anpord 0.1.20 → 0.1.21

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.
Files changed (57) hide show
  1. package/dist/api-context.cjs +1 -1
  2. package/dist/api-context.d.cts +1 -1
  3. package/dist/api-context.d.mts +1 -1
  4. package/dist/api-context.mjs +1 -1
  5. package/dist/{api-mocks-C3SsliJZ.mjs → api-mocks-BYWvJp9v.mjs} +3 -3
  6. package/dist/{api-mocks-xXKzUFc9.d.cts → api-mocks-BenVx0e8.d.cts} +11 -11
  7. package/dist/{api-mocks-xXKzUFc9.d.mts → api-mocks-BenVx0e8.d.mts} +11 -11
  8. package/dist/{api-mocks-BjnewAya.cjs → api-mocks-CgiC7T9U.cjs} +8 -2
  9. package/dist/api-runtime.cjs +1 -1
  10. package/dist/api-runtime.d.cts +1 -1
  11. package/dist/api-runtime.d.mts +1 -1
  12. package/dist/api-runtime.mjs +1 -1
  13. package/dist/api.cjs +1 -1
  14. package/dist/api.d.cts +1 -1
  15. package/dist/api.d.mts +1 -1
  16. package/dist/api.mjs +1 -1
  17. package/dist/bin.cjs +4005 -50
  18. package/dist/bin.mjs +4008 -53
  19. package/dist/{client-D4SBJEsG.cjs → client-BZDftC1x.cjs} +1 -1
  20. package/dist/{client-ClWUDcPC.d.mts → client-C7iEhiDm.d.mts} +5352 -2328
  21. package/dist/{client-BD2VyKja.d.cts → client-CsGxLUZq.d.cts} +5352 -2328
  22. package/dist/{client-DCSRP55Q.mjs → client-D0l3kwWV.mjs} +1 -1
  23. package/dist/{compiler-DSN62KNe.cjs → compiler-Cy85vNbp.cjs} +2 -2
  24. package/dist/{compiler-BKchnKps.mjs → compiler-DSQRHvXS.mjs} +2 -2
  25. package/dist/config.cjs +1 -1
  26. package/dist/config.d.cts +1 -1
  27. package/dist/config.d.mts +1 -1
  28. package/dist/config.mjs +1 -1
  29. package/dist/{eval-turns-DZTHit19.mjs → eval-turns-BC-GMHD-.mjs} +2 -1
  30. package/dist/{eval-turns-weR_7JIp.cjs → eval-turns-DcJ8_1qr.cjs} +7 -0
  31. package/dist/{eval-validations-D0Cz6xSg.cjs → eval-validations-CZm7vVtQ.cjs} +12 -0
  32. package/dist/{eval-validations-CUG2PvR_.mjs → eval-validations-DNO-qwkU.mjs} +1 -1
  33. package/dist/eval.cjs +1 -1
  34. package/dist/eval.d.cts +1 -1
  35. package/dist/eval.d.mts +1 -1
  36. package/dist/eval.mjs +1 -1
  37. package/dist/{evals-WMHOpPxd.d.cts → evals-Dw2Ek-mx.d.cts} +115 -112
  38. package/dist/{evals-WMHOpPxd.d.mts → evals-Dw2Ek-mx.d.mts} +115 -112
  39. package/dist/{evals-api-B0lnr_R8.mjs → evals-api-CK3F1KoI.mjs} +113 -14
  40. package/dist/{evals-api-DuDr8UdV.cjs → evals-api-CNoLlXL5.cjs} +124 -13
  41. package/dist/index.cjs +2 -2
  42. package/dist/index.d.cts +3 -3
  43. package/dist/index.d.mts +3 -3
  44. package/dist/index.mjs +2 -2
  45. package/dist/{runtime-B5it0Siy.cjs → runtime-Bfs2j0jW.cjs} +2 -2
  46. package/dist/{runtime-ChWJywJY.mjs → runtime-CYr8RbsW.mjs} +2 -2
  47. package/dist/source.d.cts +1 -1
  48. package/dist/source.d.mts +1 -1
  49. package/dist/{types-Cw5k37Uk.d.mts → types-D-h67MVJ.d.mts} +5 -4
  50. package/dist/{types-CUVZ1A_F.d.cts → types-Dvd_z5ge.d.cts} +5 -4
  51. package/dist/validator-runtime.cjs +2 -2
  52. package/dist/validator-runtime.d.cts +1 -1
  53. package/dist/validator-runtime.d.mts +1 -1
  54. package/dist/validator-runtime.mjs +2 -2
  55. package/dist/validators.cjs +1 -1
  56. package/dist/validators.mjs +1 -1
  57. package/package.json +3 -1
@@ -1,4 +1,4 @@
1
- import { _ as CredentialConnection, b as CredentialValues, f as ApiKeyAuthentication, g as NotFound, h as Forbidden, m as Conflict, p as BadRequest, t as PublicEvalsGroup, v as CredentialIntegration, y as CredentialScope } from "./evals-api-B0lnr_R8.mjs";
1
+ import { S as CredentialValues, _ as Forbidden, b as CredentialIntegration, g as Conflict, h as BadRequest, m as ApiKeyAuthentication, t as PublicEvalsGroup, v as NotFound, x as CredentialScope, y as CredentialConnection } from "./evals-api-CK3F1KoI.mjs";
2
2
  import { Effect, Layer, Redacted, Schema } from "effect";
3
3
  import { FetchHttpClient, HttpApi, HttpApiClient, HttpApiEndpoint, HttpApiGroup, HttpClient, HttpClientRequest, OpenApi } from "@effect/platform";
