Add files using upload-large-folder tool

src_code_for_reproducibility/docs/source/conf.py

@@ -0,0 +1,48 @@
+# Configuration file for the Sphinx documentation builder.
+import os
+import sys
+sys.path.insert(0, os.path.abspath('../..'))
+
+# -- Project information -----------------------------------------------------
+project = 'llm_negotiation'
+copyright = '2023, Your Name'
+author = 'Your Name'
+
+# -- General configuration ---------------------------------------------------
+extensions = [
+    'sphinx.ext.autodoc',
+    'sphinx.ext.viewcode',
+    'sphinx.ext.napoleon',
+    'sphinx.ext.autosummary',
+    'sphinx.ext.intersphinx',
+    'sphinx.ext.mathjax',
+    'sphinxcontrib.mermaid',
+    'sphinx_rtd_theme',
+]
+
+templates_path = ['_templates']
+exclude_patterns = []
+
+# -- Options for HTML output -------------------------------------------------
+html_theme = 'sphinx_rtd_theme'
+html_static_path = ['_static']
+
+# -- Napoleon settings -------------------------------------------------------
+napoleon_google_docstring = True
+napoleon_numpy_docstring = False
+napoleon_include_init_with_doc = True
+napoleon_include_private_with_doc = False
+napoleon_include_special_with_doc = True
+napoleon_use_admonition_for_examples = False
+napoleon_use_admonition_for_notes = False
+napoleon_use_admonition_for_references = False
+napoleon_use_ivar = False
+napoleon_use_param = True
+napoleon_use_rtype = True
+napoleon_preprocess_types = False
+napoleon_type_aliases = None
+napoleon_attr_annotations = True
+
+# -- Path setup --------------------------------------------------------------
+# Make sure the project's modules can be found by Sphinx
+sys.path.insert(0, os.path.abspath('../../src'))

src_code_for_reproducibility/docs/source/index.rst

@@ -0,0 +1,22 @@
+Welcome to LLM Negotiation's documentation!
+===========================================
+This library is a collection of tools for training and evaluating LLM-based agents in multi-agent environments. It is designed to be easy to use and extend. 
+
+.. toctree::
+   :maxdepth: 3
+   :caption: Contents:
+
+   installation
+   marl_standard
+   environments
+   launch
+   usage
+   modules
+   contributing
+
+Indices and tables
+==================
+
+* :ref:`genindex`
+* :ref:`modindex`
+* :ref:`search`

src_code_for_reproducibility/docs/source/src.environments.dond.rst

@@ -0,0 +1,19 @@
+src.environments.dond package
+=============================
+
+.. automodule:: src.environments.dond
+   :members:
+   :undoc-members:
+   :show-inheritance:
+
+Submodules
+----------
+
+.. toctree::
+   :maxdepth: 4
+
+   src.environments.dond.dond_agent
+   src.environments.dond.dond_game
+   src.environments.dond.dond_log_funcs
+   src.environments.dond.dond_statistics_funcs
+   src.environments.dond.dond_training_data_funcs

src_code_for_reproducibility/docs/source/src.utils.rst

@@ -0,0 +1,24 @@
+src.utils package
+=================
+
+.. automodule:: src.utils
+   :members:
+   :undoc-members:
+   :show-inheritance:
+
+Submodules
+----------
+
+.. toctree::
+   :maxdepth: 4
+
+   src.utils.common_imports
+   src.utils.export_ppo_training_set
+   src.utils.extra_stats
+   src.utils.inherit_args
+   src.utils.log_gpu_usage
+   src.utils.log_statistics
+   src.utils.model_to_cpu
+   src.utils.parallel_shuffle
+   src.utils.quick_stats
+   src.utils.update_start_epoch

src_code_for_reproducibility/markov_games/diplomacy/diplomacy_agent.py

@@ -0,0 +1,259 @@
+from typing import Dict, List, Tuple, Optional, Any
+import copy
+
+class DiplomacyAgent:
+    """Agent handler for Diplomacy game that follows the MARL standard.
+    
+    This class is responsible for parsing LLM output into valid Diplomacy orders,
+    managing the agent state, and providing information for logging.
+    """
+    
+    def __init__(self, policy_id: str, power_name: str, random_valid_move=False):
+        """Initialize the agent handler for a power in the Diplomacy game.
+        
+        Args:
+            power_name: The name of the power this agent controls (e.g., 'FRANCE', 'ENGLAND')
+            policy_id: The identifier for the policy this agent uses
+            random_valid_move: If True, will select random valid moves instead of using LLM (default: False)
+        """
+        self.policy_id = policy_id
+        self.power_name = power_name
+        self.orders = []
+        self.wait = True
+        self.processing_state = "WAITING_FOR_ORDERS"
+        self.parsed_orders = []
+        self.order_status = {}
+        self.message_history = []
+        self.random_valid_move = random_valid_move
+        
+    def step(self, observation_from_env, policy_output=None):
+        """Update the agent state based on the observation and LLM output.
+        
+        Args:
+            observation_from_env: The observation from the environment
+            policy_output: The output from the LLM
+            
+        Returns:
+            policy_id: The policy identifier
+            policy_input: The input to the policy
+            action: The official action to be sent to the environment
+            done: Whether the LLM action is ready to be sent to the environment
+            info: Additional information about the agent
+        """
+        info = {}
+        
+        # If random_valid_move is enabled, select random valid moves
+        if self.random_valid_move:
+            valid_orders = self._select_random_valid_moves(observation_from_env)
+            self.orders = valid_orders
+            self.wait = False
+            action = {
+                "orders": valid_orders,
+                "wait": False
+            }
+            return self.policy_id, {}, action, True, info
+        
+        # If no policy output, this is the initial step - prepare prompt
+        if policy_output is None:
+            # Create initial prompt for the LLM
+            phase = observation_from_env.get('phase', '')
+            units = observation_from_env.get('units', {}).get(self.power_name, [])
+            centers = observation_from_env.get('centers', {}).get(self.power_name, [])
+            orderable_locations = observation_from_env.get('orderable_locations', {})
+            
+            prompt = self._create_prompt(phase, units, centers, orderable_locations)
+            
+            return self.policy_id, {"prompt": prompt}, None, False, info
+            
+        # Process the LLM output to extract orders
+        success, parsed_orders = self._parse_llm_output(policy_output)
+        self.parsed_orders = parsed_orders
+        
+        if not success:
+            # Need more information from LLM
+            clarification_prompt = self._create_clarification_prompt(policy_output, parsed_orders)
+            return self.policy_id, {"prompt": clarification_prompt}, None, False, info
+        
+        # Validate if the orders are valid for the current phase
+        valid_orders = self._validate_orders(parsed_orders, observation_from_env)
+        
+        if valid_orders:
+            # Orders are valid, prepare action for environment
+            self.orders = valid_orders
+            self.wait = False
+            action = {
+                "orders": valid_orders,
+                "wait": False
+            }
+            return self.policy_id, {}, action, True, info
+        else:
+            # Orders are invalid, ask for new ones
+            error_prompt = self._create_error_prompt(parsed_orders, observation_from_env)
+            return self.policy_id, {"prompt": error_prompt}, None, False, info
+    
+    def _create_prompt(self, phase, units, centers, orderable_locations):
+        """Create the initial prompt for the LLM.
+        
+        Args:
+            phase: The current game phase
+            units: List of units controlled by this power
+            centers: List of supply centers controlled by this power
+            orderable_locations: List of locations where orders can be issued
+            
+        Returns:
+            A prompt string for the LLM
+        """
+        prompt = f"You are playing as {self.power_name} in Diplomacy. The current phase is {phase}.\n\n"
+        prompt += f"Your units: {', '.join(units)}\n"
+        prompt += f"Your supply centers: {', '.join(centers)}\n"
+        prompt += f"Locations you can order: {', '.join(orderable_locations)}\n\n"
+        
+        if phase.endswith('M'):  # Movement phase
+            prompt += "Please provide orders for your units in the form:\n"
+            prompt += "- A LON H (hold)\n"
+            prompt += "- F NTH - NWY (move)\n"
+            prompt += "- A WAL S F LON (support)\n"
+            prompt += "- F NWG C A NWY - EDI (convoy)\n"
+        elif phase.endswith('R'):  # Retreat phase
+            prompt += "Please provide retreat orders for your dislodged units:\n"
+            prompt += "- A PAR R MAR (retreat to MAR)\n"
+            prompt += "- A PAR D (disband)\n"
+        elif phase.endswith('A'):  # Adjustment phase
+            if len(units) < len(centers):
+                prompt += "You can build units. Please provide build orders:\n"
+                prompt += "- A PAR B (build army in PAR)\n"
+                prompt += "- F BRE B (build fleet in BRE)\n"
+                prompt += "- WAIVE (waive a build)\n"
+            elif len(units) > len(centers):
+                prompt += "You must remove units. Please provide disbandment orders:\n"
+                prompt += "- A PAR D (disband army in PAR)\n"
+                prompt += "- F BRE D (disband fleet in BRE)\n"
+        
+        prompt += "\nProvide your orders as a list, one per line."
+        return prompt
+    
+    def _parse_llm_output(self, llm_output):
+        """Parse the LLM output to extract orders.
+        
+        Args:
+            llm_output: The raw output from the LLM
+            
+        Returns:
+            success: Whether parsing was successful
+            parsed_orders: List of parsed orders
+        """
+        # Simple parsing for now - extract lines that look like orders
+        lines = llm_output.strip().split('\n')
+        orders = []
+        
+        for line in lines:
+            # Remove list markers, hyphens, etc.
+            line = line.strip('- *•').strip()
+            
+            # Skip empty lines and lines that don't look like orders
+            if not line or line.startswith('I ') or line.startswith('Let\'s'):
+                continue
+                
+            # Check if it looks like a Diplomacy order
+            if (' H' in line or ' -' in line or ' S ' in line or ' C ' in line or 
+                ' R ' in line or ' D' in line or ' B' in line or line == 'WAIVE'):
+                orders.append(line)
+        
+        return len(orders) > 0, orders
+    
+    def _validate_orders(self, orders, observation):
+        """Validate if the orders are valid for the current phase.
+        
+        Args:
+            orders: List of orders to validate
+            observation: Current observation from the environment
+            
+        Returns:
+            List of valid orders or None if invalid
+        """
+        # For simplicity, we'll assume all parsed orders are valid
+        # In a real implementation, we would use the game's validation logic
+        return orders
+    
+    def _create_clarification_prompt(self, previous_output, parsed_orders):
+        """Create a prompt asking for clarification when orders couldn't be parsed.
+        
+        Args:
+            previous_output: The previous LLM output
+            parsed_orders: Any orders that were successfully parsed
+            
+        Returns:
+            A prompt string for the LLM
+        """
+        prompt = f"I couldn't fully understand your orders for {self.power_name}. "
+        
+        if parsed_orders:
+            prompt += f"I understood these orders:\n"
+            for order in parsed_orders:
+                prompt += f"- {order}\n"
+                
+        prompt += "\nPlease provide clear, valid Diplomacy orders in the format:\n"
+        prompt += "- A LON H\n- F NTH - NWY\n- etc.\n"
+        return prompt
+    
+    def _create_error_prompt(self, invalid_orders, observation):
+        """Create a prompt when orders are invalid.
+        
+        Args:
+            invalid_orders: The invalid orders
+            observation: Current observation from the environment
+            
+        Returns:
+            A prompt string for the LLM
+        """
+        prompt = f"The following orders for {self.power_name} are invalid:\n"
+        for order in invalid_orders:
+            prompt += f"- {order}\n"
+            
+        prompt += "\nPlease provide valid orders for your units."
+        return prompt
+    
+    def get_log_info(self):
+        """Get information about the agent required to log a trajectory.
+        
+        Returns:
+            log_info: Information about the agent required to log a trajectory.
+        """
+        return {
+            "power_name": self.power_name,
+            "orders": self.orders,
+            "wait": self.wait,
+            "parsing_state": self.processing_state,
+            "message_history": self.message_history
+        }
+    
+    def render(self):
+        """Render the current state of the agent."""
+        print(f"Power: {self.power_name}")
+        print(f"Orders: {self.orders}")
+        print(f"Wait: {self.wait}")
+    
+    def close(self):
+        """Perform any necessary cleanup."""
+        pass
+
+    def _select_random_valid_moves(self, observation):
+        """Select random valid moves for all units.
+        
+        Args:
+            observation: Current observation from the environment
+            
+        Returns:
+            List of valid orders
+        """
+        import random
+        
+        possible_orders = observation.get('possible_orders', {})
+        valid_orders = []
+        
+        # For each location with possible orders, select one randomly
+        for location, orders in possible_orders.items():
+            if orders:  # If there are any possible orders for this location
+                valid_orders.append(random.choice(orders))
+        
+        return valid_orders

