@ls-stack/agent-eval 0.60.1 → 0.60.2

package/dist/{app-CxKEVlng.mjs → app-DPCFFkyQ.mjs}

@@ -1,7 +1,7 @@
-import { Et as getCaseRowCaseKey, Ot as caseRowSchema, dt as getEvalTitle, nt as updateManualScoreRequestSchema, rt as extractCacheEntries, tt as createRunRequestSchema } from "./runExecution-CjWJUUZ5.mjs";
-import { o as stageManualInputFile } from "./cli-CVBSlTD8.mjs";
-import "./src-DjOTPnDz.mjs";
-import { t as getRunnerInstance } from "./runner-DzDRasWV.mjs";
+import { Et as getCaseRowCaseKey, Ot as caseRowSchema, dt as getEvalTitle, nt as updateManualScoreRequestSchema, rt as extractCacheEntries, tt as createRunRequestSchema } from "./runExecution-Bq0Y3y_1.mjs";
+import { o as stageManualInputFile } from "./cli-CbePEEua.mjs";
+import "./src-CVM_FqPx.mjs";
+import { t as getRunnerInstance } from "./runner-Kp0JqxrU.mjs";
 import { z } from "zod";
 import { resultify } from "t-result";
 import { readFile } from "node:fs/promises";

package/dist/bin.mjs

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { t as runCli } from "./cli-CVBSlTD8.mjs";
+import { t as runCli } from "./cli-CbePEEua.mjs";
 import { spawn } from "node:child_process";
 //#region src/bin.ts
 const moduleMocksFlag = "--experimental-test-module-mocks";

package/dist/caseChild.mjs

@@ -1,4 +1,4 @@
-import { Ct as resolveLlmCallsConfig, It as runWithEvalRegistry, J as runInEvalRuntimeScope, L as configureEvalRunLogs, St as resolveApiCallsConfig, _ as createBufferedCacheStore, a as isCaseChildParentMessage, d as loadEvalModule, l as registerAgentEvalsPackageResolutionHooks, p as loadConfig, r as runCase, v as createFsCacheStore, y as getCacheRetentionOptions } from "./runExecution-CjWJUUZ5.mjs";
+import { Ct as resolveLlmCallsConfig, It as runWithEvalRegistry, J as runInEvalRuntimeScope, L as configureEvalRunLogs, St as resolveApiCallsConfig, _ as createBufferedCacheStore, a as isCaseChildParentMessage, d as loadEvalModule, l as registerAgentEvalsPackageResolutionHooks, p as loadConfig, r as runCase, v as createFsCacheStore, y as getCacheRetentionOptions } from "./runExecution-Bq0Y3y_1.mjs";
 //#region ../runner/src/caseChild.ts
 let fatalErrorReported = false;
 let disconnectExpected = false;

package/dist/{cli-CVBSlTD8.mjs → cli-CbePEEua.mjs}

@@ -1,5 +1,5 @@
-import { Ct as resolveLlmCallsConfig, Et as getCaseRowCaseKey, Ft as getEvalRegistry, Ot as caseRowSchema, St as resolveApiCallsConfig, Tt as buildEvalKey, _t as matchesTagsFilter, c as resolveArtifactPath, dt as getEvalTitle, f as resolveEvalDefaultConfig, ft as getEvalDisplayStatus, h as normalizeScoreDef, lt as applyDerivedCallAttributes, m as buildDeclaredColumnDefs, o as stripTerminalControlCodes, p as loadConfig, pt as deriveScopedSummaryFromCases, s as resolveTracePresentation, v as createFsCacheStore, xt as runSummarySchema, y as getCacheRetentionOptions } from "./runExecution-CjWJUUZ5.mjs";
-import { C as validateCharts, S as parseEvalDiscovery, _ as runTouchesEval, a as validateTagsFilters, b as deriveEvalFreshness, c as getLatestRunInfos, d as nextShortIdFromSnapshots, f as persistCaseDetail, g as recomputePersistedCaseStatus, h as recomputeEvalStatusesInRuns, i as resolveEvalTags, l as loadPersistedRunSnapshot, m as persistRunState, n as getTargetEvalKeys, o as generateRunId, p as deleteTemporaryRuns, s as getLastRunStatuses, u as loadPersistedRunSnapshots, v as buildManualInputDescriptor, x as loadIsolatedEvalRegistry, y as parseManualInputValues } from "./runOrchestration-DE2TFAS6.mjs";
+import { Ct as resolveLlmCallsConfig, Et as getCaseRowCaseKey, Ft as getEvalRegistry, Ot as caseRowSchema, St as resolveApiCallsConfig, Tt as buildEvalKey, _t as matchesTagsFilter, c as resolveArtifactPath, dt as getEvalTitle, f as resolveEvalDefaultConfig, ft as getEvalDisplayStatus, h as normalizeScoreDef, lt as applyDerivedCallAttributes, m as buildDeclaredColumnDefs, o as stripTerminalControlCodes, p as loadConfig, pt as deriveScopedSummaryFromCases, s as resolveTracePresentation, v as createFsCacheStore, xt as runSummarySchema, y as getCacheRetentionOptions } from "./runExecution-Bq0Y3y_1.mjs";
+import { C as validateCharts, S as parseEvalDiscovery, _ as runTouchesEval, a as validateTagsFilters, b as deriveEvalFreshness, c as getLatestRunInfos, d as nextShortIdFromSnapshots, f as persistCaseDetail, g as recomputePersistedCaseStatus, h as recomputeEvalStatusesInRuns, i as resolveEvalTags, l as loadPersistedRunSnapshot, m as persistRunState, n as getTargetEvalKeys, o as generateRunId, p as deleteTemporaryRuns, s as getLastRunStatuses, u as loadPersistedRunSnapshots, v as buildManualInputDescriptor, x as loadIsolatedEvalRegistry, y as parseManualInputValues } from "./runOrchestration-BpwW0AmB.mjs";
 import { parseEnv } from "node:util";
 import { resultify } from "t-result";
 import { copyFile, mkdir, readFile, rm, writeFile } from "node:fs/promises";
@@ -2232,8 +2232,8 @@ async function commandApp(args) {
 	const { serve } = await import("@hono/node-server");
 	const bundledWebDist = resolve(currentDir, "apps/web/dist");
 	if (existsSync(bundledWebDist)) process.env.AGENT_EVALS_WEB_DIST = bundledWebDist;
-	const appModule = await import("./app-CxKEVlng.mjs");
-	const runnerModule = await import("./runner-Cu1CQPTB.mjs");
+	const appModule = await import("./app-DPCFFkyQ.mjs");
+	const runnerModule = await import("./runner-XEP21_u9.mjs");
 	if (!isHonoAppModule(appModule)) throw new Error("Server app module is invalid");
 	if (!isServerRunnerModule(runnerModule)) throw new Error("Server runner module is invalid");
 	await runnerModule.initRunner({ loadEnv: args.loadEnv });

