@ls-stack/agent-eval 0.60.2 → 0.60.3

package/dist/{app-DPCFFkyQ.mjs → app-l3ynaNsb.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-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 { o as stageManualInputFile } from "./cli-BSVUCUxr.mjs";
+import "./src-D5vGo2iv.mjs";
+import { t as getRunnerInstance } from "./runner-DmkSq-QG.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-CbePEEua.mjs";
+import { t as runCli } from "./cli-BSVUCUxr.mjs";
 import { spawn } from "node:child_process";
 //#region src/bin.ts
 const moduleMocksFlag = "--experimental-test-module-mocks";

package/dist/{cli-CbePEEua.mjs → cli-BSVUCUxr.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-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 { C as parseEvalDiscovery, S as loadIsolatedEvalRegistry, _ as recomputePersistedCaseStatus, a as validateTagsFilters, b as parseManualInputValues, c as getLatestRunInfos, d as loadPersistedRunSnapshots, f as nextShortIdFromSnapshots, g as recomputeEvalStatusesInRuns, h as persistRunState, i as resolveEvalTags, l as loadPersistedCaseDetail, m as deleteTemporaryRuns, n as getTargetEvalKeys, o as generateRunId, p as persistCaseDetail, s as getLastRunStatuses, u as loadPersistedRunSnapshot, v as runTouchesEval, w as validateCharts, x as deriveEvalFreshness, y as buildManualInputDescriptor } from "./runOrchestration-C7qQISz2.mjs";
 import { parseEnv } from "node:util";
 import { resultify } from "t-result";
 import { copyFile, mkdir, readFile, rm, writeFile } from "node:fs/promises";
@@ -28,10 +28,14 @@ function resolveCaseDetailLookup(run, caseId) {
 		const caseDetail = run.caseDetails.get(lookupId);
 		if (caseDetail) return caseDetail;
 	}
-	const matchingCaseRow = run.cases.find((caseRow) => lookupIds.has(getCaseRowCaseKey(caseRow)) || lookupIds.has(caseRow.caseId));
+	const matchingCaseRow = resolveCaseRowForCaseDetailLookup(run, caseId);
 	if (matchingCaseRow === void 0) return void 0;
 	return run.caseDetails.get(getCaseRowCaseKey(matchingCaseRow));
 }
+function resolveCaseRowForCaseDetailLookup(run, caseId) {
+	const lookupIds = new Set(getCaseLookupIds(caseId));
+	return run.cases.find((caseRow) => lookupIds.has(getCaseRowCaseKey(caseRow)) || lookupIds.has(caseRow.caseId));
+}
 //#endregion
 //#region ../runner/src/configReload.ts
 /** Coordinates idle-only reloads for workspace config and `.env` in app mode. */
@@ -823,7 +827,7 @@ async function markRunTerminalFromChild(runState, event, managerContext) {
 		runState.manifest = snapshot.manifest;
 		runState.summary = snapshot.summary;
 		runState.cases = snapshot.cases;
-		runState.caseDetails = snapshot.caseDetails;
+		runState.caseDetails = /* @__PURE__ */ new Map();
 	} else if (event.type === "run.finished") {
 		runState.manifest.status = "completed";
 		runState.manifest.endedAt = (/* @__PURE__ */ new Date()).toISOString();
@@ -995,6 +999,26 @@ function createRunner({ watchForChanges = true, loadEnv = true } = {}) {
 		if (typeof config.concurrency !== "number" || !Number.isFinite(config.concurrency)) return 1;
 		return Math.max(1, Math.floor(config.concurrency));
 	}
+	function getCaseDetailFileId(run, caseRow) {
+		const caseKey = getCaseRowCaseKey(caseRow);
+		return run.cases.some((existing) => existing.caseId === caseRow.caseId && getCaseRowCaseKey(existing) !== caseKey) ? caseKey : caseRow.caseId;
+	}
+	function hydrateCaseDetailForRow(run, caseRow) {
+		const caseKey = getCaseRowCaseKey(caseRow);
+		const cached = run.caseDetails.get(caseKey);
+		if (cached !== void 0) return cached;
