@elizaos/cli 1.6.0-alpha.4 → 1.6.0-beta.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/dist/commands/agent/utils/validation.d.ts +31 -3
- package/dist/commands/agent/utils/validation.d.ts.map +1 -1
- package/dist/commands/create/types.d.ts +8 -7
- package/dist/commands/create/types.d.ts.map +1 -1
- package/dist/commands/create/utils/validation.d.ts +2 -2
- package/dist/commands/create/utils/validation.d.ts.map +1 -1
- package/dist/commands/report/src/report-schema.d.ts +167 -11
- package/dist/commands/report/src/report-schema.d.ts.map +1 -1
- package/dist/commands/scenario/src/MockEngine.d.ts +11 -11
- package/dist/commands/scenario/src/matrix-schema.d.ts +70 -5
- package/dist/commands/scenario/src/matrix-schema.d.ts.map +1 -1
- package/dist/commands/scenario/src/schema.d.ts +2007 -175
- package/dist/commands/scenario/src/schema.d.ts.map +1 -1
- package/dist/index.js +15087 -23793
- package/dist/index.js.map +5 -73
- package/dist/templates/plugin-quick-starter/package.json +4 -4
- package/dist/templates/plugin-starter/package.json +4 -4
- package/dist/templates/project-starter/package.json +7 -7
- package/dist/templates/project-tee-starter/package.json +5 -5
- package/dist/utils/get-config.d.ts +194 -27
- package/dist/utils/get-config.d.ts.map +1 -1
- package/dist/version.d.ts +2 -2
- package/dist/version.d.ts.map +1 -1
- package/dist/version.js +2 -2
- package/package.json +8 -9
- package/templates/plugin-quick-starter/package.json +4 -4
- package/templates/plugin-starter/package.json +4 -4
- package/templates/project-starter/package.json +7 -7
- package/templates/project-tee-starter/package.json +5 -5
|
@@ -19,197 +19,612 @@ export declare const EnhancedEvaluationResultSchema: z.ZodObject<{
|
|
|
19
19
|
success: z.ZodBoolean;
|
|
20
20
|
summary: z.ZodString;
|
|
21
21
|
details: z.ZodRecord<z.ZodString, z.ZodAny>;
|
|
22
|
-
}, z.
|
|
22
|
+
}, "strip", z.ZodTypeAny, {
|
|
23
|
+
success: boolean;
|
|
24
|
+
evaluator_type: string;
|
|
25
|
+
summary: string;
|
|
26
|
+
details: Record<string, any>;
|
|
27
|
+
}, {
|
|
28
|
+
success: boolean;
|
|
29
|
+
evaluator_type: string;
|
|
30
|
+
summary: string;
|
|
31
|
+
details: Record<string, any>;
|
|
32
|
+
}>;
|
|
23
33
|
export declare const CapabilityCheckSchema: z.ZodObject<{
|
|
24
34
|
capability: z.ZodString;
|
|
25
35
|
achieved: z.ZodBoolean;
|
|
26
36
|
reasoning: z.ZodString;
|
|
27
|
-
}, z.
|
|
37
|
+
}, "strip", z.ZodTypeAny, {
|
|
38
|
+
capability: string;
|
|
39
|
+
achieved: boolean;
|
|
40
|
+
reasoning: string;
|
|
41
|
+
}, {
|
|
42
|
+
capability: string;
|
|
43
|
+
achieved: boolean;
|
|
44
|
+
reasoning: string;
|
|
45
|
+
}>;
|
|
28
46
|
export declare const LLMJudgeResultSchema: z.ZodObject<{
|
|
29
47
|
qualitative_summary: z.ZodString;
|
|
30
48
|
capability_checklist: z.ZodArray<z.ZodObject<{
|
|
31
49
|
capability: z.ZodString;
|
|
32
50
|
achieved: z.ZodBoolean;
|
|
33
51
|
reasoning: z.ZodString;
|
|
34
|
-
}, z.
|
|
35
|
-
|
|
36
|
-
|
|
52
|
+
}, "strip", z.ZodTypeAny, {
|
|
53
|
+
capability: string;
|
|
54
|
+
achieved: boolean;
|
|
55
|
+
reasoning: string;
|
|
56
|
+
}, {
|
|
57
|
+
capability: string;
|
|
58
|
+
achieved: boolean;
|
|
59
|
+
reasoning: string;
|
|
60
|
+
}>, "many">;
|
|
61
|
+
}, "strip", z.ZodTypeAny, {
|
|
62
|
+
qualitative_summary: string;
|
|
63
|
+
capability_checklist: {
|
|
64
|
+
capability: string;
|
|
65
|
+
achieved: boolean;
|
|
66
|
+
reasoning: string;
|
|
67
|
+
}[];
|
|
68
|
+
}, {
|
|
69
|
+
qualitative_summary: string;
|
|
70
|
+
capability_checklist: {
|
|
71
|
+
capability: string;
|
|
72
|
+
achieved: boolean;
|
|
73
|
+
reasoning: string;
|
|
74
|
+
}[];
|
|
75
|
+
}>;
|
|
76
|
+
export declare const EvaluationSchema: z.ZodDiscriminatedUnion<"type", [z.ZodObject<z.objectUtil.extendShape<{
|
|
77
|
+
type: z.ZodString;
|
|
78
|
+
}, {
|
|
37
79
|
type: z.ZodLiteral<"string_contains">;
|
|
38
80
|
value: z.ZodString;
|
|
39
81
|
case_sensitive: z.ZodOptional<z.ZodBoolean>;
|
|
40
|
-
}, z.
|
|
82
|
+
}>, "strip", z.ZodTypeAny, {
|
|
83
|
+
value: string;
|
|
84
|
+
type: "string_contains";
|
|
85
|
+
case_sensitive?: boolean | undefined;
|
|
86
|
+
}, {
|
|
87
|
+
value: string;
|
|
88
|
+
type: "string_contains";
|
|
89
|
+
case_sensitive?: boolean | undefined;
|
|
90
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
91
|
+
type: z.ZodString;
|
|
92
|
+
}, {
|
|
41
93
|
type: z.ZodLiteral<"regex_match">;
|
|
42
94
|
pattern: z.ZodString;
|
|
43
|
-
}, z.
|
|
95
|
+
}>, "strip", z.ZodTypeAny, {
|
|
96
|
+
type: "regex_match";
|
|
97
|
+
pattern: string;
|
|
98
|
+
}, {
|
|
99
|
+
type: "regex_match";
|
|
100
|
+
pattern: string;
|
|
101
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
102
|
+
type: z.ZodString;
|
|
103
|
+
}, {
|
|
44
104
|
type: z.ZodLiteral<"file_exists">;
|
|
45
105
|
path: z.ZodString;
|
|
46
|
-
}, z.
|
|
106
|
+
}>, "strip", z.ZodTypeAny, {
|
|
107
|
+
path: string;
|
|
108
|
+
type: "file_exists";
|
|
109
|
+
}, {
|
|
110
|
+
path: string;
|
|
111
|
+
type: "file_exists";
|
|
112
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
113
|
+
type: z.ZodString;
|
|
114
|
+
}, {
|
|
47
115
|
type: z.ZodLiteral<"trajectory_contains_action">;
|
|
48
116
|
action: z.ZodString;
|
|
49
|
-
}, z.
|
|
117
|
+
}>, "strip", z.ZodTypeAny, {
|
|
118
|
+
type: "trajectory_contains_action";
|
|
119
|
+
action: string;
|
|
120
|
+
}, {
|
|
121
|
+
type: "trajectory_contains_action";
|
|
122
|
+
action: string;
|
|
123
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
124
|
+
type: z.ZodString;
|
|
125
|
+
}, {
|
|
50
126
|
type: z.ZodLiteral<"llm_judge">;
|
|
51
127
|
prompt: z.ZodString;
|
|
52
128
|
expected: z.ZodString;
|
|
53
129
|
model_type: z.ZodOptional<z.ZodString>;
|
|
54
130
|
temperature: z.ZodOptional<z.ZodNumber>;
|
|
55
131
|
json_schema: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodAny>>;
|
|
56
|
-
capabilities: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
57
|
-
}, z.
|
|
132
|
+
capabilities: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
133
|
+
}>, "strip", z.ZodTypeAny, {
|
|
134
|
+
type: "llm_judge";
|
|
135
|
+
expected: string;
|
|
136
|
+
prompt: string;
|
|
137
|
+
model_type?: string | undefined;
|
|
138
|
+
temperature?: number | undefined;
|
|
139
|
+
json_schema?: Record<string, any> | undefined;
|
|
140
|
+
capabilities?: string[] | undefined;
|
|
141
|
+
}, {
|
|
142
|
+
type: "llm_judge";
|
|
143
|
+
expected: string;
|
|
144
|
+
prompt: string;
|
|
145
|
+
model_type?: string | undefined;
|
|
146
|
+
temperature?: number | undefined;
|
|
147
|
+
json_schema?: Record<string, any> | undefined;
|
|
148
|
+
capabilities?: string[] | undefined;
|
|
149
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
150
|
+
type: z.ZodString;
|
|
151
|
+
}, {
|
|
58
152
|
type: z.ZodLiteral<"execution_time">;
|
|
59
153
|
max_duration_ms: z.ZodNumber;
|
|
60
154
|
min_duration_ms: z.ZodOptional<z.ZodNumber>;
|
|
61
155
|
target_duration_ms: z.ZodOptional<z.ZodNumber>;
|
|
62
|
-
}, z.
|
|
156
|
+
}>, "strip", z.ZodTypeAny, {
|
|
157
|
+
type: "execution_time";
|
|
158
|
+
max_duration_ms: number;
|
|
159
|
+
min_duration_ms?: number | undefined;
|
|
160
|
+
target_duration_ms?: number | undefined;
|
|
161
|
+
}, {
|
|
162
|
+
type: "execution_time";
|
|
163
|
+
max_duration_ms: number;
|
|
164
|
+
min_duration_ms?: number | undefined;
|
|
165
|
+
target_duration_ms?: number | undefined;
|
|
166
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
167
|
+
type: z.ZodString;
|
|
168
|
+
}, {
|
|
63
169
|
type: z.ZodLiteral<"conversation_length">;
|
|
64
170
|
min_turns: z.ZodOptional<z.ZodNumber>;
|
|
65
171
|
max_turns: z.ZodOptional<z.ZodNumber>;
|
|
66
172
|
optimal_turns: z.ZodOptional<z.ZodNumber>;
|
|
67
|
-
target_range: z.ZodOptional<z.ZodArray<z.ZodNumber
|
|
68
|
-
}, z.
|
|
173
|
+
target_range: z.ZodEffects<z.ZodOptional<z.ZodArray<z.ZodNumber, "many">>, number[] | undefined, number[] | undefined>;
|
|
174
|
+
}>, "strip", z.ZodTypeAny, {
|
|
175
|
+
type: "conversation_length";
|
|
176
|
+
min_turns?: number | undefined;
|
|
177
|
+
max_turns?: number | undefined;
|
|
178
|
+
optimal_turns?: number | undefined;
|
|
179
|
+
target_range?: number[] | undefined;
|
|
180
|
+
}, {
|
|
181
|
+
type: "conversation_length";
|
|
182
|
+
min_turns?: number | undefined;
|
|
183
|
+
max_turns?: number | undefined;
|
|
184
|
+
optimal_turns?: number | undefined;
|
|
185
|
+
target_range?: number[] | undefined;
|
|
186
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
187
|
+
type: z.ZodString;
|
|
188
|
+
}, {
|
|
69
189
|
type: z.ZodLiteral<"conversation_flow">;
|
|
70
|
-
required_patterns: z.ZodArray<z.ZodEnum<
|
|
71
|
-
question_then_answer: "question_then_answer";
|
|
72
|
-
problem_then_solution: "problem_then_solution";
|
|
73
|
-
clarification_cycle: "clarification_cycle";
|
|
74
|
-
empathy_then_solution: "empathy_then_solution";
|
|
75
|
-
escalation_pattern: "escalation_pattern";
|
|
76
|
-
}>>;
|
|
190
|
+
required_patterns: z.ZodArray<z.ZodEnum<["question_then_answer", "problem_then_solution", "clarification_cycle", "empathy_then_solution", "escalation_pattern"]>, "many">;
|
|
77
191
|
flow_quality_threshold: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
78
|
-
}, z.
|
|
192
|
+
}>, "strip", z.ZodTypeAny, {
|
|
193
|
+
type: "conversation_flow";
|
|
194
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
195
|
+
flow_quality_threshold: number;
|
|
196
|
+
}, {
|
|
197
|
+
type: "conversation_flow";
|
|
198
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
199
|
+
flow_quality_threshold?: number | undefined;
|
|
200
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
201
|
+
type: z.ZodString;
|
|
202
|
+
}, {
|
|
79
203
|
type: z.ZodLiteral<"user_satisfaction">;
|
|
80
204
|
satisfaction_threshold: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
81
205
|
indicators: z.ZodOptional<z.ZodObject<{
|
|
82
|
-
positive: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
83
|
-
negative: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
84
|
-
}, z.
|
|
85
|
-
|
|
86
|
-
|
|
87
|
-
|
|
88
|
-
|
|
89
|
-
|
|
90
|
-
}
|
|
206
|
+
positive: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
207
|
+
negative: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
208
|
+
}, "strip", z.ZodTypeAny, {
|
|
209
|
+
positive?: string[] | undefined;
|
|
210
|
+
negative?: string[] | undefined;
|
|
211
|
+
}, {
|
|
212
|
+
positive?: string[] | undefined;
|
|
213
|
+
negative?: string[] | undefined;
|
|
214
|
+
}>>;
|
|
215
|
+
measurement_method: z.ZodDefault<z.ZodOptional<z.ZodEnum<["sentiment_analysis", "keyword_analysis", "llm_judge"]>>>;
|
|
216
|
+
}>, "strip", z.ZodTypeAny, {
|
|
217
|
+
type: "user_satisfaction";
|
|
218
|
+
satisfaction_threshold: number;
|
|
219
|
+
measurement_method: "llm_judge" | "sentiment_analysis" | "keyword_analysis";
|
|
220
|
+
indicators?: {
|
|
221
|
+
positive?: string[] | undefined;
|
|
222
|
+
negative?: string[] | undefined;
|
|
223
|
+
} | undefined;
|
|
224
|
+
}, {
|
|
225
|
+
type: "user_satisfaction";
|
|
226
|
+
satisfaction_threshold?: number | undefined;
|
|
227
|
+
indicators?: {
|
|
228
|
+
positive?: string[] | undefined;
|
|
229
|
+
negative?: string[] | undefined;
|
|
230
|
+
} | undefined;
|
|
231
|
+
measurement_method?: "llm_judge" | "sentiment_analysis" | "keyword_analysis" | undefined;
|
|
232
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
233
|
+
type: z.ZodString;
|
|
234
|
+
}, {
|
|
91
235
|
type: z.ZodLiteral<"context_retention">;
|
|
92
|
-
test_memory_of: z.ZodArray<z.ZodString>;
|
|
236
|
+
test_memory_of: z.ZodArray<z.ZodString, "many">;
|
|
93
237
|
retention_turns: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
94
238
|
memory_accuracy_threshold: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
95
|
-
}, z.
|
|
239
|
+
}>, "strip", z.ZodTypeAny, {
|
|
240
|
+
type: "context_retention";
|
|
241
|
+
test_memory_of: string[];
|
|
242
|
+
retention_turns: number;
|
|
243
|
+
memory_accuracy_threshold: number;
|
|
244
|
+
}, {
|
|
245
|
+
type: "context_retention";
|
|
246
|
+
test_memory_of: string[];
|
|
247
|
+
retention_turns?: number | undefined;
|
|
248
|
+
memory_accuracy_threshold?: number | undefined;
|
|
249
|
+
}>]>;
|
|
96
250
|
declare const PluginConfigSchema: z.ZodObject<{
|
|
97
251
|
name: z.ZodString;
|
|
98
252
|
version: z.ZodOptional<z.ZodString>;
|
|
99
253
|
config: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodAny>>;
|
|
100
254
|
enabled: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
101
|
-
}, z.
