@ls-stack/agent-eval 0.17.0 → 0.18.0
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.
- package/dist/{app-DTotEBoY.mjs → app-hAlVvT-Q.mjs} +23 -4
- package/dist/apps/web/dist/assets/index-BovjyzD8.css +1 -0
- package/dist/apps/web/dist/assets/index-BxcwUS7V.js +118 -0
- package/dist/apps/web/dist/index.html +2 -2
- package/dist/bin.mjs +1 -1
- package/dist/{cli-CULTt3Xp.mjs → cli-3zANEAhG.mjs} +3 -3
- package/dist/index.d.mts +63 -350
- package/dist/index.mjs +4 -4
- package/dist/runChild.mjs +1 -1
- package/dist/{runOrchestration-D2okEB3I.mjs → runOrchestration-BBg_VUH5.mjs} +416 -1710
- package/dist/{runner-DyM0Gp8G.mjs → runner-DxlahWDo.mjs} +1 -1
- package/dist/{runner-BSXZiQIi.mjs → runner-RmZPRz-h.mjs} +2 -2
- package/dist/src-BC4OrajN.mjs +3 -0
- package/package.json +1 -1
- package/skills/agent-eval/SKILL.md +5 -3
- package/dist/apps/web/dist/assets/index-C5IRkeUz.js +0 -118
- package/dist/apps/web/dist/assets/index-Cn9WoTj5.css +0 -1
- package/dist/src-CNf3xwVw.mjs +0 -3
|
@@ -25,8 +25,8 @@
|
|
|
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-
|
|
29
|
-
<link rel="stylesheet" crossorigin href="/assets/index-
|
|
28
|
+
<script type="module" crossorigin src="/assets/index-BxcwUS7V.js"></script>
|
|
29
|
+
<link rel="stylesheet" crossorigin href="/assets/index-BovjyzD8.css">
|
|
30
30
|
</head>
|
|
31
31
|
<body>
|
|
32
32
|
<div id="root"></div>
|
package/dist/bin.mjs
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import {
|
|
1
|
+
import { An as getEvalRegistry, L as runSummarySchema, M as getEvalDisplayStatus, N as deriveScopedSummaryFromCases, S as createFsCacheStore, X as resolveLlmCallsConfig, Y as resolveApiCallsConfig, _ as loadEvalModule, a as loadPersistedRunSnapshot, b as buildDeclaredColumnDefs, c as persistCaseDetail, d as recomputePersistedCaseStatus, f as runTouchesEval, g as setLatestRunInfoMap, h as getTargetEvalIds, i as getLatestRunInfos, j as getEvalTitle, l as persistRunState, m as buildEvalSummary, n as generateRunId, o as loadPersistedRunSnapshots, p as resolveArtifactPath, r as getLastRunStatuses, s as nextShortIdFromSnapshots, u as recomputeEvalStatusesInRuns, v as parseEvalMetas, x as normalizeScoreDef, y as loadConfig } from "./runOrchestration-BBg_VUH5.mjs";
|
|
2
2
|
import { createHash } from "node:crypto";
|
|
3
3
|
import { mkdir, readFile, rm, writeFile } from "node:fs/promises";
|
|
4
4
|
import { dirname, join, relative, resolve } from "node:path";
|
|
@@ -1029,8 +1029,8 @@ async function commandApp(args) {
|
|
|
1029
1029
|
const { serve } = await import("@hono/node-server");
|
|
1030
1030
|
const bundledWebDist = resolve(currentDir, "apps/web/dist");
|
|
1031
1031
|
if (existsSync(bundledWebDist)) process.env.AGENT_EVALS_WEB_DIST = bundledWebDist;
|
|
1032
|
-
const appModule = await import("./app-
|
|
1033
|
-
const runnerModule = await import("./runner-
|
|
1032
|
+
const appModule = await import("./app-hAlVvT-Q.mjs");
|
|
1033
|
+
const runnerModule = await import("./runner-DxlahWDo.mjs");
|
|
1034
1034
|
if (!isHonoAppModule(appModule)) throw new Error("Server app module is invalid");
|
|
1035
1035
|
if (!isServerRunnerModule(runnerModule)) throw new Error("Server runner module is invalid");
|
|
1036
1036
|
await runnerModule.initRunner();
|
package/dist/index.d.mts
CHANGED
|
@@ -685,6 +685,14 @@ declare const runLogPhaseSchema: z$1.ZodEnum<{
|
|
|
685
685
|
/** Eval runner phase that emitted a captured case log. */
|
|
686
686
|
type RunLogPhase = z$1.infer<typeof runLogPhaseSchema>;
|
|
687
687
|
/** Schema for one persisted log entry captured during a case run. */
|
|
688
|
+
declare const runLogLocationSchema: z$1.ZodObject<{
|
|
689
|
+
file: z$1.ZodString;
|
|
690
|
+
line: z$1.ZodNumber;
|
|
691
|
+
column: z$1.ZodNumber;
|
|
692
|
+
}, z$1.core.$strip>;
|
|
693
|
+
/** Best-effort source location for one captured case log. */
|
|
694
|
+
type RunLogLocation = z$1.infer<typeof runLogLocationSchema>;
|
|
695
|
+
/** Schema for one persisted log entry captured during a case run. */
|
|
688
696
|
declare const runLogEntrySchema: z$1.ZodObject<{
|
|
689
697
|
timestamp: z$1.ZodString;
|
|
690
698
|
level: z$1.ZodEnum<{
|
|
@@ -702,6 +710,11 @@ declare const runLogEntrySchema: z$1.ZodObject<{
|
|
|
702
710
|
message: z$1.ZodString;
|
|
703
711
|
args: z$1.ZodDefault<z$1.ZodArray<z$1.ZodUnknown>>;
|
|
704
712
|
truncated: z$1.ZodDefault<z$1.ZodBoolean>;
|
|
713
|
+
location: z$1.ZodOptional<z$1.ZodObject<{
|
|
714
|
+
file: z$1.ZodString;
|
|
715
|
+
line: z$1.ZodNumber;
|
|
716
|
+
column: z$1.ZodNumber;
|
|
717
|
+
}, z$1.core.$strip>>;
|
|
705
718
|
source: z$1.ZodOptional<z$1.ZodString>;
|
|
706
719
|
}, z$1.core.$strip>;
|
|
707
720
|
/** Persisted log entry captured during a case run. */
|
|
@@ -964,6 +977,11 @@ declare const caseDetailSchema: z$1.ZodObject<{
|
|
|
964
977
|
message: z$1.ZodString;
|
|
965
978
|
args: z$1.ZodDefault<z$1.ZodArray<z$1.ZodUnknown>>;
|
|
966
979
|
truncated: z$1.ZodDefault<z$1.ZodBoolean>;
|
|
980
|
+
location: z$1.ZodOptional<z$1.ZodObject<{
|
|
981
|
+
file: z$1.ZodString;
|
|
982
|
+
line: z$1.ZodNumber;
|
|
983
|
+
column: z$1.ZodNumber;
|
|
984
|
+
}, z$1.core.$strip>>;
|
|
967
985
|
source: z$1.ZodOptional<z$1.ZodString>;
|
|
968
986
|
}, z$1.core.$strip>>>;
|
|
969
987
|
error: z$1.ZodNullable<z$1.ZodObject<{
|
|
@@ -2531,18 +2549,21 @@ type CacheDebugKeyFile = z$1.infer<typeof cacheDebugKeyFileSchema>;
|
|
|
2531
2549
|
//#endregion
|
|
2532
2550
|
//#region ../shared/src/utils/extractCacheHits.d.ts
|
|
2533
2551
|
/**
|
|
2534
|
-
* Single cache
|
|
2535
|
-
*
|
|
2552
|
+
* Single cache activity entry rendered as one row in the case drawer's Cache
|
|
2553
|
+
* tab.
