@ls-stack/agent-eval 0.58.2 → 0.58.4
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-sGeXC4AT.mjs} +63 -11
- package/dist/apps/web/dist/assets/index-BXFsxHVc.js +377 -0
- package/dist/apps/web/dist/index.html +1 -1
- package/dist/bin.mjs +1 -1
- package/dist/caseChild.mjs +4 -3
- package/dist/{cli-HBwXIJsg.mjs → cli-Bf5RzM8O.mjs} +7 -6
- package/dist/index.d.mts +129 -123
- package/dist/index.mjs +3 -3
- package/dist/runChild.mjs +5 -4
- package/dist/{runExecution-pHJ0_TzH.mjs → runExecution-CLkC-4Z1.mjs} +40 -19
- package/dist/{runOrchestration-ngVXShH4.mjs → runOrchestration-BS-WxTee.mjs} +1 -1
- package/dist/{runner-BnZMGBla.mjs → runner-Bz5ZPqmm.mjs} +1 -1
- package/dist/{runner-D_pz2NON.mjs → runner-DW-11txl.mjs} +2 -2
- package/dist/{src-AeXGBJ26.mjs → src-BjMMDm_O.mjs} +2 -2
- package/package.json +1 -1
- package/skills/agent-eval/SKILL.md +7 -6
- package/dist/apps/web/dist/assets/index-BMWBZw_u.js +0 -377
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
|
-
avg: "avg";
|
|
2247
2244
|
min: "min";
|
|
2248
2245
|
max: "max";
|
|
2249
2246
|
sum: "sum";
|
|
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
|
-
avg: "avg";
|
|
2279
2276
|
min: "min";
|
|
2280
2277
|
max: "max";
|
|
2281
2278
|
sum: "sum";
|
|
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
|
-
avg: "avg";
|
|
2290
2287
|
min: "min";
|
|
2291
2288
|
max: "max";
|
|
2292
2289
|
sum: "sum";
|
|
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
|
-
avg: "avg";
|
|
2303
2300
|
min: "min";
|
|
2304
2301
|
max: "max";
|
|
2305
2302
|
sum: "sum";
|
|
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
|
-
avg: "avg";
|
|
2343
2340
|
min: "min";
|
|
2344
2341
|
max: "max";
|
|
2345
2342
|
sum: "sum";
|
|
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
|
-
avg: "avg";
|
|
2354
2351
|
min: "min";
|
|
2355
2352
|
max: "max";
|
|
2356
2353
|
sum: "sum";
|
|
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
|
-
avg: "avg";
|
|
2367
2364
|
min: "min";
|
|
2368
2365
|
max: "max";
|
|
2369
2366
|
sum: "sum";
|
|
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
|
-
avg: "avg";
|
|
2470
2467
|
min: "min";
|
|
2471
2468
|
max: "max";
|
|
2472
2469
|
sum: "sum";
|
|
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
|
-
avg: "avg";
|
|
2481
2478
|
min: "min";
|
|
2482
2479
|
max: "max";
|
|
2483
2480
|
sum: "sum";
|
|
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
|
-
avg: "avg";
|
|
2494
2491
|
min: "min";
|
|
2495
2492
|
max: "max";
|
|
2496
2493
|
sum: "sum";
|
|
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
|
-
avg: "avg";
|
|
2521
2518
|
min: "min";
|
|
2522
2519
|
max: "max";
|
|
2523
2520
|
sum: "sum";
|
|
2521
|
+
avg: "avg";
|
|
2524
2522
|
best: "best";
|
|
2525
2523
|
worst: "worst";
|
|
2526
2524
|
}>>;
|
|
@@ -2536,16 +2534,16 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2536
2534
|
metrics: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
2537
2535
|
source: z$1.ZodLiteral<"builtin">;
|
|
2538
2536
|
metric: z$1.ZodEnum<{
|
|
2539
|
-
passRate: "passRate";
|
|
2540
2537
|
durationMs: "durationMs";
|
|
2538
|
+
passRate: "passRate";
|
|
