@ls-stack/agent-eval 0.58.2 → 0.58.3
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-BxD6aHbp.mjs → app-ROCEce9X.mjs} +4 -4
- package/dist/apps/web/dist/assets/{index-BMWBZw_u.js → index-PTikBbhf.js} +65 -65
- package/dist/apps/web/dist/index.html +1 -1
- package/dist/bin.mjs +1 -1
- package/dist/caseChild.mjs +1 -1
- package/dist/{cli-HBwXIJsg.mjs → cli-SP4kEtYL.mjs} +4 -4
- package/dist/index.d.mts +116 -120
- package/dist/index.mjs +3 -3
- package/dist/runChild.mjs +2 -2
- package/dist/{runExecution-pHJ0_TzH.mjs → runExecution-CFw0MQFs.mjs} +5 -16
- package/dist/{runOrchestration-ngVXShH4.mjs → runOrchestration-CxjiQmof.mjs} +1 -1
- package/dist/{runner-BnZMGBla.mjs → runner-BlFQyvN2.mjs} +1 -1
- package/dist/{runner-D_pz2NON.mjs → runner-CY3bgsjU.mjs} +2 -2
- package/dist/{src-AeXGBJ26.mjs → src-7GbQj1sb.mjs} +2 -2
- package/package.json +1 -1
- package/skills/agent-eval/SKILL.md +5 -4
package/dist/index.d.mts
CHANGED
|
@@ -431,10 +431,8 @@ type EvalDeriveFn<TInput = unknown> = (ctx: EvalDeriveContext<TInput>) => Record
|
|
|
431
431
|
type EvalDeriveConfig<TInput = unknown> = EvalDeriveMap<TInput> | EvalDeriveFn<TInput>;
|
|
432
432
|
/** Function that records trace-derived assertions for one case. */
|
|
433
433
|
type EvalTracingAssertionsFn<TInput = unknown> = (ctx: EvalDeriveContext<TInput>) => MaybePromise$1<void>;
|
|
434
|
-
/** Keyed trace-derived assertion config for grouping related checks. */
|
|
435
|
-
type EvalTracingAssertionsMap<TInput = unknown> = Record<string, EvalTracingAssertionsFn<TInput>>;
|
|
436
434
|
/** Trace-derived assertion config accepted globally and on eval definitions. */
|
|
437
|
-
type EvalTracingAssertionsConfig<TInput = unknown> =
|
|
435
|
+
type EvalTracingAssertionsConfig<TInput = unknown> = EvalTracingAssertionsFn<TInput>;
|
|
438
436
|
/** UI overrides for a derived or scored column emitted by an eval. */
|
|
439
437
|
type EvalColumnOverride = {
|
|
440
438
|
/** Display label shown for the column in tables and detail views. */label?: string;
|
|
@@ -2044,9 +2042,9 @@ declare const traceAttributeDisplaySchema: z$1.ZodObject<{
|
|
|
2044
2042
|
subtree: "subtree";
|
|
2045
2043
|
}>>;
|
|
2046
2044
|
mode: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2047
|
-
sum: "sum";
|
|
2048
2045
|
all: "all";
|
|
2049
2046
|
last: "last";
|
|
2047
|
+
sum: "sum";
|
|
2050
2048
|
}>>;
|
|
2051
2049
|
}, z$1.core.$strip>;
|
|
2052
2050
|
/**
|
|
@@ -2080,9 +2078,9 @@ declare const traceDisplayConfigSchema: z$1.ZodObject<{
|
|
|
2080
2078
|
subtree: "subtree";
|
|
2081
2079
|
}>>;
|
|
2082
2080
|
mode: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2083
|
-
sum: "sum";
|
|
2084
2081
|
all: "all";
|
|
2085
2082
|
last: "last";
|
|
2083
|
+
sum: "sum";
|
|
2086
2084
|
}>>;
|
|
2087
2085
|
}, z$1.core.$strip>>>;
|
|
2088
2086
|
}, z$1.core.$strip>;
|
|
@@ -2120,9 +2118,9 @@ declare const traceAttributeDisplayInputSchema: z$1.ZodObject<{
|
|
|
2120
2118
|
subtree: "subtree";
|
|
2121
2119
|
}>>;
|
|
2122
2120
|
mode: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2123
|
-
sum: "sum";
|
|
2124
2121
|
all: "all";
|
|
2125
2122
|
last: "last";
|
|
2123
|
+
sum: "sum";
|
|
2126
2124
|
}>>;
|
|
2127
2125
|
transform: z$1.ZodOptional<z$1.ZodCustom<TraceAttributeTransform, TraceAttributeTransform>>;
|
|
2128
2126
|
}, z$1.core.$strip>;
|
|
@@ -2158,9 +2156,9 @@ declare const traceDisplayInputConfigSchema: z$1.ZodObject<{
|
|
|
2158
2156
|
subtree: "subtree";
|
|
2159
2157
|
}>>;
|
|
2160
2158
|
mode: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2161
|
-
sum: "sum";
|
|
2162
2159
|
all: "all";
|
|
2163
2160
|
last: "last";
|
|
2161
|
+
sum: "sum";
|
|
2164
2162
|
}>>;
|
|
2165
2163
|
transform: z$1.ZodOptional<z$1.ZodCustom<TraceAttributeTransform, TraceAttributeTransform>>;
|
|
2166
2164
|
}, z$1.core.$strip>>>;
|
|
@@ -2197,8 +2195,8 @@ declare const traceSpanSchema$1: z$1.ZodObject<{
|
|
|
2197
2195
|
status: z$1.ZodEnum<{
|
|
2198
2196
|
error: "error";
|
|
2199
2197
|
running: "running";
|
|
2200
|
-
cancelled: "cancelled";
|
|
2201
2198
|
ok: "ok";
|
|
2199
|
+
cancelled: "cancelled";
|
|
2202
2200
|
}>;
|
|
2203
2201
|
attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
|
|
2204
2202
|
error: z$1.ZodOptional<z$1.ZodObject<{
|
|
@@ -2243,10 +2241,10 @@ type EvalFreshnessStatus = z$1.infer<typeof evalFreshnessStatusSchema>;
|
|
|
2243
2241
|
* `best` selects the highest finite value and `worst` selects the lowest.