|
|
2536
2554
|
*
|
|
2537
|
-
* `
|
|
2538
|
-
*
|
|
2539
|
-
*
|
|
2540
|
-
* (
|
|
2555
|
+
* `action === 'hit'` rows reused an existing persisted cache entry.
|
|
2556
|
+
* `action === 'added'` rows came from a miss or refresh that wrote a persisted
|
|
2557
|
+
* cache entry during the run. `origin === 'caseRoot'` rows came from
|
|
2558
|
+
* `evalTracer.cache(...)` calls made directly from the case body (no
|
|
2559
|
+
* surrounding `traceSpan`), which would otherwise be invisible.
|
|
2541
2560
|
*/
|
|
2542
|
-
type
|
|
2561
|
+
type CacheActivityEntry = {
|
|
2543
2562
|
id: string;
|
|
2544
2563
|
source: 'span' | 'value';
|
|
2545
2564
|
origin: 'span' | 'caseRoot';
|
|
2565
|
+
action: 'hit' | 'added';
|
|
2566
|
+
status: 'hit' | 'miss' | 'refresh';
|
|
2546
2567
|
name: string;
|
|
2547
2568
|
namespace: string;
|
|
2548
2569
|
key: string;
|
|
@@ -2550,14 +2571,26 @@ type CacheHitEntry = {
|
|
|
2550
2571
|
age: number | undefined;
|
|
2551
2572
|
spanId: string | undefined;
|
|
2552
2573
|
};
|
|
2574
|
+
/** Cache activity row narrowed to cache hits for compatibility helpers. */
|
|
2575
|
+
type CacheHitEntry = CacheActivityEntry & {
|
|
2576
|
+
action: 'hit';
|
|
2577
|
+
status: 'hit';
|
|
2578
|
+
};
|
|
2579
|
+
/**
|
|
2580
|
+
* Collect every cache hit or cache write recorded for a case run.
|
|
2581
|
+
*
|
|
2582
|
+
* Walks `spans` for span-level cache activity (`attributes['cache.status']`)
|
|
2583
|
+
* and per-span value-cache refs (`attributes['cache.refs']`), then appends
|
|
2584
|
+
* spanless value-cache refs persisted on the case scope. Bypasses are skipped
|
|
2585
|
+
* because they do not read or write a persisted cache entry.
|
|
2586
|
+
*/
|
|
2587
|
+
declare function extractCacheEntries(spans: EvalTraceSpan[], caseCacheRefs: TraceCacheRef[]): CacheActivityEntry[];
|
|
2553
2588
|
/**
|
|
2554
2589
|
* Collect every `status === 'hit'` cache event recorded for a case run.
|
|
2555
2590
|
*
|
|
2556
|
-
*
|
|
2557
|
-
*
|
|
2558
|
-
*
|
|
2559
|
-
* statuses (`miss`/`refresh`/`bypass`) are skipped — they remain visible
|
|
2560
|
-
* inline in the Trace tab.
|
|
2591
|
+
* This compatibility helper returns only rows that reused an existing
|
|
2592
|
+
* persisted cache entry. Use `extractCacheEntries(...)` when the UI should
|
|
2593
|
+
* include cache misses and refreshes that wrote entries during the run.
|
|
2561
2594
|
*/
|
|
2562
2595
|
declare function extractCacheHits(spans: EvalTraceSpan[], caseCacheRefs: TraceCacheRef[]): CacheHitEntry[];
|
|
2563
2596
|
//#endregion
|
|
@@ -3077,344 +3110,24 @@ declare function incrementEvalOutput(key: string, delta: number): void;
|
|
|
3077
3110
|
*/
|
|
3078
3111
|
declare function evalAssert(condition: boolean, message: string): void;
|
|
3079
3112
|
//#endregion
|
|
3080
|
-
//#region ../../node_modules/.pnpm/seroval@1.5.2/node_modules/seroval/dist/types/core/constants.d.ts
|
|
3081
|
-
declare const enum SerovalConstant {
|
|
3082
|
-
Null = 0,
|
|
3083
|
-
Undefined = 1,
|
|
3084
|
-
True = 2,
|
|
3085
|
-
False = 3,
|
|
3086
|
-
NegZero = 4,
|
|
3087
|
-
Inf = 5,
|
|
3088
|
-
NegInf = 6,
|
|
3089
|
-
Nan = 7
|
|
3090
|
-
}
|
|
3091
|
-
declare const enum SerovalNodeType {
|
|
3092
|
-
Number = 0,
|
|
3093
|
-
String = 1,
|
|
3094
|
-
Constant = 2,
|
|
3095
|
-
BigInt = 3,
|
|
3096
|
-
IndexedValue = 4,
|
|
3097
|
-
Date = 5,
|
|
3098
|
-
RegExp = 6,
|
|
3099
|
-
Set = 7,
|
|
3100
|
-
Map = 8,
|
|
3101
|
-
Array = 9,
|
|
3102
|
-
Object = 10,
|
|
3103
|
-
NullConstructor = 11,
|
|
3104
|
-
Promise = 12,
|
|
3105
|
-
Error = 13,
|
|
3106
|
-
AggregateError = 14,
|
|
3107
|
-
TypedArray = 15,
|
|
3108
|
-
BigIntTypedArray = 16,
|
|
3109
|
-
WKSymbol = 17,
|
|
3110
|
-
Reference = 18,
|
|
3111
|
-
ArrayBuffer = 19,
|
|
3112
|
-
DataView = 20,
|
|
3113
|
-
Boxed = 21,
|
|
3114
|
-
PromiseConstructor = 22,
|
|
3115
|
-
PromiseSuccess = 23,
|
|
3116
|
-
PromiseFailure = 24,
|
|
3117
|
-
Plugin = 25,
|
|
3118
|
-
SpecialReference = 26,
|
|
3119
|
-
IteratorFactory = 27,
|
|
3120
|
-
IteratorFactoryInstance = 28,
|
|
3121
|
-
AsyncIteratorFactory = 29,
|
|
3122
|
-
AsyncIteratorFactoryInstance = 30,
|
|
3123
|
-
StreamConstructor = 31,
|
|
3124
|
-
