@openpond/evals 0.2.0 → 0.4.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CONTRACT.md +24 -5
- package/README.md +27 -8
- package/RELEASING.md +13 -45
- package/dist/compatibility.js +45 -0
- package/dist/conformance.js +7 -11
- package/dist/evidence/authoring.js +1 -1
- package/dist/evidence/conformance.js +1 -1
- package/dist/evidence/contracts.js +1 -2
- package/dist/evidence/eligibility.js +1 -1
- package/dist/graders.js +1 -1
- package/dist/harness.js +33 -103
- package/dist/index.js +3 -1
- package/dist/model-improvement-qualification.js +125 -0
- package/dist/review-conformance.js +209 -0
- package/dist/runs.js +33 -13
- package/dist/tasksets.js +3 -17
- package/dist/types/compatibility.d.ts +11 -0
- package/dist/types/compatibility.d.ts.map +1 -0
- package/dist/types/conformance.d.ts +10 -48
- package/dist/types/conformance.d.ts.map +1 -1
- package/dist/types/evidence/conformance.d.ts +8 -8
- package/dist/types/evidence/contracts.d.ts +8 -8
- package/dist/types/harness.d.ts +3 -471
- package/dist/types/harness.d.ts.map +1 -1
- package/dist/types/index.d.ts +3 -1
- package/dist/types/index.d.ts.map +1 -1
- package/dist/types/model-improvement-qualification.d.ts +222 -0
- package/dist/types/model-improvement-qualification.d.ts.map +1 -0
- package/dist/types/review-conformance.d.ts +607 -0
- package/dist/types/review-conformance.d.ts.map +1 -0
- package/dist/types/runs.d.ts +47 -10
- package/dist/types/runs.d.ts.map +1 -1
- package/dist/types/tasksets.d.ts +4 -35
- package/dist/types/tasksets.d.ts.map +1 -1
- package/package.json +14 -2
- package/dist/common.js +0 -67
- package/dist/sha256.js +0 -91
- package/dist/types/common.d.ts +0 -51
- package/dist/types/common.d.ts.map +0 -1
- package/dist/types/sha256.d.ts +0 -2
- package/dist/types/sha256.d.ts.map +0 -1
package/dist/types/harness.d.ts
CHANGED
|
@@ -1,394 +1,7 @@
|
|
|
1
1
|
import { z } from "zod";
|
|
2
|
-
import { ImmutableArtifactRefSchema } from "
|
|
2
|
+
import { HarnessTraceSchema, ImmutableArtifactRefSchema, type ModelAction, type ToolObservation } from "@openpond/harness";
|
|
3
3
|
import { type AttemptReceipt, type RunManifest } from "./runs.js";
|
|
4
|
-
export
|
|
5
|
-
portable: z.ZodBoolean;
|
|
6
|
-
blockers: z.ZodArray<z.ZodString>;
|
|
7
|
-
localOnlyAssetRefs: z.ZodArray<z.ZodString>;
|
|
8
|
-
hostPrivateAssetRefs: z.ZodArray<z.ZodString>;
|
|
9
|
-
}, z.core.$strict>;
|
|
10
|
-
export declare const AgentSnapshotContentSchema: z.ZodObject<{
|
|
11
|
-
schemaVersion: z.ZodLiteral<"openpond.agentSnapshot.v1">;
|
|
12
|
-
id: z.ZodString;
|
|
13
|
-
profileRelease: z.ZodNullable<z.ZodObject<{
|
|
14
|
-
id: z.ZodString;
|
|
15
|
-
contentHash: z.ZodString;
|
|
16
|
-
}, z.core.$strict>>;
|
|
17
|
-
instructions: z.ZodArray<z.ZodObject<{
|
|
18
|
-
id: z.ZodString;
|
|
19
|
-
path: z.ZodString;
|
|
20
|
-
contentHash: z.ZodString;
|
|
21
|
-
sizeBytes: z.ZodNumber;
|
|
22
|
-
mediaType: z.ZodString;
|
|
23
|
-
visibility: z.ZodEnum<{
|
|
24
|
-
policy: "policy";
|
|
25
|
-
verifier: "verifier";
|
|
26
|
-
host_private: "host_private";