src_code_for_reproducibility/markov_games/diplomacy/diplomacy_env.py

@@ -0,0 +1,230 @@
+from typing import Dict, List, Tuple, Optional, Any
+from diplomacy import Game
+import random
+
+class DiplomacyEnv:
+    """Multi-Agent Reinforcement Learning environment for Diplomacy.
+    
+    This class wraps the Diplomacy game engine to provide an interface
+    compliant with the MARL standard.
+    """
+    
+    def __init__(self, random_seed=None, map_name="standard", game_id=None, rules=None, max_steps=50):
+        """Initialize the Diplomacy environment.
+        
+        Args:
+            map_name: The name of the map to use (default: "standard")
+            game_id: Optional game ID
+            rules: Optional rules to apply to the game
+            max_steps: Maximum number of steps before forcing game end (default: 10)
+        """
+        self.random_seed = random_seed
+        self.map_name = map_name
+        self.game_id = game_id
+        self.rules = rules or []
+        self.game = None
+        self.active_powers = []
+        self.render_mode = None
+        self.max_steps = max_steps
+        self.current_steps = 0
+    
+    def reset(self):
+        """Reset the environment to an initial state and return the initial observation.
+        
+        Returns:
+            observation: A dictionary where keys are agent identifiers and values are observations.
+        """
+        # Initialize a new game
+        self.game = Game(game_id=self.game_id, map_name=self.map_name)
+        
+        # Apply rules
+        for rule in self.rules:
+            self.game.add_rule(rule)
+        
+        # Determine active powers (not eliminated)
+        self.active_powers = [name for name, power in self.game.powers.items() 
+                             if not power.is_eliminated()]
+        
+        # Reset step counter
+        self.current_steps = 0
+        
+        # Create initial observations for all powers
+        observations = {}
+        for power_name in self.active_powers:
+            observations[power_name] = self._create_observation(power_name)
+        
+        return observations
+    
+    def step(self, actions):
+        """Take a step in the environment using the provided actions.
+        
+        Args:
+            actions: A dictionary where keys are agent identifiers and values are actions.
+            
+        Returns:
+            observations: A dictionary where keys are agent identifiers and values are observations.
+            done: Whether the episode has ended.
+            info: Additional information about the environment.
+        """
+        print(f"stepping {self.current_steps}")
+        self.current_steps += 1
+        # Apply actions (orders) for each power
+        for power_name, action in actions.items():
+            if power_name in self.active_powers:
+                orders = action.get("orders", [])
+                wait = action.get("wait", True)
+                
+                # Set orders for the power
+                if orders:
+                    self.game.set_orders(power_name, orders)
+                
+                # Set wait flag
+                self.game.set_wait(power_name, wait)
+        
+        # Check if all active powers are ready to proceed
+        if self.game.does_not_wait():
+            # Process the current phase
+            self.game.process()
+            
+            
+            # Update active powers list after processing
+            self.active_powers = [name for name, power in self.game.powers.items() 
+                                 if not power.is_eliminated()]
+        
+        # Create observations for all active powers
+        observations = {}
+        for power_name in self.active_powers:
+            observations[power_name] = self._create_observation(power_name)
+        
+        # Check if the game is done (either naturally or due to max steps)
+        done = self.game.is_game_done or self.current_steps >= self.max_steps
+        
+        # Create info dict
+        info = {
+            "phase": self.game.get_current_phase(),
+            "active_powers": self.active_powers,
+            "centers": self.game.get_centers(),
+            "units": self.game.get_units(),
+            "current_steps": self.current_steps,
+            "max_steps_reached": self.current_steps >= self.max_steps
+        }
+        
+        return observations, done, info
+    
+    def _create_observation(self, power_name):
+        """Create observation for a specific power.
+        
+        Args:
+            power_name: The name of the power
+            
+        Returns:
+            An observation dictionary
+        """
+        observation = {
+            "phase": self.game.get_current_phase(),
+            "units": self.game.get_units(),
+            "centers": self.game.get_centers(),
+            "orderable_locations": self.game.get_orderable_locations(power_name),
+            "order_status": self.game.get_order_status(power_name),
+            "possible_orders": self._get_possible_orders_for_power(power_name)
+        }
+        return observation
+    
+    def _get_possible_orders_for_power(self, power_name):
+        """Get all possible orders for a power's units.
+        
+        Args:
+            power_name: The name of the power
+            
+        Returns:
+            A dictionary mapping units to their possible orders
+        """
+        all_possible_orders = self.game.get_all_possible_orders()
+        
+        # Filter for only the locations where this power has units
+        power_units = self.game.get_units(power_name)
+        power_unit_locations = [unit[2:] for unit in power_units]
+        
+        # For retreat phases, include retreating units
+        if self.game.phase_type == 'R':
+            power = self.game.get_power(power_name)
+            power_unit_locations.extend([unit[2:] for unit in power.retreats])
+        
+        # For adjustment phases, include buildable locations
+        elif self.game.phase_type == 'A':
+            power = self.game.get_power(power_name)
+            # If we have more centers than units, we can build
+            if len(power.centers) > len(power.units):
+                buildable_sites = self.game._build_sites(power)
+                power_unit_locations.extend(buildable_sites)
+            # If we have more units than centers, we need to remove
+            elif len(power.units) > len(power.centers):
+                # All units are candidates for removal
+                pass
+        
+        # Filter the possible orders to only those for this power's units/locations
+        power_possible_orders = {}
+        for loc, orders in all_possible_orders.items():
+            if loc[:3] in power_unit_locations:
+                power_possible_orders[loc] = orders
+        
+        return power_possible_orders
+    
+    def get_log_info(self):
+        """Get additional information about the environment for logging.
+        
+        Returns:
+            log_info: Information about the environment required to log the game.
+        """
+        if not self.game:
+            return {}
+            
+        return {
+            "game_id": self.game.game_id,
+            "phase": self.game.get_current_phase(),
+            "map_name": self.game.map_name,
+            "centers": self.game.get_centers(),
+            "units": self.game.get_units(),
+            "powers": {name: {
+                "units": power.units,
+                "centers": power.centers,
+                "is_eliminated": power.is_eliminated(),
+                "order_status": self.game.get_order_status(name)
+            } for name, power in self.game.powers.items()},
+            "orders": self.game.get_orders(),
+            "active_powers": self.active_powers,
+            "is_game_done": self.game.is_game_done,
+            "outcome": self.game.outcome if self.game.is_game_done else None
+        }
+    
+    def render(self, mode='human'):
+        """Render the current state of the environment.
+        
+        Args:
+            mode: The rendering mode ('human', 'svg', etc.)
+        
+        Returns:
+            The rendered image if applicable
+        """
+        self.render_mode = mode
+        if self.game:
+            if mode == 'human':
+                # Just print basic game state
+                print(f"Game: {self.game.game_id}")
+                print(f"Phase: {self.game.get_current_phase()}")
+                print(f"Active Powers: {self.active_powers}")
+                print("Supply Centers:")
+                for power_name, centers in self.game.get_centers().items():
+                    print(f"  {power_name}: {centers}")
+                print("Units:")
+                for power_name, units in self.game.get_units().items():
+                    print(f"  {power_name}: {units}")
+                return None
+            elif mode == 'svg':
+                # Return SVG representation
+                return self.game.render(output_format='svg')
+        return None
+    
+    def close(self):
+        """Perform any necessary cleanup."""
+        self.game = None