|
|
2244
2242
|
*/
|
|
2245
2243
|
declare const evalStatAggregateSchema: z$1.ZodEnum<{
|
|
2246
|
-
|
|
2244
|
+
sum: "sum";
|
|
2247
2245
|
min: "min";
|
|
2248
2246
|
max: "max";
|
|
2249
|
-
|
|
2247
|
+
avg: "avg";
|
|
2250
2248
|
best: "best";
|
|
2251
2249
|
worst: "worst";
|
|
2252
2250
|
}>;
|
|
@@ -2275,10 +2273,10 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
|
2275
2273
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2276
2274
|
kind: z$1.ZodLiteral<"duration">;
|
|
2277
2275
|
aggregate: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2278
|
-
|
|
2276
|
+
sum: "sum";
|
|
2279
2277
|
min: "min";
|
|
2280
2278
|
max: "max";
|
|
2281
|
-
|
|
2279
|
+
avg: "avg";
|
|
2282
2280
|
best: "best";
|
|
2283
2281
|
worst: "worst";
|
|
2284
2282
|
}>>;
|
|
@@ -2286,10 +2284,10 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
|
2286
2284
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2287
2285
|
kind: z$1.ZodLiteral<"cacheHits">;
|
|
2288
2286
|
aggregate: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2289
|
-
|
|
2287
|
+
sum: "sum";
|
|
2290
2288
|
min: "min";
|
|
2291
2289
|
max: "max";
|
|
2292
|
-
|
|
2290
|
+
avg: "avg";
|
|
2293
2291
|
best: "best";
|
|
2294
2292
|
worst: "worst";
|
|
2295
2293
|
}>>;
|
|
@@ -2299,10 +2297,10 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
|
2299
2297
|
key: z$1.ZodString;
|
|
2300
2298
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2301
2299
|
aggregate: z$1.ZodEnum<{
|
|
2302
|
-
|
|
2300
|
+
sum: "sum";
|
|
2303
2301
|
min: "min";
|
|
2304
2302
|
max: "max";
|
|
2305
|
-
|
|
2303
|
+
avg: "avg";
|
|
2306
2304
|
best: "best";
|
|
2307
2305
|
worst: "worst";
|
|
2308
2306
|
}>;
|
|
@@ -2339,10 +2337,10 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
|
|
|
2339
2337
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2340
2338
|
kind: z$1.ZodLiteral<"duration">;
|
|
2341
2339
|
aggregate: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2342
|
-
|
|
2340
|
+
sum: "sum";
|
|
2343
2341
|
min: "min";
|
|
2344
2342
|
max: "max";
|
|
2345
|
-
|
|
2343
|
+
avg: "avg";
|
|
2346
2344
|
best: "best";
|
|
2347
2345
|
worst: "worst";
|
|
2348
2346
|
}>>;
|
|
@@ -2350,10 +2348,10 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
|
|
|
2350
2348
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2351
2349
|
kind: z$1.ZodLiteral<"cacheHits">;
|
|
2352
2350
|
aggregate: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2353
|
-
|
|
2351
|
+
sum: "sum";
|
|
2354
2352
|
min: "min";
|
|
2355
2353
|
max: "max";
|
|
2356
|
-
|
|
2354
|
+
avg: "avg";
|
|
2357
2355
|
best: "best";
|
|
2358
2356
|
worst: "worst";
|
|
2359
2357
|
}>>;
|
|
@@ -2363,10 +2361,10 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
|
|
|
2363
2361
|
key: z$1.ZodString;
|
|
2364
2362
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2365
2363
|
aggregate: z$1.ZodEnum<{
|
|
2366
|
-
|
|
2364
|
+
sum: "sum";
|
|
2367
2365
|
min: "min";
|
|
2368
2366
|
max: "max";
|
|
2369
|
-
|
|
2367
|
+
avg: "avg";
|
|
2370
2368
|
best: "best";
|
|
2371
2369
|
worst: "worst";
|
|
2372
2370
|
}>;
|
|
@@ -2449,10 +2447,10 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2449
2447
|
caseIds: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
2450
2448
|
lastRunStatus: z$1.ZodNullable<z$1.ZodEnum<{
|
|
2451
2449
|
error: "error";
|
|
2452
|
-
pass: "pass";
|
|
2453
|
-
fail: "fail";
|
|
2454
2450
|
running: "running";
|
|
2455
2451
|
cancelled: "cancelled";
|
|
2452
|
+
pass: "pass";
|
|
2453
|
+
fail: "fail";
|
|
2456
2454
|
unscored: "unscored";
|
|
2457
2455
|
}>>;
|
|
2458
2456
|
stats: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
@@ -2466,10 +2464,10 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2466
2464
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2467
2465
|
kind: z$1.ZodLiteral<"duration">;
|
|
2468
2466
|
aggregate: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2469
|
-
|
|
2467
|
+
sum: "sum";
|
|
2470
2468
|
min: "min";
|
|
2471
2469
|
max: "max";
|
|
2472
|
-
|
|
2470
|
+
avg: "avg";
|
|
2473
2471
|
best: "best";
|
|
2474
2472
|
worst: "worst";
|
|
2475
2473
|
}>>;
|
|
@@ -2477,10 +2475,10 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2477
2475
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2478
2476
|
kind: z$1.ZodLiteral<"cacheHits">;
|
|
2479
2477
|
aggregate: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2480
|
-
|
|
2478
|
+
sum: "sum";
|
|
2481
2479
|
min: "min";
|
|
2482
2480
|
max: "max";
|
|
2483
|
-
|
|
2481
|
+
avg: "avg";
|
|
2484
2482
|
best: "best";
|
|
2485
2483
|
worst: "worst";
|
|
2486
2484
|
}>>;
|
|
@@ -2490,10 +2488,10 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2490
2488
|
key: z$1.ZodString;
|
|
2491
2489
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2492
2490
|
aggregate: z$1.ZodEnum<{
|
|
2493
|
-
|
|
2491
|
+
sum: "sum";
|
|
2494
2492
|
min: "min";
|
|
2495
2493
|
max: "max";
|
|
2496
|
-
|
|
2494
|
+
avg: "avg";
|
|
2497
2495
|
best: "best";
|
|
2498
2496
|
worst: "worst";
|
|
2499
2497
|
}>;
|
|
@@ -2517,10 +2515,10 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2517
2515
|
accent: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2518
2516
|
}, z$1.core.$strip>], "kind">>>;
|
|
2519
2517
|
defaultStatAggregate: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2520
|
-
|
|
2518
|
+
sum: "sum";
|
|
2521
2519
|
