diff --git a/ui/desktop/tests/integration/test_providers_code_exec.test.ts b/ui/desktop/tests/integration/test_providers_code_exec.test.ts
index f1d8c0c904bc..5db51778eab7 100644
--- a/ui/desktop/tests/integration/test_providers_code_exec.test.ts
+++ b/ui/desktop/tests/integration/test_providers_code_exec.test.ts
@@ -22,6 +22,13 @@ beforeAll(() => {
 
 const { testAll } = providerTest(discoverTestCases({ skipAgentic: true }));
 
+// Matches: "execute_typescript | code_execution", "get_function_details | code_execution",
+//           "tool call | execute", "tool calls | execute" (old format)
+//           "▸ execute N tool call" (new format with tool_graph)
+//           "▸ execute_typescript" (plain tool name in output)
+const codeExecPattern =
+  /(execute_typescript \| code_execution)|(get_function_details \| code_execution)|(tool calls? \| execute)|(▸.*execute.*tool call)|(▸ execute_typescript)/;
+
 testAll('invokes code_execution tool', async (tc, { expect }) => {
   const testdir = fs.mkdtempSync(path.join(os.tmpdir(), 'goose-codeexec-'));
   try {
@@ -30,16 +37,11 @@ testAll('invokes code_execution tool', async (tc, { expect }) => {
       testdir,
       "Store a memory with category 'test' and data 'hello world', then retrieve all memories from category 'test'.",
       BUILTINS,
-      { GOOSE_PROVIDER: tc.provider, GOOSE_MODEL: tc.model }
+      { GOOSE_PROVIDER: tc.provider, GOOSE_MODEL: tc.model },
+      55_000,
+      (output) => codeExecPattern.test(output)
     );
 
-    // Matches: "execute_typescript | code_execution", "get_function_details | code_execution",
-    //           "tool call | execute", "tool calls | execute" (old format)
-    //           "▸ execute N tool call" (new format with tool_graph)
-    //           "▸ execute_typescript" (plain tool name in output)
-    const codeExecPattern =
-      /(execute_typescript \| code_execution)|(get_function_details \| code_execution)|(tool calls? \| execute)|(▸.*execute.*tool call)|(▸ execute_typescript)/;
-
     expect(
       codeExecPattern.test(output),
       `Expected code_execution tool to be called\n\nFull output:\n${output}`