@ls-stack/agent-eval 0.60.3 → 0.60.4

package/dist/{app-l3ynaNsb.mjs → app-gg10KvzS.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-BSVUCUxr.mjs";
-import "./src-D5vGo2iv.mjs";
-import { t as getRunnerInstance } from "./runner-DmkSq-QG.mjs";
+import { Dt as caseRowSchema, Tt as getCaseRowCaseKey, et as createRunRequestSchema, nt as extractCacheEntries, tt as updateManualScoreRequestSchema, ut as getEvalTitle } from "./runExecution-Bu9yfdUS.mjs";
+import { o as stageManualInputFile } from "./cli-OLZIjQpx.mjs";
+import "./src-Cy3OxoZW.mjs";
+import { t as getRunnerInstance } from "./runner-SxtKn-Xh.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-BSVUCUxr.mjs";
+import { t as runCli } from "./cli-OLZIjQpx.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-Bq0Y3y_1.mjs";
+import { Ft as runWithEvalRegistry, I as configureEvalRunLogs, St as resolveLlmCallsConfig, _ as createFsCacheStore, a as isCaseChildParentMessage, d as loadEvalModule, g as createBufferedCacheStore, l as registerAgentEvalsPackageResolutionHooks, p as loadConfig, q as runInEvalRuntimeScope, r as runCase, v as getCacheRetentionOptions, xt as resolveApiCallsConfig } from "./runExecution-Bu9yfdUS.mjs";
 //#region ../runner/src/caseChild.ts
 let fatalErrorReported = false;
 let disconnectExpected = false;

package/dist/{cli-BSVUCUxr.mjs → cli-OLZIjQpx.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 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 { Dt as caseRowSchema, Pt as getEvalRegistry, St as resolveLlmCallsConfig, Tt as getCaseRowCaseKey, _ as createFsCacheStore, bt as runSummarySchema, c as resolveArtifactPath, ct as applyDerivedCallAttributes, dt as getEvalDisplayStatus, f as resolveEvalDefaultConfig, ft as deriveScopedSummaryFromCases, gt as matchesTagsFilter, m as buildDeclaredColumnDefs, o as stripTerminalControlCodes, p as loadConfig, s as resolveTracePresentation, ut as getEvalTitle, v as getCacheRetentionOptions, wt as buildEvalKey, xt as resolveApiCallsConfig } from "./runExecution-Bu9yfdUS.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-mpgZmEZ6.mjs";
 import { parseEnv } from "node:util";
 import { resultify } from "t-result";
 import { copyFile, mkdir, readFile, rm, writeFile } from "node:fs/promises";
@@ -973,7 +973,6 @@ function createRunner({ watchForChanges = true, loadEnv = true } = {}) {
 	let discoveryRefreshTimer;
 	let runHistoryRefreshTimer;
 	let cachePruneIdleTimer;
-	let registryLoadCounter = 0;
 	const configReload = createConfigReloadController({
 		getActiveRunCount,
 		closeRunnerWatchers: closeWatchers,
@@ -1019,9 +1018,8 @@ function createRunner({ watchForChanges = true, loadEnv = true } = {}) {
 		if (caseRow === void 0) return void 0;
 		return hydrateCaseDetailForRow(run, caseRow);
 	}
-	function nextRegistryLoadIsolationKey(prefix, filePath) {
-		registryLoadCounter++;
-		return `${prefix}:${String(registryLoadCounter)}:${filePath}`;
+	function getDiscoveryModuleIsolationKey(filePath) {
+		return `discovery:${filePath}`;
 	}
 	const runner = {
 		async init() {
@@ -1042,24 +1040,11 @@ function createRunner({ watchForChanges = true, loadEnv = true } = {}) {
 		async recomputeStatusesForEval(evalKey) {
 			const evalMeta = resolveEvalMeta(evalKey);
 			if (!evalMeta) return { updatedRuns: 0 };
-			const entry = (await loadIsolatedEvalRegistry({
-				evalFilePath: evalMeta.sourceFilePath,
-				sourceFingerprint: evalMeta.sourceFingerprint ?? void 0,
-				moduleIsolation: {
-					key: nextRegistryLoadIsolationKey("recompute-status", evalMeta.sourceFilePath),
-					workspaceRoot
-				},
-				runtimeScope: "env"
-			})).get(evalMeta.id);
-			if (!entry) return { updatedRuns: 0 };
 			const scoreThresholds = /* @__PURE__ */ new Map();
-			entry.use((evalDef) => {
-				for (const [key, def] of Object.entries(evalDef.scores ?? {})) {
-					const threshold = normalizeScoreDef(def).passThreshold;
-					if (threshold !== void 0) scoreThresholds.set(key, threshold);
-				}
-				for (const [key, def] of Object.entries(evalDef.manualScores ?? {})) if (def.passThreshold !== void 0) scoreThresholds.set(key, def.passThreshold);