min: "min";
|
|
2522
2520
|
max: "max";
|
|
2523
|
-
|
|
2521
|
+
avg: "avg";
|
|
2524
2522
|
best: "best";
|
|
2525
2523
|
worst: "worst";
|
|
2526
2524
|
}>>;
|
|
@@ -2541,11 +2539,11 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2541
2539
|
}>;
|
|
2542
2540
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2543
2541
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2544
|
-
success: "success";
|
|
2545
2542
|
error: "error";
|
|
2543
|
+
success: "success";
|
|
2544
|
+
warning: "warning";
|
|
2546
2545
|
accent: "accent";
|
|
2547
2546
|
accentDim: "accentDim";
|
|
2548
|
-
warning: "warning";
|
|
2549
2547
|
textMuted: "textMuted";
|
|
2550
2548
|
}>>;
|
|
2551
2549
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -2556,20 +2554,20 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2556
2554
|
source: z$1.ZodLiteral<"column">;
|
|
2557
2555
|
key: z$1.ZodString;
|
|
2558
2556
|
aggregate: z$1.ZodEnum<{
|
|
2559
|
-
|
|
2557
|
+
sum: "sum";
|
|
2560
2558
|
min: "min";
|
|
2561
2559
|
max: "max";
|
|
2562
|
-
|
|
2560
|
+
avg: "avg";
|
|
2563
2561
|
latest: "latest";
|
|
2564
2562
|
passThresholdRate: "passThresholdRate";
|
|
2565
2563
|
}>;
|
|
2566
2564
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2567
2565
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2568
|
-
success: "success";
|
|
2569
2566
|
error: "error";
|
|
2567
|
+
success: "success";
|
|
2568
|
+
warning: "warning";
|
|
2570
2569
|
accent: "accent";
|
|
2571
2570
|
accentDim: "accentDim";
|
|
2572
|
-
warning: "warning";
|
|
2573
2571
|
textMuted: "textMuted";
|
|
2574
2572
|
}>>;
|
|
2575
2573
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -2598,10 +2596,10 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2598
2596
|
source: z$1.ZodLiteral<"column">;
|
|
2599
2597
|
key: z$1.ZodString;
|
|
2600
2598
|
aggregate: z$1.ZodEnum<{
|
|
2601
|
-
|
|
2599
|
+
sum: "sum";
|
|
2602
2600
|
min: "min";
|
|
2603
2601
|
max: "max";
|
|
2604
|
-
|
|
2602
|
+
avg: "avg";
|
|
2605
2603
|
latest: "latest";
|
|
2606
2604
|
passThresholdRate: "passThresholdRate";
|
|
2607
2605
|
}>;
|
|
@@ -2698,11 +2696,11 @@ declare const caseRowSchema$1: z$1.ZodObject<{
|
|
|
2698
2696
|
tags: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
2699
2697
|
status: z$1.ZodEnum<{
|
|
2700
2698
|
error: "error";
|
|
2701
|
-
pass: "pass";
|
|
2702
|
-
fail: "fail";
|
|
2703
2699
|
running: "running";
|
|
2704
2700
|
cancelled: "cancelled";
|
|
2705
2701
|
pending: "pending";
|
|
2702
|
+
pass: "pass";
|
|
2703
|
+
fail: "fail";
|
|
2706
2704
|
}>;
|
|
2707
2705
|
durationMs: z$1.ZodNullable<z$1.ZodNumber>;
|
|
2708
2706
|
cacheHits: z$1.ZodOptional<z$1.ZodNumber>;
|
|
@@ -2840,8 +2838,8 @@ declare const scoreTraceSchema: z$1.ZodObject<{
|
|
|
2840
2838
|
status: z$1.ZodEnum<{
|
|
2841
2839
|
error: "error";
|
|
2842
2840
|
running: "running";
|
|
2843
|
-
cancelled: "cancelled";
|
|
2844
2841
|
ok: "ok";
|
|
2842
|
+
cancelled: "cancelled";
|
|
2845
2843
|
}>;
|
|
2846
2844
|
attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
|
|
2847
2845
|
error: z$1.ZodOptional<z$1.ZodObject<{
|
|
@@ -2891,9 +2889,9 @@ declare const scoreTraceSchema: z$1.ZodObject<{
|
|
|
2891
2889
|
subtree: "subtree";
|
|
2892
2890
|
}>>;
|
|
2893
2891
|
mode: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2894
|
-
sum: "sum";
|
|
2895
2892
|
all: "all";
|
|
2896
2893
|
last: "last";
|
|
2894
|
+
sum: "sum";
|
|
2897
2895
|
}>>;
|
|
2898
2896
|
}, z$1.core.$strip>>>;
|
|
2899
2897
|
}, z$1.core.$strip>;
|
|
@@ -2903,10 +2901,10 @@ declare const scoreTraceSchema: z$1.ZodObject<{
|
|
|
2903
2901
|
namespace: z$1.ZodString;
|
|
2904
2902
|
key: z$1.ZodString;
|
|
2905
2903
|
status: z$1.ZodEnum<{
|
|
2904
|
+
bypass: "bypass";
|
|
2905
|
+
refresh: "refresh";
|
|
2906
2906
|
hit: "hit";
|
|
2907
2907
|
miss: "miss";
|
|
2908
|
-
refresh: "refresh";
|
|
2909
|
-
bypass: "bypass";
|
|
2910
2908
|
}>;
|
|
2911
2909
|
read: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2912
2910
|
stored: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
@@ -2925,11 +2923,11 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
|
|
|
2925
2923
|
tags: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
2926
2924
|
status: z$1.ZodEnum<{
|
|
2927
2925
|
error: "error";
|
|
2928
|
-
pass: "pass";
|
|
2929
|
-
fail: "fail";
|
|
2930
2926
|
running: "running";
|
|
2931
2927
|
cancelled: "cancelled";
|
|
2932
2928
|
pending: "pending";
|
|
2929
|
+
pass: "pass";
|
|
2930
|
+
fail: "fail";
|
|
2933
2931
|
}>;
|
|
2934
2932
|
input: z$1.ZodUnknown;
|
|
2935
2933
|
trace: z$1.ZodArray<z$1.ZodObject<{
|
|
@@ -2943,8 +2941,8 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
|
|
|
2943
2941
|
status: z$1.ZodEnum<{
|
|
2944
2942
|
error: "error";
|
|
2945
2943
|
running: "running";
|
|
2946
|
-
cancelled: "cancelled";
|
|
2947
2944
|
ok: "ok";
|
|
2945
|
+
cancelled: "cancelled";
|
|
2948
2946
|
}>;
|
|
2949
2947
|
attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
|
|
2950
2948
|
error: z$1.ZodOptional<z$1.ZodObject<{
|
|
@@ -2994,9 +2992,9 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
|
|
|
2994
2992
|
subtree: "subtree";
|
|
2995
2993
|
}>>;
|
|
2996
2994
|
mode: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2997
|
-
sum: "sum";
|
|
2998
2995
|
all: "all";
|
|
2999
2996
|
last: "last";
|
|
2997
|
+
sum: "sum";
|
|
3000
2998