|
|
102
|
-
|
|
255
|
+
}, "strip", z.ZodTypeAny, {
|
|
256
|
+
name: string;
|
|
257
|
+
enabled: boolean;
|
|
258
|
+
version?: string | undefined;
|
|
259
|
+
config?: Record<string, any> | undefined;
|
|
260
|
+
}, {
|
|
261
|
+
name: string;
|
|
262
|
+
enabled?: boolean | undefined;
|
|
263
|
+
version?: string | undefined;
|
|
264
|
+
config?: Record<string, any> | undefined;
|
|
265
|
+
}>;
|
|
266
|
+
declare const PluginReferenceSchema: z.ZodUnion<[z.ZodString, z.ZodObject<{
|
|
103
267
|
name: z.ZodString;
|
|
104
268
|
version: z.ZodOptional<z.ZodString>;
|
|
105
269
|
config: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodAny>>;
|
|
106
270
|
enabled: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
107
|
-
}, z.
|
|
271
|
+
}, "strip", z.ZodTypeAny, {
|
|
272
|
+
name: string;
|
|
273
|
+
enabled: boolean;
|
|
274
|
+
version?: string | undefined;
|
|
275
|
+
config?: Record<string, any> | undefined;
|
|
276
|
+
}, {
|
|
277
|
+
name: string;
|
|
278
|
+
enabled?: boolean | undefined;
|
|
279
|
+
version?: string | undefined;
|
|
280
|
+
config?: Record<string, any> | undefined;
|
|
281
|
+
}>]>;
|
|
108
282
|
export declare const ScenarioSchema: z.ZodObject<{
|
|
109
283
|
name: z.ZodString;
|
|
110
284
|
description: z.ZodString;
|
|
111
|
-
plugins: z.ZodOptional<z.ZodArray<z.ZodUnion<
|
|
285
|
+
plugins: z.ZodOptional<z.ZodArray<z.ZodUnion<[z.ZodString, z.ZodObject<{
|
|
112
286
|
name: z.ZodString;
|
|
113
287
|
version: z.ZodOptional<z.ZodString>;
|
|
114
288
|
config: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodAny>>;
|
|
115
289
|
enabled: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
116
|
-
}, z.
|
|
290
|
+
}, "strip", z.ZodTypeAny, {
|
|
291
|
+
name: string;
|
|
292
|
+
enabled: boolean;
|
|
293
|
+
version?: string | undefined;
|
|
294
|
+
config?: Record<string, any> | undefined;
|
|
295
|
+
}, {
|
|
296
|
+
name: string;
|
|
297
|
+
enabled?: boolean | undefined;
|
|
298
|
+
version?: string | undefined;
|
|
299
|
+
config?: Record<string, any> | undefined;
|
|
300
|
+
}>]>, "many">>;
|
|
117
301
|
environment: z.ZodObject<{
|
|
118
|
-
type: z.ZodEnum<
|
|
119
|
-
|
|
120
|
-
|
|
121
|
-
|
|
122
|
-
|
|
302
|
+
type: z.ZodEnum<["e2b", "local"]>;
|
|
303
|
+
}, "strip", z.ZodTypeAny, {
|
|
304
|
+
type: "local" | "e2b";
|
|
305
|
+
}, {
|
|
306
|
+
type: "local" | "e2b";
|
|
307
|
+
}>;
|
|
123
308
|
setup: z.ZodOptional<z.ZodObject<{
|
|
124
309
|
mocks: z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
125
310
|
service: z.ZodOptional<z.ZodString>;
|
|
126
311
|
method: z.ZodString;
|
|
127
312
|
when: z.ZodOptional<z.ZodObject<{
|
|
128
|
-
args: z.ZodOptional<z.ZodArray<z.ZodAny>>;
|
|
313
|
+
args: z.ZodOptional<z.ZodArray<z.ZodAny, "many">>;
|
|
129
314
|
input: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodAny>>;
|
|
130
315
|
context: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodAny>>;
|
|
131
316
|
matcher: z.ZodOptional<z.ZodString>;
|
|
132
|
-
partialArgs: z.ZodOptional<z.ZodArray<z.ZodAny>>;
|
|
133
|
-
}, z.
|
|
317
|
+
partialArgs: z.ZodOptional<z.ZodArray<z.ZodAny, "many">>;
|
|
318
|
+
}, "strip", z.ZodTypeAny, {
|
|
319
|
+
context?: Record<string, any> | undefined;
|
|
320
|
+
args?: any[] | undefined;
|
|
321
|
+
input?: Record<string, any> | undefined;
|
|
322
|
+
matcher?: string | undefined;
|
|
323
|
+
partialArgs?: any[] | undefined;
|
|
324
|
+
}, {
|
|
325
|
+
context?: Record<string, any> | undefined;
|
|
326
|
+
args?: any[] | undefined;
|
|
327
|
+
input?: Record<string, any> | undefined;
|
|
328
|
+
matcher?: string | undefined;
|
|
329
|
+
partialArgs?: any[] | undefined;
|
|
330
|
+
}>>;
|
|
134
331
|
response: z.ZodAny;
|
|
135
332
|
responseFn: z.ZodOptional<z.ZodString>;
|
|
136
333
|
error: z.ZodOptional<z.ZodObject<{
|
|
137
334
|
code: z.ZodString;
|
|
138
335
|
message: z.ZodString;
|
|
139
336
|
status: z.ZodOptional<z.ZodNumber>;
|
|
140
|
-
}, z.
|
|
337
|
+
}, "strip", z.ZodTypeAny, {
|
|
338
|
+
message: string;
|
|
339
|
+
code: string;
|
|
340
|
+
status?: number | undefined;
|
|
341
|
+
}, {
|
|
342
|
+
message: string;
|
|
343
|
+
code: string;
|
|
344
|
+
status?: number | undefined;
|
|
345
|
+
}>>;
|
|
141
346
|
metadata: z.ZodOptional<z.ZodObject<{
|
|
142
347
|
delay: z.ZodOptional<z.ZodNumber>;
|
|
143
348
|
probability: z.ZodOptional<z.ZodNumber>;
|
|
144
|
-
}, z.
|
|
145
|
-
|
|
349
|
+
}, "strip", z.ZodTypeAny, {
|
|
350
|
+
delay?: number | undefined;
|
|
351
|
+
probability?: number | undefined;
|
|
352
|
+
}, {
|
|
353
|
+
delay?: number | undefined;
|
|
354
|
+
probability?: number | undefined;
|
|
355
|
+
}>>;
|
|
356
|
+
}, "strip", z.ZodTypeAny, {
|
|
357
|
+
method: string;
|
|
358
|
+
error?: {
|
|
359
|
+
message: string;
|
|
360
|
+
code: string;
|
|
361
|
+
status?: number | undefined;
|
|
362
|
+
} | undefined;
|
|
363
|
+
metadata?: {
|
|
364
|
+
delay?: number | undefined;
|
|
365
|
+
probability?: number | undefined;
|
|
366
|
+
} | undefined;
|
|
367
|
+
service?: string | undefined;
|
|
368
|
+
when?: {
|
|
369
|
+
context?: Record<string, any> | undefined;
|
|
370
|
+
args?: any[] | undefined;
|
|
371
|
+
input?: Record<string, any> | undefined;
|
|
372
|
+
matcher?: string | undefined;
|
|
373
|
+
partialArgs?: any[] | undefined;
|
|
374
|
+
} | undefined;
|
|
375
|
+
response?: any;
|
|
376
|
+
responseFn?: string | undefined;
|
|
377
|
+
}, {
|
|
378
|
+
method: string;
|
|
379
|
+
error?: {
|
|
380
|
+
message: string;
|
|
381
|
+
code: string;
|
|
382
|
+
status?: number | undefined;
|
|
383
|
+
} | undefined;
|
|
384
|
+
metadata?: {
|
|
385
|
+
delay?: number | undefined;
|
|
386
|
+
probability?: number | undefined;
|
|
387
|
+
} | undefined;
|
|
388
|
+
service?: string | undefined;
|
|
389
|
+
when?: {
|
|
390
|
+
context?: Record<string, any> | undefined;
|
|
391
|
+
args?: any[] | undefined;
|
|
392
|
+
input?: Record<string, any> | undefined;
|
|
393
|
+
matcher?: string | undefined;
|
|
394
|
+
partialArgs?: any[] | undefined;
|
|
395
|
+
} | undefined;
|
|
396
|
+
response?: any;
|
|
397
|
+
responseFn?: string | undefined;
|
|
398
|
+
}>, "many">>;
|
|
146
399
|
virtual_fs: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>;
|
|
147
|
-
}, z.
|
|
400
|
+
}, "strip", z.ZodTypeAny, {
|
|
401
|
+
mocks?: {
|
|
402
|
+
method: string;
|
|
403
|
+
error?: {
|
|
404
|
+
message: string;
|
|
405
|
+
code: string;
|
|
406
|
+
status?: number | undefined;
|
|
407
|
+
} | undefined;
|
|
408
|
+
metadata?: {
|
|
409
|
+
delay?: number | undefined;
|
|
410
|
+
probability?: number | undefined;
|
|
411
|
+
} | undefined;
|
|
412
|
+
service?: string | undefined;
|
|
413
|
+
when?: {
|
|
414
|
+
context?: Record<string, any> | undefined;
|
|
415
|
+
args?: any[] | undefined;
|
|
416
|
+
input?: Record<string, any> | undefined;
|
|
417
|
+
matcher?: string | undefined;
|
|
418
|
+
partialArgs?: any[] | undefined;
|
|
419
|
+
} | undefined;
|
|
420
|
+
response?: any;
|
|
421
|
+
responseFn?: string | undefined;
|
|
422
|
+
}[] | undefined;
|
|
423
|
+
virtual_fs?: Record<string, string> | undefined;
|
|
424
|
+
}, {
|
|
425
|
+
mocks?: {
|
|
426
|
+
method: string;
|
|
427
|
+
error?: {
|
|
428
|
+
message: string;
|
|
429
|
+
code: string;
|
|
430
|
+
status?: number | undefined;
|
|
431
|
+
} | undefined;
|
|
432
|
+
metadata?: {
|
|
433
|
+
delay?: number | undefined;
|
|
434
|
+
probability?: number | undefined;
|
|
435
|
+
} | undefined;
|
|
436
|
+
service?: string | undefined;
|
|
437
|
+
when?: {
|
|
438
|
+
context?: Record<string, any> | undefined;
|
|
439
|
+
args?: any[] | undefined;
|
|
440
|
+
input?: Record<string, any> | undefined;
|
|
441
|
+
matcher?: string | undefined;
|
|
442
|
+
partialArgs?: any[] | undefined;
|
|
443
|
+
} | undefined;
|
|
444
|
+
response?: any;
|
|
445
|
+
responseFn?: string | undefined;
|
|
446
|
+
}[] | undefined;
|
|
447
|
+
virtual_fs?: Record<string, string> | undefined;
|
|
448
|
+
}>>;
|
|
148
449
|
run: z.ZodArray<z.ZodObject<{
|
|
149
450
|
name: z.ZodOptional<z.ZodString>;
|
|
150
451
|
lang: z.ZodOptional<z.ZodString>;
|
|
151
452
|
code: z.ZodOptional<z.ZodString>;
|
|
152
453
|
input: z.ZodOptional<z.ZodString>;
|
|
153
|
-
evaluations: z.ZodArray<z.ZodDiscriminatedUnion<[z.ZodObject<{
|
|
454
|
+
evaluations: z.ZodArray<z.ZodDiscriminatedUnion<"type", [z.ZodObject<z.objectUtil.extendShape<{
|
|
455
|
+
type: z.ZodString;
|
|
456
|
+
}, {
|
|
154
457
|
type: z.ZodLiteral<"string_contains">;
|
|
155
458
|
value: z.ZodString;
|
|
156
459
|
case_sensitive: z.ZodOptional<z.ZodBoolean>;
|
|
157
|
-
}, z.
|
|
460
|
+
}>, "strip", z.ZodTypeAny, {
|
|
461
|
+
value: string;
|
|
462
|
+
type: "string_contains";
|
|
463
|
+
case_sensitive?: boolean | undefined;
|
|
464
|
+
}, {
|
|
465
|
+
value: string;
|
|
466
|
+
type: "string_contains";
|
|
467
|
+
case_sensitive?: boolean | undefined;
|
|
468
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
469
|
+
type: z.ZodString;
|
|
470
|
+
}, {
|
|
158
471
|
type: z.ZodLiteral<"regex_match">;
|
|
159
472
|
pattern: z.ZodString;
|
|
160
|
-
}, z.
|
|
473
|
+
}>, "strip", z.ZodTypeAny, {
|
|
474
|
+
type: "regex_match";
|
|
475
|
+
pattern: string;
|
|
476
|
+
}, {
|
|
477
|
+
type: "regex_match";
|
|
478
|
+
pattern: string;
|
|
479
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
480
|
+
type: z.ZodString;
|
|
481
|
+
}, {
|
|
161
482
|
type: z.ZodLiteral<"file_exists">;
|
|
162
483
|
path: z.ZodString;
|
|
163
|
-
}, z.
|
|
484
|
+
}>, "strip", z.ZodTypeAny, {
|
|
485
|
+
path: string;
|
|
486
|
+
type: "file_exists";
|
|
487
|
+
}, {
|
|
488
|
+
path: string;
|
|
489
|
+
type: "file_exists";
|
|
490
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
491
|
+
type: z.ZodString;
|
|
492
|
+
}, {
|
|
164
493
|
type: z.ZodLiteral<"trajectory_contains_action">;
|
|
165
494
|
action: z.ZodString;
|
|
166
|
-
}, z.
|
|
495
|
+
}>, "strip", z.ZodTypeAny, {
|
|
496
|
+
type: "trajectory_contains_action";
|
|
497
|
+
action: string;
|
|
498
|
+
}, {
|
|
499
|
+
type: "trajectory_contains_action";
|
|
500
|
+
action: string;
|
|
501
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
502
|
+
type: z.ZodString;
|
|
503
|
+
}, {
|
|
167
504
|
type: z.ZodLiteral<"llm_judge">;
|
|
168
505
|
prompt: z.ZodString;
|
|
169
506
|
expected: z.ZodString;
|
|
170
507
|
model_type: z.ZodOptional<z.ZodString>;
|
|
171
508
|
temperature: z.ZodOptional<z.ZodNumber>;
|
|
172
509
|
json_schema: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodAny>>;
|
|
173
|
-
capabilities: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
174
|
-
}, z.
|
|
510
|
+
capabilities: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
511
|
+
}>, "strip", z.ZodTypeAny, {
|
|
512
|
+
type: "llm_judge";
|
|
513
|
+
expected: string;
|
|
514
|
+
prompt: string;
|
|
515
|
+
model_type?: string | undefined;
|
|
516
|
+
temperature?: number | undefined;
|
|
517
|
+
json_schema?: Record<string, any> | undefined;
|
|
518
|
+
capabilities?: string[] | undefined;
|
|
519
|
+
}, {
|
|
520
|
+
type: "llm_judge";
|
|
521
|
+
expected: string;
|
|
522
|
+
prompt: string;
|
|
523
|
+
model_type?: string | undefined;
|
|
524
|
+
temperature?: number | undefined;
|
|
525
|
+
json_schema?: Record<string, any> | undefined;
|
|
526
|
+
capabilities?: string[] | undefined;
|
|
527
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
528
|
+
type: z.ZodString;
|
|
529
|
+
}, {
|
|
175
530
|
type: z.ZodLiteral<"execution_time">;
|
|
176
531
|
max_duration_ms: z.ZodNumber;
|
|
177
532
|
min_duration_ms: z.ZodOptional<z.ZodNumber>;
|
|
178
533
|
target_duration_ms: z.ZodOptional<z.ZodNumber>;
|
|
179
|
-
}, z.
|
|
534
|
+
}>, "strip", z.ZodTypeAny, {
|
|
535
|
+
type: "execution_time";
|
|
536
|
+
max_duration_ms: number;
|
|
537
|
+
min_duration_ms?: number | undefined;
|
|
538
|
+
target_duration_ms?: number | undefined;
|
|
539
|
+
}, {
|
|
540
|
+
type: "execution_time";
|
|
541
|
+
max_duration_ms: number;
|
|
542
|
+
min_duration_ms?: number | undefined;
|
|
543
|
+
target_duration_ms?: number | undefined;
|
|
544
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
545
|
+
type: z.ZodString;
|
|
546
|
+
}, {
|
|
180
547
|
type: z.ZodLiteral<"conversation_length">;
|
|
181
548
|
min_turns: z.ZodOptional<z.ZodNumber>;
|
|
182
549
|
max_turns: z.ZodOptional<z.ZodNumber>;
|
|
183
550
|
optimal_turns: z.ZodOptional<z.ZodNumber>;
|
|
184
|
-
target_range: z.ZodOptional<z.ZodArray<z.ZodNumber
|
|
185
|
-
}, z.