-			});
+			for (const columnDef of evalMeta.columnDefs) {
+				if (columnDef.isScore !== true || columnDef.passThreshold === void 0) continue;
+				scoreThresholds.set(columnDef.key, columnDef.passThreshold);
+			}
 			const updatedRuns = await recomputeEvalStatusesInRuns({
 				runs: runs.values(),
 				evalKey: evalMeta.key,
@@ -1264,7 +1249,7 @@ function createRunner({ watchForChanges = true, loadEnv = true } = {}) {
 						evalFilePath: filePath,
 						sourceFingerprint,
 						moduleIsolation: {
-							key: nextRegistryLoadIsolationKey("discovery", filePath),
+							key: getDiscoveryModuleIsolationKey(filePath),
 							workspaceRoot
 						},
 						runtimeScope: "env"
@@ -2258,8 +2243,8 @@ 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-l3ynaNsb.mjs");
-	const runnerModule = await import("./runner-C9xNJHt3.mjs");
+	const appModule = await import("./app-gg10KvzS.mjs");
+	const runnerModule = await import("./runner-C4Y0lWb1.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

@@ -2214,8 +2214,8 @@ declare const traceSpanSchema$1: z.ZodObject<{
   status: z.ZodEnum<{
     error: "error";
     running: "running";
-    cancelled: "cancelled";
     ok: "ok";
+    cancelled: "cancelled";
   }>;
   attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
   error: z.ZodOptional<z.ZodObject<{
@@ -2466,10 +2466,10 @@ declare const evalSummarySchema$1: z.ZodObject<{
   caseIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
   lastRunStatus: z.ZodNullable<z.ZodEnum<{
     error: "error";
-    pass: "pass";
-    fail: "fail";
     running: "running";
     cancelled: "cancelled";
+    pass: "pass";
+    fail: "fail";
     unscored: "unscored";
   }>>;
   stats: z.ZodOptional<z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
@@ -2558,8 +2558,8 @@ declare const evalSummarySchema$1: z.ZodObject<{
       }>;
       label: z.ZodOptional<z.ZodString>;
       color: z.ZodOptional<z.ZodEnum<{
-        success: "success";
         error: "error";
+        success: "success";
         accent: "accent";
         accentDim: "accentDim";
         warning: "warning";
@@ -2582,8 +2582,8 @@ declare const evalSummarySchema$1: z.ZodObject<{
       }>;
       label: z.ZodOptional<z.ZodString>;
       color: z.ZodOptional<z.ZodEnum<{
-        success: "success";
         error: "error";
+        success: "success";
         accent: "accent";
         accentDim: "accentDim";
         warning: "warning";
@@ -2715,10 +2715,10 @@ declare const caseRowSchema$1: z.ZodObject<{
   tags: z.ZodOptional<z.ZodArray<z.ZodString>>;
   status: z.ZodEnum<{
     error: "error";
-    pass: "pass";
-    fail: "fail";
     running: "running";
     cancelled: "cancelled";
+    pass: "pass";
+    fail: "fail";
     pending: "pending";
   }>;
   durationMs: z.ZodNullable<z.ZodNumber>;
@@ -2857,8 +2857,8 @@ declare const scoreTraceSchema: z.ZodObject<{
     status: z.ZodEnum<{
       error: "error";
       running: "running";
-      cancelled: "cancelled";
       ok: "ok";
+      cancelled: "cancelled";
     }>;
     attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
     error: z.ZodOptional<z.ZodObject<{
@@ -2920,10 +2920,10 @@ declare const scoreTraceSchema: z.ZodObject<{
     namespace: z.ZodString;
     key: z.ZodString;
     status: z.ZodEnum<{
+      bypass: "bypass";
+      refresh: "refresh";
       hit: "hit";
       miss: "miss";
-      refresh: "refresh";
-      bypass: "bypass";
     }>;
     read: z.ZodOptional<z.ZodBoolean>;
     stored: z.ZodOptional<z.ZodBoolean>;
@@ -2942,10 +2942,10 @@ declare const caseDetailSchema$1: z.ZodObject<{
   tags: z.ZodOptional<z.ZodArray<z.ZodString>>;
   status: z.ZodEnum<{
     error: "error";
-    pass: "pass";
-    fail: "fail";
     running: "running";
     cancelled: "cancelled";
+    pass: "pass";
+    fail: "fail";
     pending: "pending";
   }>;
   input: z.ZodUnknown;
@@ -2960,8 +2960,8 @@ declare const caseDetailSchema$1: z.ZodObject<{
     status: z.ZodEnum<{
       error: "error";
       running: "running";
-      cancelled: "cancelled";
       ok: "ok";
+      cancelled: "cancelled";