|
}>>;
|
|
3001
2999
|
}, z$1.core.$strip>>>;
|
|
3002
3000
|
}, z$1.core.$strip>;
|
|
@@ -3012,8 +3010,8 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
|
|
|
3012
3010
|
status: z$1.ZodEnum<{
|
|
3013
3011
|
error: "error";
|
|
3014
3012
|
running: "running";
|
|
3015
|
-
cancelled: "cancelled";
|
|
3016
3013
|
ok: "ok";
|
|
3014
|
+
cancelled: "cancelled";
|
|
3017
3015
|
}>;
|
|
3018
3016
|
attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
|
|
3019
3017
|
error: z$1.ZodOptional<z$1.ZodObject<{
|
|
@@ -3063,9 +3061,9 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
|
|
|
3063
3061
|
subtree: "subtree";
|
|
3064
3062
|
}>>;
|
|
3065
3063
|
mode: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3066
|
-
sum: "sum";
|
|
3067
3064
|
all: "all";
|
|
3068
3065
|
last: "last";
|
|
3066
|
+
sum: "sum";
|
|
3069
3067
|
}>>;
|
|
3070
3068
|
}, z$1.core.$strip>>>;
|
|
3071
3069
|
}, z$1.core.$strip>;
|
|
@@ -3075,10 +3073,10 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
|
|
|
3075
3073
|
namespace: z$1.ZodString;
|
|
3076
3074
|
key: z$1.ZodString;
|
|
3077
3075
|
status: z$1.ZodEnum<{
|
|
3076
|
+
bypass: "bypass";
|
|
3077
|
+
refresh: "refresh";
|
|
3078
3078
|
hit: "hit";
|
|
3079
3079
|
miss: "miss";
|
|
3080
|
-
refresh: "refresh";
|
|
3081
|
-
bypass: "bypass";
|
|
3082
3080
|
}>;
|
|
3083
3081
|
read: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
3084
3082
|
stored: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
@@ -3196,10 +3194,10 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
|
|
|
3196
3194
|
namespace: z$1.ZodString;
|
|
3197
3195
|
key: z$1.ZodString;
|
|
3198
3196
|
status: z$1.ZodEnum<{
|
|
3197
|
+
bypass: "bypass";
|
|
3198
|
+
refresh: "refresh";
|
|
3199
3199
|
hit: "hit";
|
|
3200
3200
|
miss: "miss";
|
|
3201
|
-
refresh: "refresh";
|
|
3202
|
-
bypass: "bypass";
|
|
3203
3201
|
}>;
|
|
3204
3202
|
read: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
3205
3203
|
stored: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
@@ -3252,10 +3250,10 @@ declare const evalChartBuiltinMetricSchema: z$1.ZodEnum<{
|
|
|
3252
3250
|
type EvalChartBuiltinMetric = z$1.infer<typeof evalChartBuiltinMetricSchema>;
|
|
3253
3251
|
/** Reducer applied to a numeric column across all cases of a single run. */
|
|
3254
3252
|
declare const evalChartAggregateSchema: z$1.ZodEnum<{
|
|
3255
|
-
|
|
3253
|
+
sum: "sum";
|
|
3256
3254
|
min: "min";
|
|
3257
3255
|
max: "max";
|
|
3258
|
-
|
|
3256
|
+
avg: "avg";
|
|
3259
3257
|
latest: "latest";
|
|
3260
3258
|
passThresholdRate: "passThresholdRate";
|
|
3261
3259
|
}>;
|
|
@@ -3266,11 +3264,11 @@ type EvalChartAggregate = z$1.infer<typeof evalChartAggregateSchema>;
|
|
|
3266
3264
|
* not emit raw hex so authored evals stay decoupled from the web theme.
|
|
3267
3265
|
*/
|
|
3268
3266
|
declare const evalChartColorSchema: z$1.ZodEnum<{
|
|
3269
|
-
success: "success";
|
|
3270
3267
|
error: "error";
|
|
3268
|
+
success: "success";
|
|
3269
|
+
warning: "warning";
|
|
3271
3270
|
accent: "accent";
|
|
3272
3271
|
accentDim: "accentDim";
|
|
3273
|
-
warning: "warning";
|
|
3274
3272
|
textMuted: "textMuted";
|
|
3275
3273
|
}>;
|
|
3276
3274
|
/** Semantic color token resolved to a theme color by the web UI. */
|
|
@@ -3295,11 +3293,11 @@ declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
|
3295
3293
|
}>;
|
|
3296
3294
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3297
3295
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3298
|
-
success: "success";
|
|
3299
3296
|
error: "error";
|
|
3297
|
+
success: "success";
|
|
3298
|
+
warning: "warning";
|
|
3300
3299
|
accent: "accent";
|
|
3301
3300
|
accentDim: "accentDim";
|
|
3302
|
-
warning: "warning";
|
|
3303
3301
|
textMuted: "textMuted";
|
|
3304
3302
|
}>>;
|
|
3305
3303
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -3310,20 +3308,20 @@ declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
|
3310
3308
|
source: z$1.ZodLiteral<"column">;
|
|
3311
3309
|
key: z$1.ZodString;
|
|
3312
3310
|
aggregate: z$1.ZodEnum<{
|
|
3313
|
-
|
|
3311
|
+
sum: "sum";
|
|
3314
3312
|
min: "min";
|
|
3315
3313
|
max: "max";
|
|
3316
|
-
|
|
3314
|
+
avg: "avg";
|
|
3317
3315
|
latest: "latest";
|
|
3318
3316
|
passThresholdRate: "passThresholdRate";
|
|
3319
3317
|
}>;
|
|
3320
3318
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3321
3319
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3322
|
-
success: "success";
|
|
3323
3320
|
error: "error";
|
|
3321
|
+
success: "success";
|
|
3322
|
+
warning: "warning";
|
|
3324
3323
|
accent: "accent";
|
|
3325
3324
|
accentDim: "accentDim";
|
|
3326
|
-
warning: "warning";
|
|
3327
3325
|
textMuted: "textMuted";
|
|
3328
3326
|
}>>;
|
|
3329
3327
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -3345,10 +3343,10 @@ declare const evalChartTooltipExtraSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObj
|
|
|
3345
3343
|
source: z$1.ZodLiteral<"column">;
|
|
3346
3344
|
key: z$1.ZodString;
|
|
3347
3345
|
aggregate: z$1.ZodEnum<{
|
|
3348
|
-
|
|
3346
|
+
sum: "sum";
|
|
3349
3347
|
min: "min";
|
|
3350
3348
|
max: "max";
|
|
3351
|
-
|
|
3349
|
+
avg: "avg";
|
|
3352
3350
|
latest: "latest";
|
|
3353
3351
|
passThresholdRate: "passThresholdRate";
|
|
3354
3352
|
}>;
|
|
@@ -3378,11 +3376,11 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
|
|
|
3378
