openmerit 0.1.4 → 0.1.6-preview.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/CHANGELOG.md +40 -0
- package/README.md +121 -386
- package/dist/core/src/index.d.ts +101 -0
- package/dist/core/src/index.js +1649 -0
- package/dist/core/src/store.d.ts +35 -0
- package/dist/core/src/store.js +102 -0
- package/dist/pi/src/index.d.ts +32 -0
- package/dist/pi/src/index.js +794 -0
- package/dist/pi/src/scheduler.d.ts +11 -0
- package/dist/pi/src/scheduler.js +137 -0
- package/dist/pi/src/wakeup.d.ts +2 -0
- package/dist/pi/src/wakeup.js +108 -0
- package/dist/protocol/src/index.d.ts +484 -0
- package/dist/protocol/src/index.js +47 -0
- package/dist/protocol/src/schemas.d.ts +576 -0
- package/dist/protocol/src/schemas.js +280 -0
- package/dist/terminal/public/app.js +297 -0
- package/dist/terminal/public/brands/anthropic.png +0 -0
- package/dist/terminal/public/brands/baai.png +0 -0
- package/dist/terminal/public/brands/baseten.png +0 -0
- package/dist/terminal/public/brands/cerebras.png +0 -0
- package/dist/terminal/public/brands/cohere.png +0 -0
- package/dist/terminal/public/brands/deepseek.ico +0 -0
- package/dist/terminal/public/brands/google.png +0 -0
- package/dist/terminal/public/brands/groq.ico +0 -0
- package/dist/terminal/public/brands/lm-studio.png +0 -0
- package/dist/terminal/public/brands/meta.ico +0 -0
- package/dist/terminal/public/brands/mistral.png +0 -0
- package/dist/terminal/public/brands/nomic.png +0 -0
- package/dist/terminal/public/brands/ollama.png +0 -0
- package/dist/terminal/public/brands/openai.png +0 -0
- package/dist/terminal/public/brands/openrouter.png +0 -0
- package/dist/terminal/public/brands/qwen.png +0 -0
- package/dist/terminal/public/brands/vllm.ico +0 -0
- package/dist/terminal/public/brands/vllm.png +0 -0
- package/dist/terminal/public/favicon.svg +1 -0
- package/dist/terminal/public/flow.css +1 -0
- package/dist/terminal/public/flow.js +770 -0
- package/dist/terminal/public/index.html +21 -0
- package/dist/terminal/public/styles.css +779 -0
- package/dist/terminal/src/activity-merge.mjs +64 -0
- package/dist/terminal/src/browser.mjs +29 -0
- package/dist/terminal/src/cli.mjs +60 -0
- package/dist/terminal/src/collect.mjs +311 -0
- package/dist/terminal/src/discovery.mjs +93 -0
- package/dist/terminal/src/hardware.mjs +57 -0
- package/dist/terminal/src/project-activity.mjs +156 -0
- package/dist/terminal/src/sample.mjs +171 -0
- package/dist/terminal/src/server.mjs +56 -0
- package/dist/terminal/src/services.mjs +62 -0
- package/dist/terminal/src/topology.mjs +30 -0
- package/docs/adapter-guide.md +189 -0
- package/docs/architecture.md +59 -0
- package/docs/automation.md +74 -0
- package/docs/budgets.md +37 -0
- package/docs/commands.md +85 -0
- package/docs/demo-backfill.md +29 -0
- package/docs/demo-fieldkit.md +47 -0
- package/docs/demo-placement.md +30 -0
- package/docs/demo-spam.md +15 -0
- package/docs/demo-support.md +42 -0
- package/docs/demo.md +57 -0
- package/docs/first-trial.md +60 -0
- package/docs/getting-started.md +65 -0
- package/docs/index.md +40 -0
- package/docs/inference-terminal.md +439 -0
- package/docs/lifecycle.md +30 -0
- package/docs/memo.md +126 -0
- package/docs/metrics-and-evidence.md +48 -0
- package/docs/operations.md +40 -0
- package/docs/pareto-spec.md +76 -0
- package/docs/pi-extension.md +54 -0
- package/docs/roadmap.md +28 -0
- package/docs/security.md +37 -0
- package/docs/site-artwork-linocut.md +23 -0
- package/docs/site-artwork-miniature-diverse.md +28 -0
- package/docs/site-artwork-miniature.md +26 -0
- package/docs/site-demo.md +177 -0
- package/docs/site-design.md +94 -0
- package/docs/site-documentation.md +83 -0
- package/docs/site-dynamic-og.md +35 -0
- package/docs/site-faq-maintenance.md +115 -0
- package/docs/site-hero-resolution.md +60 -0
- package/docs/site-illustration-sequences.md +227 -0
- package/docs/site-inference-terminal.md +203 -0
- package/docs/site-memo.md +39 -0
- package/docs/site-og-image.md +38 -0
