feat: merge geekan:main

metagpt/actions/__init__.py

@@ -13,7 +13,6 @@ from metagpt.actions.add_requirement import UserRequirement
 from metagpt.actions.debug_error import DebugError
 from metagpt.actions.design_api import WriteDesign
 from metagpt.actions.design_api_review import DesignReview
-from metagpt.actions.design_filenames import DesignFilenames
 from metagpt.actions.project_management import AssignTasks, WriteTasks
 from metagpt.actions.research import CollectLinks, WebBrowseAndSummarize, ConductResearch
 from metagpt.actions.run_code import RunCode
@@ -33,7 +32,6 @@ class ActionType(Enum):
     WRITE_PRD_REVIEW = WritePRDReview
     WRITE_DESIGN = WriteDesign
     DESIGN_REVIEW = DesignReview
-    DESIGN_FILENAMES = DesignFilenames
     WRTIE_CODE = WriteCode
     WRITE_CODE_REVIEW = WriteCodeReview
     WRITE_TEST = WriteTest

metagpt/actions/action.py

@@ -4,49 +4,60 @@
 @Time    : 2023/5/11 14:43
 @Author  : alexanderwu
 @File    : action.py
-@Modified By: mashenquan, 2023/8/20. Add function return annotations.
-@Modified By: mashenquan, 2023/9/8. Replace LLM with LLMFactory
 """
 
 from __future__ import annotations
 
-from abc import ABC
-from typing import Optional
+from typing import Any, Optional, Union
 
-from tenacity import retry, stop_after_attempt, wait_random_exponential
+from pydantic import BaseModel, Field
 
-from metagpt.actions.action_output import ActionOutput
 from metagpt.llm import LLM
-from metagpt.logs import logger
 from metagpt.provider.base_gpt_api import BaseGPTAPI
-from metagpt.provider.postprecess.llm_output_postprecess import llm_output_postprecess
-from metagpt.utils.common import OutputParser, general_after_log
+from metagpt.schema import (
+    CodeSummarizeContext,
+    CodingContext,
+    RunCodeContext,
+    TestingContext,
+)
+
+action_subclass_registry = {}
 
 
-class Action(ABC):
-    def __init__(self, name: str = "", context=None, llm: BaseGPTAPI = None):
-        self.name: str = name
-        if llm is None:
-            llm = LLM()
-        self.llm = llm
-        self.context = context
-        self.prefix = ""  # aask*时会加上prefix，作为system_message
-        self.profile = ""  # FIXME: USELESS
-        self.desc = ""  # for skill manager
-        self.nodes = ...
+class Action(BaseModel):
+    name: str = ""
+    llm: BaseGPTAPI = Field(default_factory=LLM, exclude=True)
+    context: Union[dict, CodingContext, CodeSummarizeContext, TestingContext, RunCodeContext, str, None] = ""
+    prefix = ""  # aask*时会加上prefix，作为system_message
+    desc = ""  # for skill manager
+    # node: ActionNode = Field(default_factory=ActionNode, exclude=True)
 
-        # Output, useless
-        # self.content = ""
-        # self.instruct_content = None
-        # self.env = None
+    # builtin variables
+    builtin_class_name: str = ""
 
-    # def set_env(self, env):
-    #     self.env = env
+    class Config:
+        arbitrary_types_allowed = True
 
-    def set_prefix(self, prefix, profile):
+    def __init__(self, **kwargs: Any):
+        super().__init__(**kwargs)
+
+        # deserialize child classes dynamically for inherited `action`
+        object.__setattr__(self, "builtin_class_name", self.__class__.__name__)
+        self.__fields__["builtin_class_name"].default = self.__class__.__name__
+
+    def __init_subclass__(cls, **kwargs: Any) -> None:
+        super().__init_subclass__(**kwargs)
+        action_subclass_registry[cls.__name__] = cls
+
+    def dict(self, *args, **kwargs) -> "DictStrAny":
+        obj_dict = super(Action, self).dict(*args, **kwargs)
+        if "llm" in obj_dict:
+            obj_dict.pop("llm")
+        return obj_dict
+
+    def set_prefix(self, prefix):
         """Set prefix for later usage"""
         self.prefix = prefix
-        self.profile = profile
         return self
 
     def __str__(self):
@@ -62,33 +73,6 @@ class Action(ABC):
         system_msgs.append(self.prefix)
         return await self.llm.aask(prompt, system_msgs)
 
-    @retry(
-        wait=wait_random_exponential(min=1, max=60),
-        stop=stop_after_attempt(6),
-        after=general_after_log(logger),
-    )
-    async def _aask_v1(
-        self,
-        prompt: str,
-        output_class_name: str,
-        output_data_mapping: dict,
-        system_msgs: Optional[list[str]] = None,
-        format="markdown",  # compatible to original format
-    ) -> ActionOutput:
-        content = await self.llm.aask(prompt, system_msgs)
-        logger.debug(f"llm raw output:\n{content}")
-        output_class = ActionOutput.create_model_class(output_class_name, output_data_mapping)
-
-        if format == "json":
-            parsed_data = llm_output_postprecess(output=content, schema=output_class.schema(), req_key="[/CONTENT]")
-
-        else:  # using markdown parser
-            parsed_data = OutputParser.parse_data_with_mapping(content, output_data_mapping)
-
-        logger.debug(f"parsed_data:\n{parsed_data}")
-        instruct_content = output_class(**parsed_data)
-        return ActionOutput(content, instruct_content)
-
-    async def run(self, *args, **kwargs) -> str | ActionOutput | None:
+    async def run(self, *args, **kwargs):
         """Run action"""
         raise NotImplementedError("The run method should be implemented in a subclass.")

metagpt/actions/action_node.py

@@ -4,24 +4,26 @@
 @Time    : 2023/12/11 18:45
 @Author  : alexanderwu
 @File    : action_node.py
+
+NOTE: You should use typing.List instead of list to do type annotation. Because in the markdown extraction process,
+  we can use typing to extract the type of the node, but we cannot use built-in list to extract.
 """
 import json
-import re
-from typing import Any, Dict, List, Optional, Type
+from typing import Any, Dict, List, Optional, Tuple, Type
 
 from pydantic import BaseModel, create_model, root_validator, validator
 from tenacity import retry, stop_after_attempt, wait_random_exponential
 
-from metagpt.actions import ActionOutput
 from metagpt.llm import BaseGPTAPI
 from metagpt.logs import logger
-from metagpt.utils.common import OutputParser
-from metagpt.utils.custom_decoder import CustomDecoder
+from metagpt.provider.postprecess.llm_output_postprecess import llm_output_postprecess
+from metagpt.utils.common import OutputParser, general_after_log
+
+TAG = "CONTENT"
+
+LANGUAGE_CONSTRAINT = "Language: Please use the same language as the user input."
+FORMAT_CONSTRAINT = f"Format: output wrapped inside [{TAG}][/{TAG}] like format example, nothing else."
 
-CONSTRAINT = """
-- Language: Please use the same language as the user input.
-- Format: output wrapped inside [CONTENT][/CONTENT] as format example, nothing else.
-"""
 
 SIMPLE_TEMPLATE = """
 ## context
@@ -32,21 +34,21 @@ SIMPLE_TEMPLATE = """
 ## format example
 {example}
 
-## nodes: "<node>: <type>  # <comment>"
+## nodes: "<node>: <type>  # <instruction>"
 {instruction}
 
 ## constraint
 {constraint}
 
 ## action
-Based on the 'context' content, fill in the {node_name} using the 'format example' format above."
+Follow instructions of nodes, generate output and make sure it follows the format example.
 """
 
 
-def dict_to_markdown(d, prefix="###", postfix="\n"):
+def dict_to_markdown(d, prefix="- ", kv_sep="\n", postfix="\n"):
     markdown_str = ""
     for key, value in d.items():
-        markdown_str += f"{prefix} {key}: {value}{postfix}"
+        markdown_str += f"{prefix}{key}{kv_sep}{value}{postfix}"
     return markdown_str
 
 
@@ -76,7 +78,7 @@ class ActionNode:
         key: str,
         expected_type: Type,
         instruction: str,
-        example: str,
+        example: Any,
         content: str = "",
         children: dict[str, "ActionNode"] = None,
     ):
@@ -111,22 +113,22 @@ class ActionNode:
         obj.add_children(nodes)
         return obj
 
-    def get_children_mapping(self) -> Dict[str, Type]:
+    def get_children_mapping(self) -> Dict[str, Tuple[Type, Any]]:
         """获得子ActionNode的字典，以key索引"""
         return {k: (v.expected_type, ...) for k, v in self.children.items()}
 
-    def get_self_mapping(self) -> Dict[str, Type]:
+    def get_self_mapping(self) -> Dict[str, Tuple[Type, Any]]:
         """get self key: type mapping"""
         return {self.key: (self.expected_type, ...)}
 
-    def get_mapping(self, mode="children") -> Dict[str, Type]:
+    def get_mapping(self, mode="children") -> Dict[str, Tuple[Type, Any]]:
         """get key: type mapping under mode"""
         if mode == "children" or (mode == "auto" and self.children):
             return self.get_children_mapping()
         return self.get_self_mapping()
 
     @classmethod
-    def create_model_class(cls, class_name: str, mapping: Dict[str, Type]):
+    def create_model_class(cls, class_name: str, mapping: Dict[str, Tuple[Type, Any]]):
         """基于pydantic v1的模型动态生成，用来检验结果类型正确性"""
         new_class = create_model(class_name, **mapping)
 
@@ -148,29 +150,6 @@ class ActionNode:
         new_class.__root_validator_check_missing_fields = classmethod(check_missing_fields)
         return new_class
 
-    @classmethod
-    def create_model_class_v2(cls, class_name: str, mapping: Dict[str, Type]):
-        """基于pydantic v2的模型动态生成，用来检验结果类型正确性，待验证"""
-        new_class = create_model(class_name, **mapping)
-
-        @model_validator(mode="before")
-        def check_missing_fields(data):
-            required_fields = set(mapping.keys())
-            missing_fields = required_fields - set(data.keys())
-            if missing_fields:
-                raise ValueError(f"Missing fields: {missing_fields}")
-            return data
-
-        @field_validator("*")
-        def check_name(v: Any, field: str) -> Any:
-            if field not in mapping.keys():
-                raise ValueError(f"Unrecognized block: {field}")
-            return v
-
-        new_class.__model_validator_check_missing_fields = classmethod(check_missing_fields)
-        new_class.__field_validator_check_name = classmethod(check_name)
-        return new_class
-
     def create_children_class(self):
         """使用object内有的字段直接生成model_class"""
         class_name = f"{self.key}_AN"
@@ -197,46 +176,46 @@ class ActionNode:
             return node_dict
 
         # 遍历子节点并递归调用 to_dict 方法
-        for child_key, child_node in self.children.items():
+        for _, child_node in self.children.items():
             node_dict.update(child_node.to_dict(format_func))
 
         return node_dict
 
-    def compile_to(self, i: Dict, to) -> str:
-        if to == "json":
+    def compile_to(self, i: Dict, schema, kv_sep) -> str:
+        if schema == "json":
             return json.dumps(i, indent=4)
-        elif to == "markdown":
-            return dict_to_markdown(i)
+        elif schema == "markdown":
+            return dict_to_markdown(i, kv_sep=kv_sep)
         else:
             return str(i)
 
-    def tagging(self, text, to, tag="") -> str:
+    def tagging(self, text, schema, tag="") -> str:
         if not tag:
             return text
-        if to == "json":
+        if schema == "json":
             return f"[{tag}]\n" + text + f"\n[/{tag}]"
-        else:
+        else:  # markdown
             return f"[{tag}]\n" + text + f"\n[/{tag}]"
 
-    def _compile_f(self, to, mode, tag, format_func) -> str:
+    def _compile_f(self, schema, mode, tag, format_func, kv_sep) -> str:
         nodes = self.to_dict(format_func=format_func, mode=mode)
-        text = self.compile_to(nodes, to)
-        return self.tagging(text, to, tag)
+        text = self.compile_to(nodes, schema, kv_sep)
+        return self.tagging(text, schema, tag)
 
-    def compile_instruction(self, to="raw", mode="children", tag="") -> str:
+    def compile_instruction(self, schema="markdown", mode="children", tag="") -> str:
         """compile to raw/json/markdown template with all/root/children nodes"""
         format_func = lambda i: f"{i.expected_type}  # {i.instruction}"
-        return self._compile_f(to, mode, tag, format_func)
+        return self._compile_f(schema, mode, tag, format_func, kv_sep=": ")
 
-    def compile_example(self, to="raw", mode="children", tag="") -> str:
+    def compile_example(self, schema="json", mode="children", tag="") -> str:
         """compile to raw/json/markdown examples with all/root/children nodes"""
 
         # 这里不能使用f-string，因为转译为str后再json.dumps会额外加上引号，无法作为有效的example
         # 错误示例："File list": "['main.py', 'const.py', 'game.py']", 注意这里值不是list，而是str
         format_func = lambda i: i.example
-        return self._compile_f(to, mode, tag, format_func)
+        return self._compile_f(schema, mode, tag, format_func, kv_sep="\n")
 
-    def compile(self, context, to="json", mode="children", template=SIMPLE_TEMPLATE) -> str:
+    def compile(self, context, schema="json", mode="children", template=SIMPLE_TEMPLATE) -> str:
         """
         mode: all/root/children
             mode="children": 编译所有子节点为一个统一模板，包括instruction与example
@@ -245,48 +224,47 @@ class ActionNode:
         """
 
         # FIXME: json instruction会带来格式问题，如："Project name": "web_2048  # 项目名称使用下划线",
-        self.instruction = self.compile_instruction(to="markdown", mode=mode)
-        self.example = self.compile_example(to=to, tag="CONTENT", mode=mode)
-        node_name = "nodes" if template != SIMPLE_TEMPLATE else f'"{list(self.children.keys())[0]}" node'
+        # compile example暂时不支持markdown
+        self.instruction = self.compile_instruction(schema="markdown", mode=mode)
+        self.example = self.compile_example(schema=schema, tag=TAG, mode=mode)
+        # nodes = ", ".join(self.to_dict(mode=mode).keys())
+        constraints = [LANGUAGE_CONSTRAINT, FORMAT_CONSTRAINT]
+        constraint = "\n".join(constraints)
+
         prompt = template.format(
             context=context,
             example=self.example,
             instruction=self.instruction,
-            constraint=CONSTRAINT,
-            node_name=node_name,
+            constraint=constraint,
         )
         return prompt
 