package/dist/index.d.mts

@@ -1941,10 +1941,10 @@ type ColumnKind = z.infer<typeof columnKindSchema>;
 declare const columnFormatSchema: z.ZodEnum<{
   number: "number";
   boolean: "boolean";
-  duration: "duration";
-  json: "json";
   file: "file";
+  duration: "duration";
   markdown: "markdown";
+  json: "json";
   image: "image";
   html: "html";
   pdf: "pdf";
@@ -1968,10 +1968,10 @@ declare const columnDefSchema: z.ZodObject<{
   format: z.ZodOptional<z.ZodEnum<{
     number: "number";
     boolean: "boolean";
-    duration: "duration";
-    json: "json";
     file: "file";
+    duration: "duration";
     markdown: "markdown";
+    json: "json";
     image: "image";
     html: "html";
     pdf: "pdf";
@@ -2061,9 +2061,9 @@ declare const traceAttributeDisplaySchema: z.ZodObject<{
     subtree: "subtree";
   }>>;
   mode: z.ZodOptional<z.ZodEnum<{
+    sum: "sum";
     all: "all";
     last: "last";
-    sum: "sum";
   }>>;
 }, z.core.$strip>;
 /**
@@ -2097,9 +2097,9 @@ declare const traceDisplayConfigSchema: z.ZodObject<{
       subtree: "subtree";
     }>>;
     mode: z.ZodOptional<z.ZodEnum<{
+      sum: "sum";
       all: "all";
       last: "last";
-      sum: "sum";
     }>>;
   }, z.core.$strip>>>;
 }, z.core.$strip>;
@@ -2137,9 +2137,9 @@ declare const traceAttributeDisplayInputSchema: z.ZodObject<{
     subtree: "subtree";
   }>>;
   mode: z.ZodOptional<z.ZodEnum<{
+    sum: "sum";
     all: "all";
     last: "last";
-    sum: "sum";
   }>>;
   transform: z.ZodOptional<z.ZodCustom<TraceAttributeTransform, TraceAttributeTransform>>;
 }, z.core.$strip>;
@@ -2175,9 +2175,9 @@ declare const traceDisplayInputConfigSchema: z.ZodObject<{
       subtree: "subtree";
     }>>;
     mode: z.ZodOptional<z.ZodEnum<{
+      sum: "sum";
       all: "all";
       last: "last";
-      sum: "sum";
     }>>;
     transform: z.ZodOptional<z.ZodCustom<TraceAttributeTransform, TraceAttributeTransform>>;
   }, z.core.$strip>>>;
@@ -2214,8 +2214,8 @@ declare const traceSpanSchema$1: z.ZodObject<{
   status: z.ZodEnum<{
     error: "error";
     running: "running";
-    ok: "ok";
     cancelled: "cancelled";
+    ok: "ok";
   }>;
   attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
   error: z.ZodOptional<z.ZodObject<{
@@ -2260,10 +2260,10 @@ type EvalFreshnessStatus = z.infer<typeof evalFreshnessStatusSchema>;
  * `best` selects the highest finite value and `worst` selects the lowest.
  */
 declare const evalStatAggregateSchema: z.ZodEnum<{
-  sum: "sum";
   avg: "avg";
   min: "min";
   max: "max";
+  sum: "sum";
   best: "best";
   worst: "worst";
 }>;
@@ -2292,10 +2292,10 @@ declare const evalStatItemSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
   hideIfNoValue: z.ZodOptional<z.ZodBoolean>;
   kind: z.ZodLiteral<"duration">;
   aggregate: z.ZodOptional<z.ZodEnum<{
-    sum: "sum";
     avg: "avg";
     min: "min";
     max: "max";
+    sum: "sum";
     best: "best";
     worst: "worst";
   }>>;
@@ -2303,10 +2303,10 @@ declare const evalStatItemSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
   hideIfNoValue: z.ZodOptional<z.ZodBoolean>;
   kind: z.ZodLiteral<"cacheHits">;
   aggregate: z.ZodOptional<z.ZodEnum<{
-    sum: "sum";
     avg: "avg";
     min: "min";
     max: "max";
+    sum: "sum";
     best: "best";
     worst: "worst";
   }>>;
@@ -2316,20 +2316,20 @@ declare const evalStatItemSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
   key: z.ZodString;
   label: z.ZodOptional<z.ZodString>;
   aggregate: z.ZodEnum<{
-    sum: "sum";
     avg: "avg";
     min: "min";
     max: "max";
+    sum: "sum";
     best: "best";
     worst: "worst";
   }>;
   format: z.ZodOptional<z.ZodEnum<{
     number: "number";
     boolean: "boolean";
-    duration: "duration";
-    json: "json";
     file: "file";
+    duration: "duration";
     markdown: "markdown";
+    json: "json";
     image: "image";
     html: "html";
     pdf: "pdf";
@@ -2356,10 +2356,10 @@ declare const evalStatsConfigSchema: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodOb
   hideIfNoValue: z.ZodOptional<z.ZodBoolean>;
   kind: z.ZodLiteral<"duration">;
   aggregate: z.ZodOptional<z.ZodEnum<{
-    sum: "sum";
     avg: "avg";
     min: "min";
     max: "max";
+    sum: "sum";
     best: "best";
     worst: "worst";
   }>>;
@@ -2367,10 +2367,10 @@ declare const evalStatsConfigSchema: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodOb
   hideIfNoValue: z.ZodOptional<z.ZodBoolean>;
   kind: z.ZodLiteral<"cacheHits">;
   aggregate: z.ZodOptional<z.ZodEnum<{
-    sum: "sum";
     avg: "avg";
     min: "min";
     max: "max";
+    sum: "sum";
     best: "best";
     worst: "worst";
   }>>;
@@ -2380,20 +2380,20 @@ declare const evalStatsConfigSchema: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodOb
   key: z.ZodString;
   label: z.ZodOptional<z.ZodString>;
   aggregate: z.ZodEnum<{
-    sum: "sum";
     avg: "avg";