src_code_for_reproducibility/markov_games/diplomacy/diplomacy_logging.py

@@ -0,0 +1,360 @@
+import os
+import json
+from utils.common_imports import *
+
+
+
+def diplomacy_log_match(
+        path,
+        agents_log_info,
+        env_log_info,
+        metrics_func=None,
+        metrics_func_args=None
+        ):
+    """
+    Logs the Diplomacy game data and generates HTML visualizations using the get_log_info methods.
+    
+    Args:
+        path (str): Base path to save the data.
+        agents_log_info (list): List of agent information dictionaries containing the get_log_info results.
+        env_log_info (dict): Environment information from its get_log_info method.
+        metrics_func (str, optional): Name of the function to calculate metrics.
+        metrics_func_args (dict, optional): Arguments for the metrics function.
+    """
+    # Create directory structure
+    os.makedirs(path, exist_ok=True)
+    
+    # Save the environment log info
+    env_log_path = os.path.join(path, "env_log.json")
+    with open(env_log_path, "w") as f:
+        json.dump(env_log_info, f, indent=4, default=_json_serialize)
+    
+    # Process each agent's log info
+    for agent_log in agents_log_info:
+        power_name = agent_log["power_name"]
+        
+        # Define paths for raw data and statistics subfolders
+        power_path = os.path.join(path, power_name)
+        raw_data_path = os.path.join(power_path, "raw_data")
+        statistics_path = os.path.join(power_path, "statistics")
+        
+        # Ensure directories exist
+        os.makedirs(raw_data_path, exist_ok=True)
+        os.makedirs(statistics_path, exist_ok=True)
+        
+        # Determine the next available file number for raw data
+        raw_files = os.listdir(raw_data_path)
+        raw_numbers = [int(f.split('_')[-1].split('.')[0]) for f in raw_files if f.startswith("log_")]
+        next_raw_number = max(raw_numbers, default=0) + 1
+        raw_file = os.path.join(raw_data_path, f"log_{next_raw_number}.json")
+        
+        # Save agent log info
+        with open(raw_file, "w") as f:
+            json.dump(agent_log, f, indent=4, default=_json_serialize)
+        
+        # Log metrics if a metrics function is provided
+        if metrics_func:
+            metrics_files = os.listdir(statistics_path)
+            metrics_numbers = [int(f.split('_')[-1].split('.')[0]) for f in metrics_files if f.startswith("metrics_")]
+            next_metrics_number = max(metrics_numbers, default=0) + 1
+            metrics_file = os.path.join(statistics_path, f"metrics_{next_metrics_number}.json")
+            
+            metrics = globals()[metrics_func](agent_log, info, **metrics_func_args)
+            with open(metrics_file, "w") as f:
+                json.dump(metrics, f, indent=4)
+    
+    # Generate the HTML visualization
+    html_content = generate_diplomacy_html(agents_log_info, env_log_info)
+    
+    # Ensure the html directory exists
+    html_path = os.path.join(path, "html")
+    os.makedirs(html_path, exist_ok=True)
+    
+    # Determine the next available file number for HTML
+    html_files = os.listdir(html_path)
+    html_numbers = [int(f.split('_')[-1].split('.')[0]) for f in html_files if f.startswith("game_summary_")]
+    next_html_number = max(html_numbers, default=0) + 1
+    html_file = os.path.join(html_path, f"game_summary_{next_html_number}.html")
+    
+    # Save the HTML content to a file
+    with open(html_file, "w") as f:
+        f.write(html_content)
+
+def generate_diplomacy_html(agent_infos, env_info):
+    """
+    Generate HTML visualization for a Diplomacy game.
+    
+    Args:
+        agent_infos (list): List of agent information dictionaries from get_log_info.
+        env_info (dict): Environment information from get_log_info.
+        
+    Returns:
+        str: HTML content for the game visualization.
+    """
+    # Extract game information
+    game_id = env_info.get("game_id", "Unknown")
+    phase = env_info.get("phase", "Unknown")
+    map_name = env_info.get("map_name", "standard")
+    is_game_done = env_info.get("is_game_done", False)
+    outcome = env_info.get("outcome", [])
+    
+    centers = env_info.get("centers", {})
+    units = env_info.get("units", {})
+    
+    # HTML head and style
+    html_content = """
+    <!DOCTYPE html>
+    <html lang="en">
+    <head>
+        <meta charset="UTF-8">
+        <meta name="viewport" content="width=device-width, initial-scale=1.0">
+        <title>Diplomacy Game {game_id}</title>
+        <style>
+            body {{
+                font-family: 'Arial', sans-serif;
+                background-color: #f5f5f5;
+                color: #333333;
+                margin: 0;
+                padding: 20px;
+            }}
+            .container {{
+                display: grid;
+                grid-template-columns: repeat(3, 1fr);
+                grid-gap: 20px;
+                margin-bottom: 30px;
+            }}
+            .central-info {{
+                grid-column: span 3;
+                background: #fff;
+                padding: 20px;
+                border-radius: 10px;
+                box-shadow: 0 4px 12px rgba(0, 0, 0, 0.1);
+                margin-bottom: 20px;
+            }}
+            .power-column {{
+                background: #fff;
+                padding: 15px;
+                border-radius: 10px;
+                box-shadow: 0 4px 12px rgba(0, 0, 0, 0.1);
+            }}
+            .message {{
+                margin-bottom: 15px;
+                padding: 12px;
+                border-radius: 8px;
+                box-shadow: 0 1px 4px rgba(0, 0, 0, 0.1);
+            }}
+            .user {{
+                background: rgba(235, 245, 255, 0.8);
+                border-left: 4px solid #007bff;
+            }}
+            .assistant {{
+                background: rgba(240, 255, 240, 0.8);
+                border-right: 4px solid #28a745;
+            }}
+            .orders {{
+                background: rgba(255, 248, 225, 0.8);
+                border-left: 4px solid #ffc107;
+            }}
+            .role {{
+                font-weight: bold;
+                margin-bottom: 5px;
+                color: #333333;
+            }}
+            .power-name {{
+                text-align: center;
+                font-size: 1.4em;
+                margin-bottom: 15px;
+                color: #000;
+                font-weight: 600;
+                text-transform: uppercase;
+                letter-spacing: 1px;
+            }}
+            .game-info {{
+                display: grid;
+                grid-template-columns: repeat(2, 1fr);
+                grid-gap: 15px;
+            }}
+            .info-card {{
+                background: #f9f9f9;
+                padding: 15px;
+                border-radius: 8px;
+                box-shadow: 0 1px 3px rgba(0, 0, 0, 0.1);
+            }}
+            .supply-centers, .units-list {{
+                display: flex;
+                flex-wrap: wrap;
+                justify-content: space-between;
+            }}
+            .supply-center, .unit {{
+                flex: 0 0 30%;
+                margin-bottom: 10px;
+                padding: 8px;
+                background: #f0f0f0;
+                border-radius: 5px;
+                text-align: center;
+            }}
+            h2 {{
+                border-bottom: 2px solid #eee;
+                padding-bottom: 10px;
+                margin-top: 0;
+            }}
+            .outcome {{
+                background: #e8f5e9;
+                padding: 15px;
+                border-radius: 8px;
+                margin-top: 15px;
+                font-weight: bold;
+                text-align: center;
+            }}
+            .austria {{ border-top: 5px solid #ff5050; }}
+            .england {{ border-top: 5px solid #5050ff; }}
+            .france {{ border-top: 5px solid #50c0ff; }}
+            .germany {{ border-top: 5px solid #808080; }}
+            .italy {{ border-top: 5px solid #50ff50; }}
+            .russia {{ border-top: 5px solid #ffffff; border: 1px solid #ccc; }}
+            .turkey {{ border-top: 5px solid #c0c000; }}
+        </style>
+    </head>
+    <body>
+        <div class="central-info">
+            <h2>Game Information</h2>
+            <div class="game-info">
+                <div class="info-card">
+                    <h3>Game Details</h3>
+                    <p><strong>Game ID:</strong> {game_id}</p>
+                    <p><strong>Phase:</strong> {phase}</p>
+                    <p><strong>Map:</strong> {map_name}</p>
+                    <p><strong>Status:</strong> {status}</p>
+                </div>
+                <div class="info-card">
+                    <h3>Supply Centers</h3>
+                    <div class="supply-centers">
+    """.format(
+        game_id=game_id,
+        phase=phase,
+        map_name=map_name,
+        status="Completed" if is_game_done else "Active"
+    )
+
+    # Add supply center information
+    for power, power_centers in centers.items():
+        html_content += f"""
+                        <div class="supply-center">
+                            <strong>{power}:</strong> {len(power_centers)}
+                        </div>
+        """
+
+    html_content += """
+                    </div>
+                </div>
+            </div>
+    """
+
+    # Add outcome if game is done
+    if is_game_done and outcome:
+        winners = outcome[1:] if len(outcome) > 1 else ["Draw"]
+        html_content += f"""
+            <div class="outcome">
+                <h3>Game Outcome</h3>
+                <p>Winners: {', '.join(winners)}</p>
+            </div>
+        """
+
+    html_content += """
+        </div>
+        <div class="container">
+    """
+
+    # Add each power's information
+    for agent_log in agent_infos:
+        power_name = agent_log["power_name"]
+        power_class = power_name.lower()
+        orders = agent_log.get("orders", [])
+        message_history = agent_log.get("message_history", [])
+        
+        html_content += f"""
+            <div class="power-column {power_class}">
+                <div class="power-name">{power_name}</div>
+                
+                <div class="info-card">
+                    <h3>Units</h3>
+                    <ul>
+        """
+        
+        # Add units information
+        power_units = units.get(power_name, [])
+        for unit in power_units:
+            html_content += f"<li>{unit}</li>"
+        
+        html_content += """
+                    </ul>
+                </div>
+                
+                <div class="message orders">
+                    <div class="role">Final Orders</div>
+                    <ul>
+        """
+        
+        # Add orders
+        for order in orders:
+            html_content += f"<li>{order}</li>"
+        
+        html_content += """
+                    </ul>
+                </div>
+        """
+        
+        # Add message history
+        for message in message_history:
+            if isinstance(message, dict):
+                # Skip system messages or handle differently
+                if message.get("role") == "system":
+                    continue
+                
+                role = message.get("role", "unknown")
+                content = message.get("content", "")
+                
+                role_class = "user" if role == "user" else "assistant"
+                role_display = "Environment" if role == "user" else f"LLM ({power_name})"
+                
+                # Escape HTML characters in content
+                content = content.replace("<", "&lt;").replace(">", "&gt;").replace("\n", "<br>")
+                
+                html_content += f"""
+                <div class="message {role_class}">
+                    <div class="role">{role_display}</div>
+                    <p>{content}</p>
+                </div>
+                """
+            elif isinstance(message, str):
+                # Simple string messages (may be used in some implementations)
+                html_content += f"""
+                <div class="message">
+                    <p>{message}</p>
+                </div>
+                """
+        
+        html_content += """
+            </div>
+        """
+
+    html_content += """
+        </div>
+    </body>
+    </html>
+    """
+    
+    return html_content
+
+def _json_serialize(obj):
+    """
+    A helper function to convert non-JSON-serializable objects
+    (like OrderResult) into strings or dicts.
+    """
+    # Check for the specific object types you know are problematic
+    if obj.__class__.__name__ == "OrderResult":
+        # Return a string representation or a dict
+        return str(obj)
+    
+    # Fallback: attempt to convert anything else to string
+    return str(obj)