|
|
27
|
-
}>;
|
|
28
|
-
}, z.core.$strict>>;
|
|
29
|
-
skills: z.ZodArray<z.ZodObject<{
|
|
30
|
-
id: z.ZodString;
|
|
31
|
-
path: z.ZodString;
|
|
32
|
-
contentHash: z.ZodString;
|
|
33
|
-
sizeBytes: z.ZodNumber;
|
|
34
|
-
mediaType: z.ZodString;
|
|
35
|
-
visibility: z.ZodEnum<{
|
|
36
|
-
policy: "policy";
|
|
37
|
-
verifier: "verifier";
|
|
38
|
-
host_private: "host_private";
|
|
39
|
-
}>;
|
|
40
|
-
}, z.core.$strict>>;
|
|
41
|
-
agents: z.ZodArray<z.ZodObject<{
|
|
42
|
-
id: z.ZodString;
|
|
43
|
-
path: z.ZodString;
|
|
44
|
-
contentHash: z.ZodString;
|
|
45
|
-
sizeBytes: z.ZodNumber;
|
|
46
|
-
mediaType: z.ZodString;
|
|
47
|
-
visibility: z.ZodEnum<{
|
|
48
|
-
policy: "policy";
|
|
49
|
-
verifier: "verifier";
|
|
50
|
-
host_private: "host_private";
|
|
51
|
-
}>;
|
|
52
|
-
}, z.core.$strict>>;
|
|
53
|
-
toolDeclarations: z.ZodArray<z.ZodObject<{
|
|
54
|
-
name: z.ZodString;
|
|
55
|
-
description: z.ZodString;
|
|
56
|
-
inputSchema: z.ZodRecord<z.ZodString, z.ZodUnknown>;
|
|
57
|
-
inputSchemaHash: z.ZodString;
|
|
58
|
-
sideEffect: z.ZodEnum<{
|
|
59
|
-
read: "read";
|
|
60
|
-
write: "write";
|
|
61
|
-
}>;
|
|
62
|
-
timeoutMs: z.ZodNumber;
|
|
63
|
-
}, z.core.$strict>>;
|
|
64
|
-
capabilityRequirements: z.ZodArray<z.ZodObject<{
|
|
65
|
-
id: z.ZodString;
|
|
66
|
-
required: z.ZodBoolean;
|
|
67
|
-
scopes: z.ZodArray<z.ZodString>;
|
|
68
|
-
}, z.core.$strict>>;
|
|
69
|
-
dependencyLock: z.ZodObject<{
|
|
70
|
-
id: z.ZodString;
|
|
71
|
-
path: z.ZodString;
|
|
72
|
-
contentHash: z.ZodString;
|
|
73
|
-
sizeBytes: z.ZodNumber;
|
|
74
|
-
mediaType: z.ZodString;
|
|
75
|
-
visibility: z.ZodEnum<{
|
|
76
|
-
policy: "policy";
|
|
77
|
-
verifier: "verifier";
|
|
78
|
-
host_private: "host_private";
|
|
79
|
-
}>;
|
|
80
|
-
}, z.core.$strict>;
|
|
81
|
-
portability: z.ZodObject<{
|
|
82
|
-
portable: z.ZodBoolean;
|
|
83
|
-
blockers: z.ZodArray<z.ZodString>;
|
|
84
|
-
localOnlyAssetRefs: z.ZodArray<z.ZodString>;
|
|
85
|
-
hostPrivateAssetRefs: z.ZodArray<z.ZodString>;
|
|
86
|
-
}, z.core.$strict>;
|
|
87
|
-
metadata: z.ZodDefault<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
88
|
-
}, z.core.$strict>;
|
|
89
|
-
export declare const AgentSnapshotSchema: z.ZodObject<{
|
|
90
|
-
schemaVersion: z.ZodLiteral<"openpond.agentSnapshot.v1">;
|
|
91
|
-
id: z.ZodString;
|
|
92
|
-
profileRelease: z.ZodNullable<z.ZodObject<{
|
|
93
|
-
id: z.ZodString;
|
|
94
|
-
contentHash: z.ZodString;
|
|
95
|
-
}, z.core.$strict>>;
|
|
96
|
-
instructions: z.ZodArray<z.ZodObject<{
|
|
97
|
-
id: z.ZodString;
|
|
98
|
-
path: z.ZodString;
|
|
99
|
-
contentHash: z.ZodString;
|
|
100
|
-
sizeBytes: z.ZodNumber;
|
|
101
|
-
mediaType: z.ZodString;
|
|
102
|
-
visibility: z.ZodEnum<{
|
|
103
|
-
policy: "policy";
|
|
104
|
-
verifier: "verifier";
|
|
105
|
-
host_private: "host_private";
|
|
106
|
-
}>;
|
|
107
|
-
}, z.core.$strict>>;
|
|
108
|
-
skills: z.ZodArray<z.ZodObject<{
|
|
109
|
-
id: z.ZodString;
|
|
110
|
-
path: z.ZodString;
|
|
111
|
-
contentHash: z.ZodString;
