From f54bb159b5ee18cdf3e64545a860ca67e74b97a4 Mon Sep 17 00:00:00 2001 From: better629 Date: Mon, 8 Apr 2024 19:38:37 +0800 Subject: [PATCH] use WerewolfEnv --- examples/werewolf_game/start_game.py | 48 ++--------- .../ext/werewolf/actions/moderator_actions.py | 84 +------------------ metagpt/ext/werewolf/roles/base_player.py | 27 +++--- metagpt/ext/werewolf/roles/human_player.py | 5 +- metagpt/ext/werewolf/roles/moderator.py | 38 ++++----- metagpt/ext/werewolf/roles/werewolf.py | 4 +- metagpt/ext/werewolf/roles/witch.py | 12 +-- metagpt/ext/werewolf/werewolf_game.py | 48 ++++------- .../actions/test_experience_operation.py | 1 - 9 files changed, 67 insertions(+), 200 deletions(-) diff --git a/examples/werewolf_game/start_game.py b/examples/werewolf_game/start_game.py index fdd17256a..72023eed7 100644 --- a/examples/werewolf_game/start_game.py +++ b/examples/werewolf_game/start_game.py @@ -1,51 +1,12 @@ import asyncio -import random import fire from metagpt.ext.werewolf.roles import Guard, Moderator, Seer, Villager, Werewolf, Witch -from metagpt.ext.werewolf.roles.human_player import prepare_human_player from metagpt.ext.werewolf.werewolf_game import WerewolfGame from metagpt.logs import logger -def init_game_setup( - shuffle=True, - add_human=False, - use_reflection=True, - use_experience=False, - use_memory_selection=False, - new_experience_version="", -): - roles = [Villager, Villager, Werewolf, Werewolf, Guard, Seer, Witch] - if shuffle: - # random.seed(2023) - random.shuffle(roles) - if add_human: - assigned_role_idx = random.randint(0, len(roles) - 1) - assigned_role = roles[assigned_role_idx] - roles[assigned_role_idx] = prepare_human_player(assigned_role) - - players = [ - role( - name=f"Player{i+1}", - use_reflection=use_reflection, - use_experience=use_experience, - use_memory_selection=use_memory_selection, - new_experience_version=new_experience_version, - ) - for i, role in enumerate(roles) - ] - - if add_human: - logger.info(f"You are assigned {players[assigned_role_idx].name}({players[assigned_role_idx].profile})") - - game_setup = ["Game setup:"] + [f"{player.name}: {player.profile}," for player in players] - game_setup = "\n".join(game_setup) - - return game_setup, players - - async def start_game( investment: float = 3.0, n_round: int = 5, @@ -57,7 +18,10 @@ async def start_game( new_experience_version: str = "", ): game = WerewolfGame() - game_setup, players = init_game_setup( + game_setup, players = game.env.init_game_setup( + role_uniq_objs=[Villager, Werewolf, Guard, Seer, Witch], + num_werewolf=2, + num_villager=2, shuffle=shuffle, add_human=add_human, use_reflection=use_reflection, @@ -65,10 +29,12 @@ async def start_game( use_memory_selection=use_memory_selection, new_experience_version=new_experience_version, ) + logger.info(f"game_setup\n{game_setup}") + players = [Moderator()] + players game.hire(players) game.invest(investment) - game.start_project(game_setup) + game.run_project(game_setup) await game.run(n_round=n_round) diff --git a/metagpt/ext/werewolf/actions/moderator_actions.py b/metagpt/ext/werewolf/actions/moderator_actions.py index 6153c66c7..42bd0fc4e 100644 --- a/metagpt/ext/werewolf/actions/moderator_actions.py +++ b/metagpt/ext/werewolf/actions/moderator_actions.py @@ -1,90 +1,12 @@ from metagpt.actions import Action - -STEP_INSTRUCTIONS = { - # 上帝需要介入的全部步骤和对应指令 - # The 1-st night - 0: { - "content": "It’s dark, everyone close your eyes. I will talk with you/your team secretly at night.", - "send_to": "Moderator", # for moderator to continuen speaking - "restricted_to": "", - }, - 1: { - "content": "Guard, please open your eyes!", - "send_to": "Moderator", # for moderator to continuen speaking - "restricted_to": "", - }, - 2: { - "content": """Guard, now tell me who you protect tonight? - You only choose one from the following living options please: {living_players}. - Or you can pass. For example: Protect ...""", - "send_to": "Guard", - "restricted_to": "Moderator,Guard", - }, - 3: {"content": "Guard, close your eyes", "send_to": "Moderator", "restricted_to": ""}, - 4: {"content": "Werewolves, please open your eyes!", "send_to": "Moderator", "restricted_to": ""}, - 5: { - "content": """Werewolves, I secretly tell you that {werewolf_players} are - all of the 2 werewolves! Keep in mind you are teammates. The rest players are not werewolves. - choose one from the following living options please: - {living_players}. For example: Kill ...""", - "send_to": "Werewolf", - "restricted_to": "Moderator,Werewolf", - }, - 6: {"content": "Werewolves, close your eyes", "send_to": "Moderator", "restricted_to": ""}, - 7: {"content": "Witch, please open your eyes!", "send_to": "Moderator", "restricted_to": ""}, - 8: { - "content": """Witch, tonight {player_hunted} has been killed by the werewolves. - You have a bottle of antidote, would you like to save him/her? If so, say "Save", else, say "Pass".""", - "send_to": "Witch", - "restricted_to": "Moderator,Witch", - }, # 要先判断女巫是否有解药，再去询问女巫是否使用解药救人 - 9: { - "content": """Witch, you also have a bottle of poison, would you like to use it to kill one of the living players? - Choose one from the following living options: {living_players}. - If so, say ONLY "Poison PlayerX", replace PlayerX with the actual player name, else, say "Pass".""", - "send_to": "Witch", - "restricted_to": "Moderator,Witch", - }, # - 10: {"content": "Witch, close your eyes", "send_to": "Moderator", "restricted_to": ""}, - 11: {"content": "Seer, please open your eyes!", "send_to": "Moderator", "restricted_to": ""}, - 12: { - "content": """Seer, you can check one player's identity. Who are you going to verify its identity tonight? - Choose only one from the following living options:{living_players}.""", - "send_to": "Seer", - "restricted_to": "Moderator,Seer", - }, - 13: {"content": "Seer, close your eyes", "send_to": "Moderator", "restricted_to": ""}, - # The 1-st daytime - 14: { - "content": """It's daytime. Everyone woke up except those who had been killed.""", - "send_to": "Moderator", - "restricted_to": "", - }, - 15: {"content": "{player_current_dead} was killed last night!", "send_to": "Moderator", "restricted_to": ""}, - 16: { - "content": """Living players: {living_players}, now freely talk about the current situation based on your observation and - reflection with a few sentences. Decide whether to reveal your identity based on your reflection.""", - "send_to": "", # send to all to speak in daytime - "restricted_to": "", - }, - 17: { - "content": """Now vote and tell me who you think is the werewolf. Don’t mention your role. - You only choose one from the following living options please: - {living_players}. Say ONLY: I vote to eliminate ...""", - "send_to": "", - "restricted_to": "", - }, - 18: {"content": """{player_current_dead} was eliminated.""", "send_to": "Moderator", "restricted_to": ""}, -} +from metagpt.environment.werewolf.werewolf_ext_env import STEP_INSTRUCTIONS class InstructSpeak(Action): name: str = "InstructSpeak" async def run(self, step_idx, living_players, werewolf_players, player_hunted, player_current_dead): - instruction_info = STEP_INSTRUCTIONS.get( - step_idx, {"content": "Unknown instruction.", "send_to": "", "restricted_to": ""} - ) + instruction_info = STEP_INSTRUCTIONS.get(step_idx, {"content": "Unknown instruction.", "send_to": {}}) content = instruction_info["content"] if "{living_players}" in content and "{werewolf_players}" in content: content = content.format(living_players=living_players, werewolf_players=werewolf_players) @@ -98,7 +20,7 @@ class InstructSpeak(Action): player_current_dead = "No one" if not player_current_dead else player_current_dead content = content.format(player_current_dead=player_current_dead) - return content, instruction_info["send_to"], instruction_info["restricted_to"] + return content, instruction_info["send_to"] class ParseSpeak(Action): diff --git a/metagpt/ext/werewolf/roles/base_player.py b/metagpt/ext/werewolf/roles/base_player.py index 548a48177..6f459526e 100644 --- a/metagpt/ext/werewolf/roles/base_player.py +++ b/metagpt/ext/werewolf/roles/base_player.py @@ -57,23 +57,23 @@ class BasePlayer(Role): await super()._observe() # 只有发给全体的（""）或发给自己的（self.profile）消息需要走下面的_react流程， # 其他的收听到即可，不用做动作 - self._rc.news = [msg for msg in self._rc.news if msg.send_to in ["", self.profile]] - return len(self._rc.news) + self.rc.news = [msg for msg in self.rc.news if msg.send_to in ["", self.profile]] + return len(self.rc.news) async def _think(self): - news = self._rc.news[0] + news = self.rc.news[0] assert news.cause_by == InstructSpeak # 消息为来自Moderator的指令时，才去做动作 - if not news.restricted_to: + if not news.send_to: # 消息接收范围为全体角色的，做公开发言（发表投票观点也算发言） - self._rc.todo = Speak() - elif self.profile in news.restricted_to.split(","): + self.rc.todo = Speak() + elif self.profile in news.send_to: # FIXME: hard code to split, restricted为"Moderator"或"Moderator,角色profile" # Moderator加密发给自己的，意味着要执行角色的特殊动作 - self._rc.todo = self.special_actions[0]() + self.rc.todo = self.special_actions[0]() async def _act(self): # todo为_think时确定的，有两种情况，Speak或Protect - todo = self._rc.todo + todo = self.rc.todo logger.info(f"{self._setting}: ready to {str(todo)}") # 可以用这个函数获取该角色的全部记忆和最新的instruction @@ -107,21 +107,20 @@ class BasePlayer(Role): reflection=reflection, experiences=experiences, ) - restricted_to = "" + send_to = "" elif isinstance(todo, NighttimeWhispers): rsp = await todo.run( profile=self.profile, name=self.name, context=memories, reflection=reflection, experiences=experiences ) - restricted_to = f"Moderator,{self.profile}" # 给Moderator发送使用特殊技能的加密消息 + send_to = {"Moderator", self.profile} # 给Moderator发送使用特殊技能的加密消息 msg = Message( content=rsp, role=self.profile, sent_from=self.name, cause_by=type(todo), - send_to="", - restricted_to=restricted_to, + send_to=send_to, ) self.experiences.append( @@ -140,7 +139,7 @@ class BasePlayer(Role): return msg def get_all_memories(self) -> str: - memories = self._rc.memory.get() + memories = self.rc.memory.get() time_stamp_pattern = r"[0-9]+ \| " # NOTE: 