     }>;
     attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
     error: z.ZodOptional<z.ZodObject<{
@@ -3029,8 +3029,8 @@ declare const caseDetailSchema$1: z.ZodObject<{
       status: z.ZodEnum<{
         error: "error";
         running: "running";
-        cancelled: "cancelled";
         ok: "ok";
+        cancelled: "cancelled";
       }>;
       attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
       error: z.ZodOptional<z.ZodObject<{
@@ -3092,10 +3092,10 @@ declare const caseDetailSchema$1: z.ZodObject<{
       namespace: z.ZodString;
       key: z.ZodString;
       status: z.ZodEnum<{
+        bypass: "bypass";
+        refresh: "refresh";
         hit: "hit";
         miss: "miss";
-        refresh: "refresh";
-        bypass: "bypass";
       }>;
       read: z.ZodOptional<z.ZodBoolean>;
       stored: z.ZodOptional<z.ZodBoolean>;
@@ -3213,10 +3213,10 @@ declare const caseDetailSchema$1: z.ZodObject<{
     namespace: z.ZodString;
     key: z.ZodString;
     status: z.ZodEnum<{
+      bypass: "bypass";
+      refresh: "refresh";
       hit: "hit";
       miss: "miss";
-      refresh: "refresh";
-      bypass: "bypass";
     }>;
     read: z.ZodOptional<z.ZodBoolean>;
     stored: z.ZodOptional<z.ZodBoolean>;
@@ -3283,8 +3283,8 @@ type EvalChartAggregate = z.infer<typeof evalChartAggregateSchema>;
  * not emit raw hex so authored evals stay decoupled from the web theme.
  */
 declare const evalChartColorSchema: z.ZodEnum<{
-  success: "success";
   error: "error";
+  success: "success";
   accent: "accent";
   accentDim: "accentDim";
   warning: "warning";
@@ -3312,8 +3312,8 @@ declare const evalChartMetricSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
   }>;
   label: z.ZodOptional<z.ZodString>;
   color: z.ZodOptional<z.ZodEnum<{
-    success: "success";
     error: "error";
+    success: "success";
     accent: "accent";
     accentDim: "accentDim";
     warning: "warning";
@@ -3336,8 +3336,8 @@ declare const evalChartMetricSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
   }>;
   label: z.ZodOptional<z.ZodString>;
   color: z.ZodOptional<z.ZodEnum<{
-    success: "success";
     error: "error";
+    success: "success";
     accent: "accent";
     accentDim: "accentDim";
     warning: "warning";
@@ -3395,8 +3395,8 @@ declare const evalChartConfigSchema: z.ZodObject<{
     }>;
     label: z.ZodOptional<z.ZodString>;
     color: z.ZodOptional<z.ZodEnum<{
-      success: "success";
       error: "error";
+      success: "success";
       accent: "accent";
       accentDim: "accentDim";
       warning: "warning";
@@ -3419,8 +3419,8 @@ declare const evalChartConfigSchema: z.ZodObject<{
     }>;
     label: z.ZodOptional<z.ZodString>;
     color: z.ZodOptional<z.ZodEnum<{
-      success: "success";
       error: "error";
+      success: "success";
       accent: "accent";
       accentDim: "accentDim";
       warning: "warning";
@@ -3485,8 +3485,8 @@ declare const evalChartsConfigSchema: z.ZodArray<z.ZodObject<{
     }>;
     label: z.ZodOptional<z.ZodString>;
     color: z.ZodOptional<z.ZodEnum<{
-      success: "success";
       error: "error";
+      success: "success";
       accent: "accent";
       accentDim: "accentDim";
       warning: "warning";
@@ -3509,8 +3509,8 @@ declare const evalChartsConfigSchema: z.ZodArray<z.ZodObject<{
     }>;
     label: z.ZodOptional<z.ZodString>;
     color: z.ZodOptional<z.ZodEnum<{
-      success: "success";
       error: "error";
+      success: "success";
       accent: "accent";
       accentDim: "accentDim";
       warning: "warning";
@@ -3589,9 +3589,9 @@ declare const runManifestSchema$1: z.ZodObject<{
     median: "median";
   }>>>;
   cacheMode: z.ZodOptional<z.ZodEnum<{
-    refresh: "refresh";
-    bypass: "bypass";
     use: "use";
+    bypass: "bypass";
+    refresh: "refresh";
   }>>;
 }, z.core.$strip>;
 /** Persisted lifecycle metadata for a single eval run. */
@@ -4567,9 +4567,9 @@ declare function extractApiCalls(spans: EvalTraceSpan$1[], config: ResolvedApiCa
  * - `refresh`: never read, always write (forces re-execution and overwrites).