-    @retry(wait=wait_random_exponential(min=1, max=10), stop=stop_after_attempt(6))
+    @retry(
+        wait=wait_random_exponential(min=1, max=20),
+        stop=stop_after_attempt(6),
+        after=general_after_log(logger),
+    )
     async def _aask_v1(
         self,
         prompt: str,
         output_class_name: str,
         output_data_mapping: dict,
         system_msgs: Optional[list[str]] = None,
-        format="markdown",  # compatible to original format
-    ) -> ActionOutput:
+        schema="markdown",  # compatible to original format
+    ) -> (str, BaseModel):
+        """Use ActionOutput to wrap the output of aask"""
         content = await self.llm.aask(prompt, system_msgs)
-        logger.debug(content)
-        output_class = ActionOutput.create_model_class(output_class_name, output_data_mapping)
-
-        if format == "json":
-            pattern = r"\[CONTENT\](\s*\{.*?\}\s*)\[/CONTENT\]"
-            matches = re.findall(pattern, content, re.DOTALL)
-
-            for match in matches:
-                if match:
-                    content = match
-                    break
-
-            parsed_data = CustomDecoder(strict=False).decode(content)
+        logger.debug(f"llm raw output:\n{content}")
+        output_class = self.create_model_class(output_class_name, output_data_mapping)
 
+        if schema == "json":
+            parsed_data = llm_output_postprecess(output=content, schema=output_class.schema(), req_key=f"[/{TAG}]")
         else:  # using markdown parser
             parsed_data = OutputParser.parse_data_with_mapping(content, output_data_mapping)
 
-        logger.debug(parsed_data)
+        logger.debug(f"parsed_data:\n{parsed_data}")
         instruct_content = output_class(**parsed_data)
-        return ActionOutput(content, instruct_content)
+        return content, instruct_content
 
     def get(self, key):
         return self.instruct_content.dict()[key]
@@ -302,23 +280,22 @@ class ActionNode:
     def set_context(self, context):
         self.set_recursive("context", context)
 
-    async def simple_fill(self, to, mode):
-        prompt = self.compile(context=self.context, to=to, mode=mode)
+    async def simple_fill(self, schema, mode):
+        prompt = self.compile(context=self.context, schema=schema, mode=mode)
         mapping = self.get_mapping(mode)
 
         class_name = f"{self.key}_AN"
-        print(prompt)
-        output = await self._aask_v1(prompt, class_name, mapping, format=to)
-        self.content = output.content
-        self.instruct_content = output.instruct_content
+        content, scontent = await self._aask_v1(prompt, class_name, mapping, schema=schema)
+        self.content = content
+        self.instruct_content = scontent
         return self
 
-    async def fill(self, context, llm, to="json", mode="auto", strgy="simple"):
+    async def fill(self, context, llm, schema="json", mode="auto", strgy="simple"):
         """Fill the node(s) with mode.
 
         :param context: Everything we should know when filling node.
         :param llm: Large Language Model with pre-defined system message.
-        :param to: json/markdown, determine example and output format.
+        :param schema: json/markdown, determine example and output format.
          - json: it's easy to open source LLM with json format
          - markdown: when generating code, markdown is always better
         :param mode: auto/children/root
@@ -334,12 +311,12 @@ class ActionNode:
         self.set_context(context)
 
         if strgy == "simple":
-            return await self.simple_fill(to, mode)
+            return await self.simple_fill(schema, mode)
         elif strgy == "complex":
             # 这里隐式假设了拥有children
             tmp = {}
             for _, i in self.children.items():
-                child = await i.simple_fill(to, mode)
+                child = await i.simple_fill(schema, mode)
                 tmp.update(child.instruct_content.dict())
             cls = self.create_children_class()
             self.instruct_content = cls(**tmp)

metagpt/actions/action_output.py

@@ -4,40 +4,15 @@
 @Time    : 2023/7/11 10:03
 @Author  : chengmaoyu
 @File    : action_output
-@Modified By: mashenquan, 2023/8/20. Allow 'instruct_content' to be blank.
 """
 
-from typing import Dict, Optional, Type
-
-from pydantic import BaseModel, create_model, root_validator, validator
+from pydantic import BaseModel
 
 
 class ActionOutput:
     content: str
-    instruct_content: Optional[BaseModel] = None
+    instruct_content: BaseModel
 
-    def __init__(self, content: str, instruct_content: BaseModel = None):
+    def __init__(self, content: str, instruct_content: BaseModel):
         self.content = content
         self.instruct_content = instruct_content
-
-    @classmethod
-    def create_model_class(cls, class_name: str, mapping: Dict[str, Type]):
-        new_class = create_model(class_name, **mapping)
-
-        @validator("*", allow_reuse=True)
-        def check_name(v, field):
-            if field.name not in mapping.keys():
-                raise ValueError(f"Unrecognized block: {field.name}")
-            return v
-
-        @root_validator(pre=True, allow_reuse=True)
-        def check_missing_fields(values):
-            required_fields = set(mapping.keys())
-            missing_fields = required_fields - set(values.keys())
-            if missing_fields:
-                raise ValueError(f"Missing fields: {missing_fields}")
-            return values
-
-        new_class.__validator_check_name = classmethod(check_name)
-        new_class.__root_validator_check_missing_fields = classmethod(check_missing_fields)
-        return new_class

metagpt/actions/analyze_dep_libs.py

@@ -1,37 +0,0 @@
-#!/usr/bin/env python
-# -*- coding: utf-8 -*-
-"""
-@Time    : 2023/5/19 12:01
-@Author  : alexanderwu
-@File    : analyze_dep_libs.py
-"""
-
-from metagpt.actions import Action
-
-PROMPT = """You are an AI developer, trying to write a program that generates code for users based on their intentions.
-
-For the user's prompt:
-
----
-The API is: {prompt}
----
-
-We decide the generated files are: {filepaths_string}
-
-Now that we have a file list, we need to understand the shared dependencies they have.
-Please list and briefly describe the shared contents between the files we are generating, including exported variables, 
-data patterns, id names of all DOM elements that javascript functions will use, message names and function names.
-Focus only on the names of shared dependencies, do not add any other explanations.
-"""
-
-
-class AnalyzeDepLibs(Action):
-    def __init__(self, name, context=None, llm=None):
-        super().__init__(name, context, llm)
-        self.desc = "Analyze the runtime dependencies of the program based on the context"
-
-    async def run(self, requirement, filepaths_string):
-        # prompt = f"Below is the product requirement document (PRD):\n\n{prd}\n\n{PROMPT}"
-        prompt = PROMPT.format(prompt=requirement, filepaths_string=filepaths_string)
-        design_filenames = await self._aask(prompt)
-        return design_filenames

metagpt/actions/debug_error.py

@@ -10,11 +10,14 @@
 """
 import re
 
+from pydantic import Field
+
 from metagpt.actions.action import Action
 from metagpt.config import CONFIG
 from metagpt.const import TEST_CODES_FILE_REPO, TEST_OUTPUTS_FILE_REPO
+from metagpt.llm import LLM, BaseGPTAPI
 from metagpt.logs import logger
-from metagpt.schema import RunCodeResult
+from metagpt.schema import RunCodeContext, RunCodeResult
 from metagpt.utils.common import CodeParser
 from metagpt.utils.file_repository import FileRepository
 
@@ -47,8 +50,9 @@ Now you should start rewriting the code:
 
 
 class DebugError(Action):
-    def __init__(self, name="DebugError", context=None, llm=None):
-        super().__init__(name, context, llm)
+    name: str = "DebugError"
+    context: RunCodeContext = Field(default_factory=RunCodeContext)
+    llm: BaseGPTAPI = Field(default_factory=LLM)
 
     async def run(self, *args, **kwargs) -> str:
         output_doc = await FileRepository.get_file(

metagpt/actions/design_api.py

@@ -11,6 +11,9 @@
 """
 import json
 from pathlib import Path
+from typing import Optional
+
+from pydantic import Field
 
 from metagpt.actions import Action, ActionOutput
 from metagpt.actions.design_api_an import DESIGN_API_NODE
@@ -22,16 +25,13 @@ from metagpt.const import (
     SYSTEM_DESIGN_FILE_REPO,
     SYSTEM_DESIGN_PDF_FILE_REPO,
 )
+from metagpt.llm import LLM
 from metagpt.logs import logger
-from metagpt.schema import Document, Documents
+from metagpt.provider.base_gpt_api import BaseGPTAPI
+from metagpt.schema import Document, Documents, Message
 from metagpt.utils.file_repository import FileRepository
-
-# from metagpt.utils.get_template import get_template
 from metagpt.utils.mermaid import mermaid_to_file
 
-# from typing import List
-
-
 NEW_REQ_TEMPLATE = """
 ### Legacy Content
 {old_design}
@@ -42,15 +42,16 @@ NEW_REQ_TEMPLATE = """
 
 
 class WriteDesign(Action):
-    def __init__(self, name, context=None, llm=None):
-        super().__init__(name, context, llm)
-        self.desc = (
-            "Based on the PRD, think about the system design, and design the corresponding APIs, "
-            "data structures, library tables, processes, and paths. Please provide your design, feedback "
-            "clearly and in detail."
-        )
+    name: str = ""
+    context: Optional[str] = None
+    llm: BaseGPTAPI = Field(default_factory=LLM)
+    desc: str = (
+        "Based on the PRD, think about the system design, and design the corresponding APIs, "
+        "data structures, library tables, processes, and paths. Please provide your design, feedback "
+        "clearly and in detail."
+    )
 
-    async def run(self, with_messages, format=CONFIG.prompt_format):
+    async def run(self, with_messages: Message, schema: str = CONFIG.prompt_schema):
         # Use `git diff` to identify which PRD documents have been modified in the `docs/prds` directory.
         prds_file_repo = CONFIG.git_repo.new_file_repository(PRDS_FILE_REPO)
         changed_prds = prds_file_repo.changed_files
@@ -80,13 +81,13 @@ class WriteDesign(Action):
         # leaving room for global optimization in subsequent steps.
         return ActionOutput(content=changed_files.json(), instruct_content=changed_files)
 
-    async def _new_system_design(self, context, format=CONFIG.prompt_format):
-        node = await DESIGN_API_NODE.fill(context=context, llm=self.llm, to=format)
+    async def _new_system_design(self, context, schema=CONFIG.prompt_schema):
+        node = await DESIGN_API_NODE.fill(context=context, llm=self.llm, schema=schema)
         return node
 
-    async def _merge(self, prd_doc, system_design_doc, format=CONFIG.prompt_format):
+    async def _merge(self, prd_doc, system_design_doc, schema=CONFIG.prompt_schema):
         context = NEW_REQ_TEMPLATE.format(old_design=system_design_doc.content, context=prd_doc.content)
-        node = await DESIGN_API_NODE.fill(context=context, llm=self.llm, to=format)
+        node = await DESIGN_API_NODE.fill(context=context, llm=self.llm, schema=schema)
         system_design_doc.content = node.instruct_content.json(ensure_ascii=False)
         return system_design_doc
 

metagpt/actions/design_api_an.py

@@ -5,6 +5,8 @@
 @Author  : alexanderwu
 @File    : design_api_an.py
 """
+from typing import List
+
 from metagpt.actions.action_node import ActionNode
 from metagpt.logs import logger
 from metagpt.utils.mermaid import MMC1, MMC2
@@ -22,7 +24,7 @@ PROJECT_NAME = ActionNode(
 
 FILE_LIST = ActionNode(
     key="File list",
-    expected_type=list[str],
+    expected_type=List[str],
     instruction="Only need relative paths. ALWAYS write a main.py or app.py here",
     example=["main.py", "game.py"],
 )

metagpt/actions/design_filenames.py

@@ -1,30 +0,0 @@
-#!/usr/bin/env python
-# -*- coding: utf-8 -*-
-"""
-@Time    : 2023/5/19 11:50
-@Author  : alexanderwu
-@File    : design_filenames.py
-"""
-from metagpt.actions import Action
-from metagpt.logs import logger
-
-PROMPT = """You are an AI developer, trying to write a program that generates code for users based on their intentions.
-When given their intentions, provide a complete and exhaustive list of file paths needed to write the program for the user.
-Only list the file paths you will write and return them as a Python string list.
-Do not add any other explanations, just return a Python string list."""
-
-
-class DesignFilenames(Action):
-    def __init__(self, name, context=None, llm=None):
-        super().__init__(name, context, llm)
-        self.desc = (
-            "Based on the PRD, consider system design, and carry out the basic design of the corresponding "
-            "APIs, data structures, and database tables. Please give your design, feedback clearly and in detail."
-        )
-
-    async def run(self, prd):
-        prompt = f"The following is the Product Requirement Document (PRD):\n\n{prd}\n\n{PROMPT}"
-        design_filenames = await self._aask(prompt)
-        logger.debug(prompt)
-        logger.debug(design_filenames)
-        return design_filenames

metagpt/actions/detail_mining.py

@@ -1,51 +0,0 @@
-#!/usr/bin/env python
-# -*- coding: utf-8 -*-
-"""
-@Time    : 2023/9/12 17:45
-@Author  : fisherdeng
-@File    : detail_mining.py
-"""
-from metagpt.actions import Action, ActionOutput
-
-PROMPT_TEMPLATE = """
-##TOPIC
-{topic}
-
-##RECORD
-{record}
-
-##Format example
-{format_example}
------
-
-Task: Refer to the "##TOPIC" (discussion objectives) and "##RECORD" (discussion records) to further inquire about the details that interest you, within a word limit of 150 words.
-Special Note 1: Your intention is solely to ask questions without endorsing or negating any individual's viewpoints.
-Special Note 2: This output should only include the topic "##OUTPUT". Do not add, remove, or modify the topic. Begin the output with '##OUTPUT', followed by an immediate line break, and then proceed to provide the content in the specified format as outlined in the "##Format example" section.
-Special Note 3: The output should be in the same language as the input.
-"""
-FORMAT_EXAMPLE = """
-
-##
-
-##OUTPUT
-...(Please provide the specific details you would like to inquire about here.)
-
-##
-
-##
-"""
-OUTPUT_MAPPING = {
-    "OUTPUT": (str, ...),
-}
-
-
-class DetailMining(Action):
-    """This class allows LLM to further mine noteworthy details based on specific "##TOPIC"(discussion topic) and "##RECORD" (discussion records), thereby deepening the discussion."""
-
-    def __init__(self, name="", context=None, llm=None):
-        super().__init__(name, context, llm)
-
-    async def run(self, topic, record) -> ActionOutput:
-        prompt = PROMPT_TEMPLATE.format(topic=topic, record=record, format_example=FORMAT_EXAMPLE)
-        rsp = await self._aask_v1(prompt, "detail_mining", OUTPUT_MAPPING)
-        return rsp

metagpt/actions/fix_bug.py

@@ -10,5 +10,7 @@ from metagpt.actions import Action
 class FixBug(Action):
     """Fix bug action without any implementation details"""
 
