anpord 0.1.13 → 0.1.17
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/README.md +1 -1
- 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-BXZmPUwV.d.cts +101 -0
- package/dist/api-mocks-BXZmPUwV.d.mts +101 -0
- package/dist/api-mocks-CuVw86Qw.cjs +68 -0
- package/dist/api-mocks-wNUnpzJf.mjs +33 -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 +9 -26
- package/dist/cli-runtime.mjs +5 -22
- package/dist/{client-BirFO3jz.d.cts → client-BVoIRPot.d.cts} +1082 -1082
- package/dist/{client-BfdjFtlV.mjs → client-DE3YNtWc.mjs} +3 -90
- package/dist/{client-C-P2I00z.cjs → client-DziJjMFh.cjs} +4 -91
- package/dist/{client-BgRs8JvW.d.mts → client-_OgJhHpJ.d.mts} +1083 -1083
- package/dist/{compiler-D96wI7UP.mjs → compiler-C62ss3a4.mjs} +161 -61
- package/dist/{compiler-uwb1pkdK.cjs → compiler-CYwJNlT4.cjs} +202 -72
- 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-0FkhUOIB.d.cts} +2 -2
- package/dist/{eval-judges-DPPUftbh.d.mts → eval-judges-0FkhUOIB.d.mts} +2 -2
- package/dist/{eval-judges-BkmiVLLq.cjs → eval-judges-1zF4PVqV.cjs} +1 -1
- package/dist/{eval-judges-CwTvTq7N.mjs → eval-judges-Bee_ABJt.mjs} +1 -1
- package/dist/{eval-validations-_MlnQ_uv.cjs → eval-validations-BdlPhudg.cjs} +8 -2
- package/dist/{eval-validations-BRdoZrTQ.mjs → eval-validations-BvX9dkYh.mjs} +3 -3
- package/dist/eval.cjs +4 -1
- package/dist/eval.d.cts +275 -25
- package/dist/eval.d.mts +275 -25
- package/dist/eval.mjs +2 -2
- package/dist/{evals-B6aT3xgH.d.cts → evals-CFxc4IrF.d.cts} +124 -124
- package/dist/{evals-B6aT3xgH.d.mts → evals-CFxc4IrF.d.mts} +124 -124
- package/dist/{evals-DouMsrsM.cjs → evals-api-CqhPZCr5.cjs} +104 -73
- package/dist/{evals-Dcc_x6jW.mjs → evals-api-CsdU741J.mjs} +96 -5
- package/dist/index.cjs +8 -8
- package/dist/index.d.cts +12 -8
- package/dist/index.d.mts +13 -9
- package/dist/index.mjs +7 -7
- package/dist/mcp-runtime.cjs +9 -18
- package/dist/mcp-runtime.mjs +5 -14
- package/dist/mock-journal-Bnaozlx4.cjs +39 -0
- package/dist/mock-journal-rbsXfDRs.mjs +22 -0
- package/dist/runtime-CXcsGvwP.cjs +292 -0
- package/dist/runtime-DDUrsAI-.mjs +269 -0
- package/dist/source.d.cts +1 -1
- package/dist/source.d.mts +1 -1
- package/dist/{types-Codjaq-o.d.mts → types-BjQf9IIi.d.mts} +21 -11
- package/dist/{types-ISsteq95.d.cts → types-C2nRDJ95.d.cts} +20 -10
- 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.cjs +1 -1
- package/dist/validators.d.cts +1 -1
- package/dist/validators.d.mts +1 -1
- package/dist/validators.mjs +1 -1
- package/package.json +21 -1
|
@@ -79,81 +79,81 @@ 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 kind: "command" | "judge" | "code";
|
|
83
82
|
readonly message: string;
|
|
84
|
-
readonly index: number;
|
|
85
83
|
readonly name: string;
|
|
86
|
-
readonly
|
|
87
|
-
readonly
|
|
88
|
-
|
|
84
|
+
readonly kind: "judge" | "command" | "code";
|
|
85
|
+
readonly index: number;
|
|
86
|
+
readonly error: {
|
|
87
|
+
readonly format: "text" | "json";
|
|
89
88
|
readonly truncated: boolean;
|
|
89
|
+
readonly text: string;
|
|
90
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
91
|
+
} | null;
|
|
92
|
+
readonly output: {
|
|
90
93
|
readonly format: "text" | "json";
|
|
94
|
+
readonly truncated: boolean;
|
|
95
|
+
readonly text: string;
|
|
91
96
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
92
97
|
};
|
|
93
98
|
readonly input: {
|
|
94
|
-
readonly text: string;
|
|
95
|
-
readonly truncated: boolean;
|
|
96
99
|
readonly format: "text" | "json";
|
|
97
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
98
|
-
};
|
|
99
|
-
readonly error: {
|
|
100
|
-
readonly text: string;
|
|
101
100
|