  */
 declare const cacheModeSchema: z.ZodEnum<{
-  refresh: "refresh";
-  bypass: "bypass";
   use: "use";
+  bypass: "bypass";
+  refresh: "refresh";
 }>;
 /** Mode controlling how cached spans behave during a run. */
 type CacheMode = z.infer<typeof cacheModeSchema>;
@@ -4583,17 +4583,17 @@ declare const spanCacheOptionsSchema: z.ZodObject<{
 type SpanCacheOptions = z.infer<typeof spanCacheOptionsSchema>;
 /** Category of operation stored in the eval cache. */
 declare const cacheOperationTypeSchema: z.ZodEnum<{
-  value: "value";
   span: "span";
+  value: "value";
 }>;
 /** Category of operation stored in the eval cache. */
 type CacheOperationType = z.infer<typeof cacheOperationTypeSchema>;
 /** Status of a cache lookup recorded on a span or case scope. */
 declare const cacheStatusSchema: z.ZodEnum<{
+  bypass: "bypass";
+  refresh: "refresh";
   hit: "hit";
   miss: "miss";
-  refresh: "refresh";
-  bypass: "bypass";
 }>;
 /** Status of a cache lookup recorded on a span or case scope. */
 type CacheStatus = z.infer<typeof cacheStatusSchema>;
@@ -4610,10 +4610,10 @@ declare const traceCacheRefSchema: z.ZodObject<{
   namespace: z.ZodString;
   key: z.ZodString;
   status: z.ZodEnum<{
+    bypass: "bypass";
+    refresh: "refresh";
     hit: "hit";
     miss: "miss";
-    refresh: "refresh";
-    bypass: "bypass";
   }>;
   read: z.ZodOptional<z.ZodBoolean>;
   stored: z.ZodOptional<z.ZodBoolean>;
@@ -4721,8 +4721,8 @@ declare const cacheRecordingSchema: z.ZodObject<{
   finalStatus: z.ZodOptional<z.ZodEnum<{
     error: "error";
     running: "running";
-    cancelled: "cancelled";
     ok: "ok";
+    cancelled: "cancelled";
   }>>;
   finalError: z.ZodOptional<z.ZodObject<{
     name: z.ZodOptional<z.ZodString>;
@@ -4809,8 +4809,8 @@ declare const cacheEntrySchema: z.ZodObject<{
   key: z.ZodString;
   namespace: z.ZodString;
   operationType: z.ZodOptional<z.ZodEnum<{
-    value: "value";
     span: "span";
+    value: "value";
   }>>;
   operationName: z.ZodOptional<z.ZodString>;
   spanName: z.ZodOptional<z.ZodString>;
@@ -4822,8 +4822,8 @@ declare const cacheEntrySchema: z.ZodObject<{
     finalStatus: z.ZodOptional<z.ZodEnum<{
       error: "error";
       running: "running";
-      cancelled: "cancelled";
       ok: "ok";
+      cancelled: "cancelled";
     }>>;
     finalError: z.ZodOptional<z.ZodObject<{
       name: z.ZodOptional<z.ZodString>;
@@ -4916,8 +4916,8 @@ declare const cacheDebugKeyEntrySchema: z.ZodObject<{
   key: z.ZodString;
   namespace: z.ZodString;
   operationType: z.ZodEnum<{
-    value: "value";
     span: "span";
+    value: "value";
   }>;
   operationName: z.ZodString;
   storedAt: z.ZodString;
@@ -4927,8 +4927,8 @@ declare const cacheDebugKeyEntrySchema: z.ZodObject<{
     key: z.ZodString;
     namespace: z.ZodString;
     operationType: z.ZodOptional<z.ZodEnum<{
-      value: "value";
       span: "span";
+      value: "value";
     }>>;
     operationName: z.ZodOptional<z.ZodString>;
     spanName: z.ZodOptional<z.ZodString>;
@@ -4940,8 +4940,8 @@ declare const cacheDebugKeyEntrySchema: z.ZodObject<{
       finalStatus: z.ZodOptional<z.ZodEnum<{
         error: "error";
         running: "running";
-        cancelled: "cancelled";
         ok: "ok";
+        cancelled: "cancelled";
       }>>;
       finalError: z.ZodOptional<z.ZodObject<{
         name: z.ZodOptional<z.ZodString>;
@@ -5034,8 +5034,8 @@ declare const cacheEntryWithDebugKeySchema$1: z.ZodObject<{