|
|
551
|
+
target_range: z.ZodEffects<z.ZodOptional<z.ZodArray<z.ZodNumber, "many">>, number[] | undefined, number[] | undefined>;
|
|
552
|
+
}>, "strip", z.ZodTypeAny, {
|
|
553
|
+
type: "conversation_length";
|
|
554
|
+
min_turns?: number | undefined;
|
|
555
|
+
max_turns?: number | undefined;
|
|
556
|
+
optimal_turns?: number | undefined;
|
|
557
|
+
target_range?: number[] | undefined;
|
|
558
|
+
}, {
|
|
559
|
+
type: "conversation_length";
|
|
560
|
+
min_turns?: number | undefined;
|
|
561
|
+
max_turns?: number | undefined;
|
|
562
|
+
optimal_turns?: number | undefined;
|
|
563
|
+
target_range?: number[] | undefined;
|
|
564
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
565
|
+
type: z.ZodString;
|
|
566
|
+
}, {
|
|
186
567
|
type: z.ZodLiteral<"conversation_flow">;
|
|
187
|
-
required_patterns: z.ZodArray<z.ZodEnum<
|
|
188
|
-
question_then_answer: "question_then_answer";
|
|
189
|
-
problem_then_solution: "problem_then_solution";
|
|
190
|
-
clarification_cycle: "clarification_cycle";
|
|
191
|
-
empathy_then_solution: "empathy_then_solution";
|
|
192
|
-
escalation_pattern: "escalation_pattern";
|
|
193
|
-
}>>;
|
|
568
|
+
required_patterns: z.ZodArray<z.ZodEnum<["question_then_answer", "problem_then_solution", "clarification_cycle", "empathy_then_solution", "escalation_pattern"]>, "many">;
|
|
194
569
|
flow_quality_threshold: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
195
|
-
}, z.
|
|
570
|
+
}>, "strip", z.ZodTypeAny, {
|
|
571
|
+
type: "conversation_flow";
|
|
572
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
573
|
+
flow_quality_threshold: number;
|
|
574
|
+
}, {
|
|
575
|
+
type: "conversation_flow";
|
|
576
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
577
|
+
flow_quality_threshold?: number | undefined;
|
|
578
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
579
|
+
type: z.ZodString;
|
|
580
|
+
}, {
|
|
196
581
|
type: z.ZodLiteral<"user_satisfaction">;
|
|
197
582
|
satisfaction_threshold: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
198
583
|
indicators: z.ZodOptional<z.ZodObject<{
|
|
199
|
-
positive: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
200
|
-
negative: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
201
|
-
}, z.
|
|
202
|
-
|
|
203
|
-
|
|
204
|
-
|
|
205
|
-
|
|
206
|
-
|
|
207
|
-
|
|
584
|
+
positive: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
585
|
+
negative: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
586
|
+
}, "strip", z.ZodTypeAny, {
|
|
587
|
+
positive?: string[] | undefined;
|
|
588
|
+
negative?: string[] | undefined;
|
|
589
|
+
}, {
|
|
590
|
+
positive?: string[] | undefined;
|
|
591
|
+
negative?: string[] | undefined;
|
|
592
|
+
}>>;
|
|
593
|
+
measurement_method: z.ZodDefault<z.ZodOptional<z.ZodEnum<["sentiment_analysis", "keyword_analysis", "llm_judge"]>>>;
|
|
594
|
+
}>, "strip", z.ZodTypeAny, {
|
|
595
|
+
type: "user_satisfaction";
|
|
596
|
+
satisfaction_threshold: number;
|
|
597
|
+
measurement_method: "llm_judge" | "sentiment_analysis" | "keyword_analysis";
|
|
598
|
+
indicators?: {
|
|
599
|
+
positive?: string[] | undefined;
|
|
600
|
+
negative?: string[] | undefined;
|
|
601
|
+
} | undefined;
|
|
602
|
+
}, {
|
|
603
|
+
type: "user_satisfaction";
|
|
604
|
+
satisfaction_threshold?: number | undefined;
|
|
605
|
+
indicators?: {
|
|
606
|
+
positive?: string[] | undefined;
|
|
607
|
+
negative?: string[] | undefined;
|
|
608
|
+
} | undefined;
|
|
609
|
+
measurement_method?: "llm_judge" | "sentiment_analysis" | "keyword_analysis" | undefined;
|
|
610
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
611
|
+
type: z.ZodString;
|
|
612
|
+
}, {
|
|
208
613
|
type: z.ZodLiteral<"context_retention">;
|
|
209
|
-
test_memory_of: z.ZodArray<z.ZodString>;
|
|
614
|
+
test_memory_of: z.ZodArray<z.ZodString, "many">;
|
|
210
615
|
retention_turns: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
211
616
|
memory_accuracy_threshold: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
212
|
-
}, z.
|
|
617
|
+
}>, "strip", z.ZodTypeAny, {
|
|
618
|
+
type: "context_retention";
|
|
619
|
+
test_memory_of: string[];
|
|
620
|
+
retention_turns: number;
|
|
621
|
+
memory_accuracy_threshold: number;
|
|
622
|
+
}, {
|
|
623
|
+
type: "context_retention";
|
|
624
|
+
test_memory_of: string[];
|
|
625
|
+
retention_turns?: number | undefined;
|
|
626
|
+
memory_accuracy_threshold?: number | undefined;
|
|
627
|
+
}>]>, "many">;
|
|
213
628
|
conversation: z.ZodOptional<z.ZodObject<{
|
|
214
629
|
max_turns: z.ZodNumber;
|
|
215
630
|
timeout_per_turn_ms: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
@@ -221,165 +636,1512 @@ export declare const ScenarioSchema: z.ZodObject<{
|
|
|
221
636
|
persona: z.ZodString;
|
|
222
637
|
objective: z.ZodString;
|
|
223
638
|
style: z.ZodOptional<z.ZodString>;
|
|
224
|
-
constraints: z.ZodDefault<z.ZodOptional<z.ZodArray<z.ZodString>>>;
|
|
639
|
+
constraints: z.ZodDefault<z.ZodOptional<z.ZodArray<z.ZodString, "many">>>;
|
|
225
640
|
emotional_state: z.ZodOptional<z.ZodString>;
|
|
226
|
-
knowledge_level: z.ZodDefault<z.ZodOptional<z.ZodEnum<
|
|
227
|
-
|
|
228
|
-
|
|
229
|
-
|
|
230
|
-
|
|
231
|
-
|
|
641
|
+
knowledge_level: z.ZodDefault<z.ZodOptional<z.ZodEnum<["beginner", "intermediate", "expert"]>>>;
|
|
642
|
+
}, "strip", z.ZodTypeAny, {
|
|
643
|
+
model_type: string;
|
|
644
|
+
temperature: number;
|
|
645
|
+
max_tokens: number;
|
|
646
|
+
persona: string;
|
|
647
|
+
objective: string;
|
|
648
|
+
constraints: string[];
|
|
649
|
+
knowledge_level: "beginner" | "intermediate" | "expert";
|
|
650
|
+
style?: string | undefined;
|
|
651
|
+
emotional_state?: string | undefined;
|
|
652
|
+
}, {
|
|
653
|
+
persona: string;
|
|
654
|
+
objective: string;
|
|
655
|
+
style?: string | undefined;
|
|
656
|
+
model_type?: string | undefined;
|
|
657
|
+
temperature?: number | undefined;
|
|
658
|
+
max_tokens?: number | undefined;
|
|
659
|
+
constraints?: string[] | undefined;
|
|
660
|
+
emotional_state?: string | undefined;
|
|
661
|
+
knowledge_level?: "beginner" | "intermediate" | "expert" | undefined;
|
|
662
|
+
}>;
|
|
232
663
|
termination_conditions: z.ZodDefault<z.ZodOptional<z.ZodArray<z.ZodObject<{
|
|
233
|
-
type: z.ZodEnum<
|
|
234
|
-
max_turns_reached: "max_turns_reached";
|
|
235
|
-
user_expresses_satisfaction: "user_expresses_satisfaction";
|
|
236
|
-
agent_provides_solution: "agent_provides_solution";
|
|
237
|
-
conversation_stuck: "conversation_stuck";
|
|
238
|
-
escalation_needed: "escalation_needed";
|
|
239
|
-
goal_achieved: "goal_achieved";
|
|
240
|
-
custom_condition: "custom_condition";
|
|
241
|
-
}>;
|
|
664
|
+
type: z.ZodEnum<["max_turns_reached", "user_expresses_satisfaction", "agent_provides_solution", "conversation_stuck", "escalation_needed", "goal_achieved", "custom_condition"]>;
|
|
242
665
|
description: z.ZodOptional<z.ZodString>;
|
|
243
|
-
keywords: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
666
|
+
keywords: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
244
667
|
llm_judge: z.ZodOptional<z.ZodObject<{
|
|
245
668
|
prompt: z.ZodString;
|
|
246
669
|
threshold: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
247
|
-
}, z.
|
|
248
|
-
|
|
249
|
-
|
|
670
|
+
}, "strip", z.ZodTypeAny, {
|
|
671
|
+
prompt: string;
|
|
672
|
+
threshold: number;
|
|
673
|
+
}, {
|
|
674
|
+
prompt: string;
|
|
675
|
+
threshold?: number | undefined;
|
|
676
|
+
}>>;
|
|
677
|
+
}, "strip", z.ZodTypeAny, {
|
|
678
|
+
type: "max_turns_reached" | "user_expresses_satisfaction" | "agent_provides_solution" | "conversation_stuck" | "escalation_needed" | "goal_achieved" | "custom_condition";
|
|
679
|
+
description?: string | undefined;
|
|
680
|
+
llm_judge?: {
|
|
681
|
+
prompt: string;
|
|
682
|
+
threshold: number;
|
|
683
|
+
} | undefined;
|
|
684
|
+
keywords?: string[] | undefined;
|
|
685
|
+
}, {
|
|
686
|
+
type: "max_turns_reached" | "user_expresses_satisfaction" | "agent_provides_solution" | "conversation_stuck" | "escalation_needed" | "goal_achieved" | "custom_condition";
|
|
687
|
+
description?: string | undefined;
|
|
688
|
+
llm_judge?: {
|
|
689
|
+
prompt: string;
|
|
690
|
+
threshold?: number | undefined;
|
|
691
|
+
} | undefined;
|
|
692
|
+
keywords?: string[] | undefined;
|
|
693
|
+
}>, "many">>>;
|
|
694
|
+
turn_evaluations: z.ZodDefault<z.ZodOptional<z.ZodArray<z.ZodDiscriminatedUnion<"type", [z.ZodObject<z.objectUtil.extendShape<{
|
|
695
|
+
type: z.ZodString;
|
|
696
|
+
}, {
|
|
250
697
|
type: z.ZodLiteral<"string_contains">;
|
|
251
698
|
value: z.ZodString;
|
|
252
699
|
case_sensitive: z.ZodOptional<z.ZodBoolean>;
|
|
253
|
-
}, z.
|
|
700
|
+
}>, "strip", z.ZodTypeAny, {
|
|
701
|
+
value: string;
|
|
702
|
+
type: "string_contains";
|
|
703
|
+
case_sensitive?: boolean | undefined;
|
|
704
|
+
}, {
|
|
705
|
+
value: string;
|
|
706
|
+
type: "string_contains";
|
|
707
|
+
case_sensitive?: boolean | undefined;
|
|
708
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
709
|
+
type: z.ZodString;
|
|
710
|
+
}, {
|
|
254
711
|
type: z.ZodLiteral<"regex_match">;
|
|
255
712
|
pattern: z.ZodString;
|
|
256
|
-
}, z.
|
|
713
|
+
}>, "strip", z.ZodTypeAny, {
|
|
714
|
+
type: "regex_match";
|
|
715
|
+
pattern: string;
|
|
716
|
+
}, {
|
|
717
|
+
type: "regex_match";
|
|
718
|
+
pattern: string;
|
|
719
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
720
|
+
type: z.ZodString;
|
|
721
|
+
}, {
|
|
257
722
|
type: z.ZodLiteral<"file_exists">;
|
|
258
723
|
path: z.ZodString;
|
|
259
|
-
}, z.
|
|
724
|
+
}>, "strip", z.ZodTypeAny, {
|
|
725
|
+
path: string;
|
|
726
|
+
type: "file_exists";
|
|
727
|
+
}, {
|
|
728
|
+
path: string;
|
|
729
|
+
type: "file_exists";
|
|
730
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
731
|
+
type: z.ZodString;
|
|
732
|
+
}, {
|
|
260
733
|
type: z.ZodLiteral<"trajectory_contains_action">;
|
|
261
734
|
action: z.ZodString;
|
|
262
|
-
}, z.
|
|
735
|
+
}>, "strip", z.ZodTypeAny, {
|
|
736
|
+
type: "trajectory_contains_action";
|
|
737
|
+
action: string;
|
|
738
|
+
}, {
|
|
739
|
+
type: "trajectory_contains_action";
|
|
740
|
+
action: string;
|
|
741
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
742
|
+
type: z.ZodString;
|
|
743
|
+
}, {
|
|
263
744
|
type: z.ZodLiteral<"llm_judge">;
|
|
264
745
|
prompt: z.ZodString;
|
|
265
746
|
expected: z.ZodString;
|
|
266
747
|
model_type: z.ZodOptional<z.ZodString>;
|
|
267
748
|
temperature: z.ZodOptional<z.ZodNumber>;
|
|
268
749
|
json_schema: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodAny>>;
|
|
269
|
-
capabilities: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
270
|
-
}, z.
|
|
750
|
+
capabilities: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
751
|
+
}>, "strip", z.ZodTypeAny, {
|
|
752
|
+
type: "llm_judge";
|
|
753
|
+
expected: string;
|
|
754
|
+
prompt: string;
|
|
755
|
+
model_type?: string | undefined;
|
|
756
|
+
temperature?: number | undefined;
|
|
757
|
+
json_schema?: Record<string, any> | undefined;
|
|
758
|
+
capabilities?: string[] | undefined;
|
|
759
|
+
}, {
|
|
760
|
+
type: "llm_judge";
|
|
761
|
+
expected: string;
|
|
762
|
+
prompt: string;
|
|
763
|
+
model_type?: string | undefined;
|
|
764
|
+
temperature?: number | undefined;
|
|
765
|
+
json_schema?: Record<string, any> | undefined;
|
|
766
|
+
capabilities?: string[] | undefined;
|
|
767
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
768
|
+
type: z.ZodString;
|
|
769
|
+
}, {
|
|
271
770
|
type: z.ZodLiteral<"execution_time">;
|
|
272
771
|
max_duration_ms: z.ZodNumber;
|
|
273
772
|
min_duration_ms: z.ZodOptional<z.ZodNumber>;
|
|
274
773
|
target_duration_ms: z.ZodOptional<z.ZodNumber>;
|
|
275
|
-
}, z.
|
|
774
|
+
}>, "strip", z.ZodTypeAny, {
|
|
775
|
+
type: "execution_time";
|
|
776
|
+
max_duration_ms: number;
|
|
777
|
+
min_duration_ms?: number | undefined;
|
|
778
|
+
target_duration_ms?: number | undefined;
|
|
779
|
+
}, {
|
|
780
|
+
type: "execution_time";
|
|
781
|
+
max_duration_ms: number;
|
|
782
|
+
min_duration_ms?: number | undefined;
|
|
783
|
+
target_duration_ms?: number | undefined;
|
|
784
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
785
|
+
type: z.ZodString;
|
|
786
|
+
}, {
|
|
276
787
|
type: z.ZodLiteral<"conversation_length">;
|
|
277
788
|
min_turns: z.ZodOptional<z.ZodNumber>;
|
|
278
789
|
max_turns: z.ZodOptional<z.ZodNumber>;
|
|
279
790
|
optimal_turns: z.ZodOptional<z.ZodNumber>;
|
|
280
|
-
target_range: z.ZodOptional<z.ZodArray<z.ZodNumber
|
|
281
|
-
}, z.
