@alis-build/alis-evals-v1 1.24.0 → 1.26.0

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.
@@ -1,7 +1,31 @@
1
- // @generated by protoc-gen-es v2.13.0 with parameter "target=js+dts,import_extension=js,json_types=true"
1
+ // @generated by protoc-gen-es v2.16.0 with parameter "target=js+dts,import_extension=js,json_types=true"
2
2
  // @generated from file alis/evals/v1/evaluation.proto (package alis.evals.v1, syntax proto3)
3
3
  /* eslint-disable */
4
4
 
5
+ // Runs post-deploy tests and evaluations and records the outcome of each
6
+ // finished suite as a `Run` (`runs/{run_id}`).
7
+ //
8
+ // `TestService` has one RPC per kind of run. Each starts a
9
+ // `google.longrunning.Operation` whose response lists the names of the runs
10
+ // it produced:
11
+ //
12
+ // - `RunIntegrationTest`: integration cases scored by deterministic checks.
13
+ // - `RunLoadTest`: generated traffic at a `RunLoadTestRequest.Mode` preset,
14
+ // with Service Level Objective (SLO) checks on latency, errors, and
15
+ // throughput.
16
+ // - `RunAgentEval`: agent evaluation cases scored by deterministic checks,
17
+ // LLM-as-judge scores, rubric dimensions, and ADK prebuilt metrics.
18
+ // - `RunInfraObservation`: Cloud Monitoring snapshots of declared Cloud Run
19
+ // and Spanner targets, without generating load.
20
+ //
21
+ // Requests select cases by id: `suite.case` selects one case, `suite` selects
22
+ // every case in that suite, and an empty list runs every registered case. An
23
+ // operation completes successfully once execution finishes, even when cases
24
+ // fail, so read `Run.status` for the outcome. Runs, cases, checks, and
25
+ // metrics all report a `Status`. Each completed run is published as a
26
+ // `RunPublishedEvent` for persistence and downstream consumers such as
27
+ // BigQuery or Pub/Sub.
28
+
5
29
  import type { GenEnum, GenFile, GenMessage, GenService } from "@bufbuild/protobuf/codegenv2";
6
30
  import type { JsonObject, Message } from "@bufbuild/protobuf";
7
31
  import type { Duration, DurationJson, StructJson, Timestamp, TimestampJson } from "@bufbuild/protobuf/wkt";
@@ -14,52 +38,86 @@ import type { OperationSchema } from "@alis-build/google-longrunning/google/long
14
38
  export declare const file_alis_evals_v1_evaluation: GenFile;
15
39
 
16
40
  /**
41
+ * Represents the outcome of a completed suite within a test or evaluation job.
42
+ * This is the common persistence and event payload for all run kinds.
43
+ * Every branch lists each executed case with its rolled-up status and all
44
+ * evaluated checks, passed and failed alike. Multiple Runs may share the same
45
+ * parent operation.
46
+ *
17
47
  * @generated from message alis.evals.v1.Run
18
48
  */
19
49
  export declare type Run = Message<"alis.evals.v1.Run"> & {
20
50
  /**
51
+ * The resource name of the run.
52
+ *
53
+ * Format: `runs/{run_id}`
54
+ *
55
+ * `{run_id}` is independent of the long-running operation identifier.
56
+ *
21
57
  * @generated from field: string name = 2;
22
58
  */
23
59
  name: string;
24
60
 
25
61
  /**
62
+ * The batch ID of the run.
63
+ * This is populated when batch id is provided in the request.
64
+ *
26
65
  * @generated from field: optional string batch_id = 3;
27
66
  */
28
67
  batchId?: string | undefined;
29
68
 
30
69
  /**
70
+ * The type of run. Set explicitly so consumers can filter without inspecting
71
+ * the data branch.
72
+ *
31
73
  * @generated from field: alis.evals.v1.Run.Type type = 4;
32
74
  */
33
75
  type: Run_Type;
34
76
 
35
77
  /**
78
+ * The rolled-up outcome for the run.
79
+ *
80
+ * Precedence: FAILED if any case or check failed, errored, or scored
81
+ * below threshold; else PASSED when at least one case ran successfully.
82
+ *
36
83
  * @generated from field: alis.evals.v1.Status status = 5;
37
84
  */
38
85
  status: Status;
39
86
 
40
87
  /**
88
+ * Branch-specific result detail. Exactly one branch is set; it corresponds to
89
+ * `type`.
90
+ *
41
91
  * @generated from oneof alis.evals.v1.Run.data
42
92
  */
43
93
  data: {
44
94
  /**
95
+ * Integration test case results (RunIntegrationTest).
96
+ *
45
97
  * @generated from field: alis.evals.v1.IntegrationTestResults integration_test = 6;
46
98
  */
47
99
  value: IntegrationTestResults;
48
100
  case: "integrationTest";
49
101
  } | {
50
102
  /**
103
+ * Load test case results (RunLoadTest).
104
+ *
51
105
  * @generated from field: alis.evals.v1.LoadTestResults load_test = 7;
52
106
  */
53
107
  value: LoadTestResults;
54
108
  case: "loadTest";
55
109
  } | {
56
110
  /**
111
+ * Agent evaluation case results (RunAgentEval).
112
+ *
57
113
  * @generated from field: alis.evals.v1.AgentEvalResults agent_eval = 8;
58
114
  */
59
115
  value: AgentEvalResults;
60
116
  case: "agentEval";
61
117
  } | {
62
118
  /**
119
+ * Infrastructure observation case results (RunInfraObservation).
120
+ *
63
121
  * @generated from field: alis.evals.v1.InfraObservationResults infra_observation = 9;
64
122
  */
65
123
  value: InfraObservationResults;
@@ -67,106 +125,169 @@ export declare type Run = Message<"alis.evals.v1.Run"> & {
67
125
  } | { case: undefined; value?: undefined };
68
126
 
69
127
  /**
128
+ * Execution wall-clock start. Distinct from resource lifecycle timestamps.
129
+ *
70
130
  * @generated from field: google.protobuf.Timestamp start_time = 21;
71
131
  */
72
132
  startTime?: Timestamp | undefined;
73
133
 
74
134
  /**
135
+ * Execution wall-clock end.
136
+ *
75
137
  * @generated from field: google.protobuf.Timestamp end_time = 22;
76
138
  */
77
139
  endTime?: Timestamp | undefined;
78
140
 
79
141
  /**
142
+ * The parent long-running operation for this run.
143
+ *
144
+ * Format: `operations/{operation_id}`
145
+ *
80
146
  * @generated from field: string operation = 23;
81
147
  */
82
148
  operation: string;
83
149
 
84
150
  /**
151
+ * Optional. Populated when the run could not complete. Root-cause summary
152
+ * for triage.
153
+ *
85
154
  * @generated from field: optional google.rpc.Status error = 24;
86
155
  */
87
156
  error?: Status$1 | undefined;
88
157
 
89
158
  /**
159
+ * The time when this run resource was materialized.
160
+ *
90
161
  * @generated from field: google.protobuf.Timestamp create_time = 25;
91
162
  */
92
163
  createTime?: Timestamp | undefined;
93
164
 
94
165
  /**
166
+ * The project ID of the environment where the run was executed.
167
+ * Example: marvel-sm-dev-123, marvel-sm-prod-456
168
+ *
95
169
  * @generated from field: string google_project_id = 26;
96
170
  */
97
171
  googleProjectId: string;
98
172
  };
99
173
 
100
174
  /**
175
+ * Represents the outcome of a completed suite within a test or evaluation job.
176
+ * This is the common persistence and event payload for all run kinds.
177
+ * Every branch lists each executed case with its rolled-up status and all
178
+ * evaluated checks, passed and failed alike. Multiple Runs may share the same
179
+ * parent operation.
180
+ *
101
181
  * @generated from message alis.evals.v1.Run
102
182
  */
103
183
  export declare type RunJson = {
104
184
  /**
185
+ * The resource name of the run.
186
+ *
187
+ * Format: `runs/{run_id}`
188
+ *
189
+ * `{run_id}` is independent of the long-running operation identifier.
190
+ *
105
191
  * @generated from field: string name = 2;
106
192
  */
107
193
  name?: string;
108
194
 
109
195
  /**
196
+ * The batch ID of the run.
197
+ * This is populated when batch id is provided in the request.
198
+ *
110
199
  * @generated from field: optional string batch_id = 3;
111
200
  */
112
201
  batchId?: string;
113
202
 
114
203
  /**
204
+ * The type of run. Set explicitly so consumers can filter without inspecting
205
+ * the data branch.
206
+ *
115
207
  * @generated from field: alis.evals.v1.Run.Type type = 4;
116
208
  */
117
209
  type?: Run_TypeJson;
118
210
 
119
211
  /**
212
+ * The rolled-up outcome for the run.
213
+ *
214
+ * Precedence: FAILED if any case or check failed, errored, or scored
215
+ * below threshold; else PASSED when at least one case ran successfully.
216
+ *
120
217
  * @generated from field: alis.evals.v1.Status status = 5;
121
218
  */
122
219
  status?: StatusJson;
123
220
 
124
221
  /**
222
+ * Integration test case results (RunIntegrationTest).
223
+ *
125
224
  * @generated from field: alis.evals.v1.IntegrationTestResults integration_test = 6;
126
225
  */
127
226
  integrationTest?: IntegrationTestResultsJson;
128
227
 
129
228
  /**
229
+ * Load test case results (RunLoadTest).
230
+ *
130
231
  * @generated from field: alis.evals.v1.LoadTestResults load_test = 7;
131
232
  */
132
233
  loadTest?: LoadTestResultsJson;
133
234
 
134
235
  /**
236
+ * Agent evaluation case results (RunAgentEval).
237
+ *
135
238
  * @generated from field: alis.evals.v1.AgentEvalResults agent_eval = 8;
136
239
  */
137
240
  agentEval?: AgentEvalResultsJson;
138
241
 
139
242
  /**
243
+ * Infrastructure observation case results (RunInfraObservation).
244
+ *
140
245
  * @generated from field: alis.evals.v1.InfraObservationResults infra_observation = 9;
141
246
  */
142
247
  infraObservation?: InfraObservationResultsJson;
143
248
 
144
249
  /**
250
+ * Execution wall-clock start. Distinct from resource lifecycle timestamps.
251
+ *
145
252
  * @generated from field: google.protobuf.Timestamp start_time = 21;
146
253
  */
147
254
  startTime?: TimestampJson;
148
255
 
149
256
  /**
257
+ * Execution wall-clock end.
258
+ *
150
259
  * @generated from field: google.protobuf.Timestamp end_time = 22;
151
260
  */
152
261
  endTime?: TimestampJson;
153
262
 
154
263
  /**
264
+ * The parent long-running operation for this run.
265
+ *
266
+ * Format: `operations/{operation_id}`
267
+ *
155
268
  * @generated from field: string operation = 23;
156
269
  */
157
270
  operation?: string;
158
271
 
159
272
  /**
273
+ * Optional. Populated when the run could not complete. Root-cause summary
274
+ * for triage.
275
+ *
160
276
  * @generated from field: optional google.rpc.Status error = 24;
161
277
  */
162
278
  error?: StatusJson$1;
163
279
 
164
280
  /**
281
+ * The time when this run resource was materialized.
282
+ *
165
283
  * @generated from field: google.protobuf.Timestamp create_time = 25;
166
284
  */
167
285
  createTime?: TimestampJson;
168
286
 
169
287
  /**
288
+ * The project ID of the environment where the run was executed.
289
+ * Example: marvel-sm-dev-123, marvel-sm-prod-456
290
+ *
170
291
  * @generated from field: string google_project_id = 26;
171
292
  */
172
293
  googleProjectId?: string;
@@ -179,36 +300,50 @@ export declare type RunJson = {
179
300
  export declare const RunSchema: GenMessage<Run, {jsonType: RunJson}>;
180
301
 
181
302
  /**
303
+ * Identifies which kind of test or evaluation produced a Run.
304
+ *
182
305
  * @generated from enum alis.evals.v1.Run.Type
183
306
  */
184
307
  export enum Run_Type {
185
308
  /**
309
+ * Default value. This value is unused.
310
+ *
186
311
  * @generated from enum value: TYPE_UNSPECIFIED = 0;
187
312
  */
188
313
  TYPE_UNSPECIFIED = 0,
189
314
 
190
315
  /**
316
+ * An integration test run (RunIntegrationTest).
317
+ *
191
318
  * @generated from enum value: INTEGRATION_TEST = 1;
192
319
  */
193
320
  INTEGRATION_TEST = 1,
194
321
 
195
322
  /**
323
+ * A load test run (RunLoadTest).
324
+ *
196
325
  * @generated from enum value: LOAD_TEST = 2;
197
326
  */
198
327
  LOAD_TEST = 2,
199
328
 
200
329
  /**
330
+ * An agent evaluation run (RunAgentEval).
331
+ *
201
332
  * @generated from enum value: AGENT_EVAL = 3;
202
333
  */
203
334
  AGENT_EVAL = 3,
204
335
 
205
336
  /**
337
+ * An infrastructure observation run (RunInfraObservation).
338
+ *
206
339
  * @generated from enum value: INFRA_OBSERVATION = 4;
207
340
  */
208
341
  INFRA_OBSERVATION = 4,
209
342
  }
210
343
 
211
344
  /**
345
+ * Identifies which kind of test or evaluation produced a Run.
346
+ *
212
347
  * @generated from enum alis.evals.v1.Run.Type
213
348
  */
214
349
  export declare type Run_TypeJson = "TYPE_UNSPECIFIED" | "INTEGRATION_TEST" | "LOAD_TEST" | "AGENT_EVAL" | "INFRA_OBSERVATION";
@@ -219,20 +354,30 @@ export declare type Run_TypeJson = "TYPE_UNSPECIFIED" | "INTEGRATION_TEST" | "LO
219
354
  export declare const Run_TypeSchema: GenEnum<Run_Type, Run_TypeJson>;
220
355
 
221
356
  /**
357
+ * Result detail for an integration test run.
358
+ *
222
359
  * @generated from message alis.evals.v1.IntegrationTestResults
223
360
  */
224
361
  export declare type IntegrationTestResults = Message<"alis.evals.v1.IntegrationTestResults"> & {
225
362
  /**
363
+ * Every case executed in this run, in execution order. Each entry
364
+ * includes its rolled-up status and all checks evaluated for that case.
365
+ *
226
366
  * @generated from field: repeated alis.evals.v1.IntegrationTestResults.Case cases = 1;
227
367
  */
228
368
  cases: IntegrationTestResults_Case[];
229
369
  };
230
370
 
231
371
  /**
372
+ * Result detail for an integration test run.
373
+ *
232
374
  * @generated from message alis.evals.v1.IntegrationTestResults
233
375
  */
234
376
  export declare type IntegrationTestResultsJson = {
235
377
  /**
378
+ * Every case executed in this run, in execution order. Each entry
379
+ * includes its rolled-up status and all checks evaluated for that case.
380
+ *
236
381
  * @generated from field: repeated alis.evals.v1.IntegrationTestResults.Case cases = 1;
237
382
  */
238
383
  cases?: IntegrationTestResults_CaseJson[];
@@ -245,50 +390,72 @@ export declare type IntegrationTestResultsJson = {
245
390
  export declare const IntegrationTestResultsSchema: GenMessage<IntegrationTestResults, {jsonType: IntegrationTestResultsJson}>;
246
391
 
247
392
  /**
393
+ * The result of one integration case.
394
+ *
248
395
  * @generated from message alis.evals.v1.IntegrationTestResults.Case
249
396
  */
250
397
  export declare type IntegrationTestResults_Case = Message<"alis.evals.v1.IntegrationTestResults.Case"> & {
251
398
  /**
399
+ * The case id.
400
+ * Example: example-v1.example-case-1
401
+ *
252
402
  * @generated from field: string id = 1;
253
403
  */
254
404
  id: string;
255
405
 
256
406
  /**
407
+ * The rolled-up status for the case.
408
+ *
257
409
  * @generated from field: alis.evals.v1.Status status = 2;
258
410
  */
259
411
  status: Status;
260
412
 
261
413
  /**
414
+ * All checks evaluated for this case, in execution order.
415
+ *
262
416
  * @generated from field: repeated alis.evals.v1.IntegrationTestResults.Case.Check checks = 3;
263
417
  */
264
418
  checks: IntegrationTestResults_Case_Check[];
265
419
 
266
420
  /**
421
+ * Wall-clock duration for this case.
422
+ *
267
423
  * @generated from field: google.protobuf.Duration duration = 4;
268
424
  */
269
425
  duration?: Duration | undefined;
270
426
  };
271
427
 
272
428
  /**
429
+ * The result of one integration case.
430
+ *
273
431
  * @generated from message alis.evals.v1.IntegrationTestResults.Case
274
432
  */
275
433
  export declare type IntegrationTestResults_CaseJson = {
276
434
  /**
435
+ * The case id.
436
+ * Example: example-v1.example-case-1
437
+ *
277
438
  * @generated from field: string id = 1;
278
439
  */
279
440
  id?: string;
280
441
 
281
442
  /**
443
+ * The rolled-up status for the case.
444
+ *
282
445
  * @generated from field: alis.evals.v1.Status status = 2;
283
446
  */
284
447
  status?: StatusJson;
285
448
 
286
449
  /**
450
+ * All checks evaluated for this case, in execution order.
451
+ *
287
452
  * @generated from field: repeated alis.evals.v1.IntegrationTestResults.Case.Check checks = 3;
288
453
  */
289
454
  checks?: IntegrationTestResults_Case_CheckJson[];
290
455
 
291
456
  /**
457
+ * Wall-clock duration for this case.
458
+ *
292
459
  * @generated from field: google.protobuf.Duration duration = 4;
293
460
  */
294
461
  duration?: DurationJson;
@@ -301,40 +468,62 @@ export declare type IntegrationTestResults_CaseJson = {
301
468
  export declare const IntegrationTestResults_CaseSchema: GenMessage<IntegrationTestResults_Case, {jsonType: IntegrationTestResults_CaseJson}>;
302
469
 
303
470
  /**
471
+ * One deterministic check within a case. Checks are rule-based
472
+ * pass/fail validations (for example response shape, status code, or
473
+ * field equality).
474
+ *
304
475
  * @generated from message alis.evals.v1.IntegrationTestResults.Case.Check
305
476
  */
306
477
  export declare type IntegrationTestResults_Case_Check = Message<"alis.evals.v1.IntegrationTestResults.Case.Check"> & {
307
478
  /**
479
+ * A stable identifier for the check (for example a metric or assertion id).
480
+ *
308
481
  * @generated from field: string id = 1;
309
482
  */
310
483
  id: string;
311
484
 
312
485
  /**
486
+ * The outcome of the check.
487
+ *
313
488
  * @generated from field: alis.evals.v1.Status status = 2;
314
489
  */
315
490
  status: Status;
316
491
 
317
492
  /**
493
+ * Human-readable detail: validation failures, expected versus actual,
494
+ * or error text. Empty when the check passed with no extra context.
495
+ *
318
496
  * @generated from field: string message = 3;
319
497
  */
320
498
  message: string;
321
499
  };
322
500
 
323
501
  /**
502
+ * One deterministic check within a case. Checks are rule-based
503
+ * pass/fail validations (for example response shape, status code, or
504
+ * field equality).
505
+ *
324
506
  * @generated from message alis.evals.v1.IntegrationTestResults.Case.Check
325
507
  */
326
508
  export declare type IntegrationTestResults_Case_CheckJson = {
327
509
  /**
510
+ * A stable identifier for the check (for example a metric or assertion id).
511
+ *
328
512
  * @generated from field: string id = 1;
329
513
  */
330
514
  id?: string;
331
515
 
332
516
  /**
517
+ * The outcome of the check.
518
+ *
333
519
  * @generated from field: alis.evals.v1.Status status = 2;
334
520
  */
335
521
  status?: StatusJson;
336
522
 
337
523
  /**
524
+ * Human-readable detail: validation failures, expected versus actual,
525
+ * or error text. Empty when the check passed with no extra context.
526
+ *
338
527
  * @generated from field: string message = 3;
339
528
  */
340
529
  message?: string;
@@ -347,70 +536,120 @@ export declare type IntegrationTestResults_Case_CheckJson = {
347
536
  export declare const IntegrationTestResults_Case_CheckSchema: GenMessage<IntegrationTestResults_Case_Check, {jsonType: IntegrationTestResults_Case_CheckJson}>;
348
537
 
349
538
  /**
539
+ * Request latency distribution, in milliseconds.
540
+ *
541
+ * Percentile fields give the latency at or below which that fraction of
542
+ * requests completed. For example, p95_ms is the latency experienced by
543
+ * the fastest 95% of requests; the remaining 5% were slower. Percentiles
544
+ * surface tail latency that a mean would hide and are the usual basis for
545
+ * SLO checks.
546
+ *
547
+ * Package-level so load summaries, Cloud Run metrics, and Spanner metrics
548
+ * share an identical BQ column shape.
549
+ *
350
550
  * @generated from message alis.evals.v1.LatencyPercentiles
351
551
  */
352
552
  export declare type LatencyPercentiles = Message<"alis.evals.v1.LatencyPercentiles"> & {
353
553
  /**
554
+ * Median latency. Half of requests completed at or below this value.
555
+ *
354
556
  * @generated from field: double p50_ms = 1;
355
557
  */
356
558
  p50Ms: number;
357
559
 
