@ls-stack/agent-eval 0.4.0 → 0.6.0

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/index.d.mts CHANGED
@@ -1,39 +1,39 @@
1
- import { z } from "zod/v4";
1
+ import { z, z as z$1 } from "zod/v4";
2
2
 
3
3
  //#region ../shared/src/schemas/display.d.ts
4
- declare const scalarCellSchema: z.ZodUnion<readonly [z.ZodString, z.ZodNumber, z.ZodBoolean, z.ZodNull]>;
4
+ declare const scalarCellSchema: z$1.ZodUnion<readonly [z$1.ZodString, z$1.ZodNumber, z$1.ZodBoolean, z$1.ZodNull]>;
5
5
  /** Primitive table cell value supported by the eval UI. */
6
- type ScalarCell = z.infer<typeof scalarCellSchema>;
7
- declare const jsonCellSchema: z.ZodType<string | number | boolean | null | Record<string, unknown> | unknown[]>;
6
+ type ScalarCell = z$1.infer<typeof scalarCellSchema>;
7
+ declare const jsonCellSchema: z$1.ZodType<string | number | boolean | null | Record<string, unknown> | unknown[]>;
8
8
  /** JSON-safe value supported by `format: 'json'` columns. */
9
- type JsonCell = z.infer<typeof jsonCellSchema>;
10
- declare const repoFileRefSchema: z.ZodObject<{
11
- source: z.ZodLiteral<"repo">;
12
- path: z.ZodString;
13
- mimeType: z.ZodOptional<z.ZodString>;
14
- }, z.core.$strip>;
9
+ type JsonCell = z$1.infer<typeof jsonCellSchema>;
10
+ declare const repoFileRefSchema: z$1.ZodObject<{
11
+ source: z$1.ZodLiteral<"repo">;
12
+ path: z$1.ZodString;
13
+ mimeType: z$1.ZodOptional<z$1.ZodString>;
14
+ }, z$1.core.$strip>;
15
15
  /** Reference to a file that lives in the authored workspace. */
16
- type RepoFileRef = z.infer<typeof repoFileRefSchema>;
17
- declare const runArtifactRefSchema: z.ZodObject<{
18
- source: z.ZodLiteral<"run">;
19
- artifactId: z.ZodString;
20
- mimeType: z.ZodString;
21
- fileName: z.ZodOptional<z.ZodString>;
22
- }, z.core.$strip>;
16
+ type RepoFileRef = z$1.infer<typeof repoFileRefSchema>;
17
+ declare const runArtifactRefSchema: z$1.ZodObject<{
18
+ source: z$1.ZodLiteral<"run">;
19
+ artifactId: z$1.ZodString;
20
+ mimeType: z$1.ZodString;
21
+ fileName: z$1.ZodOptional<z$1.ZodString>;
22
+ }, z$1.core.$strip>;
23
23
  /** Reference to a generated artifact stored under a specific run. */
24
- type RunArtifactRef = z.infer<typeof runArtifactRefSchema>;
25
- declare const fileRefSchema: z.ZodUnion<readonly [z.ZodObject<{
26
- source: z.ZodLiteral<"repo">;
27
- path: z.ZodString;
28
- mimeType: z.ZodOptional<z.ZodString>;
29
- }, z.core.$strip>, z.ZodObject<{
30
- source: z.ZodLiteral<"run">;
31
- artifactId: z.ZodString;
32
- mimeType: z.ZodString;
33
- fileName: z.ZodOptional<z.ZodString>;
34
- }, z.core.$strip>]>;
24
+ type RunArtifactRef = z$1.infer<typeof runArtifactRefSchema>;
25
+ declare const fileRefSchema: z$1.ZodUnion<readonly [z$1.ZodObject<{
26
+ source: z$1.ZodLiteral<"repo">;
27
+ path: z$1.ZodString;
28
+ mimeType: z$1.ZodOptional<z$1.ZodString>;
29
+ }, z$1.core.$strip>, z$1.ZodObject<{
30
+ source: z$1.ZodLiteral<"run">;
31
+ artifactId: z$1.ZodString;
32
+ mimeType: z$1.ZodString;
33
+ fileName: z$1.ZodOptional<z$1.ZodString>;
34
+ }, z$1.core.$strip>]>;
35
35
  /** File reference supported by media and file columns. */
36
- type FileRef = z.infer<typeof fileRefSchema>;
36
+ type FileRef = z$1.infer<typeof fileRefSchema>;
37
37
  /** Numeric presentation options for values rendered with `format: 'number'`. */
38
38
  type NumberDisplayOptions = {
39
39
  /** Number notation used when rendering the value. */notation?: 'standard' | 'compact'; /** Compact style used when `notation: 'compact'` is enabled. */
@@ -43,17 +43,17 @@ type NumberDisplayOptions = {
43
43
  decimalPlaces?: number;
44
44
  };
45
45
  /** Schema for numeric presentation options used by number-formatted values. */
46
- declare const numberDisplayOptionsSchema: z.ZodType<NumberDisplayOptions>;
46
+ declare const numberDisplayOptionsSchema: z$1.ZodType<NumberDisplayOptions>;
47
47
  /** Schema for the supported column rendering kinds in list views. */
48
- declare const columnKindSchema: z.ZodEnum<{
48
+ declare const columnKindSchema: z$1.ZodEnum<{
49
49
  string: "string";
50
50
  number: "number";
51
51
  boolean: "boolean";
52
52
  }>;
53
53
  /** Display kind used by a column definition in the UI. */
54
- type ColumnKind = z.infer<typeof columnKindSchema>;
54
+ type ColumnKind = z$1.infer<typeof columnKindSchema>;
55
55
  /** Schema for the built-in column formatting presets. */
56
- declare const columnFormatSchema: z.ZodEnum<{
56
+ declare const columnFormatSchema: z$1.ZodEnum<{
57
57
  number: "number";
58
58
  boolean: "boolean";
59
59
  file: "file";
@@ -68,17 +68,17 @@ declare const columnFormatSchema: z.ZodEnum<{
68
68
  stars: "stars";
69
69
  }>;
70
70
  /** Formatting preset applied to a column value in the UI. */
71
- type ColumnFormat = z.infer<typeof columnFormatSchema>;
71
+ type ColumnFormat = z$1.infer<typeof columnFormatSchema>;
72
72
  /** Schema describing a rendered column in the eval results table. */
73
- declare const columnDefSchema: z.ZodObject<{
74
- key: z.ZodString;
75
- label: z.ZodString;
76
- kind: z.ZodEnum<{
73
+ declare const columnDefSchema: z$1.ZodObject<{
74
+ key: z$1.ZodString;
75
+ label: z$1.ZodString;
76
+ kind: z$1.ZodEnum<{
77
77
  string: "string";
78
78
  number: "number";
79
79
  boolean: "boolean";
80
80
  }>;
81
- format: z.ZodOptional<z.ZodEnum<{
81
+ format: z$1.ZodOptional<z$1.ZodEnum<{
82
82
  number: "number";
83
83
  boolean: "boolean";
84
84
  file: "file";
@@ -92,34 +92,34 @@ declare const columnDefSchema: z.ZodObject<{
92
92
  passFail: "passFail";
93
93
  stars: "stars";
94
94
  }>>;
95
- numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
96
- isScore: z.ZodOptional<z.ZodBoolean>;
97
- isManualScore: z.ZodOptional<z.ZodBoolean>;
98
- passThreshold: z.ZodOptional<z.ZodNumber>;
99
- maxStars: z.ZodOptional<z.ZodNumber>;
100
- hideInTable: z.ZodOptional<z.ZodBoolean>;
101
- sortable: z.ZodOptional<z.ZodBoolean>;
102
- align: z.ZodOptional<z.ZodEnum<{
95
+ numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
96
+ isScore: z$1.ZodOptional<z$1.ZodBoolean>;
97
+ isManualScore: z$1.ZodOptional<z$1.ZodBoolean>;
98
+ passThreshold: z$1.ZodOptional<z$1.ZodNumber>;
99
+ maxStars: z$1.ZodOptional<z$1.ZodNumber>;
100
+ hideInTable: z$1.ZodOptional<z$1.ZodBoolean>;
101
+ sortable: z$1.ZodOptional<z$1.ZodBoolean>;
102
+ align: z$1.ZodOptional<z$1.ZodEnum<{
103
103
  left: "left";
104
104
  center: "center";
105
105
  right: "right";
106
106
  }>>;
107
- }, z.core.$strip>;
107
+ }, z$1.core.$strip>;
108
108
  /** Column definition exposed to the UI for eval and case tables. */
109
- type ColumnDef = z.infer<typeof columnDefSchema>;
109
+ type ColumnDef = z$1.infer<typeof columnDefSchema>;
110
110
  /** Schema for any supported value that can populate a table cell. */
111
- declare const cellValueSchema: z.ZodUnion<readonly [z.ZodType<string | number | boolean | Record<string, unknown> | unknown[] | null, unknown, z.core.$ZodTypeInternals<string | number | boolean | Record<string, unknown> | unknown[] | null, unknown>>, z.ZodUnion<readonly [z.ZodObject<{
112
- source: z.ZodLiteral<"repo">;
113
- path: z.ZodString;
114
- mimeType: z.ZodOptional<z.ZodString>;
115
- }, z.core.$strip>, z.ZodObject<{
116
- source: z.ZodLiteral<"run">;
117
- artifactId: z.ZodString;
118
- mimeType: z.ZodString;
119
- fileName: z.ZodOptional<z.ZodString>;
120
- }, z.core.$strip>]>]>;
111
+ declare const cellValueSchema: 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<{
112
+ source: z$1.ZodLiteral<"repo">;
113
+ path: z$1.ZodString;
114
+ mimeType: z$1.ZodOptional<z$1.ZodString>;
115
+ }, z$1.core.$strip>, z$1.ZodObject<{
116
+ source: z$1.ZodLiteral<"run">;
117
+ artifactId: z$1.ZodString;
118
+ mimeType: z$1.ZodString;
119
+ fileName: z$1.ZodOptional<z$1.ZodString>;
120
+ }, z$1.core.$strip>]>]>;
121
121
  /** Value stored in a rendered eval result table cell. */
122
- type CellValue = z.infer<typeof cellValueSchema>;
122
+ type CellValue = z$1.infer<typeof cellValueSchema>;
123
123
  //#endregion
124
124
  //#region ../shared/src/schemas/trace.d.ts
125
125
  /**
@@ -128,9 +128,9 @@ type CellValue = z.infer<typeof cellValueSchema>;
128
128
  * The value is intentionally open-ended so external tracers can preserve their
129
129
  * native span kinds instead of collapsing them into the built-in categories.
130
130
  */
131
- declare const traceSpanKindSchema: z.ZodString;
131
+ declare const traceSpanKindSchema: z$1.ZodString;
132
132
  /** Schema for the supported presentation formats of trace attributes. */
133
- declare const traceAttributeDisplayFormatSchema: z.ZodEnum<{
133
+ declare const traceAttributeDisplayFormatSchema: z$1.ZodEnum<{
134
134
  string: "string";
135
135
  number: "number";
136
136
  duration: "duration";
@@ -141,42 +141,42 @@ declare const traceAttributeDisplayFormatSchema: z.ZodEnum<{
141
141
  *
142
142
  * This affects presentation only and does not change the stored value.
143
143
  */
144
- type TraceAttributeDisplayFormat = z.infer<typeof traceAttributeDisplayFormatSchema>;
144
+ type TraceAttributeDisplayFormat = z$1.infer<typeof traceAttributeDisplayFormatSchema>;
145
145
  /** Schema for the UI locations where a trace attribute can appear. */
146
- declare const traceAttributeDisplayPlacementSchema: z.ZodEnum<{
146
+ declare const traceAttributeDisplayPlacementSchema: z$1.ZodEnum<{
147
147
  tree: "tree";
148
148
  detail: "detail";
149
149
  section: "section";
150
150
  }>;
151
151
  /** UI locations where a trace attribute may be rendered. */
152
- type TraceAttributeDisplayPlacement = z.infer<typeof traceAttributeDisplayPlacementSchema>;
152
+ type TraceAttributeDisplayPlacement = z$1.infer<typeof traceAttributeDisplayPlacementSchema>;
153
153
  /** Schema for resolved trace display rules sent to the UI. */
154
- declare const traceAttributeDisplaySchema: z.ZodObject<{
155
- key: z.ZodOptional<z.ZodString>;
156
- path: z.ZodString;
157
- label: z.ZodOptional<z.ZodString>;
158
- format: z.ZodOptional<z.ZodEnum<{
154
+ declare const traceAttributeDisplaySchema: z$1.ZodObject<{
155
+ key: z$1.ZodOptional<z$1.ZodString>;
156
+ path: z$1.ZodString;
157
+ label: z$1.ZodOptional<z$1.ZodString>;
158
+ format: z$1.ZodOptional<z$1.ZodEnum<{
159
159
  string: "string";
160
160
  number: "number";
161
161
  duration: "duration";
162
162
  json: "json";
163
163
  }>>;
164
- numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
165
- placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
164
+ numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
165
+ placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
166
166
  tree: "tree";
167
167
  detail: "detail";
168
168
  section: "section";
169
169
  }>>>;
170
- scope: z.ZodOptional<z.ZodEnum<{
170
+ scope: z$1.ZodOptional<z$1.ZodEnum<{
171
171
  self: "self";
172
172
  subtree: "subtree";
173
173
  }>>;
174
- mode: z.ZodOptional<z.ZodEnum<{
174
+ mode: z$1.ZodOptional<z$1.ZodEnum<{
175
175
  all: "all";
176
176
  last: "last";
177
177
  sum: "sum";
178
178
  }>>;
179
- }, z.core.$strip>;
179
+ }, z$1.core.$strip>;
180
180
  /**
181
181
  * Resolved trace display rule consumed by the UI.
182
182
  *
@@ -184,38 +184,38 @@ declare const traceAttributeDisplaySchema: z.ZodObject<{
184
184
  * control whether the value comes from the current span only or from the full
185
185
  * subtree, and how multiple matches are combined.
186
186
  */
187
- type TraceAttributeDisplay = z.infer<typeof traceAttributeDisplaySchema>;
187
+ type TraceAttributeDisplay = z$1.infer<typeof traceAttributeDisplaySchema>;
188
188
  /** Schema for trace display config after transforms have been resolved. */
189
- declare const traceDisplayConfigSchema: z.ZodObject<{
190
- attributes: z.ZodOptional<z.ZodArray<z.ZodObject<{
191
- key: z.ZodOptional<z.ZodString>;
192
- path: z.ZodString;
193
- label: z.ZodOptional<z.ZodString>;
194
- format: z.ZodOptional<z.ZodEnum<{
189
+ declare const traceDisplayConfigSchema: z$1.ZodObject<{
190
+ attributes: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
191
+ key: z$1.ZodOptional<z$1.ZodString>;
192
+ path: z$1.ZodString;
193
+ label: z$1.ZodOptional<z$1.ZodString>;
194
+ format: z$1.ZodOptional<z$1.ZodEnum<{
195
195
  string: "string";
196
196
  number: "number";
197
197
  duration: "duration";
198
198
  json: "json";
199
199
  }>>;
200
- numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
201
- placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
200
+ numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
201
+ placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
202
202
  tree: "tree";
203
203
  detail: "detail";
204
204
  section: "section";
205
205
  }>>>;
206
- scope: z.ZodOptional<z.ZodEnum<{
206
+ scope: z$1.ZodOptional<z$1.ZodEnum<{
207
207
  self: "self";
208
208
  subtree: "subtree";
209
209
  }>>;
210
- mode: z.ZodOptional<z.ZodEnum<{
210
+ mode: z$1.ZodOptional<z$1.ZodEnum<{
211
211
  all: "all";
212
212
  last: "last";
213
213
  sum: "sum";
214
214
  }>>;
215
- }, z.core.$strip>>>;
216
- }, z.core.$strip>;
215
+ }, z$1.core.$strip>>>;
216
+ }, z$1.core.$strip>;
217
217
  /** UI-ready trace display configuration attached to case details. */
218
- type TraceDisplayConfig = z.infer<typeof traceDisplayConfigSchema>;
218
+ type TraceDisplayConfig = z$1.infer<typeof traceDisplayConfigSchema>;
219
219
  /** Context passed to a `traceDisplay` transform while resolving a span value. */
220
220
  type TraceAttributeTransformContext = {
221
221
  value: unknown;
@@ -227,33 +227,33 @@ type TraceAttributeTransformContext = {
227
227
  */
228
228
  type TraceAttributeTransform = (ctx: TraceAttributeTransformContext) => unknown;
229
229
  /** Schema for authored trace display rules accepted from user config. */
230
- declare const traceAttributeDisplayInputSchema: z.ZodObject<{
231
- key: z.ZodOptional<z.ZodString>;
232
- path: z.ZodString;
233
- label: z.ZodOptional<z.ZodString>;
234
- format: z.ZodOptional<z.ZodEnum<{
230
+ declare const traceAttributeDisplayInputSchema: z$1.ZodObject<{
231
+ key: z$1.ZodOptional<z$1.ZodString>;
232
+ path: z$1.ZodString;
233
+ label: z$1.ZodOptional<z$1.ZodString>;
234
+ format: z$1.ZodOptional<z$1.ZodEnum<{
235
235
  string: "string";
236
236
  number: "number";
237
237
  duration: "duration";
238
238
  json: "json";
239
239
  }>>;
240
- numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
241
- placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
240
+ numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
241
+ placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
242
242
  tree: "tree";
243
243
  detail: "detail";
244
244
  section: "section";
245
245
  }>>>;
246
- scope: z.ZodOptional<z.ZodEnum<{
246
+ scope: z$1.ZodOptional<z$1.ZodEnum<{
247
247
  self: "self";
248
248
  subtree: "subtree";
249
249
  }>>;
250
- mode: z.ZodOptional<z.ZodEnum<{
250
+ mode: z$1.ZodOptional<z$1.ZodEnum<{
251
251
  all: "all";
252
252
  last: "last";
253
253
  sum: "sum";
254
254
  }>>;
255
- transform: z.ZodOptional<z.ZodCustom<TraceAttributeTransform, TraceAttributeTransform>>;
256
- }, z.core.$strip>;
255
+ transform: z$1.ZodOptional<z$1.ZodCustom<TraceAttributeTransform, TraceAttributeTransform>>;
256
+ }, z$1.core.$strip>;
257
257
  /**
258
258
  * Authored trace display rule accepted in eval definitions and config files.
259
259
  *
@@ -262,75 +262,112 @@ declare const traceAttributeDisplayInputSchema: z.ZodObject<{
262
262
  * `format: 'number'` values. `transform` runs in the
263
263
  * runner before the UI receives the resolved trace payload.