+		const detail = loadPersistedCaseDetail(run.runDir, getCaseDetailFileId(run, caseRow));
+		if (detail === null) return void 0;
+		run.caseDetails.set(detail.caseKey ?? detail.caseId, detail);
+		return detail;
+	}
+	function hydrateCaseDetailForLookup(run, caseId) {
+		const cached = resolveCaseDetailLookup(run, caseId);
+		if (cached !== void 0) return cached;
+		const caseRow = resolveCaseRowForCaseDetailLookup(run, caseId);
+		if (caseRow === void 0) return void 0;
+		return hydrateCaseDetailForRow(run, caseRow);
+	}
 	function nextRegistryLoadIsolationKey(prefix, filePath) {
 		registryLoadCounter++;
 		return `${prefix}:${String(registryLoadCounter)}:${filePath}`;
@@ -1041,6 +1065,7 @@ function createRunner({ watchForChanges = true, loadEnv = true } = {}) {
 				evalKey: evalMeta.key,
 				evalExists: evals.has(evalMeta.key),
 				scoreThresholds,
+				getCaseDetail: hydrateCaseDetailForRow,
 				persistCaseDetail
 			});
 			emitDiscoveryEvent();
@@ -1052,6 +1077,7 @@ function createRunner({ watchForChanges = true, loadEnv = true } = {}) {
 				updated: false,
 				reason: "Run not found"
 			};
+			hydrateCaseDetailForLookup(run, caseId);
 			return recalculateDerivedAttributesForCase({
 				run,
 				caseId,
@@ -1107,7 +1133,7 @@ function createRunner({ watchForChanges = true, loadEnv = true } = {}) {
 				updated: false,
 				reason: "Manual score not found"
 			};
-			const caseDetail = run.caseDetails.get(getCaseRowCaseKey(caseRow));
+			const caseDetail = hydrateCaseDetailForRow(run, caseRow);
 			if (!caseDetail) return {
 				updated: false,
 				reason: "Case detail not found"
@@ -1480,7 +1506,7 @@ function createRunner({ watchForChanges = true, loadEnv = true } = {}) {
 		getCaseDetail(runId, caseId) {
 			const run = runs.get(runId);
 			if (!run) return void 0;
-			return resolveCaseDetailLookup(run, caseId);
+			return hydrateCaseDetailForLookup(run, caseId);
 		},
 		subscribe(runId, listener) {
 			const run = runs.get(runId);
@@ -2232,8 +2258,8 @@ 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-DPCFFkyQ.mjs");
-	const runnerModule = await import("./runner-XEP21_u9.mjs");
+	const appModule = await import("./app-l3ynaNsb.mjs");
+	const runnerModule = await import("./runner-C9xNJHt3.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

@@ -1942,7 +1942,6 @@ declare const columnFormatSchema: z.ZodEnum<{
   number: "number";
   boolean: "boolean";
   file: "file";
-  duration: "duration";
   markdown: "markdown";
   json: "json";
   image: "image";
@@ -1950,6 +1949,7 @@ declare const columnFormatSchema: z.ZodEnum<{
   pdf: "pdf";
   audio: "audio";
   video: "video";
+  duration: "duration";
   percent: "percent";
   passFail: "passFail";
   stars: "stars";
@@ -1969,7 +1969,6 @@ declare const columnDefSchema: z.ZodObject<{
     number: "number";
     boolean: "boolean";
     file: "file";
-    duration: "duration";
     markdown: "markdown";
     json: "json";
     image: "image";
@@ -1977,6 +1976,7 @@ declare const columnDefSchema: z.ZodObject<{
     pdf: "pdf";
     audio: "audio";
     video: "video";
+    duration: "duration";
     percent: "percent";
     passFail: "passFail";
     stars: "stars";
@@ -2022,8 +2022,8 @@ type CellValue = z.infer<typeof cellValueSchema>; //#endregion
 declare const traceAttributeDisplayFormatSchema: z.ZodEnum<{
   string: "string";
   number: "number";
-  duration: "duration";
   json: "json";
+  duration: "duration";
 }>;
 /**
  * Formatting hint for trace attribute values rendered by the UI.