3376
|
}>;
|
|
3379
3377
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3380
3378
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3381
|
-
success: "success";
|
|
3382
3379
|
error: "error";
|
|
3380
|
+
success: "success";
|
|
3381
|
+
warning: "warning";
|
|
3383
3382
|
accent: "accent";
|
|
3384
3383
|
accentDim: "accentDim";
|
|
3385
|
-
warning: "warning";
|
|
3386
3384
|
textMuted: "textMuted";
|
|
3387
3385
|
}>>;
|
|
3388
3386
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -3393,20 +3391,20 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
|
|
|
3393
3391
|
source: z$1.ZodLiteral<"column">;
|
|
3394
3392
|
key: z$1.ZodString;
|
|
3395
3393
|
aggregate: z$1.ZodEnum<{
|
|
3396
|
-
|
|
3394
|
+
sum: "sum";
|
|
3397
3395
|
min: "min";
|
|
3398
3396
|
max: "max";
|
|
3399
|
-
|
|
3397
|
+
avg: "avg";
|
|
3400
3398
|
latest: "latest";
|
|
3401
3399
|
passThresholdRate: "passThresholdRate";
|
|
3402
3400
|
}>;
|
|
3403
3401
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3404
3402
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3405
|
-
success: "success";
|
|
3406
3403
|
error: "error";
|
|
3404
|
+
success: "success";
|
|
3405
|
+
warning: "warning";
|
|
3407
3406
|
accent: "accent";
|
|
3408
3407
|
accentDim: "accentDim";
|
|
3409
|
-
warning: "warning";
|
|
3410
3408
|
textMuted: "textMuted";
|
|
3411
3409
|
}>>;
|
|
3412
3410
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -3435,10 +3433,10 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
|
|
|
3435
3433
|
source: z$1.ZodLiteral<"column">;
|
|
3436
3434
|
key: z$1.ZodString;
|
|
3437
3435
|
aggregate: z$1.ZodEnum<{
|
|
3438
|
-
|
|
3436
|
+
sum: "sum";
|
|
3439
3437
|
min: "min";
|
|
3440
3438
|
max: "max";
|
|
3441
|
-
|
|
3439
|
+
avg: "avg";
|
|
3442
3440
|
latest: "latest";
|
|
3443
3441
|
passThresholdRate: "passThresholdRate";
|
|
3444
3442
|
}>;
|
|
@@ -3468,11 +3466,11 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
|
|
|
3468
3466
|
}>;
|
|
3469
3467
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3470
3468
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3471
|
-
success: "success";
|
|
3472
3469
|
error: "error";
|
|
3470
|
+
success: "success";
|
|
3471
|
+
warning: "warning";
|
|
3473
3472
|
accent: "accent";
|
|
3474
3473
|
accentDim: "accentDim";
|
|
3475
|
-
warning: "warning";
|
|
3476
3474
|
textMuted: "textMuted";
|
|
3477
3475
|
}>>;
|
|
3478
3476
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -3483,20 +3481,20 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
|
|
|
3483
3481
|
source: z$1.ZodLiteral<"column">;
|
|
3484
3482
|
key: z$1.ZodString;
|
|
3485
3483
|
aggregate: z$1.ZodEnum<{
|
|
3486
|
-
|
|
3484
|
+
sum: "sum";
|
|
3487
3485
|
min: "min";
|
|
3488
3486
|
max: "max";
|
|
3489
|
-
|
|
3487
|
+
avg: "avg";
|
|
3490
3488
|
latest: "latest";
|
|
3491
3489
|
passThresholdRate: "passThresholdRate";
|
|
3492
3490
|
}>;
|
|
3493
3491
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3494
3492
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3495
|
-
success: "success";
|
|
3496
3493
|
error: "error";
|
|
3494
|
+
success: "success";
|
|
3495
|
+
warning: "warning";
|
|
3497
3496
|
accent: "accent";
|
|
3498
3497
|
accentDim: "accentDim";
|
|
3499
|
-
warning: "warning";
|
|
3500
3498
|
textMuted: "textMuted";
|
|
3501
3499
|
}>>;
|
|
3502
3500
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -3525,10 +3523,10 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
|
|
|
3525
3523
|
source: z$1.ZodLiteral<"column">;
|
|
3526
3524
|
key: z$1.ZodString;
|
|
3527
3525
|
aggregate: z$1.ZodEnum<{
|
|
3528
|
-
|
|
3526
|
+
sum: "sum";
|
|
3529
3527
|
min: "min";
|
|
3530
3528
|
max: "max";
|
|
3531
|
-
|
|
3529
|
+
avg: "avg";
|
|
3532
3530
|
latest: "latest";
|
|
3533
3531
|
passThresholdRate: "passThresholdRate";
|
|
3534
3532
|
}>;
|
|
@@ -3556,9 +3554,9 @@ declare const runManifestSchema$1: z$1.ZodObject<{
|
|
|
3556
3554
|
evalSourceFingerprints: z$1.ZodDefault<z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodString>>>;
|
|
3557
3555
|
target: z$1.ZodObject<{
|
|
3558
3556
|
mode: z$1.ZodEnum<{
|
|
3559
|
-
caseIds: "caseIds";
|
|
3560
3557
|
all: "all";
|
|
3561
3558
|
evalIds: "evalIds";
|
|
3559
|
+
caseIds: "caseIds";
|
|
3562
3560
|
}>;
|
|
3563
3561
|
evalKeys: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
3564
3562
|
files: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
@@ -3572,9 +3570,9 @@ declare const runManifestSchema$1: z$1.ZodObject<{
|
|
|
3572
3570
|
median: "median";
|
|
3573
3571
|
}>>>;
|
|
3574
3572
|
cacheMode: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3575
|
-
refresh: "refresh";
|
|
3576
|
-
bypass: "bypass";
|
|
3577
3573
|
use: "use";
|
|
3574
|
+
bypass: "bypass";
|
|
3575
|
+
refresh: "refresh";
|
|
3578
3576
|
}>>;
|
|
3579
3577
|
}, z$1.core.$strip>;
|
|
3580
3578
|
/** Persisted lifecycle metadata for a single eval run. */
|
|
@@ -3723,11 +3721,9 @@ type EvalDeriveConfig$1<TInput = unknown> = EvalDeriveMap$1<TInput> | EvalDerive
|
|
|
3723
3721
|
/** Schema for keyed or object-returning trace-derived output config. */
|
|
3724
3722
|
/** Function that records trace-derived assertions for one case. */
|
|
3725
3723
|
type EvalTracingAssertionsFn$1<TInput = unknown> = (ctx: EvalDeriveContext$1<TInput>) => MaybePromise<void>;
|
|
3726
|
-
