@archon-claw/cli 0.0.2

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (80) hide show
  1. package/dist/agent.d.ts +2 -0
  2. package/dist/agent.js +152 -0
  3. package/dist/cli.d.ts +2 -0
  4. package/dist/cli.js +141 -0
  5. package/dist/config.d.ts +2 -0
  6. package/dist/config.js +161 -0
  7. package/dist/eval/assertions.d.ts +9 -0
  8. package/dist/eval/assertions.js +137 -0
  9. package/dist/eval/execute.d.ts +13 -0
  10. package/dist/eval/execute.js +260 -0
  11. package/dist/eval/formatter.d.ts +10 -0
  12. package/dist/eval/formatter.js +62 -0
  13. package/dist/eval/judge.d.ts +7 -0
  14. package/dist/eval/judge.js +116 -0
  15. package/dist/eval/runner.d.ts +9 -0
  16. package/dist/eval/runner.js +156 -0
  17. package/dist/eval/types.d.ts +67 -0
  18. package/dist/eval/types.js +1 -0
  19. package/dist/llm.d.ts +7 -0
  20. package/dist/llm.js +52 -0
  21. package/dist/mcp-manager.d.ts +51 -0
  22. package/dist/mcp-manager.js +268 -0
  23. package/dist/pending-tool-results.d.ts +4 -0
  24. package/dist/pending-tool-results.js +39 -0
  25. package/dist/public/assets/chat-input-BBnVJs9h.js +151 -0
  26. package/dist/public/assets/chat-input-CISJdhF2.css +1 -0
  27. package/dist/public/assets/embed-DhIUBDdf.js +1 -0
  28. package/dist/public/assets/main-Bfvj6DnV.js +16 -0
  29. package/dist/public/embed/widget.js +233 -0
  30. package/dist/public/embed.html +14 -0
  31. package/dist/public/index.html +14 -0
  32. package/dist/scaffold.d.ts +2 -0
  33. package/dist/scaffold.js +82 -0
  34. package/dist/schemas.d.ts +899 -0
  35. package/dist/schemas.js +134 -0
  36. package/dist/server.d.ts +3 -0
  37. package/dist/server.js +258 -0
  38. package/dist/session.d.ts +8 -0
  39. package/dist/session.js +70 -0
  40. package/dist/templates/agent/model.json +6 -0
  41. package/dist/templates/agent/system-prompt.md +9 -0
  42. package/dist/templates/agent/tool-impls/greeting.impl.js +9 -0
  43. package/dist/templates/agent/tools/greeting.json +14 -0
  44. package/dist/templates/workspace/.claude/skills/create-agent/SKILL.md +90 -0
  45. package/dist/templates/workspace/.claude/skills/create-dataset/SKILL.md +57 -0
  46. package/dist/templates/workspace/.claude/skills/create-eval-case/SKILL.md +159 -0
  47. package/dist/templates/workspace/.claude/skills/create-eval-judge/SKILL.md +128 -0
  48. package/dist/templates/workspace/.claude/skills/create-mcp-config/SKILL.md +151 -0
  49. package/dist/templates/workspace/.claude/skills/create-model-config/SKILL.md +45 -0
  50. package/dist/templates/workspace/.claude/skills/create-skill/SKILL.md +63 -0
  51. package/dist/templates/workspace/.claude/skills/create-system-prompt/SKILL.md +168 -0
  52. package/dist/templates/workspace/.claude/skills/create-tool/SKILL.md +56 -0
  53. package/dist/templates/workspace/.claude/skills/create-tool-impl/SKILL.md +83 -0
  54. package/dist/templates/workspace/.claude/skills/create-tool-test/SKILL.md +117 -0
  55. package/dist/templates/workspace/.claude/skills/create-tool-ui/SKILL.md +218 -0
  56. package/dist/test-runner.d.ts +22 -0
  57. package/dist/test-runner.js +166 -0
  58. package/dist/types.d.ts +75 -0
  59. package/dist/types.js +1 -0
  60. package/dist/validator/index.d.ts +16 -0
  61. package/dist/validator/index.js +54 -0
  62. package/dist/validator/plugin.d.ts +21 -0
  63. package/dist/validator/plugin.js +1 -0
  64. package/dist/validator/plugins/agent-dir.d.ts +2 -0
  65. package/dist/validator/plugins/agent-dir.js +171 -0
  66. package/dist/validator/plugins/agent-skill.d.ts +2 -0
  67. package/dist/validator/plugins/agent-skill.js +31 -0
  68. package/dist/validator/plugins/dataset.d.ts +2 -0
  69. package/dist/validator/plugins/dataset.js +20 -0
  70. package/dist/validator/plugins/mcp.d.ts +2 -0
  71. package/dist/validator/plugins/mcp.js +20 -0
  72. package/dist/validator/plugins/model.d.ts +2 -0
  73. package/dist/validator/plugins/model.js +20 -0
  74. package/dist/validator/plugins/system-prompt.d.ts +2 -0
  75. package/dist/validator/plugins/system-prompt.js +25 -0
  76. package/dist/validator/plugins/tool.d.ts +2 -0
  77. package/dist/validator/plugins/tool.js +20 -0
  78. package/dist/validator/zod-utils.d.ts +3 -0
  79. package/dist/validator/zod-utils.js +7 -0
  80. package/package.json +41 -0
@@ -0,0 +1,899 @@
1
+ import { z } from "zod";
2
+ export declare const modelConfigSchema: z.ZodObject<{
3
+ $schema: z.ZodOptional<z.ZodString>;
4
+ provider: z.ZodEnum<["anthropic", "openai", "bailian"]>;
5
+ model: z.ZodString;
6
+ maxTokens: z.ZodOptional<z.ZodNumber>;