358
560
  /**
561
+ * 95th percentile latency. 95% of requests completed at or below this
562
+ * value; the slowest 5% exceeded it.
563
+ *
359
564
  * @generated from field: double p95_ms = 2;
360
565
  */
361
566
  p95Ms: number;
362
567
 
363
568
  /**
569
+ * 99th percentile latency. 99% of requests completed at or below this
570
+ * value; the slowest 1% exceeded it.
571
+ *
364
572
  * @generated from field: double p99_ms = 3;
365
573
  */
366
574
  p99Ms: number;
367
575
 
368
576
  /**
577
+ * The fastest observed request latency.
578
+ *
369
579
  * @generated from field: double min_ms = 4;
370
580
  */
371
581
  minMs: number;
372
582
 
373
583
  /**
584
+ * The arithmetic mean of observed request latencies.
585
+ *
374
586
  * @generated from field: double mean_ms = 5;
375
587
  */
376
588
  meanMs: number;
377
589
 
378
590
  /**
591
+ * The slowest observed request latency.
592
+ *
379
593
  * @generated from field: double max_ms = 6;
380
594
  */
381
595
  maxMs: number;
382
596
  };
383
597
 
384
598
  /**
599
+ * Request latency distribution, in milliseconds.
600
+ *
601
+ * Percentile fields give the latency at or below which that fraction of
602
+ * requests completed. For example, p95_ms is the latency experienced by
603
+ * the fastest 95% of requests; the remaining 5% were slower. Percentiles
604
+ * surface tail latency that a mean would hide and are the usual basis for
605
+ * SLO checks.
606
+ *
607
+ * Package-level so load summaries, Cloud Run metrics, and Spanner metrics
608
+ * share an identical BQ column shape.
609
+ *
385
610
  * @generated from message alis.evals.v1.LatencyPercentiles
386
611
  */
387
612
  export declare type LatencyPercentilesJson = {
388
613
  /**
614
+ * Median latency. Half of requests completed at or below this value.
615
+ *
389
616
  * @generated from field: double p50_ms = 1;
390
617
  */
391
618
  p50Ms?: number | "NaN" | "Infinity" | "-Infinity";
392
619
 
393
620
  /**
621
+ * 95th percentile latency. 95% of requests completed at or below this
622
+ * value; the slowest 5% exceeded it.
623
+ *
394
624
  * @generated from field: double p95_ms = 2;
395
625
  */
396
626
  p95Ms?: number | "NaN" | "Infinity" | "-Infinity";
397
627
 
398
628
  /**
629
+ * 99th percentile latency. 99% of requests completed at or below this
630
+ * value; the slowest 1% exceeded it.
631
+ *
399
632
  * @generated from field: double p99_ms = 3;
400
633
  */
401
634
  p99Ms?: number | "NaN" | "Infinity" | "-Infinity";
402
635
 
403
636
  /**
637
+ * The fastest observed request latency.
638
+ *
404
639
  * @generated from field: double min_ms = 4;
405
640
  */
406
641
  minMs?: number | "NaN" | "Infinity" | "-Infinity";
407
642
 
408
643
  /**
644
+ * The arithmetic mean of observed request latencies.
645
+ *
409
646
  * @generated from field: double mean_ms = 5;
410
647
  */
411
648
  meanMs?: number | "NaN" | "Infinity" | "-Infinity";
412
649
 
413
650
  /**
651
+ * The slowest observed request latency.
652
+ *
414
653
  * @generated from field: double max_ms = 6;
415
654
  */
416
655
  maxMs?: number | "NaN" | "Infinity" | "-Infinity";
@@ -423,20 +662,40 @@ export declare type LatencyPercentilesJson = {
423
662
  export declare const LatencyPercentilesSchema: GenMessage<LatencyPercentiles, {jsonType: LatencyPercentilesJson}>;
424
663
 
425
664
  /**
665
+ * Result detail for a load test run. The runner generates sustained traffic
666
+ * against each selected load case, records per-case performance in the case's
667
+ * summary, and evaluates the configured Service Level Objectives (SLOs). An
668
+ * SLO defines an acceptable bound on a metric; every evaluated check is listed
669
+ * with its observed value, limit, and outcome, passed and failed alike.
670
+ *
426
671
  * @generated from message alis.evals.v1.LoadTestResults
427
672
  */
428
673
  export declare type LoadTestResults = Message<"alis.evals.v1.LoadTestResults"> & {
429
674
  /**
675
+ * Every load case executed in this run, in execution order. Each entry
676
+ * includes its rolled-up status, the performance summary for its load
677
+ * window, and all SLO checks evaluated against that summary.
678
+ *
430
679
  * @generated from field: repeated alis.evals.v1.LoadTestResults.Case cases = 1;
431
680
  */
432
681
  cases: LoadTestResults_Case[];
433
682
  };
434
683
 
435
684
  /**
685
+ * Result detail for a load test run. The runner generates sustained traffic
686
+ * against each selected load case, records per-case performance in the case's
687
+ * summary, and evaluates the configured Service Level Objectives (SLOs). An
688
+ * SLO defines an acceptable bound on a metric; every evaluated check is listed
689
+ * with its observed value, limit, and outcome, passed and failed alike.
690
+ *
436
691
  * @generated from message alis.evals.v1.LoadTestResults
437
692
  */
438
693
  export declare type LoadTestResultsJson = {
439
694
  /**
695
+ * Every load case executed in this run, in execution order. Each entry
696
+ * includes its rolled-up status, the performance summary for its load
697
+ * window, and all SLO checks evaluated against that summary.
698
+ *
440
699
  * @generated from field: repeated alis.evals.v1.LoadTestResults.Case cases = 1;
441
700
  */
442
701
  cases?: LoadTestResults_CaseJson[];
@@ -449,30 +708,42 @@ export declare type LoadTestResultsJson = {
449
708
  export declare const LoadTestResultsSchema: GenMessage<LoadTestResults, {jsonType: LoadTestResultsJson}>;
450
709
 
451
710
  /**
711
+ * Key/value entry for string-valued associative fields on the wire.
712
+ *
452
713
  * @generated from message alis.evals.v1.LoadTestResults.StringEntry
453
714
  */
454
715
  export declare type LoadTestResults_StringEntry = Message<"alis.evals.v1.LoadTestResults.StringEntry"> & {
455
716
  /**
717
+ * The entry key.
718
+ *
456
719
  * @generated from field: string key = 1;
457
720
  */
458
721
  key: string;
459
722
 
460
723
  /**
724
+ * The string value associated with `key`.
725
+ *
461
726
  * @generated from field: string value = 2;
462
727
  */
463
728
  value: string;
464
729
  };
465
730
 
466
731
  /**
732
+ * Key/value entry for string-valued associative fields on the wire.
733
+ *
467
734
  * @generated from message alis.evals.v1.LoadTestResults.StringEntry
468
735
  */
469
736
  export declare type LoadTestResults_StringEntryJson = {
470
737
  /**
738
+ * The entry key.
739
+ *
471
740
  * @generated from field: string key = 1;
472
741
  */
473
742
  key?: string;
474
743
 
475
744
  /**
745
+ * The string value associated with `key`.
746
+ *
476
747
  * @generated from field: string value = 2;
477
748
  */
478
749
  value?: string;
@@ -485,30 +756,42 @@ export declare type LoadTestResults_StringEntryJson = {
485
756
  export declare const LoadTestResults_StringEntrySchema: GenMessage<LoadTestResults_StringEntry, {jsonType: LoadTestResults_StringEntryJson}>;
486
757
 
487
758
  /**
759
+ * Key/value entry for int64-valued associative fields on the wire.
760
+ *
488
761
  * @generated from message alis.evals.v1.LoadTestResults.Int64Entry
489
762
  */
490
763
  export declare type LoadTestResults_Int64Entry = Message<"alis.evals.v1.LoadTestResults.Int64Entry"> & {
491
764
  /**
765
+ * The entry key.
766
+ *
492
767
  * @generated from field: string key = 1;
493
768
  */
494
769
  key: string;
495
770
 
496
771
  /**
772
+ * The int64 value associated with `key`.
773
+ *
497
774
  * @generated from field: int64 value = 2;
498
775
  */
499
776
  value: bigint;
500
777
  };
501
778
 
502
779
  /**
780
+ * Key/value entry for int64-valued associative fields on the wire.
781
+ *
503
782
  * @generated from message alis.evals.v1.LoadTestResults.Int64Entry
504
783
  */
505
784
  export declare type LoadTestResults_Int64EntryJson = {
506
785
  /**
786
+ * The entry key.
787
+ *
507
788
  * @generated from field: string key = 1;
508
789
  */
509
790
  key?: string;
510
791
 
511
792
  /**
793
+ * The int64 value associated with `key`.
794
+ *
512
795
  * @generated from field: int64 value = 2;
513
796
  */
514
797
  value?: string;
@@ -521,100 +804,158 @@ export declare type LoadTestResults_Int64EntryJson = {
521
804
  export declare const LoadTestResults_Int64EntrySchema: GenMessage<LoadTestResults_Int64Entry, {jsonType: LoadTestResults_Int64EntryJson}>;
522
805
 
523
806
  /**
807
+ * The result of one load case.
808
+ *
524
809
  * @generated from message alis.evals.v1.LoadTestResults.Case
525
810
  */
526
811
  export declare type LoadTestResults_Case = Message<"alis.evals.v1.LoadTestResults.Case"> & {
527
812
  /**
813
+ * The case id.
814
+ * Example: files-v2-load.list-files
815
+ *
528
816
  * @generated from field: string id = 1;
529
817
  */
530
818
  id: string;
531
819
 
532
820
  /**
821
+ * The rolled-up status for the case. FAILED when any SLO check failed or
822
+ * load generation could not complete.
823
+ *
533
824
  * @generated from field: alis.evals.v1.Status status = 2;
534
825
  */
535
826
  status: Status;
536
827
 
537
828
  /**
829
+ * Performance metrics observed during this case's load window. Always
830
+ * populated for executed cases.
831
+ *
538
832
  * @generated from field: alis.evals.v1.LoadTestResults.Summary summary = 3;
539
833
  */
540
834
  summary?: LoadTestResults_Summary | undefined;
541
835
 
542
836
  /**
837
+ * All SLO checks evaluated for this case, in declaration order.
838
+ *
543
839
  * @generated from field: repeated alis.evals.v1.LoadTestResults.SloCheck checks = 4;
544
840
  */
545
841
  checks: LoadTestResults_SloCheck[];
546
842
 
547
843
  /**
844
+ * Author-declared labels for filtering and grouping in reports (for example
845
+ * model name, RPC method, or dataset slice). Empty when the case declares
846
+ * no tags.
847
+ *
548
848
  * @generated from field: repeated alis.evals.v1.LoadTestResults.StringEntry tags = 5;
549
849
  */
550
850
  tags: LoadTestResults_StringEntry[];
551
851
 
552
852
  /**
853
+ * Server-side Cloud Run metrics for declared targets over this case's
854
+ * observation window. Empty when the suite declares no Cloud Run targets.
855
+ *
553
856
  * @generated from field: repeated alis.evals.v1.CloudRunTargetSnapshot cloud_run = 6;
554
857
  */
555
858
  cloudRun: CloudRunTargetSnapshot[];
556
859
 
557
860
  /**
861
+ * Server-side Spanner metrics for declared targets over this case's
862
+ * observation window. Empty when the suite declares no Spanner targets.
863
+ *
558
864
  * @generated from field: repeated alis.evals.v1.SpannerTargetSnapshot spanner = 7;
559
865
  */
560
866
  spanner: SpannerTargetSnapshot[];
561
867
 
562
868
  /**
869
+ * Infrastructure SLO check outcomes. Diagnostics-only in v1: always empty
870
+ * on the wire until infra SLO evaluators ship.
871
+ *
563
872
  * @generated from field: repeated alis.evals.v1.InfraSloCheck infra_checks = 8;
564
873
  */
565
874
  infraChecks: InfraSloCheck[];
566
875
 
567
876
  /**
877
+ * Developer and framework validation outcomes for this case in declaration
878
+ * order. Integration cases continue using checks instead.
879
+ *
568
880
  * @generated from field: repeated alis.evals.v1.Validation validations = 9;
569
881
  */
570
882
  validations: Validation[];
571
883
  };
572
884
 
573
885
  /**
886
+ * The result of one load case.
887
+ *
574
888
  * @generated from message alis.evals.v1.LoadTestResults.Case
575
889
  */
576
890
  export declare type LoadTestResults_CaseJson = {
577
891
  /**
892
+ * The case id.
893
+ * Example: files-v2-load.list-files
894
+ *
578
895
  * @generated from field: string id = 1;
579
896
  */
580
897
  id?: string;
581
898
 
582
899
  /**
900
+ * The rolled-up status for the case. FAILED when any SLO check failed or
901
+ * load generation could not complete.
902
+ *
583
903
  * @generated from field: alis.evals.v1.Status status = 2;
584
904
  */
585
905
  status?: StatusJson;
586
906
 
587
907
  /**
908
+ * Performance metrics observed during this case's load window. Always
909
+ * populated for executed cases.
910
+ *
588
911
  * @generated from field: alis.evals.v1.LoadTestResults.Summary summary = 3;
589
912
  */
590
913
  summary?: LoadTestResults_SummaryJson;
591
914
 
592
915
  /**
916
+ * All SLO checks evaluated for this case, in declaration order.
917
+ *
593
918
  * @generated from field: repeated alis.evals.v1.LoadTestResults.SloCheck checks = 4;
594
919
  */
595
920
  checks?: LoadTestResults_SloCheckJson[];
596
921
 
597
922
  /**
923
+ * Author-declared labels for filtering and grouping in reports (for example
924
+ * model name, RPC method, or dataset slice). Empty when the case declares
925
+ * no tags.
926
+ *
598
927
  * @generated from field: repeated alis.evals.v1.LoadTestResults.StringEntry tags = 5;
599
928
  */
600
929
  tags?: LoadTestResults_StringEntryJson[];
601
930
 
602
931
  /**
932
+ * Server-side Cloud Run metrics for declared targets over this case's
933
+ * observation window. Empty when the suite declares no Cloud Run targets.
934
+ *
603
935
  * @generated from field: repeated alis.evals.v1.CloudRunTargetSnapshot cloud_run = 6;
604
936
  */
605
937
  cloudRun?: CloudRunTargetSnapshotJson[];
606
938
 
607
939
  /**
940
+ * Server-side Spanner metrics for declared targets over this case's
941
+ * observation window. Empty when the suite declares no Spanner targets.
942
+ *
608
943
  * @generated from field: repeated alis.evals.v1.SpannerTargetSnapshot spanner = 7;
609
944
  */
610
945
  spanner?: SpannerTargetSnapshotJson[];
611
946
 
612
947
  /**
948
+ * Infrastructure SLO check outcomes. Diagnostics-only in v1: always empty
949
+ * on the wire until infra SLO evaluators ship.
950
+ *
613
951
  * @generated from field: repeated alis.evals.v1.InfraSloCheck infra_checks = 8;
614
952
  */
615
953
  infraChecks?: InfraSloCheckJson[];
616
954
 
617
955
  /**
956
+ * Developer and framework validation outcomes for this case in declaration
957
+ * order. Integration cases continue using checks instead.
958
+ *
618
959
  * @generated from field: repeated alis.evals.v1.Validation validations = 9;
619
960
  */
620
961
  validations?: ValidationJson[];
@@ -627,160 +968,250 @@ export declare type LoadTestResults_CaseJson = {
627
968
  export declare const LoadTestResults_CaseSchema: GenMessage<LoadTestResults_Case, {jsonType: LoadTestResults_CaseJson}>;
628
969
 
629
970
  /**
971
+ * Performance summary for one load case's load window: the resolved load
972
+ * settings and the observed request counts, errors, throughput, and latency.
973
+ *
630
974
  * @generated from message alis.evals.v1.LoadTestResults.Summary
631
975
  */
632
976
  export declare type LoadTestResults_Summary = Message<"alis.evals.v1.LoadTestResults.Summary"> & {
633
977
  /**
978
+ * The load intensity mode used for this run.
979
+ *
634
980
  * @generated from field: alis.evals.v1.RunLoadTestRequest.Mode mode = 1;
635
981
  */
636
982
  mode: RunLoadTestRequest_Mode;
637
983
 
638
984
  /**
985
+ * The request rate the generator was configured to sustain, in queries
986
+ * (requests) per second. Resolved from the mode preset and any suite-level
987
+ * overrides; recorded so observed throughput can be interpreted and runs
988
+ * compared over time.
989
+ *
639
990
  * @generated from field: double target_qps = 2;
640
991
  */
641
992
  targetQps: number;
642
993
 
643
994
  /**
995
+ * The number of concurrent workers used to generate load.
996
+ *
644
997
  * @generated from field: int32 concurrency = 3;
645
998
  */
646
999
  concurrency: number;
647
1000
 
648
1001
  /**
1002
+ * Total wall-clock duration of the case's load window, excluding warmup.
1003
+ *
649
1004
  * @generated from field: google.protobuf.Duration duration = 4;
650
1005
  */
651
1006
  duration?: Duration | undefined;
652
1007
 
653
1008
  /**
1009
+ * Total requests sent during the load window.
1010
+ *
654
1011
  * @generated from field: int64 request_count = 5;
655
1012
  */
656
1013
  requestCount: bigint;
657
1014
 
658
1015
  /**
1016
+ * Total request errors observed during the load window (transport
1017
+ * failures, non-OK statuses, timeouts). Distinct from SLO checks, which
1018
+ * are threshold comparisons evaluated after metrics are aggregated.
1019
+ *
659
1020
  * @generated from field: int64 error_count = 6;
660
1021
  */
661
1022
  errorCount: bigint;
662
1023
 
663
1024
  /**
1025
+ * Observed steady-state throughput in queries (requests) per second.
1026
+ *
664
1027
  * @generated from field: double actual_qps = 7;
665
1028
  */
666
1029
  actualQps: number;
667
1030
 
668
1031
  /**
1032
+ * Request latency distribution for the load window.
1033
+ *
669
1034
  * @generated from field: alis.evals.v1.LatencyPercentiles latency = 8;
670
1035
  */
671
1036
  latency?: LatencyPercentiles | undefined;
672
1037
 
673
1038
  /**
1039
+ * Request errors grouped by canonical gRPC status code name (for example
1040
+ * `UNAVAILABLE` or `DEADLINE_EXCEEDED`). Values sum to error_count.
1041
+ *
674
1042
  * @generated from field: repeated alis.evals.v1.LoadTestResults.Int64Entry errors_by_code = 9;
675
1043
  */
676
1044
  errorsByCode: LoadTestResults_Int64Entry[];
677
1045
 
678
1046
  /**
1047
+ * Scheduled request ticks that were not dispatched because all workers
1048
+ * were busy. Distinct from error_count; surfaces generator saturation.
1049
+ *
679
1050
  * @generated from field: int64 dropped_count = 10;
680
1051
  */
681
1052
  droppedCount: bigint;
682
1053
 
683
1054
  /**
1055
+ * Semantic assertion passes during the load window. Distinct from
1056
+ * transport error_count — a request may succeed at the wire layer yet
1057
+ * fail a check (for example an unexpected eval score).
1058
+ *
684
1059
  * @generated from field: int64 check_passed_count = 11;
685
1060
  */
686
1061
  checkPassedCount: bigint;
687
1062
 
688
1063
  /**
1064
+ * Semantic assertion failures during the load window.
1065
+ *
689
1066
  * @generated from field: int64 check_failed_count = 12;
690
1067
  */
691
1068
  checkFailedCount: bigint;
692
1069
 
693
1070
  /**
1071
+ * Streaming RPC metrics when the case exercised client or server streaming.
1072
+ * Omitted for unary-only cases.
1073
+ *
694
1074
  * @generated from field: optional alis.evals.v1.LoadTestResults.StreamSummary stream = 13;
695
1075
  */
696
1076
  stream?: LoadTestResults_StreamSummary | undefined;
697
1077
 
698
1078
  /**
1079
+ * Resolved QPS stage configuration. Empty when the profile used a
1080
+ * constant target_qps for the entire window.
1081
+ *
699
1082
  * @generated from field: repeated alis.evals.v1.LoadTestResults.LoadStage qps_stages = 14;
700
1083
  */
701
1084
  qpsStages: LoadTestResults_LoadStage[];
702
1085
 
703
1086
  /**
1087
+ * Resolved concurrency stage configuration. Empty when the profile used
1088
+ * a constant concurrency for the entire window.
1089
+ *
704
1090
  * @generated from field: repeated alis.evals.v1.LoadTestResults.LoadStage concurrency_stages = 15;
705
1091
  */
706
1092
  concurrencyStages: LoadTestResults_LoadStage[];
707
1093
  };
708
1094
 
709
1095
  /**
1096
+ * Performance summary for one load case's load window: the resolved load
1097
+ * settings and the observed request counts, errors, throughput, and latency.
1098
+ *
710
1099
  * @generated from message alis.evals.v1.LoadTestResults.Summary
711
1100
  */
712
1101
  export declare type LoadTestResults_SummaryJson = {
713
1102
  /**
1103
+ * The load intensity mode used for this run.
1104
+ *
714
1105
  * @generated from field: alis.evals.v1.RunLoadTestRequest.Mode mode = 1;
715
1106
  */
716
1107
  mode?: RunLoadTestRequest_ModeJson;
717
1108
 
