Refactor AdbResponse to introduce a top-level output field in android_env.

android_world/env/adb_utils.py

@@ -203,7 +203,7 @@ def check_ok(response: adb_pb2.AdbResponse, message=None) -> None:
     else:
       raise RuntimeError(
           f'ADB command failed with status {response.status}:'
-          f' {response.generic.output.decode()}.'
+          f' {response.output.decode()}.'
       )
 
 
@@ -238,7 +238,7 @@ def start_activity(
     logging.error('Failed to launch activity: %r', activity)
     return response
 
-  logging.debug('Launch package output %r', response.generic.output)
+  logging.debug('Launch package output %r', response.output)
   return response
 
 
@@ -719,7 +719,7 @@ def close_recents(env: env_interface.AndroidEnvInterface):
   response = issue_generic_request('shell dumpsys activity recents', env)
   if response.status != adb_pb2.AdbResponse.Status.OK:
     return
-  recents_ids = re.findall(r'id=(\d+)', response.generic.output.decode())
+  recents_ids = re.findall(r'id=(\d+)', response.output.decode())
   for recents_id in recents_ids:
     issue_generic_request(['shell', 'am', 'stack', 'remove', recents_id], env)
 
@@ -910,7 +910,7 @@ def get_api_level(env: env_interface.AndroidEnvInterface) -> int:
   )
   if version.status != adb_pb2.AdbResponse.Status.OK:
     raise RuntimeError('Failed to get API level.')
-  return int(version.generic.output)
+  return int(version.output)
 
 
 def _toggle_svc(
@@ -1081,10 +1081,10 @@ def check_airplane_mode(env: env_interface.AndroidEnvInterface) -> bool:
   if response.status != adb_pb2.AdbResponse.Status.OK:
     raise RuntimeError(
         f'ADB command failed with status {response.status}:'
-        f' {response.generic.output.decode()}.'
+        f' {response.output.decode()}.'
     )
 
-  return response.generic.output.decode().replace('\r', '').strip('\n') == '1'
+  return response.output.decode().replace('\r', '').strip('\n') == '1'
 
 
 def extract_broadcast_data(raw_output: str) -> Optional[str]:
@@ -1155,7 +1155,7 @@ def get_clipboard_contents(env: env_interface.AndroidEnvInterface) -> str:
   if res.status != adb_pb2.AdbResponse.Status.OK:
     raise RuntimeError('Failed to get clipboard content.')
 
-  output_str = res.generic.output.decode('utf-8')
+  output_str = res.output.decode('utf-8')
   result = _extract_clipper_output(output_str)
 
   press_back_button(env)
@@ -1221,7 +1221,7 @@ def set_clipboard_contents(
   output_str = issue_generic_request(
       ['shell', 'am', 'broadcast', '-a', 'clipper.set', '-e', 'text', content],
       env,
-  ).generic.output.decode('utf-8')
+  ).output.decode('utf-8')
   _extract_clipper_output(output_str)
   press_back_button(env)
 
@@ -1288,7 +1288,7 @@ def get_call_state(
   adb_args = ['shell', 'dumpsys', 'telephony.registry']
   response = issue_generic_request(adb_args, env, timeout_sec)
 
-  output = response.generic.output.decode('utf-8')
+  output = response.output.decode('utf-8')
   state_match = re.search(r'mCallState=(\d)', output)
 
   state = 'UNKNOWN'
@@ -1540,7 +1540,7 @@ def get_all_settings(env: env_interface.AndroidEnvInterface) -> dict[str, str]:
   settings = {}
   for adb_command in adb_commands:
     response = issue_generic_request(adb_command, env)
-    lines = response.generic.output.decode().split('\n')
+    lines = response.output.decode().split('\n')
     for line in lines:
       if not line:
         continue
@@ -1602,9 +1602,7 @@ def get_screen_size(env: env_interface.AndroidEnvInterface) -> tuple[int, int]:
   """
   adb_command = ['shell', 'wm size']
   adb_response = issue_generic_request(adb_command, env)
-  return _parse_screen_size_response(
-      adb_response.generic.output.decode('utf-8')
-  )
+  return _parse_screen_size_response(adb_response.output.decode('utf-8'))
 