/** Keyed trace-derived assertion config for grouping related checks. */
|
|
3727
|
-
type EvalTracingAssertionsMap$1<TInput = unknown> = Record<string, EvalTracingAssertionsFn$1<TInput>>;
|
|
3728
3724
|
/** Trace-derived assertion config accepted globally and on eval definitions. */
|
|
3729
|
-
type EvalTracingAssertionsConfig$1<TInput = unknown> =
|
|
3730
|
-
/** Schema for
|
|
3725
|
+
type EvalTracingAssertionsConfig$1<TInput = unknown> = EvalTracingAssertionsFn$1<TInput>;
|
|
3726
|
+
/** Schema for trace-derived assertion config. */
|
|
3731
3727
|
/** UI overrides for a derived or scored column emitted by an eval. */
|
|
3732
3728
|
type EvalColumnOverride$1 = {
|
|
3733
3729
|
/** Display label shown for the column in tables and detail views. */label?: string;
|
|
@@ -4523,9 +4519,9 @@ declare function extractApiCalls(spans: EvalTraceSpan$1[], config: ResolvedApiCa
|
|
|
4523
4519
|
* - `refresh`: never read, always write (forces re-execution and overwrites).
|
|
4524
4520
|
*/
|
|
4525
4521
|
declare const cacheModeSchema: z$1.ZodEnum<{
|
|
4526
|
-
refresh: "refresh";
|
|
4527
|
-
bypass: "bypass";
|
|
4528
4522
|
use: "use";
|
|
4523
|
+
bypass: "bypass";
|
|
4524
|
+
refresh: "refresh";
|
|
4529
4525
|
}>;
|
|
4530
4526
|
/** Mode controlling how cached spans behave during a run. */
|
|
4531
4527
|
type CacheMode = z$1.infer<typeof cacheModeSchema>;
|
|
@@ -4539,17 +4535,17 @@ declare const spanCacheOptionsSchema: z$1.ZodObject<{
|
|
|
4539
4535
|
type SpanCacheOptions = z$1.infer<typeof spanCacheOptionsSchema>;
|
|
4540
4536
|
/** Category of operation stored in the eval cache. */
|
|
4541
4537
|
declare const cacheOperationTypeSchema: z$1.ZodEnum<{
|
|
4542
|
-
value: "value";
|
|
4543
4538
|
span: "span";
|
|
4539
|
+
value: "value";
|
|
4544
4540
|
}>;
|
|
4545
4541
|
/** Category of operation stored in the eval cache. */
|
|
4546
4542
|
type CacheOperationType = z$1.infer<typeof cacheOperationTypeSchema>;
|
|
4547
4543
|
/** Status of a cache lookup recorded on a span or case scope. */
|
|
4548
4544
|
declare const cacheStatusSchema: z$1.ZodEnum<{
|
|
4545
|
+
bypass: "bypass";
|
|
4546
|
+
refresh: "refresh";
|
|
4549
4547
|
hit: "hit";
|
|
4550
4548
|
miss: "miss";
|
|
4551
|
-
refresh: "refresh";
|
|
4552
|
-
bypass: "bypass";
|
|
4553
4549
|
}>;
|
|
4554
4550
|
/** Status of a cache lookup recorded on a span or case scope. */
|
|
4555
4551
|
type CacheStatus = z$1.infer<typeof cacheStatusSchema>;
|
|
@@ -4566,10 +4562,10 @@ declare const traceCacheRefSchema: z$1.ZodObject<{
|
|
|
4566
4562
|
namespace: z$1.ZodString;
|
|
4567
4563
|
key: z$1.ZodString;
|
|
4568
4564
|
status: z$1.ZodEnum<{
|
|
4565
|
+
bypass: "bypass";
|
|
4566
|
+
refresh: "refresh";
|
|
4569
4567
|
hit: "hit";
|
|
4570
4568
|
miss: "miss";
|
|
4571
|
-
refresh: "refresh";
|
|
4572
|
-
bypass: "bypass";
|
|
4573
4569
|
}>;
|
|
4574
4570
|
read: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
4575
4571
|
stored: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
@@ -4677,8 +4673,8 @@ declare const cacheRecordingSchema: z$1.ZodObject<{
|
|
|
4677
4673
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4678
4674
|
error: "error";
|
|
4679
4675
|
running: "running";
|
|
4680
|
-
cancelled: "cancelled";
|
|
4681
4676
|
ok: "ok";
|
|
4677
|
+
cancelled: "cancelled";
|
|
4682
4678
|
}>>;
|
|
4683
4679
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
4684
4680
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4765,8 +4761,8 @@ declare const cacheEntrySchema: z$1.ZodObject<{
|
|
|
4765
4761
|
key: z$1.ZodString;
|
|
4766
4762
|
namespace: z$1.ZodString;
|
|
4767
4763
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4768
|
-
value: "value";
|
|
4769
4764
|
span: "span";
|
|
4765
|
+
value: "value";
|
|
4770
4766
|
}>>;
|
|
4771
4767
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4772
4768
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4778,8 +4774,8 @@ declare const cacheEntrySchema: z$1.ZodObject<{
|
|
|
4778
4774
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4779
4775
|
error: "error";
|
|
4780
4776
|
running: "running";
|
|
4781
|
-
cancelled: "cancelled";
|
|
4782
4777
|
ok: "ok";
|
|
4778
|
+
cancelled: "cancelled";
|
|
4783
4779
|
}>>;
|
|
4784
4780
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
4785
4781
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4872,8 +4868,8 @@ declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
|
|
|
4872
4868
|
key: z$1.ZodString;
|
|
4873
4869
|
namespace: z$1.ZodString;
|
|
4874
4870
|
operationType: z$1.ZodEnum<{
|
|
4875
|
-
value: "value";
|
|
4876
4871
|
span: "span";
|
|
4872
|
+
value: "value";
|
|
4877
4873
|
}>;
|
|
4878
4874
|
operationName: z$1.ZodString;
|
|
4879
4875
|
storedAt: z$1.ZodString;
|
|
@@ -4883,8 +4879,8 @@ declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
|
|
|
4883
4879
|
key: z$1.ZodString;
|
|
4884
4880
|
namespace: z$1.ZodString;
|
|
4885
4881
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4886
|
-
value: "value";
|
|
4887
4882
|
span: "span";
|
|
4883
|
+
value: "value";
|
|
4888
4884
|
}>>;
|
|
4889
4885
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4890
4886
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4896,8 +4892,8 @@ declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
|
|
|
4896
4892
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4897
4893
|
error: "error";