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";
|
|
2546
2544
|
accent: "accent";
|
|
2547
|
-
accentDim: "accentDim";
|
|
2548
2545
|
warning: "warning";
|
|
2546
|
+
accentDim: "accentDim";
|
|
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
|
-
avg: "avg";
|
|
2560
2557
|
min: "min";
|
|
2561
2558
|
max: "max";
|
|
2562
2559
|
sum: "sum";
|
|
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";
|
|
2570
2568
|
accent: "accent";
|
|
2571
|
-
accentDim: "accentDim";
|
|
2572
2569
|
warning: "warning";
|
|
2570
|
+
accentDim: "accentDim";
|
|
2573
2571
|
textMuted: "textMuted";
|
|
2574
2572
|
}>>;
|
|
2575
2573
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -2590,18 +2588,18 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2590
2588
|
tooltipExtras: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
2591
2589
|
source: z$1.ZodLiteral<"builtin">;
|
|
2592
2590
|
metric: z$1.ZodEnum<{
|
|
2593
|
-
passRate: "passRate";
|
|
2594
2591
|
durationMs: "durationMs";
|
|
2592
|
+
passRate: "passRate";
|
|
2595
2593
|
}>;
|
|
2596
2594
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2597
2595
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
2598
2596
|
source: z$1.ZodLiteral<"column">;
|
|
2599
2597
|
key: z$1.ZodString;
|
|
2600
2598
|
aggregate: z$1.ZodEnum<{
|
|
2601
|
-
avg: "avg";
|
|
2602
2599
|
min: "min";
|
|
2603
2600
|
max: "max";
|
|
2604
2601
|
sum: "sum";
|
|
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>;
|
|
@@ -2781,9 +2779,9 @@ declare const runLogLevelSchema: z$1.ZodEnum<{
|
|
|
2781
2779
|
type RunLogLevel = z$1.infer<typeof runLogLevelSchema>;
|
|
2782
2780
|
/** Eval runner phase that emitted a captured case log. */
|
|
2783
2781
|
declare const runLogPhaseSchema: z$1.ZodEnum<{
|
|
2782
|
+
tracingAssertions: "tracingAssertions";
|
|
2784
2783
|
eval: "eval";
|
|
2785
2784
|
derive: "derive";
|
|
2786
|
-
tracingAssertions: "tracingAssertions";
|
|
2787
2785
|
outputsSchema: "outputsSchema";
|
|
2788
2786
|
scorer: "scorer";
|
|
2789
2787
|
}>;
|
|
@@ -2808,9 +2806,9 @@ declare const runLogEntrySchema: z$1.ZodObject<{
|
|
|
2808
2806
|
warn: "warn";
|
|
2809
2807
|
}>;
|
|
2810
2808
|
phase: z$1.ZodEnum<{
|
|
2809
|
+
tracingAssertions: "tracingAssertions";
|
|
2811
2810
|
eval: "eval";
|
|
2812
2811
|
derive: "derive";
|
|
2813
|
-
tracingAssertions: "tracingAssertions";
|
|
2814
2812
|
outputsSchema: "outputsSchema";
|
|
2815
2813
|
scorer: "scorer";
|
|
2816
2814
|
}>;
|
|
@@ -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>;
|
|
@@ -3167,9 +3165,9 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
|
|
|
3167
3165
|
warn: "warn";
|
|
3168
3166
|
}>;
|
|
3169
3167
|
phase: z$1.ZodEnum<{
|
|
3168
|
+
tracingAssertions: "tracingAssertions";
|
|
3170
3169
|
eval: "eval";
|
|
3171
3170
|
derive: "derive";
|
|
3172
|
-
tracingAssertions: "tracingAssertions";
|
|
3173
3171
|
outputsSchema: "outputsSchema";
|
|
3174
3172
|
scorer: "scorer";
|
|
3175
3173
|
}>;
|
|
@@ -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>;
|
|
@@ -3242,8 +3240,8 @@ type EvalChartType = z$1.infer<typeof evalChartTypeSchema>;
|
|
|
3242
3240
|
* than from a per-case column.