4
4
  //#region ../schema/src/domain/prompts.ts
@@ -1,6 +1,6 @@
1
1
  const require_eval_judges = require("./eval-judges-1zF4PVqV.cjs");
2
- const require_api_mocks = require("./api-mocks-BjnewAya.cjs");
3
- const require_evals_api = require("./evals-api-DuDr8UdV.cjs");
2
+ const require_api_mocks = require("./api-mocks-CgiC7T9U.cjs");
3
+ const require_evals_api = require("./evals-api-CNoLlXL5.cjs");
4
4
  const require_source = require("./source-aW5nSo2h.cjs");
5
5
  let effect = require("effect");
6
6
  let node_fs_promises = require("node:fs/promises");
@@ -1,6 +1,6 @@
1
1
  import { t as EvalJudge } from "./eval-judges-Bee_ABJt.mjs";
2
- import { r as API_PROGRAM } from "./api-mocks-C3SsliJZ.mjs";
3
- import { a as EnvName, c as ProfilePath, i as EvalValidator, l as profileFitsHarness, n as PublicStartEvalRequest, o as HarnessProfile, s as PROFILE_LIMITS, u as EvalSourceFiles } from "./evals-api-B0lnr_R8.mjs";
2
+ import { r as API_PROGRAM } from "./api-mocks-BYWvJp9v.mjs";
3
+ import { c as HarnessProfile, d as profileFitsHarness, f as EvalSourceFiles, l as PROFILE_LIMITS, n as PublicStartEvalRequest, o as EvalValidator, s as EnvName, u as ProfilePath } from "./evals-api-CK3F1KoI.mjs";
4
4
  import { r as repo } from "./source-BP11pzVV.mjs";
5
5
  import { Data, Effect, Option, Schema } from "effect";
6
6
  import { mkdtemp, readFile, readdir, rm, stat, writeFile } from "node:fs/promises";
package/dist/config.cjs CHANGED
@@ -1,5 +1,5 @@
1
1
  Object.defineProperty(exports, Symbol.toStringTag, { value: "Module" });
2
- const require_client = require("./client-D4SBJEsG.cjs");
2
+ const require_client = require("./client-BZDftC1x.cjs");
3
3
  let effect = require("effect");
4
4
  //#region src/client/config.ts
5
5
  const apiKeyConfig = effect.Config.redacted("ANPORD_API_KEY");
package/dist/config.d.cts CHANGED
@@ -1,4 +1,4 @@
1
- import { t as AnpordApi } from "./client-BD2VyKja.cjs";
1
+ import { t as AnpordApi } from "./client-CsGxLUZq.cjs";
2
2
  import { Config, Layer } from "effect";
3
3
  //#region src/client/config.d.ts
4
4
  declare const apiKeyConfig: Config.Config<import("effect/Redacted").Redacted<string>>;
package/dist/config.d.mts CHANGED
@@ -1,4 +1,4 @@
1
- import { t as AnpordApi } from "./client-ClWUDcPC.mjs";
1
+ import { t as AnpordApi } from "./client-C7iEhiDm.mjs";
2
2
  import { Config, Layer } from "effect";
3
3
  //#region src/client/config.d.ts
4
4
  declare const apiKeyConfig: Config.Config<import("effect/Redacted").Redacted<string>>;
package/dist/config.mjs CHANGED
@@ -1,4 +1,4 @@
1
- import { n as DEFAULT_BASE_URL, r as layer } from "./client-DCSRP55Q.mjs";
1
+ import { n as DEFAULT_BASE_URL, r as layer } from "./client-D0l3kwWV.mjs";
2
2
  import { Config, Effect, Layer } from "effect";
3
3
  //#region src/client/config.ts
4
4
  const apiKeyConfig = Config.redacted("ANPORD_API_KEY");
@@ -1,6 +1,7 @@
1
1
  import { Schema } from "effect";
2
2
  //#region ../schema/src/domain/eval-turns.ts
3
3
  const text = Schema.String.pipe(Schema.minLength(1));
4
+ const DEFAULT_USER_MODEL = "gpt-5.4-mini";
4
5
  const EvalSimulatedUser = Schema.Struct({
5
6
  kind: Schema.Literal("simulated"),
6
7
  goal: text.pipe(Schema.maxLength(2e3)),
@@ -19,4 +20,4 @@ Schema.Struct({
19
20
  });
20
21
  Schema.Literal("user-done", "max-turns", "no-user", "failed");
21
22
  //#endregion
22
- export { EvalSimulatedUser as n, EvalUser as r, EvalScriptedUser as t };
23
+ export { EvalUser as i, EvalScriptedUser as n, EvalSimulatedUser as r, DEFAULT_USER_MODEL as t };
@@ -1,6 +1,7 @@
1
1
  let effect = require("effect");
2
2
  //#region ../schema/src/domain/eval-turns.ts
3
3
  const text = effect.Schema.String.pipe(effect.Schema.minLength(1));
4
+ const DEFAULT_USER_MODEL = "gpt-5.4-mini";
4
5
  const EvalSimulatedUser = effect.Schema.Struct({
5
6
  kind: effect.Schema.Literal("simulated"),
6
7
