@ls-stack/agent-eval 0.3.0 → 0.5.0

package/dist/index.d.mts

@@ -122,19 +122,13 @@ declare const cellValueSchema: z.ZodUnion<readonly [z.ZodType<string | number |
 type CellValue = z.infer<typeof cellValueSchema>;
 //#endregion
 //#region ../shared/src/schemas/trace.d.ts
-/** Schema for the semantic categories used to classify trace spans. */
-declare const traceSpanKindSchema: z.ZodEnum<{
-  eval: "eval";
-  agent: "agent";
-  llm: "llm";
-  tool: "tool";
-  retrieval: "retrieval";
-  scorer: "scorer";
-  checkpoint: "checkpoint";
-  custom: "custom";
-}>;
-/** Semantic category used to classify a trace span in the UI. */
-type TraceSpanKind = z.infer<typeof traceSpanKindSchema>;
+/**
+ * Schema for span categories recorded in traces.
+ *
+ * The value is intentionally open-ended so external tracers can preserve their
+ * native span kinds instead of collapsing them into the built-in categories.
+ */
+declare const traceSpanKindSchema: z.ZodString;
 /** Schema for the supported presentation formats of trace attributes. */
 declare const traceAttributeDisplayFormatSchema: z.ZodEnum<{
   string: "string";
@@ -301,21 +295,21 @@ declare const traceDisplayInputConfigSchema: z.ZodObject<{
 }, z.core.$strip>;
 /** Trace display configuration authored by users in config or eval files. */
 type TraceDisplayInputConfig = z.infer<typeof traceDisplayInputConfigSchema>;
+/** Schema for an error attached to a trace span. */
+declare const traceSpanErrorSchema: z.ZodObject<{
+  name: z.ZodOptional<z.ZodString>;
+  message: z.ZodString;
+  stack: z.ZodOptional<z.ZodString>;
+  capturedAt: z.ZodOptional<z.ZodString>;
+}, z.core.$catchall<z.ZodUnknown>>;
+/** Error payload stored on a trace span. */
+type EvalTraceSpanError = z.infer<typeof traceSpanErrorSchema>;
 /** Schema for a persisted trace span captured during case execution. */
 declare const traceSpanSchema: z.ZodObject<{
   id: z.ZodString;
   parentId: z.ZodNullable<z.ZodString>;
   caseId: z.ZodString;
-  kind: z.ZodEnum<{
-    eval: "eval";
-    agent: "agent";
-    llm: "llm";
-    tool: "tool";
-    retrieval: "retrieval";
-    scorer: "scorer";
-    checkpoint: "checkpoint";
-    custom: "custom";
-  }>;
+  kind: z.ZodString;
   name: z.ZodString;
   startedAt: z.ZodString;
   endedAt: z.ZodNullable<z.ZodString>;
@@ -330,7 +324,14 @@ declare const traceSpanSchema: z.ZodObject<{
     name: z.ZodOptional<z.ZodString>;
     message: z.ZodString;
     stack: z.ZodOptional<z.ZodString>;
-  }, z.core.$strip>>;
+    capturedAt: z.ZodOptional<z.ZodString>;
+  }, z.core.$catchall<z.ZodUnknown>>>;
+  errors: z.ZodOptional<z.ZodArray<z.ZodObject<{
+    name: z.ZodOptional<z.ZodString>;
+    message: z.ZodString;
+    stack: z.ZodOptional<z.ZodString>;
+    capturedAt: z.ZodOptional<z.ZodString>;
+  }, z.core.$catchall<z.ZodUnknown>>>>;
 }, z.core.$strip>;
 /** Persisted trace span shape stored for each eval case run. */
 type EvalTraceSpan = z.infer<typeof traceSpanSchema>;
@@ -650,16 +651,7 @@ declare const scoreTraceSchema: z.ZodObject<{
     id: z.ZodString;
     parentId: z.ZodNullable<z.ZodString>;
     caseId: z.ZodString;
-    kind: z.ZodEnum<{
-      custom: "custom";
-      eval: "eval";
-      agent: "agent";
-      llm: "llm";
-      tool: "tool";
-      retrieval: "retrieval";
-      scorer: "scorer";
-      checkpoint: "checkpoint";
-    }>;
+    kind: z.ZodString;
     name: z.ZodString;