@@ -2047,8 +2047,8 @@ declare const traceAttributeDisplaySchema: z.ZodObject<{
   format: z.ZodOptional<z.ZodEnum<{
     string: "string";
     number: "number";
-    duration: "duration";
     json: "json";
+    duration: "duration";
   }>>;
   numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
   placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
@@ -2083,8 +2083,8 @@ declare const traceDisplayConfigSchema: z.ZodObject<{
     format: z.ZodOptional<z.ZodEnum<{
       string: "string";
       number: "number";
-      duration: "duration";
       json: "json";
+      duration: "duration";
     }>>;
     numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
     placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
@@ -2123,8 +2123,8 @@ declare const traceAttributeDisplayInputSchema: z.ZodObject<{
   format: z.ZodOptional<z.ZodEnum<{
     string: "string";
     number: "number";
-    duration: "duration";
     json: "json";
+    duration: "duration";
   }>>;
   numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
   placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
@@ -2161,8 +2161,8 @@ declare const traceDisplayInputConfigSchema: z.ZodObject<{
     format: z.ZodOptional<z.ZodEnum<{
       string: "string";
       number: "number";
-      duration: "duration";
       json: "json";
+      duration: "duration";
     }>>;
     numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
     placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
@@ -2327,7 +2327,6 @@ declare const evalStatItemSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
     number: "number";
     boolean: "boolean";
     file: "file";
-    duration: "duration";
     markdown: "markdown";
     json: "json";
     image: "image";
@@ -2335,6 +2334,7 @@ declare const evalStatItemSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
     pdf: "pdf";
     audio: "audio";
     video: "video";
+    duration: "duration";
     percent: "percent";
     passFail: "passFail";
     stars: "stars";
@@ -2391,7 +2391,6 @@ declare const evalStatsConfigSchema: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodOb
     number: "number";
     boolean: "boolean";
     file: "file";
-    duration: "duration";
     markdown: "markdown";
     json: "json";
     image: "image";
@@ -2399,6 +2398,7 @@ declare const evalStatsConfigSchema: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodOb
     pdf: "pdf";
     audio: "audio";
     video: "video";
+    duration: "duration";
     percent: "percent";
     passFail: "passFail";
     stars: "stars";
@@ -2437,7 +2437,6 @@ declare const evalSummarySchema$1: z.ZodObject<{
       number: "number";
       boolean: "boolean";
       file: "file";
-      duration: "duration";
       markdown: "markdown";
       json: "json";
       image: "image";
@@ -2445,6 +2444,7 @@ declare const evalSummarySchema$1: z.ZodObject<{
       pdf: "pdf";
       audio: "audio";
       video: "video";
+      duration: "duration";
       percent: "percent";
       passFail: "passFail";
       stars: "stars";
@@ -2518,7 +2518,6 @@ declare const evalSummarySchema$1: z.ZodObject<{
       number: "number";
       boolean: "boolean";
       file: "file";
-      duration: "duration";
       markdown: "markdown";
       json: "json";
       image: "image";
@@ -2526,6 +2525,7 @@ declare const evalSummarySchema$1: z.ZodObject<{
       pdf: "pdf";
       audio: "audio";
       video: "video";
+      duration: "duration";
       percent: "percent";
       passFail: "passFail";
       stars: "stars";
@@ -2559,8 +2559,8 @@ declare const evalSummarySchema$1: z.ZodObject<{
       label: z.ZodOptional<z.ZodString>;
       color: z.ZodOptional<z.ZodEnum<{
         success: "success";
-        accent: "accent";
         error: "error";
+        accent: "accent";
         accentDim: "accentDim";
         warning: "warning";
         textMuted: "textMuted";
@@ -2583,8 +2583,8 @@ declare const evalSummarySchema$1: z.ZodObject<{
