@ls-stack/agent-eval 0.58.0 → 0.58.1

package/dist/{app-L9GdY28I.mjs → app-DhMIbjlE.mjs}

@@ -1,7 +1,7 @@
-import { et as createRunRequestSchema, nt as extractCacheEntries, tt as updateManualScoreRequestSchema, ut as getEvalTitle } from "./runExecution-C4kAOhC1.mjs";
-import { o as stageManualInputFile } from "./cli-Cf37PZKi.mjs";
-import "./src-303BocMW.mjs";
-import { t as getRunnerInstance } from "./runner-bjd_UB9i.mjs";
+import { et as createRunRequestSchema, nt as extractCacheEntries, tt as updateManualScoreRequestSchema, ut as getEvalTitle } from "./runExecution-d42Lm0i5.mjs";
+import { o as stageManualInputFile } from "./cli-_g2qOMK6.mjs";
+import "./src-CdZsOn6y.mjs";
+import { t as getRunnerInstance } from "./runner-MSr8sAWm.mjs";
 import { z } from "zod/v4";
 import { readFile } from "node:fs/promises";
 import { dirname, isAbsolute, join, relative, resolve, sep } from "node:path";

package/dist/bin.mjs

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { t as runCli } from "./cli-Cf37PZKi.mjs";
+import { t as runCli } from "./cli-_g2qOMK6.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 { I as configureEvalRunLogs, Pt as runWithEvalRegistry, St as resolveLlmCallsConfig, _ as createBufferedCacheStore, a as isCaseChildParentMessage, d as loadEvalModule, l as registerAgentEvalsPackageResolutionHooks, p as loadConfig, q as runInEvalRuntimeScope, r as runCase, v as createFsCacheStore, xt as resolveApiCallsConfig } from "./runExecution-C4kAOhC1.mjs";
+import { I as configureEvalRunLogs, Pt as runWithEvalRegistry, St as resolveLlmCallsConfig, _ as createBufferedCacheStore, a as isCaseChildParentMessage, d as loadEvalModule, l as registerAgentEvalsPackageResolutionHooks, p as loadConfig, q as runInEvalRuntimeScope, r as runCase, v as createFsCacheStore, xt as resolveApiCallsConfig } from "./runExecution-d42Lm0i5.mjs";
 //#region ../runner/src/caseChild.ts
 let fatalErrorReported = false;
 let disconnectExpected = false;

package/dist/{cli-Cf37PZKi.mjs → cli-_g2qOMK6.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-C4kAOhC1.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-5xEiQxiS.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-d42Lm0i5.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-CvmFeOmT.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";
@@ -2172,8 +2172,8 @@ 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-L9GdY28I.mjs");
-	const runnerModule = await import("./runner-JIykMlve.mjs");
+	const appModule = await import("./app-DhMIbjlE.mjs");
+	const runnerModule = await import("./runner-BKogjiYd.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();