 
 def get_logical_screen_size(
@@ -1627,7 +1625,7 @@ def get_logical_screen_size(
       'shell dumpsys input | grep logicalFrame', env
   )
   if response.status:
-    raw_output = response.generic.output.decode('utf-8')
+    raw_output = response.output.decode('utf-8')
     pattern = r'logicalFrame=\[0, 0, (\d+), (\d+)\]'
     matches = re.findall(pattern, raw_output)
     for m in matches:
@@ -1654,7 +1652,7 @@ def get_physical_frame_boundary(
       'shell dumpsys input | grep physicalFrame', env
   )
   if response.status:
-    raw_output = response.generic.output.decode('utf-8')
+    raw_output = response.output.decode('utf-8')
     pattern = r'physicalFrame=\[(\d+), (\d+), (\d+), (\d+)\]'
     matches = re.findall(pattern, raw_output)
     for m in matches:
@@ -1690,7 +1688,7 @@ def get_orientation(
       'shell dumpsys window | grep mCurrentRotation', env
   )
   if response.status:
-    raw_output = response.generic.output.decode('utf-8')
+    raw_output = response.output.decode('utf-8')
     pattern = r'mCurrentRotation=ROTATION_(\d+)'
     matches = re.findall(pattern, raw_output)
     for m in matches:
@@ -1763,7 +1761,7 @@ def set_root_if_needed(
   """
   response = issue_generic_request(['shell', 'whoami'], env, timeout_sec)
 
-  if response.generic.output.decode('utf-8').strip() == 'root':
+  if response.output.decode('utf-8').strip() == 'root':
     return response
 
   return issue_generic_request(['root'], env, timeout_sec)
@@ -1777,4 +1775,4 @@ def uiautomator_dump(env, timeout_sec: Optional[float] = 30) -> str:
   read_args = 'shell cat /sdcard/window_dump.xml'
   response = issue_generic_request(read_args, env, timeout_sec=timeout_sec)
 
-  return response.generic.output.decode('utf-8')
+  return response.output.decode('utf-8')

android_world/env/adb_utils_test.py

@@ -48,7 +48,7 @@ def test_get_call_state_idle(self):
   mForegroundCallState=0"""
     # Setup
     mock_dumpsys_response = adb_pb2.AdbResponse()
-    mock_dumpsys_response.generic.output = dumpsys_content.encode('utf-8')
+    mock_dumpsys_response.output = dumpsys_content.encode('utf-8')
     self.mock_issue_generic_request.return_value = mock_dumpsys_response
 
     # Act
@@ -65,7 +65,7 @@ def test_get_call_state_ringing(self):
   mForegroundCallState=0"""
     # Setup
     mock_dumpsys_response = adb_pb2.AdbResponse()
-    mock_dumpsys_response.generic.output = dumpsys_content.encode('utf-8')
+    mock_dumpsys_response.output = dumpsys_content.encode('utf-8')
     self.mock_issue_generic_request.return_value = mock_dumpsys_response
 
     # Act
@@ -76,13 +76,13 @@ def test_get_call_state_ringing(self):
 
   def test_call_emulator(self):
     mock_response = adb_pb2.AdbResponse()
-    mock_response.generic.output = b'Success'
+    mock_response.output = b'Success'
     self.mock_issue_generic_request.return_value = mock_response
 
     phone_number = '+123456789'
     result = adb_utils.call_emulator(self.mock_env, phone_number)
 
-    self.assertEqual(result.generic.output.decode(), 'Success')
+    self.assertEqual(result.output.decode(), 'Success')
 
   @mock.patch.object(adb_utils, 'get_call_state', autospec=True)
   def test_end_call_if_active(self, mock_get_call_state):
@@ -97,38 +97,38 @@ def test_clear_android_emulator_call_log(self):
 
   def test_call_phone_number(self):
     mock_response = adb_pb2.AdbResponse()
-    mock_response.generic.output = b'Success'
+    mock_response.output = b'Success'
     self.mock_issue_generic_request.return_value = mock_response
 
     phone_number = '123456789'
     result = adb_utils.call_phone_number(self.mock_env, phone_number)
 