     min: "min";
     max: "max";
+    sum: "sum";
     best: "best";
     worst: "worst";
   }>;
   format: z.ZodOptional<z.ZodEnum<{
     number: "number";
     boolean: "boolean";
-    duration: "duration";
-    json: "json";
     file: "file";
+    duration: "duration";
     markdown: "markdown";
+    json: "json";
     image: "image";
     html: "html";
     pdf: "pdf";
@@ -2436,10 +2436,10 @@ declare const evalSummarySchema$1: z.ZodObject<{
     format: z.ZodOptional<z.ZodEnum<{
       number: "number";
       boolean: "boolean";
-      duration: "duration";
-      json: "json";
       file: "file";
+      duration: "duration";
       markdown: "markdown";
+      json: "json";
       image: "image";
       html: "html";
       pdf: "pdf";
@@ -2466,10 +2466,10 @@ declare const evalSummarySchema$1: z.ZodObject<{
   caseIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
   lastRunStatus: z.ZodNullable<z.ZodEnum<{
     error: "error";
-    running: "running";
-    cancelled: "cancelled";
     pass: "pass";
     fail: "fail";
+    running: "running";
+    cancelled: "cancelled";
     unscored: "unscored";
   }>>;
   stats: z.ZodOptional<z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
@@ -2483,10 +2483,10 @@ declare const evalSummarySchema$1: z.ZodObject<{
     hideIfNoValue: z.ZodOptional<z.ZodBoolean>;
     kind: z.ZodLiteral<"duration">;
     aggregate: z.ZodOptional<z.ZodEnum<{
-      sum: "sum";
       avg: "avg";
       min: "min";
       max: "max";
+      sum: "sum";
       best: "best";
       worst: "worst";
     }>>;
@@ -2494,10 +2494,10 @@ declare const evalSummarySchema$1: z.ZodObject<{
     hideIfNoValue: z.ZodOptional<z.ZodBoolean>;
     kind: z.ZodLiteral<"cacheHits">;
     aggregate: z.ZodOptional<z.ZodEnum<{
-      sum: "sum";
       avg: "avg";
       min: "min";
       max: "max";
+      sum: "sum";
       best: "best";
       worst: "worst";
     }>>;
@@ -2507,20 +2507,20 @@ declare const evalSummarySchema$1: z.ZodObject<{
     key: z.ZodString;
     label: z.ZodOptional<z.ZodString>;
     aggregate: z.ZodEnum<{
-      sum: "sum";
       avg: "avg";
       min: "min";
       max: "max";
+      sum: "sum";
       best: "best";
       worst: "worst";
     }>;
     format: z.ZodOptional<z.ZodEnum<{
       number: "number";
       boolean: "boolean";
-      duration: "duration";
-      json: "json";
       file: "file";
+      duration: "duration";
       markdown: "markdown";
+      json: "json";
       image: "image";
       html: "html";
       pdf: "pdf";
@@ -2534,10 +2534,10 @@ declare const evalSummarySchema$1: z.ZodObject<{
     accent: z.ZodOptional<z.ZodBoolean>;
   }, z.core.$strip>], "kind">>>;
   defaultStatAggregate: z.ZodOptional<z.ZodEnum<{
-    sum: "sum";
     avg: "avg";
     min: "min";
     max: "max";
+    sum: "sum";
     best: "best";
     worst: "worst";
   }>>;
@@ -2558,11 +2558,11 @@ declare const evalSummarySchema$1: z.ZodObject<{
       }>;
       label: z.ZodOptional<z.ZodString>;
       color: z.ZodOptional<z.ZodEnum<{
-        error: "error";
         success: "success";
-        warning: "warning";
         accent: "accent";
+        error: "error";
         accentDim: "accentDim";
+        warning: "warning";
         textMuted: "textMuted";
       }>>;
       axis: z.ZodOptional<z.ZodEnum<{
@@ -2573,20 +2573,20 @@ declare const evalSummarySchema$1: z.ZodObject<{
       source: z.ZodLiteral<"column">;
       key: z.ZodString;
       aggregate: z.ZodEnum<{
-        sum: "sum";
         avg: "avg";
         min: "min";
         max: "max";
+        sum: "sum";
         latest: "latest";
         passThresholdRate: "passThresholdRate";
       }>;
       label: z.ZodOptional<z.ZodString>;
       color: z.ZodOptional<z.ZodEnum<{
-        error: "error";
         success: "success";
-        warning: "warning";
         accent: "accent";
+        error: "error";
         accentDim: "accentDim";
+        warning: "warning";
         textMuted: "textMuted";
       }>>;
       axis: z.ZodOptional<z.ZodEnum<{
@@ -2615,10 +2615,10 @@ declare const evalSummarySchema$1: z.ZodObject<{
       source: z.ZodLiteral<"column">;
       key: z.ZodString;
       aggregate: z.ZodEnum<{
-        sum: "sum";
         avg: "avg";
         min: "min";
         max: "max";
+        sum: "sum";
         latest: "latest";
         passThresholdRate: "passThresholdRate";
       }>;
@@ -2715,10 +2715,10 @@ declare const caseRowSchema$1: z.ZodObject<{
   tags: z.ZodOptional<z.ZodArray<z.ZodString>>;
   status: z.ZodEnum<{
     error: "error";
-    running: "running";
-    cancelled: "cancelled";
     pass: "pass";
     fail: "fail";
+    running: "running";
+    cancelled: "cancelled";
     pending: "pending";
   }>;
   durationMs: z.ZodNullable<z.ZodNumber>;
@@ -2748,10 +2748,10 @@ declare const caseRowSchema$1: z.ZodObject<{
     format: z.ZodOptional<z.ZodEnum<{