|
|
112
|
-
sizeBytes: z.ZodNumber;
|
|
113
|
-
mediaType: z.ZodString;
|
|
114
|
-
visibility: z.ZodEnum<{
|
|
115
|
-
policy: "policy";
|
|
116
|
-
verifier: "verifier";
|
|
117
|
-
host_private: "host_private";
|
|
118
|
-
}>;
|
|
119
|
-
}, z.core.$strict>>;
|
|
120
|
-
agents: z.ZodArray<z.ZodObject<{
|
|
121
|
-
id: z.ZodString;
|
|
122
|
-
path: z.ZodString;
|
|
123
|
-
contentHash: z.ZodString;
|
|
124
|
-
sizeBytes: z.ZodNumber;
|
|
125
|
-
mediaType: z.ZodString;
|
|
126
|
-
visibility: z.ZodEnum<{
|
|
127
|
-
policy: "policy";
|
|
128
|
-
verifier: "verifier";
|
|
129
|
-
host_private: "host_private";
|
|
130
|
-
}>;
|
|
131
|
-
}, z.core.$strict>>;
|
|
132
|
-
toolDeclarations: z.ZodArray<z.ZodObject<{
|
|
133
|
-
name: z.ZodString;
|
|
134
|
-
description: z.ZodString;
|
|
135
|
-
inputSchema: z.ZodRecord<z.ZodString, z.ZodUnknown>;
|
|
136
|
-
inputSchemaHash: z.ZodString;
|
|
137
|
-
sideEffect: z.ZodEnum<{
|
|
138
|
-
read: "read";
|
|
139
|
-
write: "write";
|
|
140
|
-
}>;
|
|
141
|
-
timeoutMs: z.ZodNumber;
|
|
142
|
-
}, z.core.$strict>>;
|
|
143
|
-
capabilityRequirements: z.ZodArray<z.ZodObject<{
|
|
144
|
-
id: z.ZodString;
|
|
145
|
-
required: z.ZodBoolean;
|
|
146
|
-
scopes: z.ZodArray<z.ZodString>;
|
|
147
|
-
}, z.core.$strict>>;
|
|
148
|
-
dependencyLock: z.ZodObject<{
|
|
149
|
-
id: z.ZodString;
|
|
150
|
-
path: z.ZodString;
|
|
151
|
-
contentHash: z.ZodString;
|
|
152
|
-
sizeBytes: z.ZodNumber;
|
|
153
|
-
mediaType: z.ZodString;
|
|
154
|
-
visibility: z.ZodEnum<{
|
|
155
|
-
policy: "policy";
|
|
156
|
-
verifier: "verifier";
|
|
157
|
-
host_private: "host_private";
|
|
158
|
-
}>;
|
|
159
|
-
}, z.core.$strict>;
|
|
160
|
-
portability: z.ZodObject<{
|
|
161
|
-
portable: z.ZodBoolean;
|
|
162
|
-
blockers: z.ZodArray<z.ZodString>;
|
|
163
|
-
localOnlyAssetRefs: z.ZodArray<z.ZodString>;
|
|
164
|
-
hostPrivateAssetRefs: z.ZodArray<z.ZodString>;
|
|
165
|
-
}, z.core.$strict>;
|
|
166
|
-
metadata: z.ZodDefault<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
167
|
-
contentHash: z.ZodString;
|
|
168
|
-
}, z.core.$strict>;
|
|
169
|
-
export declare const LifecycleContractSchema: z.ZodObject<{
|
|
170
|
-
create: z.ZodLiteral<true>;
|
|
171
|
-
reset: z.ZodLiteral<true>;
|
|
172
|
-
step: z.ZodLiteral<true>;
|
|
173
|
-
collect: z.ZodLiteral<true>;
|
|
174
|
-
destroy: z.ZodLiteral<true>;
|
|
175
|
-
resetScope: z.ZodEnum<{
|
|
176
|
-
task: "task";
|
|
177
|
-
attempt: "attempt";
|
|
178
|
-
}>;
|
|
179
|
-
}, z.core.$strict>;
|
|
180
|
-
export declare const GraderInterfaceContractSchema: z.ZodObject<{
|
|
181
|
-
visibleEvidence: z.ZodArray<z.ZodString>;
|
|
182
|
-
privilegedEvidence: z.ZodArray<z.ZodString>;
|
|
183
|
-
privateVerifierIsolation: z.ZodBoolean;
|
|
184
|
-
}, z.core.$strict>;
|
|
185
|
-
export declare const HarnessReleaseContentSchema: z.ZodObject<{
|
|
186
|
-
schemaVersion: z.ZodLiteral<"openpond.harnessRelease.v1">;
|
|
187
|
-
id: z.ZodString;
|
|
188
|
-
agentSnapshot: z.ZodObject<{
|
|
189
|
-
id: z.ZodString;
|
|
190
|
-
contentHash: z.ZodString;
