anpord 0.1.13 → 0.1.16
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/api-context.cjs +18 -0
- package/dist/api-context.d.cts +8 -0
- package/dist/api-context.d.mts +8 -0
- package/dist/api-context.mjs +17 -0
- package/dist/api-mocks-B3qngoqS.cjs +68 -0
- package/dist/api-mocks-BfKw_hXc.mjs +33 -0
- package/dist/api-mocks-Bz3G00ve.d.cts +101 -0
- package/dist/api-mocks-Bz3G00ve.d.mts +101 -0
- package/dist/api-runtime.cjs +4 -0
- package/dist/api-runtime.d.cts +13 -0
- package/dist/api-runtime.d.mts +13 -0
- package/dist/api-runtime.mjs +2 -0
- package/dist/api.cjs +5 -0
- package/dist/api.d.cts +4 -0
- package/dist/api.d.mts +4 -0
- package/dist/api.mjs +2 -0
- package/dist/bin.cjs +8 -8
- package/dist/bin.mjs +6 -6
- package/dist/cli-runtime.cjs +1 -1
- package/dist/cli-runtime.mjs +1 -1
- package/dist/{client-BgRs8JvW.d.mts → client-C8tBcLYs.d.mts} +17 -17
- package/dist/{client-BirFO3jz.d.cts → client-CMkx-LKC.d.cts} +16 -16
- package/dist/{client-C-P2I00z.cjs → client-CfYNoI3L.cjs} +4 -91
- package/dist/{client-BfdjFtlV.mjs → client-YWKJd3f-.mjs} +3 -90
- package/dist/{compiler-uwb1pkdK.cjs → compiler-4FKsNxZ-.cjs} +89 -33
- package/dist/{compiler-D96wI7UP.mjs → compiler-Dsltkkj5.mjs} +77 -21
- package/dist/config.cjs +1 -1
- package/dist/config.d.cts +1 -1
- package/dist/config.d.mts +1 -1
- package/dist/config.mjs +1 -1
- package/dist/define-NZvmQuIv.d.cts +34 -0
- package/dist/define-NZvmQuIv.d.mts +34 -0
- package/dist/{eval-judges-DPPUftbh.d.cts → eval-judges-JiC8cUCE.d.cts} +1 -1
- package/dist/{eval-judges-DPPUftbh.d.mts → eval-judges-JiC8cUCE.d.mts} +1 -1
- package/dist/{eval-validations-BRdoZrTQ.mjs → eval-validations-Bmagd2y7.mjs} +2 -2
- package/dist/{eval-validations-_MlnQ_uv.cjs → eval-validations-CqMUt5ra.cjs} +7 -1
- package/dist/eval.cjs +1 -1
- package/dist/eval.d.cts +42 -21
- package/dist/eval.d.mts +42 -21
- package/dist/eval.mjs +1 -1
- package/dist/{evals-B6aT3xgH.d.cts → evals-B2UuSop0.d.cts} +136 -136
- package/dist/{evals-B6aT3xgH.d.mts → evals-B2UuSop0.d.mts} +136 -136
- package/dist/{types-Codjaq-o.d.mts → evals-api-CfAA3cew.d.mts} +12 -112
- package/dist/{types-ISsteq95.d.cts → evals-api-Clbj_nEN.d.cts} +11 -111
- package/dist/{evals-Dcc_x6jW.mjs → evals-api-DDv4gXU1.mjs} +95 -4
- package/dist/{evals-DouMsrsM.cjs → evals-api-kOuyIqvB.cjs} +103 -72
- package/dist/index.cjs +2 -2
- package/dist/index.d.cts +6 -6
- package/dist/index.d.mts +7 -7
- package/dist/index.mjs +2 -2
- package/dist/mcp-runtime.cjs +1 -1
- package/dist/mcp-runtime.mjs +1 -1
- package/dist/runtime-C0_JPkef.mjs +269 -0
- package/dist/runtime-YVJP5P2r.cjs +292 -0
- package/dist/source.d.cts +1 -1
- package/dist/source.d.mts +1 -1
- package/dist/types-BHokRQpg.d.mts +112 -0
- package/dist/types-UKEi5CCa.d.cts +112 -0
- package/dist/validator-runtime.cjs +6 -1
- package/dist/validator-runtime.d.cts +1 -1
- package/dist/validator-runtime.d.mts +1 -1
- package/dist/validator-runtime.mjs +6 -1
- package/dist/validators.d.cts +1 -1
- package/dist/validators.d.mts +1 -1
- package/package.json +21 -1
|
@@ -79,80 +79,80 @@ declare const EvalJournalEntry: Schema.Union<[Schema.Struct<{
|
|
|
79
79
|
type EvalJournalEntry = typeof EvalJournalEntry.Type;
|
|
80
80
|
declare const EvalTrial: Schema.Struct<{
|
|
81
81
|
validations: Schema.optional<Schema.refine<readonly {
|
|
82
|
-
readonly
|
|
82
|
+
readonly id: string;
|
|
83
|
+
readonly name: string;
|
|
84
|
+
readonly startedAt: number | null;
|
|
85
|
+
readonly status: "passed" | "running" | "failed" | "queued" | "error" | "skipped";
