error";
         success: "success";
+        error: "error";
         warning: "warning";
         accent: "accent";
         accentDim: "accentDim";
@@ -2555,16 +2556,16 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
       key: z$1.ZodString;
       aggregate: z$1.ZodEnum<{
         sum: "sum";
+        avg: "avg";
         min: "min";
         max: "max";
-        avg: "avg";
         latest: "latest";
         passThresholdRate: "passThresholdRate";
       }>;
       label: z$1.ZodOptional<z$1.ZodString>;
       color: z$1.ZodOptional<z$1.ZodEnum<{
-        error: "error";
         success: "success";
+        error: "error";
         warning: "warning";
         accent: "accent";
         accentDim: "accentDim";
@@ -2597,9 +2598,9 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
       key: z$1.ZodString;
       aggregate: z$1.ZodEnum<{
         sum: "sum";
+        avg: "avg";
         min: "min";
         max: "max";
-        avg: "avg";
         latest: "latest";
         passThresholdRate: "passThresholdRate";
       }>;
@@ -2698,9 +2699,9 @@ 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>;
@@ -2901,10 +2902,10 @@ 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 +2926,9 @@ 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 +3074,10 @@ 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>;
@@ -3194,10 +3195,10 @@ 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>;
@@ -3251,9 +3252,9 @@ 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<{
   sum: "sum";
+  avg: "avg";
   min: "min";
   max: "max";
-  avg: "avg";
   latest: "latest";
   passThresholdRate: "passThresholdRate";
 }>;
@@ -3264,8 +3265,8 @@ 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<{
-  error: "error";
   success: "success";
+  error: "error";
   warning: "warning";
   accent: "accent";
   accentDim: "accentDim";
@@ -3293,8 +3294,8 @@ declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   }>;
   label: z$1.ZodOptional<z$1.ZodString>;
   color: z$1.ZodOptional<z$1.ZodEnum<{
-    error: "error";
     success: "success";
+    error: "error";
     warning: "warning";
     accent: "accent";
     accentDim: "accentDim";
@@ -3309,16 +3310,16 @@ declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   key: z$1.ZodString;
   aggregate: z$1.ZodEnum<{
     sum: "sum";
+    avg: "avg";
     min: "min";
     max: "max";
-    avg: "avg";
     latest: "latest";
     passThresholdRate: "passThresholdRate";
   }>;
   label: z$1.ZodOptional<z$1.ZodString>;
   color: z$1.ZodOptional<z$1.ZodEnum<{
-    error: "error";
     success: "success";
+    error: "error";
     warning: "warning";
     accent: "accent";
     accentDim: "accentDim";
@@ -3344,9 +3345,9 @@ declare const evalChartTooltipExtraSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObj
   key: z$1.ZodString;
   aggregate: z$1.ZodEnum<{
     sum: "sum";
+    avg: "avg";
     min: "min";
     max: "max";
-    avg: "avg";
     latest: "latest";
     passThresholdRate: "passThresholdRate";
   }>;
@@ -3376,8 +3377,8 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
     }>;
     label: z$1.ZodOptional<z$1.ZodString>;
     color: z$1.ZodOptional<z$1.ZodEnum<{
-      error: "error";
       success: "success";
+      error: "error";
       warning: "warning";
       accent: "accent";
       accentDim: "accentDim";
@@ -3392,16 +3393,16 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
       sum: "sum";
+      avg: "avg";
       min: "min";
       max: "max";
-      avg: "avg";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
     label: z$1.ZodOptional<z$1.ZodString>;
     color: z$1.ZodOptional<z$1.ZodEnum<{
-      error: "error";
       success: "success";
+      error: "error";
       warning: "warning";
       accent: "accent";
       accentDim: "accentDim";
@@ -3434,9 +3435,9 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
       sum: "sum";
+      avg: "avg";
       min: "min";
       max: "max";
-      avg: "avg";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
@@ -3466,8 +3467,8 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
     }>;
     label: z$1.ZodOptional<z$1.ZodString>;
     color: z$1.ZodOptional<z$1.ZodEnum<{
-      error: "error";
       success: "success";
+      error: "error";
       warning: "warning";
       accent: "accent";
       accentDim: "accentDim";
@@ -3482,16 +3483,16 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
       sum: "sum";
+      avg: "avg";
       min: "min";
       max: "max";
-      avg: "avg";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
     label: z$1.ZodOptional<z$1.ZodString>;
     color: z$1.ZodOptional<z$1.ZodEnum<{
-      error: "error";
       success: "success";
+      error: "error";
       warning: "warning";
       accent: "accent";
       accentDim: "accentDim";
@@ -3524,9 +3525,9 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
       sum: "sum";
+      avg: "avg";
       min: "min";
       max: "max";
-      avg: "avg";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
@@ -3555,8 +3556,8 @@ 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 +3571,9 @@ 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. */
@@ -3694,10 +3695,11 @@ 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 the names of every span with `kind: 'tool'` or `kind: 'tool_call'`. */
+  listToolCallSpanNames: () => string[]; /** Return whether a tool-call span has a name exactly matching `name`. */
+  hasToolCallSpan: (name: string) => boolean; /** Return whether a tool-call span name appears exactly `expectedCalls` times. */
+  hasNToolCallSpans: (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[];
@@ -4290,15 +4292,26 @@ type AgentEvalsConfig$1 = {
     /** Disable the cache entirely; spans with `cache` options execute as if uncached. */enabled?: boolean; /** Override the directory used to persist cache entries. */
     dir?: string;
     /**
-     * Default maximum entries retained for each cache namespace. Defaults to
-     * `100`; non-positive or non-finite values fall back to the default.