264
264
  */
265
- type TraceAttributeDisplayInput = z.infer<typeof traceAttributeDisplayInputSchema>;
265
+ type TraceAttributeDisplayInput = z$1.infer<typeof traceAttributeDisplayInputSchema>;
266
266
  /** Schema for authored trace display config in eval or workspace config. */
267
- declare const traceDisplayInputConfigSchema: z.ZodObject<{
268
- attributes: z.ZodOptional<z.ZodArray<z.ZodObject<{
269
- key: z.ZodOptional<z.ZodString>;
270
- path: z.ZodString;
271
- label: z.ZodOptional<z.ZodString>;
272
- format: z.ZodOptional<z.ZodEnum<{
267
+ declare const traceDisplayInputConfigSchema: z$1.ZodObject<{
268
+ attributes: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
269
+ key: z$1.ZodOptional<z$1.ZodString>;
270
+ path: z$1.ZodString;
271
+ label: z$1.ZodOptional<z$1.ZodString>;
272
+ format: z$1.ZodOptional<z$1.ZodEnum<{
273
273
  string: "string";
274
274
  number: "number";
275
275
  duration: "duration";
276
276
  json: "json";
277
277
  }>>;
278
- numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
279
- placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
278
+ numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
279
+ placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
280
280
  tree: "tree";
281
281
  detail: "detail";
282
282
  section: "section";
283
283
  }>>>;
284
- scope: z.ZodOptional<z.ZodEnum<{
284
+ scope: z$1.ZodOptional<z$1.ZodEnum<{
285
285
  self: "self";
286
286
  subtree: "subtree";
287
287
  }>>;
288
- mode: z.ZodOptional<z.ZodEnum<{
288
+ mode: z$1.ZodOptional<z$1.ZodEnum<{
289
289
  all: "all";
290
290
  last: "last";
291
291
  sum: "sum";
292
292
  }>>;
293
- transform: z.ZodOptional<z.ZodCustom<TraceAttributeTransform, TraceAttributeTransform>>;
294
- }, z.core.$strip>>>;
295
- }, z.core.$strip>;
293
+ transform: z$1.ZodOptional<z$1.ZodCustom<TraceAttributeTransform, TraceAttributeTransform>>;
294
+ }, z$1.core.$strip>>>;
295
+ }, z$1.core.$strip>;
296
296
  /** Trace display configuration authored by users in config or eval files. */
297
- type TraceDisplayInputConfig = z.infer<typeof traceDisplayInputConfigSchema>;
297
+ type TraceDisplayInputConfig = z$1.infer<typeof traceDisplayInputConfigSchema>;
298
+ /** Schema for an error attached to a trace span. */
299
+ declare const traceSpanErrorSchema: z$1.ZodObject<{
300
+ name: z$1.ZodOptional<z$1.ZodString>;
301
+ message: z$1.ZodString;
302
+ stack: z$1.ZodOptional<z$1.ZodString>;
303
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
304
+ }, z$1.core.$catchall<z$1.ZodUnknown>>;
305
+ /** Error payload stored on a trace span. */
306
+ type EvalTraceSpanError = z$1.infer<typeof traceSpanErrorSchema>;
307
+ /** Schema for a warning attached to a trace span. */
308
+ declare const traceSpanWarningSchema: z$1.ZodObject<{
309
+ name: z$1.ZodOptional<z$1.ZodString>;
310
+ message: z$1.ZodString;
311
+ stack: z$1.ZodOptional<z$1.ZodString>;
312
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
313
+ }, z$1.core.$catchall<z$1.ZodUnknown>>;
314
+ /** Warning payload stored on a trace span. */
315
+ type EvalTraceSpanWarning = z$1.infer<typeof traceSpanWarningSchema>;
298
316
  /** Schema for a persisted trace span captured during case execution. */
299
- declare const traceSpanSchema: z.ZodObject<{
300
- id: z.ZodString;
301
- parentId: z.ZodNullable<z.ZodString>;
302
- caseId: z.ZodString;
303
- kind: z.ZodString;
304
- name: z.ZodString;
305
- startedAt: z.ZodString;
306
- endedAt: z.ZodNullable<z.ZodString>;
307
- status: z.ZodEnum<{
317
+ declare const traceSpanSchema: z$1.ZodObject<{
318
+ id: z$1.ZodString;
319
+ parentId: z$1.ZodNullable<z$1.ZodString>;
320
+ caseId: z$1.ZodString;
321
+ kind: z$1.ZodString;
322
+ name: z$1.ZodString;
323
+ startedAt: z$1.ZodString;
324
+ endedAt: z$1.ZodNullable<z$1.ZodString>;
325
+ status: z$1.ZodEnum<{
308
326
  error: "error";
309
327
  running: "running";
310
328
  ok: "ok";
311
329
  cancelled: "cancelled";
312
330
  }>;
313
- attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
314
- error: z.ZodOptional<z.ZodObject<{
315
- name: z.ZodOptional<z.ZodString>;
316
- message: z.ZodString;
317
- stack: z.ZodOptional<z.ZodString>;
318
- }, z.core.$strip>>;
319
- }, z.core.$strip>;
331
+ attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
332
+ error: z$1.ZodOptional<z$1.ZodObject<{
333
+ name: z$1.ZodOptional<z$1.ZodString>;
334
+ message: z$1.ZodString;
335
+ stack: z$1.ZodOptional<z$1.ZodString>;
336
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
337
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>;
338
+ errors: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
339
+ name: z$1.ZodOptional<z$1.ZodString>;
340
+ message: z$1.ZodString;
341
+ stack: z$1.ZodOptional<z$1.ZodString>;
342
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
343
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
344
+ warning: z$1.ZodOptional<z$1.ZodObject<{
345
+ name: z$1.ZodOptional<z$1.ZodString>;
346
+ message: z$1.ZodString;
347
+ stack: z$1.ZodOptional<z$1.ZodString>;
348
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
349
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>;
350
+ warnings: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
351
+ name: z$1.ZodOptional<z$1.ZodString>;
352
+ message: z$1.ZodString;
353
+ stack: z$1.ZodOptional<z$1.ZodString>;
354
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
355
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
356
+ }, z$1.core.$strip>;
320
357
  /** Persisted trace span shape stored for each eval case run. */
321
- type EvalTraceSpan = z.infer<typeof traceSpanSchema>;
358
+ type EvalTraceSpan = z$1.infer<typeof traceSpanSchema>;
322
359
  //#endregion
323
360
  //#region ../shared/src/schemas/eval.d.ts
324
361
  /** Freshness signal derived from the latest relevant run plus git state. */
325
- declare const evalFreshnessStatusSchema: z.ZodEnum<{
362
+ declare const evalFreshnessStatusSchema: z$1.ZodEnum<{
326
363
  fresh: "fresh";
327
364
  stale: "stale";
328
365
  outdated: "outdated";
329
366
  }>;
330
367
  /** Freshness signal derived from the latest relevant run plus git state. */
331
- type EvalFreshnessStatus = z.infer<typeof evalFreshnessStatusSchema>;
368
+ type EvalFreshnessStatus = z$1.infer<typeof evalFreshnessStatusSchema>;
332
369
  /** Reducer used to collapse a column's per-case values into a single stat. */
333
- declare const evalStatAggregateSchema: z.ZodEnum<{
370
+ declare const evalStatAggregateSchema: z$1.ZodEnum<{
334
371
  avg: "avg";
335
372
  min: "min";
336
373
  max: "max";
@@ -338,30 +375,30 @@ declare const evalStatAggregateSchema: z.ZodEnum<{
338
375
  last: "last";
339
376
  }>;
340
377
  /** Reducer used to collapse a column's per-case values into a single stat. */
341
- type EvalStatAggregate = z.infer<typeof evalStatAggregateSchema>;
378
+ type EvalStatAggregate = z$1.infer<typeof evalStatAggregateSchema>;
342
379
  /**
343
380
  * One entry in the EvalCard stats row. Built-in kinds use latest run totals;
344
381
  * `column` aggregates a score or numeric output column across the latest run.
345
382
  */
346
- declare const evalStatItemSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
347
- kind: z.ZodLiteral<"cases">;
348
- }, z.core.$strip>, z.ZodObject<{
349
- kind: z.ZodLiteral<"passRate">;
350
- accent: z.ZodOptional<z.ZodBoolean>;
351
- }, z.core.$strip>, z.ZodObject<{
352
- kind: z.ZodLiteral<"duration">;
353
- }, z.core.$strip>, z.ZodObject<{
354
- kind: z.ZodLiteral<"column">;
355
- key: z.ZodString;
356
- label: z.ZodOptional<z.ZodString>;
357
- aggregate: z.ZodEnum<{
383
+ declare const evalStatItemSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
384
+ kind: z$1.ZodLiteral<"cases">;
385
+ }, z$1.core.$strip>, z$1.ZodObject<{
386
+ kind: z$1.ZodLiteral<"passRate">;
387
+ accent: z$1.ZodOptional<z$1.ZodBoolean>;
388
+ }, z$1.core.$strip>, z$1.ZodObject<{
389
+ kind: z$1.ZodLiteral<"duration">;
390
+ }, z$1.core.$strip>, z$1.ZodObject<{
391
+ kind: z$1.ZodLiteral<"column">;
392
+ key: z$1.ZodString;
393
+ label: z$1.ZodOptional<z$1.ZodString>;
394
+ aggregate: z$1.ZodEnum<{
358
395
  avg: "avg";
359
396
  min: "min";
360
397
  max: "max";
361
398
  sum: "sum";
362
399
  last: "last";
363
400
  }>;
364
- format: z.ZodOptional<z.ZodEnum<{
401
+ format: z$1.ZodOptional<z$1.ZodEnum<{
365
402
  number: "number";
366
403
  boolean: "boolean";
367
404
  file: "file";
@@ -375,30 +412,30 @@ declare const evalStatItemSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
375
412
  passFail: "passFail";
376
413
  stars: "stars";
377
414
  }>>;
378
- accent: z.ZodOptional<z.ZodBoolean>;
379
- }, z.core.$strip>], "kind">;
415
+ accent: z$1.ZodOptional<z$1.ZodBoolean>;
416
+ }, z$1.core.$strip>], "kind">;
380
417
  /** Single stat rendered in the EvalCard stats row. */
381
- type EvalStatItem = z.infer<typeof evalStatItemSchema>;
418
+ type EvalStatItem = z$1.infer<typeof evalStatItemSchema>;
382
419
  /** Ordered list of stats rendered in the EvalCard stats row. */
383
- declare const evalStatsConfigSchema: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
384
- kind: z.ZodLiteral<"cases">;
385
- }, z.core.$strip>, z.ZodObject<{
386
- kind: z.ZodLiteral<"passRate">;
387
- accent: z.ZodOptional<z.ZodBoolean>;
388
- }, z.core.$strip>, z.ZodObject<{
389
- kind: z.ZodLiteral<"duration">;
390
- }, z.core.$strip>, z.ZodObject<{
391
- kind: z.ZodLiteral<"column">;
392
- key: z.ZodString;
393
- label: z.ZodOptional<z.ZodString>;
394
- aggregate: z.ZodEnum<{
420
+ declare const evalStatsConfigSchema: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
421
+ kind: z$1.ZodLiteral<"cases">;
422
+ }, z$1.core.$strip>, z$1.ZodObject<{
423
+ kind: z$1.ZodLiteral<"passRate">;
424
+ accent: z$1.ZodOptional<z$1.ZodBoolean>;
425
+ }, z$1.core.$strip>, z$1.ZodObject<{
426
+ kind: z$1.ZodLiteral<"duration">;
427
+ }, z$1.core.$strip>, z$1.ZodObject<{
428
+ kind: z$1.ZodLiteral<"column">;
429
+ key: z$1.ZodString;
430
+ label: z$1.ZodOptional<z$1.ZodString>;
431
+ aggregate: z$1.ZodEnum<{
395
432
  avg: "avg";
396
433
  min: "min";
397
434
  max: "max";
398
435
  sum: "sum";
399
436
  last: "last";
400
437
  }>;
401
- format: z.ZodOptional<z.ZodEnum<{
438
+ format: z$1.ZodOptional<z$1.ZodEnum<{
402
439
  number: "number";
403
440
  boolean: "boolean";
404
441
  file: "file";
@@ -412,34 +449,34 @@ declare const evalStatsConfigSchema: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodOb
412
449
  passFail: "passFail";
413
450
  stars: "stars";
414
451
  }>>;
415
- accent: z.ZodOptional<z.ZodBoolean>;
416
- }, z.core.$strip>], "kind">>;
452
+ accent: z$1.ZodOptional<z$1.ZodBoolean>;
453
+ }, z$1.core.$strip>], "kind">>;
417
454
  /** Ordered list of stats rendered in the EvalCard stats row. */
418
- type EvalStatsConfig = z.infer<typeof evalStatsConfigSchema>;
455
+ type EvalStatsConfig = z$1.infer<typeof evalStatsConfigSchema>;
419
456
  /** Schema summarizing a discovered eval for list and overview screens. */
420
- declare const evalSummarySchema: z.ZodObject<{
421
- id: z.ZodString;
422
- title: z.ZodOptional<z.ZodString>;
423
- filePath: z.ZodString;
424
- stale: z.ZodBoolean;
425
- outdated: z.ZodBoolean;
426
- freshnessStatus: z.ZodEnum<{
457
+ declare const evalSummarySchema: z$1.ZodObject<{
458
+ id: z$1.ZodString;
459
+ title: z$1.ZodOptional<z$1.ZodString>;
460
+ filePath: z$1.ZodString;
461
+ stale: z$1.ZodBoolean;
462
+ outdated: z$1.ZodBoolean;
463
+ freshnessStatus: z$1.ZodEnum<{
427
464
  fresh: "fresh";
428
465
  stale: "stale";
429
466
  outdated: "outdated";
430
467
  }>;
431
- latestRunAt: z.ZodNullable<z.ZodString>;
432
- latestRunCommitSha: z.ZodNullable<z.ZodString>;
433
- currentCommitSha: z.ZodNullable<z.ZodString>;
434
- columnDefs: z.ZodArray<z.ZodObject<{
435
- key: z.ZodString;
436
- label: z.ZodString;
437
- kind: z.ZodEnum<{
468
+ latestRunAt: z$1.ZodNullable<z$1.ZodString>;
469
+ latestRunCommitSha: z$1.ZodNullable<z$1.ZodString>;
470
+ currentCommitSha: z$1.ZodNullable<z$1.ZodString>;
471
+ columnDefs: z$1.ZodArray<z$1.ZodObject<{
472
+ key: z$1.ZodString;
473
+ label: z$1.ZodString;
474
+ kind: z$1.ZodEnum<{
438
475
  string: "string";
439
476
  number: "number";
440
477
  boolean: "boolean";
441
478
  }>;
442
- format: z.ZodOptional<z.ZodEnum<{
479
+ format: z$1.ZodOptional<z$1.ZodEnum<{
443
480
  number: "number";
444
481
  boolean: "boolean";
445
482
  file: "file";
@@ -453,21 +490,21 @@ declare const evalSummarySchema: z.ZodObject<{
453
490
  passFail: "passFail";
454
491
  stars: "stars";
455
492
  }>>;
456
- numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
457
- isScore: z.ZodOptional<z.ZodBoolean>;
458
- isManualScore: z.ZodOptional<z.ZodBoolean>;
459
- passThreshold: z.ZodOptional<z.ZodNumber>;
460
- maxStars: z.ZodOptional<z.ZodNumber>;
461
- hideInTable: z.ZodOptional<z.ZodBoolean>;
462
- sortable: z.ZodOptional<z.ZodBoolean>;
463
- align: z.ZodOptional<z.ZodEnum<{
493
+ numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
494
+ isScore: z$1.ZodOptional<z$1.ZodBoolean>;
495
+ isManualScore: z$1.ZodOptional<z$1.ZodBoolean>;
496
+ passThreshold: z$1.ZodOptional<z$1.ZodNumber>;
497
+ maxStars: z$1.ZodOptional<z$1.ZodNumber>;
498
+ hideInTable: z$1.ZodOptional<z$1.ZodBoolean>;
499
+ sortable: z$1.ZodOptional<z$1.ZodBoolean>;
500
+ align: z$1.ZodOptional<z$1.ZodEnum<{
464
501
  left: "left";
465
502
  center: "center";
466
503
  right: "right";
467
504
  }>>;
468
- }, z.core.$strip>>;
469
- caseCount: z.ZodNullable<z.ZodNumber>;
470
- lastRunStatus: z.ZodNullable<z.ZodEnum<{
505
+ }, z$1.core.$strip>>;
506
+ caseCount: z$1.ZodNullable<z$1.ZodNumber>;
507
+ lastRunStatus: z$1.ZodNullable<z$1.ZodEnum<{
471
508
  error: "error";
472
509
  pass: "pass";
473
510
  fail: "fail";
@@ -475,25 +512,25 @@ declare const evalSummarySchema: z.ZodObject<{
475
512
  cancelled: "cancelled";
476
513
  unscored: "unscored";
477
514
  }>>;
478
- stats: z.ZodOptional<z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
479
- kind: z.ZodLiteral<"cases">;
480
- }, z.core.$strip>, z.ZodObject<{
481
- kind: z.ZodLiteral<"passRate">;
482
- accent: z.ZodOptional<z.ZodBoolean>;
483
- }, z.core.$strip>, z.ZodObject<{
484
- kind: z.ZodLiteral<"duration">;
485
- }, z.core.$strip>, z.ZodObject<{
486
- kind: z.ZodLiteral<"column">;
487
- key: z.ZodString;
488
- label: z.ZodOptional<z.ZodString>;
489
- aggregate: z.ZodEnum<{
515
+ stats: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
516
+ kind: z$1.ZodLiteral<"cases">;
517
+ }, z$1.core.$strip>, z$1.ZodObject<{
518
+ kind: z$1.ZodLiteral<"passRate">;
519
+ accent: z$1.ZodOptional<z$1.ZodBoolean>;
520
+ }, z$1.core.$strip>, z$1.ZodObject<{
521
+ kind: z$1.ZodLiteral<"duration">;
522
+ }, z$1.core.$strip>, z$1.ZodObject<{
523
+ kind: z$1.ZodLiteral<"column">;
524
+ key: z$1.ZodString;
525
+ label: z$1.ZodOptional<z$1.ZodString>;
526
+ aggregate: z$1.ZodEnum<{
490
527
  avg: "avg";
491
528
  min: "min";
492
529
  max: "max";
493
530
  sum: "sum";
494
531
  last: "last";
495
532
  }>;
496
- format: z.ZodOptional<z.ZodEnum<{
533
+ format: z$1.ZodOptional<z$1.ZodEnum<{
497
534
  number: "number";
498
535
  boolean: "boolean";
499
536
  file: "file";
@@ -507,23 +544,23 @@ declare const evalSummarySchema: z.ZodObject<{
507
544
  passFail: "passFail";
508
545
  stars: "stars";
509
546
  }>>;
510
- accent: z.ZodOptional<z.ZodBoolean>;
511
- }, z.core.$strip>], "kind">>>;
512
- charts: z.ZodOptional<z.ZodArray<z.ZodObject<{
513
- heading: z.ZodOptional<z.ZodString>;
514
- type: z.ZodEnum<{
547
+ accent: z$1.ZodOptional<z$1.ZodBoolean>;
548
+ }, z$1.core.$strip>], "kind">>>;
549
+ charts: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
550
+ heading: z$1.ZodOptional<z$1.ZodString>;
551
+ type: z$1.ZodEnum<{
515
552
  area: "area";
516
553
  line: "line";
517
554
  bar: "bar";
518
555
  }>;
519
- metrics: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
520
- source: z.ZodLiteral<"builtin">;
521
- metric: z.ZodEnum<{
556
+ metrics: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
557
+ source: z$1.ZodLiteral<"builtin">;
558
+ metric: z$1.ZodEnum<{
522
559
  passRate: "passRate";
523
560
  durationMs: "durationMs";
524
561
  }>;
525
- label: z.ZodOptional<z.ZodString>;
526
- color: z.ZodOptional<z.ZodEnum<{
562
+ label: z$1.ZodOptional<z$1.ZodString>;
563
+ color: z$1.ZodOptional<z$1.ZodEnum<{
527
564
  success: "success";
528
565
  accent: "accent";
529
566
  error: "error";
@@ -531,14 +568,14 @@ declare const evalSummarySchema: z.ZodObject<{
531
568
  warning: "warning";