除Moderator外，其他角色使用memory，只能用m.sent_from（玩家名）不能用m.role（玩家角色），因为他们不知道说话者的身份 memories = [f"{m.sent_from}: {re.sub(time_stamp_pattern, '', m.content)}" for m in memories] # regex去掉时间戳 @@ -148,7 +147,7 @@ class BasePlayer(Role): return memories def get_latest_instruction(self) -> str: - return self._rc.important_memory[-1].content # 角色监听着Moderator的InstructSpeak，是其重要记忆，直接获取即可 + return self.rc.important_memory[-1].content # 角色监听着Moderator的InstructSpeak，是其重要记忆，直接获取即可 def set_status(self, new_status): self.status = new_status diff --git a/metagpt/ext/werewolf/roles/human_player.py b/metagpt/ext/werewolf/roles/human_player.py index e47588b34..830bf06cf 100644 --- a/metagpt/ext/werewolf/roles/human_player.py +++ b/metagpt/ext/werewolf/roles/human_player.py @@ -5,7 +5,7 @@ from metagpt.schema import Message async def _act(self): - todo = self._rc.todo + todo = self.rc.todo memories = self.get_all_memories() @@ -29,8 +29,7 @@ async def _act(self): role=self.profile, sent_from=self.name, cause_by=msg_cause_by, - send_to="", - restricted_to=msg_restricted_to, # 给Moderator及自身阵营发送加密消息 + send_to=msg_restricted_to, # 给Moderator及自身阵营发送加密消息 ) logger.info(f"{self._setting}: {rsp}") diff --git a/metagpt/ext/werewolf/roles/moderator.py b/metagpt/ext/werewolf/roles/moderator.py index 0b841fd22..f0de03959 100644 --- a/metagpt/ext/werewolf/roles/moderator.py +++ b/metagpt/ext/werewolf/roles/moderator.py @@ -4,9 +4,9 @@ from datetime import datetime from metagpt.actions.add_requirement import UserRequirement from metagpt.const import DEFAULT_WORKSPACE_ROOT +from metagpt.environment.werewolf.werewolf_ext_env import STEP_INSTRUCTIONS from metagpt.ext.werewolf.actions import Hunt, Poison, Protect, Save, Verify from metagpt.ext.werewolf.actions.moderator_actions import ( - STEP_INSTRUCTIONS, AnnounceGameResult, InstructSpeak, ParseSpeak, @@ -64,14 +64,14 @@ class Moderator(Role): def update_player_status(self, player_names: list[str]): if not player_names: return - roles_in_env = self._rc.env.get_roles() + roles_in_env = self.rc.env.get_roles() for role_setting, role in roles_in_env.items(): for player_name in player_names: if player_name in role_setting: role.set_status(new_status=1) # 更新为死亡 def _record_all_experiences(self): - roles_in_env = self._rc.env.get_roles() + roles_in_env = self.rc.env.get_roles() timestamp = datetime.now().strftime("%Y_%m_%d_%H_%M_%S") for _, role in roles_in_env.items(): if role == self: @@ -104,7 +104,7 @@ class Moderator(Role): # default return msg_content = "Understood" - restricted_to = "" + send_to = set() msg_cause_by = latest_msg.cause_by if msg_cause_by == Hunt: @@ -116,13 +116,13 @@ class Moderator(Role): msg_content = f"{target} is a werewolf" else: msg_content = f"{target} is a good guy" - restricted_to = "Moderator,Seer" + send_to = {"Moderator", "Seer"} elif msg_cause_by == Save: if "pass" in latest_msg_content.lower(): pass elif not self.witch_antidote_left: msg_content = "You have no antidote left and thus can not save the player" - restricted_to = "Moderator,Witch" + send_to = {"Moderator", "Witch"} else: self.witch_antidote_left -= 1 self.is_hunted_player_saved = True @@ -131,12 +131,12 @@ class Moderator(Role): pass elif not self.witch_poison_left: msg_content = "You have no poison left and thus can not poison the player" - restricted_to = "Moderator,Witch" + send_to = {"Moderator", "Witch"} else: self.witch_poison_left -= 1 self.player_poisoned = target # "" if not poisoned and "PlayerX" if poisoned - return msg_content, restricted_to + return msg_content, send_to def _update_game_states(self, memories): step_idx = self.step_idx % len(STEP_INSTRUCTIONS) @@ -198,27 +198,27 @@ class Moderator(Role): async def _think(self): if self.winner is not None: - self._rc.todo = AnnounceGameResult() + self.rc.todo = AnnounceGameResult() return - latest_msg = self._rc.memory.get()[-1] + latest_msg = self.rc.memory.get()[-1] if latest_msg.role in ["User"]: # 上一轮消息是用户指令，解析用户指令，开始游戏 game_setup = latest_msg.content self._parse_game_setup(game_setup) - self._rc.todo = InstructSpeak() + self.rc.todo = InstructSpeak() elif latest_msg.role in [self.profile]: # 1. 上一轮消息是Moderator自己的指令，继续发出指令，一个事情可以分几条消息来说 # 2. 上一轮消息是Moderator自己的解析消息，一个阶段结束，发出新一个阶段的指令 - self._rc.todo = InstructSpeak() + self.rc.todo = InstructSpeak() else: # 上一轮消息是游戏角色的发言，解析角色的发言 - self._rc.todo = ParseSpeak() + self.rc.todo = ParseSpeak() async def _act(self): - todo = self._rc.todo + todo = self.rc.todo logger.info(f"{self._setting} ready to {todo}") memories = self.get_all_memories(mode="msg") @@ -231,7 +231,7 @@ class Moderator(Role): # 根据_think的结果，执行InstructSpeak还是ParseSpeak, 并将结果返回 if isinstance(todo, InstructSpeak): - msg_content, msg_to_send_to, msg_restriced_to = await self._instruct_speak() + msg_content, msg_to_send_to = await self._instruct_speak() # msg_content = f"Step {self.step_idx}: {msg_content}" # HACK: 加一个unique的step_idx避免记忆的自动去重 msg = Message( content=msg_content, @@ -239,19 +239,17 @@ class Moderator(Role): sent_from=self.name, cause_by=InstructSpeak, send_to=msg_to_send_to, - restricted_to=msg_restriced_to, ) elif isinstance(todo, ParseSpeak): - msg_content, msg_restriced_to = await self._parse_speak(memories) + msg_content, msg_to_send_to = await self._parse_speak(memories) # msg_content = f"Step {self.step_idx}: {msg_content}" # HACK: 加一个unique的step_idx避免记忆的自动去重 msg = Message( content=msg_content, role=self.profile, sent_from=self.name, cause_by=ParseSpeak, - send_to="", - restricted_to=msg_restriced_to, + send_to=msg_to_send_to, ) elif isinstance(todo, AnnounceGameResult): @@ -263,7 +261,7 @@ class Moderator(Role): return msg def get_all_memories(self, mode="str") -> str: - memories = self._rc.memory.get() + memories = self.rc.memory.get() if mode == "str": memories = [f"{m.sent_from}({m.role}): {m.content}" for m in memories] memories = "\n".join(memories) diff --git a/metagpt/ext/werewolf/roles/werewolf.py b/metagpt/ext/werewolf/roles/werewolf.py index 0a0c8a002..4d98eb8fc 100644 --- a/metagpt/ext/werewolf/roles/werewolf.py +++ b/metagpt/ext/werewolf/roles/werewolf.py @@ -10,5 +10,5 @@ class Werewolf(BasePlayer): async def _think(self): """狼人白天发言时需要伪装，与其他角色不同，因此需要重写_think""" await super()._think() - if isinstance(self._rc.todo, Speak): - self._rc.todo = Impersonate() + if isinstance(self.rc.todo, Speak): + self.rc.todo = Impersonate() diff --git a/metagpt/ext/werewolf/roles/witch.py b/metagpt/ext/werewolf/roles/witch.py index 70375052f..207669d64 100644 --- a/metagpt/ext/werewolf/roles/witch.py +++ b/metagpt/ext/werewolf/roles/witch.py @@ -9,18 +9,18 @@ class Witch(BasePlayer): async def _think(self): """女巫涉及两个特殊技能，因此在此需要改写_think进行路由""" - news = self._rc.news[0] + news = self.rc.news[0] assert news.cause_by == InstructSpeak # 消息为来自Moderator的指令时，才去做动作 - if not news.restricted_to: + if not news.send_to: # 消息接收范围为全体角色的，做公开发言（发表投票观点也算发言） - self._rc.todo = Speak() - elif self.profile in news.restricted_to.split(","): + self.rc.todo = Speak() + elif self.profile in news.send_to: # FIXME: hard code to split, restricted为"Moderator"或"Moderator,角色profile" # Moderator加密发给自己的，意味着要执行角色的特殊动作 # 这里用关键词进行动作的选择，需要Moderator侧的指令进行配合 if "save" in news.content.lower(): - self._rc.todo = Save() + self.rc.todo = Save() elif "poison" in news.content.lower(): - self._rc.todo = Poison() + self.rc.todo = Poison() else: raise ValueError("Moderator's instructions must include save or poison keyword") diff --git a/metagpt/ext/werewolf/werewolf_game.py b/metagpt/ext/werewolf/werewolf_game.py index 6609df949..091429e84 100644 --- a/metagpt/ext/werewolf/werewolf_game.py +++ b/metagpt/ext/werewolf/werewolf_game.py @@ -1,42 +1,26 @@ +from typing import Any, Optional + from metagpt.actions.add_requirement import UserRequirement -from metagpt.environment import Environment +from metagpt.context import Context +from metagpt.environment.werewolf.werewolf_env import WerewolfEnv from metagpt.schema import Message from metagpt.team import Team -class WerewolfEnvironment(Environment): - timestamp: int = 0 - - def publish_message(self, message: Message, add_timestamp: bool = True): - """向当前环境发布信息 - Post information to the current environment - """ - # self.message_queue.put(message) - if add_timestamp: - # 因消息内容可能重复，例如，连续两晚杀同一个人， - # 因此需要加一个unique的time_stamp以使得相同的message在加入记忆时不被自动去重 - message.content = f"{self.timestamp} | " + message.content - self.memory.add(message) - self.history += f"\n{message}" - - async def run(self, k=1): - """处理一次所有信息的运行，各角色顺序执行 - Process all Role runs at once - """ - for _ in range(k): - for role in self.roles.values(): - await role.run() - self.timestamp += 1 - - class WerewolfGame(Team): """Use the "software company paradigm" to hold a werewolf game""" - environment = WerewolfEnvironment() + env: Optional[WerewolfEnv] = None - def start_project(self, idea): - """Start a project from user instruction.""" + def __init__(self, context: Context = None, **data: Any): + super(Team, self).__init__(**data) + ctx = context or Context() + if not self.env: + self.env = WerewolfEnv(context=ctx) + else: + self.env.context = ctx # The `env` object is allocated by deserialization + + def run_project(self, idea): + """Run a project from user instruction.""" self.idea = idea - self.environment.publish_message( - Message(role="User", content=idea, cause_by=UserRequirement, restricted_to="Moderator") - ) + self.env.publish_message(Message(role="User", content=idea, cause_by=UserRequirement, send_to={"Moderator"})) diff --git a/tests/metagpt/ext/werewolf/actions/test_experience_operation.py b/tests/metagpt/ext/werewolf/actions/test_experience_operation.py index a91cb5d09..fc2f544a8 100644 --- a/tests/metagpt/ext/werewolf/actions/test_experience_operation.py +++ b/tests/metagpt/ext/werewolf/actions/test_experience_operation.py @@ -122,7 +122,6 @@ class TestActualRetrieve: action = RetrieveExperiences(collection_name=self.collection_name) all_experiences = action.collection.get() logger.info(f"{len(all_experiences['metadatas'])=}") - print(*["metadatas"][-5:], sep="\n") @pytest.mark.asyncio async def test_retrieve_werewolf_experience(self):