src_code_for_reproducibility/markov_games/ipd/__init__.py

@@ -0,0 +1,7 @@
+from .Ipd_hard_coded_agents import AlwaysCooperateIPDAgent, AlwaysDefectIPDAgent
+
+__all__ = [
+	"AlwaysCooperateIPDAgent",
+	"AlwaysDefectIPDAgent",
+]
+

src_code_for_reproducibility/markov_games/ipd/ipd_agent.py

@@ -0,0 +1,115 @@
+import copy
+import json
+import random
+import re
+from collections.abc import Callable
+from copy import deepcopy
+from dataclasses import dataclass, field
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+from mllm.markov_games.agent import Agent
+from mllm.markov_games.rollout_tree import AgentActLog, ChatTurn
+
+
+@dataclass
+class IPDAgentState:
+    """
+    TOWRITE
+    """
+
+    nb_retries: int
+    round_nb: int
+    chat_counter: int
+    chat_history: List[ChatTurn]
+
+
+@dataclass
+class IPDAgent(Agent):
+    seed: int
+    agent_id: str
+    agent_name: str
+    policy: Callable[[List[Dict]], str]
+    intro_prompt: str  # Introduction prompt explaining the game rules
+    goal_prompt: str  # Prompt explaining the agent's goal
+    strategy_prompt: str  # Prompt suggesting a strategy to the agent
+    max_errors: int  # Maximum number of errors allowed before default action
+    allow_reasoning: bool  # Whether to allow reasoning in the response
+    max_reasoning_chars: int  # Maximum number of characters for reasoning
+    cooperate_string: str  # string parsed as playing cooperate by simulation
+    defect_string: str  # string parsed as playing defect by simulation
+
+    def __post_init__(self):
+        self.state = IPDAgentState(
+            nb_retries=0, round_nb=0, chat_counter=0, chat_history=[]
+        )
+
+    async def act(self, observation) -> Tuple[Any, AgentActLog]:
+        """
+        TOWRITE
+        """
+
+        action = None
+        action_is_ready = False
+        round_nb = observation.round_nb
+
+        # If it's the first round, we need to send the intro prompt
+        if round_nb == 0 and self.state.chat_counter == 0:
+            self.state.chat_history.append(
+                ChatTurn(
+                    agent_id=self.agent_id,
+                    role="user",
+                    content=self.intro_prompt,
+                    is_state_end=True,
+                )
+            )
+
+        # If new round
+        if round_nb > self.state.round_nb:
+            coagent_action = observation.last_coagent_move
+            user_message = f"Last round, the other agent played {coagent_action}."
+            self.state.chat_history.append(
+                ChatTurn(
+                    agent_id=self.agent_id,
+                    role="user",
+                    content=user_message,
+                    is_state_end=True,
+                )
+            )
+
+        # If not new round, try to get valid action from policy
+        output_chat_turn: ChatTurn = await self.policy(
+            state=self.state.chat_history,
+            agent_id=self.agent_id,
+            regex=f"({self.cooperate_string}|{self.defect_string})",
+        )
+        self.state.chat_history.append(output_chat_turn)
+        action = output_chat_turn.content
+
+        agent_step_log = AgentActLog(
+            chat_turns=self.state.chat_history[self.state.chat_counter :], info=None
+        )
+        self.state.chat_counter = len(self.state.chat_history)
+        self.state.round_nb = round_nb
+
+        return action, agent_step_log
+
+    def get_safe_copy(self):
+        """
+        Return a safe copy of the agent.
+        """
+        agent_copy = copy.copy(self)
+        agent_copy.state = copy.deepcopy(self.state)
+        return agent_copy
+
+    def reset(self):
+        self.state = IPDAgentState()
+        raise NotImplementedError
+
+    def render(self):
+        pass
+
+    def close(self):
+        pass
+
+    def get_agent_info(self):
+        pass

src_code_for_reproducibility/markov_games/ipd/ipd_simulation.py

@@ -0,0 +1,162 @@
+import copy
+import random
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple
+
+import numpy as np
+
+from mllm.markov_games.markov_game import Simulation
+from mllm.markov_games.rollout_tree import SimulationStepLog
+from mllm.utils.get_coagent_id import get_coagent_id
+
+
+@dataclass
+class IPDState:
+    """
+    State of the Iterated Prisoner's Dilemma game.
+    """
+
+    round_nb: int = 0
+    done: bool = False
+    last_moves: Dict[str, str] | None = None
+
+
+@dataclass
+class IPDObs:
+    """
+    Observation in Iterated Prisoner's Dilemma game.
+    """
+
+    round_nb: int
+    last_coagent_move: str | None
+
+
+class IPD(Simulation):
+    """
+    Iterated Prisoner's Dilemma simulation following the standard.
+
+    In each round of the game, two agents simultaneously choose to either cooperate (C) or defect (D).
+    The payoffs are as follows:
+    - If both cooperate: Both receive the "reward" (usually 3 points)
+    - If both defect: Both receive the "punishment" (usually 1 point)
+    - If one cooperates and one defects: The defector receives the "temptation" (usually 5 points)
+      and the cooperator receives the "sucker" payoff (usually 0 points)
+
+    The game is played for a specified number of rounds.
+    """
+
+    def __init__(
+        self,
+        agent_ids: List[str],
+        agent_names: List[str],
+        seed: int,
+        rounds_per_game: int,
+        reward: float,  # Both cooperate
+        punishment: float,  # Both defect
+        temptation: float,  # Defector's reward when other cooperates
+        sucker: float,  # Cooperator's reward when other defects
+        cooperate_actions: List[str],
+        defect_actions: List[str],
+    ):
+        self.agent_ids = agent_ids
+        self.agent_names = agent_names
+        self.seed = seed
+        self.rounds_per_game = rounds_per_game
+        self.reward = reward
+        self.punishment = punishment
+        self.temptation = temptation
+        self.sucker = sucker
+        self.cooperate_actions = cooperate_actions
+        self.defect_actions = defect_actions
+        self.state = IPDState()
+
+    def step(self, actions: Dict[str, str]) -> Tuple[bool, SimulationStepLog]:
+        """
+        Take a step in the environment using the provided actions.
+        Here, the observations are just the states of the game.
+
+        Args:
+            actions (dict): A dictionary where keys are agent identifiers and values are actions ('C' or 'D').
+
+        Returns:
+            observations (dict): A dictionary where keys are agent identifiers and values are observations.
+            done (bool): Whether the episode has ended.
+            info (dict): Additional information about the environment.
+        """
+
+        # Calculate rewards using payoff matrix
+        agent0_action = actions[self.agent_ids[0]]
+        agent1_action = actions[self.agent_ids[1]]
+
+        # Normalize actions to standard cooperate/defect/gibberish format
+        def normalize_action(action):
+            if action in self.cooperate_actions:
+                return "C"
+            elif action in self.defect_actions:
+                return "D"
+            else:
+                return "D"
+
+        norm_action0 = normalize_action(agent0_action)
+        norm_action1 = normalize_action(agent1_action)
+
+        payoffs = {
+            ("C", "C"): [self.reward, self.reward],
+            ("C", "D"): [self.sucker, self.temptation],
+            ("D", "C"): [self.temptation, self.sucker],
+            ("D", "D"): [self.punishment, self.punishment],
+        }
+
+        round_rewards = {
+            self.agent_ids[0]: payoffs[(norm_action0, norm_action1)][0],
+            self.agent_ids[1]: payoffs[(norm_action0, norm_action1)][1],
+        }
+
+        # Update game state
+        self.state.round_nb += 1
+        self.state.last_moves = copy.deepcopy(actions)
+        done = self.state.round_nb >= self.rounds_per_game
+        step_log = SimulationStepLog(
+            rewards=round_rewards,
+            info={
+                "actions": {
+                    self.agent_ids[0]: norm_action0,
+                    self.agent_ids[1]: norm_action1,
+                }
+            },
+        )
+
+        return done, step_log
+
+    def get_obs(self):
+        """Returns all agent observations in dict
+        Returns:
+            observations
+        """
+        observations = {}
+        for agent_id in self.agent_ids:
+            observations[agent_id] = self.get_obs_agent(agent_id)
+        return observations
+
+    def get_obs_agent(self, agent_id):
+        """Returns observation for agent_id"""
+        if self.state.last_moves != None:
+            other_id = get_coagent_id(self.agent_ids, agent_id)
+            last_coagent_move = self.state.last_moves[other_id]
+        else:
+            last_coagent_move = None
+        obs = IPDObs(round_nb=self.state.round_nb, last_coagent_move=last_coagent_move)
+        return obs
+
+    def reset(self):
+        """Returns initial observations and states"""
+        self.state = IPDState()
+        return self.get_obs()
+
+    def get_safe_copy(self):
+        """
+        Return a safe copy of the simulation.
+        """
+        simulation_copy = copy.copy(self)
+        simulation_copy.state = copy.deepcopy(self.state)
+        return simulation_copy

