@wix/evalforge-types 0.47.0 → 0.49.0
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/README.md +1 -1
- package/build/index.js +60 -29
- package/build/index.js.map +3 -3
- package/build/index.mjs +57 -26
- package/build/index.mjs.map +3 -3
- package/build/types/evaluation/eval-run.d.ts +14 -6
- package/build/types/project/project.d.ts +3 -0
- package/build/types/scenario/test-scenario.d.ts +3 -0
- package/build/types/target/index.d.ts +1 -1
- package/build/types/target/preset.d.ts +54 -0
- package/package.json +2 -2
- package/build/types/target/skills-group.d.ts +0 -37
package/README.md
CHANGED
|
@@ -9,7 +9,7 @@ Shared TypeScript types and [Zod](https://zod.dev/) schemas for the EvalForge pl
|
|
|
9
9
|
| Module | Description |
|
|
10
10
|
|--------|-------------|
|
|
11
11
|
| `common` | Base entity schemas (`BaseEntity`, `TenantEntity`), `Rule`, shared utilities |
|
|
12
|
-
| `target` | Testable entities — `Agent`, `Skill`, `SkillVersion`, `
|
|
12
|
+
| `target` | Testable entities — `Agent`, `Skill`, `SkillVersion`, `Preset`, `Mcp`, `SubAgent` |
|
|
13
13
|
| `scenario` | Test scenario definitions and environment configuration |
|
|
14
14
|
| `suite` | Test suite organization |
|
|
15
15
|
| `evaluation` | Eval run schemas, configs, results, and statuses |
|
package/build/index.js
CHANGED
|
@@ -59,11 +59,11 @@ __export(index_exports, {
|
|
|
59
59
|
CreateCustomAssertionInputSchema: () => CreateCustomAssertionInputSchema,
|
|
60
60
|
CreateEvalRunInputSchema: () => CreateEvalRunInputSchema,
|
|
61
61
|
CreateMcpInputSchema: () => CreateMcpInputSchema,
|
|
62
|
+
CreatePresetInputSchema: () => CreatePresetInputSchema,
|
|
62
63
|
CreateProjectInputSchema: () => CreateProjectInputSchema,
|
|
63
64
|
CreateRuleInputSchema: () => CreateRuleInputSchema,
|
|
64
65
|
CreateSkillInputSchema: () => CreateSkillInputSchema,
|
|
65
66
|
CreateSkillVersionInputSchema: () => CreateSkillVersionInputSchema,
|
|
66
|
-
CreateSkillsGroupInputSchema: () => CreateSkillsGroupInputSchema,
|
|
67
67
|
CreateSubAgentInputSchema: () => CreateSubAgentInputSchema,
|
|
68
68
|
CreateTemplateInputSchema: () => CreateTemplateInputSchema,
|
|
69
69
|
CreateTestScenarioInputSchema: () => CreateTestScenarioInputSchema,
|
|
@@ -115,6 +115,7 @@ __export(index_exports, {
|
|
|
115
115
|
OpenAIModel: () => Model,
|
|
116
116
|
OpenAIModelSchema: () => OpenAIModelSchema,
|
|
117
117
|
PlaywrightNLTestSchema: () => PlaywrightNLTestSchema,
|
|
118
|
+
PresetSchema: () => PresetSchema,
|
|
118
119
|
ProjectSchema: () => ProjectSchema,
|
|
119
120
|
PromptResultSchema: () => PromptResultSchema,
|
|
120
121
|
RUN_COMMAND_LABELS: () => RUN_COMMAND_LABELS,
|
|
@@ -135,7 +136,6 @@ __export(index_exports, {
|
|
|
135
136
|
SkillWasCalledAssertionSchema: () => SkillWasCalledAssertionSchema,
|
|
136
137
|
SkillWasCalledConfigSchema: () => SkillWasCalledConfigSchema,
|
|
137
138
|
SkillWithLatestVersionSchema: () => SkillWithLatestVersionSchema,
|
|
138
|
-
SkillsGroupSchema: () => SkillsGroupSchema,
|
|
139
139
|
SubAgentSchema: () => SubAgentSchema,
|
|
140
140
|
TRACE_EVENT_PREFIX: () => TRACE_EVENT_PREFIX,
|
|
141
141
|
TargetSchema: () => TargetSchema,
|
|
@@ -166,10 +166,10 @@ __export(index_exports, {
|
|
|
166
166
|