  goal: text.pipe(effect.Schema.maxLength(2e3)),
@@ -19,6 +20,12 @@ effect.Schema.Struct({
19
20
  });
20
21
  effect.Schema.Literal("user-done", "max-turns", "no-user", "failed");
21
22
  //#endregion
23
+ Object.defineProperty(exports, "DEFAULT_USER_MODEL", {
24
+ enumerable: true,
25
+ get: function() {
26
+ return DEFAULT_USER_MODEL;
27
+ }
28
+ });
22
29
  Object.defineProperty(exports, "EvalScriptedUser", {
23
30
  enumerable: true,
24
31
  get: function() {
@@ -110,6 +110,12 @@ const validationExecution = (identity, startedAt) => ({
110
110
  logs: []
111
111
  });
112
112
  //#endregion
113
+ Object.defineProperty(exports, "EvalValidation", {
114
+ enumerable: true,
115
+ get: function() {
116
+ return EvalValidation;
117
+ }
118
+ });
113
119
  Object.defineProperty(exports, "EvalValidations", {
114
120
  enumerable: true,
115
121
  get: function() {
@@ -122,6 +128,12 @@ Object.defineProperty(exports, "VALIDATION_FRAME", {
122
128
  return VALIDATION_FRAME;
123
129
  }
124
130
  });
131
+ Object.defineProperty(exports, "VALIDATION_TEXT_LIMIT", {
132
+ enumerable: true,
133
+ get: function() {
134
+ return VALIDATION_TEXT_LIMIT;
135
+ }
136
+ });
125
137
  Object.defineProperty(exports, "ValidationValue", {
126
138
  enumerable: true,
127
139
  get: function() {
@@ -110,4 +110,4 @@ const validationExecution = (identity, startedAt) => ({
110
110
  logs: []
111
111
  });
112
112
  //#endregion
113
- export { validationCapture as a, unavailableValue as i, VALIDATION_FRAME as n, validationExecution as o, ValidationValue as r, validationSnapshot as s, EvalValidations as t };
113
+ export { ValidationValue as a, validationExecution as c, VALIDATION_TEXT_LIMIT as i, validationSnapshot as l, EvalValidations as n, unavailableValue as o, VALIDATION_FRAME as r, validationCapture as s, EvalValidation as t };
package/dist/eval.cjs CHANGED
@@ -1,4 +1,4 @@
1
1
  Object.defineProperty(exports, Symbol.toStringTag, { value: "Module" });
2
- const require_compiler = require("./compiler-DSN62KNe.cjs");
2
+ const require_compiler = require("./compiler-Cy85vNbp.cjs");
3
3
  exports.compileDefinition = require_compiler.compileDefinition;
4
4
  exports.compileEval = require_compiler.compileEval;
package/dist/eval.d.cts CHANGED
@@ -1,4 +1,4 @@
1
- import { h as PublicStartEvalRequest, i as EvalDefinition } from "./types-CUVZ1A_F.cjs";
1
+ import { h as PublicStartEvalRequest, i as EvalDefinition } from "./types-Dvd_z5ge.cjs";
2
2
  import { Effect } from "effect";
3
3
  //#region src/evals/compiler.d.ts
4
4
  declare const compileDefinition: (definition: EvalDefinition) => Promise<PublicStartEvalRequest>;
package/dist/eval.d.mts CHANGED
@@ -1,4 +1,4 @@
1
- import { h as PublicStartEvalRequest, i as EvalDefinition } from "./types-Cw5k37Uk.mjs";
1
+ import { h as PublicStartEvalRequest, i as EvalDefinition } from "./types-D-h67MVJ.mjs";
2
2
  import { Effect } from "effect";
3
3
  //#region src/evals/compiler.d.ts
4
4
  declare const compileDefinition: (definition: EvalDefinition) => Promise<PublicStartEvalRequest>;
package/dist/eval.mjs CHANGED
@@ -1,2 +1,2 @@
1
- import { n as compileEval, t as compileDefinition } from "./compiler-BKchnKps.mjs";
1
+ import { n as compileEval, t as compileDefinition } from "./compiler-DSQRHvXS.mjs";
2
2
  export { compileDefinition, compileEval };
@@ -2,7 +2,7 @@ import { Schema } from "effect";
2
2
  //#region ../schema/src/domain/evals.d.ts
3
3
  declare const EvalHarness: Schema.Literal<["codex", "opencode", "pi", "fx", "claude", "gemini", "qwen", "cursor", "command"]>;
4
4
  type EvalHarness = typeof EvalHarness.Type;
5
- declare const EvalSandbox: Schema.Literal<["daytona", "e2b", "upstash", "modal", "cloudflare", "vercel"]>;
5
+ declare const EvalSandbox: Schema.Literal<["daytona", "e2b", "upstash", "modal", "cloudflare", "vercel", "local"]>;
6
6
  type EvalSandbox = typeof EvalSandbox.Type;
7
7