|
|
791
|
+
target_range: z.ZodEffects<z.ZodOptional<z.ZodArray<z.ZodNumber, "many">>, number[] | undefined, number[] | undefined>;
|
|
792
|
+
}>, "strip", z.ZodTypeAny, {
|
|
793
|
+
type: "conversation_length";
|
|
794
|
+
min_turns?: number | undefined;
|
|
795
|
+
max_turns?: number | undefined;
|
|
796
|
+
optimal_turns?: number | undefined;
|
|
797
|
+
target_range?: number[] | undefined;
|
|
798
|
+
}, {
|
|
799
|
+
type: "conversation_length";
|
|
800
|
+
min_turns?: number | undefined;
|
|
801
|
+
max_turns?: number | undefined;
|
|
802
|
+
optimal_turns?: number | undefined;
|
|
803
|
+
target_range?: number[] | undefined;
|
|
804
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
805
|
+
type: z.ZodString;
|
|
806
|
+
}, {
|
|
282
807
|
type: z.ZodLiteral<"conversation_flow">;
|
|
283
|
-
required_patterns: z.ZodArray<z.ZodEnum<
|
|
284
|
-
question_then_answer: "question_then_answer";
|
|
285
|
-
problem_then_solution: "problem_then_solution";
|
|
286
|
-
clarification_cycle: "clarification_cycle";
|
|
287
|
-
empathy_then_solution: "empathy_then_solution";
|
|
288
|
-
escalation_pattern: "escalation_pattern";
|
|
289
|
-
}>>;
|
|
808
|
+
required_patterns: z.ZodArray<z.ZodEnum<["question_then_answer", "problem_then_solution", "clarification_cycle", "empathy_then_solution", "escalation_pattern"]>, "many">;
|
|
290
809
|
flow_quality_threshold: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
291
|
-
}, z.
|
|
810
|
+
}>, "strip", z.ZodTypeAny, {
|
|
811
|
+
type: "conversation_flow";
|
|
812
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
813
|
+
flow_quality_threshold: number;
|
|
814
|
+
}, {
|
|
815
|
+
type: "conversation_flow";
|
|
816
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
817
|
+
flow_quality_threshold?: number | undefined;
|
|
818
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
819
|
+
type: z.ZodString;
|
|
820
|
+
}, {
|
|
292
821
|
type: z.ZodLiteral<"user_satisfaction">;
|
|
293
822
|
satisfaction_threshold: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
294
823
|
indicators: z.ZodOptional<z.ZodObject<{
|
|
295
|
-
positive: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
296
|
-
negative: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
297
|
-
}, z.
|
|
298
|
-
|
|
299
|
-
|
|
300
|
-
|
|
301
|
-
|
|
302
|
-
|
|
303
|
-
|
|
824
|
+
positive: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
825
|
+
negative: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
826
|
+
}, "strip", z.ZodTypeAny, {
|
|
827
|
+
positive?: string[] | undefined;
|
|
828
|
+
negative?: string[] | undefined;
|
|
829
|
+
}, {
|
|
830
|
+
positive?: string[] | undefined;
|
|
831
|
+
negative?: string[] | undefined;
|
|
832
|
+
}>>;
|
|
833
|
+
measurement_method: z.ZodDefault<z.ZodOptional<z.ZodEnum<["sentiment_analysis", "keyword_analysis", "llm_judge"]>>>;
|
|
834
|
+
}>, "strip", z.ZodTypeAny, {
|
|
835
|
+
type: "user_satisfaction";
|
|
836
|
+
satisfaction_threshold: number;
|
|
837
|
+
measurement_method: "llm_judge" | "sentiment_analysis" | "keyword_analysis";
|
|
838
|
+
indicators?: {
|
|
839
|
+
positive?: string[] | undefined;
|
|
840
|
+
negative?: string[] | undefined;
|
|
841
|
+
} | undefined;
|
|
842
|
+
}, {
|
|
843
|
+
type: "user_satisfaction";
|
|
844
|
+
satisfaction_threshold?: number | undefined;
|
|
845
|
+
indicators?: {
|
|
846
|
+
positive?: string[] | undefined;
|
|
847
|
+
negative?: string[] | undefined;
|
|
848
|
+
} | undefined;
|
|
849
|
+
measurement_method?: "llm_judge" | "sentiment_analysis" | "keyword_analysis" | undefined;
|
|
850
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
851
|
+
type: z.ZodString;
|
|
852
|
+
}, {
|
|
304
853
|
type: z.ZodLiteral<"context_retention">;
|
|
305
|
-
test_memory_of: z.ZodArray<z.ZodString>;
|
|
854
|
+
test_memory_of: z.ZodArray<z.ZodString, "many">;
|
|
306
855
|
retention_turns: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
307
856
|
memory_accuracy_threshold: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
308
|
-
}, z.
|
|
309
|
-
|
|
857
|
+
}>, "strip", z.ZodTypeAny, {
|
|
858
|
+
type: "context_retention";
|
|
859
|
+
test_memory_of: string[];
|
|
860
|
+
retention_turns: number;
|
|
861
|
+
memory_accuracy_threshold: number;
|
|
862
|
+
}, {
|
|
863
|
+
type: "context_retention";
|
|
864
|
+
test_memory_of: string[];
|
|
865
|
+
retention_turns?: number | undefined;
|
|
866
|
+
memory_accuracy_threshold?: number | undefined;
|
|
867
|
+
}>]>, "many">>>;
|
|
868
|
+
final_evaluations: z.ZodDefault<z.ZodOptional<z.ZodArray<z.ZodDiscriminatedUnion<"type", [z.ZodObject<z.objectUtil.extendShape<{
|
|
869
|
+
type: z.ZodString;
|
|
870
|
+
}, {
|
|
310
871
|
type: z.ZodLiteral<"string_contains">;
|
|
311
872
|
value: z.ZodString;
|
|
312
873
|
case_sensitive: z.ZodOptional<z.ZodBoolean>;
|
|
313
|
-
}, z.
|
|
874
|
+
}>, "strip", z.ZodTypeAny, {
|
|
875
|
+
value: string;
|
|
876
|
+
type: "string_contains";
|
|
877
|
+
case_sensitive?: boolean | undefined;
|
|
878
|
+
}, {
|
|
879
|
+
value: string;
|
|
880
|
+
type: "string_contains";
|
|
881
|
+
case_sensitive?: boolean | undefined;
|
|
882
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
883
|
+
type: z.ZodString;
|
|
884
|
+
}, {
|
|
314
885
|
type: z.ZodLiteral<"regex_match">;
|
|
315
886
|
pattern: z.ZodString;
|
|
316
|
-
}, z.
|
|
887
|
+
}>, "strip", z.ZodTypeAny, {
|
|
888
|
+
type: "regex_match";
|
|
889
|
+
pattern: string;
|
|
890
|
+
}, {
|
|
891
|
+
type: "regex_match";
|
|
892
|
+
pattern: string;
|
|
893
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
894
|
+
type: z.ZodString;
|
|
895
|
+
}, {
|
|
317
896
|
type: z.ZodLiteral<"file_exists">;
|
|
318
897
|
path: z.ZodString;
|
|
319
|
-
}, z.
|
|
898
|
+
}>, "strip", z.ZodTypeAny, {
|
|
899
|
+
path: string;
|
|
900
|
+
type: "file_exists";
|
|
901
|
+
}, {
|
|
902
|
+
path: string;
|
|
903
|
+
type: "file_exists";
|
|
904
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
905
|
+
type: z.ZodString;
|
|
906
|
+
}, {
|
|
320
907
|
type: z.ZodLiteral<"trajectory_contains_action">;
|
|
321
908
|
action: z.ZodString;
|
|
322
|
-
}, z.
|
|
909
|
+
}>, "strip", z.ZodTypeAny, {
|
|
910
|
+
type: "trajectory_contains_action";
|
|
911
|
+
action: string;
|
|
912
|
+
}, {
|
|
913
|
+
type: "trajectory_contains_action";
|
|
914
|
+
action: string;
|
|
915
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
916
|
+
type: z.ZodString;
|
|
917
|
+
}, {
|
|
323
918
|
type: z.ZodLiteral<"llm_judge">;
|
|
324
919
|
prompt: z.ZodString;
|
|
325
920
|
expected: z.ZodString;
|
|
326
921
|
model_type: z.ZodOptional<z.ZodString>;
|
|
327
922
|
temperature: z.ZodOptional<z.ZodNumber>;
|
|
328
923
|
json_schema: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodAny>>;
|
|
329
|
-
capabilities: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
330
|
-
}, z.
|
|
924
|
+
capabilities: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
925
|
+
}>, "strip", z.ZodTypeAny, {
|
|
926
|
+
type: "llm_judge";
|
|
927
|
+
expected: string;
|
|
928
|
+
prompt: string;
|
|
929
|
+
model_type?: string | undefined;
|
|
930
|
+
temperature?: number | undefined;
|
|
931
|
+
json_schema?: Record<string, any> | undefined;
|
|
932
|
+
capabilities?: string[] | undefined;
|
|
933
|
+
}, {
|
|
934
|
+
type: "llm_judge";
|
|
935
|
+
expected: string;
|
|
936
|
+
prompt: string;
|
|
937
|
+
model_type?: string | undefined;
|
|
938
|
+
temperature?: number | undefined;
|
|
939
|
+
json_schema?: Record<string, any> | undefined;
|
|
940
|
+
capabilities?: string[] | undefined;
|
|
941
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
942
|
+
type: z.ZodString;
|
|
943
|
+
}, {
|
|
331
944
|
type: z.ZodLiteral<"execution_time">;
|
|
332
945
|
max_duration_ms: z.ZodNumber;
|
|
333
946
|
min_duration_ms: z.ZodOptional<z.ZodNumber>;
|
|
334
947
|
target_duration_ms: z.ZodOptional<z.ZodNumber>;
|
|
335
|
-
}, z.
|
|
948
|
+
}>, "strip", z.ZodTypeAny, {
|
|
949
|
+
type: "execution_time";
|
|
950
|
+
max_duration_ms: number;
|
|
951
|
+
min_duration_ms?: number | undefined;
|
|
952
|
+
target_duration_ms?: number | undefined;
|
|
953
|
+
}, {
|
|
954
|
+
type: "execution_time";
|
|
955
|
+
max_duration_ms: number;
|
|
956
|
+
min_duration_ms?: number | undefined;
|
|
957
|
+
target_duration_ms?: number | undefined;
|
|
958
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
959
|
+
type: z.ZodString;
|
|
960
|
+
}, {
|
|
336
961
|
type: z.ZodLiteral<"conversation_length">;
|
|
337
962
|
min_turns: z.ZodOptional<z.ZodNumber>;
|
|
338
963
|
max_turns: z.ZodOptional<z.ZodNumber>;
|
|
339
964
|
optimal_turns: z.ZodOptional<z.ZodNumber>;
|
|
340
|
-
target_range: z.ZodOptional<z.ZodArray<z.ZodNumber
|
|
341
|
-
}, z.
|
|
965
|
+
target_range: z.ZodEffects<z.ZodOptional<z.ZodArray<z.ZodNumber, "many">>, number[] | undefined, number[] | undefined>;
|
|
966
|
+
}>, "strip", z.ZodTypeAny, {
|
|
967
|
+
type: "conversation_length";
|
|
968
|
+
min_turns?: number | undefined;
|
|
969
|
+
max_turns?: number | undefined;
|
|
970
|
+
optimal_turns?: number | undefined;
|
|
971
|
+
target_range?: number[] | undefined;
|
|
972
|
+
}, {
|
|
973
|
+
type: "conversation_length";
|
|
974
|
+
min_turns?: number | undefined;
|
|
975
|
+
max_turns?: number | undefined;
|
|
976
|
+
optimal_turns?: number | undefined;
|
|
977
|
+
target_range?: number[] | undefined;
|
|
978
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
979
|
+
type: z.ZodString;
|
|
980
|
+
}, {
|
|
342
981
|
type: z.ZodLiteral<"conversation_flow">;
|
|
343
|
-
required_patterns: z.ZodArray<z.ZodEnum<
|
|
344
|
-
question_then_answer: "question_then_answer";
|
|
345
|
-
problem_then_solution: "problem_then_solution";
|
|
346
|
-
clarification_cycle: "clarification_cycle";
|
|
347
|
-
empathy_then_solution: "empathy_then_solution";
|
|
348
|
-
escalation_pattern: "escalation_pattern";
|
|
349
|
-
}>>;
|
|
982
|
+
required_patterns: z.ZodArray<z.ZodEnum<["question_then_answer", "problem_then_solution", "clarification_cycle", "empathy_then_solution", "escalation_pattern"]>, "many">;
|
|
350
983
|
flow_quality_threshold: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
351
|
-
}, z.
|
|
984
|
+
}>, "strip", z.ZodTypeAny, {
|
|
985
|
+
type: "conversation_flow";
|
|
986
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
987
|
+
flow_quality_threshold: number;
|
|
988
|
+
}, {
|
|
989
|
+
type: "conversation_flow";
|
|
990
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
991
|
+
flow_quality_threshold?: number | undefined;
|
|
992
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
993
|
+
type: z.ZodString;
|
|
994
|
+
}, {
|
|
352
995
|
type: z.ZodLiteral<"user_satisfaction">;
|
|
353
996
|
satisfaction_threshold: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
354
997
|
indicators: z.ZodOptional<z.ZodObject<{
|
|
355
|
-
positive: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
356
|
-
negative: z.ZodOptional<z.ZodArray<z.ZodString>>;
|
|
357
|
-
}, z.
|
|
358
|
-
|
|
359
|
-
|
|
360
|
-
|
|
361
|
-
|
|
362
|
-
|
|
363
|
-
|
|
998
|
+
positive: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
999
|
+
negative: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
|
|
1000
|
+
}, "strip", z.ZodTypeAny, {
|
|
1001
|
+
positive?: string[] | undefined;
|
|
1002
|
+
negative?: string[] | undefined;
|
|
1003
|
+
}, {
|
|
1004
|
+
positive?: string[] | undefined;
|
|
1005
|
+
negative?: string[] | undefined;
|
|
1006
|
+
}>>;
|
|
1007
|
+
measurement_method: z.ZodDefault<z.ZodOptional<z.ZodEnum<["sentiment_analysis", "keyword_analysis", "llm_judge"]>>>;
|
|
1008
|
+
}>, "strip", z.ZodTypeAny, {
|
|
1009
|
+
type: "user_satisfaction";
|
|
1010
|
+
satisfaction_threshold: number;
|
|
1011
|
+
measurement_method: "llm_judge" | "sentiment_analysis" | "keyword_analysis";
|
|
1012
|
+
indicators?: {
|
|
1013
|
+
positive?: string[] | undefined;
|
|
1014
|
+
negative?: string[] | undefined;
|
|
1015
|
+
} | undefined;
|
|
1016
|
+
}, {
|
|
1017
|
+
type: "user_satisfaction";
|
|
1018
|
+
satisfaction_threshold?: number | undefined;
|
|
1019
|
+
indicators?: {
|
|
1020
|
+
positive?: string[] | undefined;
|
|
1021
|
+
negative?: string[] | undefined;
|
|
1022
|
+
} | undefined;
|
|
1023
|
+
measurement_method?: "llm_judge" | "sentiment_analysis" | "keyword_analysis" | undefined;
|
|
1024
|
+
}>, z.ZodObject<z.objectUtil.extendShape<{
|
|
1025
|
+
type: z.ZodString;
|
|
1026
|
+
}, {
|
|
364
1027
|
type: z.ZodLiteral<"context_retention">;
|
|
365
|
-
test_memory_of: z.ZodArray<z.ZodString>;
|
|
1028
|
+
test_memory_of: z.ZodArray<z.ZodString, "many">;
|
|
366
1029
|
retention_turns: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
367
1030
|
memory_accuracy_threshold: z.ZodDefault<z.ZodOptional<z.ZodNumber>>;
|
|
368
|
-
}, z.