package/dist/index.d.mts

@@ -1455,7 +1455,9 @@ type CacheScopeContext = {
 /** Active recording frame captured while a cached operation body executes. */
 type CacheRecordingFrame = {
   /** Length of `scope.spans` immediately before the cached body started. */baseSpanIndex: number; /** Parent id used when recording and replaying direct child spans. */
-  replayParentSpanId: string | null; /** Ordered observable effects recorded during the cached body. */
+  replayParentSpanId: string | null; /** Spans created by this cache body's async execution branch. */
+  spanIds: Set<string>; /** Non-cache attributes written to the replay parent by this async branch. */
+  finalAttributes: Record<string, unknown>; /** Ordered observable effects recorded during the cached body. */
   ops: CacheRecordingOp$1[];
 };
 /** Mutable per-case runtime state stored in async local storage. */
@@ -1480,11 +1482,6 @@ type EvalCaseScope = {
   logs: RunLogEntry$1[];
   spans: EvalTraceSpan$2[];
   checkpoints: Map<string, unknown>;
-  /**
-   * Stack of active cache recorders. Ops are written to the top-most frame
-   * when it exists and `replayingDepth === 0`.
-   */
-  recordingStack: CacheRecordingFrame[];
   /**
    * Incremented while replaying a cached operation, so nested SDK calls do not
    * accidentally double-record ops into outer recorders.
@@ -2017,8 +2014,8 @@ declare const traceAttributeDisplaySchema: z$1.ZodObject<{
     subtree: "subtree";
   }>>;
   mode: z$1.ZodOptional<z$1.ZodEnum<{
-    all: "all";
     sum: "sum";
+    all: "all";
     last: "last";
   }>>;
 }, z$1.core.$strip>;
@@ -2053,8 +2050,8 @@ declare const traceDisplayConfigSchema: z$1.ZodObject<{
       subtree: "subtree";
     }>>;
     mode: z$1.ZodOptional<z$1.ZodEnum<{
-      all: "all";
       sum: "sum";
+      all: "all";
       last: "last";
     }>>;
   }, z$1.core.$strip>>>;
@@ -2093,8 +2090,8 @@ declare const traceAttributeDisplayInputSchema: z$1.ZodObject<{
     subtree: "subtree";
   }>>;
   mode: z$1.ZodOptional<z$1.ZodEnum<{
-    all: "all";
     sum: "sum";
+    all: "all";
     last: "last";
   }>>;
   transform: z$1.ZodOptional<z$1.ZodCustom<TraceAttributeTransform, TraceAttributeTransform>>;
@@ -2131,8 +2128,8 @@ declare const traceDisplayInputConfigSchema: z$1.ZodObject<{
       subtree: "subtree";
     }>>;
     mode: z$1.ZodOptional<z$1.ZodEnum<{
-      all: "all";
       sum: "sum";
+      all: "all";
       last: "last";
     }>>;
     transform: z$1.ZodOptional<z$1.ZodCustom<TraceAttributeTransform, TraceAttributeTransform>>;
@@ -2217,9 +2214,9 @@ type EvalFreshnessStatus = z$1.infer<typeof evalFreshnessStatusSchema>;
  */
 declare const evalStatAggregateSchema: z$1.ZodEnum<{
   avg: "avg";
-  sum: "sum";
   min: "min";
   max: "max";
+  sum: "sum";
   best: "best";
   worst: "worst";
 }>;
@@ -2249,9 +2246,9 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   kind: z$1.ZodLiteral<"duration">;
   aggregate: z$1.ZodOptional<z$1.ZodEnum<{
     avg: "avg";
-    sum: "sum";
     min: "min";
     max: "max";
+    sum: "sum";
     best: "best";
     worst: "worst";
   }>>;
@@ -2260,9 +2257,9 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   kind: z$1.ZodLiteral<"cacheHits">;
   aggregate: z$1.ZodOptional<z$1.ZodEnum<{
     avg: "avg";
-    sum: "sum";
     min: "min";
     max: "max";
+    sum: "sum";
     best: "best";
     worst: "worst";
   }>>;
@@ -2273,9 +2270,9 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   label: z$1.ZodOptional<z$1.ZodString>;
   aggregate: z$1.ZodEnum<{
     avg: "avg";
-    sum: "sum";
     min: "min";
     max: "max";
+    sum: "sum";
     best: "best";
     worst: "worst";
   }>;
@@ -2313,9 +2310,9 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
   kind: z$1.ZodLiteral<"duration">;
   aggregate: z$1.ZodOptional<z$1.ZodEnum<{
     avg: "avg";
-    sum: "sum";
     min: "min";
     max: "max";
+    sum: "sum";
     best: "best";
     worst: "worst";
   }>>;
@@ -2324,9 +2321,9 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
   kind: z$1.ZodLiteral<"cacheHits">;
   aggregate: z$1.ZodOptional<z$1.ZodEnum<{
     avg: "avg";
-    sum: "sum";
     min: "min";
     max: "max";
+    sum: "sum";
     best: "best";
     worst: "worst";
   }>>;
@@ -2337,9 +2334,9 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
   label: z$1.ZodOptional<z$1.ZodString>;
   aggregate: z$1.ZodEnum<{
     avg: "avg";
-    sum: "sum";
     min: "min";
     max: "max";
+    sum: "sum";
     best: "best";
     worst: "worst";
   }>;
@@ -2422,10 +2419,10 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
   caseIds: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
   lastRunStatus: z$1.ZodNullable<z$1.ZodEnum<{
     error: "error";
-    running: "running";
-    cancelled: "cancelled";
     pass: "pass";
     fail: "fail";
+    running: "running";
+    cancelled: "cancelled";
     unscored: "unscored";
   }>>;
   stats: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
@@ -2440,9 +2437,9 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
     kind: z$1.ZodLiteral<"duration">;
     aggregate: z$1.ZodOptional<z$1.ZodEnum<{
       avg: "avg";