|
|
83
86
|
readonly message: string;
|
|
87
|
+
readonly kind: "command" | "judge" | "code";
|
|
84
88
|
readonly index: number;
|
|
85
|
-
readonly name: string;
|
|
86
|
-
readonly exitCode: number | null;
|
|
87
|
-
readonly output: {
|
|
88
|
-
readonly text: string;
|
|
89
|
-
readonly truncated: boolean;
|
|
90
|
-
readonly format: "text" | "json";
|
|
91
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
92
|
-
};
|
|
93
|
-
readonly input: {
|
|
94
|
-
readonly text: string;
|
|
95
|
-
readonly truncated: boolean;
|
|
96
|
-
readonly format: "text" | "json";
|
|
97
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
98
|
-
};
|
|
99
89
|
readonly error: {
|
|
100
|
-
readonly text: string;
|
|
101
90
|
readonly truncated: boolean;
|
|
91
|
+
readonly text: string;
|
|
102
92
|
readonly format: "text" | "json";
|
|
103
93
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
104
94
|
} | null;
|
|
105
|
-
readonly status: "queued" | "running" | "passed" | "failed" | "error" | "skipped";
|
|
106
|
-
readonly id: string;
|
|
107
|
-
readonly startedAt: number | null;
|
|
108
95
|
readonly durationMs: number | null;
|
|
96
|
+
readonly exitCode: number | null;
|
|
109
97
|
readonly truncated: boolean;
|
|
110
98
|
readonly judgment?: {
|
|
111
|
-
readonly name: string;
|
|
112
99
|
readonly model: string;
|
|
100
|
+
readonly name: string;
|
|
113
101
|
readonly threshold: number;
|
|
102
|
+
readonly reason: string;
|
|
114
103
|
readonly error: string | null;
|
|
115
104
|
readonly durationMs: number;
|
|
116
105
|
readonly evaluator: string;
|
|
117
106
|
readonly score: number | null;
|
|
118
107
|
readonly choice: string | null;
|
|
119
|
-
readonly reason: string;
|
|
120
108
|
} | undefined;
|
|
121
|
-
readonly
|
|
109
|
+
readonly input: {
|
|
110
|
+
readonly truncated: boolean;
|
|
122
111
|
readonly text: string;
|
|
112
|
+
readonly format: "text" | "json";
|
|
113
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
114
|
+
};
|
|
115
|
+
readonly output: {
|
|
116
|
+
readonly truncated: boolean;
|
|
117
|
+
readonly text: string;
|
|
118
|
+
readonly format: "text" | "json";
|
|
119
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
120
|
+
};
|
|
121
|
+
readonly metadata?: {
|
|
123
122
|
readonly truncated: boolean;
|
|
123
|
+
readonly text: string;
|
|
124
124
|
readonly format: "text" | "json";
|
|
125
125
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
126
126
|
} | undefined;
|
|
127
127
|
readonly calls: readonly {
|
|
128
|
+
readonly startedAt: number;
|
|
128
129
|
readonly index: number;
|
|
129
|
-
readonly
|
|
130
|
-
readonly text: string;
|
|
130
|
+
readonly error: {
|
|
131
131
|
readonly truncated: boolean;
|
|
132
|
+
readonly text: string;
|
|
132
133
|
readonly format: "text" | "json";
|
|
133
134
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
134
|
-
};
|
|
135
|
+
} | null;
|
|
136
|
+
readonly durationMs: number | null;
|
|
135
137
|
readonly input: {
|
|
136
|
-
readonly text: string;
|
|
137
138
|
readonly truncated: boolean;
|
|
139
|
+
readonly text: string;
|
|
138
140
|
readonly format: "text" | "json";
|
|
139
141
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
140
142
|
};
|
|
141
|
-
readonly
|
|
142
|
-
readonly text: string;
|
|
143
|
+
readonly output: {
|
|
143
144
|
readonly truncated: boolean;
|
|
145
|
+
readonly text: string;
|
|
144
146
|
readonly format: "text" | "json";
|
|
145
147
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
146
|
-
}