   key: z.ZodString;
   namespace: z.ZodString;
   operationType: z.ZodOptional<z.ZodEnum<{
-    value: "value";
     span: "span";
+    value: "value";
   }>>;
   operationName: z.ZodOptional<z.ZodString>;
   spanName: z.ZodOptional<z.ZodString>;
@@ -5047,8 +5047,8 @@ declare const cacheEntryWithDebugKeySchema$1: z.ZodObject<{
     finalStatus: z.ZodOptional<z.ZodEnum<{
       error: "error";
       running: "running";
-      cancelled: "cancelled";
       ok: "ok";
+      cancelled: "cancelled";
     }>>;
     finalError: z.ZodOptional<z.ZodObject<{
       name: z.ZodOptional<z.ZodString>;
@@ -5132,8 +5132,8 @@ declare const cacheEntryWithDebugKeySchema$1: z.ZodObject<{
     key: z.ZodString;
     namespace: z.ZodString;
     operationType: z.ZodEnum<{
-      value: "value";
       span: "span";
+      value: "value";
     }>;
     operationName: z.ZodString;
     storedAt: z.ZodString;
@@ -5143,8 +5143,8 @@ declare const cacheEntryWithDebugKeySchema$1: z.ZodObject<{
       key: z.ZodString;
       namespace: z.ZodString;
       operationType: z.ZodOptional<z.ZodEnum<{
-        value: "value";
         span: "span";
+        value: "value";
       }>>;
       operationName: z.ZodOptional<z.ZodString>;
       spanName: z.ZodOptional<z.ZodString>;
@@ -5156,8 +5156,8 @@ declare const cacheEntryWithDebugKeySchema$1: z.ZodObject<{
         finalStatus: z.ZodOptional<z.ZodEnum<{
           error: "error";
           running: "running";
-          cancelled: "cancelled";
           ok: "ok";
+          cancelled: "cancelled";
         }>>;
         finalError: z.ZodOptional<z.ZodObject<{
           name: z.ZodOptional<z.ZodString>;
@@ -5250,8 +5250,8 @@ declare const cacheFileSchema: z.ZodObject<{
     key: z.ZodString;
     namespace: z.ZodString;
     operationType: z.ZodOptional<z.ZodEnum<{
-      value: "value";
       span: "span";
+      value: "value";
     }>>;
     operationName: z.ZodOptional<z.ZodString>;
     spanName: z.ZodOptional<z.ZodString>;
@@ -5263,8 +5263,8 @@ declare const cacheFileSchema: z.ZodObject<{
       finalStatus: z.ZodOptional<z.ZodEnum<{
         error: "error";
         running: "running";
-        cancelled: "cancelled";
         ok: "ok";
+        cancelled: "cancelled";
       }>>;
       finalError: z.ZodOptional<z.ZodObject<{
         name: z.ZodOptional<z.ZodString>;
@@ -5356,8 +5356,8 @@ declare const cacheDebugKeyFileSchema: z.ZodObject<{
     key: z.ZodString;
     namespace: z.ZodString;
     operationType: z.ZodEnum<{
-      value: "value";
       span: "span";
+      value: "value";
     }>;
     operationName: z.ZodString;
     storedAt: z.ZodString;
@@ -5367,8 +5367,8 @@ declare const cacheDebugKeyFileSchema: z.ZodObject<{
       key: z.ZodString;
       namespace: z.ZodString;
       operationType: z.ZodOptional<z.ZodEnum<{
-        value: "value";
         span: "span";
+        value: "value";
       }>>;
       operationName: z.ZodOptional<z.ZodString>;
       spanName: z.ZodOptional<z.ZodString>;
@@ -5380,8 +5380,8 @@ declare const cacheDebugKeyFileSchema: z.ZodObject<{
         finalStatus: z.ZodOptional<z.ZodEnum<{
           error: "error";
           running: "running";
-          cancelled: "cancelled";
           ok: "ok";
+          cancelled: "cancelled";
         }>>;
         finalError: z.ZodOptional<z.ZodObject<{
           name: z.ZodOptional<z.ZodString>;
@@ -5579,9 +5579,9 @@ declare const createRunRequestSchema$1: z.ZodObject<{
   temporary: z.ZodOptional<z.ZodBoolean>;