|
|
3243
3241
|
*/
|
|
3244
3242
|
declare const evalChartBuiltinMetricSchema: z$1.ZodEnum<{
|
|
3245
|
-
passRate: "passRate";
|
|
3246
3243
|
durationMs: "durationMs";
|
|
3244
|
+
passRate: "passRate";
|
|
3247
3245
|
}>;
|
|
3248
3246
|
/**
|
|
3249
3247
|
* Run-level metric sourced from the aggregated `RunSummary` for a run, rather
|
|
@@ -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
|
-
avg: "avg";
|
|
3256
3253
|
min: "min";
|
|
3257
3254
|
max: "max";
|
|
3258
3255
|
sum: "sum";
|
|
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";
|
|
3271
3269
|
accent: "accent";
|
|
3272
|
-
accentDim: "accentDim";
|
|
3273
3270
|
warning: "warning";
|
|
3271
|
+
accentDim: "accentDim";
|
|
3274
3272
|
textMuted: "textMuted";
|
|
3275
3273
|
}>;
|
|
3276
3274
|
/** Semantic color token resolved to a theme color by the web UI. */
|
|
@@ -3290,16 +3288,16 @@ type EvalChartAxis = z$1.infer<typeof evalChartAxisSchema>;
|
|
|
3290
3288
|
declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
3291
3289
|
source: z$1.ZodLiteral<"builtin">;
|
|
3292
3290
|
metric: z$1.ZodEnum<{
|
|
3293
|
-
passRate: "passRate";
|
|
3294
3291
|
durationMs: "durationMs";
|
|
3292
|
+
passRate: "passRate";
|
|
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";
|
|
3300
3298
|
accent: "accent";
|
|
3301
|
-
accentDim: "accentDim";
|
|
3302
3299
|
warning: "warning";
|
|
3300
|
+
accentDim: "accentDim";
|
|
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
|
-
avg: "avg";
|
|
3314
3311
|
min: "min";
|
|
3315
3312
|
max: "max";
|
|
3316
3313
|
sum: "sum";
|
|
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";
|
|
3324
3322
|
accent: "accent";
|
|
3325
|
-
accentDim: "accentDim";
|
|
3326
3323
|
warning: "warning";
|
|
3324
|
+
accentDim: "accentDim";
|
|
3327
3325
|
textMuted: "textMuted";
|
|
3328
3326
|
}>>;
|
|
3329
3327
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -3337,18 +3335,18 @@ type EvalChartMetric = z$1.infer<typeof evalChartMetricSchema>;
|
|
|
3337
3335
|
declare const evalChartTooltipExtraSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
3338
3336
|
source: z$1.ZodLiteral<"builtin">;
|
|
3339
3337
|
metric: z$1.ZodEnum<{
|
|
3340
|
-
passRate: "passRate";
|
|
3341
3338
|
durationMs: "durationMs";
|
|
3339
|
+
passRate: "passRate";
|
|
3342
3340
|
}>;
|
|
3343
3341
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3344
3342
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
3345
3343
|
source: z$1.ZodLiteral<"column">;
|
|
3346
3344
|
key: z$1.ZodString;
|
|
3347
3345
|
aggregate: z$1.ZodEnum<{
|
|
3348
|
-
avg: "avg";
|
|
3349
3346
|
min: "min";
|
|
3350
3347
|
max: "max";
|
|
3351
3348
|
sum: "sum";
|
|
3349
|
+
avg: "avg";
|
|
3352
3350
|
latest: "latest";
|
|
3353
3351
|
passThresholdRate: "passThresholdRate";
|
|
3354
3352
|
}>;
|
|
@@ -3373,16 +3371,16 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
|
|
|
3373
3371
|
metrics: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
3374
3372
|
source: z$1.ZodLiteral<"builtin">;
|
|
3375
3373
|
metric: z$1.ZodEnum<{
|
|
3376
|
-
passRate: "passRate";
|
|
3377
3374
|
durationMs: "durationMs";
|
|
3375
|
+
passRate: "passRate";
|
|
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";
|
|
3383
3381
|
accent: "accent";
|
|
3384
|
-
accentDim: "accentDim";
|
|
3385
3382
|
warning: "warning";
|
|
3383
|
+
accentDim: "accentDim";
|
|
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
|
-
avg: "avg";
|
|
3397
3394
|
min: "min";
|
|
3398
3395
|
max: "max";
|
|
3399
3396
|
sum: "sum";
|
|
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";
|
|
3407
3405
|
accent: "accent";
|
|
3408
|
-
accentDim: "accentDim";
|
|
3409
3406
|
warning: "warning";
|
|
3407
|
+
accentDim: "accentDim";
|
|
3410
3408
|
textMuted: "textMuted";
|
|
3411
3409
|
}>>;
|
|
3412
3410