|
|
1031
|
+
}>, "strip", z.ZodTypeAny, {
|
|
1032
|
+
type: "context_retention";
|
|
1033
|
+
test_memory_of: string[];
|
|
1034
|
+
retention_turns: number;
|
|
1035
|
+
memory_accuracy_threshold: number;
|
|
1036
|
+
}, {
|
|
1037
|
+
type: "context_retention";
|
|
1038
|
+
test_memory_of: string[];
|
|
1039
|
+
retention_turns?: number | undefined;
|
|
1040
|
+
memory_accuracy_threshold?: number | undefined;
|
|
1041
|
+
}>]>, "many">>>;
|
|
369
1042
|
debug_options: z.ZodDefault<z.ZodOptional<z.ZodObject<{
|
|
370
1043
|
log_user_simulation: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
371
1044
|
log_turn_decisions: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
372
1045
|
export_full_transcript: z.ZodDefault<z.ZodOptional<z.ZodBoolean>>;
|
|
373
|
-
}, z.
|
|
374
|
-
|
|
375
|
-
|
|
1046
|
+
}, "strip", z.ZodTypeAny, {
|
|
1047
|
+
log_user_simulation: boolean;
|
|
1048
|
+
log_turn_decisions: boolean;
|
|
1049
|
+
export_full_transcript: boolean;
|
|
1050
|
+
}, {
|
|
1051
|
+
log_user_simulation?: boolean | undefined;
|
|
1052
|
+
log_turn_decisions?: boolean | undefined;
|
|
1053
|
+
export_full_transcript?: boolean | undefined;
|
|
1054
|
+
}>>>;
|
|
1055
|
+
}, "strip", z.ZodTypeAny, {
|
|
1056
|
+
max_turns: number;
|
|
1057
|
+
timeout_per_turn_ms: number;
|
|
1058
|
+
total_timeout_ms: number;
|
|
1059
|
+
user_simulator: {
|
|
1060
|
+
model_type: string;
|
|
1061
|
+
temperature: number;
|
|
1062
|
+
max_tokens: number;
|
|
1063
|
+
persona: string;
|
|
1064
|
+
objective: string;
|
|
1065
|
+
constraints: string[];
|
|
1066
|
+
knowledge_level: "beginner" | "intermediate" | "expert";
|
|
1067
|
+
style?: string | undefined;
|
|
1068
|
+
emotional_state?: string | undefined;
|
|
1069
|
+
};
|
|
1070
|
+
termination_conditions: {
|
|
1071
|
+
type: "max_turns_reached" | "user_expresses_satisfaction" | "agent_provides_solution" | "conversation_stuck" | "escalation_needed" | "goal_achieved" | "custom_condition";
|
|
1072
|
+
description?: string | undefined;
|
|
1073
|
+
llm_judge?: {
|
|
1074
|
+
prompt: string;
|
|
1075
|
+
threshold: number;
|
|
1076
|
+
} | undefined;
|
|
1077
|
+
keywords?: string[] | undefined;
|
|
1078
|
+
}[];
|
|
1079
|
+
turn_evaluations: ({
|
|
1080
|
+
value: string;
|
|
1081
|
+
type: "string_contains";
|
|
1082
|
+
case_sensitive?: boolean | undefined;
|
|
1083
|
+
} | {
|
|
1084
|
+
type: "regex_match";
|
|
1085
|
+
pattern: string;
|
|
1086
|
+
} | {
|
|
1087
|
+
path: string;
|
|
1088
|
+
type: "file_exists";
|
|
1089
|
+
} | {
|
|
1090
|
+
type: "trajectory_contains_action";
|
|
1091
|
+
action: string;
|
|
1092
|
+
} | {
|
|
1093
|
+
type: "llm_judge";
|
|
1094
|
+
expected: string;
|
|
1095
|
+
prompt: string;
|
|
1096
|
+
model_type?: string | undefined;
|
|
1097
|
+
temperature?: number | undefined;
|
|
1098
|
+
json_schema?: Record<string, any> | undefined;
|
|
1099
|
+
capabilities?: string[] | undefined;
|
|
1100
|
+
} | {
|
|
1101
|
+
type: "execution_time";
|
|
1102
|
+
max_duration_ms: number;
|
|
1103
|
+
min_duration_ms?: number | undefined;
|
|
1104
|
+
target_duration_ms?: number | undefined;
|
|
1105
|
+
} | {
|
|
1106
|
+
type: "conversation_length";
|
|
1107
|
+
min_turns?: number | undefined;
|
|
1108
|
+
max_turns?: number | undefined;
|
|
1109
|
+
optimal_turns?: number | undefined;
|
|
1110
|
+
target_range?: number[] | undefined;
|
|
1111
|
+
} | {
|
|
1112
|
+
type: "conversation_flow";
|
|
1113
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
1114
|
+
flow_quality_threshold: number;
|
|
1115
|
+
} | {
|
|
1116
|
+
type: "user_satisfaction";
|
|
1117
|
+
satisfaction_threshold: number;
|
|
1118
|
+
measurement_method: "llm_judge" | "sentiment_analysis" | "keyword_analysis";
|
|
1119
|
+
indicators?: {
|
|
1120
|
+
positive?: string[] | undefined;
|
|
1121
|
+
negative?: string[] | undefined;
|
|
1122
|
+
} | undefined;
|
|
1123
|
+
} | {
|
|
1124
|
+
type: "context_retention";
|
|
1125
|
+
test_memory_of: string[];
|
|
1126
|
+
retention_turns: number;
|
|
1127
|
+
memory_accuracy_threshold: number;
|
|
1128
|
+
})[];
|
|
1129
|
+
final_evaluations: ({
|
|
1130
|
+
value: string;
|
|
1131
|
+
type: "string_contains";
|
|
1132
|
+
case_sensitive?: boolean | undefined;
|
|
1133
|
+
} | {
|
|
1134
|
+
type: "regex_match";
|
|
1135
|
+
pattern: string;
|
|
1136
|
+
} | {
|
|
1137
|
+
path: string;
|
|
1138
|
+
type: "file_exists";
|
|
1139
|
+
} | {
|
|
1140
|
+
type: "trajectory_contains_action";
|
|
1141
|
+
action: string;
|
|
1142
|
+
} | {
|
|
1143
|
+
type: "llm_judge";
|
|
1144
|
+
expected: string;
|
|
1145
|
+
prompt: string;
|
|
1146
|
+
model_type?: string | undefined;
|
|
1147
|
+
temperature?: number | undefined;
|
|
1148
|
+
json_schema?: Record<string, any> | undefined;
|
|
1149
|
+
capabilities?: string[] | undefined;
|
|
1150
|
+
} | {
|
|
1151
|
+
type: "execution_time";
|
|
1152
|
+
max_duration_ms: number;
|
|
1153
|
+
min_duration_ms?: number | undefined;
|
|
1154
|
+
target_duration_ms?: number | undefined;
|
|
1155
|
+
} | {
|
|
1156
|
+
type: "conversation_length";
|
|
1157
|
+
min_turns?: number | undefined;
|
|
1158
|
+
max_turns?: number | undefined;
|
|
1159
|
+
optimal_turns?: number | undefined;
|
|
1160
|
+
target_range?: number[] | undefined;
|
|
1161
|
+
} | {
|
|
1162
|
+
type: "conversation_flow";
|
|
1163
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
1164
|
+
flow_quality_threshold: number;
|
|
1165
|
+
} | {
|
|
1166
|
+
type: "user_satisfaction";
|
|
1167
|
+
satisfaction_threshold: number;
|
|
1168
|
+
measurement_method: "llm_judge" | "sentiment_analysis" | "keyword_analysis";
|
|
1169
|
+
indicators?: {
|
|
1170
|
+
positive?: string[] | undefined;
|
|
1171
|
+
negative?: string[] | undefined;
|
|
1172
|
+
} | undefined;
|
|
1173
|
+
} | {
|
|
1174
|
+
type: "context_retention";
|
|
1175
|
+
test_memory_of: string[];
|
|
1176
|
+
retention_turns: number;
|
|
1177
|
+
memory_accuracy_threshold: number;
|
|
1178
|
+
})[];
|
|
1179
|
+
debug_options: {
|
|
1180
|
+
log_user_simulation: boolean;
|
|
1181
|
+
log_turn_decisions: boolean;
|
|
1182
|
+
export_full_transcript: boolean;
|
|
1183
|
+
};
|
|
1184
|
+
}, {
|
|
1185
|
+
max_turns: number;
|
|
1186
|
+
user_simulator: {
|
|
1187
|
+
persona: string;
|
|
1188
|
+
objective: string;
|
|
1189
|
+
style?: string | undefined;
|
|
1190
|
+
model_type?: string | undefined;
|
|
1191
|
+
temperature?: number | undefined;
|
|
1192
|
+
max_tokens?: number | undefined;
|
|
1193
|
+
constraints?: string[] | undefined;
|
|
1194
|
+
emotional_state?: string | undefined;
|
|
1195
|
+
knowledge_level?: "beginner" | "intermediate" | "expert" | undefined;
|
|
1196
|
+
};
|
|
1197
|
+
timeout_per_turn_ms?: number | undefined;
|
|
1198
|
+
total_timeout_ms?: number | undefined;
|
|
1199
|
+
termination_conditions?: {
|
|
1200
|
+
type: "max_turns_reached" | "user_expresses_satisfaction" | "agent_provides_solution" | "conversation_stuck" | "escalation_needed" | "goal_achieved" | "custom_condition";
|
|
1201
|
+
description?: string | undefined;
|
|
1202
|
+
llm_judge?: {
|
|
1203
|
+
prompt: string;
|
|
1204
|
+
threshold?: number | undefined;
|
|
1205
|
+
} | undefined;
|
|
1206
|
+
keywords?: string[] | undefined;
|
|
1207
|
+
}[] | undefined;
|
|
1208
|
+
turn_evaluations?: ({
|
|
1209
|
+
value: string;
|
|
1210
|
+
type: "string_contains";
|
|
1211
|
+
case_sensitive?: boolean | undefined;
|
|
1212
|
+
} | {
|
|
1213
|
+
type: "regex_match";
|
|
1214
|
+
pattern: string;
|
|
1215
|
+
} | {
|
|
1216
|
+
path: string;
|
|
1217
|
+
type: "file_exists";
|
|
1218
|
+
} | {
|
|
1219
|
+
type: "trajectory_contains_action";
|
|
1220
|
+
action: string;
|
|
1221
|
+
} | {
|
|
1222
|
+
type: "llm_judge";
|
|
1223
|
+
expected: string;
|
|
1224
|
+
prompt: string;
|
|
1225
|
+
model_type?: string | undefined;
|
|
1226
|
+
temperature?: number | undefined;
|
|
1227
|
+
json_schema?: Record<string, any> | undefined;
|
|
1228
|
+
capabilities?: string[] | undefined;
|
|
1229
|
+
} | {
|
|
1230
|
+
type: "execution_time";
|
|
1231
|
+
max_duration_ms: number;
|
|
1232
|
+
min_duration_ms?: number | undefined;
|
|
1233
|
+
target_duration_ms?: number | undefined;
|
|
1234
|
+
} | {
|
|
1235
|
+
type: "conversation_length";
|
|
1236
|
+
min_turns?: number | undefined;
|
|
1237
|
+
max_turns?: number | undefined;
|
|
1238
|
+
optimal_turns?: number | undefined;
|
|
1239
|
+
target_range?: number[] | undefined;
|
|
1240
|
+
} | {
|
|
1241
|
+
type: "conversation_flow";
|
|
1242
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
1243
|
+
flow_quality_threshold?: number | undefined;
|
|
1244
|
+
} | {
|
|
1245
|
+
type: "user_satisfaction";
|
|
1246
|
+
satisfaction_threshold?: number | undefined;
|
|
1247
|
+
indicators?: {
|
|
1248
|
+
positive?: string[] | undefined;
|
|
1249
|
+
negative?: string[] | undefined;
|
|
1250
|
+
} | undefined;
|
|
1251
|
+
measurement_method?: "llm_judge" | "sentiment_analysis" | "keyword_analysis" | undefined;
|
|
1252
|
+
} | {
|
|
1253
|
+
type: "context_retention";
|
|
1254
|
+
test_memory_of: string[];
|
|
1255
|
+
retention_turns?: number | undefined;
|
|
1256
|
+
memory_accuracy_threshold?: number | undefined;
|
|
1257
|
+
})[] | undefined;
|
|
1258
|
+
final_evaluations?: ({
|
|
1259
|
+
value: string;
|
|
1260
|
+
type: "string_contains";
|
|
1261
|
+
case_sensitive?: boolean | undefined;
|
|
1262
|
+
} | {
|
|
1263
|
+
type: "regex_match";
|
|
1264
|
+
pattern: string;
|
|
1265
|
+
} | {
|
|
1266
|
+
path: string;
|
|
1267
|
+
type: "file_exists";
|
|
1268
|
+
} | {
|
|
1269
|
+
type: "trajectory_contains_action";
|
|
1270
|
+
action: string;
|
|
1271
|
+
} | {
|
|
1272
|
+
type: "llm_judge";
|
|
1273
|
+
expected: string;
|
|
1274
|
+
prompt: string;
|
|
1275
|
+
model_type?: string | undefined;
|
|
1276
|
+
temperature?: number | undefined;
|
|
1277
|
+
json_schema?: Record<string, any> | undefined;
|
|
1278
|
+
capabilities?: string[] | undefined;
|
|
1279
|
+
} | {
|
|
1280
|
+
type: "execution_time";
|
|
1281
|
+
max_duration_ms: number;
|
|
1282
|
+
min_duration_ms?: number | undefined;
|
|
1283
|
+
target_duration_ms?: number | undefined;
|
|
1284
|
+
} | {
|
|
1285
|
+
type: "conversation_length";
|
|
1286
|
+
min_turns?: number | undefined;
|
|
1287
|
+
max_turns?: number | undefined;
|
|
1288
|
+
optimal_turns?: number | undefined;
|
|
1289
|
+
target_range?: number[] | undefined;
|
|
1290
|
+
} | {
|
|
1291
|
+
type: "conversation_flow";
|
|
1292
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
1293
|
+
flow_quality_threshold?: number | undefined;
|
|
1294
|
+
} | {
|
|
1295
|
+
type: "user_satisfaction";
|
|
1296
|
+
satisfaction_threshold?: number | undefined;
|
|
1297
|
+
indicators?: {
|
|
1298
|
+
positive?: string[] | undefined;
|
|
1299
|
+
negative?: string[] | undefined;
|
|
1300
|
+
} | undefined;
|
|
1301
|
+
measurement_method?: "llm_judge" | "sentiment_analysis" | "keyword_analysis" | undefined;
|
|
1302
|
+
} | {
|
|
1303
|
+
type: "context_retention";
|
|
1304
|
+
test_memory_of: string[];
|
|
1305
|
+
retention_turns?: number | undefined;
|
|
1306
|
+
memory_accuracy_threshold?: number | undefined;
|
|
1307
|
+
})[] | undefined;
|
|
1308
|
+
debug_options?: {
|
|
1309
|
+
log_user_simulation?: boolean | undefined;
|
|
1310
|
+
log_turn_decisions?: boolean | undefined;
|
|
1311
|
+
export_full_transcript?: boolean | undefined;
|
|
1312
|
+
} | undefined;
|
|
1313
|
+
}>>;
|
|
1314
|
+
}, "strip", z.ZodTypeAny, {
|
|
1315
|
+
evaluations: ({
|
|
1316
|
+
value: string;
|
|
1317
|
+
type: "string_contains";
|
|
1318
|
+
case_sensitive?: boolean | undefined;
|
|
1319
|
+
} | {
|
|
1320
|
+
type: "regex_match";
|
|
1321
|
+
pattern: string;
|
|
1322
|
+
} | {
|
|
1323
|
+
path: string;
|
|
1324
|
+
type: "file_exists";
|
|
1325
|
+
} | {
|
|
1326
|
+
type: "trajectory_contains_action";
|
|
1327
|
+
action: string;
|
|
1328
|
+
} | {
|
|
1329
|
+
type: "llm_judge";
|
