@ls-stack/agent-eval 0.35.0 → 0.35.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (14) hide show
  1. package/dist/{app-CcZv9l_q.mjs → app-BlNzXWDM.mjs} +4 -4
  2. package/dist/apps/web/dist/assets/{index-sWPMWjFJ.js → index-CwyTEhGB.js} +41 -41
  3. package/dist/apps/web/dist/index.html +1 -1
  4. package/dist/bin.mjs +1 -1
  5. package/dist/{cli-CVwIjcsX.mjs → cli-Dg3abrOv.mjs} +3 -3
  6. package/dist/index.d.mts +61 -54
  7. package/dist/index.mjs +3 -3
  8. package/dist/runChild.mjs +1 -1
  9. package/dist/{runOrchestration-DoslE_Oo.mjs → runOrchestration-V1TxX8es.mjs} +12 -7
  10. package/dist/{runner-ChHgWruW.mjs → runner-BCs5rzej.mjs} +2 -2
  11. package/dist/{runner-DA_o115w.mjs → runner-znY6PY1M.mjs} +1 -1
  12. package/dist/src-DBypR4TV.mjs +3 -0
  13. package/package.json +3 -3
  14. package/dist/src-Bcc2ZHK8.mjs +0 -3
package/dist/apps/web/dist/index.html CHANGED
@@ -25,7 +25,7 @@
25
25
  href="https://fonts.googleapis.com/css2?family=Geist:wght@400;500;600&family=JetBrains+Mono:wght@400;500&display=swap"
26
26
  rel="stylesheet"
27
27
  />
28
- <script type="module" crossorigin src="/assets/index-sWPMWjFJ.js"></script>
28
+ <script type="module" crossorigin src="/assets/index-CwyTEhGB.js"></script>
29
29
  <link rel="stylesheet" crossorigin href="/assets/index-BJpxc61J.css">
30
30
  </head>
31
31
  <body>
package/dist/bin.mjs CHANGED
@@ -1,5 +1,5 @@
1
1
  #!/usr/bin/env node
2
- import { t as runCli } from "./cli-CVwIjcsX.mjs";
2
+ import { t as runCli } from "./cli-Dg3abrOv.mjs";
3
3
  import { spawn } from "node:child_process";
4
4
  //#region src/bin.ts
5
5
  const moduleMocksFlag = "--experimental-test-module-mocks";
package/dist/{cli-CVwIjcsX.mjs → cli-Dg3abrOv.mjs} RENAMED
@@ -1,4 +1,4 @@
1
- import { C as loadConfig, D as createFsCacheStore, E as validateCharts, H as getEvalDisplayStatus, S as resolveEvalDefaultConfig, T as normalizeScoreDef, U as deriveScopedSummaryFromCases, V as getEvalTitle, _ as buildManualInputDescriptor, a as getLastRunStatuses, b as loadEvalModule, bt as getCaseRowEvalKey, c as loadPersistedRunSnapshots, d as persistRunState, f as recomputeEvalStatusesInRuns, g as resolveArtifactPath, h as resolveTracePresentation, i as generateRunId, l as nextShortIdFromSnapshots, m as runTouchesEval, mt as resolveLlmCallsConfig, n as getTargetEvalKeys, o as getLatestRunInfos, p as recomputePersistedCaseStatus, pr as getEvalRegistry, pt as resolveApiCallsConfig, q as runSummarySchema, s as loadPersistedRunSnapshot, u as persistCaseDetail, v as parseManualInputValues, vt as buildEvalKey, w as buildDeclaredColumnDefs, x as parseEvalDiscovery, y as deriveEvalFreshness, yt as getCaseRowCaseKey, z as applyDerivedCallAttributes } from "./runOrchestration-DoslE_Oo.mjs";
1
+ import { C as loadConfig, D as createFsCacheStore, E as validateCharts, H as getEvalDisplayStatus, S as resolveEvalDefaultConfig, T as normalizeScoreDef, U as deriveScopedSummaryFromCases, V as getEvalTitle, _ as buildManualInputDescriptor, a as getLastRunStatuses, b as loadEvalModule, bt as getCaseRowEvalKey, c as loadPersistedRunSnapshots, d as persistRunState, f as recomputeEvalStatusesInRuns, g as resolveArtifactPath, h as resolveTracePresentation, i as generateRunId, l as nextShortIdFromSnapshots, m as runTouchesEval, mt as resolveLlmCallsConfig, n as getTargetEvalKeys, o as getLatestRunInfos, p as recomputePersistedCaseStatus, pr as getEvalRegistry, pt as resolveApiCallsConfig, q as runSummarySchema, s as loadPersistedRunSnapshot, u as persistCaseDetail, v as parseManualInputValues, vt as buildEvalKey, w as buildDeclaredColumnDefs, x as parseEvalDiscovery, y as deriveEvalFreshness, yt as getCaseRowCaseKey, z as applyDerivedCallAttributes } from "./runOrchestration-V1TxX8es.mjs";
2
2
  import { createHash, randomUUID } from "node:crypto";
3
3
  import { copyFile, mkdir, readFile, rm, writeFile } from "node:fs/promises";
4
4
  import { basename, dirname, extname, isAbsolute, join, relative, resolve, sep } from "node:path";
@@ -1940,8 +1940,8 @@ async function commandApp(args) {
1940
1940
  const { serve } = await import("@hono/node-server");
1941
1941
  const bundledWebDist = resolve(currentDir, "apps/web/dist");
1942
1942
  if (existsSync(bundledWebDist)) process.env.AGENT_EVALS_WEB_DIST = bundledWebDist;
1943
- const appModule = await import("./app-CcZv9l_q.mjs");
1944
- const runnerModule = await import("./runner-DA_o115w.mjs");
1943
+ const appModule = await import("./app-BlNzXWDM.mjs");
1944
+ const runnerModule = await import("./runner-znY6PY1M.mjs");
1945
1945
  if (!isHonoAppModule(appModule)) throw new Error("Server app module is invalid");
1946
1946
  if (!isServerRunnerModule(runnerModule)) throw new Error("Server runner module is invalid");
1947
1947
  await runnerModule.initRunner();
package/dist/index.d.mts CHANGED
@@ -2626,6 +2626,13 @@ type CacheSerializationOptions = {
2626
2626
  * and set items are omitted instead of being written to cache files.