532
569
  textMuted: "textMuted";
533
570
  }>>;
534
- axis: z.ZodOptional<z.ZodEnum<{
571
+ axis: z$1.ZodOptional<z$1.ZodEnum<{
535
572
  left: "left";
536
573
  right: "right";
537
574
  }>>;
538
- }, z.core.$strip>, z.ZodObject<{
539
- source: z.ZodLiteral<"column">;
540
- key: z.ZodString;
541
- aggregate: z.ZodEnum<{
575
+ }, z$1.core.$strip>, z$1.ZodObject<{
576
+ source: z$1.ZodLiteral<"column">;
577
+ key: z$1.ZodString;
578
+ aggregate: z$1.ZodEnum<{
542
579
  avg: "avg";
543
580
  min: "min";
544
581
  max: "max";
@@ -546,8 +583,8 @@ declare const evalSummarySchema: z.ZodObject<{
546
583
  latest: "latest";
547
584
  passThresholdRate: "passThresholdRate";
548
585
  }>;
549
- label: z.ZodOptional<z.ZodString>;
550
- color: z.ZodOptional<z.ZodEnum<{
586
+ label: z$1.ZodOptional<z$1.ZodString>;
587
+ color: z$1.ZodOptional<z$1.ZodEnum<{
551
588
  success: "success";
552
589
  accent: "accent";
553
590
  error: "error";
@@ -555,32 +592,32 @@ declare const evalSummarySchema: z.ZodObject<{
555
592
  warning: "warning";
556
593
  textMuted: "textMuted";
557
594
  }>>;
558
- axis: z.ZodOptional<z.ZodEnum<{
595
+ axis: z$1.ZodOptional<z$1.ZodEnum<{
559
596
  left: "left";
560
597
  right: "right";
561
598
  }>>;
562
- }, z.core.$strip>], "source">>;
563
- yDomain: z.ZodOptional<z.ZodObject<{
564
- left: z.ZodOptional<z.ZodObject<{
565
- min: z.ZodOptional<z.ZodNumber>;
566
- max: z.ZodOptional<z.ZodNumber>;
567
- }, z.core.$strip>>;
568
- right: z.ZodOptional<z.ZodObject<{
569
- min: z.ZodOptional<z.ZodNumber>;
570
- max: z.ZodOptional<z.ZodNumber>;
571
- }, z.core.$strip>>;
572
- }, z.core.$strip>>;
573
- tooltipExtras: z.ZodOptional<z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
574
- source: z.ZodLiteral<"builtin">;
575
- metric: z.ZodEnum<{
599
+ }, z$1.core.$strip>], "source">>;
600
+ yDomain: z$1.ZodOptional<z$1.ZodObject<{
601
+ left: z$1.ZodOptional<z$1.ZodObject<{
602
+ min: z$1.ZodOptional<z$1.ZodNumber>;
603
+ max: z$1.ZodOptional<z$1.ZodNumber>;
604
+ }, z$1.core.$strip>>;
605
+ right: z$1.ZodOptional<z$1.ZodObject<{
606
+ min: z$1.ZodOptional<z$1.ZodNumber>;
607
+ max: z$1.ZodOptional<z$1.ZodNumber>;
608
+ }, z$1.core.$strip>>;
609
+ }, z$1.core.$strip>>;
610
+ tooltipExtras: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
611
+ source: z$1.ZodLiteral<"builtin">;
612
+ metric: z$1.ZodEnum<{
576
613
  passRate: "passRate";
577
614
  durationMs: "durationMs";
578
615
  }>;
579
- label: z.ZodOptional<z.ZodString>;
580
- }, z.core.$strip>, z.ZodObject<{
581
- source: z.ZodLiteral<"column">;
582
- key: z.ZodString;
583
- aggregate: z.ZodEnum<{
616
+ label: z$1.ZodOptional<z$1.ZodString>;
617
+ }, z$1.core.$strip>, z$1.ZodObject<{
618
+ source: z$1.ZodLiteral<"column">;
619
+ key: z$1.ZodString;
620
+ aggregate: z$1.ZodEnum<{
584
621
  avg: "avg";
585
622
  min: "min";
586
623
  max: "max";
@@ -588,17 +625,17 @@ declare const evalSummarySchema: z.ZodObject<{
588
625
  latest: "latest";
589
626
  passThresholdRate: "passThresholdRate";
590
627
  }>;
591
- label: z.ZodOptional<z.ZodString>;
592
- }, z.core.$strip>], "source">>>;
593
- }, z.core.$strip>>>;
594
- }, z.core.$strip>;
628
+ label: z$1.ZodOptional<z$1.ZodString>;
629
+ }, z$1.core.$strip>], "source">>>;
630
+ }, z$1.core.$strip>>>;
631
+ }, z$1.core.$strip>;
595
632
  /** Metadata shown for one discovered eval in the explorer UI. */
596
- type EvalSummary = z.infer<typeof evalSummarySchema>;
633
+ type EvalSummary = z$1.infer<typeof evalSummarySchema>;
597
634
  /** Schema for one case row in an eval run result table. */
598
- declare const caseRowSchema: z.ZodObject<{
599
- caseId: z.ZodString;
600
- evalId: z.ZodString;
601
- status: z.ZodEnum<{
635
+ declare const caseRowSchema: z$1.ZodObject<{
636
+ caseId: z$1.ZodString;
637
+ evalId: z$1.ZodString;
638
+ status: z$1.ZodEnum<{
602
639
  error: "error";
603
640
  pass: "pass";
604
641
  fail: "fail";
@@ -606,88 +643,107 @@ declare const caseRowSchema: z.ZodObject<{
606
643
  cancelled: "cancelled";
607
644
  pending: "pending";
608
645
  }>;
609
- latencyMs: z.ZodNullable<z.ZodNumber>;
610
- costUsd: z.ZodOptional<z.ZodNullable<z.ZodNumber>>;
611
- columns: z.ZodRecord<z.ZodString, z.ZodUnion<readonly [z.ZodType<string | number | boolean | Record<string, unknown> | unknown[] | null, unknown, z.core.$ZodTypeInternals<string | number | boolean | Record<string, unknown> | unknown[] | null, unknown>>, z.ZodUnion<readonly [z.ZodObject<{
612
- source: z.ZodLiteral<"repo">;
613
- path: z.ZodString;
614
- mimeType: z.ZodOptional<z.ZodString>;
615
- }, z.core.$strip>, z.ZodObject<{
616
- source: z.ZodLiteral<"run">;
617
- artifactId: z.ZodString;
618
- mimeType: z.ZodString;
619
- fileName: z.ZodOptional<z.ZodString>;
620
- }, z.core.$strip>]>]>>;
621
- trial: z.ZodNumber;
622
- }, z.core.$strip>;
646
+ latencyMs: z$1.ZodNullable<z$1.ZodNumber>;
647
+ costUsd: z$1.ZodOptional<z$1.ZodNullable<z$1.ZodNumber>>;
648
+ 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<{
649
+ source: z$1.ZodLiteral<"repo">;
650
+ path: z$1.ZodString;
651
+ mimeType: z$1.ZodOptional<z$1.ZodString>;
652
+ }, z$1.core.$strip>, z$1.ZodObject<{
653
+ source: z$1.ZodLiteral<"run">;
654
+ artifactId: z$1.ZodString;
655
+ mimeType: z$1.ZodString;
656
+ fileName: z$1.ZodOptional<z$1.ZodString>;
657
+ }, z$1.core.$strip>]>]>>;
658
+ trial: z$1.ZodNumber;
659
+ }, z$1.core.$strip>;
623
660
  /** Flattened per-case row rendered in run tables and streamed updates. */
624
- type CaseRow = z.infer<typeof caseRowSchema>;
661
+ type CaseRow = z$1.infer<typeof caseRowSchema>;
625
662
  /** Structured assertion failure metadata captured for one case run. */
626
- declare const assertionFailureSchema: z.ZodObject<{
627
- message: z.ZodString;
628
- stack: z.ZodOptional<z.ZodString>;
629
- }, z.core.$strip>;
663
+ declare const assertionFailureSchema: z$1.ZodObject<{
664
+ message: z$1.ZodString;
665
+ stack: z$1.ZodOptional<z$1.ZodString>;
666
+ }, z$1.core.$strip>;
630
667
  /** Assertion failure metadata captured for one case run. */
631
- type AssertionFailure = z.infer<typeof assertionFailureSchema>;
668
+ type AssertionFailure = z$1.infer<typeof assertionFailureSchema>;
632
669
  /** Trace payload captured while computing one score for a case. */
633
- declare const scoreTraceSchema: z.ZodObject<{
634
- trace: z.ZodArray<z.ZodObject<{
635
- id: z.ZodString;
636
- parentId: z.ZodNullable<z.ZodString>;
637
- caseId: z.ZodString;
638
- kind: z.ZodString;
639
- name: z.ZodString;
640
- startedAt: z.ZodString;
641
- endedAt: z.ZodNullable<z.ZodString>;
642
- status: z.ZodEnum<{
670
+ declare const scoreTraceSchema: z$1.ZodObject<{
671
+ trace: z$1.ZodArray<z$1.ZodObject<{
672
+ id: z$1.ZodString;
673
+ parentId: z$1.ZodNullable<z$1.ZodString>;
674
+ caseId: z$1.ZodString;
675
+ kind: z$1.ZodString;
676
+ name: z$1.ZodString;
677
+ startedAt: z$1.ZodString;
678
+ endedAt: z$1.ZodNullable<z$1.ZodString>;
679
+ status: z$1.ZodEnum<{
643
680
  error: "error";
644
681
  running: "running";
645
682
  cancelled: "cancelled";
646
683
  ok: "ok";
647
684
  }>;
648
- attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
649
- error: z.ZodOptional<z.ZodObject<{
650
- name: z.ZodOptional<z.ZodString>;
651
- message: z.ZodString;
652
- stack: z.ZodOptional<z.ZodString>;
653
- }, z.core.$strip>>;
654
- }, z.core.$strip>>;
655
- traceDisplay: z.ZodObject<{
656
- attributes: z.ZodOptional<z.ZodArray<z.ZodObject<{
657
- key: z.ZodOptional<z.ZodString>;
658
- path: z.ZodString;
659
- label: z.ZodOptional<z.ZodString>;
660
- format: z.ZodOptional<z.ZodEnum<{
685
+ attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
686
+ error: z$1.ZodOptional<z$1.ZodObject<{
687
+ name: z$1.ZodOptional<z$1.ZodString>;
688
+ message: z$1.ZodString;
689
+ stack: z$1.ZodOptional<z$1.ZodString>;
690
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
691
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>;
692
+ errors: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
693
+ name: z$1.ZodOptional<z$1.ZodString>;
694
+ message: z$1.ZodString;
695
+ stack: z$1.ZodOptional<z$1.ZodString>;
696
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
697
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
698
+ warning: z$1.ZodOptional<z$1.ZodObject<{
699
+ name: z$1.ZodOptional<z$1.ZodString>;
700
+ message: z$1.ZodString;
701
+ stack: z$1.ZodOptional<z$1.ZodString>;
702
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
703
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>;
704
+ warnings: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
705
+ name: z$1.ZodOptional<z$1.ZodString>;
706
+ message: z$1.ZodString;
707
+ stack: z$1.ZodOptional<z$1.ZodString>;
708
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
709
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
710
+ }, z$1.core.$strip>>;
711
+ traceDisplay: z$1.ZodObject<{
712
+ attributes: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
713
+ key: z$1.ZodOptional<z$1.ZodString>;
714
+ path: z$1.ZodString;
715
+ label: z$1.ZodOptional<z$1.ZodString>;
716
+ format: z$1.ZodOptional<z$1.ZodEnum<{
661
717
  string: "string";
662
718
  number: "number";
663
719
  duration: "duration";
664
720
  json: "json";
665
721
  }>>;
666
- numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
667
- placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
722
+ numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
723
+ placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
668
724
  tree: "tree";
669
725
  detail: "detail";
670
726
  section: "section";
671
727
  }>>>;
672
- scope: z.ZodOptional<z.ZodEnum<{
728
+ scope: z$1.ZodOptional<z$1.ZodEnum<{
673
729
  self: "self";
674
730
  subtree: "subtree";
675
731
  }>>;
676
- mode: z.ZodOptional<z.ZodEnum<{
732
+ mode: z$1.ZodOptional<z$1.ZodEnum<{
677
733
  sum: "sum";
678
734
  last: "last";
679
735
  all: "all";
680
736
  }>>;
681
- }, z.core.$strip>>>;
682
- }, z.core.$strip>;
683
- }, z.core.$strip>;
737
+ }, z$1.core.$strip>>>;
738
+ }, z$1.core.$strip>;
739
+ }, z$1.core.$strip>;
684
740
  /** Trace payload captured while computing one score for a case. */
685
- type ScoreTrace = z.infer<typeof scoreTraceSchema>;
741
+ type ScoreTrace = z$1.infer<typeof scoreTraceSchema>;
686
742
  /** Schema for the detailed payload shown when opening a specific case. */
687
- declare const caseDetailSchema: z.ZodObject<{
688
- caseId: z.ZodString;
689
- evalId: z.ZodString;
690
- status: z.ZodEnum<{
743
+ declare const caseDetailSchema: z$1.ZodObject<{
744
+ caseId: z$1.ZodString;
745
+ evalId: z$1.ZodString;
746
+ status: z$1.ZodEnum<{
691
747
  error: "error";
692
748
  pass: "pass";
693
749
  fail: "fail";
@@ -695,148 +751,186 @@ declare const caseDetailSchema: z.ZodObject<{
695
751
  cancelled: "cancelled";
696
752
  pending: "pending";
697
753
  }>;
698
- input: z.ZodUnknown;
699
- trace: z.ZodArray<z.ZodObject<{
700
- id: z.ZodString;
701
- parentId: z.ZodNullable<z.ZodString>;
702
- caseId: z.ZodString;
703
- kind: z.ZodString;
704
- name: z.ZodString;
705
- startedAt: z.ZodString;
706
- endedAt: z.ZodNullable<z.ZodString>;
707
- status: z.ZodEnum<{
754
+ input: z$1.ZodUnknown;
755
+ trace: z$1.ZodArray<z$1.ZodObject<{
756
+ id: z$1.ZodString;
757
+ parentId: z$1.ZodNullable<z$1.ZodString>;
758
+ caseId: z$1.ZodString;
759
+ kind: z$1.ZodString;
760
+ name: z$1.ZodString;
761
+ startedAt: z$1.ZodString;
762
+ endedAt: z$1.ZodNullable<z$1.ZodString>;
763
+ status: z$1.ZodEnum<{
708
764
  error: "error";
709
765
  running: "running";
710
766
  cancelled: "cancelled";
711
767
  ok: "ok";
712
768
  }>;
713
- attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
714
- error: z.ZodOptional<z.ZodObject<{
715
- name: z.ZodOptional<z.ZodString>;
716
- message: z.ZodString;
717
- stack: z.ZodOptional<z.ZodString>;
718
- }, z.core.$strip>>;
719
- }, z.core.$strip>>;
720
- traceDisplay: z.ZodObject<{
721
- attributes: z.ZodOptional<z.ZodArray<z.ZodObject<{
722
- key: z.ZodOptional<z.ZodString>;
723
- path: z.ZodString;
724
- label: z.ZodOptional<z.ZodString>;
725
- format: z.ZodOptional<z.ZodEnum<{
769
+ attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
770
+ error: z$1.ZodOptional<z$1.ZodObject<{
771
+ name: z$1.ZodOptional<z$1.ZodString>;
772
+ message: z$1.ZodString;
773
+ stack: z$1.ZodOptional<z$1.ZodString>;
774
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
775
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>;
776
+ errors: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
777
+ name: z$1.ZodOptional<z$1.ZodString>;
778
+ message: z$1.ZodString;
779
+ stack: z$1.ZodOptional<z$1.ZodString>;
780
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
781
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
782
+ warning: z$1.ZodOptional<z$1.ZodObject<{
783
+ name: z$1.ZodOptional<z$1.ZodString>;
784
+ message: z$1.ZodString;
785
+ stack: z$1.ZodOptional<z$1.ZodString>;
786
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
787
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>;
788
+ warnings: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
789
+ name: z$1.ZodOptional<z$1.ZodString>;
790
+ message: z$1.ZodString;
791
+ stack: z$1.ZodOptional<z$1.ZodString>;
792
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
793
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
794
+ }, z$1.core.$strip>>;
795
+ traceDisplay: z$1.ZodObject<{
796
+ attributes: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
797
+ key: z$1.ZodOptional<z$1.ZodString>;
798
+ path: z$1.ZodString;
799
+ label: z$1.ZodOptional<z$1.ZodString>;
800
+ format: z$1.ZodOptional<z$1.ZodEnum<{
726
801
  string: "string";
727
802
  number: "number";
728
803
  duration: "duration";
729
804
  json: "json";
730
805
  }>>;
731
- numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
732
- placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
806
+ numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
807
+ placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
733
808
  tree: "tree";
734
809
  detail: "detail";
735
810
  section: "section";
736
811
  }>>>;
737
- scope: z.ZodOptional<z.ZodEnum<{
812
+ scope: z$1.ZodOptional<z$1.ZodEnum<{
738
813
  self: "self";
739
814
  subtree: "subtree";
740
815
  }>>;
741
- mode: z.ZodOptional<z.ZodEnum<{
816
+ mode: z$1.ZodOptional<z$1.ZodEnum<{
742
817
  sum: "sum";
743
818
  last: "last";
744
819
  all: "all";
745
820
  }>>;
746
- }, z.core.$strip>>>;
747
- }, z.core.$strip>;
748
- scoringTraces: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodObject<{
749
- trace: z.ZodArray<z.ZodObject<{
750
- id: z.ZodString;
751
- parentId: z.ZodNullable<z.ZodString>;
752
- caseId: z.ZodString;
753
- kind: z.ZodString;
754
- name: z.ZodString;
755
- startedAt: z.ZodString;
756
- endedAt: z.ZodNullable<z.ZodString>;
757
- status: z.ZodEnum<{
821
+ }, z$1.core.$strip>>>;
822
+ }, z$1.core.$strip>;
823
+ scoringTraces: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodObject<{
824
+ trace: z$1.ZodArray<z$1.ZodObject<{
825
+ id: z$1.ZodString;
826
+ parentId: z$1.ZodNullable<z$1.ZodString>;
827
+ caseId: z$1.ZodString;
828
+ kind: z$1.ZodString;
829
+ name: z$1.ZodString;
830
+ startedAt: z$1.ZodString;
831
+ endedAt: z$1.ZodNullable<z$1.ZodString>;
832
+ status: z$1.ZodEnum<{
758
833
  error: "error";
759
834
  running: "running";
760
835
  cancelled: "cancelled";
761
836
  ok: "ok";
762
837
  }>;
763
- attributes: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
764
- error: z.ZodOptional<z.ZodObject<{
765
- name: z.ZodOptional<z.ZodString>;
766
- message: z.ZodString;
767
- stack: z.ZodOptional<z.ZodString>;
768
- }, z.core.$strip>>;
769
- }, z.core.$strip>>;
770
- traceDisplay: z.ZodObject<{
771
- attributes: z.ZodOptional<z.ZodArray<z.ZodObject<{
772
- key: z.ZodOptional<z.ZodString>;
773
- path: z.ZodString;
774
- label: z.ZodOptional<z.ZodString>;
775
- format: z.ZodOptional<z.ZodEnum<{
838
+ attributes: z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>>;
839
+ error: z$1.ZodOptional<z$1.ZodObject<{
840
+ name: z$1.ZodOptional<z$1.ZodString>;
841
+ message: z$1.ZodString;
842
+ stack: z$1.ZodOptional<z$1.ZodString>;
843
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
844
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>;
845
+ errors: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
846
+ name: z$1.ZodOptional<z$1.ZodString>;
847
+ message: z$1.ZodString;
848
+ stack: z$1.ZodOptional<z$1.ZodString>;
849
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
850