- package/docs/site-og-workshop.md +21 -0
- package/docs/site-section-artwork.md +56 -0
- package/docs/site-skill-review.md +57 -0
- package/docs/site-terminal-preview.md +85 -0
- package/docs/testing.md +118 -0
- package/docs/troubleshooting.md +55 -0
- package/docs/ux-reference.md +32 -0
- package/package.json +74 -42
- package/benchmark/invoice_ocr/data/invoice_01_ground_truth.json +0 -38
- package/benchmark/invoice_ocr/data/invoice_01_row_2.jpg +0 -0
- package/benchmark/invoice_ocr/data/invoice_02_ground_truth.json +0 -32
- package/benchmark/invoice_ocr/data/invoice_02_row_5.jpg +0 -0
- package/benchmark/invoice_ocr/data/invoice_03_ground_truth.json +0 -26
- package/benchmark/invoice_ocr/data/invoice_03_row_6.jpg +0 -0
- package/benchmark/invoice_ocr/data/invoice_04_ground_truth.json +0 -26
- package/benchmark/invoice_ocr/data/invoice_04_row_7.jpg +0 -0
- package/benchmark/invoice_ocr/data/invoice_05_ground_truth.json +0 -38
- package/benchmark/invoice_ocr/data/invoice_05_row_947.jpg +0 -0
- package/benchmark/invoice_ocr/data/invoice_06_ground_truth.json +0 -38
- package/benchmark/invoice_ocr/data/invoice_06_row_948.jpg +0 -0
- package/benchmark/invoice_ocr/data/invoice_07_ground_truth.json +0 -20
- package/benchmark/invoice_ocr/data/invoice_07_row_949.jpg +0 -0
- package/benchmark/invoice_ocr/data/invoice_08_ground_truth.json +0 -38
- package/benchmark/invoice_ocr/data/invoice_08_row_1888.jpg +0 -0
- package/benchmark/invoice_ocr/data/invoice_09_ground_truth.json +0 -26
- package/benchmark/invoice_ocr/data/invoice_09_row_1890.jpg +0 -0
- package/benchmark/invoice_ocr/data/invoice_10_ground_truth.json +0 -20
- package/benchmark/invoice_ocr/data/invoice_10_row_1892.jpg +0 -0
- package/benchmark/invoice_ocr/data/manifest.json +0 -97
- package/dist/benchmarks.js +0 -98
- package/dist/catalog.js +0 -61
- package/dist/cli.js +0 -188
- package/dist/daemon.js +0 -407
- package/dist/diagnostics.js +0 -227
- package/dist/frontier.js +0 -56
- package/dist/harness.js +0 -1
- package/dist/integrations.js +0 -19
- package/dist/invoice-eval.js +0 -33
- package/dist/invoice-score.js +0 -124
- package/dist/judge.js +0 -43
- package/dist/llm.js +0 -207
- package/dist/pi-config.js +0 -46
- package/dist/pi-trials.js +0 -373
- package/dist/policy.js +0 -185
- package/dist/providers.js +0 -1
- package/dist/recommend.js +0 -76
- package/dist/routes.js +0 -74
- package/dist/standalone.js +0 -224
- package/dist/store.js +0 -89
- package/dist/strategist.js +0 -68
- package/dist/task-input.js +0 -54
- package/dist/traces.js +0 -127
- package/dist/trials.js +0 -140
- package/dist/types.js +0 -2
- package/examples/invoice-prompt.txt +0 -19
- package/examples/task.example.json +0 -7
- package/extension/openmerit.ts +0 -947
- package/instructions/OPENMERIT.md +0 -63
- package/instructions/openmerit.policy.json +0 -37
- package/rules.md +0 -43
|
@@ -0,0 +1,576 @@
|
|
|
1
|
+
import { Type, type TSchema } from "typebox";
|
|
2
|
+
import type { IntentKind } from "./index.ts";
|
|
3
|
+
export declare const OPENMERIT_SCHEMA_DIALECT: "https://json-schema.org/draft/2020-12/schema";
|
|
4
|
+
export declare const OPENMERIT_RESULT_SCHEMA_VERSION: "0.2";
|
|
5
|
+
export declare const EvidenceReferenceSchema: Type.TObject<{
|
|
6
|
+
id: Type.TString;
|
|
7
|
+
source: Type.TUnion<Type.TLiteral<string>[]>;
|
|
8
|
+
uri: Type.TString;
|
|
9
|
+
mediaType: Type.TOptional<Type.TString>;
|
|
10
|
+
digest: Type.TOptional<Type.TString>;
|
|
11
|
+
}>;
|
|
12
|
+
export declare const MetricObjectiveSchema: Type.TObject<{
|
|
13
|
+
metricId: Type.TUnion<Type.TLiteral<string>[]>;
|
|
14
|
+
required: Type.TBoolean;
|
|
15
|
+
minimumSamples: Type.TInteger;
|
|
16
|
+
tolerance: Type.TNumber;
|
|
17
|
+
samplingPlan: Type.TObject<{
|
|
18
|
+
strategy: Type.TUnion<Type.TLiteral<string>[]>;
|
|
19
|
+
representativeCaseCount: Type.TInteger;
|
|
20
|
+
repetitionsPerCase: Type.TInteger;
|
|
21
|
+
aggregation: Type.TUnion<Type.TLiteral<string>[]>;