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -3427,18 +3425,18 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
|
|
|
3427
3425
|
tooltipExtras: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
3428
3426
|
source: z$1.ZodLiteral<"builtin">;
|
|
3429
3427
|
metric: z$1.ZodEnum<{
|
|
3430
|
-
passRate: "passRate";
|
|
3431
3428
|
durationMs: "durationMs";
|
|
3429
|
+
passRate: "passRate";
|
|
3432
3430
|
}>;
|
|
3433
3431
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3434
3432
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
3435
3433
|
source: z$1.ZodLiteral<"column">;
|
|
3436
3434
|
key: z$1.ZodString;
|
|
3437
3435
|
aggregate: z$1.ZodEnum<{
|
|
3438
|
-
avg: "avg";
|
|
3439
3436
|
min: "min";
|
|
3440
3437
|
max: "max";
|
|
3441
3438
|
sum: "sum";
|
|
3439
|
+
avg: "avg";
|
|
3442
3440
|
latest: "latest";
|
|
3443
3441
|
passThresholdRate: "passThresholdRate";
|
|
3444
3442
|
}>;
|
|
@@ -3463,16 +3461,16 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
|
|
|
3463
3461
|
metrics: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
3464
3462
|
source: z$1.ZodLiteral<"builtin">;
|
|
3465
3463
|
metric: z$1.ZodEnum<{
|
|
3466
|
-
passRate: "passRate";
|
|
3467
3464
|
durationMs: "durationMs";
|
|
3465
|
+
passRate: "passRate";
|
|
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";
|
|
3473
3471
|
accent: "accent";
|
|
3474
|
-
accentDim: "accentDim";
|
|
3475
3472
|
warning: "warning";
|
|
3473
|
+
accentDim: "accentDim";
|
|
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
|
-
avg: "avg";
|
|
3487
3484
|
min: "min";
|
|
3488
3485
|
max: "max";
|
|
3489
3486
|
sum: "sum";
|
|
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";
|
|
3497
3495
|
accent: "accent";
|
|
3498
|
-
accentDim: "accentDim";
|
|
3499
3496
|
warning: "warning";
|
|
3497
|
+
accentDim: "accentDim";
|
|
3500
3498
|
textMuted: "textMuted";
|
|
3501
3499
|
}>>;
|
|
3502
3500
|
axis: z$1.ZodOptional<z$1.ZodEnum<{
|
|
@@ -3517,18 +3515,18 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
|
|
|
3517
3515
|
tooltipExtras: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
3518
3516
|
source: z$1.ZodLiteral<"builtin">;
|
|
3519
3517
|
metric: z$1.ZodEnum<{
|
|
3520
|
-
passRate: "passRate";
|
|
3521
3518
|
durationMs: "durationMs";
|
|
3519
|
+
passRate: "passRate";
|
|
3522
3520
|
}>;
|
|
3523
3521
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3524
3522
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
3525
3523
|
source: z$1.ZodLiteral<"column">;
|
|
3526
3524
|
key: z$1.ZodString;
|
|
3527
3525
|
aggregate: z$1.ZodEnum<{
|
|
3528
|
-
avg: "avg";
|
|
3529
3526
|
min: "min";
|
|
3530
3527
|
max: "max";
|
|
3531
3528
|
sum: "sum";
|
|
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. */
|
|
@@ -3670,8 +3668,8 @@ type TrialSelectionMode = z$1.infer<typeof trialSelectionModeSchema>;
|
|
|
3670
3668
|
/** Built-in eval-level output/column keys. */
|
|
3671
3669
|
/** Removal config for built-in eval-level outputs and UI metadata. */
|
|
3672
3670
|
declare const removeDefaultConfigSchema: z$1.ZodUnion<readonly [z$1.ZodLiteral<true>, z$1.ZodArray<z$1.ZodEnum<{
|
|
3673
|
-
costUsd: "costUsd";
|
|
3674
3671
|
apiCalls: "apiCalls";
|
|
3672
|
+
costUsd: "costUsd";
|
|
3675
3673
|
llmTurns: "llmTurns";
|
|
3676
3674
|
inputTokens: "inputTokens";
|
|
3677
3675
|
outputTokens: "outputTokens";
|
|
@@ -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;
|
|
@@ -4294,15 +4290,26 @@ type AgentEvalsConfig$1 = {
|
|
|
4294
4290
|
/** Disable the cache entirely; spans with `cache` options execute as if uncached. */enabled?: boolean; /** Override the directory used to persist cache entries. */
|
|
4295
4291
|
dir?: string;
|
|
4296
4292
|
/**
|
|
4297
|
-
*
|
|
4298
|
-
*
|
|
4293
|
+
* Maximum entries retained per cache namespace.