-      sum: "sum";
       min: "min";
       max: "max";
+      sum: "sum";
       best: "best";
       worst: "worst";
     }>>;
@@ -2451,9 +2448,9 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
     kind: z$1.ZodLiteral<"cacheHits">;
     aggregate: z$1.ZodOptional<z$1.ZodEnum<{
       avg: "avg";
-      sum: "sum";
       min: "min";
       max: "max";
+      sum: "sum";
       best: "best";
       worst: "worst";
     }>>;
@@ -2464,9 +2461,9 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
     label: z$1.ZodOptional<z$1.ZodString>;
     aggregate: z$1.ZodEnum<{
       avg: "avg";
-      sum: "sum";
       min: "min";
       max: "max";
+      sum: "sum";
       best: "best";
       worst: "worst";
     }>;
@@ -2491,9 +2488,9 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
   }, z$1.core.$strip>], "kind">>>;
   defaultStatAggregate: z$1.ZodOptional<z$1.ZodEnum<{
     avg: "avg";
-    sum: "sum";
     min: "min";
     max: "max";
+    sum: "sum";
     best: "best";
     worst: "worst";
   }>>;
@@ -2530,9 +2527,9 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
       key: z$1.ZodString;
       aggregate: z$1.ZodEnum<{
         avg: "avg";
-        sum: "sum";
         min: "min";
         max: "max";
+        sum: "sum";
         latest: "latest";
         passThresholdRate: "passThresholdRate";
       }>;
@@ -2572,9 +2569,9 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
       key: z$1.ZodString;
       aggregate: z$1.ZodEnum<{
         avg: "avg";
-        sum: "sum";
         min: "min";
         max: "max";
+        sum: "sum";
         latest: "latest";
         passThresholdRate: "passThresholdRate";
       }>;
@@ -2671,11 +2668,11 @@ declare const caseRowSchema$1: z$1.ZodObject<{
   tags: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
   status: z$1.ZodEnum<{
     error: "error";
-    pending: "pending";
-    running: "running";
-    cancelled: "cancelled";
     pass: "pass";
     fail: "fail";
+    running: "running";
+    cancelled: "cancelled";
+    pending: "pending";
   }>;
   durationMs: z$1.ZodNullable<z$1.ZodNumber>;
   cacheHits: z$1.ZodOptional<z$1.ZodNumber>;
@@ -2862,8 +2859,8 @@ declare const scoreTraceSchema: z$1.ZodObject<{
         subtree: "subtree";
       }>>;
       mode: z$1.ZodOptional<z$1.ZodEnum<{
-        all: "all";
         sum: "sum";
+        all: "all";
         last: "last";
       }>>;
     }, z$1.core.$strip>>>;
@@ -2874,10 +2871,10 @@ declare const scoreTraceSchema: z$1.ZodObject<{
     namespace: z$1.ZodString;
     key: z$1.ZodString;
     status: z$1.ZodEnum<{
-      bypass: "bypass";
-      refresh: "refresh";
       hit: "hit";
       miss: "miss";
+      refresh: "refresh";
+      bypass: "bypass";
     }>;
     read: z$1.ZodOptional<z$1.ZodBoolean>;
     stored: z$1.ZodOptional<z$1.ZodBoolean>;
@@ -2896,11 +2893,11 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
   tags: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
   status: z$1.ZodEnum<{
     error: "error";
-    pending: "pending";
-    running: "running";
-    cancelled: "cancelled";
     pass: "pass";
     fail: "fail";
+    running: "running";
+    cancelled: "cancelled";
+    pending: "pending";
   }>;
   input: z$1.ZodUnknown;
   trace: z$1.ZodArray<z$1.ZodObject<{
@@ -2965,8 +2962,8 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
         subtree: "subtree";
       }>>;
       mode: z$1.ZodOptional<z$1.ZodEnum<{
-        all: "all";
         sum: "sum";
+        all: "all";
         last: "last";
       }>>;
     }, z$1.core.$strip>>>;
@@ -3034,8 +3031,8 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
           subtree: "subtree";
         }>>;
         mode: z$1.ZodOptional<z$1.ZodEnum<{
-          all: "all";
           sum: "sum";
+          all: "all";
           last: "last";
         }>>;
       }, z$1.core.$strip>>>;
@@ -3046,10 +3043,10 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
       namespace: z$1.ZodString;
       key: z$1.ZodString;
       status: z$1.ZodEnum<{
-        bypass: "bypass";
-        refresh: "refresh";
         hit: "hit";
         miss: "miss";
+        refresh: "refresh";
+        bypass: "bypass";
       }>;
       read: z$1.ZodOptional<z$1.ZodBoolean>;
       stored: z$1.ZodOptional<z$1.ZodBoolean>;
@@ -3166,10 +3163,10 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
     namespace: z$1.ZodString;
     key: z$1.ZodString;
     status: z$1.ZodEnum<{
-      bypass: "bypass";
-      refresh: "refresh";
       hit: "hit";
       miss: "miss";
+      refresh: "refresh";
+      bypass: "bypass";
     }>;
     read: z$1.ZodOptional<z$1.ZodBoolean>;
     stored: z$1.ZodOptional<z$1.ZodBoolean>;
@@ -3223,9 +3220,9 @@ type EvalChartBuiltinMetric = z$1.infer<typeof evalChartBuiltinMetricSchema>;
 /** Reducer applied to a numeric column across all cases of a single run. */
 declare const evalChartAggregateSchema: z$1.ZodEnum<{
   avg: "avg";
-  sum: "sum";
   min: "min";
   max: "max";
+  sum: "sum";
   latest: "latest";
   passThresholdRate: "passThresholdRate";
 }>;
@@ -3281,9 +3278,9 @@ declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