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
851
+ warning: z$1.ZodOptional<z$1.ZodObject<{
852
+ name: z$1.ZodOptional<z$1.ZodString>;
853
+ message: z$1.ZodString;
854
+ stack: z$1.ZodOptional<z$1.ZodString>;
855
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
856
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>;
857
+ warnings: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
858
+ name: z$1.ZodOptional<z$1.ZodString>;
859
+ message: z$1.ZodString;
860
+ stack: z$1.ZodOptional<z$1.ZodString>;
861
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
862
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
863
+ }, z$1.core.$strip>>;
864
+ traceDisplay: z$1.ZodObject<{
865
+ attributes: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
866
+ key: z$1.ZodOptional<z$1.ZodString>;
867
+ path: z$1.ZodString;
868
+ label: z$1.ZodOptional<z$1.ZodString>;
869
+ format: z$1.ZodOptional<z$1.ZodEnum<{
776
870
  string: "string";
777
871
  number: "number";
778
872
  duration: "duration";
779
873
  json: "json";
780
874
  }>>;
781
- numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
782
- placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
875
+ numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
876
+ placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
783
877
  tree: "tree";
784
878
  detail: "detail";
785
879
  section: "section";
786
880
  }>>>;
787
- scope: z.ZodOptional<z.ZodEnum<{
881
+ scope: z$1.ZodOptional<z$1.ZodEnum<{
788
882
  self: "self";
789
883
  subtree: "subtree";
790
884
  }>>;
791
- mode: z.ZodOptional<z.ZodEnum<{
885
+ mode: z$1.ZodOptional<z$1.ZodEnum<{
792
886
  sum: "sum";
793
887
  last: "last";
794
888
  all: "all";
795
889
  }>>;
796
- }, z.core.$strip>>>;
797
- }, z.core.$strip>;
798
- }, z.core.$strip>>>;
799
- columns: z.ZodRecord<z.ZodString, z.ZodUnion<readonly [z.ZodType<string | number | boolean | Record<string, unknown> | unknown[] | null, unknown, z.core.$ZodTypeInternals<string | number | boolean | Record<string, unknown> | unknown[] | null, unknown>>, z.ZodUnion<readonly [z.ZodObject<{
800
- source: z.ZodLiteral<"repo">;
801
- path: z.ZodString;
802
- mimeType: z.ZodOptional<z.ZodString>;
803
- }, z.core.$strip>, z.ZodObject<{
804
- source: z.ZodLiteral<"run">;
805
- artifactId: z.ZodString;
806
- mimeType: z.ZodString;
807
- fileName: z.ZodOptional<z.ZodString>;
808
- }, z.core.$strip>]>]>>;
809
- assertionFailures: z.ZodArray<z.ZodUnion<readonly [z.ZodObject<{
810
- message: z.ZodString;
811
- stack: z.ZodOptional<z.ZodString>;
812
- }, z.core.$strip>, z.ZodPipe<z.ZodString, z.ZodTransform<{
890
+ }, z$1.core.$strip>>>;
891
+ }, z$1.core.$strip>;
892
+ }, z$1.core.$strip>>>;
893
+ 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<{
894
+ source: z$1.ZodLiteral<"repo">;
895
+ path: z$1.ZodString;
896
+ mimeType: z$1.ZodOptional<z$1.ZodString>;
897
+ }, z$1.core.$strip>, z$1.ZodObject<{
898
+ source: z$1.ZodLiteral<"run">;
899
+ artifactId: z$1.ZodString;
900
+ mimeType: z$1.ZodString;
901
+ fileName: z$1.ZodOptional<z$1.ZodString>;
902
+ }, z$1.core.$strip>]>]>>;
903
+ assertionFailures: z$1.ZodArray<z$1.ZodUnion<readonly [z$1.ZodObject<{
904
+ message: z$1.ZodString;
905
+ stack: z$1.ZodOptional<z$1.ZodString>;
906
+ }, z$1.core.$strip>, z$1.ZodPipe<z$1.ZodString, z$1.ZodTransform<{
813
907
  message: string;
814
908
  stack?: string | undefined;
815
909
  }, string>>]>>;
816
- error: z.ZodNullable<z.ZodObject<{
817
- name: z.ZodOptional<z.ZodString>;
818
- message: z.ZodString;
819
- stack: z.ZodOptional<z.ZodString>;
820
- }, z.core.$strip>>;
821
- trial: z.ZodNumber;
822
- }, z.core.$strip>;
910
+ error: z$1.ZodNullable<z$1.ZodObject<{
911
+ name: z$1.ZodOptional<z$1.ZodString>;
912
+ message: z$1.ZodString;
913
+ stack: z$1.ZodOptional<z$1.ZodString>;
914
+ }, z$1.core.$strip>>;
915
+ trial: z$1.ZodNumber;
916
+ }, z$1.core.$strip>;
823
917
  /** Full case payload including inputs, trace, outputs, and failures. */
824
- type CaseDetail = z.infer<typeof caseDetailSchema>;
918
+ type CaseDetail = z$1.infer<typeof caseDetailSchema>;
825
919
  //#endregion
826
920
  //#region ../shared/src/schemas/chart.d.ts
827
921
  /** Chart type rendered for a single eval history chart. */
828
- declare const evalChartTypeSchema: z.ZodEnum<{
922
+ declare const evalChartTypeSchema: z$1.ZodEnum<{
829
923
  area: "area";
830
924
  line: "line";
831
925
  bar: "bar";
832
926
  }>;
833
927
  /** Chart type rendered for a single eval history chart. */
834
- type EvalChartType = z.infer<typeof evalChartTypeSchema>;
928
+ type EvalChartType = z$1.infer<typeof evalChartTypeSchema>;
835
929
  /**
836
930
  * Run-level metric sourced from the aggregated `RunSummary` for a run, rather
837
931
  * than from a per-case column.
838
932
  */
839
- declare const evalChartBuiltinMetricSchema: z.ZodEnum<{
933
+ declare const evalChartBuiltinMetricSchema: z$1.ZodEnum<{
840
934
  passRate: "passRate";
841
935
  durationMs: "durationMs";
842
936
  }>;
@@ -844,9 +938,9 @@ declare const evalChartBuiltinMetricSchema: z.ZodEnum<{
844
938
  * Run-level metric sourced from the aggregated `RunSummary` for a run, rather
845
939
  * than from a per-case column.
846
940
  */
847
- type EvalChartBuiltinMetric = z.infer<typeof evalChartBuiltinMetricSchema>;
941
+ type EvalChartBuiltinMetric = z$1.infer<typeof evalChartBuiltinMetricSchema>;
848
942
  /** Reducer applied to a numeric column across all cases of a single run. */
849
- declare const evalChartAggregateSchema: z.ZodEnum<{
943
+ declare const evalChartAggregateSchema: z$1.ZodEnum<{
850
944
  avg: "avg";
851
945
  sum: "sum";
852
946
  min: "min";
@@ -855,12 +949,12 @@ declare const evalChartAggregateSchema: z.ZodEnum<{
855
949
  passThresholdRate: "passThresholdRate";
856
950
  }>;
857
951
  /** Reducer applied to a numeric column across all cases of a single run. */
858
- type EvalChartAggregate = z.infer<typeof evalChartAggregateSchema>;
952
+ type EvalChartAggregate = z$1.infer<typeof evalChartAggregateSchema>;
859
953
  /**
860
954
  * Semantic color token resolved to a theme color by the web UI. The SDK does
861
955
  * not emit raw hex so authored evals stay decoupled from the web theme.
862
956
  */
863
- declare const evalChartColorSchema: z.ZodEnum<{
957
+ declare const evalChartColorSchema: z$1.ZodEnum<{
864
958
  accent: "accent";
865
959
  accentDim: "accentDim";
866
960
  success: "success";
@@ -869,27 +963,27 @@ declare const evalChartColorSchema: z.ZodEnum<{
869
963
  textMuted: "textMuted";
870
964
  }>;
871
965
  /** Semantic color token resolved to a theme color by the web UI. */
872
- type EvalChartColor = z.infer<typeof evalChartColorSchema>;
966
+ type EvalChartColor = z$1.infer<typeof evalChartColorSchema>;
873
967
  /** Y-axis placement for a plotted series on a dual-axis chart. */
874
- declare const evalChartAxisSchema: z.ZodEnum<{
968
+ declare const evalChartAxisSchema: z$1.ZodEnum<{
875
969
  left: "left";
876
970
  right: "right";
877
971
  }>;
878
972
  /** Y-axis placement for a plotted series on a dual-axis chart. */
879
- type EvalChartAxis = z.infer<typeof evalChartAxisSchema>;
973
+ type EvalChartAxis = z$1.infer<typeof evalChartAxisSchema>;
880
974
  /**
881
975
  * One plotted series on an eval history chart. `builtin` metrics come from the
882
976
  * per-run `RunSummary`; `column` metrics aggregate a per-case score or
883
977
  * `setEvalOutput` column across the run using `aggregate`.
884
978
  */
885
- declare const evalChartMetricSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
886
- source: z.ZodLiteral<"builtin">;
887
- metric: z.ZodEnum<{
979
+ declare const evalChartMetricSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
980
+ source: z$1.ZodLiteral<"builtin">;
981
+ metric: z$1.ZodEnum<{
888
982
  passRate: "passRate";
889
983
  durationMs: "durationMs";
890
984
  }>;
891
- label: z.ZodOptional<z.ZodString>;
892
- color: z.ZodOptional<z.ZodEnum<{
985
+ label: z$1.ZodOptional<z$1.ZodString>;
986
+ color: z$1.ZodOptional<z$1.ZodEnum<{
893
987
  accent: "accent";
894
988
  accentDim: "accentDim";
895
989
  success: "success";
@@ -897,14 +991,14 @@ declare const evalChartMetricSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
897
991
  warning: "warning";
898
992
  textMuted: "textMuted";
899
993
  }>>;
900
- axis: z.ZodOptional<z.ZodEnum<{
994
+ axis: z$1.ZodOptional<z$1.ZodEnum<{
901
995
  left: "left";
902
996
  right: "right";
903
997
  }>>;
904
- }, z.core.$strip>, z.ZodObject<{
905
- source: z.ZodLiteral<"column">;
906
- key: z.ZodString;
907
- aggregate: z.ZodEnum<{
998
+ }, z$1.core.$strip>, z$1.ZodObject<{
999
+ source: z$1.ZodLiteral<"column">;
1000
+ key: z$1.ZodString;
1001
+ aggregate: z$1.ZodEnum<{
908
1002
  avg: "avg";
909
1003
  sum: "sum";
910
1004
  min: "min";
@@ -912,8 +1006,8 @@ declare const evalChartMetricSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
912
1006
  latest: "latest";
913
1007
  passThresholdRate: "passThresholdRate";
914
1008
  }>;
915
- label: z.ZodOptional<z.ZodString>;
916
- color: z.ZodOptional<z.ZodEnum<{
1009
+ label: z$1.ZodOptional<z$1.ZodString>;
1010
+ color: z$1.ZodOptional<z$1.ZodEnum<{
917
1011
  accent: "accent";
918
1012
  accentDim: "accentDim";
919
1013
  success: "success";
@@ -921,25 +1015,25 @@ declare const evalChartMetricSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
921
1015
  warning: "warning";
922
1016
  textMuted: "textMuted";
923
1017
  }>>;
924
- axis: z.ZodOptional<z.ZodEnum<{
1018
+ axis: z$1.ZodOptional<z$1.ZodEnum<{
925
1019
  left: "left";
926
1020
  right: "right";
927
1021
  }>>;
928
- }, z.core.$strip>], "source">;
1022
+ }, z$1.core.$strip>], "source">;
929
1023
  /** One plotted series on an eval history chart. */
930
- type EvalChartMetric = z.infer<typeof evalChartMetricSchema>;
1024
+ type EvalChartMetric = z$1.infer<typeof evalChartMetricSchema>;
931
1025
  /** Extra field rendered only in the tooltip, not plotted as a series. */
932
- declare const evalChartTooltipExtraSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
933
- source: z.ZodLiteral<"builtin">;
934
- metric: z.ZodEnum<{
1026
+ declare const evalChartTooltipExtraSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
1027
+ source: z$1.ZodLiteral<"builtin">;
1028
+ metric: z$1.ZodEnum<{
935
1029
  passRate: "passRate";
936
1030
  durationMs: "durationMs";
937
1031
  }>;
938
- label: z.ZodOptional<z.ZodString>;
939
- }, z.core.$strip>, z.ZodObject<{
940
- source: z.ZodLiteral<"column">;
941
- key: z.ZodString;
942
- aggregate: z.ZodEnum<{
1032
+ label: z$1.ZodOptional<z$1.ZodString>;
1033
+ }, z$1.core.$strip>, z$1.ZodObject<{
1034
+ source: z$1.ZodLiteral<"column">;
1035
+ key: z$1.ZodString;
1036
+ aggregate: z$1.ZodEnum<{
943
1037
  avg: "avg";
944
1038
  sum: "sum";
945
1039
  min: "min";
@@ -947,30 +1041,30 @@ declare const evalChartTooltipExtraSchema: z.ZodDiscriminatedUnion<[z.ZodObject<
947
1041
  latest: "latest";
948
1042
  passThresholdRate: "passThresholdRate";
949
1043
  }>;
950
- label: z.ZodOptional<z.ZodString>;
951
- }, z.core.$strip>], "source">;
1044
+ label: z$1.ZodOptional<z$1.ZodString>;
1045
+ }, z$1.core.$strip>], "source">;
952
1046
  /** Extra field rendered only in the tooltip, not plotted as a series. */
953
- type EvalChartTooltipExtra = z.infer<typeof evalChartTooltipExtraSchema>;
1047
+ type EvalChartTooltipExtra = z$1.infer<typeof evalChartTooltipExtraSchema>;
954
1048
  /**
955
1049
  * Authored configuration for one eval history chart rendered in `EvalCard`.
956
1050
  * Authors declare a list of these via `EvalDefinition.charts` — the UI renders
957
1051
  * each entry as its own chart frame, stacked in authoring order.
958
1052
  */
959
- declare const evalChartConfigSchema: z.ZodObject<{
960
- heading: z.ZodOptional<z.ZodString>;
961
- type: z.ZodEnum<{
1053
+ declare const evalChartConfigSchema: z$1.ZodObject<{
1054
+ heading: z$1.ZodOptional<z$1.ZodString>;
1055
+ type: z$1.ZodEnum<{
962
1056
  area: "area";
963
1057
  line: "line";
964
1058
  bar: "bar";
965
1059
  }>;
966
- metrics: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
967
- source: z.ZodLiteral<"builtin">;
968
- metric: z.ZodEnum<{
1060
+ metrics: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
1061
+ source: z$1.ZodLiteral<"builtin">;
1062
+ metric: z$1.ZodEnum<{
969
1063
  passRate: "passRate";
970
1064
  durationMs: "durationMs";
971
1065
  }>;
972
- label: z.ZodOptional<z.ZodString>;
973
- color: z.ZodOptional<z.ZodEnum<{
1066
+ label: z$1.ZodOptional<z$1.ZodString>;
1067
+ color: z$1.ZodOptional<z$1.ZodEnum<{
974
1068
  accent: "accent";
975
1069
  accentDim: "accentDim";
976
1070
  success: "success";
@@ -978,14 +1072,14 @@ declare const evalChartConfigSchema: z.ZodObject<{
978
1072
  warning: "warning";
979
1073
  textMuted: "textMuted";
980
1074
  }>>;
981
- axis: z.ZodOptional<z.ZodEnum<{
1075
+ axis: z$1.ZodOptional<z$1.ZodEnum<{
982
1076
  left: "left";
983
1077
  right: "right";
984
1078
  }>>;
985
- }, z.core.$strip>, z.ZodObject<{
986
- source: z.ZodLiteral<"column">;
987
- key: z.ZodString;
988
- aggregate: z.ZodEnum<{
1079
+ }, z$1.core.$strip>, z$1.ZodObject<{
1080
+ source: z$1.ZodLiteral<"column">;
1081
+ key: z$1.ZodString;
1082
+ aggregate: z$1.ZodEnum<{
989
1083
  avg: "avg";
990
1084
  sum: "sum";
991
1085
  min: "min";
@@ -993,8 +1087,8 @@ declare const evalChartConfigSchema: z.ZodObject<{
993
1087
  latest: "latest";
994
1088
  passThresholdRate: "passThresholdRate";
995
1089
  }>;
996
- label: z.ZodOptional<z.ZodString>;
997
- color: z.ZodOptional<z.ZodEnum<{
1090
+ label: z$1.ZodOptional<z$1.ZodString>;
1091
+ color: z$1.ZodOptional<z$1.ZodEnum<{
998
1092
  accent: "accent";
999
1093
  accentDim: "accentDim";
1000
1094
  success: "success";
@@ -1002,32 +1096,32 @@ declare const evalChartConfigSchema: z.ZodObject<{
1002
1096
  warning: "warning";
1003
1097
  textMuted: "textMuted";
1004
1098
  }>>;
1005
- axis: z.ZodOptional<z.ZodEnum<{
1099
+ axis: z$1.ZodOptional<z$1.ZodEnum<{
1006
1100
  left: "left";
1007
1101
  right: "right";
1008
1102
  }>>;
1009
- }, z.core.$strip>], "source">>;
1010
- yDomain: z.ZodOptional<z.ZodObject<{
1011
- left: z.ZodOptional<z.ZodObject<{
1012
- min: z.ZodOptional<z.ZodNumber>;
1013
- max: z.ZodOptional<z.ZodNumber>;
1014
- }, z.core.$strip>>;
1015
- right: z.ZodOptional<z.ZodObject<{
1016
- min: z.ZodOptional<z.ZodNumber>;
1017
- max: z.ZodOptional<z.ZodNumber>;
1018
- }, z.core.$strip>>;
1019
- }, z.core.$strip>>;
1020
- tooltipExtras: z.ZodOptional<z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
1021
- source: z.ZodLiteral<"builtin">;
1022
- metric: z.ZodEnum<{
1103
+ }, z$1.core.$strip>], "source">>;
1104
+ yDomain: z$1.ZodOptional<z$1.ZodObject<{
1105
+ left: z$1.ZodOptional<z$1.ZodObject<{
1106
+ min: z$1.ZodOptional<z$1.ZodNumber>;
1107
+ max: z$1.ZodOptional<z$1.ZodNumber>;
1108
+ }, z$1.core.$strip>>;
1109
+ right: z$1.ZodOptional<z$1.ZodObject<{
1110
+ min: z$1.ZodOptional<z$1.ZodNumber>;
1111
+ max: z$1.ZodOptional<z$1.ZodNumber>;
1112
+ }, z$1.core.$strip>>;
1113
+ }, z$1.core.$strip>>;
1114
+ tooltipExtras: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
1115
+ source: z$1.ZodLiteral<"builtin">;
1116
+ metric: z$1.ZodEnum<{
1023
1117
  passRate: "passRate";
1024
1118
  durationMs: "durationMs";
1025
1119
  }>;
1026
- label: z.ZodOptional<z.ZodString>;
1027
- }, z.core.$strip>, z.ZodObject<{
1028
- source: z.ZodLiteral<"column">;
1029
- key: z.ZodString;
1030
- aggregate: z.ZodEnum<{
1120
+ label: z$1.ZodOptional<z$1.ZodString>;
1121
+ }, z$1.core.$strip>, z$1.ZodObject<{
1122
+ source: z$1.ZodLiteral<"column">;
1123
+ key: z$1.ZodString;
1124
+ aggregate: z$1.ZodEnum<{
1031
1125
  avg: "avg";
1032
1126
  sum: "sum";
1033
1127
  min: "min";
@@ -1035,30 +1129,30 @@ declare const evalChartConfigSchema: z.ZodObject<{
1035
1129
  latest: "latest";
1036
1130
  passThresholdRate: "passThresholdRate";
1037
1131
  }>;
1038
- label: z.ZodOptional<z.ZodString>;
1039
- }, z.core.$strip>], "source">>>;
1040
- }, z.core.$strip>;
1132
+ label: z$1.ZodOptional<z$1.ZodString>;
1133
+ }, z$1.core.$strip>], "source">>>;
1134
+ }, z$1.core.$strip>;
1041
1135
  /** Authored configuration for one eval history chart. */
1042
- type EvalChartConfig = z.infer<typeof evalChartConfigSchema>;
1136
+ type EvalChartConfig = z$1.infer<typeof evalChartConfigSchema>;
1043
1137
  /**
1044
1138
  * Ordered list of history charts rendered for an eval. Opt-in: when omitted or
1045
1139
  * empty, the UI renders no history chart at all.
