@wix/evalforge-evaluator 0.29.0 → 0.31.0

package/build/index.js

@@ -6599,6 +6599,48 @@ 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, 60);
+        return desc2.length < String(args.description).length ? `${desc2}...` : desc2;
+      }
+      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 ${toolName}...`;
+}
 async function pushTraceEvent(url2, event, routeHeader, authToken) {
   try {
     const headers = {
@@ -6825,6 +6867,9 @@ async function executeWithClaudeCode(skill, scenario, options) {
   console.log("[SDK-DEBUG] ============================================");
   let traceStepNumber = 0;
   const traceContext = options.traceContext;
+  let lastAction = "Starting...";
+  let lastToolName;
+  let lastFilePath;
   const maxTurns = options.maxTurns ?? 10;
   console.error(
     "[DEBUG-H5] Claude SDK query START",
@@ -6931,6 +6976,13 @@ async function executeWithClaudeCode(skill, scenario, options) {
     if (traceContext) {
       heartbeatHandle = setInterval(() => {
         const elapsedMs = Date.now() - executionStartTime;
+        let progressMessage = lastAction;
+        if (lastToolName && lastFilePath) {
+          progressMessage = `${lastToolName}: ${lastFilePath}`;
+        } else if (lastToolName) {
+          progressMessage = `Using ${lastToolName}...`;
+        }
+        progressMessage += ` (${Math.round(elapsedMs / 1e3)}s)`;
         const progressEvent = {
           evalRunId: traceContext.evalRunId,
           scenarioId: traceContext.scenarioId,
@@ -6939,7 +6991,9 @@ async function executeWithClaudeCode(skill, scenario, options) {
           targetName: traceContext.targetName,
           stepNumber: traceStepNumber,
           type: import_evalforge_types.LiveTraceEventType.PROGRESS,
-          outputPreview: `Executing... (${Math.round(elapsedMs / 1e3)}s elapsed, ${messageCount} messages)`,
+          outputPreview: progressMessage,
+          toolName: lastToolName,
+          filePath: lastFilePath,
           elapsedMs,
           timestamp: (/* @__PURE__ */ new Date()).toISOString(),
           isComplete: false
@@ -6984,6 +7038,22 @@ IMPORTANT: This is an automated evaluation run. Execute the requested changes im
             // Not complete yet
           );
           if (traceEvent) {
+            lastToolName = traceEvent.toolName;
+            lastFilePath = traceEvent.filePath;
+            if (traceEvent.type === import_evalforge_types.LiveTraceEventType.THINKING) {
+              lastAction = "Thinking...";
+            } else if (traceEvent.type === import_evalforge_types.LiveTraceEventType.TOOL_USE) {
+              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) {
+              lastAction = `Reading: ${traceEvent.filePath || "file"}`;
+            } else if (traceEvent.type === import_evalforge_types.LiveTraceEventType.COMPLETION) {
+              lastAction = "Processing response...";
+            }
             emitTraceEvent(
               traceEvent,
               traceContext.tracePushUrl,