readonly truncated: boolean;
|
|
102
|
-
readonly
|
|
101
|
+
readonly text: string;
|
|
103
102
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
104
|
-
}
|
|
105
|
-
readonly status: "queued" | "running" | "passed" | "failed" | "error" | "skipped";
|
|
103
|
+
};
|
|
106
104
|
readonly id: string;
|
|
107
105
|
readonly startedAt: number | null;
|
|
106
|
+
readonly status: "error" | "passed" | "running" | "failed" | "queued" | "skipped";
|
|
108
107
|
readonly durationMs: number | null;
|
|
108
|
+
readonly exitCode: number | null;
|
|
109
109
|
readonly truncated: boolean;
|
|
110
110
|
readonly judgment?: {
|
|
111
111
|
readonly name: string;
|
|
112
112
|
readonly model: string;
|
|
113
113
|
readonly threshold: number;
|
|
114
114
|
readonly error: string | null;
|
|
115
|
+
readonly reason: string;
|
|
115
116
|
readonly durationMs: number;
|
|
116
117
|
readonly evaluator: string;
|
|
117
118
|
readonly score: number | null;
|
|
118
119
|
readonly choice: string | null;
|
|
119
|
-
readonly reason: string;
|
|
120
120
|
} | undefined;
|
|
121
121
|
readonly metadata?: {
|
|
122
|
-
readonly text: string;
|
|
123
|
-
readonly truncated: boolean;
|
|
124
122
|
readonly format: "text" | "json";
|
|
123
|
+
readonly truncated: boolean;
|
|
124
|
+
readonly text: string;
|
|
125
125
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
126
126
|
} | undefined;
|
|
127
127
|
readonly calls: readonly {
|
|
128
128
|
readonly index: number;
|
|
129
|
-
readonly
|
|
130
|
-
readonly
|
|
129
|
+
readonly error: {
|
|
130
|
+
readonly format: "text" | "json";
|
|
131
131
|
readonly truncated: boolean;
|
|
132
|
+
readonly text: string;
|
|
133
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
134
|
+
} | null;
|
|
135
|
+
readonly output: {
|
|
132
136
|
readonly format: "text" | "json";
|
|
137
|
+
readonly truncated: boolean;
|
|
138
|
+
readonly text: string;
|
|
133
139
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
134
140
|
};
|
|
135
141
|
readonly input: {
|
|
136
|
-
readonly text: string;
|
|
137
|
-
readonly truncated: boolean;
|
|
138
142
|
readonly format: "text" | "json";
|
|
139
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
140
|
-
};
|
|
141
|
-
readonly error: {
|
|
142
|
-
readonly text: string;
|
|
143
143
|
readonly truncated: boolean;
|
|
144
|
-
readonly
|
|
144
|
+
readonly text: string;
|
|
145
145
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
146
|
-
}
|
|
146
|
+
};
|
|
147
147
|
readonly startedAt: number;
|
|
148
148
|
readonly durationMs: number | null;
|
|
149
|
-
readonly method: "answer" | "transcript" | "readText" | "exists" | "exec" | "cli.calls" | "mcp.calls";
|
|
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
|
-
readonly truncated: boolean;
|
|
156
154
|
readonly format: "text" | "json";
|
|
155
|
+
readonly truncated: boolean;
|
|
156
|
+
readonly text: string;
|
|
157
157
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
158
158
|
};
|
|
159
159
|
readonly index: number;
|
|
@@ -207,7 +207,7 @@ declare const EvalTrial: Schema.Struct<{
|
|
|
207
207
|
}>>;
|
|
208
208
|
calls: Schema.filter<Schema.Array$<Schema.Struct<{
|
|
209
209
|
index: Schema.refine<number, typeof Schema.NonNegative>;
|
|
210
|
-
method: Schema.Literal<["answer", "transcript", "readText", "exists", "exec", "cli.calls", "mcp.calls"]>;
|
|
210
|
+
method: Schema.Literal<["answer", "transcript", "readText", "exists", "exec", "cli.calls", "api.calls", "api.url", "mcp.calls"]>;
|
|
211
211
|
startedAt: typeof Schema.Number;
|
|
212
212
|
durationMs: Schema.NullOr<Schema.refine<number, typeof Schema.NonNegative>>;
|
|
213
213
|