718
1109
  /**
1110
+ * The request rate the generator was configured to sustain, in queries
1111
+ * (requests) per second. Resolved from the mode preset and any suite-level
1112
+ * overrides; recorded so observed throughput can be interpreted and runs
1113
+ * compared over time.
1114
+ *
719
1115
  * @generated from field: double target_qps = 2;
720
1116
  */
721
1117
  targetQps?: number | "NaN" | "Infinity" | "-Infinity";
722
1118
 
723
1119
  /**
1120
+ * The number of concurrent workers used to generate load.
1121
+ *
724
1122
  * @generated from field: int32 concurrency = 3;
725
1123
  */
726
1124
  concurrency?: number;
727
1125
 
728
1126
  /**
1127
+ * Total wall-clock duration of the case's load window, excluding warmup.
1128
+ *
729
1129
  * @generated from field: google.protobuf.Duration duration = 4;
730
1130
  */
731
1131
  duration?: DurationJson;
732
1132
 
733
1133
  /**
1134
+ * Total requests sent during the load window.
1135
+ *
734
1136
  * @generated from field: int64 request_count = 5;
735
1137
  */
736
1138
  requestCount?: string;
737
1139
 
738
1140
  /**
1141
+ * Total request errors observed during the load window (transport
1142
+ * failures, non-OK statuses, timeouts). Distinct from SLO checks, which
1143
+ * are threshold comparisons evaluated after metrics are aggregated.
1144
+ *
739
1145
  * @generated from field: int64 error_count = 6;
740
1146
  */
741
1147
  errorCount?: string;
742
1148
 
743
1149
  /**
1150
+ * Observed steady-state throughput in queries (requests) per second.
1151
+ *
744
1152
  * @generated from field: double actual_qps = 7;
745
1153
  */
746
1154
  actualQps?: number | "NaN" | "Infinity" | "-Infinity";
747
1155
 
748
1156
  /**
1157
+ * Request latency distribution for the load window.
1158
+ *
749
1159
  * @generated from field: alis.evals.v1.LatencyPercentiles latency = 8;
750
1160
  */
751
1161
  latency?: LatencyPercentilesJson;
752
1162
 
753
1163
  /**
1164
+ * Request errors grouped by canonical gRPC status code name (for example
1165
+ * `UNAVAILABLE` or `DEADLINE_EXCEEDED`). Values sum to error_count.
1166
+ *
754
1167
  * @generated from field: repeated alis.evals.v1.LoadTestResults.Int64Entry errors_by_code = 9;
755
1168
  */
756
1169
  errorsByCode?: LoadTestResults_Int64EntryJson[];
757
1170
 
758
1171
  /**
1172
+ * Scheduled request ticks that were not dispatched because all workers
1173
+ * were busy. Distinct from error_count; surfaces generator saturation.
1174
+ *
759
1175
  * @generated from field: int64 dropped_count = 10;
760
1176
  */
761
1177
  droppedCount?: string;
762
1178
 
763
1179
  /**
1180
+ * Semantic assertion passes during the load window. Distinct from
1181
+ * transport error_count — a request may succeed at the wire layer yet
1182
+ * fail a check (for example an unexpected eval score).
1183
+ *
764
1184
  * @generated from field: int64 check_passed_count = 11;
765
1185
  */
766
1186
  checkPassedCount?: string;
767
1187
 
768
1188
  /**
1189
+ * Semantic assertion failures during the load window.
1190
+ *
769
1191
  * @generated from field: int64 check_failed_count = 12;
770
1192
  */
771
1193
  checkFailedCount?: string;
772
1194
 
773
1195
  /**
1196
+ * Streaming RPC metrics when the case exercised client or server streaming.
1197
+ * Omitted for unary-only cases.
1198
+ *
774
1199
  * @generated from field: optional alis.evals.v1.LoadTestResults.StreamSummary stream = 13;
775
1200
  */
776
1201
  stream?: LoadTestResults_StreamSummaryJson;
777
1202
 
778
1203
  /**
1204
+ * Resolved QPS stage configuration. Empty when the profile used a
1205
+ * constant target_qps for the entire window.
1206
+ *
779
1207
  * @generated from field: repeated alis.evals.v1.LoadTestResults.LoadStage qps_stages = 14;
780
1208
  */
781
1209
  qpsStages?: LoadTestResults_LoadStageJson[];
782
1210
 
783
1211
  /**
1212
+ * Resolved concurrency stage configuration. Empty when the profile used
1213
+ * a constant concurrency for the entire window.
1214
+ *
784
1215
  * @generated from field: repeated alis.evals.v1.LoadTestResults.LoadStage concurrency_stages = 15;
785
1216
  */
786
1217
  concurrencyStages?: LoadTestResults_LoadStageJson[];
@@ -793,30 +1224,44 @@ export declare type LoadTestResults_SummaryJson = {
793
1224
  export declare const LoadTestResults_SummarySchema: GenMessage<LoadTestResults_Summary, {jsonType: LoadTestResults_SummaryJson}>;
794
1225
 
795
1226
  /**
1227
+ * One step in a staged load profile (QPS or concurrency depending on context).
1228
+ *
796
1229
  * @generated from message alis.evals.v1.LoadTestResults.LoadStage
797
1230
  */
798
1231
  export declare type LoadTestResults_LoadStage = Message<"alis.evals.v1.LoadTestResults.LoadStage"> & {
799
1232
  /**
1233
+ * Wall-clock duration of this stage.
1234
+ *
800
1235
  * @generated from field: google.protobuf.Duration duration = 1;
801
1236
  */
802
1237
  duration?: Duration | undefined;
803
1238
 
804
1239
  /**
1240
+ * Target value for this stage: queries per second when nested under
1241
+ * qps_stages, or worker count when nested under concurrency_stages.
1242
+ *
805
1243
  * @generated from field: double target = 2;
806
1244
  */
807
1245
  target: number;
808
1246
  };
809
1247
 
810
1248
  /**
1249
+ * One step in a staged load profile (QPS or concurrency depending on context).
1250
+ *
811
1251
  * @generated from message alis.evals.v1.LoadTestResults.LoadStage
812
1252
  */
813
1253
  export declare type LoadTestResults_LoadStageJson = {
814
1254
  /**
1255
+ * Wall-clock duration of this stage.
1256
+ *
815
1257
  * @generated from field: google.protobuf.Duration duration = 1;
816
1258
  */
817
1259
  duration?: DurationJson;
818
1260
 
819
1261
  /**
1262
+ * Target value for this stage: queries per second when nested under
1263
+ * qps_stages, or worker count when nested under concurrency_stages.
1264
+ *
820
1265
  * @generated from field: double target = 2;
821
1266
  */
822
1267
  target?: number | "NaN" | "Infinity" | "-Infinity";
@@ -829,60 +1274,88 @@ export declare type LoadTestResults_LoadStageJson = {
829
1274
  export declare const LoadTestResults_LoadStageSchema: GenMessage<LoadTestResults_LoadStage, {jsonType: LoadTestResults_LoadStageJson}>;
830
1275
 
831
1276
  /**
1277
+ * Aggregate streaming metrics for a load case that exercised streaming RPCs.
1278
+ *
832
1279
  * @generated from message alis.evals.v1.LoadTestResults.StreamSummary
833
1280
  */
834
1281
  export declare type LoadTestResults_StreamSummary = Message<"alis.evals.v1.LoadTestResults.StreamSummary"> & {
835
1282
  /**
1283
+ * Number of streaming invocations measured during the load window.
1284
+ *
836
1285
  * @generated from field: int64 stream_count = 1;
837
1286
  */
838
1287
  streamCount: bigint;
839
1288
 
840
1289
  /**
1290
+ * Total messages sent across all measured streams.
1291
+ *
841
1292
  * @generated from field: int64 messages_sent_total = 2;
842
1293
  */
843
1294
  messagesSentTotal: bigint;
844
1295
 
845
1296
  /**
1297
+ * Time-to-first-byte / send-phase latency distribution in milliseconds.
1298
+ * For client streaming, this spans stream open through the last successful
1299
+ * Send (or through the first send error).
1300
+ *
846
1301
  * @generated from field: alis.evals.v1.LatencyPercentiles ttfb = 3;
847
1302
  */
848
1303
  ttfb?: LatencyPercentiles | undefined;
849
1304
 
850
1305
  /**
1306
+ * CloseAndRecv / response-phase latency distribution in milliseconds.
1307
+ *
851
1308
  * @generated from field: alis.evals.v1.LatencyPercentiles response_latency = 4;
852
1309
  */
853
1310
  responseLatency?: LatencyPercentiles | undefined;
854
1311
 
855
1312
  /**
1313
+ * End-to-end stream call duration distribution in milliseconds.
1314
+ *
856
1315
  * @generated from field: alis.evals.v1.LatencyPercentiles total_duration = 5;
857
1316
  */
858
1317
  totalDuration?: LatencyPercentiles | undefined;
859
1318
  };
860
1319
 
861
1320
  /**
1321
+ * Aggregate streaming metrics for a load case that exercised streaming RPCs.
1322
+ *
862
1323
  * @generated from message alis.evals.v1.LoadTestResults.StreamSummary
863
1324
  */
864
1325
  export declare type LoadTestResults_StreamSummaryJson = {
865
1326
  /**
1327
+ * Number of streaming invocations measured during the load window.
1328
+ *
866
1329
  * @generated from field: int64 stream_count = 1;
867
1330
  */
868
1331
  streamCount?: string;
869
1332
 
870
1333
  /**
1334
+ * Total messages sent across all measured streams.
1335
+ *
871
1336
  * @generated from field: int64 messages_sent_total = 2;
872
1337
  */
873
1338
  messagesSentTotal?: string;
874
1339
 
875
1340
  /**
1341
+ * Time-to-first-byte / send-phase latency distribution in milliseconds.
1342
+ * For client streaming, this spans stream open through the last successful
1343
+ * Send (or through the first send error).
1344
+ *
876
1345
  * @generated from field: alis.evals.v1.LatencyPercentiles ttfb = 3;
877
1346
  */
878
1347
  ttfb?: LatencyPercentilesJson;
879
1348
 
880
1349
  /**
1350
+ * CloseAndRecv / response-phase latency distribution in milliseconds.
1351
+ *
881
1352
  * @generated from field: alis.evals.v1.LatencyPercentiles response_latency = 4;
882
1353
  */
883
1354
  responseLatency?: LatencyPercentilesJson;
884
1355
 
885
1356
  /**
1357
+ * End-to-end stream call duration distribution in milliseconds.
1358
+ *
886
1359
  * @generated from field: alis.evals.v1.LatencyPercentiles total_duration = 5;
887
1360
  */
888
1361
  totalDuration?: LatencyPercentilesJson;
@@ -895,70 +1368,118 @@ export declare type LoadTestResults_StreamSummaryJson = {
895
1368
  export declare const LoadTestResults_StreamSummarySchema: GenMessage<LoadTestResults_StreamSummary, {jsonType: LoadTestResults_StreamSummaryJson}>;
896
1369
 
897
1370
  /**
1371
+ * One Service Level Objective (SLO) check outcome.
1372
+ *
1373
+ * An SLO expresses an acceptable bound on a load-test metric (for example
1374
+ * "p99 latency must stay below 500 ms" or "error rate must stay below 1%").
1375
+ * Every configured check is recorded with the measured value, the configured
1376
+ * limit, and whether the bound held, so consumers can see headroom on passed
1377
+ * checks as well as breaches.
1378
+ *
898
1379
  * @generated from message alis.evals.v1.LoadTestResults.SloCheck
899
1380
  */
900
1381
  export declare type LoadTestResults_SloCheck = Message<"alis.evals.v1.LoadTestResults.SloCheck"> & {
901
1382
  /**
1383
+ * A stable identifier for the SLO check or metric (for example
1384
+ * `latency.p99_ms` or `error_rate`).
1385
+ *
902
1386
  * @generated from field: string id = 1;
903
1387
  */
904
1388
  id: string;
905
1389
 
906
1390
  /**
1391
+ * The outcome of the check. FAILED when the observed value crossed the
1392
+ * limit or the metric could not be measured.
1393
+ *
907
1394
  * @generated from field: alis.evals.v1.Status status = 2;
908
1395
  */
909
1396
  status: Status;
910
1397
 
911
1398
  /**
1399
+ * Human-readable detail (for example expected versus observed values or
1400
+ * measurement errors). Empty when the check passed with no extra context.
1401
+ *
912
1402
  * @generated from field: string message = 3;
913
1403
  */
914
1404
  message: string;
915
1405
 
916
1406
  /**
1407
+ * The metric value measured during the run (for example 612.4 when unit
1408
+ * is `ms`).
1409
+ *
917
1410
  * @generated from field: double observed = 4;
918
1411
  */
919
1412
  observed: number;
920
1413
 
921
1414
  /**
1415
+ * The configured SLO bound (for example 500.0 when unit is `ms`).
1416
+ *
922
1417
  * @generated from field: double limit = 5;
923
1418
  */
924
1419
  limit: number;
925
1420
 
926
1421
  /**
1422
+ * The unit shared by observed and limit (for example `ms`, `rps`, or `%`).
1423
+ *
927
1424
  * @generated from field: string unit = 6;
928
1425
  */
929
1426
  unit: string;
930
1427
  };
931
1428
 
932
1429
  /**
1430
+ * One Service Level Objective (SLO) check outcome.
1431
+ *
1432
+ * An SLO expresses an acceptable bound on a load-test metric (for example
1433
+ * "p99 latency must stay below 500 ms" or "error rate must stay below 1%").
1434
+ * Every configured check is recorded with the measured value, the configured
1435
+ * limit, and whether the bound held, so consumers can see headroom on passed
1436
+ * checks as well as breaches.
1437
+ *
933
1438
  * @generated from message alis.evals.v1.LoadTestResults.SloCheck
934
1439
  */
935
1440
  export declare type LoadTestResults_SloCheckJson = {
936
1441
  /**
1442
+ * A stable identifier for the SLO check or metric (for example
1443
+ * `latency.p99_ms` or `error_rate`).
1444
+ *
937
1445
  * @generated from field: string id = 1;
938
1446
  */
939
1447
  id?: string;
940
1448
 
941
1449
  /**
1450
+ * The outcome of the check. FAILED when the observed value crossed the
1451
+ * limit or the metric could not be measured.
1452
+ *
942
1453
  * @generated from field: alis.evals.v1.Status status = 2;
943
1454
  */
944
1455
  status?: StatusJson;
945
1456
 
946
1457
  /**
1458
+ * Human-readable detail (for example expected versus observed values or
1459
+ * measurement errors). Empty when the check passed with no extra context.
1460
+ *
947
1461
  * @generated from field: string message = 3;
948
1462
  */
949
1463
  message?: string;
950
1464
 
951
1465
  /**
1466
+ * The metric value measured during the run (for example 612.4 when unit
1467
+ * is `ms`).
1468
+ *
952
1469
  * @generated from field: double observed = 4;
953
1470
  */
954
1471
  observed?: number | "NaN" | "Infinity" | "-Infinity";
955
1472
 
956
1473
  /**
1474
+ * The configured SLO bound (for example 500.0 when unit is `ms`).
1475
+ *
957
1476
  * @generated from field: double limit = 5;
958
1477
  */
959
1478
  limit?: number | "NaN" | "Infinity" | "-Infinity";
960
1479
 
961
1480
  /**
1481
+ * The unit shared by observed and limit (for example `ms`, `rps`, or `%`).
1482
+ *
962
1483
  * @generated from field: string unit = 6;
963
1484
  */
964
1485
  unit?: string;
@@ -971,30 +1492,48 @@ export declare type LoadTestResults_SloCheckJson = {
971
1492
  export declare const LoadTestResults_SloCheckSchema: GenMessage<LoadTestResults_SloCheck, {jsonType: LoadTestResults_SloCheckJson}>;
972
1493
 
973
1494
  /**
1495
+ * Result detail for an agent evaluation run. Every executed case is listed
1496
+ * with its rolled-up status and a flat list of metrics. JudgeInfo summarizes
1497
+ * the scoring model used across the run when LLM-as-judge metrics are present.
1498
+ *
974
1499
  * @generated from message alis.evals.v1.AgentEvalResults
975
1500
  */
976
1501
  export declare type AgentEvalResults = Message<"alis.evals.v1.AgentEvalResults"> & {
977
1502
  /**
1503
+ * Every case executed in this run, in execution order.
1504
+ *
978
1505
  * @generated from field: repeated alis.evals.v1.AgentEvalResults.Case cases = 1;
979
1506
  */
980
1507
  cases: AgentEvalResults_Case[];
981
1508
 
982
1509
  /**
1510
+ * Judge model provenance and call statistics for this run. Unset when the
1511
+ * run used no LLM-as-judge metrics.
1512
+ *
983
1513
  * @generated from field: optional alis.evals.v1.AgentEvalResults.JudgeInfo judge = 2;
984
1514
  */
985
1515
  judge?: AgentEvalResults_JudgeInfo | undefined;
986
1516
  };
987
1517
 
988
1518
  /**
1519
+ * Result detail for an agent evaluation run. Every executed case is listed
1520
+ * with its rolled-up status and a flat list of metrics. JudgeInfo summarizes
1521
+ * the scoring model used across the run when LLM-as-judge metrics are present.
1522
+ *
989
1523
  * @generated from message alis.evals.v1.AgentEvalResults
990
1524
  */
991
1525
  export declare type AgentEvalResultsJson = {
992
1526
  /**
1527
+ * Every case executed in this run, in execution order.
1528
+ *
993
1529
  * @generated from field: repeated alis.evals.v1.AgentEvalResults.Case cases = 1;
994
1530
  */
995
1531
  cases?: AgentEvalResults_CaseJson[];
996
1532
 
997
1533
  /**
1534
+ * Judge model provenance and call statistics for this run. Unset when the
1535
+ * run used no LLM-as-judge metrics.
1536
+ *
998
1537
  * @generated from field: optional alis.evals.v1.AgentEvalResults.JudgeInfo judge = 2;
999
1538
  */
1000
1539
  judge?: AgentEvalResults_JudgeInfoJson;
@@ -1007,50 +1546,82 @@ export declare type AgentEvalResultsJson = {
1007
1546
  export declare const AgentEvalResultsSchema: GenMessage<AgentEvalResults, {jsonType: AgentEvalResultsJson}>;
1008
1547
 
1009
1548
  /**
1549
+ * Provenance and statistics for the LLM-as-judge used in this run.
1550
+ *
1551
+ * The judge is a separate model invoked to score metrics that cannot be
1552
+ * checked deterministically (for example tone, reasoning quality, or policy
1553
+ * adherence).
1554
+ *
1010
1555
  * @generated from message alis.evals.v1.AgentEvalResults.JudgeInfo
1011
1556
  */
1012
1557
  export declare type AgentEvalResults_JudgeInfo = Message<"alis.evals.v1.AgentEvalResults.JudgeInfo"> & {
1013
1558
  /**
1559
+ * The judge model identifier (for example `gemini-2.5-pro`).
1560
+ *
1014
1561
  * @generated from field: string model = 1;
1015
1562
  */
1016
1563
  model: string;
1017
1564
 
1018
1565
  /**
1566
+ * The pinned model version, when available.
1567
+ *
1019
1568
  * @generated from field: optional string model_version = 2;
1020
1569
  */
1021
1570
  modelVersion?: string | undefined;
1022
1571
 
1023
1572
  /**
1573
+ * Total judge API calls made during this run (one or more per metric).
1574
+ *
1024
1575
  * @generated from field: int64 judge_call_count = 3;
1025
1576
  */
1026
1577
  judgeCallCount: bigint;
1027
1578
 
1028
1579
  /**
1580
+ * Judge infrastructure failures (timeout, API error). Distinct from
1581
+ * metric failures, where the judge ran successfully but assigned a score
1582
+ * below threshold.
1583
+ *
1029
1584
  * @generated from field: int64 judge_error_count = 4;
1030
1585
  */
1031
1586
  judgeErrorCount: bigint;
1032
1587
  };
1033
1588
 
1034
1589
  /**
1590
+ * Provenance and statistics for the LLM-as-judge used in this run.
1591
+ *
1592
+ * The judge is a separate model invoked to score metrics that cannot be
1593
+ * checked deterministically (for example tone, reasoning quality, or policy
1594
+ * adherence).
1595
+ *
1035
1596
  * @generated from message alis.evals.v1.AgentEvalResults.JudgeInfo
1036
1597
  */
1037
1598
  export declare type AgentEvalResults_JudgeInfoJson = {
1038
1599
  /**
1600
+ * The judge model identifier (for example `gemini-2.5-pro`).
1601
+ *
1039
1602
  * @generated from field: string model = 1;
1040
1603
  */
1041
1604
  model?: string;
1042
1605
 
1043
1606
  /**
1607
+ * The pinned model version, when available.
1608
+ *
1044
1609
  * @generated from field: optional string model_version = 2;
1045
1610
  */
1046
1611
  modelVersion?: string;
1047
1612
 
1048
1613
  /**
1614
+ * Total judge API calls made during this run (one or more per metric).
1615
+ *
1049
1616
  * @generated from field: int64 judge_call_count = 3;
1050
1617
  */
1051
1618
  judgeCallCount?: string;
1052
1619
 