|
|
191
|
-
}, z.core.$strict>;
|
|
192
|
-
program: z.ZodObject<{
|
|
193
|
-
id: z.ZodString;
|
|
194
|
-
path: z.ZodString;
|
|
195
|
-
contentHash: z.ZodString;
|
|
196
|
-
sizeBytes: z.ZodNumber;
|
|
197
|
-
mediaType: z.ZodString;
|
|
198
|
-
visibility: z.ZodEnum<{
|
|
199
|
-
policy: "policy";
|
|
200
|
-
verifier: "verifier";
|
|
201
|
-
host_private: "host_private";
|
|
202
|
-
}>;
|
|
203
|
-
}, z.core.$strict>;
|
|
204
|
-
environment: z.ZodObject<{
|
|
205
|
-
protocolVersion: z.ZodLiteral<"openpond.environment.v1">;
|
|
206
|
-
kind: z.ZodEnum<{
|
|
207
|
-
text: "text";
|
|
208
|
-
agent: "agent";
|
|
209
|
-
work: "work";
|
|
210
|
-
custom_program: "custom_program";
|
|
211
|
-
}>;
|
|
212
|
-
entrypoint: z.ZodString;
|
|
213
|
-
stateful: z.ZodBoolean;
|
|
214
|
-
deterministicSeeds: z.ZodBoolean;
|
|
215
|
-
lifecycle: z.ZodArray<z.ZodEnum<{
|
|
216
|
-
create: "create";
|
|
217
|
-
reset: "reset";
|
|
218
|
-
step: "step";
|
|
219
|
-
collect: "collect";
|
|
220
|
-
destroy: "destroy";
|
|
221
|
-
}>>;
|
|
222
|
-
networkPolicy: z.ZodEnum<{
|
|
223
|
-
none: "none";
|
|
224
|
-
declared_read_only: "declared_read_only";
|
|
225
|
-
declared_scoped: "declared_scoped";
|
|
226
|
-
}>;
|
|
227
|
-
defaultTimeoutMs: z.ZodNumber;
|
|
228
|
-
}, z.core.$strict>;
|
|
229
|
-
tools: z.ZodArray<z.ZodObject<{
|
|
230
|
-
name: z.ZodString;
|
|
231
|
-
description: z.ZodString;
|
|
232
|
-
inputSchema: z.ZodRecord<z.ZodString, z.ZodUnknown>;
|
|
233
|
-
inputSchemaHash: z.ZodString;
|
|
234
|
-
sideEffect: z.ZodEnum<{
|
|
235
|
-
read: "read";
|
|
236
|
-
write: "write";
|
|
237
|
-
}>;
|
|
238
|
-
timeoutMs: z.ZodNumber;
|
|
239
|
-
}, z.core.$strict>>;
|
|
240
|
-
lifecycle: z.ZodObject<{
|
|
241
|
-
create: z.ZodLiteral<true>;
|
|
242
|
-
reset: z.ZodLiteral<true>;
|
|
243
|
-
step: z.ZodLiteral<true>;
|
|
244
|
-
collect: z.ZodLiteral<true>;
|
|
245
|
-
destroy: z.ZodLiteral<true>;
|
|
246
|
-
resetScope: z.ZodEnum<{
|
|
247
|
-
task: "task";
|
|
248
|
-
attempt: "attempt";
|
|
249
|
-
}>;
|
|
250
|
-
}, z.core.$strict>;
|
|
251
|
-
graderInterface: z.ZodObject<{
|
|
252
|
-
visibleEvidence: z.ZodArray<z.ZodString>;
|
|
253
|
-
privilegedEvidence: z.ZodArray<z.ZodString>;
|
|
254
|
-
privateVerifierIsolation: z.ZodBoolean;
|
|
255
|
-
}, z.core.$strict>;
|
|
256
|
-
policy: z.ZodObject<{
|
|
257
|
-
policyVisibleFields: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
258
|
-
privilegedFields: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
259
|
-
hiddenGraderRefs: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
260
|
-
connectedAppScopes: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
261
|
-
}, z.core.$strict>;
|
|
262
|
-
files: z.ZodArray<z.ZodObject<{
|
|
263
|
-
id: z.ZodString;
|
|
264
|
-
path: z.ZodString;
|
|
265
|
-
contentHash: z.ZodString;
|
|
266
|
-
sizeBytes: z.ZodNumber;
|
|
267
|
-
mediaType: z.ZodString;
|
|
268
|
-
visibility: z.ZodEnum<{
|
|
269
|
-
policy: "policy";
|
|
270
|
-
verifier: "verifier";
|
|
271
|
-
host_private: "host_private";
|
|
272