   key: z$1.ZodString;
   aggregate: z$1.ZodEnum<{
     avg: "avg";
-    sum: "sum";
     min: "min";
     max: "max";
+    sum: "sum";
     latest: "latest";
     passThresholdRate: "passThresholdRate";
   }>;
@@ -3316,9 +3313,9 @@ declare const evalChartTooltipExtraSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObj
   key: z$1.ZodString;
   aggregate: z$1.ZodEnum<{
     avg: "avg";
-    sum: "sum";
     min: "min";
     max: "max";
+    sum: "sum";
     latest: "latest";
     passThresholdRate: "passThresholdRate";
   }>;
@@ -3364,9 +3361,9 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
       avg: "avg";
-      sum: "sum";
       min: "min";
       max: "max";
+      sum: "sum";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
@@ -3406,9 +3403,9 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
       avg: "avg";
-      sum: "sum";
       min: "min";
       max: "max";
+      sum: "sum";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
@@ -3454,9 +3451,9 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
       avg: "avg";
-      sum: "sum";
       min: "min";
       max: "max";
+      sum: "sum";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
@@ -3496,9 +3493,9 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
     key: z$1.ZodString;
     aggregate: z$1.ZodEnum<{
       avg: "avg";
-      sum: "sum";
       min: "min";
       max: "max";
+      sum: "sum";
       latest: "latest";
       passThresholdRate: "passThresholdRate";
     }>;
@@ -3514,10 +3511,10 @@ declare const runManifestSchema$1: z$1.ZodObject<{
   shortId: z$1.ZodString;
   status: z$1.ZodEnum<{
     error: "error";
-    pending: "pending";
     running: "running";
-    completed: "completed";
     cancelled: "cancelled";
+    pending: "pending";
+    completed: "completed";
   }>;
   temporary: z$1.ZodDefault<z$1.ZodOptional<z$1.ZodBoolean>>;
   startedAt: z$1.ZodString;
@@ -3526,9 +3523,9 @@ declare const runManifestSchema$1: z$1.ZodObject<{
   evalSourceFingerprints: z$1.ZodDefault<z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodString>>>;
   target: z$1.ZodObject<{
     mode: z$1.ZodEnum<{
+      caseIds: "caseIds";
       all: "all";
       evalIds: "evalIds";
-      caseIds: "caseIds";
     }>;
     evalKeys: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
     files: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
@@ -3542,9 +3539,9 @@ declare const runManifestSchema$1: z$1.ZodObject<{
     median: "median";
   }>>>;
   cacheMode: z$1.ZodOptional<z$1.ZodEnum<{
-    use: "use";
-    bypass: "bypass";
     refresh: "refresh";
+    bypass: "bypass";
+    use: "use";
   }>>;
 }, z$1.core.$strip>;
 /** Persisted lifecycle metadata for a single eval run. */
@@ -3554,10 +3551,10 @@ declare const runSummarySchema$1: z$1.ZodObject<{
   runId: z$1.ZodString;
   status: z$1.ZodEnum<{
     error: "error";
-    pending: "pending";
     running: "running";
-    completed: "completed";
     cancelled: "cancelled";
+    pending: "pending";
+    completed: "completed";
   }>;
   totalCases: z$1.ZodNumber;
   passedCases: z$1.ZodNumber;
@@ -4469,9 +4466,9 @@ declare function extractApiCalls(spans: EvalTraceSpan$1[], config: ResolvedApiCa
  * - `refresh`: never read, always write (forces re-execution and overwrites).
  */
 declare const cacheModeSchema: z$1.ZodEnum<{
-  use: "use";
-  bypass: "bypass";
   refresh: "refresh";
+  bypass: "bypass";
+  use: "use";
 }>;
 /** Mode controlling how cached spans behave during a run. */
 type CacheMode = z$1.infer<typeof cacheModeSchema>;
@@ -4492,10 +4489,10 @@ declare const cacheOperationTypeSchema: z$1.ZodEnum<{
 type CacheOperationType = z$1.infer<typeof cacheOperationTypeSchema>;
 /** Status of a cache lookup recorded on a span or case scope. */
 declare const cacheStatusSchema: z$1.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$1.infer<typeof cacheStatusSchema>;
@@ -4512,10 +4509,10 @@ declare const traceCacheRefSchema: z$1.ZodObject<{
   namespace: z$1.ZodString;
   key: z$1.ZodString;
   status: z$1.ZodEnum<{
-    bypass: "bypass";
-    refresh: "refresh";
     hit: "hit";
     miss: "miss";
+    refresh: "refresh";
+    bypass: "bypass";
   }>;
   read: z$1.ZodOptional<z$1.ZodBoolean>;
   stored: z$1.ZodOptional<z$1.ZodBoolean>;
@@ -5467,9 +5464,9 @@ type ConfigReloadState = z$1.infer<typeof configReloadStateSchema$1>;
 declare const createRunRequestSchema$1: z$1.ZodObject<{
   target: z$1.ZodObject<{
     mode: z$1.ZodEnum<{
+      caseIds: "caseIds";
       all: "all";
       evalIds: "evalIds";
-      caseIds: "caseIds";
     }>;
     evalKeys: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
     files: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
@@ -5481,9 +5478,9 @@ declare const createRunRequestSchema$1: z$1.ZodObject<{
   temporary: z$1.ZodOptional<z$1.ZodBoolean>;
   cache: z$1.ZodOptional<z$1.ZodObject<{
     mode: z$1.ZodDefault<z$1.ZodEnum<{
-      use: "use";
-      bypass: "bypass";
       refresh: "refresh";
+      bypass: "bypass";
+      use: "use";
     }>>;
   }, z$1.core.$strip>>;
   manualInputs: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;