1053
1620
  /**
1621
+ * Judge infrastructure failures (timeout, API error). Distinct from
1622
+ * metric failures, where the judge ran successfully but assigned a score
1623
+ * below threshold.
1624
+ *
1054
1625
  * @generated from field: int64 judge_error_count = 4;
1055
1626
  */
1056
1627
  judgeErrorCount?: string;
@@ -1063,70 +1634,102 @@ export declare type AgentEvalResults_JudgeInfoJson = {
1063
1634
  export declare const AgentEvalResults_JudgeInfoSchema: GenMessage<AgentEvalResults_JudgeInfo, {jsonType: AgentEvalResults_JudgeInfoJson}>;
1064
1635
 
1065
1636
  /**
1637
+ * The result of one agent evaluation case.
1638
+ *
1066
1639
  * @generated from message alis.evals.v1.AgentEvalResults.Case
1067
1640
  */
1068
1641
  export declare type AgentEvalResults_Case = Message<"alis.evals.v1.AgentEvalResults.Case"> & {
1069
1642
  /**
1643
+ * The case id (ADK evalId).
1644
+ *
1070
1645
  * @generated from field: string id = 1;
1071
1646
  */
1072
1647
  id: string;
1073
1648
 
1074
1649
  /**
1650
+ * The rolled-up status for the case.
1651
+ *
1075
1652
  * @generated from field: alis.evals.v1.Status status = 2;
1076
1653
  */
1077
1654
  status: Status;
1078
1655
 
1079
1656
  /**
1657
+ * Wall-clock duration for this case.
1658
+ *
1080
1659
  * @generated from field: google.protobuf.Duration duration = 3;
1081
1660
  */
1082
1661
  duration?: Duration | undefined;
1083
1662
 
1084
1663
  /**
1664
+ * ADK session identifier for triage and replay.
1665
+ *
1085
1666
  * @generated from field: string session_id = 4;
1086
1667
  */
1087
1668
  sessionId: string;
1088
1669
 
1089
1670
  /**
1671
+ * All metrics evaluated for this case, in execution order. Includes
1672
+ * deterministic checks, LLM-as-judge scores, and ADK prebuilt metrics.
1673
+ *
1090
1674
  * @generated from field: repeated alis.evals.v1.AgentEvalResults.Case.Metric metrics = 5;
1091
1675
  */
1092
1676
  metrics: AgentEvalResults_Case_Metric[];
1093
1677
 
1094
1678
  /**
1679
+ * Developer and framework validation outcomes for this case in declaration
1680
+ * order. Integration cases continue using checks instead.
1681
+ *
1095
1682
  * @generated from field: repeated alis.evals.v1.Validation validations = 6;
1096
1683
  */
1097
1684
  validations: Validation[];
1098
1685
  };
1099
1686
 
1100
1687
  /**
1688
+ * The result of one agent evaluation case.
1689
+ *
1101
1690
  * @generated from message alis.evals.v1.AgentEvalResults.Case
1102
1691
  */
1103
1692
  export declare type AgentEvalResults_CaseJson = {
1104
1693
  /**
1694
+ * The case id (ADK evalId).
1695
+ *
1105
1696
  * @generated from field: string id = 1;
1106
1697
  */
1107
1698
  id?: string;
1108
1699
 
1109
1700
  /**
1701
+ * The rolled-up status for the case.
1702
+ *
1110
1703
  * @generated from field: alis.evals.v1.Status status = 2;
1111
1704
  */
1112
1705
  status?: StatusJson;
1113
1706
 
1114
1707
  /**
1708
+ * Wall-clock duration for this case.
1709
+ *
1115
1710
  * @generated from field: google.protobuf.Duration duration = 3;
1116
1711
  */
1117
1712
  duration?: DurationJson;
1118
1713
 
1119
1714
  /**
1715
+ * ADK session identifier for triage and replay.
1716
+ *
1120
1717
  * @generated from field: string session_id = 4;
1121
1718
  */
1122
1719
  sessionId?: string;
1123
1720
 
1124
1721
  /**
1722
+ * All metrics evaluated for this case, in execution order. Includes
1723
+ * deterministic checks, LLM-as-judge scores, and ADK prebuilt metrics.
1724
+ *
1125
1725
  * @generated from field: repeated alis.evals.v1.AgentEvalResults.Case.Metric metrics = 5;
1126
1726
  */
1127
1727
  metrics?: AgentEvalResults_Case_MetricJson[];
1128
1728
 
1129
1729
  /**
1730
+ * Developer and framework validation outcomes for this case in declaration
1731
+ * order. Integration cases continue using checks instead.
1732
+ *
1130
1733
  * @generated from field: repeated alis.evals.v1.Validation validations = 6;
1131
1734
  */
1132
1735
  validations?: ValidationJson[];
@@ -1139,70 +1742,102 @@ export declare type AgentEvalResults_CaseJson = {
1139
1742
  export declare const AgentEvalResults_CaseSchema: GenMessage<AgentEvalResults_Case, {jsonType: AgentEvalResults_CaseJson}>;
1140
1743
 
1141
1744
  /**
1745
+ * One evaluated metric within a case.
1746
+ *
1142
1747
  * @generated from message alis.evals.v1.AgentEvalResults.Case.Metric
1143
1748
  */
1144
1749
  export declare type AgentEvalResults_Case_Metric = Message<"alis.evals.v1.AgentEvalResults.Case.Metric"> & {
1145
1750
  /**
1751
+ * A stable identifier for the metric (ADK metricName).
1752
+ *
1146
1753
  * @generated from field: string id = 1;
1147
1754
  */
1148
1755
  id: string;
1149
1756
 
1150
1757
  /**
1758
+ * The outcome of the metric.
1759
+ *
1151
1760
  * @generated from field: alis.evals.v1.Status status = 2;
1152
1761
  */
1153
1762
  status: Status;
1154
1763
 
1155
1764
  /**
1765
+ * The observed score, when applicable. Omitted for binary metrics.
1766
+ *
1156
1767
  * @generated from field: optional double score = 3;
1157
1768
  */
1158
1769
  score?: number | undefined;
1159
1770
 
1160
1771
  /**
1772
+ * The minimum score required for this metric to pass, when applicable.
1773
+ *
1161
1774
  * @generated from field: optional double threshold = 4;
1162
1775
  */
1163
1776
  threshold?: number | undefined;
1164
1777
 
1165
1778
  /**
1779
+ * Human-readable detail: validation failures, score below threshold,
1780
+ * or not-evaluated reason. Empty when the metric passed with no extra
1781
+ * context.
1782
+ *
1166
1783
  * @generated from field: string message = 5;
1167
1784
  */
1168
1785
  message: string;
1169
1786
 
1170
1787
  /**
1788
+ * Per-rubric breakdown when the metric defines multiple rubric dimensions.
1789
+ *
1171
1790
  * @generated from field: repeated alis.evals.v1.AgentEvalResults.Case.Metric.RubricScore rubric = 6;
1172
1791
  */
1173
1792
  rubric: AgentEvalResults_Case_Metric_RubricScore[];
1174
1793
  };
1175
1794
 
1176
1795
  /**
1796
+ * One evaluated metric within a case.
1797
+ *
1177
1798
  * @generated from message alis.evals.v1.AgentEvalResults.Case.Metric
1178
1799
  */
1179
1800
  export declare type AgentEvalResults_Case_MetricJson = {
1180
1801
  /**
1802
+ * A stable identifier for the metric (ADK metricName).
1803
+ *
1181
1804
  * @generated from field: string id = 1;
1182
1805
  */
1183
1806
  id?: string;
1184
1807
 
1185
1808
  /**
1809
+ * The outcome of the metric.
1810
+ *
1186
1811
  * @generated from field: alis.evals.v1.Status status = 2;
1187
1812
  */
1188
1813
  status?: StatusJson;
1189
1814
 
1190
1815
  /**
1816
+ * The observed score, when applicable. Omitted for binary metrics.
1817
+ *
1191
1818
  * @generated from field: optional double score = 3;
1192
1819
  */
1193
1820
  score?: number | "NaN" | "Infinity" | "-Infinity";
1194
1821
 
1195
1822
  /**
1823
+ * The minimum score required for this metric to pass, when applicable.
1824
+ *
1196
1825
  * @generated from field: optional double threshold = 4;
1197
1826
  */
1198
1827
  threshold?: number | "NaN" | "Infinity" | "-Infinity";
1199
1828
 
1200
1829
  /**
1830
+ * Human-readable detail: validation failures, score below threshold,
1831
+ * or not-evaluated reason. Empty when the metric passed with no extra
1832
+ * context.
1833
+ *
1201
1834
  * @generated from field: string message = 5;
1202
1835
  */
1203
1836
  message?: string;
1204
1837
 
1205
1838
  /**
1839
+ * Per-rubric breakdown when the metric defines multiple rubric dimensions.
1840
+ *
1206
1841
  * @generated from field: repeated alis.evals.v1.AgentEvalResults.Case.Metric.RubricScore rubric = 6;
1207
1842
  */
1208
1843
  rubric?: AgentEvalResults_Case_Metric_RubricScoreJson[];
@@ -1215,50 +1850,70 @@ export declare type AgentEvalResults_Case_MetricJson = {
1215
1850
  export declare const AgentEvalResults_Case_MetricSchema: GenMessage<AgentEvalResults_Case_Metric, {jsonType: AgentEvalResults_Case_MetricJson}>;
1216
1851
 
1217
1852
  /**
1853
+ * The result for one rubric dimension within a metric.
1854
+ *
1218
1855
  * @generated from message alis.evals.v1.AgentEvalResults.Case.Metric.RubricScore
1219
1856
  */
1220
1857
  export declare type AgentEvalResults_Case_Metric_RubricScore = Message<"alis.evals.v1.AgentEvalResults.Case.Metric.RubricScore"> & {
1221
1858
  /**
1859
+ * The identifier of the rubric dimension.
1860
+ *
1222
1861
  * @generated from field: string id = 1;
1223
1862
  */
1224
1863
  id: string;
1225
1864
 
1226
1865
  /**
1866
+ * The outcome assigned to this rubric dimension.
1867
+ *
1227
1868
  * @generated from field: alis.evals.v1.Status status = 2;
1228
1869
  */
1229
1870
  status: Status;
1230
1871
 
1231
1872
  /**
1873
+ * The observed score for this rubric dimension, when applicable.
1874
+ *
1232
1875
  * @generated from field: optional double score = 3;
1233
1876
  */
1234
1877
  score?: number | undefined;
1235
1878
 
1236
1879
  /**
1880
+ * The rationale for the score, when applicable.
1881
+ *
1237
1882
  * @generated from field: optional string rationale = 4;
1238
1883
  */
1239
1884
  rationale?: string | undefined;
1240
1885
  };
1241
1886
 
1242
1887
  /**
1888
+ * The result for one rubric dimension within a metric.
1889
+ *
1243
1890
  * @generated from message alis.evals.v1.AgentEvalResults.Case.Metric.RubricScore
1244
1891
  */
1245
1892
  export declare type AgentEvalResults_Case_Metric_RubricScoreJson = {
1246
1893
  /**
1894
+ * The identifier of the rubric dimension.
1895
+ *
1247
1896
  * @generated from field: string id = 1;
1248
1897
  */
1249
1898
  id?: string;
1250
1899
 
1251
1900
  /**
1901
+ * The outcome assigned to this rubric dimension.
1902
+ *
1252
1903
  * @generated from field: alis.evals.v1.Status status = 2;
1253
1904
  */
1254
1905
  status?: StatusJson;
1255
1906
 
1256
1907
  /**
1908
+ * The observed score for this rubric dimension, when applicable.
1909
+ *
1257
1910
  * @generated from field: optional double score = 3;
1258
1911
  */
1259
1912
  score?: number | "NaN" | "Infinity" | "-Infinity";
1260
1913
 
1261
1914
  /**
1915
+ * The rationale for the score, when applicable.
1916
+ *
1262
1917
  * @generated from field: optional string rationale = 4;
1263
1918
  */
1264
1919
  rationale?: string;
@@ -1271,50 +1926,76 @@ export declare type AgentEvalResults_Case_Metric_RubricScoreJson = {
1271
1926
  export declare const AgentEvalResults_Case_Metric_RubricScoreSchema: GenMessage<AgentEvalResults_Case_Metric_RubricScore, {jsonType: AgentEvalResults_Case_Metric_RubricScoreJson}>;
1272
1927
 
1273
1928
  /**
1929
+ * Declared Cloud Run target for infrastructure observation.
1930
+ *
1931
+ * Identifies the Monitoring resource filters used when fetching server-side
1932
+ * Cloud Run metrics. Referenced from CloudRunTargetSnapshot.target.
1933
+ *
1274
1934
  * @generated from message alis.evals.v1.CloudRunTargetRef
1275
1935
  */
1276
1936
  export declare type CloudRunTargetRef = Message<"alis.evals.v1.CloudRunTargetRef"> & {
1277
1937
  /**
1938
+ * Google Cloud project ID hosting the Cloud Run service.
1939
+ *
1278
1940
  * @generated from field: string project_id = 1;
1279
1941
  */
1280
1942
  projectId: string;
1281
1943
 
1282
1944
  /**
1945
+ * Cloud Run region (for example `europe-west1`).
1946
+ *
1283
1947
  * @generated from field: string region = 2;
1284
1948
  */
1285
1949
  region: string;
1286
1950
 
1287
1951
  /**
1952
+ * Cloud Run service name (for example `search-v1`).
1953
+ *
1288
1954
  * @generated from field: string service_name = 3;
1289
1955
  */
1290
1956
  serviceName: string;
1291
1957
 
1292
1958
  /**
1959
+ * Optional revision filter. Empty aggregates all revisions.
1960
+ *
1293
1961
  * @generated from field: optional string revision = 4;
1294
1962
  */
1295
1963
  revision?: string | undefined;
1296
1964
  };
1297
1965
 
1298
1966
  /**
1967
+ * Declared Cloud Run target for infrastructure observation.
1968
+ *
1969
+ * Identifies the Monitoring resource filters used when fetching server-side
1970
+ * Cloud Run metrics. Referenced from CloudRunTargetSnapshot.target.
1971
+ *
1299
1972
  * @generated from message alis.evals.v1.CloudRunTargetRef
1300
1973
  */
1301
1974
  export declare type CloudRunTargetRefJson = {
1302
1975
  /**
1976
+ * Google Cloud project ID hosting the Cloud Run service.
1977
+ *
1303
1978
  * @generated from field: string project_id = 1;
1304
1979
  */
1305
1980
  projectId?: string;
1306
1981
 
1307
1982
  /**
1983
+ * Cloud Run region (for example `europe-west1`).
1984
+ *
1308
1985
  * @generated from field: string region = 2;
1309
1986
  */
1310
1987
  region?: string;
1311
1988
 
1312
1989
  /**
1990
+ * Cloud Run service name (for example `search-v1`).
1991
+ *
1313
1992
  * @generated from field: string service_name = 3;
1314
1993
  */
1315
1994
  serviceName?: string;
1316
1995
 
1317
1996
  /**
1997
+ * Optional revision filter. Empty aggregates all revisions.
1998
+ *
1318
1999
  * @generated from field: optional string revision = 4;
1319
2000
  */
1320
2001
  revision?: string;
@@ -1327,50 +2008,76 @@ export declare type CloudRunTargetRefJson = {
1327
2008
  export declare const CloudRunTargetRefSchema: GenMessage<CloudRunTargetRef, {jsonType: CloudRunTargetRefJson}>;
1328
2009
 
1329
2010
  /**
2011
+ * Declared Spanner target for infrastructure observation.
2012
+ *
2013
+ * Identifies the Monitoring resource filters used when fetching server-side
2014
+ * Spanner metrics. Referenced from SpannerTargetSnapshot.target.
2015
+ *
1330
2016
  * @generated from message alis.evals.v1.SpannerTargetRef
1331
2017
  */
1332
2018
  export declare type SpannerTargetRef = Message<"alis.evals.v1.SpannerTargetRef"> & {
1333
2019
  /**
2020
+ * Google Cloud project ID hosting the Spanner instance.
2021
+ *
1334
2022
  * @generated from field: string project_id = 1;
1335
2023
  */
1336
2024
  projectId: string;
1337
2025
 
1338
2026
  /**
2027
+ * Spanner instance ID (for example `prod-spanner`).
2028
+ *
1339
2029
  * @generated from field: string instance_id = 2;
1340
2030
  */
1341
2031
  instanceId: string;
1342
2032
 
1343
2033
  /**
2034
+ * Spanner instance location (for example `europe-west1`).
2035
+ *
1344
2036
  * @generated from field: string location = 3;
1345
2037
  */
1346
2038
  location: string;
1347
2039
 
1348
2040
  /**
2041
+ * Required database within the instance (for example `orders`).
2042
+ *
1349
2043
  * @generated from field: string database = 4;
1350
2044
  */
1351
2045
  database: string;
1352
2046
  };
1353
2047
 
1354
2048
  /**
2049
+ * Declared Spanner target for infrastructure observation.
2050
+ *
2051
+ * Identifies the Monitoring resource filters used when fetching server-side
2052
+ * Spanner metrics. Referenced from SpannerTargetSnapshot.target.
2053
+ *
1355
2054
  * @generated from message alis.evals.v1.SpannerTargetRef
1356
2055
  */
1357
2056
  export declare type SpannerTargetRefJson = {
1358
2057
  /**
2058
+ * Google Cloud project ID hosting the Spanner instance.
2059
+ *
1359
2060
  * @generated from field: string project_id = 1;
1360
2061
  */
1361
2062
  projectId?: string;
1362
2063
 
1363
2064
  /**
2065
+ * Spanner instance ID (for example `prod-spanner`).
2066
+ *
1364
2067
  * @generated from field: string instance_id = 2;
1365
2068
  */
1366
2069
  instanceId?: string;
1367
2070
 
1368
2071
  /**
2072
+ * Spanner instance location (for example `europe-west1`).
2073
+ *
1369
2074
  * @generated from field: string location = 3;
1370
2075
  */
1371
2076
  location?: string;
1372
2077
 
1373
2078
  /**
2079
+ * Required database within the instance (for example `orders`).
2080
+ *
1374
2081
  * @generated from field: string database = 4;
1375
2082
  */
1376
2083
  database?: string;
@@ -1383,80 +2090,122 @@ export declare type SpannerTargetRefJson = {
1383
2090
  export declare const SpannerTargetRefSchema: GenMessage<SpannerTargetRef, {jsonType: SpannerTargetRefJson}>;
1384
2091
 
1385
2092
  /**
2093
+ * Observed Cloud Run server-side metrics for one target and window.
2094
+ *
1386
2095
  * @generated from message alis.evals.v1.CloudRunMetrics
1387
2096
  */
1388
2097
  export declare type CloudRunMetrics = Message<"alis.evals.v1.CloudRunMetrics"> & {
1389
2098
  /**
2099
+ * Server-side request count. Zero when the service saw no traffic.
2100
+ *
1390
2101
  * @generated from field: int64 request_count = 1;
1391
2102
  */
1392
2103
  requestCount: bigint;
1393
2104
 
1394
2105
  /**
2106
+ * Server-side request latency distribution (run.googleapis.com/request_latencies).
2107
+ * Unset when this metric type could not be fetched.
2108
+ *
1395
2109
  * @generated from field: optional alis.evals.v1.LatencyPercentiles latency = 2;
1396
2110
  */
1397
2111
  latency?: LatencyPercentiles | undefined;
1398
2112
 
1399
2113
  /**
2114
+ * Fraction of server-side responses with 5xx status (0.0–1.0). Unset when
2115
+ * this metric type could not be fetched.
2116
+ *
1400
2117
  * @generated from field: optional double error_5xx_rate = 3;
1401
2118
  */
1402
2119
  error5xxRate?: number | undefined;
1403
2120
 
1404
2121
  /**
2122
+ * Peak concurrent instance count during the window. Unset when this metric
2123
+ * type could not be fetched.
2124
+ *
1405
2125
  * @generated from field: optional double max_instance_count = 4;
1406
2126
  */
1407
2127
  maxInstanceCount?: number | undefined;
1408
2128
 
1409
2129
  /**
2130
+ * Container CPU utilization 99th percentile during the window (0.0–1.0).
2131
+ * Unset when this metric type could not be fetched.
2132
+ *
1410
2133
  * @generated from field: optional double cpu_utilization_p99 = 5;
1411
2134
  */
1412
2135
  cpuUtilizationP99?: number | undefined;
1413
2136
 
1414
2137
  /**
2138
+ * Container memory utilization 99th percentile during the window (0.0–1.0).
2139
+ * Unset when this metric type could not be fetched.
2140
+ *
1415
2141
  * @generated from field: optional double memory_utilization_p99 = 6;
1416
2142
  */
1417
2143
  memoryUtilizationP99?: number | undefined;
1418
2144
 
1419
2145
  /**
2146
+ * Container startup latency 99th percentile in milliseconds, when available.
2147
+ *
1420
2148
  * @generated from field: optional double startup_latency_p99 = 7;
1421
2149
  */
1422
2150
  startupLatencyP99?: number | undefined;
1423
2151
  };
1424
2152
 
1425
2153
  /**
2154
+ * Observed Cloud Run server-side metrics for one target and window.
2155
+ *
1426
2156
  * @generated from message alis.evals.v1.CloudRunMetrics
1427
2157
  */
1428
2158
  export declare type CloudRunMetricsJson = {
1429
2159
  /**
2160
+ * Server-side request count. Zero when the service saw no traffic.
2161
+ *
1430
2162
  * @generated from field: int64 request_count = 1;
1431
2163
  */
1432
2164
  requestCount?: string;
1433
2165
 
