@ls-stack/agent-eval 0.47.0 → 0.51.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/{app-BZmhhSFZ.mjs → app-CzLj4ZX0.mjs} +4 -4
- package/dist/apps/web/dist/assets/index-C5SveD-X.css +1 -0
- package/dist/apps/web/dist/assets/index-DwgyYZgf.js +373 -0
- package/dist/apps/web/dist/index.html +2 -2
- package/dist/bin.mjs +1 -1
- package/dist/{cli-vdJYkEVk.mjs → cli-Cvs7tc2v.mjs} +3 -3
- package/dist/index.d.mts +90 -49
- package/dist/index.mjs +3 -3
- package/dist/runChild.mjs +1 -1
- package/dist/{runOrchestration-BFdxG9ws.mjs → runOrchestration-o38J7uZO.mjs} +161 -110
- package/dist/{runner-DJWn_7p0.mjs → runner-LdMiDmAN.mjs} +2 -2
- package/dist/{runner--aH0jO4Z.mjs → runner-iWtmKx9z.mjs} +1 -1
- package/dist/{src-BRqs3kSA.mjs → src-Jahivm6d.mjs} +2 -2
- package/package.json +3 -3
- package/skills/agent-eval/SKILL.md +8 -2
- package/dist/apps/web/dist/assets/index-B5JrV3_C.css +0 -1
- package/dist/apps/web/dist/assets/index-DB61h-lP.js +0 -369
|
@@ -25,8 +25,8 @@
|
|
|
25
25
|
href="https://fonts.googleapis.com/css2?family=Geist:wght@400;500;600&family=JetBrains+Mono:wght@400;500&display=swap"
|
|
26
26
|
rel="stylesheet"
|
|
27
27
|
/>
|
|
28
|
-
<script type="module" crossorigin src="/assets/index-
|
|
29
|
-
<link rel="stylesheet" crossorigin href="/assets/index-
|
|
28
|
+
<script type="module" crossorigin src="/assets/index-DwgyYZgf.js"></script>
|
|
29
|
+
<link rel="stylesheet" crossorigin href="/assets/index-C5SveD-X.css">
|
|
30
30
|
</head>
|
|
31
31
|
<body>
|
|
32
32
|
<div id="root"></div>
|
package/dist/bin.mjs
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { A as validateCharts, At as buildEvalKey, C as deriveEvalFreshness, Ct as getEvalDisplayStatus, D as loadConfig, Dt as runSummarySchema, E as resolveEvalDefaultConfig, Lt as getEvalRegistry, O as buildDeclaredColumnDefs, Ot as resolveApiCallsConfig, S as parseManualInputValues, St as getEvalTitle, T as parseEvalDiscovery, Tt as matchesTagsFilter, _ as recomputePersistedCaseStatus, a as validateTagsFilters, b as resolveArtifactPath, bt as applyDerivedCallAttributes, c as getLastRunStatuses, d as loadPersistedRunSnapshots, f as nextShortIdFromSnapshots, g as recomputeEvalStatusesInRuns, h as persistRunState, i as resolveEvalTags, j as createFsCacheStore, jt as getCaseRowCaseKey, k as normalizeScoreDef, kt as resolveLlmCallsConfig, l as getLatestRunInfos, m as deleteTemporaryRuns, n as getTargetEvalKeys, o as stripTerminalControlCodes, p as persistCaseDetail, s as generateRunId, u as loadPersistedRunSnapshot, v as runTouchesEval, w as loadEvalModule, wt as deriveScopedSummaryFromCases, x as buildManualInputDescriptor, y as resolveTracePresentation } from "./runOrchestration-
|
|
1
|
+
import { A as validateCharts, At as buildEvalKey, C as deriveEvalFreshness, Ct as getEvalDisplayStatus, D as loadConfig, Dt as runSummarySchema, E as resolveEvalDefaultConfig, Lt as getEvalRegistry, O as buildDeclaredColumnDefs, Ot as resolveApiCallsConfig, S as parseManualInputValues, St as getEvalTitle, T as parseEvalDiscovery, Tt as matchesTagsFilter, _ as recomputePersistedCaseStatus, a as validateTagsFilters, b as resolveArtifactPath, bt as applyDerivedCallAttributes, c as getLastRunStatuses, d as loadPersistedRunSnapshots, f as nextShortIdFromSnapshots, g as recomputeEvalStatusesInRuns, h as persistRunState, i as resolveEvalTags, j as createFsCacheStore, jt as getCaseRowCaseKey, k as normalizeScoreDef, kt as resolveLlmCallsConfig, l as getLatestRunInfos, m as deleteTemporaryRuns, n as getTargetEvalKeys, o as stripTerminalControlCodes, p as persistCaseDetail, s as generateRunId, u as loadPersistedRunSnapshot, v as runTouchesEval, w as loadEvalModule, wt as deriveScopedSummaryFromCases, x as buildManualInputDescriptor, y as resolveTracePresentation } from "./runOrchestration-o38J7uZO.mjs";
|
|
2
2
|
import { copyFile, mkdir, readFile, rm, writeFile } from "node:fs/promises";
|
|
3
3
|
import { basename, dirname, extname, isAbsolute, join, relative, resolve, sep } from "node:path";
|
|
4
4
|
import { createHash, randomUUID } from "node:crypto";
|
|
@@ -2095,8 +2095,8 @@ async function commandApp(args) {
|
|
|
2095
2095
|
const { serve } = await import("@hono/node-server");
|
|
2096
2096
|
const bundledWebDist = resolve(currentDir, "apps/web/dist");