|
|
22
|
+
rationale: Type.TString;
|
|
23
|
+
}>;
|
|
24
|
+
constraint: Type.TOptional<Type.TObject<{
|
|
25
|
+
operator: Type.TUnion<Type.TLiteral<string>[]>;
|
|
26
|
+
value: Type.TNumber;
|
|
27
|
+
}>>;
|
|
28
|
+
}>;
|
|
29
|
+
export declare const ApplicationLlmTargetSchema: Type.TObject<{
|
|
30
|
+
id: Type.TString;
|
|
31
|
+
kind: Type.TLiteral<"application_llm_call">;
|
|
32
|
+
name: Type.TString;
|
|
33
|
+
applicationId: Type.TString;
|
|
34
|
+
callSites: Type.TArray<Type.TString>;
|
|
35
|
+
routeKey: Type.TString;
|
|
36
|
+
confirmedAt: Type.TString;
|
|
37
|
+
}>;
|
|
38
|
+
export declare const TaskProfileSchema: Type.TObject<{
|
|
39
|
+
id: Type.TString;
|
|
40
|
+
revision: Type.TInteger;
|
|
41
|
+
name: Type.TString;
|
|
42
|
+
goal: Type.TString;
|
|
43
|
+
applicationTarget: Type.TObject<{
|
|
44
|
+
id: Type.TString;
|
|
45
|
+
kind: Type.TLiteral<"application_llm_call">;
|
|
46
|
+
name: Type.TString;
|
|
47
|
+
applicationId: Type.TString;
|
|
48
|
+
callSites: Type.TArray<Type.TString>;
|
|
49
|
+
routeKey: Type.TString;
|
|
50
|
+
confirmedAt: Type.TString;
|
|
51
|
+
}>;
|
|
52
|
+
objectives: Type.TArray<Type.TObject<{
|
|
53
|
+
metricId: Type.TUnion<Type.TLiteral<string>[]>;
|
|
54
|
+
required: Type.TBoolean;
|
|
55
|
+
minimumSamples: Type.TInteger;
|
|
56
|
+
tolerance: Type.TNumber;
|
|
57
|
+
samplingPlan: Type.TObject<{
|
|
58
|
+
strategy: Type.TUnion<Type.TLiteral<string>[]>;
|
|
59
|
+
representativeCaseCount: Type.TInteger;
|
|
60
|
+
repetitionsPerCase: Type.TInteger;
|
|
61
|
+
aggregation: Type.TUnion<Type.TLiteral<string>[]>;
|
|
62
|
+
rationale: Type.TString;
|
|
63
|
+
}>;
|
|
64
|
+
constraint: Type.TOptional<Type.TObject<{
|
|
65
|
+
operator: Type.TUnion<Type.TLiteral<string>[]>;
|
|
66
|
+
value: Type.TNumber;
|
|
67
|
+
}>>;
|
|
68
|
+
}>>;
|
|
69
|
+
createdAt: Type.TString;
|
|
70
|
+
confirmedAt: Type.TOptional<Type.TString>;
|
|
71
|
+
}>;
|
|
72
|
+
export declare const ObservabilityCoverageSchema: Type.TObject<{
|
|
73
|
+
targetId: Type.TString;
|
|
74
|
+
status: Type.TUnion<Type.TLiteral<string>[]>;
|
|
75
|
+
requiredMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
|
|
76
|
+
coveredMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
|
|
77
|
+
missingMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
|
|
78
|
+
checkedAt: Type.TString;
|
|
79
|
+
}>;
|
|
80
|
+
export declare const SupervisedGraduationPolicySchema: Type.TObject<{
|
|
81
|
+
mode: Type.TLiteral<"supervised_graduation">;
|
|
82
|
+
evaluationBudget: Type.TObject<{
|
|
83
|
+
currency: Type.TString;
|
|
84
|
+
maximumSpend: Type.TNumber;
|
|
85
|
+
maximumCandidateCount: Type.TOptional<Type.TInteger>;
|
|
86
|
+
maximumRunsPerCandidate: Type.TOptional<Type.TInteger>;
|
|
87
|
+
expiresAt: Type.TOptional<Type.TString>;
|
|
88
|
+
}>;
|
|
89
|
+
automaticSwapsEnabled: Type.TBoolean;
|
|
90
|
+
confirmationRequiredUntilVerifiedSwaps: Type.TInteger;
|
|
91
|
+
requirePostSwapVerification: Type.TLiteral<true>;
|
|
92
|
+
rollbackOnRegression: Type.TBoolean;
|
|
93
|
+
checkPolicy: Type.TObject<{
|
|
94
|
+
baselineAssessment: Type.TObject<{
|
|
95
|
+
afterCompletedTasks: Type.TOptional<Type.TInteger>;
|
|
96
|
+
afterElapsedSeconds: Type.TOptional<Type.TInteger>;
|
|
97
|
+
}>;
|
|
98
|
+
frontierReassessment: Type.TObject<{
|
|
99
|
+
afterCompletedTasks: Type.TOptional<Type.TInteger>;
|
|
100
|
+
afterElapsedSeconds: Type.TOptional<Type.TInteger>;
|
|
101
|
+
onModelCatalogChange: Type.TBoolean;
|
|
102
|
+
}>;
|
|
103
|
+
regressionCheck: Type.TObject<{
|
|
104
|
+
afterCompletedTasks: Type.TOptional<Type.TInteger>;
|
|
105
|
+
afterElapsedSeconds: Type.TOptional<Type.TInteger>;
|
|
106
|
+
thresholds: Type.TArray<Type.TObject<{
|
|
107
|
+
metricId: Type.TUnion<Type.TLiteral<string>[]>;
|
|
108
|
+
relativeChangeAtLeast: Type.TNumber;
|
|
109
|
+
}>>;
|
|
110
|
+
}>;
|
|
111
|
+
postSwapVerification: Type.TObject<{
|
|
112
|
+
