Skip to content

Game Coordinator

Coordinator is the centerpiece of the game orchestration. It provides an interface between the agents and the worlds.

In detail it handles:

  1. World initialiazation
  2. Registration of new agents in the game
  3. Agent-World communication (message verification and forwarding)
  4. Recording (and storing) trajectories of agents (optional)
  5. Detection of episode ends (either by reaching timout or agents reaching their respective goals)
  6. Assigning rewards for each action and at the end of each episode
  7. Removing agents from the game
  8. Registering the GameReset requests and handelling the game resets.

To facilitate the communication the coordinator uses a TCP server to which agents connect. The communication is asynchronous and depends of the

AIDojoCoordinator.coordinator.AgentServer(actions_queue, agent_response_queues, max_connections)

Bases: Protocol

Class used for serving the agents when connecting to the game run by the GameCoordinator.

Attributes:

Name Type Description
actions_queue Queue

Queue for actions from agents.

answers_queues dict

Mapping of agent addresses to their response queues.

max_connections int

Maximum allowed concurrent agent connections.

current_connections int

Current number of connected agents.

logger Logger

Logger for the AgentServer.

Initialize the AgentServer.

Parameters:

Name Type Description Default
actions_queue Queue

Queue for actions from agents.

required
agent_response_queues dict

Mapping of agent addresses to their response queues.

required
max_connections int

Maximum allowed concurrent agent connections.

required
Source code in AIDojoCoordinator/coordinator.py
26
27
28
29
30
31
32
33
34
35
36
37
38
39
def __init__(self, actions_queue, agent_response_queues, max_connections):
    """
    Initialize the AgentServer.

    Args:
        actions_queue (asyncio.Queue): Queue for actions from agents.
        agent_response_queues (dict): Mapping of agent addresses to their response queues.
        max_connections (int): Maximum allowed concurrent agent connections.
    """
    self.actions_queue = actions_queue
    self.answers_queues = agent_response_queues
    self.max_connections = max_connections
    self.current_connections = 0
    self.logger = logging.getLogger("AIDojo-AgentServer")

__call__(reader, writer) async

Allow the server instance to be called as a coroutine.

Parameters:

Name Type Description Default
reader StreamReader

Stream reader for the agent.

required
writer StreamWriter

Stream writer for the agent.

required
Source code in AIDojoCoordinator/coordinator.py
126
127
128
129
130
131
132
133
134
async def __call__(self, reader, writer):
    """
    Allow the server instance to be called as a coroutine.

    Args:
        reader (asyncio.StreamReader): Stream reader for the agent.
        writer (asyncio.StreamWriter): Stream writer for the agent.
    """
    await self.handle_new_agent(reader, writer)

handle_agent_quit(peername) async

Helper function to handle agent disconnection.

Parameters:

Name Type Description Default
peername tuple

The address of the disconnecting agent.

required
Source code in AIDojoCoordinator/coordinator.py
41
42
43
44
45
46
47
48
49
50
51
async def handle_agent_quit(self, peername:tuple):
    """
    Helper function to handle agent disconnection.

    Args:
        peername (tuple): The address of the disconnecting agent.
    """
    # Send a quit message to the Coordinator
    self.logger.info(f"\tHandling agent quit for {peername}.")
    quit_message = Action(ActionType.QuitGame, parameters={}).to_json()
    await self.actions_queue.put((peername, quit_message))

handle_new_agent(reader, writer) async

Handle a new agent connection.

Parameters:

Name Type Description Default
reader StreamReader

Stream reader for the agent.

required
writer StreamWriter

Stream writer for the agent.

required
Source code in AIDojoCoordinator/coordinator.py
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
async def handle_new_agent(self, reader, writer):
    """
    Handle a new agent connection.