1046
1140
  */
1047
- declare const evalChartsConfigSchema: z.ZodArray<z.ZodObject<{
1048
- heading: z.ZodOptional<z.ZodString>;
1049
- type: z.ZodEnum<{
1141
+ declare const evalChartsConfigSchema: z$1.ZodArray<z$1.ZodObject<{
1142
+ heading: z$1.ZodOptional<z$1.ZodString>;
1143
+ type: z$1.ZodEnum<{
1050
1144
  area: "area";
1051
1145
  line: "line";
1052
1146
  bar: "bar";
1053
1147
  }>;
1054
- metrics: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
1055
- source: z.ZodLiteral<"builtin">;
1056
- metric: z.ZodEnum<{
1148
+ metrics: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
1149
+ source: z$1.ZodLiteral<"builtin">;
1150
+ metric: z$1.ZodEnum<{
1057
1151
  passRate: "passRate";
1058
1152
  durationMs: "durationMs";
1059
1153
  }>;
1060
- label: z.ZodOptional<z.ZodString>;
1061
- color: z.ZodOptional<z.ZodEnum<{
1154
+ label: z$1.ZodOptional<z$1.ZodString>;
1155
+ color: z$1.ZodOptional<z$1.ZodEnum<{
1062
1156
  accent: "accent";
1063
1157
  accentDim: "accentDim";
1064
1158
  success: "success";
@@ -1066,14 +1160,14 @@ declare const evalChartsConfigSchema: z.ZodArray<z.ZodObject<{
1066
1160
  warning: "warning";
1067
1161
  textMuted: "textMuted";
1068
1162
  }>>;
1069
- axis: z.ZodOptional<z.ZodEnum<{
1163
+ axis: z$1.ZodOptional<z$1.ZodEnum<{
1070
1164
  left: "left";
1071
1165
  right: "right";
1072
1166
  }>>;
1073
- }, z.core.$strip>, z.ZodObject<{
1074
- source: z.ZodLiteral<"column">;
1075
- key: z.ZodString;
1076
- aggregate: z.ZodEnum<{
1167
+ }, z$1.core.$strip>, z$1.ZodObject<{
1168
+ source: z$1.ZodLiteral<"column">;
1169
+ key: z$1.ZodString;
1170
+ aggregate: z$1.ZodEnum<{
1077
1171
  avg: "avg";
1078
1172
  sum: "sum";
1079
1173
  min: "min";
@@ -1081,8 +1175,8 @@ declare const evalChartsConfigSchema: z.ZodArray<z.ZodObject<{
1081
1175
  latest: "latest";
1082
1176
  passThresholdRate: "passThresholdRate";
1083
1177
  }>;
1084
- label: z.ZodOptional<z.ZodString>;
1085
- color: z.ZodOptional<z.ZodEnum<{
1178
+ label: z$1.ZodOptional<z$1.ZodString>;
1179
+ color: z$1.ZodOptional<z$1.ZodEnum<{
1086
1180
  accent: "accent";
1087
1181
  accentDim: "accentDim";
1088
1182
  success: "success";
@@ -1090,32 +1184,32 @@ declare const evalChartsConfigSchema: z.ZodArray<z.ZodObject<{
1090
1184
  warning: "warning";
1091
1185
  textMuted: "textMuted";
1092
1186
  }>>;
1093
- axis: z.ZodOptional<z.ZodEnum<{
1187
+ axis: z$1.ZodOptional<z$1.ZodEnum<{
1094
1188
  left: "left";
1095
1189
  right: "right";
1096
1190
  }>>;
1097
- }, z.core.$strip>], "source">>;
1098
- yDomain: z.ZodOptional<z.ZodObject<{
1099
- left: z.ZodOptional<z.ZodObject<{
1100
- min: z.ZodOptional<z.ZodNumber>;
1101
- max: z.ZodOptional<z.ZodNumber>;
1102
- }, z.core.$strip>>;
1103
- right: z.ZodOptional<z.ZodObject<{
1104
- min: z.ZodOptional<z.ZodNumber>;
1105
- max: z.ZodOptional<z.ZodNumber>;
1106
- }, z.core.$strip>>;
1107
- }, z.core.$strip>>;
1108
- tooltipExtras: z.ZodOptional<z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
1109
- source: z.ZodLiteral<"builtin">;
1110
- metric: z.ZodEnum<{
1191
+ }, z$1.core.$strip>], "source">>;
1192
+ yDomain: z$1.ZodOptional<z$1.ZodObject<{
1193
+ left: z$1.ZodOptional<z$1.ZodObject<{
1194
+ min: z$1.ZodOptional<z$1.ZodNumber>;
1195
+ max: z$1.ZodOptional<z$1.ZodNumber>;
1196
+ }, z$1.core.$strip>>;
1197
+ right: z$1.ZodOptional<z$1.ZodObject<{
1198
+ min: z$1.ZodOptional<z$1.ZodNumber>;
1199
+ max: z$1.ZodOptional<z$1.ZodNumber>;
1200
+ }, z$1.core.$strip>>;
1201
+ }, z$1.core.$strip>>;
1202
+ tooltipExtras: z$1.ZodOptional<z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
1203
+ source: z$1.ZodLiteral<"builtin">;
1204
+ metric: z$1.ZodEnum<{
1111
1205
  passRate: "passRate";
1112
1206
  durationMs: "durationMs";
1113
1207
  }>;
1114
- label: z.ZodOptional<z.ZodString>;
1115
- }, z.core.$strip>, z.ZodObject<{
1116
- source: z.ZodLiteral<"column">;
1117
- key: z.ZodString;
1118
- aggregate: z.ZodEnum<{
1208
+ label: z$1.ZodOptional<z$1.ZodString>;
1209
+ }, z$1.core.$strip>, z$1.ZodObject<{
1210
+ source: z$1.ZodLiteral<"column">;
1211
+ key: z$1.ZodString;
1212
+ aggregate: z$1.ZodEnum<{
1119
1213
  avg: "avg";
1120
1214
  sum: "sum";
1121
1215
  min: "min";
@@ -1123,70 +1217,70 @@ declare const evalChartsConfigSchema: z.ZodArray<z.ZodObject<{
1123
1217
  latest: "latest";
1124
1218
  passThresholdRate: "passThresholdRate";
1125
1219
  }>;
1126
- label: z.ZodOptional<z.ZodString>;
1127
- }, z.core.$strip>], "source">>>;
1128
- }, z.core.$strip>>;
1220
+ label: z$1.ZodOptional<z$1.ZodString>;
1221
+ }, z$1.core.$strip>], "source">>>;
1222
+ }, z$1.core.$strip>>;
1129
1223
  /** Ordered list of history charts rendered for an eval. */
1130
- type EvalChartsConfig = z.infer<typeof evalChartsConfigSchema>;
1224
+ type EvalChartsConfig = z$1.infer<typeof evalChartsConfigSchema>;
1131
1225
  //#endregion
1132
1226
  //#region ../shared/src/schemas/run.d.ts
1133
1227
  /** Schema for persisted metadata about a single run invocation. */
1134
- declare const runManifestSchema: z.ZodObject<{
1135
- id: z.ZodString;
1136
- shortId: z.ZodString;
1137
- status: z.ZodEnum<{
1228
+ declare const runManifestSchema: z$1.ZodObject<{
1229
+ id: z$1.ZodString;
1230
+ shortId: z$1.ZodString;
1231
+ status: z$1.ZodEnum<{
1138
1232
  pending: "pending";
1139
1233
  running: "running";
1140
1234
  completed: "completed";
1141
1235
  cancelled: "cancelled";
1142
1236
  error: "error";
1143
1237
  }>;
1144
- startedAt: z.ZodString;
1145
- endedAt: z.ZodNullable<z.ZodString>;
1146
- commitSha: z.ZodDefault<z.ZodOptional<z.ZodNullable<z.ZodString>>>;
1147
- evalSourceFingerprints: z.ZodDefault<z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>>;
1148
- target: z.ZodObject<{
1149
- mode: z.ZodEnum<{
1238
+ startedAt: z$1.ZodString;
1239
+ endedAt: z$1.ZodNullable<z$1.ZodString>;
1240
+ commitSha: z$1.ZodDefault<z$1.ZodOptional<z$1.ZodNullable<z$1.ZodString>>>;
1241
+ evalSourceFingerprints: z$1.ZodDefault<z$1.ZodOptional<z$1.ZodRecord<z$1.ZodString, z$1.ZodString>>>;
1242
+ target: z$1.ZodObject<{
1243
+ mode: z$1.ZodEnum<{
1150
1244
  all: "all";
1151
1245
  evalIds: "evalIds";
1152
1246
  caseIds: "caseIds";
1153
1247
  }>;
1154
- evalIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
1155
- caseIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
1156
- }, z.core.$strip>;
1157
- trials: z.ZodNumber;
1158
- trialSelection: z.ZodDefault<z.ZodOptional<z.ZodEnum<{
1248
+ evalIds: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
1249
+ caseIds: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
1250
+ }, z$1.core.$strip>;
1251
+ trials: z$1.ZodNumber;
1252
+ trialSelection: z$1.ZodDefault<z$1.ZodOptional<z$1.ZodEnum<{
1159
1253
  lowestScore: "lowestScore";
1160
1254
  median: "median";
1161
1255
  }>>>;
1162
- cacheMode: z.ZodOptional<z.ZodEnum<{
1256
+ cacheMode: z$1.ZodOptional<z$1.ZodEnum<{
1163
1257
  use: "use";
1164
1258
  bypass: "bypass";
1165
1259
  refresh: "refresh";
1166
1260
  }>>;
1167
- }, z.core.$strip>;
1261
+ }, z$1.core.$strip>;
1168
1262
  /** Persisted lifecycle metadata for a single eval run. */
1169
- type RunManifest = z.infer<typeof runManifestSchema>;
1263
+ type RunManifest = z$1.infer<typeof runManifestSchema>;
1170
1264
  /** Schema for aggregate metrics computed over a completed or active run. */
1171
- declare const runSummarySchema: z.ZodObject<{
1172
- runId: z.ZodString;
1173
- status: z.ZodEnum<{
1265
+ declare const runSummarySchema: z$1.ZodObject<{
1266
+ runId: z$1.ZodString;
1267
+ status: z$1.ZodEnum<{
1174
1268
  pending: "pending";
1175
1269
  running: "running";
1176
1270
  completed: "completed";
1177
1271
  cancelled: "cancelled";
1178
1272
  error: "error";
1179
1273
  }>;
1180
- totalCases: z.ZodNumber;
1181
- passedCases: z.ZodNumber;
1182
- failedCases: z.ZodNumber;
1183
- errorCases: z.ZodNumber;
1184
- cancelledCases: z.ZodNumber;
1185
- totalDurationMs: z.ZodNullable<z.ZodNumber>;
1186
- errorMessage: z.ZodDefault<z.ZodNullable<z.ZodString>>;
1187
- }, z.core.$strip>;
1274
+ totalCases: z$1.ZodNumber;
1275
+ passedCases: z$1.ZodNumber;
1276
+ failedCases: z$1.ZodNumber;
1277
+ errorCases: z$1.ZodNumber;
1278
+ cancelledCases: z$1.ZodNumber;
1279
+ totalDurationMs: z$1.ZodNullable<z$1.ZodNumber>;
1280
+ errorMessage: z$1.ZodDefault<z$1.ZodNullable<z$1.ZodString>>;
1281
+ }, z$1.core.$strip>;
1188
1282
  /** Roll-up statistics for one run. */
1189
- type RunSummary = z.infer<typeof runSummarySchema>;
1283
+ type RunSummary = z$1.infer<typeof runSummarySchema>;
1190
1284
  //#endregion
1191
1285
  //#region ../shared/src/status.d.ts
1192
1286
  /**
@@ -1270,7 +1364,7 @@ type EvalTitleLike = {
1270
1364
  declare function getEvalTitle(evalLike: EvalTitleLike): string;
1271
1365
  //#endregion
1272
1366
  //#region ../shared/src/schemas/sse.d.ts
1273
- declare const sseEventTypeSchema: z.ZodEnum<{
1367
+ declare const sseEventTypeSchema: z$1.ZodEnum<{
1274
1368
  "discovery.updated": "discovery.updated";
1275
1369
  "run.started": "run.started";
1276
1370
  "run.summary": "run.summary";
@@ -1283,55 +1377,55 @@ declare const sseEventTypeSchema: z.ZodEnum<{
1283
1377
  "run.error": "run.error";
1284
1378
  }>;
1285
1379
  /** Server-sent event name emitted by the runner or backend. */
1286
- type SseEventType = z.infer<typeof sseEventTypeSchema>;
1380
+ type SseEventType = z$1.infer<typeof sseEventTypeSchema>;
1287
1381
  /** Schema for the SSE envelope used to stream run updates to clients. */
1288
- declare const sseEnvelopeSchema: z.ZodObject<{
1289
- type: z.ZodString;
1290
- runId: z.ZodOptional<z.ZodString>;
1291
- timestamp: z.ZodString;
1292
- payload: z.ZodUnknown;
1293
- }, z.core.$strip>;
1382
+ declare const sseEnvelopeSchema: z$1.ZodObject<{
1383
+ type: z$1.ZodString;
1384
+ runId: z$1.ZodOptional<z$1.ZodString>;
1385
+ timestamp: z$1.ZodString;
1386
+ payload: z$1.ZodUnknown;
1387
+ }, z$1.core.$strip>;
1294
1388
  /** Wire format for a streamed event emitted during eval execution. */
1295
- type SseEnvelope = z.infer<typeof sseEnvelopeSchema>;
1389
+ type SseEnvelope = z$1.infer<typeof sseEnvelopeSchema>;
1296
1390
  //#endregion
1297
1391
  //#region ../shared/src/schemas/api.d.ts
1298
1392
  /** Schema for the API request that starts a new eval run. */
1299
- declare const createRunRequestSchema: z.ZodObject<{
1300
- target: z.ZodObject<{
1301
- mode: z.ZodEnum<{
1393
+ declare const createRunRequestSchema: z$1.ZodObject<{
1394
+ target: z$1.ZodObject<{
1395
+ mode: z$1.ZodEnum<{
1302
1396
  all: "all";
1303
1397
  evalIds: "evalIds";
1304
1398
  caseIds: "caseIds";
1305
1399
  }>;
1306
- evalIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
1307
- caseIds: z.ZodOptional<z.ZodArray<z.ZodString>>;
1308
- }, z.core.$strip>;
1309
- trials: z.ZodNumber;
1310
- cache: z.ZodOptional<z.ZodObject<{
1311
- mode: z.ZodDefault<z.ZodEnum<{
1400
+ evalIds: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
1401
+ caseIds: z$1.ZodOptional<z$1.ZodArray<z$1.ZodString>>;
1402
+ }, z$1.core.$strip>;
1403
+ trials: z$1.ZodNumber;
1404
+ cache: z$1.ZodOptional<z$1.ZodObject<{
1405
+ mode: z$1.ZodDefault<z$1.ZodEnum<{
1312
1406
  use: "use";
1313
1407
  bypass: "bypass";
1314
1408
  refresh: "refresh";
1315
1409
  }>>;
1316
- }, z.core.$strip>>;
1317
- }, z.core.$strip>;
1410
+ }, z$1.core.$strip>>;
1411
+ }, z$1.core.$strip>;
1318
1412
  /** Request payload accepted by the run creation endpoint. */
1319
- type CreateRunRequest = z.infer<typeof createRunRequestSchema>;
1413
+ type CreateRunRequest = z$1.infer<typeof createRunRequestSchema>;
1320
1414
  /** Schema for updating a UI-authored manual score on one persisted case. */
1321
- declare const updateManualScoreRequestSchema: z.ZodObject<{
1322
- value: z.ZodNullable<z.ZodNumber>;
1323
- }, z.core.$strip>;
1415
+ declare const updateManualScoreRequestSchema: z$1.ZodObject<{
1416
+ value: z$1.ZodNullable<z$1.ZodNumber>;
1417
+ }, z$1.core.$strip>;
1324
1418
  /** Request payload accepted by the manual score update endpoint. */
1325
- type UpdateManualScoreRequest = z.infer<typeof updateManualScoreRequestSchema>;
1419
+ type UpdateManualScoreRequest = z$1.infer<typeof updateManualScoreRequestSchema>;
1326
1420
  //#endregion
1327
1421
  //#region ../shared/src/schemas/config.d.ts
1328
1422
  /** Strategy used to collapse repeated trials into one stored case result. */
1329
- declare const trialSelectionModeSchema: z.ZodEnum<{
1423
+ declare const trialSelectionModeSchema: z$1.ZodEnum<{
1330
1424
  lowestScore: "lowestScore";
1331
1425
  median: "median";
1332
1426
  }>;
1333
1427
  /** Strategy used to collapse repeated trials into one stored case result. */
1334
- type TrialSelectionMode = z.infer<typeof trialSelectionModeSchema>;
1428
+ type TrialSelectionMode = z$1.infer<typeof trialSelectionModeSchema>;
1335
1429
  /** Top-level config authored in `agent-evals.config.ts`. */
1336
1430
  type AgentEvalsConfig = {
1337
1431
  /** Root directory used to resolve all relative paths. Defaults to `process.cwd()`. */workspaceRoot?: string; /** Glob patterns (relative to `workspaceRoot`) used to discover eval files. */
@@ -1376,51 +1470,51 @@ type AgentEvalsConfig = {
1376
1470
  };
1377
1471
  };
1378
1472
  /** Zod schema for validating `agent-evals.config.ts` input. */
1379
- declare const agentEvalsConfigSchema: z.ZodObject<{
1380
- workspaceRoot: z.ZodOptional<z.ZodString>;
1381
- include: z.ZodArray<z.ZodString>;
1382
- defaultTrials: z.ZodOptional<z.ZodNumber>;
1383
- trialSelection: z.ZodOptional<z.ZodEnum<{
1473
+ declare const agentEvalsConfigSchema: z$1.ZodObject<{
1474
+ workspaceRoot: z$1.ZodOptional<z$1.ZodString>;
1475
+ include: z$1.ZodArray<z$1.ZodString>;
1476
+ defaultTrials: z$1.ZodOptional<z$1.ZodNumber>;
1477
+ trialSelection: z$1.ZodOptional<z$1.ZodEnum<{
1384
1478
  lowestScore: "lowestScore";
1385
1479
  median: "median";
1386
1480
  }>>;
1387
- concurrency: z.ZodOptional<z.ZodNumber>;
1388
- staleAfterDays: z.ZodOptional<z.ZodNumber>;
1389
- traceDisplay: z.ZodOptional<z.ZodObject<{
1390
- attributes: z.ZodOptional<z.ZodArray<z.ZodObject<{
1391
- key: z.ZodOptional<z.ZodString>;
1392
- path: z.ZodString;
1393
- label: z.ZodOptional<z.ZodString>;
1394
- format: z.ZodOptional<z.ZodEnum<{
1481
+ concurrency: z$1.ZodOptional<z$1.ZodNumber>;
1482
+ staleAfterDays: z$1.ZodOptional<z$1.ZodNumber>;
1483
+ traceDisplay: z$1.ZodOptional<z$1.ZodObject<{
1484
+ attributes: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
1485
+ key: z$1.ZodOptional<z$1.ZodString>;
1486
+ path: z$1.ZodString;
1487
+ label: z$1.ZodOptional<z$1.ZodString>;
1488
+ format: z$1.ZodOptional<z$1.ZodEnum<{
1395
1489
  string: "string";
1396
1490
  number: "number";
1397
1491
  duration: "duration";
1398
1492
  json: "json";
1399
1493
  }>>;
1400
- numberFormat: z.ZodOptional<z.ZodType<NumberDisplayOptions, unknown, z.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
1401
- placements: z.ZodOptional<z.ZodArray<z.ZodEnum<{
1494
+ numberFormat: z$1.ZodOptional<z$1.ZodType<NumberDisplayOptions, unknown, z$1.core.$ZodTypeInternals<NumberDisplayOptions, unknown>>>;
1495
+ placements: z$1.ZodOptional<z$1.ZodArray<z$1.ZodEnum<{
1402
1496
  tree: "tree";
1403
1497
  detail: "detail";
1404
1498
  section: "section";
1405
1499
  }>>>;
1406
- scope: z.ZodOptional<z.ZodEnum<{
1500
+ scope: z$1.ZodOptional<z$1.ZodEnum<{
1407
1501
  self: "self";
1408
1502
  subtree: "subtree";
1409
1503
  }>>;
1410
- mode: z.ZodOptional<z.ZodEnum<{
1504
+ mode: z$1.ZodOptional<z$1.ZodEnum<{
1411
1505
  all: "all";
1412
1506
  last: "last";
1413
1507
  sum: "sum";
1414
1508
  }>>;
1415
- transform: z.ZodOptional<z.ZodCustom<TraceAttributeTransform, TraceAttributeTransform>>;
1416
- }, z.core.$strip>>>;
1417
- }, z.core.$strip>>;
1418
- cache: z.ZodOptional<z.ZodObject<{
1419
- enabled: z.ZodOptional<z.ZodBoolean>;
1420
- dir: z.ZodOptional<z.ZodString>;
1421
- maxEntriesPerEval: z.ZodPipe<z.ZodTransform<number | undefined, unknown>, z.ZodOptional<z.ZodNumber>>;
1422
- }, z.core.$strip>>;
1423
- }, z.core.$strip>;
1509
+ transform: z$1.ZodOptional<z$1.ZodCustom<TraceAttributeTransform, TraceAttributeTransform>>;
1510
+ }, z$1.core.$strip>>>;
1511
+ }, z$1.core.$strip>>;
1512
+ cache: z$1.ZodOptional<z$1.ZodObject<{
1513
+ enabled: z$1.ZodOptional<z$1.ZodBoolean>;
1514
+ dir: z$1.ZodOptional<z$1.ZodString>;
1515
+ maxEntriesPerEval: z$1.ZodPipe<z$1.ZodTransform<number | undefined, unknown>, z$1.ZodOptional<z$1.ZodNumber>>;
1516
+ }, z$1.core.$strip>>;
1517
+ }, z$1.core.$strip>;
1424
1518
  //#endregion
1425
1519
  //#region ../shared/src/schemas/cache.d.ts
1426
1520
  /**
@@ -1430,162 +1524,305 @@ declare const agentEvalsConfigSchema: z.ZodObject<{
1430
1524
  * - `bypass`: never read, never write.