2627
2627
  */
2628
2628
  preserveUndefined?: boolean;
2629
+ /**
2630
+ * Compress large nested strings/JSON blobs with gzip wrappers.
2631
+ *
2632
+ * Enabled by default for reusable cache files. Disable for output artifacts
2633
+ * that need synchronous browser-side deserialization.
2634
+ */
2635
+ compress?: boolean;
2629
2636
  };
2630
2637
  /**
2631
2638
  * Serialize one cached value while keeping plain JSON as plain JSON.
@@ -2908,13 +2915,13 @@ type ColumnKind = z$1.infer<typeof columnKindSchema>;
2908
2915
  declare const columnFormatSchema: z$1.ZodEnum<{
2909
2916
  number: "number";
2910
2917
  boolean: "boolean";
2918
+ duration: "duration";
2919
+ json: "json";
2911
2920
  file: "file";
2912
2921
  markdown: "markdown";
2913
- json: "json";
2914
2922
  image: "image";
2915
2923
  audio: "audio";
2916
2924
  video: "video";
2917
- duration: "duration";
2918
2925
  percent: "percent";
2919
2926
  passFail: "passFail";
2920
2927
  stars: "stars";
@@ -2933,13 +2940,13 @@ declare const columnDefSchema: z$1.ZodObject<{
2933
2940
  format: z$1.ZodOptional<z$1.ZodEnum<{
2934
2941
  number: "number";
2935
2942
  boolean: "boolean";
2943
+ duration: "duration";
2944
+ json: "json";
2936
2945
  file: "file";
2937
2946
  markdown: "markdown";
2938
- json: "json";
2939
2947
  image: "image";
2940
2948
  audio: "audio";
2941
2949
  video: "video";
2942
- duration: "duration";
2943
2950
  percent: "percent";
2944
2951
  passFail: "passFail";
2945
2952
  stars: "stars";
@@ -2984,8 +2991,8 @@ declare const traceSpanKindSchema: z$1.ZodString;
2984
2991
  declare const traceAttributeDisplayFormatSchema: z$1.ZodEnum<{
2985
2992
  string: "string";
2986
2993
  number: "number";
2987
- json: "json";
2988
2994
  duration: "duration";
2995
+ json: "json";
2989
2996
  }>;
2990
2997
  /**
2991
2998
  * Formatting hint for trace attribute values rendered by the UI.
@@ -3009,8 +3016,8 @@ declare const traceAttributeDisplaySchema: z$1.ZodObject<{
3009
3016
  format: z$1.ZodOptional<z$1.ZodEnum<{
3010
3017
  string: "string";
3011
3018
  number: "number";
3012
- json: "json";
3013
3019
  duration: "duration";
3020
+ json: "json";
3014
3021
  }>>;
3015
3022
  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
3016
3023
  placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
@@ -3045,8 +3052,8 @@ declare const traceDisplayConfigSchema: z$1.ZodObject<{
3045
3052
  format: z$1.ZodOptional<z$1.ZodEnum<{
3046
3053
  string: "string";
3047
3054
  number: "number";
3048
- json: "json";
3049
3055
  duration: "duration";
3056
+ json: "json";
3050
3057
  }>>;
3051
3058
  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
3052
3059
  placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
@@ -3085,8 +3092,8 @@ declare const traceAttributeDisplayInputSchema: z$1.ZodObject<{
3085
3092
  format: z$1.ZodOptional<z$1.ZodEnum<{
3086
3093
  string: "string";
3087
3094
  number: "number";
3088
- json: "json";
3089
3095
  duration: "duration";
3096
+ json: "json";
3090
3097
  }>>;
3091
3098
  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
3092
3099
  placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
@@ -3123,8 +3130,8 @@ declare const traceDisplayInputConfigSchema: z$1.ZodObject<{
3123
3130
  format: z$1.ZodOptional<z$1.ZodEnum<{
3124
3131
  string: "string";
3125
3132
  number: "number";
3126
- json: "json";
3127
3133
  duration: "duration";
3134
+ json: "json";
3128
3135
  }>>;
3129
3136
  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
3130
3137
  placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
@@ -3255,13 +3262,13 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
3255
3262
  format: z$1.ZodOptional<z$1.ZodEnum<{
3256
3263
  number: "number";
3257
3264
  boolean: "boolean";
3265
+ duration: "duration";
3266
+ json: "json";
3258
3267
  file: "file";
3259
3268
  markdown: "markdown";
3260
- json: "json";
3261
3269
  image: "image";
3262
3270
  audio: "audio";
3263
3271
  video: "video";
3264
- duration: "duration";
3265
3272
  percent: "percent";
3266
3273
  passFail: "passFail";
3267
3274
  stars: "stars";
@@ -3297,13 +3304,13 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
3297
3304
  format: z$1.ZodOptional<z$1.ZodEnum<{
3298
3305
  number: "number";
3299
3306
  boolean: "boolean";
3307
+ duration: "duration";
3308
+ json: "json";
3300
3309
  file: "file";
3301
3310
  markdown: "markdown";
3302
- json: "json";
3303
3311
  image: "image";
3304
3312
  audio: "audio";
3305
3313
  video: "video";
3306
- duration: "duration";