     startedAt: z.ZodString;
     endedAt: z.ZodNullable<z.ZodString>;
@@ -674,7 +666,14 @@ declare const scoreTraceSchema: z.ZodObject<{
       name: z.ZodOptional<z.ZodString>;
       message: z.ZodString;
       stack: z.ZodOptional<z.ZodString>;
-    }, z.core.$strip>>;
+      capturedAt: z.ZodOptional<z.ZodString>;
+    }, z.core.$catchall<z.ZodUnknown>>>;
+    errors: z.ZodOptional<z.ZodArray<z.ZodObject<{
+      name: z.ZodOptional<z.ZodString>;
+      message: z.ZodString;
+      stack: z.ZodOptional<z.ZodString>;
+      capturedAt: z.ZodOptional<z.ZodString>;
+    }, z.core.$catchall<z.ZodUnknown>>>>;
   }, z.core.$strip>>;
   traceDisplay: z.ZodObject<{
     attributes: z.ZodOptional<z.ZodArray<z.ZodObject<{
@@ -724,16 +723,7 @@ declare const caseDetailSchema: z.ZodObject<{
     id: z.ZodString;
     parentId: z.ZodNullable<z.ZodString>;
     caseId: z.ZodString;
-    kind: z.ZodEnum<{
-      custom: "custom";
-      eval: "eval";
-      agent: "agent";
-      llm: "llm";
-      tool: "tool";
-      retrieval: "retrieval";
-      scorer: "scorer";
-      checkpoint: "checkpoint";
-    }>;
+    kind: z.ZodString;
     name: z.ZodString;
     startedAt: z.ZodString;
     endedAt: z.ZodNullable<z.ZodString>;
@@ -748,7 +738,14 @@ declare const caseDetailSchema: z.ZodObject<{
       name: z.ZodOptional<z.ZodString>;
       message: z.ZodString;
       stack: z.ZodOptional<z.ZodString>;
-    }, z.core.$strip>>;
+      capturedAt: z.ZodOptional<z.ZodString>;
+    }, z.core.$catchall<z.ZodUnknown>>>;
+    errors: z.ZodOptional<z.ZodArray<z.ZodObject<{
+      name: z.ZodOptional<z.ZodString>;
+      message: z.ZodString;
+      stack: z.ZodOptional<z.ZodString>;
+      capturedAt: z.ZodOptional<z.ZodString>;
+    }, z.core.$catchall<z.ZodUnknown>>>>;
   }, z.core.$strip>>;
   traceDisplay: z.ZodObject<{
     attributes: z.ZodOptional<z.ZodArray<z.ZodObject<{
@@ -783,16 +780,7 @@ declare const caseDetailSchema: z.ZodObject<{
       id: z.ZodString;
       parentId: z.ZodNullable<z.ZodString>;
       caseId: z.ZodString;
-      kind: z.ZodEnum<{
-        custom: "custom";
-        eval: "eval";
-        agent: "agent";
-        llm: "llm";
-        tool: "tool";
-        retrieval: "retrieval";
-        scorer: "scorer";
-        checkpoint: "checkpoint";
-      }>;
+      kind: z.ZodString;
       name: z.ZodString;
       startedAt: z.ZodString;
       endedAt: z.ZodNullable<z.ZodString>;
@@ -807,7 +795,14 @@ declare const caseDetailSchema: z.ZodObject<{
         name: z.ZodOptional<z.ZodString>;
         message: z.ZodString;
         stack: z.ZodOptional<z.ZodString>;
-      }, z.core.$strip>>;
+        capturedAt: z.ZodOptional<z.ZodString>;
+      }, z.core.$catchall<z.ZodUnknown>>>;
+      errors: z.ZodOptional<z.ZodArray<z.ZodObject<{
+        name: z.ZodOptional<z.ZodString>;
+        message: z.ZodString;
+        stack: z.ZodOptional<z.ZodString>;
+        capturedAt: z.ZodOptional<z.ZodString>;
+      }, z.core.$catchall<z.ZodUnknown>>>>;
     }, z.core.$strip>>;
     traceDisplay: z.ZodObject<{
       attributes: z.ZodOptional<z.ZodArray<z.ZodObject<{
@@ -1491,16 +1486,7 @@ declare const cacheListItemSchema: z.ZodObject<{
   key: z.ZodString;
   namespace: z.ZodString;
   spanName: z.ZodString;
-  spanKind: z.ZodEnum<{
-    eval: "eval";
-    agent: "agent";
-    llm: "llm";
-    tool: "tool";
-    retrieval: "retrieval";