7
+ temperature: z.ZodOptional<z.ZodNumber>;
8
+ apiKey: z.ZodOptional<z.ZodString>;
9
+ baseURL: z.ZodOptional<z.ZodString>;
10
+ }, "strict", z.ZodTypeAny, {
11
+ provider: "anthropic" | "openai" | "bailian";
12
+ model: string;
13
+ $schema?: string | undefined;
14
+ maxTokens?: number | undefined;
15
+ temperature?: number | undefined;
16
+ apiKey?: string | undefined;
17
+ baseURL?: string | undefined;
18
+ }, {
19
+ provider: "anthropic" | "openai" | "bailian";
20
+ model: string;
21
+ $schema?: string | undefined;
22
+ maxTokens?: number | undefined;
23
+ temperature?: number | undefined;
24
+ apiKey?: string | undefined;
25
+ baseURL?: string | undefined;
26
+ }>;
27
+ export type ModelConfig = z.infer<typeof modelConfigSchema>;
28
+ export declare const jsonSchemaPropertySchema: z.ZodType<JSONSchemaProperty>;
29
+ export interface JSONSchemaProperty {
30
+ type: "string" | "number" | "boolean" | "array" | "object";
31
+ description?: string;
32
+ enum?: (string | number)[];
33
+ items?: JSONSchemaProperty;
34
+ properties?: Record<string, JSONSchemaProperty>;
35
+ required?: string[];
36
+ }
37
+ export declare const jsonSchemaObjectSchema: z.ZodObject<{
38
+ type: z.ZodLiteral<"object">;
39
+ properties: z.ZodRecord<z.ZodString, z.ZodType<JSONSchemaProperty, z.ZodTypeDef, JSONSchemaProperty>>;
40
+ required: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
41
+ }, "strip", z.ZodTypeAny, {
42
+ type: "object";
43
+ properties: Record<string, JSONSchemaProperty>;
44
+ required?: string[] | undefined;
45
+ }, {
46
+ type: "object";
47
+ properties: Record<string, JSONSchemaProperty>;
48
+ required?: string[] | undefined;
49
+ }>;
50
+ export type JSONSchema = z.infer<typeof jsonSchemaObjectSchema>;
51
+ export declare const toolSchemaSchema: z.ZodEffects<z.ZodEffects<z.ZodObject<{
52
+ $schema: z.ZodOptional<z.ZodString>;
53
+ name: z.ZodString;
54
+ description: z.ZodString;
55
+ input_schema: z.ZodObject<{
56
+ type: z.ZodLiteral<"object">;
57
+ properties: z.ZodRecord<z.ZodString, z.ZodType<JSONSchemaProperty, z.ZodTypeDef, JSONSchemaProperty>>;
58
+ required: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
59
+ }, "strip", z.ZodTypeAny, {
60
+ type: "object";
61
+ properties: Record<string, JSONSchemaProperty>;
62
+ required?: string[] | undefined;
63
+ }, {
64
+ type: "object";
65
+ properties: Record<string, JSONSchemaProperty>;
66
+ required?: string[] | undefined;
67
+ }>;
68
+ execution_target: z.ZodOptional<z.ZodEnum<["server", "client", "host"]>>;
69
+ handler: z.ZodOptional<z.ZodString>;
70
+ }, "strict", z.ZodTypeAny, {
71
+ description: string;
72
+ name: string;
73
+ input_schema: {
74
+ type: "object";
75
+ properties: Record<string, JSONSchemaProperty>;
76
+ required?: string[] | undefined;
77
+ };
78
+ $schema?: string | undefined;
79
+ execution_target?: "server" | "client" | "host" | undefined;
80
+ handler?: string | undefined;
81
+ }, {
82
+ description: string;
83
+ name: string;
84
+ input_schema: {
85
+ type: "object";
86
+ properties: Record<string, JSONSchemaProperty>;
87
+ required?: string[] | undefined;
88
+ };
89
+ $schema?: string | undefined;
90
+ execution_target?: "server" | "client" | "host" | undefined;
91
+ handler?: string | undefined;
92
+ }>, {
93
+ description: string;
94
+ name: string;
95
+ input_schema: {
96
+ type: "object";
97
+ properties: Record<string, JSONSchemaProperty>;
98
+ required?: string[] | undefined;
99
+ };
100
+ $schema?: string | undefined;
101
+ execution_target?: "server" | "client" | "host" | undefined;
102
+ handler?: string | undefined;
103
+ }, {
104
+ description: string;
105
+ name: string;
106
+ input_schema: {
107
+ type: "object";
108
+ properties: Record<string, JSONSchemaProperty>;
109
+ required?: string[] | undefined;
110
+ };
111
+ $schema?: string | undefined;
112
+ execution_target?: "server" | "client" | "host" | undefined;
113
+ handler?: string | undefined;
114
+ }>, {
115
+ description: string;
116
+ name: string;
117
+ input_schema: {
118
+ type: "object";
119
+ properties: Record<string, JSONSchemaProperty>;
120
+ required?: string[] | undefined;
121
+ };
122
+ $schema?: string | undefined;
123
+ execution_target?: "server" | "client" | "host" | undefined;
124
+ handler?: string | undefined;
125
+ }, {
126
+ description: string;
127
+ name: string;
128
+ input_schema: {
129
+ type: "object";
130
+ properties: Record<string, JSONSchemaProperty>;
131
+ required?: string[] | undefined;
132
+ };
133
+ $schema?: string | undefined;
134
+ execution_target?: "server" | "client" | "host" | undefined;
