@ls-stack/agent-eval 0.40.0 → 0.41.0

package/dist/index.mjs

@@ -1,4 +1,4 @@
-import { $ as apiCallMetricFormatSchema, $n as evalLog, $t as evalChartTypeSchema, A as configReloadStateSchema, An as jsonCellSchema, At as evalStatsConfigSchema, B as simulateTokenAllocation, Bn as hashCacheKeySync, Bt as manualInputJsonFieldSchema, Cn as traceSpanSchema, Ct as assertionFailureSchema, Dn as columnFormatSchema, Dt as evalFreshnessStatusSchema, En as columnDefSchema, Et as discoveryIssueSchema, F as extractCacheEntries, Fn as buildTraceTree, Ft as runLogPhaseSchema, G as deriveScopedSummaryFromCases, Gn as repoFile, Gt as manualInputTextFieldSchema, H as getNestedAttribute, Hn as deserializeCacheValue, Ht as manualInputNumberFieldSchema, I as extractCacheHits, In as captureEvalSpanError, It as scoreTraceSchema, J as runManifestSchema, Jn as evalExpect, Jt as evalChartBuiltinMetricSchema, K as deriveStatusFromCaseRows, Kn as manualInputFileValueSchema, Kt as evalChartAggregateSchema, L as extractApiCalls, Ln as evalSpan, Lt as manualInputBooleanFieldSchema, M as createRunRequestSchema, Mn as repoFileRefSchema, Mt as runLogEntrySchema, N as updateManualScoreRequestSchema, Nn as runArtifactRefSchema, Nt as runLogLevelSchema, On as columnKindSchema, Ot as evalStatAggregateSchema, P as sseEnvelopeSchema, Pn as z, Pt as runLogLocationSchema, Q as agentEvalsConfigSchema, Qn as evalAssert, Qt as evalChartTooltipExtraSchema, R as extractLlmCalls, Rn as evalTracer, Rt as manualInputDescriptorSchema, Sn as traceSpanKindSchema, St as getCaseRowEvalKey, Tn as cellValueSchema, Tt as caseRowSchema, U as getEvalTitle, Un as serializeCacheRecording, Ut as manualInputSelectFieldSchema, V as applyDerivedCallAttributes, Vn as deserializeCacheRecording, Vt as manualInputMultilineFieldSchema, W as getEvalDisplayStatus, Wn as serializeCacheValue, Wt as manualInputSelectOptionSchema, X as DEFAULT_API_CALLS_CONFIG, Xn as appendToEvalOutput, Xt as evalChartConfigSchema, Y as runSummarySchema, Yn as EvalAssertionError, Yt as evalChartColorSchema, Z as DEFAULT_LLM_CALLS_CONFIG, Zt as evalChartMetricSchema, _n as traceAttributeDisplayPlacementSchema, _t as runLogsConfigSchema, an as cacheFileSchema, ar as mergeEvalOutput, at as evalColumnsSchema, bn as traceDisplayInputConfigSchema, bt as buildEvalKey, cn as cacheOperationTypeSchema, cr as runInEvalScope, ct as llmCallMetricFormatSchema, dn as cacheStatusSchema, dr as setScopeCacheContext, dt as llmCallPricingRateSchema, en as evalChartsConfigSchema, er as evalTime, et as apiCallMetricPlacementSchema, fn as serializedCacheSpanSchema, fr as startEvalBackgroundJob, ft as llmCallPricingSchema, gn as traceAttributeDisplayInputSchema, gt as resolveLlmCallsConfig, hn as traceAttributeDisplayFormatSchema, ht as resolveApiCallsConfig, in as cacheEntryWithDebugKeySchema, ir as isInEvalScope, it as evalColumnOverrideSchema, j as configReloadStatusSchema, jn as numberDisplayOptionsSchema, jt as evalSummarySchema, kn as fileRefSchema, kt as evalStatItemSchema, ln as cacheRecordingOpSchema, lr as runInExistingEvalScope, lt as llmCallMetricPlacementSchema, mn as traceCacheRefSchema, mr as getEvalRegistry, mt as removeDefaultConfigSchema, nn as cacheDebugKeyFileSchema, nr as getEvalCaseInput, nt as apiCallsConfigSchema, on as cacheListItemSchema, or as nextEvalId, ot as evalDeriveConfigSchema, pn as spanCacheOptionsSchema, pr