       number: "number";
       boolean: "boolean";
-      duration: "duration";
-      json: "json";
       file: "file";
+      duration: "duration";
       markdown: "markdown";
+      json: "json";
       image: "image";
       html: "html";
       pdf: "pdf";
@@ -2857,8 +2857,8 @@ declare const scoreTraceSchema: z.ZodObject<{
     status: z.ZodEnum<{
       error: "error";
       running: "running";
-      ok: "ok";
       cancelled: "cancelled";
+      ok: "ok";
     }>;
     attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
     error: z.ZodOptional<z.ZodObject<{
@@ -2908,9 +2908,9 @@ declare const scoreTraceSchema: z.ZodObject<{
         subtree: "subtree";
       }>>;
       mode: z.ZodOptional<z.ZodEnum<{
+        sum: "sum";
         all: "all";
         last: "last";
-        sum: "sum";
       }>>;
     }, z.core.$strip>>>;
   }, z.core.$strip>;
@@ -2942,10 +2942,10 @@ declare const caseDetailSchema$1: z.ZodObject<{
   tags: z.ZodOptional<z.ZodArray<z.ZodString>>;
   status: z.ZodEnum<{
     error: "error";
-    running: "running";
-    cancelled: "cancelled";
     pass: "pass";
     fail: "fail";
+    running: "running";
+    cancelled: "cancelled";
     pending: "pending";
   }>;
   input: z.ZodUnknown;
@@ -2960,8 +2960,8 @@ declare const caseDetailSchema$1: z.ZodObject<{
     status: z.ZodEnum<{
       error: "error";
       running: "running";
-      ok: "ok";
       cancelled: "cancelled";
+      ok: "ok";
     }>;
     attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
     error: z.ZodOptional<z.ZodObject<{
@@ -3011,9 +3011,9 @@ declare const caseDetailSchema$1: z.ZodObject<{
         subtree: "subtree";
       }>>;
       mode: z.ZodOptional<z.ZodEnum<{
+        sum: "sum";
         all: "all";
         last: "last";
-        sum: "sum";
       }>>;
     }, z.core.$strip>>>;
   }, z.core.$strip>;
@@ -3029,8 +3029,8 @@ declare const caseDetailSchema$1: z.ZodObject<{
       status: z.ZodEnum<{
         error: "error";
         running: "running";
-        ok: "ok";
         cancelled: "cancelled";
+        ok: "ok";
       }>;
       attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
       error: z.ZodOptional<z.ZodObject<{
@@ -3080,9 +3080,9 @@ declare const caseDetailSchema$1: z.ZodObject<{
           subtree: "subtree";
         }>>;
         mode: z.ZodOptional<z.ZodEnum<{
+          sum: "sum";
           all: "all";
           last: "last";
-          sum: "sum";
         }>>;
       }, z.core.$strip>>>;
     }, z.core.$strip>;
@@ -3126,10 +3126,10 @@ declare const caseDetailSchema$1: z.ZodObject<{
     format: z.ZodOptional<z.ZodEnum<{
       number: "number";
       boolean: "boolean";
-      duration: "duration";
-      json: "json";
       file: "file";
+      duration: "duration";
       markdown: "markdown";
+      json: "json";
       image: "image";
       html: "html";
       pdf: "pdf";
@@ -3269,10 +3269,10 @@ declare const evalChartBuiltinMetricSchema: z.ZodEnum<{
 type EvalChartBuiltinMetric = z.infer<typeof evalChartBuiltinMetricSchema>;
 /** Reducer applied to a numeric column across all cases of a single run. */
 declare const evalChartAggregateSchema: z.ZodEnum<{
-  sum: "sum";
   avg: "avg";
   min: "min";
   max: "max";
+  sum: "sum";
   latest: "latest";
   passThresholdRate: "passThresholdRate";
 }>;
@@ -3283,11 +3283,11 @@ type EvalChartAggregate = z.infer<typeof evalChartAggregateSchema>;
  * not emit raw hex so authored evals stay decoupled from the web theme.
  */
 declare const evalChartColorSchema: z.ZodEnum<{
-  error: "error";
   success: "success";
-  warning: "warning";
   accent: "accent";
+  error: "error";
   accentDim: "accentDim";
+  warning: "warning";
   textMuted: "textMuted";
 }>;
 /** Semantic color token resolved to a theme color by the web UI. */
@@ -3312,11 +3312,11 @@ declare const evalChartMetricSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
   }>;
   label: z.ZodOptional<z.ZodString>;
   color: z.ZodOptional<z.ZodEnum<{
-    error: "error";
     success: "success";
-    warning: "warning";
     accent: "accent";
+    error: "error";
     accentDim: "accentDim";
+    warning: "warning";
     textMuted: "textMuted";
   }>>;
   axis: z.ZodOptional<z.ZodEnum<{
@@ -3327,20 +3327,20 @@ declare const evalChartMetricSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
   source: z.ZodLiteral<"column">;
   key: z.ZodString;
   aggregate: z.ZodEnum<{
-    sum: "sum";
     avg: "avg";
     min: "min";
     max: "max";
+    sum: "sum";
     latest: "latest";
     passThresholdRate: "passThresholdRate";
   }>;
   label: z.ZodOptional<z.ZodString>;
   color: z.ZodOptional<z.ZodEnum<{
-    error: "error";
     success: "success";
-    warning: "warning";
     accent: "accent";
+    error: "error";
     accentDim: "accentDim";
+    warning: "warning";
     textMuted: "textMuted";
   }>>;
   axis: z.ZodOptional<z.ZodEnum<{