|
|
147
|
-
readonly
|
|
148
|
-
readonly durationMs: number | null;
|
|
149
|
-
readonly method: "answer" | "transcript" | "readText" | "exists" | "exec" | "cli.calls" | "mcp.calls";
|
|
148
|
+
};
|
|
149
|
+
readonly method: "answer" | "transcript" | "readText" | "exists" | "exec" | "cli.calls" | "api.calls" | "api.url" | "mcp.calls";
|
|
150
150
|
}[];
|
|
151
151
|
readonly logs: readonly {
|
|
152
152
|
readonly at: number;
|
|
153
153
|
readonly value: {
|
|
154
|
-
readonly text: string;
|
|
155
154
|
readonly truncated: boolean;
|
|
155
|
+
readonly text: string;
|
|
156
156
|
readonly format: "text" | "json";
|
|
157
157
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
158
158
|
};
|
|
@@ -207,7 +207,7 @@ declare const EvalTrial: Schema.Struct<{
|
|
|
207
207
|
}>>;
|
|
208
208
|
calls: Schema.filter<Schema.Array$<Schema.Struct<{
|
|
209
209
|
index: Schema.refine<number, typeof Schema.NonNegative>;
|
|
210
|
-
method: Schema.Literal<["answer", "transcript", "readText", "exists", "exec", "cli.calls", "mcp.calls"]>;
|
|
210
|
+
method: Schema.Literal<["answer", "transcript", "readText", "exists", "exec", "cli.calls", "api.calls", "api.url", "mcp.calls"]>;
|
|
211
211
|
startedAt: typeof Schema.Number;
|
|
212
212
|
durationMs: Schema.NullOr<Schema.refine<number, typeof Schema.NonNegative>>;
|
|
213
213
|
input: Schema.Struct<{
|
|
@@ -417,80 +417,80 @@ declare const EvalCell: Schema.Struct<{
|
|
|
417
417
|
taskIndex: typeof Schema.Int;
|
|
418
418
|
trials: Schema.Array$<Schema.Struct<{
|
|
419
419
|
validations: Schema.optional<Schema.refine<readonly {
|
|
420
|
-
readonly
|
|
420
|
+
readonly id: string;
|
|
421
|
+
readonly name: string;
|
|
422
|
+
readonly startedAt: number | null;
|
|
423
|
+
readonly status: "passed" | "running" | "failed" | "queued" | "error" | "skipped";
|
|
421
424
|
readonly message: string;
|
|
425
|
+
readonly kind: "command" | "judge" | "code";
|
|
422
426
|
readonly index: number;
|
|
423
|
-
readonly name: string;
|
|
424
|
-
readonly exitCode: number | null;
|
|
425
|
-
readonly output: {
|
|
426
|
-
readonly text: string;
|
|
427
|
-
readonly truncated: boolean;
|
|
428
|
-
readonly format: "text" | "json";
|
|
429
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
430
|
-
};
|
|
431
|
-
readonly input: {
|
|
432
|
-
readonly text: string;
|
|
433
|
-
readonly truncated: boolean;
|
|
434
|
-
readonly format: "text" | "json";
|
|
435
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
436
|
-
};
|
|
437
427
|
readonly error: {
|
|
438
|
-
readonly text: string;
|
|
439
428
|
readonly truncated: boolean;
|
|
429
|
+
readonly text: string;
|
|
440
430
|
readonly format: "text" | "json";
|
|
441
431
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
442
432
|
} | null;
|
|
443
|
-
readonly status: "queued" | "running" | "passed" | "failed" | "error" | "skipped";
|
|
444
|
-
readonly id: string;
|
|
445
|
-
readonly startedAt: number | null;
|
|
446
433
|
readonly durationMs: number | null;
|
|
434
|
+
readonly exitCode: number | null;
|
|
447
435
|
readonly truncated: boolean;
|
|
448
436
|
readonly judgment?: {
|
|
449
|
-
readonly name: string;
|
|
450
437
|
readonly model: string;
|
|
438
|
+
readonly name: string;
|
|
451
439
|
readonly threshold: number;
|
|
440
|
+
readonly reason: string;
|
|
452
441
|
readonly error: string | null;
|
|
453
442
|
readonly durationMs: number;
|
|
454
443
|
readonly evaluator: string;
|
|
455
444
|
readonly score: number | null;
|
|
456
445
|
readonly choice: string | null;