src_code_for_reproducibility/markov_games/ipd/ipd_statistics.py

@@ -0,0 +1,18 @@
+from __future__ import annotations
+
+from typing import Dict, Callable, List, Tuple
+
+from mllm.markov_games.rollout_tree import SimulationStepLog
+
+
+def avg_reward(sl: SimulationStepLog) -> List[Tuple[str, float]]:
+    for aid in sl.rewards.keys():
+        if "buffer" in str(aid) and "live" not in str(aid):
+            return None
+    # One value per agent at each step
+    rewards_dict = {f"reward-{aid}": float(v) for aid, v in (sl.rewards or {}).items()}
+    return [(key, value) for key, value in rewards_dict.items() if value is not None]
+
+stat_functs: list[Callable[[SimulationStepLog], List[Tuple[str, float]]]] = [
+    avg_reward,
+]

src_code_for_reproducibility/markov_games/negotiation/README.md

@@ -0,0 +1,40 @@
+## Negotiation Games: core mechanics and variants
+
+This family of games feature two agents who, in each round, may briefly communicate and then simultaneously propose how to split a fixed resource (most commonly 10 coins). Rewards are the amount kept multiplied by an agent’s per-unit value. The starting speaker alternates deterministically across rounds.
+
+Communication is optional and variant-dependent: some settings encourage rich messaging to share private information, while others remove messaging entirely to focus on allocation behavior.
+
+Proportional splitting is used when the two proposals exceed the available total: allocations are scaled proportionally rather than discarded. This preserves a useful learning signal even when agents over-claim.
+
+### Variants (in increasing difficulty) 
+
+- No‑Press Split 
+  - Single item type (coins)
+  - No communication; agents go straight to making split proposals, with the starting player alternating deterministically.
+  - Motivation: mirrors no‑communication setups (e.g., Advantage Alignment) while keeping the split decision nontrivial.
+  - Deterministic Mode: values are fixed and public: one agent values coins at 10, the other at 1 (alternates each round).
+  - Stochastic Mode: values are random and uncorrelated. 
+
+- Trust-and-Split RPS (TAS-RPS)
+  - Single item type (coins)
+  - Each round, a rock–paper–scissors hand draw creates a strong asymmetry: the winner’s per-coin value is 10, the loser’s is 1.
+  - Each agent initially sees only their own hand and must communicate to coordinate an optimal split.
+  - Motivation: enforce large value disparity so one’s own value reveals little about the other’s (avoiding ceiling effects) and incentivize meaningful communication.
+ 
+- Trust-and-Split (TAS)
+  - Single item type (coins); each round, each agent’s per-coin value is independently sampled in a broad range (e.g., 1–20).
+  - Each agent observes only their own value; they may use short messages to share and negotiate.
+  - Motivation: a simple blend that tests whether agents learn to exchange private information and coordinate proportional, value-aware splits.
+
+- Deal-or-No-Deal (DOND)
+  - Introduced in [Deal or No Deal? End-to-End Learning for Negotiation Dialogues](https://arxiv.org/pdf/1706.05125)
+  - Multiple item types (typically "books", "hats" and "balls") with limited stocks; each agent has its own per-type values.
+  - A deal pays out only if both proposals exactly agree and respect the stock; otherwise no deal (zero reward) that round.
+  - Motivation: a known benchmark closer to real-world bargaining, where both parties must explicitly agree.
+
+
+
+
+
+
+

src_code_for_reproducibility/markov_games/negotiation/dond_agent.py

@@ -0,0 +1,61 @@
+import copy
+import re
+from collections.abc import Callable
+from dataclasses import dataclass
+from typing import Any, Dict, List, Tuple
+
+from mllm.markov_games.agent import Agent
+from mllm.markov_games.rollout_tree import AgentActLog, ChatTurn
+from mllm.markov_games.negotiation.dond_simulation import (
+    DealNoDealObs,
+)
+from mllm.markov_games.negotiation.nego_simulation import Split
+from mllm.markov_games.negotiation.nego_agent import NegotiationAgent, NegotiationAgentState
+
+class DealNoDealAgent(NegotiationAgent):
+    def __init__(
+        self,
+        *args,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        self.intro_prompt = (
+                "You are {agent_id}. You are playing an iterated game. "
+                "At each round, you and other agent will try to distribute among yourselves items of types {item_types}. "
+                "You only know how much you value each item type, but not the other agent's values. "
+                "You can communicate with the other agent by sending up to {quota_messages_per_agent_per_round} short messages per round. "
+                "Each round, after exchanging messages, you and the other agent will submit a private proposal. "
+                "A deal is accepted only if both proposals match exactly and are within stock; otherwise no deal (0 points for both at that round). "
+                "The values of the items of the other agent at the previous round are revealed to you after each round. "
+                "Your goal is: {goal}."
+            )
+        self.new_round_prompt = ("New round {round_nb}. Items: {stock}. Your values: {values}. ")
+        self.last_round_prompt = ("Last round, other agent's values: {previous_values_coagent}. ")
+        self.send_split_prompt = ("Respond with <split>...</split> where you propose how many items of each type you want to keep.")
+        
+    def get_message_regex(self, observation: DealNoDealObs) -> str:
+        return r"<message>[\s\S]{0,400}</message>"
+    
+    def get_split_regex(self, observation: DealNoDealObs) -> str:
+        parts = []
+        for t in observation.item_types:
+            s = int(observation.quantities.get(t, 0))
+            allowed = "|".join(str(k) for k in range(0, s + 1))
+            rng = f"({allowed})"
+            parts.append(fr"<{t}>{rng}</{t}>")
+        items_block = "".join(parts)
+        return fr"(<split>{items_block}</split>)"
+    
+    def get_split_action(self, policy_output: str, observation: DealNoDealObs) -> Split:
+        import re as _re
+        allocations: Dict[str, int] = {}
+        for t in observation.item_types:
+            m = _re.search(fr"<{t}>([0-9]+)</{t}>", policy_output)
+            if m:
+                allocations[t] = int(m.group(1))
+            else:
+                allocations[t] = 0
+        return Split(items_given_to_self=allocations)
+ 
+
+