input: Schema.Struct<{
|
|
@@ -417,81 +417,81 @@ declare const EvalCell: Schema.Struct<{
|
|
|
417
417
|
taskIndex: typeof Schema.Int;
|
|
418
418
|
trials: Schema.Array$<Schema.Struct<{
|
|
419
419
|
validations: Schema.optional<Schema.refine<readonly {
|
|
420
|
-
readonly kind: "command" | "judge" | "code";
|
|
421
420
|
readonly message: string;
|
|
422
|
-
readonly index: number;
|
|
423
421
|
readonly name: string;
|
|
424
|
-
readonly
|
|
425
|
-
readonly
|
|
426
|
-
|
|
422
|
+
readonly kind: "judge" | "command" | "code";
|
|
423
|
+
readonly index: number;
|
|
424
|
+
readonly error: {
|
|
425
|
+
readonly format: "text" | "json";
|
|
427
426
|
readonly truncated: boolean;
|
|
427
|
+
readonly text: string;
|
|
428
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
429
|
+
} | null;
|
|
430
|
+
readonly output: {
|
|
428
431
|
readonly format: "text" | "json";
|
|
432
|
+
readonly truncated: boolean;
|
|
433
|
+
readonly text: string;
|
|
429
434
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
430
435
|
};
|
|
431
436
|
readonly input: {
|
|
432
|
-
readonly text: string;
|
|
433
|
-
readonly truncated: boolean;
|
|
434
437
|
readonly format: "text" | "json";
|
|
435
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
436
|
-
};
|
|
437
|
-
readonly error: {
|
|
438
|
-
readonly text: string;
|
|
439
438
|
readonly truncated: boolean;
|
|
440
|
-
readonly
|
|
439
|
+
readonly text: string;
|
|
441
440
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
442
|
-
}
|
|
443
|
-
readonly status: "queued" | "running" | "passed" | "failed" | "error" | "skipped";
|
|
441
|
+
};
|
|
444
442
|
readonly id: string;
|
|
445
443
|
readonly startedAt: number | null;
|
|
444
|
+
readonly status: "error" | "passed" | "running" | "failed" | "queued" | "skipped";
|
|
446
445
|
readonly durationMs: number | null;
|
|
446
|
+
readonly exitCode: number | null;
|
|
447
447
|
readonly truncated: boolean;
|
|
448
448
|
readonly judgment?: {
|
|
449
449
|
readonly name: string;
|
|
450
450
|
readonly model: string;
|
|
451
451
|
readonly threshold: number;
|
|
452
452
|
readonly error: string | null;
|
|
453
|
+
readonly reason: string;
|
|
453
454
|
readonly durationMs: number;
|
|
454
455
|
readonly evaluator: string;
|
|
455
456
|
readonly score: number | null;
|
|
456
457
|
readonly choice: string | null;
|
|
457
|
-
readonly reason: string;
|
|
458
458
|
} | undefined;
|
|
459
459
|
readonly metadata?: {
|
|
460
|
-
readonly text: string;
|
|
461
|
-
readonly truncated: boolean;
|
|
462
460
|
readonly format: "text" | "json";
|
|
461
|
+
readonly truncated: boolean;
|
|
462
|
+
readonly text: string;
|
|
463
463
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
464
464
|
} | undefined;
|
|
465
465
|
readonly calls: readonly {
|
|
466
466
|
readonly index: number;
|
|
467
|
-
readonly
|
|
468
|
-
readonly
|
|
467
|
+
readonly error: {
|
|
468
|
+
readonly format: "text" | "json";
|
|
469
469
|
readonly truncated: boolean;
|
|
470
|
+
readonly text: string;
|
|
471
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
472
|
+
} | null;
|
|
473
|
+
readonly output: {
|
|
470
474
|
readonly format: "text" | "json";
|
|
475
|
+
readonly truncated: boolean;
|
|
476
|
+
readonly text: string;
|
|
471
477
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
472
478
|
};
|
|
473
479
|
readonly input: {
|
|
474
|
-
readonly text: string;
|
|
475
|
-
readonly truncated: boolean;
|
|
476
480
|
readonly format: "text" | "json";
|
|
477
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
478
|
-
};
|
|
479
|
-
readonly error: {
|
|
480
|
-
readonly text: string;
|
|
481
481
|
readonly truncated: boolean;