-    scorer: "scorer";
-    checkpoint: "checkpoint";
-    custom: "custom";
-  }>;
+  spanKind: z.ZodString;
   storedAt: z.ZodString;
   codeFingerprint: z.ZodString;
   sizeBytes: z.ZodNumber;
@@ -1509,15 +1495,12 @@ declare const cacheListItemSchema: z.ZodObject<{
 type CacheListItem = z.infer<typeof cacheListItemSchema>;
 /** Serialized nested span captured while recording a cached operation. */
 type SerializedCacheSpan = {
-  kind: TraceSpanKind;
+  kind: string;
   name: string;
   attributes?: Record<string, unknown>;
   status: 'running' | 'ok' | 'error' | 'cancelled';
-  error?: {
-    name?: string;
-    message: string;
-    stack?: string;
-  };
+  error?: EvalTraceSpanError;
+  errors?: EvalTraceSpanError[];
   children: SerializedCacheSpan[];
 };
 /** Zod schema for `SerializedCacheSpan`, defined lazily for recursion. */
@@ -1550,6 +1533,24 @@ type CacheRecordingOp = z.infer<typeof cacheRecordingOpSchema>;
 declare const cacheRecordingSchema: z.ZodObject<{
   returnValue: z.ZodUnknown;
   finalAttributes: z.ZodRecord<z.ZodString, z.ZodUnknown>;
+  finalStatus: z.ZodOptional<z.ZodEnum<{
+    error: "error";
+    running: "running";
+    ok: "ok";
+    cancelled: "cancelled";
+  }>>;
+  finalError: z.ZodOptional<z.ZodObject<{
+    name: z.ZodOptional<z.ZodString>;
+    message: z.ZodString;
+    stack: z.ZodOptional<z.ZodString>;
+    capturedAt: z.ZodOptional<z.ZodString>;
+  }, z.core.$catchall<z.ZodUnknown>>>;
+  finalErrors: z.ZodOptional<z.ZodArray<z.ZodObject<{
+    name: z.ZodOptional<z.ZodString>;
+    message: z.ZodString;
+    stack: z.ZodOptional<z.ZodString>;
+    capturedAt: z.ZodOptional<z.ZodString>;
+  }, z.core.$catchall<z.ZodUnknown>>>>;
   ops: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
     kind: z.ZodLiteral<"setOutput">;
     key: z.ZodString;
@@ -1575,21 +1576,30 @@ declare const cacheEntrySchema: z.ZodObject<{
   key: z.ZodString;
   namespace: z.ZodString;
   spanName: z.ZodString;
-  spanKind: z.ZodEnum<{
-    eval: "eval";
-    agent: "agent";
-    llm: "llm";
-    tool: "tool";
-    retrieval: "retrieval";
-    scorer: "scorer";
-    checkpoint: "checkpoint";
-    custom: "custom";
-  }>;
+  spanKind: z.ZodString;
   storedAt: z.ZodString;
   codeFingerprint: z.ZodString;
   recording: z.ZodObject<{
     returnValue: z.ZodUnknown;
     finalAttributes: z.ZodRecord<z.ZodString, z.ZodUnknown>;
+    finalStatus: z.ZodOptional<z.ZodEnum<{
+      error: "error";
+      running: "running";
+      ok: "ok";
+      cancelled: "cancelled";
+    }>>;
+    finalError: z.ZodOptional<z.ZodObject<{
+      name: z.ZodOptional<z.ZodString>;
+      message: z.ZodString;
+      stack: z.ZodOptional<z.ZodString>;
+      capturedAt: z.ZodOptional<z.ZodString>;
+    }, z.core.$catchall<z.ZodUnknown>>>;
+    finalErrors: z.ZodOptional<z.ZodArray<z.ZodObject<{
+      name: z.ZodOptional<z.ZodString>;
+      message: z.ZodString;
+      stack: z.ZodOptional<z.ZodString>;
+      capturedAt: z.ZodOptional<z.ZodString>;
+    }, z.core.$catchall<z.ZodUnknown>>>>;
     ops: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
       kind: z.ZodLiteral<"setOutput">;
       key: z.ZodString;
@@ -1619,21 +1629,30 @@ declare const cacheFileSchema: z.ZodObject<{
     key: z.ZodString;
     namespace: z.ZodString;
     spanName: z.ZodString;
-    spanKind: z.ZodEnum<{
-      eval: "eval";
-      agent: "agent";