|
|
2097
2097
|
if (existsSync(bundledWebDist)) process.env.AGENT_EVALS_WEB_DIST = bundledWebDist;
|
|
2098
|
-
const appModule = await import("./app-
|
|
2099
|
-
const runnerModule = await import("./runner
|
|
2098
|
+
const appModule = await import("./app-CzLj4ZX0.mjs");
|
|
2099
|
+
const runnerModule = await import("./runner-iWtmKx9z.mjs");
|
|
2100
2100
|
if (!isHonoAppModule(appModule)) throw new Error("Server app module is invalid");
|
|
2101
2101
|
if (!isServerRunnerModule(runnerModule)) throw new Error("Server runner module is invalid");
|
|
2102
2102
|
await runnerModule.initRunner();
|
package/dist/index.d.mts
CHANGED
|
@@ -152,6 +152,9 @@ declare const evalStatsConfigSchema$1: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z
|
|
|
152
152
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
153
153
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
154
154
|
kind: z$1.ZodLiteral<"duration">;
|
|
155
|
+
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
156
|
+
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
157
|
+
kind: z$1.ZodLiteral<"cacheHits">;
|
|
155
158
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
156
159
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
157
160
|
kind: z$1.ZodLiteral<"column">;
|
|
@@ -222,6 +225,7 @@ declare const runLogEntrySchema$1: z$1.ZodObject<{
|
|
|
222
225
|
file: z$1.ZodString;
|
|
223
226
|
line: z$1.ZodNumber;
|
|
224
227
|
column: z$1.ZodNumber;
|
|
228
|
+
stack: z$1.ZodOptional<z$1.ZodString>;
|
|
225
229
|
}, z$1.core.$strip>>;
|
|
226
230
|
source: z$1.ZodOptional<z$1.ZodString>;
|
|
227
231
|
}, z$1.core.$strip>;
|
|
@@ -1012,13 +1016,15 @@ type EvalDefinitionBase<TInput = unknown, TOutputs extends EvalOutputs = EvalOut
|
|
|
1012
1016
|
* Opt-in: when omitted (or empty) the EvalCard renders no stats row at all.
|
|
1013
1017
|
* When provided, the stats render in order, left to right.
|
|
1014
1018
|
*
|
|
1015
|
-
* Built-in kinds (`cases`, `passRate`, `duration`, `
|
|
1016
|
-
* latest run summary. `
|
|
1017
|
-
*
|
|
1018
|
-
*
|
|
1019
|
-
*
|
|
1020
|
-
*
|
|
1021
|
-
*
|
|
1019
|
+
* Built-in kinds (`cases`, `passRate`, `duration`, `cacheHits`) read from
|
|
1020
|
+
* the latest run summary. `cacheHits` counts Agent Eval operation-level cache
|
|
1021
|
+
* hits over total cache operations, not LLM provider prompt-cache read
|
|
1022
|
+
* tokens. `kind: 'column'` aggregates a score or numeric output column across
|
|
1023
|
+
* the latest run's cases — `key` must match one of the eval's score or column
|
|
1024
|
+
* keys, and only finite numeric values participate in the reduction. When no
|
|
1025
|
+
* case has a numeric value for the key the stat renders an em dash, or hides
|
|
1026
|
+
* when `hideIfNoValue` is true. `label`, `format`, and `numberFormat` default
|
|
1027
|
+
* to the matching `ColumnDef`.
|
|
1022
1028
|
*/
|
|
1023
1029
|
stats?: EvalStatsConfig$1;
|
|
1024
1030
|
/**
|
|
@@ -1302,8 +1308,6 @@ type EvalCaseScope = {
|
|
|
1302
1308
|
logs: RunLogEntry$1[];
|
|
1303
1309
|
spans: EvalTraceSpan$2[];
|
|
1304
1310
|
checkpoints: Map<string, unknown>;
|
|
1305
|
-
spanStack: string[];
|
|
1306
|
-
activeSpanStack: EvalTraceSpan$2[];
|
|
1307
1311
|
/**
|
|
1308
1312
|
* Stack of active cache recorders. Ops are written to the top-most frame
|
|
1309
1313
|
* when it exists and `replayingDepth === 0`.
|
|
@@ -1984,8 +1988,8 @@ declare const traceSpanSchema$1: z$1.ZodObject<{
|
|
|
1984
1988
|
status: z$1.ZodEnum<{
|
|
1985
1989
|
error: "error";
|
|
1986
1990
|
running: "running";
|
|
1987
|
-
ok: "ok";
|
|
1988
1991
|
cancelled: "cancelled";
|
|
1992
|
+
ok: "ok";
|
|
1989
1993
|
}>;
|
|
1990
1994
|
attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
|
|
1991
1995
|
error: z$1.ZodOptional<z$1.ZodObject<{
|
|
@@ -2036,6 +2040,8 @@ declare const evalStatAggregateSchema: z$1.ZodEnum<{
|
|
|
2036
2040
|
type EvalStatAggregate = z$1.infer<typeof evalStatAggregateSchema>;
|
|
2037
2041
|
/**
|
|
2038
2042
|
* One entry in the EvalCard stats row. Built-in kinds use latest run totals;
|
|
2043
|
+
* `cacheHits` counts Agent Eval operation-level cache hits from spans and
|
|
2044
|
+
* `evalTracer.cache(...)` refs, not LLM provider prompt-cache read tokens.