@@ -3362,10 +3362,10 @@ declare const evalChartTooltipExtraSchema: z.ZodDiscriminatedUnion<[z.ZodObject<
   source: z.ZodLiteral<"column">;
   key: z.ZodString;
   aggregate: z.ZodEnum<{
-    sum: "sum";
     avg: "avg";
     min: "min";
     max: "max";
+    sum: "sum";
     latest: "latest";
     passThresholdRate: "passThresholdRate";
   }>;
@@ -3395,11 +3395,11 @@ declare const evalChartConfigSchema: z.ZodObject<{
     }>;
     label: z.ZodOptional<z.ZodString>;
     color: z.ZodOptional<z.ZodEnum<{
-      error: "error";
       success: "success";
-      warning: "warning";
       accent: "accent";
+      error: "error";
       accentDim: "accentDim";
+      warning: "warning";
       textMuted: "textMuted";
     }>>;
     axis: z.ZodOptional<z.ZodEnum<{
@@ -3410,20 +3410,20 @@ declare const evalChartConfigSchema: z.ZodObject<{
     source: z.ZodLiteral<"column">;
     key: z.ZodString;
     aggregate: z.ZodEnum<{
-      sum: "sum";
       avg: "avg";
       min: "min";
       max: "max";
+      sum: "sum";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
     label: z.ZodOptional<z.ZodString>;
     color: z.ZodOptional<z.ZodEnum<{
-      error: "error";
       success: "success";
-      warning: "warning";
       accent: "accent";
+      error: "error";
       accentDim: "accentDim";
+      warning: "warning";
       textMuted: "textMuted";
     }>>;
     axis: z.ZodOptional<z.ZodEnum<{
@@ -3452,10 +3452,10 @@ declare const evalChartConfigSchema: z.ZodObject<{
     source: z.ZodLiteral<"column">;
     key: z.ZodString;
     aggregate: z.ZodEnum<{
-      sum: "sum";
       avg: "avg";
       min: "min";
       max: "max";
+      sum: "sum";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
@@ -3485,11 +3485,11 @@ declare const evalChartsConfigSchema: z.ZodArray<z.ZodObject<{
     }>;
     label: z.ZodOptional<z.ZodString>;
     color: z.ZodOptional<z.ZodEnum<{
-      error: "error";
       success: "success";
-      warning: "warning";
       accent: "accent";
+      error: "error";
       accentDim: "accentDim";
+      warning: "warning";
       textMuted: "textMuted";
     }>>;
     axis: z.ZodOptional<z.ZodEnum<{
@@ -3500,20 +3500,20 @@ declare const evalChartsConfigSchema: z.ZodArray<z.ZodObject<{
     source: z.ZodLiteral<"column">;
     key: z.ZodString;
     aggregate: z.ZodEnum<{
-      sum: "sum";
       avg: "avg";
       min: "min";
       max: "max";
+      sum: "sum";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
     label: z.ZodOptional<z.ZodString>;
     color: z.ZodOptional<z.ZodEnum<{
-      error: "error";
       success: "success";
-      warning: "warning";
       accent: "accent";
+      error: "error";
       accentDim: "accentDim";
+      warning: "warning";
       textMuted: "textMuted";
     }>>;
     axis: z.ZodOptional<z.ZodEnum<{
@@ -3542,10 +3542,10 @@ declare const evalChartsConfigSchema: z.ZodArray<z.ZodObject<{
     source: z.ZodLiteral<"column">;
     key: z.ZodString;
     aggregate: z.ZodEnum<{
-      sum: "sum";
       avg: "avg";
       min: "min";
       max: "max";
+      sum: "sum";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
@@ -3573,8 +3573,8 @@ declare const runManifestSchema$1: z.ZodObject<{
   evalSourceFingerprints: z.ZodDefault<z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>>;
   target: z.ZodObject<{
     mode: z.ZodEnum<{
-      all: "all";
       caseIds: "caseIds";
+      all: "all";
       evalIds: "evalIds";
     }>;
     evalKeys: z.ZodOptional<z.ZodArray<z.ZodString>>;
@@ -4669,10 +4669,10 @@ declare const cacheRecordingOpSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
     format: z.ZodOptional<z.ZodEnum<{
       number: "number";
       boolean: "boolean";
-      duration: "duration";
-      json: "json";
       file: "file";
+      duration: "duration";
       markdown: "markdown";
+      json: "json";
       image: "image";
       html: "html";
       pdf: "pdf";
@@ -4721,8 +4721,8 @@ declare const cacheRecordingSchema: z.ZodObject<{
   finalStatus: z.ZodOptional<z.ZodEnum<{
     error: "error";
     running: "running";
-    ok: "ok";
     cancelled: "cancelled";
+    ok: "ok";
   }>>;
   finalError: z.ZodOptional<z.ZodObject<{
     name: z.ZodOptional<z.ZodString>;
@@ -4757,10 +4757,10 @@ declare const cacheRecordingSchema: z.ZodObject<{
       format: z.ZodOptional<z.ZodEnum<{
         number: "number";
         boolean: "boolean";
-        duration: "duration";
-        json: "json";
         file: "file";
+        duration: "duration";
         markdown: "markdown";
+        json: "json";
         image: "image";
         html: "html";
         pdf: "pdf";
@@ -4822,8 +4822,8 @@ declare const cacheEntrySchema: z.ZodObject<{
     finalStatus: z.ZodOptional<z.ZodEnum<{
       error: "error";
       running: "running";
-      ok: "ok";
       cancelled: "cancelled";
+      ok: "ok";
     }>>;