3307
3314
  percent: "percent";
3308
3315
  passFail: "passFail";
3309
3316
  stars: "stars";
@@ -3340,13 +3347,13 @@ declare const evalSummarySchema: z$1.ZodObject<{
3340
3347
  format: z$1.ZodOptional<z$1.ZodEnum<{
3341
3348
  number: "number";
3342
3349
  boolean: "boolean";
3350
+ duration: "duration";
3351
+ json: "json";
3343
3352
  file: "file";
3344
3353
  markdown: "markdown";
3345
- json: "json";
3346
3354
  image: "image";
3347
3355
  audio: "audio";
3348
3356
  video: "video";
3349
- duration: "duration";
3350
3357
  percent: "percent";
3351
3358
  passFail: "passFail";
3352
3359
  stars: "stars";
@@ -3398,13 +3405,13 @@ declare const evalSummarySchema: z$1.ZodObject<{
3398
3405
  format: z$1.ZodOptional<z$1.ZodEnum<{
3399
3406
  number: "number";
3400
3407
  boolean: "boolean";
3408
+ duration: "duration";
3409
+ json: "json";
3401
3410
  file: "file";
3402
3411
  markdown: "markdown";
3403
- json: "json";
3404
3412
  image: "image";
3405
3413
  audio: "audio";
3406
3414
  video: "video";
3407
- duration: "duration";
3408
3415
  percent: "percent";
3409
3416
  passFail: "passFail";
3410
3417
  stars: "stars";
@@ -3429,8 +3436,8 @@ declare const evalSummarySchema: z$1.ZodObject<{
3429
3436
  }>;
3430
3437
  label: z$1.ZodOptional<z$1.ZodString>;
3431
3438
  color: z$1.ZodOptional<z$1.ZodEnum<{
3432
- success: "success";
3433
3439
  error: "error";
3440
+ success: "success";
3434
3441
  warning: "warning";
3435
3442
  accent: "accent";
3436
3443
  accentDim: "accentDim";
@@ -3453,8 +3460,8 @@ declare const evalSummarySchema: z$1.ZodObject<{
3453
3460
  }>;
3454
3461
  label: z$1.ZodOptional<z$1.ZodString>;
3455
3462
  color: z$1.ZodOptional<z$1.ZodEnum<{
3456
- success: "success";
3457
3463
  error: "error";
3464
+ success: "success";
3458
3465
  warning: "warning";
3459
3466
  accent: "accent";
3460
3467
  accentDim: "accentDim";
@@ -3587,9 +3594,9 @@ declare const caseRowSchema: z$1.ZodObject<{
3587
3594
  error: "error";
3588
3595
  running: "running";
3589
3596
  cancelled: "cancelled";
3597
+ pending: "pending";
3590
3598
  pass: "pass";
3591
3599
  fail: "fail";
3592
- pending: "pending";
3593
3600
  }>;
3594
3601
  durationMs: z$1.ZodNullable<z$1.ZodNumber>;
3595
3602
  costUsd: z$1.ZodOptional<z$1.ZodNullable<z$1.ZodNumber>>;
@@ -3718,8 +3725,8 @@ declare const scoreTraceSchema: z$1.ZodObject<{
3718
3725
  format: z$1.ZodOptional<z$1.ZodEnum<{
3719
3726
  string: "string";
3720
3727
  number: "number";
3721
- json: "json";
3722
3728
  duration: "duration";
3729
+ json: "json";
3723
3730
  }>>;
3724
3731
  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
3725
3732
  placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
@@ -3751,9 +3758,9 @@ declare const caseDetailSchema: z$1.ZodObject<{
3751
3758
  error: "error";
3752
3759
  running: "running";
3753
3760
  cancelled: "cancelled";
3761
+ pending: "pending";
3754
3762
  pass: "pass";
3755
3763
  fail: "fail";
3756
- pending: "pending";
3757
3764
  }>;
3758
3765
  input: z$1.ZodUnknown;
3759
3766
  trace: z$1.ZodArray<z$1.ZodObject<{
@@ -3804,8 +3811,8 @@ declare const caseDetailSchema: z$1.ZodObject<{
3804
3811
  format: z$1.ZodOptional<z$1.ZodEnum<{
3805
3812
  string: "string";
3806
3813
  number: "number";
3807
- json: "json";
3808
3814
  duration: "duration";
3815
+ json: "json";
3809
3816
  }>>;
3810
3817
  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
3811
3818
  placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
@@ -3873,8 +3880,8 @@ declare const caseDetailSchema: z$1.ZodObject<{
3873
3880
  format: z$1.ZodOptional<z$1.ZodEnum<{
3874
3881
  string: "string";
3875
3882
  number: "number";
3876
- json: "json";
3877
3883
  duration: "duration";
3884
+ json: "json";
3878
3885
  }>>;
3879
3886
  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
3880
3887
  placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
@@ -3949,10 +3956,10 @@ declare const caseDetailSchema: z$1.ZodObject<{
3949
3956
  namespace: z$1.ZodString;
3950
3957
  key: z$1.ZodString;
3951
3958
  status: z$1.ZodEnum<{
3959
+ bypass: "bypass";
3960
+ refresh: "refresh";
3952
3961
  hit: "hit";
3953
3962
  miss: "miss";
3954
- refresh: "refresh";
3955
- bypass: "bypass";
3956
3963
  }>;
3957
3964
  read: z$1.ZodOptional<z$1.ZodBoolean>;
3958
3965
  stored: z$1.ZodOptional<z$1.ZodBoolean>;
@@ -4037,8 +4044,8 @@ type EvalChartAggregate = z$1.infer<typeof evalChartAggregateSchema>;
4037
4044
  * not emit raw hex so authored evals stay decoupled from the web theme.