135
+ handler?: string | undefined;
136
+ }>;
137
+ export type ToolSchema = z.infer<typeof toolSchemaSchema>;
138
+ export declare const datasetSchema: z.ZodArray<z.ZodUnknown, "many">;
139
+ export declare const agentSkillFrontmatterSchema: z.ZodObject<{
140
+ name: z.ZodString;
141
+ description: z.ZodString;
142
+ }, "strict", z.ZodTypeAny, {
143
+ description: string;
144
+ name: string;
145
+ }, {
146
+ description: string;
147
+ name: string;
148
+ }>;
149
+ export declare const assertionTypeSchema: z.ZodEnum<["contains", "not-contains", "regex", "length-min", "length-max", "json-valid", "tool-called", "tool-not-called", "tool-called-with"]>;
150
+ export type AssertionType = z.infer<typeof assertionTypeSchema>;
151
+ export declare const assertionSchema: z.ZodObject<{
152
+ type: z.ZodEnum<["contains", "not-contains", "regex", "length-min", "length-max", "json-valid", "tool-called", "tool-not-called", "tool-called-with"]>;
153
+ value: z.ZodString;
154
+ }, "strip", z.ZodTypeAny, {
155
+ value: string;
156
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
157
+ }, {
158
+ value: string;
159
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
160
+ }>;
161
+ export type Assertion = z.infer<typeof assertionSchema>;
162
+ export declare const evalTurnToolCallSchema: z.ZodObject<{
163
+ name: z.ZodString;
164
+ args: z.ZodRecord<z.ZodString, z.ZodUnknown>;
165
+ result: z.ZodUnknown;
166
+ }, "strip", z.ZodTypeAny, {
167
+ name: string;
168
+ args: Record<string, unknown>;
169
+ result?: unknown;
170
+ }, {
171
+ name: string;
172
+ args: Record<string, unknown>;
173
+ result?: unknown;
174
+ }>;
175
+ export type EvalTurnToolCall = z.infer<typeof evalTurnToolCallSchema>;
176
+ export declare const evalTurnSchema: z.ZodObject<{
177
+ role: z.ZodEnum<["user", "assistant"]>;
178
+ content: z.ZodString;
179
+ assertions: z.ZodOptional<z.ZodArray<z.ZodObject<{
180
+ type: z.ZodEnum<["contains", "not-contains", "regex", "length-min", "length-max", "json-valid", "tool-called", "tool-not-called", "tool-called-with"]>;
181
+ value: z.ZodString;
182
+ }, "strip", z.ZodTypeAny, {
183
+ value: string;
184
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
185
+ }, {
186
+ value: string;
187
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
188
+ }>, "many">>;
189
+ judge: z.ZodOptional<z.ZodBoolean>;
190
+ expectedOutput: z.ZodOptional<z.ZodString>;
191
+ toolCalls: z.ZodOptional<z.ZodArray<z.ZodObject<{
192
+ name: z.ZodString;
193
+ args: z.ZodRecord<z.ZodString, z.ZodUnknown>;
194
+ result: z.ZodUnknown;
195
+ }, "strip", z.ZodTypeAny, {
196
+ name: string;
197
+ args: Record<string, unknown>;
198
+ result?: unknown;
199
+ }, {
200
+ name: string;
201
+ args: Record<string, unknown>;
202
+ result?: unknown;
203
+ }>, "many">>;
204
+ }, "strip", z.ZodTypeAny, {
205
+ role: "user" | "assistant";
206
+ content: string;
207
+ assertions?: {
208
+ value: string;
209
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
210
+ }[] | undefined;
211
+ judge?: boolean | undefined;
212
+ expectedOutput?: string | undefined;
213
+ toolCalls?: {
214
+ name: string;
215
+ args: Record<string, unknown>;
216
+ result?: unknown;
217
+ }[] | undefined;
218
+ }, {
219
+ role: "user" | "assistant";
220
+ content: string;
221
+ assertions?: {
222
+ value: string;
223
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
224
+ }[] | undefined;
225
+ judge?: boolean | undefined;
226
+ expectedOutput?: string | undefined;
227
+ toolCalls?: {
228
+ name: string;
229
+ args: Record<string, unknown>;
230
+ result?: unknown;
231
+ }[] | undefined;
232
+ }>;
233
+ export type EvalTurn = z.infer<typeof evalTurnSchema>;
234
+ export declare const evalModeSchema: z.ZodEnum<["single", "injected", "sequential"]>;
235
+ export type EvalMode = z.infer<typeof evalModeSchema>;
236
+ export declare const evalCaseSchema: z.ZodObject<{
237
+ name: z.ZodString;
238
+ mode: z.ZodEnum<["single", "injected", "sequential"]>;
239
+ turns: z.ZodArray<z.ZodObject<{
240
+ role: z.ZodEnum<["user", "assistant"]>;
241
+ content: z.ZodString;
242
+ assertions: z.ZodOptional<z.ZodArray<z.ZodObject<{
243
+ type: z.ZodEnum<["contains", "not-contains", "regex", "length-min", "length-max", "json-valid", "tool-called", "tool-not-called", "tool-called-with"]>;
244
+ value: z.ZodString;
245
+ }, "strip", z.ZodTypeAny, {
246
+ value: string;
247
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
248
+ }, {
249
+ value: string;
250
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
251
+ }>, "many">>;
