@m4trix/evals 0.30.0 → 0.32.0

package/dist/index.d.ts

@@ -279,6 +279,8 @@ interface EvaluateMeta {
      * `runDatasetWith` / `runDatasetJobsWithSharedConcurrency` was invoked). Shared across all jobs in a batch.
      */
     triggerTimestamp: number;
+    /** Same instant as {@link triggerTimestamp}, as an ISO 8601 string (`toISOString()`). */
+    triggeredAt: string;
     /**
      * Identifier of the current test-case execution shared across all evaluators
      * for this specific test-case run.

package/dist/index.js

@@ -1302,6 +1302,23 @@ function isRunConfigLike(value) {
 function isTestCaseLike(value) {
   return hasMethod(value, "getName") && hasMethod(value, "getTags") && hasMethod(value, "getInput");
 }
+function collectTestCasesFromExportValues(exports) {
+  const out = [];
+  for (const value of exports) {
+    if (isTestCaseLike(value)) {
+      out.push(value);
+      continue;
+    }
+    if (Array.isArray(value)) {
+      for (const item of value) {
+        if (isTestCaseLike(item)) {
+          out.push(item);
+        }
+      }
+    }
+  }
+  return out;
+}
 async function walkDirectory(rootDir, excludeDirectories) {
   const out = [];
   async function walk(currentDir) {
@@ -1410,7 +1427,7 @@ async function collectTestCasesFromFiles(config) {
   const found = await Promise.all(
     matched.map(async (absolutePath) => {
       const exports = await loadModuleExports(absolutePath);
-      const testCases = exports.filter(isTestCaseLike);
+      const testCases = collectTestCasesFromExportValues(exports);
       const relPath = relative(config.rootDir, absolutePath);
       return testCases.map((testCase) => ({
         id: toId("test-case", relPath, testCase.getName()),
@@ -1567,6 +1584,7 @@ function processOneEvaluation(task, unit, totalEvaluations, publishEvent, persis
               meta: {
                 triggerId: task.triggerId,
                 triggerTimestamp: task.triggerTimestamp,
+                triggeredAt: new Date(task.triggerTimestamp).toISOString(),
                 runId: evaluatorRunId,
                 datasetName: task.dataset.getDisplayLabel(),
                 testCaseId: testCaseItem.id,