4038
4045
  */
4039
4046
  declare const evalChartColorSchema: z$1.ZodEnum<{
4040
- success: "success";
4041
4047
  error: "error";
4048
+ success: "success";
4042
4049
  warning: "warning";
4043
4050
  accent: "accent";
4044
4051
  accentDim: "accentDim";
@@ -4066,8 +4073,8 @@ declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
4066
4073
  }>;
4067
4074
  label: z$1.ZodOptional<z$1.ZodString>;
4068
4075
  color: z$1.ZodOptional<z$1.ZodEnum<{
4069
- success: "success";
4070
4076
  error: "error";
4077
+ success: "success";
4071
4078
  warning: "warning";
4072
4079
  accent: "accent";
4073
4080
  accentDim: "accentDim";
@@ -4090,8 +4097,8 @@ declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
4090
4097
  }>;
4091
4098
  label: z$1.ZodOptional<z$1.ZodString>;
4092
4099
  color: z$1.ZodOptional<z$1.ZodEnum<{
4093
- success: "success";
4094
4100
  error: "error";
4101
+ success: "success";
4095
4102
  warning: "warning";
4096
4103
  accent: "accent";
4097
4104
  accentDim: "accentDim";
@@ -4149,8 +4156,8 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
4149
4156
  }>;
4150
4157
  label: z$1.ZodOptional<z$1.ZodString>;
4151
4158
  color: z$1.ZodOptional<z$1.ZodEnum<{
4152
- success: "success";
4153
4159
  error: "error";
4160
+ success: "success";
4154
4161
  warning: "warning";
4155
4162
  accent: "accent";
4156
4163
  accentDim: "accentDim";
@@ -4173,8 +4180,8 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
4173
4180
  }>;
4174
4181
  label: z$1.ZodOptional<z$1.ZodString>;
4175
4182
  color: z$1.ZodOptional<z$1.ZodEnum<{
4176
- success: "success";
4177
4183
  error: "error";
4184
+ success: "success";
4178
4185
  warning: "warning";
4179
4186
  accent: "accent";
4180
4187
  accentDim: "accentDim";
@@ -4239,8 +4246,8 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
4239
4246
  }>;
4240
4247
  label: z$1.ZodOptional<z$1.ZodString>;
4241
4248
  color: z$1.ZodOptional<z$1.ZodEnum<{
4242
- success: "success";
4243
4249
  error: "error";
4250
+ success: "success";
4244
4251
  warning: "warning";
4245
4252
  accent: "accent";
4246
4253
  accentDim: "accentDim";
@@ -4263,8 +4270,8 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
4263
4270
  }>;
4264
4271
  label: z$1.ZodOptional<z$1.ZodString>;
4265
4272
  color: z$1.ZodOptional<z$1.ZodEnum<{
4266
- success: "success";
4267
4273
  error: "error";
4274
+ success: "success";
4268
4275
  warning: "warning";
4269
4276
  accent: "accent";
4270
4277
  accentDim: "accentDim";
@@ -4341,9 +4348,9 @@ declare const runManifestSchema: z$1.ZodObject<{
4341
4348
  median: "median";
4342
4349
  }>>>;
4343
4350
  cacheMode: z$1.ZodOptional<z$1.ZodEnum<{
4344
- refresh: "refresh";
4345
- bypass: "bypass";
4346
4351
  use: "use";
4352
+ bypass: "bypass";
4353
+ refresh: "refresh";
4347
4354
  }>>;
4348
4355
  }, z$1.core.$strip>;
4349
4356
  /** Persisted lifecycle metadata for a single eval run. */
@@ -4574,8 +4581,8 @@ declare const llmCallMetricFormatSchema: z$1.ZodEnum<{
4574
4581
  string: "string";
4575
4582
  number: "number";
4576
4583
  boolean: "boolean";
4577
- json: "json";
4578
4584
  duration: "duration";
4585
+ json: "json";
4579
4586
  }>;
4580
4587
  /** Render format applied to an LLM-call metric value. */
4581
4588
  type LlmCallMetricFormat = z$1.infer<typeof llmCallMetricFormatSchema>;
@@ -4584,8 +4591,8 @@ declare const apiCallMetricFormatSchema: z$1.ZodEnum<{
4584
4591
  string: "string";
4585
4592
  number: "number";
4586
4593
  boolean: "boolean";
4587
- json: "json";
4588
4594
  duration: "duration";
4595
+ json: "json";
4589
4596
  }>;
4590
4597
  /** Render format applied to an API-call metric value. */
4591
4598
  type ApiCallMetricFormat = z$1.infer<typeof apiCallMetricFormatSchema>;
@@ -4654,8 +4661,8 @@ declare const llmCallMetricSchema: z$1.ZodObject<{
4654
4661
  string: "string";
4655
4662
  number: "number";
4656
4663
  boolean: "boolean";
4657
- json: "json";
4658
4664
  duration: "duration";
4665
+ json: "json";
4659
4666
  }>>;
4660
4667
  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
4661
4668
  placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
@@ -4683,8 +4690,8 @@ declare const apiCallMetricSchema: z$1.ZodObject<{
4683
4690
  string: "string";
4684
4691
  number: "number";
4685
4692
  boolean: "boolean";
4686
- json: "json";
4687
4693
  duration: "duration";
4694
+ json: "json";
4688
4695
  }>>;
4689
4696
  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
4690
4697
  placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
@@ -4797,8 +4804,8 @@ declare const llmCallsConfigSchema: z$1.ZodObject<{
4797
4804
  string: "string";
4798
4805
  number: "number";
4799
4806
  boolean: "boolean";
4800
- json: "json";
4801
4807
  duration: "duration";
4808
+ json: "json";
4802
4809
  }>>;
4803
4810
  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
4804
4811
  placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
@@ -4833,8 +4840,8 @@ declare const apiCallsConfigSchema: z$1.ZodObject<{
