@m4trix/evals 0.28.0 → 0.30.0

package/dist/cli-simple.js

@@ -986,8 +986,11 @@ function processOneEvaluation(task, unit, totalEvaluations, publishEvent, persis
               output,
               meta: {
                 triggerId: task.triggerId,
+                triggerTimestamp: task.triggerTimestamp,
                 runId: evaluatorRunId,
                 datasetName: task.dataset.getDisplayLabel(),
+                testCaseId: testCaseItem.id,
+                testCaseName: getTestCaseDisplayLabel(testCaseItem.testCase),
                 repetitionId,
                 repetitionIndex,
                 repetitionCount,
@@ -1464,6 +1467,7 @@ var EffectRunner = class {
     const globalConcurrency = Math.max(1, request.globalConcurrency);
     const sem = Effect.unsafeMakeSemaphore(globalConcurrency);
     const triggerId = request.triggerId ?? `trg-${randomUUID()}`;
+    const triggerTimestamp = request.triggerTimestamp ?? Date.now();
     const snapshots = [];
     for (const job of request.jobs) {
       snapshots.push(
@@ -1471,6 +1475,7 @@ var EffectRunner = class {
           datasetId: job.datasetId,
           evaluatorIds: job.evaluatorIds,
           triggerId,
+          triggerTimestamp,
           maxConcurrency: this.config.maxConcurrency ?? 1,
           globalEvaluationSemaphore: sem,
           runConfigName: job.runConfigName,
@@ -1508,6 +1513,7 @@ var EffectRunner = class {
       datasetId: request.datasetId,
       evaluatorIds: request.evaluatorIds,
       triggerId: request.triggerId,
+      triggerTimestamp: request.triggerTimestamp ?? Date.now(),
       maxConcurrency: request.concurrency ?? this.config.maxConcurrency ?? 1,
       repetitions: request.repetitions,
       runConfigName,
@@ -1535,6 +1541,7 @@ var EffectRunner = class {
     const totalEvaluations = selectedTestCases.length * repetitions;
     const runConfigTags = [...params.runConfigTags ?? []];
     const triggerId = params.triggerId ?? `trg-${randomUUID()}`;
+    const triggerTimestamp = params.triggerTimestamp ?? Date.now();
     const runId = `run-${randomUUID()}`;
     const artifactPath = createArtifactPath(this.config.artifactDirectory, params.datasetId, runId);
     const snapshot = {
@@ -1578,6 +1585,7 @@ var EffectRunner = class {
       Queue.offer(this.runQueue, {
         runId,
         triggerId,
+        triggerTimestamp,
         datasetId: params.datasetId,
         dataset: dataset.dataset,
         evaluators: selectedEvaluators,
@@ -2010,6 +2018,7 @@ function RunView({
   runConfigNames,
   concurrency,
   experimentName,
+  triggerTimestamp,
   onComplete
 }) {
   const [phase, setPhase] = useState("loading");
@@ -2180,7 +2189,8 @@ function RunView({
     const snapshots = await runner.runDatasetJobsWithSharedConcurrency({
       jobs,
       globalConcurrency: concurrency,
-      experimentName
+      experimentName,
+      triggerTimestamp
     });
     for (let i = 0; i < snapshots.length; i += 1) {
       const snap = snapshots[i];
@@ -2237,7 +2247,7 @@ function RunView({
     setPhase("completed");
     const exitCode = failedTestCases > 0 ? 1 : 0;
     setTimeout(() => onComplete(void 0, exitCode), 200);
-  }, [runner, runConfigNames, concurrency, experimentName, onComplete]);
+  }, [runner, runConfigNames, concurrency, experimentName, triggerTimestamp, onComplete]);
   useEffect(() => {
     void runEval();
   }, [runEval]);
@@ -2725,7 +2735,7 @@ function formatEvaluatorScoreLine(name, scores, passed, metrics, options) {
   }
   return lines;
 }
-async function runSimpleEvalRunConfigsPlain(runner, runConfigNames, concurrency, experimentName) {
+async function runSimpleEvalRunConfigsPlain(runner, runConfigNames, concurrency, experimentName, triggerTimestamp) {
   const jobs = await runner.expandRunConfigNamesToJobs(runConfigNames);
   if (jobs.length === 0) {
     throw new Error("No jobs expanded from RunConfigs.");
@@ -2935,7 +2945,8 @@ async function runSimpleEvalRunConfigsPlain(runner, runConfigNames, concurrency,
   const snapshots = await runner.runDatasetJobsWithSharedConcurrency({
     jobs,
     globalConcurrency: concurrency,
-    experimentName
+    experimentName,
+    triggerTimestamp
   });
   for (let i = 0; i < snapshots.length; i += 1) {
     const snap = snapshots[i];
@@ -3035,7 +3046,7 @@ async function runSimpleEvalRunConfigsPlain(runner, runConfigNames, concurrency,
   }
   return failedTestCasesTotal > 0 ? 1 : 0;
 }
-async function runSimpleEvalRunConfigsInk(runner, runConfigNames, concurrency, experimentName) {
+async function runSimpleEvalRunConfigsInk(runner, runConfigNames, concurrency, experimentName, triggerTimestamp) {
   return new Promise((resolve5, reject) => {
     const app = render(
       React.createElement(RunView, {
@@ -3043,6 +3054,7 @@ async function runSimpleEvalRunConfigsInk(runner, runConfigNames, concurrency, e
         runConfigNames,
         concurrency,
         experimentName,
+        triggerTimestamp,
         onComplete: (err, exitCode) => {
           app.unmount();
           if (err) {
@@ -3100,11 +3112,13 @@ async function main() {
   try {
     if (args.command === "run") {
       const concurrency = args.concurrency ?? getDefaultConcurrency();
+      const triggerTimestamp = Date.now();
       const exitCode = await (useInk ? runSimpleEvalRunConfigsInk : runSimpleEvalRunConfigsPlain)(
         runner,
         args.runConfigNames,
         concurrency,
-        args.experimentName
+        args.experimentName,
+        triggerTimestamp
       );
       if (args.ci && exitCode !== 0) {
         process.exit(1);