       label: z.ZodOptional<z.ZodString>;
       color: z.ZodOptional<z.ZodEnum<{
         success: "success";
-        accent: "accent";
         error: "error";
+        accent: "accent";
         accentDim: "accentDim";
         warning: "warning";
         textMuted: "textMuted";
@@ -2749,7 +2749,6 @@ declare const caseRowSchema$1: z.ZodObject<{
       number: "number";
       boolean: "boolean";
       file: "file";
-      duration: "duration";
       markdown: "markdown";
       json: "json";
       image: "image";
@@ -2757,6 +2756,7 @@ declare const caseRowSchema$1: z.ZodObject<{
       pdf: "pdf";
       audio: "audio";
       video: "video";
+      duration: "duration";
       percent: "percent";
       passFail: "passFail";
       stars: "stars";
@@ -2894,8 +2894,8 @@ declare const scoreTraceSchema: z.ZodObject<{
       format: z.ZodOptional<z.ZodEnum<{
         string: "string";
         number: "number";
-        duration: "duration";
         json: "json";
+        duration: "duration";
       }>>;
       numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
       placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
@@ -2997,8 +2997,8 @@ declare const caseDetailSchema$1: z.ZodObject<{
       format: z.ZodOptional<z.ZodEnum<{
         string: "string";
         number: "number";
-        duration: "duration";
         json: "json";
+        duration: "duration";
       }>>;
       numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
       placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
@@ -3066,8 +3066,8 @@ declare const caseDetailSchema$1: z.ZodObject<{
         format: z.ZodOptional<z.ZodEnum<{
           string: "string";
           number: "number";
-          duration: "duration";
           json: "json";
+          duration: "duration";
         }>>;
         numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
         placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
@@ -3127,7 +3127,6 @@ declare const caseDetailSchema$1: z.ZodObject<{
       number: "number";
       boolean: "boolean";
       file: "file";
-      duration: "duration";
       markdown: "markdown";
       json: "json";
       image: "image";
@@ -3135,6 +3134,7 @@ declare const caseDetailSchema$1: z.ZodObject<{
       pdf: "pdf";
       audio: "audio";
       video: "video";
+      duration: "duration";
       percent: "percent";
       passFail: "passFail";
       stars: "stars";
@@ -3284,8 +3284,8 @@ type EvalChartAggregate = z.infer<typeof evalChartAggregateSchema>;
  */
 declare const evalChartColorSchema: z.ZodEnum<{
   success: "success";
-  accent: "accent";
   error: "error";
+  accent: "accent";
   accentDim: "accentDim";
   warning: "warning";
   textMuted: "textMuted";
@@ -3313,8 +3313,8 @@ declare const evalChartMetricSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
   label: z.ZodOptional<z.ZodString>;
   color: z.ZodOptional<z.ZodEnum<{
     success: "success";
-    accent: "accent";
     error: "error";
+    accent: "accent";
     accentDim: "accentDim";
     warning: "warning";
     textMuted: "textMuted";
@@ -3337,8 +3337,8 @@ declare const evalChartMetricSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
   label: z.ZodOptional<z.ZodString>;
   color: z.ZodOptional<z.ZodEnum<{
     success: "success";
-    accent: "accent";
     error: "error";
+    accent: "accent";
     accentDim: "accentDim";
     warning: "warning";
     textMuted: "textMuted";
@@ -3396,8 +3396,8 @@ declare const evalChartConfigSchema: z.ZodObject<{
     label: z.ZodOptional<z.ZodString>;
     color: z.ZodOptional<z.ZodEnum<{
       success: "success";
-      accent: "accent";
       error: "error";
+      accent: "accent";
       accentDim: "accentDim";
       warning: "warning";
       textMuted: "textMuted";
@@ -3420,8 +3420,8 @@ declare const evalChartConfigSchema: z.ZodObject<{
     label: z.ZodOptional<z.ZodString>;
     color: z.ZodOptional<z.ZodEnum<{
       success: "success";
-      accent: "accent";