-     */
-    maxEntriesPerNamespace?: number;
-    /**
-     * Exact namespace-specific retention caps. Values override
-     * `maxEntriesPerNamespace` for matching namespaces.
+     * Maximum entries retained per cache namespace.
+     *
+     * Pass a number to set the default cap for every namespace. Pass an object
+     * to set a default cap plus exact namespace-specific caps. Non-positive or
+     * non-finite values fall back to the default.
+     *
+     * @example
+     * ```ts
+     * cache: {
+     *   maxEntries: {
+     *     default: 50,
+     *     namespaces: { 'receipt-audit.receipt-audit-context': 200 },
+     *   },
+     * }
+     * ```
      */
-    maxEntriesByNamespace?: Record<string, number>;
+    maxEntries?: number | {
+      default?: number;
+      namespaces?: Record<string, number>;
+    };
     /**
      * Milliseconds the runner waits after becoming idle before pruning indexed
      * cache entries. Defaults to `5000`; non-positive or non-finite values use
@@ -4309,8 +4322,7 @@ type AgentEvalsConfig$1 = {
      * Minimum milliseconds between `lastAccessedAt` index rewrites for repeated
      * cache hits. Defaults to four hours. Set to `0` to record every hit.
      */
-    lastAccessedAtUpdateIntervalMs?: number; /** Legacy alias for `maxEntriesPerNamespace`, retained so older config files keep working. */
-    maxEntriesPerEval?: number;
+    lastAccessedAtUpdateIntervalMs?: number;
   };
 };
 /** Zod schema for validating `agent-evals.config.ts` input. */
@@ -4519,9 +4531,9 @@ 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>;
@@ -4535,17 +4547,17 @@ 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>;
@@ -4562,10 +4574,10 @@ 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>;
@@ -4761,8 +4773,8 @@ 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>;
@@ -4868,8 +4880,8 @@ 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;
@@ -4879,8 +4891,8 @@ 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>;
@@ -4986,8 +4998,8 @@ 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>;
@@ -5084,8 +5096,8 @@ 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;
@@ -5095,8 +5107,8 @@ 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>;
@@ -5202,8 +5214,8 @@ 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>;
@@ -5308,8 +5320,8 @@ 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;
@@ -5319,8 +5331,8 @@ 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>;
@@ -5494,8 +5506,8 @@ 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. */
@@ -5503,8 +5515,8 @@ 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;
@@ -5518,8 +5530,8 @@ 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>>;
@@ -5531,9 +5543,9 @@ 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>>;

package/dist/index.mjs

@@ -1,4 +1,4 @@
-import { $ as setScopeCacheContext, A as repoFile, B as evalTime, C as evalTracer, D as deserializeCacheValue, E as deserializeCacheRecording, F as EvalRuntimeUsageError, Ft as getEvalRegistry, H as getEvalCaseInput, I as appendToEvalOutput, J as runInEvalRuntimeScope, K as mergeEvalOutput, M as readManualInputFile, N as evalExpect, O as serializeCacheRecording, P as EvalAssertionError, Q as setEvalOutput, R as evalAssert, S as evalSpan, T as hashCacheKeySync, U as incrementEvalOutput, V as getCurrentScope, W as isInEvalScope, X as runInExistingEvalScope, Y as runInEvalScope, at as extractApiCalls, b as buildTraceTree, ct as simulateTokenAllocation, et as startEvalBackgroundJob, it as extractCacheHits, j as manualInputFileValueSchema, k as serializeCacheValue, ot as extractLlmCalls, q as nextEvalId, rt as extractCacheEntries, st as simulateLlmCallCost, ut as getNestedAttribute, w as hashCacheKey, x as captureEvalSpanError, y as z, z as evalLog } from "./runExecution-CFw0MQFs.mjs";