-      llm: "llm";
-      tool: "tool";
-      retrieval: "retrieval";
-      scorer: "scorer";
-      checkpoint: "checkpoint";
-      custom: "custom";
-    }>;
+    spanKind: z.ZodString;
     storedAt: z.ZodString;
     codeFingerprint: z.ZodString;
     recording: z.ZodObject<{
       returnValue: z.ZodUnknown;
       finalAttributes: z.ZodRecord<z.ZodString, z.ZodUnknown>;
+      finalStatus: z.ZodOptional<z.ZodEnum<{
+        error: "error";
+        running: "running";
+        ok: "ok";
+        cancelled: "cancelled";
+      }>>;
+      finalError: z.ZodOptional<z.ZodObject<{
+        name: z.ZodOptional<z.ZodString>;
+        message: z.ZodString;
+        stack: z.ZodOptional<z.ZodString>;
+        capturedAt: z.ZodOptional<z.ZodString>;
+      }, z.core.$catchall<z.ZodUnknown>>>;
+      finalErrors: z.ZodOptional<z.ZodArray<z.ZodObject<{
+        name: z.ZodOptional<z.ZodString>;
+        message: z.ZodString;
+        stack: z.ZodOptional<z.ZodString>;
+        capturedAt: z.ZodOptional<z.ZodString>;
+      }, z.core.$catchall<z.ZodUnknown>>>>;
       ops: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
         kind: z.ZodLiteral<"setOutput">;
         key: z.ZodString;
@@ -1703,7 +1722,7 @@ type EvalTraceTree = {
   spans: EvalTraceSpan[];
   rootSpans: EvalTraceSpan[];
   findSpan: (name: string) => EvalTraceSpan | undefined;
-  findSpansByKind: (kind: EvalTraceSpan['kind']) => EvalTraceSpan[];
+  findSpansByKind: (kind: string) => EvalTraceSpan[];
   flattenDfs: () => EvalTraceSpan[];
   checkpoints: Map<string, unknown>;
 };
@@ -1963,7 +1982,7 @@ type TraceSpanTimestamp = Date | string;
 type TraceExternalSpanStartInfo = {
   /** Stable span id from the upstream tracer. Generated when omitted. */id?: string; /** Parent span id from the upstream tracer. Defaults to the active eval span. */
   parentId?: string | null; /** Semantic category used by the trace UI. */
-  kind: EvalTraceSpan['kind']; /** Display name for the span. */
+  kind: string; /** Display name for the span. */
   name: string; /** Span start time. Defaults to now. */
   startedAt?: TraceSpanTimestamp; /** Initial span attributes. Later updates merge into this object. */
   attributes?: Record<string, unknown>;
@@ -1984,7 +2003,7 @@ type TraceExternalSpanEndInfo = TraceExternalSpanUpdateInfo & {
 type TraceExternalSpanRecordInfo = {
   /** Stable span id from the upstream tracer. Generated when omitted. */id?: string; /** Parent span id from the upstream tracer. Defaults to the active eval span. */
   parentId?: string | null; /** Semantic category used by the trace UI. */
-  kind: EvalTraceSpan['kind']; /** Display name for the span. */
+  kind: string; /** Display name for the span. */
   name: string; /** Span start time. Defaults to now. */
   startedAt?: TraceSpanTimestamp; /** Span end time. Defaults to the start time. */
   endedAt?: TraceSpanTimestamp | null; /** Final span status. Defaults to `ok`. */
@@ -2007,8 +2026,15 @@ declare function recordExternalSpan(info: TraceExternalSpanRecordInfo): string;
  * Calls are no-ops when executed outside of `evalTracer.span(...)`.
  */
 declare const evalSpan: TraceActiveSpan;
+/**
+ * Attach one or more recoverable errors to the active eval span.
+ *
+ * The active span is marked as `error` even if its callback later completes
+ * without throwing. Calls outside `evalTracer.span(...)` are ignored.