StreamNext = 32,
|
|
3125
|
-
StreamThrow = 33,
|
|
3126
|
-
StreamReturn = 34,
|
|
3127
|
-
Sequence = 35
|
|
3128
|
-
}
|
|
3129
|
-
declare const enum SerovalObjectFlags {
|
|
3130
|
-
None = 0,
|
|
3131
|
-
NonExtensible = 1,
|
|
3132
|
-
Sealed = 2,
|
|
3133
|
-
Frozen = 3
|
|
3134
|
-
}
|
|
3135
|
-
declare const enum Symbols {
|
|
3136
|
-
AsyncIterator = 0,
|
|
3137
|
-
HasInstance = 1,
|
|
3138
|
-
IsConcatSpreadable = 2,
|
|
3139
|
-
Iterator = 3,
|
|
3140
|
-
Match = 4,
|
|
3141
|
-
MatchAll = 5,
|
|
3142
|
-
Replace = 6,
|
|
3143
|
-
Search = 7,
|
|
3144
|
-
Species = 8,
|
|
3145
|
-
Split = 9,
|
|
3146
|
-
ToPrimitive = 10,
|
|
3147
|
-
ToStringTag = 11,
|
|
3148
|
-
Unscopables = 12
|
|
3149
|
-
}
|
|
3150
|
-
declare const enum ErrorConstructorTag {
|
|
3151
|
-
Error = 0,
|
|
3152
|
-
EvalError = 1,
|
|
3153
|
-
RangeError = 2,
|
|
3154
|
-
ReferenceError = 3,
|
|
3155
|
-
SyntaxError = 4,
|
|
3156
|
-
TypeError = 5,
|
|
3157
|
-
URIError = 6
|
|
3158
|
-
}
|
|
3159
|
-
//#endregion
|
|
3160
|
-
//#region ../../node_modules/.pnpm/seroval@1.5.2/node_modules/seroval/dist/types/core/special-reference.d.ts
|
|
3161
|
-
declare const enum SpecialReference {
|
|
3162
|
-
MapSentinel = 0,
|
|
3163
|
-
PromiseConstructor = 1,
|
|
3164
|
-
PromiseSuccess = 2,
|
|
3165
|
-
PromiseFailure = 3,
|
|
3166
|
-
StreamConstructor = 4,
|
|
3167
|
-
ArrayBufferConstructor = 5
|
|
3168
|
-
}
|
|
3169
|
-
//#endregion
|
|
3170
|
-
//#region ../../node_modules/.pnpm/seroval@1.5.2/node_modules/seroval/dist/types/core/types.d.ts
|
|
3171
|
-
interface SerovalBaseNode {
|
|
3172
|
-
t: SerovalNodeType;
|
|
3173
|
-
i: number | undefined;
|
|
3174
|
-
s: unknown;
|
|
3175
|
-
c: string | undefined;
|
|
3176
|
-
m: string | undefined;
|
|
3177
|
-
p: SerovalObjectRecordNode | undefined;
|
|
3178
|
-
e: SerovalMapRecordNode | undefined;
|
|
3179
|
-
a: (SerovalNode | 0)[] | undefined;
|
|
3180
|
-
f: SerovalNode | undefined;
|
|
3181
|
-
b: number | undefined;
|
|
3182
|
-
o: SerovalObjectFlags | undefined;
|
|
3183
|
-
l: number | undefined;
|
|
3184
|
-
}
|
|
3185
|
-
type SerovalObjectRecordKey = string | SerovalNode;
|
|
3186
|
-
interface SerovalObjectRecordNode {
|
|
3187
|
-
k: SerovalObjectRecordKey[];
|
|
3188
|
-
v: SerovalNode[];
|
|
3189
|
-
}
|
|
3190
|
-
interface SerovalMapRecordNode {
|
|
3191
|
-
k: SerovalNode[];
|
|
3192
|
-
v: SerovalNode[];
|
|
3193
|
-
}
|
|
3194
|
-
interface SerovalNumberNode extends SerovalBaseNode {
|
|
3195
|
-
t: SerovalNodeType.Number;
|
|
3196
|
-
s: number;
|
|
3197
|
-
}
|
|
3198
|
-
interface SerovalStringNode extends SerovalBaseNode {
|
|
3199
|
-
t: SerovalNodeType.String;
|
|
3200
|
-
s: string;
|
|
3201
|
-
}
|
|
3202
|
-
interface SerovalConstantNode extends SerovalBaseNode {
|
|
3203
|
-
t: SerovalNodeType.Constant;
|
|
3204
|
-
s: SerovalConstant;
|
|
3205
|
-
}
|
|
3206
|
-
type SerovalPrimitiveNode = SerovalNumberNode | SerovalStringNode | SerovalConstantNode;
|
|
3207
|
-
interface SerovalIndexedValueNode extends SerovalBaseNode {
|
|
3208
|
-
t: SerovalNodeType.IndexedValue;
|
|
3209
|
-
i: number;
|
|
3210
|
-
}
|
|
3211
|
-
interface SerovalBigIntNode extends SerovalBaseNode {
|
|
3212
|
-
t: SerovalNodeType.BigInt;
|
|
3213
|
-
s: string;
|
|
3214
|
-
}
|
|
3215
|
-
interface SerovalDateNode extends SerovalBaseNode {
|
|
3216
|
-
t: SerovalNodeType.Date;
|
|
3217
|
-
i: number;
|
|
3218
|
-
s: string;
|
|
3219
|
-
}
|
|
3220
|
-
interface SerovalRegExpNode extends SerovalBaseNode {
|
|
3221
|
-
t: SerovalNodeType.RegExp;
|
|
3222
|
-
i: number;
|
|
3223
|
-
c: string;
|
|
3224
|
-
m: string;
|
|
3225
|
-
}
|
|
3226
|
-
interface SerovalArrayBufferNode extends SerovalBaseNode {
|
|
3227
|
-
t: SerovalNodeType.ArrayBuffer;
|
|
3228
|
-
i: number;
|
|
3229
|
-
s: string;
|
|
3230
|
-
f: SerovalNodeWithID;
|
|
3231
|
-
}
|
|
3232
|
-
interface SerovalTypedArrayNode extends SerovalBaseNode {
|
|
3233
|
-
t: SerovalNodeType.TypedArray;
|
|
3234
|
-
i: number;
|
|
3235
|
-
c: string;
|
|
3236
|
-
f: SerovalNode;
|
|
3237
|
-
b: number;
|
|
3238
|
-
l: number;
|
|
3239
|
-
}
|
|
3240
|
-
interface SerovalBigIntTypedArrayNode extends SerovalBaseNode {
|
|
3241
|
-
t: SerovalNodeType.BigIntTypedArray;