afterCompletedTasks: Type.TOptional<Type.TInteger>;
|
|
113
|
+
afterElapsedSeconds: Type.TOptional<Type.TInteger>;
|
|
114
|
+
}>;
|
|
115
|
+
cooldownSeconds: Type.TInteger;
|
|
116
|
+
execution: Type.TObject<{
|
|
117
|
+
mode: Type.TUnion<Type.TLiteral<string>[]>;
|
|
118
|
+
infrastructureChangesAllowed: Type.TBoolean;
|
|
119
|
+
}>;
|
|
120
|
+
}>;
|
|
121
|
+
}>;
|
|
122
|
+
export declare const MetricRecordSchema: Type.TObject<{
|
|
123
|
+
targetId: Type.TString;
|
|
124
|
+
metricId: Type.TUnion<Type.TLiteral<string>[]>;
|
|
125
|
+
state: Type.TUnion<Type.TLiteral<string>[]>;
|
|
126
|
+
scope: Type.TUnion<Type.TLiteral<string>[]>;
|
|
127
|
+
direction: Type.TUnion<Type.TLiteral<string>[]>;
|
|
128
|
+
method: Type.TUnion<Type.TLiteral<string>[]>;
|
|
129
|
+
aggregation: Type.TUnion<Type.TLiteral<string>[]>;
|
|
130
|
+
value: Type.TOptional<Type.TNumber>;
|
|
131
|
+
unit: Type.TString;
|
|
132
|
+
sampleCount: Type.TInteger;
|
|
133
|
+
interval: Type.TOptional<Type.TObject<{
|
|
134
|
+
lower: Type.TNumber;
|
|
135
|
+
upper: Type.TNumber;
|
|
136
|
+
confidenceLevel: Type.TOptional<Type.TNumber>;
|
|
137
|
+
}>>;
|
|
138
|
+
window: Type.TOptional<Type.TObject<{
|
|
139
|
+
startedAt: Type.TString;
|
|
140
|
+
endedAt: Type.TString;
|
|
141
|
+
}>>;
|
|
142
|
+
evidence: Type.TArray<Type.TObject<{
|
|
143
|
+
id: Type.TString;
|
|
144
|
+
source: Type.TUnion<Type.TLiteral<string>[]>;
|
|
145
|
+
uri: Type.TString;
|
|
146
|
+
mediaType: Type.TOptional<Type.TString>;
|
|
147
|
+
digest: Type.TOptional<Type.TString>;
|
|
148
|
+
}>>;
|
|
149
|
+
note: Type.TOptional<Type.TString>;
|
|
150
|
+
}>;
|
|
151
|
+
export declare const CandidateDescriptorSchema: Type.TObject<{
|
|
152
|
+
targetId: Type.TString;
|
|
153
|
+
candidateId: Type.TString;
|
|
154
|
+
modelId: Type.TString;
|
|
155
|
+
modelVersion: Type.TOptional<Type.TString>;
|
|
156
|
+
reasons: Type.TArray<Type.TString>;
|
|
157
|
+
}>;
|
|
158
|
+
export declare const PreliminaryModelLeadSchema: Type.TObject<{
|
|
159
|
+
targetId: Type.TString;
|
|
160
|
+
modelId: Type.TString;
|
|
161
|
+
modelVersion: Type.TOptional<Type.TString>;
|
|
162
|
+
estimatedPrice: Type.TObject<{
|
|
163
|
+
currency: Type.TString;
|
|
164
|
+
inputPerMillionTokens: Type.TNumber;
|
|
165
|
+
outputPerMillionTokens: Type.TNumber;
|
|
166
|
+
sourceUri: Type.TString;
|
|
167
|
+
checkedAt: Type.TString;
|
|
168
|
+
}>;
|
|
169
|
+
reputation: Type.TObject<{
|
|
170
|
+
summary: Type.TString;
|
|
171
|
+
sourceUri: Type.TString;
|
|
172
|
+
checkedAt: Type.TString;
|
|
173
|
+
}>;
|
|
174
|
+
}>;
|
|
175
|
+
export declare const CandidateAssessmentSchema: Type.TObject<{
|
|
176
|
+
id: Type.TString;
|
|
177
|
+
targetId: Type.TString;
|
|
178
|
+
taskProfileId: Type.TString;
|
|
179
|
+
taskProfileRevision: Type.TInteger;
|
|
180
|
+
candidateId: Type.TString;
|
|
181
|
+
modelId: Type.TString;
|
|
182
|
+
modelVersion: Type.TOptional<Type.TString>;
|
|
183
|
+
productRevision: Type.TString;
|
|
184
|
+
harnessId: Type.TString;
|
|
185
|
+
startedAt: Type.TString;
|
|
186
|
+
completedAt: Type.TString;
|
|
187
|
+
metrics: Type.TArray<Type.TObject<{
|
|
188
|
+
targetId: Type.TString;
|
|
189
|
+
metricId: Type.TUnion<Type.TLiteral<string>[]>;
|
|
190
|
+
state: Type.TUnion<Type.TLiteral<string>[]>;
|
|
191
|
+
scope: Type.TUnion<Type.TLiteral<string>[]>;
|
|
192
|
+
direction: Type.TUnion<Type.TLiteral<string>[]>;
|
|
193
|
+
method: Type.TUnion<Type.TLiteral<string>[]>;
|
|
194
|
+
aggregation: Type.TUnion<Type.TLiteral<string>[]>;
|
|
195
|
+
value: Type.TOptional<Type.TNumber>;
|
|
196
|
+
unit: Type.TString;
|
|
197
|
+
sampleCount: Type.TInteger;
|
|
198
|
+
interval: Type.TOptional<Type.TObject<{
|
|
199
|
+
lower: Type.TNumber;
|
|
200
|
+
upper: Type.TNumber;
|
|
201
|
+
confidenceLevel: Type.TOptional<Type.TNumber>;
|
|
202
|
+
}>>;
|
|
203
|
+
window: Type.TOptional<Type.TObject<{
|
|
204
|
+
startedAt: Type.TString;
|
|
205
|
+
endedAt: Type.TString;