  declare const EvalTrialStatus: Schema.Literal<["queued", "running", "passed", "failed", "void"]>;
8
8
  type EvalTrialStatus = typeof EvalTrialStatus.Type;
@@ -27,7 +27,7 @@ declare const EvalTask: Schema.Struct<{
27
27
  name: typeof Schema.String;
28
28
  version: typeof Schema.String;
29
29
  }>>>;
30
- sandbox: Schema.Literal<["daytona", "e2b", "upstash", "modal", "cloudflare", "vercel"]>;
30
+ sandbox: Schema.Literal<["daytona", "e2b", "upstash", "modal", "cloudflare", "vercel", "local"]>;
31
31
  }>;
32
32
  type EvalTask = typeof EvalTask.Type;
33
33
  declare const EvalUsage: Schema.Struct<{
@@ -87,84 +87,84 @@ type EvalArtifact = typeof EvalArtifact.Type;
87
87
  declare const EvalTrial: Schema.Struct<{
88
88
  artifacts: Schema.optional<Schema.Array$<Schema.Struct<{
89
89
  path: typeof Schema.String;
90
- sha256: Schema.filter<typeof Schema.String>;
91
90
  byteSize: Schema.filter<typeof Schema.Int>;
91
+ sha256: Schema.filter<typeof Schema.String>;
92
92
  }>>>;
93
93
  validations: Schema.optional<Schema.refine<readonly {
94
94
  readonly name: string;
95
95
  readonly id: string;
96
96
  readonly kind: "command" | "judge" | "code";
97
- readonly status: "passed" | "running" | "failed" | "queued" | "error" | "skipped";
97
+ readonly status: "passed" | "running" | "failed" | "error" | "queued" | "skipped";
98
98
  readonly message: string;
99
99
  readonly startedAt: number | null;
100
100
  readonly index: number;
101
- readonly error: {
101
+ readonly exitCode: number | null;
102
+ readonly output: {
103
+ readonly text: string;
102
104
  readonly truncated: boolean;
105
+ readonly format: "text" | "json";
106
+ readonly state: "captured" | "disabled" | "unavailable";
107
+ };
108
+ readonly input: {
103
109
  readonly text: string;
110
+ readonly truncated: boolean;
111
+ readonly format: "text" | "json";
112
+ readonly state: "captured" | "disabled" | "unavailable";
113
+ };
114
+ readonly error: {
115
+ readonly text: string;
116
+ readonly truncated: boolean;
104
117
  readonly format: "text" | "json";
105
118
  readonly state: "captured" | "disabled" | "unavailable";
106
119
  } | null;
107
120
  readonly durationMs: number | null;
108
- readonly exitCode: number | null;
109
121
  readonly truncated: boolean;
110
122
  readonly judgment?: {
111
123
  readonly name: string;
112
124
  readonly model: string;
113
125
  readonly threshold: number;
114
- readonly reason: string;
115
126
  readonly error: string | null;
127
+ readonly reason: string;
116
128
  readonly durationMs: number;
117
129
  readonly evaluator: string;
118
130
  readonly score: number | null;
119
131
  readonly choice: string | null;
120
132
  } | undefined;
121
- readonly input: {
122
- readonly truncated: boolean;
123
- readonly text: string;
124
- readonly format: "text" | "json";
125
- readonly state: "captured" | "disabled" | "unavailable";
126
- };
127
- readonly output: {
128
- readonly truncated: boolean;
129
- readonly text: string;
130
- readonly format: "text" | "json";
131
- readonly state: "captured" | "disabled" | "unavailable";
132
- };
133
133
  readonly metadata?: {
134
- readonly truncated: boolean;
135
134
  readonly text: string;
135
+ readonly truncated: boolean;
136
136
  readonly format: "text" | "json";
137
137
  readonly state: "captured" | "disabled" | "unavailable";
138
138
  } | undefined;
139
139
  readonly calls: readonly {
140
140
  readonly startedAt: number;
141
141
  readonly index: number;
142
- readonly error: {
143
- readonly truncated: boolean;
142
+ readonly output: {
144
143
  readonly text: string;
144
+ readonly truncated: boolean;
145
145
  readonly format: "text" | "json";
146
146
  readonly state: "captured" | "disabled" | "unavailable";
147
- } | null;
148
- readonly durationMs: number | null;
147
+ };
149
148
  readonly input: {
150
- readonly truncated: boolean;
151
149