|
|
3242
|
-
i: number;
|
|
3243
|
-
c: string;
|
|
3244
|
-
f: SerovalNode;
|
|
3245
|
-
b: number;
|
|
3246
|
-
l: number;
|
|
3247
|
-
}
|
|
3248
|
-
type SerovalSemiPrimitiveNode = SerovalBigIntNode | SerovalDateNode | SerovalRegExpNode | SerovalTypedArrayNode | SerovalBigIntTypedArrayNode;
|
|
3249
|
-
interface SerovalSetNode extends SerovalBaseNode {
|
|
3250
|
-
t: SerovalNodeType.Set;
|
|
3251
|
-
i: number;
|
|
3252
|
-
a: SerovalNode[];
|
|
3253
|
-
}
|
|
3254
|
-
interface SerovalMapNode extends SerovalBaseNode {
|
|
3255
|
-
t: SerovalNodeType.Map;
|
|
3256
|
-
i: number;
|
|
3257
|
-
e: SerovalMapRecordNode;
|
|
3258
|
-
f: SerovalNodeWithID;
|
|
3259
|
-
}
|
|
3260
|
-
interface SerovalArrayNode extends SerovalBaseNode {
|
|
3261
|
-
t: SerovalNodeType.Array;
|
|
3262
|
-
a: (SerovalNode | 0)[];
|
|
3263
|
-
i: number;
|
|
3264
|
-
o: SerovalObjectFlags;
|
|
3265
|
-
}
|
|
3266
|
-
interface SerovalObjectNode extends SerovalBaseNode {
|
|
3267
|
-
t: SerovalNodeType.Object;
|
|
3268
|
-
p: SerovalObjectRecordNode;
|
|
3269
|
-
i: number;
|
|
3270
|
-
o: SerovalObjectFlags;
|
|
3271
|
-
}
|
|
3272
|
-
interface SerovalNullConstructorNode extends SerovalBaseNode {
|
|
3273
|
-
t: SerovalNodeType.NullConstructor;
|
|
3274
|
-
p: SerovalObjectRecordNode;
|
|
3275
|
-
i: number;
|
|
3276
|
-
o: SerovalObjectFlags;
|
|
3277
|
-
}
|
|
3278
|
-
interface SerovalPromiseNode extends SerovalBaseNode {
|
|
3279
|
-
t: SerovalNodeType.Promise;
|
|
3280
|
-
s: 0 | 1;
|
|
3281
|
-
f: SerovalNode;
|
|
3282
|
-
i: number;
|
|
3283
|
-
}
|
|
3284
|
-
interface SerovalErrorNode extends SerovalBaseNode {
|
|
3285
|
-
t: SerovalNodeType.Error;
|
|
3286
|
-
s: ErrorConstructorTag;
|
|
3287
|
-
m: string;
|
|
3288
|
-
p: SerovalObjectRecordNode | undefined;
|
|
3289
|
-
i: number;
|
|
3290
|
-
}
|
|
3291
|
-
interface SerovalAggregateErrorNode extends SerovalBaseNode {
|
|
3292
|
-
t: SerovalNodeType.AggregateError;
|
|
3293
|
-
i: number;
|
|
3294
|
-
m: string;
|
|
3295
|
-
p: SerovalObjectRecordNode | undefined;
|
|
3296
|
-
}
|
|
3297
|
-
interface SerovalWKSymbolNode extends SerovalBaseNode {
|
|
3298
|
-
t: SerovalNodeType.WKSymbol;
|
|
3299
|
-
i: number;
|
|
3300
|
-
s: Symbols;
|
|
3301
|
-
}
|
|
3302
|
-
interface SerovalReferenceNode extends SerovalBaseNode {
|
|
3303
|
-
t: SerovalNodeType.Reference;
|
|
3304
|
-
i: number;
|
|
3305
|
-
s: string;
|
|
3306
|
-
}
|
|
3307
|
-
interface SerovalDataViewNode extends SerovalBaseNode {
|
|
3308
|
-
t: SerovalNodeType.DataView;
|
|
3309
|
-
i: number;
|
|
3310
|
-
f: SerovalNode;
|
|
3311
|
-
b: number;
|
|
3312
|
-
l: number;
|
|
3313
|
-
}
|
|
3314
|
-
interface SerovalBoxedNode extends SerovalBaseNode {
|
|
3315
|
-
t: SerovalNodeType.Boxed;
|
|
3316
|
-
i: number;
|
|
3317
|
-
f: SerovalNode;
|
|
3318
|
-
}
|
|
3319
|
-
interface SerovalPromiseConstructorNode extends SerovalBaseNode {
|
|
3320
|
-
t: SerovalNodeType.PromiseConstructor;
|
|
3321
|
-
i: number;
|
|
3322
|
-
s: number;
|
|
3323
|
-
f: SerovalNodeWithID;
|
|
3324
|
-
}
|
|
3325
|
-
interface SerovalPromiseResolveNode extends SerovalBaseNode {
|
|
3326
|
-
t: SerovalNodeType.PromiseSuccess;
|
|
3327
|
-
i: number;
|
|
3328
|
-
a: [resolver: SerovalNodeWithID, resolved: SerovalNode];
|
|
3329
|
-
}
|
|
3330
|
-
interface SerovalPromiseRejectNode extends SerovalBaseNode {
|
|
3331
|
-
t: SerovalNodeType.PromiseFailure;
|
|
3332
|
-
i: number;
|
|
3333
|
-
a: [resolver: SerovalNodeWithID, resolved: SerovalNode];
|
|
3334
|
-
}
|
|
3335
|
-
interface SerovalPluginNode extends SerovalBaseNode {
|
|
3336
|
-
t: SerovalNodeType.Plugin;
|
|
3337
|
-
i: number;
|
|
3338
|
-
s: Record<string, SerovalNode>;
|
|
3339
|
-
c: string;
|
|
3340
|
-
}
|
|
3341
|
-
/**
|
|
3342
|
-
* Represents special values as placeholders
|
|
3343
|
-
*/
|
|
3344
|
-
interface SerovalSpecialReferenceNode extends SerovalBaseNode {
|
|
3345
|
-
t: SerovalNodeType.SpecialReference;
|
|
3346
|
-
i: number;
|
|
3347
|
-
s: SpecialReference;
|
|
3348
|
-
}
|
|
3349
|
-
interface SerovalIteratorFactoryNode extends SerovalBaseNode {
|
|
3350
|
-
t: SerovalNodeType.IteratorFactory;
|
|
3351
|
-
i: number;
|
|
3352
|
-
f: SerovalNodeWithID;
|
|
3353
|
-
}
|
|
3354
