@ls-stack/agent-eval 0.60.1 → 0.60.3

package/dist/{app-CxKEVlng.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-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-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-CVBSlTD8.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/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-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-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 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-CxKEVlng.mjs");
-	const runnerModule = await import("./runner-Cu1CQPTB.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 });