|
|
2039
2045
|
* `column` aggregates a score or numeric output column across the latest run.
|
|
2040
2046
|
*/
|
|
2041
2047
|
declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
@@ -2048,6 +2054,9 @@ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
|
2048
2054
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
2049
2055
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2050
2056
|
kind: z$1.ZodLiteral<"duration">;
|
|
2057
|
+
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
2058
|
+
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2059
|
+
kind: z$1.ZodLiteral<"cacheHits">;
|
|
2051
2060
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
2052
2061
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2053
2062
|
kind: z$1.ZodLiteral<"column">;
|
|
@@ -2090,6 +2099,9 @@ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1
|
|
|
2090
2099
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
2091
2100
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2092
2101
|
kind: z$1.ZodLiteral<"duration">;
|
|
2102
|
+
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
2103
|
+
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2104
|
+
kind: z$1.ZodLiteral<"cacheHits">;
|
|
2093
2105
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
2094
2106
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2095
2107
|
kind: z$1.ZodLiteral<"column">;
|
|
@@ -2193,6 +2205,9 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2193
2205
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
2194
2206
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2195
2207
|
kind: z$1.ZodLiteral<"duration">;
|
|
2208
|
+
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
2209
|
+
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2210
|
+
kind: z$1.ZodLiteral<"cacheHits">;
|
|
2196
2211
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
2197
2212
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
2198
2213
|
kind: z$1.ZodLiteral<"column">;
|
|
@@ -2239,8 +2254,8 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2239
2254
|
}>;
|
|
2240
2255
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2241
2256
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2242
|
-
error: "error";
|
|
2243
2257
|
success: "success";
|
|
2258
|
+
error: "error";
|
|
2244
2259
|
warning: "warning";
|
|
2245
2260
|
accent: "accent";
|
|
2246
2261
|
accentDim: "accentDim";
|
|
@@ -2263,8 +2278,8 @@ declare const evalSummarySchema$1: z$1.ZodObject<{
|
|
|
2263
2278
|
}>;
|
|
2264
2279
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2265
2280
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2266
|
-
error: "error";
|
|
2267
2281
|
success: "success";
|
|
2282
|
+
error: "error";
|
|
2268
2283
|
warning: "warning";
|
|
2269
2284
|
accent: "accent";
|
|
2270
2285
|
accentDim: "accentDim";
|
|
@@ -2396,13 +2411,15 @@ declare const caseRowSchema$1: z$1.ZodObject<{
|
|
|
2396
2411
|
tags: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
2397
2412
|
status: z$1.ZodEnum<{
|
|
2398
2413
|
error: "error";
|
|
2414
|
+
pending: "pending";
|
|
2399
2415
|
running: "running";
|
|
2400
2416
|
cancelled: "cancelled";
|
|
2401
2417
|
pass: "pass";
|
|
2402
2418
|
fail: "fail";
|
|
2403
|
-
pending: "pending";
|
|
2404
2419
|
}>;
|
|
2405
2420
|
durationMs: z$1.ZodNullable<z$1.ZodNumber>;
|
|
2421
|
+
cacheHits: z$1.ZodOptional<z$1.ZodNumber>;
|
|
2422
|
+
cacheOperations: z$1.ZodOptional<z$1.ZodNumber>;
|
|
2406
2423
|
costUsd: z$1.ZodOptional<z$1.ZodNullable<z$1.ZodNumber>>;
|
|
2407
2424
|
columns: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnion<readonly [z$1.ZodType<string | number | boolean | Record<string, unknown> | unknown[] | null, unknown, z$1.core.$ZodTypeInternals<string | number | boolean | Record<string, unknown> | unknown[] | null, unknown>>, z$1.ZodUnion<readonly [z$1.ZodObject<{
|
|
2408
2425
|
source: z$1.ZodLiteral<"repo">;
|
|
@@ -2449,8 +2466,9 @@ declare const runLogLocationSchema: z$1.ZodObject<{
|
|
|
2449
2466
|
file: z$1.ZodString;
|
|
2450
2467
|
line: z$1.ZodNumber;
|
|
2451
2468
|
column: z$1.ZodNumber;
|
|
2469
|
+
stack: z$1.ZodOptional<z$1.ZodString>;