+    name: str = "FixBug"
+
     async def run(self, *args, **kwargs):
         raise NotImplementedError

metagpt/actions/generate_questions.py

@@ -0,0 +1,25 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+"""
+@Time    : 2023/9/12 17:45
+@Author  : fisherdeng
+@File    : generate_questions.py
+"""
+from metagpt.actions import Action
+from metagpt.actions.action_node import ActionNode
+
+QUESTIONS = ActionNode(
+    key="Questions",
+    expected_type=list[str],
+    instruction="Task: Refer to the context to further inquire about the details that interest you, within a word limit"
+    " of 150 words. Please provide the specific details you would like to inquire about here",
+    example=["1. What ...", "2. How ...", "3. ..."],
+)
+
+
+class GenerateQuestions(Action):
+    """This class allows LLM to further mine noteworthy details based on specific "##TOPIC"(discussion topic) and
+    "##RECORD" (discussion records), thereby deepening the discussion."""
+
+    async def run(self, context):
+        return await QUESTIONS.fill(context=context, llm=self.llm)

metagpt/actions/prepare_documents.py

@@ -9,31 +9,41 @@
 """
 import shutil
 from pathlib import Path
+from typing import Optional
+
+from pydantic import Field
 
 from metagpt.actions import Action, ActionOutput
 from metagpt.config import CONFIG
-from metagpt.const import DEFAULT_WORKSPACE_ROOT, DOCS_FILE_REPO, REQUIREMENT_FILENAME
+from metagpt.const import DOCS_FILE_REPO, REQUIREMENT_FILENAME
+from metagpt.llm import LLM
+from metagpt.provider.base_gpt_api import BaseGPTAPI
 from metagpt.schema import Document
 from metagpt.utils.file_repository import FileRepository
 from metagpt.utils.git_repository import GitRepository
 
 
 class PrepareDocuments(Action):
-    def __init__(self, name="", context=None, llm=None):
-        super().__init__(name, context, llm)
+    """PrepareDocuments Action: initialize project folder and add new requirements to docs/requirements.txt."""
+
+    name: str = "PrepareDocuments"
+    context: Optional[str] = None
+    llm: BaseGPTAPI = Field(default_factory=LLM)
+
+    def _init_repo(self):
+        """Initialize the Git environment."""
+        path = CONFIG.project_path
+        if not path:
+            name = CONFIG.project_name or FileRepository.new_filename()
+            path = Path(CONFIG.workspace_path) / name
+
+        if path.exists() and not CONFIG.inc:
+            shutil.rmtree(path)
+        CONFIG.git_repo = GitRepository(local_path=path, auto_init=True)
 
     async def run(self, with_messages, **kwargs):
-        if not CONFIG.git_repo:
-            # Create and initialize the workspace folder, initialize the Git environment.
-            project_name = CONFIG.project_name or FileRepository.new_filename()
-            workdir = CONFIG.project_path
-            if not workdir and CONFIG.workspace_path:
-                workdir = Path(CONFIG.workspace_path) / project_name
-            workdir = Path(workdir or DEFAULT_WORKSPACE_ROOT / project_name)
-            if not CONFIG.inc and workdir.exists():
-                shutil.rmtree(workdir)
-            CONFIG.git_repo = GitRepository()
-            CONFIG.git_repo.open(local_path=workdir, auto_init=True)
+        """Create and initialize the workspace folder, initialize the Git environment."""
+        self._init_repo()
 
         # Write the newly added requirements from the main parameter idea to `docs/requirement.txt`.
         doc = Document(root_path=DOCS_FILE_REPO, filename=REQUIREMENT_FILENAME, content=with_messages[0].content)

metagpt/actions/prepare_interview.py

@@ -6,35 +6,18 @@
 @File    : prepare_interview.py
 """
 from metagpt.actions import Action
+from metagpt.actions.action_node import ActionNode
 
-PROMPT_TEMPLATE = """
-# Context
-{context}
-
-## Format example
----
-Q1: question 1 here
-References:
-  - point 1
-  - point 2
-
-Q2: question 2 here...
----
-
------
-Role: You are an interviewer of our company who is well-knonwn in frontend or backend develop;
+QUESTIONS = ActionNode(
+    key="Questions",
+    expected_type=list[str],
+    instruction="""Role: You are an interviewer of our company who is well-knonwn in frontend or backend develop;
 Requirement: Provide a list of questions for the interviewer to ask the interviewee, by reading the resume of the interviewee in the context.
-Attention: Provide as markdown block as the format above, at least 10 questions.
-"""
-
-# prepare for a interview
+Attention: Provide as markdown block as the format above, at least 10 questions.""",
+    example=["1. What ...", "2. How ..."],
+)
 
 
 class PrepareInterview(Action):
-    def __init__(self, name, context=None, llm=None):
-        super().__init__(name, context, llm)
-
     async def run(self, context):
-        prompt = PROMPT_TEMPLATE.format(context=context)
-        question_list = await self._aask_v1(prompt)
-        return question_list
+        return await QUESTIONS.fill(context=context, llm=self.llm)

metagpt/actions/project_management.py

@@ -9,7 +9,11 @@
         2. Move the document storage operations related to WritePRD from the save operation of WriteDesign.
         3. According to the design in Section 2.2.3.5.4 of RFC 135, add incremental iteration functionality.
 """
+
 import json
+from typing import Optional
+
+from pydantic import Field
 
 from metagpt.actions import ActionOutput
 from metagpt.actions.action import Action
@@ -21,7 +25,9 @@ from metagpt.const import (
     TASK_FILE_REPO,
     TASK_PDF_FILE_REPO,
 )
+from metagpt.llm import LLM
 from metagpt.logs import logger
+from metagpt.provider.base_gpt_api import BaseGPTAPI
 from metagpt.schema import Document, Documents
 from metagpt.utils.file_repository import FileRepository
 
@@ -35,10 +41,11 @@ NEW_REQ_TEMPLATE = """
 
 
 class WriteTasks(Action):
-    def __init__(self, name="CreateTasks", context=None, llm=None):
-        super().__init__(name, context, llm)
+    name: str = "CreateTasks"
+    context: Optional[str] = None
+    llm: BaseGPTAPI = Field(default_factory=LLM)
 
-    async def run(self, with_messages, format=CONFIG.prompt_format):
+    async def run(self, with_messages, schema=CONFIG.prompt_schema):
         system_design_file_repo = CONFIG.git_repo.new_file_repository(SYSTEM_DESIGN_FILE_REPO)
         changed_system_designs = system_design_file_repo.changed_files
 
@@ -85,16 +92,16 @@ class WriteTasks(Action):
         await self._save_pdf(task_doc=task_doc)
         return task_doc
 
-    async def _run_new_tasks(self, context, format=CONFIG.prompt_format):
-        node = await PM_NODE.fill(context, self.llm, format)
+    async def _run_new_tasks(self, context, schema=CONFIG.prompt_schema):
+        node = await PM_NODE.fill(context, self.llm, schema)
         # prompt_template, format_example = get_template(templates, format)
         # prompt = prompt_template.format(context=context, format_example=format_example)
         # rsp = await self._aask_v1(prompt, "task", OUTPUT_MAPPING, format=format)
         return node
 
-    async def _merge(self, system_design_doc, task_doc, format=CONFIG.prompt_format) -> Document:
+    async def _merge(self, system_design_doc, task_doc, schema=CONFIG.prompt_schema) -> Document:
         context = NEW_REQ_TEMPLATE.format(context=system_design_doc.content, old_tasks=task_doc.content)
-        node = await PM_NODE.fill(context, self.llm, format)
+        node = await PM_NODE.fill(context, self.llm, schema)
         task_doc.content = node.instruct_content.json(ensure_ascii=False)
         return task_doc
 

metagpt/actions/project_management_an.py

@@ -5,26 +5,28 @@
 @Author  : alexanderwu
 @File    : project_management_an.py
 """
+from typing import List
+
 from metagpt.actions.action_node import ActionNode
 from metagpt.logs import logger
 
 REQUIRED_PYTHON_PACKAGES = ActionNode(
     key="Required Python packages",
-    expected_type=list[str],
+    expected_type=List[str],
     instruction="Provide required Python packages in requirements.txt format.",
     example=["flask==1.1.2", "bcrypt==3.2.0"],
 )
 
 REQUIRED_OTHER_LANGUAGE_PACKAGES = ActionNode(
     key="Required Other language third-party packages",
-    expected_type=list[str],
+    expected_type=List[str],
     instruction="List down the required packages for languages other than Python.",
     example=["No third-party dependencies required"],
 )
 
 LOGIC_ANALYSIS = ActionNode(
     key="Logic Analysis",
-    expected_type=list[list[str]],
+    expected_type=List[List[str]],
     instruction="Provide a list of files with the classes/methods/functions to be implemented, "
     "including dependency analysis and imports.",
     example=[
@@ -35,7 +37,7 @@ LOGIC_ANALYSIS = ActionNode(
 
 TASK_LIST = ActionNode(
     key="Task list",
-    expected_type=list[str],
+    expected_type=List[str],
     instruction="Break down the tasks into a list of filenames, prioritized by dependency order.",
     example=["game.py", "main.py"],
 )

metagpt/actions/run_code.py

@@ -18,10 +18,13 @@
 import subprocess
 from typing import Tuple
 
+from pydantic import Field
+
 from metagpt.actions.action import Action
 from metagpt.config import CONFIG
+from metagpt.llm import LLM, BaseGPTAPI
 from metagpt.logs import logger
-from metagpt.schema import RunCodeResult
+from metagpt.schema import RunCodeContext, RunCodeResult
 from metagpt.utils.exceptions import handle_exception
 
 PROMPT_TEMPLATE = """
@@ -74,8 +77,9 @@ standard errors:
 
 
 class RunCode(Action):
-    def __init__(self, name="RunCode", context=None, llm=None):
-        super().__init__(name, context, llm)
+    name: str = "RunCode"
+    context: RunCodeContext = Field(default_factory=RunCodeContext)
+    llm: BaseGPTAPI = Field(default_factory=LLM)
 
     @classmethod
     @handle_exception

metagpt/actions/search_and_summarize.py

@@ -4,14 +4,19 @@
 @Time    : 2023/5/23 17:26
 @Author  : alexanderwu
 @File    : search_google.py
-@Modified By: mashenquan, 2023/8/20. Remove global configuration `CONFIG`, enable configuration support for business isolation.
 """
+from typing import Optional
+
 import pydantic
+from pydantic import Field, root_validator
 
 from metagpt.actions import Action
-from metagpt.config import CONFIG
+from metagpt.config import CONFIG, Config
+from metagpt.llm import LLM
 from metagpt.logs import logger
+from metagpt.provider.base_gpt_api import BaseGPTAPI
 from metagpt.schema import Message
+from metagpt.tools import SearchEngineType
 from metagpt.tools.search_engine import SearchEngine
 
 SEARCH_AND_SUMMARIZE_SYSTEM = """### Requirements
@@ -55,7 +60,6 @@ SEARCH_AND_SUMMARIZE_PROMPT = """
 
 """
 
-
 SEARCH_AND_SUMMARIZE_SALES_SYSTEM = """## Requirements
 1. Please summarize the latest dialogue based on the reference information (secondary) and dialogue history (primary). Do not include text that is irrelevant to the conversation.
 - The context is for reference only. If it is irrelevant to the user's search request history, please reduce its reference and usage.
@@ -102,16 +106,31 @@ You are a member of a professional butler team and will provide helpful suggesti
 
 
 class SearchAndSummarize(Action):
-    def __init__(self, name="", context=None, llm=None, engine=None, search_func=None):
-        self.engine = engine or CONFIG.search_engine
+    name: str = ""
+    content: Optional[str] = None
+    llm: BaseGPTAPI = Field(default_factory=LLM)
+    config: None = Field(default_factory=Config)
+    engine: Optional[SearchEngineType] = CONFIG.search_engine
+    search_func: Optional[str] = None
+    search_engine: SearchEngine = None
 
+    result = ""
+
+    @root_validator
+    def validate_engine_and_run_func(cls, values):
+        engine = values.get("engine")
+        search_func = values.get("search_func")
+        config = Config()
+
+        if engine is None:
+            engine = config.search_engine
         try:
-            self.search_engine = SearchEngine(self.engine, run_func=search_func)
+            search_engine = SearchEngine(engine=engine, run_func=search_func)
         except pydantic.ValidationError:
-            self.search_engine = None
+            search_engine = None
 
-        self.result = ""
-        super().__init__(name, context, llm)
+        values["search_engine"] = search_engine
+        return values
 
     async def run(self, context: list[Message], system_text=SEARCH_AND_SUMMARIZE_SYSTEM) -> str:
         if self.search_engine is None:
@@ -130,8 +149,7 @@ class SearchAndSummarize(Action):
         system_prompt = [system_text]
 
         prompt = SEARCH_AND_SUMMARIZE_PROMPT.format(
-            # PREFIX = self.prefix,
-            ROLE=self.profile,
+            ROLE=self.prefix,
             CONTEXT=rsp,
             QUERY_HISTORY="\n".join([str(i) for i in context[:-1]]),
             QUERY=str(context[-1]),

metagpt/actions/skill_action.py

@@ -12,6 +12,7 @@ import ast
 import importlib
 import traceback
 from copy import deepcopy
+from typing import Dict, Optional
 
 from metagpt.actions import Action, ActionOutput
 from metagpt.learn.skill_loader import Skill
@@ -19,12 +20,10 @@ from metagpt.logs import logger
 
 
 class ArgumentsParingAction(Action):
-    def __init__(self, last_talk: str, skill: Skill, context=None, llm=None, **kwargs):
-        super(ArgumentsParingAction, self).__init__(name="", context=context, llm=llm)
-        self.skill = skill
-        self.ask = last_talk
-        self.rsp = None
-        self.args = None
+    skill: Skill
+    ask: str
+    rsp: Optional[ActionOutput]
+    args: Optional[Dict]
 
