@wix/evalforge-evaluator 0.187.0 → 0.189.0

package/build/index.mjs

@@ -7353,8 +7353,7 @@ async function fetchEvaluationData(api, projectId2, evalRunId2) {
 import {
   AssertionResultStatus,
   DEFAULT_JUDGE_MODEL,
-  LiveTraceEventType as LiveTraceEventType4,
-  formatTraceEventLine
+  LiveTraceEventType as LiveTraceEventType4
 } from "@wix/evalforge-types";
 import {
   evaluateAssertions as evaluateAssertionsBase
@@ -7485,10 +7484,13 @@ function installWithCache(workDir, exec, cacheBase, pm) {
     );
   }
 }
-async function installDependencies(workDir, exec = execFileSync, cacheBase) {
+async function installDependencies(workDir, onProgress, options = {}) {
   if (!existsSync(path.join(workDir, "package.json"))) {
     return;
   }
+  const exec = options.exec ?? execFileSync;
+  const cacheBase = options.cacheBase;
+  onProgress("Installing dependencies...");
   const pm = detectPackageManager(workDir);
   if (cacheBase) {
     installWithCache(workDir, exec, cacheBase, pm);
@@ -7564,7 +7566,8 @@ function writeWixEnvFile(workDir) {
     console.warn("[environment] Failed to read wix.config.json");
   }
 }
-async function prepareWorkingDirectory(config, evalRunId2, targetId, scenarioId, template) {
+async function prepareWorkingDirectory(config, evalRunId2, targetId, scenarioId, onProgress, options = {}) {
+  const template = options.template;
   const baseDir = config.evaluationsDir ?? path2.join(tmpdir(), "evalforge-evaluations");
   const nodeModulesCacheDir = path2.join(baseDir, "_node_modules_cache");
   if (template) {
@@ -7578,10 +7581,14 @@ async function prepareWorkingDirectory(config, evalRunId2, targetId, scenarioId,
       rmSync(workDir2, { recursive: true });
     }
     mkdirSync2(workDir2, { recursive: true });
+    onProgress("Fetching template files...");
     await fetchAndWriteTemplateFiles(template, workDir2);
     console.log(`Template files written to ${workDir2}`);
     writeWixEnvFile(workDir2);
-    await installDependencies(workDir2, void 0, nodeModulesCacheDir);
+    await installDependencies(workDir2, onProgress, {
+      cacheBase: nodeModulesCacheDir
+    });
+    onProgress("Environment ready");
     return workDir2;
   }
   const workDir = path2.join(baseDir, `${evalRunId2}_${targetId}_${scenarioId}`);
@@ -7593,6 +7600,13 @@ async function prepareWorkingDirectory(config, evalRunId2, targetId, scenarioId,
   return workDir;
 }
 
+// src/run-scenario/agents/shared/trace-emit.ts
+import { TRACE_EVENT_PREFIX } from "@wix/evalforge-types";
+function emitTraceEvent(event, pushEvent) {
+  console.log(`${TRACE_EVENT_PREFIX}${JSON.stringify(event)}`);
+  pushEvent?.(event);
+}
+
 // src/run-scenario/run-agent-with-context.ts
 import { randomUUID as randomUUID4 } from "crypto";
 
@@ -8051,13 +8065,6 @@ function buildConversation(timestampedMessages) {
   return messages;
 }
 
-// src/run-scenario/agents/shared/trace-emit.ts
-import { TRACE_EVENT_PREFIX } from "@wix/evalforge-types";
-function emitTraceEvent(event, pushEvent) {
-  console.log(`${TRACE_EVENT_PREFIX}${JSON.stringify(event)}`);
-  pushEvent?.(event);
-}
-
 // src/run-scenario/agents/claude-code/execute.ts
 var DEFAULT_MODEL = ClaudeModel.CLAUDE_4_5_SONNET_1_0;
 async function* buildPromptStream(triggerPrompt, images) {
@@ -12045,6 +12052,22 @@ async function runAgentWithContext(config, evalRunId2, scenario, evalData, workD
 async function runScenario(config, evalRunId2, scenario, evalData, template, resolvedAssertions, pushEvent, apiClient, projectId2) {
   const targetId = evalData.evalRun.presetId ?? evalData.agent?.id ?? evalData.evalRun.id;
   const targetName = evalData.presetName ?? evalData.agent?.name ?? "";
+  const emitSetupProgress = (outputPreview) => emitTraceEvent(
+    {
+      evalRunId: evalRunId2,
+      scenarioId: scenario.id,
+      scenarioName: scenario.name,
+      targetId,
+      targetName,
+      stepNumber: 0,
+      type: LiveTraceEventType4.PROGRESS,
+      outputPreview,
+      elapsedMs: 0,
+      timestamp: (/* @__PURE__ */ new Date()).toISOString(),
+      isComplete: false
+    },
+    pushEvent
+  );
   let provisionedSite;
   if (apiClient && projectId2 && scenario.siteSetup && scenario.siteSetup.mode !== "none") {
     provisionedSite = await apiClient.provisionScenarioSite(
@@ -12075,29 +12098,14 @@ async function runScenario(config, evalRunId2, scenario, evalData, template, res
 
 Site ID: ${provisionedSite.id}` : scenario.triggerPrompt;
   try {
-    if (template) {
-      console.log(
-        formatTraceEventLine({
-          evalRunId: evalRunId2,
-          scenarioId: scenario.id,
-          scenarioName: scenario.name,
-          targetId,
-          targetName,
-          stepNumber: 0,
-          type: LiveTraceEventType4.PROGRESS,
-          outputPreview: "Setting up environment (installing dependencies)...",
-          elapsedMs: 0,
-          timestamp: (/* @__PURE__ */ new Date()).toISOString(),
-          isComplete: false
-        })
-      );
-    }
+    emitSetupProgress("Setting up environment...");
     const workDir = await prepareWorkingDirectory(
       config,
       evalRunId2,
       targetId,
       scenario.id,
-      template
+      emitSetupProgress,
+      { template }
     );
     const partialResult = await runAgentWithContext(
       config,