1431
1525
  * - `refresh`: never read, always write (forces re-execution and overwrites).
1432
1526
  */
1433
- declare const cacheModeSchema: z.ZodEnum<{
1527
+ declare const cacheModeSchema: z$1.ZodEnum<{
1434
1528
  use: "use";
1435
1529
  bypass: "bypass";
1436
1530
  refresh: "refresh";
1437
1531
  }>;
1438
1532
  /** Mode controlling how cached spans behave during a run. */
1439
- type CacheMode = z.infer<typeof cacheModeSchema>;
1533
+ type CacheMode = z$1.infer<typeof cacheModeSchema>;
1440
1534
  /** Options accepted by an `evalTracer.span` call to opt the span into caching. */
1441
- declare const spanCacheOptionsSchema: z.ZodObject<{
1442
- key: z.ZodUnknown;
1443
- namespace: z.ZodOptional<z.ZodString>;
1444
- }, z.core.$strip>;
1535
+ declare const spanCacheOptionsSchema: z$1.ZodObject<{
1536
+ key: z$1.ZodUnknown;
1537
+ namespace: z$1.ZodOptional<z$1.ZodString>;
1538
+ }, z$1.core.$strip>;
1445
1539
  /** Options accepted by an `evalTracer.span` call to opt the span into caching. */
1446
- type SpanCacheOptions = z.infer<typeof spanCacheOptionsSchema>;
1540
+ type SpanCacheOptions = z$1.infer<typeof spanCacheOptionsSchema>;
1541
+ /** Category of operation stored in the eval cache. */
1542
+ declare const cacheOperationTypeSchema: z$1.ZodEnum<{
1543
+ span: "span";
1544
+ value: "value";
1545
+ }>;
1546
+ /** Category of operation stored in the eval cache. */
1547
+ type CacheOperationType = z$1.infer<typeof cacheOperationTypeSchema>;
1447
1548
  /** Summary of a single persisted cache entry, used by list/delete endpoints. */
1448
- declare const cacheListItemSchema: z.ZodObject<{
1449
- key: z.ZodString;
1450
- namespace: z.ZodString;
1451
- spanName: z.ZodString;
1452
- spanKind: z.ZodString;
1453
- storedAt: z.ZodString;
1454
- codeFingerprint: z.ZodString;
1455
- sizeBytes: z.ZodNumber;
1456
- }, z.core.$strip>;
1549
+ declare const cacheListItemSchema: z$1.ZodObject<{
1550
+ key: z$1.ZodString;
1551
+ namespace: z$1.ZodString;
1552
+ operationType: z$1.ZodEnum<{
1553
+ span: "span";
1554
+ value: "value";
1555
+ }>;
1556
+ operationName: z$1.ZodString;
1557
+ spanName: z$1.ZodOptional<z$1.ZodString>;
1558
+ spanKind: z$1.ZodOptional<z$1.ZodString>;
1559
+ storedAt: z$1.ZodString;
1560
+ codeFingerprint: z$1.ZodString;
1561
+ sizeBytes: z$1.ZodNumber;
1562
+ }, z$1.core.$strip>;
1457
1563
  /** Summary row for a single cache entry. */
1458
- type CacheListItem = z.infer<typeof cacheListItemSchema>;
1564
+ type CacheListItem = z$1.infer<typeof cacheListItemSchema>;
1459
1565
  /** Serialized nested span captured while recording a cached operation. */
1460
1566
  type SerializedCacheSpan = {
1461
1567
  kind: string;
1462
1568
  name: string;
1463
1569
  attributes?: Record<string, unknown>;
1464
1570
  status: 'running' | 'ok' | 'error' | 'cancelled';
1465
- error?: {
1466
- name?: string;
1467
- message: string;
1468
- stack?: string;
1469
- };
1571
+ error?: EvalTraceSpanError;
1572
+ errors?: EvalTraceSpanError[];
1573
+ warning?: EvalTraceSpanWarning;
1574
+ warnings?: EvalTraceSpanWarning[];
1470
1575
  children: SerializedCacheSpan[];
1471
1576
  };
1472
1577
  /** Zod schema for `SerializedCacheSpan`, defined lazily for recursion. */
1473
- declare const serializedCacheSpanSchema: z.ZodType<SerializedCacheSpan>;
1578
+ declare const serializedCacheSpanSchema: z$1.ZodType<SerializedCacheSpan>;
1474
1579
  /**
1475
1580
  * One captured operation performed while a cached span's body executed.
1476
1581
  *
1477
1582
  * Operations are replayed in order against a fresh scope on cache hit to
1478
1583
  * reproduce the observable effects of the original run.
1479
1584
  */
1480
- declare const cacheRecordingOpSchema: z.ZodDiscriminatedUnion<[z.ZodObject<{
1481
- kind: z.ZodLiteral<"setOutput">;
1482
- key: z.ZodString;
1483
- value: z.ZodUnknown;
1484
- }, z.core.$strip>, z.ZodObject<{
1485
- kind: z.ZodLiteral<"incrementOutput">;
1486
- key: z.ZodString;
1487
- delta: z.ZodNumber;
1488
- }, z.core.$strip>, z.ZodObject<{
1489
- kind: z.ZodLiteral<"checkpoint">;
1490
- name: z.ZodString;
1491
- data: z.ZodUnknown;
1492
- }, z.core.$strip>, z.ZodObject<{
1493
- kind: z.ZodLiteral<"subSpan">;
1494
- span: z.ZodType<SerializedCacheSpan, unknown, z.core.$ZodTypeInternals<SerializedCacheSpan, unknown>>;
1495
- }, z.core.$strip>], "kind">;
1585
+ declare const cacheRecordingOpSchema: z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
1586
+ kind: z$1.ZodLiteral<"setOutput">;
1587
+ key: z$1.ZodString;
1588
+ value: z$1.ZodUnknown;
1589
+ }, z$1.core.$strip>, z$1.ZodObject<{
1590
+ kind: z$1.ZodLiteral<"appendOutput">;
1591
+ key: z$1.ZodString;
1592
+ value: z$1.ZodUnknown;
1593
+ }, z$1.core.$strip>, z$1.ZodObject<{
1594
+ kind: z$1.ZodLiteral<"mergeOutput">;
1595
+ key: z$1.ZodString;
1596
+ patch: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
1597
+ }, z$1.core.$strip>, z$1.ZodObject<{
1598
+ kind: z$1.ZodLiteral<"incrementOutput">;
1599
+ key: z$1.ZodString;
1600
+ delta: z$1.ZodNumber;
1601
+ }, z$1.core.$strip>, z$1.ZodObject<{
1602
+ kind: z$1.ZodLiteral<"checkpoint">;
1603
+ name: z$1.ZodString;
1604
+ data: z$1.ZodUnknown;
1605
+ }, z$1.core.$strip>, z$1.ZodObject<{
1606
+ kind: z$1.ZodLiteral<"subSpan">;
1607
+ span: z$1.ZodType<SerializedCacheSpan, unknown, z$1.core.$ZodTypeInternals<SerializedCacheSpan, unknown>>;
1608
+ }, z$1.core.$strip>], "kind">;
1496
1609
  /** Single effect captured by a cache recording. */
1497
- type CacheRecordingOp = z.infer<typeof cacheRecordingOpSchema>;
1610
+ type CacheRecordingOp = z$1.infer<typeof cacheRecordingOpSchema>;
1498
1611
  /** Captured observable effects + return value of a cached span body. */
1499
- declare const cacheRecordingSchema: z.ZodObject<{
1500
- returnValue: z.ZodUnknown;
1501
- finalAttributes: z.ZodRecord<z.ZodString, z.ZodUnknown>;
1502
- ops: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
1503
- kind: z.ZodLiteral<"setOutput">;
1504
- key: z.ZodString;
1505
- value: z.ZodUnknown;
1506
- }, z.core.$strip>, z.ZodObject<{
1507
- kind: z.ZodLiteral<"incrementOutput">;
1508
- key: z.ZodString;
1509
- delta: z.ZodNumber;
1510
- }, z.core.$strip>, z.ZodObject<{
1511
- kind: z.ZodLiteral<"checkpoint">;
1512
- name: z.ZodString;
1513
- data: z.ZodUnknown;
1514
- }, z.core.$strip>, z.ZodObject<{
1515
- kind: z.ZodLiteral<"subSpan">;
1516
- span: z.ZodType<SerializedCacheSpan, unknown, z.core.$ZodTypeInternals<SerializedCacheSpan, unknown>>;
1517
- }, z.core.$strip>], "kind">>;
1518
- }, z.core.$strip>;
1612
+ declare const cacheRecordingSchema: z$1.ZodObject<{
1613
+ returnValue: z$1.ZodUnknown;
1614
+ finalAttributes: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
1615
+ finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
1616
+ error: "error";
1617
+ running: "running";
1618
+ ok: "ok";
1619
+ cancelled: "cancelled";
1620
+ }>>;
1621
+ finalError: z$1.ZodOptional<z$1.ZodObject<{
1622
+ name: z$1.ZodOptional<z$1.ZodString>;
1623
+ message: z$1.ZodString;
1624
+ stack: z$1.ZodOptional<z$1.ZodString>;
1625
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
1626
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>;
1627
+ finalErrors: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
1628
+ name: z$1.ZodOptional<z$1.ZodString>;
1629
+ message: z$1.ZodString;
1630
+ stack: z$1.ZodOptional<z$1.ZodString>;
1631
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
1632
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
1633
+ finalWarning: z$1.ZodOptional<z$1.ZodObject<{
1634
+ name: z$1.ZodOptional<z$1.ZodString>;
1635
+ message: z$1.ZodString;
1636
+ stack: z$1.ZodOptional<z$1.ZodString>;
1637
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
1638
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>;
1639
+ finalWarnings: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
1640
+ name: z$1.ZodOptional<z$1.ZodString>;
1641
+ message: z$1.ZodString;
1642
+ stack: z$1.ZodOptional<z$1.ZodString>;
1643
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
1644
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
1645
+ ops: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
1646
+ kind: z$1.ZodLiteral<"setOutput">;
1647
+ key: z$1.ZodString;
1648
+ value: z$1.ZodUnknown;
1649
+ }, z$1.core.$strip>, z$1.ZodObject<{
1650
+ kind: z$1.ZodLiteral<"appendOutput">;
1651
+ key: z$1.ZodString;
1652
+ value: z$1.ZodUnknown;
1653
+ }, z$1.core.$strip>, z$1.ZodObject<{
1654
+ kind: z$1.ZodLiteral<"mergeOutput">;
1655
+ key: z$1.ZodString;
1656
+ patch: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
1657
+ }, z$1.core.$strip>, z$1.ZodObject<{
1658
+ kind: z$1.ZodLiteral<"incrementOutput">;
1659
+ key: z$1.ZodString;
1660
+ delta: z$1.ZodNumber;
1661
+ }, z$1.core.$strip>, z$1.ZodObject<{
1662
+ kind: z$1.ZodLiteral<"checkpoint">;
1663
+ name: z$1.ZodString;
1664
+ data: z$1.ZodUnknown;
1665
+ }, z$1.core.$strip>, z$1.ZodObject<{
1666
+ kind: z$1.ZodLiteral<"subSpan">;
1667
+ span: z$1.ZodType<SerializedCacheSpan, unknown, z$1.core.$ZodTypeInternals<SerializedCacheSpan, unknown>>;
1668
+ }, z$1.core.$strip>], "kind">>;
1669
+ }, z$1.core.$strip>;
1519
1670
  /** Captured observable effects + return value of a cached span body. */
1520
- type CacheRecording = z.infer<typeof cacheRecordingSchema>;
1671
+ type CacheRecording = z$1.infer<typeof cacheRecordingSchema>;
1521
1672
  /** Persisted cache file containing metadata and a recording. */
1522
- declare const cacheEntrySchema: z.ZodObject<{
1523
- version: z.ZodLiteral<1>;
1524
- key: z.ZodString;
1525
- namespace: z.ZodString;
1526
- spanName: z.ZodString;
1527
- spanKind: z.ZodString;
1528
- storedAt: z.ZodString;
1529
- codeFingerprint: z.ZodString;
1530
- recording: z.ZodObject<{
1531
- returnValue: z.ZodUnknown;
1532
- finalAttributes: z.ZodRecord<z.ZodString, z.ZodUnknown>;
1533
- ops: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
1534
- kind: z.ZodLiteral<"setOutput">;
1535
- key: z.ZodString;
1536
- value: z.ZodUnknown;
1537
- }, z.core.$strip>, z.ZodObject<{
1538
- kind: z.ZodLiteral<"incrementOutput">;
1539
- key: z.ZodString;
1540
- delta: z.ZodNumber;
1541
- }, z.core.$strip>, z.ZodObject<{
1542
- kind: z.ZodLiteral<"checkpoint">;
1543
- name: z.ZodString;
1544
- data: z.ZodUnknown;
1545
- }, z.core.$strip>, z.ZodObject<{
1546
- kind: z.ZodLiteral<"subSpan">;
1547
- span: z.ZodType<SerializedCacheSpan, unknown, z.core.$ZodTypeInternals<SerializedCacheSpan, unknown>>;
1548
- }, z.core.$strip>], "kind">>;
1549
- }, z.core.$strip>;
1550
- }, z.core.$strip>;
1673
+ declare const cacheEntrySchema: z$1.ZodObject<{
1674
+ version: z$1.ZodLiteral<1>;
1675
+ key: z$1.ZodString;
1676
+ namespace: z$1.ZodString;
1677
+ operationType: z$1.ZodOptional<z$1.ZodEnum<{
1678
+ span: "span";
1679
+ value: "value";
1680
+ }>>;
1681
+ operationName: z$1.ZodOptional<z$1.ZodString>;
1682
+ spanName: z$1.ZodOptional<z$1.ZodString>;
1683
+ spanKind: z$1.ZodOptional<z$1.ZodString>;
1684
+ storedAt: z$1.ZodString;
1685
+ codeFingerprint: z$1.ZodString;
1686
+ recording: z$1.ZodObject<{
1687
+ returnValue: z$1.ZodUnknown;
1688
+ finalAttributes: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
1689
+ finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
1690
+ error: "error";
1691
+ running: "running";
1692
+ ok: "ok";
1693
+ cancelled: "cancelled";
1694
+ }>>;
1695
+ finalError: z$1.ZodOptional<z$1.ZodObject<{
1696
+ name: z$1.ZodOptional<z$1.ZodString>;
1697
+ message: z$1.ZodString;
1698
+ stack: z$1.ZodOptional<z$1.ZodString>;
1699
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
1700
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>;
1701
+ finalErrors: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
1702
+ name: z$1.ZodOptional<z$1.ZodString>;
1703
+ message: z$1.ZodString;
1704
+ stack: z$1.ZodOptional<z$1.ZodString>;
1705
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
1706
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
1707
+ finalWarning: z$1.ZodOptional<z$1.ZodObject<{
1708
+ name: z$1.ZodOptional<z$1.ZodString>;
1709
+ message: z$1.ZodString;
1710
+ stack: z$1.ZodOptional<z$1.ZodString>;
1711
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
1712
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>;
1713
+ finalWarnings: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
1714
+ name: z$1.ZodOptional<z$1.ZodString>;
1715
+ message: z$1.ZodString;
1716
+ stack: z$1.ZodOptional<z$1.ZodString>;
1717
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
1718
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
1719
+ ops: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
1720
+ kind: z$1.ZodLiteral<"setOutput">;
1721
+ key: z$1.ZodString;
1722
+ value: z$1.ZodUnknown;
1723
+ }, z$1.core.$strip>, z$1.ZodObject<{
1724
+ kind: z$1.ZodLiteral<"appendOutput">;
1725
+ key: z$1.ZodString;
1726
+ value: z$1.ZodUnknown;
1727
+ }, z$1.core.$strip>, z$1.ZodObject<{
1728
+ kind: z$1.ZodLiteral<"mergeOutput">;
1729
+ key: z$1.ZodString;
1730
+ patch: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
1731
+ }, z$1.core.$strip>, z$1.ZodObject<{
1732
+ kind: z$1.ZodLiteral<"incrementOutput">;
1733
+ key: z$1.ZodString;
1734
+ delta: z$1.ZodNumber;
1735
+ }, z$1.core.$strip>, z$1.ZodObject<{
1736
+ kind: z$1.ZodLiteral<"checkpoint">;
1737
+ name: z$1.ZodString;
1738
+ data: z$1.ZodUnknown;
1739
+ }, z$1.core.$strip>, z$1.ZodObject<{
1740
+ kind: z$1.ZodLiteral<"subSpan">;