1434
2166
  /**
2167
+ * Server-side request latency distribution (run.googleapis.com/request_latencies).
2168
+ * Unset when this metric type could not be fetched.
2169
+ *
1435
2170
  * @generated from field: optional alis.evals.v1.LatencyPercentiles latency = 2;
1436
2171
  */
1437
2172
  latency?: LatencyPercentilesJson;
1438
2173
 
1439
2174
  /**
2175
+ * Fraction of server-side responses with 5xx status (0.0–1.0). Unset when
2176
+ * this metric type could not be fetched.
2177
+ *
1440
2178
  * @generated from field: optional double error_5xx_rate = 3;
1441
2179
  */
1442
2180
  error5xxRate?: number | "NaN" | "Infinity" | "-Infinity";
1443
2181
 
1444
2182
  /**
2183
+ * Peak concurrent instance count during the window. Unset when this metric
2184
+ * type could not be fetched.
2185
+ *
1445
2186
  * @generated from field: optional double max_instance_count = 4;
1446
2187
  */
1447
2188
  maxInstanceCount?: number | "NaN" | "Infinity" | "-Infinity";
1448
2189
 
1449
2190
  /**
2191
+ * Container CPU utilization 99th percentile during the window (0.0–1.0).
2192
+ * Unset when this metric type could not be fetched.
2193
+ *
1450
2194
  * @generated from field: optional double cpu_utilization_p99 = 5;
1451
2195
  */
1452
2196
  cpuUtilizationP99?: number | "NaN" | "Infinity" | "-Infinity";
1453
2197
 
1454
2198
  /**
2199
+ * Container memory utilization 99th percentile during the window (0.0–1.0).
2200
+ * Unset when this metric type could not be fetched.
2201
+ *
1455
2202
  * @generated from field: optional double memory_utilization_p99 = 6;
1456
2203
  */
1457
2204
  memoryUtilizationP99?: number | "NaN" | "Infinity" | "-Infinity";
1458
2205
 
1459
2206
  /**
2207
+ * Container startup latency 99th percentile in milliseconds, when available.
2208
+ *
1460
2209
  * @generated from field: optional double startup_latency_p99 = 7;
1461
2210
  */
1462
2211
  startupLatencyP99?: number | "NaN" | "Infinity" | "-Infinity";
@@ -1469,60 +2218,94 @@ export declare type CloudRunMetricsJson = {
1469
2218
  export declare const CloudRunMetricsSchema: GenMessage<CloudRunMetrics, {jsonType: CloudRunMetricsJson}>;
1470
2219
 
1471
2220
  /**
2221
+ * Observed Spanner server-side metrics for one target and window.
2222
+ *
1472
2223
  * @generated from message alis.evals.v1.SpannerMetrics
1473
2224
  */
1474
2225
  export declare type SpannerMetrics = Message<"alis.evals.v1.SpannerMetrics"> & {
1475
2226
  /**
2227
+ * Query count for the database during the window. Zero when no queries ran.
2228
+ *
1476
2229
  * @generated from field: int64 query_count = 1;
1477
2230
  */
1478
2231
  queryCount: bigint;
1479
2232
 
1480
2233
  /**
2234
+ * Query count where status was not OK during the window.
2235
+ *
1481
2236
  * @generated from field: int64 query_error_count = 2;
1482
2237
  */
1483
2238
  queryErrorCount: bigint;
1484
2239
 
1485
2240
  /**
2241
+ * API request latency distribution (spanner.googleapis.com/api/request_latencies).
2242
+ * Unset when this metric type could not be fetched.
2243
+ *
1486
2244
  * @generated from field: optional alis.evals.v1.LatencyPercentiles api_latency = 3;
1487
2245
  */
1488
2246
  apiLatency?: LatencyPercentiles | undefined;
1489
2247
 
1490
2248
  /**
2249
+ * Query execution latency distribution
2250
+ * (spanner.googleapis.com/query_stat/total/query_latencies). Unset when this
2251
+ * metric type could not be fetched.
2252
+ *
1491
2253
  * @generated from field: optional alis.evals.v1.LatencyPercentiles query_latency = 4;
1492
2254
  */
1493
2255
  queryLatency?: LatencyPercentiles | undefined;
1494
2256
 
1495
2257
  /**
2258
+ * Instance CPU utilization maximum during the window (0.0–1.0). Instance-
2259
+ * scoped; not attributable to a single database when multiple databases share
2260
+ * the instance. Unset when this metric type could not be fetched.
2261
+ *
1496
2262
  * @generated from field: optional double cpu_utilization_max = 5;
1497
2263
  */
1498
2264
  cpuUtilizationMax?: number | undefined;
1499
2265
  };
1500
2266
 
1501
2267
  /**
2268
+ * Observed Spanner server-side metrics for one target and window.
2269
+ *
1502
2270
  * @generated from message alis.evals.v1.SpannerMetrics
1503
2271
  */
1504
2272
  export declare type SpannerMetricsJson = {
1505
2273
  /**
2274
+ * Query count for the database during the window. Zero when no queries ran.
2275
+ *
1506
2276
  * @generated from field: int64 query_count = 1;
1507
2277
  */
1508
2278
  queryCount?: string;
1509
2279
 
1510
2280
  /**
2281
+ * Query count where status was not OK during the window.
2282
+ *
1511
2283
  * @generated from field: int64 query_error_count = 2;
1512
2284
  */
1513
2285
  queryErrorCount?: string;
1514
2286
 
1515
2287
  /**
2288
+ * API request latency distribution (spanner.googleapis.com/api/request_latencies).
2289
+ * Unset when this metric type could not be fetched.
2290
+ *
1516
2291
  * @generated from field: optional alis.evals.v1.LatencyPercentiles api_latency = 3;
1517
2292
  */
1518
2293
  apiLatency?: LatencyPercentilesJson;
1519
2294
 
1520
2295
  /**
2296
+ * Query execution latency distribution
2297
+ * (spanner.googleapis.com/query_stat/total/query_latencies). Unset when this
2298
+ * metric type could not be fetched.
2299
+ *
1521
2300
  * @generated from field: optional alis.evals.v1.LatencyPercentiles query_latency = 4;
1522
2301
  */
1523
2302
  queryLatency?: LatencyPercentilesJson;
1524
2303
 
1525
2304
  /**
2305
+ * Instance CPU utilization maximum during the window (0.0–1.0). Instance-
2306
+ * scoped; not attributable to a single database when multiple databases share
2307
+ * the instance. Unset when this metric type could not be fetched.
2308
+ *
1526
2309
  * @generated from field: optional double cpu_utilization_max = 5;
1527
2310
  */
1528
2311
  cpuUtilizationMax?: number | "NaN" | "Infinity" | "-Infinity";
@@ -1535,90 +2318,140 @@ export declare type SpannerMetricsJson = {
1535
2318
  export declare const SpannerMetricsSchema: GenMessage<SpannerMetrics, {jsonType: SpannerMetricsJson}>;
1536
2319
 
1537
2320
  /**
2321
+ * Cloud Run infrastructure snapshot for one declared target.
2322
+ *
2323
+ * Always emitted for every target declared on the suite, even when
2324
+ * request_count is zero or the fetch partially failed. Pairs configuration
2325
+ * (id, role, target) with observed metrics over window_start..window_end.
2326
+ *
1538
2327
  * @generated from message alis.evals.v1.CloudRunTargetSnapshot
1539
2328
  */
1540
2329
  export declare type CloudRunTargetSnapshot = Message<"alis.evals.v1.CloudRunTargetSnapshot"> & {
1541
2330
  /**
2331
+ * Stable target identifier from suite configuration (for example `search-v1`).
2332
+ *
1542
2333
  * @generated from field: string id = 1;
1543
2334
  */
1544
2335
  id: string;
1545
2336
 
1546
2337
  /**
2338
+ * Role of this target within the case (ENTRY or DEPENDENCY).
2339
+ *
1547
2340
  * @generated from field: alis.evals.v1.InfraTargetRole role = 2;
1548
2341
  */
1549
2342
  role: InfraTargetRole;
1550
2343
 
1551
2344
  /**
2345
+ * Declared Cloud Run resource coordinates used for the Monitoring query.
2346
+ *
1552
2347
  * @generated from field: alis.evals.v1.CloudRunTargetRef target = 3;
1553
2348
  */
1554
2349
  target?: CloudRunTargetRef | undefined;
1555
2350
 
1556
2351
  /**
2352
+ * Inclusive start of the observation window (UTC).
2353
+ *
1557
2354
  * @generated from field: google.protobuf.Timestamp window_start = 4;
1558
2355
  */
1559
2356
  windowStart?: Timestamp | undefined;
1560
2357
 
1561
2358
  /**
2359
+ * Exclusive end of the observation window (UTC).
2360
+ *
1562
2361
  * @generated from field: google.protobuf.Timestamp window_end = 5;
1563
2362
  */
1564
2363
  windowEnd?: Timestamp | undefined;
1565
2364
 
1566
2365
  /**
2366
+ * Outcome of the Monitoring fetch for this target.
2367
+ *
1567
2368
  * @generated from field: alis.evals.v1.InfraFetchStatus fetch_status = 6;
1568
2369
  */
1569
2370
  fetchStatus: InfraFetchStatus;
1570
2371
 
1571
2372
  /**
2373
+ * Human-readable detail when fetch_status is not OK or when individual
2374
+ * metrics failed partially. Unset on full success.
2375
+ *
1572
2376
  * @generated from field: optional string fetch_message = 7;
1573
2377
  */
1574
2378
  fetchMessage?: string | undefined;
1575
2379
 
1576
2380
  /**
2381
+ * Observed server-side metrics for this target and window. Populated on
2382
+ * best-effort basis when fetch_status is OK; fields may be unset when a
2383
+ * specific metric type failed.
2384
+ *
1577
2385
  * @generated from field: alis.evals.v1.CloudRunMetrics metrics = 8;
1578
2386
  */
1579
2387
  metrics?: CloudRunMetrics | undefined;
1580
2388
  };
1581
2389
 
1582
2390
  /**
2391
+ * Cloud Run infrastructure snapshot for one declared target.
2392
+ *
2393
+ * Always emitted for every target declared on the suite, even when
2394
+ * request_count is zero or the fetch partially failed. Pairs configuration
2395
+ * (id, role, target) with observed metrics over window_start..window_end.
2396
+ *
1583
2397
  * @generated from message alis.evals.v1.CloudRunTargetSnapshot
1584
2398
  */
1585
2399
  export declare type CloudRunTargetSnapshotJson = {
1586
2400
  /**
2401
+ * Stable target identifier from suite configuration (for example `search-v1`).
2402
+ *
1587
2403
  * @generated from field: string id = 1;
1588
2404
  */
1589
2405
  id?: string;
1590
2406
 
1591
2407
  /**
2408
+ * Role of this target within the case (ENTRY or DEPENDENCY).
2409
+ *
1592
2410
  * @generated from field: alis.evals.v1.InfraTargetRole role = 2;
1593
2411
  */
1594
2412
  role?: InfraTargetRoleJson;
1595
2413
 
1596
2414
  /**
2415
+ * Declared Cloud Run resource coordinates used for the Monitoring query.
2416
+ *
1597
2417
  * @generated from field: alis.evals.v1.CloudRunTargetRef target = 3;
1598
2418
  */
1599
2419
  target?: CloudRunTargetRefJson;
1600
2420
 
1601
2421
  /**
2422
+ * Inclusive start of the observation window (UTC).
2423
+ *
1602
2424
  * @generated from field: google.protobuf.Timestamp window_start = 4;
1603
2425
  */
1604
2426
  windowStart?: TimestampJson;
1605
2427
 
1606
2428
  /**
2429
+ * Exclusive end of the observation window (UTC).
2430
+ *
1607
2431
  * @generated from field: google.protobuf.Timestamp window_end = 5;
1608
2432
  */
1609
2433
  windowEnd?: TimestampJson;
1610
2434
 
1611
2435
  /**
2436
+ * Outcome of the Monitoring fetch for this target.
2437
+ *
1612
2438
  * @generated from field: alis.evals.v1.InfraFetchStatus fetch_status = 6;
1613
2439
  */
1614
2440
  fetchStatus?: InfraFetchStatusJson;
1615
2441
 
1616
2442
  /**
2443
+ * Human-readable detail when fetch_status is not OK or when individual
2444
+ * metrics failed partially. Unset on full success.
2445
+ *
1617
2446
  * @generated from field: optional string fetch_message = 7;
1618
2447
  */
1619
2448
  fetchMessage?: string;
1620
2449
 
1621
2450
  /**
2451
+ * Observed server-side metrics for this target and window. Populated on
2452
+ * best-effort basis when fetch_status is OK; fields may be unset when a
2453
+ * specific metric type failed.
2454
+ *
1622
2455
  * @generated from field: alis.evals.v1.CloudRunMetrics metrics = 8;
1623
2456
  */
1624
2457
  metrics?: CloudRunMetricsJson;
@@ -1631,90 +2464,140 @@ export declare type CloudRunTargetSnapshotJson = {
1631
2464
  export declare const CloudRunTargetSnapshotSchema: GenMessage<CloudRunTargetSnapshot, {jsonType: CloudRunTargetSnapshotJson}>;
1632
2465
 
1633
2466
  /**
2467
+ * Spanner infrastructure snapshot for one declared target.
2468
+ *
2469
+ * Always emitted for every target declared on the suite, even when
2470
+ * query_count is zero or the fetch partially failed. Pairs configuration
2471
+ * (id, role, target) with observed metrics over window_start..window_end.
2472
+ *
1634
2473
  * @generated from message alis.evals.v1.SpannerTargetSnapshot
1635
2474
  */
1636
2475
  export declare type SpannerTargetSnapshot = Message<"alis.evals.v1.SpannerTargetSnapshot"> & {
1637
2476
  /**
2477
+ * Stable target identifier from suite configuration (for example `orders-db`).
2478
+ *
1638
2479
  * @generated from field: string id = 1;
1639
2480
  */
1640
2481
  id: string;
1641
2482
 
1642
2483
  /**
2484
+ * Always INFRA_TARGET_ROLE_DEPENDENCY on the wire for Spanner targets.
2485
+ *
1643
2486
  * @generated from field: alis.evals.v1.InfraTargetRole role = 2;
1644
2487
  */
1645
2488
  role: InfraTargetRole;
1646
2489
 
1647
2490
  /**
2491
+ * Declared Spanner resource coordinates used for the Monitoring query.
2492
+ *
1648
2493
  * @generated from field: alis.evals.v1.SpannerTargetRef target = 3;
1649
2494
  */
1650
2495
  target?: SpannerTargetRef | undefined;
1651
2496
 
1652
2497
  /**
2498
+ * Inclusive start of the observation window (UTC).
2499
+ *
1653
2500
  * @generated from field: google.protobuf.Timestamp window_start = 4;
1654
2501
  */
1655
2502
  windowStart?: Timestamp | undefined;
1656
2503
 
1657
2504
  /**
2505
+ * Exclusive end of the observation window (UTC).
2506
+ *
1658
2507
  * @generated from field: google.protobuf.Timestamp window_end = 5;
1659
2508
  */
1660
2509
  windowEnd?: Timestamp | undefined;
1661
2510
 
1662
2511
  /**
2512
+ * Outcome of the Monitoring fetch for this target.
2513
+ *
1663
2514
  * @generated from field: alis.evals.v1.InfraFetchStatus fetch_status = 6;
1664
2515
  */
1665
2516
  fetchStatus: InfraFetchStatus;
1666
2517
 
1667
2518
  /**
2519
+ * Human-readable detail when fetch_status is not OK or when individual
2520
+ * metrics failed partially. Unset on full success.
2521
+ *
1668
2522
  * @generated from field: optional string fetch_message = 7;
1669
2523
  */
1670
2524
  fetchMessage?: string | undefined;
1671
2525
 
1672
2526
  /**
2527
+ * Observed server-side metrics for this target and window. Populated on
2528
+ * best-effort basis when fetch_status is OK; fields may be unset when a
2529
+ * specific metric type failed.
2530
+ *
1673
2531
  * @generated from field: alis.evals.v1.SpannerMetrics metrics = 8;
1674
2532
  */
1675
2533
  metrics?: SpannerMetrics | undefined;
1676
2534
  };
1677
2535
 
1678
2536
  /**
2537
+ * Spanner infrastructure snapshot for one declared target.
2538
+ *
2539
+ * Always emitted for every target declared on the suite, even when
2540
+ * query_count is zero or the fetch partially failed. Pairs configuration
2541
+ * (id, role, target) with observed metrics over window_start..window_end.
2542
+ *
1679
2543
  * @generated from message alis.evals.v1.SpannerTargetSnapshot
1680
2544
  */
1681
2545
  export declare type SpannerTargetSnapshotJson = {
1682
2546
  /**
2547
+ * Stable target identifier from suite configuration (for example `orders-db`).
2548
+ *
1683
2549
  * @generated from field: string id = 1;
1684
2550
  */
1685
2551
  id?: string;
1686
2552
 
1687
2553
  /**
2554
+ * Always INFRA_TARGET_ROLE_DEPENDENCY on the wire for Spanner targets.
2555
+ *
1688
2556
  * @generated from field: alis.evals.v1.InfraTargetRole role = 2;
1689
2557
  */
1690
2558
  role?: InfraTargetRoleJson;
1691
2559
 
1692
2560
  /**
2561
+ * Declared Spanner resource coordinates used for the Monitoring query.
2562
+ *
1693
2563
  * @generated from field: alis.evals.v1.SpannerTargetRef target = 3;
1694
2564
  */
1695
2565
  target?: SpannerTargetRefJson;
1696
2566
 
1697
2567
  /**
2568
+ * Inclusive start of the observation window (UTC).
2569
+ *
1698
2570
  * @generated from field: google.protobuf.Timestamp window_start = 4;
1699
2571
  */
1700
2572
  windowStart?: TimestampJson;
1701
2573
 
1702
2574
  /**
2575
+ * Exclusive end of the observation window (UTC).
2576
+ *
1703
2577
  * @generated from field: google.protobuf.Timestamp window_end = 5;
1704
2578
  */
1705
2579
  windowEnd?: TimestampJson;
1706
2580
 
1707
2581
  /**
2582
+ * Outcome of the Monitoring fetch for this target.
2583
+ *
1708
2584
  * @generated from field: alis.evals.v1.InfraFetchStatus fetch_status = 6;
1709
2585
  */
1710
2586
  fetchStatus?: InfraFetchStatusJson;
1711
2587
 
1712
2588
  /**
2589
+ * Human-readable detail when fetch_status is not OK or when individual
2590
+ * metrics failed partially. Unset on full success.
2591
+ *
1713
2592
  * @generated from field: optional string fetch_message = 7;
1714
2593
  */
1715
2594
  fetchMessage?: string;
1716
2595
 
1717
2596
  /**
2597
+ * Observed server-side metrics for this target and window. Populated on
2598
+ * best-effort basis when fetch_status is OK; fields may be unset when a
2599
+ * specific metric type failed.
2600
+ *
1718
2601
  * @generated from field: alis.evals.v1.SpannerMetrics metrics = 8;
1719
2602
  */
1720
2603
  metrics?: SpannerMetricsJson;
@@ -1727,90 +2610,138 @@ export declare type SpannerTargetSnapshotJson = {
1727
2610
  export declare const SpannerTargetSnapshotSchema: GenMessage<SpannerTargetSnapshot, {jsonType: SpannerTargetSnapshotJson}>;
1728
2611
 
1729
2612
  /**
2613
+ * One infrastructure SLO check outcome.
2614
+ *
2615
+ * Mirrors SloCheck on the client side: snapshots carry observed values;
2616
+ * infra_checks carry pass/fail against configured limits. Diagnostics-only in
2617
+ * v1: infra_checks is always empty on the wire until infra SLO evaluators ship.
2618
+ *
1730
2619
  * @generated from message alis.evals.v1.InfraSloCheck
1731
2620
  */
1732
2621
  export declare type InfraSloCheck = Message<"alis.evals.v1.InfraSloCheck"> & {
1733
2622
  /**
2623
+ * Resource kind this check evaluated (CLOUD_RUN or SPANNER).
2624
+ *
1734
2625
  * @generated from field: alis.evals.v1.InfraKind kind = 1;
1735
2626
  */
1736
2627
  kind: InfraKind;
1737
2628
 
1738
2629
  /**
2630
+ * Matches CloudRunTargetSnapshot.id or SpannerTargetSnapshot.id.
2631
+ *
1739
2632
  * @generated from field: string target_id = 2;
1740
2633
  */
1741
2634
  targetId: string;
1742
2635
 
1743
2636
  /**
2637
+ * Stable metric key (for example `latency.p99_ms` or `query_error_rate`).
2638
+ *
1744
2639
  * @generated from field: string check_id = 3;
1745
2640
  */
1746
2641
  checkId: string;
1747
2642
 
1748
2643
  /**
2644
+ * The outcome of the check. FAILED when the observed value crossed the
2645
+ * limit or the metric could not be measured.
2646
+ *
1749
2647
  * @generated from field: alis.evals.v1.Status status = 4;
1750
2648
  */
1751
2649
  status: Status;
1752
2650
 