    Args:
        reader (asyncio.StreamReader): Stream reader for the agent.
        writer (asyncio.StreamWriter): Stream writer for the agent.
    """
    # get the peername of the writer
    peername = writer.get_extra_info("peername")
    queue_created = False
    try:
        self.logger.info(f"New connection from {peername}")
        # Check if the maximum number of connections has been reached
        if self.current_connections < self.max_connections:
            # increment the count of current connections
            self.current_connections += 1
            self.logger.info(f"New agent connected: {peername}. Current connections: {self.current_connections}")
            # Ensure a queue exists for this agent
            if peername not in self.answers_queues:
                self.answers_queues[peername] = asyncio.Queue(maxsize=2)
                queue_created = True
                self.logger.info(f"Created queue for agent {peername}")
                # Handle the new agent
                while True:
                    # Step 1: Read data from the agent
                    data = await reader.read(ProtocolConfig.BUFFER_SIZE)
                    if not data:
                        self.logger.info(f"Agent {peername} disconnected.")
                        await self.handle_agent_quit(peername)
                        break

                    raw_message = data.decode().strip()
                    self.logger.debug(f"Handler received from {peername}: {raw_message}")

                    # Step 2: Forward the message to the Coordinator
                    await self.actions_queue.put((peername, raw_message))

                    # Step 3: Get a matching response from the answers queue
                    response_queue = self.answers_queues[peername]
                    response = await response_queue.get()
                    self.logger.info(f"Sending response to agent {peername}: {response}")

                    # Step 4: Send the response to the agent
                    response = str(response).encode() + ProtocolConfig.END_OF_MESSAGE
                    writer.write(response)
                    await writer.drain()
            else:
                self.logger.warning(f"Queue for agent {peername} already exists. Closing connection.")
        else:
            self.logger.info(f"Max connections reached. Rejecting new connection from {writer.get_extra_info('peername')}")
    except ConnectionResetError:
        self.logger.warning(f"Connection reset by {peername}")
        await self.handle_agent_quit(peername)
    except asyncio.CancelledError:
        self.logger.debug("Connection handling cancelled.")
        raise  # Ensure the exception propagates
    except Exception as e:
        self.logger.error(f"Unexpected error with client {peername}: {e}")
        raise
    finally:
        try:
            if peername in self.answers_queues:
                # If the queue was created, remove it
                if queue_created:
                    self.answers_queues.pop(peername)
                    self.logger.info(f"Removed queue for agent {peername}")
                self.current_connections = max(0, self.current_connections - 1)
            writer.close()
            await writer.wait_closed()
        except Exception:
            # swallow exceptions on close to avoid crash on cleanup
            pass

AIDojoCoordinator.coordinator.GameCoordinator(game_host, game_port, service_host, service_port, allowed_roles=['Attacker', 'Defender', 'Benign'], task_config_file=None)

Class for creation, and management of agent interactions in AI Dojo.

Source code in AIDojoCoordinator/coordinator.py
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
def __init__(self, game_host: str, game_port: int, service_host:str, service_port:int, allowed_roles=["Attacker", "Defender", "Benign"], task_config_file:str=None) -> None:
    self.host = game_host
    self.port = game_port
    self.logger = logging.getLogger("AIDojo-GameCoordinator")

    self._tasks = set()
    self.shutdown_flag = asyncio.Event()
    self._reset_event = asyncio.Event()
    self._episode_end_event = asyncio.Event()
    self._episode_start_event = asyncio.Event()
    self._episode_rewards_condition = asyncio.Condition()
    self._reset_done_condition = asyncio.Condition()
    self._reset_lock = asyncio.Lock()
    self._agents_lock = asyncio.Lock()

    # for accessing configuration remotely
    self._service_host = service_host
    self._service_port = service_port
    # for reading configuration locally
    self._task_config_file = task_config_file
    self.logger = logging.getLogger("AIDojo-GameCoordinator")
    self.ALLOWED_ROLES = allowed_roles
    self._cyst_objects = None
    self._cyst_object_string = None

    # prepare agent communication
    self._agent_action_queue = asyncio.Queue()
    self._agent_response_queues = {}

    # agent information
    self.agents = {}
    # step counter per agent_addr (int)
    self._agent_steps = {}
    # reset request per agent_addr (bool)
    self._reset_requests = {}
    self._agent_status = {}
    self._episode_ends = {}
    self._agent_observations = {}
    # starting per agent_addr (dict)
    self._agent_starting_position = {}
    # current state per agent_addr (GameState)
    self._agent_states = {}
    # last action played by agent (Action)
    self._agent_last_action = {}
    # False positives per agent (due to added blocks)
    self._agent_false_positives = {}
    # agent status dict {agent_addr: int}
    self._agent_rewards = {}
    # trajectories per agent_addr
    self._agent_trajectories = {}
    # false_positives per agent_addr
    self._agent_false_positives = {}

add_false_positive(agent)

Method for adding false positive to the agent.