1741
+ span: z$1.ZodType<SerializedCacheSpan, unknown, z$1.core.$ZodTypeInternals<SerializedCacheSpan, unknown>>;
1742
+ }, z$1.core.$strip>], "kind">>;
1743
+ }, z$1.core.$strip>;
1744
+ }, z$1.core.$strip>;
1551
1745
  /** Persisted cache file contents. */
1552
- type CacheEntry = z.infer<typeof cacheEntrySchema>;
1746
+ type CacheEntry = z$1.infer<typeof cacheEntrySchema>;
1553
1747
  /** Persisted per-owner cache file containing multiple cache entries. */
1554
- declare const cacheFileSchema: z.ZodObject<{
1555
- version: z.ZodLiteral<1>;
1556
- owner: z.ZodString;
1557
- entries: z.ZodRecord<z.ZodString, z.ZodObject<{
1558
- version: z.ZodLiteral<1>;
1559
- key: z.ZodString;
1560
- namespace: z.ZodString;
1561
- spanName: z.ZodString;
1562
- spanKind: z.ZodString;
1563
- storedAt: z.ZodString;
1564
- codeFingerprint: z.ZodString;
1565
- recording: z.ZodObject<{
1566
- returnValue: z.ZodUnknown;
1567
- finalAttributes: z.ZodRecord<z.ZodString, z.ZodUnknown>;
1568
- ops: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
1569
- kind: z.ZodLiteral<"setOutput">;
1570
- key: z.ZodString;
1571
- value: z.ZodUnknown;
1572
- }, z.core.$strip>, z.ZodObject<{
1573
- kind: z.ZodLiteral<"incrementOutput">;
1574
- key: z.ZodString;
1575
- delta: z.ZodNumber;
1576
- }, z.core.$strip>, z.ZodObject<{
1577
- kind: z.ZodLiteral<"checkpoint">;
1578
- name: z.ZodString;
1579
- data: z.ZodUnknown;
1580
- }, z.core.$strip>, z.ZodObject<{
1581
- kind: z.ZodLiteral<"subSpan">;
1582
- span: z.ZodType<SerializedCacheSpan, unknown, z.core.$ZodTypeInternals<SerializedCacheSpan, unknown>>;
1583
- }, z.core.$strip>], "kind">>;
1584
- }, z.core.$strip>;
1585
- }, z.core.$strip>>;
1586
- }, z.core.$strip>;
1748
+ declare const cacheFileSchema: z$1.ZodObject<{
1749
+ version: z$1.ZodLiteral<1>;
1750
+ owner: z$1.ZodString;
1751
+ entries: z$1.ZodRecord<z$1.ZodString, z$1.ZodObject<{
1752
+ version: z$1.ZodLiteral<1>;
1753
+ key: z$1.ZodString;
1754
+ namespace: z$1.ZodString;
1755
+ operationType: z$1.ZodOptional<z$1.ZodEnum<{
1756
+ span: "span";
1757
+ value: "value";
1758
+ }>>;
1759
+ operationName: z$1.ZodOptional<z$1.ZodString>;
1760
+ spanName: z$1.ZodOptional<z$1.ZodString>;
1761
+ spanKind: z$1.ZodOptional<z$1.ZodString>;
1762
+ storedAt: z$1.ZodString;
1763
+ codeFingerprint: z$1.ZodString;
1764
+ recording: z$1.ZodObject<{
1765
+ returnValue: z$1.ZodUnknown;
1766
+ finalAttributes: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
1767
+ finalStatus: z$1.ZodOptional<z$1.ZodEnum<{
1768
+ error: "error";
1769
+ running: "running";
1770
+ ok: "ok";
1771
+ cancelled: "cancelled";
1772
+ }>>;
1773
+ finalError: z$1.ZodOptional<z$1.ZodObject<{
1774
+ name: z$1.ZodOptional<z$1.ZodString>;
1775
+ message: z$1.ZodString;
1776
+ stack: z$1.ZodOptional<z$1.ZodString>;
1777
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
1778
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>;
1779
+ finalErrors: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
1780
+ name: z$1.ZodOptional<z$1.ZodString>;
1781
+ message: z$1.ZodString;
1782
+ stack: z$1.ZodOptional<z$1.ZodString>;
1783
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
1784
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
1785
+ finalWarning: z$1.ZodOptional<z$1.ZodObject<{
1786
+ name: z$1.ZodOptional<z$1.ZodString>;
1787
+ message: z$1.ZodString;
1788
+ stack: z$1.ZodOptional<z$1.ZodString>;
1789
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
1790
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>;
1791
+ finalWarnings: z$1.ZodOptional<z$1.ZodArray<z$1.ZodObject<{
1792
+ name: z$1.ZodOptional<z$1.ZodString>;
1793
+ message: z$1.ZodString;
1794
+ stack: z$1.ZodOptional<z$1.ZodString>;
1795
+ capturedAt: z$1.ZodOptional<z$1.ZodString>;
1796
+ }, z$1.core.$catchall<z$1.ZodUnknown>>>>;
1797
+ ops: z$1.ZodArray<z$1.ZodDiscriminatedUnion<[z$1.ZodObject<{
1798
+ kind: z$1.ZodLiteral<"setOutput">;
1799
+ key: z$1.ZodString;
1800
+ value: z$1.ZodUnknown;
1801
+ }, z$1.core.$strip>, z$1.ZodObject<{
1802
+ kind: z$1.ZodLiteral<"appendOutput">;
1803
+ key: z$1.ZodString;
1804
+ value: z$1.ZodUnknown;
1805
+ }, z$1.core.$strip>, z$1.ZodObject<{
1806
+ kind: z$1.ZodLiteral<"mergeOutput">;
1807
+ key: z$1.ZodString;
1808
+ patch: z$1.ZodRecord<z$1.ZodString, z$1.ZodUnknown>;
1809
+ }, z$1.core.$strip>, z$1.ZodObject<{
1810
+ kind: z$1.ZodLiteral<"incrementOutput">;
1811
+ key: z$1.ZodString;
1812
+ delta: z$1.ZodNumber;
1813
+ }, z$1.core.$strip>, z$1.ZodObject<{
1814
+ kind: z$1.ZodLiteral<"checkpoint">;
1815
+ name: z$1.ZodString;
1816
+ data: z$1.ZodUnknown;
1817
+ }, z$1.core.$strip>, z$1.ZodObject<{
1818
+ kind: z$1.ZodLiteral<"subSpan">;
1819
+ span: z$1.ZodType<SerializedCacheSpan, unknown, z$1.core.$ZodTypeInternals<SerializedCacheSpan, unknown>>;
1820
+ }, z$1.core.$strip>], "kind">>;
1821
+ }, z$1.core.$strip>;
1822
+ }, z$1.core.$strip>>;
1823
+ }, z$1.core.$strip>;
1587
1824
  /** Persisted per-owner cache file contents. */
1588
- type CacheFile = z.infer<typeof cacheFileSchema>;
1825
+ type CacheFile = z$1.infer<typeof cacheFileSchema>;
1589
1826
  //#endregion
1590
1827
  //#region ../sdk/src/types.d.ts
1591
1828
  /** Single authored eval case with its stable identifier and input payload. */
@@ -1594,6 +1831,15 @@ type EvalCase<TInput> = {
1594
1831
  input: TInput;
1595
1832
  tags?: string[];
1596
1833
  };
1834
+ /** Runtime output values collected from `setEvalOutput` and `deriveFromTracing`. */
1835
+ type EvalOutputs = Record<string, unknown>;
1836
+ /**
1837
+ * Schema used to validate and type an eval's collected runtime outputs.
1838
+ *
1839
+ * Zod schemas are supported directly. The runner validates after `execute` and
1840
+ * `deriveFromTracing` finish, before computed scores run.
1841
+ */
1842
+ type EvalOutputsSchema<TOutputs extends EvalOutputs> = z$1.ZodType<TOutputs>;
1597
1843
  /** UI overrides for a derived or scored column emitted by an eval. */
1598
1844
  type EvalColumnOverride = {
1599
1845
  /** Display label shown for the column in tables and detail views. */label?: string;
@@ -1650,13 +1896,13 @@ type EvalDeriveContext<TInput> = {
1650
1896
  case: EvalCase<TInput>;
1651
1897
  };
1652
1898
  /** Context passed to score functions after outputs have been collected. */
1653
- type EvalScoreContext<TInput> = {
1899
+ type EvalScoreContext<TInput, TOutputs extends EvalOutputs = EvalOutputs> = {
1654
1900
  input: TInput;
1655
- outputs: Record<string, unknown>;
1901
+ outputs: TOutputs;
1656
1902
  case: EvalCase<TInput>;
1657
1903
  };
1658
1904
  /** Score callback that computes a numeric result for one case. */
1659
- type EvalScoreFn<TInput> = (ctx: EvalScoreContext<TInput>) => number | Promise<number>;
1905
+ type EvalScoreFn<TInput, TOutputs extends EvalOutputs = EvalOutputs> = (ctx: EvalScoreContext<TInput, TOutputs>) => number | Promise<number>;
1660
1906
  /**
1661
1907
  * Score definition accepted by `defineEval`, with optional UI metadata.
1662
1908
  *
@@ -1665,8 +1911,8 @@ type EvalScoreFn<TInput> = (ctx: EvalScoreContext<TInput>) => number | Promise<n
1665
1911
  * score without a `passThreshold` is informational only and never causes
1666
1912
  * a case to fail on its own.
1667
1913
  */
1668
- type EvalScoreDef<TInput> = EvalScoreFn<TInput> | ({
1669
- compute: EvalScoreFn<TInput>;
1914
+ type EvalScoreDef<TInput, TOutputs extends EvalOutputs = EvalOutputs> = EvalScoreFn<TInput, TOutputs> | ({
1915
+ compute: EvalScoreFn<TInput, TOutputs>;
1670
1916
  passThreshold?: number;
1671
1917
  } & EvalColumnOverride);
1672
1918
  /**
@@ -1684,7 +1930,7 @@ type EvalManualScoreDef = EvalColumnOverride & {
1684
1930
  passThreshold?: number;
1685
1931
  };
1686
1932
  /** Complete authored eval definition consumed by `defineEval`. */
1687
- type EvalDefinition<TInput = unknown> = {
1933
+ type EvalDefinition<TInput = unknown, TOutputs extends EvalOutputs = EvalOutputs> = {
1688
1934
  id: string;
1689
1935
  title?: string;
1690
1936
  /**
@@ -1694,6 +1940,17 @@ type EvalDefinition<TInput = unknown> = {
1694
1940
  * eval once using a synthetic case with empty object input.
1695
1941
  */
1696
1942
  cases?: EvalCase<TInput>[] | (() => Promise<EvalCase<TInput>[]>);
1943
+ /**
1944
+ * Optional schema for runtime outputs collected through `setEvalOutput` and
1945
+ * `deriveFromTracing`.
1946
+ *
1947
+ * The runner validates configured output fields before scoring. For Zod
1948
+ * object schemas, only declared keys are passed to the schema; parsed fields
1949
+ * are merged back into the raw output map, so schema defaults and transforms
1950
+ * apply to configured fields while unconfigured outputs are kept unchanged.
1951
+ * Validation failures mark the case as failed and skip computed scores.
1952
+ */
1953
+ outputsSchema?: EvalOutputsSchema<TOutputs>;
1697
1954
  columns?: EvalColumns;
1698
1955
  /**
1699
1956
  * Per-eval trace attribute display rules for the UI.
@@ -1704,8 +1961,8 @@ type EvalDefinition<TInput = unknown> = {
1704
1961
  */
1705
1962
  traceDisplay?: TraceDisplayInputConfig;
1706
1963
  execute: (ctx: EvalExecuteContext<TInput>) => Promise<void> | void;
1707
- deriveFromTracing?: (ctx: EvalDeriveContext<TInput>) => Record<string, unknown> | Promise<Record<string, unknown>>;
1708
- scores?: Record<string, EvalScoreDef<TInput>>;
1964
+ deriveFromTracing?: (ctx: EvalDeriveContext<TInput>) => Partial<TOutputs> | Promise<Partial<TOutputs>>;
1965
+ scores?: Record<string, EvalScoreDef<TInput, TOutputs>>;
1709
1966
  /**
1710
1967
  * Score columns whose values are entered in the web UI after a run.
1711
1968
  *
@@ -1755,7 +2012,7 @@ type EvalDefinition<TInput = unknown> = {
1755
2012
  type EvalRegistryEntry = {
1756
2013
  id: string;
1757
2014
  title?: string;
1758
- use: <R>(fn: <TInput>(def: EvalDefinition<TInput>) => R) => R;
2015
+ use: <R>(fn: <TInput, TOutputs extends EvalOutputs>(def: EvalDefinition<TInput, TOutputs>) => R) => R;
1759
2016
  };
1760
2017
  /** Return the in-memory registry of evals defined in the current process. */
1761
2018
  declare function getEvalRegistry(): Map<string, EvalRegistryEntry>;
@@ -1763,7 +2020,7 @@ declare function getEvalRegistry(): Map<string, EvalRegistryEntry>;
1763
2020
  * Register an eval definition with the SDK so the runner can discover it
1764
2021
  * after importing the eval module.
1765
2022
  */
1766
- declare function defineEval<TInput>(definition: EvalDefinition<TInput>): void;
2023
+ declare function defineEval<TInput = unknown, TOutputs extends EvalOutputs = EvalOutputs>(definition: EvalDefinition<TInput, TOutputs>): void;
1767
2024
  //#endregion
1768
2025
  //#region ../sdk/src/repoFile.d.ts
1769
2026
  /**
@@ -1778,7 +2035,7 @@ declare function repoFile(path: string, mimeType?: string): RepoFileRef;
1778
2035
  //#endregion
1779
2036
  //#region ../sdk/src/runtime.d.ts
1780
2037
  /**
1781
- * Adapter used by the SDK to read and write cache entries for cached spans.
2038
+ * Adapter used by the SDK to read and write cache entries.
1782
2039
  *
1783
2040
  * Implementations are typically injected by the runner before the eval case
1784
2041
  * starts executing.
@@ -1794,15 +2051,16 @@ type CacheScopeContext = {
1794
2051
  evalId: string; /** Hash of the eval source file; used to invalidate on code changes. */
1795
2052
  codeFingerprint: string;
1796
2053
  };
1797
- /** Active recording frame captured while a cached span body executes. */
2054
+ /** Active recording frame captured while a cached operation body executes. */
1798
2055
  type CacheRecordingFrame = {
1799
- /** Length of `scope.spans` immediately before the cached body started. */baseSpanIndex: number; /** Id of the cached span that owns this recording. */
1800
- cachedSpanId: string; /** Ordered observable effects recorded during the cached body. */
2056
+ /** Length of `scope.spans` immediately before the cached body started. */baseSpanIndex: number; /** Parent id used when recording and replaying direct child spans. */
2057
+ replayParentSpanId: string | null; /** Ordered observable effects recorded during the cached body. */
1801
2058
  ops: CacheRecordingOp[];
1802
2059
  };
1803
2060
  /** Mutable per-case runtime state stored in async local storage. */
1804
2061
  type EvalCaseScope = {
1805
- caseId: string;
2062
+ caseId: string; /** Authored input for the current case, when provided by the runner. */
2063
+ input?: unknown;
1806
2064
  outputs: Record<string, unknown>; /** Structured assertion failures recorded for the current case. */
1807
2065
  assertionFailures: AssertionFailure[];
1808
2066
  spans: EvalTraceSpan[];
@@ -1815,7 +2073,7 @@ type EvalCaseScope = {
1815
2073
  */
1816
2074
  recordingStack: CacheRecordingFrame[];
1817
2075
  /**
1818
- * Incremented while replaying a cached span, so nested SDK calls do not
2076
+ * Incremented while replaying a cached operation, so nested SDK calls do not
1819
2077
  * accidentally double-record ops into outer recorders.
1820
2078
  */
1821
2079
  replayingDepth: number; /** Runner-provided cache adapter + mode; absent when caching is disabled. */
@@ -1834,6 +2092,15 @@ declare function getCurrentScope(): EvalCaseScope | undefined;
1834
2092
  * behavior without importing or inspecting the full eval scope.
1835
2093
  */
1836
2094
  declare function isInEvalScope(): boolean;
2095
+ /**
2096
+ * Return the authored input for the current eval case.