|
|
206
|
+
}>>;
|
|
207
|
+
evidence: Type.TArray<Type.TObject<{
|
|
208
|
+
id: Type.TString;
|
|
209
|
+
source: Type.TUnion<Type.TLiteral<string>[]>;
|
|
210
|
+
uri: Type.TString;
|
|
211
|
+
mediaType: Type.TOptional<Type.TString>;
|
|
212
|
+
digest: Type.TOptional<Type.TString>;
|
|
213
|
+
}>>;
|
|
214
|
+
note: Type.TOptional<Type.TString>;
|
|
215
|
+
}>>;
|
|
216
|
+
}>;
|
|
217
|
+
export declare const FrontierSnapshotSchema: Type.TObject<{
|
|
218
|
+
id: Type.TString;
|
|
219
|
+
targetId: Type.TString;
|
|
220
|
+
taskProfileId: Type.TString;
|
|
221
|
+
taskProfileRevision: Type.TInteger;
|
|
222
|
+
productRevision: Type.TString;
|
|
223
|
+
createdAt: Type.TString;
|
|
224
|
+
frontierCandidateIds: Type.TArray<Type.TString>;
|
|
225
|
+
dominatedCandidateIds: Type.TArray<Type.TString>;
|
|
226
|
+
ineligibleCandidateIds: Type.TArray<Type.TString>;
|
|
227
|
+
unresolvedCandidateIds: Type.TArray<Type.TString>;
|
|
228
|
+
comparisons: Type.TArray<Type.TObject<{
|
|
229
|
+
candidateId: Type.TString;
|
|
230
|
+
otherCandidateId: Type.TString;
|
|
231
|
+
state: Type.TUnion<Type.TLiteral<string>[]>;
|
|
232
|
+
reasons: Type.TArray<Type.TString>;
|
|
233
|
+
}>>;
|
|
234
|
+
assessmentIds: Type.TArray<Type.TString>;
|
|
235
|
+
calculationEvidence: Type.TArray<Type.TObject<{
|
|
236
|
+
id: Type.TString;
|
|
237
|
+
source: Type.TUnion<Type.TLiteral<string>[]>;
|
|
238
|
+
uri: Type.TString;
|
|
239
|
+
mediaType: Type.TOptional<Type.TString>;
|
|
240
|
+
digest: Type.TOptional<Type.TString>;
|
|
241
|
+
}>>;
|
|
242
|
+
}>;
|
|
243
|
+
export declare const FrontierRecommendationSchema: Type.TObject<{
|
|
244
|
+
targetId: Type.TString;
|
|
245
|
+
frontierSnapshotId: Type.TString;
|
|
246
|
+
selectedCandidateId: Type.TString;
|
|
247
|
+
currentCandidateId: Type.TString;
|
|
248
|
+
reasons: Type.TArray<Type.TString>;
|
|
249
|
+
preferenceEvidence: Type.TArray<Type.TObject<{
|
|
250
|
+
id: Type.TString;
|
|
251
|
+
source: Type.TUnion<Type.TLiteral<string>[]>;
|
|
252
|
+
uri: Type.TString;
|
|
253
|
+
mediaType: Type.TOptional<Type.TString>;
|
|
254
|
+
digest: Type.TOptional<Type.TString>;
|
|
255
|
+
}>>;
|
|
256
|
+
}>;
|
|
257
|
+
export declare const INTENT_OUTPUT_SCHEMAS: {
|
|
258
|
+
readonly establish_evals: Type.TObject<{
|
|
259
|
+
taskProfile: Type.TObject<{
|
|
260
|
+
confirmedAt: Type.TString;
|
|
261
|
+
id: Type.TString;
|
|
262
|
+
revision: Type.TInteger;
|
|
263
|
+
name: Type.TString;
|
|
264
|
+
goal: Type.TString;
|
|
265
|
+
applicationTarget: Type.TObject<{
|
|
266
|
+
id: Type.TString;
|
|
267
|
+
kind: Type.TLiteral<"application_llm_call">;
|
|
268
|
+
name: Type.TString;
|
|
269
|
+
applicationId: Type.TString;
|
|
270
|
+
callSites: Type.TArray<Type.TString>;
|
|
271
|
+
routeKey: Type.TString;
|
|
272
|
+
confirmedAt: Type.TString;
|
|
273
|
+
}>;
|
|
274
|
+
objectives: Type.TArray<Type.TObject<{
|
|
275
|
+
metricId: Type.TUnion<Type.TLiteral<string>[]>;
|
|
276
|
+
required: Type.TBoolean;
|
|
277
|
+
minimumSamples: Type.TInteger;
|
|
278
|
+
tolerance: Type.TNumber;
|
|
279
|
+
samplingPlan: Type.TObject<{
|
|
280
|
+
strategy: Type.TUnion<Type.TLiteral<string>[]>;
|
|
281
|
+
representativeCaseCount: Type.TInteger;
|
|
282
|
+
repetitionsPerCase: Type.TInteger;
|
|
283
|
+
aggregation: Type.TUnion<Type.TLiteral<string>[]>;
|
|
284
|
+
rationale: Type.TString;
|
|
285
|
+
}>;
|
|
286
|
+
constraint: Type.TOptional<Type.TObject<{
|
|
287
|
+
operator: Type.TUnion<Type.TLiteral<string>[]>;
|
|
288
|
+
value: Type.TNumber;
|
|
289
|
+
}>>;
|
|
290
|
+
}>>;
|
|
291
|
+
createdAt: Type.TString;
|
|
292
|
+
}>;
|
|
293
|
+
incumbent: Type.TObject<{
|
|
294
|
+
targetId: Type.TString;
|
|
295
|
+
candidateId: Type.TString;
|
|
296
|
+
modelId: Type.TString;
|
|
297
|
+
modelVersion: Type.TOptional<Type.TString>;
|
|
298
|
+
reasons: Type.TArray<Type.TString>;
|
|
299
|
+
}>;
|
|
300
|
+
policy: Type.TObject<{
|
|
301
|
+
mode: Type.TLiteral<"supervised_graduation">;