     finalError: z.ZodOptional<z.ZodObject<{
       name: z.ZodOptional<z.ZodString>;
@@ -4858,10 +4858,10 @@ declare const cacheEntrySchema: z.ZodObject<{
         format: z.ZodOptional<z.ZodEnum<{
           number: "number";
           boolean: "boolean";
-          duration: "duration";
-          json: "json";
           file: "file";
+          duration: "duration";
           markdown: "markdown";
+          json: "json";
           image: "image";
           html: "html";
           pdf: "pdf";
@@ -4940,8 +4940,8 @@ declare const cacheDebugKeyEntrySchema: z.ZodObject<{
       finalStatus: z.ZodOptional<z.ZodEnum<{
         error: "error";
         running: "running";
-        ok: "ok";
         cancelled: "cancelled";
+        ok: "ok";
       }>>;
       finalError: z.ZodOptional<z.ZodObject<{
         name: z.ZodOptional<z.ZodString>;
@@ -4976,10 +4976,10 @@ declare const cacheDebugKeyEntrySchema: z.ZodObject<{
           format: z.ZodOptional<z.ZodEnum<{
             number: "number";
             boolean: "boolean";
-            duration: "duration";
-            json: "json";
             file: "file";
+            duration: "duration";
             markdown: "markdown";
+            json: "json";
             image: "image";
             html: "html";
             pdf: "pdf";
@@ -5047,8 +5047,8 @@ declare const cacheEntryWithDebugKeySchema$1: z.ZodObject<{
     finalStatus: z.ZodOptional<z.ZodEnum<{
       error: "error";
       running: "running";
-      ok: "ok";
       cancelled: "cancelled";
+      ok: "ok";
     }>>;
     finalError: z.ZodOptional<z.ZodObject<{
       name: z.ZodOptional<z.ZodString>;
@@ -5083,10 +5083,10 @@ declare const cacheEntryWithDebugKeySchema$1: z.ZodObject<{
         format: z.ZodOptional<z.ZodEnum<{
           number: "number";
           boolean: "boolean";
-          duration: "duration";
-          json: "json";
           file: "file";
+          duration: "duration";
           markdown: "markdown";
+          json: "json";
           image: "image";
           html: "html";
           pdf: "pdf";
@@ -5156,8 +5156,8 @@ declare const cacheEntryWithDebugKeySchema$1: z.ZodObject<{
         finalStatus: z.ZodOptional<z.ZodEnum<{
           error: "error";
           running: "running";
-          ok: "ok";
           cancelled: "cancelled";
+          ok: "ok";
         }>>;
         finalError: z.ZodOptional<z.ZodObject<{
           name: z.ZodOptional<z.ZodString>;
@@ -5192,10 +5192,10 @@ declare const cacheEntryWithDebugKeySchema$1: z.ZodObject<{
             format: z.ZodOptional<z.ZodEnum<{
               number: "number";
               boolean: "boolean";
-              duration: "duration";
-              json: "json";
               file: "file";
+              duration: "duration";
               markdown: "markdown";
+              json: "json";
               image: "image";
               html: "html";
               pdf: "pdf";
@@ -5263,8 +5263,8 @@ declare const cacheFileSchema: z.ZodObject<{
       finalStatus: z.ZodOptional<z.ZodEnum<{
         error: "error";
         running: "running";
-        ok: "ok";
         cancelled: "cancelled";
+        ok: "ok";
       }>>;
       finalError: z.ZodOptional<z.ZodObject<{
         name: z.ZodOptional<z.ZodString>;
@@ -5299,10 +5299,10 @@ declare const cacheFileSchema: z.ZodObject<{
           format: z.ZodOptional<z.ZodEnum<{
             number: "number";
             boolean: "boolean";
-            duration: "duration";
-            json: "json";
             file: "file";
+            duration: "duration";
             markdown: "markdown";
+            json: "json";
             image: "image";
             html: "html";
             pdf: "pdf";
@@ -5380,8 +5380,8 @@ declare const cacheDebugKeyFileSchema: z.ZodObject<{
         finalStatus: z.ZodOptional<z.ZodEnum<{
           error: "error";
           running: "running";
-          ok: "ok";
           cancelled: "cancelled";
+          ok: "ok";
         }>>;
         finalError: z.ZodOptional<z.ZodObject<{
           name: z.ZodOptional<z.ZodString>;
@@ -5416,10 +5416,10 @@ declare const cacheDebugKeyFileSchema: z.ZodObject<{
             format: z.ZodOptional<z.ZodEnum<{
               number: "number";
               boolean: "boolean";
-              duration: "duration";
-              json: "json";
               file: "file";
+              duration: "duration";
               markdown: "markdown";
+              json: "json";
               image: "image";
               html: "html";
               pdf: "pdf";
@@ -5565,8 +5565,8 @@ type ConfigReloadState = z.infer<typeof configReloadStateSchema$1>;
 declare const createRunRequestSchema$1: z.ZodObject<{
   target: z.ZodObject<{
     mode: z.ZodEnum<{
-      all: "all";
       caseIds: "caseIds";
+      all: "all";
       evalIds: "evalIds";
     }>;
     evalKeys: z.ZodOptional<z.ZodArray<z.ZodString>>;