|
|
457
|
-
readonly reason: string;
|
|
458
446
|
} | undefined;
|
|
459
|
-
readonly
|
|
447
|
+
readonly input: {
|
|
448
|
+
readonly truncated: boolean;
|
|
460
449
|
readonly text: string;
|
|
450
|
+
readonly format: "text" | "json";
|
|
451
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
452
|
+
};
|
|
453
|
+
readonly output: {
|
|
454
|
+
readonly truncated: boolean;
|
|
455
|
+
readonly text: string;
|
|
456
|
+
readonly format: "text" | "json";
|
|
457
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
458
|
+
};
|
|
459
|
+
readonly metadata?: {
|
|
461
460
|
readonly truncated: boolean;
|
|
461
|
+
readonly text: string;
|
|
462
462
|
readonly format: "text" | "json";
|
|
463
463
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
464
464
|
} | undefined;
|
|
465
465
|
readonly calls: readonly {
|
|
466
|
+
readonly startedAt: number;
|
|
466
467
|
readonly index: number;
|
|
467
|
-
readonly
|
|
468
|
-
readonly text: string;
|
|
468
|
+
readonly error: {
|
|
469
469
|
readonly truncated: boolean;
|
|
470
|
+
readonly text: string;
|
|
470
471
|
readonly format: "text" | "json";
|
|
471
472
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
472
|
-
};
|
|
473
|
+
} | null;
|
|
474
|
+
readonly durationMs: number | null;
|
|
473
475
|
readonly input: {
|
|
474
|
-
readonly text: string;
|
|
475
476
|
readonly truncated: boolean;
|
|
477
|
+
readonly text: string;
|
|
476
478
|
readonly format: "text" | "json";
|
|
477
479
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
478
480
|
};
|
|
479
|
-
readonly
|
|
480
|
-
readonly text: string;
|
|
481
|
+
readonly output: {
|
|
481
482
|
readonly truncated: boolean;
|
|
483
|
+
readonly text: string;
|
|
482
484
|
readonly format: "text" | "json";
|
|
483
485
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
484
|
-
}
|
|
485
|
-
readonly
|
|
486
|
-
readonly durationMs: number | null;
|
|
487
|
-
readonly method: "answer" | "transcript" | "readText" | "exists" | "exec" | "cli.calls" | "mcp.calls";
|
|
486
|
+
};
|
|
487
|
+
readonly method: "answer" | "transcript" | "readText" | "exists" | "exec" | "cli.calls" | "api.calls" | "api.url" | "mcp.calls";
|
|
488
488
|
}[];
|
|
489
489
|
readonly logs: readonly {
|
|
490
490
|
readonly at: number;
|
|
491
491
|
readonly value: {
|
|
492
|
-
readonly text: string;
|
|
493
492
|
readonly truncated: boolean;
|
|
493
|
+
readonly text: string;
|
|
494
494
|
readonly format: "text" | "json";
|
|
495
495
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
496
496
|
};
|
|
@@ -545,7 +545,7 @@ declare const EvalCell: Schema.Struct<{
|
|
|
545
545
|
}>>;
|
|
546
546
|
calls: Schema.filter<Schema.Array$<Schema.Struct<{
|
|
547
547
|
index: Schema.refine<number, typeof Schema.NonNegative>;
|
|
548
|
-
method: Schema.Literal<["answer", "transcript", "readText", "exists", "exec", "cli.calls", "mcp.calls"]>;
|
|
548
|
+
method: Schema.Literal<["answer", "transcript", "readText", "exists", "exec", "cli.calls", "api.calls", "api.url", "mcp.calls"]>;
|
|
549
549
|
startedAt: typeof Schema.Number;
|
|
550
550
|
durationMs: Schema.NullOr<Schema.refine<number, typeof Schema.NonNegative>>;
|
|
551
551
|
input: Schema.Struct<{
|
|
@@ -736,80 +736,80 @@ declare const EvalRun: Schema.Struct<{
|
|
|
736
736
|
taskIndex: typeof Schema.Int;
|
|
737
737
|
trials: Schema.Array$<Schema.Struct<{
|
|
738
738
|
validations: Schema.optional<Schema.refine<readonly {
|
|
739
|
-
readonly
|
|
739
|
+
readonly id: string;
|
|
740
|
+
readonly name: string;
|
|
741
|
+