       error: "error";
+      accent: "accent";
       accentDim: "accentDim";
       warning: "warning";
       textMuted: "textMuted";
@@ -3486,8 +3486,8 @@ declare const evalChartsConfigSchema: z.ZodArray<z.ZodObject<{
     label: z.ZodOptional<z.ZodString>;
     color: z.ZodOptional<z.ZodEnum<{
       success: "success";
-      accent: "accent";
       error: "error";
+      accent: "accent";
       accentDim: "accentDim";
       warning: "warning";
       textMuted: "textMuted";
@@ -3510,8 +3510,8 @@ declare const evalChartsConfigSchema: z.ZodArray<z.ZodObject<{
     label: z.ZodOptional<z.ZodString>;
     color: z.ZodOptional<z.ZodEnum<{
       success: "success";
-      accent: "accent";
       error: "error";
+      accent: "accent";
       accentDim: "accentDim";
       warning: "warning";
       textMuted: "textMuted";
@@ -3808,8 +3808,8 @@ declare const llmCallMetricFormatSchema$1: z.ZodEnum<{
   string: "string";
   number: "number";
   boolean: "boolean";
-  duration: "duration";
   json: "json";
+  duration: "duration";
 }>;
 /** Render format applied to an LLM-call metric value. */
 type LlmCallMetricFormat = z.infer<typeof llmCallMetricFormatSchema$1>;
@@ -3818,8 +3818,8 @@ declare const apiCallMetricFormatSchema$1: z.ZodEnum<{
   string: "string";
   number: "number";
   boolean: "boolean";
-  duration: "duration";
   json: "json";
+  duration: "duration";
 }>;
 /** Render format applied to an API-call metric value. */
 type ApiCallMetricFormat = z.infer<typeof apiCallMetricFormatSchema$1>;
@@ -3888,8 +3888,8 @@ declare const llmCallMetricSchema: z.ZodObject<{
     string: "string";
     number: "number";
     boolean: "boolean";
-    duration: "duration";
     json: "json";
+    duration: "duration";
   }>>;
   numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
   placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
@@ -3917,8 +3917,8 @@ declare const apiCallMetricSchema: z.ZodObject<{
     string: "string";
     number: "number";
     boolean: "boolean";
-    duration: "duration";
     json: "json";
+    duration: "duration";
   }>>;
   numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
   placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
@@ -4031,8 +4031,8 @@ declare const llmCallsConfigSchema: z.ZodObject<{
       string: "string";
       number: "number";
       boolean: "boolean";
-      duration: "duration";
       json: "json";
+      duration: "duration";
     }>>;
     numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
     placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
@@ -4067,8 +4067,8 @@ declare const apiCallsConfigSchema: z.ZodObject<{
       string: "string";
       number: "number";
       boolean: "boolean";
-      duration: "duration";
       json: "json";
+      duration: "duration";
     }>>;
     numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
     placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
@@ -4670,7 +4670,6 @@ declare const cacheRecordingOpSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
       number: "number";
       boolean: "boolean";
       file: "file";
-      duration: "duration";
       markdown: "markdown";
       json: "json";
       image: "image";
@@ -4678,6 +4677,7 @@ declare const cacheRecordingOpSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
       pdf: "pdf";
       audio: "audio";
       video: "video";
+      duration: "duration";
       percent: "percent";
       passFail: "passFail";
       stars: "stars";
@@ -4758,7 +4758,6 @@ declare const cacheRecordingSchema: z.ZodObject<{
         number: "number";
         boolean: "boolean";
         file: "file";
-        duration: "duration";
         markdown: "markdown";
         json: "json";
         image: "image";
@@ -4766,6 +4765,7 @@ declare const cacheRecordingSchema: z.ZodObject<{
         pdf: "pdf";
         audio: "audio";
         video: "video";
+        duration: "duration";
         percent: "percent";