UpdateAgentInputSchema: () => UpdateAgentInputSchema,
|
|
167
167
|
UpdateCustomAssertionInputSchema: () => UpdateCustomAssertionInputSchema,
|
|
168
168
|
UpdateMcpInputSchema: () => UpdateMcpInputSchema,
|
|
169
|
+
UpdatePresetInputSchema: () => UpdatePresetInputSchema,
|
|
169
170
|
UpdateProjectInputSchema: () => UpdateProjectInputSchema,
|
|
170
171
|
UpdateRuleInputSchema: () => UpdateRuleInputSchema,
|
|
171
172
|
UpdateSkillInputSchema: () => UpdateSkillInputSchema,
|
|
172
|
-
UpdateSkillsGroupInputSchema: () => UpdateSkillsGroupInputSchema,
|
|
173
173
|
UpdateSubAgentInputSchema: () => UpdateSubAgentInputSchema,
|
|
174
174
|
UpdateTemplateInputSchema: () => UpdateTemplateInputSchema,
|
|
175
175
|
UpdateTestScenarioInputSchema: () => UpdateTestScenarioInputSchema,
|
|
@@ -1326,34 +1326,51 @@ var SkillWithLatestVersionSchema = SkillSchema.extend({
|
|
|
1326
1326
|
latestVersion: SkillVersionSchema.optional()
|
|
1327
1327
|
});
|
|
1328
1328
|
|
|
1329
|
-
// src/target/
|
|
1329
|
+
// src/target/sub-agent.ts
|
|
1330
1330
|
var import_zod8 = require("zod");
|
|
1331
|
-
var
|
|
1332
|
-
/**
|
|
1333
|
-
|
|
1331
|
+
var SubAgentSchema = TargetSchema.extend({
|
|
1332
|
+
/** The full sub-agent markdown content (YAML frontmatter + body) */
|
|
1333
|
+
subAgentMd: import_zod8.z.string()
|
|
1334
1334
|
});
|
|
1335
|
-
var
|
|
1335
|
+
var SubAgentInputBaseSchema = SubAgentSchema.omit({
|
|
1336
1336
|
id: true,
|
|
1337
1337
|
createdAt: true,
|
|
1338
1338
|
updatedAt: true,
|
|
1339
1339
|
deleted: true
|
|
1340
1340
|
});
|
|
1341
|
-
var
|
|
1341
|
+
var CreateSubAgentInputSchema = SubAgentInputBaseSchema;
|
|
1342
|
+
var UpdateSubAgentInputSchema = SubAgentInputBaseSchema.partial();
|
|
1342
1343
|
|
|
1343
|
-
// src/target/
|
|
1344
|
+
// src/target/preset.ts
|
|
1344
1345
|
var import_zod9 = require("zod");
|
|
1345
|
-
var
|
|
1346
|
-
/**
|
|
1347
|
-
|
|
1348
|
-
|
|
1349
|
-
|
|
1346
|
+
var PresetSchema = TenantEntitySchema.extend({
|
|
1347
|
+
/** Agent ID for this preset */
|
|
1348
|
+
agentId: import_zod9.z.string(),
|
|
1349
|
+
/** Skill IDs included in this preset */
|
|
1350
|
+
skillIds: import_zod9.z.array(import_zod9.z.string()).default([]),
|
|
1351
|
+
/** Optional map of skillId → skillVersionId for version pinning */
|
|
1352
|
+
skillVersions: import_zod9.z.record(import_zod9.z.string(), import_zod9.z.string()).optional(),
|
|
1353
|
+
/** MCP server IDs included in this preset */
|
|
1354
|
+
mcpIds: import_zod9.z.array(import_zod9.z.string()).default([]),
|
|
1355
|
+
/** Sub-agent IDs included in this preset */
|
|
1356
|
+
subAgentIds: import_zod9.z.array(import_zod9.z.string()).default([]),
|
|
1357
|
+
/** Rule IDs included in this preset */
|
|
1358
|
+
ruleIds: import_zod9.z.array(import_zod9.z.string()).default([])
|
|
1359
|
+
});
|
|
1360
|
+
var atLeastOneEntity = (data) => (data.skillIds?.length ?? 0) > 0 || (data.mcpIds?.length ?? 0) > 0 || (data.subAgentIds?.length ?? 0) > 0 || (data.ruleIds?.length ?? 0) > 0;
|
|
1361
|
+
var AT_LEAST_ONE_ENTITY_MESSAGE = "At least one of skillIds, mcpIds, subAgentIds, or ruleIds must be non-empty";
|
|
1362
|
+
var CreatePresetInputSchema = PresetSchema.omit({
|
|
1350
1363
|
id: true,
|
|
1351
1364
|
createdAt: true,
|