252
+ judge: z.ZodOptional<z.ZodBoolean>;
253
+ expectedOutput: z.ZodOptional<z.ZodString>;
254
+ toolCalls: z.ZodOptional<z.ZodArray<z.ZodObject<{
255
+ name: z.ZodString;
256
+ args: z.ZodRecord<z.ZodString, z.ZodUnknown>;
257
+ result: z.ZodUnknown;
258
+ }, "strip", z.ZodTypeAny, {
259
+ name: string;
260
+ args: Record<string, unknown>;
261
+ result?: unknown;
262
+ }, {
263
+ name: string;
264
+ args: Record<string, unknown>;
265
+ result?: unknown;
266
+ }>, "many">>;
267
+ }, "strip", z.ZodTypeAny, {
268
+ role: "user" | "assistant";
269
+ content: string;
270
+ assertions?: {
271
+ value: string;
272
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
273
+ }[] | undefined;
274
+ judge?: boolean | undefined;
275
+ expectedOutput?: string | undefined;
276
+ toolCalls?: {
277
+ name: string;
278
+ args: Record<string, unknown>;
279
+ result?: unknown;
280
+ }[] | undefined;
281
+ }, {
282
+ role: "user" | "assistant";
283
+ content: string;
284
+ assertions?: {
285
+ value: string;
286
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
287
+ }[] | undefined;
288
+ judge?: boolean | undefined;
289
+ expectedOutput?: string | undefined;
290
+ toolCalls?: {
291
+ name: string;
292
+ args: Record<string, unknown>;
293
+ result?: unknown;
294
+ }[] | undefined;
295
+ }>, "many">;
296
+ assertions: z.ZodOptional<z.ZodArray<z.ZodObject<{
297
+ type: z.ZodEnum<["contains", "not-contains", "regex", "length-min", "length-max", "json-valid", "tool-called", "tool-not-called", "tool-called-with"]>;
298
+ value: z.ZodString;
299
+ }, "strip", z.ZodTypeAny, {
300
+ value: string;
301
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
302
+ }, {
303
+ value: string;
304
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
305
+ }>, "many">>;
306
+ expectedOutput: z.ZodOptional<z.ZodString>;
307
+ tags: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
308
+ tools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
309
+ judge: z.ZodOptional<z.ZodString>;
310
+ }, "strip", z.ZodTypeAny, {
311
+ name: string;
312
+ mode: "single" | "injected" | "sequential";
313
+ turns: {
314
+ role: "user" | "assistant";
315
+ content: string;
316
+ assertions?: {
317
+ value: string;
318
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
319
+ }[] | undefined;
320
+ judge?: boolean | undefined;
321
+ expectedOutput?: string | undefined;
322
+ toolCalls?: {
323
+ name: string;
324
+ args: Record<string, unknown>;
325
+ result?: unknown;
326
+ }[] | undefined;
327
+ }[];
328
+ assertions?: {
329
+ value: string;
330
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
331
+ }[] | undefined;
332
+ judge?: string | undefined;
333
+ expectedOutput?: string | undefined;
334
+ tags?: string[] | undefined;
335
+ tools?: string[] | undefined;
336
+ }, {
337
+ name: string;
338
+ mode: "single" | "injected" | "sequential";
339
+ turns: {
340
+ role: "user" | "assistant";
341
+ content: string;
342
+ assertions?: {
343
+ value: string;
344
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
345
+ }[] | undefined;
346
+ judge?: boolean | undefined;
347
+ expectedOutput?: string | undefined;
348
+ toolCalls?: {
349
+ name: string;
350
+ args: Record<string, unknown>;
351
+ result?: unknown;
352
+ }[] | undefined;
353
+ }[];
354
+ assertions?: {
355
+ value: string;
356
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
357
+ }[] | undefined;
358
+ judge?: string | undefined;
359
+ expectedOutput?: string | undefined;
360
+ tags?: string[] | undefined;
361
+ tools?: string[] | undefined;
362
+ }>;
363
+ export type EvalCase = z.infer<typeof evalCaseSchema>;
364
+ export declare const evalFileSchema: z.ZodObject<{
365
+ $schema: z.ZodOptional<z.ZodString>;
366
+ name: z.ZodString;
367
+ description: z.ZodOptional<z.ZodString>;
368
+ cases: z.ZodArray<z.ZodObject<{
369
+ name: z.ZodString;
370
+ mode: z.ZodEnum<["single", "injected", "sequential"]>;
371
+ turns: z.ZodArray<z.ZodObject<{
372
+ role: z.ZodEnum<["user", "assistant"]>;
373
+ content: z.ZodString;
374
+ assertions: z.ZodOptional<z.ZodArray<z.ZodObject<{
375
+ type: z.ZodEnum<["contains", "not-contains", "regex", "length-min", "length-max", "json-valid", "tool-called", "tool-not-called", "tool-called-with"]>;
376
+ value: z.ZodString;
377
+ }, "strip", z.ZodTypeAny, {
378
+ value: string;
379
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
380
+ }, {
381
+ value: string;
382
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
383
+ }>, "many">>;
384
+ judge: z.ZodOptional<z.ZodBoolean>;
385
+ expectedOutput: z.ZodOptional<z.ZodString>;