         passFail: "passFail";
         stars: "stars";
@@ -4859,7 +4859,6 @@ declare const cacheEntrySchema: z.ZodObject<{
           number: "number";
           boolean: "boolean";
           file: "file";
-          duration: "duration";
           markdown: "markdown";
           json: "json";
           image: "image";
@@ -4867,6 +4866,7 @@ declare const cacheEntrySchema: z.ZodObject<{
           pdf: "pdf";
           audio: "audio";
           video: "video";
+          duration: "duration";
           percent: "percent";
           passFail: "passFail";
           stars: "stars";
@@ -4977,7 +4977,6 @@ declare const cacheDebugKeyEntrySchema: z.ZodObject<{
             number: "number";
             boolean: "boolean";
             file: "file";
-            duration: "duration";
             markdown: "markdown";
             json: "json";
             image: "image";
@@ -4985,6 +4984,7 @@ declare const cacheDebugKeyEntrySchema: z.ZodObject<{
             pdf: "pdf";
             audio: "audio";
             video: "video";
+            duration: "duration";
             percent: "percent";
             passFail: "passFail";
             stars: "stars";
@@ -5084,7 +5084,6 @@ declare const cacheEntryWithDebugKeySchema$1: z.ZodObject<{
           number: "number";
           boolean: "boolean";
           file: "file";
-          duration: "duration";
           markdown: "markdown";
           json: "json";
           image: "image";
@@ -5092,6 +5091,7 @@ declare const cacheEntryWithDebugKeySchema$1: z.ZodObject<{
           pdf: "pdf";
           audio: "audio";
           video: "video";
+          duration: "duration";
           percent: "percent";
           passFail: "passFail";
           stars: "stars";
@@ -5193,7 +5193,6 @@ declare const cacheEntryWithDebugKeySchema$1: z.ZodObject<{
               number: "number";
               boolean: "boolean";
               file: "file";
-              duration: "duration";
               markdown: "markdown";
               json: "json";
               image: "image";
@@ -5201,6 +5200,7 @@ declare const cacheEntryWithDebugKeySchema$1: z.ZodObject<{
               pdf: "pdf";
               audio: "audio";
               video: "video";
+              duration: "duration";
               percent: "percent";
               passFail: "passFail";
               stars: "stars";
@@ -5300,7 +5300,6 @@ declare const cacheFileSchema: z.ZodObject<{
             number: "number";
             boolean: "boolean";
             file: "file";
-            duration: "duration";
             markdown: "markdown";
             json: "json";
             image: "image";
@@ -5308,6 +5307,7 @@ declare const cacheFileSchema: z.ZodObject<{
             pdf: "pdf";
             audio: "audio";
             video: "video";
+            duration: "duration";
             percent: "percent";
             passFail: "passFail";
             stars: "stars";
@@ -5417,7 +5417,6 @@ declare const cacheDebugKeyFileSchema: z.ZodObject<{
               number: "number";
               boolean: "boolean";
               file: "file";
-              duration: "duration";
               markdown: "markdown";
               json: "json";
               image: "image";
@@ -5425,6 +5424,7 @@ declare const cacheDebugKeyFileSchema: z.ZodObject<{
               pdf: "pdf";
               audio: "audio";
               video: "video";
+              duration: "duration";
               percent: "percent";
               passFail: "passFail";
               stars: "stars";

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-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";
+import { a as materializeManualInputFiles, i as isManualInputFileValue, n as createRunner, o as stageManualInputFile, r as cleanupStagedManualInputFiles, s as stageManualInputFileFromPath, t as runCli } from "./cli-BSVUCUxr.mjs";
+import { n as matchesEvalTags, t as defineEval } from "./src-D5vGo2iv.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-Bq0Y3y_1.mjs";