|
-
interface SerovalIteratorFactoryInstanceNode extends SerovalBaseNode {
|
|
3355
|
-
t: SerovalNodeType.IteratorFactoryInstance;
|
|
3356
|
-
a: [instance: SerovalNodeWithID, sequence: SerovalNodeWithID];
|
|
3357
|
-
}
|
|
3358
|
-
interface SerovalAsyncIteratorFactoryNode extends SerovalBaseNode {
|
|
3359
|
-
t: SerovalNodeType.AsyncIteratorFactory;
|
|
3360
|
-
i: number;
|
|
3361
|
-
a: [promise: SerovalNodeWithID, symbol: SerovalNodeWithID];
|
|
3362
|
-
}
|
|
3363
|
-
interface SerovalAsyncIteratorFactoryInstanceNode extends SerovalBaseNode {
|
|
3364
|
-
t: SerovalNodeType.AsyncIteratorFactoryInstance;
|
|
3365
|
-
a: [instance: SerovalNodeWithID, sequence: SerovalNodeWithID];
|
|
3366
|
-
}
|
|
3367
|
-
interface SerovalStreamConstructorNode extends SerovalBaseNode {
|
|
3368
|
-
t: SerovalNodeType.StreamConstructor;
|
|
3369
|
-
i: number;
|
|
3370
|
-
a: SerovalNode[];
|
|
3371
|
-
f: SerovalNodeWithID;
|
|
3372
|
-
}
|
|
3373
|
-
interface SerovalStreamNextNode extends SerovalBaseNode {
|
|
3374
|
-
t: SerovalNodeType.StreamNext;
|
|
3375
|
-
i: number;
|
|
3376
|
-
f: SerovalNode;
|
|
3377
|
-
}
|
|
3378
|
-
interface SerovalStreamThrowNode extends SerovalBaseNode {
|
|
3379
|
-
t: SerovalNodeType.StreamThrow;
|
|
3380
|
-
i: number;
|
|
3381
|
-
f: SerovalNode;
|
|
3382
|
-
}
|
|
3383
|
-
interface SerovalStreamReturnNode extends SerovalBaseNode {
|
|
3384
|
-
t: SerovalNodeType.StreamReturn;
|
|
3385
|
-
i: number;
|
|
3386
|
-
f: SerovalNode;
|
|
3387
|
-
}
|
|
3388
|
-
interface SerovalSequenceNode extends SerovalBaseNode {
|
|
3389
|
-
t: SerovalNodeType.Sequence;
|
|
3390
|
-
i: number;
|
|
3391
|
-
s: number;
|
|
3392
|
-
a: SerovalNode[];
|
|
3393
|
-
l: number;
|
|
3394
|
-
}
|
|
3395
|
-
type SerovalSyncNode = SerovalPrimitiveNode | SerovalIndexedValueNode | SerovalSemiPrimitiveNode | SerovalSetNode | SerovalMapNode | SerovalArrayNode | SerovalObjectNode | SerovalNullConstructorNode | SerovalPromiseNode | SerovalErrorNode | SerovalAggregateErrorNode | SerovalWKSymbolNode | SerovalReferenceNode | SerovalArrayBufferNode | SerovalDataViewNode | SerovalBoxedNode | SerovalPluginNode | SerovalSpecialReferenceNode | SerovalIteratorFactoryNode | SerovalIteratorFactoryInstanceNode | SerovalAsyncIteratorFactoryNode | SerovalAsyncIteratorFactoryInstanceNode | SerovalSequenceNode;
|
|
3396
|
-
type SerovalAsyncNode = SerovalPromiseNode | SerovalPromiseConstructorNode | SerovalPromiseResolveNode | SerovalPromiseRejectNode | SerovalStreamConstructorNode | SerovalStreamNextNode | SerovalStreamThrowNode | SerovalStreamReturnNode;
|
|
3397
|
-
type SerovalNode = SerovalSyncNode | SerovalAsyncNode;
|
|
3398
|
-
type SerovalNodeWithID = Extract<SerovalNode, {
|
|
3399
|
-
i: number;
|
|
3400
|
-
}>;
|
|
3401
|
-
//#endregion
|
|
3402
|
-
//#region ../../node_modules/.pnpm/seroval@1.5.2/node_modules/seroval/dist/types/core/tree/index.d.ts
|
|
3403
|
-
interface SerovalJSON {
|
|
3404
|
-
t: SerovalNode;
|
|
3405
|
-
f: number;
|
|
3406
|
-
m: number[];
|
|
3407
|
-
}
|
|
3408
|
-
//#endregion
|
|
3409
3113
|
//#region ../sdk/src/cacheSerialization.d.ts
|
|
3410
3114
|
declare const serializedCacheValueMarker = "__agentEvalsCacheSerialization";
|
|
3411
|
-
declare const
|
|
3412
|
-
|
|
3413
|
-
type
|
|
3414
|
-
[serializedCacheValueMarker]: typeof
|
|
3415
|
-
|
|
3115
|
+
declare const jsonSafeCacheValueVersion = "json-safe-v1";
|
|
3116
|
+
type JsonSafeCacheValueType = 'ArrayBuffer' | 'BigInt' | 'Blob' | 'CompressedJson' | 'CompressedString' | 'Date' | 'Error' | 'File' | 'Float64Array' | 'Headers' | 'Map' | 'Number' | 'Object' | 'RegExp' | 'Set' | 'URL' | 'URLSearchParams' | 'Undefined';
|
|
3117
|
+
type JsonSafeSerializedCacheValue = {
|
|
3118
|
+
[serializedCacheValueMarker]: typeof jsonSafeCacheValueVersion;
|
|
3119
|
+
codec?: 'gzip';
|
|
3120
|
+
length?: number;
|
|
3121
|
+
type: JsonSafeCacheValueType;
|
|
3122
|
+
value?: unknown;
|
|
3416
3123
|
};
|
|
3417
|
-
/**
|
|
3124
|
+
/** JSON-safe persisted representation for one rich cached value. */
|
|
3125
|
+
type SerializedCacheValue = JsonSafeSerializedCacheValue;
|
|
3126
|
+
/**
|
|
3127
|
+
* Serialize one cached value while keeping plain JSON as plain JSON.