4833
4840
  string: "string";
4834
4841
  number: "number";
4835
4842
  boolean: "boolean";
4836
- json: "json";
4837
4843
  duration: "duration";
4844
+ json: "json";
4838
4845
  }>>;
4839
4846
  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
4840
4847
  placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
@@ -5135,8 +5142,8 @@ declare const agentEvalsConfigSchema: z$1.ZodObject<{
5135
5142
  format: z$1.ZodOptional<z$1.ZodEnum<{
5136
5143
  string: "string";
5137
5144
  number: "number";
5138
- json: "json";
5139
5145
  duration: "duration";
5146
+ json: "json";
5140
5147
  }>>;
5141
5148
  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
5142
5149
  placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
@@ -5183,13 +5190,13 @@ declare const agentEvalsConfigSchema: z$1.ZodObject<{
5183
5190
  format: z$1.ZodOptional<z$1.ZodEnum<{
5184
5191
  number: "number";
5185
5192
  boolean: "boolean";
5193
+ duration: "duration";
5194
+ json: "json";
5186
5195
  file: "file";
5187
5196
  markdown: "markdown";
5188
- json: "json";
5189
5197
  image: "image";
5190
5198
  audio: "audio";
5191
5199
  video: "video";
5192
- duration: "duration";
5193
5200
  percent: "percent";
5194
5201
  passFail: "passFail";
5195
5202
  stars: "stars";
@@ -5248,8 +5255,8 @@ declare const agentEvalsConfigSchema: z$1.ZodObject<{
5248
5255
  string: "string";
5249
5256
  number: "number";
5250
5257
  boolean: "boolean";
5251
- json: "json";
5252
5258
  duration: "duration";
5259
+ json: "json";
5253
5260
  }>>;
5254
5261
  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
5255
5262
  placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
@@ -5293,8 +5300,8 @@ declare const agentEvalsConfigSchema: z$1.ZodObject<{
5293
5300
  string: "string";
5294
5301
  number: "number";
5295
5302
  boolean: "boolean";
5296
- json: "json";
5297
5303
  duration: "duration";
5304
+ json: "json";
5298
5305
  }>>;
5299
5306
  numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
5300
5307
  placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
@@ -5527,9 +5534,9 @@ declare function extractApiCalls(spans: EvalTraceSpan[], config: ResolvedApiCall
5527
5534
  * - `refresh`: never read, always write (forces re-execution and overwrites).
5528
5535
  */
5529
5536
  declare const cacheModeSchema: z$1.ZodEnum<{
5530
- refresh: "refresh";
5531
- bypass: "bypass";
5532
5537
  use: "use";
5538
+ bypass: "bypass";
5539
+ refresh: "refresh";
5533
5540
  }>;
5534
5541
  /** Mode controlling how cached spans behave during a run. */
5535
5542
  type CacheMode = z$1.infer<typeof cacheModeSchema>;
@@ -5550,10 +5557,10 @@ declare const cacheOperationTypeSchema: z$1.ZodEnum<{
5550
5557
  type CacheOperationType = z$1.infer<typeof cacheOperationTypeSchema>;
5551
5558
  /** Status of a cache lookup recorded on a span or case scope. */
5552
5559
  declare const cacheStatusSchema: z$1.ZodEnum<{
5560
+ bypass: "bypass";
5561
+ refresh: "refresh";
5553
5562
  hit: "hit";
5554
5563
  miss: "miss";
5555
- refresh: "refresh";
5556
- bypass: "bypass";
5557
5564
  }>;
5558
5565
  /** Status of a cache lookup recorded on a span or case scope. */
5559
5566
  type CacheStatus = z$1.infer<typeof cacheStatusSchema>;
@@ -5570,10 +5577,10 @@ declare const traceCacheRefSchema: z$1.ZodObject<{
5570
5577
  namespace: z$1.ZodString;
5571
5578
  key: z$1.ZodString;
5572
5579
  status: z$1.ZodEnum<{
5580
+ bypass: "bypass";
5581
+ refresh: "refresh";
5573
5582
  hit: "hit";
5574
5583
  miss: "miss";
5575
- refresh: "refresh";
5576
- bypass: "bypass";
5577
5584
  }>;
5578
5585
  read: z$1.ZodOptional<z$1.ZodBoolean>;
5579
5586
  stored: z$1.ZodOptional<z$1.ZodBoolean>;
@@ -6089,9 +6096,9 @@ declare const createRunRequestSchema: z$1.ZodObject<{
6089
6096
  trials: z$1.ZodNumber;
6090
6097
  cache: z$1.ZodOptional<z$1.ZodObject<{
6091
6098
  mode: z$1.ZodDefault<z$1.ZodEnum<{
6092
- refresh: "refresh";
6093
- bypass: "bypass";
6094
6099
  use: "use";
6100
+ bypass: "bypass";
6101
+ refresh: "refresh";
6095
6102
  }>>;
6096
6103
  }, z$1.core.$strip>>;
6097
6104
  manualInputs: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
package/dist/index.mjs CHANGED
@@ -1,4 +1,4 @@
1
- import { $ as apiCallMetricSchema, $n as getCurrentScope, $t as cacheDebugKeyEntrySchema, A as createRunRequestSchema, An as repoFileRefSchema, At as runLogEntrySchema, B as getNestedAttribute, Bn as deserializeCacheValue, Bt as manualInputNumberFieldSchema, Cn as cellValueSchema, Ct as caseRowSchema, Dn as fileRefSchema, Dt as evalStatItemSchema, En as columnKindSchema, Et as evalStatAggregateSchema, F as extractApiCalls, Fn as evalSpan, Ft as manualInputBooleanFieldSchema, G as deriveStatusFromChildStatuses, Gn as readManualInputFile, Gt as evalChartAxisSchema, H as getEvalDisplayStatus, Hn as serializeCacheValue, Ht as manualInputSelectOptionSchema, I as extractLlmCalls, In as evalTracer, It as manualInputDescriptorSchema, J as