+ */
+declare function captureEvalSpanError(errorOrErrors: unknown, ...additionalErrors: readonly unknown[]): void;
 type TraceSpanInfoBase = {
-  kind: EvalTraceSpan['kind'];
+  kind: string;
   name: string;
   attributes?: Record<string, unknown>;
 };
@@ -2117,7 +2143,8 @@ type EvalRunner = {
   cancelRun(id: string): void; /** Return full details for a single case in a run, when available. */
   getCaseDetail(runId: string, caseId: string): CaseDetail | undefined; /** Subscribe to streamed events for a specific run. */
   subscribe(runId: string, listener: (event: SseEnvelope) => void): () => void; /** Subscribe to discovery updates triggered by file changes or manual refresh. */
-  subscribeDiscovery(listener: (event: SseEnvelope) => void): () => void; /** Resolve the workspace root backing this runner instance. */
+  subscribeDiscovery(listener: (event: SseEnvelope) => void): () => void; /** Stop background filesystem watchers owned by this runner instance. */
+  close(): Promise<void>; /** Resolve the workspace root backing this runner instance. */
   getWorkspaceRoot(): string; /** Resolve a persisted artifact path when artifact storage is supported. */
   getArtifactPath(artifactId: string): string | undefined; /** Return summaries for every persisted cache entry in the workspace. */
   listCache(): Promise<CacheListItem[]>;
@@ -2175,4 +2202,4 @@ declare function createRunner({
  */
 declare function runCli(argv: string[]): Promise<void>;
 //#endregion
-export { type AgentEvalsConfig, type AssertionFailure, type CacheAdapter, type CacheEntry, type CacheFile, type CacheListItem, type CacheMode, type CacheRecording, type CacheRecordingFrame, type CacheRecordingOp, type CacheScopeContext, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type CreateRunRequest, type DerivedStatus, EvalAssertionError, type EvalCase, type EvalCaseScope, type EvalChartAggregate, type EvalChartAxis, type EvalChartBuiltinMetric, type EvalChartColor, type EvalChartConfig, type EvalChartMetric, type EvalChartTooltipExtra, type EvalChartType, type EvalChartsConfig, type EvalColumnOverride, type EvalColumns, type EvalDefinition, type EvalDeriveContext, type EvalDisplayStatus, type EvalExecuteContext, type EvalFreshnessStatus, type EvalManualScoreDef, type EvalRunner, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, type EvalTraceSpan, type EvalTraceTree, type FileRef, type JsonCell, type NumberDisplayOptions, type RepoFileRef, type RunArtifactRef, type RunInEvalScopeOptions, type RunManifest, type RunSummary, type ScalarCell, type ScopedCaseSummary, type ScoreTrace, type SerializedCacheSpan, type SpanCacheOptions, type SseEnvelope, type SseEventType, type TraceActiveSpan, type TraceAttributeDisplay, type TraceAttributeDisplayFormat, type TraceAttributeDisplayInput, type TraceAttributeDisplayPlacement, type TraceAttributeTransform, type TraceAttributeTransformContext, type TraceDisplayConfig, type TraceDisplayInputConfig, type TraceSpanInfo, type TraceSpanKind, type TrialSelectionMode, type UpdateManualScoreRequest, agentEvalsConfigSchema, assertionFailureSchema, buildTraceTree, cacheEntrySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheRecordingOpSchema, cacheRecordingSchema, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, fileRefSchema, getCurrentScope, getEvalDisplayStatus, getEvalRegistry, getEvalTitle, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, numberDisplayOptionsSchema, repoFile, repoFileRefSchema, runArtifactRefSchema, runCli, runInEvalScope, runManifestSchema, runSummarySchema, scoreTraceSchema, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanKindSchema, traceSpanSchema, trialSelectionModeSchema, updateManualScoreRequestSchema };