|
1352
1365
|
updatedAt: true,
|
|
1353
1366
|
deleted: true
|
|
1354
|
-
});
|
|
1355
|
-
var
|
|
1356
|
-
|
|
1367
|
+
}).refine(atLeastOneEntity, { message: AT_LEAST_ONE_ENTITY_MESSAGE });
|
|
1368
|
+
var UpdatePresetInputSchema = PresetSchema.omit({
|
|
1369
|
+
id: true,
|
|
1370
|
+
createdAt: true,
|
|
1371
|
+
updatedAt: true,
|
|
1372
|
+
deleted: true
|
|
1373
|
+
}).partial();
|
|
1357
1374
|
|
|
1358
1375
|
// src/test/index.ts
|
|
1359
1376
|
var import_zod20 = require("zod");
|
|
@@ -1791,7 +1808,9 @@ var TestScenarioSchema = TenantEntitySchema.extend({
|
|
|
1791
1808
|
/** IDs of saved assertions to evaluate (from assertions table) - legacy, use assertionLinks */
|
|
1792
1809
|
assertionIds: import_zod24.z.array(import_zod24.z.string()).optional(),
|
|
1793
1810
|
/** Linked assertions with per-scenario parameter values */
|
|
1794
|
-
assertionLinks: import_zod24.z.array(ScenarioAssertionLinkSchema).optional()
|
|
1811
|
+
assertionLinks: import_zod24.z.array(ScenarioAssertionLinkSchema).optional(),
|
|
1812
|
+
/** Tags for categorisation and filtering */
|
|
1813
|
+
tags: import_zod24.z.array(import_zod24.z.string()).optional()
|
|
1795
1814
|
});
|
|
1796
1815
|
var CreateTestScenarioInputSchema = TestScenarioSchema.omit({
|
|
1797
1816
|
id: true,
|
|
@@ -2056,11 +2075,13 @@ var FailureAnalysisSchema = import_zod28.z.object({
|
|
|
2056
2075
|
var EvalRunSchema = TenantEntitySchema.extend({
|
|
2057
2076
|
/** Agent ID for this run */
|
|
2058
2077
|
agentId: import_zod28.z.string().optional(),
|
|
2059
|
-
/**
|
|
2060
|
-
|
|
2078
|
+
/** Preset ID that originated this run (optional) */
|
|
2079
|
+
presetId: import_zod28.z.string().optional(),
|
|
2080
|
+
/** Skill IDs for this run */
|
|
2081
|
+
skillIds: import_zod28.z.array(import_zod28.z.string()).optional(),
|
|
2061
2082
|
/** Map of skillId to skillVersionId for this run */
|
|
2062
2083
|
skillVersions: import_zod28.z.record(import_zod28.z.string(), import_zod28.z.string()).optional(),
|
|
2063
|
-
/** Scenario IDs to run */
|
|
2084
|
+
/** Scenario IDs to run (always present — resolved server-side from tags when needed) */
|
|
2064
2085
|
scenarioIds: import_zod28.z.array(import_zod28.z.string()),
|
|
2065
2086
|
/** Current status */
|
|
2066
2087
|
status: EvalStatusSchema,
|
|
@@ -2095,7 +2116,9 @@ var EvalRunSchema = TenantEntitySchema.extend({
|
|
|
2095
2116
|
/** Sub-agent IDs to enable for this run (optional) */
|
|
2096
2117
|
subAgentIds: import_zod28.z.array(import_zod28.z.string()).optional(),
|
|
2097
2118
|
/** Rule IDs to enable for this run (optional) */
|
|
2098
|
-
ruleIds: import_zod28.z.array(import_zod28.z.string()).optional()
|
|
2119
|
+
ruleIds: import_zod28.z.array(import_zod28.z.string()).optional(),
|
|
2120
|
+
/** Tags used to select scenarios for this run (for traceability) */
|
|
2121
|
+
tags: import_zod28.z.array(import_zod28.z.string()).optional()
|
|
2099
2122
|
});
|
|
2100
2123
|
var CreateEvalRunInputSchema = EvalRunSchema.omit({
|
|
2101
2124
|
id: true,
|
|
@@ -2106,8 +2129,15 @@ var CreateEvalRunInputSchema = EvalRunSchema.omit({
|
|
|
2106
2129
|
results: true,
|
|
2107
2130
|
aggregateMetrics: true,
|
|
2108
2131
|
startedAt: true,
|
|
2109
|
-
completedAt: true
|
|
2110