|
1330
|
+
expected: string;
|
|
1331
|
+
prompt: string;
|
|
1332
|
+
model_type?: string | undefined;
|
|
1333
|
+
temperature?: number | undefined;
|
|
1334
|
+
json_schema?: Record<string, any> | undefined;
|
|
1335
|
+
capabilities?: string[] | undefined;
|
|
1336
|
+
} | {
|
|
1337
|
+
type: "execution_time";
|
|
1338
|
+
max_duration_ms: number;
|
|
1339
|
+
min_duration_ms?: number | undefined;
|
|
1340
|
+
target_duration_ms?: number | undefined;
|
|
1341
|
+
} | {
|
|
1342
|
+
type: "conversation_length";
|
|
1343
|
+
min_turns?: number | undefined;
|
|
1344
|
+
max_turns?: number | undefined;
|
|
1345
|
+
optimal_turns?: number | undefined;
|
|
1346
|
+
target_range?: number[] | undefined;
|
|
1347
|
+
} | {
|
|
1348
|
+
type: "conversation_flow";
|
|
1349
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
1350
|
+
flow_quality_threshold: number;
|
|
1351
|
+
} | {
|
|
1352
|
+
type: "user_satisfaction";
|
|
1353
|
+
satisfaction_threshold: number;
|
|
1354
|
+
measurement_method: "llm_judge" | "sentiment_analysis" | "keyword_analysis";
|
|
1355
|
+
indicators?: {
|
|
1356
|
+
positive?: string[] | undefined;
|
|
1357
|
+
negative?: string[] | undefined;
|
|
1358
|
+
} | undefined;
|
|
1359
|
+
} | {
|
|
1360
|
+
type: "context_retention";
|
|
1361
|
+
test_memory_of: string[];
|
|
1362
|
+
retention_turns: number;
|
|
1363
|
+
memory_accuracy_threshold: number;
|
|
1364
|
+
})[];
|
|
1365
|
+
name?: string | undefined;
|
|
1366
|
+
code?: string | undefined;
|
|
1367
|
+
input?: string | undefined;
|
|
1368
|
+
lang?: string | undefined;
|
|
1369
|
+
conversation?: {
|
|
1370
|
+
max_turns: number;
|
|
1371
|
+
timeout_per_turn_ms: number;
|
|
1372
|
+
total_timeout_ms: number;
|
|
1373
|
+
user_simulator: {
|
|
1374
|
+
model_type: string;
|
|
1375
|
+
temperature: number;
|
|
1376
|
+
max_tokens: number;
|
|
1377
|
+
persona: string;
|
|
1378
|
+
objective: string;
|
|
1379
|
+
constraints: string[];
|
|
1380
|
+
knowledge_level: "beginner" | "intermediate" | "expert";
|
|
1381
|
+
style?: string | undefined;
|
|
1382
|
+
emotional_state?: string | undefined;
|
|
1383
|
+
};
|
|
1384
|
+
termination_conditions: {
|
|
1385
|
+
type: "max_turns_reached" | "user_expresses_satisfaction" | "agent_provides_solution" | "conversation_stuck" | "escalation_needed" | "goal_achieved" | "custom_condition";
|
|
1386
|
+
description?: string | undefined;
|
|
1387
|
+
llm_judge?: {
|
|
1388
|
+
prompt: string;
|
|
1389
|
+
threshold: number;
|
|
1390
|
+
} | undefined;
|
|
1391
|
+
keywords?: string[] | undefined;
|
|
1392
|
+
}[];
|
|
1393
|
+
turn_evaluations: ({
|
|
1394
|
+
value: string;
|
|
1395
|
+
type: "string_contains";
|
|
1396
|
+
case_sensitive?: boolean | undefined;
|
|
1397
|
+
} | {
|
|
1398
|
+
type: "regex_match";
|
|
1399
|
+
pattern: string;
|
|
1400
|
+
} | {
|
|
1401
|
+
path: string;
|
|
1402
|
+
type: "file_exists";
|
|
1403
|
+
} | {
|
|
1404
|
+
type: "trajectory_contains_action";
|
|
1405
|
+
action: string;
|
|
1406
|
+
} | {
|
|
1407
|
+
type: "llm_judge";
|
|
1408
|
+
expected: string;
|
|
1409
|
+
prompt: string;
|
|
1410
|
+
model_type?: string | undefined;
|
|
1411
|
+
temperature?: number | undefined;
|
|
1412
|
+
json_schema?: Record<string, any> | undefined;
|
|
1413
|
+
capabilities?: string[] | undefined;
|
|
1414
|
+
} | {
|
|
1415
|
+
type: "execution_time";
|
|
1416
|
+
max_duration_ms: number;
|
|
1417
|
+
min_duration_ms?: number | undefined;
|
|
1418
|
+
target_duration_ms?: number | undefined;
|
|
1419
|
+
} | {
|
|
1420
|
+
type: "conversation_length";
|
|
1421
|
+
min_turns?: number | undefined;
|
|
1422
|
+
max_turns?: number | undefined;
|
|
1423
|
+
optimal_turns?: number | undefined;
|
|
1424
|
+
target_range?: number[] | undefined;
|
|
1425
|
+
} | {
|
|
1426
|
+
type: "conversation_flow";
|
|
1427
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
1428
|
+
flow_quality_threshold: number;
|
|
1429
|
+
} | {
|
|
1430
|
+
type: "user_satisfaction";
|
|
1431
|
+
satisfaction_threshold: number;
|
|
1432
|
+
measurement_method: "llm_judge" | "sentiment_analysis" | "keyword_analysis";
|
|
1433
|
+
indicators?: {
|
|
1434
|
+
positive?: string[] | undefined;
|
|
1435
|
+
negative?: string[] | undefined;
|
|
1436
|
+
} | undefined;
|
|
1437
|
+
} | {
|
|
1438
|
+
type: "context_retention";
|
|
1439
|
+
test_memory_of: string[];
|
|
1440
|
+
retention_turns: number;
|
|
1441
|
+
memory_accuracy_threshold: number;
|
|
1442
|
+
})[];
|
|
1443
|
+
final_evaluations: ({
|
|
1444
|
+
value: string;
|
|
1445
|
+
type: "string_contains";
|
|
1446
|
+
case_sensitive?: boolean | undefined;
|
|
1447
|
+
} | {
|
|
1448
|
+
type: "regex_match";
|
|
1449
|
+
pattern: string;
|
|
1450
|
+
} | {
|
|
1451
|
+
path: string;
|
|
1452
|
+
type: "file_exists";
|
|
1453
|
+
} | {
|
|
1454
|
+
type: "trajectory_contains_action";
|
|
1455
|
+
action: string;
|
|
1456
|
+
} | {
|
|
1457
|
+
type: "llm_judge";
|
|
1458
|
+
expected: string;
|
|
1459
|
+
prompt: string;
|
|
1460
|
+
model_type?: string | undefined;
|
|
1461
|
+
temperature?: number | undefined;
|
|
1462
|
+
json_schema?: Record<string, any> | undefined;
|
|
1463
|
+
capabilities?: string[] | undefined;
|
|
1464
|
+
} | {
|
|
1465
|
+
type: "execution_time";
|
|
1466
|
+
max_duration_ms: number;
|
|
1467
|
+
min_duration_ms?: number | undefined;
|
|
1468
|
+
target_duration_ms?: number | undefined;
|
|
1469
|
+
} | {
|
|
1470
|
+
type: "conversation_length";
|
|
1471
|
+
min_turns?: number | undefined;
|
|
1472
|
+
max_turns?: number | undefined;
|
|
1473
|
+
optimal_turns?: number | undefined;
|
|
1474
|
+
target_range?: number[] | undefined;
|
|
1475
|
+
} | {
|
|
1476
|
+
type: "conversation_flow";
|
|
1477
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
1478
|
+
flow_quality_threshold: number;
|
|
1479
|
+
} | {
|
|
1480
|
+
type: "user_satisfaction";
|
|
1481
|
+
satisfaction_threshold: number;
|
|
1482
|
+
measurement_method: "llm_judge" | "sentiment_analysis" | "keyword_analysis";
|
|
1483
|
+
indicators?: {
|
|
1484
|
+
positive?: string[] | undefined;
|
|
1485
|
+
negative?: string[] | undefined;
|
|
1486
|
+
} | undefined;
|
|
1487
|
+
} | {
|
|
1488
|
+
type: "context_retention";
|
|
1489
|
+
test_memory_of: string[];
|
|
1490
|
+
retention_turns: number;
|
|
1491
|
+
memory_accuracy_threshold: number;
|
|
1492
|
+
})[];
|
|
1493
|
+
debug_options: {
|
|
1494
|
+
log_user_simulation: boolean;
|
|
1495
|
+
log_turn_decisions: boolean;
|
|
1496
|
+
export_full_transcript: boolean;
|
|
1497
|
+
};
|
|
1498
|
+
} | undefined;
|
|
1499
|
+
}, {
|
|
1500
|
+
evaluations: ({
|
|
1501
|
+
value: string;
|
|
1502
|
+
type: "string_contains";
|
|
1503
|
+
case_sensitive?: boolean | undefined;
|
|
1504
|
+
} | {
|
|
1505
|
+
type: "regex_match";
|
|
1506
|
+
pattern: string;
|
|
1507
|
+
} | {
|
|
1508
|
+
path: string;
|
|
1509
|
+
type: "file_exists";
|
|
1510
|
+
} | {
|
|
1511
|
+
type: "trajectory_contains_action";
|
|
1512
|
+
action: string;
|
|
1513
|
+
} | {
|
|
1514
|
+
type: "llm_judge";
|
|
1515
|
+
expected: string;
|
|
1516
|
+
prompt: string;
|
|
1517
|
+
model_type?: string | undefined;
|
|
1518
|
+
temperature?: number | undefined;
|
|
1519
|
+
json_schema?: Record<string, any> | undefined;
|
|
1520
|
+
capabilities?: string[] | undefined;
|
|
1521
|
+
} | {
|
|
1522
|
+
type: "execution_time";
|
|
1523
|
+
max_duration_ms: number;
|
|
1524
|
+
min_duration_ms?: number | undefined;
|
|
1525
|
+
target_duration_ms?: number | undefined;
|
|
1526
|
+
} | {
|
|
1527
|
+
type: "conversation_length";
|
|
1528
|
+
min_turns?: number | undefined;
|
|
1529
|
+
max_turns?: number | undefined;
|
|
1530
|
+
optimal_turns?: number | undefined;
|
|
1531
|
+
target_range?: number[] | undefined;
|
|
1532
|
+
} | {
|
|
1533
|
+
type: "conversation_flow";
|
|
1534
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
1535
|
+
flow_quality_threshold?: number | undefined;
|
|
1536
|
+
} | {
|
|
1537
|
+
type: "user_satisfaction";
|
|
1538
|
+
satisfaction_threshold?: number | undefined;
|
|
1539
|
+
indicators?: {
|
|
1540
|
+
positive?: string[] | undefined;
|
|
1541
|
+
negative?: string[] | undefined;
|
|
1542
|
+
} | undefined;
|
|
1543
|
+
measurement_method?: "llm_judge" | "sentiment_analysis" | "keyword_analysis" | undefined;
|
|
1544
|
+
} | {
|
|
1545
|
+
type: "context_retention";
|
|
1546
|
+
test_memory_of: string[];
|
|
1547
|
+
retention_turns?: number | undefined;
|
|
1548
|
+
memory_accuracy_threshold?: number | undefined;
|
|
1549
|
+
})[];
|
|
1550
|
+
name?: string | undefined;
|
|
1551
|
+
code?: string | undefined;
|
|
1552
|
+
input?: string | undefined;
|
|
1553
|
+
lang?: string | undefined;
|
|
1554
|
+
conversation?: {
|
|
1555
|
+
max_turns: number;
|
|
1556
|
+
user_simulator: {
|
|
1557
|
+
persona: string;
|
|
1558
|
+
objective: string;
|
|
1559
|
+
style?: string | undefined;
|
|
1560
|
+
model_type?: string | undefined;
|
|
1561
|
+
temperature?: number | undefined;
|
|
1562
|
+
max_tokens?: number | undefined;
|
|
1563
|
+
constraints?: string[] | undefined;
|
|
1564
|
+
emotional_state?: string | undefined;
|
|
1565
|
+
knowledge_level?: "beginner" | "intermediate" | "expert" | undefined;
|
|
1566
|
+
};
|
|
1567
|
+
timeout_per_turn_ms?: number | undefined;
|
|
1568
|
+
total_timeout_ms?: number | undefined;
|
|
1569
|
+
termination_conditions?: {
|
|
1570
|
+
type: "max_turns_reached" | "user_expresses_satisfaction" | "agent_provides_solution" | "conversation_stuck" | "escalation_needed" | "goal_achieved" | "custom_condition";
|
|
1571
|
+
description?: string | undefined;
|
|
1572
|
+
llm_judge?: {
|
|
1573
|
+
prompt: string;
|
|
1574
|
+
threshold?: number | undefined;
|
|
1575
|
+
} | undefined;
|
|
1576
|
+
keywords?: string[] | undefined;
|
|
1577
|
+
}[] | undefined;
|
|
1578
|
+
turn_evaluations?: ({
|
|
1579
|
+
value: string;
|
|
1580
|
+
type: "string_contains";
|
|
1581
|
+
case_sensitive?: boolean | undefined;
|
|
1582
|
+
} | {
|
|
1583
|
+
type: "regex_match";
|
|
1584
|
+
pattern: string;
|
|
1585
|
+
} | {
|
|
1586
|
+
path: string;
|
|
1587
|
+
type: "file_exists";
|
|
1588
|
+
} | {
|
|
1589
|
+
type: "trajectory_contains_action";
|
|
1590
|
+
action: string;
|
|
1591
|
+
} | {
|
|
1592
|
+
type: "llm_judge";
|
|
1593
|
+
expected: string;
|
|
1594
|
+
prompt: string;
|
|
1595
|
+
model_type?: string | undefined;
|
|
1596
|
+
temperature?: number | undefined;
|
|
1597
|
+
json_schema?: Record<string, any> | undefined;
|
|
1598
|
+
capabilities?: string[] | undefined;
|
|
1599
|
+
} | {
|
|
1600
|
+
type: "execution_time";
|
|
1601
|
+
max_duration_ms: number;
|
|
1602
|
+
min_duration_ms?: number | undefined;
|
|
1603
|
+
target_duration_ms?: number | undefined;
|
|
1604
|
+
} | {
|
|
1605
|
+
type: "conversation_length";
|
|
1606
|
+
min_turns?: number | undefined;
|
|
1607
|
+
max_turns?: number | undefined;
|
|
1608
|
+
optimal_turns?: number | undefined;
|
|
1609
|
+
target_range?: number[] | undefined;
|
|
1610
|
+
} | {
|
|
1611
|
+
type: "conversation_flow";
|
|
1612
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
1613
|
+
flow_quality_threshold?: number | undefined;
|
|
1614
|
+
} | {
|
|
1615
|
+
type: "user_satisfaction";
|
|
1616
|
+
satisfaction_threshold?: number | undefined;
|
|
1617
|
+
indicators?: {
|
|
1618
|
+
positive?: string[] | undefined;
|
|
1619
|
+
negative?: string[] | undefined;
|
|
1620
|
+
} | undefined;
|
|
1621
|
+
measurement_method?: "llm_judge" | "sentiment_analysis" | "keyword_analysis" | undefined;
|
|
1622
|
+
} | {
|
|
1623
|
+
type: "context_retention";
|
|
1624
|
+
test_memory_of: string[];
|
|
1625
|
+
retention_turns?: number | undefined;
|
|
1626
|
+
memory_accuracy_threshold?: number | undefined;
|
|
1627
|
+
})[] | undefined;
|
|
1628
|
+
final_evaluations?: ({
|
|
1629
|
+
value: string;
|
|
1630
|
+
type: "string_contains";
|
|
1631
|
+
case_sensitive?: boolean | undefined;
|
|
1632
|
+
} | {