386
+ toolCalls: z.ZodOptional<z.ZodArray<z.ZodObject<{
387
+ name: z.ZodString;
388
+ args: z.ZodRecord<z.ZodString, z.ZodUnknown>;
389
+ result: z.ZodUnknown;
390
+ }, "strip", z.ZodTypeAny, {
391
+ name: string;
392
+ args: Record<string, unknown>;
393
+ result?: unknown;
394
+ }, {
395
+ name: string;
396
+ args: Record<string, unknown>;
397
+ result?: unknown;
398
+ }>, "many">>;
399
+ }, "strip", z.ZodTypeAny, {
400
+ role: "user" | "assistant";
401
+ content: string;
402
+ assertions?: {
403
+ value: string;
404
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
405
+ }[] | undefined;
406
+ judge?: boolean | undefined;
407
+ expectedOutput?: string | undefined;
408
+ toolCalls?: {
409
+ name: string;
410
+ args: Record<string, unknown>;
411
+ result?: unknown;
412
+ }[] | undefined;
413
+ }, {
414
+ role: "user" | "assistant";
415
+ content: string;
416
+ assertions?: {
417
+ value: string;
418
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
419
+ }[] | undefined;
420
+ judge?: boolean | undefined;
421
+ expectedOutput?: string | undefined;
422
+ toolCalls?: {
423
+ name: string;
424
+ args: Record<string, unknown>;
425
+ result?: unknown;
426
+ }[] | undefined;
427
+ }>, "many">;
428
+ assertions: z.ZodOptional<z.ZodArray<z.ZodObject<{
429
+ type: z.ZodEnum<["contains", "not-contains", "regex", "length-min", "length-max", "json-valid", "tool-called", "tool-not-called", "tool-called-with"]>;
430
+ value: z.ZodString;
431
+ }, "strip", z.ZodTypeAny, {
432
+ value: string;
433
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
434
+ }, {
435
+ value: string;
436
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
437
+ }>, "many">>;
438
+ expectedOutput: z.ZodOptional<z.ZodString>;
439
+ tags: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
440
+ tools: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
441
+ judge: z.ZodOptional<z.ZodString>;
442
+ }, "strip", z.ZodTypeAny, {
443
+ name: string;
444
+ mode: "single" | "injected" | "sequential";
445
+ turns: {
446
+ role: "user" | "assistant";
447
+ content: string;
448
+ assertions?: {
449
+ value: string;
450
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
451
+ }[] | undefined;
452
+ judge?: boolean | undefined;
453
+ expectedOutput?: string | undefined;
454
+ toolCalls?: {
455
+ name: string;
456
+ args: Record<string, unknown>;
457
+ result?: unknown;
458
+ }[] | undefined;
459
+ }[];
460
+ assertions?: {
461
+ value: string;
462
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
463
+ }[] | undefined;
464
+ judge?: string | undefined;
465
+ expectedOutput?: string | undefined;
466
+ tags?: string[] | undefined;
467
+ tools?: string[] | undefined;
468
+ }, {
469
+ name: string;
470
+ mode: "single" | "injected" | "sequential";
471
+ turns: {
472
+ role: "user" | "assistant";
473
+ content: string;
474
+ assertions?: {
475
+ value: string;
476
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
477
+ }[] | undefined;
478
+ judge?: boolean | undefined;
479
+ expectedOutput?: string | undefined;
480
+ toolCalls?: {
481
+ name: string;
482
+ args: Record<string, unknown>;
483
+ result?: unknown;
484
+ }[] | undefined;
485
+ }[];
486
+ assertions?: {
487
+ value: string;
488
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
489
+ }[] | undefined;
490
+ judge?: string | undefined;
491
+ expectedOutput?: string | undefined;
492
+ tags?: string[] | undefined;
493
+ tools?: string[] | undefined;
494
+ }>, "many">;
495
+ }, "strip", z.ZodTypeAny, {
496
+ name: string;
497
+ cases: {
498
+ name: string;
499
+ mode: "single" | "injected" | "sequential";
500
+ turns: {
501
+ role: "user" | "assistant";
502
+ content: string;
503
+ assertions?: {
504
+ value: string;
505
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
506
+ }[] | undefined;
507
+ judge?: boolean | undefined;
508
+ expectedOutput?: string | undefined;
509
+ toolCalls?: {
510
+ name: string;
511
+ args: Record<string, unknown>;
512
+ result?: unknown;
513
+ }[] | undefined;
514
+ }[];
515
+ assertions?: {
516
+ value: string;
517
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
518
+ }[] | undefined;
519
+ judge?: string | undefined;
520
+ expectedOutput?: string | undefined;
521
+ tags?: string[] | undefined;
522
+ tools?: string[] | undefined;
523
+ }[];
524
+ $schema?: string | undefined;
525
+ description?: string | undefined;
526
+ }, {
527
+ name: string;
528
+ cases: {
529
+ name: string;
530
+ mode: "single" | "injected" | "sequential";
531
+ turns: {
532
+ role: "user" | "assistant";
533
+ content: string;