DEFAULT_API_CALLS_CONFIG, Jn as advanceEvalTime, Jt as evalChartConfigSchema, K as runManifestSchema, Kn as evalExpect, Kt as evalChartBuiltinMetricSchema, L as simulateLlmCallCost, Ln as hashCacheKey, Lt as manualInputFieldDescriptorSchema, M as sseEnvelopeSchema, Mn as z, Mt as runLogLocationSchema, N as extractCacheEntries, Nn as buildTraceTree, Nt as runLogPhaseSchema, O as configReloadStateSchema, On as jsonCellSchema, Ot as evalStatsConfigSchema, P as extractCacheHits, Pn as captureEvalSpanError, Pt as scoreTraceSchema, Q as apiCallMetricPlacementSchema, Qn as evalLog, Qt as evalChartsConfigSchema, R as simulateTokenAllocation, Rn as hashCacheKeySync, Rt as manualInputJsonFieldSchema, Sn as traceSpanWarningSchema, St as caseDetailSchema, Tn as columnFormatSchema, Tt as evalFreshnessStatusSchema, U as deriveScopedSummaryFromCases, Un as repoFile, Ut as manualInputTextFieldSchema, V as getEvalTitle, Vn as serializeCacheRecording, Vt as manualInputSelectFieldSchema, W as deriveStatusFromCaseRows, Wn as manualInputFileValueSchema, Wt as evalChartAggregateSchema, X as agentEvalsConfigSchema, Xt as evalChartTooltipExtraSchema, Y as DEFAULT_LLM_CALLS_CONFIG, Yn as appendToEvalOutput, Yt as evalChartMetricSchema, Z as apiCallMetricFormatSchema, Zn as evalAssert, Zt as evalChartTypeSchema, _n as traceDisplayConfigSchema, _t as buildCaseKey, an as cacheModeSchema, ar as nextEvalId, at as llmCallCostCurrencySchema, bn as traceSpanKindSchema, bt as getCaseRowEvalKey, cn as cacheRecordingSchema, cr as runInExistingEvalScope, ct as llmCallMetricSchema, dn as spanCacheOptionsSchema, dr as startEvalBackgroundJob, dt as llmCallsConfigSchema, en as cacheDebugKeyFileSchema, er as getEvalCaseInput, et as apiCallsConfigSchema, fn as traceCacheRefSchema, fr as defineEval, ft as removeDefaultConfigSchema, gn as traceAttributeDisplaySchema, gt as trialSelectionModeSchema, hn as traceAttributeDisplayPlacementSchema, ht as runLogsConfigSchema, in as cacheListItemSchema, ir as mergeEvalOutput, it as evalDeriveConfigSchema, j as updateManualScoreRequestSchema, jn as runArtifactRefSchema, jt as runLogLevelSchema, k as configReloadStatusSchema, kn as numberDisplayOptionsSchema, kt as evalSummarySchema, ln as cacheStatusSchema, lr as setEvalOutput, lt as llmCallPricingRateSchema, mn as traceAttributeDisplayInputSchema, mt as resolveLlmCallsConfig, nn as cacheEntryWithDebugKeySchema, nr as incrementEvalOutput, nt as evalColumnOverrideSchema, on as cacheOperationTypeSchema, or as runInEvalRuntimeScope, ot as llmCallMetricFormatSchema, pn as traceAttributeDisplayFormatSchema, pr as getEvalRegistry, pt as resolveApiCallsConfig, q as runSummarySchema, qn as EvalAssertionError, qt as evalChartColorSchema, rn as cacheFileSchema, rr as isInEvalScope, rt as evalColumnsSchema, sn as cacheRecordingOpSchema, sr as runInEvalScope, st as llmCallMetricPlacementSchema, tn as cacheEntrySchema, tr as getEvalStartTime, tt as defaultConfigKeySchema, un as serializedCacheSpanSchema, ur as setScopeCacheContext, ut as llmCallPricingSchema, vn as traceDisplayInputConfigSchema, vt as buildEvalKey, wn as columnDefSchema, wt as discoveryIssueSchema, xn as traceSpanSchema, xt as assertionFailureSchema, yn as traceSpanErrorSchema, yt as getCaseRowCaseKey, z as applyDerivedCallAttributes, zn as deserializeCacheRecording, zt as manualInputMultilineFieldSchema } from "./runOrchestration-DoslE_Oo.mjs";
2
- import { a as materializeManualInputFiles, i as isManualInputFileValue, n as createRunner, o as stageManualInputFile, r as cleanupStagedManualInputFiles, s as stageManualInputFileFromPath, t as runCli } from "./cli-CVwIjcsX.mjs";
3
- import "./src-Bcc2ZHK8.mjs";
1
+ import { $ as apiCallMetricSchema, $n as getCurrentScope, $t as cacheDebugKeyEntrySchema, A as createRunRequestSchema, An as repoFileRefSchema, At as runLogEntrySchema, B as getNestedAttribute, Bn as deserializeCacheValue, Bt as manualInputNumberFieldSchema, Cn as cellValueSchema, Ct as caseRowSchema, Dn as fileRefSchema, Dt as evalStatItemSchema, En as columnKindSchema, Et as evalStatAggregateSchema, F as extractApiCalls, Fn as evalSpan, Ft as manualInputBooleanFieldSchema, G as deriveStatusFromChildStatuses, Gn as readManualInputFile, Gt as evalChartAxisSchema, H as getEvalDisplayStatus, Hn as serializeCacheValue, Ht as manualInputSelectOptionSchema, I as extractLlmCalls, In as evalTracer, It as manualInputDescriptorSchema, J as DEFAULT_API_CALLS_CONFIG, Jn as advanceEvalTime, Jt as evalChartConfigSchema, K as runManifestSchema, Kn as evalExpect, Kt as evalChartBuiltinMetricSchema, L as simulateLlmCallCost, Ln as hashCacheKey, Lt as manualInputFieldDescriptorSchema, M as sseEnvelopeSchema, Mn as z, Mt as runLogLocationSchema, N as extractCacheEntries, Nn as buildTraceTree, Nt as runLogPhaseSchema, O as configReloadStateSchema, On as jsonCellSchema, Ot as evalStatsConfigSchema, P as extractCacheHits, Pn as captureEvalSpanError, Pt as scoreTraceSchema, Q as