  readonly text: string;
150
+ readonly truncated: boolean;
152
151
  readonly format: "text" | "json";
153
152
  readonly state: "captured" | "disabled" | "unavailable";
154
153
  };
155
- readonly output: {
156
- readonly truncated: boolean;
154
+ readonly error: {
157
155
  readonly text: string;
156
+ readonly truncated: boolean;
158
157
  readonly format: "text" | "json";
159
158
  readonly state: "captured" | "disabled" | "unavailable";
160
- };
159
+ } | null;
160
+ readonly durationMs: number | null;
161
161
  readonly method: "answer" | "transcript" | "turns" | "readText" | "exists" | "exec" | "cli.calls" | "api.calls" | "api.url" | "mcp.calls";
162
162
  }[];
163
163
  readonly logs: readonly {
164
164
  readonly at: number;
165
165
  readonly value: {
166
- readonly truncated: boolean;
167
166
  readonly text: string;
167
+ readonly truncated: boolean;
168
168
  readonly format: "text" | "json";
169
169
  readonly state: "captured" | "disabled" | "unavailable";
170
170
  };
@@ -430,84 +430,84 @@ declare const EvalCell: Schema.Struct<{
430
430
  trials: Schema.Array$<Schema.Struct<{
431
431
  artifacts: Schema.optional<Schema.Array$<Schema.Struct<{
432
432
  path: typeof Schema.String;
433
- sha256: Schema.filter<typeof Schema.String>;
434
433
  byteSize: Schema.filter<typeof Schema.Int>;
434
+ sha256: Schema.filter<typeof Schema.String>;
435
435
  }>>>;
436
436
  validations: Schema.optional<Schema.refine<readonly {
437
437
  readonly name: string;
438
438
  readonly id: string;
439
439
  readonly kind: "command" | "judge" | "code";
440
- readonly status: "passed" | "running" | "failed" | "queued" | "error" | "skipped";
440
+ readonly status: "passed" | "running" | "failed" | "error" | "queued" | "skipped";
441
441
  readonly message: string;
442
442
  readonly startedAt: number | null;
443
443
  readonly index: number;
444
- readonly error: {
444
+ readonly exitCode: number | null;
445
+ readonly output: {
446
+ readonly text: string;
445
447
  readonly truncated: boolean;
448
+ readonly format: "text" | "json";
449
+ readonly state: "captured" | "disabled" | "unavailable";
450
+ };
451
+ readonly input: {
452
+ readonly text: string;
453
+ readonly truncated: boolean;
454
+ readonly format: "text" | "json";
455
+ readonly state: "captured" | "disabled" | "unavailable";
456
+ };
457
+ readonly error: {
446
458
  readonly text: string;
459
+ readonly truncated: boolean;
447
460
  readonly format: "text" | "json";
448
461
  readonly state: "captured" | "disabled" | "unavailable";
449
462
  } | null;
450
463
  readonly durationMs: number | null;
451
- readonly exitCode: number | null;
452
464
  readonly truncated: boolean;
453
465
  readonly judgment?: {
454
466
  readonly name: string;
455
467
  readonly model: string;
456
468
  readonly threshold: number;
457
- readonly reason: string;
458
469
  readonly error: string | null;
470
+ readonly reason: string;
459
471
  readonly durationMs: number;
460
472
  readonly evaluator: string;
461
473
  readonly score: number | null;
462
474
  readonly choice: string | null;
463
475
  } | undefined;
464
- readonly input: {
465
- readonly truncated: boolean;
466
- readonly text: string;
467
- readonly format: "text" | "json";
468
- readonly state: "captured" | "disabled" | "unavailable";
469
- };
470
- readonly output: {
471
- readonly truncated: boolean;
472
- readonly text: string;
473
- readonly format: "text" | "json";
474
- readonly state: "captured" | "disabled" | "unavailable";
475
- };
476
476
  readonly metadata?: {
477
- readonly truncated: boolean;
478
477
  readonly text: string;
478
+ readonly truncated: boolean;
479
479
  readonly format: "text" | "json";
480
480
  readonly state: "captured" | "disabled" | "unavailable";
481
481
  } | undefined;
482
482
  readonly calls: readonly {
483
483
  readonly startedAt: number;
484
484
  readonly index: number;