|
|
2452
2470
|
}, z$1.core.$strip>;
|
|
2453
|
-
/** Best-effort source location for one
|
|
2471
|
+
/** Best-effort source location and captured stack for one case log. */
|
|
2454
2472
|
type RunLogLocation = z$1.infer<typeof runLogLocationSchema>;
|
|
2455
2473
|
/** Schema for one persisted log entry captured during a case run. */
|
|
2456
2474
|
declare const runLogEntrySchema: z$1.ZodObject<{
|
|
@@ -2474,6 +2492,7 @@ declare const runLogEntrySchema: z$1.ZodObject<{
|
|
|
2474
2492
|
file: z$1.ZodString;
|
|
2475
2493
|
line: z$1.ZodNumber;
|
|
2476
2494
|
column: z$1.ZodNumber;
|
|
2495
|
+
stack: z$1.ZodOptional<z$1.ZodString>;
|
|
2477
2496
|
}, z$1.core.$strip>>;
|
|
2478
2497
|
source: z$1.ZodOptional<z$1.ZodString>;
|
|
2479
2498
|
}, z$1.core.$strip>;
|
|
@@ -2492,8 +2511,8 @@ declare const scoreTraceSchema: z$1.ZodObject<{
|
|
|
2492
2511
|
status: z$1.ZodEnum<{
|
|
2493
2512
|
error: "error";
|
|
2494
2513
|
running: "running";
|
|
2495
|
-
ok: "ok";
|
|
2496
2514
|
cancelled: "cancelled";
|
|
2515
|
+
ok: "ok";
|
|
2497
2516
|
}>;
|
|
2498
2517
|
attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
|
|
2499
2518
|
error: z$1.ZodOptional<z$1.ZodObject<{
|
|
@@ -2561,11 +2580,11 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
|
|
|
2561
2580
|
tags: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
2562
2581
|
status: z$1.ZodEnum<{
|
|
2563
2582
|
error: "error";
|
|
2583
|
+
pending: "pending";
|
|
2564
2584
|
running: "running";
|
|
2565
2585
|
cancelled: "cancelled";
|
|
2566
2586
|
pass: "pass";
|
|
2567
2587
|
fail: "fail";
|
|
2568
|
-
pending: "pending";
|
|
2569
2588
|
}>;
|
|
2570
2589
|
input: z$1.ZodUnknown;
|
|
2571
2590
|
trace: z$1.ZodArray<z$1.ZodObject<{
|
|
@@ -2579,8 +2598,8 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
|
|
|
2579
2598
|
status: z$1.ZodEnum<{
|
|
2580
2599
|
error: "error";
|
|
2581
2600
|
running: "running";
|
|
2582
|
-
ok: "ok";
|
|
2583
2601
|
cancelled: "cancelled";
|
|
2602
|
+
ok: "ok";
|
|
2584
2603
|
}>;
|
|
2585
2604
|
attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
|
|
2586
2605
|
error: z$1.ZodOptional<z$1.ZodObject<{
|
|
@@ -2648,8 +2667,8 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
|
|
|
2648
2667
|
status: z$1.ZodEnum<{
|
|
2649
2668
|
error: "error";
|
|
2650
2669
|
running: "running";
|
|
2651
|
-
ok: "ok";
|
|
2652
2670
|
cancelled: "cancelled";
|
|
2671
|
+
ok: "ok";
|
|
2653
2672
|
}>;
|
|
2654
2673
|
attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
|
|
2655
2674
|
error: z$1.ZodOptional<z$1.ZodObject<{
|
|
@@ -2746,6 +2765,7 @@ declare const caseDetailSchema$1: z$1.ZodObject<{
|
|
|
2746
2765
|
file: z$1.ZodString;
|
|
2747
2766
|
line: z$1.ZodNumber;
|
|
2748
2767
|
column: z$1.ZodNumber;
|
|
2768
|
+
stack: z$1.ZodOptional<z$1.ZodString>;
|
|
2749
2769
|
}, z$1.core.$strip>>;
|
|
2750
2770
|
source: z$1.ZodOptional<z$1.ZodString>;
|
|
2751
2771
|
}, z$1.core.$strip>>>;
|
|
@@ -2831,8 +2851,8 @@ type EvalChartAggregate = z$1.infer<typeof evalChartAggregateSchema>;
|
|
|
2831
2851
|
* not emit raw hex so authored evals stay decoupled from the web theme.
|
|
2832
2852
|
*/
|
|
2833
2853
|
declare const evalChartColorSchema: z$1.ZodEnum<{
|
|
2834
|
-
error: "error";
|
|
2835
2854
|
success: "success";
|
|
2855
|
+
error: "error";
|
|
2836
2856
|
warning: "warning";
|
|
2837
2857
|
accent: "accent";
|
|
2838
2858
|
accentDim: "accentDim";
|
|
@@ -2860,8 +2880,8 @@ declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
|
2860
2880
|
}>;
|
|
2861
2881
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2862
2882
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2863
|
-
error: "error";
|
|
2864
2883
|
success: "success";
|
|
2884
|
+
error: "error";
|
|
2865
2885
|
warning: "warning";
|
|
2866
2886
|
accent: "accent";
|
|
2867
2887
|
accentDim: "accentDim";
|
|
@@ -2884,8 +2904,8 @@ declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
|
|
|
2884
2904
|
}>;
|
|
2885
2905
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2886
2906
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2887
|
-
error: "error";
|
|
2888
2907
|
success: "success";
