@wix/evalforge-evaluator 0.30.0 → 0.32.0

package/build/index.js

@@ -6599,6 +6599,49 @@ function emitTraceEvent(event, tracePushUrl, routeHeader, authToken) {
     });
   }
 }
+function extractToolActionDescription(toolName, toolArgs) {
+  if (!toolName) {
+    return "Using tool...";
+  }
+  if (toolArgs) {
+    try {
+      const args = JSON.parse(toolArgs);
+      if ((toolName === "Task" || toolName === "dispatch_agent") && args.description) {
+        const desc2 = String(args.description).slice(0, 55);
+        const truncatedDesc = desc2.length < String(args.description).length ? `${desc2}...` : desc2;
+        return `Task: ${truncatedDesc}`;
+      }
+      if ((toolName === "Bash" || toolName === "bash" || toolName === "execute") && args.command) {
+        const cmd = String(args.command).slice(0, 50);
+        return `Running: ${cmd}${String(args.command).length > 50 ? "..." : ""}`;
+      }
+      if (toolName === "Search" || toolName === "search" || toolName === "Grep") {
+        const query = args.query || args.pattern || args.search;
+        if (query) {
+          return `Searching: ${String(query).slice(0, 40)}`;
+        }
+      }
+      if (toolName === "LS" || toolName === "ls" || toolName === "ListFiles") {
+        const path9 = args.path || args.directory || ".";
+        return `Listing: ${String(path9).slice(0, 50)}`;
+      }
+      if ((toolName === "Read" || toolName === "read" || toolName === "View") && (args.file_path || args.path || args.target_file)) {
+        const filePath = String(
+          args.file_path || args.path || args.target_file
+        ).slice(0, 50);
+        return `Reading: ${filePath}`;
+      }
+      if ((toolName === "Write" || toolName === "Edit" || toolName === "write") && (args.file_path || args.path || args.target_file)) {
+        const filePath = String(
+          args.file_path || args.path || args.target_file
+        ).slice(0, 50);
+        return `Writing: ${filePath}`;
+      }
+    } catch {
+    }
+  }
+  return `Using ${toolName}...`;
+}
 async function pushTraceEvent(url2, event, routeHeader, authToken) {
   try {
     const headers = {
@@ -6932,15 +6975,27 @@ async function executeWithClaudeCode(skill, scenario, options) {
       }, SDK_TIMEOUT_MS);
     });
     if (traceContext) {
+      let lastReportedAction = "";
+      let sameActionCount = 0;
       heartbeatHandle = setInterval(() => {
         const elapsedMs = Date.now() - executionStartTime;
         let progressMessage = lastAction;
-        if (lastToolName && lastFilePath) {
+        if (lastAction === lastReportedAction) {
+          sameActionCount++;
+        } else {
+          sameActionCount = 1;
+          lastReportedAction = lastAction;
+        }
+        const isTaskTool = lastToolName === "Task" || lastToolName === "dispatch_agent";
+        if (isTaskTool && sameActionCount > 1) {
+          progressMessage = `Waiting for ${lastAction}`;
+        } else if (lastToolName && lastFilePath) {
           progressMessage = `${lastToolName}: ${lastFilePath}`;
-        } else if (lastToolName) {
+        } else if (lastToolName && !isTaskTool) {
           progressMessage = `Using ${lastToolName}...`;
         }
-        progressMessage += ` (${Math.round(elapsedMs / 1e3)}s)`;
+        const elapsedSec = Math.round(elapsedMs / 1e3);
+        progressMessage += ` (${elapsedSec}s, step ${traceStepNumber})`;
         const progressEvent = {
           evalRunId: traceContext.evalRunId,
           scenarioId: traceContext.scenarioId,
@@ -7001,7 +7056,10 @@ IMPORTANT: This is an automated evaluation run. Execute the requested changes im
             if (traceEvent.type === import_evalforge_types.LiveTraceEventType.THINKING) {
               lastAction = "Thinking...";
             } else if (traceEvent.type === import_evalforge_types.LiveTraceEventType.TOOL_USE) {
-              lastAction = `Using ${traceEvent.toolName || "tool"}...`;
+              lastAction = extractToolActionDescription(
+                traceEvent.toolName,
+                traceEvent.toolArgs
+              );
             } else if (traceEvent.type === import_evalforge_types.LiveTraceEventType.FILE_WRITE) {
               lastAction = `Writing: ${traceEvent.filePath || "file"}`;
             } else if (traceEvent.type === import_evalforge_types.LiveTraceEventType.FILE_READ) {