|
|
1633
|
+
type: "regex_match";
|
|
1634
|
+
pattern: string;
|
|
1635
|
+
} | {
|
|
1636
|
+
path: string;
|
|
1637
|
+
type: "file_exists";
|
|
1638
|
+
} | {
|
|
1639
|
+
type: "trajectory_contains_action";
|
|
1640
|
+
action: string;
|
|
1641
|
+
} | {
|
|
1642
|
+
type: "llm_judge";
|
|
1643
|
+
expected: string;
|
|
1644
|
+
prompt: string;
|
|
1645
|
+
model_type?: string | undefined;
|
|
1646
|
+
temperature?: number | undefined;
|
|
1647
|
+
json_schema?: Record<string, any> | undefined;
|
|
1648
|
+
capabilities?: string[] | undefined;
|
|
1649
|
+
} | {
|
|
1650
|
+
type: "execution_time";
|
|
1651
|
+
max_duration_ms: number;
|
|
1652
|
+
min_duration_ms?: number | undefined;
|
|
1653
|
+
target_duration_ms?: number | undefined;
|
|
1654
|
+
} | {
|
|
1655
|
+
type: "conversation_length";
|
|
1656
|
+
min_turns?: number | undefined;
|
|
1657
|
+
max_turns?: number | undefined;
|
|
1658
|
+
optimal_turns?: number | undefined;
|
|
1659
|
+
target_range?: number[] | undefined;
|
|
1660
|
+
} | {
|
|
1661
|
+
type: "conversation_flow";
|
|
1662
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
1663
|
+
flow_quality_threshold?: number | undefined;
|
|
1664
|
+
} | {
|
|
1665
|
+
type: "user_satisfaction";
|
|
1666
|
+
satisfaction_threshold?: number | undefined;
|
|
1667
|
+
indicators?: {
|
|
1668
|
+
positive?: string[] | undefined;
|
|
1669
|
+
negative?: string[] | undefined;
|
|
1670
|
+
} | undefined;
|
|
1671
|
+
measurement_method?: "llm_judge" | "sentiment_analysis" | "keyword_analysis" | undefined;
|
|
1672
|
+
} | {
|
|
1673
|
+
type: "context_retention";
|
|
1674
|
+
test_memory_of: string[];
|
|
1675
|
+
retention_turns?: number | undefined;
|
|
1676
|
+
memory_accuracy_threshold?: number | undefined;
|
|
1677
|
+
})[] | undefined;
|
|
1678
|
+
debug_options?: {
|
|
1679
|
+
log_user_simulation?: boolean | undefined;
|
|
1680
|
+
log_turn_decisions?: boolean | undefined;
|
|
1681
|
+
export_full_transcript?: boolean | undefined;
|
|
1682
|
+
} | undefined;
|
|
1683
|
+
} | undefined;
|
|
1684
|
+
}>, "many">;
|
|
376
1685
|
judgment: z.ZodObject<{
|
|
377
|
-
strategy: z.ZodEnum<
|
|
378
|
-
|
|
379
|
-
|
|
380
|
-
|
|
381
|
-
|
|
382
|
-
}
|
|
1686
|
+
strategy: z.ZodEnum<["all_pass", "any_pass"]>;
|
|
1687
|
+
}, "strip", z.ZodTypeAny, {
|
|
1688
|
+
strategy: "all_pass" | "any_pass";
|
|
1689
|
+
}, {
|
|
1690
|
+
strategy: "all_pass" | "any_pass";
|
|
1691
|
+
}>;
|
|
1692
|
+
}, "strip", z.ZodTypeAny, {
|
|
1693
|
+
name: string;
|
|
1694
|
+
run: {
|
|
1695
|
+
evaluations: ({
|
|
1696
|
+
value: string;
|
|
1697
|
+
type: "string_contains";
|
|
1698
|
+
case_sensitive?: boolean | undefined;
|
|
1699
|
+
} | {
|
|
1700
|
+
type: "regex_match";
|
|
1701
|
+
pattern: string;
|
|
1702
|
+
} | {
|
|
1703
|
+
path: string;
|
|
1704
|
+
type: "file_exists";
|
|
1705
|
+
} | {
|
|
1706
|
+
type: "trajectory_contains_action";
|
|
1707
|
+
action: string;
|
|
1708
|
+
} | {
|
|
1709
|
+
type: "llm_judge";
|
|
1710
|
+
expected: string;
|
|
1711
|
+
prompt: string;
|
|
1712
|
+
model_type?: string | undefined;
|
|
1713
|
+
temperature?: number | undefined;
|
|
1714
|
+
json_schema?: Record<string, any> | undefined;
|
|
1715
|
+
capabilities?: string[] | undefined;
|
|
1716
|
+
} | {
|
|
1717
|
+
type: "execution_time";
|
|
1718
|
+
max_duration_ms: number;
|
|
1719
|
+
min_duration_ms?: number | undefined;
|
|
1720
|
+
target_duration_ms?: number | undefined;
|
|
1721
|
+
} | {
|
|
1722
|
+
type: "conversation_length";
|
|
1723
|
+
min_turns?: number | undefined;
|
|
1724
|
+
max_turns?: number | undefined;
|
|
1725
|
+
optimal_turns?: number | undefined;
|
|
1726
|
+
target_range?: number[] | undefined;
|
|
1727
|
+
} | {
|
|
1728
|
+
type: "conversation_flow";
|
|
1729
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
1730
|
+
flow_quality_threshold: number;
|
|
1731
|
+
} | {
|
|
1732
|
+
type: "user_satisfaction";
|
|
1733
|
+
satisfaction_threshold: number;
|
|
1734
|
+
measurement_method: "llm_judge" | "sentiment_analysis" | "keyword_analysis";
|
|
1735
|
+
indicators?: {
|
|
1736
|
+
positive?: string[] | undefined;
|
|
1737
|
+
negative?: string[] | undefined;
|
|
1738
|
+
} | undefined;
|
|
1739
|
+
} | {
|
|
1740
|
+
type: "context_retention";
|
|
1741
|
+
test_memory_of: string[];
|
|
1742
|
+
retention_turns: number;
|
|
1743
|
+
memory_accuracy_threshold: number;
|
|
1744
|
+
})[];
|
|
1745
|
+
name?: string | undefined;
|
|
1746
|
+
code?: string | undefined;
|
|
1747
|
+
input?: string | undefined;
|
|
1748
|
+
lang?: string | undefined;
|
|
1749
|
+
conversation?: {
|
|
1750
|
+
max_turns: number;
|
|
1751
|
+
timeout_per_turn_ms: number;
|
|
1752
|
+
total_timeout_ms: number;
|
|
1753
|
+
user_simulator: {
|
|
1754
|
+
model_type: string;
|
|
1755
|
+
temperature: number;
|
|
1756
|
+
max_tokens: number;
|
|
1757
|
+
persona: string;
|
|
1758
|
+
objective: string;
|
|
1759
|
+
constraints: string[];
|
|
1760
|
+
knowledge_level: "beginner" | "intermediate" | "expert";
|
|
1761
|
+
style?: string | undefined;
|
|
1762
|
+
emotional_state?: string | undefined;
|
|
1763
|
+
};
|
|
1764
|
+
termination_conditions: {
|
|
1765
|
+
type: "max_turns_reached" | "user_expresses_satisfaction" | "agent_provides_solution" | "conversation_stuck" | "escalation_needed" | "goal_achieved" | "custom_condition";
|
|
1766
|
+
description?: string | undefined;
|
|
1767
|
+
llm_judge?: {
|
|
1768
|
+
prompt: string;
|
|
1769
|
+
threshold: number;
|
|
1770
|
+
} | undefined;
|
|
1771
|
+
keywords?: string[] | undefined;
|
|
1772
|
+
}[];
|
|
1773
|
+
turn_evaluations: ({
|
|
1774
|
+
value: string;
|
|
1775
|
+
type: "string_contains";
|
|
1776
|
+
case_sensitive?: boolean | undefined;
|
|
1777
|
+
} | {
|
|
1778
|
+
type: "regex_match";
|
|
1779
|
+
pattern: string;
|
|
1780
|
+
} | {
|
|
1781
|
+
path: string;
|
|
1782
|
+
type: "file_exists";
|
|
1783
|
+
} | {
|
|
1784
|
+
type: "trajectory_contains_action";
|
|
1785
|
+
action: string;
|
|
1786
|
+
} | {
|
|
1787
|
+
type: "llm_judge";
|
|
1788
|
+
expected: string;
|
|
1789
|
+
prompt: string;
|
|
1790
|
+
model_type?: string | undefined;
|
|
1791
|
+
temperature?: number | undefined;
|
|
1792
|
+
json_schema?: Record<string, any> | undefined;
|
|
1793
|
+
capabilities?: string[] | undefined;
|
|
1794
|
+
} | {
|
|
1795
|
+
type: "execution_time";
|
|
1796
|
+
max_duration_ms: number;
|
|
1797
|
+
min_duration_ms?: number | undefined;
|
|
1798
|
+
target_duration_ms?: number | undefined;
|
|
1799
|
+
} | {
|
|
1800
|
+
type: "conversation_length";
|
|
1801
|
+
min_turns?: number | undefined;
|
|
1802
|
+
max_turns?: number | undefined;
|
|
1803
|
+
optimal_turns?: number | undefined;
|
|
1804
|
+
target_range?: number[] | undefined;
|
|
1805
|
+
} | {
|
|
1806
|
+
type: "conversation_flow";
|
|
1807
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
1808
|
+
flow_quality_threshold: number;
|
|
1809
|
+
} | {
|
|
1810
|
+
type: "user_satisfaction";
|
|
1811
|
+
satisfaction_threshold: number;
|
|
1812
|
+
measurement_method: "llm_judge" | "sentiment_analysis" | "keyword_analysis";
|
|
1813
|
+
indicators?: {
|
|
1814
|
+
positive?: string[] | undefined;
|
|
1815
|
+
negative?: string[] | undefined;
|
|
1816
|
+
} | undefined;
|
|
1817
|
+
} | {
|
|
1818
|
+
type: "context_retention";
|
|
1819
|
+
test_memory_of: string[];
|
|
1820
|
+
retention_turns: number;
|
|
1821
|
+
memory_accuracy_threshold: number;
|
|
1822
|
+
})[];
|
|
1823
|
+
final_evaluations: ({
|
|
1824
|
+
value: string;
|
|
1825
|
+
type: "string_contains";
|
|
1826
|
+
case_sensitive?: boolean | undefined;
|
|
1827
|
+
} | {
|
|
1828
|
+
type: "regex_match";
|
|
1829
|
+
pattern: string;
|
|
1830
|
+
} | {
|
|
1831
|
+
path: string;
|
|
1832
|
+
type: "file_exists";
|
|
1833
|
+
} | {
|
|
1834
|
+
type: "trajectory_contains_action";
|
|
1835
|
+
action: string;
|
|
1836
|
+
} | {
|
|
1837
|
+
type: "llm_judge";
|
|
1838
|
+
expected: string;
|
|
1839
|
+
prompt: string;
|
|
1840
|
+
model_type?: string | undefined;
|
|
1841
|
+
temperature?: number | undefined;
|
|
1842
|
+
json_schema?: Record<string, any> | undefined;
|
|
1843
|
+
capabilities?: string[] | undefined;
|
|
1844
|
+
} | {
|
|
1845
|
+
type: "execution_time";
|
|
1846
|
+
max_duration_ms: number;
|
|
1847
|
+
min_duration_ms?: number | undefined;
|
|
1848
|
+
target_duration_ms?: number | undefined;
|
|
1849
|
+
} | {
|
|
1850
|
+
type: "conversation_length";
|
|
1851
|
+
min_turns?: number | undefined;
|
|
1852
|
+
max_turns?: number | undefined;
|
|
1853
|
+
optimal_turns?: number | undefined;
|
|
1854
|
+
target_range?: number[] | undefined;
|
|
1855
|
+
} | {
|
|
1856
|
+
type: "conversation_flow";
|
|
1857
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
1858
|
+
flow_quality_threshold: number;
|
|
1859
|
+
} | {
|
|
1860
|
+
type: "user_satisfaction";
|
|
1861
|
+
satisfaction_threshold: number;
|
|
1862
|
+
measurement_method: "llm_judge" | "sentiment_analysis" | "keyword_analysis";
|
|
1863
|
+
indicators?: {
|
|
1864
|
+
positive?: string[] | undefined;
|
|
1865
|
+
negative?: string[] | undefined;
|
|
1866
|
+
} | undefined;
|
|
1867
|
+
} | {
|
|
1868
|
+
type: "context_retention";
|
|
1869
|
+
test_memory_of: string[];
|
|
1870
|
+
retention_turns: number;
|
|
1871
|
+
memory_accuracy_threshold: number;
|
|
1872
|
+
})[];
|
|
1873
|
+
debug_options: {
|
|
1874
|
+
log_user_simulation: boolean;
|
|
1875
|
+
log_turn_decisions: boolean;
|
|
1876
|
+
export_full_transcript: boolean;
|
|
1877
|
+
};
|
|
1878
|
+
} | undefined;
|
|
1879
|
+
}[];
|
|
1880
|
+
description: string;
|
|
1881
|
+
environment: {
|
|
1882
|
+
type: "local" | "e2b";
|
|
1883
|
+
};
|
|
1884
|
+
judgment: {
|
|
1885
|
+
strategy: "all_pass" | "any_pass";
|
|
1886
|
+
};
|
|
1887
|
+
plugins?: (string | {
|
|
1888
|
+
name: string;
|
|
1889
|
+
enabled: boolean;
|
|
1890
|
+
version?: string | undefined;
|
|
1891
|
+
config?: Record<string, any> | undefined;
|
|
1892
|
+
})[] | undefined;
|
|
1893
|
+
setup?: {
|
|
1894
|
+
mocks?: {
|
|
1895
|
+
method: string;
|
|
1896
|
+
error?: {
|
|
1897
|
+
message: string;
|
|
1898
|
+
code: string;
|
|
1899
|
+
status?: number | undefined;
|
|
1900
|
+
} | undefined;
|
|
1901
|
+
metadata?: {
|
|
1902
|
+
delay?: number | undefined;
|
|
1903
|
+
probability?: number | undefined;
|
|
1904
|
+
} | undefined;
|
|
1905
|
+
service?: string | undefined;
|
|
1906
|
+
when?: {
|
|
1907
|
+
context?: Record<string, any> | undefined;
|
|
1908
|
+
args?: any[] | undefined;
|
|
1909
|
+
input?: Record<string, any> | undefined;
|
|
1910
|
+
matcher?: string | undefined;
|
|
1911
|
+
partialArgs?: any[] | undefined;
|
|
1912
|
+
} | undefined;
|
|
1913
|
+
response?: any;
|
|
1914
|
+
responseFn?: string | undefined;
|
|
1915
|
+
}[] | undefined;
|
|
1916
|
+
virtual_fs?: Record<string, string> | undefined;
|
|
1917
|
+
} | undefined;
|
|
1918
|
+
}, {
|
|
1919
|
+
name: string;
|
|
1920
|
+
run: {
|
|
1921
|
+
evaluations: ({
|
|
1922
|
+
value: string;
|
|
1923
|
+
type: "string_contains";
|
|
1924
|
+
case_sensitive?: boolean | undefined;
|
|
1925
|
+
} | {
|
|
1926
|
+
type: "regex_match";
|
|
1927
|
+
pattern: string;
|
|
1928
|
+
} | {
|
|
1929
|
+
path: string;
|
|
1930
|
+
type: "file_exists";
|
|
1931
|
+
} | {
|
|
1932
|
+
type: "trajectory_contains_action";
|
|
1933
|
+
action: string;
|
|
1934
|
+
} | {
|
|
1935
|
+
type: "llm_judge";
|
|
1936
|
+
expected: string;
|
|
1937
|
+
prompt: string;
|
|
1938
|
+
model_type?: string | undefined;
|
|
1939
|
+
temperature?: number | undefined;
|
|
1940
|
+
json_schema?: Record<string, any> | undefined;
|
|
1941
|
+
capabilities?: string[] | undefined;
|
|
1942
|
+
} | {
|
|
1943
|
+
type: "execution_time";
|
|
1944
|
+
max_duration_ms: number;
|
|
1945
|
+
min_duration_ms?: number | undefined;
|
|
1946
|
+
target_duration_ms?: number | undefined;
|
|
1947
|
+
} | {
|
|
1948
|
+
type: "conversation_length";