as defineEval, pt as llmCallsConfigSchema, q as deriveStatusFromChildStatuses, qn as readManualInputFile, qt as evalChartAxisSchema, rn as cacheEntrySchema, rr as incrementEvalOutput, rt as defaultConfigKeySchema, sn as cacheModeSchema, sr as runInEvalRuntimeScope, st as llmCallCostCurrencySchema, tn as cacheDebugKeyEntrySchema, tr as getCurrentScope, tt as apiCallMetricSchema, un as cacheRecordingSchema, ur as setEvalOutput, ut as llmCallMetricSchema, vn as traceAttributeDisplaySchema, vt as trialSelectionModeSchema, wn as traceSpanWarningSchema, wt as caseDetailSchema, xn as traceSpanErrorSchema, xt as getCaseRowCaseKey, yn as traceDisplayConfigSchema, yt as buildCaseKey, z as simulateLlmCallCost, zn as hashCacheKey, zt as manualInputFieldDescriptorSchema } from "./runOrchestration-CTzVNrDP.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-Sg8UdOnm.mjs";
-import "./src-BZzPFS8r.mjs";
-export { DEFAULT_API_CALLS_CONFIG, DEFAULT_LLM_CALLS_CONFIG, EvalAssertionError, agentEvalsConfigSchema, apiCallMetricFormatSchema, apiCallMetricPlacementSchema, apiCallMetricSchema, apiCallsConfigSchema, appendToEvalOutput, applyDerivedCallAttributes, assertionFailureSchema, buildCaseKey, buildEvalKey, buildTraceTree, cacheDebugKeyEntrySchema, cacheDebugKeyFileSchema, cacheEntrySchema, cacheEntryWithDebugKeySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheOperationTypeSchema, cacheRecordingOpSchema, cacheRecordingSchema, cacheStatusSchema, captureEvalSpanError, caseDetailSchema, caseRowSchema, cellValueSchema, cleanupStagedManualInputFiles, columnDefSchema, columnFormatSchema, columnKindSchema, configReloadStateSchema, configReloadStatusSchema, createRunRequestSchema, createRunner, defaultConfigKeySchema, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, deserializeCacheRecording, deserializeCacheValue, discoveryIssueSchema, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalColumnOverrideSchema, evalColumnsSchema, evalDeriveConfigSchema, evalExpect, evalFreshnessStatusSchema, evalLog, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTime, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, fileRefSchema, getCaseRowCaseKey, getCaseRowEvalKey, getCurrentScope, getEvalCaseInput, getEvalDisplayStatus, getEvalRegistry, getEvalTitle, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, isManualInputFileValue, jsonCellSchema, llmCallCostCurrencySchema, llmCallMetricFormatSchema, llmCallMetricPlacementSchema, llmCallMetricSchema, llmCallPricingRateSchema, llmCallPricingSchema, llmCallsConfigSchema, manualInputBooleanFieldSchema, manualInputDescriptorSchema, manualInputFieldDescriptorSchema, manualInputFileValueSchema, manualInputJsonFieldSchema, manualInputMultilineFieldSchema, manualInputNumberFieldSchema, manualInputSelectFieldSchema, manualInputSelectOptionSchema, manualInputTextFieldSchema, materializeManualInputFiles, mergeEvalOutput, nextEvalId, numberDisplayOptionsSchema, readManualInputFile, removeDefaultConfigSchema, repoFile, repoFileRefSchema, resolveApiCallsConfig, resolveLlmCallsConfig, runArtifactRefSchema, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, runLogEntrySchema, runLogLevelSchema, runLogLocationSchema, runLogPhaseSchema, runLogsConfigSchema, runManifestSchema, runSummarySchema, scoreTraceSchema, serializeCacheRecording, serializeCacheValue, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, simulateLlmCallCost, simulateTokenAllocation, spanCacheOptionsSchema, sseEnvelopeSchema, stageManualInputFile, stageManualInputFileFromPath, startEvalBackgroundJob, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceCacheRefSchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanErrorSchema, traceSpanKindSchema, traceSpanSchema, traceSpanWarningSchema, trialSelectionModeSchema, updateManualScoreRequestSchema, z };