|
|
3128
|
+
*
|
|
3129
|
+
* Rich runtime values use small tagged wrappers.
|
|
3130
|
+
*/
|
|
3418
3131
|
declare function serializeCacheValue(value: unknown): Promise<unknown>;
|
|
3419
3132
|
/** Revive one cached value, while preserving legacy JSON-round-tripped data. */
|
|
3420
3133
|
declare function deserializeCacheValue(value: unknown): unknown;
|
|
@@ -3689,8 +3402,8 @@ type EvalRunner = {
|
|
|
3689
3402
|
/**
|
|
3690
3403
|
* Return the full persisted cache entry for `namespace` + `key`, including
|
|
3691
3404
|
* its recording and optional raw-key debug metadata. Returns `null` when no
|
|
3692
|
-
* entry matches. Used by the case drawer's Cache
|
|
3693
|
-
*
|
|
3405
|
+
* entry matches. Used by the case drawer's Cache tab to lazily fetch the
|
|
3406
|
+
* cached return value when a row is expanded.
|
|
3694
3407
|
*/
|
|
3695
3408
|
getCacheEntry(namespace: string, key: string): Promise<CacheEntryWithDebugKey | null>;
|
|
3696
3409
|
/**
|
|
@@ -3747,4 +3460,4 @@ declare function createRunner({
|
|
|
3747
3460
|
*/
|
|
3748
3461
|
declare function runCli(argv: string[]): Promise<void>;
|
|
3749
3462
|
//#endregion
|
|
3750
|
-
export { type AgentEvalsConfig, type ApiCallEntry, type ApiCallMetric, type ApiCallMetricFormat, type ApiCallMetricPlacement, type ApiCallMetricValue, type ApiCallsConfigInput, type AssertionFailure, type CacheAdapter, type CacheDebugKeyEntry, type CacheDebugKeyFile, type CacheDebugKeyWrite, type CacheEntry, type CacheEntryWithDebugKey, type CacheFile, type CacheHitEntry, type CacheKeyHashInput, type CacheKeyHashOptions, type CacheListItem, type CacheMode, type CacheOperationType, type CacheRecording, type CacheRecordingFrame, type CacheRecordingOp, type CacheScopeContext, type CacheStatus, type CaptureEvalSpanErrorLevel, type CaptureEvalSpanErrorOptions, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type CreateRunRequest, DEFAULT_API_CALLS_CONFIG, DEFAULT_LLM_CALLS_CONFIG, type DerivedStatus, EvalAssertionError, type EvalCase, type EvalCaseScope, type EvalChartAggregate, type EvalChartAxis, type EvalChartBuiltinMetric, type EvalChartColor, type EvalChartConfig, type EvalChartMetric, type EvalChartTooltipExtra, type EvalChartType, type EvalChartsConfig, type EvalColumnOverride, type EvalColumns, type EvalDefinition, type EvalDeriveContext, type EvalDisplayStatus, type EvalExecuteContext, type EvalFreshnessStatus, type EvalManualScoreDef, type EvalOutputs, type EvalOutputsSchema, type EvalRunner, type EvalRuntimeScope, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalSetOutput, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, type EvalTraceSpan, type EvalTraceSpanError, type EvalTraceSpanWarning, type EvalTraceTree, type FileRef, type JsonCell, type LlmCallEntry, type LlmCallMetric, type LlmCallMetricFormat, type LlmCallMetricPlacement, type LlmCallMetricValue, type LlmCallsConfigInput, type NumberDisplayOptions, type RepoFileRef, type ResolvedApiCallMetric, type ResolvedApiCallsConfig, type ResolvedLlmCallMetric, type ResolvedLlmCallsConfig, type RunArtifactRef, type RunInEvalScopeOptions, type RunLogEntry, type RunLogLevel, type RunLogPhase, type RunLogsConfigInput, type RunManifest, type RunSummary, type ScalarCell, type ScopedCaseSummary, type ScoreTrace, type SerializedCacheSpan, type SerializedCacheValue, type SpanCacheOptions, type SseEnvelope, type SseEventType, type TraceActiveSpan, type TraceAttributeDisplay, type TraceAttributeDisplayFormat, type TraceAttributeDisplayInput, type TraceAttributeDisplayPlacement, type TraceAttributeTransform, type TraceAttributeTransformContext, type TraceCacheInfo, type TraceCacheRef, type TraceDisplayConfig, type TraceDisplayInputConfig, type TraceSpanInfo, type TrialSelectionMode, type UpdateManualScoreRequest, agentEvalsConfigSchema, apiCallMetricFormatSchema, apiCallMetricPlacementSchema, apiCallMetricSchema, apiCallsConfigSchema, appendToEvalOutput, assertionFailureSchema, buildTraceTree, cacheDebugKeyEntrySchema, cacheDebugKeyFileSchema, cacheEntrySchema, cacheEntryWithDebugKeySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheOperationTypeSchema, cacheRecordingOpSchema, cacheRecordingSchema, cacheStatusSchema, captureEvalSpanError, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, deserializeCacheRecording, deserializeCacheValue, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalLog, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, extractApiCalls, extractCacheHits, extractLlmCalls, fileRefSchema, getCurrentScope, getEvalCaseInput, getEvalDisplayStatus, getEvalRegistry, getEvalTitle, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, llmCallMetricFormatSchema, llmCallMetricPlacementSchema, llmCallMetricSchema, llmCallsConfigSchema, mergeEvalOutput, nextEvalId, numberDisplayOptionsSchema, repoFile, repoFileRefSchema, resolveApiCallsConfig, resolveLlmCallsConfig, runArtifactRefSchema, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, runLogEntrySchema, runLogLevelSchema, runLogPhaseSchema, runLogsConfigSchema, runManifestSchema, runSummarySchema, scoreTraceSchema, serializeCacheRecording, serializeCacheValue, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, startEvalBackgroundJob, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceCacheRefSchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanErrorSchema, traceSpanKindSchema, traceSpanSchema, traceSpanWarningSchema, trialSelectionModeSchema, updateManualScoreRequestSchema, z };
|
|
3463
|
+
