def _world_function():
world_generator = utils.get_world_fn_attr(
self._world_module, overworld_name, "generate_world"
)
overworld = world_generator(self.opt, [overworld_agent])
while not overworld.episode_done() and not self.system_done:
world_type = overworld.parley()
if world_type is None:
time.sleep(0.5)
continue
if world_type == self.manager.EXIT_STR:
self.manager._remove_agent(overworld_agent.id)
return world_type
else:
continue
# perform onboarding
if onboard_map is not None:
onboard_type = onboard_map.get(world_type)
if onboard_type:
onboard_id = 'onboard-{}-{}'.format(overworld_agent.id, time.time())
agent = self.manager._create_agent(onboard_id, overworld_agent.id)
agent.data = overworld_agent.data
agent_state.set_active_agent(agent)
agent_state.assign_agent_to_task(agent, onboard_id)
_, onboard_data = self._run_world(task, onboard_type, [agent])
agent_state.onboard_data = onboard_data
agent_state.data = agent.data
self.manager.add_agent_to_pool(agent_state, world_type)
log_utils.print_and_log(logging.INFO, 'onboarding/overworld complete')
return world_type
def on_disconnect(*args):
"""
Disconnect event is a no-op for us, as the server reconnects automatically
on a retry.
"""
log_utils.print_and_log(
logging.INFO, 'World server disconnected: {}'.format(args))
self.alive = False
self._ensure_closed()
def _done_callback(fut):
e = fut.exception()
if e is not None:
log_utils.print_and_log(
logging.ERROR,
'World {} had error {}'.format(task_id, repr(e)),
should_print=True,
)
traceback.print_exc(file=sys.stdout)
if self.debug:
raise e
def on_error(ws, error):
try:
if error.errno == errno.ECONNREFUSED:
self._ensure_closed()
self.use_socket = False
raise Exception("Socket refused connection, cancelling")
else:
log_utils.print_and_log(
logging.WARN,
'Socket logged error: {}'.format(repr(error)))
except BaseException:
if type(error) is websocket.WebSocketConnectionClosedException:
return # Connection closed is noop
log_utils.print_and_log(
logging.WARN,
'Socket logged error: {} Restarting'.format(repr(error)),
)
self._ensure_closed()
def on_message(*args):
"""
Incoming message handler for messages from the FB user.
"""
packet_dict = json.loads(args[1])
if packet_dict['type'] == 'conn_success':
self.alive = True
return # No action for successful connection
if packet_dict['type'] == 'pong':
self.last_pong = time.time()
return # No further action for pongs
message_data = packet_dict['content']
log_utils.print_and_log(
logging.DEBUG,
'Message data received: {}'.format(message_data))
for message_packet in message_data['entry']:
for message in message_packet['messaging']:
self.message_callback(message)
def shutdown(self):
"""
Handle any client shutdown cleanup.
"""
try:
self.is_running = False
self.world_runner.shutdown()
if not self.bypass_server_setup:
self.socket.keep_running = False
self._expire_all_conversations()
except BaseException as e:
log_utils.print_and_log(logging.ERROR,
f'world ended in error: {e}')
finally:
if not self.bypass_server_setup:
server_utils.delete_server(self.server_task_name,
self.opt['local'])
def run_socket(*args):
url_base_name = self.server_url.split('https://')[1]
while self.keep_running:
try:
sock_addr = "wss://{}/".format(url_base_name)
self.ws = websocket.WebSocketApp(
sock_addr,
on_message=on_message,
on_error=on_error,
on_close=on_disconnect,
)
self.ws.on_open = on_socket_open
self.ws.run_forever(ping_interval=1, ping_timeout=0.9)
except Exception as e:
log_utils.print_and_log(
logging.WARN,
'Socket error {}, attempting restart'.format(repr(e)),
)
time.sleep(0.2)
def _done_callback(fut):
"""
Log and raise exception of task world, if there is one.
Additionally, set active agent to overworld agent.
"""
e = fut.exception()
if e is not None:
log_utils.print_and_log(
logging.ERROR,
'World {} had error {}'.format(world_type, repr(e)),
should_print=True,
)
traceback.print_exc(file=sys.stdout)
for agent in agents:
self.observe_message(
agent.id,
'Sorry, this world closed. Returning to overworld.')
else:
log_utils.print_and_log(
logging.INFO,
'World {} had no error'.format(world_type),
should_print=True,
)
self.active_worlds[task_id] = None
for agent in agents:
self.after_agent_removed(agent.id)
agent_state = self.get_agent_state(agent.id)
agent_state.data = agent.data
next_task = agent.data.get("next_task")
log_utils.print_and_log(logging.INFO,
"Next task: {}".format(next_task))
if next_task is None:
self._launch_overworld(agent.id)
overworld_agent = agent_state.get_overworld_agent()
overworld_agent.data = agent_state.data
agent_state.set_active_agent(overworld_agent)
elif next_task == self.EXIT_STR:
self._remove_agent(agent.id)
else:
self.add_agent_to_pool(agent_state, next_task)
def on_socket_open(*args):
log_utils.print_and_log(logging.DEBUG,
'Socket open: {}'.format(args))
self._send_world_alive()
def start_task(self):
"""
Begin handling task.