package/dist/index.mjs

@@ -1,4 +1,4 @@
-import { $ as startEvalBackgroundJob, A as repoFile, B as getCurrentScope, C as evalTracer, D as deserializeCacheValue, E as deserializeCacheRecording, F as appendToEvalOutput, G as mergeEvalOutput, H as incrementEvalOutput, J as runInEvalScope, K as nextEvalId, L as evalAssert, M as readManualInputFile, N as evalExpect, Nt as getEvalRegistry, O as serializeCacheRecording, P as EvalAssertionError, Q as setScopeCacheContext, R as evalLog, S as evalSpan, T as hashCacheKeySync, U as isInEvalScope, V as getEvalCaseInput, Y as runInExistingEvalScope, Z as setEvalOutput, at as extractLlmCalls, b as buildTraceTree, it as extractApiCalls, j as manualInputFileValueSchema, k as serializeCacheValue, lt as getNestedAttribute, nt as extractCacheEntries, ot as simulateLlmCallCost, q as runInEvalRuntimeScope, rt as extractCacheHits, st as simulateTokenAllocation, w as hashCacheKey, x as captureEvalSpanError, y as z, z as evalTime } from "./runExecution-C4kAOhC1.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-Cf37PZKi.mjs";
-import { n as matchesEvalTags, t as defineEval } from "./src-303BocMW.mjs";
+import { $ as startEvalBackgroundJob, A as repoFile, B as getCurrentScope, C as evalTracer, D as deserializeCacheValue, E as deserializeCacheRecording, F as appendToEvalOutput, G as mergeEvalOutput, H as incrementEvalOutput, J as runInEvalScope, K as nextEvalId, L as evalAssert, M as readManualInputFile, N as evalExpect, Nt as getEvalRegistry, O as serializeCacheRecording, P as EvalAssertionError, Q as setScopeCacheContext, R as evalLog, S as evalSpan, T as hashCacheKeySync, U as isInEvalScope, V as getEvalCaseInput, Y as runInExistingEvalScope, Z as setEvalOutput, at as extractLlmCalls, b as buildTraceTree, it as extractApiCalls, j as manualInputFileValueSchema, k as serializeCacheValue, lt as getNestedAttribute, nt as extractCacheEntries, ot as simulateLlmCallCost, q as runInEvalRuntimeScope, rt as extractCacheHits, st as simulateTokenAllocation, w as hashCacheKey, x as captureEvalSpanError, y as z, z as evalTime } from "./runExecution-d42Lm0i5.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-_g2qOMK6.mjs";
+import { n as matchesEvalTags, t as defineEval } from "./src-CdZsOn6y.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

@@ -1,5 +1,5 @@
-import { At as evalChartsConfigSchema, Ct as buildEvalKey, Dt as evalStatAggregateSchema, I as configureEvalRunLogs, Ot as evalStatsConfigSchema, bt as runSummarySchema, et as createRunRequestSchema, jt as columnDefSchema, kt as manualInputDescriptorSchema, l as registerAgentEvalsPackageResolutionHooks, p as loadConfig, v as createFsCacheStore, yt as runManifestSchema } from "./runExecution-C4kAOhC1.mjs";
-import { S as parseEvalDiscovery, m as persistRunState, r as getTargetEvals$1, t as executeRun } from "./runOrchestration-5xEiQxiS.mjs";
+import { At as evalChartsConfigSchema, Ct as buildEvalKey, Dt as evalStatAggregateSchema, I as configureEvalRunLogs, Ot as evalStatsConfigSchema, bt as runSummarySchema, et as createRunRequestSchema, jt as columnDefSchema, kt as manualInputDescriptorSchema, l as registerAgentEvalsPackageResolutionHooks, p as loadConfig, v as createFsCacheStore, yt as runManifestSchema } from "./runExecution-d42Lm0i5.mjs";
+import { S as parseEvalDiscovery, m as persistRunState, r as getTargetEvals$1, t as executeRun } from "./runOrchestration-CvmFeOmT.mjs";
 import { z } from "zod/v4";
 import { readFile } from "node:fs/promises";
 import { relative } from "node:path";