|
|
4294
|
+
*
|
|
4295
|
+
* Pass a number to set the default cap for every namespace. Pass an object
|
|
4296
|
+
* to set a default cap plus exact namespace-specific caps. Non-positive or
|
|
4297
|
+
* non-finite values fall back to the default.
|
|
4298
|
+
*
|
|
4299
|
+
* @example
|
|
4300
|
+
* ```ts
|
|
4301
|
+
* cache: {
|
|
4302
|
+
* maxEntries: {
|
|
4303
|
+
* default: 50,
|
|
4304
|
+
* namespaces: { 'receipt-audit.receipt-audit-context': 200 },
|
|
4305
|
+
* },
|
|
4306
|
+
* }
|
|
4307
|
+
* ```
|
|
4299
4308
|
*/
|
|
4300
|
-
|
|
4301
|
-
|
|
4302
|
-
|
|
4303
|
-
|
|
4304
|
-
*/
|
|
4305
|
-
maxEntriesByNamespace?: Record<string, number>;
|
|
4309
|
+
maxEntries?: number | {
|
|
4310
|
+
default?: number;
|
|
4311
|
+
namespaces?: Record<string, number>;
|
|
4312
|
+
};
|
|
4306
4313
|
/**
|
|
4307
4314
|
* Milliseconds the runner waits after becoming idle before pruning indexed
|
|
4308
4315
|
* cache entries. Defaults to `5000`; non-positive or non-finite values use
|
|
@@ -4313,8 +4320,7 @@ type AgentEvalsConfig$1 = {
|
|
|
4313
4320
|
* Minimum milliseconds between `lastAccessedAt` index rewrites for repeated
|
|
4314
4321
|
* cache hits. Defaults to four hours. Set to `0` to record every hit.
|
|
4315
4322
|
*/
|
|
4316
|
-
lastAccessedAtUpdateIntervalMs?: number;
|
|
4317
|
-
maxEntriesPerEval?: number;
|
|
4323
|
+
lastAccessedAtUpdateIntervalMs?: number;
|
|
4318
4324
|
};
|
|
4319
4325
|
};
|
|
4320
4326
|
/** Zod schema for validating `agent-evals.config.ts` input. */
|
|
@@ -4523,9 +4529,9 @@ declare function extractApiCalls(spans: EvalTraceSpan$1[], config: ResolvedApiCa
|
|
|
4523
4529
|
* - `refresh`: never read, always write (forces re-execution and overwrites).