Source code in AIDojoCoordinator/coordinator.py
857
858
859
860
861
862
863
864
865
866
def add_false_positive(self, agent:tuple)->None:
    """
    Method for adding false positive to the agent.
    """
    self.logger.debug(f"Adding false positive to {agent}")
    if agent in self._agent_false_positives:
        self._agent_false_positives[agent] += 1
    else:
        self._agent_false_positives[agent] = 1
    self.logger.debug(f"False positives for {agent}: {self._agent_false_positives[agent]}")

convert_msg_dict_to_json(msg_dict)

Helper function to create text-base messge from a dictionary. Used in the Agent-Game communication.

Source code in AIDojoCoordinator/coordinator.py
215
216
217
218
219
220
221
222
223
224
225
226
def convert_msg_dict_to_json(self, msg_dict:dict)->str:
    """
    Helper function to create text-base messge from a dictionary. Used in the Agent-Game communication.
    """
    try:
        # Convert message into string representation
        output_message = json.dumps(msg_dict)
    except Exception as e:
        self.logger.error(f"Error when converting msg to JSON:{e}")
        raise e
        # Send to anwer_queue
    return output_message

create_agent_queue(agent_addr) async

Creates a queue for the given agent address if it doesn't already exist.

Source code in AIDojoCoordinator/coordinator.py
207
208
209
210
211
212
213
async def create_agent_queue(self, agent_addr:tuple)->None:
    """
    Creates a queue for the given agent address if it doesn't already exist.
    """
    if agent_addr not in self._agent_response_queues:
        self._agent_response_queues[agent_addr] = asyncio.Queue()
        self.logger.info(f"Created queue for agent {agent_addr}. {len(self._agent_response_queues)} queues in total.")

goal_check(agent_addr)

Check if the goal conditons were satisfied in a given game state

Source code in AIDojoCoordinator/coordinator.py
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
def goal_check(self, agent_addr:tuple)->bool:
    """
    Check if the goal conditons were satisfied in a given game state
    """
    def goal_dict_satistfied(goal_dict:dict, known_dict: dict)-> bool:
        """
        Helper function for checking if a goal dictionary condition is satisfied
        """
        # check if we have all IPs that should have some values (are keys in goal_dict)
        if goal_dict.keys() <= known_dict.keys():
            try:
                # Check if values (sets) for EACH key (host) in goal_dict are subsets of known_dict, keep matching_keys
                matching_keys = [host for host in goal_dict.keys() if goal_dict[host]<= known_dict[host]]
                # Check we have the amount of mathing keys as in the goal_dict
                if len(matching_keys) == len(goal_dict.keys()):
                    return True
            except KeyError:
                # some keys are missing in the known_dict
                return False
        return False
    self.logger.debug(f"Checking goal for agent {agent_addr}.")
    goal_conditions = self._win_conditions_per_role[self.agents[agent_addr][1]]
    state = self._agent_states[agent_addr]
    # For each part of the state of the game, check if the conditions are met
    goal_reached = {}    
    goal_reached["networks"] = set(goal_conditions["known_networks"]) <= set(state.known_networks)
    goal_reached["known_hosts"] = set(goal_conditions["known_hosts"]) <= set(state.known_hosts)
    goal_reached["controlled_hosts"] = set(goal_conditions["controlled_hosts"]) <= set(state.controlled_hosts)
    goal_reached["services"] = goal_dict_satistfied(goal_conditions["known_services"], state.known_services)
    goal_reached["data"] = goal_dict_satistfied(goal_conditions["known_data"], state.known_data)
    goal_reached["known_blocks"] = goal_dict_satistfied(goal_conditions["known_blocks"], state.known_blocks)
    self.logger.debug(f"\t{goal_reached}")
    return all(goal_reached.values())

is_agent_benign(agent_addr)

Check if the agent is benign (defender, normal)

Source code in AIDojoCoordinator/coordinator.py
938
939
940
941
942
943
944
def is_agent_benign(self, agent_addr:tuple)->bool:
    """
    Check if the agent is benign (defender, normal)
    """
    if agent_addr not in self.agents:
        return False
    return self.agents[agent_addr][1].lower() in ["defender", "benign"]

register_agent(agent_id, agent_role, agent_initial_view) async

Domain specific method of the environment. Creates the initial state of the agent.

