@wix/evalforge-types 0.34.0 → 0.36.0

package/build/index.mjs

@@ -983,9 +983,18 @@ var TargetSchema = TenantEntitySchema.extend({
 
 // src/target/agent.ts
 import { z as z5 } from "zod";
+var AgentRunCommand = /* @__PURE__ */ ((AgentRunCommand2) => {
+  AgentRunCommand2["CLAUDE"] = "claude";
+  return AgentRunCommand2;
+})(AgentRunCommand || {});
+var AVAILABLE_RUN_COMMANDS = Object.values(AgentRunCommand);
+var RUN_COMMAND_LABELS = {
+  ["claude" /* CLAUDE */]: "Claude Code"
+};
+var AgentRunCommandSchema = z5.nativeEnum(AgentRunCommand);
 var AgentSchema = TargetSchema.extend({
   /** Command to run the agent */
-  runCommand: z5.string(),
+  runCommand: AgentRunCommandSchema,
   /** Optional model configuration override */
   modelConfig: ModelConfigSchema.optional()
 });
@@ -1324,9 +1333,15 @@ var LlmJudgeAssertionSchema = z20.object({
   maxTokens: z20.number().int().optional(),
   temperature: z20.number().min(0).max(1).optional()
 });
+var TimeAssertionSchema = z20.object({
+  type: z20.literal("time_limit"),
+  /** Maximum allowed duration in milliseconds */
+  maxDurationMs: z20.number().int().positive()
+});
 var AssertionSchema = z20.union([
   SkillWasCalledAssertionSchema,
   BuildPassedAssertionSchema,
+  TimeAssertionSchema,
   LlmJudgeAssertionSchema
 ]);
 
@@ -1372,6 +1387,7 @@ import { z as z22 } from "zod";
 var AssertionTypeSchema = z22.enum([
   "skill_was_called",
   "build_passed",
+  "time_limit",
   "llm_judge"
 ]);
 var AssertionParameterTypeSchema = z22.enum([
@@ -1412,6 +1428,10 @@ var BuildPassedConfigSchema = z22.strictObject({
   /** Expected exit code (default: 0) */
   expectedExitCode: z22.number().int().optional()
 });
+var TimeConfigSchema = z22.strictObject({
+  /** Maximum allowed duration in milliseconds */
+  maxDurationMs: z22.number().int().positive()
+});
 var LlmJudgeConfigSchema = z22.object({
   /**
    * Prompt template with placeholders:
@@ -1441,7 +1461,9 @@ var AssertionConfigSchema = z22.union([
   LlmJudgeConfigSchema,
   // requires prompt - check first
   SkillWasCalledConfigSchema,
-  // requires skillName
+  // requires skillNames
+  TimeConfigSchema,
+  // requires maxDurationMs, uses strictObject
   BuildPassedConfigSchema,
   // all optional, uses strictObject to reject unknown keys
   z22.object({})
@@ -1466,6 +1488,8 @@ function validateAssertionConfig(type, config) {
       return SkillWasCalledConfigSchema.safeParse(config).success;
     case "build_passed":
       return BuildPassedConfigSchema.safeParse(config).success;
+    case "time_limit":
+      return TimeConfigSchema.safeParse(config).success;
     case "llm_judge":
       return LlmJudgeConfigSchema.safeParse(config).success;
     default:
@@ -1969,6 +1993,7 @@ var UpdateTemplateInputSchema = CreateTemplateInputSchema.partial();
 var SYSTEM_ASSERTION_IDS = {
   SKILL_WAS_CALLED: "system:skill_was_called",
   BUILD_PASSED: "system:build_passed",
+  TIME_LIMIT: "system:time_limit",
   LLM_JUDGE: "system:llm_judge"
 };
 function isSystemAssertionId(id) {
@@ -2025,6 +2050,21 @@ var SYSTEM_ASSERTIONS = {
       }
     ]
   },
+  [SYSTEM_ASSERTION_IDS.TIME_LIMIT]: {
+    id: SYSTEM_ASSERTION_IDS.TIME_LIMIT,
+    name: "Time Limit",
+    description: "Check that the scenario completed within a maximum duration",
+    type: "time_limit",
+    parameters: [
+      {
+        name: "maxDurationMs",
+        label: "Max Duration (ms)",
+        type: "number",
+        required: true,
+        defaultValue: 3e5
+      }
+    ]
+  },
   [SYSTEM_ASSERTION_IDS.LLM_JUDGE]: {
     id: SYSTEM_ASSERTION_IDS.LLM_JUDGE,
     name: "LLM Judge",
@@ -2076,6 +2116,9 @@ function getSystemAssertion(id) {
 var export_ClaudeModel = import_types.ClaudeModel;
 export {
   AVAILABLE_MODEL_IDS,
+  AVAILABLE_RUN_COMMANDS,
+  AgentRunCommand,
+  AgentRunCommandSchema,
   AgentSchema,
   AllowedCommands,
   ApiCallSchema,
@@ -2153,6 +2196,7 @@ export {
   PlaywrightNLTestSchema,
   ProjectSchema,
   PromptResultSchema,
+  RUN_COMMAND_LABELS,
   SEMVER_REGEX,
   SKILL_FOLDER_NAME_REGEX,
   SYSTEM_ASSERTIONS,
@@ -2182,6 +2226,8 @@ export {
   TestSuiteSchema,
   TestType,
   TestTypeSchema,
+  TimeAssertionSchema,
+  TimeConfigSchema,
   TokenUsageSchema,
   ToolTestSchema,
   TriggerMetadataSchema,