|
|
4524
4530
|
*/
|
|
4525
4531
|
declare const cacheModeSchema: z$1.ZodEnum<{
|
|
4526
|
-
refresh: "refresh";
|
|
4527
|
-
bypass: "bypass";
|
|
4528
4532
|
use: "use";
|
|
4533
|
+
bypass: "bypass";
|
|
4534
|
+
refresh: "refresh";
|
|
4529
4535
|
}>;
|
|
4530
4536
|
/** Mode controlling how cached spans behave during a run. */
|
|
4531
4537
|
type CacheMode = z$1.infer<typeof cacheModeSchema>;
|
|
@@ -4539,17 +4545,17 @@ declare const spanCacheOptionsSchema: z$1.ZodObject<{
|
|
|
4539
4545
|
type SpanCacheOptions = z$1.infer<typeof spanCacheOptionsSchema>;
|
|
4540
4546
|
/** Category of operation stored in the eval cache. */
|
|
4541
4547
|
declare const cacheOperationTypeSchema: z$1.ZodEnum<{
|
|
4542
|
-
value: "value";
|
|
4543
4548
|
span: "span";
|
|
4549
|
+
value: "value";
|
|
4544
4550
|
}>;
|
|
4545
4551
|
/** Category of operation stored in the eval cache. */
|
|
4546
4552
|
type CacheOperationType = z$1.infer<typeof cacheOperationTypeSchema>;
|
|
4547
4553
|
/** Status of a cache lookup recorded on a span or case scope. */
|
|
4548
4554
|
declare const cacheStatusSchema: z$1.ZodEnum<{
|
|
4555
|
+
bypass: "bypass";
|
|
4556
|
+
refresh: "refresh";
|
|
4549
4557
|
hit: "hit";
|
|
4550
4558
|
miss: "miss";
|
|
4551
|
-
refresh: "refresh";
|
|
4552
|
-
bypass: "bypass";
|
|
4553
4559
|
}>;
|
|
4554
4560
|
/** Status of a cache lookup recorded on a span or case scope. */
|
|
4555
4561
|
type CacheStatus = z$1.infer<typeof cacheStatusSchema>;
|
|
@@ -4566,10 +4572,10 @@ declare const traceCacheRefSchema: z$1.ZodObject<{
|
|
|
4566
4572
|
namespace: z$1.ZodString;
|
|
4567
4573
|
key: z$1.ZodString;
|
|
4568
4574
|
status: z$1.ZodEnum<{
|
|
4575
|
+
bypass: "bypass";
|
|
4576
|
+
refresh: "refresh";
|
|
4569
4577
|
hit: "hit";
|
|
4570
4578
|
miss: "miss";
|
|
4571
|
-
refresh: "refresh";
|
|
4572
|
-
bypass: "bypass";
|
|
4573
4579
|
}>;
|
|
4574
4580
|
read: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
4575
4581
|
stored: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
@@ -4677,8 +4683,8 @@ declare const cacheRecordingSchema: z$1.ZodObject<{
|
|
|
4677
4683
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4678
4684
|
error: "error";
|
|
4679
4685
|
running: "running";
|
|
4680
|
-
cancelled: "cancelled";
|
|
4681
4686
|
ok: "ok";
|
|
4687
|
+
cancelled: "cancelled";
|
|
4682
4688
|
}>>;
|
|
4683
4689
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
4684
4690
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4765,8 +4771,8 @@ declare const cacheEntrySchema: z$1.ZodObject<{
|
|
|
4765
4771
|
key: z$1.ZodString;
|
|
4766
4772
|
namespace: z$1.ZodString;
|
|
4767
4773
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4768
|
-
value: "value";
|
|
4769
4774
|
span: "span";
|
|
4775
|
+
value: "value";
|
|
4770
4776
|
}>>;
|
|
4771
4777
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4772
4778
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4778,8 +4784,8 @@ declare const cacheEntrySchema: z$1.ZodObject<{
|
|
|
4778
4784
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4779
4785
|
error: "error";
|
|
4780
4786
|
running: "running";
|
|
4781
|
-
cancelled: "cancelled";
|
|
4782
4787
|
ok: "ok";
|
|
4788
|
+
cancelled: "cancelled";
|
|
4783
4789
|
}>>;
|
|
4784
4790
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
4785
4791
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4872,8 +4878,8 @@ declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
|
|
|
4872
4878
|
key: z$1.ZodString;
|
|
4873
4879
|
namespace: z$1.ZodString;
|
|
4874
4880
|
operationType: z$1.ZodEnum<{
|
|
4875
|
-
value: "value";
|
|
4876
4881
|
span: "span";
|
|
4882
|
+
value: "value";
|
|
4877
4883
|
}>;
|
|
4878
4884
|
operationName: z$1.ZodString;
|
|
4879
4885
|
storedAt: z$1.ZodString;
|
|
@@ -4883,8 +4889,8 @@ declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
|
|
|
4883
4889
|
key: z$1.ZodString;