|
-
}>;
|
|
273
|
-
}, z.core.$strict>>;
|
|
274
|
-
metadata: z.ZodDefault<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
275
|
-
}, z.core.$strict>;
|
|
276
|
-
export declare const HarnessReleaseSchema: z.ZodObject<{
|
|
277
|
-
schemaVersion: z.ZodLiteral<"openpond.harnessRelease.v1">;
|
|
278
|
-
id: z.ZodString;
|
|
279
|
-
agentSnapshot: z.ZodObject<{
|
|
280
|
-
id: z.ZodString;
|
|
281
|
-
contentHash: z.ZodString;
|
|
282
|
-
}, z.core.$strict>;
|
|
283
|
-
program: z.ZodObject<{
|
|
284
|
-
id: z.ZodString;
|
|
285
|
-
path: z.ZodString;
|
|
286
|
-
contentHash: z.ZodString;
|
|
287
|
-
sizeBytes: z.ZodNumber;
|
|
288
|
-
mediaType: z.ZodString;
|
|
289
|
-
visibility: z.ZodEnum<{
|
|
290
|
-
policy: "policy";
|
|
291
|
-
verifier: "verifier";
|
|
292
|
-
host_private: "host_private";
|
|
293
|
-
}>;
|
|
294
|
-
}, z.core.$strict>;
|
|
295
|
-
environment: z.ZodObject<{
|
|
296
|
-
protocolVersion: z.ZodLiteral<"openpond.environment.v1">;
|
|
297
|
-
kind: z.ZodEnum<{
|
|
298
|
-
text: "text";
|
|
299
|
-
agent: "agent";
|
|
300
|
-
work: "work";
|
|
301
|
-
custom_program: "custom_program";
|
|
302
|
-
}>;
|
|
303
|
-
entrypoint: z.ZodString;
|
|
304
|
-
stateful: z.ZodBoolean;
|
|
305
|
-
deterministicSeeds: z.ZodBoolean;
|
|
306
|
-
lifecycle: z.ZodArray<z.ZodEnum<{
|
|
307
|
-
create: "create";
|
|
308
|
-
reset: "reset";
|
|
309
|
-
step: "step";
|
|
310
|
-
collect: "collect";
|
|
311
|
-
destroy: "destroy";
|
|
312
|
-
}>>;
|
|
313
|
-
networkPolicy: z.ZodEnum<{
|
|
314
|
-
none: "none";
|
|
315
|
-
declared_read_only: "declared_read_only";
|
|
316
|
-
declared_scoped: "declared_scoped";
|
|
317
|
-
}>;
|
|
318
|
-
defaultTimeoutMs: z.ZodNumber;
|
|
319
|
-
}, z.core.$strict>;
|
|
320
|
-
tools: z.ZodArray<z.ZodObject<{
|
|
321
|
-
name: z.ZodString;
|
|
322
|
-
description: z.ZodString;
|
|
323
|
-
inputSchema: z.ZodRecord<z.ZodString, z.ZodUnknown>;
|
|
324
|
-
inputSchemaHash: z.ZodString;
|
|
325
|
-
sideEffect: z.ZodEnum<{
|
|
326
|
-
read: "read";
|
|
327
|
-
write: "write";
|
|
328
|
-
}>;
|
|
329
|
-
timeoutMs: z.ZodNumber;
|
|
330
|
-
}, z.core.$strict>>;
|
|
331
|
-
lifecycle: z.ZodObject<{
|
|
332
|
-
create: z.ZodLiteral<true>;
|
|
333
|
-
reset: z.ZodLiteral<true>;
|
|
334
|
-
step: z.ZodLiteral<true>;
|
|
335
|
-
collect: z.ZodLiteral<true>;
|
|
336
|
-
destroy: z.ZodLiteral<true>;
|
|
337
|
-
resetScope: z.ZodEnum<{
|
|
338
|
-
task: "task";
|
|
339
|
-
attempt: "attempt";
|
|
340
|
-
}>;
|
|
341
|
-
}, z.core.$strict>;
|
|
342
|
-
graderInterface: z.ZodObject<{
|
|
343
|
-
visibleEvidence: z.ZodArray<z.ZodString>;
|
|
344
|
-
privilegedEvidence: z.ZodArray<z.ZodString>;
|
|
345
|
-
privateVerifierIsolation: z.ZodBoolean;
|
|
346
|
-
}, z.core.$strict>;
|
|
347
|
-
policy: z.ZodObject<{
|
|
348
|
-
policyVisibleFields: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
349
|
-
privilegedFields: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
350
|
-
hiddenGraderRefs: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
351
|
-
connectedAppScopes: z.ZodDefault<z.ZodArray<z.ZodString>>;