+import { $ as getCurrentScope, B as deserializeCacheValue, F as evalSpan, G as readManualInputFile, H as serializeCacheValue, I as evalTracer, J as appendToEvalOutput, K as evalExpect, L as hashCacheKey, M as z, N as buildTraceTree, P as captureEvalSpanError, Q as evalTime, R as hashCacheKeySync, Rt as getEvalRegistry, U as repoFile, V as serializeCacheRecording, W as manualInputFileValueSchema, X as evalAssert, Z as evalLog, _t as extractLlmCalls, at as nextEvalId, ct as runInExistingEvalScope, dt as startEvalBackgroundJob, et as getEvalCaseInput, gt as extractApiCalls, ht as extractCacheHits, it as mergeEvalOutput, lt as setEvalOutput, mt as extractCacheEntries, nt as isInEvalScope, ot as runInEvalRuntimeScope, q as EvalAssertionError, st as runInEvalScope, tt as incrementEvalOutput, ut as setScopeCacheContext, vt as simulateLlmCallCost, xt as getNestedAttribute, yt as simulateTokenAllocation, z as deserializeCacheRecording } from "./runOrchestration-DhTiT4V0.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-Bk5g-bat.mjs";
+import { n as matchesEvalTags, t as defineEval } from "./src-t6OVp1li.mjs";
+export { EvalAssertionError, 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, z };

package/dist/runChild.mjs

@@ -1,8 +1,8 @@
-import { At as evalStatsConfigSchema, C as parseEvalDiscovery, En as columnDefSchema, J as runManifestSchema, M as createRunRequestSchema, Rt as manualInputDescriptorSchema, T as loadConfig, Y as runSummarySchema, Zn as configureEvalRunLogs, bt as buildEvalKey, en as evalChartsConfigSchema, k as createFsCacheStore, p as persistRunState, r as getTargetEvals$1, t as executeRun } from "./runOrchestration-CTzVNrDP.mjs";
-import { createHash } from "node:crypto";
+import { At as buildEvalKey, D as loadConfig, Dt as runSummarySchema, Et as runManifestSchema, Ft as evalChartsConfigSchema, It as columnDefSchema, Nt as evalStatsConfigSchema, Pt as manualInputDescriptorSchema, T as parseEvalDiscovery, Y as configureEvalRunLogs, ft as createRunRequestSchema, h as persistRunState, j as createFsCacheStore, r as getTargetEvals$1, t as executeRun } from "./runOrchestration-DhTiT4V0.mjs";
+import { z } from "zod/v4";
 import { readFile } from "node:fs/promises";
 import { relative } from "node:path";
-import { z } from "zod/v4";
+import { createHash } from "node:crypto";
 import { glob } from "glob";
 //#region ../runner/src/runChild.ts
 const evalMetaSchema = z.object({
@@ -10,6 +10,7 @@ const evalMetaSchema = z.object({
 	id: z.string(),
 	title: z.string().optional(),
 	filePath: z.string(),
+	tags: z.array(z.string()).default([]),
 	sourceFilePath: z.string(),
 	sourceFingerprint: z.string().nullable(),
 	columnDefs: z.array(columnDefSchema),
@@ -109,6 +110,7 @@ async function discoverRunEvals(params) {
 				id: meta.id,
 				title: meta.title,
 				filePath: relativeFilePath,
+				tags: [],
 				sourceFilePath: meta.filePath,
 				sourceFingerprint,
 				columnDefs: [],