|
|
4884
4890
|
namespace: z$1.ZodString;
|
|
4885
4891
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4886
|
-
value: "value";
|
|
4887
4892
|
span: "span";
|
|
4893
|
+
value: "value";
|
|
4888
4894
|
}>>;
|
|
4889
4895
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4890
4896
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4896,8 +4902,8 @@ declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
|
|
|
4896
4902
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4897
4903
|
error: "error";
|
|
4898
4904
|
running: "running";
|
|
4899
|
-
cancelled: "cancelled";
|
|
4900
4905
|
ok: "ok";
|
|
4906
|
+
cancelled: "cancelled";
|
|
4901
4907
|
}>>;
|
|
4902
4908
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
4903
4909
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4990,8 +4996,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
4990
4996
|
key: z$1.ZodString;
|
|
4991
4997
|
namespace: z$1.ZodString;
|
|
4992
4998
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4993
|
-
value: "value";
|
|
4994
4999
|
span: "span";
|
|
5000
|
+
value: "value";
|
|
4995
5001
|
}>>;
|
|
4996
5002
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4997
5003
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5003,8 +5009,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
5003
5009
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
5004
5010
|
error: "error";
|
|
5005
5011
|
running: "running";
|
|
5006
|
-
cancelled: "cancelled";
|
|
5007
5012
|
ok: "ok";
|
|
5013
|
+
cancelled: "cancelled";
|
|
5008
5014
|
}>>;
|
|
5009
5015
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
5010
5016
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5088,8 +5094,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
5088
5094
|
key: z$1.ZodString;
|
|
5089
5095
|
namespace: z$1.ZodString;
|
|
5090
5096
|
operationType: z$1.ZodEnum<{
|
|
5091
|
-
value: "value";
|
|
5092
5097
|
span: "span";
|
|
5098
|
+
value: "value";
|
|
5093
5099
|
}>;
|
|
5094
5100
|
operationName: z$1.ZodString;
|
|
5095
5101
|
storedAt: z$1.ZodString;
|
|
@@ -5099,8 +5105,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
5099
5105
|
key: z$1.ZodString;
|
|
5100
5106
|
namespace: z$1.ZodString;
|
|
5101
5107
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
5102
|
-
value: "value";
|
|
5103
5108
|
span: "span";
|
|
5109
|
+
value: "value";
|
|
5104
5110
|
}>>;
|
|
5105
5111
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
5106
5112
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5112,8 +5118,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
5112
5118
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
5113
5119
|
error: "error";
|
|
5114
5120
|
running: "running";
|
|
5115
|
-
cancelled: "cancelled";
|
|
5116
5121
|
ok: "ok";
|
|
5122
|
+
cancelled: "cancelled";
|
|
5117
5123
|
}>>;
|
|
5118
5124
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
5119
5125
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5206,8 +5212,8 @@ declare const cacheFileSchema: z$1.ZodObject<{
|
|
|
5206
5212
|
key: z$1.ZodString;
|
|
5207
5213
|
namespace: z$1.ZodString;
|
|
5208
5214
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
5209
|
-
value: "value";
|
|
5210
5215
|
span: "span";
|
|
5216
|
+
value: "value";
|
|
5211
5217
|
}>>;
|
|
5212
5218
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
5213
5219
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5219,8 +5225,8 @@ declare const cacheFileSchema: z$1.ZodObject<{
|
|
|
5219
5225
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
5220
5226
|
error: "error";
|
|
5221
5227
|
running: "running";
|
|
5222
|
-
cancelled: "cancelled";
|
|
5223
5228
|
ok: "ok";
|
|
5229
|
+
cancelled: "cancelled";
|
|
5224
5230
|
}>>;
|
|
5225
5231
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
5226
5232
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5312,8 +5318,8 @@ declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