     @property
     def prompt(self):
@@ -70,25 +69,23 @@ class ArgumentsParingAction(Action):
 
 
 class SkillAction(Action):
-    def __init__(self, skill: Skill, args: dict, context=None, llm=None, **kwargs):
-        super(SkillAction, self).__init__(name="", context=context, llm=llm)
-        self._skill = skill
-        self._args = args
-        self.rsp = None
+    skill: Skill
+    args: Dict
+    rsp: str = ""
 
     async def run(self, *args, **kwargs) -> str | ActionOutput | None:
         """Run action"""
         options = deepcopy(kwargs)
-        if self._args:
-            for k in self._args.keys():
+        if self.args:
+            for k in self.args.keys():
                 if k in options:
                     options.pop(k)
         try:
-            self.rsp = await self.find_and_call_function(self._skill.name, args=self._args, **options)
+            self.rsp = await self.find_and_call_function(self.skill.name, args=self.args, **options)
         except Exception as e:
             logger.exception(f"{e}, traceback:{traceback.format_exc()}")
             self.rsp = f"Error: {e}"
-        return ActionOutput(content=self.rsp, instruct_content=self._skill.json())
+        return ActionOutput(content=self.rsp, instruct_content=self.skill.json())
 
     @staticmethod
     async def find_and_call_function(function_name, args, **kwargs):

metagpt/actions/summarize_code.py

@@ -7,12 +7,15 @@
 """
 from pathlib import Path
 
+from pydantic import Field
 from tenacity import retry, stop_after_attempt, wait_random_exponential
 
 from metagpt.actions.action import Action
 from metagpt.config import CONFIG
 from metagpt.const import SYSTEM_DESIGN_FILE_REPO, TASK_FILE_REPO
+from metagpt.llm import LLM, BaseGPTAPI
 from metagpt.logs import logger
+from metagpt.schema import CodeSummarizeContext
 from metagpt.utils.file_repository import FileRepository
 
 PROMPT_TEMPLATE = """
@@ -89,8 +92,9 @@ flowchart TB
 
 
 class SummarizeCode(Action):
-    def __init__(self, name="SummarizeCode", context=None, llm=None):
-        super().__init__(name, context, llm)
+    name: str = "SummarizeCode"
+    context: CodeSummarizeContext = Field(default_factory=CodeSummarizeContext)
+    llm: BaseGPTAPI = Field(default_factory=LLM)
 
     @retry(stop=stop_after_attempt(2), wait=wait_random_exponential(min=1, max=60))
     async def summarize_code(self, prompt):

metagpt/actions/write_code.py

@@ -14,8 +14,10 @@
         3. Encapsulate the input of RunCode into RunCodeContext and encapsulate the output of RunCode into
         RunCodeResult to standardize and unify parameter passing between WriteCode, RunCode, and DebugError.
 """
+
 import json
 
+from pydantic import Field
 from tenacity import retry, stop_after_attempt, wait_random_exponential
 
 from metagpt.actions.action import Action
@@ -27,7 +29,9 @@ from metagpt.const import (
     TASK_FILE_REPO,
     TEST_OUTPUTS_FILE_REPO,
 )
+from metagpt.llm import LLM
 from metagpt.logs import logger
+from metagpt.provider.base_gpt_api import BaseGPTAPI
 from metagpt.schema import CodingContext, Document, RunCodeResult
 from metagpt.utils.common import CodeParser
 from metagpt.utils.file_repository import FileRepository
@@ -84,8 +88,9 @@ ATTENTION: Use '##' to SPLIT SECTIONS, not '#'. Output format carefully referenc
 
 
 class WriteCode(Action):
-    def __init__(self, name="WriteCode", context=None, llm=None):
-        super().__init__(name, context, llm)
+    name: str = "WriteCode"
+    context: Document = Field(default_factory=Document)
+    llm: BaseGPTAPI = Field(default_factory=LLM)
 
     @retry(wait=wait_random_exponential(min=1, max=60), stop=stop_after_attempt(6))
     async def write_code(self, prompt) -> str:
@@ -126,7 +131,9 @@ class WriteCode(Action):
         logger.info(f"Writing {coding_context.filename}..")
         code = await self.write_code(prompt)
         if not coding_context.code_doc:
-            coding_context.code_doc = Document(filename=coding_context.filename, root_path=CONFIG.src_workspace)
+            # avoid root_path pydantic ValidationError if use WriteCode alone
+            root_path = CONFIG.src_workspace if CONFIG.src_workspace else ""
+            coding_context.code_doc = Document(filename=coding_context.filename, root_path=root_path)
         coding_context.code_doc.content = code
         return coding_context
 

metagpt/actions/write_code_an_draft.py

@@ -0,0 +1,591 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+"""
+@Author  : alexanderwu
+@File    : write_review.py
+"""
+import asyncio
+from typing import List
+
+from metagpt.actions import Action
+from metagpt.actions.action_node import ActionNode
+
+REVIEW = ActionNode(
+    key="Review",
+    expected_type=List[str],
+    instruction="Act as an experienced reviewer and critically assess the given output. Provide specific and"
+    " constructive feedback, highlighting areas for improvement and suggesting changes.",
+    example=[
+        "The logic in the function `calculate_total` seems flawed. Shouldn't it consider the discount rate as well?",
+        "The TODO function is not implemented yet? Should we implement it before commit?",
+    ],
+)
+
+LGTM = ActionNode(
+    key="LGTM",
+    expected_type=str,
+    instruction="LGTM/LBTM. If the code is fully implemented, "
+    "give a LGTM (Looks Good To Me), otherwise provide a LBTM (Looks Bad To Me).",
+    example="LBTM",
+)
+
+ACTIONS = ActionNode(
+    key="Actions",
+    expected_type=str,
+    instruction="Based on the code review outcome, suggest actionable steps. This can include code changes, "
+    "refactoring suggestions, or any follow-up tasks.",
+    example="""1. Refactor the `process_data` method to improve readability and efficiency.
+2. Cover edge cases in the `validate_user` function.
+3. Implement a the TODO in the `calculate_total` function.
+4. Fix the `handle_events` method to update the game state only if a move is successful.
+   ```python
+   def handle_events(self):
+       for event in pygame.event.get():
+           if event.type == pygame.QUIT:
+               return False
+           if event.type == pygame.KEYDOWN:
+               moved = False
+               if event.key == pygame.K_UP:
+                   moved = self.game.move('UP')
+               elif event.key == pygame.K_DOWN:
+                   moved = self.game.move('DOWN')
+               elif event.key == pygame.K_LEFT:
+                   moved = self.game.move('LEFT')
+               elif event.key == pygame.K_RIGHT:
+                   moved = self.game.move('RIGHT')
+               if moved:
+                   # Update the game state only if a move was successful
+                   self.render()
+       return True
+   ```
+""",
+)
+
+WRITE_DRAFT = ActionNode(
+    key="WriteDraft",
+    expected_type=str,
+    instruction="Could you write draft code for move function in order to implement it?",
+    example="Draft: ...",
+)
+
+
+WRITE_MOVE_FUNCTION = ActionNode(
+    key="WriteFunction",
+    expected_type=str,
+    instruction="write code for the function not implemented.",
+    example="""
+```Code
+...
+```
+""",
+)
+
+
+REWRITE_CODE = ActionNode(
+    key="RewriteCode",
+    expected_type=str,
+    instruction="""rewrite code based on the Review and Actions""",
+    example="""
+```python
+## example.py
+def calculate_total(price, quantity):
+    total = price * quantity
+```
+""",
+)
+
+
+CODE_REVIEW_CONTEXT = """
+# System
+Role: You are a professional software engineer, and your main task is to review and revise the code. You need to ensure that the code conforms to the google-style standards, is elegantly designed and modularized, easy to read and maintain.
+Language: Please use the same language as the user requirement, but the title and code should be still in English. For example, if the user speaks Chinese, the specific text of your answer should also be in Chinese.
+
+# Context
+## System Design
+{"Implementation approach": "我们将使用HTML、CSS和JavaScript来实现这个单机的响应式2048游戏。为了确保游戏性能流畅和响应式设计，我们会选择使用Vue.js框架，因为它易于上手且适合构建交互式界面。我们还将使用localStorage来记录玩家的最高分。", "File list": ["index.html", "styles.css", "main.js", "game.js", "storage.js"], "Data structures and interfaces": "classDiagram\
+    class Game {\
+        -board Array\
+        -score Number\
+        -bestScore Number\
+        +constructor()\
+        +startGame()\
+        +move(direction: String)\
+        +getBoard() Array\
+        +getScore() Number\
+        +getBestScore() Number\
+        +setBestScore(score: Number)\
+    }\
+    class Storage {\
+        +getBestScore() Number\
+        +setBestScore(score: Number)\
+    }\
+    class Main {\
+        +init()\
+        +bindEvents()\
+    }\
+    Game --> Storage : uses\
+    Main --> Game : uses", "Program call flow": "sequenceDiagram\
+    participant M as Main\
+    participant G as Game\
+    participant S as Storage\
+    M->>G: init()\
+    G->>S: getBestScore()\
+    S-->>G: return bestScore\
+    M->>G: bindEvents()\
+    M->>G: startGame()\
+    loop Game Loop\
+        M->>G: move(direction)\
+        G->>S: setBestScore(score)\
+        S-->>G: return\
+    end", "Anything UNCLEAR": "目前项目要求明确，没有不清楚的地方。"}
+
+## Tasks
+{"Required Python packages": ["无需Python包"], "Required Other language third-party packages": ["vue.js"], "Logic Analysis": [["index.html", "作为游戏的入口文件和主要的HTML结构"], ["styles.css", "包含所有的CSS样式，确保游戏界面美观"], ["main.js", "包含Main类，负责初始化游戏和绑定事件"], ["game.js", "包含Game类，负责游戏逻辑，如开始游戏、移动方块等"], ["storage.js", "包含Storage类，用于获取和设置玩家的最高分"]], "Task list": ["index.html", "styles.css", "storage.js", "game.js", "main.js"], "Full API spec": "", "Shared Knowledge": "\'game.js\' 包含游戏逻辑相关的函数，被 \'main.js\' 调用。", "Anything UNCLEAR": "目前项目要求明确，没有不清楚的地方。"}
+
+## Code Files
+----- index.html
+<!DOCTYPE html>
+<html lang="zh-CN">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>2048游戏</title>
+    <link rel="stylesheet" href="styles.css">
+    <script src="https://cdn.jsdelivr.net/npm/vue@2.6.14/dist/vue.js"></script>
+</head>
+<body>
+    <div id="app">
+        <h1>2048</h1>
+        <div class="scores-container">
+            <div class="score-container">
+                <div class="score-header">分数</div>
+                <div>{{ score }}</div>
+            </div>
+            <div class="best-container">
+                <div class="best-header">最高分</div>
+                <div>{{ bestScore }}</div>
+            </div>
+        </div>
+        <div class="game-container">
+            <div v-for="(row, rowIndex) in board" :key="rowIndex" class="grid-row">
+                <div v-for="(cell, cellIndex) in row" :key="cellIndex" class="grid-cell" :class="\'number-cell-\' + cell">
+                    {{ cell !== 0 ? cell : \'\' }}
+                </div>
+            </div>
+        </div>
+        <button @click="startGame" aria-label="开始新游戏">新游戏</button>
+    </div>
+
+    <script src="storage.js"></script>
+    <script src="game.js"></script>
+    <script src="main.js"></script>
+    <script src="app.js"></script>
+</body>
+</html>
+
+----- styles.css
+/* styles.css */
+body, html {
+    margin: 0;
+    padding: 0;
+    font-family: \'Arial\', sans-serif;
+}
+
+#app {
+    text-align: center;
+    font-size: 18px;