+export { type AgentEvalsConfig, type AssertionFailure, type CacheAdapter, type CacheEntry, type CacheFile, type CacheListItem, type CacheMode, type CacheRecording, type CacheRecordingFrame, type CacheRecordingOp, type CacheScopeContext, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type CreateRunRequest, type DerivedStatus, EvalAssertionError, type EvalCase, type EvalCaseScope, type EvalChartAggregate, type EvalChartAxis, type EvalChartBuiltinMetric, type EvalChartColor, type EvalChartConfig, type EvalChartMetric, type EvalChartTooltipExtra, type EvalChartType, type EvalChartsConfig, type EvalColumnOverride, type EvalColumns, type EvalDefinition, type EvalDeriveContext, type EvalDisplayStatus, type EvalExecuteContext, type EvalFreshnessStatus, type EvalManualScoreDef, type EvalRunner, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, type EvalTraceSpan, type EvalTraceSpanError, type EvalTraceTree, type FileRef, type JsonCell, type NumberDisplayOptions, type RepoFileRef, type RunArtifactRef, type RunInEvalScopeOptions, type RunManifest, type RunSummary, type ScalarCell, type ScopedCaseSummary, type ScoreTrace, type SerializedCacheSpan, type SpanCacheOptions, type SseEnvelope, type SseEventType, type TraceActiveSpan, type TraceAttributeDisplay, type TraceAttributeDisplayFormat, type TraceAttributeDisplayInput, type TraceAttributeDisplayPlacement, type TraceAttributeTransform, type TraceAttributeTransformContext, type TraceDisplayConfig, type TraceDisplayInputConfig, type TraceSpanInfo, type TrialSelectionMode, type UpdateManualScoreRequest, agentEvalsConfigSchema, assertionFailureSchema, buildTraceTree, cacheEntrySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheRecordingOpSchema, cacheRecordingSchema, captureEvalSpanError, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, fileRefSchema, getCurrentScope, getEvalDisplayStatus, getEvalRegistry, getEvalTitle, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, numberDisplayOptionsSchema, repoFile, repoFileRefSchema, runArtifactRefSchema, runCli, runInEvalScope, runManifestSchema, runSummarySchema, scoreTraceSchema, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanErrorSchema, traceSpanKindSchema, traceSpanSchema, trialSelectionModeSchema, updateManualScoreRequestSchema };

package/dist/index.mjs

@@ -1,3 +1,3 @@
-import { $ as fileRefSchema, A as evalSummarySchema, B as evalChartsConfigSchema, C as assertionFailureSchema, D as evalStatAggregateSchema, E as evalFreshnessStatusSchema, F as evalChartColorSchema, G as traceDisplayConfigSchema, H as traceAttributeDisplayInputSchema, I as evalChartConfigSchema, J as traceSpanSchema, K as traceDisplayInputConfigSchema, L as evalChartMetricSchema, M as evalChartAggregateSchema, N as evalChartAxisSchema, O as evalStatItemSchema, P as evalChartBuiltinMetricSchema, Q as columnKindSchema, R as evalChartTooltipExtraSchema, S as spanCacheOptionsSchema, T as caseRowSchema, U as traceAttributeDisplayPlacementSchema, V as traceAttributeDisplayFormatSchema, W as traceAttributeDisplaySchema, X as columnDefSchema, Y as cellValueSchema, Z as columnFormatSchema, _ as cacheListItemSchema, _t as repoFile, a as sseEnvelopeSchema, at as evalSpan, b as cacheRecordingSchema, c as deriveScopedSummaryFromCases, ct as hashCacheKeySync, d as runManifestSchema, dt as getCurrentScope, et as jsonCellSchema, f as runSummarySchema, ft as incrementEvalOutput, g as cacheFileSchema, gt as setScopeCacheContext, h as cacheEntrySchema, ht as setEvalOutput, i as updateManualScoreRequestSchema, it as buildTraceTree, j as scoreTraceSchema, k as evalStatsConfigSchema, l as deriveStatusFromCaseRows, lt as EvalAssertionError, m as trialSelectionModeSchema, mt as runInEvalScope, n as createRunner, nt as repoFileRefSchema, o as getEvalTitle, ot as evalTracer, p as agentEvalsConfigSchema, pt as isInEvalScope, q as traceSpanKindSchema, r as createRunRequestSchema, rt as runArtifactRefSchema, s as getEvalDisplayStatus, st as hashCacheKey, t as runCli, tt as numberDisplayOptionsSchema, u as deriveStatusFromChildStatuses, ut as evalAssert, v as cacheModeSchema, vt as defineEval, w as caseDetailSchema, x as serializedCacheSpanSchema, y as cacheRecordingOpSchema, yt as getEvalRegistry, z as evalChartTypeSchema } from "./cli-Dw9et3_Z.mjs";
-import "./src-CXclO9ZI.mjs";