apiCallMetricPlacementSchema, Qn as evalLog, Qt as evalChartsConfigSchema, R as simulateTokenAllocation, Rn as hashCacheKeySync, Rt as manualInputJsonFieldSchema, Sn as traceSpanWarningSchema, St as caseDetailSchema, Tn as columnFormatSchema, Tt as evalFreshnessStatusSchema, U as deriveScopedSummaryFromCases, Un as repoFile, Ut as manualInputTextFieldSchema, V as getEvalTitle, Vn as serializeCacheRecording, Vt as manualInputSelectFieldSchema, W as deriveStatusFromCaseRows, Wn as manualInputFileValueSchema, Wt as evalChartAggregateSchema, X as agentEvalsConfigSchema, Xt as evalChartTooltipExtraSchema, Y as DEFAULT_LLM_CALLS_CONFIG, Yn as appendToEvalOutput, Yt as evalChartMetricSchema, Z as apiCallMetricFormatSchema, Zn as evalAssert, Zt as evalChartTypeSchema, _n as traceDisplayConfigSchema, _t as buildCaseKey, an as cacheModeSchema, ar as nextEvalId, at as llmCallCostCurrencySchema, bn as traceSpanKindSchema, bt as getCaseRowEvalKey, cn as cacheRecordingSchema, cr as runInExistingEvalScope, ct as llmCallMetricSchema, dn as spanCacheOptionsSchema, dr as startEvalBackgroundJob, dt as llmCallsConfigSchema, en as cacheDebugKeyFileSchema, er as getEvalCaseInput, et as apiCallsConfigSchema, fn as traceCacheRefSchema, fr as defineEval, ft as removeDefaultConfigSchema, gn as traceAttributeDisplaySchema, gt as trialSelectionModeSchema, hn as traceAttributeDisplayPlacementSchema, ht as runLogsConfigSchema, in as cacheListItemSchema, ir as mergeEvalOutput, it as evalDeriveConfigSchema, j as updateManualScoreRequestSchema, jn as runArtifactRefSchema, jt as runLogLevelSchema, k as configReloadStatusSchema, kn as numberDisplayOptionsSchema, kt as evalSummarySchema, ln as cacheStatusSchema, lr as setEvalOutput, lt as llmCallPricingRateSchema, mn as traceAttributeDisplayInputSchema, mt as resolveLlmCallsConfig, nn as cacheEntryWithDebugKeySchema, nr as incrementEvalOutput, nt as evalColumnOverrideSchema, on as cacheOperationTypeSchema, or as runInEvalRuntimeScope, ot as llmCallMetricFormatSchema, pn as traceAttributeDisplayFormatSchema, pr as getEvalRegistry, pt as resolveApiCallsConfig, q as runSummarySchema, qn as EvalAssertionError, qt as evalChartColorSchema, rn as cacheFileSchema, rr as isInEvalScope, rt as evalColumnsSchema, sn as cacheRecordingOpSchema, sr as runInEvalScope, st as llmCallMetricPlacementSchema, tn as cacheEntrySchema, tr as getEvalStartTime, tt as defaultConfigKeySchema, un as serializedCacheSpanSchema, ur as setScopeCacheContext, ut as llmCallPricingSchema, vn as traceDisplayInputConfigSchema, vt as buildEvalKey, wn as columnDefSchema, wt as discoveryIssueSchema, xn as traceSpanSchema, xt as assertionFailureSchema, yn as traceSpanErrorSchema, yt as getCaseRowCaseKey, z as applyDerivedCallAttributes, zn as deserializeCacheRecording, zt as manualInputMultilineFieldSchema } from "./runOrchestration-V1TxX8es.mjs";
2
+ import { a as materializeManualInputFiles, i as isManualInputFileValue, n as createRunner, o as stageManualInputFile, r as cleanupStagedManualInputFiles, s as stageManualInputFileFromPath, t as runCli } from "./cli-Dg3abrOv.mjs";
3
+ import "./src-DBypR4TV.mjs";
4
4
  export { DEFAULT_API_CALLS_CONFIG, DEFAULT_LLM_CALLS_CONFIG, EvalAssertionError, advanceEvalTime, 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, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, fileRefSchema, getCaseRowCaseKey, getCaseRowEvalKey, getCurrentScope, getEvalCaseInput, getEvalDisplayStatus, getEvalRegistry, getEvalStartTime, 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 };
package/dist/runChild.mjs CHANGED
@@ -1,4 +1,4 @@
1
- import { A as createRunRequestSchema, C as loadConfig, D as createFsCacheStore, It as manualInputDescriptorSchema, K as runManifestSchema, Ot as evalStatsConfigSchema, Qt as evalChartsConfigSchema, Xn as configureEvalRunLogs, q as runSummarySchema, r as getTargetEvals$1, t as executeRun, vt as buildEvalKey, wn as columnDefSchema, x as parseEvalDiscovery } from "./runOrchestration-DoslE_Oo.mjs";
1
+ import { A as createRunRequestSchema, C as loadConfig, D as createFsCacheStore, It as manualInputDescriptorSchema, K as runManifestSchema, Ot as evalStatsConfigSchema, Qt as evalChartsConfigSchema, Xn as configureEvalRunLogs, q as runSummarySchema, r as getTargetEvals$1, t as executeRun, vt as buildEvalKey, wn as columnDefSchema, x as parseEvalDiscovery } from "./runOrchestration-V1TxX8es.mjs";
2
2
  import { createHash } from "node:crypto";
3
3
  import { readFile } from "node:fs/promises";
4
4
  import { relative } from "node:path";
package/dist/{runOrchestration-DoslE_Oo.mjs → runOrchestration-V1TxX8es.mjs} RENAMED
@@ -913,13 +913,16 @@ async function cloneCacheValue(value, options = void 0) {
913
913
  return deserializeCacheValue(await serializeCacheValue(value, options));
914
914
  }
915
915
  function normalizeCacheSerializationOptions(options) {
916
- return { preserveUndefined: options?.preserveUndefined === true };
916
+ return {
917