|
|
352
|
-
}, z.core.$strict>;
|
|
353
|
-
files: z.ZodArray<z.ZodObject<{
|
|
354
|
-
id: z.ZodString;
|
|
355
|
-
path: z.ZodString;
|
|
356
|
-
contentHash: z.ZodString;
|
|
357
|
-
sizeBytes: z.ZodNumber;
|
|
358
|
-
mediaType: z.ZodString;
|
|
359
|
-
visibility: z.ZodEnum<{
|
|
360
|
-
policy: "policy";
|
|
361
|
-
verifier: "verifier";
|
|
362
|
-
host_private: "host_private";
|
|
363
|
-
}>;
|
|
364
|
-
}, z.core.$strict>>;
|
|
365
|
-
metadata: z.ZodDefault<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
366
|
-
contentHash: z.ZodString;
|
|
367
|
-
}, z.core.$strict>;
|
|
368
|
-
export declare const ModelActionSchema: z.ZodObject<{
|
|
369
|
-
id: z.ZodString;
|
|
370
|
-
turn: z.ZodNumber;
|
|
371
|
-
kind: z.ZodEnum<{
|
|
372
|
-
message: "message";
|
|
373
|
-
terminal: "terminal";
|
|
374
|
-
tool_call: "tool_call";
|
|
375
|
-
}>;
|
|
376
|
-
name: z.ZodNullable<z.ZodString>;
|
|
377
|
-
arguments: z.ZodRecord<z.ZodString, z.ZodUnknown>;
|
|
378
|
-
content: z.ZodNullable<z.ZodString>;
|
|
379
|
-
}, z.core.$strict>;
|
|
380
|
-
export declare const ToolObservationSchema: z.ZodObject<{
|
|
381
|
-
actionId: z.ZodString;
|
|
382
|
-
turn: z.ZodNumber;
|
|
383
|
-
terminal: z.ZodBoolean;
|
|
384
|
-
output: z.ZodRecord<z.ZodString, z.ZodUnknown>;
|
|
385
|
-
artifactRefs: z.ZodArray<z.ZodObject<{
|
|
386
|
-
id: z.ZodString;
|
|
387
|
-
contentHash: z.ZodString;
|
|
388
|
-
mediaType: z.ZodDefault<z.ZodNullable<z.ZodString>>;
|
|
389
|
-
sizeBytes: z.ZodDefault<z.ZodNullable<z.ZodNumber>>;
|
|
390
|
-
}, z.core.$strict>>;
|
|
391
|
-
}, z.core.$strict>;
|
|
4
|
+
export * from "@openpond/harness/harness";
|
|
392
5
|
export type HarnessLease = {
|
|
393
6
|
id: string;
|
|
394
7
|
metadata: Record<string, unknown>;
|
|
@@ -419,7 +32,7 @@ export type HarnessExecutionInput = {
|
|
|
419
32
|
};
|
|
420
33
|
export type HarnessExecutionResult = {
|
|
421
34
|
receipt: AttemptReceipt;
|
|
422
|
-
trace:
|
|
35
|
+
trace: z.infer<typeof HarnessTraceSchema>;
|
|
423
36
|
output: Record<string, unknown>;
|
|
424
37
|
};
|
|
425
38
|
export interface HarnessExecutor {
|
|
@@ -443,85 +56,4 @@ export declare function executeRuntimeProtocol(input: {
|
|
|
443
56
|
signal?: AbortSignal;
|
|
444
57
|
now?: () => string;
|
|
445
58
|
}): Promise<HarnessExecutionResult>;
|
|
446
|
-
export declare function createAgentSnapshot(input: z.input<typeof AgentSnapshotContentSchema>): AgentSnapshot;
|
|
447
|
-
export declare function createHarnessRelease(input: z.input<typeof HarnessReleaseContentSchema>): HarnessRelease;
|
|
448
|
-
export declare const HarnessLifecycleEventSchema: z.ZodObject<{
|
|
449
|
-
sequence: z.ZodNumber;
|
|
450
|
-
type: z.ZodEnum<{
|
|
451
|
-
reset: "reset";
|
|
452
|
-
terminal: "terminal";
|
|
453
|
-
created: "created";
|
|
454
|
-
action: "action";
|
|
455
|
-
observation: "observation";
|
|
456
|
-
failure: "failure";
|
|
457
|
-
collected: "collected";
|
|
458
|
-
destroyed: "destroyed";
|
|
459
|
-
}>;
|
|
460
|
-
payloadHash: z.ZodString;
|
|
461
|
-