534
+ assertions?: {
535
+ value: string;
536
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
537
+ }[] | undefined;
538
+ judge?: boolean | undefined;
539
+ expectedOutput?: string | undefined;
540
+ toolCalls?: {
541
+ name: string;
542
+ args: Record<string, unknown>;
543
+ result?: unknown;
544
+ }[] | undefined;
545
+ }[];
546
+ assertions?: {
547
+ value: string;
548
+ type: "contains" | "not-contains" | "regex" | "length-min" | "length-max" | "json-valid" | "tool-called" | "tool-not-called" | "tool-called-with";
549
+ }[] | undefined;
550
+ judge?: string | undefined;
551
+ expectedOutput?: string | undefined;
552
+ tags?: string[] | undefined;
553
+ tools?: string[] | undefined;
554
+ }[];
555
+ $schema?: string | undefined;
556
+ description?: string | undefined;
557
+ }>;
558
+ export type EvalFile = z.infer<typeof evalFileSchema>;
559
+ export declare const judgeDimensionSchema: z.ZodObject<{
560
+ key: z.ZodString;
561
+ label: z.ZodString;
562
+ weight: z.ZodNumber;
563
+ type: z.ZodOptional<z.ZodEnum<["numeric", "binary"]>>;
564
+ min: z.ZodOptional<z.ZodNumber>;
565
+ max: z.ZodOptional<z.ZodNumber>;
566
+ }, "strip", z.ZodTypeAny, {
567
+ key: string;
568
+ label: string;
569
+ weight: number;
570
+ type?: "numeric" | "binary" | undefined;
571
+ min?: number | undefined;
572
+ max?: number | undefined;
573
+ }, {
574
+ key: string;
575
+ label: string;
576
+ weight: number;
577
+ type?: "numeric" | "binary" | undefined;
578
+ min?: number | undefined;
579
+ max?: number | undefined;
580
+ }>;
581
+ export type JudgeDimension = z.infer<typeof judgeDimensionSchema>;
582
+ export declare const judgeConfigSchema: z.ZodObject<{
583
+ $schema: z.ZodOptional<z.ZodString>;
584
+ model: z.ZodOptional<z.ZodObject<{
585
+ $schema: z.ZodOptional<z.ZodString>;
586
+ provider: z.ZodEnum<["anthropic", "openai", "bailian"]>;
587
+ model: z.ZodString;
588
+ maxTokens: z.ZodOptional<z.ZodNumber>;
589
+ temperature: z.ZodOptional<z.ZodNumber>;
590
+ apiKey: z.ZodOptional<z.ZodString>;
591
+ baseURL: z.ZodOptional<z.ZodString>;
592
+ }, "strict", z.ZodTypeAny, {
593
+ provider: "anthropic" | "openai" | "bailian";
594
+ model: string;
595
+ $schema?: string | undefined;
596
+ maxTokens?: number | undefined;
597
+ temperature?: number | undefined;
598
+ apiKey?: string | undefined;
599
+ baseURL?: string | undefined;
600
+ }, {
601
+ provider: "anthropic" | "openai" | "bailian";
602
+ model: string;
603
+ $schema?: string | undefined;
604
+ maxTokens?: number | undefined;
605
+ temperature?: number | undefined;
606
+ apiKey?: string | undefined;
607
+ baseURL?: string | undefined;
608
+ }>>;
609
+ dimensions: z.ZodArray<z.ZodObject<{
610
+ key: z.ZodString;
611
+ label: z.ZodString;
612
+ weight: z.ZodNumber;
613
+ type: z.ZodOptional<z.ZodEnum<["numeric", "binary"]>>;
614
+ min: z.ZodOptional<z.ZodNumber>;
615
+ max: z.ZodOptional<z.ZodNumber>;
616
+ }, "strip", z.ZodTypeAny, {
617
+ key: string;
618
+ label: string;
619
+ weight: number;
620
+ type?: "numeric" | "binary" | undefined;
621
+ min?: number | undefined;
622
+ max?: number | undefined;
623
+ }, {
624
+ key: string;
625
+ label: string;
626
+ weight: number;
627
+ type?: "numeric" | "binary" | undefined;
628
+ min?: number | undefined;
629
+ max?: number | undefined;
630
+ }>, "many">;
631
+ promptTemplate: z.ZodOptional<z.ZodString>;
632
+ turnPromptTemplate: z.ZodOptional<z.ZodString>;
633
+ }, "strip", z.ZodTypeAny, {
634
+ dimensions: {
635
+ key: string;
636
+ label: string;
637
+ weight: number;
638
+ type?: "numeric" | "binary" | undefined;
639
+ min?: number | undefined;
640
+ max?: number | undefined;
641
+ }[];
642
+ $schema?: string | undefined;
643
+ model?: {
644
+ provider: "anthropic" | "openai" | "bailian";
645
+ model: string;
646
+ $schema?: string | undefined;
647
+ maxTokens?: number | undefined;
648
+ temperature?: number | undefined;
649
+ apiKey?: string | undefined;
650
+ baseURL?: string | undefined;
651
+ } | undefined;
652
+ promptTemplate?: string | undefined;
653
+ turnPromptTemplate?: string | undefined;
654
+ }, {
655
+ dimensions: {
656
+ key: string;
657
+ label: string;
658
+ weight: number;
659
+ type?: "numeric" | "binary" | undefined;
660
+ min?: number | undefined;
661
+ max?: number | undefined;
662
+ }[];
663
+ $schema?: string | undefined;
664
+ model?: {
665
+ provider: "anthropic" | "openai" | "bailian";
666
+ model: string;
667
+ $schema?: string | undefined;
668
+ maxTokens?: number | undefined;
669
+ temperature?: number | undefined;
670
+ apiKey?: string | undefined;
