@ls-stack/agent-eval 0.55.1 → 0.56.0

package/dist/{cli-BR3wMZMx.mjs → cli-DQO2Fpt2.mjs}

@@ -1,5 +1,5 @@
-import { Ct as buildEvalKey, Nt as getEvalRegistry, St as resolveLlmCallsConfig, bt as runSummarySchema, c as resolveArtifactPath, ct as applyDerivedCallAttributes, dt as getEvalDisplayStatus, f as resolveEvalDefaultConfig, ft as deriveScopedSummaryFromCases, gt as matchesTagsFilter, h as normalizeScoreDef, m as buildDeclaredColumnDefs, o as stripTerminalControlCodes, p as loadConfig, s as resolveTracePresentation, ut as getEvalTitle, v as createFsCacheStore, wt as getCaseRowCaseKey, xt as resolveApiCallsConfig } from "./runExecution-Sw38bCaq.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-DJsdLYeZ.mjs";
+import { Ct as buildEvalKey, Nt as getEvalRegistry, St as resolveLlmCallsConfig, bt as runSummarySchema, c as resolveArtifactPath, ct as applyDerivedCallAttributes, dt as getEvalDisplayStatus, f as resolveEvalDefaultConfig, ft as deriveScopedSummaryFromCases, gt as matchesTagsFilter, h as normalizeScoreDef, m as buildDeclaredColumnDefs, o as stripTerminalControlCodes, p as loadConfig, s as resolveTracePresentation, ut as getEvalTitle, v as createFsCacheStore, wt as getCaseRowCaseKey, xt as resolveApiCallsConfig } from "./runExecution-6lrtj48K.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-BYaN2mzS.mjs";
 import { copyFile, mkdir, readFile, rm, writeFile } from "node:fs/promises";
 import { basename, dirname, extname, isAbsolute, join, relative, resolve, sep } from "node:path";
 import { createHash, randomUUID } from "node:crypto";
@@ -794,6 +794,7 @@ async function markRunErrored(runState, message, managerContext) {
 		payload: { message }
 	});
 	managerContext.emitDiscoveryEvent();
+	managerContext.onRunTerminal();
 }
 async function markRunTerminalFromChild(runState, event, managerContext) {
 	const snapshot = await loadPersistedRunSnapshot(runState.runDir);
@@ -815,6 +816,7 @@ async function markRunTerminalFromChild(runState, event, managerContext) {
 	}
 	managerContext.emitEvent(runState, event);
 	managerContext.emitDiscoveryEvent();
+	managerContext.onRunTerminal();
 }
 //#endregion
 //#region ../runner/src/runnerStateHydration.ts
@@ -883,6 +885,7 @@ function getWatchRootsForIncludePatterns(params) {
 }
 //#endregion
 //#region ../runner/src/runner.ts
+const defaultCachePruneIdleDelayMs = 5e3;
 /** Create an in-memory eval runner bound to the current workspace config. */
 function createRunner({ watchForChanges = true } = {}) {
 	let config;
@@ -903,6 +906,7 @@ function createRunner({ watchForChanges = true } = {}) {
 	let runHistoryWatcher;
 	let discoveryRefreshTimer;
 	let runHistoryRefreshTimer;
+	let cachePruneIdleTimer;
 	let registryLoadCounter = 0;
 	const configReload = createConfigReloadController({
 		getActiveRunCount,
@@ -942,6 +946,9 @@ function createRunner({ watchForChanges = true } = {}) {
 		async clearCache(filter) {
 			await cacheStore.clear(filter);
 		},
+		async repairCache() {
+			return cacheStore.repair();
+		},
 		async recomputeStatusesForEval(evalKey) {
 			const evalMeta = resolveEvalMeta(evalKey);
 			if (!evalMeta) return { updatedRuns: 0 };
@@ -1252,6 +1259,7 @@ function createRunner({ watchForChanges = true } = {}) {
 			emitDiscoveryEvent();
 		},
 		async startRun(request) {
+			cancelCacheRetentionPrune();
 			const tagsFilterError = validateTagsFilters(request.target.tagsFilter);
 			if (tagsFilterError !== null) throw new Error(tagsFilterError);
 			const deletedTemporaryRuns = await deleteTemporaryRuns({
@@ -1349,7 +1357,8 @@ function createRunner({ watchForChanges = true } = {}) {
 					workspaceRoot,
 					evals,
 					emitEvent,
-					emitDiscoveryEvent
+					emitDiscoveryEvent,
+					onRunTerminal: scheduleCacheRetentionPrune
 				}
 			});
 			if (deletedTemporaryRuns > 0) emitDiscoveryEvent();
@@ -1398,6 +1407,7 @@ function createRunner({ watchForChanges = true } = {}) {
 				payload: run.summary
 			});
 			emitDiscoveryEvent();
+			scheduleCacheRetentionPrune();
 		},
 		getCaseDetail(runId, caseId) {
 			const run = runs.get(runId);
@@ -1419,6 +1429,7 @@ function createRunner({ watchForChanges = true } = {}) {
 			};
 		},
 		async close() {
+			cancelCacheRetentionPrune();
 			await Promise.all([closeWatchers(), configReload.close()]);
 		},
 		getWorkspaceRoot() {
@@ -1522,6 +1533,28 @@ function createRunner({ watchForChanges = true } = {}) {
 	function getActiveRunCount() {
 		return [...runs.values()].filter((run) => run.manifest.status === "running").length;
 	}
+	function getCachePruneIdleDelayMs() {
+		const configured = config.cache?.pruneIdleDelayMs;
+		if (configured === void 0 || !Number.isFinite(configured) || configured <= 0) return defaultCachePruneIdleDelayMs;
+		return Math.floor(configured);
+	}
+	function cancelCacheRetentionPrune() {
+		if (cachePruneIdleTimer === void 0) return;
+		clearTimeout(cachePruneIdleTimer);
+		cachePruneIdleTimer = void 0;
+	}
+	function scheduleCacheRetentionPrune() {
+		cancelCacheRetentionPrune();
+		cachePruneIdleTimer = setTimeout(() => {
+			cachePruneIdleTimer = void 0;
+			pruneCacheRetentionIfIdle();
+		}, getCachePruneIdleDelayMs());
+		cachePruneIdleTimer.unref();
+	}
+	async function pruneCacheRetentionIfIdle() {
+		if (getActiveRunCount() > 0) return;
+		await cacheStore.pruneRetention();
+	}
 	function emitDiscoveryEvent() {
 		const lastRunStatuses = getLastRunStatuses({
 			runs: runs.values(),
@@ -1657,19 +1690,20 @@ Flags:
   `);
 		return;
 	}
-	if (topic === "cache" || topic === "cache list" || topic === "cache clear") {
+	if (topic === "cache" || topic === "cache list" || topic === "cache clear" || topic === "cache repair") {
 		console.info(`
-agent-evals cache - Manage cached operation entries
+agent-evals cache - Manage cached namespace/key entries
 