Source code in AIDojoCoordinator/coordinator.py
755
756
757
758
759
async def register_agent(self, agent_id:tuple, agent_role:str, agent_initial_view:dict)->GameState:
    """
    Domain specific method of the environment. Creates the initial state of the agent.
    """
    raise NotImplementedError

remove_agent(agent_id, agent_state) async

Domain specific method of the environment. Creates the initial state of the agent.

Source code in AIDojoCoordinator/coordinator.py
761
762
763
764
765
async def remove_agent(self, agent_id:tuple, agent_state:GameState)->bool:
    """
    Domain specific method of the environment. Creates the initial state of the agent.
    """
    raise NotImplementedError

run()

Wrapper for ayncio run function. Starts all tasks in AIDojo

Source code in AIDojoCoordinator/coordinator.py
228
229
230
231
232
233
234
235
236
237
def run(self)->None:
    """
    Wrapper for ayncio run function. Starts all tasks in AIDojo
    """
    try:
        asyncio.run(self.start_tasks())
    except Exception as e:
        self.logger.error(f"Unexpected error: {e}")
    finally:
        self.logger.info(f"{__class__.__name__} has exited.")

run_game() async

Task responsible for reading messages from the agent queue and processing them based on the ActionType.

Source code in AIDojoCoordinator/coordinator.py
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
async def run_game(self):
    """
    Task responsible for reading messages from the agent queue and processing them based on the ActionType.
    """
    while not self.shutdown_flag.is_set():
        # Read message from the queue
        agent_addr, message = await self._agent_action_queue.get()
        if message is not None:
            self.logger.info(f"Coordinator received from agent {agent_addr}: {message}.")
            try:  # Convert message to Action
                action = Action.from_json(message)
                self.logger.debug(f"\tConverted to: {action}.")
            except Exception as e:
                self.logger.error(
                    f"Error when converting msg to Action using Action.from_json():{e}, {message}"
                )
            match action.type:  # process action based on its type
                case ActionType.JoinGame:
                    self.logger.debug(f"About agent {agent_addr}. Start processing of ActionType.JoinGame by {agent_addr}")
                    self.logger.debug(f"{action.type}, {action.type.value}, {action.type == ActionType.JoinGame}")
                    self._spawn_task(self._process_join_game_action, agent_addr, action)
                case ActionType.QuitGame:
                    self.logger.debug(f"About agent {agent_addr}. Start processing of ActionType.QuitGame by {agent_addr}")
                    self._spawn_task(self._process_quit_game_action, agent_addr)
                case ActionType.ResetGame:
                    self.logger.debug(f"About agent {agent_addr}. Start processing of ActionType.ResetGame by {agent_addr}")
                    self._spawn_task(self._process_reset_game_action, agent_addr, action)
                case ActionType.ExfiltrateData | ActionType.FindData | ActionType.ScanNetwork | ActionType.FindServices | ActionType.ExploitService:
                    self.logger.debug(f"About agent {agent_addr}. Start processing of {action.type} by {agent_addr}")
                    self._spawn_task(self._process_game_action, agent_addr, action)
                case ActionType.BlockIP:
                    self.logger.debug(f"About agent {agent_addr}. Start processing of {action.type} by {agent_addr}")
                    self._spawn_task(self._process_game_action, agent_addr, action)
                case _:
                    self.logger.warning(f"About agent {agent_addr}. Unsupported action type: {action}!")
    self.logger.info("\tAction processing task stopped.")

shutdown_signal_handler() async

Handle shutdown signals.