src_code_for_reproducibility/markov_games/negotiation/nego_simulation.py

@@ -0,0 +1,241 @@
+"""
+Negotiation simulation environment
+other agent is set at the start of every round. Even though current agent changes over message turns in a round.
+"""
+import copy
+from abc import abstractmethod
+from dataclasses import dataclass
+from typing import Any, Dict, List, Tuple
+
+from numpy.random import default_rng
+
+from mllm.markov_games.rollout_tree import SimulationStepLog
+from mllm.markov_games.simulation import Simulation
+from mllm.utils.get_coagent_id import get_coagent_id
+
+AgentId = str
+
+
+@dataclass
+class Split:
+    items_given_to_self: Dict[str, int]
+
+
+@dataclass
+class Message:
+    message: str
+
+
+@dataclass  # gets extended by variants
+class NegotiationState:
+    round_nb: int
+    last_message: str
+    current_agent: AgentId
+    quantities: Dict[str, int]
+    values: Dict[AgentId, Dict[str, float]]
+    splits: Dict[AgentId, Split | None]
+    nb_messages_sent: Dict[AgentId, int]
+    previous_values: Dict[AgentId, Dict[str, float]] | None
+    previous_splits: Dict[AgentId, Dict[str, int] | None] | None
+    previous_points: Dict[AgentId, float] | None
+    previous_quantities: Dict[str, int] | None
+    split_phase: bool
+
+
+@dataclass  # gets extended by variants
+class NegotiationObs:
+    round_nb: int
+    last_message: str
+    quota_messages_per_agent_per_round: int
+    current_agent: AgentId
+    other_agent: str
+    quantities: Dict[str, int]
+    item_types: List[str]
+    value: Dict[str, int]
+    split_phase: bool
+    last_split_agent: Dict[str, int] | None
+    last_value_agent: Dict[str, int] | None
+    last_points_agent: float | None
+    last_split_coagent: Dict[str, int] | None
+    last_value_coagent: Dict[str, int] | None
+    last_points_coagent: float | None
+    last_quantities: Dict[str, int] | None
+
+
+def compute_tas_style_rewards(
+    agent_ids: List[AgentId],
+    values: Dict[AgentId, float],
+    splits: Dict[AgentId, Split],
+    quantities: Dict[str, int],
+) -> Dict[AgentId, float]:
+    """
+    TAS-like reward computation: if sum of proposed coins exceeds max_coins,
+    allocate proportionally. Otherwise, use proposed amounts directly.
+    Rewards are quantity_kept * per-coin value for each agent.
+    """
+    a0, a1 = agent_ids[0], agent_ids[1]
+    r0, r1 = 0.0, 0.0
+
+    for item in quantities:
+        max_item = quantities[item]
+        item_to_self_0 = int(
+            (splits[a0].items_given_to_self.get(item, 0))
+            if splits[a0] is not None
+            else 0
+        )
+        item_to_self_1 = int(
+            (splits[a1].items_given_to_self.get(item, 0))
+            if splits[a1] is not None
+            else 0
+        )
+        denom = max(int(max_item), item_to_self_0 + item_to_self_1)
+        q0 = float(max_item) * float(item_to_self_0) / float(denom)
+        q1 = float(max_item) * float(item_to_self_1) / float(denom)
+        if type(values[a0]) is not dict:
+            r0 += q0 * float(values[a0])
+            r1 += q1 * float(values[a1])
+        else:
+            r0 += q0 * float(values[a0][item])
+            r1 += q1 * float(values[a1][item])
+    return {a0: r0, a1: r1}
+
+
+class NegotiationSimulation(Simulation):
+    def __init__(
+        self,
+        agent_ids: List[AgentId],
+        agent_names: List[str],
+        seed: int,
+        nb_of_rounds: int,
+        quota_messages_per_agent_per_round: int,
+        item_types: List[str] | None = None,
+    ):
+        self.seed = seed
+        self.rng = default_rng(self.seed)
+        self.agent_ids = list(agent_ids)
+        self.agent_names = agent_names
+        self.agent_id_to_name = {
+            agent_id: agent_name for agent_id, agent_name in zip(agent_ids, agent_names)
+        }
+        self.nb_of_rounds = int(nb_of_rounds)
+        self.quota_messages_per_agent_per_round = int(
+            quota_messages_per_agent_per_round
+        )
+        if item_types is not None:
+            self.item_types = [item.lower() for item in item_types]
+        else:
+            self.item_types = ["coins"]
+        self.state: NegotiationState | None = None
+        self._starting_agent_index = self.rng.choice([0, 1])
+        self.reset()
+
+    def _other(self, agent_id: AgentId) -> AgentId:
+        return get_coagent_id(self.agent_ids, agent_id)
+
+    @abstractmethod
+    def set_new_round_of_variant(self):
+        pass
+
+    @abstractmethod
+    def get_info_of_variant(
+        self, state: NegotiationState, actions: Dict[AgentId, Any]
+    ) -> Dict[str, Any]:
+        pass
+
+    def step(self, actions: Any) -> Tuple[bool, SimulationStepLog]:
+        """
+        Returns terminated, step_log
+        """
+        assert self.state is not None
+        current_agent = self.state.current_agent
+        a0, a1 = self.agent_ids[0], self.agent_ids[1]
+        action = actions.get(current_agent)
+
+        # Split phase: require both splits in the same timestep
+        if self.state.split_phase:
+            action_a0 = actions.get(a0)
+            action_a1 = actions.get(a1)
+            have_both_splits = isinstance(action_a0, Split) and isinstance(
+                action_a1, Split
+            )
+            if not have_both_splits:
+                rewards = {agent_id: 0.0 for agent_id in self.agent_ids}
+                return False, SimulationStepLog(
+                    rewards=rewards, info={"type": "waiting_for_splits"}
+                )
+
+            # Record splits
+            self.state.splits[a0] = action_a0
+            self.state.splits[a1] = action_a1
+
+            # Compute rewards and end round
+            rewards = self.get_rewards(self.state.splits)
+
+            # Info
+            info = self.get_info_of_variant(self.state, actions)
+
+            # Prepare next round
+            # Alternate starting agent
+            self.state.round_nb += 1
+            self._starting_agent_index = 1 - self._starting_agent_index
+            self.state.current_agent = self.agent_ids[self._starting_agent_index]
+            self.state.previous_values = copy.deepcopy(self.state.values)
+            self.state.previous_splits = copy.deepcopy(self.state.splits)
+            self.state.previous_quantities = copy.deepcopy(self.state.quantities)
+            self.state.previous_points = copy.deepcopy(rewards)
+            self.state.last_message = ""
+            self.set_new_round_of_variant()  # variant specific
+            self.state.splits = {agent_id: None for agent_id in self.agent_ids}
+            self.state.nb_messages_sent = {agent_id: 0 for agent_id in self.agent_ids}
+            is_last_timestep_in_round = True
+            done = self.state.round_nb >= self.nb_of_rounds
+
+        # Message phase
+        elif isinstance(action, Message):
+            self.state.last_message = action.message
+            self.state.nb_messages_sent[current_agent] += 1
+
+            # Move turn to other agent
+            self.state.current_agent = self._other(current_agent)
+
+            # If both agents have reached their message quota, enter split phase
+            if all(
+                self.state.nb_messages_sent[agent_id]
+                >= self.quota_messages_per_agent_per_round
+                for agent_id in self.agent_ids
+            ):
+                self.state.split_phase = True
+            is_last_timestep_in_round = False
+            done = False
+            rewards = {agent_id: 0.0 for agent_id in self.agent_ids}
+            info = {"type": "message"}
+
+        info[
+            "is_last_timestep_in_round"
+        ] = is_last_timestep_in_round  # Used later to group round timesteps if needed
+        return done, SimulationStepLog(rewards=rewards, info=info)
+
+    def get_obs(self):
+        """Returns all agent observations in dict"""
+        return {agent_id: self.get_obs_agent(agent_id) for agent_id in self.agent_ids}
+
+    @abstractmethod
+    def get_rewards(self, splits: Dict[AgentId, Split]) -> Dict[AgentId, float]:
+        pass
+
+    @abstractmethod
+    def get_obs_agent(self, agent_id):
+        pass
+
+    def get_state(self):
+        return self.state
+
+    def get_safe_copy(self):
+        """Return a safe copy of the simulation."""
+        simulation_copy = copy.copy(self)
+        simulation_copy.state = copy.deepcopy(self.state)
+        return simulation_copy
+
+    @abstractmethod
+    def reset(self) -> dict[AgentId, NegotiationObs]:
+        pass