485
- readonly error: {
486
- readonly truncated: boolean;
485
+ readonly output: {
487
486
  readonly text: string;
487
+ readonly truncated: boolean;
488
488
  readonly format: "text" | "json";
489
489
  readonly state: "captured" | "disabled" | "unavailable";
490
- } | null;
491
- readonly durationMs: number | null;
490
+ };
492
491
  readonly input: {
493
- readonly truncated: boolean;
494
492
  readonly text: string;
493
+ readonly truncated: boolean;
495
494
  readonly format: "text" | "json";
496
495
  readonly state: "captured" | "disabled" | "unavailable";
497
496
  };
498
- readonly output: {
499
- readonly truncated: boolean;
497
+ readonly error: {
500
498
  readonly text: string;
499
+ readonly truncated: boolean;
501
500
  readonly format: "text" | "json";
502
501
  readonly state: "captured" | "disabled" | "unavailable";
503
- };
502
+ } | null;
503
+ readonly durationMs: number | null;
504
504
  readonly method: "answer" | "transcript" | "turns" | "readText" | "exists" | "exec" | "cli.calls" | "api.calls" | "api.url" | "mcp.calls";
505
505
  }[];
506
506
  readonly logs: readonly {
507
507
  readonly at: number;
508
508
  readonly value: {
509
- readonly truncated: boolean;
510
509
  readonly text: string;
510
+ readonly truncated: boolean;
511
511
  readonly format: "text" | "json";
512
512
  readonly state: "captured" | "disabled" | "unavailable";
513
513
  };
@@ -754,84 +754,84 @@ declare const EvalRun: Schema.Struct<{
754
754
  trials: Schema.Array$<Schema.Struct<{
755
755
  artifacts: Schema.optional<Schema.Array$<Schema.Struct<{
756
756
  path: typeof Schema.String;
757
- sha256: Schema.filter<typeof Schema.String>;
758
757
  byteSize: Schema.filter<typeof Schema.Int>;
758
+ sha256: Schema.filter<typeof Schema.String>;
759
759
  }>>>;
760
760
  validations: Schema.optional<Schema.refine<readonly {
761
761
  readonly name: string;
762
762
  readonly id: string;
763
763
  readonly kind: "command" | "judge" | "code";
764
- readonly status: "passed" | "running" | "failed" | "queued" | "error" | "skipped";
764
+ readonly status: "passed" | "running" | "failed" | "error" | "queued" | "skipped";
765
765
  readonly message: string;
766
766
  readonly startedAt: number | null;
767
767
  readonly index: number;
768
- readonly error: {
768
+ readonly exitCode: number | null;
769
+ readonly output: {
770
+ readonly text: string;
769
771
  readonly truncated: boolean;
772
+ readonly format: "text" | "json";
773
+ readonly state: "captured" | "disabled" | "unavailable";
774
+ };
775
+ readonly input: {
770
776
  readonly text: string;
777
+ readonly truncated: boolean;
778
+ readonly format: "text" | "json";
779
+ readonly state: "captured" | "disabled" | "unavailable";
780
+ };
781
+ readonly error: {
782
+ readonly text: string;
783
+ readonly truncated: boolean;
771
784
  readonly format: "text" | "json";
772
785
  readonly state: "captured" | "disabled" | "unavailable";
773
786
  } | null;
774
787
  readonly durationMs: number | null;
775
- readonly exitCode: number | null;
776
788
  readonly truncated: boolean;
777
789
  readonly judgment?: {
778
790
  readonly name: string;
779
791
  readonly model: string;
780
792
  readonly threshold: number;
781
- readonly reason: string;
782
793
  readonly error: string | null;
794
+ readonly reason: string;
783
795
  readonly durationMs: number;
784
796
  readonly evaluator: string;
785
797
  readonly score: number | null;
786
798
  readonly choice: string | null;
787
799
  } | undefined;
788
- readonly input: {
789
- readonly truncated: boolean;
790
- readonly text: string;
791
- readonly format: "text" | "json";
792
- readonly state: "captured" | "disabled" | "unavailable";
793
- };
794
- readonly output: {
795
- readonly truncated: boolean;
796
- readonly text: string;
797
- readonly format: "text" | "json";
798
- readonly state: "captured" | "disabled" | "unavailable";