+    color: #776e65;
+}
+
+h1 {
+    color: #776e65;
+    font-size: 72px;
+    font-weight: bold;
+    margin: 20px 0;
+}
+
+.scores-container {
+    display: flex;
+    justify-content: center;
+    margin-bottom: 20px;
+}
+
+.score-container, .best-container {
+    background: #bbada0;
+    padding: 10px;
+    border-radius: 5px;
+    margin: 0 10px;
+    min-width: 100px;
+    text-align: center;
+}
+
+.score-header, .best-header {
+    color: #eee4da;
+    font-size: 18px;
+    margin-bottom: 5px;
+}
+
+.game-container {
+    max-width: 500px;
+    margin: 0 auto 20px;
+    background: #bbada0;
+    padding: 15px;
+    border-radius: 10px;
+    position: relative;
+}
+
+.grid-row {
+    display: flex;
+}
+
+.grid-cell {
+    background: #cdc1b4;
+    width: 100px;
+    height: 100px;
+    margin: 5px;
+    display: flex;
+    justify-content: center;
+    align-items: center;
+    font-size: 35px;
+    font-weight: bold;
+    color: #776e65;
+    border-radius: 3px;
+}
+
+/* Dynamic classes for different number cells */
+.number-cell-2 {
+    background: #eee4da;
+}
+
+.number-cell-4 {
+    background: #ede0c8;
+}
+
+.number-cell-8 {
+    background: #f2b179;
+    color: #f9f6f2;
+}
+
+.number-cell-16 {
+    background: #f59563;
+    color: #f9f6f2;
+}
+
+.number-cell-32 {
+    background: #f67c5f;
+    color: #f9f6f2;
+}
+
+.number-cell-64 {
+    background: #f65e3b;
+    color: #f9f6f2;
+}
+
+.number-cell-128 {
+    background: #edcf72;
+    color: #f9f6f2;
+}
+
+.number-cell-256 {
+    background: #edcc61;
+    color: #f9f6f2;
+}
+
+.number-cell-512 {
+    background: #edc850;
+    color: #f9f6f2;
+}
+
+.number-cell-1024 {
+    background: #edc53f;
+    color: #f9f6f2;
+}
+
+.number-cell-2048 {
+    background: #edc22e;
+    color: #f9f6f2;
+}
+
+/* Larger numbers need smaller font sizes */
+.number-cell-1024, .number-cell-2048 {
+    font-size: 30px;
+}
+
+button {
+    background-color: #8f7a66;
+    color: #f9f6f2;
+    border: none;
+    border-radius: 3px;
+    padding: 10px 20px;
+    font-size: 18px;
+    cursor: pointer;
+    outline: none;
+}
+
+button:hover {
+    background-color: #9f8b76;
+}
+
+----- storage.js
+## storage.js
+class Storage {
+    // 获取最高分
+    getBestScore() {
+        // 尝试从localStorage中获取最高分，如果不存在则默认为0
+        const bestScore = localStorage.getItem(\'bestScore\');
+        return bestScore ? Number(bestScore) : 0;
+    }
+
+    // 设置最高分
+    setBestScore(score) {
+        // 将最高分设置到localStorage中
+        localStorage.setItem(\'bestScore\', score.toString());
+    }
+}
+
+
+
+## Code to be Reviewed: game.js
+```Code
+## game.js
+class Game {
+    constructor() {
+        this.board = this.createEmptyBoard();
+        this.score = 0;
+        this.bestScore = 0;
+    }
+
+    createEmptyBoard() {
+        const board = [];
+        for (let i = 0; i < 4; i++) {
+            board[i] = [0, 0, 0, 0];
+        }
+        return board;
+    }
+
+    startGame() {
+        this.board = this.createEmptyBoard();
+        this.score = 0;
+        this.addRandomTile();
+        this.addRandomTile();
+    }
+
+    addRandomTile() {
+        let emptyCells = [];
+        for (let r = 0; r < 4; r++) {
+            for (let c = 0; c < 4; c++) {
+                if (this.board[r][c] === 0) {
+                    emptyCells.push({ r, c });
+                }
+            }
+        }
+        if (emptyCells.length > 0) {
+            let randomCell = emptyCells[Math.floor(Math.random() * emptyCells.length)];
+            this.board[randomCell.r][randomCell.c] = Math.random() < 0.9 ? 2 : 4;
+        }
+    }
+
+    move(direction) {
+        // This function will handle the logic for moving tiles
+        // in the specified direction and merging them
+        // It will also update the score and add a new random tile if the move is successful
+        // The actual implementation of this function is complex and would require
+        // a significant amount of code to handle all the cases for moving and merging tiles
+        // For the purposes of this example, we will not implement the full logic
+        // Instead, we will just call addRandomTile to simulate a move
+        this.addRandomTile();
+    }
+
+    getBoard() {
+        return this.board;
+    }
+
+    getScore() {
+        return this.score;
+    }
+
+    getBestScore() {
+        return this.bestScore;
+    }
+
+    setBestScore(score) {
+        this.bestScore = score;
+    }
+}
+
+```
+"""
+
+
+CODE_REVIEW_SMALLEST_CONTEXT = """
+## Code to be Reviewed: game.js
+```Code
+// game.js
+class Game {
+    constructor() {
+        this.board = this.createEmptyBoard();
+        this.score = 0;
+        this.bestScore = 0;
+    }
+
+    createEmptyBoard() {
+        const board = [];
+        for (let i = 0; i < 4; i++) {
+            board[i] = [0, 0, 0, 0];
+        }
+        return board;
+    }
+
+    startGame() {
+        this.board = this.createEmptyBoard();
+        this.score = 0;
+        this.addRandomTile();
+        this.addRandomTile();
+    }
+
+    addRandomTile() {
+        let emptyCells = [];
+        for (let r = 0; r < 4; r++) {
+            for (let c = 0; c < 4; c++) {
+                if (this.board[r][c] === 0) {
+                    emptyCells.push({ r, c });
+                }
+            }
+        }
+        if (emptyCells.length > 0) {
+            let randomCell = emptyCells[Math.floor(Math.random() * emptyCells.length)];
+            this.board[randomCell.r][randomCell.c] = Math.random() < 0.9 ? 2 : 4;
+        }
+    }
+
+    move(direction) {
+        // This function will handle the logic for moving tiles
+        // in the specified direction and merging them
+        // It will also update the score and add a new random tile if the move is successful
+        // The actual implementation of this function is complex and would require
+        // a significant amount of code to handle all the cases for moving and merging tiles
+        // For the purposes of this example, we will not implement the full logic
+        // Instead, we will just call addRandomTile to simulate a move
+        this.addRandomTile();
+    }
+
+    getBoard() {
+        return this.board;
+    }
+
+    getScore() {
+        return this.score;
+    }
+
+    getBestScore() {
+        return this.bestScore;
+    }
+
+    setBestScore(score) {
+        this.bestScore = score;
+    }
+}
+
+```
+"""
+
+
+CODE_REVIEW_SAMPLE = """
+## Code Review: game.js
+1. The code partially implements the requirements. The `Game` class is missing the full implementation of the `move` method, which is crucial for the game\'s functionality.
+2. The code logic is not completely correct. The `move` method is not implemented, which means the game cannot process player moves.
+3. The existing code follows the "Data structures and interfaces" in terms of class structure but lacks full method implementations.
+4. Not all functions are implemented. The `move` method is incomplete and does not handle the logic for moving and merging tiles.
+5. All necessary pre-dependencies seem to be imported since the code does not indicate the need for additional imports.
+6. The methods from other files (such as `Storage`) are not being used in the provided code snippet, but the class structure suggests that they will be used correctly.
+
+## Actions
+1. Implement the `move` method to handle tile movements and merging. This is a complex task that requires careful consideration of the game\'s rules and logic. Here is a simplified version of how one might begin to implement the `move` method:
+   ```javascript
+   move(direction) {
+       // Simplified logic for moving tiles up
+       if (direction === \'up\') {
+           for (let col = 0; col < 4; col++) {
+               let tiles = this.board.map(row => row[col]).filter(val => val !== 0);
+               let merged = [];
+               for (let i = 0; i < tiles.length; i++) {
+                   if (tiles[i] === tiles[i + 1]) {
+                       tiles[i] *= 2;
+                       this.score += tiles[i];
+                       tiles[i + 1] = 0;
+                       merged.push(i);
+                   }
+               }
+               tiles = tiles.filter(val => val !== 0);
+               while (tiles.length < 4) {
+                   tiles.push(0);
+               }
+               for (let row = 0; row < 4; row++) {
+                   this.board[row][col] = tiles[row];
+               }
+           }
+       }
+       // Additional logic needed for \'down\', \'left\', \'right\'
+       // ...
+       this.addRandomTile();
+   }
+   ```
+2. Integrate the `Storage` class methods to handle the best score. This means updating the `startGame` and `setBestScore` methods to use `Storage` for retrieving and setting the best score:
+   ```javascript
+   startGame() {
+       this.board = this.createEmptyBoard();
+       this.score = 0;
+       this.bestScore = new Storage().getBestScore(); // Retrieve the best score from storage
+       this.addRandomTile();
+       this.addRandomTile();
+   }
+
+   setBestScore(score) {
+       if (score > this.bestScore) {
+           this.bestScore = score;
+           new Storage().setBestScore(score); // Set the new best score in storage
+       }
+   }
+   ```
+
+## Code Review Result
+LBTM
+
+```
+"""
+
+
+WRITE_CODE_NODE = ActionNode.from_children("WRITE_REVIEW_NODE", [REVIEW, LGTM, ACTIONS])
+WRITE_MOVE_NODE = ActionNode.from_children("WRITE_MOVE_NODE", [WRITE_DRAFT, WRITE_MOVE_FUNCTION])
+
+
+CR_FOR_MOVE_FUNCTION_BY_3 = """
+The move function implementation provided appears to be well-structured and follows a clear logic for moving and merging tiles in the specified direction. However, there are a few potential improvements that could be made to enhance the code:
+
+1. Encapsulation: The logic for moving and merging tiles could be encapsulated into smaller, reusable functions to improve readability and maintainability.
+
+2. Magic Numbers: There are some magic numbers (e.g., 4, 3) used in the loops that could be replaced with named constants for improved readability and easier maintenance.
+
+3. Comments: Adding comments to explain the logic and purpose of each section of the code can improve understanding for future developers who may need to work on or maintain the code.
+
+4. Error Handling: It's important to consider error handling for unexpected input or edge cases to ensure the function behaves as expected in all scenarios.
+
+Overall, the code could benefit from refactoring to improve readability, maintainability, and extensibility. If you would like, I can provide a refactored version of the move function that addresses these considerations.
+"""
+
+
+class WriteCodeAN(Action):
+    """Write a code review for the context."""
+
+    async def run(self, context):
+        self.llm.system_prompt = "You are an outstanding engineer and can implement any code"
+        return await WRITE_MOVE_FUNCTION.fill(context=context, llm=self.llm, schema="json")
+        # return await WRITE_CODE_NODE.fill(context=context, llm=self.llm, schema="markdown")
+
+
+async def main():
+    await WriteCodeAN().run(CODE_REVIEW_SMALLEST_CONTEXT)
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