|
|
|
5312
5318
|
key: z$1.ZodString;
|
|
5313
5319
|
namespace: z$1.ZodString;
|
|
5314
5320
|
operationType: z$1.ZodEnum<{
|
|
5315
|
-
value: "value";
|
|
5316
5321
|
span: "span";
|
|
5322
|
+
value: "value";
|
|
5317
5323
|
}>;
|
|
5318
5324
|
operationName: z$1.ZodString;
|
|
5319
5325
|
storedAt: z$1.ZodString;
|
|
@@ -5323,8 +5329,8 @@ declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
|
|
|
5323
5329
|
key: z$1.ZodString;
|
|
5324
5330
|
namespace: z$1.ZodString;
|
|
5325
5331
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
5326
|
-
value: "value";
|
|
5327
5332
|
span: "span";
|
|
5333
|
+
value: "value";
|
|
5328
5334
|
}>>;
|
|
5329
5335
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
5330
5336
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5336,8 +5342,8 @@ declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
|
|
|
5336
5342
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
5337
5343
|
error: "error";
|
|
5338
5344
|
running: "running";
|
|
5339
|
-
cancelled: "cancelled";
|
|
5340
5345
|
ok: "ok";
|
|
5346
|
+
cancelled: "cancelled";
|
|
5341
5347
|
}>>;
|
|
5342
5348
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
5343
5349
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -5498,8 +5504,8 @@ type SseEnvelope = z$1.infer<typeof sseEnvelopeSchema$1>; //#endregion
|
|
|
5498
5504
|
//#region src/schemas/api.d.ts
|
|
5499
5505
|
/** Lifecycle state for an app config reload triggered by `agent-evals.config.ts`. */
|
|
5500
5506
|
declare const configReloadStatusSchema: z$1.ZodEnum<{
|
|
5501
|
-
pending: "pending";
|
|
5502
5507
|
idle: "idle";
|
|
5508
|
+
pending: "pending";
|
|
5503
5509
|
reloading: "reloading";
|
|
5504
5510
|
}>;
|
|
5505
5511
|
/** Status for config reloads in the long-running app server. */
|
|
@@ -5507,8 +5513,8 @@ type ConfigReloadStatus = z$1.infer<typeof configReloadStatusSchema>;
|
|
|
5507
5513
|
/** UI/API-visible state for config reloads in `agent-evals app`. */
|
|
5508
5514
|
declare const configReloadStateSchema$1: z$1.ZodObject<{
|
|
5509
5515
|
status: z$1.ZodEnum<{
|
|
5510
|
-
pending: "pending";
|
|
5511
5516
|
idle: "idle";
|
|
5517
|
+
pending: "pending";
|
|
5512
5518
|
reloading: "reloading";
|
|
5513
5519
|
}>;
|
|
5514
5520
|
activeRunCount: z$1.ZodNumber;
|
|
@@ -5521,9 +5527,9 @@ type ConfigReloadState = z$1.infer<typeof configReloadStateSchema$1>;
|
|
|
5521
5527
|
declare const createRunRequestSchema$1: z$1.ZodObject<{
|
|
5522
5528
|
target: z$1.ZodObject<{
|
|
5523
5529
|
mode: z$1.ZodEnum<{
|
|
5524
|
-
caseIds: "caseIds";
|
|
5525
5530
|
all: "all";
|
|
5526
5531
|
evalIds: "evalIds";
|
|
5532
|
+
caseIds: "caseIds";
|
|
5527
5533
|
}>;
|
|
5528
5534
|
evalKeys: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
5529
5535
|
files: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
@@ -5535,9 +5541,9 @@ declare const createRunRequestSchema$1: z$1.ZodObject<{
|
|
|
5535
5541
|
temporary: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
5536
5542
|
cache: z$1.ZodOptional<z$1.ZodObject<{
|
|
5537
5543
|
mode: z$1.ZodDefault<z$1.ZodEnum<{
|
|
5538
|
-
refresh: "refresh";
|
|
5539
|
-
bypass: "bypass";
|
|
5540
5544
|
use: "use";
|
|
5545
|
+
bypass: "bypass";
|
|
5546
|
+
refresh: "refresh";
|
|
5541
5547
|
}>>;
|
|
5542
5548
|
}, z$1.core.$strip>>;
|
|
5543
5549
|
manualInputs: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
|
|
@@ -7286,4 +7292,4 @@ declare function defineEval<TInput = unknown, TOutputs extends EvalOutputs = Eva
|
|
|
7286
7292
|
/** Return whether the active eval case has tags matching the typed input. */
|
|
7287
7293
|
declare function matchesEvalTags(input: EvalTagMatchInput): boolean;
|
|
7288
7294
|
//#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
|
|
7295
|
+
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 };
|