package/dist/{runExecution-C4kAOhC1.mjs → runExecution-d42Lm0i5.mjs}

@@ -2651,6 +2651,7 @@ const scopeStorage = new AsyncLocalStorage();
 const runtimeScopeStorage = new AsyncLocalStorage();
 const evalClockStorage = new AsyncLocalStorage();
 const activeSpanStackStorage = new AsyncLocalStorage();
+const recordingStackStorage = new AsyncLocalStorage();
 let activeEvalScopeCount = 0;
 let activeEvalRuntimeScopeCount = 0;
 let consoleCaptureEnabled = true;
@@ -2803,6 +2804,20 @@ async function runWithActiveSpan(span, fn) {
 	const currentStack = activeSpanStackStorage.getStore() ?? [];
 	return await activeSpanStackStorage.run([...currentStack, span], fn);
 }
+/** Execute a callback with a cache recording frame scoped to this async branch. */
+async function runWithCacheRecordingFrame(frame, fn) {
+	const currentStack = recordingStackStorage.getStore() ?? [];
+	return await recordingStackStorage.run([...currentStack, frame], fn);
+}
+function getCurrentCacheRecordingFrame(scope) {
+	if (scope.replayingDepth > 0) return void 0;
+	return recordingStackStorage.getStore()?.at(-1);
+}
+/** Mark a span as created by the active cache recorder, when one exists. */
+function recordSpanForActiveCacheRecording(scope, spanId) {
+	if (scope.replayingDepth > 0) return;
+	for (const frame of recordingStackStorage.getStore() ?? []) frame.spanIds.add(spanId);
+}
 /**
 * Return the current eval runner phase for this async execution.
 *
@@ -3110,7 +3125,6 @@ async function runInEvalScope(caseId, fn, options = {}) {
 		logs: [],
 		spans: [],
 		checkpoints: /* @__PURE__ */ new Map(),
-		recordingStack: [],
 		replayingDepth: 0,
 		cacheContext: options.cacheContext,
 		caseCacheRefs: [],
@@ -3150,10 +3164,16 @@ function nextEvalId() {
 	scope.nextEvalIdCounter++;
 	return `${scope.idPrefix}-${scope.nextEvalIdCounter}`;
 }
-function recordOpIfActive(scope, op) {
-	if (scope.replayingDepth > 0) return;
-	const top = scope.recordingStack.at(-1);
-	if (top) top.ops.push(op);
+function recordCacheRecordingOpIfActive(scope, op) {
+	getCurrentCacheRecordingFrame(scope)?.ops.push(op);
+}
+function recordCacheRecordingAttributesIfActive(scope, span, attributes) {
+	const frames = recordingStackStorage.getStore();
+	if (scope.replayingDepth > 0 || frames === void 0) return;
+	for (const [key, value] of Object.entries(attributes)) {
+		if (key.startsWith("cache.")) continue;
+		for (const frame of frames) if (span.id === frame.replayParentSpanId) frame.finalAttributes[key] = value;
+	}
 }
 function normalizeEvalOutputOptions(options) {
 	if (options === void 0) return void 0;
@@ -3185,7 +3205,7 @@ function setEvalOutput(key, value, options = void 0) {
 	scope.outputs[key] = value;
 	const column = normalizeEvalOutputOptions(options);
 	if (column !== void 0) scope.outputColumnOverrides[key] = column;
-	recordOpIfActive(scope, {
+	recordCacheRecordingOpIfActive(scope, {
 		kind: "setOutput",
 		key,
 		value,
@@ -3205,7 +3225,7 @@ function appendToEvalOutput(key, value) {
 	if (existing === void 0) scope.outputs[key] = [value];
 	else if (Array.isArray(existing)) scope.outputs[key] = [...copyArray$1(existing), value];
 	else scope.outputs[key] = [existing, value];
-	recordOpIfActive(scope, {
+	recordCacheRecordingOpIfActive(scope, {
 		kind: "appendOutput",
 		key,
 		value
@@ -3223,7 +3243,7 @@ function mergeEvalOutput(key, patch) {
 	const existing = scope.outputs[key];
 	if (existing === void 0) {
 		scope.outputs[key] = { ...patch };
-		recordOpIfActive(scope, {
+		recordCacheRecordingOpIfActive(scope, {
 			kind: "mergeOutput",
 			key,
 			patch
@@ -3238,7 +3258,7 @@ function mergeEvalOutput(key, patch) {
 		...existing,
 		...patch
 	};
-	recordOpIfActive(scope, {
+	recordCacheRecordingOpIfActive(scope, {
 		kind: "mergeOutput",
 		key,
 		patch
@@ -3256,7 +3276,7 @@ function incrementEvalOutput(key, delta) {
 	const existing = scope.outputs[key];
 	if (existing === void 0) {
 		scope.outputs[key] = delta;
-		recordOpIfActive(scope, {
+		recordCacheRecordingOpIfActive(scope, {
 			kind: "incrementOutput",
 			key,
 			delta
@@ -3268,7 +3288,7 @@ function incrementEvalOutput(key, delta) {
 		return;
 	}
 	scope.outputs[key] = existing + delta;
-	recordOpIfActive(scope, {
+	recordCacheRecordingOpIfActive(scope, {
 		kind: "incrementOutput",
 		key,
 		delta
@@ -3675,10 +3695,6 @@ async function materializeExternalJsonValues(value, store) {
 	if (!isRecordLike$3(value)) return value;
 	return Object.fromEntries(await Promise.all(Object.entries(value).map(async ([key, entryValue]) => [key, await materializeExternalJsonValues(entryValue, store)])));
 }
-/** Clone one value through the same serialization path used for cache data. */
-async function cloneCacheValue(value, options = void 0) {
-	return deserializeCacheValue(await serializeCacheValue(value, options));
-}
 function normalizeCacheSerializationOptions(options) {
 	return {
 		compress: options?.compress !== false,
@@ -4109,29 +4125,6 @@ function valueKind$1(value) {
 function copyArray(value) {
 	return value.map((item) => item);
 }
-function stripCacheAttributes(attributes) {
-	if (!attributes) return {};
-	const result = {};
-	for (const [key, value] of Object.entries(attributes)) if (!key.startsWith("cache.")) result[key] = value;
-	return result;
-}
-async function snapshotNonCacheAttributes(span) {
-	const snapshot = await cloneCacheValue(stripCacheAttributes(span?.attributes));
-	return isRecordLike$2(snapshot) ? snapshot : {};
-}
-function diffNonCacheAttributes(before, after) {
-	const result = {};
