fix(controller): credential-extract — IIFE auto-wrap for bare arrow fns + log eval result

src/controller/service.ts

@@ -288,6 +288,22 @@ const validateAndFixJavaScript = (code: string): string => {
     (_match, selector: string) => `.matches(\`${selector}\`)`
   );
 
+  // Auto-wrap bare async function expressions in IIFE so they actually
+  // execute. The LLM commonly emits `async () => { ... }` (or
+  // `async function () { ... }`) without trailing `()`, which evaluates
+  // to a function reference and silently does nothing — this is the
+  // root cause behind credential-extraction failures where
+  // `clipboard.writeText` etc. never runs. Detect that the trimmed code
+  // is exactly one such expression with no trailing call and wrap it.
+  const trimmed = fixedCode.trim();
+  const startsAsyncArrow = /^async\s*(?:\([^)]*\)|[A-Za-z_$][\w$]*)\s*=>/.test(
+    trimmed
+  );
+  const startsAsyncFn = /^async\s+function\b/.test(trimmed);
+  if ((startsAsyncArrow || startsAsyncFn) && !/\)\s*;?\s*$/.test(trimmed)) {
+    fixedCode = `(${trimmed})()`;
+  }
+
   return fixedCode;
 };
 
@@ -2591,6 +2607,7 @@ You will be given a query and the markdown of a webpage that has been filtered t
     });
 
     type EvaluateAction = z.infer<typeof EvaluateActionSchema>;
+    const evaluateLogger = this.logger;
     this.registry.action(
       'Execute browser JavaScript on the current page and return the result.',
       { param_model: EvaluateActionSchema }
@@ -2606,6 +2623,11 @@ You will be given a query and the markdown of a webpage that has been filtered t
       }
 
       const validatedCode = validateAndFixJavaScript(params.code);
+      if (validatedCode !== params.code) {
+        evaluateLogger.info(
+          `🛠 evaluate: auto-fixed input (len ${params.code.length} → ${validatedCode.length})`
+        );
+      }
 
       const payload = (await page.evaluate(
         async ({ code }: { code: string }) => {
@@ -2671,6 +2693,18 @@ You will be given a query and the markdown of a webpage that has been filtered t
         rendered = `${rendered.slice(0, maxChars - 50)}\n... [Truncated after 20000 characters]`;
       }
 
+      // Log the evaluate return value so callers (and trace consumers) can
+      // see whether the executed code actually produced a result. Truncate
+      // aggressively — the agent already has the full value via
+      // extracted_content; this log is for debugging the "code ran but did
+      // nothing" failure mode (e.g. unwrapped async fn expressions).
+      const previewMax = 500;
+      const preview =
+        rendered.length > previewMax
+          ? `${rendered.slice(0, previewMax)}... [+${rendered.length - previewMax} chars]`
+          : rendered;
+      evaluateLogger.info(`🧪 evaluate result (${rendered.length} chars): ${preview}`);
+
       const maxMemoryChars = 10000;
       const includeExtractedContentOnlyOnce = rendered.length >= maxMemoryChars;
       const longTermMemory = includeExtractedContentOnlyOnce

test/controller.test.ts

@@ -2308,6 +2308,59 @@ describe('Regression Coverage', () => {
     expect(result.error).toBeNull();
   });
 
+  describe('evaluate auto-wraps bare function expressions in IIFE', () => {
+    const buildPage = () => ({
+      evaluate: vi.fn(async (_handler: unknown, args: { code: string }) => ({
+        ok: true,
+        result: args.code,
+      })),
+      url: vi.fn(() => 'https://example.com'),
+    });
+
+    const runEvaluate = async (code: string) => {
+      const controller = new Controller();
+      const page = buildPage();
+      const browserSession = {
+        get_current_page: vi.fn(async () => page),
+      };
+      const result = await controller.registry.execute_action(
+        'evaluate',
+        { code },
+        { browser_session: browserSession as any }
+      );
+      const evaluatedCode = (page.evaluate.mock.calls[0]?.[1] as {
+        code: string;
+      })?.code;
+      return { result, evaluatedCode };
+    };
+
+    it('wraps a bare async arrow expression in an IIFE', async () => {
+      const { result, evaluatedCode } = await runEvaluate(
+        `async () => {\n  return { username: document.querySelector('input[type=email]')?.value };\n}`
+      );
+      expect(result.error).toBeNull();
+      expect(evaluatedCode?.startsWith('(async () => {')).toBe(true);
+      expect(evaluatedCode?.endsWith('})()')).toBe(true);
+    });
+
+    it('wraps a bare async function expression in an IIFE', async () => {
+      const { evaluatedCode } = await runEvaluate(
+        `async function () { return 1; }`
+      );
+      expect(evaluatedCode).toBe(`(async function () { return 1; })()`);
+    });
+
+    it('leaves a non-function expression untouched', async () => {
+      const { evaluatedCode } = await runEvaluate(`document.title`);
+      expect(evaluatedCode).toBe('document.title');
+    });
+
+    it('does not double-wrap an existing IIFE', async () => {
+      const { evaluatedCode } = await runEvaluate(`(async () => 1)()`);
+      expect(evaluatedCode).toBe(`(async () => 1)()`);
+    });
+  });
+
   it('evaluate returns action error on JavaScript failure', async () => {
     const controller = new Controller();
     const page = {