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, z as evalLog } from "./runExecution-CjWJUUZ5.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-CVBSlTD8.mjs";
-import { n as matchesEvalTags, t as defineEval } from "./src-DjOTPnDz.mjs";
+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, z as evalLog } from "./runExecution-Bq0Y3y_1.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-CbePEEua.mjs";
+import { n as matchesEvalTags, t as defineEval } from "./src-CVM_FqPx.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 };

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, y as getCacheRetentionOptions } from "./runExecution-CjWJUUZ5.mjs";
-import { S as parseEvalDiscovery, m as persistRunState, r as getTargetEvals$1, t as executeRun } from "./runOrchestration-DE2TFAS6.mjs";
+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, y as getCacheRetentionOptions } from "./runExecution-Bq0Y3y_1.mjs";
+import { S as parseEvalDiscovery, m as persistRunState, r as getTargetEvals$1, t as executeRun } from "./runOrchestration-BpwW0AmB.mjs";
 import { z } from "zod";
 import { readFile } from "node:fs/promises";
 import { relative } from "node:path";

package/dist/{runExecution-CjWJUUZ5.mjs → runExecution-Bq0Y3y_1.mjs}

@@ -289,7 +289,7 @@ z.object({
 	key: z.string(),
 	namespace: z.string(),
 	storedAt: z.string(),
-	/** Last successful cache hit time. `null` means the entry has not been hit yet. */
+	/** Last successful cache read or write time. Legacy entries may be `null`. */
 	lastAccessedAt: z.string().nullable()
 });
 z.object({
@@ -5395,7 +5395,7 @@ function createFsCacheStore(options) {
 				const index = await readNamespaceIndex(cacheDir, entry.namespace);
 				index.entries[entry.key] = {
 					storedAt: entry.storedAt,
-					lastAccessedAt: null,
+					lastAccessedAt: entry.storedAt,
 					blobRefs: await collectExternalJsonBlobRefs(entry, blobDirs)
 				};
 				await writeNamespaceIndex(cacheDir, index);