-import { S as parseEvalDiscovery, m as persistRunState, r as getTargetEvals$1, t as executeRun } from "./runOrchestration-BpwW0AmB.mjs";
+import { C as parseEvalDiscovery, h as persistRunState, r as getTargetEvals$1, t as executeRun } from "./runOrchestration-C7qQISz2.mjs";
 import { z } from "zod";
 import { readFile } from "node:fs/promises";
 import { relative } from "node:path";

package/dist/{runOrchestration-BpwW0AmB.mjs → runOrchestration-C7qQISz2.mjs}

@@ -2,7 +2,7 @@ import { Dt as caseDetailSchema, Et as getCaseRowCaseKey, It as runWithEvalRegis
 import { Result, resultify } from "t-result";
 import { readFile, readdir, rm, writeFile } from "node:fs/promises";
 import { dirname, join } from "node:path";
-import { existsSync } from "node:fs";
+import { existsSync, readFileSync } from "node:fs";
 import { fileURLToPath } from "node:url";
 import { spawn } from "node:child_process";
 //#region ../runner/src/chartValidation.ts
@@ -670,7 +670,7 @@ async function recomputeEvalStatusesInRuns(params) {
 		let changed = false;
 		for (const caseRow of run.cases) {
 			if (caseRow.evalKey !== params.evalKey) continue;
-			const caseDetail = run.caseDetails.get(getCaseRowCaseKey(caseRow));
+			const caseDetail = params.getCaseDetail?.(run, caseRow) ?? run.caseDetails.get(getCaseRowCaseKey(caseRow));
 			const nextStatus = recomputePersistedCaseStatus(caseRow, caseDetail, params.scoreThresholds);
 			if (caseRow.status === nextStatus) continue;
 			caseRow.status = nextStatus;
@@ -725,14 +725,22 @@ function nextShortIdFromSnapshots(snapshots) {
 	}
 	return maxNum + 1;
 }
-async function loadPersistedRunSnapshots(localStateDir) {
+/**
+* Load persisted run metadata from the local state directory.
+*
+* Case details are skipped by default so long-running app processes can keep
+* run history in memory without retaining every trace payload. Pass
+* `includeCaseDetails` only for narrow maintenance flows that need full
+* details for every case.
+*/
+async function loadPersistedRunSnapshots(localStateDir, options = {}) {
 	const runsDir = join(localStateDir, "runs");
 	const entriesResult = await resultify(() => readdir(runsDir, { withFileTypes: true }));
 	if (entriesResult.error) return [];
 	const snapshots = [];
 	const runDirs = entriesResult.value.filter((entry) => entry.isDirectory()).map((entry) => join(runsDir, entry.name)).toSorted();
 	for (const runDir of runDirs) {
-		const snapshot = await loadPersistedRunSnapshot(runDir);
+		const snapshot = await loadPersistedRunSnapshot(runDir, options);
 		if (!snapshot) continue;
 		snapshots.push(snapshot);
 	}
@@ -766,7 +774,14 @@ function getLatestRunInfos(params) {
 function toLastRunStatus$1(status) {
 	return status === "pending" ? null : status;
 }
-async function loadPersistedRunSnapshot(runDir) {
+/**
+* Load one persisted run snapshot from disk.
+*
+* The returned snapshot includes manifest, summary, and case rows. Case
+* details are loaded only when `includeCaseDetails` is true; otherwise callers
+* should use `loadPersistedCaseDetail` for the specific case being inspected.
+*/
+async function loadPersistedRunSnapshot(runDir, options = {}) {
 	const manifest = await readParsedJsonFile(join(runDir, "run.json"), { safeParse: runManifestSchema.safeParse.bind(runManifestSchema) });
 	if (!manifest) return null;
 	const summary = await readParsedJsonFile(join(runDir, "summary.json"), { safeParse: runSummarySchema.safeParse.bind(runSummarySchema) });
@@ -776,9 +791,18 @@ async function loadPersistedRunSnapshot(runDir) {
 		manifest,
 		summary,
 		cases: await readCaseRows(runDir),
-		caseDetails: await readCaseDetails(runDir)
+		caseDetails: options.includeCaseDetails === true ? await readCaseDetails(runDir) : /* @__PURE__ */ new Map()
 	};
 }
+/**
+* Load one persisted case detail by its artifact file id.
+*
+* Returns `null` when the file is missing, invalid JSON, or no longer matches
+* the current case-detail schema.