|
|
4898
4894
|
running: "running";
|
|
4899
|
-
cancelled: "cancelled";
|
|
4900
4895
|
ok: "ok";
|
|
4896
|
+
cancelled: "cancelled";
|
|
4901
4897
|
}>>;
|
|
4902
4898
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
4903
4899
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4990,8 +4986,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
4990
4986
|
key: z$1.ZodString;
|
|
4991
4987
|
namespace: z$1.ZodString;
|
|
4992
4988
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4993
|
-
value: "value";
|
|
4994
4989
|
span: "span";
|
|
4990
|
+
value: "value";
|
|
4995
4991
|
}>>;
|
|
4996
4992
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4997
4993
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5003,8 +4999,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
5003
4999
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
5004
5000
|
error: "error";
|
|
5005
5001
|
running: "running";
|
|
5006
|
-
cancelled: "cancelled";
|
|
5007
5002
|
ok: "ok";
|
|
5003
|
+
cancelled: "cancelled";
|
|
5008
5004
|
}>>;
|
|
5009
5005
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
5010
5006
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5088,8 +5084,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
5088
5084
|
key: z$1.ZodString;
|
|
5089
5085
|
namespace: z$1.ZodString;
|
|
5090
5086
|
operationType: z$1.ZodEnum<{
|
|
5091
|
-
value: "value";
|
|
5092
5087
|
span: "span";
|
|
5088
|
+
value: "value";
|
|
5093
5089
|
}>;
|
|
5094
5090
|
operationName: z$1.ZodString;
|
|
5095
5091
|
storedAt: z$1.ZodString;
|
|
@@ -5099,8 +5095,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
5099
5095
|
key: z$1.ZodString;
|
|
5100
5096
|
namespace: z$1.ZodString;
|
|
5101
5097
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
5102
|
-
value: "value";
|
|
5103
5098
|
span: "span";
|
|
5099
|
+
value: "value";
|
|
5104
5100
|
}>>;
|
|
5105
5101
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
5106
5102
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5112,8 +5108,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
5112
5108
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
5113
5109
|
error: "error";
|
|
5114
5110
|
running: "running";
|
|
5115
|
-
cancelled: "cancelled";
|
|
5116
5111
|
ok: "ok";
|
|
5112
|
+
cancelled: "cancelled";
|
|
5117
5113
|
}>>;
|
|
5118
5114
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
5119
5115
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5206,8 +5202,8 @@ declare const cacheFileSchema: z$1.ZodObject<{
|
|
|
5206
5202
|
key: z$1.ZodString;
|
|
5207
5203
|
namespace: z$1.ZodString;
|
|
5208
5204
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
5209
|
-
value: "value";
|
|
5210
5205
|
span: "span";
|
|
5206
|
+
value: "value";
|
|
5211
5207
|
}>>;
|
|
5212
5208
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
5213
5209
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5219,8 +5215,8 @@ declare const cacheFileSchema: z$1.ZodObject<{
|
|
|
5219
5215
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
5220
5216
|
error: "error";
|
|
5221
5217
|
running: "running";
|
|
5222
|
-
cancelled: "cancelled";
|
|
5223
5218
|
ok: "ok";
|
|
5219
|
+
cancelled: "cancelled";
|
|
5224
5220
|
}>>;
|
|
5225
5221
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
5226
5222
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5312,8 +5308,8 @@ declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
|
|
|
5312
5308
|
key: z$1.ZodString;
|
|
5313
5309
|
namespace: z$1.ZodString;
|
|
5314
5310
|
operationType: z$1.ZodEnum<{
|
|
5315
|
-
value: "value";
|
|
5316
5311
|
span: "span";
|
|
5312
|
+
value: "value";
|
|
5317
5313
|
}>;
|
|
5318
5314
|
operationName: z$1.ZodString;
|
|
5319
5315
|
storedAt: z$1.ZodString;
|
|
@@ -5323,8 +5319,8 @@ declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
|
|
|
5323
5319
|
key: z$1.ZodString;
|
|
5324
5320
|
namespace: z$1.ZodString;
|
|
5325
5321
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
5326
|
-
value: "value";
|
|
5327
5322
|
span: "span";
|
|
5323
|
+
value: "value";
|
|
5328
5324
|
}>>;
|
|
5329
5325
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
5330
5326
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5336,8 +5332,8 @@ declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
|
|
|
5336
5332
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
5337
5333
|
error: "error";
|
|
5338
5334
|
running: "running";
|
|
5339
|
-
cancelled: "cancelled";
|
|
5340
5335
|
ok: "ok";
|
|
5336
|
+
cancelled: "cancelled";
|
|
5341
5337
|
}>>;
|
|
5342
5338
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
5343
5339
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5498,8 +5494,8 @@ type SseEnvelope = z$1.infer<typeof sseEnvelopeSchema$1>; //#endregion
|
|
|
5498
5494
|
//#region src/schemas/api.d.ts
|
|
5499
5495
|
/** Lifecycle state for an app config reload triggered by `agent-evals.config.ts`. */
|
|
5500
5496
|
declare const configReloadStatusSchema: z$1.ZodEnum<{
|
|
5501
|
-
pending: "pending";
|
|
5502
5497
|
idle: "idle";
|
|
5498
|
+
pending: "pending";
|
|
5503
5499
|
reloading: "reloading";
|
|
5504
5500
|
}>;
|
|
5505
5501
|
/** Status for config reloads in the long-running app server. */
|
|
@@ -5507,8 +5503,8 @@ type ConfigReloadStatus = z$1.infer<typeof configReloadStatusSchema>;
|
|
|
5507
5503
|