   cache: z.ZodOptional<z.ZodObject<{
     mode: z.ZodDefault<z.ZodEnum<{
-      refresh: "refresh";
-      bypass: "bypass";
       use: "use";
+      bypass: "bypass";
+      refresh: "refresh";
     }>>;
   }, z.core.$strip>>;
   manualInputs: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;

package/dist/index.mjs

@@ -1,4 +1,4 @@
-import { $ as setScopeCacheContext, A as repoFile, B as evalTime, C as evalTracer, D as deserializeCacheValue, E as deserializeCacheRecording, F as EvalRuntimeUsageError, Ft as getEvalRegistry, H as getEvalCaseInput, I as appendToEvalOutput, J as runInEvalRuntimeScope, K as mergeEvalOutput, M as readManualInputFile, N as evalExpect, O as serializeCacheRecording, P as EvalAssertionError, Q as setEvalOutput, R as evalAssert, S as evalSpan, T as hashCacheKeySync, U as incrementEvalOutput, V as getCurrentScope, W as isInEvalScope, X as runInExistingEvalScope, Y as runInEvalScope, at as extractApiCalls, b as buildTraceTree, ct as simulateTokenAllocation, et as startEvalBackgroundJob, it as extractCacheHits, j as manualInputFileValueSchema, k as serializeCacheValue, ot as extractLlmCalls, q as nextEvalId, rt as extractCacheEntries, st as simulateLlmCallCost, ut as getNestedAttribute, w as hashCacheKey, x as captureEvalSpanError, 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-BSVUCUxr.mjs";
-import { n as matchesEvalTags, t as defineEval } from "./src-D5vGo2iv.mjs";
+import { $ as startEvalBackgroundJob, A as manualInputFileValueSchema, B as getCurrentScope, C as hashCacheKey, D as serializeCacheRecording, E as deserializeCacheValue, F as appendToEvalOutput, G as mergeEvalOutput, H as incrementEvalOutput, J as runInEvalScope, K as nextEvalId, L as evalAssert, M as evalExpect, N as EvalAssertionError, O as serializeCacheValue, P as EvalRuntimeUsageError, Pt as getEvalRegistry, Q as setScopeCacheContext, R as evalLog, S as evalTracer, T as deserializeCacheRecording, U as isInEvalScope, V as getEvalCaseInput, Y as runInExistingEvalScope, Z as setEvalOutput, at as extractLlmCalls, b as captureEvalSpanError, it as extractApiCalls, j as readManualInputFile, k as repoFile, lt as getNestedAttribute, nt as extractCacheEntries, ot as simulateLlmCallCost, q as runInEvalRuntimeScope, rt as extractCacheHits, st as simulateTokenAllocation, w as hashCacheKeySync, x as evalSpan, y as buildTraceTree, z as evalTime } from "./runExecution-Bu9yfdUS.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-OLZIjQpx.mjs";
+import { n as matchesEvalTags, t as defineEval } from "./src-Cy3OxoZW.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 { C as parseEvalDiscovery, h as persistRunState, r as getTargetEvals$1, t as executeRun } from "./runOrchestration-C7qQISz2.mjs";
+import { At as manualInputDescriptorSchema, I as configureEvalRunLogs, Mt as columnDefSchema, Ot as evalStatAggregateSchema, _ as createFsCacheStore, bt as runSummarySchema, et as createRunRequestSchema, jt as evalChartsConfigSchema, kt as evalStatsConfigSchema, l as registerAgentEvalsPackageResolutionHooks, p as loadConfig, v as getCacheRetentionOptions, wt as buildEvalKey, yt as runManifestSchema } from "./runExecution-Bu9yfdUS.mjs";
+import { C as parseEvalDiscovery, h as persistRunState, r as getTargetEvals$1, t as executeRun } from "./runOrchestration-mpgZmEZ6.mjs";
 import { z } from "zod";
 import { readFile } from "node:fs/promises";
 import { relative } from "node:path";