799
- };
800
800
  readonly metadata?: {
801
- readonly truncated: boolean;
802
801
  readonly text: string;
802
+ readonly truncated: boolean;
803
803
  readonly format: "text" | "json";
804
804
  readonly state: "captured" | "disabled" | "unavailable";
805
805
  } | undefined;
806
806
  readonly calls: readonly {
807
807
  readonly startedAt: number;
808
808
  readonly index: number;
809
- readonly error: {
810
- readonly truncated: boolean;
809
+ readonly output: {
811
810
  readonly text: string;
811
+ readonly truncated: boolean;
812
812
  readonly format: "text" | "json";
813
813
  readonly state: "captured" | "disabled" | "unavailable";
814
- } | null;
815
- readonly durationMs: number | null;
814
+ };
816
815
  readonly input: {
817
- readonly truncated: boolean;
818
816
  readonly text: string;
817
+ readonly truncated: boolean;
819
818
  readonly format: "text" | "json";
820
819
  readonly state: "captured" | "disabled" | "unavailable";
821
820
  };
822
- readonly output: {
823
- readonly truncated: boolean;
821
+ readonly error: {
824
822
  readonly text: string;
823
+ readonly truncated: boolean;
825
824
  readonly format: "text" | "json";
826
825
  readonly state: "captured" | "disabled" | "unavailable";
827
- };
826
+ } | null;
827
+ readonly durationMs: number | null;
828
828
  readonly method: "answer" | "transcript" | "turns" | "readText" | "exists" | "exec" | "cli.calls" | "api.calls" | "api.url" | "mcp.calls";
829
829
  }[];
830
830
  readonly logs: readonly {
831
831
  readonly at: number;
832
832
  readonly value: {
833
- readonly truncated: boolean;
834
833
  readonly text: string;
834
+ readonly truncated: boolean;
835
835
  readonly format: "text" | "json";
836
836
  readonly state: "captured" | "disabled" | "unavailable";
837
837
  };
@@ -1023,6 +1023,9 @@ declare const EvalRun: Schema.Struct<{
1023
1023
  incomplete: typeof Schema.Boolean;
1024
1024
  knownActualUsd: typeof Schema.Number;
1025
1025
  }>>;
1026
+ executedBy: Schema.optionalWith<Schema.NullOr<Schema.Literal<["client"]>>, {
1027
+ default: () => null;
1028
+ }>;
1026
1029
  failure: Schema.NullOr<typeof Schema.String>;
1027
1030
  finishedAt: Schema.NullOr<Schema.transformOrFail<Schema.SchemaClass<string, string, never>, typeof Schema.DateTimeUtcFromSelf, never>>;
1028
1031
  id: typeof Schema.String;
@@ -1037,7 +1040,7 @@ declare const EvalRun: Schema.Struct<{
1037
1040
  name: typeof Schema.String;
1038
1041
  version: typeof Schema.String;
1039
1042
  }>>>;
1040
- sandbox: Schema.Literal<["daytona", "e2b", "upstash", "modal", "cloudflare", "vercel"]>;
1043
+ sandbox: Schema.Literal<["daytona", "e2b", "upstash", "modal", "cloudflare", "vercel", "local"]>;
1041
1044
  }>>;
1042
1045
  }>;
1043
1046
  type EvalRun = typeof EvalRun.Type;
@@ -1057,7 +1060,7 @@ declare const EvalRunSummary: Schema.Struct<{
1057
1060
  name: typeof Schema.String;
1058
1061
  version: typeof Schema.String;
1059
1062
  }>>>;
1060
- sandbox: Schema.Literal<["daytona", "e2b", "upstash", "modal", "cloudflare", "vercel"]>;
1063
+ sandbox: Schema.Literal<["daytona", "e2b", "upstash", "modal", "cloudflare", "vercel", "local"]>;
1061
1064
  }>>;
1062
1065
  commandMax: Schema.NullOr<typeof Schema.Int>;
1063
1066
  commandMin: Schema.NullOr<typeof Schema.Int>;
@@ -1101,84 +1104,84 @@ declare const EvalCellHistoryEntry: Schema.Struct<{
1101
1104
  trials: Schema.Array$<Schema.Struct<{
1102
1105
  artifacts: Schema.optional<Schema.Array$<Schema.Struct<{
1103
1106
  path: typeof Schema.String;
1104
- sha256: Schema.filter<typeof Schema.String>;
1105
1107
  byteSize: Schema.filter<typeof Schema.Int>;
1108
+ sha256: Schema.filter<typeof Schema.String>;
1106
1109
  }>>>;
1107
1110
  validations: Schema.optional<Schema.refine<readonly {
1108
1111
  readonly name: string;
1109
1112
  readonly id: string;
1110
1113
  readonly kind: "command" | "judge" | "code";