|
|
302
|
+
evaluationBudget: Type.TObject<{
|
|
303
|
+
currency: Type.TString;
|
|
304
|
+
maximumSpend: Type.TNumber;
|
|
305
|
+
maximumCandidateCount: Type.TOptional<Type.TInteger>;
|
|
306
|
+
maximumRunsPerCandidate: Type.TOptional<Type.TInteger>;
|
|
307
|
+
expiresAt: Type.TOptional<Type.TString>;
|
|
308
|
+
}>;
|
|
309
|
+
automaticSwapsEnabled: Type.TBoolean;
|
|
310
|
+
confirmationRequiredUntilVerifiedSwaps: Type.TInteger;
|
|
311
|
+
requirePostSwapVerification: Type.TLiteral<true>;
|
|
312
|
+
rollbackOnRegression: Type.TBoolean;
|
|
313
|
+
checkPolicy: Type.TObject<{
|
|
314
|
+
baselineAssessment: Type.TObject<{
|
|
315
|
+
afterCompletedTasks: Type.TOptional<Type.TInteger>;
|
|
316
|
+
afterElapsedSeconds: Type.TOptional<Type.TInteger>;
|
|
317
|
+
}>;
|
|
318
|
+
frontierReassessment: Type.TObject<{
|
|
319
|
+
afterCompletedTasks: Type.TOptional<Type.TInteger>;
|
|
320
|
+
afterElapsedSeconds: Type.TOptional<Type.TInteger>;
|
|
321
|
+
onModelCatalogChange: Type.TBoolean;
|
|
322
|
+
}>;
|
|
323
|
+
regressionCheck: Type.TObject<{
|
|
324
|
+
afterCompletedTasks: Type.TOptional<Type.TInteger>;
|
|
325
|
+
afterElapsedSeconds: Type.TOptional<Type.TInteger>;
|
|
326
|
+
thresholds: Type.TArray<Type.TObject<{
|
|
327
|
+
metricId: Type.TUnion<Type.TLiteral<string>[]>;
|
|
328
|
+
relativeChangeAtLeast: Type.TNumber;
|
|
329
|
+
}>>;
|
|
330
|
+
}>;
|
|
331
|
+
postSwapVerification: Type.TObject<{
|
|
332
|
+
afterCompletedTasks: Type.TOptional<Type.TInteger>;
|
|
333
|
+
afterElapsedSeconds: Type.TOptional<Type.TInteger>;
|
|
334
|
+
}>;
|
|
335
|
+
cooldownSeconds: Type.TInteger;
|
|
336
|
+
execution: Type.TObject<{
|
|
337
|
+
mode: Type.TUnion<Type.TLiteral<string>[]>;
|
|
338
|
+
infrastructureChangesAllowed: Type.TBoolean;
|
|
339
|
+
}>;
|
|
340
|
+
}>;
|
|
341
|
+
}>;
|
|
342
|
+
observability: Type.TObject<{
|
|
343
|
+
targetId: Type.TString;
|
|
344
|
+
status: Type.TUnion<Type.TLiteral<string>[]>;
|
|
345
|
+
requiredMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
|
|
346
|
+
coveredMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
|
|
347
|
+
missingMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
|
|
348
|
+
checkedAt: Type.TString;
|
|
349
|
+
}>;
|
|
350
|
+
preliminaryModelLeads: Type.TOptional<Type.TArray<Type.TObject<{
|
|
351
|
+
targetId: Type.TString;
|
|
352
|
+
modelId: Type.TString;
|
|
353
|
+
modelVersion: Type.TOptional<Type.TString>;
|
|
354
|
+
estimatedPrice: Type.TObject<{
|
|
355
|
+
currency: Type.TString;
|
|
356
|
+
inputPerMillionTokens: Type.TNumber;
|
|
357
|
+
outputPerMillionTokens: Type.TNumber;
|
|
358
|
+
sourceUri: Type.TString;
|
|
359
|
+
checkedAt: Type.TString;
|
|
360
|
+
}>;
|
|
361
|
+
reputation: Type.TObject<{
|
|
362
|
+
summary: Type.TString;
|
|
363
|
+
sourceUri: Type.TString;
|
|
364
|
+
checkedAt: Type.TString;
|
|
365
|
+
}>;
|
|
366
|
+
}>>>;
|
|
367
|
+
}>;
|
|
368
|
+
readonly instrument_observability: Type.TObject<{
|
|
369
|
+
targetId: Type.TString;
|
|
370
|
+
coveredMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
|
|
371
|
+
missingMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
|
|
372
|
+
}>;
|
|
373
|
+
readonly run_assessment: Type.TObject<{
|
|
374
|
+
targetId: Type.TString;
|
|
375
|
+
baselineEvidenceSufficient: Type.TBoolean;
|
|
376
|
+
metrics: Type.TArray<Type.TObject<{
|
|
377
|
+
targetId: Type.TString;
|
|
378
|
+
metricId: Type.TUnion<Type.TLiteral<string>[]>;
|
|
379
|
+
state: Type.TUnion<Type.TLiteral<string>[]>;
|
|
380
|
+
scope: Type.TUnion<Type.TLiteral<string>[]>;
|
|
381
|
+
direction: Type.TUnion<Type.TLiteral<string>[]>;
|
|
382
|
+
method: Type.TUnion<Type.TLiteral<string>[]>;
|
|
383
|
+
aggregation: Type.TUnion<Type.TLiteral<string>[]>;
|
|
384
|
+
value: Type.TOptional<Type.TNumber>;
|
|
385
|
+
unit: Type.TString;
|
|
386
|
+
sampleCount: Type.TInteger;
|
|
387
|
+
interval: Type.TOptional<Type.TObject<{
|
|
388
|
+
lower: Type.TNumber;