|
|
482
|
-
readonly
|
|
482
|
+
readonly text: string;
|
|
483
483
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
484
|
-
}
|
|
484
|
+
};
|
|
485
485
|
readonly startedAt: number;
|
|
486
486
|
readonly durationMs: number | null;
|
|
487
|
-
readonly method: "answer" | "transcript" | "readText" | "exists" | "exec" | "cli.calls" | "mcp.calls";
|
|
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
|
-
readonly truncated: boolean;
|
|
494
492
|
readonly format: "text" | "json";
|
|
493
|
+
readonly truncated: boolean;
|
|
494
|
+
readonly text: string;
|
|
495
495
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
496
496
|
};
|
|
497
497
|
readonly index: number;
|
|
@@ -545,7 +545,7 @@ declare const EvalCell: Schema.Struct<{
|
|
|
545
545
|
}>>;
|
|
546
546
|
calls: Schema.filter<Schema.Array$<Schema.Struct<{
|
|
547
547
|
index: Schema.refine<number, typeof Schema.NonNegative>;
|
|
548
|
-
method: Schema.Literal<["answer", "transcript", "readText", "exists", "exec", "cli.calls", "mcp.calls"]>;
|
|
548
|
+
method: Schema.Literal<["answer", "transcript", "readText", "exists", "exec", "cli.calls", "api.calls", "api.url", "mcp.calls"]>;
|
|
549
549
|
startedAt: typeof Schema.Number;
|
|
550
550
|
durationMs: Schema.NullOr<Schema.refine<number, typeof Schema.NonNegative>>;
|
|
551
551
|
input: Schema.Struct<{
|
|
@@ -736,81 +736,81 @@ declare const EvalRun: Schema.Struct<{
|
|
|
736
736
|
taskIndex: typeof Schema.Int;
|
|
737
737
|
trials: Schema.Array$<Schema.Struct<{
|
|
738
738
|
validations: Schema.optional<Schema.refine<readonly {
|
|
739
|
-
readonly kind: "command" | "judge" | "code";
|
|
740
739
|
readonly message: string;
|
|
741
|
-
readonly index: number;
|
|
742
740
|
readonly name: string;
|
|
743
|
-
readonly
|
|
744
|
-
readonly
|
|
745
|
-
|
|
741
|
+
readonly kind: "judge" | "command" | "code";
|
|
742
|
+
readonly index: number;
|
|
743
|
+
readonly error: {
|
|
744
|
+
readonly format: "text" | "json";
|
|
746
745
|
readonly truncated: boolean;
|
|
746
|
+
readonly text: string;
|
|
747
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
748
|
+
} | null;
|
|
749
|
+
readonly output: {
|
|
747
750
|
readonly format: "text" | "json";
|
|
751
|
+
readonly truncated: boolean;
|
|
752
|
+
readonly text: string;
|
|
748
753
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
749
754
|
};
|
|
750
755
|
readonly input: {
|
|
751
|
-
readonly text: string;
|
|
752
|
-
readonly truncated: boolean;
|
|
753
756
|
readonly format: "text" | "json";
|
|
754
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
755
|
-
};
|
|
756
|
-
readonly error: {
|
|
757
|
-
readonly text: string;
|
|
758
757
|
readonly truncated: boolean;
|
|
759
|
-
readonly
|
|
758
|
+
readonly text: string;
|
|
760
759
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
761
|
-
}
|
|
762
|
-
readonly status: "queued" | "running" | "passed" | "failed" | "error" | "skipped";
|
|
760
|
+
};
|
|
763
761
|
readonly id: string;
|
|
764
762
|
readonly startedAt: number | null;
|
|
763
|
+
readonly status: "error" | "passed" | "running" | "failed" | "queued" | "skipped";
|
|
765
764
|
readonly durationMs: number | null;
|
|
765
|
+
readonly exitCode: number | null;
|
|
766
766
|
readonly truncated: boolean;
|
|
767
767
|
readonly judgment?: {
|
|
768
768
|
readonly name: string;
|
|
769
769
|
readonly model: string;
|
|
770
770
|
readonly threshold: number;
|
|
771
771
|
readonly error: string | null;
|
|
772
|
+
readonly reason: string;
|
|
772
773
|
readonly durationMs: number;
|
|
773
774
|
readonly evaluator: string;
|
|
774
775
|
readonly score: number | null;
|
|
775