src_code_for_reproducibility/markov_games/negotiation/negotiation_statistics.py

@@ -0,0 +1,244 @@
+from __future__ import annotations
+
+from typing import Callable, Dict, List, Tuple
+
+from mllm.markov_games.negotiation.nego_simulation import Split
+from mllm.markov_games.rollout_tree import SimulationStepLog
+
+
+def avg_reward(sl: SimulationStepLog) -> List[Tuple[str, float]]:
+    """Average (per-step) reward for each agent and overall.
+
+    What it computes:
+            - Returns the raw reward for every (non-buffer) agent at the current
+                simulation step.
+            - Adds an aggregate key ``all_agents`` which is the simple arithmetic
+                mean across the agents present in ``sl.rewards``.
+
+    Rationale / motivation:
+            Monitoring the reward stream at each step helps:
+                * Diagnose reward shaping issues (e.g., unintended negative drift).
+                * Provide a fairness snapshot (are rewards systematically skewed?).
+                * Supply a ubiquitous baseline metric used by other higher‑level
+                    summaries (efficiency, surplus allocation, etc.).
+
+    Return shape:
+            { agent_id: float, ..., "all_agents": float }
+            If any agent id contains the substring "buffer" we treat this step as
+            an implementation artifact (e.g., rollout buffer) and return ``None``
+            to avoid polluting aggregates.
+    """
+    for aid in sl.rewards.keys():
+        if "buffer" in str(aid) and "live" not in str(aid):
+            return None
+    # One value per agent at each step
+    rewards_dict = {f"reward-{aid}": float(v) for aid, v in (sl.rewards or {}).items()}
+    return [(key, value) for key, value in rewards_dict.items() if value is not None]
+
+
+def split_efficiency(sl: SimulationStepLog) -> List[Tuple[str, float]] | None:
+    """Final‑round allocation efficiency relative to an upper bound.
+
+    What it computes (only on the last timestep of a negotiation round):
+            - Uses ``info['values']`` (per‑agent per‑item valuations) and
+                ``info['quantities']`` (available item counts) to form a greedy
+                *upper bound* on achievable total reward: allocate each unit of an
+                item to the single agent who values that item most.
+            - Compares the actually realized sum of rewards at that final
+                timestep to this constructed maximum.
+            - Emits a single scalar under key ``"all_agents"`` equal to
+                achieved / theoretical_max.
+
+    Motivation:
+            Efficiency (a core welfare notion) distinguishes between coordination
+            failures (low efficiency) versus strategic distributional disputes
+            (high efficiency but uneven splits). Tracking this per round helps
+            evaluate whether models learn to identify and realize joint surplus.
+
+    Notes / caveats:
+            - Only defined for 2+ non‑buffer agents; if a buffer agent is present
+                returns ``None`` to exclude spurious steps.
+            - Requires the environment to have populated ``values`` and
+                ``quantities``; otherwise returns ``None``.
+            - This is an optimistic bound (not necessarily reachable under
+                protocol constraints) but is simple, fast, and comparable across
+                runs.
+    """
+    info = sl.info or {}
+    if not info or not info.get("is_last_timestep_in_round"):
+        return None
+    quantities = info.get("quantities") or {}
+    values = info.get("values") or {}
+    if not values or not quantities:
+        return None
+    agent_ids = list(sl.rewards.keys())
+    if type(values[agent_ids[0]]) is dict:
+        item_keys = list(values.values())[0].keys()
+        max_vals, max_quantities = [], []
+        for item in item_keys:
+            max_val = max(float(agent_vals[item]) for agent_vals in values.values())
+            max_vals.append(max_val)
+            max_quantities.append(quantities[item])
+    else:
+        max_vals = [max(float(v) for v in values.values())]
+        max_quantities = [quantities[item] for item in quantities.keys()]
+    for aid in sl.rewards.keys():
+        if "buffer" in str(aid) and "live" not in str(aid):
+            return None
+    achieved = sum(float(v) for v in sl.rewards.values())
+    max_reward = sum(d * v for d, v in zip(max_quantities, max_vals))
+    # Efficiency is a global metric; emit same value for a special key "all"
+    return [("split_efficiency", achieved / max_reward)]
+
+
+def _extract_items_from_split(raw_split: Dict) -> Dict[str, float] | None:
+    """Return a mapping item->proposal amount from a split structure.
+
+    Supports both generic negotiation splits with nested structure
+    { 'items_given_to_self': {item: qty, ...}}
+    and TAS coin-only variants which may already be a flat mapping {'coins': qty}.
+    """
+
+    if raw_split is None:
+        return {}
+    elif isinstance(raw_split, Split):
+        return {k: float(v) for k, v in raw_split.items_given_to_self.items()}
+    elif isinstance(raw_split, dict):
+        if "items_given_to_self" in raw_split and isinstance(
+            raw_split["items_given_to_self"], dict
+        ):
+            return {k: float(v) for k, v in raw_split["items_given_to_self"].items()}
+        # Fallback: assume already flat mapping of items
+        elif hasattr(raw_split, "items_given_to_self"):
+            return {k: float(v) for k, v in raw_split["items_given_to_self"].items()}
+        return {
+            k: float(v) for k, v in raw_split.items() if isinstance(v, (int, float))
+        }
+    return {}
+
+
+def _average_proposal_relative_value(
+    sl: SimulationStepLog,
+    metric_name: str,
+    comparator: Callable[[float, float], bool],
+    opposite_comparator: Callable[[float, float], bool],
+) -> Dict[str, float | None] | None:
+    """Shared implementation for proposal size conditioned on relative value.
+
+    Parameters:
+            comparator: returns True when agent_0's value relation (e.g. < or >)
+                                    to agent_1 holds for an item and we should collect agent_0's
+                                    proposed quantity for that item.
+            opposite_comparator: inverse relation used to collect agent_1's items.
+
+    Behavior:
+            - Executes only on final timestep of a round (where the definitive
+                proposal / allocation is known via ``info['splits']``).
+            - For each item, classifies which agent's value satisfies the chosen
+                relation and records that agent's proposed quantity from the split.
+            - Averages (mean) across all qualifying items per agent; if no items
+                qualify for an agent returns ``None`` for that agent id.
+            - Adds ``all_agents`` mean across the numeric (non-None) agent values.
+
+    Why this matters:
+            Distinguishing how much an agent *asks for* when it subjectively
+            values items more (or less) than its counterpart reveals patterns of
+            opportunism vs. concession. This is especially useful when raw reward
+            differences are subtle but allocation *intent* differs.
+    """
+    info = sl.info or {}
+    if not info or not info.get("is_last_timestep_in_round"):
+        return None
+    quantities = info.get("quantities") or {}
+    splits = info.get("splits") or {}
+    values = info.get("values") or {}
+    agent_ids: List[str] = list(sl.rewards.keys())
+    if len(agent_ids) != 2:
+        return None  # Only defined for 2-agent case.
+    for aid in agent_ids:
+        if "buffer" in str(aid) and "live" not in str(aid):
+            return None
+    # Extract per-agent item proposals robustly
+    split_items = {aid: _extract_items_from_split(splits.get(aid)) for aid in agent_ids}
+    agent_0_vals: List[float] = []
+    agent_1_vals: List[float] = []
+    for item in quantities.keys():
+        # Values may be either a float (same for all items) or dict per item
+        v0_raw = values[agent_ids[0]]
+        v1_raw = values[agent_ids[1]]
+        v0 = float(v0_raw[item]) if isinstance(v0_raw, dict) else float(v0_raw)
+        v1 = float(v1_raw[item]) if isinstance(v1_raw, dict) else float(v1_raw)
+        if comparator(v0, v1):
+            agent_0_vals.append(split_items[agent_ids[0]].get(item, 0.0))
+        elif opposite_comparator(v0, v1):
+            agent_1_vals.append(split_items[agent_ids[1]].get(item, 0.0))
+    out: Dict[str, float | None] = {}
+    out[f"{metric_name}-{agent_ids[0]}"] = (
+        sum(agent_0_vals) / len(agent_0_vals) if agent_0_vals else None
+    )
+    out[f"{metric_name}-{agent_ids[1]}"] = (
+        sum(agent_1_vals) / len(agent_1_vals) if agent_1_vals else None
+    )
+
+    return [(key, value) for key, value in out.items() if value is not None]
+
+
+def average_proposal_when_agent_values_item_lower(
+    sl: SimulationStepLog,
+) -> List[Tuple[str, float | None]] | None:
+    """Mean quantity an agent proposes for items it values *less* than opponent.
+
+    Interpretation:
+        A higher value implies the agent still claims (or is allocated) a
+        notable share of items where it has a comparative *disadvantage* in
+        valuation, signaling either strategic over-claiming or protocol-driven
+        egalitarian splits. Conversely, very low numbers can indicate
+        efficient specialization or excessive concession.
+
+    Returns:
+        Mapping { agent_id: float | None, "all_agents": float | None } where
+        None indicates no qualifying items for that agent in the round.
+    """
+    return _average_proposal_relative_value(
+        sl,
+        "average_proposal_when_agent_values_item_lower",
+        lambda a, b: a < b,
+        lambda a, b: a > b,
+    )
+
+
+def average_proposal_when_agent_values_item_higher(
+    sl: SimulationStepLog,
+) -> List[Tuple[str, float | None]] | None:
+    """Mean quantity an agent proposes for items it values *more* than opponent.
+
+    Interpretation:
+        Captures how aggressively an agent claims items where it holds a
+        comparative *advantage*. Elevated values can reflect rational
+        specialization (efficient exploitation of comparative advantage) or
+        potentially unfair grabs if paired with low concession in the lower
+        valuation metric. Comparing this with the 'lower' counterpart helps
+        profile negotiation style (cooperative vs. exploitative).
+
+    Returns:
+        Mapping { agent_id: float | None, "all_agents": float | None } where
+        None indicates no qualifying items.
+    """
+    return _average_proposal_relative_value(
+        sl,
+        "average_proposal_when_agent_values_item_higher",
+        lambda a, b: a > b,
+        lambda a, b: a < b,
+    )
+
+
+# Explicit list of metric functions exported for rendering. Helper functions
+# starting with '_' are intentionally excluded. Update this list when adding
+# new public statistics so render.py can rely on it instead of introspecting
+# every callable in the module.
+stat_functs: list[Callable[[SimulationStepLog], List[Tuple[str, float]]]] = [
+    avg_reward,
+    average_proposal_when_agent_values_item_lower,
+    average_proposal_when_agent_values_item_higher,
+    split_efficiency,
+]