metadata: z.ZodDefault<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
462
|
-
}, z.core.$strict>;
|
|
463
|
-
export declare const HarnessTraceSchema: z.ZodObject<{
|
|
464
|
-
schemaVersion: z.ZodLiteral<"openpond.harnessTrace.v1">;
|
|
465
|
-
manifest: z.ZodObject<{
|
|
466
|
-
id: z.ZodString;
|
|
467
|
-
contentHash: z.ZodString;
|
|
468
|
-
}, z.core.$strict>;
|
|
469
|
-
taskId: z.ZodString;
|
|
470
|
-
seed: z.ZodString;
|
|
471
|
-
events: z.ZodArray<z.ZodObject<{
|
|
472
|
-
sequence: z.ZodNumber;
|
|
473
|
-
type: z.ZodEnum<{
|
|
474
|
-
reset: "reset";
|
|
475
|
-
terminal: "terminal";
|
|
476
|
-
created: "created";
|
|
477
|
-
action: "action";
|
|
478
|
-
observation: "observation";
|
|
479
|
-
failure: "failure";
|
|
480
|
-
collected: "collected";
|
|
481
|
-
destroyed: "destroyed";
|
|
482
|
-
}>;
|
|
483
|
-
payloadHash: z.ZodString;
|
|
484
|
-
metadata: z.ZodDefault<z.ZodRecord<z.ZodString, z.ZodUnknown>>;
|
|
485
|
-
}, z.core.$strict>>;
|
|
486
|
-
actions: z.ZodArray<z.ZodObject<{
|
|
487
|
-
id: z.ZodString;
|
|
488
|
-
turn: z.ZodNumber;
|
|
489
|
-
kind: z.ZodEnum<{
|
|
490
|
-
message: "message";
|
|
491
|
-
terminal: "terminal";
|
|
492
|
-
tool_call: "tool_call";
|
|
493
|
-
}>;
|
|
494
|
-
name: z.ZodNullable<z.ZodString>;
|
|
495
|
-
arguments: z.ZodRecord<z.ZodString, z.ZodUnknown>;
|
|
496
|
-
content: z.ZodNullable<z.ZodString>;
|
|
497
|
-
}, z.core.$strict>>;
|
|
498
|
-
observations: z.ZodArray<z.ZodObject<{
|
|
499
|
-
actionId: z.ZodString;
|
|
500
|
-
turn: z.ZodNumber;
|
|
501
|
-
terminal: z.ZodBoolean;
|
|
502
|
-
output: z.ZodRecord<z.ZodString, z.ZodUnknown>;
|
|
503
|
-
artifactRefs: z.ZodArray<z.ZodObject<{
|
|
504
|
-
id: z.ZodString;
|
|
505
|
-
contentHash: z.ZodString;
|
|
506
|
-
mediaType: z.ZodDefault<z.ZodNullable<z.ZodString>>;
|
|
507
|
-
sizeBytes: z.ZodDefault<z.ZodNullable<z.ZodNumber>>;
|
|
508
|
-
}, z.core.$strict>>;
|
|
509
|
-
}, z.core.$strict>>;
|
|
510
|
-
terminal: z.ZodBoolean;
|
|
511
|
-
failureClass: z.ZodNullable<z.ZodEnum<{
|
|
512
|
-
policy_failure: "policy_failure";
|
|
513
|
-
grader_failure: "grader_failure";
|
|
514
|
-
environment_failure: "environment_failure";
|
|
515
|
-
infrastructure_failure: "infrastructure_failure";
|
|
516
|
-
timeout: "timeout";
|
|
517
|
-
cancelled: "cancelled";
|
|
518
|
-
}>>;
|
|
519
|
-
output: z.ZodRecord<z.ZodString, z.ZodUnknown>;
|
|
520
|
-
contentHash: z.ZodString;
|
|
521
|
-
}, z.core.$strict>;
|
|
522
|
-
export type AgentSnapshot = z.infer<typeof AgentSnapshotSchema>;
|
|
523
|
-
export type HarnessRelease = z.infer<typeof HarnessReleaseSchema>;
|
|
524
|
-
export type ModelAction = z.infer<typeof ModelActionSchema>;
|
|
525
|
-
export type ToolObservation = z.infer<typeof ToolObservationSchema>;
|
|
526
|
-
export type HarnessTrace = z.infer<typeof HarnessTraceSchema>;
|
|
527
59
|
//# sourceMappingURL=harness.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"harness.d.ts","sourceRoot":"","sources":["../../../../src/harness.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAExB,OAAO,
|
|