|
|
2908
|
+
error: "error";
|
|
2889
2909
|
warning: "warning";
|
|
2890
2910
|
accent: "accent";
|
|
2891
2911
|
accentDim: "accentDim";
|
|
@@ -2943,8 +2963,8 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
|
|
|
2943
2963
|
}>;
|
|
2944
2964
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2945
2965
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2946
|
-
error: "error";
|
|
2947
2966
|
success: "success";
|
|
2967
|
+
error: "error";
|
|
2948
2968
|
warning: "warning";
|
|
2949
2969
|
accent: "accent";
|
|
2950
2970
|
accentDim: "accentDim";
|
|
@@ -2967,8 +2987,8 @@ declare const evalChartConfigSchema: z$1.ZodObject<{
|
|
|
2967
2987
|
}>;
|
|
2968
2988
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
2969
2989
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
2970
|
-
error: "error";
|
|
2971
2990
|
success: "success";
|
|
2991
|
+
error: "error";
|
|
2972
2992
|
warning: "warning";
|
|
2973
2993
|
accent: "accent";
|
|
2974
2994
|
accentDim: "accentDim";
|
|
@@ -3033,8 +3053,8 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
|
|
|
3033
3053
|
}>;
|
|
3034
3054
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3035
3055
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3036
|
-
error: "error";
|
|
3037
3056
|
success: "success";
|
|
3057
|
+
error: "error";
|
|
3038
3058
|
warning: "warning";
|
|
3039
3059
|
accent: "accent";
|
|
3040
3060
|
accentDim: "accentDim";
|
|
@@ -3057,8 +3077,8 @@ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
|
|
|
3057
3077
|
}>;
|
|
3058
3078
|
label: z$1.ZodOptional<z$1.ZodString>;
|
|
3059
3079
|
color: z$1.ZodOptional<z$1.ZodEnum<{
|
|
3060
|
-
error: "error";
|
|
3061
3080
|
success: "success";
|
|
3081
|
+
error: "error";
|
|
3062
3082
|
warning: "warning";
|
|
3063
3083
|
accent: "accent";
|
|
3064
3084
|
accentDim: "accentDim";
|
|
@@ -3109,10 +3129,10 @@ declare const runManifestSchema$1: z$1.ZodObject<{
|
|
|
3109
3129
|
shortId: z$1.ZodString;
|
|
3110
3130
|
status: z$1.ZodEnum<{
|
|
3111
3131
|
error: "error";
|
|
3112
|
-
running: "running";
|
|
3113
|
-
cancelled: "cancelled";
|
|
3114
3132
|
pending: "pending";
|
|
3133
|
+
running: "running";
|
|
3115
3134
|
completed: "completed";
|
|
3135
|
+
cancelled: "cancelled";
|
|
3116
3136
|
}>;
|
|
3117
3137
|
temporary: z$1.ZodDefault<z$1.ZodOptional<z$1.ZodBoolean>>;
|
|
3118
3138
|
startedAt: z$1.ZodString;
|
|
@@ -3122,8 +3142,8 @@ declare const runManifestSchema$1: z$1.ZodObject<{
|
|
|
3122
3142
|
target: z$1.ZodObject<{
|
|
3123
3143
|
mode: z$1.ZodEnum<{
|
|
3124
3144
|
all: "all";
|
|
3125
|
-
caseIds: "caseIds";
|
|
3126
3145
|
evalIds: "evalIds";
|
|
3146
|
+
caseIds: "caseIds";
|
|
3127
3147
|
}>;
|
|
3128
3148
|
evalKeys: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
3129
3149
|
files: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
@@ -3149,10 +3169,10 @@ declare const runSummarySchema$1: z$1.ZodObject<{
|
|
|
3149
3169
|
runId: z$1.ZodString;
|
|
3150
3170
|
status: z$1.ZodEnum<{
|
|
3151
3171
|
error: "error";
|
|
3152
|
-
running: "running";
|
|
3153
|
-
cancelled: "cancelled";
|
|
3154
3172
|
pending: "pending";
|
|
3173
|
+
running: "running";
|
|
3155
3174
|
completed: "completed";
|
|
3175
|
+
cancelled: "cancelled";
|
|
3156
3176
|
}>;
|
|
3157
3177
|
totalCases: z$1.ZodNumber;
|
|
3158
3178
|
passedCases: z$1.ZodNumber;
|
|
@@ -3189,6 +3209,21 @@ type ScopedCaseSummary = {
|
|
|
3189
3209
|
pendingCases: number;
|
|
3190
3210
|
runningCases: number;
|
|
3191
3211
|
totalDurationMs: number | null;
|
|
3212
|
+
/**
|
|
3213
|
+
* Sum of Agent Eval operation-level cache hits across the scoped case rows.
|
|
3214
|
+
*
|
|
3215
|
+
* Missing values from older run artifacts count as zero. This is separate
|
|
3216
|
+
* from LLM prompt-cache token reads such as `cachedInputTokens`.
|
|
3217
|
+
*/
|
|
3218
|
+
cacheHits: number;
|
|
3219
|
+
/**
|
|
3220
|
+
* Sum of Agent Eval operation-level cache activity entries across the scoped
|
|
3221
|
+
* case rows.
|
|
3222
|
+
*
|
|
3223
|
+
* This is the denominator for `cacheHits`. Missing values from older run
|
|
3224
|
+
* artifacts count as zero.