-	for (const [key, value] of Object.entries(after)) if (!cacheAttributeValuesEqual(before[key], value)) result[key] = value;
-	return result;
-}
-function cacheAttributeValuesEqual(left, right) {
-	if (Object.is(left, right)) return true;
-	try {
-		return JSON.stringify(left) === JSON.stringify(right);
-	} catch {
-		return false;
-	}
-}
 function appendCacheRef(span, ref) {
 	if (span === void 0) return;
 	const existing = span.attributes?.["cache.refs"];
@@ -4150,7 +4143,7 @@ function recordCacheRef(scope, span, ref) {
 	}
 	scope.caseCacheRefs.push(ref);
 }
-function serializeSubSpanTree(scope, spanId) {
+function serializeSubSpanTree(scope, spanId, spanIds) {
 	const original = scope.spans.find((s) => s.id === spanId);
 	if (!original) return {
 		kind: "custom",
@@ -4163,7 +4156,7 @@ function serializeSubSpanTree(scope, spanId) {
 		warnings: void 0,
 		children: []
 	};
-	const children = scope.spans.filter((s) => s.parentId === spanId).map((child) => serializeSubSpanTree(scope, child.id));
+	const children = scope.spans.filter((s) => s.parentId === spanId && spanIds.has(s.id)).map((child) => serializeSubSpanTree(scope, child.id, spanIds));
 	return {
 		kind: original.kind,
 		name: original.name,
@@ -4179,9 +4172,9 @@ function serializeSubSpanTree(scope, spanId) {
 function appendSubSpanOps(scope, frame) {
 	for (let i = frame.baseSpanIndex; i < scope.spans.length; i++) {
 		const candidate = scope.spans[i];
-		if (candidate?.parentId === frame.replayParentSpanId) frame.ops.push({
+		if (candidate?.parentId === frame.replayParentSpanId && frame.spanIds.has(candidate.id)) frame.ops.push({
 			kind: "subSpan",
-			span: serializeSubSpanTree(scope, candidate.id)
+			span: serializeSubSpanTree(scope, candidate.id, frame.spanIds)
 		});
 	}
 }
@@ -4437,25 +4430,21 @@ function createTraceCache(generateSpanId) {
 			key: keyHash,
 			status: "bypass"
 		});
-		const beforeAttributes = await snapshotNonCacheAttributes(activeSpan);
 		const frame = {
 			baseSpanIndex: scope.spans.length,
 			replayParentSpanId: activeSpan?.id ?? null,
+			spanIds: /* @__PURE__ */ new Set(),
+			finalAttributes: {},
 			ops: []
 		};
-		scope.recordingStack.push(frame);
-		let bodyResult;
-		try {
-			bodyResult = await fn();
-		} finally {
-			scope.recordingStack.pop();
-		}
+		const bodyResult = await runWithCacheRecordingFrame(frame, async () => {
+			return await fn();
+		});
 		appendSubSpanOps(scope, frame);
 		if (canStore) {
-			const finalAttributes = diffNonCacheAttributes(beforeAttributes, await snapshotNonCacheAttributes(activeSpan));
 			const recording = {
 				returnValue: bodyResult,
-				finalAttributes,
+				finalAttributes: frame.finalAttributes,
 				ops: frame.ops
 			};
 			await cacheCtx.adapter.write({
@@ -4514,6 +4503,13 @@ function mergeSpanAttributes(span, attributes) {
 		...span.attributes,
 		...attributes
 	};
+	const scope = getCurrentScope();
+	if (scope !== void 0) recordCacheRecordingAttributesIfActive(scope, span, attributes);
+}
+function copyNonCacheAttributes(attributes) {
+	const result = {};
+	for (const [key, value] of Object.entries(attributes ?? {})) if (!key.startsWith("cache.")) result[key] = value;
+	return result;
 }
 function isRecordLike$1(value) {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
@@ -4688,6 +4684,7 @@ function startExternalSpan(info) {
 		status: "running",
 		attributes: info.attributes
 	});
+	recordSpanForActiveCacheRecording(scope, id);
 	return createExternalSpanHandle(id);
 }
 function updateExternalSpan(info) {
@@ -4746,6 +4743,7 @@ function recordExternalSpan(info) {
 		warning: info.warning,
 		warnings: info.warnings
 	});
+	recordSpanForActiveCacheRecording(scope, id);
 	return id;
 }
 /**
@@ -4831,6 +4829,7 @@ async function traceSpanInternal(info, fn) {
 		attributes: info.attributes
 	};
 	scope.spans.push(spanRecord);
+	recordSpanForActiveCacheRecording(scope, id);
 	const activeSpan = createSpanHandle(spanRecord);
 	return await runWithActiveSpan(spanRecord, async () => {
 		try {
@@ -4880,21 +4879,19 @@ async function traceSpanInternal(info, fn) {
 				const frame = {
 					baseSpanIndex: scope.spans.length,
 					replayParentSpanId: id,
+					spanIds: /* @__PURE__ */ new Set(),
+					finalAttributes: copyNonCacheAttributes(spanRecord.attributes),
 					ops: []
 				};
-				scope.recordingStack.push(frame);
-				let bodyResult;
-				try {
-					bodyResult = await fn(activeSpan);
-				} finally {
-					scope.recordingStack.pop();
-				}
+				const bodyResult = await runWithCacheRecordingFrame(frame, async () => {
+					return await fn(activeSpan);
+				});
 				appendSubSpanOps(scope, frame);
 				finishSpanWithoutThrownError(spanRecord, realStartedAt);
 				if (canStore) {
 					const recording = {
 						returnValue: bodyResult,
-						finalAttributes: stripCacheAttributes(spanRecord.attributes),
+						finalAttributes: frame.finalAttributes,
 						finalStatus: spanRecord.status,
 						finalError: spanRecord.error,
 						finalErrors: spanRecord.errors,
@@ -4998,14 +4995,12 @@ const evalTracer = {
 			status: "ok",
 			attributes: { value: data }
 		});
-		if (scope.replayingDepth === 0) {
-			const top = scope.recordingStack.at(-1);
-			if (top) top.ops.push({
-				kind: "checkpoint",
-				name,
-				data
-			});
-		}
+		recordSpanForActiveCacheRecording(scope, id);
+		recordCacheRecordingOpIfActive(scope, {
+			kind: "checkpoint",
+			name,
+			data
+		});
 	}
 };
 /** Build a queryable trace tree helper from a flat span list and checkpoints. */