export { type AgentEvalsConfig, type ApiCallEntry, type ApiCallMetric, type ApiCallMetricFormat, type ApiCallMetricPlacement, type ApiCallMetricValue, type ApiCallsConfigInput, type AssertionFailure, type CacheActivityEntry, type CacheAdapter, type CacheDebugKeyEntry, type CacheDebugKeyFile, type CacheDebugKeyWrite, type CacheEntry, type CacheEntryWithDebugKey, type CacheFile, type CacheHitEntry, type CacheKeyHashInput, type CacheKeyHashOptions, type CacheListItem, type CacheMode, type CacheOperationType, type CacheRecording, type CacheRecordingFrame, type CacheRecordingOp, type CacheScopeContext, type CacheStatus, type CaptureEvalSpanErrorLevel, type CaptureEvalSpanErrorOptions, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type CreateRunRequest, DEFAULT_API_CALLS_CONFIG, DEFAULT_LLM_CALLS_CONFIG, type DerivedStatus, EvalAssertionError, type EvalCase, type EvalCaseScope, type EvalChartAggregate, type EvalChartAxis, type EvalChartBuiltinMetric, type EvalChartColor, type EvalChartConfig, type EvalChartMetric, type EvalChartTooltipExtra, type EvalChartType, type EvalChartsConfig, type EvalColumnOverride, type EvalColumns, type EvalDefinition, type EvalDeriveContext, type EvalDisplayStatus, type EvalExecuteContext, type EvalFreshnessStatus, type EvalManualScoreDef, type EvalOutputs, type EvalOutputsSchema, type EvalRunner, type EvalRuntimeScope, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalSetOutput, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, type EvalTraceSpan, type EvalTraceSpanError, type EvalTraceSpanWarning, type EvalTraceTree, type FileRef, type JsonCell, type LlmCallEntry, type LlmCallMetric, type LlmCallMetricFormat, type LlmCallMetricPlacement, type LlmCallMetricValue, type LlmCallsConfigInput, type NumberDisplayOptions, type RepoFileRef, type ResolvedApiCallMetric, type ResolvedApiCallsConfig, type ResolvedLlmCallMetric, type ResolvedLlmCallsConfig, type RunArtifactRef, type RunInEvalScopeOptions, type RunLogEntry, type RunLogLevel, type RunLogLocation, type RunLogPhase, type RunLogsConfigInput, type RunManifest, type RunSummary, type ScalarCell, type ScopedCaseSummary, type ScoreTrace, type SerializedCacheSpan, type SerializedCacheValue, type SpanCacheOptions, type SseEnvelope, type SseEventType, type TraceActiveSpan, type TraceAttributeDisplay, type TraceAttributeDisplayFormat, type TraceAttributeDisplayInput, type TraceAttributeDisplayPlacement, type TraceAttributeTransform, type TraceAttributeTransformContext, type TraceCacheInfo, type TraceCacheRef, type TraceDisplayConfig, type TraceDisplayInputConfig, type TraceSpanInfo, type TrialSelectionMode, type UpdateManualScoreRequest, agentEvalsConfigSchema, apiCallMetricFormatSchema, apiCallMetricPlacementSchema, apiCallMetricSchema, apiCallsConfigSchema, appendToEvalOutput, assertionFailureSchema, buildTraceTree, cacheDebugKeyEntrySchema, cacheDebugKeyFileSchema, cacheEntrySchema, cacheEntryWithDebugKeySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheOperationTypeSchema, cacheRecordingOpSchema, cacheRecordingSchema, cacheStatusSchema, captureEvalSpanError, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, deserializeCacheRecording, deserializeCacheValue, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalLog, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, fileRefSchema, getCurrentScope, getEvalCaseInput, getEvalDisplayStatus, getEvalRegistry, getEvalTitle, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, llmCallMetricFormatSchema, llmCallMetricPlacementSchema, llmCallMetricSchema, llmCallsConfigSchema, mergeEvalOutput, nextEvalId, numberDisplayOptionsSchema, repoFile, repoFileRefSchema, resolveApiCallsConfig, resolveLlmCallsConfig, runArtifactRefSchema, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, runLogEntrySchema, runLogLevelSchema, runLogLocationSchema, runLogPhaseSchema, runLogsConfigSchema, runManifestSchema, runSummarySchema, scoreTraceSchema, serializeCacheRecording, serializeCacheValue, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, startEvalBackgroundJob, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceCacheRefSchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanErrorSchema, traceSpanKindSchema, traceSpanSchema, traceSpanWarningSchema, trialSelectionModeSchema, updateManualScoreRequestSchema, z };
|
package/dist/index.mjs
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { $ as
|
|
2
|
-
import { n as createRunner, t as runCli } from "./cli-
|
|
3
|
-
import "./src-
|
|
4
|
-
export { DEFAULT_API_CALLS_CONFIG, DEFAULT_LLM_CALLS_CONFIG, EvalAssertionError, agentEvalsConfigSchema, apiCallMetricFormatSchema, apiCallMetricPlacementSchema, apiCallMetricSchema, apiCallsConfigSchema, appendToEvalOutput, assertionFailureSchema, buildTraceTree, cacheDebugKeyEntrySchema, cacheDebugKeyFileSchema, cacheEntrySchema, cacheEntryWithDebugKeySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheOperationTypeSchema, cacheRecordingOpSchema, cacheRecordingSchema, cacheStatusSchema, captureEvalSpanError, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, deserializeCacheRecording, deserializeCacheValue, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalLog, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, extractApiCalls, extractCacheHits, extractLlmCalls, fileRefSchema, getCurrentScope, getEvalCaseInput, getEvalDisplayStatus, getEvalRegistry, getEvalTitle, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, llmCallMetricFormatSchema, llmCallMetricPlacementSchema, llmCallMetricSchema, llmCallsConfigSchema, mergeEvalOutput, nextEvalId, numberDisplayOptionsSchema, repoFile, repoFileRefSchema, resolveApiCallsConfig, resolveLlmCallsConfig, runArtifactRefSchema, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, runLogEntrySchema, runLogLevelSchema, runLogPhaseSchema, runLogsConfigSchema, runManifestSchema, runSummarySchema, scoreTraceSchema, serializeCacheRecording, serializeCacheValue, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, startEvalBackgroundJob, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceCacheRefSchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanErrorSchema, traceSpanKindSchema, traceSpanSchema, traceSpanWarningSchema, trialSelectionModeSchema, updateManualScoreRequestSchema, z };