package/dist/{runExecution-Bq0Y3y_1.mjs → runExecution-Bu9yfdUS.mjs}

@@ -7239,4 +7239,4 @@ function recordAssertionFailure(scope, failure) {
 	});
 }
 //#endregion
-export { setScopeCacheContext as $, repoFile as A, evalStatsConfigSchema as At, evalTime as B, evalTracer as C, resolveLlmCallsConfig as Ct, deserializeCacheValue as D, caseDetailSchema as Dt, deserializeCacheRecording as E, getCaseRowCaseKey as Et, EvalRuntimeUsageError as F, getEvalRegistry as Ft, matchesEvalTags as G, getEvalCaseInput as H, appendToEvalOutput as I, runWithEvalRegistry as It, runInEvalRuntimeScope as J, mergeEvalOutput as K, configureEvalRunLogs as L, readManualInputFile as M, evalChartsConfigSchema as Mt, evalExpect as N, columnDefSchema as Nt, serializeCacheRecording as O, caseRowSchema as Ot, EvalAssertionError as P, defineEval as Pt, setEvalOutput as Q, evalAssert as R, evalSpan as S, resolveApiCallsConfig as St, hashCacheKeySync as T, buildEvalKey as Tt, incrementEvalOutput as U, getCurrentScope as V, isInEvalScope as W, runInExistingEvalScope as X, runInEvalScope as Y, runWithEvalClock as Z, createBufferedCacheStore as _, matchesTagsFilter as _t, isCaseChildParentMessage as a, extractApiCalls as at, buildTraceTree as b, runManifestSchema as bt, resolveArtifactPath as c, simulateTokenAllocation as ct, loadEvalModule as d, getEvalTitle as dt, startEvalBackgroundJob as et, resolveEvalDefaultConfig as f, getEvalDisplayStatus as ft, commitPendingCacheWrites as g, dedupeEvalTags as gt, normalizeScoreDef as h, deriveStatusFromChildStatuses as ht, isCaseChildMessage as i, extractCacheHits as it, manualInputFileValueSchema as j, manualInputDescriptorSchema as jt, serializeCacheValue as k, evalStatAggregateSchema as kt, registerAgentEvalsPackageResolutionHooks as l, applyDerivedCallAttributes as lt, buildDeclaredColumnDefs as m, deriveStatusFromCaseRows as mt, resolveRunnableEvalCases as n, updateManualScoreRequestSchema as nt, stripTerminalControlCodes as o, extractLlmCalls as ot, loadConfig as p, deriveScopedSummaryFromCases as pt, nextEvalId as q, runCase as r, extractCacheEntries as rt, resolveTracePresentation as s, simulateLlmCallCost as st, filterEvalCases as t, createRunRequestSchema as tt, runWithModuleIsolation as u, getNestedAttribute as ut, createFsCacheStore as v, validateEvalTagName as vt, hashCacheKey as w, buildCaseKey as wt, captureEvalSpanError as x, runSummarySchema as xt, getCacheRetentionOptions as y, validateTagsFilterExpression as yt, evalLog as z };
+export { startEvalBackgroundJob as $, manualInputFileValueSchema as A, manualInputDescriptorSchema as At, getCurrentScope as B, hashCacheKey as C, buildCaseKey as Ct, serializeCacheRecording as D, caseRowSchema as Dt, deserializeCacheValue as E, caseDetailSchema as Et, appendToEvalOutput as F, runWithEvalRegistry as Ft, mergeEvalOutput as G, incrementEvalOutput as H, configureEvalRunLogs as I, runInEvalScope as J, nextEvalId as K, evalAssert as L, evalExpect as M, columnDefSchema as Mt, EvalAssertionError as N, defineEval as Nt, serializeCacheValue as O, evalStatAggregateSchema as Ot, EvalRuntimeUsageError as P, getEvalRegistry as Pt, setScopeCacheContext as Q, evalLog as R, evalTracer as S, resolveLlmCallsConfig as St, deserializeCacheRecording as T, getCaseRowCaseKey as Tt, isInEvalScope as U, getEvalCaseInput as V, matchesEvalTags as W, runWithEvalClock as X, runInExistingEvalScope as Y, setEvalOutput as Z, createFsCacheStore as _, validateEvalTagName as _t, isCaseChildParentMessage as