2097
+ *
2098
+ * Pass a dot-separated path to read nested values, for example
2099
+ * `getEvalCaseInput('customer.tier')`. Calls outside an eval case scope return
2100
+ * `undefined` so shared workflow code can safely use this helper.
2101
+ */
2102
+ declare function getEvalCaseInput(): unknown;
2103
+ declare function getEvalCaseInput(path: string): unknown;
1837
2104
  /**
1838
2105
  * Attach cache context (adapter, mode, eval id, fingerprint) to a scope.
1839
2106
  *
@@ -1843,7 +2110,8 @@ declare function isInEvalScope(): boolean;
1843
2110
  declare function setScopeCacheContext(scope: EvalCaseScope, context: CacheScopeContext): void;
1844
2111
  /** Optional inputs accepted when starting a new eval case scope. */
1845
2112
  type RunInEvalScopeOptions = {
1846
- /** Cache adapter + mode attached to the scope before `fn` runs. */cacheContext?: CacheScopeContext;
2113
+ /** Authored input for the active eval case. */input?: unknown; /** Cache adapter + mode attached to the scope before `fn` runs. */
2114
+ cacheContext?: CacheScopeContext;
1847
2115
  };
1848
2116
  /**
1849
2117
  * Execute a callback inside a fresh eval case scope and capture its outputs,
@@ -1861,6 +2129,20 @@ declare function runInEvalScope<T>(caseId: string, fn: () => Promise<T> | T, opt
1861
2129
  * refs, and native `Blob`/`File` instances for media or file columns.
1862
2130
  */
1863
2131
  declare function setEvalOutput(key: string, value: unknown): void;
2132
+ /**
2133
+ * Append an item to an output array in the current case scope.
2134
+ *
2135
+ * Missing values become `[value]`, existing arrays receive the item, and
2136
+ * existing scalar/object values are preserved as `[existing, value]`.
2137
+ */
2138
+ declare function appendToEvalOutput(key: string, value: unknown): void;
2139
+ /**
2140
+ * Shallow-merge object fields into an output value in the current case scope.
2141
+ *
2142
+ * Missing values become a copy of `patch`. Non-object existing values are
2143
+ * recorded as assertion failures instead of being replaced.
2144
+ */
2145
+ declare function mergeEvalOutput(key: string, patch: Record<string, unknown>): void;
1864
2146
  /**
1865
2147
  * Add a numeric delta to an output value in the current case scope.
1866
2148
  *
@@ -1876,17 +2158,69 @@ declare function incrementEvalOutput(key: string, delta: number): void;
1876
2158
  */
1877
2159
  declare function evalAssert(condition: boolean, message: string): void;
1878
2160
  //#endregion
1879
- //#region ../sdk/src/tracer.d.ts
2161
+ //#region ../sdk/src/traceDiagnostics.d.ts
2162
+ /** Severity used when attaching a recoverable diagnostic to an active span. */
2163
+ type CaptureEvalSpanErrorLevel = 'error' | 'warning';
2164
+ /** Options accepted by `captureEvalSpanError(...)`. */
2165
+ type CaptureEvalSpanErrorOptions = {
2166
+ /**
2167
+ * Captured diagnostic severity.
2168
+ *
2169
+ * `error` marks the active span as errored. `warning` records the diagnostic
2170
+ * without changing an otherwise successful span's status.
2171
+ */
2172
+ level?: CaptureEvalSpanErrorLevel;
2173
+ };
2174
+ //#endregion
2175
+ //#region ../sdk/src/cacheRecording.d.ts
2176
+ /** Cache reference appended to the active span by `evalTracer.cache(...)`. */
2177
+ type TraceCacheRef = {
2178
+ type: 'value';
2179
+ name: string;
2180
+ namespace: string;
2181
+ key: string;
2182
+ status: 'hit' | 'miss' | 'refresh' | 'bypass';
2183
+ storedAt?: string;
2184
+ age?: number;
2185
+ };
2186
+ //#endregion
2187
+ //#region ../sdk/src/valueCache.d.ts
2188
+ /** Info accepted by `evalTracer.cache(info, fn)` for spanless value caching. */
2189
+ type TraceCacheInfo = {
2190
+ /** Display name used for cache listings and the default namespace. */name: string; /** Arbitrary JSON-safe value used to derive the cache key. */
2191
+ key: unknown; /** Override the default namespace (`${evalId}__${name}`). */
2192
+ namespace?: string;
2193
+ };
2194
+ //#endregion
2195
+ //#region ../sdk/src/cacheKey.d.ts
2196
+ type CacheKeyHashInput = {
2197
+ namespace: string;
2198
+ codeFingerprint: string;
2199
+ key: unknown;
2200
+ };
1880
2201
  /**
1881
- * Mutable handle for the current span.
2202
+ * Hash the components of a cache key into a deterministic hex digest.
1882
2203
  *
1883
- * Prefer the ambient `evalSpan` export for most code so helpers deeper in the call
1884
- * stack can annotate the active span without receiving an injected argument.
2204
+ * Native `Blob` and `File` values are read asynchronously and hashed by
2205
+ * content. Use `hashCacheKeySync` only when the key contains no async values.
1885
2206
  */
2207
+ declare function hashCacheKey(input: CacheKeyHashInput): Promise<string>;
2208
+ /**
2209
+ * Synchronously hash cache key components. This supports JSON-like data and
2210
+ * in-memory binary values such as `Buffer`, `ArrayBuffer`, and typed arrays,
2211
+ * but cannot content-hash native `Blob` or `File` values.
2212
+ */
2213
+ declare function hashCacheKeySync(input: CacheKeyHashInput): string;
2214
+ //#endregion
2215
+ //#region ../sdk/src/tracer.d.ts
2216
+ /** Mutable handle for the current span. Prefer ambient `evalSpan` in helpers. */
1886
2217
  type TraceActiveSpan = {
1887
2218
  /** Rename the active span after it has been created. */setName(value: string): void; /** Set a single attribute on the active span. Later writes replace the same key. */
1888
2219
  setAttribute(key: string, value: unknown): void; /** Merge multiple attributes into the active span. */
1889
- setAttributes(value: Record<string, unknown>): void;
2220
+ setAttributes(value: Record<string, unknown>): void; /** Add a numeric delta to one attribute. */
2221
+ incrementAttribute(key: string, delta: number): void; /** Append one item to an attribute array, preserving an existing scalar. */
2222
+ appendToAttribute(key: string, value: unknown): void; /** Shallow-merge object fields into one attribute. */
2223
+ mergeAttribute(key: string, patch: Record<string, unknown>): void;
1890
2224
  };
1891
2225
  /** Timestamp accepted by the external span lifecycle API. */
1892
2226
  type TraceSpanTimestamp = Date | string;
@@ -1905,7 +2239,9 @@ type TraceExternalSpanUpdateInfo = {
1905
2239
  name?: string; /** Attributes to merge into the recorded span. */
1906
2240
  attributes?: Record<string, unknown>; /** Optional status override, useful when the upstream tracer emits one. */
1907
2241
  status?: EvalTraceSpan['status']; /** Optional error payload to attach to the span. */
1908
- error?: EvalTraceSpan['error'];
2242
+ error?: EvalTraceSpan['error']; /** Optional latest warning payload to attach to the span. */
2243
+ warning?: EvalTraceSpan['warning']; /** Optional warning payloads to attach to the span. */
2244
+ warnings?: EvalTraceSpan['warnings'];
1909
2245
  };
1910
2246
  /** Info accepted by `evalTracer.endSpan(info)` for externally managed spans. */
1911
2247
  type TraceExternalSpanEndInfo = TraceExternalSpanUpdateInfo & {
@@ -1921,7 +2257,9 @@ type TraceExternalSpanRecordInfo = {
1921
2257
  endedAt?: TraceSpanTimestamp | null; /** Final span status. Defaults to `ok`. */
1922
2258
  status?: EvalTraceSpan['status']; /** Final span attributes. */
1923
2259
  attributes?: Record<string, unknown>; /** Optional error payload to attach to the span. */
1924
- error?: EvalTraceSpan['error'];
2260
+ error?: EvalTraceSpan['error']; /** Optional latest warning payload to attach to the span. */
2261
+ warning?: EvalTraceSpan['warning']; /** Optional warning payloads to attach to the span. */
2262
+ warnings?: EvalTraceSpan['warnings'];
1925
2263
  };
1926
2264
  /** Mutable handle returned by `evalTracer.startSpan(...)`. */
1927
2265
  type TraceExternalSpanHandle = TraceActiveSpan & {
@@ -1938,6 +2276,15 @@ declare function recordExternalSpan(info: TraceExternalSpanRecordInfo): string;
1938
2276
  * Calls are no-ops when executed outside of `evalTracer.span(...)`.
1939
2277
  */
1940
2278
  declare const evalSpan: TraceActiveSpan;
2279
+ /**
2280
+ * Attach one or more recoverable errors to the active eval span.
2281
+ *
2282
+ * By default the active span is marked as `error` even if its callback later
2283
+ * completes without throwing. Pass `'warning'` or `{ level: 'warning' }` as the
2284
+ * final argument to record the diagnostic without changing span status. Calls
2285
+ * outside `evalTracer.span(...)` are ignored.
2286
+ */
2287
+ declare function captureEvalSpanError(errorOrErrors: unknown, ...additionalErrorsOrOptions: readonly unknown[]): void;
1941
2288
  type TraceSpanInfoBase = {
1942
2289
  kind: string;
1943
2290
  name: string;
@@ -1969,6 +2316,16 @@ declare function traceSpan(info: TraceSpanInfoCached, fn: (span: TraceActiveSpan
1969
2316
  */
1970
2317
  declare const evalTracer: {
1971
2318
  /** Run a callback inside a new trace span and record its lifecycle. */span: typeof traceSpan;
2319
+ /**
2320
+ * Cache a pure value without creating a trace span.
2321
+ *
2322
+ * When called inside an active span, the span receives a `cache.refs` entry
2323
+ * describing the value cache status for this run.
2324
+ */
2325
+ cache: {
2326
+ <T>(info: TraceCacheInfo, fn: () => Promise<T> | T): Promise<T>;
2327
+ (info: TraceCacheInfo, fn: () => unknown): Promise<unknown>;
2328
+ };
1972
2329
  /**
1973
2330
  * Start a span whose lifecycle is controlled by an external tracer/exporter.
1974
2331
  *
@@ -2001,24 +2358,6 @@ declare const evalTracer: {
2001
2358
  };
2002
2359
  /** Build a queryable trace tree helper from a flat span list and checkpoints. */
2003
2360
  declare function buildTraceTree(spans: EvalTraceSpan[], checkpoints: Map<string, unknown>): EvalTraceTree;
2004
- type CacheKeyHashInput = {
2005
- namespace: string;
2006
- codeFingerprint: string;
2007
- key: unknown;
2008
- };
2009
- /**
2010
- * Hash the components of a cache key into a deterministic hex digest.
2011
- *
2012
- * Native `Blob` and `File` values are read asynchronously and hashed by
2013
- * content. Use `hashCacheKeySync` only when the key contains no async values.
2014
- */
2015
- declare function hashCacheKey(input: CacheKeyHashInput): Promise<string>;
2016
- /**
2017
- * Synchronously hash cache key components. This supports JSON-like data and
2018
- * in-memory binary values such as `Buffer`, `ArrayBuffer`, and typed arrays,
2019
- * but cannot content-hash native `Blob` or `File` values.
2020
- */
2021
- declare function hashCacheKeySync(input: CacheKeyHashInput): string;
2022
2361
  //#endregion
2023
2362
  //#region ../runner/src/cacheStore.d.ts
2024
2363
  /** Filter accepted by `FsCacheStore.clear` to narrow the set of entries removed. */
@@ -2048,7 +2387,8 @@ type EvalRunner = {
2048
2387
  cancelRun(id: string): void; /** Return full details for a single case in a run, when available. */
2049
2388
  getCaseDetail(runId: string, caseId: string): CaseDetail | undefined; /** Subscribe to streamed events for a specific run. */
2050
2389
  subscribe(runId: string, listener: (event: SseEnvelope) => void): () => void; /** Subscribe to discovery updates triggered by file changes or manual refresh. */
2051
- subscribeDiscovery(listener: (event: SseEnvelope) => void): () => void; /** Resolve the workspace root backing this runner instance. */
2390
+ subscribeDiscovery(listener: (event: SseEnvelope) => void): () => void; /** Stop background filesystem watchers owned by this runner instance. */
2391
+ close(): Promise<void>; /** Resolve the workspace root backing this runner instance. */
2052
2392
  getWorkspaceRoot(): string; /** Resolve a persisted artifact path when artifact storage is supported. */
2053
2393
  getArtifactPath(artifactId: string): string | undefined; /** Return summaries for every persisted cache entry in the workspace. */
2054
2394
  listCache(): Promise<CacheListItem[]>;
@@ -2106,4 +2446,4 @@ declare function createRunner({
2106
2446
  */
2107
2447
  declare function runCli(argv: string[]): Promise<void>;
2108
2448
  //#endregion
2109
- export { type AgentEvalsConfig, type AssertionFailure, type CacheAdapter, type CacheEntry, type CacheFile, type CacheListItem, type CacheMode, type CacheRecording, type CacheRecordingFrame, type CacheRecordingOp, type CacheScopeContext, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type CreateRunRequest, type DerivedStatus, EvalAssertionError, type EvalCase, type EvalCaseScope, type EvalChartAggregate, type EvalChartAxis, type EvalChartBuiltinMetric, type EvalChartColor, type EvalChartConfig, type EvalChartMetric, type EvalChartTooltipExtra, type EvalChartType, type EvalChartsConfig, type EvalColumnOverride, type EvalColumns, type EvalDefinition, type EvalDeriveContext, type EvalDisplayStatus, type EvalExecuteContext, type EvalFreshnessStatus, type EvalManualScoreDef, type EvalRunner, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, type EvalTraceSpan, type EvalTraceTree, type FileRef, type JsonCell, type NumberDisplayOptions, type RepoFileRef, type RunArtifactRef, type RunInEvalScopeOptions, type RunManifest, type RunSummary, type ScalarCell, type ScopedCaseSummary, type ScoreTrace, type SerializedCacheSpan, type SpanCacheOptions, type SseEnvelope, type SseEventType, type TraceActiveSpan, type TraceAttributeDisplay, type TraceAttributeDisplayFormat, type TraceAttributeDisplayInput, type TraceAttributeDisplayPlacement, type TraceAttributeTransform, type TraceAttributeTransformContext, type TraceDisplayConfig, type TraceDisplayInputConfig, type TraceSpanInfo, type TrialSelectionMode, type UpdateManualScoreRequest, agentEvalsConfigSchema, assertionFailureSchema, buildTraceTree, cacheEntrySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheRecordingOpSchema, cacheRecordingSchema, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, fileRefSchema, getCurrentScope, getEvalDisplayStatus, getEvalRegistry, getEvalTitle, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, numberDisplayOptionsSchema, repoFile, repoFileRefSchema, runArtifactRefSchema, runCli, runInEvalScope, runManifestSchema, runSummarySchema, scoreTraceSchema, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanKindSchema, traceSpanSchema, trialSelectionModeSchema, updateManualScoreRequestSchema };
2449
+ export { type AgentEvalsConfig, type AssertionFailure, type CacheAdapter, type CacheEntry, type CacheFile, type CacheListItem, type CacheMode, type CacheOperationType, type CacheRecording, type CacheRecordingFrame, type CacheRecordingOp, type CacheScopeContext, type CaptureEvalSpanErrorLevel, type CaptureEvalSpanErrorOptions, type CaseDetail, type CaseRow, type CellValue, type ColumnDef, type ColumnFormat, type ColumnKind, type CreateRunRequest, type DerivedStatus, EvalAssertionError, type EvalCase, type EvalCaseScope, type EvalChartAggregate, type EvalChartAxis, type EvalChartBuiltinMetric, type EvalChartColor, type EvalChartConfig, type EvalChartMetric, type EvalChartTooltipExtra, type EvalChartType, type EvalChartsConfig, type EvalColumnOverride, type EvalColumns, type EvalDefinition, type EvalDeriveContext, type EvalDisplayStatus, type EvalExecuteContext, type EvalFreshnessStatus, type EvalManualScoreDef, type EvalOutputs, type EvalOutputsSchema, type EvalRunner, type EvalScoreContext, type EvalScoreDef, type EvalScoreFn, type EvalStatAggregate, type EvalStatItem, type EvalStatsConfig, type EvalSummary, type EvalTraceSpan, type EvalTraceSpanError, type EvalTraceSpanWarning, type EvalTraceTree, type FileRef, type JsonCell, type NumberDisplayOptions, type RepoFileRef, type RunArtifactRef, type RunInEvalScopeOptions, type RunManifest, type RunSummary, type ScalarCell, type ScopedCaseSummary, type ScoreTrace, type SerializedCacheSpan, type SpanCacheOptions, type SseEnvelope, type SseEventType, type TraceActiveSpan, type TraceAttributeDisplay, type TraceAttributeDisplayFormat, type TraceAttributeDisplayInput, type TraceAttributeDisplayPlacement, type TraceAttributeTransform, type TraceAttributeTransformContext, type TraceCacheInfo, type TraceCacheRef, type TraceDisplayConfig, type TraceDisplayInputConfig, type TraceSpanInfo, type TrialSelectionMode, type UpdateManualScoreRequest, agentEvalsConfigSchema, appendToEvalOutput, assertionFailureSchema, buildTraceTree, cacheEntrySchema, cacheFileSchema, cacheListItemSchema, cacheModeSchema, cacheOperationTypeSchema, cacheRecordingOpSchema, cacheRecordingSchema, captureEvalSpanError, caseDetailSchema, caseRowSchema, cellValueSchema, columnDefSchema, columnFormatSchema, columnKindSchema, createRunRequestSchema, createRunner, defineEval, deriveScopedSummaryFromCases, deriveStatusFromCaseRows, deriveStatusFromChildStatuses, evalAssert, evalChartAggregateSchema, evalChartAxisSchema, evalChartBuiltinMetricSchema, evalChartColorSchema, evalChartConfigSchema, evalChartMetricSchema, evalChartTooltipExtraSchema, evalChartTypeSchema, evalChartsConfigSchema, evalFreshnessStatusSchema, evalSpan, evalStatAggregateSchema, evalStatItemSchema, evalStatsConfigSchema, evalSummarySchema, evalTracer, fileRefSchema, getCurrentScope, getEvalCaseInput, getEvalDisplayStatus, getEvalRegistry, getEvalTitle, hashCacheKey, hashCacheKeySync, incrementEvalOutput, isInEvalScope, jsonCellSchema, mergeEvalOutput, numberDisplayOptionsSchema, repoFile, repoFileRefSchema, runArtifactRefSchema, runCli, runInEvalScope, runManifestSchema, runSummarySchema, scoreTraceSchema, serializedCacheSpanSchema, setEvalOutput, setScopeCacheContext, spanCacheOptionsSchema, sseEnvelopeSchema, traceAttributeDisplayFormatSchema, traceAttributeDisplayInputSchema, traceAttributeDisplayPlacementSchema, traceAttributeDisplaySchema, traceDisplayConfigSchema, traceDisplayInputConfigSchema, traceSpanErrorSchema, traceSpanKindSchema, traceSpanSchema, traceSpanWarningSchema, trialSelectionModeSchema, updateManualScoreRequestSchema, z };