-    self.assertEqual(result.generic.output.decode(), 'Success')
+    self.assertEqual(result.output.decode(), 'Success')
 
   def test_text_emulator(self):
     mock_response = adb_pb2.AdbResponse()
-    mock_response.generic.output = b'Success'
+    mock_response.output = b'Success'
     self.mock_issue_generic_request.return_value = mock_response
 
     phone_number = '+123456789'
     message = 'Hello, world!'
     result = adb_utils.text_emulator(self.mock_env, phone_number, message)
 
-    self.assertEqual(result.generic.output.decode(), 'Success')
+    self.assertEqual(result.output.decode(), 'Success')
 
 
 class AdbSettingsTest(AdbTestSetup):
 
   def test_set_default_app(self):
     mock_response = adb_pb2.AdbResponse()
-    mock_response.generic.output = b'Success'
+    mock_response.output = b'Success'
     self.mock_issue_generic_request.return_value = mock_response
 
     setting_key = 'sms_default_application'
     package_name = 'com.example.app'
     result = adb_utils.set_default_app(self.mock_env, setting_key, package_name)
 
-    self.assertEqual(result.generic.output.decode(), 'Success')
+    self.assertEqual(result.output.decode(), 'Success')
 
   def test_successful_put_operation(self):
     self.mock_env.execute_adb_call.return_value = adb_pb2.AdbResponse()

android_world/task_evals/common_validators/file_validators.py

@@ -217,7 +217,7 @@ def is_successful(self, env: interface.AsyncEnv) -> float:
         ],
         env.controller,
     )
-    file_contents = res.generic.output.decode().replace("\r", "").strip()
+    file_contents = res.output.decode().replace("\r", "").strip()
     match = fuzzy_match_lib.fuzzy_match(file_contents, self.params["text"])
     if not match:
       logging.info("%s does not match %s", file_contents, self.params["text"])

android_world/task_evals/common_validators/file_validators_test.py

@@ -39,7 +39,7 @@ def test_is_successful(self):
     self.mock_check_file_or_folder_exists.return_value = True
 
     mock_response_cat = adb_pb2.AdbResponse()
-    mock_response_cat.generic.output = b"Hello World"
+    mock_response_cat.output = b"Hello World"
     self.mock_issue_generic_request.return_value = mock_response_cat
 
     env = mock.MagicMock()
@@ -68,7 +68,7 @@ def test_is_successful(
       self,
   ):
     response_ls_deleted = adb_pb2.AdbResponse()
-    response_ls_deleted.generic.output = b"another_note.md\n"
+    response_ls_deleted.output = b"another_note.md\n"
     self.mock_check_file_or_folder_exists.side_effect = [
         True,  # File exists.
         False,  # File doesn't exist.
@@ -88,7 +88,7 @@ def test_is_successful(
   def test_is_successful_subfolder(self):
     # Create mock adb response for 'ls' command when note is deleted
     mock_response_ls_deleted = adb_pb2.AdbResponse()
-    mock_response_ls_deleted.generic.output = b"another_note.md\n"
+    mock_response_ls_deleted.output = b"another_note.md\n"
     self.mock_check_file_or_folder_exists.side_effect = [
         True,  # File exists.
         False,  # File doesn't exist.
@@ -109,9 +109,7 @@ def test_is_successful_subfolder(self):
   def test_is_not_successful(self):
     # Create mock adb response for 'ls' command when note still exists
     mock_response_ls_still_exists = adb_pb2.AdbResponse()
-    mock_response_ls_still_exists.generic.output = (
-        b"test_note.md\nanother_note.md\n"
-    )
+    mock_response_ls_still_exists.output = b"test_note.md\nanother_note.md\n"
     self.mock_check_file_or_folder_exists.side_effect = [
         True,  # File exists.
         True,  # File still exists.
@@ -143,7 +141,7 @@ def setUp(self):
   def test_is_successful(self):
     # Create mock adb response for 'ls' command when note is deleted
     mock_response_ls_deleted = adb_pb2.AdbResponse()
-    mock_response_ls_deleted.generic.output = b"another_note.md\n"
+    mock_response_ls_deleted.output = b"another_note.md\n"
 
     self.mock_check_file_or_folder_exists.side_effect = [
         True,  # Source file exists.
@@ -166,9 +164,7 @@ def test_is_successful(self):
   def test_is_not_successful(self):
     # Create mock adb response for 'ls' command when note still exists
     mock_response_ls_still_exists = adb_pb2.AdbResponse()
-    mock_response_ls_still_exists.generic.output = (
-        b"test_note.md\nanother_note.md\n"
-    )
+    mock_response_ls_still_exists.output = b"test_note.md\nanother_note.md\n"
 
     self.mock_check_file_or_folder_exists.side_effect = [
         True,  # Source file exists.