|
|
3225
|
+
*/
|
|
3226
|
+
cacheOperations: number;
|
|
3192
3227
|
};
|
|
3193
3228
|
//#endregion
|
|
3194
3229
|
//#region src/evalStatus.d.ts
|
|
@@ -4042,8 +4077,8 @@ declare const spanCacheOptionsSchema: z$1.ZodObject<{
|
|
|
4042
4077
|
type SpanCacheOptions = z$1.infer<typeof spanCacheOptionsSchema>;
|
|
4043
4078
|
/** Category of operation stored in the eval cache. */
|
|
4044
4079
|
declare const cacheOperationTypeSchema: z$1.ZodEnum<{
|
|
4045
|
-
span: "span";
|
|
4046
4080
|
value: "value";
|
|
4081
|
+
span: "span";
|
|
4047
4082
|
}>;
|
|
4048
4083
|
/** Category of operation stored in the eval cache. */
|
|
4049
4084
|
type CacheOperationType = z$1.infer<typeof cacheOperationTypeSchema>;
|
|
@@ -4086,8 +4121,8 @@ declare const cacheListItemSchema$1: z$1.ZodObject<{
|
|
|
4086
4121
|
key: z$1.ZodString;
|
|
4087
4122
|
namespace: z$1.ZodString;
|
|
4088
4123
|
operationType: z$1.ZodEnum<{
|
|
4089
|
-
span: "span";
|
|
4090
4124
|
value: "value";
|
|
4125
|
+
span: "span";
|
|
4091
4126
|
}>;
|
|
4092
4127
|
operationName: z$1.ZodString;
|
|
4093
4128
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4149,8 +4184,8 @@ declare const cacheRecordingSchema: z$1.ZodObject<{
|
|
|
4149
4184
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4150
4185
|
error: "error";
|
|
4151
4186
|
running: "running";
|
|
4152
|
-
ok: "ok";
|
|
4153
4187
|
cancelled: "cancelled";
|
|
4188
|
+
ok: "ok";
|
|
4154
4189
|
}>>;
|
|
4155
4190
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
4156
4191
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4209,8 +4244,8 @@ declare const cacheEntrySchema: z$1.ZodObject<{
|
|
|
4209
4244
|
key: z$1.ZodString;
|
|
4210
4245
|
namespace: z$1.ZodString;
|
|
4211
4246
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4212
|
-
span: "span";
|
|
4213
4247
|
value: "value";
|
|
4248
|
+
span: "span";
|
|
4214
4249
|
}>>;
|
|
4215
4250
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4216
4251
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4222,8 +4257,8 @@ declare const cacheEntrySchema: z$1.ZodObject<{
|
|
|
4222
4257
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4223
4258
|
error: "error";
|
|
4224
4259
|
running: "running";
|
|
4225
|
-
ok: "ok";
|
|
4226
4260
|
cancelled: "cancelled";
|
|
4261
|
+
ok: "ok";
|
|
4227
4262
|
}>>;
|
|
4228
4263
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
4229
4264
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4288,8 +4323,8 @@ declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
|
|
|
4288
4323
|
key: z$1.ZodString;
|
|
4289
4324
|
namespace: z$1.ZodString;
|
|
4290
4325
|
operationType: z$1.ZodEnum<{
|
|
4291
|
-
span: "span";
|
|
4292
4326
|
value: "value";
|
|
4327
|
+
span: "span";
|
|
4293
4328
|
}>;
|
|
4294
4329
|
operationName: z$1.ZodString;
|
|
4295
4330
|
storedAt: z$1.ZodString;
|
|
@@ -4299,8 +4334,8 @@ declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
|
|
|
4299
4334
|
key: z$1.ZodString;
|
|
4300
4335
|
namespace: z$1.ZodString;
|
|
4301
4336
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4302
|
-
span: "span";
|
|
4303
4337
|
value: "value";
|
|
4338
|
+
span: "span";
|
|
4304
4339
|
}>>;
|
|
4305
4340
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4306
4341
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4312,8 +4347,8 @@ declare const cacheDebugKeyEntrySchema: z$1.ZodObject<{
|
|
|
4312
4347
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4313
4348
|
error: "error";
|
|
4314
4349
|
running: "running";
|
|
4315
|
-
ok: "ok";
|
|
4316
4350
|
cancelled: "cancelled";
|
|
4351
|
+
ok: "ok";
|
|
4317
4352
|
}>>;
|
|
4318
4353
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
4319
4354
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4378,8 +4413,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
4378
4413
|
key: z$1.ZodString;
|
|
4379
4414
|
namespace: z$1.ZodString;
|
|
4380
4415
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4381
|
-
span: "span";
|
|
4382
4416
|
value: "value";
|
|
4417
|
+
span: "span";
|
|
4383
4418
|
}>>;
|
|
4384
4419
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4385
4420
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4391,8 +4426,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
4391
4426
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4392
4427
|
error: "error";
|
|
4393
4428
|
running: "running";
|
|
4394
|
-
ok: "ok";
|
|
4395
4429
|
cancelled: "cancelled";
|
|
4430
|
+
ok: "ok";
|
|
4396
4431
|
}>>;
|
|
4397
4432
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
4398
4433
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4448,8 +4483,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
4448
4483
|
key: z$1.ZodString;
|
|
4449
4484
|
namespace: z$1.ZodString;
|
|
4450
4485
|
operationType: z$1.ZodEnum<{
|
|
4451
|
-
span: "span";
|
|
4452
4486
|
value: "value";
|
|
4487
|
+
span: "span";
|
|
4453
4488
|
}>;
|
|
4454
4489
|
operationName: z$1.ZodString;