+*/
+function loadPersistedCaseDetail(runDir, fileId) {
+	return readParsedJsonFileSync(join(runDir, "case-details", `${encodeCaseDetailFileName(fileId)}.json`), { safeParse: caseDetailSchema.safeParse.bind(caseDetailSchema) });
+}
 async function readParsedJsonFile(filePath, schema) {
 	const fileResult = await resultify(() => readFile(filePath, "utf-8"));
 	if (fileResult.error) return null;
@@ -788,6 +812,15 @@ async function readParsedJsonFile(filePath, schema) {
 	if (!parsed.success) return null;
 	return parsed.data;
 }
+function readParsedJsonFileSync(filePath, schema) {
+	const fileResult = resultify(() => readFileSync(filePath, "utf-8"));
+	if (fileResult.error) return null;
+	const jsonResult = resultify(() => JSON.parse(fileResult.value));
+	if (jsonResult.error) return null;
+	const parsed = schema.safeParse(jsonResult.value);
+	if (!parsed.success) return null;
+	return parsed.data;
+}
 async function readCaseRows(runDir) {
 	const fileResult = await resultify(() => readFile(join(runDir, "cases.jsonl"), "utf-8"));
 	if (fileResult.error) return [];
@@ -1660,4 +1693,4 @@ function toLastRunStatus(status) {
 	return status === "pending" ? null : status;
 }
 //#endregion
-export { validateCharts as C, parseEvalDiscovery as S, runTouchesEval as _, validateTagsFilters as a, deriveEvalFreshness as b, getLatestRunInfos as c, nextShortIdFromSnapshots as d, persistCaseDetail as f, recomputePersistedCaseStatus as g, recomputeEvalStatusesInRuns as h, resolveEvalTags as i, loadPersistedRunSnapshot as l, persistRunState as m, getTargetEvalKeys as n, generateRunId as o, deleteTemporaryRuns as p, getTargetEvals as r, getLastRunStatuses as s, executeRun as t, loadPersistedRunSnapshots as u, buildManualInputDescriptor as v, loadIsolatedEvalRegistry as x, parseManualInputValues as y };
+export { parseEvalDiscovery as C, loadIsolatedEvalRegistry as S, recomputePersistedCaseStatus as _, validateTagsFilters as a, parseManualInputValues as b, getLatestRunInfos as c, loadPersistedRunSnapshots as d, nextShortIdFromSnapshots as f, recomputeEvalStatusesInRuns as g, persistRunState as h, resolveEvalTags as i, loadPersistedCaseDetail as l, deleteTemporaryRuns as m, getTargetEvalKeys as n, generateRunId as o, persistCaseDetail as p, getTargetEvals as r, getLastRunStatuses as s, executeRun as t, loadPersistedRunSnapshot as u, runTouchesEval as v, validateCharts as w, deriveEvalFreshness as x, buildManualInputDescriptor as y };

package/dist/{runner-XEP21_u9.mjs → runner-C9xNJHt3.mjs}

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

package/dist/{runner-Kp0JqxrU.mjs → runner-DmkSq-QG.mjs}

@@ -1,5 +1,5 @@
-import { n as createRunner } from "./cli-CbePEEua.mjs";
-import "./src-CVM_FqPx.mjs";
+import { n as createRunner } from "./cli-BSVUCUxr.mjs";
+import "./src-D5vGo2iv.mjs";
 //#region ../../apps/server/src/runner.ts
 let runnerInstance = null;
 function getRunnerInstance({ loadEnv = true } = {}) {

package/dist/{src-CVM_FqPx.mjs → src-D5vGo2iv.mjs}

@@ -1,5 +1,5 @@
 import { G as matchesEvalTags$1, Pt as defineEval$1 } from "./runExecution-Bq0Y3y_1.mjs";
-import "./cli-CbePEEua.mjs";
+import "./cli-BSVUCUxr.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": "0.60.2",
+  "version": "0.60.3",
   "type": "module",
   "bin": {
     "agent-evals": "./dist/bin.mjs"