671
+ baseURL?: string | undefined;
672
+ } | undefined;
673
+ promptTemplate?: string | undefined;
674
+ turnPromptTemplate?: string | undefined;
675
+ }>;
676
+ export type JudgeConfig = z.infer<typeof judgeConfigSchema>;
677
+ export declare const mcpToolsFilterSchema: z.ZodObject<{
678
+ include: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
679
+ exclude: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
680
+ rename: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>;
681
+ }, "strict", z.ZodTypeAny, {
682
+ include?: string[] | undefined;
683
+ exclude?: string[] | undefined;
684
+ rename?: Record<string, string> | undefined;
685
+ }, {
686
+ include?: string[] | undefined;
687
+ exclude?: string[] | undefined;
688
+ rename?: Record<string, string> | undefined;
689
+ }>;
690
+ export declare const mcpServerConfigSchema: z.ZodEffects<z.ZodObject<{
691
+ transport: z.ZodOptional<z.ZodEnum<["stdio", "sse", "streamable-http"]>>;
692
+ command: z.ZodOptional<z.ZodString>;
693
+ args: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
694
+ env: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>;
695
+ cwd: z.ZodOptional<z.ZodString>;
696
+ url: z.ZodOptional<z.ZodString>;
697
+ headers: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>;
698
+ enabled: z.ZodOptional<z.ZodBoolean>;
699
+ timeout: z.ZodOptional<z.ZodNumber>;
700
+ tools: z.ZodOptional<z.ZodObject<{
701
+ include: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
702
+ exclude: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
703
+ rename: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>;
704
+ }, "strict", z.ZodTypeAny, {
705
+ include?: string[] | undefined;
706
+ exclude?: string[] | undefined;
707
+ rename?: Record<string, string> | undefined;
708
+ }, {
709
+ include?: string[] | undefined;
710
+ exclude?: string[] | undefined;
711
+ rename?: Record<string, string> | undefined;
712
+ }>>;
713
+ }, "strict", z.ZodTypeAny, {
714
+ args?: string[] | undefined;
715
+ tools?: {
716
+ include?: string[] | undefined;
717
+ exclude?: string[] | undefined;
718
+ rename?: Record<string, string> | undefined;
719
+ } | undefined;
720
+ transport?: "stdio" | "sse" | "streamable-http" | undefined;
721
+ command?: string | undefined;
722
+ env?: Record<string, string> | undefined;
723
+ cwd?: string | undefined;
724
+ url?: string | undefined;
725
+ headers?: Record<string, string> | undefined;
726
+ enabled?: boolean | undefined;
727
+ timeout?: number | undefined;
728
+ }, {
729
+ args?: string[] | undefined;
730
+ tools?: {
731
+ include?: string[] | undefined;
732
+ exclude?: string[] | undefined;
733
+ rename?: Record<string, string> | undefined;
734
+ } | undefined;
735
+ transport?: "stdio" | "sse" | "streamable-http" | undefined;
736
+ command?: string | undefined;
737
+ env?: Record<string, string> | undefined;
738
+ cwd?: string | undefined;
739
+ url?: string | undefined;
740
+ headers?: Record<string, string> | undefined;
741
+ enabled?: boolean | undefined;
742
+ timeout?: number | undefined;
743
+ }>, {
744
+ args?: string[] | undefined;
745
+ tools?: {
746
+ include?: string[] | undefined;
747
+ exclude?: string[] | undefined;
748
+ rename?: Record<string, string> | undefined;
749
+ } | undefined;
750
+ transport?: "stdio" | "sse" | "streamable-http" | undefined;
751
+ command?: string | undefined;
752
+ env?: Record<string, string> | undefined;
753
+ cwd?: string | undefined;
754
+ url?: string | undefined;
755
+ headers?: Record<string, string> | undefined;
756
+ enabled?: boolean | undefined;
757
+ timeout?: number | undefined;
758
+ }, {
759
+ args?: string[] | undefined;
760
+ tools?: {
761
+ include?: string[] | undefined;
762
+ exclude?: string[] | undefined;
763
+ rename?: Record<string, string> | undefined;
764
+ } | undefined;
765
+ transport?: "stdio" | "sse" | "streamable-http" | undefined;
766
+ command?: string | undefined;
767
+ env?: Record<string, string> | undefined;
768
+ cwd?: string | undefined;
769
+ url?: string | undefined;
770
+ headers?: Record<string, string> | undefined;
771
+ enabled?: boolean | undefined;
772
+ timeout?: number | undefined;
773
+ }>;
774
+ export declare const mcpConfigSchema: z.ZodObject<{
775
+ $schema: z.ZodOptional<z.ZodString>;
776
+ mcpServers: z.ZodRecord<z.ZodString, z.ZodEffects<z.ZodObject<{
777
+ transport: z.ZodOptional<z.ZodEnum<["stdio", "sse", "streamable-http"]>>;
778
+ command: z.ZodOptional<z.ZodString>;
779
+ args: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
780