package/dist/{runOrchestration-DE2TFAS6.mjs → runOrchestration-BpwW0AmB.mjs}

@@ -1,4 +1,4 @@
-import { Dt as caseDetailSchema, Et as getCaseRowCaseKey, It as runWithEvalRegistry, J as runInEvalRuntimeScope, Ot as caseRowSchema, Z as runWithEvalClock, _t as matchesTagsFilter, bt as runManifestSchema, d as loadEvalModule, f as resolveEvalDefaultConfig, g as commitPendingCacheWrites, gt as dedupeEvalTags, ht as deriveStatusFromChildStatuses, i as isCaseChildMessage, m as buildDeclaredColumnDefs, mt as deriveStatusFromCaseRows, n as resolveRunnableEvalCases, o as stripTerminalControlCodes, pt as deriveScopedSummaryFromCases, t as filterEvalCases, u as runWithModuleIsolation, vt as validateEvalTagName, wt as buildCaseKey, xt as runSummarySchema, yt as validateTagsFilterExpression } from "./runExecution-CjWJUUZ5.mjs";
+import { Dt as caseDetailSchema, Et as getCaseRowCaseKey, It as runWithEvalRegistry, J as runInEvalRuntimeScope, Ot as caseRowSchema, Z as runWithEvalClock, _t as matchesTagsFilter, bt as runManifestSchema, d as loadEvalModule, f as resolveEvalDefaultConfig, g as commitPendingCacheWrites, gt as dedupeEvalTags, ht as deriveStatusFromChildStatuses, i as isCaseChildMessage, m as buildDeclaredColumnDefs, mt as deriveStatusFromCaseRows, n as resolveRunnableEvalCases, o as stripTerminalControlCodes, pt as deriveScopedSummaryFromCases, t as filterEvalCases, u as runWithModuleIsolation, vt as validateEvalTagName, wt as buildCaseKey, xt as runSummarySchema, yt as validateTagsFilterExpression } from "./runExecution-Bq0Y3y_1.mjs";
 import { Result, resultify } from "t-result";
 import { readFile, readdir, rm, writeFile } from "node:fs/promises";
 import { dirname, join } from "node:path";

package/dist/{runner-DzDRasWV.mjs → runner-Kp0JqxrU.mjs}

@@ -1,5 +1,5 @@
-import { n as createRunner } from "./cli-CVBSlTD8.mjs";
-import "./src-DjOTPnDz.mjs";
+import { n as createRunner } from "./cli-CbePEEua.mjs";
+import "./src-CVM_FqPx.mjs";
 //#region ../../apps/server/src/runner.ts
 let runnerInstance = null;
 function getRunnerInstance({ loadEnv = true } = {}) {

package/dist/{runner-Cu1CQPTB.mjs → runner-XEP21_u9.mjs}

@@ -1,2 +1,2 @@
-import { n as initRunner, t as getRunnerInstance } from "./runner-DzDRasWV.mjs";
+import { n as initRunner, t as getRunnerInstance } from "./runner-Kp0JqxrU.mjs";
 export { getRunnerInstance, initRunner };

package/dist/{src-DjOTPnDz.mjs → src-CVM_FqPx.mjs}

@@ -1,5 +1,5 @@
-import { G as matchesEvalTags$1, Pt as defineEval$1 } from "./runExecution-CjWJUUZ5.mjs";
-import "./cli-CVBSlTD8.mjs";
+import { G as matchesEvalTags$1, Pt as defineEval$1 } from "./runExecution-Bq0Y3y_1.mjs";
+import "./cli-CbePEEua.mjs";
 //#region src/index.ts
 /** Register an eval definition with typed tag support. */
 function defineEval(definition) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ls-stack/agent-eval",
-  "version": "0.60.1",
+  "version": "0.60.2",
   "type": "module",
   "bin": {
     "agent-evals": "./dist/bin.mjs"
@@ -32,9 +32,9 @@
   "devDependencies": {
     "@types/node": "^24.7.2",
     "typescript": "^5.9.2",
-    "@agent-evals/sdk": "0.0.1",
+    "@agent-evals/runner": "0.0.1",
     "@agent-evals/shared": "0.0.1",
-    "@agent-evals/runner": "0.0.1"
+    "@agent-evals/sdk": "0.0.1"
   },
   "scripts": {
     "build": "pnpm --filter @agent-evals/web build && pnpm --filter @agent-evals/shared build && pnpm --filter @agent-evals/sdk build && pnpm --filter @agent-evals/runner build && tsdown --filter cli-js && tsdown --filter cli-types",

package/skills/agent-eval/SKILL.md

@@ -568,6 +568,8 @@ Mental model:
   runner stays idle for `cache.pruneIdleDelayMs ?? 5000` milliseconds. Configure
   `cache.maxEntries` as a number for the default cap, or as
   `{ default, namespaces }` for exact namespace-specific caps.
+  Writes initialize the row's last access time to the stored time; later cache
+  hits refresh that timestamp at the configured access-time update interval.
 - Unindexed legacy cache files are ignored by normal lookup/listing. Use
   `agent-evals cache repair` to remove unindexed cache files, stale index rows,
   debug sidecars, and unreferenced blob files.