package/dist/{runOrchestration-5xEiQxiS.mjs → runOrchestration-CvmFeOmT.mjs}

@@ -1,4 +1,4 @@
-import { Et as caseRowSchema, Pt as runWithEvalRegistry, Tt as caseDetailSchema, X as runWithEvalClock, _t as validateEvalTagName, bt as runSummarySchema, d as loadEvalModule, f as resolveEvalDefaultConfig, ft as deriveScopedSummaryFromCases, g as commitPendingCacheWrites, gt as matchesTagsFilter, 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, wt as getCaseRowCaseKey, yt as runManifestSchema } from "./runExecution-C4kAOhC1.mjs";
+import { Et as caseRowSchema, Pt as runWithEvalRegistry, Tt as caseDetailSchema, X as runWithEvalClock, _t as validateEvalTagName, bt as runSummarySchema, d as loadEvalModule, f as resolveEvalDefaultConfig, ft as deriveScopedSummaryFromCases, g as commitPendingCacheWrites, gt as matchesTagsFilter, 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, wt as getCaseRowCaseKey, yt as runManifestSchema } from "./runExecution-d42Lm0i5.mjs";
 import { readFile, readdir, rm, writeFile } from "node:fs/promises";
 import { dirname, join } from "node:path";
 import { existsSync } from "node:fs";

package/dist/{runner-JIykMlve.mjs → runner-BKogjiYd.mjs}

@@ -1,2 +1,2 @@
-import { n as initRunner, t as getRunnerInstance } from "./runner-bjd_UB9i.mjs";
+import { n as initRunner, t as getRunnerInstance } from "./runner-MSr8sAWm.mjs";
 export { getRunnerInstance, initRunner };

package/dist/{runner-bjd_UB9i.mjs → runner-MSr8sAWm.mjs}

@@ -1,5 +1,5 @@
-import { n as createRunner } from "./cli-Cf37PZKi.mjs";
-import "./src-303BocMW.mjs";
+import { n as createRunner } from "./cli-_g2qOMK6.mjs";
+import "./src-CdZsOn6y.mjs";
 //#region ../../apps/server/src/runner.ts
 let runnerInstance = null;
 function getRunnerInstance() {

package/dist/{src-303BocMW.mjs → src-CdZsOn6y.mjs}

@@ -1,5 +1,5 @@
-import { Mt as defineEval$1, W as matchesEvalTags$1 } from "./runExecution-C4kAOhC1.mjs";
-import "./cli-Cf37PZKi.mjs";
+import { Mt as defineEval$1, W as matchesEvalTags$1 } from "./runExecution-d42Lm0i5.mjs";
+import "./cli-_g2qOMK6.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": "0.58.0",
+  "version": "0.58.1",
   "type": "module",
   "bin": {
     "agent-evals": "./dist/bin.mjs"
@@ -33,8 +33,8 @@
     "@types/node": "^24.7.2",
     "typescript": "^5.9.2",
     "@agent-evals/runner": "0.0.1",
-    "@agent-evals/sdk": "0.0.1",
-    "@agent-evals/shared": "0.0.1"
+    "@agent-evals/shared": "0.0.1",
+    "@agent-evals/sdk": "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",