|
|
4455
4490
|
storedAt: z$1.ZodString;
|
|
@@ -4459,8 +4494,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
4459
4494
|
key: z$1.ZodString;
|
|
4460
4495
|
namespace: z$1.ZodString;
|
|
4461
4496
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4462
|
-
span: "span";
|
|
4463
4497
|
value: "value";
|
|
4498
|
+
span: "span";
|
|
4464
4499
|
}>>;
|
|
4465
4500
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4466
4501
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4472,8 +4507,8 @@ declare const cacheEntryWithDebugKeySchema$1: z$1.ZodObject<{
|
|
|
4472
4507
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4473
4508
|
error: "error";
|
|
4474
4509
|
running: "running";
|
|
4475
|
-
ok: "ok";
|
|
4476
4510
|
cancelled: "cancelled";
|
|
4511
|
+
ok: "ok";
|
|
4477
4512
|
}>>;
|
|
4478
4513
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
4479
4514
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4538,8 +4573,8 @@ declare const cacheFileSchema: z$1.ZodObject<{
|
|
|
4538
4573
|
key: z$1.ZodString;
|
|
4539
4574
|
namespace: z$1.ZodString;
|
|
4540
4575
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4541
|
-
span: "span";
|
|
4542
4576
|
value: "value";
|
|
4577
|
+
span: "span";
|
|
4543
4578
|
}>>;
|
|
4544
4579
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4545
4580
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4551,8 +4586,8 @@ declare const cacheFileSchema: z$1.ZodObject<{
|
|
|
4551
4586
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4552
4587
|
error: "error";
|
|
4553
4588
|
running: "running";
|
|
4554
|
-
ok: "ok";
|
|
4555
4589
|
cancelled: "cancelled";
|
|
4590
|
+
ok: "ok";
|
|
4556
4591
|
}>>;
|
|
4557
4592
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
4558
4593
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4616,8 +4651,8 @@ declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
|
|
|
4616
4651
|
key: z$1.ZodString;
|
|
4617
4652
|
namespace: z$1.ZodString;
|
|
4618
4653
|
operationType: z$1.ZodEnum<{
|
|
4619
|
-
span: "span";
|
|
4620
4654
|
value: "value";
|
|
4655
|
+
span: "span";
|
|
4621
4656
|
}>;
|
|
4622
4657
|
operationName: z$1.ZodString;
|
|
4623
4658
|
storedAt: z$1.ZodString;
|
|
@@ -4627,8 +4662,8 @@ declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
|
|
|
4627
4662
|
key: z$1.ZodString;
|
|
4628
4663
|
namespace: z$1.ZodString;
|
|
4629
4664
|
operationType: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4630
|
-
span: "span";
|
|
4631
4665
|
value: "value";
|
|
4666
|
+
span: "span";
|
|
4632
4667
|
}>>;
|
|
4633
4668
|
operationName: z$1.ZodOptional<z$1.ZodString>;
|
|
4634
4669
|
spanName: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4640,8 +4675,8 @@ declare const cacheDebugKeyFileSchema: z$1.ZodObject<{
|
|
|
4640
4675
|
finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
|
|
4641
4676
|
error: "error";
|
|
4642
4677
|
running: "running";
|
|
4643
|
-
ok: "ok";
|
|
4644
4678
|
cancelled: "cancelled";
|
|
4679
|
+
ok: "ok";
|
|
4645
4680
|
}>>;
|
|
4646
4681
|
finalError: z$1.ZodOptional<z$1.ZodObject<{
|
|
4647
4682
|
name: z$1.ZodOptional<z$1.ZodString>;
|
|
@@ -4798,8 +4833,8 @@ declare const createRunRequestSchema$1: z$1.ZodObject<{
|
|
|
4798
4833
|
target: z$1.ZodObject<{
|
|
4799
4834
|
mode: z$1.ZodEnum<{
|
|
4800
4835
|
all: "all";
|
|
4801
|
-
caseIds: "caseIds";
|
|
4802
4836
|
evalIds: "evalIds";
|
|
4837
|
+
caseIds: "caseIds";
|
|
4803
4838
|
}>;
|
|
4804
4839
|
evalKeys: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
4805
4840
|
files: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
|
|
@@ -5148,6 +5183,9 @@ declare const evalSummarySchema: z$1.ZodObject<{
|
|
|
5148
5183
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
5149
5184
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
5150
5185
|
kind: z$1.ZodLiteral<"duration">;
|
|
5186
|
+
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
5187
|
+
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
5188
|
+
kind: z$1.ZodLiteral<"cacheHits">;
|
|
5151
5189
|
}, z$1.core.$strip>, z$1.ZodObject<{
|
|
5152
5190
|
hideIfNoValue: z$1.ZodOptional<z$1.ZodBoolean>;
|
|
5153
5191
|
kind: z$1.ZodLiteral<"column">;
|
|
@@ -5358,6 +5396,8 @@ declare const caseRowSchema: z$1.ZodObject<{
|
|
|
5358
5396
|
pending: "pending";
|
|
5359
5397
|
}>;
|
|
5360
5398
|
durationMs: z$1.ZodNullable<z$1.ZodNumber>;
|
|
5399
|
+
cacheHits: z$1.ZodOptional<z$1.ZodNumber>;
|
|
5400
|
+
cacheOperations: z$1.ZodOptional<z$1.ZodNumber>;
|
|
5361
5401
|
costUsd: z$1.ZodOptional<z$1.ZodNullable<z$1.ZodNumber>>;
|
|
5362
5402
|
columns: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnion<readonly [z$1.ZodType<string | number | boolean | Record<string, unknown> | unknown[] | null, unknown, z$1.core.$ZodTypeInternals<string | number | boolean | Record<string, unknown> | unknown[] | null, unknown>>, z$1.ZodUnion<readonly [z$1.ZodObject<{
|
|
5363
5403
|
source: z$1.ZodLiteral<"repo">;
|
|