1
|
+
{"version":3,"file":"harness.d.ts","sourceRoot":"","sources":["../../../../src/harness.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,CAAC,EAAE,MAAM,KAAK,CAAC;AAExB,OAAO,EAGL,kBAAkB,EAClB,0BAA0B,EAI1B,KAAK,WAAW,EAChB,KAAK,eAAe,EACrB,MAAM,mBAAmB,CAAC;AAE3B,OAAO,EAEL,KAAK,cAAc,EACnB,KAAK,WAAW,EACjB,MAAM,WAAW,CAAC;AAEnB,cAAc,2BAA2B,CAAC;AAE1C,MAAM,MAAM,YAAY,GAAG;IAAE,EAAE,EAAE,MAAM,CAAC;IAAC,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAA;CAAE,CAAC;AAE7E,MAAM,WAAW,cAAc;IAC7B,MAAM,CAAC,KAAK,EAAE;QACZ,QAAQ,EAAE,WAAW,CAAC;QACtB,MAAM,EAAE,MAAM,CAAC;QACf,IAAI,EAAE,MAAM,CAAC;QACb,MAAM,EAAE,WAAW,CAAC;KACrB,GAAG,OAAO,CAAC,YAAY,CAAC,CAAC;IAC1B,KAAK,CACH,KAAK,EAAE,YAAY,EACnB,KAAK,EAAE;QAAE,IAAI,EAAE,MAAM,CAAC;QAAC,MAAM,EAAE,WAAW,CAAA;KAAE,GAC3C,OAAO,CAAC,IAAI,CAAC,CAAC;IACjB,IAAI,CACF,KAAK,EAAE,YAAY,EACnB,MAAM,EAAE,WAAW,EACnB,MAAM,EAAE,WAAW,GAClB,OAAO,CAAC,eAAe,CAAC,CAAC;IAC5B,OAAO,CACL,KAAK,EAAE,YAAY,EACnB,MAAM,EAAE,WAAW,GAClB,OAAO,CAAC;QACT,YAAY,EAAE,CAAC,CAAC,KAAK,CAAC,OAAO,0BAA0B,CAAC,EAAE,CAAC;QAC3D,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;KACnC,CAAC,CAAC;IACH,OAAO,CAAC,KAAK,EAAE,YAAY,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;CAC7C;AAED,MAAM,MAAM,qBAAqB,GAAG;IAClC,QAAQ,EAAE,WAAW,CAAC;IACtB,MAAM,EAAE,MAAM,CAAC;IACf,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,CAAC,EAAE,WAAW,CAAC;CACtB,CAAC;AAEF,MAAM,MAAM,sBAAsB,GAAG;IACnC,OAAO,EAAE,cAAc,CAAC;IACxB,KAAK,EAAE,CAAC,CAAC,KAAK,CAAC,OAAO,kBAAkB,CAAC,CAAC;IAC1C,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;CACjC,CAAC;AAEF,MAAM,WAAW,eAAe;IAC9B,OAAO,CAAC,KAAK,EAAE,qBAAqB,GAAG,OAAO,CAAC,sBAAsB,CAAC,CAAC;CACxE;AAED,MAAM,WAAW,gBAAgB;IAC/B,GAAG,CAAC,KAAK,EAAE;QACT,QAAQ,EAAE,WAAW,CAAC;QACtB,OAAO,EAAE,MAAM,EAAE,CAAC;QAClB,KAAK,EAAE,MAAM,EAAE,CAAC;QAChB,WAAW,EAAE,MAAM,CAAC;QACpB,MAAM,CAAC,EAAE,WAAW,CAAC;KACtB,GAAG,OAAO,CAAC,sBAAsB,EAAE,CAAC,CAAC;CACvC;AAED,wBAAsB,sBAAsB,CAAC,KAAK,EAAE;IAClD,QAAQ,EAAE,WAAW,CAAC;IACtB,MAAM,EAAE,MAAM,CAAC;IACf,IAAI,EAAE,MAAM,CAAC;IACb,OAAO,EAAE,WAAW,EAAE,CAAC;IACvB,OAAO,EAAE,cAAc,CAAC;IACxB,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB,GAAG,CAAC,EAAE,MAAM,MAAM,CAAC;CACpB,GAAG,OAAO,CAAC,sBAAsB,CAAC,CA8JlC"}
|
package/dist/types/index.d.ts
CHANGED
|
@@ -1,7 +1,9 @@
|
|
|
1
|
-
export * from "./
|
|
1
|
+
export * from "./compatibility.js";
|
|
2
2
|
export * from "./evidence/index.js";
|
|
3
3
|
export * from "./graders.js";
|
|
4
4
|
export * from "./harness.js";
|
|
5
5
|
export * from "./runs.js";
|
|
6
|
+
export * from "./model-improvement-qualification.js";
|
|
7
|
+
export * from "./review-conformance.js";
|
|
6
8
|
export * from "./tasksets.js";
|
|
7
9
|
//# sourceMappingURL=index.d.ts.map
|
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../../src/index.ts"],"names":[],"mappings":"AAAA,cAAc,
|
|
1
|
+
{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../../../src/index.ts"],"names":[],"mappings":"AAAA,cAAc,oBAAoB,CAAC;AACnC,cAAc,qBAAqB,CAAC;AACpC,cAAc,cAAc,CAAC;AAC7B,cAAc,cAAc,CAAC;AAC7B,cAAc,WAAW,CAAC;AAC1B,cAAc,sCAAsC,CAAC;AACrD,cAAc,yBAAyB,CAAC;AACxC,cAAc,eAAe,CAAC"}
|