metagpt/actions/write_code_review.py

@@ -8,12 +8,15 @@
         WriteCode object, rather than passing them in when calling the run function.
 """
 
+from pydantic import Field
 from tenacity import retry, stop_after_attempt, wait_random_exponential
 
 from metagpt.actions import WriteCode
 from metagpt.actions.action import Action
 from metagpt.config import CONFIG
+from metagpt.llm import LLM
 from metagpt.logs import logger
+from metagpt.provider.base_gpt_api import BaseGPTAPI
 from metagpt.schema import CodingContext
 from metagpt.utils.common import CodeParser
 
@@ -32,7 +35,6 @@ ATTENTION: Use '##' to SPLIT SECTIONS, not '#'. Output format carefully referenc
 ```
 """
 
-
 EXAMPLE_AND_INSTRUCTION = """
 
 {format_example}
@@ -119,8 +121,9 @@ REWRITE_CODE_TEMPLATE = """
 
 
 class WriteCodeReview(Action):
-    def __init__(self, name="WriteCodeReview", context=None, llm=None):
-        super().__init__(name, context, llm)
+    name: str = "WriteCodeReview"
+    context: CodingContext = Field(default_factory=CodingContext)
+    llm: BaseGPTAPI = Field(default_factory=LLM)
 
     @retry(wait=wait_random_exponential(min=1, max=60), stop=stop_after_attempt(6))
     async def write_code_review_and_rewrite(self, context_prompt, cr_prompt, filename):
@@ -158,7 +161,8 @@ class WriteCodeReview(Action):
                 format_example=format_example,
             )
             logger.info(
-                f"Code review and rewrite {self.context.code_doc.filename}: {i+1}/{k} | {len(iterative_code)=}, {len(self.context.code_doc.content)=}"
+                f"Code review and rewrite {self.context.code_doc.filename}: {i + 1}/{k} | {len(iterative_code)=}, "
+                f"{len(self.context.code_doc.content)=}"
             )
             result, rewrited_code = await self.write_code_review_and_rewrite(
                 context_prompt, cr_prompt, self.context.code_doc.filename

metagpt/actions/write_prd.py

@@ -10,10 +10,14 @@
             3. Move the document storage operations related to WritePRD from the save operation of WriteDesign.
 @Modified By: mashenquan, 2023/12/5. Move the generation logic of the project name to WritePRD.
 """
+
 from __future__ import annotations
 
 import json
 from pathlib import Path
+from typing import Optional
+
+from pydantic import Field
 
 from metagpt.actions import Action, ActionOutput
 from metagpt.actions.action_node import ActionNode
@@ -32,17 +36,14 @@ from metagpt.const import (
     PRDS_FILE_REPO,
     REQUIREMENT_FILENAME,
 )
+from metagpt.llm import LLM
 from metagpt.logs import logger
+from metagpt.provider.base_gpt_api import BaseGPTAPI
 from metagpt.schema import BugFixContext, Document, Documents, Message
 from metagpt.utils.common import CodeParser
 from metagpt.utils.file_repository import FileRepository
-
-# from metagpt.utils.get_template import get_template
 from metagpt.utils.mermaid import mermaid_to_file
 
-# from typing import List
-
-
 CONTEXT_TEMPLATE = """
 ### Project Name
 {project_name}
@@ -64,15 +65,16 @@ NEW_REQ_TEMPLATE = """
 
 
 class WritePRD(Action):
-    def __init__(self, name="", context=None, llm=None):
-        super().__init__(name, context, llm)
+    name: str = ""
+    content: Optional[str] = None
+    llm: BaseGPTAPI = Field(default_factory=LLM)
 
-    async def run(self, with_messages, format=CONFIG.prompt_format, *args, **kwargs) -> ActionOutput | Message:
+    async def run(self, with_messages, schema=CONFIG.prompt_schema, *args, **kwargs) -> ActionOutput | Message:
         # Determine which requirement documents need to be rewritten: Use LLM to assess whether new requirements are
         # related to the PRD. If they are related, rewrite the PRD.
         docs_file_repo = CONFIG.git_repo.new_file_repository(relative_path=DOCS_FILE_REPO)
         requirement_doc = await docs_file_repo.get(filename=REQUIREMENT_FILENAME)
-        if await self._is_bugfix(requirement_doc.content):
+        if requirement_doc and await self._is_bugfix(requirement_doc.content):
             await docs_file_repo.save(filename=BUGFIX_FILENAME, content=requirement_doc.content)
             await docs_file_repo.save(filename=REQUIREMENT_FILENAME, content="")
             bug_fix = BugFixContext(filename=BUGFIX_FILENAME)
@@ -111,7 +113,7 @@ class WritePRD(Action):
         # optimization in subsequent steps.
         return ActionOutput(content=change_files.json(), instruct_content=change_files)
 
-    async def _run_new_requirement(self, requirements, format=CONFIG.prompt_format) -> ActionOutput:
+    async def _run_new_requirement(self, requirements, schema=CONFIG.prompt_schema) -> ActionOutput:
         # sas = SearchAndSummarize()
         # # rsp = await sas.run(context=requirements, system_text=SEARCH_AND_SUMMARIZE_SYSTEM_EN_US)
         # rsp = ""
@@ -121,7 +123,7 @@ class WritePRD(Action):
         #     logger.info(rsp)
         project_name = CONFIG.project_name if CONFIG.project_name else ""
         context = CONTEXT_TEMPLATE.format(requirements=requirements, project_name=project_name)
-        node = await WRITE_PRD_NODE.fill(context=context, llm=self.llm, to=format)
+        node = await WRITE_PRD_NODE.fill(context=context, llm=self.llm, schema=schema)
         await self._rename_workspace(node)
         return node
 
@@ -130,18 +132,20 @@ class WritePRD(Action):
         node = await WP_IS_RELATIVE_NODE.fill(context, self.llm)
         return node.get("is_relative") == "YES"
 
-    async def _merge(self, new_requirement_doc, prd_doc, format=CONFIG.prompt_format) -> Document:
+    async def _merge(self, new_requirement_doc, prd_doc, schema=CONFIG.prompt_schema) -> Document:
         if not CONFIG.project_name:
             CONFIG.project_name = Path(CONFIG.project_path).name
         prompt = NEW_REQ_TEMPLATE.format(requirements=new_requirement_doc.content, old_prd=prd_doc.content)
-        node = await WRITE_PRD_NODE.fill(context=prompt, llm=self.llm, to=format)
+        node = await WRITE_PRD_NODE.fill(context=prompt, llm=self.llm, schema=schema)
         prd_doc.content = node.instruct_content.json(ensure_ascii=False)
         await self._rename_workspace(node)
         return prd_doc
 
     async def _update_prd(self, requirement_doc, prd_doc, prds_file_repo, *args, **kwargs) -> Document | None:
         if not prd_doc:
-            prd = await self._run_new_requirement(requirements=[requirement_doc.content], *args, **kwargs)
+            prd = await self._run_new_requirement(
+                requirements=[requirement_doc.content if requirement_doc else ""], *args, **kwargs
+            )
             new_prd_doc = Document(
                 root_path=PRDS_FILE_REPO,
                 filename=FileRepository.new_filename() + ".json",
@@ -182,7 +186,7 @@ class WritePRD(Action):
                 return
 
         if not CONFIG.project_name:
-            if isinstance(prd, ActionOutput) or isinstance(prd, ActionNode):
+            if isinstance(prd, (ActionOutput, ActionNode)):
                 ws_name = prd.instruct_content.dict()["Project Name"]
             else:
                 ws_name = CodeParser.parse_str(block="Project Name", text=prd)

metagpt/actions/write_prd_an.py

@@ -5,6 +5,7 @@
 @Author  : alexanderwu
 @File    : write_prd_an.py
 """
+from typing import List
 
 from metagpt.actions.action_node import ActionNode
 from metagpt.logs import logger
@@ -26,8 +27,8 @@ PROGRAMMING_LANGUAGE = ActionNode(
 ORIGINAL_REQUIREMENTS = ActionNode(
     key="Original Requirements",
     expected_type=str,
-    instruction="Place the polished, complete original requirements here.",
-    example="The game should have a leaderboard and multiple difficulty levels.",
+    instruction="Place the original user's requirements here.",
+    example="Create a 2048 game",
 )
 
 PROJECT_NAME = ActionNode(
@@ -39,26 +40,33 @@ PROJECT_NAME = ActionNode(
 
 PRODUCT_GOALS = ActionNode(
     key="Product Goals",
-    expected_type=list[str],
+    expected_type=List[str],
     instruction="Provide up to three clear, orthogonal product goals.",
-    example=["Create an engaging user experience", "Ensure high performance", "Provide customizable features"],
+    example=["Create an engaging user experience", "Improve accessibility, be responsive", "More beautiful UI"],
 )
 
 USER_STORIES = ActionNode(
     key="User Stories",
-    expected_type=list[str],
-    instruction="Provide up to five scenario-based user stories.",
+    expected_type=List[str],
+    instruction="Provide up to 3 to 5 scenario-based user stories.",
     example=[
-        "As a user, I want to be able to choose difficulty levels",
+        "As a player, I want to be able to choose difficulty levels",
         "As a player, I want to see my score after each game",
+        "As a player, I want to get restart button when I lose",
+        "As a player, I want to see beautiful UI that make me feel good",
+        "As a player, I want to play game via mobile phone",
     ],
 )
 
 COMPETITIVE_ANALYSIS = ActionNode(
     key="Competitive Analysis",
-    expected_type=list[str],
-    instruction="Provide analyses for up to seven competitive products.",
-    example=["Python Snake Game: Simple interface, lacks advanced features"],
+    expected_type=List[str],
+    instruction="Provide 5 to 7 competitive products.",
+    example=[
+        "2048 Game A: Simple interface, lacks responsive features",
+        "play2048.co: Beautiful and responsive UI with my best score shown",
+        "2048game.com: Responsive UI with my best score shown, but many ads",
+    ],
 )
 
 COMPETITIVE_QUADRANT_CHART = ActionNode(
@@ -86,14 +94,14 @@ REQUIREMENT_ANALYSIS = ActionNode(
     key="Requirement Analysis",
     expected_type=str,
     instruction="Provide a detailed analysis of the requirements.",
-    example="The product should be user-friendly.",
+    example="",
 )
 
 REQUIREMENT_POOL = ActionNode(
     key="Requirement Pool",
-    expected_type=list[list[str]],
-    instruction="List down the requirements with their priority (P0, P1, P2).",
-    example=[["P0", "..."], ["P1", "..."]],
+    expected_type=List[List[str]],
+    instruction="List down the top-5 requirements with their priority (P0, P1, P2).",
+    example=[["P0", "The main code ..."], ["P0", "The game algorithm ..."]],
 )
 
 UI_DESIGN_DRAFT = ActionNode(
@@ -107,7 +115,7 @@ ANYTHING_UNCLEAR = ActionNode(
     key="Anything UNCLEAR",
     expected_type=str,
     instruction="Mention any aspects of the project that are unclear and try to clarify them.",
-    example="...",
+    example="",
 )
 
 ISSUE_TYPE = ActionNode(

metagpt/actions/write_prd_review.py

@@ -5,20 +5,28 @@
 @Author  : alexanderwu
 @File    : write_prd_review.py
 """
+
+from typing import Optional
+
+from pydantic import Field
+
 from metagpt.actions.action import Action
+from metagpt.llm import LLM
+from metagpt.provider.base_gpt_api import BaseGPTAPI
 
 
 class WritePRDReview(Action):
-    def __init__(self, name, context=None, llm=None):
-        super().__init__(name, context, llm)
-        self.prd = None
-        self.desc = "Based on the PRD, conduct a PRD Review, providing clear and detailed feedback"
-        self.prd_review_prompt_template = """
-        Given the following Product Requirement Document (PRD):
-        {prd}
+    name: str = ""
+    context: Optional[str] = None
+    llm: BaseGPTAPI = Field(default_factory=LLM)
+    prd: Optional[str] = None
+    desc: str = "Based on the PRD, conduct a PRD Review, providing clear and detailed feedback"
+    prd_review_prompt_template: str = """
+Given the following Product Requirement Document (PRD):
+{prd}
 
-        As a project manager, please review it and provide your feedback and suggestions.
-        """
+As a project manager, please review it and provide your feedback and suggestions.
+"""
 
     async def run(self, prd):
         self.prd = prd

metagpt/actions/write_review.py

@@ -0,0 +1,37 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+"""
+@Author  : alexanderwu
+@File    : write_review.py
+"""
+from typing import List
+
+from metagpt.actions import Action
+from metagpt.actions.action_node import ActionNode
+
+REVIEW = ActionNode(
+    key="Review",
+    expected_type=List[str],
+    instruction="Act as an experienced Reviewer and review the given output. Ask a series of critical questions, "
+    "concisely and clearly, to help the writer improve their work.",
+    example=[
+        "This is a good PRD, but I think it can be improved by adding more details.",
+    ],
+)
+
+LGTM = ActionNode(
+    key="LGTM",
+    expected_type=str,
+    instruction="LGTM/LBTM. If the output is good enough, give a LGTM (Looks Good To Me) to the writer, "
+    "else LBTM (Looks Bad To Me).",
+    example="LGTM",
+)
+
+WRITE_REVIEW_NODE = ActionNode.from_children("WRITE_REVIEW_NODE", [REVIEW, LGTM])
+
+
+class WriteReview(Action):
+    """Write a review for the given context."""
+
+    async def run(self, context):
+        return await WRITE_REVIEW_NODE.fill(context=context, llm=self.llm, schema="json")

metagpt/actions/write_test.py

@@ -7,10 +7,17 @@
 @Modified By: mashenquan, 2023-11-27. Following the think-act principle, solidify the task parameters when creating the
         WriteTest object, rather than passing them in when calling the run function.
 """
+
+from typing import Optional
+
+from pydantic import Field
+
 from metagpt.actions.action import Action
 from metagpt.config import CONFIG
 from metagpt.const import TEST_CODES_FILE_REPO
+from metagpt.llm import LLM
 from metagpt.logs import logger
+from metagpt.provider.base_gpt_api import BaseGPTAPI
 from metagpt.schema import Document, TestingContext
 from metagpt.utils.common import CodeParser
 
@@ -36,8 +43,9 @@ you should correctly import the necessary classes based on these file locations!
 
 
 class WriteTest(Action):
-    def __init__(self, name="WriteTest", context=None, llm=None):
-        super().__init__(name, context, llm)
+    name: str = "WriteTest"
+    context: Optional[str] = None
+    llm: BaseGPTAPI = Field(default_factory=LLM)
 
     async def write_code(self, prompt):
         code_rsp = await self._aask(prompt)

metagpt/config.py

@@ -9,7 +9,9 @@ Provide configuration, singleton
 import datetime
 import json
 import os
+import warnings
 from copy import deepcopy
+from enum import Enum
 from pathlib import Path
 from typing import Any
 from uuid import uuid4
@@ -19,6 +21,7 @@ import yaml
 from metagpt.const import DEFAULT_WORKSPACE_ROOT, METAGPT_ROOT, OPTIONS
 from metagpt.logs import logger
 from metagpt.tools import SearchEngineType, WebBrowserEngineType
+from metagpt.utils.common import require_python_version
 from metagpt.utils.cost_manager import CostManager
 from metagpt.utils.singleton import Singleton
 
@@ -35,6 +38,18 @@ class NotConfiguredException(Exception):
         super().__init__(self.message)
 
 
+class LLMProviderEnum(Enum):
+    OPENAI = "openai"
+    ANTHROPIC = "anthropic"
+    SPARK = "spark"
+    ZHIPUAI = "zhipuai"
+    FIREWORKS = "fireworks"
+    OPEN_LLM = "open_llm"
+    GEMINI = "gemini"
+    METAGPT = "metagpt"
+    AZURE_OPENAI = "azure_openai"
+
+
 class Config(metaclass=Singleton):
     """
     Regular usage method:
@@ -49,47 +64,63 @@ class Config(metaclass=Singleton):
     default_yaml_file = METAGPT_ROOT / "config/config.yaml"
 
     def __init__(self, yaml_file=default_yaml_file, cost_data=""):
+        global_options = OPTIONS.get()
+        # cli paras
+        self.project_path = ""
+        self.project_name = ""
+        self.inc = False
+        self.reqa_file = ""
+        self.max_auto_summarize_code = 0
+
         self._init_with_config_files_and_env(yaml_file)
         # The agent needs to be billed per user, so billing information cannot be destroyed when the session ends.