a, extractLlmCalls as at, captureEvalSpanError as b, runSummarySchema as bt, resolveArtifactPath as c, applyDerivedCallAttributes as ct, loadEvalModule as d, getEvalDisplayStatus as dt, createRunRequestSchema as et, resolveEvalDefaultConfig as f, deriveScopedSummaryFromCases as ft, createBufferedCacheStore as g, matchesTagsFilter as gt, commitPendingCacheWrites as h, dedupeEvalTags as ht, isCaseChildMessage as i, extractApiCalls as it, readManualInputFile as j, evalChartsConfigSchema as jt, repoFile as k, evalStatsConfigSchema as kt, registerAgentEvalsPackageResolutionHooks as l, getNestedAttribute as lt, buildDeclaredColumnDefs as m, deriveStatusFromChildStatuses as mt, resolveRunnableEvalCases as n, extractCacheEntries as nt, stripTerminalControlCodes as o, simulateLlmCallCost as ot, loadConfig as p, deriveStatusFromCaseRows as pt, runInEvalRuntimeScope as q, runCase as r, extractCacheHits as rt, resolveTracePresentation as s, simulateTokenAllocation as st, filterEvalCases as t, updateManualScoreRequestSchema as tt, runWithModuleIsolation as u, getEvalTitle as ut, getCacheRetentionOptions as v, validateTagsFilterExpression as vt, hashCacheKeySync as w, buildEvalKey as wt, evalSpan as x, resolveApiCallsConfig as xt, buildTraceTree as y, runManifestSchema as yt, evalTime as z };

package/dist/{runOrchestration-C7qQISz2.mjs → runOrchestration-mpgZmEZ6.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-Bq0Y3y_1.mjs";
+import { Ct as buildCaseKey, Dt as caseRowSchema, Et as caseDetailSchema, Ft as runWithEvalRegistry, Tt as getCaseRowCaseKey, X as runWithEvalClock, _t as validateEvalTagName, bt as runSummarySchema, d as loadEvalModule, f as resolveEvalDefaultConfig, ft as deriveScopedSummaryFromCases, gt as matchesTagsFilter, h as commitPendingCacheWrites, ht as dedupeEvalTags, i as isCaseChildMessage, m as buildDeclaredColumnDefs, mt as deriveStatusFromChildStatuses, n as resolveRunnableEvalCases, o as stripTerminalControlCodes, pt as deriveStatusFromCaseRows, q as runInEvalRuntimeScope, t as filterEvalCases, u as runWithModuleIsolation, vt as validateTagsFilterExpression, yt as runManifestSchema } from "./runExecution-Bu9yfdUS.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-C9xNJHt3.mjs → runner-C4Y0lWb1.mjs}

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

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

@@ -1,5 +1,5 @@
-import { n as createRunner } from "./cli-BSVUCUxr.mjs";
-import "./src-D5vGo2iv.mjs";
+import { n as createRunner } from "./cli-OLZIjQpx.mjs";
+import "./src-Cy3OxoZW.mjs";
 //#region ../../apps/server/src/runner.ts
 let runnerInstance = null;
 function getRunnerInstance({ loadEnv = true } = {}) {

package/dist/{src-D5vGo2iv.mjs → src-Cy3OxoZW.mjs}

@@ -1,5 +1,5 @@
-import { G as matchesEvalTags$1, Pt as defineEval$1 } from "./runExecution-Bq0Y3y_1.mjs";
-import "./cli-BSVUCUxr.mjs";
+import { Nt as defineEval$1, W as matchesEvalTags$1 } from "./runExecution-Bu9yfdUS.mjs";
+import "./cli-OLZIjQpx.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": "0.60.3",
+  "version": "0.60.4",
   "type": "module",
   "bin": {
     "agent-evals": "./dist/bin.mjs"
@@ -32,9 +32,9 @@
   "devDependencies": {
     "@types/node": "^24.7.2",
     "typescript": "^5.9.2",
-    "@agent-evals/runner": "0.0.1",
     "@agent-evals/shared": "0.0.1",
-    "@agent-evals/sdk": "0.0.1"
+    "@agent-evals/sdk": "0.0.1",
+    "@agent-evals/runner": "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",