1753
2651
  /**
2652
+ * Human-readable detail. Unset when the check passed with no extra context.
2653
+ *
1754
2654
  * @generated from field: optional string message = 5;
1755
2655
  */
1756
2656
  message?: string | undefined;
1757
2657
 
1758
2658
  /**
2659
+ * The metric value measured during the observation window (for example 612.4
2660
+ * when unit is `ms`).
2661
+ *
1759
2662
  * @generated from field: double observed = 6;
1760
2663
  */
1761
2664
  observed: number;
1762
2665
 
1763
2666
  /**
2667
+ * The configured SLO bound (for example 500.0 when unit is `ms`).
2668
+ *
1764
2669
  * @generated from field: double limit = 7;
1765
2670
  */
1766
2671
  limit: number;
1767
2672
 
1768
2673
  /**
2674
+ * Unit shared by observed and limit (for example `ms`, `%`, `count`).
2675
+ *
1769
2676
  * @generated from field: string unit = 8;
1770
2677
  */
1771
2678
  unit: string;
1772
2679
  };
1773
2680
 
1774
2681
  /**
2682
+ * One infrastructure SLO check outcome.
2683
+ *
2684
+ * Mirrors SloCheck on the client side: snapshots carry observed values;
2685
+ * infra_checks carry pass/fail against configured limits. Diagnostics-only in
2686
+ * v1: infra_checks is always empty on the wire until infra SLO evaluators ship.
2687
+ *
1775
2688
  * @generated from message alis.evals.v1.InfraSloCheck
1776
2689
  */
1777
2690
  export declare type InfraSloCheckJson = {
1778
2691
  /**
2692
+ * Resource kind this check evaluated (CLOUD_RUN or SPANNER).
2693
+ *
1779
2694
  * @generated from field: alis.evals.v1.InfraKind kind = 1;
1780
2695
  */
1781
2696
  kind?: InfraKindJson;
1782
2697
 
1783
2698
  /**
2699
+ * Matches CloudRunTargetSnapshot.id or SpannerTargetSnapshot.id.
2700
+ *
1784
2701
  * @generated from field: string target_id = 2;
1785
2702
  */
1786
2703
  targetId?: string;
1787
2704
 
1788
2705
  /**
2706
+ * Stable metric key (for example `latency.p99_ms` or `query_error_rate`).
2707
+ *
1789
2708
  * @generated from field: string check_id = 3;
1790
2709
  */
1791
2710
  checkId?: string;
1792
2711
 
1793
2712
  /**
2713
+ * The outcome of the check. FAILED when the observed value crossed the
2714
+ * limit or the metric could not be measured.
2715
+ *
1794
2716
  * @generated from field: alis.evals.v1.Status status = 4;
1795
2717
  */
1796
2718
  status?: StatusJson;
1797
2719
 
1798
2720
  /**
2721
+ * Human-readable detail. Unset when the check passed with no extra context.
2722
+ *
1799
2723
  * @generated from field: optional string message = 5;
1800
2724
  */
1801
2725
  message?: string;
1802
2726
 
1803
2727
  /**
2728
+ * The metric value measured during the observation window (for example 612.4
2729
+ * when unit is `ms`).
2730
+ *
1804
2731
  * @generated from field: double observed = 6;
1805
2732
  */
1806
2733
  observed?: number | "NaN" | "Infinity" | "-Infinity";
1807
2734
 
1808
2735
  /**
2736
+ * The configured SLO bound (for example 500.0 when unit is `ms`).
2737
+ *
1809
2738
  * @generated from field: double limit = 7;
1810
2739
  */
1811
2740
  limit?: number | "NaN" | "Infinity" | "-Infinity";
1812
2741
 
1813
2742
  /**
2743
+ * Unit shared by observed and limit (for example `ms`, `%`, `count`).
2744
+ *
1814
2745
  * @generated from field: string unit = 8;
1815
2746
  */
1816
2747
  unit?: string;
@@ -1823,20 +2754,36 @@ export declare type InfraSloCheckJson = {
1823
2754
  export declare const InfraSloCheckSchema: GenMessage<InfraSloCheck, {jsonType: InfraSloCheckJson}>;
1824
2755
 
1825
2756
  /**
2757
+ * Result detail for a standalone infrastructure observation run.
2758
+ *
2759
+ * Produced by RunInfraObservation without load generation. Each case fetches
2760
+ * Cloud Monitoring snapshots over a resolved lookback window for declared
2761
+ * Cloud Run and Spanner targets.
2762
+ *
1826
2763
  * @generated from message alis.evals.v1.InfraObservationResults
1827
2764
  */
1828
2765
  export declare type InfraObservationResults = Message<"alis.evals.v1.InfraObservationResults"> & {
1829
2766
  /**
2767
+ * Every infra observation case executed in this run, in execution order.
2768
+ *
1830
2769
  * @generated from field: repeated alis.evals.v1.InfraObservationResults.Case cases = 1;
1831
2770
  */
1832
2771
  cases: InfraObservationResults_Case[];
1833
2772
  };
1834
2773
 
1835
2774
  /**
2775
+ * Result detail for a standalone infrastructure observation run.
2776
+ *
2777
+ * Produced by RunInfraObservation without load generation. Each case fetches
2778
+ * Cloud Monitoring snapshots over a resolved lookback window for declared
2779
+ * Cloud Run and Spanner targets.
2780
+ *
1836
2781
  * @generated from message alis.evals.v1.InfraObservationResults
1837
2782
  */
1838
2783
  export declare type InfraObservationResultsJson = {
1839
2784
  /**
2785
+ * Every infra observation case executed in this run, in execution order.
2786
+ *
1840
2787
  * @generated from field: repeated alis.evals.v1.InfraObservationResults.Case cases = 1;
1841
2788
  */
1842
2789
  cases?: InfraObservationResults_CaseJson[];
@@ -1849,100 +2796,154 @@ export declare type InfraObservationResultsJson = {
1849
2796
  export declare const InfraObservationResultsSchema: GenMessage<InfraObservationResults, {jsonType: InfraObservationResultsJson}>;
1850
2797
 
1851
2798
  /**
2799
+ * The result of one infrastructure observation case.
2800
+ *
1852
2801
  * @generated from message alis.evals.v1.InfraObservationResults.Case
1853
2802
  */
1854
2803
  export declare type InfraObservationResults_Case = Message<"alis.evals.v1.InfraObservationResults.Case"> & {
1855
2804
  /**
2805
+ * The case id.
2806
+ * Example: peak-hours.observe-search
2807
+ *
1856
2808
  * @generated from field: string id = 1;
1857
2809
  */
1858
2810
  id: string;
1859
2811
 
1860
2812
  /**
2813
+ * The rolled-up status for the case. Diagnostics-only in v1: PASSED when
2814
+ * observation completes; FAILED only on runner or setup errors, not on
2815
+ * infra threshold breach until infra SLO evaluators ship.
2816
+ *
1861
2817
  * @generated from field: alis.evals.v1.Status status = 2;
1862
2818
  */
1863
2819
  status: Status;
1864
2820
 
1865
2821
  /**
2822
+ * Resolved lookback duration for this case.
2823
+ *
1866
2824
  * @generated from field: google.protobuf.Duration lookback = 3;
1867
2825
  */
1868
2826
  lookback?: Duration | undefined;
1869
2827
 
1870
2828
  /**
2829
+ * Inclusive start of the settled observation window (UTC).
2830
+ *
1871
2831
  * @generated from field: google.protobuf.Timestamp window_start = 4;
1872
2832
  */
1873
2833
  windowStart?: Timestamp | undefined;
1874
2834
 
1875
2835
  /**
2836
+ * Exclusive end of the settled observation window (UTC).
2837
+ *
1876
2838
  * @generated from field: google.protobuf.Timestamp window_end = 5;
1877
2839
  */
1878
2840
  windowEnd?: Timestamp | undefined;
1879
2841
 
1880
2842
  /**
2843
+ * Server-side Cloud Run metrics for declared targets over this case's
2844
+ * observation window. Empty when the suite declares no Cloud Run targets.
2845
+ *
1881
2846
  * @generated from field: repeated alis.evals.v1.CloudRunTargetSnapshot cloud_run = 6;
1882
2847
  */
1883
2848
  cloudRun: CloudRunTargetSnapshot[];
1884
2849
 
1885
2850
  /**
2851
+ * Server-side Spanner metrics for declared targets over this case's
2852
+ * observation window. Empty when the suite declares no Spanner targets.
2853
+ *
1886
2854
  * @generated from field: repeated alis.evals.v1.SpannerTargetSnapshot spanner = 7;
1887
2855
  */
1888
2856
  spanner: SpannerTargetSnapshot[];
1889
2857
 
1890
2858
  /**
2859
+ * Infrastructure SLO check outcomes. Diagnostics-only in v1: always empty
2860
+ * on the wire until infra SLO evaluators ship.
2861
+ *
1891
2862
  * @generated from field: repeated alis.evals.v1.InfraSloCheck infra_checks = 8;
1892
2863
  */
1893
2864
  infraChecks: InfraSloCheck[];
1894
2865
 
1895
2866
  /**
2867
+ * Developer and framework validation outcomes for this case in declaration
2868
+ * order. Integration cases continue using checks instead.
2869
+ *
1896
2870
  * @generated from field: repeated alis.evals.v1.Validation validations = 9;
1897
2871
  */
1898
2872
  validations: Validation[];
1899
2873
  };
1900
2874
 
1901
2875
  /**
2876
+ * The result of one infrastructure observation case.
2877
+ *
1902
2878
  * @generated from message alis.evals.v1.InfraObservationResults.Case
1903
2879
  */
1904
2880
  export declare type InfraObservationResults_CaseJson = {
1905
2881
  /**
2882
+ * The case id.
2883
+ * Example: peak-hours.observe-search
2884
+ *
1906
2885
  * @generated from field: string id = 1;
1907
2886
  */
1908
2887
  id?: string;
1909
2888
 
1910
2889
  /**
2890
+ * The rolled-up status for the case. Diagnostics-only in v1: PASSED when
2891
+ * observation completes; FAILED only on runner or setup errors, not on
2892
+ * infra threshold breach until infra SLO evaluators ship.
2893
+ *
1911
2894
  * @generated from field: alis.evals.v1.Status status = 2;
1912
2895
  */
1913
2896
  status?: StatusJson;
1914
2897
 
1915
2898
  /**
2899
+ * Resolved lookback duration for this case.
2900
+ *
1916
2901
  * @generated from field: google.protobuf.Duration lookback = 3;
1917
2902
  */
1918
2903
  lookback?: DurationJson;
1919
2904
 
1920
2905
  /**
2906
+ * Inclusive start of the settled observation window (UTC).
2907
+ *
1921
2908
  * @generated from field: google.protobuf.Timestamp window_start = 4;
1922
2909
  */
1923
2910
  windowStart?: TimestampJson;
1924
2911
 
1925
2912
  /**
2913
+ * Exclusive end of the settled observation window (UTC).
2914
+ *
1926
2915
  * @generated from field: google.protobuf.Timestamp window_end = 5;
1927
2916
  */
1928
2917
  windowEnd?: TimestampJson;
1929
2918
 
1930
2919
  /**
2920
+ * Server-side Cloud Run metrics for declared targets over this case's
2921
+ * observation window. Empty when the suite declares no Cloud Run targets.
2922
+ *
1931
2923
  * @generated from field: repeated alis.evals.v1.CloudRunTargetSnapshot cloud_run = 6;
1932
2924
  */
1933
2925
  cloudRun?: CloudRunTargetSnapshotJson[];
1934
2926
 
1935
2927
  /**
2928
+ * Server-side Spanner metrics for declared targets over this case's
2929
+ * observation window. Empty when the suite declares no Spanner targets.
2930
+ *
1936
2931
  * @generated from field: repeated alis.evals.v1.SpannerTargetSnapshot spanner = 7;
1937
2932
  */
1938
2933
  spanner?: SpannerTargetSnapshotJson[];
1939
2934
 
1940
2935
  /**
2936
+ * Infrastructure SLO check outcomes. Diagnostics-only in v1: always empty
2937
+ * on the wire until infra SLO evaluators ship.
2938
+ *
1941
2939
  * @generated from field: repeated alis.evals.v1.InfraSloCheck infra_checks = 8;
1942
2940
  */
1943
2941
  infraChecks?: InfraSloCheckJson[];
1944
2942
 
1945
2943
  /**
2944
+ * Developer and framework validation outcomes for this case in declaration
2945
+ * order. Integration cases continue using checks instead.
2946
+ *
1946
2947
  * @generated from field: repeated alis.evals.v1.Validation validations = 9;
1947
2948
  */
1948
2949
  validations?: ValidationJson[];
@@ -1955,30 +2956,58 @@ export declare type InfraObservationResults_CaseJson = {
1955
2956
  export declare const InfraObservationResults_CaseSchema: GenMessage<InfraObservationResults_Case, {jsonType: InfraObservationResults_CaseJson}>;
1956
2957
 
1957
2958
  /**
2959
+ * The request message for RunIntegrationTest.
2960
+ *
1958
2961
  * @generated from message alis.evals.v1.RunIntegrationTestRequest
1959
2962
  */
1960
2963
  export declare type RunIntegrationTestRequest = Message<"alis.evals.v1.RunIntegrationTestRequest"> & {
1961
2964
  /**
2965
+ * The batch ID to associate with the run.
2966
+ *
1962
2967
  * @generated from field: optional string batch_id = 1;
1963
2968
  */
1964
2969
  batchId?: string | undefined;
1965
2970
 
1966
2971
  /**
2972
+ * The case ids to execute.
2973
+ *
2974
+ * If empty, all cases registered for integration runs are executed.
2975
+ * If non-empty, only registered cases whose id matches an entry are
2976
+ * executed.
2977
+ *
2978
+ * A qualified id `suite.case` selects one case. A suite id `suite`
2979
+ * selects all cases in that suite; shared setup and teardown for the suite
2980
+ * run once per suite.
2981
+ *
1967
2982
  * @generated from field: repeated string case_ids = 2;
1968
2983
  */
1969
2984
  caseIds: string[];
1970
2985
  };
1971
2986
 
1972
2987
  /**
2988
+ * The request message for RunIntegrationTest.
2989
+ *
1973
2990
  * @generated from message alis.evals.v1.RunIntegrationTestRequest
1974
2991
  */
1975
2992
  export declare type RunIntegrationTestRequestJson = {
1976
2993
  /**
2994
+ * The batch ID to associate with the run.
2995
+ *
1977
2996
  * @generated from field: optional string batch_id = 1;
1978
2997
  */
1979
2998
  batchId?: string;
1980
2999
 
1981
3000
  /**
3001
+ * The case ids to execute.
3002
+ *
3003
+ * If empty, all cases registered for integration runs are executed.
3004
+ * If non-empty, only registered cases whose id matches an entry are
3005
+ * executed.
3006
+ *
3007
+ * A qualified id `suite.case` selects one case. A suite id `suite`
3008
+ * selects all cases in that suite; shared setup and teardown for the suite
3009
+ * run once per suite.
3010
+ *
1982
3011
  * @generated from field: repeated string case_ids = 2;
1983
3012
  */
1984
3013
  caseIds?: string[];
@@ -1991,50 +3020,76 @@ export declare type RunIntegrationTestRequestJson = {
1991
3020
  export declare const RunIntegrationTestRequestSchema: GenMessage<RunIntegrationTestRequest, {jsonType: RunIntegrationTestRequestJson}>;
1992
3021
 
1993
3022
  /**
3023
+ * Metadata for a RunIntegrationTest long-running operation while it is in
3024
+ * progress.
3025
+ *
1994
3026
  * @generated from message alis.evals.v1.RunIntegrationTestMetadata
1995
3027
  */
1996
3028
  export declare type RunIntegrationTestMetadata = Message<"alis.evals.v1.RunIntegrationTestMetadata"> & {
1997
3029
  /**
3030
+ * The number of cases selected for this operation after filter resolution.
3031
+ *
1998
3032
  * @generated from field: int32 case_count = 1;
1999
3033
  */
2000
3034
  caseCount: number;
2001
3035
 
2002
3036
  /**
3037
+ * The number of cases finished so far (passed, failed, or errored).
3038
+ * This value increases monotonically during the run.
3039
+ *
2003
3040
  * @generated from field: int32 completed_case_count = 2;
2004
3041
  */
2005
3042
  completedCaseCount: number;
2006
3043
 
2007
3044
  /**
3045
+ * The number of suites selected for this operation.
3046
+ *
2008
3047
  * @generated from field: int32 suite_count = 3;
2009
3048
  */
2010
3049
  suiteCount: number;
2011
3050
 
2012
3051
  /**
3052
+ * The number of suites finished so far. This value increases monotonically
3053
+ * during the run.
3054
+ *
2013
3055
  * @generated from field: int32 completed_suite_count = 4;
2014
3056
  */
2015
3057
  completedSuiteCount: number;
2016
3058
  };
2017
3059
 
2018
3060
  /**
3061
+ * Metadata for a RunIntegrationTest long-running operation while it is in
3062
+ * progress.
3063
+ *
2019
3064
  * @generated from message alis.evals.v1.RunIntegrationTestMetadata
2020
3065
  */
2021
3066
  export declare type RunIntegrationTestMetadataJson = {
2022
3067
  /**
3068
+ * The number of cases selected for this operation after filter resolution.
3069
+ *
2023
3070
  * @generated from field: int32 case_count = 1;
2024
3071
  */
2025
3072
  caseCount?: number;
2026
3073
 
2027
3074
  /**
3075
+ * The number of cases finished so far (passed, failed, or errored).
3076
+ * This value increases monotonically during the run.
3077
+ *
2028
3078
  * @generated from field: int32 completed_case_count = 2;
2029
3079
  */
2030
3080
  completedCaseCount?: number;
2031
3081
 
2032
3082
  /**
3083
+ * The number of suites selected for this operation.
3084
+ *
2033
3085
  * @generated from field: int32 suite_count = 3;
2034
3086
  */
2035
3087
  suiteCount?: number;
2036
3088
 
2037
3089
  /**
3090
+ * The number of suites finished so far. This value increases monotonically
3091
+ * during the run.
3092
+ *
2038
3093
  * @generated from field: int32 completed_suite_count = 4;
2039
3094
  */
2040
3095
  completedSuiteCount?: number;
@@ -2047,20 +3102,36 @@ export declare type RunIntegrationTestMetadataJson = {
2047
3102
  export declare const RunIntegrationTestMetadataSchema: GenMessage<RunIntegrationTestMetadata, {jsonType: RunIntegrationTestMetadataJson}>;
2048
3103
 
2049
3104
  /**
3105
+ * The response message for RunIntegrationTest, stored in the completed
3106
+ * Operation.
3107
+ *
2050
3108
  * @generated from message alis.evals.v1.RunIntegrationTestResponse
2051
3109
  */
2052
3110
  export declare type RunIntegrationTestResponse = Message<"alis.evals.v1.RunIntegrationTestResponse"> & {
2053
3111
  /**
3112
+ * The resource names of all runs produced by this operation.
3113
+ *
3114
+ * Each entry has the format `runs/{run_id}`. Single-suite jobs return a
3115
+ * list of length one.
3116
+ *
2054
3117
  * @generated from field: repeated string runs = 1;
2055
3118
  */
2056
3119
  runs: string[];
2057
3120
  };
2058
3121
 
2059
3122
  /**
3123
+ * The response message for RunIntegrationTest, stored in the completed
3124
+ * Operation.
3125
+ *
2060
3126
  * @generated from message alis.evals.v1.RunIntegrationTestResponse
2061
3127
  */
2062
3128
  export declare type RunIntegrationTestResponseJson = {
2063
3129
  /**
3130
+ * The resource names of all runs produced by this operation.
3131
+ *
3132
+ * Each entry has the format `runs/{run_id}`. Single-suite jobs return a
3133
+ * list of length one.
3134
+ *
2064
3135
  * @generated from field: repeated string runs = 1;
2065
3136
  */
2066
3137
  runs?: string[];
@@ -2073,40 +3144,78 @@ export declare type RunIntegrationTestResponseJson = {
2073
3144
  export declare const RunIntegrationTestResponseSchema: GenMessage<RunIntegrationTestResponse, {jsonType: RunIntegrationTestResponseJson}>;
2074
3145
 
2075
3146
  /**
3147
+ * The request message for RunLoadTest.
3148
+ *
2076
3149
  * @generated from message alis.evals.v1.RunLoadTestRequest
2077
3150
  */
2078
3151
  export declare type RunLoadTestRequest = Message<"alis.evals.v1.RunLoadTestRequest"> & {
2079
3152
  /**
3153
+ * The load intensity preset. Controls concurrency, request rate, and
3154
+ * duration according to server-side configuration; suites may override the
3155
+ * preset's parameters at registration. The resolved values are recorded in
3156
+ * each case's Summary.
3157
+ *
2080
3158
  * @generated from field: alis.evals.v1.RunLoadTestRequest.Mode mode = 1;
2081
3159
  */
2082
3160
  mode: RunLoadTestRequest_Mode;
2083
3161
 
2084
3162
  /**
3163
+ * The load case ids to execute.
3164
+ *
3165
+ * If empty, all cases registered for load tests are executed.
3166
+ * If non-empty, only registered cases whose id matches an entry are
3167
+ * executed.
3168
+ *
3169
+ * A qualified id `suite.case` selects one case. A suite id `suite`
3170
+ * selects all cases in that suite; shared setup and teardown for the suite
3171
+ * run once per suite.
3172
+ *
2085
3173
  * @generated from field: repeated string case_ids = 2;
2086
3174
  */
2087
3175
  caseIds: string[];
2088
3176
 