+ compress: options?.compress !== false,
918
+ preserveUndefined: options?.preserveUndefined === true
919
+ };
917
920
  }
918
921
  async function serializeJsonSafeValue(value, refs, depth, config) {
919
922
  if (value === void 0) return config.preserveUndefined ? jsonSafeValue("Undefined") : void 0;
920
923
  if (typeof value === "bigint") return jsonSafeValue("BigInt", value.toString());
921
924
  if (typeof value === "number") return serializeNumber(value);
922
- if (typeof value === "string") return serializeString(value, depth);
925
+ if (typeof value === "string") return serializeString(value, depth, config);
923
926
  if (value instanceof Date) return jsonSafeValue("Date", value.toISOString());
924
927
  if (value instanceof Map) return serializeMap(value, refs, depth, config);
925
928
  if (value instanceof Set) return serializeSet(value, refs, depth, config);
@@ -959,7 +962,7 @@ async function serializeJsonSafeValue(value, refs, depth, config) {
959
962
  if (serializedItem !== void 0) items.push(serializedItem);
960
963
  }
961
964
  refs.delete(value);
962
- return compressNestedJsonValue(items, depth) ?? items;
965
+ return compressNestedJsonValue(items, depth, config) ?? items;
963
966
  }
964
967
  const entries = [];
965
968
  for (const [key, entryValue] of Object.entries(value)) {
@@ -968,7 +971,7 @@ async function serializeJsonSafeValue(value, refs, depth, config) {
968
971
  }
969
972
  refs.delete(value);
970
973
  const serialized = hasSerializationMarkerKey(value) ? jsonSafeValue("Object", entries) : Object.fromEntries(entries);
971
- return compressNestedJsonValue(serialized, depth) ?? serialized;
974
+ return compressNestedJsonValue(serialized, depth, config) ?? serialized;
972
975
  }
973
976
  function serializeNumber(value) {
974
977
  if (Number.isNaN(value)) return jsonSafeValue("Number", "NaN");
@@ -977,8 +980,9 @@ function serializeNumber(value) {
977
980
  if (Object.is(value, -0)) return jsonSafeValue("Number", "-0");
978
981
  return value;
979
982
  }
980
- function serializeString(value, depth) {
983
+ function serializeString(value, depth, config) {
981
984
  if (depth === 0) return value;
985
+ if (!config.compress) return value;
982
986
  return compressNestedStringValue(value) ?? value;
983
987
  }
984
988
  function isDenseNumberArray(value) {
@@ -1018,8 +1022,9 @@ function compressNestedStringValue(value) {
1018
1022
  };
1019
1023
  return compressionIsWorthIt(serialized, rawSize) ? serialized : void 0;
1020
1024
  }
1021
- function compressNestedJsonValue(value, depth) {
1025
+ function compressNestedJsonValue(value, depth, config) {
1022
1026
  if (depth === 0) return void 0;
1027
+ if (!config.compress) return void 0;
1023
1028
  const raw = JSON.stringify(value);
1024
1029
  const rawSize = Buffer$1.byteLength(raw);
1025
1030
  if (rawSize < compressedJsonMinBytes) return void 0;
@@ -5077,7 +5082,7 @@ function buildDeclaredColumnDefs(overrides, scores, manualScores) {
5077
5082
  async function toCellValue(value) {
5078
5083
  const fileRef = fileRefSchema.safeParse(value);
5079
5084
  if (fileRef.success) return fileRef.data;
5080
- const serialized = await serializeCacheValue(value, { preserveUndefined: true });
5085
+ const serialized = await serializeCacheValue(value, { compress: false });
5081
5086
  const parsed = jsonCellSchema.safeParse(serialized);
5082
5087
  if (parsed.success) return parsed.data;
5083
5088
  }
package/dist/{runner-ChHgWruW.mjs → runner-BCs5rzej.mjs} RENAMED
@@ -1,5 +1,5 @@
1
- import { n as createRunner } from "./cli-CVwIjcsX.mjs";
2
- import "./src-Bcc2ZHK8.mjs";
1
+ import { n as createRunner } from "./cli-Dg3abrOv.mjs";
2
+ import "./src-DBypR4TV.mjs";
3
3
  //#region ../../apps/server/src/runner.ts
4
4
  let runnerInstance = null;
5
5
  function getRunnerInstance() {
package/dist/{runner-DA_o115w.mjs → runner-znY6PY1M.mjs} RENAMED
@@ -1,2 +1,2 @@
1
- import { n as initRunner, t as getRunnerInstance } from "./runner-ChHgWruW.mjs";
1
+ import { n as initRunner, t as getRunnerInstance } from "./runner-BCs5rzej.mjs";
2
2
  export { getRunnerInstance, initRunner };
package/dist/src-DBypR4TV.mjs ADDED
@@ -0,0 +1,3 @@
1
+ import "./runOrchestration-V1TxX8es.mjs";
2
+ import "./cli-Dg3abrOv.mjs";
3
+ export {};
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@ls-stack/agent-eval",
3
- "version": "0.35.0",
3
+ "version": "0.35.1",
4
4
  "type": "module",
5
5
  "bin": {
6
6
  "agent-evals": "./dist/bin.mjs"
@@ -32,8 +32,8 @@
32
32
  "@types/node": "^24.7.2",
33
33
  "typescript": "^5.9.2",
34
34
  "@agent-evals/runner": "0.0.1",
35
- "@agent-evals/shared": "0.0.1",
36
- "@agent-evals/sdk": "0.0.1"
35
+ "@agent-evals/sdk": "0.0.1",
36
+ "@agent-evals/shared": "0.0.1"
37
37
  },
38
38
  "scripts": {
39
39
  "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",
package/dist/src-Bcc2ZHK8.mjs DELETED
@@ -1,3 +0,0 @@
1
- import "./runOrchestration-DoslE_Oo.mjs";
2
- import "./cli-CVwIjcsX.mjs";
3
- export {};