a program that moderate a Mafia(Werewolf) game with 7 players between GPT models.

to see communication, team work and social/logical abilities of GPT model.(it is fun and educational)

the roles are :

• 1 Medical
• 1 Seer
• 3 Villagers
• 2 Werewolves

install openai via pip :

pip install openai

set your Openai API key as an enivronmental variable with key : OPENAI_API_KEY

• you can configure your model(Default is 'gpt-3.5-turbo') and your rate_limit and token_limit in api.py

• make DEBUG=True in api.py if you want to see all messages passed to GPT model.

python run.py

it will save a log file in the bellow format in records/ folder.

every event in the game is a dictionary.

{
  'event':"<event_name>",
  'content':"<...>"
}

events are :

• roles :

'content':{
  'player':"<player_number>"
  'role':"<role_string>"
}

• cycle :

'content':"<day/night>"

• speech :

'content':{
  'player':"<player_number>"
  'context':"<speech_string>"
}

• voted :

'content':{
  'player':"<player_number>"
  'voted_to_player':"<player_number>"
  'reason':"<complete response>"
}

• vote_start :

'content':{
  'player':"<player_number>"
  'voted_to_player':"<player_number>"
  'reason':"<complete response>"
}

• healed :

'content':{
  'player':"<player_number>"
  'reason':"<complete response>"
}

• targeted :

'content':{
  'player':"<player_number>"
  'reason':"<complete response>"
}

• killed :

'content':{
  'player':"<player_number>"
}

• inquiried :

'content':{
  'player':"<player_number>"
  'context':"<True: if player is werewolf>"
  'reason':"<complete response>"
}

• notetaking :

'content':{
  'player':"<player_number>"
  'context':"<speech_string>"
}

• end :

'content':{
  'winner':"<Werewolves/Villagers>"
}

• Game Introduction('/intro_prompt.txt') : rules and the role of the player
• Game Reports('/report_prompt.txt) :
  • Speeches in last round
  • Game status : alive players in seperation of the teams.
  • Players status : alivness of players based on their number
  • Agent previous SPEACIAL actions
  • Agent previous VOTES
  • Notes that have taken by the agent
• Commands('/game.py') :
  • Speak command
  • Vote command
  • targetting/healing/killing/inquiring command

• optimizing prompts
• a web interface to play natively a log-file as an animation
• allowing a player/players to play with GPT (Probably tricky prompts will lead to game hacking.)
• allowing more chat models or different GPT tempretures to compete