readonly startedAt: number | null;
|
|
742
|
+
readonly status: "passed" | "running" | "failed" | "queued" | "error" | "skipped";
|
|
740
743
|
readonly message: string;
|
|
744
|
+
readonly kind: "command" | "judge" | "code";
|
|
741
745
|
readonly index: number;
|
|
742
|
-
readonly name: string;
|
|
743
|
-
readonly exitCode: number | null;
|
|
744
|
-
readonly output: {
|
|
745
|
-
readonly text: string;
|
|
746
|
-
readonly truncated: boolean;
|
|
747
|
-
readonly format: "text" | "json";
|
|
748
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
749
|
-
};
|
|
750
|
-
readonly input: {
|
|
751
|
-
readonly text: string;
|
|
752
|
-
readonly truncated: boolean;
|
|
753
|
-
readonly format: "text" | "json";
|
|
754
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
755
|
-
};
|
|
756
746
|
readonly error: {
|
|
757
|
-
readonly text: string;
|
|
758
747
|
readonly truncated: boolean;
|
|
748
|
+
readonly text: string;
|
|
759
749
|
readonly format: "text" | "json";
|
|
760
750
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
761
751
|
} | null;
|
|
762
|
-
readonly status: "queued" | "running" | "passed" | "failed" | "error" | "skipped";
|
|
763
|
-
readonly id: string;
|
|
764
|
-
readonly startedAt: number | null;
|
|
765
752
|
readonly durationMs: number | null;
|
|
753
|
+
readonly exitCode: number | null;
|
|
766
754
|
readonly truncated: boolean;
|
|
767
755
|
readonly judgment?: {
|
|
768
|
-
readonly name: string;
|
|
769
756
|
readonly model: string;
|
|
757
|
+
readonly name: string;
|
|
770
758
|
readonly threshold: number;
|
|
759
|
+
readonly reason: string;
|
|
771
760
|
readonly error: string | null;
|
|
772
761
|
readonly durationMs: number;
|
|
773
762
|
readonly evaluator: string;
|
|
774
763
|
readonly score: number | null;
|
|
775
764
|
readonly choice: string | null;
|
|
776
|
-
readonly reason: string;
|
|
777
765
|
} | undefined;
|
|
778
|
-
readonly
|
|
766
|
+
readonly input: {
|
|
767
|
+
readonly truncated: boolean;
|
|
779
768
|
readonly text: string;
|
|
769
|
+
readonly format: "text" | "json";
|
|
770
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
771
|
+
};
|
|
772
|
+
readonly output: {
|
|
773
|
+
readonly truncated: boolean;
|
|
774
|
+
readonly text: string;
|
|
775
|
+
readonly format: "text" | "json";
|
|
776
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
777
|
+
};
|
|
778
|
+
readonly metadata?: {
|
|
780
779
|
readonly truncated: boolean;
|
|
780
|
+
readonly text: string;
|
|
781
781
|
readonly format: "text" | "json";
|
|
782
782
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
783
783
|
} | undefined;
|
|
784
784
|
readonly calls: readonly {
|
|
785
|
+
readonly startedAt: number;
|
|
785
786
|
readonly index: number;
|
|
786
|
-
readonly
|
|
787
|
-
readonly text: string;
|
|
787
|
+
readonly error: {
|
|
788
788
|
readonly truncated: boolean;
|
|
789
|
+
readonly text: string;
|
|
789
790
|
readonly format: "text" | "json";
|
|
790
791
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
791
|
-
};
|
|
792
|
+
} | null;
|
|
793
|
+
readonly durationMs: number | null;
|
|
792
794
|
readonly input: {
|
|
793
|
-
readonly text: string;
|
|
794
795
|
readonly truncated: boolean;
|
|
796
|
+
readonly text: string;
|
|
795
797
|
readonly format: "text" | "json";
|
|
796
798
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
797
799
|
};
|
|
798
|
-
readonly
|
|
799
|
-
readonly text: string;
|
|
800
|
+
readonly output: {
|
|
800
801
|
readonly truncated: boolean;
|
|
802
|
+
readonly text: string;