|
|
1949
|
+
min_turns?: number | undefined;
|
|
1950
|
+
max_turns?: number | undefined;
|
|
1951
|
+
optimal_turns?: number | undefined;
|
|
1952
|
+
target_range?: number[] | undefined;
|
|
1953
|
+
} | {
|
|
1954
|
+
type: "conversation_flow";
|
|
1955
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
1956
|
+
flow_quality_threshold?: number | undefined;
|
|
1957
|
+
} | {
|
|
1958
|
+
type: "user_satisfaction";
|
|
1959
|
+
satisfaction_threshold?: number | undefined;
|
|
1960
|
+
indicators?: {
|
|
1961
|
+
positive?: string[] | undefined;
|
|
1962
|
+
negative?: string[] | undefined;
|
|
1963
|
+
} | undefined;
|
|
1964
|
+
measurement_method?: "llm_judge" | "sentiment_analysis" | "keyword_analysis" | undefined;
|
|
1965
|
+
} | {
|
|
1966
|
+
type: "context_retention";
|
|
1967
|
+
test_memory_of: string[];
|
|
1968
|
+
retention_turns?: number | undefined;
|
|
1969
|
+
memory_accuracy_threshold?: number | undefined;
|
|
1970
|
+
})[];
|
|
1971
|
+
name?: string | undefined;
|
|
1972
|
+
code?: string | undefined;
|
|
1973
|
+
input?: string | undefined;
|
|
1974
|
+
lang?: string | undefined;
|
|
1975
|
+
conversation?: {
|
|
1976
|
+
max_turns: number;
|
|
1977
|
+
user_simulator: {
|
|
1978
|
+
persona: string;
|
|
1979
|
+
objective: string;
|
|
1980
|
+
style?: string | undefined;
|
|
1981
|
+
model_type?: string | undefined;
|
|
1982
|
+
temperature?: number | undefined;
|
|
1983
|
+
max_tokens?: number | undefined;
|
|
1984
|
+
constraints?: string[] | undefined;
|
|
1985
|
+
emotional_state?: string | undefined;
|
|
1986
|
+
knowledge_level?: "beginner" | "intermediate" | "expert" | undefined;
|
|
1987
|
+
};
|
|
1988
|
+
timeout_per_turn_ms?: number | undefined;
|
|
1989
|
+
total_timeout_ms?: number | undefined;
|
|
1990
|
+
termination_conditions?: {
|
|
1991
|
+
type: "max_turns_reached" | "user_expresses_satisfaction" | "agent_provides_solution" | "conversation_stuck" | "escalation_needed" | "goal_achieved" | "custom_condition";
|
|
1992
|
+
description?: string | undefined;
|
|
1993
|
+
llm_judge?: {
|
|
1994
|
+
prompt: string;
|
|
1995
|
+
threshold?: number | undefined;
|
|
1996
|
+
} | undefined;
|
|
1997
|
+
keywords?: string[] | undefined;
|
|
1998
|
+
}[] | undefined;
|
|
1999
|
+
turn_evaluations?: ({
|
|
2000
|
+
value: string;
|
|
2001
|
+
type: "string_contains";
|
|
2002
|
+
case_sensitive?: boolean | undefined;
|
|
2003
|
+
} | {
|
|
2004
|
+
type: "regex_match";
|
|
2005
|
+
pattern: string;
|
|
2006
|
+
} | {
|
|
2007
|
+
path: string;
|
|
2008
|
+
type: "file_exists";
|
|
2009
|
+
} | {
|
|
2010
|
+
type: "trajectory_contains_action";
|
|
2011
|
+
action: string;
|
|
2012
|
+
} | {
|
|
2013
|
+
type: "llm_judge";
|
|
2014
|
+
expected: string;
|
|
2015
|
+
prompt: string;
|
|
2016
|
+
model_type?: string | undefined;
|
|
2017
|
+
temperature?: number | undefined;
|
|
2018
|
+
json_schema?: Record<string, any> | undefined;
|
|
2019
|
+
capabilities?: string[] | undefined;
|
|
2020
|
+
} | {
|
|
2021
|
+
type: "execution_time";
|
|
2022
|
+
max_duration_ms: number;
|
|
2023
|
+
min_duration_ms?: number | undefined;
|
|
2024
|
+
target_duration_ms?: number | undefined;
|
|
2025
|
+
} | {
|
|
2026
|
+
type: "conversation_length";
|
|
2027
|
+
min_turns?: number | undefined;
|
|
2028
|
+
max_turns?: number | undefined;
|
|
2029
|
+
optimal_turns?: number | undefined;
|
|
2030
|
+
target_range?: number[] | undefined;
|
|
2031
|
+
} | {
|
|
2032
|
+
type: "conversation_flow";
|
|
2033
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
2034
|
+
flow_quality_threshold?: number | undefined;
|
|
2035
|
+
} | {
|
|
2036
|
+
type: "user_satisfaction";
|
|
2037
|
+
satisfaction_threshold?: number | undefined;
|
|
2038
|
+
indicators?: {
|
|
2039
|
+
positive?: string[] | undefined;
|
|
2040
|
+
negative?: string[] | undefined;
|
|
2041
|
+
} | undefined;
|
|
2042
|
+
measurement_method?: "llm_judge" | "sentiment_analysis" | "keyword_analysis" | undefined;
|
|
2043
|
+
} | {
|
|
2044
|
+
type: "context_retention";
|
|
2045
|
+
test_memory_of: string[];
|
|
2046
|
+
retention_turns?: number | undefined;
|
|
2047
|
+
memory_accuracy_threshold?: number | undefined;
|
|
2048
|
+
})[] | undefined;
|
|
2049
|
+
final_evaluations?: ({
|
|
2050
|
+
value: string;
|
|
2051
|
+
type: "string_contains";
|
|
2052
|
+
case_sensitive?: boolean | undefined;
|
|
2053
|
+
} | {
|
|
2054
|
+
type: "regex_match";
|
|
2055
|
+
pattern: string;
|
|
2056
|
+
} | {
|
|
2057
|
+
path: string;
|
|
2058
|
+
type: "file_exists";
|
|
2059
|
+
} | {
|
|
2060
|
+
type: "trajectory_contains_action";
|
|
2061
|
+
action: string;
|
|
2062
|
+
} | {
|
|
2063
|
+
type: "llm_judge";
|
|
2064
|
+
expected: string;
|
|
2065
|
+
prompt: string;
|
|
2066
|
+
model_type?: string | undefined;
|
|
2067
|
+
temperature?: number | undefined;
|
|
2068
|
+
json_schema?: Record<string, any> | undefined;
|
|
2069
|
+
capabilities?: string[] | undefined;
|
|
2070
|
+
} | {
|
|
2071
|
+
type: "execution_time";
|
|
2072
|
+
max_duration_ms: number;
|
|
2073
|
+
min_duration_ms?: number | undefined;
|
|
2074
|
+
target_duration_ms?: number | undefined;
|
|
2075
|
+
} | {
|
|
2076
|
+
type: "conversation_length";
|
|
2077
|
+
min_turns?: number | undefined;
|
|
2078
|
+
max_turns?: number | undefined;
|
|
2079
|
+
optimal_turns?: number | undefined;
|
|
2080
|
+
target_range?: number[] | undefined;
|
|
2081
|
+
} | {
|
|
2082
|
+
type: "conversation_flow";
|
|
2083
|
+
required_patterns: ("question_then_answer" | "problem_then_solution" | "clarification_cycle" | "empathy_then_solution" | "escalation_pattern")[];
|
|
2084
|
+
flow_quality_threshold?: number | undefined;
|
|
2085
|
+
} | {
|
|
2086
|
+
type: "user_satisfaction";
|
|
2087
|
+
satisfaction_threshold?: number | undefined;
|
|
2088
|
+
indicators?: {
|
|
2089
|
+
positive?: string[] | undefined;
|
|
2090
|
+
negative?: string[] | undefined;
|
|
2091
|
+
} | undefined;
|
|
2092
|
+
measurement_method?: "llm_judge" | "sentiment_analysis" | "keyword_analysis" | undefined;
|
|
2093
|
+
} | {
|
|
2094
|
+
type: "context_retention";
|
|
2095
|
+
test_memory_of: string[];
|
|
2096
|
+
retention_turns?: number | undefined;
|
|
2097
|
+
memory_accuracy_threshold?: number | undefined;
|
|
2098
|
+
})[] | undefined;
|
|
2099
|
+
debug_options?: {
|
|
2100
|
+
log_user_simulation?: boolean | undefined;
|
|
2101
|
+
log_turn_decisions?: boolean | undefined;
|
|
2102
|
+
export_full_transcript?: boolean | undefined;
|
|
2103
|
+
} | undefined;
|
|
2104
|
+
} | undefined;
|
|
2105
|
+
}[];
|
|
2106
|
+
description: string;
|
|
2107
|
+
environment: {
|
|
2108
|
+
type: "local" | "e2b";
|
|
2109
|
+
};
|
|
2110
|
+
judgment: {
|
|
2111
|
+
strategy: "all_pass" | "any_pass";
|
|
2112
|
+
};
|
|
2113
|
+
plugins?: (string | {
|
|
2114
|
+
name: string;
|
|
2115
|
+
enabled?: boolean | undefined;
|
|
2116
|
+
version?: string | undefined;
|
|
2117
|
+
config?: Record<string, any> | undefined;
|
|
2118
|
+
})[] | undefined;
|
|
2119
|
+
setup?: {
|
|
2120
|
+
mocks?: {
|
|
2121
|
+
method: string;
|
|
2122
|
+
error?: {
|
|
2123
|
+
message: string;
|
|
2124
|
+
code: string;
|
|
2125
|
+
status?: number | undefined;
|
|
2126
|
+
} | undefined;
|
|
2127
|
+
metadata?: {
|
|
2128
|
+
delay?: number | undefined;
|
|
2129
|
+
probability?: number | undefined;
|
|
2130
|
+
} | undefined;
|
|
2131
|
+
service?: string | undefined;
|
|
2132
|
+
when?: {
|
|
2133
|
+
context?: Record<string, any> | undefined;
|
|
2134
|
+
args?: any[] | undefined;
|
|
2135
|
+
input?: Record<string, any> | undefined;
|
|
2136
|
+
matcher?: string | undefined;
|
|
2137
|
+
partialArgs?: any[] | undefined;
|
|
2138
|
+
} | undefined;
|
|
2139
|
+
response?: any;
|
|
2140
|
+
responseFn?: string | undefined;
|
|
2141
|
+
}[] | undefined;
|
|
2142
|
+
virtual_fs?: Record<string, string> | undefined;
|
|
2143
|
+
} | undefined;
|
|
2144
|
+
}>;
|
|
383
2145
|
export type Scenario = z.infer<typeof ScenarioSchema>;
|
|
384
2146
|
export type Evaluation = z.infer<typeof EvaluationSchema>;
|
|
385
2147
|
export type PluginConfig = z.infer<typeof PluginConfigSchema>;
|
|
@@ -442,24 +2204,94 @@ export declare const ScenarioRunResultSchema: z.ZodObject<{
|
|
|
442
2204
|
execution_time_seconds: z.ZodNumber;
|
|
443
2205
|
llm_calls: z.ZodNumber;
|
|
444
2206
|
total_tokens: z.ZodNumber;
|
|
445
|
-
}, z.
|
|
2207
|
+
}, "strip", z.ZodNumber, z.objectOutputType<{
|
|
2208
|
+
execution_time_seconds: z.ZodNumber;
|
|
2209
|
+
llm_calls: z.ZodNumber;
|
|
2210
|
+
total_tokens: z.ZodNumber;
|
|
2211
|
+
}, z.ZodNumber, "strip">, z.objectInputType<{
|
|
2212
|
+
execution_time_seconds: z.ZodNumber;
|
|
2213
|
+
llm_calls: z.ZodNumber;
|
|
2214
|
+
total_tokens: z.ZodNumber;
|
|
2215
|
+
}, z.ZodNumber, "strip">>;
|
|
446
2216
|
final_agent_response: z.ZodOptional<z.ZodString>;
|
|
447
2217
|
evaluations: z.ZodArray<z.ZodObject<{
|
|
448
2218
|
evaluator_type: z.ZodString;
|
|
449
2219
|
success: z.ZodBoolean;
|
|
450
2220
|
summary: z.ZodString;
|
|
451
2221
|
details: z.ZodRecord<z.ZodString, z.ZodAny>;
|
|
452
|
-
}, z.
|
|
2222
|
+
}, "strip", z.ZodTypeAny, {
|
|
2223
|
+
success: boolean;
|
|
2224
|
+
evaluator_type: string;
|
|
2225
|
+
summary: string;
|
|
2226
|
+
details: Record<string, any>;
|
|
2227
|
+
}, {
|
|
2228
|
+
success: boolean;
|
|
2229
|
+
evaluator_type: string;
|
|
2230
|
+
summary: string;
|
|
2231
|
+
details: Record<string, any>;
|
|
2232
|
+
}>, "many">;
|
|
453
2233
|
trajectory: z.ZodArray<z.ZodObject<{
|
|
454
|
-
type: z.ZodEnum<
|
|
455
|
-
|
|
456
|
-
thought: "thought";
|
|
457
|
-
observation: "observation";
|
|
458
|
-
}>;
|
|
459
|
-
timestamp: z.ZodString;
|
|
2234
|
+
type: z.ZodEnum<["thought", "action", "observation"]>;
|
|
2235
|
+
timestamp: z.ZodEffects<z.ZodString, string, string>;
|
|
460
2236
|
content: z.ZodAny;
|
|
461
|
-
}, z.
|
|
2237
|
+
}, "strip", z.ZodTypeAny, {
|
|
2238
|
+
type: "action" | "thought" | "observation";
|
|
2239
|
+
timestamp: string;
|
|
2240
|
+
content?: any;
|
|
2241
|
+
}, {
|
|
2242
|
+
type: "action" | "thought" | "observation";
|
|
2243
|
+
timestamp: string;
|
|
2244
|
+
content?: any;
|
|
2245
|
+
}>, "many">;
|
|
462
2246
|
error: z.ZodNullable<z.ZodString>;
|
|
463
|
-
}, z.
|
|
2247
|
+
}, "strip", z.ZodTypeAny, {
|
|
2248
|
+
error: string | null;
|
|
2249
|
+
evaluations: {
|
|
2250
|
+
success: boolean;
|
|
2251
|
+
evaluator_type: string;
|
|
2252
|
+
summary: string;
|
|
2253
|
+
details: Record<string, any>;
|
|
2254
|
+
}[];
|
|
2255
|
+
run_id: string;
|
|
2256
|
+
matrix_combination_id: string;
|
|
2257
|
+
parameters: Record<string, any>;
|
|
2258
|
+
metrics: {
|
|
2259
|
+
execution_time_seconds: number;
|
|
2260
|
+
llm_calls: number;
|
|
2261
|
+
total_tokens: number;
|
|
2262
|
+
} & {
|
|
2263
|
+
[k: string]: number;
|
|
2264
|
+
};
|
|
2265
|
+
trajectory: {
|
|
2266
|
+
type: "action" | "thought" | "observation";
|
|
2267
|
+
timestamp: string;
|
|
2268
|
+
content?: any;
|
|
2269
|
+
}[];
|
|
2270
|
+
final_agent_response?: string | undefined;
|
|
2271
|
+
}, {
|
|
2272
|
+
error: string | null;
|
|
2273
|
+
evaluations: {
|
|
2274
|
+
success: boolean;
|
|
2275
|
+
evaluator_type: string;
|
|
2276
|
+
summary: string;
|
|
2277
|
+
details: Record<string, any>;
|
|
2278
|
+
}[];
|
|
2279
|
+
run_id: string;
|
|
2280
|
+
matrix_combination_id: string;
|
|
2281
|
+
parameters: Record<string, any>;
|
|
2282
|
+
metrics: {
|
|
2283
|
+
execution_time_seconds: number;
|
|
2284
|
+
llm_calls: number;
|
|
2285
|
+
total_tokens: number;
|
|
2286
|
+
} & {
|
|
2287
|
+
[k: string]: number;
|
|
2288
|
+
};
|
|
2289
|
+
trajectory: {
|
|
2290
|
+
type: "action" | "thought" | "observation";
|
|
2291
|
+
timestamp: string;
|
|
2292
|
+
content?: any;
|
|
2293
|
+
}[];
|
|
2294
|
+
final_agent_response?: string | undefined;
|
|
2295
|
+
}>;
|
|
464
2296
|
export {};
|
|
465
2297
|
//# sourceMappingURL=schema.d.ts.map
|