1111
- readonly status: "passed" | "running" | "failed" | "queued" | "error" | "skipped";
1114
+ readonly status: "passed" | "running" | "failed" | "error" | "queued" | "skipped";
1112
1115
  readonly message: string;
1113
1116
  readonly startedAt: number | null;
1114
1117
  readonly index: number;
1115
- readonly error: {
1118
+ readonly exitCode: number | null;
1119
+ readonly output: {
1120
+ readonly text: string;
1116
1121
  readonly truncated: boolean;
1122
+ readonly format: "text" | "json";
1123
+ readonly state: "captured" | "disabled" | "unavailable";
1124
+ };
1125
+ readonly input: {
1126
+ readonly text: string;
1127
+ readonly truncated: boolean;
1128
+ readonly format: "text" | "json";
1129
+ readonly state: "captured" | "disabled" | "unavailable";
1130
+ };
1131
+ readonly error: {
1117
1132
  readonly text: string;
1133
+ readonly truncated: boolean;
1118
1134
  readonly format: "text" | "json";
1119
1135
  readonly state: "captured" | "disabled" | "unavailable";
1120
1136
  } | null;
1121
1137
  readonly durationMs: number | null;
1122
- readonly exitCode: number | null;
1123
1138
  readonly truncated: boolean;
1124
1139
  readonly judgment?: {
1125
1140
  readonly name: string;
1126
1141
  readonly model: string;
1127
1142
  readonly threshold: number;
1128
- readonly reason: string;
1129
1143
  readonly error: string | null;
1144
+ readonly reason: string;
1130
1145
  readonly durationMs: number;
1131
1146
  readonly evaluator: string;
1132
1147
  readonly score: number | null;
1133
1148
  readonly choice: string | null;
1134
1149
  } | undefined;
1135
- readonly input: {
1136
- readonly truncated: boolean;
1137
- readonly text: string;
1138
- readonly format: "text" | "json";
1139
- readonly state: "captured" | "disabled" | "unavailable";
1140
- };
1141
- readonly output: {
1142
- readonly truncated: boolean;
1143
- readonly text: string;
1144
- readonly format: "text" | "json";
1145
- readonly state: "captured" | "disabled" | "unavailable";
1146
- };
1147
1150
  readonly metadata?: {
1148
- readonly truncated: boolean;
1149
1151
  readonly text: string;
1152
+ readonly truncated: boolean;
1150
1153
  readonly format: "text" | "json";
1151
1154
  readonly state: "captured" | "disabled" | "unavailable";
1152
1155
  } | undefined;
1153
1156
  readonly calls: readonly {
1154
1157
  readonly startedAt: number;
1155
1158
  readonly index: number;
1156
- readonly error: {
1157
- readonly truncated: boolean;
1159
+ readonly output: {
1158
1160
  readonly text: string;
1161
+ readonly truncated: boolean;
1159
1162
  readonly format: "text" | "json";
1160
1163
  readonly state: "captured" | "disabled" | "unavailable";
1161
- } | null;
1162
- readonly durationMs: number | null;
1164
+ };
1163
1165
  readonly input: {
1164
- readonly truncated: boolean;
1165
1166
  readonly text: string;
1167
+ readonly truncated: boolean;
1166
1168
  readonly format: "text" | "json";
1167
1169
  readonly state: "captured" | "disabled" | "unavailable";
1168
1170
  };
1169
- readonly output: {
1170
- readonly truncated: boolean;
1171
+ readonly error: {
1171
1172
  readonly text: string;
1173
+ readonly truncated: boolean;
1172
1174
  readonly format: "text" | "json";
1173
1175
  readonly state: "captured" | "disabled" | "unavailable";
1174
- };
1176
+ } | null;
1177
+ readonly durationMs: number | null;
1175
1178
  readonly method: "answer" | "transcript" | "turns" | "readText" | "exists" | "exec" | "cli.calls" | "api.calls" | "api.url" | "mcp.calls";
1176
1179
  }[];
1177
1180
  readonly logs: readonly {
1178
1181
  readonly at: number;
1179
1182
  readonly value: {
1180
- readonly truncated: boolean;
1181
1183
  readonly text: string;
1184
+ readonly truncated: boolean;
1182
1185
  readonly format: "text" | "json";
1183
1186
  readonly state: "captured" | "disabled" | "unavailable";
1184
1187
  };