776
|
readonly choice: string | null;
|
|
776
|
-
readonly reason: string;
|
|
777
777
|
} | undefined;
|
|
778
778
|
readonly metadata?: {
|
|
779
|
-
readonly text: string;
|
|
780
|
-
readonly truncated: boolean;
|
|
781
779
|
readonly format: "text" | "json";
|
|
780
|
+
readonly truncated: boolean;
|
|
781
|
+
readonly text: string;
|
|
782
782
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
783
783
|
} | undefined;
|
|
784
784
|
readonly calls: readonly {
|
|
785
785
|
readonly index: number;
|
|
786
|
-
readonly
|
|
787
|
-
readonly
|
|
786
|
+
readonly error: {
|
|
787
|
+
readonly format: "text" | "json";
|
|
788
788
|
readonly truncated: boolean;
|
|
789
|
+
readonly text: string;
|
|
790
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
791
|
+
} | null;
|
|
792
|
+
readonly output: {
|
|
789
793
|
readonly format: "text" | "json";
|
|
794
|
+
readonly truncated: boolean;
|
|
795
|
+
readonly text: string;
|
|
790
796
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
791
797
|
};
|
|
792
798
|
readonly input: {
|
|
793
|
-
readonly text: string;
|
|
794
|
-
readonly truncated: boolean;
|
|
795
799
|
readonly format: "text" | "json";
|
|
796
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
797
|
-
};
|
|
798
|
-
readonly error: {
|
|
799
|
-
readonly text: string;
|
|
800
800
|
readonly truncated: boolean;
|
|
801
|
-
readonly
|
|
801
|
+
readonly text: string;
|
|
802
802
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
803
|
-
}
|
|
803
|
+
};
|
|
804
804
|
readonly startedAt: number;
|
|
805
805
|
readonly durationMs: number | null;
|
|
806
|
-
readonly method: "answer" | "transcript" | "readText" | "exists" | "exec" | "cli.calls" | "mcp.calls";
|
|
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
|
-
readonly truncated: boolean;
|
|
813
811
|
readonly format: "text" | "json";
|
|
812
|
+
readonly truncated: boolean;
|
|
813
|
+
readonly text: string;
|
|
814
814
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
815
815
|
};
|
|
816
816
|
readonly index: number;
|
|
@@ -864,7 +864,7 @@ declare const EvalRun: Schema.Struct<{
|
|
|
864
864
|
}>>;
|
|
865
865
|
calls: Schema.filter<Schema.Array$<Schema.Struct<{
|
|
866
866
|
index: Schema.refine<number, typeof Schema.NonNegative>;
|
|
867
|
-
method: Schema.Literal<["answer", "transcript", "readText", "exists", "exec", "cli.calls", "mcp.calls"]>;
|
|
867
|
+
method: Schema.Literal<["answer", "transcript", "readText", "exists", "exec", "cli.calls", "api.calls", "api.url", "mcp.calls"]>;
|
|
868
868
|
startedAt: typeof Schema.Number;
|
|
869
869
|
durationMs: Schema.NullOr<Schema.refine<number, typeof Schema.NonNegative>>;
|
|
870
870
|
input: Schema.Struct<{
|
|
@@ -1078,81 +1078,81 @@ declare const EvalCellHistoryEntry: Schema.Struct<{
|
|
|
1078
1078
|
runId: typeof Schema.String;
|
|
1079
1079
|
trials: Schema.Array$<Schema.Struct<{
|
|
1080
1080
|
validations: Schema.optional<Schema.refine<readonly {
|
|
1081
|
-
readonly kind: "command" | "judge" | "code";
|
|
1082
1081
|
readonly message: string;
|
|
1083
|
-
readonly index: number;
|
|
1084
1082
|
readonly name: string;
|
|
1085
|
-
readonly
|
|
1086
|
-
readonly
|
|
1087
|
-
|
|
1083
|
+
readonly kind: "judge" | "command" | "code";
|
|
1084
|
+
readonly index: number;
|
|
1085
|
+
readonly error: {
|
|
1086
|
+
readonly format: "text" | "json";
|
|
1088
1087
|
readonly truncated: boolean;
|
|
1088
|
+
readonly text: string;
|
|
1089
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
1090
|
+
} | null;
|
|
1091
|
+
readonly output: {
|
|
1089
1092
|
readonly format: "text" | "json";
|
|
1093
|
+
readonly truncated: boolean;
|
