feat(evaluate) - Run evaluateInEngine262 in child_process with strict timeout

src/test/manual/test-evaluate-in-engine262.ts

@@ -7,12 +7,7 @@
  * Usage: bun run src/test/manual/test-evaluate-in-engine262.ts ["your JavaScript code"]
  */
 
-import { createEvaluateInEngine262Tool } from "../../agent-tools";
-
-interface EvaluateResult {
-  importantSections: string[][];
-  otherSections: string[][];
-}
+import { createEvaluateInEngine262Tool } from "../../agent-tools/index.js";
 
 async function main() {
   // Get test code from command line or use default
@@ -39,13 +34,15 @@ async function main() {
 
   console.log("Executing tool...\n");
   try {
-    const result = await evaluateTool.func({ code: testCode });
+    const result = await evaluateTool({ code: testCode });
 
-    // Parse and verify results
-    const parsed: EvaluateResult = JSON.parse(result);
+    // Verify results
+    if ("error" in result) {
+      throw new Error(result.error);
+    }
 
-    const importantCount = parsed.importantSections.length;
-    const otherCount = parsed.otherSections.length;
+    const importantCount = result.importantSections.length;
+    const otherCount = result.otherSections.length;
     const totalCount = importantCount + otherCount;
 
     console.log(`\n✓ Captured ${totalCount} marks`);
@@ -53,9 +50,9 @@ async function main() {
     console.log("");
 
     // Flatten and dedupe section IDs
-    const importantIds = new Set(parsed.importantSections.flat());
+    const importantIds = new Set(result.importantSections);
     const otherIds = new Set(
-      parsed.otherSections.flat().filter((id) => !importantIds.has(id)),
+      result.otherSections.filter((id: string) => !importantIds.has(id)),
     );
 
     const totalUnique = importantIds.size + otherIds.size;

src/test/manual/test-evaluate-timeout.ts

@@ -0,0 +1,67 @@
+/**
+ * Test timeout functionality for evaluateInEngine262 tool
+ * Tests that code execution properly times out after the specified duration
+ */
+
+import { createEvaluateInEngine262Tool } from "../../agent-tools/evaluateInEngine262.js";
+
+async function testTimeout() {
+  console.log("=== Testing evaluateInEngine262 Timeout ===\n");
+
+  // Test 1: Quick code should complete before timeout
+  console.log("Test 1: Quick code (should succeed)...");
+  const quickTool = createEvaluateInEngine262Tool(5000); // 5 second timeout
+  const start1 = Date.now();
+  const quickResult = await quickTool({ code: "console.log('hello'); 1 + 1" });
+  const elapsed1 = Date.now() - start1;
+
+  if (quickResult.error) {
+    throw new Error(`Quick code failed: ${quickResult.error}`);
+  }
+  console.log(`✓ Completed in ${elapsed1}ms`);
+  console.log(
+    `  Console output: ${JSON.stringify(quickResult.consoleOutput)}\n`,
+  );
+
+  // Test 2: Long-running code should timeout
+  console.log("Test 2: Long-running code (should timeout after 500ms)...");
+  const slowTool = createEvaluateInEngine262Tool(500); // 500ms timeout
+  const start2 = Date.now();
+  const slowResult = await slowTool({
+    code: `
+      // Busy-wait loop that takes ~5 seconds
+      const start = Date.now();
+      while (Date.now() - start < 5000) {
+        // Busy wait - no yielding
+        for (let i = 0; i < 1000; i++) {
+          Math.sqrt(i);
+        }
+      }
+      console.log("completed");
+    `,
+  });
+  const elapsed2 = Date.now() - start2;
+
+  if (!slowResult.error) {
+    throw new Error("Expected timeout error but code completed successfully");
+  }
+  if (!slowResult.error.includes("timeout")) {
+    throw new Error(`Expected timeout error but got: ${slowResult.error}`);
+  }
+  console.log(`✓ Timed out in ${elapsed2}ms`);
+  console.log(`  Error message: ${slowResult.error}\n`);
+
+  // Note: worker.terminate() sends signal but CPU-bound loops may not stop immediately.
+  // The important thing is that we got a timeout error before the 5-second loop completed.
+  if (elapsed2 > 5000) {
+    throw new Error(`Timeout took too long: ${elapsed2}ms (expected < 5000ms)`);
+  }
+
+  console.log(`  (Note: Total time includes worker termination cleanup)`);
+  console.log("=== All timeout tests passed! ✓ ===");
+}
+
+testTimeout().catch((error) => {
+  console.error("\nTest failed:", error);
+  process.exit(1);
+});