2089
3177
  /**
3178
+ * The batch ID to associate with the runs produced by this operation.
3179
+ *
2090
3180
  * @generated from field: optional string batch_id = 3;
2091
3181
  */
2092
3182
  batchId?: string | undefined;
2093
3183
  };
2094
3184
 
2095
3185
  /**
3186
+ * The request message for RunLoadTest.
3187
+ *
2096
3188
  * @generated from message alis.evals.v1.RunLoadTestRequest
2097
3189
  */
2098
3190
  export declare type RunLoadTestRequestJson = {
2099
3191
  /**
3192
+ * The load intensity preset. Controls concurrency, request rate, and
3193
+ * duration according to server-side configuration; suites may override the
3194
+ * preset's parameters at registration. The resolved values are recorded in
3195
+ * each case's Summary.
3196
+ *
2100
3197
  * @generated from field: alis.evals.v1.RunLoadTestRequest.Mode mode = 1;
2101
3198
  */
2102
3199
  mode?: RunLoadTestRequest_ModeJson;
2103
3200
 
2104
3201
  /**
3202
+ * The load case ids to execute.
3203
+ *
3204
+ * If empty, all cases registered for load tests are executed.
3205
+ * If non-empty, only registered cases whose id matches an entry are
3206
+ * executed.
3207
+ *
3208
+ * A qualified id `suite.case` selects one case. A suite id `suite`
3209
+ * selects all cases in that suite; shared setup and teardown for the suite
3210
+ * run once per suite.
3211
+ *
2105
3212
  * @generated from field: repeated string case_ids = 2;
2106
3213
  */
2107
3214
  caseIds?: string[];
2108
3215
 
2109
3216
  /**
3217
+ * The batch ID to associate with the runs produced by this operation.
3218
+ *
2110
3219
  * @generated from field: optional string batch_id = 3;
2111
3220
  */
2112
3221
  batchId?: string;
@@ -2119,41 +3228,62 @@ export declare type RunLoadTestRequestJson = {
2119
3228
  export declare const RunLoadTestRequestSchema: GenMessage<RunLoadTestRequest, {jsonType: RunLoadTestRequestJson}>;
2120
3229
 
2121
3230
  /**
3231
+ * Load intensity presets, ordered as an ascending scale from minimal to
3232
+ * ludicrous. Higher modes apply more concurrent load; the exact parameters
3233
+ * are environment-specific.
3234
+ *
2122
3235
  * @generated from enum alis.evals.v1.RunLoadTestRequest.Mode
2123
3236
  */
2124
3237
  export enum RunLoadTestRequest_Mode {
2125
3238
  /**
3239
+ * Default value. The server rejects this value if sent by clients.
3240
+ *
2126
3241
  * @generated from enum value: MODE_UNSPECIFIED = 0;
2127
3242
  */
2128
3243
  MODE_UNSPECIFIED = 0,
2129
3244
 
2130
3245
  /**
3246
+ * Minimal load for smoke or baseline measurement.
3247
+ *
2131
3248
  * @generated from enum value: MINIMAL = 1;
2132
3249
  */
2133
3250
  MINIMAL = 1,
2134
3251
 
2135
3252
  /**
3253
+ * Conservative load with headroom for shared infrastructure.
3254
+ *
2136
3255
  * @generated from enum value: CONSERVATIVE = 2;
2137
3256
  */
2138
3257
  CONSERVATIVE = 2,
2139
3258
 
2140
3259
  /**
3260
+ * Moderate load suitable for routine capacity checks.
3261
+ *
2141
3262
  * @generated from enum value: MODERATE = 3;
2142
3263
  */
2143
3264
  MODERATE = 3,
2144
3265
 
2145
3266
  /**
3267
+ * Strong load short of ludicrous; useful for capacity ceiling probes.
3268
+ *
2146
3269
  * @generated from enum value: HIGH = 4;
2147
3270
  */
2148
3271
  HIGH = 4,
2149
3272
 
2150
3273
  /**
3274
+ * Highest load the target environment is expected to sustain without
3275
+ * degrading shared infrastructure.
3276
+ *
2151
3277
  * @generated from enum value: LUDICROUS = 5;
2152
3278
  */
2153
3279
  LUDICROUS = 5,
2154
3280
  }
2155
3281
 
2156
3282
  /**
3283
+ * Load intensity presets, ordered as an ascending scale from minimal to
3284
+ * ludicrous. Higher modes apply more concurrent load; the exact parameters
3285
+ * are environment-specific.
3286
+ *
2157
3287
  * @generated from enum alis.evals.v1.RunLoadTestRequest.Mode
2158
3288
  */
2159
3289
  export declare type RunLoadTestRequest_ModeJson = "MODE_UNSPECIFIED" | "MINIMAL" | "CONSERVATIVE" | "MODERATE" | "HIGH" | "LUDICROUS";
@@ -2164,50 +3294,76 @@ export declare type RunLoadTestRequest_ModeJson = "MODE_UNSPECIFIED" | "MINIMAL"
2164
3294
  export declare const RunLoadTestRequest_ModeSchema: GenEnum<RunLoadTestRequest_Mode, RunLoadTestRequest_ModeJson>;
2165
3295
 
2166
3296
  /**
3297
+ * Metadata for a RunLoadTest long-running operation while it is in progress.
3298
+ *
2167
3299
  * @generated from message alis.evals.v1.RunLoadTestMetadata
2168
3300
  */
2169
3301
  export declare type RunLoadTestMetadata = Message<"alis.evals.v1.RunLoadTestMetadata"> & {
2170
3302
  /**
3303
+ * The number of load cases selected for this operation after filter
3304
+ * resolution. Each case runs one load window against its target.
3305
+ *
2171
3306
  * @generated from field: int32 case_count = 1;
2172
3307
  */
2173
3308
  caseCount: number;
2174
3309
 
2175
3310
  /**
3311
+ * The number of cases whose load window has finished so far. This value
3312
+ * increases monotonically during the run.
3313
+ *
2176
3314
  * @generated from field: int32 completed_case_count = 2;
2177
3315
  */
2178
3316
  completedCaseCount: number;
2179
3317
 
2180
3318
  /**
3319
+ * The number of suites selected for this operation.
3320
+ *
2181
3321
  * @generated from field: int32 suite_count = 3;
2182
3322
  */
2183
3323
  suiteCount: number;
2184
3324
 
2185
3325
  /**
3326
+ * The number of suites finished so far. This value increases monotonically
3327
+ * during the run.
3328
+ *
2186
3329
  * @generated from field: int32 completed_suite_count = 4;
2187
3330
  */
2188
3331
  completedSuiteCount: number;
2189
3332
  };
2190
3333
 
2191
3334
  /**
3335
+ * Metadata for a RunLoadTest long-running operation while it is in progress.
3336
+ *
2192
3337
  * @generated from message alis.evals.v1.RunLoadTestMetadata
2193
3338
  */
2194
3339
  export declare type RunLoadTestMetadataJson = {
2195
3340
  /**
3341
+ * The number of load cases selected for this operation after filter
3342
+ * resolution. Each case runs one load window against its target.
3343
+ *
2196
3344
  * @generated from field: int32 case_count = 1;
2197
3345
  */
2198
3346
  caseCount?: number;
2199
3347
 
2200
3348
  /**
3349
+ * The number of cases whose load window has finished so far. This value
3350
+ * increases monotonically during the run.
3351
+ *
2201
3352
  * @generated from field: int32 completed_case_count = 2;
2202
3353
  */
2203
3354
  completedCaseCount?: number;
2204
3355
 
2205
3356
  /**
3357
+ * The number of suites selected for this operation.
3358
+ *
2206
3359
  * @generated from field: int32 suite_count = 3;
2207
3360
  */
2208
3361
  suiteCount?: number;
2209
3362
 
2210
3363
  /**
3364
+ * The number of suites finished so far. This value increases monotonically
3365
+ * during the run.
3366
+ *
2211
3367
  * @generated from field: int32 completed_suite_count = 4;
2212
3368
  */
2213
3369
  completedSuiteCount?: number;
@@ -2220,20 +3376,32 @@ export declare type RunLoadTestMetadataJson = {
2220
3376
  export declare const RunLoadTestMetadataSchema: GenMessage<RunLoadTestMetadata, {jsonType: RunLoadTestMetadataJson}>;
2221
3377
 
2222
3378
  /**
3379
+ * The response message for RunLoadTest, stored in the completed Operation.
3380
+ *
2223
3381
  * @generated from message alis.evals.v1.RunLoadTestResponse
2224
3382
  */
2225
3383
  export declare type RunLoadTestResponse = Message<"alis.evals.v1.RunLoadTestResponse"> & {
2226
3384
  /**
3385
+ * The resource names of all runs produced by this operation.
3386
+ *
3387
+ * Each entry has the format `runs/{run_id}`.
3388
+ *
2227
3389
  * @generated from field: repeated string runs = 1;
2228
3390
  */
2229
3391
  runs: string[];
2230
3392
  };
2231
3393
 
2232
3394
  /**
3395
+ * The response message for RunLoadTest, stored in the completed Operation.
3396
+ *
2233
3397
  * @generated from message alis.evals.v1.RunLoadTestResponse
2234
3398
  */
2235
3399
  export declare type RunLoadTestResponseJson = {
2236
3400
  /**
3401
+ * The resource names of all runs produced by this operation.
3402
+ *
3403
+ * Each entry has the format `runs/{run_id}`.
3404
+ *
2237
3405
  * @generated from field: repeated string runs = 1;
2238
3406
  */
2239
3407
  runs?: string[];
@@ -2246,30 +3414,74 @@ export declare type RunLoadTestResponseJson = {
2246
3414
  export declare const RunLoadTestResponseSchema: GenMessage<RunLoadTestResponse, {jsonType: RunLoadTestResponseJson}>;
2247
3415
 
2248
3416
  /**
3417
+ * The request message for RunAgentEval.
3418
+ *
2249
3419
  * @generated from message alis.evals.v1.RunAgentEvalRequest
2250
3420
  */
2251
3421
  export declare type RunAgentEvalRequest = Message<"alis.evals.v1.RunAgentEvalRequest"> & {
2252
3422
  /**
3423
+ * The case ids to execute.
3424
+ *
3425
+ * If empty, all cases registered for agent evaluation are executed.
3426
+ * If non-empty, only registered cases whose id matches an entry are
3427
+ * executed.
3428
+ *
3429
+ * A qualified id `suite.case` selects one case. A suite id `suite`
3430
+ * selects all cases in that suite; shared setup and teardown for the suite
3431
+ * run once per suite.
3432
+ *
2253
3433
  * @generated from field: repeated string case_ids = 1;
2254
3434
  */
2255
3435
  caseIds: string[];
2256
3436
 
2257
3437
  /**
3438
+ * Initial ADK session state applied to every case in this run.
3439
+ *
3440
+ * Mirrors ADK SessionInput.state: a JSON object of arbitrary values
3441
+ * (strings, numbers, bools, lists, nested objects). Agents interpret
3442
+ * keys according to their own session-state contracts.
3443
+ *
3444
+ * Unset or empty means no run-level state is supplied. When set, these
3445
+ * entries are merged into each case's session input before inference.
3446
+ * A key already set on the case wins over the same key here.
3447
+ *
2258
3448
  * @generated from field: google.protobuf.Struct session_state = 2;
2259
3449
  */
2260
3450
  sessionState?: JsonObject | undefined;
2261
3451
  };
2262
3452
 
2263
3453
  /**
3454
+ * The request message for RunAgentEval.
3455
+ *
2264
3456
  * @generated from message alis.evals.v1.RunAgentEvalRequest
2265
3457
  */
2266
3458
  export declare type RunAgentEvalRequestJson = {
2267
3459
  /**
3460
+ * The case ids to execute.
3461
+ *
3462
+ * If empty, all cases registered for agent evaluation are executed.
3463
+ * If non-empty, only registered cases whose id matches an entry are
3464
+ * executed.
3465
+ *
3466
+ * A qualified id `suite.case` selects one case. A suite id `suite`
3467
+ * selects all cases in that suite; shared setup and teardown for the suite
3468
+ * run once per suite.
3469
+ *
2268
3470
  * @generated from field: repeated string case_ids = 1;
2269
3471
  */
2270
3472
  caseIds?: string[];
2271
3473
 
2272
3474
  /**
3475
+ * Initial ADK session state applied to every case in this run.
3476
+ *
3477
+ * Mirrors ADK SessionInput.state: a JSON object of arbitrary values
3478
+ * (strings, numbers, bools, lists, nested objects). Agents interpret
3479
+ * keys according to their own session-state contracts.
3480
+ *
3481
+ * Unset or empty means no run-level state is supplied. When set, these
3482
+ * entries are merged into each case's session input before inference.
3483
+ * A key already set on the case wins over the same key here.
3484
+ *
2273
3485
  * @generated from field: google.protobuf.Struct session_state = 2;
2274
3486
  */
2275
3487
  sessionState?: StructJson;
@@ -2282,50 +3494,74 @@ export declare type RunAgentEvalRequestJson = {
2282
3494
  export declare const RunAgentEvalRequestSchema: GenMessage<RunAgentEvalRequest, {jsonType: RunAgentEvalRequestJson}>;
2283
3495
 
2284
3496
  /**
3497
+ * Metadata for a RunAgentEval long-running operation while it is in progress.
3498
+ *
2285
3499
  * @generated from message alis.evals.v1.RunAgentEvalMetadata
2286
3500
  */
2287
3501
  export declare type RunAgentEvalMetadata = Message<"alis.evals.v1.RunAgentEvalMetadata"> & {
2288
3502
  /**
3503
+ * The number of cases selected for this operation after filter resolution.
3504
+ *
2289
3505
  * @generated from field: int32 case_count = 1;
2290
3506
  */
2291
3507
  caseCount: number;
2292
3508
 
2293
3509
  /**
3510
+ * The number of cases finished so far (passed, failed, or errored).
3511
+ * This value increases monotonically during the run.
3512
+ *
2294
3513
  * @generated from field: int32 completed_case_count = 2;
2295
3514
  */
2296
3515
  completedCaseCount: number;
2297
3516
 
2298
3517
  /**
3518
+ * The number of suites selected for this operation.
3519
+ *
2299
3520
  * @generated from field: int32 suite_count = 3;
2300
3521
  */
2301
3522
  suiteCount: number;
2302
3523
 
2303
3524
  /**
3525
+ * The number of suites finished so far. This value increases monotonically
3526
+ * during the run.
3527
+ *
2304
3528
  * @generated from field: int32 completed_suite_count = 4;
2305
3529
  */
2306
3530
  completedSuiteCount: number;
2307
3531
  };
2308
3532
 
2309
3533
  /**
3534
+ * Metadata for a RunAgentEval long-running operation while it is in progress.
3535
+ *
2310
3536
  * @generated from message alis.evals.v1.RunAgentEvalMetadata
2311
3537
  */
2312
3538
  export declare type RunAgentEvalMetadataJson = {
2313
3539
  /**
3540
+ * The number of cases selected for this operation after filter resolution.
3541
+ *
2314
3542
  * @generated from field: int32 case_count = 1;
2315
3543
  */
2316
3544
  caseCount?: number;
2317
3545
 
2318
3546
  /**
3547
+ * The number of cases finished so far (passed, failed, or errored).
3548
+ * This value increases monotonically during the run.
3549
+ *
2319
3550
  * @generated from field: int32 completed_case_count = 2;
2320
3551
  */
2321
3552
  completedCaseCount?: number;
2322
3553
 
2323
3554
  /**
3555
+ * The number of suites selected for this operation.
3556
+ *
2324
3557
  * @generated from field: int32 suite_count = 3;
2325
3558
  */
2326
3559
  suiteCount?: number;
2327
3560
 
2328
3561
  /**
3562
+ * The number of suites finished so far. This value increases monotonically
3563
+ * during the run.
3564
+ *
2329
3565
  * @generated from field: int32 completed_suite_count = 4;
2330
3566
  */
2331
3567
  completedSuiteCount?: number;
@@ -2338,20 +3574,32 @@ export declare type RunAgentEvalMetadataJson = {
2338
3574
  export declare const RunAgentEvalMetadataSchema: GenMessage<RunAgentEvalMetadata, {jsonType: RunAgentEvalMetadataJson}>;
2339
3575
 
2340
3576
  /**
3577
+ * The response message for RunAgentEval, stored in the completed Operation.
3578
+ *
2341
3579
  * @generated from message alis.evals.v1.RunAgentEvalResponse
2342
3580
  */
2343
3581
  export declare type RunAgentEvalResponse = Message<"alis.evals.v1.RunAgentEvalResponse"> & {
2344
3582
  /**
3583
+ * The resource names of all runs produced by this operation.
3584
+ *
3585
+ * Each entry has the format `runs/{run_id}`.
3586
+ *
2345
3587
  * @generated from field: repeated string runs = 1;
2346
3588
  */
2347
3589
  runs: string[];
2348
3590
  };
2349
3591
 
2350
3592
  /**
3593
+ * The response message for RunAgentEval, stored in the completed Operation.
3594
+ *
2351
3595
  * @generated from message alis.evals.v1.RunAgentEvalResponse
2352
3596
  */
2353
3597
  export declare type RunAgentEvalResponseJson = {
2354
3598
  /**
3599
+ * The resource names of all runs produced by this operation.
3600
+ *
3601
+ * Each entry has the format `runs/{run_id}`.
3602
+ *
2355
3603
  * @generated from field: repeated string runs = 1;
2356
3604
  */
2357
3605
  runs?: string[];
@@ -2364,40 +3612,74 @@ export declare type RunAgentEvalResponseJson = {
2364
3612
  export declare const RunAgentEvalResponseSchema: GenMessage<RunAgentEvalResponse, {jsonType: RunAgentEvalResponseJson}>;
2365
3613
 
2366
3614
  /**
3615
+ * The request message for RunInfraObservation.
3616
+ *
2367
3617
  * @generated from message alis.evals.v1.RunInfraObservationRequest
2368
3618
  */
2369
3619
  export declare type RunInfraObservationRequest = Message<"alis.evals.v1.RunInfraObservationRequest"> & {
2370
3620
  /**
3621
+ * The infra observation case ids to execute.
3622
+ *
3623
+ * If empty, all cases registered for infra observation are executed.
3624
+ * If non-empty, only registered cases whose id matches an entry are
3625
+ * executed.
3626
+ *
3627
+ * A qualified id `suite.case` selects one case. A suite id `suite`
3628
+ * selects all cases in that suite; shared setup and teardown for the suite
3629
+ * run once per suite.
3630
+ *
2371
3631
  * @generated from field: repeated string case_ids = 1;
2372
3632
  */
2373
3633
  caseIds: string[];
2374
3634
 
2375
3635
  /**
3636
+ * The batch ID to associate with the runs produced by this operation.
3637
+ *
2376
3638
  * @generated from field: optional string batch_id = 2;
2377
3639
  */
2378
3640
  batchId?: string | undefined;
2379
3641
 
2380
3642
  /**
3643
+ * Highest-precedence lookback override when set. When unset, the server
3644
+ * resolves lookback from per-case and suite defaults.
3645
+ *
2381
3646
  * @generated from field: optional google.protobuf.Duration lookback = 3;
2382
3647
  */
2383
3648
  lookback?: Duration | undefined;
2384
3649
  };
2385
3650
 
2386
3651
  /**
3652
+ * The request message for RunInfraObservation.
3653
+ *
2387
3654
  * @generated from message alis.evals.v1.RunInfraObservationRequest
2388
3655
  */
2389
3656
  export declare type RunInfraObservationRequestJson = {
2390
3657
  /**
3658
+ * The infra observation case ids to execute.
3659
+ *
3660
+ * If empty, all cases registered for infra observation are executed.
3661
+ * If non-empty, only registered cases whose id matches an entry are
3662
+ * executed.
3663
+ *
3664
+ * A qualified id `suite.case` selects one case. A suite id `suite`
3665
+ * selects all cases in that suite; shared setup and teardown for the suite
3666
+ * run once per suite.
3667
+ *
2391
3668
  * @generated from field: repeated string case_ids = 1;
2392
3669
  */
2393
3670
  caseIds?: string[];
2394
3671
 
2395
3672
  /**
3673
+ * The batch ID to associate with the runs produced by this operation.
3674
+ *
2396
3675
  * @generated from field: optional string batch_id = 2;
2397
3676
  */
2398
3677
  batchId?: string;
2399
3678
 
2400
3679
  /**
3680
+ * Highest-precedence lookback override when set. When unset, the server
3681
+ * resolves lookback from per-case and suite defaults.
3682
+ *
2401
3683
  * @generated from field: optional google.protobuf.Duration lookback = 3;
2402
3684
  */
2403
3685
  lookback?: DurationJson;
@@ -2410,50 +3692,76 @@ export declare type RunInfraObservationRequestJson = {
2410
3692
  export declare const RunInfraObservationRequestSchema: GenMessage<RunInfraObservationRequest, {jsonType: RunInfraObservationRequestJson}>;
2411
3693
 