|
|
801
803
|
readonly format: "text" | "json";
|
|
802
804
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
803
|
-
}
|
|
804
|
-
readonly
|
|
805
|
-
readonly durationMs: number | null;
|
|
806
|
-
readonly method: "answer" | "transcript" | "readText" | "exists" | "exec" | "cli.calls" | "mcp.calls";
|
|
805
|
+
};
|
|
806
|
+
readonly method: "answer" | "transcript" | "readText" | "exists" | "exec" | "cli.calls" | "api.calls" | "api.url" | "mcp.calls";
|
|
807
807
|
}[];
|
|
808
808
|
readonly logs: readonly {
|
|
809
809
|
readonly at: number;
|
|
810
810
|
readonly value: {
|
|
811
|
-
readonly text: string;
|
|
812
811
|
readonly truncated: boolean;
|
|
812
|
+
readonly text: string;
|
|
813
813
|
readonly format: "text" | "json";
|
|
814
814
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
815
815
|
};
|
|
@@ -864,7 +864,7 @@ declare const EvalRun: Schema.Struct<{
|
|
|
864
864
|
}>>;
|
|
865
865
|
calls: Schema.filter<Schema.Array$<Schema.Struct<{
|
|
866
866
|
index: Schema.refine<number, typeof Schema.NonNegative>;
|
|
867
|
-
method: Schema.Literal<["answer", "transcript", "readText", "exists", "exec", "cli.calls", "mcp.calls"]>;
|
|
867
|
+
method: Schema.Literal<["answer", "transcript", "readText", "exists", "exec", "cli.calls", "api.calls", "api.url", "mcp.calls"]>;
|
|
868
868
|
startedAt: typeof Schema.Number;
|
|
869
869
|
durationMs: Schema.NullOr<Schema.refine<number, typeof Schema.NonNegative>>;
|
|
870
870
|
input: Schema.Struct<{
|
|
@@ -1078,80 +1078,80 @@ declare const EvalCellHistoryEntry: Schema.Struct<{
|
|
|
1078
1078
|
runId: typeof Schema.String;
|
|
1079
1079
|
trials: Schema.Array$<Schema.Struct<{
|
|
1080
1080
|
validations: Schema.optional<Schema.refine<readonly {
|
|
1081
|
-
readonly
|
|
1081
|
+
readonly id: string;
|
|
1082
|
+
readonly name: string;
|
|
1083
|
+
readonly startedAt: number | null;
|
|
1084
|
+
readonly status: "passed" | "running" | "failed" | "queued" | "error" | "skipped";
|
|
1082
1085
|
readonly message: string;
|
|
1086
|
+
readonly kind: "command" | "judge" | "code";
|
|
1083
1087
|
readonly index: number;
|
|
1084
|
-
readonly name: string;
|
|
1085
|
-
readonly exitCode: number | null;
|
|
1086
|
-
readonly output: {
|
|
1087
|
-
readonly text: string;
|
|
1088
|
-
readonly truncated: boolean;
|
|
1089
|
-
readonly format: "text" | "json";
|
|
1090
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
1091
|
-
};
|
|
1092
|
-
readonly input: {
|
|
1093
|
-
readonly text: string;
|
|
1094
|
-
readonly truncated: boolean;
|
|
1095
|
-
readonly format: "text" | "json";
|
|
1096
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
1097
|
-
};
|
|
1098
1088
|
readonly error: {
|
|
1099
|
-
readonly text: string;
|
|
1100
1089
|
readonly truncated: boolean;
|
|
1090
|
+
readonly text: string;
|
|
1101
1091
|
readonly format: "text" | "json";
|
|
1102
1092
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
1103
1093
|
} | null;
|
|
1104
|
-
readonly status: "queued" | "running" | "passed" | "failed" | "error" | "skipped";
|
|
1105
|
-
readonly id: string;
|
|
1106
|
-
readonly startedAt: number | null;
|
|
1107
1094
|
readonly durationMs: number | null;
|
|
1095
|
+
readonly exitCode: number | null;
|
|
1108
1096
|
readonly truncated: boolean;
|
|
1109
1097
|
readonly judgment?: {
|
|
1110
|
-
readonly name: string;
|
|
1111
1098
|
readonly model: string;
|
|
1099
|
+
readonly name: string;
|
|
1112
1100
|
readonly threshold: number;
|
|
1101
|
+
readonly reason: string;
|
|
1113
1102