@@ -5567,6 +5607,7 @@ declare const caseDetailSchema: z$1.ZodObject<{
|
|
|
5567
5607
|
file: z$1.ZodString;
|
|
5568
5608
|
line: z$1.ZodNumber;
|
|
5569
5609
|
column: z$1.ZodNumber;
|
|
5610
|
+
stack: z$1.ZodOptional<z$1.ZodString>;
|
|
5570
5611
|
}, z$1.core.$strip>>;
|
|
5571
5612
|
source: z$1.ZodOptional<z$1.ZodString>;
|
|
5572
5613
|
}, z$1.core.$strip>>>;
|
package/dist/index.mjs
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { $ as getCurrentScope, B as deserializeCacheValue, F as evalSpan, G as readManualInputFile, H as serializeCacheValue, I as evalTracer, J as appendToEvalOutput, K as evalExpect, L as hashCacheKey, Lt as getEvalRegistry, M as z, N as buildTraceTree, P as captureEvalSpanError, Q as evalTime, R as hashCacheKeySync, U as repoFile, V as serializeCacheRecording, W as manualInputFileValueSchema, X as evalAssert, Z as evalLog, _t as extractLlmCalls, at as nextEvalId, ct as runInExistingEvalScope, dt as startEvalBackgroundJob, et as getEvalCaseInput, gt as extractApiCalls, ht as extractCacheHits, it as mergeEvalOutput, lt as setEvalOutput, mt as extractCacheEntries, nt as isInEvalScope, ot as runInEvalRuntimeScope, q as EvalAssertionError, st as runInEvalScope, tt as incrementEvalOutput, ut as setScopeCacheContext, vt as simulateLlmCallCost, xt as getNestedAttribute, yt as simulateTokenAllocation, z as deserializeCacheRecording } from "./runOrchestration-
|
|
2
|
-
import { a as materializeManualInputFiles, i as isManualInputFileValue, n as createRunner, o as stageManualInputFile, r as cleanupStagedManualInputFiles, s as stageManualInputFileFromPath, t as runCli } from "./cli-
|
|
3
|
-
import { n as matchesEvalTags, t as defineEval } from "./src-
|
|
1
|
+
import { $ as getCurrentScope, B as deserializeCacheValue, F as evalSpan, G as readManualInputFile, H as serializeCacheValue, I as evalTracer, J as appendToEvalOutput, K as evalExpect, L as hashCacheKey, Lt as getEvalRegistry, M as z, N as buildTraceTree, P as captureEvalSpanError, Q as evalTime, R as hashCacheKeySync, U as repoFile, V as serializeCacheRecording, W as manualInputFileValueSchema, X as evalAssert, Z as evalLog, _t as extractLlmCalls, at as nextEvalId, ct as runInExistingEvalScope, dt as startEvalBackgroundJob, et as getEvalCaseInput, gt as extractApiCalls, ht as extractCacheHits, it as mergeEvalOutput, lt as setEvalOutput, mt as extractCacheEntries, nt as isInEvalScope, ot as runInEvalRuntimeScope, q as EvalAssertionError, st as runInEvalScope, tt as incrementEvalOutput, ut as setScopeCacheContext, vt as simulateLlmCallCost, xt as getNestedAttribute, yt as simulateTokenAllocation, z as deserializeCacheRecording } from "./runOrchestration-o38J7uZO.mjs";
|
|
2
|
+
import { a as materializeManualInputFiles, i as isManualInputFileValue, n as createRunner, o as stageManualInputFile, r as cleanupStagedManualInputFiles, s as stageManualInputFileFromPath, t as runCli } from "./cli-Cvs7tc2v.mjs";
|
|
3
|
+
import { n as matchesEvalTags, t as defineEval } from "./src-Jahivm6d.mjs";
|
|
4
4
|
export { EvalAssertionError, appendToEvalOutput, buildTraceTree, captureEvalSpanError, cleanupStagedManualInputFiles, createRunner, defineEval, deserializeCacheRecording, deserializeCacheValue, evalAssert, evalExpect, evalLog, evalSpan, evalTime, evalTracer, extractApiCalls, extractCacheEntries, extractCacheHits, extractLlmCalls, getCurrentScope, getEvalCaseInput, getEvalRegistry, getNestedAttribute, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, isManualInputFileValue, manualInputFileValueSchema, matchesEvalTags, materializeManualInputFiles, mergeEvalOutput, nextEvalId, readManualInputFile, repoFile, runCli, runInEvalRuntimeScope, runInEvalScope, runInExistingEvalScope, serializeCacheRecording, serializeCacheValue, setEvalOutput, setScopeCacheContext, simulateLlmCallCost, simulateTokenAllocation, stageManualInputFile, stageManualInputFileFromPath, startEvalBackgroundJob, z };
|
package/dist/runChild.mjs
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { At as buildEvalKey, D as loadConfig, Dt as runSummarySchema, Et as runManifestSchema, Ft as columnDefSchema, Mt as evalStatsConfigSchema, Nt as manualInputDescriptorSchema, Pt as evalChartsConfigSchema, T as parseEvalDiscovery, Y as configureEvalRunLogs, ft as createRunRequestSchema, h as persistRunState, j as createFsCacheStore, r as getTargetEvals$1, t as executeRun } from "./runOrchestration-
|
|
1
|
+
import { At as buildEvalKey, D as loadConfig, Dt as runSummarySchema, Et as runManifestSchema, Ft as columnDefSchema, Mt as evalStatsConfigSchema, Nt as manualInputDescriptorSchema, Pt as evalChartsConfigSchema, T as parseEvalDiscovery, Y as configureEvalRunLogs, ft as createRunRequestSchema, h as persistRunState, j as createFsCacheStore, r as getTargetEvals$1, t as executeRun } from "./runOrchestration-o38J7uZO.mjs";
|
|
2
2
|
import { z } from "zod/v4";
|
|
3
3
|
import { readFile } from "node:fs/promises";
|
|
4
4
|
import { relative } from "node:path";
|