2412
3694
  /**
3695
+ * Metadata for a RunInfraObservation long-running operation while it is in
3696
+ * progress.
3697
+ *
2413
3698
  * @generated from message alis.evals.v1.RunInfraObservationMetadata
2414
3699
  */
2415
3700
  export declare type RunInfraObservationMetadata = Message<"alis.evals.v1.RunInfraObservationMetadata"> & {
2416
3701
  /**
3702
+ * The number of infra observation cases selected after filter resolution.
3703
+ *
2417
3704
  * @generated from field: int32 case_count = 1;
2418
3705
  */
2419
3706
  caseCount: number;
2420
3707
 
2421
3708
  /**
3709
+ * The number of cases finished so far. This value increases monotonically
3710
+ * during the run.
3711
+ *
2422
3712
  * @generated from field: int32 completed_case_count = 2;
2423
3713
  */
2424
3714
  completedCaseCount: number;
2425
3715
 
2426
3716
  /**
3717
+ * The number of suites selected for this operation.
3718
+ *
2427
3719
  * @generated from field: int32 suite_count = 3;
2428
3720
  */
2429
3721
  suiteCount: number;
2430
3722
 
2431
3723
  /**
3724
+ * The number of suites finished so far. This value increases monotonically
3725
+ * during the run.
3726
+ *
2432
3727
  * @generated from field: int32 completed_suite_count = 4;
2433
3728
  */
2434
3729
  completedSuiteCount: number;
2435
3730
  };
2436
3731
 
2437
3732
  /**
3733
+ * Metadata for a RunInfraObservation long-running operation while it is in
3734
+ * progress.
3735
+ *
2438
3736
  * @generated from message alis.evals.v1.RunInfraObservationMetadata
2439
3737
  */
2440
3738
  export declare type RunInfraObservationMetadataJson = {
2441
3739
  /**
3740
+ * The number of infra observation cases selected after filter resolution.
3741
+ *
2442
3742
  * @generated from field: int32 case_count = 1;
2443
3743
  */
2444
3744
  caseCount?: number;
2445
3745
 
2446
3746
  /**
3747
+ * The number of cases finished so far. This value increases monotonically
3748
+ * during the run.
3749
+ *
2447
3750
  * @generated from field: int32 completed_case_count = 2;
2448
3751
  */
2449
3752
  completedCaseCount?: number;
2450
3753
 
2451
3754
  /**
3755
+ * The number of suites selected for this operation.
3756
+ *
2452
3757
  * @generated from field: int32 suite_count = 3;
2453
3758
  */
2454
3759
  suiteCount?: number;
2455
3760
 
2456
3761
  /**
3762
+ * The number of suites finished so far. This value increases monotonically
3763
+ * during the run.
3764
+ *
2457
3765
  * @generated from field: int32 completed_suite_count = 4;
2458
3766
  */
2459
3767
  completedSuiteCount?: number;
@@ -2466,20 +3774,34 @@ export declare type RunInfraObservationMetadataJson = {
2466
3774
  export declare const RunInfraObservationMetadataSchema: GenMessage<RunInfraObservationMetadata, {jsonType: RunInfraObservationMetadataJson}>;
2467
3775
 
2468
3776
  /**
3777
+ * The response message for RunInfraObservation, stored in the completed
3778
+ * Operation.
3779
+ *
2469
3780
  * @generated from message alis.evals.v1.RunInfraObservationResponse
2470
3781
  */
2471
3782
  export declare type RunInfraObservationResponse = Message<"alis.evals.v1.RunInfraObservationResponse"> & {
2472
3783
  /**
3784
+ * The resource names of all runs produced by this operation.
3785
+ *
3786
+ * Each entry has the format `runs/{run_id}`.
3787
+ *
2473
3788
  * @generated from field: repeated string runs = 1;
2474
3789
  */
2475
3790
  runs: string[];
2476
3791
  };
2477
3792
 
2478
3793
  /**
3794
+ * The response message for RunInfraObservation, stored in the completed
3795
+ * Operation.
3796
+ *
2479
3797
  * @generated from message alis.evals.v1.RunInfraObservationResponse
2480
3798
  */
2481
3799
  export declare type RunInfraObservationResponseJson = {
2482
3800
  /**
3801
+ * The resource names of all runs produced by this operation.
3802
+ *
3803
+ * Each entry has the format `runs/{run_id}`.
3804
+ *
2483
3805
  * @generated from field: repeated string runs = 1;
2484
3806
  */
2485
3807
  runs?: string[];
@@ -2492,40 +3814,56 @@ export declare type RunInfraObservationResponseJson = {
2492
3814
  export declare const RunInfraObservationResponseSchema: GenMessage<RunInfraObservationResponse, {jsonType: RunInfraObservationResponseJson}>;
2493
3815
 
2494
3816
  /**
3817
+ * General validation outcome for specialized eval case branches.
3818
+ *
2495
3819
  * @generated from message alis.evals.v1.Validation
2496
3820
  */
2497
3821
  export declare type Validation = Message<"alis.evals.v1.Validation"> & {
2498
3822
  /**
3823
+ * Stable validation identifier (for example a rule id or `_evals.case`).
3824
+ *
2499
3825
  * @generated from field: string id = 1;
2500
3826
  */
2501
3827
  id: string;
2502
3828
 
2503
3829
  /**
3830
+ * Validation outcome.
3831
+ *
2504
3832
  * @generated from field: alis.evals.v1.Status status = 2;
2505
3833
  */
2506
3834
  status: Status;
2507
3835
 
2508
3836
  /**
3837
+ * Human-readable detail when status is not PASSED. Empty on pass.
3838
+ *
2509
3839
  * @generated from field: string message = 3;
2510
3840
  */
2511
3841
  message: string;
2512
3842
  };
2513
3843
 
2514
3844
  /**
3845
+ * General validation outcome for specialized eval case branches.
3846
+ *
2515
3847
  * @generated from message alis.evals.v1.Validation
2516
3848
  */
2517
3849
  export declare type ValidationJson = {
2518
3850
  /**
3851
+ * Stable validation identifier (for example a rule id or `_evals.case`).
3852
+ *
2519
3853
  * @generated from field: string id = 1;
2520
3854
  */
2521
3855
  id?: string;
2522
3856
 
2523
3857
  /**
3858
+ * Validation outcome.
3859
+ *
2524
3860
  * @generated from field: alis.evals.v1.Status status = 2;
2525
3861
  */
2526
3862
  status?: StatusJson;
2527
3863
 
2528
3864
  /**
3865
+ * Human-readable detail when status is not PASSED. Empty on pass.
3866
+ *
2529
3867
  * @generated from field: string message = 3;
2530
3868
  */
2531
3869
  message?: string;
@@ -2538,20 +3876,32 @@ export declare type ValidationJson = {
2538
3876
  export declare const ValidationSchema: GenMessage<Validation, {jsonType: ValidationJson}>;
2539
3877
 
2540
3878
  /**
3879
+ * Event published when a completed test or evaluation run is ready for
3880
+ * persistence and downstream consumers (for example BigQuery or Pub/Sub).
3881
+ *
2541
3882
  * @generated from message alis.evals.v1.RunPublishedEvent
2542
3883
  */
2543
3884
  export declare type RunPublishedEvent = Message<"alis.evals.v1.RunPublishedEvent"> & {
2544
3885
  /**
3886
+ * The completed run, including per-case status and check results for the
3887
+ * suite that finished.
3888
+ *
2545
3889
  * @generated from field: alis.evals.v1.Run run = 1;
2546
3890
  */
2547
3891
  run?: Run | undefined;
2548
3892
  };
2549
3893
 
2550
3894
  /**
3895
+ * Event published when a completed test or evaluation run is ready for
3896
+ * persistence and downstream consumers (for example BigQuery or Pub/Sub).
3897
+ *
2551
3898
  * @generated from message alis.evals.v1.RunPublishedEvent
2552
3899
  */
2553
3900
  export declare type RunPublishedEventJson = {
2554
3901
  /**
3902
+ * The completed run, including per-case status and check results for the
3903
+ * suite that finished.
3904
+ *
2555
3905
  * @generated from field: alis.evals.v1.Run run = 1;
2556
3906
  */
2557
3907
  run?: RunJson;
@@ -2564,26 +3914,44 @@ export declare type RunPublishedEventJson = {
2564
3914
  export declare const RunPublishedEventSchema: GenMessage<RunPublishedEvent, {jsonType: RunPublishedEventJson}>;
2565
3915
 
2566
3916
  /**
3917
+ * Role of an infrastructure target within a load or observation case.
3918
+ *
3919
+ * ENTRY marks the service under direct test (for example the Cloud Run service
3920
+ * receiving generated load). DEPENDENCY marks downstream services or data stores
3921
+ * observed for correlation (for example internal Cloud Run hops or Spanner).
3922
+ *
2567
3923
  * @generated from enum alis.evals.v1.InfraTargetRole
2568
3924
  */
2569
3925
  export enum InfraTargetRole {
2570
3926
  /**
3927
+ * Default value. Must not appear on completed results.
3928
+ *
2571
3929
  * @generated from enum value: INFRA_TARGET_ROLE_UNSPECIFIED = 0;
2572
3930
  */
2573
3931
  UNSPECIFIED = 0,
2574
3932
 
2575
3933
  /**
3934
+ * The primary service receiving traffic or acting as the case entrypoint.
3935
+ *
2576
3936
  * @generated from enum value: INFRA_TARGET_ROLE_ENTRY = 1;
2577
3937
  */
2578
3938
  ENTRY = 1,
2579
3939
 
2580
3940
  /**
3941
+ * A downstream service or datastore observed alongside the entry target.
3942
+ *
2581
3943
  * @generated from enum value: INFRA_TARGET_ROLE_DEPENDENCY = 2;
2582
3944
  */
2583
3945
  DEPENDENCY = 2,
2584
3946
  }
2585
3947
 
2586
3948
  /**
3949
+ * Role of an infrastructure target within a load or observation case.
3950
+ *
3951
+ * ENTRY marks the service under direct test (for example the Cloud Run service
3952
+ * receiving generated load). DEPENDENCY marks downstream services or data stores
3953
+ * observed for correlation (for example internal Cloud Run hops or Spanner).
3954
+ *
2587
3955
  * @generated from enum alis.evals.v1.InfraTargetRole
2588
3956
  */
2589
3957
  export declare type InfraTargetRoleJson = "INFRA_TARGET_ROLE_UNSPECIFIED" | "INFRA_TARGET_ROLE_ENTRY" | "INFRA_TARGET_ROLE_DEPENDENCY";
@@ -2594,36 +3962,58 @@ export declare type InfraTargetRoleJson = "INFRA_TARGET_ROLE_UNSPECIFIED" | "INF
2594
3962
  export declare const InfraTargetRoleSchema: GenEnum<InfraTargetRole, InfraTargetRoleJson>;
2595
3963
 
2596
3964
  /**
3965
+ * Outcome of fetching infrastructure metrics for one target from Cloud
3966
+ * Monitoring.
3967
+ *
3968
+ * Partial per-metric failure may still yield OK when at least one metric
3969
+ * succeeded; details are recorded in fetch_message on the snapshot.
3970
+ *
2597
3971
  * @generated from enum alis.evals.v1.InfraFetchStatus
2598
3972
  */
2599
3973
  export enum InfraFetchStatus {
2600
3974
  /**
3975
+ * Default value. Must not appear on completed results.
3976
+ *
2601
3977
  * @generated from enum value: INFRA_FETCH_STATUS_UNSPECIFIED = 0;
2602
3978
  */
2603
3979
  UNSPECIFIED = 0,
2604
3980
 
2605
3981
  /**
3982
+ * At least one metric was fetched successfully for this target.
3983
+ *
2606
3984
  * @generated from enum value: INFRA_FETCH_STATUS_OK = 1;
2607
3985
  */
2608
3986
  OK = 1,
2609
3987
 
2610
3988
  /**
3989
+ * The Monitoring API was unreachable or returned no usable data.
3990
+ *
2611
3991
  * @generated from enum value: INFRA_FETCH_STATUS_UNAVAILABLE = 2;
2612
3992
  */
2613
3993
  UNAVAILABLE = 2,
2614
3994
 
2615
3995
  /**
3996
+ * The fetch exceeded its per-target timeout.
3997
+ *
2616
3998
  * @generated from enum value: INFRA_FETCH_STATUS_TIMEOUT = 3;
2617
3999
  */
2618
4000
  TIMEOUT = 3,
2619
4001
 
2620
4002
  /**
4003
+ * The caller lacked permission to read metrics for this target.
4004
+ *
2621
4005
  * @generated from enum value: INFRA_FETCH_STATUS_PERMISSION_DENIED = 4;
2622
4006
  */
2623
4007
  PERMISSION_DENIED = 4,
2624
4008
  }
2625
4009
 
2626
4010
  /**
4011
+ * Outcome of fetching infrastructure metrics for one target from Cloud
4012
+ * Monitoring.
4013
+ *
4014
+ * Partial per-metric failure may still yield OK when at least one metric
4015
+ * succeeded; details are recorded in fetch_message on the snapshot.
4016
+ *
2627
4017
  * @generated from enum alis.evals.v1.InfraFetchStatus
2628
4018
  */
2629
4019
  export declare type InfraFetchStatusJson = "INFRA_FETCH_STATUS_UNSPECIFIED" | "INFRA_FETCH_STATUS_OK" | "INFRA_FETCH_STATUS_UNAVAILABLE" | "INFRA_FETCH_STATUS_TIMEOUT" | "INFRA_FETCH_STATUS_PERMISSION_DENIED";
@@ -2634,26 +4024,36 @@ export declare type InfraFetchStatusJson = "INFRA_FETCH_STATUS_UNSPECIFIED" | "I
2634
4024
  export declare const InfraFetchStatusSchema: GenEnum<InfraFetchStatus, InfraFetchStatusJson>;
2635
4025
 
2636
4026
  /**
4027
+ * Infrastructure resource kind referenced by infra SLO checks and snapshots.
4028
+ *
2637
4029
  * @generated from enum alis.evals.v1.InfraKind
2638
4030
  */
2639
4031
  export enum InfraKind {
2640
4032
  /**
4033
+ * Default value. Must not appear on completed results.
4034
+ *
2641
4035
  * @generated from enum value: INFRA_KIND_UNSPECIFIED = 0;
2642
4036
  */
2643
4037
  UNSPECIFIED = 0,
2644
4038
 
2645
4039
  /**
4040
+ * A Google Cloud Run service revision scope.
4041
+ *
2646
4042
  * @generated from enum value: INFRA_KIND_CLOUD_RUN = 1;
2647
4043
  */
2648
4044
  CLOUD_RUN = 1,
2649
4045
 
2650
4046
  /**
4047
+ * A Cloud Spanner instance and database scope.
4048
+ *
2651
4049
  * @generated from enum value: INFRA_KIND_SPANNER = 2;
2652
4050
  */
2653
4051
  SPANNER = 2,
2654
4052
  }
2655
4053
 
2656
4054
  /**
4055
+ * Infrastructure resource kind referenced by infra SLO checks and snapshots.
4056
+ *
2657
4057
  * @generated from enum alis.evals.v1.InfraKind
2658
4058
  */
2659
4059
  export declare type InfraKindJson = "INFRA_KIND_UNSPECIFIED" | "INFRA_KIND_CLOUD_RUN" | "INFRA_KIND_SPANNER";
@@ -2664,31 +4064,44 @@ export declare type InfraKindJson = "INFRA_KIND_UNSPECIFIED" | "INFRA_KIND_CLOUD
2664
4064
  export declare const InfraKindSchema: GenEnum<InfraKind, InfraKindJson>;
2665
4065
 
2666
4066
  /**
4067
+ * Outcome status for runs, cases, checks, metrics, and rubric dimensions.
4068
+ *
2667
4069
  * @generated from enum alis.evals.v1.Status
2668
4070
  */
2669
4071
  export enum Status {
2670
4072
  /**
4073
+ * Default value. Must not appear on completed results.
4074
+ *
2671
4075
  * @generated from enum value: STATUS_UNSPECIFIED = 0;
2672
4076
  */
2673
4077
  STATUS_UNSPECIFIED = 0,
2674
4078
 
2675
4079
  /**
4080
+ * All applicable checks passed and execution completed.
4081
+ *
2676
4082
  * @generated from enum value: PASSED = 1;
2677
4083
  */
2678
4084
  PASSED = 1,
2679
4085
 
2680
4086
  /**
4087
+ * Execution completed but at least one check failed or scored below threshold.
4088
+ *
2681
4089
  * @generated from enum value: FAILED = 2;
2682
4090
  */
2683
4091
  FAILED = 2,
2684
4092
 
2685
4093
  /**
4094
+ * The item was skipped, non-gating, monitoring-only, or preconditions were
4095
+ * not met. Does not imply pass or fail.
4096
+ *
2686
4097
  * @generated from enum value: NOT_EVALUATED = 3;
2687
4098
  */
2688
4099
  NOT_EVALUATED = 3,
2689
4100
  }
2690
4101
 
2691
4102
  /**
4103
+ * Outcome status for runs, cases, checks, metrics, and rubric dimensions.
4104
+ *
2692
4105
  * @generated from enum alis.evals.v1.Status
2693
4106
  */
2694
4107
  export declare type StatusJson = "STATUS_UNSPECIFIED" | "PASSED" | "FAILED" | "NOT_EVALUATED";
@@ -2699,10 +4112,25 @@ export declare type StatusJson = "STATUS_UNSPECIFIED" | "PASSED" | "FAILED" | "N
2699
4112
  export declare const StatusSchema: GenEnum<Status, StatusJson>;
2700
4113
 
2701
4114
  /**
4115
+ * Executes post-deploy tests and evaluations. Each RPC starts a long-running
4116
+ * operation that runs bundled cases for one kind of test. One operation may
4117
+ * produce multiple Run resources (one per completed suite). When the operation
4118
+ * completes, the response returns the resource names of all resulting Runs
4119
+ * (`runs/{run_id}`). Each full Run payload is published for persistence and
4120
+ * event consumers (for example BigQuery or Pub/Sub).
4121
+ *
2702
4122
  * @generated from service alis.evals.v1.TestService
2703
4123
  */
2704
4124
  export declare const TestService: GenService<{
2705
4125
  /**
4126
+ * Runs integration-scope cases: one or a few service boundaries, in-process
4127
+ * or sibling gRPC calls, typically fast and suitable for post-deploy smoke.
4128
+ *
4129
+ * Returns a long-running operation. Poll GetOperation or wait on the operation
4130
+ * name until done. The operation completes successfully when execution finishes,
4131
+ * even if individual cases fail; inspect each Run status in the published
4132
+ * payload to determine pass or fail.
4133
+ *
2706
4134
  * @generated from rpc alis.evals.v1.TestService.RunIntegrationTest
2707
4135
  */
2708
4136
  runIntegrationTest: {
@@ -2711,6 +4139,18 @@ export declare const TestService: GenService<{
2711
4139
  output: typeof OperationSchema;
2712
4140
  },
2713
4141
  /**
4142
+ * Runs a load test: generates traffic at the requested intensity against
4143
+ * each selected load case, records per-case performance in
4144
+ * LoadTestResults.cases, and evaluates configured Service Level Objectives
4145
+ * (SLOs). An SLO is a pass/fail bound on a metric such as latency, error
4146
+ * rate, or throughput; every evaluated check appears in the case's checks
4147
+ * with its observed value, limit, and status.
4148
+ *
4149
+ * Returns a long-running operation. Poll GetOperation or wait on the operation
4150
+ * name until done. The operation completes successfully when execution finishes,
4151
+ * even if thresholds are breached; inspect each Run status in the published
4152
+ * payload to determine pass or fail.
4153
+ *
2714
4154
  * @generated from rpc alis.evals.v1.TestService.RunLoadTest
2715
4155
  */
2716
4156
  runLoadTest: {
@@ -2719,6 +4159,15 @@ export declare const TestService: GenService<{
2719
4159
  output: typeof OperationSchema;
2720
4160
  },
2721
4161
  /**
4162
+ * Runs behavioral agent evaluation cases. Each case reports a flat list of
4163
+ * metrics (deterministic checks, LLM-as-judge scores, rubric dimensions, and
4164
+ * ADK prebuilt metrics) with per-metric status, score, and threshold.
4165
+ *
4166
+ * Returns a long-running operation. Poll GetOperation or wait on the operation
4167
+ * name until done. The operation completes successfully when execution finishes,
4168
+ * even if individual cases fail; inspect each Run status in the published
4169
+ * payload to determine pass or fail.
4170
+ *
2722
4171
  * @generated from rpc alis.evals.v1.TestService.RunAgentEval
2723
4172
  */
2724
4173
  runAgentEval: {
@@ -2727,6 +4176,14 @@ export declare const TestService: GenService<{
2727
4176
  output: typeof OperationSchema;
2728
4177
  },
2729
4178
  /**
4179
+ * Observes server-side infrastructure metrics without generating load.
4180
+ * Fetches Cloud Monitoring snapshots for declared infrastructure targets, such as Cloud Run and Spanner
4181
+ * over a configurable lookback window (for example peak-hour observation via an external scheduler).
4182
+ *
4183
+ * Returns a long-running operation. Poll GetOperation or wait on the operation
4184
+ * name until done. The operation completes successfully when execution finishes;
4185
+ * inspect each Run status and snapshot fetch_status in the published payload.
4186
+ *
2730
4187
  * @generated from rpc alis.evals.v1.TestService.RunInfraObservation
2731
4188
  */
2732
4189
  runInfraObservation: {