+ env: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>;
781
+ cwd: z.ZodOptional<z.ZodString>;
782
+ url: z.ZodOptional<z.ZodString>;
783
+ headers: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>;
784
+ enabled: z.ZodOptional<z.ZodBoolean>;
785
+ timeout: z.ZodOptional<z.ZodNumber>;
786
+ tools: z.ZodOptional<z.ZodObject<{
787
+ include: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
788
+ exclude: z.ZodOptional<z.ZodArray<z.ZodString, "many">>;
789
+ rename: z.ZodOptional<z.ZodRecord<z.ZodString, z.ZodString>>;
790
+ }, "strict", z.ZodTypeAny, {
791
+ include?: string[] | undefined;
792
+ exclude?: string[] | undefined;
793
+ rename?: Record<string, string> | undefined;
794
+ }, {
795
+ include?: string[] | undefined;
796
+ exclude?: string[] | undefined;
797
+ rename?: Record<string, string> | undefined;
798
+ }>>;
799
+ }, "strict", z.ZodTypeAny, {
800
+ args?: string[] | undefined;
801
+ tools?: {
802
+ include?: string[] | undefined;
803
+ exclude?: string[] | undefined;
804
+ rename?: Record<string, string> | undefined;
805
+ } | undefined;
806
+ transport?: "stdio" | "sse" | "streamable-http" | undefined;
807
+ command?: string | undefined;
808
+ env?: Record<string, string> | undefined;
809
+ cwd?: string | undefined;
810
+ url?: string | undefined;
811
+ headers?: Record<string, string> | undefined;
812
+ enabled?: boolean | undefined;
813
+ timeout?: number | undefined;
814
+ }, {
815
+ args?: string[] | undefined;
816
+ tools?: {
817
+ include?: string[] | undefined;
818
+ exclude?: string[] | undefined;
819
+ rename?: Record<string, string> | undefined;
820
+ } | undefined;
821
+ transport?: "stdio" | "sse" | "streamable-http" | undefined;
822
+ command?: string | undefined;
823
+ env?: Record<string, string> | undefined;
824
+ cwd?: string | undefined;
825
+ url?: string | undefined;
826
+ headers?: Record<string, string> | undefined;
827
+ enabled?: boolean | undefined;
828
+ timeout?: number | undefined;
829
+ }>, {
830
+ args?: string[] | undefined;
831
+ tools?: {
832
+ include?: string[] | undefined;
833
+ exclude?: string[] | undefined;
834
+ rename?: Record<string, string> | undefined;
835
+ } | undefined;
836
+ transport?: "stdio" | "sse" | "streamable-http" | undefined;
837
+ command?: string | undefined;
838
+ env?: Record<string, string> | undefined;
839
+ cwd?: string | undefined;
840
+ url?: string | undefined;
841
+ headers?: Record<string, string> | undefined;
842
+ enabled?: boolean | undefined;
843
+ timeout?: number | undefined;
844
+ }, {
845
+ args?: string[] | undefined;
846
+ tools?: {
847
+ include?: string[] | undefined;
848
+ exclude?: string[] | undefined;
849
+ rename?: Record<string, string> | undefined;
850
+ } | undefined;
851
+ transport?: "stdio" | "sse" | "streamable-http" | undefined;
852
+ command?: string | undefined;
853
+ env?: Record<string, string> | undefined;
854
+ cwd?: string | undefined;
855
+ url?: string | undefined;
856
+ headers?: Record<string, string> | undefined;
857
+ enabled?: boolean | undefined;
858
+ timeout?: number | undefined;
859
+ }>>;
860
+ }, "strict", z.ZodTypeAny, {
861
+ mcpServers: Record<string, {
862
+ args?: string[] | undefined;
863
+ tools?: {
864
+ include?: string[] | undefined;
865
+ exclude?: string[] | undefined;
866
+ rename?: Record<string, string> | undefined;
867
+ } | undefined;
868
+ transport?: "stdio" | "sse" | "streamable-http" | undefined;
869
+ command?: string | undefined;
870
+ env?: Record<string, string> | undefined;
871
+ cwd?: string | undefined;
872
+ url?: string | undefined;
873
+ headers?: Record<string, string> | undefined;
874
+ enabled?: boolean | undefined;
875
+ timeout?: number | undefined;
876
+ }>;
877
+ $schema?: string | undefined;
878
+ }, {
879
+ mcpServers: Record<string, {
880
+ args?: string[] | undefined;
881
+ tools?: {
882
+ include?: string[] | undefined;
883
+ exclude?: string[] | undefined;
884
+ rename?: Record<string, string> | undefined;
885
+ } | undefined;
886
+ transport?: "stdio" | "sse" | "streamable-http" | undefined;
887
+ command?: string | undefined;
888
+ env?: Record<string, string> | undefined;
889
+ cwd?: string | undefined;
890
+ url?: string | undefined;
891
+ headers?: Record<string, string> | undefined;
892
+ enabled?: boolean | undefined;
893
+ timeout?: number | undefined;
894
+ }>;
895
+ $schema?: string | undefined;
896
+ }>;
897
+ export type McpToolsFilter = z.infer<typeof mcpToolsFilterSchema>;
898
+ export type McpServerConfig = z.infer<typeof mcpServerConfigSchema>;
899
+ export type McpConfig = z.infer<typeof mcpConfigSchema>;