|
readonly error: string | null;
|
|
1114
1103
|
readonly durationMs: number;
|
|
1115
1104
|
readonly evaluator: string;
|
|
1116
1105
|
readonly score: number | null;
|
|
1117
1106
|
readonly choice: string | null;
|
|
1118
|
-
readonly reason: string;
|
|
1119
1107
|
} | undefined;
|
|
1120
|
-
readonly
|
|
1108
|
+
readonly input: {
|
|
1109
|
+
readonly truncated: boolean;
|
|
1121
1110
|
readonly text: string;
|
|
1111
|
+
readonly format: "text" | "json";
|
|
1112
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
1113
|
+
};
|
|
1114
|
+
readonly output: {
|
|
1115
|
+
readonly truncated: boolean;
|
|
1116
|
+
readonly text: string;
|
|
1117
|
+
readonly format: "text" | "json";
|
|
1118
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
1119
|
+
};
|
|
1120
|
+
readonly metadata?: {
|
|
1122
1121
|
readonly truncated: boolean;
|
|
1122
|
+
readonly text: string;
|
|
1123
1123
|
readonly format: "text" | "json";
|
|
1124
1124
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
1125
1125
|
} | undefined;
|
|
1126
1126
|
readonly calls: readonly {
|
|
1127
|
+
readonly startedAt: number;
|
|
1127
1128
|
readonly index: number;
|
|
1128
|
-
readonly
|
|
1129
|
-
readonly text: string;
|
|
1129
|
+
readonly error: {
|
|
1130
1130
|
readonly truncated: boolean;
|
|
1131
|
+
readonly text: string;
|
|
1131
1132
|
readonly format: "text" | "json";
|
|
1132
1133
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
1133
|
-
};
|
|
1134
|
+
} | null;
|
|
1135
|
+
readonly durationMs: number | null;
|
|
1134
1136
|
readonly input: {
|
|
1135
|
-
readonly text: string;
|
|
1136
1137
|
readonly truncated: boolean;
|
|
1138
|
+
readonly text: string;
|
|
1137
1139
|
readonly format: "text" | "json";
|
|
1138
1140
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
1139
1141
|
};
|
|
1140
|
-
readonly
|
|
1141
|
-
readonly text: string;
|
|
1142
|
+
readonly output: {
|
|
1142
1143
|
readonly truncated: boolean;
|
|
1144
|
+
readonly text: string;
|
|
1143
1145
|
readonly format: "text" | "json";
|
|
1144
1146
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
1145
|
-
}
|
|
1146
|
-
readonly
|
|
1147
|
-
readonly durationMs: number | null;
|
|
1148
|
-
readonly method: "answer" | "transcript" | "readText" | "exists" | "exec" | "cli.calls" | "mcp.calls";
|
|
1147
|
+
};
|
|
1148
|
+
readonly method: "answer" | "transcript" | "readText" | "exists" | "exec" | "cli.calls" | "api.calls" | "api.url" | "mcp.calls";
|
|
1149
1149
|
}[];
|
|
1150
1150
|
readonly logs: readonly {
|
|
1151
1151
|
readonly at: number;
|
|
1152
1152
|
readonly value: {
|
|
1153
|
-
readonly text: string;
|
|
1154
1153
|
readonly truncated: boolean;
|
|
1154
|
+
readonly text: string;
|
|
1155
1155
|
readonly format: "text" | "json";
|
|
1156
1156
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
1157
1157
|
};
|
|
@@ -1206,7 +1206,7 @@ declare const EvalCellHistoryEntry: Schema.Struct<{
|
|
|
1206
1206
|
}>>;
|
|
1207
1207
|
calls: Schema.filter<Schema.Array$<Schema.Struct<{
|
|
1208
1208
|
index: Schema.refine<number, typeof Schema.NonNegative>;
|
|
1209
|
-
method: Schema.Literal<["answer", "transcript", "readText", "exists", "exec", "cli.calls", "mcp.calls"]>;
|
|
1209
|
+
method: Schema.Literal<["answer", "transcript", "readText", "exists", "exec", "cli.calls", "api.calls", "api.url", "mcp.calls"]>;
|
|
1210
1210
|
startedAt: typeof Schema.Number;
|
|
1211
1211
|
durationMs: Schema.NullOr<Schema.refine<number, typeof Schema.NonNegative>>;
|
|
1212
1212
|
input: Schema.Struct<{
|