|
-
|
|
2132
|
+
completedAt: true,
|
|
2133
|
+
scenarioIds: true
|
|
2134
|
+
}).extend({
|
|
2135
|
+
/** Optional on input — backend resolves from tags when not provided */
|
|
2136
|
+
scenarioIds: import_zod28.z.array(import_zod28.z.string()).optional()
|
|
2137
|
+
}).refine(
|
|
2138
|
+
(data) => data.scenarioIds && data.scenarioIds.length > 0 || data.tags && data.tags.length > 0,
|
|
2139
|
+
{ message: "Either scenarioIds or tags must be provided" }
|
|
2140
|
+
);
|
|
2111
2141
|
var EvaluationProgressSchema = import_zod28.z.object({
|
|
2112
2142
|
runId: import_zod28.z.string(),
|
|
2113
2143
|
targetId: import_zod28.z.string(),
|
|
@@ -2281,7 +2311,8 @@ var ProjectSchema = BaseEntitySchema.extend({
|
|
|
2281
2311
|
appId: import_zod31.z.string().optional().describe("The ID of the app in Dev Center"),
|
|
2282
2312
|
appSecret: import_zod31.z.string().optional().describe("The secret of the app in Dev Center"),
|
|
2283
2313
|
useWixAuth: import_zod31.z.boolean().optional().describe("Enable Wix CLI/MCP auth for evaluations"),
|
|
2284
|
-
useBase44Auth: import_zod31.z.boolean().optional().describe("Enable Base44 auth for evaluations")
|
|
2314
|
+
useBase44Auth: import_zod31.z.boolean().optional().describe("Enable Base44 auth for evaluations"),
|
|
2315
|
+
scenarioTags: import_zod31.z.array(import_zod31.z.string()).optional().describe("Project-level tag vocabulary for scenarios")
|
|
2285
2316
|
});
|
|
2286
2317
|
var CreateProjectInputSchema = ProjectSchema.omit({
|
|
2287
2318
|
id: true,
|
|
@@ -2493,11 +2524,11 @@ function getSystemAssertion(id) {
|
|
|
2493
2524
|
CreateCustomAssertionInputSchema,
|
|
2494
2525
|
CreateEvalRunInputSchema,
|
|
2495
2526
|
CreateMcpInputSchema,
|
|
2527
|
+
CreatePresetInputSchema,
|
|
2496
2528
|
CreateProjectInputSchema,
|
|
2497
2529
|
CreateRuleInputSchema,
|
|
2498
2530
|
CreateSkillInputSchema,
|
|
2499
2531
|
CreateSkillVersionInputSchema,
|
|
2500
|
-
CreateSkillsGroupInputSchema,
|
|
2501
2532
|
CreateSubAgentInputSchema,
|
|
2502
2533
|
CreateTemplateInputSchema,
|
|
2503
2534
|
CreateTestScenarioInputSchema,
|
|
@@ -2549,6 +2580,7 @@ function getSystemAssertion(id) {
|
|
|
2549
2580
|
OpenAIModel,
|
|
2550
2581
|
OpenAIModelSchema,
|
|
2551
2582
|
PlaywrightNLTestSchema,
|
|
2583
|
+
PresetSchema,
|
|
2552
2584
|
ProjectSchema,
|
|
2553
2585
|
PromptResultSchema,
|
|
2554
2586
|
RUN_COMMAND_LABELS,
|
|
@@ -2569,7 +2601,6 @@ function getSystemAssertion(id) {
|
|
|
2569
2601
|
SkillWasCalledAssertionSchema,
|
|
2570
2602
|
SkillWasCalledConfigSchema,
|
|
2571
2603
|
SkillWithLatestVersionSchema,
|
|
2572
|
-
SkillsGroupSchema,
|
|
2573
2604
|
SubAgentSchema,
|
|
2574
2605
|
TRACE_EVENT_PREFIX,
|
|
2575
2606
|
TargetSchema,
|
|
@@ -2600,10 +2631,10 @@ function getSystemAssertion(id) {
|
|
|
2600
2631
|
UpdateAgentInputSchema,
|
|
2601
2632
|
UpdateCustomAssertionInputSchema,
|
|
2602
2633
|
UpdateMcpInputSchema,
|
|
2634
|
+
UpdatePresetInputSchema,
|
|
2603
2635
|
UpdateProjectInputSchema,
|
|
2604
2636
|
UpdateRuleInputSchema,
|
|
2605
2637
|
UpdateSkillInputSchema,
|
|
2606
|
-
UpdateSkillsGroupInputSchema,
|
|
2607
2638
|
UpdateSubAgentInputSchema,
|
|
2608
2639
|
UpdateTemplateInputSchema,
|
|
2609
2640
|
UpdateTestScenarioInputSchema,
|