|
1094
|
+
readonly text: string;
|
|
1090
1095
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
1091
1096
|
};
|
|
1092
1097
|
readonly input: {
|
|
1093
|
-
readonly text: string;
|
|
1094
|
-
readonly truncated: boolean;
|
|
1095
1098
|
readonly format: "text" | "json";
|
|
1096
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
1097
|
-
};
|
|
1098
|
-
readonly error: {
|
|
1099
|
-
readonly text: string;
|
|
1100
1099
|
readonly truncated: boolean;
|
|
1101
|
-
readonly
|
|
1100
|
+
readonly text: string;
|
|
1102
1101
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
1103
|
-
}
|
|
1104
|
-
readonly status: "queued" | "running" | "passed" | "failed" | "error" | "skipped";
|
|
1102
|
+
};
|
|
1105
1103
|
readonly id: string;
|
|
1106
1104
|
readonly startedAt: number | null;
|
|
1105
|
+
readonly status: "error" | "passed" | "running" | "failed" | "queued" | "skipped";
|
|
1107
1106
|
readonly durationMs: number | null;
|
|
1107
|
+
readonly exitCode: number | null;
|
|
1108
1108
|
readonly truncated: boolean;
|
|
1109
1109
|
readonly judgment?: {
|
|
1110
1110
|
readonly name: string;
|
|
1111
1111
|
readonly model: string;
|
|
1112
1112
|
readonly threshold: number;
|
|
1113
1113
|
readonly error: string | null;
|
|
1114
|
+
readonly reason: string;
|
|
1114
1115
|
readonly durationMs: number;
|
|
1115
1116
|
readonly evaluator: string;
|
|
1116
1117
|
readonly score: number | null;
|
|
1117
1118
|
readonly choice: string | null;
|
|
1118
|
-
readonly reason: string;
|
|
1119
1119
|
} | undefined;
|
|
1120
1120
|
readonly metadata?: {
|
|
1121
|
-
readonly text: string;
|
|
1122
|
-
readonly truncated: boolean;
|
|
1123
1121
|
readonly format: "text" | "json";
|
|
1122
|
+
readonly truncated: boolean;
|
|
1123
|
+
readonly text: string;
|
|
1124
1124
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
1125
1125
|
} | undefined;
|
|
1126
1126
|
readonly calls: readonly {
|
|
1127
1127
|
readonly index: number;
|
|
1128
|
-
readonly
|
|
1129
|
-
readonly
|
|
1128
|
+
readonly error: {
|
|
1129
|
+
readonly format: "text" | "json";
|
|
1130
1130
|
readonly truncated: boolean;
|
|
1131
|
+
readonly text: string;
|
|
1132
|
+
readonly state: "captured" | "disabled" | "unavailable";
|
|
1133
|
+
} | null;
|
|
1134
|
+
readonly output: {
|
|
1131
1135
|
readonly format: "text" | "json";
|
|
1136
|
+
readonly truncated: boolean;
|
|
1137
|
+
readonly text: string;
|
|
1132
1138
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
1133
1139
|
};
|
|
1134
1140
|
readonly input: {
|
|
1135
|
-
readonly text: string;
|
|
1136
|
-
readonly truncated: boolean;
|
|
1137
1141
|
readonly format: "text" | "json";
|
|
1138
|
-
readonly state: "captured" | "disabled" | "unavailable";
|
|
1139
|
-
};
|
|
1140
|
-
readonly error: {
|
|
1141
|
-
readonly text: string;
|
|
1142
1142
|
readonly truncated: boolean;
|
|
1143
|
-
readonly
|
|
1143
|
+
readonly text: string;
|
|
1144
1144
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
1145
|
-
}
|
|
1145
|
+
};
|
|
1146
1146
|
readonly startedAt: number;
|
|
1147
1147
|
readonly durationMs: number | null;
|
|
1148
|
-
readonly method: "answer" | "transcript" | "readText" | "exists" | "exec" | "cli.calls" | "mcp.calls";
|
|
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
|
-
readonly truncated: boolean;
|
|
1155
1153
|
readonly format: "text" | "json";
|
|
1154
|
+
readonly truncated: boolean;
|
|
1155
|
+
readonly text: string;
|
|
1156
1156
|
readonly state: "captured" | "disabled" | "unavailable";
|
|
1157
1157
|
};
|
|
1158
1158
|
readonly index: number;
|
|
@@ -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<{
|