Periodically check to see when enough agents are in the agent pool to start an
instance of the task. Continue doing this until the desired number of
conversations is had.
"""
self.running = True
print("Starting task")
while self.running:
# Loop forever until the server is shut down
with self.agent_pool_change_condition:
valid_pools = self._get_unique_pool()
for world_type, agent_pool in valid_pools.items():
# check if agent has exceeded max time in pool
world_config = self.task_configs[world_type]
if world_config.max_time_in_pool is not None:
self.check_timeout_in_pool(
world_type,
agent_pool,
world_config.max_time_in_pool,
world_config.backup_task,
)
needed_agents = self.max_agents_for[world_type]
if len(agent_pool) >= needed_agents:
log_utils.print_and_log(logging.INFO,
'starting pool',
should_print=True)
# enough agents in pool to start new conversation
self.conversation_index += 1
task_id = 't_{}'.format(self.conversation_index)
# Add the required number of valid agents to the conv
agent_states = [w for w in agent_pool[:needed_agents]]
agents = []
for state in agent_states:
agent = self._create_agent(task_id, state.get_id())
agent.onboard_data = state.onboard_data
agent.data = state.data
state.assign_agent_to_task(agent, task_id)
state.set_active_agent(agent)
agents.append(agent)
# reset wait message state
state.stored_data['seen_wait_message'] = False
assign_role_function = utils.get_assign_roles_fn(
self.world_module, self.taskworld_map[world_type])
if assign_role_function is None:
assign_role_function = utils.default_assign_roles_fn
assign_role_function(agents)
# Allow task creator to filter out workers and run
# versions of the task that require fewer agents
agents = [a for a in agents if a.disp_id is not None]
for a in agents:
# Remove selected workers from the agent pool
self.remove_agent_from_pool(
self.get_agent_state(a.id),
world_type=world_type,
mark_removed=False,
)
for a in agents:
partner_list = agents.copy()
partner_list.remove(a)
a.message_partners = partner_list
done_callback = self._get_done_callback_for_agents(
task_id, world_type, agents)
# launch task world.
future = self.world_runner.launch_task_world(
task_id, self.taskworld_map[world_type], agents)
future.add_done_callback(done_callback)
self.active_worlds[task_id] = future
time.sleep(utils.THREAD_MEDIUM_SLEEP)
def _log_debug(self, text: str):
time = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
log_utils.print_and_log(logging.DEBUG,
f'{time}: {text}',
should_print=True)
def _manager_loop_fn(self):
"""
An iteration of the manager's main loop to launch worlds.
"""
with self.agent_pool_change_condition:
valid_pools = self._get_unique_pool()
for world_type, agent_pool in valid_pools.items():
# check if agent has exceeded max time in pool
world_config = self.task_configs[world_type]
if world_config.max_time_in_pool is not None:
self.check_timeout_in_pool(
world_type,
agent_pool,
world_config.max_time_in_pool,
world_config.backup_task,
)
needed_agents = self.max_agents_for[world_type]
if len(agent_pool) >= needed_agents:
log_utils.print_and_log(
logging.INFO, 'starting pool', should_print=True
)
# enough agents in pool to start new conversation
self.conversation_index += 1
task_id = 't_{}'.format(self.conversation_index)
# Add the required number of valid agents to the conv
agent_states = [w for w in agent_pool[:needed_agents]]
agents = []
for state in agent_states:
agent = self._create_agent(task_id, state.get_id())
agent.onboard_data = state.onboard_data
agent.data = state.data
state.assign_agent_to_task(agent, task_id)
state.set_active_agent(agent)
agents.append(agent)
# reset wait message state
state.stored_data['seen_wait_message'] = False
assign_role_function = utils.get_assign_roles_fn(
self.world_module, self.taskworld_map[world_type]
)
if assign_role_function is None:
assign_role_function = utils.default_assign_roles_fn
assign_role_function(agents)
# Allow task creator to filter out workers and run
# versions of the task that require fewer agents
for a in agents:
# Remove selected workers from the agent pool
self.remove_agent_from_pool(
self.get_agent_state(a.id),
world_type=world_type,
mark_removed=False,
)
for a in agents:
partner_list = agents.copy()
partner_list.remove(a)
a.message_partners = partner_list
done_callback = self._get_done_callback_for_agents(
task_id, world_type, agents
)
# launch task world.
future = self.world_runner.launch_task_world(
task_id, self.taskworld_map[world_type], agents
)
future.add_done_callback(done_callback)
self.active_worlds[task_id] = future
def _world_fn():
log_utils.print_and_log(
logging.INFO, 'Starting task {}...'.format(task_name)
)
return self._run_world(task, world_name, agents)