 Usage:
   agent-evals cache list [flags]
   agent-evals cache clear --eval <id>
   agent-evals cache clear --all
+  agent-evals cache repair [flags]
 
 Flags:
   --eval <id>                Clear entries for specific eval(s) (comma-separated)
   --all                      Confirm clearing every cached entry
-  --json                     Output cache listing as JSON
+  --json                     Output cache listing or repair summary as JSON
   --no-env                   Disable automatic .env loading
   --help, -h                 Show this help
   `);
@@ -1683,9 +1717,10 @@ Commands:
   list                       List discovered evals
   run                        Run evals
   show-runs [id|latest]      Show saved run artifact file paths
-  cache list                 List cached operation entries
+  cache list                 List cached namespace/key entries
   cache clear --eval <id>    Clear cache entries for one eval
   cache clear --all          Clear every cached entry
+  cache repair               Remove unindexed/orphaned cache files
   help                       Show this help
 
 Options:
@@ -1961,7 +1996,7 @@ function parseArgs(argv) {
 	let cursor = 1;
 	if (args.command === "cache") {
 		const sub = normalizedArgv[cursor];
-		if (sub === "list" || sub === "clear") {
+		if (sub === "list" || sub === "clear" || sub === "repair") {
 			args.subcommand = sub;
 			args.helpTopic = `cache ${sub}`;
 			cursor++;
@@ -2136,8 +2171,8 @@ 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-BD0D9-7k.mjs");
-	const runnerModule = await import("./runner-dB69WsnM.mjs");
+	const appModule = await import("./app-Bpe6Monh.mjs");
+	const runnerModule = await import("./runner-C3CiS2o7.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();
@@ -2311,10 +2346,8 @@ async function commandCache(args) {
 		for (const entry of entries) {
 			console.info(`  ${entry.namespace}`);
 			console.info(`    key: ${entry.key}`);
-			const operationLabel = entry.operationType === "span" ? `${entry.operationName} (span ${entry.spanKind ?? "unknown"})` : `${entry.operationName} (value)`;
-			console.info(`    operation: ${operationLabel}`);
 			console.info(`    stored: ${entry.storedAt}`);
-			console.info(`    size: ${String(entry.sizeBytes)} bytes`);
+			console.info(`    last accessed: ${entry.lastAccessedAt}`);
 			console.info("");
 		}
 		return;
@@ -2342,6 +2375,20 @@ async function commandCache(args) {
 		console.info("Refusing to clear cache without --eval <id> or --all. Use one of these flags to confirm.");
 		process.exit(1);
 	}
+	if (args.subcommand === "repair") {
+		const summary = await runner.repairCache();
+		if (args.json) {
+			console.info(JSON.stringify(summary, null, 2));
+			return;
+		}
+		console.info("Cache repair complete.");
+		console.info(`Removed cache files: ${String(summary.removedCacheFiles)}`);
+		console.info(`Removed debug files: ${String(summary.removedDebugFiles)}`);
+		console.info(`Removed blob files: ${String(summary.removedBlobFiles)}`);
+		console.info(`Removed index rows: ${String(summary.removedIndexRows)}`);
+		console.info(`Rewritten indexes: ${String(summary.rewrittenIndexes)}`);
+		return;
+	}
 	printHelp(args.helpTopic);
 }
 function getSortedRunSnapshots(runner) {