android_world/task_evals/common_validators/sms_validators.py

@@ -77,13 +77,9 @@ def parse_message(row: str) -> dict[str, str]:
 
 def _decode_messages_from_response(response: adb_pb2.AdbResponse) -> list[str]:
   """Decodes the ADB response into a list of messages."""
-  if (
-      response.generic.output.decode()
-      .replace("\r", "")
-      .startswith("No result found.")
-  ):
+  if response.output.decode().replace("\r", "").startswith("No result found."):
     return []
-  messages = response.generic.output.split(b"\nRow:")
+  messages = response.output.split(b"\nRow:")
   for i, m in enumerate(messages):
     if i > 0:
       messages[i] = b"Row:" + m
@@ -229,7 +225,7 @@ def get_android_time(self, env: env_interface.AndroidEnvInterface) -> int:
     adb_output = adb_utils.issue_generic_request(
         ["shell", "date", "+%s"], env
     )  # Fetch UNIX timestamp from Android
-    return int(adb_output.generic.output.strip()) * 1000
+    return int(adb_output.output.strip()) * 1000
 
   def initialize_task(self, env: interface.AsyncEnv) -> None:
     super().initialize_task(env)

android_world/task_evals/common_validators/sms_validators_test.py

@@ -118,15 +118,13 @@ def setUp(self):
   def test_is_successful(self):
     # From shell date +%s
     mock_response_time = adb_pb2.AdbResponse()
-    mock_response_time.generic.output = '{}'.format(
-        str(int(time.time()))
-    ).encode()
+    mock_response_time.output = '{}'.format(str(int(time.time()))).encode()
 
     # Make stale message.
     one_day_s = 24 * 60 * 60
     mock_response_sms0 = adb_pb2.AdbResponse()
     date0_ms = str(int((time.time() - one_day_s) * 1000))
-    mock_response_sms0.generic.output = (
+    mock_response_sms0.output = (
         'Row: 0, address=1234567890, body=Hello World, service_center=NULL,'
         ' date={}'.format(
             date0_ms
@@ -136,7 +134,7 @@ def test_is_successful(self):
     # Successful message.
     mock_response_sms1 = adb_pb2.AdbResponse()
     date1_ms = str(int(time.time() * 1000))
-    mock_response_sms1.generic.output = (
+    mock_response_sms1.output = (
         'Row: 0, address=1234567890, body=Hello World, service_center=NULL,'
         ' date={}'.format(
             date1_ms
@@ -163,15 +161,13 @@ def test_is_successful(self):
   def test_initialize_task_message_already_sent(self):
     # From shell date +%s
     mock_response_time = adb_pb2.AdbResponse()
-    mock_response_time.generic.output = '{}'.format(
-        str(int(time.time()))
-    ).encode()
+    mock_response_time.output = '{}'.format(str(int(time.time()))).encode()
 
     # Make stale message.
     one_s = 1
     mock_response_sms0 = adb_pb2.AdbResponse()
     date0_ms = str(int((time.time() - one_s) * 1000))
-    mock_response_sms0.generic.output = (
+    mock_response_sms0.output = (
         'Row: 0, address=1234567890, body=Hello World, service_center=NULL,'
         ' date={}'.format(
             date0_ms
@@ -181,7 +177,7 @@ def test_initialize_task_message_already_sent(self):
     # Successful message.
     mock_response_sms1 = adb_pb2.AdbResponse()
     date1_ms = str(int(time.time() * 1000))
-    mock_response_sms1.generic.output = (
+    mock_response_sms1.output = (
         'Row: 0, address=1234567890, body=Hello World, service_center=NULL,'
         ' date={}'.format(
             date1_ms