-export { EvalAssertionError, agentEvalsConfigSchema, assertionFailureSchema, buildTraceTree, cacheEntrySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheRecordingOpSchema, cacheRecordingSchema, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, fileRefSchema, getCurrentScope, getEvalDisplayStatus, getEvalRegistry, getEvalTitle, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, numberDisplayOptionsSchema, repoFile, repoFileRefSchema, runArtifactRefSchema, runCli, runInEvalScope, runManifestSchema, runSummarySchema, scoreTraceSchema, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanKindSchema, traceSpanSchema, trialSelectionModeSchema, updateManualScoreRequestSchema };
+import { $ as columnKindSchema, A as evalSummarySchema, B as evalChartsConfigSchema, C as assertionFailureSchema, D as evalStatAggregateSchema, E as evalFreshnessStatusSchema, F as evalChartColorSchema, G as traceDisplayConfigSchema, H as traceAttributeDisplayInputSchema, I as evalChartConfigSchema, J as traceSpanKindSchema, K as traceDisplayInputConfigSchema, L as evalChartMetricSchema, M as evalChartAggregateSchema, N as evalChartAxisSchema, O as evalStatItemSchema, P as evalChartBuiltinMetricSchema, Q as columnFormatSchema, R as evalChartTooltipExtraSchema, S as spanCacheOptionsSchema, T as caseRowSchema, U as traceAttributeDisplayPlacementSchema, V as traceAttributeDisplayFormatSchema, W as traceAttributeDisplaySchema, X as cellValueSchema, Y as traceSpanSchema, Z as columnDefSchema, _ as cacheListItemSchema, _t as setEvalOutput, a as sseEnvelopeSchema, at as buildTraceTree, b as cacheRecordingSchema, bt as defineEval, c as deriveScopedSummaryFromCases, ct as evalTracer, d as runManifestSchema, dt as EvalAssertionError, et as fileRefSchema, f as runSummarySchema, ft as evalAssert, g as cacheFileSchema, gt as runInEvalScope, h as cacheEntrySchema, ht as isInEvalScope, i as updateManualScoreRequestSchema, it as runArtifactRefSchema, j as scoreTraceSchema, k as evalStatsConfigSchema, l as deriveStatusFromCaseRows, lt as hashCacheKey, m as trialSelectionModeSchema, mt as incrementEvalOutput, n as createRunner, nt as numberDisplayOptionsSchema, o as getEvalTitle, ot as captureEvalSpanError, p as agentEvalsConfigSchema, pt as getCurrentScope, q as traceSpanErrorSchema, r as createRunRequestSchema, rt as repoFileRefSchema, s as getEvalDisplayStatus, st as evalSpan, t as runCli, tt as jsonCellSchema, u as deriveStatusFromChildStatuses, ut as hashCacheKeySync, v as cacheModeSchema, vt as setScopeCacheContext, w as caseDetailSchema, x as serializedCacheSpanSchema, xt as getEvalRegistry, y as cacheRecordingOpSchema, yt as repoFile, z as evalChartTypeSchema } from "./cli-C5FL7C4G.mjs";
+import "./src-gqm1z1Nu.mjs";
+export { EvalAssertionError, agentEvalsConfigSchema, assertionFailureSchema, buildTraceTree, cacheEntrySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheRecordingOpSchema, cacheRecordingSchema, captureEvalSpanError, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, fileRefSchema, getCurrentScope, getEvalDisplayStatus, getEvalRegistry, getEvalTitle, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, numberDisplayOptionsSchema, repoFile, repoFileRefSchema, runArtifactRefSchema, runCli, runInEvalScope, runManifestSchema, runSummarySchema, scoreTraceSchema, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanErrorSchema, traceSpanKindSchema, traceSpanSchema, trialSelectionModeSchema, updateManualScoreRequestSchema };

package/dist/{runner-CToL8eJs.mjs → runner-Cdlvk56X.mjs}

@@ -1,2 +1,2 @@
-import { n as initRunner, t as getRunnerInstance } from "./runner-kSiHsl91.mjs";
+import { n as initRunner, t as getRunnerInstance } from "./runner-K2bN8KRS.mjs";
 export { getRunnerInstance, initRunner };

package/dist/{runner-kSiHsl91.mjs → runner-K2bN8KRS.mjs}

@@ -1,5 +1,5 @@
-import { n as createRunner } from "./cli-Dw9et3_Z.mjs";
-import "./src-CXclO9ZI.mjs";
+import { n as createRunner } from "./cli-C5FL7C4G.mjs";
+import "./src-gqm1z1Nu.mjs";
 //#region ../../apps/server/src/runner.ts
 let runnerInstance = null;
 function getRunnerInstance() {

package/dist/src-gqm1z1Nu.mjs

@@ -0,0 +1,2 @@
+import "./cli-C5FL7C4G.mjs";
+export {};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ls-stack/agent-eval",
-  "version": "0.3.0",
+  "version": "0.5.0",
   "type": "module",
   "bin": {
     "agent-evals": "./dist/bin.mjs"