|
|
389
|
+
upper: Type.TNumber;
|
|
390
|
+
confidenceLevel: Type.TOptional<Type.TNumber>;
|
|
391
|
+
}>>;
|
|
392
|
+
window: Type.TOptional<Type.TObject<{
|
|
393
|
+
startedAt: Type.TString;
|
|
394
|
+
endedAt: Type.TString;
|
|
395
|
+
}>>;
|
|
396
|
+
evidence: Type.TArray<Type.TObject<{
|
|
397
|
+
id: Type.TString;
|
|
398
|
+
source: Type.TUnion<Type.TLiteral<string>[]>;
|
|
399
|
+
uri: Type.TString;
|
|
400
|
+
mediaType: Type.TOptional<Type.TString>;
|
|
401
|
+
digest: Type.TOptional<Type.TString>;
|
|
402
|
+
}>>;
|
|
403
|
+
note: Type.TOptional<Type.TString>;
|
|
404
|
+
}>>;
|
|
405
|
+
}>;
|
|
406
|
+
readonly discover_candidates: Type.TObject<{
|
|
407
|
+
targetId: Type.TString;
|
|
408
|
+
candidates: Type.TArray<Type.TObject<{
|
|
409
|
+
targetId: Type.TString;
|
|
410
|
+
candidateId: Type.TString;
|
|
411
|
+
modelId: Type.TString;
|
|
412
|
+
modelVersion: Type.TOptional<Type.TString>;
|
|
413
|
+
reasons: Type.TArray<Type.TString>;
|
|
414
|
+
}>>;
|
|
415
|
+
}>;
|
|
416
|
+
readonly run_challenger_trials: Type.TObject<{
|
|
417
|
+
targetId: Type.TString;
|
|
418
|
+
assessments: Type.TArray<Type.TObject<{
|
|
419
|
+
id: Type.TString;
|
|
420
|
+
targetId: Type.TString;
|
|
421
|
+
taskProfileId: Type.TString;
|
|
422
|
+
taskProfileRevision: Type.TInteger;
|
|
423
|
+
candidateId: Type.TString;
|
|
424
|
+
modelId: Type.TString;
|
|
425
|
+
modelVersion: Type.TOptional<Type.TString>;
|
|
426
|
+
productRevision: Type.TString;
|
|
427
|
+
harnessId: Type.TString;
|
|
428
|
+
startedAt: Type.TString;
|
|
429
|
+
completedAt: Type.TString;
|
|
430
|
+
metrics: Type.TArray<Type.TObject<{
|
|
431
|
+
targetId: Type.TString;
|
|
432
|
+
metricId: Type.TUnion<Type.TLiteral<string>[]>;
|
|
433
|
+
state: Type.TUnion<Type.TLiteral<string>[]>;
|
|
434
|
+
scope: Type.TUnion<Type.TLiteral<string>[]>;
|
|
435
|
+
direction: Type.TUnion<Type.TLiteral<string>[]>;
|
|
436
|
+
method: Type.TUnion<Type.TLiteral<string>[]>;
|
|
437
|
+
aggregation: Type.TUnion<Type.TLiteral<string>[]>;
|
|
438
|
+
value: Type.TOptional<Type.TNumber>;
|
|
439
|
+
unit: Type.TString;
|
|
440
|
+
sampleCount: Type.TInteger;
|
|
441
|
+
interval: Type.TOptional<Type.TObject<{
|
|
442
|
+
lower: Type.TNumber;
|
|
443
|
+
upper: Type.TNumber;
|
|
444
|
+
confidenceLevel: Type.TOptional<Type.TNumber>;
|
|
445
|
+
}>>;
|
|
446
|
+
window: Type.TOptional<Type.TObject<{
|
|
447
|
+
startedAt: Type.TString;
|
|
448
|
+
endedAt: Type.TString;
|
|
449
|
+
}>>;
|
|
450
|
+
evidence: Type.TArray<Type.TObject<{
|
|
451
|
+
id: Type.TString;
|
|
452
|
+
source: Type.TUnion<Type.TLiteral<string>[]>;
|
|
453
|
+
uri: Type.TString;
|
|
454
|
+
mediaType: Type.TOptional<Type.TString>;
|
|
455
|
+
digest: Type.TOptional<Type.TString>;
|
|
456
|
+
}>>;
|
|
457
|
+
note: Type.TOptional<Type.TString>;
|
|
458
|
+
}>>;
|
|
459
|
+
}>>;
|
|
460
|
+
experimentManifest: Type.TObject<{
|
|
461
|
+
id: Type.TString;
|
|
462
|
+
source: Type.TUnion<Type.TLiteral<string>[]>;
|
|
463
|
+
uri: Type.TString;
|
|
464
|
+
mediaType: Type.TOptional<Type.TString>;
|
|
465
|
+
digest: Type.TOptional<Type.TString>;
|
|
466
|
+
}>;
|
|
467
|
+
}>;
|
|
468
|
+
readonly calculate_frontier: Type.TObject<{
|
|
469
|
+
targetId: Type.TString;
|
|
470
|
+
assessments: Type.TArray<Type.TObject<{
|
|
471
|
+
id: Type.TString;
|
|
472
|
+
targetId: Type.TString;
|
|
473
|
+
taskProfileId: Type.TString;
|
|
474
|
+
taskProfileRevision: Type.TInteger;
|
|
475
|
+
candidateId: Type.TString;
|
|
476
|
+
modelId: Type.TString;
|
|
477
|
+
modelVersion: Type.TOptional<Type.TString>;
|
|
478
|
+
productRevision: Type.TString;
|
|
479
|
+
harnessId: Type.TString;
|
|
480
|
+
startedAt: Type.TString;
|
|
481
|
+
completedAt: Type.TString;
|
|
482
|
+
metrics: Type.TArray<Type.TObject<{
|
|
483
|
+
targetId: Type.TString;
|
|
484
|
+
metricId: Type.TUnion<Type.TLiteral<string>[]>;
|
|
485
|
+
state: Type.TUnion<Type.TLiteral<string>[]>;
|
|
486
|
+