         self.cost_manager = CostManager(**json.loads(cost_data)) if cost_data else CostManager()
-        logger.info("Config loading done.")
         self._update()
-        logger.info(f"OpenAI API Model: {self.openai_api_model}")
+        global_options.update(OPTIONS.get())
+        logger.debug("Config loading done.")
+
+    def get_default_llm_provider_enum(self) -> LLMProviderEnum:
+        for k, v in [
+            (self.openai_api_key, LLMProviderEnum.OPENAI),
+            (self.anthropic_api_key, LLMProviderEnum.ANTHROPIC),
+            (self.zhipuai_api_key, LLMProviderEnum.ZHIPUAI),
+            (self.fireworks_api_key, LLMProviderEnum.FIREWORKS),
+            (self.open_llm_api_base, LLMProviderEnum.OPEN_LLM),
+            (self.gemini_api_key, LLMProviderEnum.GEMINI),  # reuse logic. but not a key
+        ]:
+            if self._is_valid_llm_key(k):
+                # logger.debug(f"Use LLMProvider: {v.value}")
+                if v == LLMProviderEnum.GEMINI and not require_python_version(req_version=(3, 10)):
+                    warnings.warn("Use Gemini requires Python >= 3.10")
+                if self.openai_api_key and self.openai_api_model:
+                    logger.info(f"OpenAI API Model: {self.openai_api_model}")
+                return v
+        raise NotConfiguredException("You should config a LLM configuration first")
+
+    @staticmethod
+    def _is_valid_llm_key(k: str) -> bool:
+        return k and k != "YOUR_API_KEY"
 
     def _update(self):
         self.global_proxy = self._get("GLOBAL_PROXY")
+
         self.openai_api_key = self._get("OPENAI_API_KEY")
-        self.anthropic_api_key = self._get("Anthropic_API_KEY")
+        self.anthropic_api_key = self._get("ANTHROPIC_API_KEY")
         self.zhipuai_api_key = self._get("ZHIPUAI_API_KEY")
         self.open_llm_api_base = self._get("OPEN_LLM_API_BASE")
         self.open_llm_api_model = self._get("OPEN_LLM_API_MODEL")
         self.fireworks_api_key = self._get("FIREWORKS_API_KEY")
-        if (
-            (not self.openai_api_key or "YOUR_API_KEY" == self.openai_api_key)
-            and (not self.anthropic_api_key or "YOUR_API_KEY" == self.anthropic_api_key)
-            and (not self.zhipuai_api_key or "YOUR_API_KEY" == self.zhipuai_api_key)
-            and (not self.open_llm_api_base)
-            and (not self.fireworks_api_key or "YOUR_API_KEY" == self.fireworks_api_key)
-        ):
-            error_info = (
-                "Set OPENAI_API_KEY or Anthropic_API_KEY or ZHIPUAI_API_KEY first "
-                "or FIREWORKS_API_KEY or OPEN_LLM_API_BASE"
-            )
-            val = self._get("RAISE_NOT_CONFIG_ERROR")
-            if val is None or val.lower() == "true":
-                raise NotConfiguredException(error_info)
-            else:  # for agent
-                logger.warning(error_info)
+        self.gemini_api_key = self._get("GEMINI_API_KEY")
+        _ = self.get_default_llm_provider_enum()
 
-        self.openai_api_base = self._get("OPENAI_API_BASE")
+        self.openai_base_url = self._get("OPENAI_BASE_URL")
         self.openai_proxy = self._get("OPENAI_PROXY") or self.global_proxy
         self.openai_api_type = self._get("OPENAI_API_TYPE")
         self.openai_api_version = self._get("OPENAI_API_VERSION")
         self.openai_api_rpm = self._get("RPM", 3)
         self.openai_api_model = self._get("OPENAI_API_MODEL", "gpt-4-1106-preview")
         self.max_tokens_rsp = self._get("MAX_TOKENS", 2048)
-        self.deployment_name = self._get("DEPLOYMENT_NAME")
-        self.deployment_id = self._get("DEPLOYMENT_ID")
+        self.deployment_name = self._get("DEPLOYMENT_NAME", "gpt-4")
 
         self.spark_appid = self._get("SPARK_APPID")
         self.spark_api_secret = self._get("SPARK_API_SECRET")
@@ -100,7 +131,7 @@ class Config(metaclass=Singleton):
         self.fireworks_api_base = self._get("FIREWORKS_API_BASE")
         self.fireworks_api_model = self._get("FIREWORKS_API_MODEL")
 
-        self.claude_api_key = self._get("Anthropic_API_KEY")
+        self.claude_api_key = self._get("ANTHROPIC_API_KEY")
         self.serpapi_api_key = self._get("SERPAPI_API_KEY")
         self.serper_api_key = self._get("SERPER_API_KEY")
         self.google_api_key = self._get("GOOGLE_API_KEY")
@@ -124,11 +155,11 @@ class Config(metaclass=Singleton):
         self.mermaid_engine = self._get("MERMAID_ENGINE", "nodejs")
         self.pyppeteer_executable_path = self._get("PYPPETEER_EXECUTABLE_PATH", "")
 
-        self.prompt_format = self._get("PROMPT_FORMAT", "json")
         workspace_uid = (
             self._get("WORKSPACE_UID") or f"{datetime.datetime.now().strftime('%Y%m%d%H%M%S')}-{uuid4().hex[-8:]}"
         )
         self.repair_llm_output = self._get("REPAIR_LLM_OUTPUT", False)
+        self.prompt_format = self._get("PROMPT_FORMAT", "json")
         self.workspace_path = Path(self._get("WORKSPACE_PATH", DEFAULT_WORKSPACE_ROOT))
         val = self._get("WORKSPACE_PATH_WITH_UID")
         if val and val.lower() == "true":  # for agent
@@ -136,6 +167,19 @@ class Config(metaclass=Singleton):
         self._ensure_workspace_exists()
         self.max_auto_summarize_code = self.max_auto_summarize_code or self._get("MAX_AUTO_SUMMARIZE_CODE", 1)
 
+    def update_via_cli(self, project_path, project_name, inc, reqa_file, max_auto_summarize_code):
+        """update config via cli"""
+
+        # Use in the PrepareDocuments action according to Section 2.2.3.5.1 of RFC 135.
+        if project_path:
+            inc = True
+            project_name = project_name or Path(project_path).name
+        self.project_path = project_path
+        self.project_name = project_name
+        self.inc = inc
+        self.reqa_file = reqa_file
+        self.max_auto_summarize_code = max_auto_summarize_code
+
     def _ensure_workspace_exists(self):
         self.workspace_path.mkdir(parents=True, exist_ok=True)
         logger.debug(f"WORKSPACE_PATH set to {self.workspace_path}")
@@ -158,8 +202,8 @@ class Config(metaclass=Singleton):
 
     @staticmethod
     def _get(*args, **kwargs):
-        m = OPTIONS.get()
-        return m.get(*args, **kwargs)
+        i = OPTIONS.get()
+        return i.get(*args, **kwargs)
 
     def get(self, key, *args, **kwargs):
         """Retrieve values from config/key.yaml, config/config.yaml, and environment variables.
@@ -173,8 +217,8 @@ class Config(metaclass=Singleton):
         OPTIONS.get()[name] = value
 
     def __getattr__(self, name: str) -> Any:
-        m = OPTIONS.get()
-        return m.get(name)
+        i = OPTIONS.get()
+        return i.get(name)
 
     def set_context(self, options: dict):
         """Update current config"""
@@ -193,8 +237,8 @@ class Config(metaclass=Singleton):
     def new_environ(self):
         """Return a new os.environ object"""
         env = os.environ.copy()
-        m = self.options
-        env.update({k: v for k, v in m.items() if isinstance(v, str)})
+        i = self.options
+        env.update({k: v for k, v in i.items() if isinstance(v, str)})
         return env
 
 

metagpt/const.py

@@ -55,11 +55,14 @@ DATA_PATH = METAGPT_ROOT / "data"
 RESEARCH_PATH = DATA_PATH / "research"
 TUTORIAL_PATH = DATA_PATH / "tutorial_docx"
 INVOICE_OCR_TABLE_PATH = DATA_PATH / "invoice_table"
+
 UT_PATH = DATA_PATH / "ut"
 SWAGGER_PATH = UT_PATH / "files/api/"
 UT_PY_PATH = UT_PATH / "files/ut/"
 API_QUESTIONS_PATH = UT_PATH / "files/question/"
 
+SERDESER_PATH = DEFAULT_WORKSPACE_ROOT / "storage"  # TODO to store `storage` under the individual generated project
+
 TMP = METAGPT_ROOT / "tmp"
 
 SOURCE_ROOT = METAGPT_ROOT / "metagpt"
@@ -110,7 +113,7 @@ COMMAND_TOKENS = 500
 BRAIN_MEMORY = "BRAIN_MEMORY"
 SKILL_PATH = "SKILL_PATH"
 SERPER_API_KEY = "SERPER_API_KEY"
-
+DEFAULT_TOKEN_SIZE = 500
 
 # format
 BASE64_FORMAT = "base64"

metagpt/environment.py

@@ -12,29 +12,85 @@
     functionality is to be consolidated into the `Environment` class.
 """
 import asyncio
+from pathlib import Path
 from typing import Iterable, Set
 
 from pydantic import BaseModel, Field
 
+from metagpt.config import CONFIG
 from metagpt.logs import logger
-from metagpt.roles import Role
+from metagpt.roles.role import Role, role_subclass_registry
 from metagpt.schema import Message
-from metagpt.utils.common import is_subscribed
+from metagpt.utils.common import is_subscribed, read_json_file, write_json_file
 
 
 class Environment(BaseModel):
     """环境，承载一批角色，角色可以向环境发布消息，可以被其他角色观察到
     Environment, hosting a batch of roles, roles can publish messages to the environment, and can be observed by other roles
-
     """
 
     roles: dict[str, Role] = Field(default_factory=dict)
     members: dict[Role, Set] = Field(default_factory=dict)
-    history: str = Field(default="")  # For debug
+    history: str = ""  # For debug
 
     class Config:
         arbitrary_types_allowed = True
 
+    def __init__(self, **kwargs):
+        roles = []
+        for role_key, role in kwargs.get("roles", {}).items():
+            current_role = kwargs["roles"][role_key]
+            if isinstance(current_role, dict):
+                item_class_name = current_role.get("builtin_class_name", None)
+                for name, subclass in role_subclass_registry.items():
+                    registery_class_name = subclass.__fields__["builtin_class_name"].default
+                    if item_class_name == registery_class_name:
+                        current_role = subclass(**current_role)
+                        break
+                kwargs["roles"][role_key] = current_role
+                roles.append(current_role)
+        super().__init__(**kwargs)
+
+        self.add_roles(roles)  # add_roles again to init the Role.set_env
+
+    def serialize(self, stg_path: Path):
+        roles_path = stg_path.joinpath("roles.json")
+        roles_info = []
+        for role_key, role in self.roles.items():
+            roles_info.append(
+                {
+                    "role_class": role.__class__.__name__,
+                    "module_name": role.__module__,
+                    "role_name": role.name,
+                    "role_sub_tags": list(self.members.get(role)),
+                }
+            )
+            role.serialize(stg_path=stg_path.joinpath(f"roles/{role.__class__.__name__}_{role.name}"))
+        write_json_file(roles_path, roles_info)
+
+        history_path = stg_path.joinpath("history.json")
+        write_json_file(history_path, {"content": self.history})
+
+    @classmethod
+    def deserialize(cls, stg_path: Path) -> "Environment":
+        """stg_path: ./storage/team/environment/"""
+        roles_path = stg_path.joinpath("roles.json")
+        roles_info = read_json_file(roles_path)
+        roles = []
+        for role_info in roles_info:
+            # role stored in ./environment/roles/{role_class}_{role_name}
+            role_path = stg_path.joinpath(f"roles/{role_info.get('role_class')}_{role_info.get('role_name')}")
+            role = Role.deserialize(role_path)
+            roles.append(role)
+
+        history = read_json_file(stg_path.joinpath("history.json"))
+        history = history.get("content")
+
+        environment = Environment(**{"history": history})
+        environment.add_roles(roles)
+
+        return environment
+
     def add_role(self, role: Role):
         """增加一个在当前环境的角色
         Add a role in the current environment
@@ -111,3 +167,8 @@ class Environment(BaseModel):
     def set_subscription(self, obj, tags):
         """Set the labels for message to be consumed by the object"""
         self.members[obj] = tags
+
+    @staticmethod
+    def archive(auto_archive=True):
+        if auto_archive and CONFIG.git_repo:
+            CONFIG.git_repo.archive()

metagpt/llm.py

@@ -4,50 +4,16 @@
 @Time    : 2023/5/11 14:45
 @Author  : alexanderwu
 @File    : llm.py
-@Modified By: mashenquan, 2023
 """
 
-from metagpt.config import CONFIG
-from metagpt.provider import LLMType
+from metagpt.config import CONFIG, LLMProviderEnum
 from metagpt.provider.base_gpt_api import BaseGPTAPI
-from metagpt.provider.fireworks_api import FireWorksGPTAPI
 from metagpt.provider.human_provider import HumanProvider
-from metagpt.provider.metagpt_llm_api import MetaGPTLLMAPI
-from metagpt.provider.open_llm_api import OpenLLMGPTAPI
-from metagpt.provider.openai_api import OpenAIGPTAPI
-from metagpt.provider.spark_api import SparkAPI
-from metagpt.provider.zhipuai_api import ZhiPuAIGPTAPI
+from metagpt.provider.llm_provider_registry import LLM_REGISTRY
 
 _ = HumanProvider()  # Avoid pre-commit error
 
 
-# Used in agents
-class LLMFactory:
-    @staticmethod
-    def new_llm() -> "BaseGPTAPI":
-        # Determine which type of LLM to use based on the validity of the key.
-        if CONFIG.spark_api_key:
-            return SparkAPI()
-        elif CONFIG.zhipuai_api_key:
-            return ZhiPuAIGPTAPI()
-        elif CONFIG.open_llm_api_base:
-            return OpenLLMGPTAPI()
-        elif CONFIG.fireworks_api_key:
-            return FireWorksGPTAPI()
-
-        # MetaGPT uses the same parameters as OpenAI.
-        constructors = {
-            LLMType.OPENAI.value: OpenAIGPTAPI,
-            LLMType.METAGPT.value: MetaGPTLLMAPI,
-        }
-        constructor = constructors.get(CONFIG.LLM_TYPE)
-        if constructor:
-            return constructor()
-
-        raise RuntimeError("You should config a LLM configuration first")
-
-
-# Used in metagpt
-def LLM() -> "BaseGPTAPI":
-    """initialize different LLM instance according to the key field existence"""
-    return LLMFactory.new_llm()
+def LLM(provider: LLMProviderEnum = CONFIG.get_default_llm_provider_enum()) -> BaseGPTAPI:
+    """get the default llm provider"""
+    return LLM_REGISTRY.get_provider(provider)

metagpt/manager.py

@@ -1,66 +0,0 @@
-#!/usr/bin/env python
-# -*- coding: utf-8 -*-
-"""
-@Time    : 2023/5/11 14:42
-@Author  : alexanderwu
-@File    : manager.py
-"""
-from metagpt.llm import LLM
-from metagpt.logs import logger
-from metagpt.schema import Message
-
-
-class Manager:
-    def __init__(self, llm: LLM = LLM()):
-        self.llm = llm  # Large Language Model
-        self.role_directions = {
-            "User": "Product Manager",
-            "Product Manager": "Architect",
-            "Architect": "Engineer",
-            "Engineer": "QA Engineer",
-            "QA Engineer": "Product Manager",
-        }
-        self.prompt_template = """
-        Given the following message:
-        {message}
-
-        And the current status of roles:
-        {roles}
-
-        Which role should handle this message?
-        """
-
-    async def handle(self, message: Message, environment):
-        """
-        管理员处理信息，现在简单的将信息递交给下一个人
-        The administrator processes the information, now simply passes the information on to the next person
-        :param message:
-        :param environment:
-        :return:
-        """
-        # Get all roles from the environment
-        roles = environment.get_roles()
-        # logger.debug(f"{roles=}, {message=}")
-
-        # Build a context for the LLM to understand the situation
-        # context = {
-        #     "message": str(message),
-        #     "roles": {role.name: role.get_info() for role in roles},
-        # }
-        # Ask the LLM to decide which role should handle the message
-        # chosen_role_name = self.llm.ask(self.prompt_template.format(context))
-
-        # FIXME: 现在通过简单的字典决定流向，但之后还是应该有思考过程
-        # The direction of flow is now determined by a simple dictionary, but there should still be a thought process afterwards
-        next_role_profile = self.role_directions[message.role]
-        # logger.debug(f"{next_role_profile}")
-        for _, role in roles.items():
-            if next_role_profile == role.profile:
-                next_role = role