src_code_for_reproducibility/markov_games/negotiation/tas_rps_simulation.py

@@ -0,0 +1,248 @@
+"""
+Trust-and-Split simulation.
+
+This environment models a simple bargaining game over 10 coins with messaging.
+Agents are assigned rock/paper/scissors hands, with the winner getting value 10 per coin
+and the loser getting value 1 per coin. Agents alternate sending messages for a fixed
+number of turns per round and then each submits a split proposal indicating how many
+coins they keep for themselves. Rewards are proportional if the proposed totals exceed 10.
+"""
+
+import copy
+from dataclasses import dataclass
+from typing import Any, Dict, List, Literal, Tuple
+
+from numpy.random import default_rng
+
+from mllm.markov_games.negotiation.nego_simulation import (
+    Message,
+    NegotiationObs,
+    NegotiationSimulation,
+    NegotiationState,
+    Split,
+    compute_tas_style_rewards,
+)
+from mllm.markov_games.rollout_tree import SimulationStepLog
+
+AgentId = str
+
+
+def _get_rps_winner(
+    hand1: Literal["rock", "paper", "scissors"],
+    hand2: Literal["rock", "paper", "scissors"],
+) -> Literal["rock", "paper", "scissors"]:
+    """Determine winner of rock-paper-scissors between two hands."""
+    if hand1 == hand2:
+        raise ValueError("Hands should be different")
+    if (
+        (hand1 == "rock" and hand2 == "scissors")
+        or (hand1 == "paper" and hand2 == "rock")
+        or (hand1 == "scissors" and hand2 == "paper")
+    ):
+        return hand1
+    else:
+        return hand2
+
+
+@dataclass
+class TrustAndSplitRPSState(NegotiationState):
+    hands: Dict[
+        AgentId, Literal["rock", "paper", "scissors"]
+    ]  # rock, paper, or scissors
+    previous_hands: Dict[AgentId, Literal["rock", "paper", "scissors"]] | None
+
+
+@dataclass
+class TrustAndSplitRPSObs(NegotiationObs):
+    hand: Literal["rock", "paper", "scissors"]
+    last_hand_agent: Literal["rock", "paper", "scissors"] | None
+    last_hand_coagent: Literal["rock", "paper", "scissors"] | None
+    last_hand_value_coagent: Literal["upper", "lower"] | None
+
+
+class TrustAndSplitRPSSimulation(NegotiationSimulation):
+    def __init__(
+        self,
+        alternating_hands: bool = False,
+        alternating_mix_ratio: float = None,
+        *args,
+        **kwargs,
+    ):
+        self.alternating_hands = alternating_hands
+        self.alternating_mix_ratio = alternating_mix_ratio
+        super().__init__(*args, **kwargs)
+        if self.alternating_mix_ratio is not None:
+            if self.rng.random() < self.alternating_mix_ratio:
+                self.alternating_hands = True
+            else:
+                self.alternating_hands = False
+
+    def _sample_hands_and_values(
+        self,
+        alternate_hands: bool = False,
+    ) -> Tuple[Dict[AgentId, str], Dict[AgentId, float]]:
+        hands = ["rock", "paper", "scissors"]
+        if alternate_hands:
+            previous_hands = list(self.state.previous_hands.values())
+            hand1, hand2 = self.rng.choice(hands, size=2, replace=False)
+            winner = _get_rps_winner(hand1, hand2)
+            loser = hand1 if winner == hand2 else hand2
+            previous_winner = _get_rps_winner(previous_hands[0], previous_hands[1])
+            agent_hands, values = {}, {}
+            for agent_id in self.agent_ids:
+                if self.state.previous_hands[agent_id] == previous_winner:
+                    agent_hands[agent_id] = loser
+                    values[agent_id] = 1.0
+                else:
+                    agent_hands[agent_id] = winner
+                    values[agent_id] = 10.0
+            return agent_hands, values
+        else:
+            # Assign different hands to each agent
+            hand1, hand2 = self.rng.choice(hands, size=2, replace=False)
+
+            agent_hands = {self.agent_ids[0]: hand1, self.agent_ids[1]: hand2}
+
+            # Determine winner and assign values
+            winner = _get_rps_winner(hand1, hand2)
+            values = {}
+            for agent_id in self.agent_ids:
+                if agent_hands[agent_id] == winner:
+                    values[agent_id] = 10.0  # Winner gets value 10
+                else:
+                    values[agent_id] = 1.0  # Loser gets value 1
+
+            return agent_hands, values
+
+    def set_new_round_of_variant(self):
+        self.state.previous_hands = copy.deepcopy(self.state.hands)
+        new_hands, new_values = self._sample_hands_and_values(
+            alternate_hands=self.alternating_hands
+        )
+        self.state.hands = new_hands
+        self.state.values = new_values
+        # Quantities are constant in TAS
+        self.state.quantities = {"coins": 10}
+        self.state.split_phase = False
+
+    def get_info_of_variant(
+        self, state: NegotiationState, actions: Dict[AgentId, Any]
+    ) -> Dict[str, Any]:
+        return {
+            "quantities": copy.deepcopy(state.quantities),
+            "hands": copy.deepcopy(state.hands),
+            "values": copy.deepcopy(state.values),
+            "previous_hands": copy.deepcopy(state.previous_hands),
+            "previous_values": copy.deepcopy(state.previous_values),
+            "splits": copy.deepcopy(state.splits),
+        }
+
+    def get_rewards(self, splits: Dict[AgentId, Split]) -> Dict[AgentId, float]:
+        return compute_tas_style_rewards(
+            self.agent_ids, self.state.values, splits, self.state.quantities
+        )
+
+    def get_obs_agent(self, agent_id):
+        """Returns observation for agent_id"""
+        other_id = self._other(agent_id)
+        last_value_coagent = (
+            None
+            if self.state.previous_values is None
+            else self.state.previous_values.get(other_id)
+        )
+        last_hand_coagent = (
+            None
+            if self.state.previous_hands is None
+            else self.state.previous_hands.get(other_id)
+        )
+        last_points_coagent = (
+            None
+            if self.state.previous_points is None
+            else round(self.state.previous_points.get(other_id), 1)
+        )
+        last_value_agent = (
+            None
+            if self.state.previous_values is None
+            else self.state.previous_values.get(agent_id)
+        )
+        last_hand_agent = (
+            None
+            if self.state.previous_hands is None
+            else self.state.previous_hands.get(agent_id)
+        )
+        last_points_agent = (
+            None
+            if self.state.previous_points is None
+            else round(self.state.previous_points.get(agent_id), 1)
+        )
+        last_split_coagent = None
+        last_split_agent = None
+        if self.state.previous_splits is not None:
+            last_split_coagent = self.state.previous_splits[
+                other_id
+            ].items_given_to_self["coins"]
+            last_split_agent = self.state.previous_splits[agent_id].items_given_to_self[
+                "coins"
+            ]
+        if last_hand_agent is None or last_hand_coagent is None:
+            last_hand_value_coagent = None
+        else:
+            winner = _get_rps_winner(last_hand_agent, last_hand_coagent)
+            last_hand_value_coagent = (
+                "upper" if winner == last_hand_coagent else "lower"
+            )
+        obs = TrustAndSplitRPSObs(
+            round_nb=self.state.round_nb,
+            last_message=self.state.last_message,
+            quota_messages_per_agent_per_round=self.quota_messages_per_agent_per_round,
+            current_agent=self.state.current_agent,
+            other_agent=self.agent_id_to_name[other_id],
+            quantities={"coins": 10},
+            item_types=self.item_types,
+            value=self.state.values[agent_id],
+            split_phase=self.state.split_phase,
+            last_split_agent=last_split_agent,
+            last_value_agent=last_value_agent,
+            last_points_agent=last_points_agent,
+            last_split_coagent=last_split_coagent,
+            last_value_coagent=last_value_coagent,
+            last_points_coagent=last_points_coagent,
+            hand=self.state.hands[agent_id],
+            last_hand_coagent=last_hand_coagent,
+            last_hand_agent=last_hand_agent,
+            last_quantities=self.state.previous_quantities,
+            last_hand_value_coagent=last_hand_value_coagent,
+        )
+        return obs
+
+    def get_state(self):
+        return self.state
+
+    def get_safe_copy(self):
+        """Return a safe copy of the simulation."""
+        simulation_copy = copy.copy(self)
+        simulation_copy.state = copy.deepcopy(self.state)
+        return simulation_copy
+
+    def reset(self):
+        """Initialize and return initial observations"""
+        # Decide starting agent alternating across resets for determinism
+        start_agent = self.agent_ids[self._starting_agent_index]
+        hands, values = self._sample_hands_and_values()
+        self.state = TrustAndSplitRPSState(
+            round_nb=0,
+            last_message="",
+            current_agent=start_agent,
+            quantities={"coins": 10},
+            values=values,
+            splits={aid: None for aid in self.agent_ids},
+            nb_messages_sent={aid: 0 for aid in self.agent_ids},
+            previous_values=None,
+            previous_splits=None,
+            previous_points=None,
+            split_phase=False,
+            hands=hands,
+            previous_hands=None,
+            previous_quantities=None,
+        )
+        return self.get_obs()

src_code_for_reproducibility/markov_games/negotiation/tas_simple_agent.py

@@ -0,0 +1,90 @@
+from mllm.markov_games.negotiation.nego_agent import NegotiationAgent
+from mllm.markov_games.negotiation.nego_simulation import Split
+from mllm.markov_games.negotiation.tas_simple_simulation import TrustAndSplitSimpleObs
+
+
+class TrustAndSplitSimpleAgent(NegotiationAgent):
+    def __init__(
+        self,
+        num_message_chars,
+        message_start_end_format: bool = False,
+        proposal_start_end_format: bool = False,
+        *args,
+        **kwargs,
+    ):
+        self.num_message_chars = num_message_chars
+        self.message_start_end_format = message_start_end_format
+        self.proposal_start_end_format = proposal_start_end_format
+        super().__init__(*args, **kwargs)
+        self.intro_prompt = (
+            "Welcome to an iterated game. You are {agent}. The other agent is {other_agent}.\n"
+            "Setup:\n"
+            "1. The game has multiple independent rounds.\n"
+            "2. In each round, there are 10 coins to split between the two agents.\n"
+            "3. Both agents are assigned a per-coin value between 1 and 10 (inclusive) in each round.\n"
+            "4. You can only observe your own per-coin value.\n"
+            "5. Because assignments are random, both agents are equally likely to have same expected per-coin value.\n"
+            "\n"
+            "Protocol:\n"
+            "1. At the start of the round, one agent begins the conversation. The starting role alternates each round.\n"
+            "2. Agents exchange a short chat ({quota_messages_per_agent_per_round} messages per round per agent) to negotiate how to split the coins.\n"
+            "   - Use this chat to communicate your private per-coin value to make informed proposals.\n"
+            "3. After the chat, both agents simultaneously propose how many coins they keep.\n"
+            "4. If the total sum of proposals is less than or equal to 10, both agents receive their proposals.\n"
+            "5. If the total sum of proposals exceeds 10, the coins are allocated proportionally.\n"
+            "6. Your points for the round = (coins you receive) x (your per-coin value for that round). \n"
+            "7. Points are accumulated across rounds.\n"
+            "Your goal: {goal}\n"
+        )
+        self.new_round_prompt = (
+            "A New Round Begins\n"
+            "Your per-coin value is {value}. You don't know {other_agent}'s value yet.\n"
+        )
+        self.last_round_prompt = "In the previous round, {other_agent} had a {last_value_str_coagent} value and proposed {last_split_coagent} coins.\n"
+        if self.proposal_start_end_format:
+            self.send_split_prompt = (
+                "Submit your proposal\n"
+                "Respond with <<proposal_start>> x <<proposal_end>> where x is an integer in [0, 10]."
+            )
+        else:
+            self.send_split_prompt = (
+                "Submit your proposal\n"
+                "Respond with <coins_to_self> x </coins_to_self> where x is an integer in [0, 10]."
+            )
+        self.wait_for_message_prompt = "Wait for {other_agent} to send a message..."
+        # self.wait_for_message_prompt = ""
+        self.last_message_prompt = "{other_agent} said: {last_message}"
+        if self.message_start_end_format:
+            self.send_message_prompt = f"Send your message now in <<message_start>>...<<message_end>> (<={self.num_message_chars} chars)."
+        else:
+            self.send_message_prompt = f"Send your message now in <message>...</message> (<={self.num_message_chars} chars)."
+
+    def get_message_regex(self, observation: TrustAndSplitSimpleObs) -> str:
+        if self.message_start_end_format:
+            return (
+                rf"<<message_start>>[\s\S]{{0,{self.num_message_chars}}}<<message_end>>"
+            )
+        else:
+            return rf"<message>[\s\S]{{0,{self.num_message_chars}}}</message>"
+
+    def get_split_regex(self, observation: TrustAndSplitSimpleObs) -> str:
+        if self.proposal_start_end_format:
+            return r"<<proposal_start>> ?(10|[0-9]) ?<<proposal_end>>"
+        else:
+            return r"<coins_to_self> ?(10|[0-9]) ?</coins_to_self>"
+
+    def get_split_action(
+        self, policy_output: str, observation: TrustAndSplitSimpleObs
+    ) -> Split:
+        import re as _re
+
+        if self.proposal_start_end_format:
+            m = _re.search(
+                r"<<proposal_start>> ?(10|[0-9]) ?<<proposal_end>>", policy_output
+            )
+        else:
+            m = _re.search(
+                r"<coins_to_self> ?(10|[0-9]) ?</coins_to_self>", policy_output
+            )
+        coins_int = int(m.group(1)) if m else int(policy_output)
+        return Split(items_given_to_self={"coins": coins_int})