scope: Type.TUnion<Type.TLiteral<string>[]>;
|
|
487
|
+
direction: Type.TUnion<Type.TLiteral<string>[]>;
|
|
488
|
+
method: Type.TUnion<Type.TLiteral<string>[]>;
|
|
489
|
+
aggregation: Type.TUnion<Type.TLiteral<string>[]>;
|
|
490
|
+
value: Type.TOptional<Type.TNumber>;
|
|
491
|
+
unit: Type.TString;
|
|
492
|
+
sampleCount: Type.TInteger;
|
|
493
|
+
interval: Type.TOptional<Type.TObject<{
|
|
494
|
+
lower: Type.TNumber;
|
|
495
|
+
upper: Type.TNumber;
|
|
496
|
+
confidenceLevel: Type.TOptional<Type.TNumber>;
|
|
497
|
+
}>>;
|
|
498
|
+
window: Type.TOptional<Type.TObject<{
|
|
499
|
+
startedAt: Type.TString;
|
|
500
|
+
endedAt: Type.TString;
|
|
501
|
+
}>>;
|
|
502
|
+
evidence: Type.TArray<Type.TObject<{
|
|
503
|
+
id: Type.TString;
|
|
504
|
+
source: Type.TUnion<Type.TLiteral<string>[]>;
|
|
505
|
+
uri: Type.TString;
|
|
506
|
+
mediaType: Type.TOptional<Type.TString>;
|
|
507
|
+
digest: Type.TOptional<Type.TString>;
|
|
508
|
+
}>>;
|
|
509
|
+
note: Type.TOptional<Type.TString>;
|
|
510
|
+
}>>;
|
|
511
|
+
}>>;
|
|
512
|
+
frontierSnapshot: Type.TObject<{
|
|
513
|
+
id: Type.TString;
|
|
514
|
+
targetId: Type.TString;
|
|
515
|
+
taskProfileId: Type.TString;
|
|
516
|
+
taskProfileRevision: Type.TInteger;
|
|
517
|
+
productRevision: Type.TString;
|
|
518
|
+
createdAt: Type.TString;
|
|
519
|
+
frontierCandidateIds: Type.TArray<Type.TString>;
|
|
520
|
+
dominatedCandidateIds: Type.TArray<Type.TString>;
|
|
521
|
+
ineligibleCandidateIds: Type.TArray<Type.TString>;
|
|
522
|
+
unresolvedCandidateIds: Type.TArray<Type.TString>;
|
|
523
|
+
comparisons: Type.TArray<Type.TObject<{
|
|
524
|
+
candidateId: Type.TString;
|
|
525
|
+
otherCandidateId: Type.TString;
|
|
526
|
+
state: Type.TUnion<Type.TLiteral<string>[]>;
|
|
527
|
+
reasons: Type.TArray<Type.TString>;
|
|
528
|
+
}>>;
|
|
529
|
+
assessmentIds: Type.TArray<Type.TString>;
|
|
530
|
+
calculationEvidence: Type.TArray<Type.TObject<{
|
|
531
|
+
id: Type.TString;
|
|
532
|
+
source: Type.TUnion<Type.TLiteral<string>[]>;
|
|
533
|
+
uri: Type.TString;
|
|
534
|
+
mediaType: Type.TOptional<Type.TString>;
|
|
535
|
+
digest: Type.TOptional<Type.TString>;
|
|
536
|
+
}>>;
|
|
537
|
+
}>;
|
|
538
|
+
recommendation: Type.TOptional<Type.TObject<{
|
|
539
|
+
targetId: Type.TString;
|
|
540
|
+
frontierSnapshotId: Type.TString;
|
|
541
|
+
selectedCandidateId: Type.TString;
|
|
542
|
+
currentCandidateId: Type.TString;
|
|
543
|
+
reasons: Type.TArray<Type.TString>;
|
|
544
|
+
preferenceEvidence: Type.TArray<Type.TObject<{
|
|
545
|
+
id: Type.TString;
|
|
546
|
+
source: Type.TUnion<Type.TLiteral<string>[]>;
|
|
547
|
+
uri: Type.TString;
|
|
548
|
+
mediaType: Type.TOptional<Type.TString>;
|
|
549
|
+
digest: Type.TOptional<Type.TString>;
|
|
550
|
+
}>>;
|
|
551
|
+
}>>;
|
|
552
|
+
}>;
|
|
553
|
+
readonly investigate_regression: Type.TObject<{
|
|
554
|
+
targetId: Type.TString;
|
|
555
|
+
regressionDetected: Type.TBoolean;
|
|
556
|
+
affectedMetricIds: Type.TArray<Type.TUnion<Type.TLiteral<string>[]>>;
|
|
557
|
+
likelyCause: Type.TOptional<Type.TString>;
|
|
558
|
+
}>;
|
|
559
|
+
readonly apply_model_swap: Type.TObject<{
|
|
560
|
+
targetId: Type.TString;
|
|
561
|
+
appliedCandidateId: Type.TString;
|
|
562
|
+
previousCandidateId: Type.TOptional<Type.TString>;
|
|
563
|
+
}>;
|
|
564
|
+
readonly verify_model_swap: Type.TObject<{
|
|
565
|
+
targetId: Type.TString;
|
|
566
|
+
verified: Type.TBoolean;
|
|
567
|
+
regressionDetected: Type.TBoolean;
|
|
568
|
+
}>;
|
|
569
|
+
readonly rollback_model_swap: Type.TObject<{
|
|
570
|
+
targetId: Type.TString;
|
|
571
|
+
restoredCandidateId: Type.TString;
|
|
572
|
+
}>;
|
|
573
|
+
};
|
|
574
|
+
export declare function intentOutputSchemaFor(kind: IntentKind): TSchema;
|
|
575
|
+
export declare function validateIntentOutputSchema(kind: IntentKind, output: unknown): readonly string[];
|
|
576
|
+
export declare function completionToolSchemaFor(kind: IntentKind): TSchema;
|