Source code in AIDojoCoordinator/coordinator.py
202
203
204
205
async def shutdown_signal_handler(self):
    """Handle shutdown signals."""
    self.logger.info("Shutdown signal received. Setting shutdown flag.")
    self.shutdown_flag.set()

start_tasks() async

High level funciton to start all the other asynchronous tasks. - Reads the conf of the coordinator - Creates queues - Start the main part of the coordinator - Start a server that listens for agents

Source code in AIDojoCoordinator/coordinator.py
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
async def start_tasks(self):
    """
    High level funciton to start all the other asynchronous tasks.
    - Reads the conf of the coordinator
    - Creates queues
    - Start the main part of the coordinator
    - Start a server that listens for agents
    """
    loop = asyncio.get_running_loop()

    # Set up signal handlers for graceful shutdown
    loop.add_signal_handler(
        signal.SIGINT, lambda: asyncio.create_task(self.shutdown_signal_handler())
    )
    loop.add_signal_handler(
        signal.SIGTERM, lambda: asyncio.create_task(self.shutdown_signal_handler())
    )


    # initialize the game objects
    if self._service_host: #get the task config using REST API
        self.logger.info(f"Fetching task configuration from {self._service_host}:{self._service_port}")
        await self._fetch_initialization_objects()
    elif self._task_config_file: # load task config locally from a file
        self.logger.info(f"Loading task configuration from file: {self._task_config_file}")
        self._load_initialization_objects()
    else:
        raise ValueError("Task configuration not specified")


    # Read configuration
    self._starting_positions_per_role = self._get_starting_position_per_role()
    self._win_conditions_per_role = self._get_win_condition_per_role()
    self._goal_description_per_role = self._get_goal_description_per_role()
    self._steps_limit_per_role = self._get_max_steps_per_role()
    self.logger.debug(f"Timeouts set to:{self._steps_limit_per_role}")
    if self.task_config.get_use_global_defender():
        self._global_defender = GlobalDefender()
    else:
        self._global_defender = None
    self._use_dynamic_ips = self.task_config.get_use_dynamic_addresses()
    self.logger.info(f"Change IP every episode set to: {self._use_dynamic_ips}")
    self._rewards = self.task_config.get_rewards(["step", "success", "fail", "false_positive"])
    self.logger.info(f"Rewards set to:{self._rewards}")
    self._min_required_players = self.task_config.get_required_num_players()
    self.logger.info(f"Min player requirement set to:{self._min_required_players}")
    # start server for agent communication
    self._spawn_task(self.start_tcp_server)

    # start episode rewards task
    self._spawn_task(self._assign_rewards_episode_end)

    # start episode rewards task
    self._spawn_task(self._reset_game)

    # start action processing task
    self._spawn_task(self.run_game)

    while not self.shutdown_flag.is_set():
        # just wait until user terminates
        await asyncio.sleep(1)
    self.logger.debug("Final cleanup started")
    # make sure there are no running tasks left
    for task in self._tasks:
        task.cancel()  # Cancel each active task
    await asyncio.gather(*self._tasks, return_exceptions=True)  # Wait for all tasks to finish
    self.logger.info("All tasks shut down.")

start_tcp_server() async

Starts TPC sever for the agent communication.

Source code in AIDojoCoordinator/coordinator.py
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
async def start_tcp_server(self):
    """
    Starts TPC sever for the agent communication.
    """
    try:
        self.logger.info("Starting the server listening for agents")
        server = await asyncio.start_server(
            AgentServer(
                self._agent_action_queue,
                self._agent_response_queues,
                max_connections=self._min_required_players
            ),
            self.host,
            self.port
        )
        addrs = ", ".join(str(sock.getsockname()) for sock in server.sockets)
        self.logger.info(f"\tServing on {addrs}")
        while not self.shutdown_flag.is_set():
            await asyncio.sleep(1)
    except asyncio.CancelledError:
        self.logger.debug("\tStopping TCP server task.")
    except Exception as e:
        self.logger.error(f"TCP server failed: {e}")
    finally:
        server.close()
        await server.wait_closed()
        self.logger.info("\tTCP server task stopped")