-import { a as materializeManualInputFiles, i as isManualInputFileValue, n as createRunner, o as stageManualInputFile, r as cleanupStagedManualInputFiles, s as stageManualInputFileFromPath, t as runCli } from "./cli-SP4kEtYL.mjs";
-import { n as matchesEvalTags, t as defineEval } from "./src-7GbQj1sb.mjs";
+import { $ as setEvalOutput, A as serializeCacheValue, B as evalLog, C as evalSpan, D as deserializeCacheRecording, E as hashCacheKeySync, F as EvalAssertionError, G as isInEvalScope, H as getCurrentScope, I as EvalRuntimeUsageError, It as getEvalRegistry, J as nextEvalId, L as appendToEvalOutput, M as manualInputFileValueSchema, N as readManualInputFile, O as deserializeCacheValue, P as evalExpect, S as captureEvalSpanError, T as hashCacheKey, U as getEvalCaseInput, V as evalTime, W as incrementEvalOutput, X as runInEvalScope, Y as runInEvalRuntimeScope, Z as runInExistingEvalScope, at as extractCacheHits, b as z, ct as simulateLlmCallCost, dt as getNestedAttribute, et as setScopeCacheContext, it as extractCacheEntries, j as repoFile, k as serializeCacheRecording, lt as simulateTokenAllocation, ot as extractApiCalls, q as mergeEvalOutput, st as extractLlmCalls, tt as startEvalBackgroundJob, w as evalTracer, x as buildTraceTree, z as evalAssert } from "./runExecution-BMnJXWhN.mjs";
+import { a as materializeManualInputFiles, i as isManualInputFileValue, n as createRunner, o as stageManualInputFile, r as cleanupStagedManualInputFiles, s as stageManualInputFileFromPath, t as runCli } from "./cli-ClAkjTvo.mjs";
+import { n as matchesEvalTags, t as defineEval } from "./src-DfzidkYr.mjs";
 export { EvalAssertionError, EvalRuntimeUsageError, 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 };

package/dist/runChild.mjs

@@ -1,5 +1,5 @@
-import { At as evalStatsConfigSchema, L as configureEvalRunLogs, Mt as evalChartsConfigSchema, Nt as columnDefSchema, Tt as buildEvalKey, bt as runManifestSchema, jt as manualInputDescriptorSchema, kt as evalStatAggregateSchema, l as registerAgentEvalsPackageResolutionHooks, p as loadConfig, tt as createRunRequestSchema, v as createFsCacheStore, xt as runSummarySchema } from "./runExecution-CFw0MQFs.mjs";
-import { S as parseEvalDiscovery, m as persistRunState, r as getTargetEvals$1, t as executeRun } from "./runOrchestration-CxjiQmof.mjs";
+import { At as evalStatAggregateSchema, Et as buildEvalKey, Mt as manualInputDescriptorSchema, Nt as evalChartsConfigSchema, Pt as columnDefSchema, R as configureEvalRunLogs, St as runSummarySchema, jt as evalStatsConfigSchema, l as registerAgentEvalsPackageResolutionHooks, nt as createRunRequestSchema, p as loadConfig, v as createFsCacheStore, xt as runManifestSchema, y as getCacheRetentionOptions } from "./runExecution-BMnJXWhN.mjs";
+import { S as parseEvalDiscovery, m as persistRunState, r as getTargetEvals$1, t as executeRun } from "./runOrchestration-CvbTAoEb.mjs";
 import { z } from "zod/v4";
 import { readFile } from "node:fs/promises";
 import { relative } from "node:path";
@@ -137,11 +137,12 @@ async function main() {
 	registerAgentEvalsPackageResolutionHooks();
 	const config = await loadConfig();
 	configureEvalRunLogs({ captureConsole: config.runLogs?.captureConsole !== false });
+	const cacheRetentionOptions = getCacheRetentionOptions(config.cache);
 	const cacheStore = createFsCacheStore({
 		workspaceRoot: context.workspaceRoot,
 		dir: config.cache?.dir,
-		maxEntriesPerNamespace: config.cache?.maxEntriesPerNamespace ?? config.cache?.maxEntriesPerEval,
-		maxEntriesByNamespace: config.cache?.maxEntriesByNamespace,
+		maxEntriesPerNamespace: cacheRetentionOptions.maxEntriesPerNamespace,
+		maxEntriesByNamespace: cacheRetentionOptions.maxEntriesByNamespace,
 		lastAccessedAtUpdateIntervalMs: config.cache?.lastAccessedAtUpdateIntervalMs
 	});
 	const evalMetas = await discoverRunEvals({