|
|
1
|
+
import { $ as assertionFailureSchema, $t as runArtifactRefSchema, A as getNestedAttribute, An as getEvalRegistry, At as cacheRecordingSchema, B as agentEvalsConfigSchema, Bt as traceDisplayInputConfigSchema, C as createRunRequestSchema, Cn as runInEvalScope, Ct as cacheEntrySchema, D as extractCacheHits, Dn as startEvalBackgroundJob, Dt as cacheModeSchema, E as extractCacheEntries, En as setScopeCacheContext, Et as cacheListItemSchema, F as deriveStatusFromChildStatuses, Ft as traceAttributeDisplayFormatSchema, G as llmCallMetricFormatSchema, Gt as cellValueSchema, H as apiCallMetricPlacementSchema, Ht as traceSpanKindSchema, I as runManifestSchema, It as traceAttributeDisplayInputSchema, J as llmCallsConfigSchema, Jt as columnKindSchema, K as llmCallMetricPlacementSchema, Kt as columnDefSchema, L as runSummarySchema, Lt as traceAttributeDisplayPlacementSchema, M as getEvalDisplayStatus, Mt as serializedCacheSpanSchema, N as deriveScopedSummaryFromCases, Nt as spanCacheOptionsSchema, O as extractApiCalls, On as repoFile, Ot as cacheOperationTypeSchema, P as deriveStatusFromCaseRows, Pt as traceCacheRefSchema, Q as trialSelectionModeSchema, Qt as repoFileRefSchema, R as DEFAULT_API_CALLS_CONFIG, Rt as traceAttributeDisplaySchema, Sn as runInEvalRuntimeScope, St as cacheDebugKeyFileSchema, T as sseEnvelopeSchema, Tn as setEvalOutput, Tt as cacheFileSchema, U as apiCallMetricSchema, Ut as traceSpanSchema, V as apiCallMetricFormatSchema, Vt as traceSpanErrorSchema, W as apiCallsConfigSchema, Wt as traceSpanWarningSchema, X as resolveLlmCallsConfig, Xt as jsonCellSchema, Y as resolveApiCallsConfig, Yt as fileRefSchema, Z as runLogsConfigSchema, Zt as numberDisplayOptionsSchema, _n as getEvalCaseInput, _t as evalChartMetricSchema, an as hashCacheKey, at as evalStatsConfigSchema, bn as mergeEvalOutput, bt as evalChartsConfigSchema, cn as deserializeCacheValue, ct as runLogLevelSchema, dn as EvalAssertionError, dt as scoreTraceSchema, en as z, et as caseDetailSchema, fn as appendToEvalOutput, ft as evalChartAggregateSchema, gn as getCurrentScope, gt as evalChartConfigSchema, hn as evalLog, ht as evalChartColorSchema, in as evalTracer, it as evalStatItemSchema, j as getEvalTitle, jt as cacheStatusSchema, k as extractLlmCalls, kn as defineEval, kt as cacheRecordingOpSchema, ln as serializeCacheRecording, lt as runLogLocationSchema, mn as evalAssert, mt as evalChartBuiltinMetricSchema, nn as captureEvalSpanError, nt as evalFreshnessStatusSchema, on as hashCacheKeySync, ot as evalSummarySchema, pt as evalChartAxisSchema, q as llmCallMetricSchema, qt as columnFormatSchema, rn as evalSpan, rt as evalStatAggregateSchema, sn as deserializeCacheRecording, st as runLogEntrySchema, tn as buildTraceTree, tt as caseRowSchema, un as serializeCacheValue, ut as runLogPhaseSchema, vn as incrementEvalOutput, vt as evalChartTooltipExtraSchema, w as updateManualScoreRequestSchema, wn as runInExistingEvalScope, wt as cacheEntryWithDebugKeySchema, xn as nextEvalId, xt as cacheDebugKeyEntrySchema, yn as isInEvalScope, yt as evalChartTypeSchema, z as DEFAULT_LLM_CALLS_CONFIG, zt as traceDisplayConfigSchema } from "./runOrchestration-BBg_VUH5.mjs";
|
|
2
|
+
import { n as createRunner, t as runCli } from "./cli-3zANEAhG.mjs";
|
|
3
|
+
import "./src-BC4OrajN.mjs";
|
|
4
|
+
export { DEFAULT_API_CALLS_CONFIG, DEFAULT_LLM_CALLS_CONFIG, EvalAssertionError, agentEvalsConfigSchema, apiCallMetricFormatSchema, apiCallMetricPlacementSchema, apiCallMetricSchema, apiCallsConfigSchema, appendToEvalOutput, assertionFailureSchema, buildTraceTree, cacheDebugKeyEntrySchema, cacheDebugKeyFileSchema, cacheEntrySchema, cacheEntryWithDebugKeySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheOperationTypeSchema, cacheRecordingOpSchema, cacheRecordingSchema, cacheStatusSchema, captureEvalSpanError, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, deserializeCacheRecording, deserializeCacheValue, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalLog, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, fileRefSchema, getCurrentScope, getEvalCaseInput, getEvalDisplayStatus, getEvalRegistry, getEvalTitle, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, llmCallMetricFormatSchema, llmCallMetricPlacementSchema, llmCallMetricSchema, llmCallsConfigSchema, mergeEvalOutput, nextEvalId, numberDisplayOptionsSchema, repoFile, repoFileRefSchema, resolveApiCallsConfig, resolveLlmCallsConfig, runArtifactRefSchema, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, runLogEntrySchema, runLogLevelSchema, runLogLocationSchema, runLogPhaseSchema, runLogsConfigSchema, runManifestSchema, runSummarySchema, scoreTraceSchema, serializeCacheRecording, serializeCacheValue, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, 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 { C as createRunRequestSchema,
|
|
1
|
+
import { C as createRunRequestSchema, I as runManifestSchema, Kt as columnDefSchema, L as runSummarySchema, S as createFsCacheStore, at as evalStatsConfigSchema, bt as evalChartsConfigSchema, pn as configureEvalRunLogs, t as executeRun, v as parseEvalMetas, y as loadConfig } from "./runOrchestration-BBg_VUH5.mjs";
|
|
2
2
|
import { createHash } from "node:crypto";
|
|
3
3
|
import { readFile } from "node:fs/promises";
|
|
4
4
|
import { relative } from "node:path";
|