/** UI/API-visible state for config reloads in `agent-evals app`. */
|
|
5508
5504
|
declare const configReloadStateSchema$1: z$1.ZodObject<{
|
|
5509
5505
|
status: z$1.ZodEnum<{
|
|
5510
|
-
pending: "pending";
|
|
5511
5506
|
idle: "idle";
|
|
5507
|
+
pending: "pending";
|
|
5512
5508
|
reloading: "reloading";
|
|
5513
5509
|
}>;
|
|
5514
5510
|
activeRunCount: z$1.ZodNumber;
|
|
@@ -5521,9 +5517,9 @@ type ConfigReloadState = z$1.infer<typeof configReloadStateSchema$1>;
|
|
|
5521
5517
|
declare const createRunRequestSchema$1: z$1.ZodObject<{
|
|
5522
5518
|
target: z$1.ZodObject<{
|
|
5523
5519
|
mode: z$1.ZodEnum<{
|
|
5524
|
-
caseIds: "caseIds";
|
|
5525
5520
|
all: "all";
|
|
5526
5521
|
evalIds: "evalIds";
|
|
5522
|
+
caseIds: "caseIds";
|
|
5527
5523
|
}>;
|
|
5528
5524
|
evalKeys: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
5529
5525
|
files: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
@@ -5535,9 +5531,9 @@ declare const createRunRequestSchema$1: z$1.ZodObject<{
|
|
|
5535
5531
|
temporary: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
5536
5532
|
cache: z$1.ZodOptional<z$1.ZodObject<{
|
|
5537
5533
|
mode: z$1.ZodDefault<z$1.ZodEnum<{
|
|
5538
|
-
refresh: "refresh";
|
|
5539
|
-
bypass: "bypass";
|
|
5540
5534
|
use: "use";
|
|
5535
|
+
bypass: "bypass";
|
|
5536
|
+
refresh: "refresh";
|
|
5541
5537
|
}>>;
|
|
5542
5538
|
}, z$1.core.$strip>>;
|
|
5543
5539
|
manualInputs: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
|
|
@@ -7286,4 +7282,4 @@ declare function defineEval<TInput = unknown, TOutputs extends EvalOutputs = Eva
|
|
|
7286
7282
|
/** Return whether the active eval case has tags matching the typed input. */
|
|
7287
7283
|
declare function matchesEvalTags(input: EvalTagMatchInput): boolean;
|
|
7288
7284
|
//#endregion
|
|
7289
|
-
export { AgentEvalTagRegistry, 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 CacheRepairSummary, type CacheScopeContext, type CacheSerializationOptions, type CacheStatus, type CallDerivedAttribute, type CallDerivedAttributeContext, type CallDerivedAttributesConfig, type CallDerivedAttributesFn, type CaptureEvalSpanErrorLevel, type CaptureEvalSpanErrorOptions, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type ConfigReloadState, type ConfigReloadStatus, type CreateRunRequest, type DefaultConfigKey, type DerivedStatus, type DiscoveryIssue, EvalAssertionError, type EvalCacheConfig, 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, EvalDefinition, type EvalDeriveConfig, type EvalDeriveContext, type EvalDeriveFn, type EvalDeriveMap, type EvalDeriveValueFn, type EvalDisplayStatus, type EvalExecuteContext, type EvalExpectation, type EvalFreshnessStatus, type EvalManualInputConfig, type EvalManualScoreDef, type EvalOutputs, type EvalOutputsSchema, type EvalRunner, type EvalRuntimeScope, EvalRuntimeUsageError, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalSetOutput, type EvalStartTime, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, EvalTag, EvalTagMatchInput, type EvalTraceTree, type EvalTracingAssertionsConfig, type EvalTracingAssertionsFn, type
|
|
7285
|
+
export { AgentEvalTagRegistry, 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 CacheRepairSummary, type CacheScopeContext, type CacheSerializationOptions, type CacheStatus, type CallDerivedAttribute, type CallDerivedAttributeContext, type CallDerivedAttributesConfig, type CallDerivedAttributesFn, type CaptureEvalSpanErrorLevel, type CaptureEvalSpanErrorOptions, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type ConfigReloadState, type ConfigReloadStatus, type CreateRunRequest, type DefaultConfigKey, type DerivedStatus, type DiscoveryIssue, EvalAssertionError, type EvalCacheConfig, 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, EvalDefinition, type EvalDeriveConfig, type EvalDeriveContext, type EvalDeriveFn, type EvalDeriveMap, type EvalDeriveValueFn, type EvalDisplayStatus, type EvalExecuteContext, type EvalExpectation, type EvalFreshnessStatus, type EvalManualInputConfig, type EvalManualScoreDef, type EvalOutputs, type EvalOutputsSchema, type EvalRunner, type EvalRuntimeScope, EvalRuntimeUsageError, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalSetOutput, type EvalStartTime, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, EvalTag, EvalTagMatchInput, type EvalTraceTree, type EvalTracingAssertionsConfig, type EvalTracingAssertionsFn, type JsonCell, type LlmCallCostBreakdown, type LlmCallCostCurrency, type LlmCallEntry, type LlmCallMetric, type LlmCallMetricFormat, type LlmCallMetricPlacement, type LlmCallMetricValue, type LlmCallPricing, type LlmCallPricingRate, type LlmCallPricingRegistry, type LlmCallSimulatedTokens, type LlmCallsConfigInput, type LlmCostScenario, type ManualInputDescriptor, type ManualInputFieldDescriptor, type ManualInputFieldKind, type ManualInputFieldOverride, type ManualInputFieldsConfig, type ManualInputFileValue, type ManualInputSelectOption, type MaterializeManualInputFilesResult, type NumberDisplayOptions, type ReadManualInputFileResult, type RemoveDefaultConfig, type ResolvedApiCallMetric, type ResolvedApiCallsConfig, type ResolvedCallDerivedAttribute, type ResolvedLlmCallCostCurrency, type ResolvedLlmCallMetric, type ResolvedLlmCallPricing, type ResolvedLlmCallsConfig, 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, 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 };
|