-                break
-        else:
-            logger.error(f"No available role can handle message: {message}.")
-            return
-
-        # Find the chosen role and handle the message
-        return await next_role.handle(message)

metagpt/memory/longterm_memory.py

@@ -5,6 +5,10 @@
 @Modified By: mashenquan, 2023/8/20. Remove global configuration `CONFIG`, enable configuration support for business isolation.
 """
 
+from typing import Optional
+
+from pydantic import Field
+
 from metagpt.logs import logger
 from metagpt.memory import Memory
 from metagpt.memory.memory_storage import MemoryStorage
@@ -18,11 +22,12 @@ class LongTermMemory(Memory):
     - update memory when it changed
     """
 
-    def __init__(self):
-        self.memory_storage: MemoryStorage = MemoryStorage()
-        super(LongTermMemory, self).__init__()
-        self.rc = None  # RoleContext
-        self.msg_from_recover = False
+    memory_storage: MemoryStorage = Field(default_factory=MemoryStorage)
+    rc: Optional["RoleContext"] = None
+    msg_from_recover: bool = False
+
+    class Config:
+        arbitrary_types_allowed = True
 
     def recover_memory(self, role_id: str, rc: "RoleContext"):
         messages = self.memory_storage.recover_memory(role_id)
@@ -38,7 +43,7 @@ class LongTermMemory(Memory):
         self.msg_from_recover = False
 
     def add(self, message: Message):
-        super(LongTermMemory, self).add(message)
+        super().add(message)
         for action in self.rc.watch:
             if message.cause_by == action and not self.msg_from_recover:
                 # currently, only add role's watching messages to its memory_storage
@@ -51,7 +56,7 @@ class LongTermMemory(Memory):
             1. find the short-term memory(stm) news
             2. furthermore, filter out similar messages based on ltm(long-term memory), get the final news
         """
-        stm_news = super(LongTermMemory, self).find_news(observed, k=k)  # shot-term memory news
+        stm_news = super().find_news(observed, k=k)  # shot-term memory news
         if not self.memory_storage.is_initialized:
             # memory_storage hasn't initialized, use default `find_news` to get stm_news
             return stm_news
@@ -65,9 +70,9 @@ class LongTermMemory(Memory):
         return ltm_news[-k:]
 
     def delete(self, message: Message):
-        super(LongTermMemory, self).delete(message)
+        super().delete(message)
         # TODO delete message in memory_storage
 
     def clear(self):
-        super(LongTermMemory, self).clear()
+        super().clear()
         self.memory_storage.clean()

metagpt/memory/memory.py

@@ -7,19 +7,50 @@
 @Modified By: mashenquan, 2023-11-1. According to RFC 116: Updated the type of index key.
 """
 from collections import defaultdict
+from pathlib import Path
 from typing import Iterable, Set
 
+from pydantic import BaseModel, Field
+
 from metagpt.schema import Message
-from metagpt.utils.common import any_to_str, any_to_str_set
+from metagpt.utils.common import (
+    any_to_str,
+    any_to_str_set,
+    read_json_file,
+    write_json_file,
+)
 
 
-class Memory:
+class Memory(BaseModel):
     """The most basic memory: super-memory"""
 
-    def __init__(self):
-        """Initialize an empty storage list and an empty index dictionary"""
-        self.storage: list[Message] = []
-        self.index: dict[str, list[Message]] = defaultdict(list)
+    storage: list[Message] = []
+    index: dict[str, list[Message]] = Field(default_factory=defaultdict(list))
+
+    def __init__(self, **kwargs):
+        index = kwargs.get("index", {})
+        new_index = defaultdict(list)
+        for action_str, value in index.items():
+            new_index[action_str] = [Message(**item_dict) for item_dict in value]
+        kwargs["index"] = new_index
+        super(Memory, self).__init__(**kwargs)
+        self.index = new_index
+
+    def serialize(self, stg_path: Path):
+        """stg_path = ./storage/team/environment/ or ./storage/team/environment/roles/{role_class}_{role_name}/"""
+        memory_path = stg_path.joinpath("memory.json")
+        storage = self.dict()
+        write_json_file(memory_path, storage)
+
+    @classmethod
+    def deserialize(cls, stg_path: Path) -> "Memory":
+        """stg_path = ./storage/team/environment/ or ./storage/team/environment/roles/{role_class}_{role_name}/"""
+        memory_path = stg_path.joinpath("memory.json")
+
+        memory_dict = read_json_file(memory_path)
+        memory = Memory(**memory_dict)
+
+        return memory
 
     def add(self, message: Message):
         """Add a new message to storage, while updating the index"""
@@ -41,6 +72,16 @@ class Memory:
         """Return all messages containing a specified content"""
         return [message for message in self.storage if content in message.content]
 
+    def delete_newest(self) -> "Message":
+        """delete the newest message from the storage"""
+        if len(self.storage) > 0:
+            newest_msg = self.storage.pop()
+            if newest_msg.cause_by and newest_msg in self.index[newest_msg.cause_by]:
+                self.index[newest_msg.cause_by].remove(newest_msg)
+        else:
+            newest_msg = None
+        return newest_msg
+
     def delete(self, message: Message):
         """Delete the specified message from storage, while updating the index"""
         self.storage.remove(message)

metagpt/memory/memory_storage.py

@@ -61,7 +61,7 @@ class MemoryStorage(FaissStore):
         return index_fpath, storage_fpath
 
     def persist(self):
-        super(MemoryStorage, self).persist()
+        super().persist()
         logger.debug(f"Agent {self.role_id} persist memory into local")
 
     def add(self, message: Message) -> bool:

metagpt/provider/__init__.py

@@ -4,23 +4,12 @@
 @Time    : 2023/5/5 22:59
 @Author  : alexanderwu
 @File    : __init__.py
-@Modified By: mashenquan, 2023-12-15. Add LLMType
 """
-from enum import Enum
 
+from metagpt.provider.fireworks_api import FireWorksGPTAPI
+from metagpt.provider.google_gemini_api import GeminiGPTAPI
+from metagpt.provider.open_llm_api import OpenLLMGPTAPI
+from metagpt.provider.openai_api import OpenAIGPTAPI
+from metagpt.provider.zhipuai_api import ZhiPuAIGPTAPI
 
-class LLMType(Enum):
-    OPENAI = "OpenAI"
-    METAGPT = "MetaGPT"
-    UNKNOWN = "UNKNOWN"
-
-    @classmethod
-    def get(cls, value):
-        for member in cls:
-            if member.value == value:
-                return member
-        return cls.UNKNOWN
-
-    @classmethod
-    def __missing__(cls, value):
-        return cls.UNKNOWN
+__all__ = ["FireWorksGPTAPI", "GeminiGPTAPI", "OpenLLMGPTAPI", "OpenAIGPTAPI", "ZhiPuAIGPTAPI"]

metagpt/provider/anthropic_api.py

@@ -14,7 +14,7 @@ from metagpt.config import CONFIG
 
 class Claude2:
     def ask(self, prompt):
-        client = Anthropic(api_key=CONFIG.claude_api_key)
+        client = Anthropic(api_key=CONFIG.anthropic_api_key)
 
         res = client.completions.create(
             model="claude-2",
@@ -24,7 +24,7 @@ class Claude2:
         return res.completion
 
     async def aask(self, prompt):
-        client = Anthropic(api_key=CONFIG.claude_api_key)
+        client = Anthropic(api_key=CONFIG.anthropic_api_key)
 
         res = client.completions.create(
             model="claude-2",

metagpt/provider/azure_openai_api.py

@@ -0,0 +1,77 @@
+# -*- coding: utf-8 -*-
+"""
+@Time    : 2023/5/5 23:08
+@Author  : alexanderwu
+@File    : openai.py
+@Modified By: mashenquan, 2023/8/20. Remove global configuration `CONFIG`, enable configuration support for business isolation;
+            Change cost control from global to company level.
+@Modified By: mashenquan, 2023/11/21. Fix bug: ReadTimeout.
+@Modified By: mashenquan, 2023/12/1. Fix bug: Unclosed connection caused by openai 0.x.
+"""
+
+
+from openai import AsyncAzureOpenAI, AzureOpenAI
+from openai._base_client import AsyncHttpxClientWrapper, SyncHttpxClientWrapper
+
+from metagpt.config import CONFIG, Config, LLMProviderEnum
+from metagpt.provider.llm_provider_registry import register_provider
+from metagpt.provider.openai_api import OpenAIGPTAPI, RateLimiter
+
+
+@register_provider(LLMProviderEnum.AZURE_OPENAI)
+class AzureOpenAIGPTAPI(OpenAIGPTAPI):
+    """
+    Check https://platform.openai.com/examples for examples
+    """
+
+    def __init__(self):
+        self.config: Config = CONFIG
+        self.__init_openai()
+        self.auto_max_tokens = False
+        # https://learn.microsoft.com/zh-cn/azure/ai-services/openai/how-to/migration?tabs=python-new%2Cdalle-fix
+        self._client = AsyncAzureOpenAI(
+            api_key=CONFIG.openai_api_key,
+            api_version=CONFIG.openai_api_version,
+            azure_endpoint=CONFIG.openai_api_base,
+        )
+        RateLimiter.__init__(self, rpm=self.rpm)
+
+    def _make_client(self):
+        kwargs, async_kwargs = self._make_client_kwargs()
+        self.client = AzureOpenAI(**kwargs)
+        self.async_client = AsyncAzureOpenAI(**async_kwargs)
+
+    def _make_client_kwargs(self) -> (dict, dict):
+        kwargs = dict(
+            api_key=self.config.openai_api_key,
+            api_version=self.config.openai_api_version,
+            azure_endpoint=self.config.openai_base_url,
+        )
+        async_kwargs = kwargs.copy()
+
+        # to use proxy, openai v1 needs http_client
+        proxy_params = self._get_proxy_params()
+        if proxy_params:
+            kwargs["http_client"] = SyncHttpxClientWrapper(**proxy_params)
+            async_kwargs["http_client"] = AsyncHttpxClientWrapper(**proxy_params)
+
+        return kwargs, async_kwargs
+
+    def _cons_kwargs(self, messages: list[dict], timeout=3, **configs) -> dict:
+        kwargs = {
+            "messages": messages,
+            "max_tokens": self.get_max_tokens(messages),
+            "n": 1,
+            "stop": None,
+            "temperature": 0.3,
+            "model": CONFIG.deployment_id,
+        }
+        if configs:
+            kwargs.update(configs)
+        try:
+            default_timeout = int(CONFIG.TIMEOUT) if CONFIG.TIMEOUT else 0
+        except ValueError:
+            default_timeout = 0
+        kwargs["timeout"] = max(default_timeout, timeout)
+
+        return kwargs

metagpt/provider/fireworks_api.py

@@ -4,18 +4,18 @@
 
 import openai
 
-from metagpt.config import CONFIG
+from metagpt.config import CONFIG, LLMProviderEnum
+from metagpt.provider.llm_provider_registry import register_provider
 from metagpt.provider.openai_api import OpenAIGPTAPI, RateLimiter
-from metagpt.utils.cost_manager import CostManager
 
 
+@register_provider(LLMProviderEnum.FIREWORKS)
 class FireWorksGPTAPI(OpenAIGPTAPI):
     def __init__(self):
         self.__init_fireworks(CONFIG)
         self.llm = openai
         self.model = CONFIG.fireworks_api_model
         self.auto_max_tokens = False
-        self._cost_manager = CostManager()
         RateLimiter.__init__(self, rpm=self.rpm)
 
     def __init_fireworks(self, config: "Config"):

metagpt/provider/google_gemini_api.py

@@ -0,0 +1,141 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+# @Desc   : Google Gemini LLM from https://ai.google.dev/tutorials/python_quickstart
+
+import google.generativeai as genai
+from google.ai import generativelanguage as glm
+from google.generativeai.generative_models import GenerativeModel
+from google.generativeai.types import content_types
+from google.generativeai.types.generation_types import (
+    AsyncGenerateContentResponse,
+    GenerateContentResponse,
+    GenerationConfig,
+)
+from tenacity import (
+    after_log,
+    retry,
+    retry_if_exception_type,
+    stop_after_attempt,
+    wait_random_exponential,
+)
+
+from metagpt.config import CONFIG, LLMProviderEnum
+from metagpt.logs import logger
+from metagpt.provider.base_gpt_api import BaseGPTAPI
+from metagpt.provider.llm_provider_registry import register_provider
+from metagpt.provider.openai_api import log_and_reraise
+
+
+class GeminiGenerativeModel(GenerativeModel):
+    """
+    Due to `https://github.com/google/generative-ai-python/pull/123`, inherit a new class.
+    Will use default GenerativeModel if it fixed.
+    """
+
+    def count_tokens(self, contents: content_types.ContentsType) -> glm.CountTokensResponse:
+        contents = content_types.to_contents(contents)
+        return self._client.count_tokens(model=self.model_name, contents=contents)
+
+    async def count_tokens_async(self, contents: content_types.ContentsType) -> glm.CountTokensResponse:
+        contents = content_types.to_contents(contents)
+        return await self._async_client.count_tokens(model=self.model_name, contents=contents)
+
+
+@register_provider(LLMProviderEnum.GEMINI)
+class GeminiGPTAPI(BaseGPTAPI):
+    """
+    Refs to `https://ai.google.dev/tutorials/python_quickstart`
+    """
+
+    def __init__(self):
+        self.use_system_prompt = False  # google gemini has no system prompt when use api
+
+        self.__init_gemini(CONFIG)
+        self.model = "gemini-pro"  # so far only one model
+        self.llm = GeminiGenerativeModel(model_name=self.model)
+
+    def __init_gemini(self, config: CONFIG):
+        genai.configure(api_key=config.gemini_api_key)
+
+    def _user_msg(self, msg: str) -> dict[str, str]:
+        # Not to change BaseGPTAPI default functions but update with Gemini's conversation format.
+        # You should follow the format.
+        return {"role": "user", "parts": [msg]}
+
+    def _assistant_msg(self, msg: str) -> dict[str, str]:
+        return {"role": "model", "parts": [msg]}
+
+    def _const_kwargs(self, messages: list[dict], stream: bool = False) -> dict:
+        kwargs = {"contents": messages, "generation_config": GenerationConfig(temperature=0.3), "stream": stream}
+        return kwargs
+
+    def _update_costs(self, usage: dict):
+        """update each request's token cost"""
+        if CONFIG.calc_usage:
+            try:
+                prompt_tokens = int(usage.get("prompt_tokens", 0))
+                completion_tokens = int(usage.get("completion_tokens", 0))
+                CONFIG.cost_manager.update_cost(prompt_tokens, completion_tokens, self.model)
+            except Exception as e:
+                logger.error(f"google gemini updats costs failed! exp: {e}")
+
+    def get_choice_text(self, resp: GenerateContentResponse) -> str:
+        return resp.text
+
+    def get_usage(self, messages: list[dict], resp_text: str) -> dict:
+        req_text = messages[-1]["parts"][0] if messages else ""
+        prompt_resp = self.llm.count_tokens(contents={"role": "user", "parts": [{"text": req_text}]})
+        completion_resp = self.llm.count_tokens(contents={"role": "model", "parts": [{"text": resp_text}]})
+        usage = {"prompt_tokens": prompt_resp.total_tokens, "completion_tokens": completion_resp.total_tokens}
+        return usage
+
+    async def aget_usage(self, messages: list[dict], resp_text: str) -> dict:
+        req_text = messages[-1]["parts"][0] if messages else ""
+        prompt_resp = await self.llm.count_tokens_async(contents={"role": "user", "parts": [{"text": req_text}]})
+        completion_resp = await self.llm.count_tokens_async(contents={"role": "model", "parts": [{"text": resp_text}]})
+        usage = {"prompt_tokens": prompt_resp.total_tokens, "completion_tokens": completion_resp.total_tokens}
+        return usage
+
+    def completion(self, messages: list[dict]) -> "GenerateContentResponse":
+        resp: GenerateContentResponse = self.llm.generate_content(**self._const_kwargs(messages))
+        usage = self.get_usage(messages, resp.text)
+        self._update_costs(usage)
+        return resp
+
+    async def _achat_completion(self, messages: list[dict]) -> "AsyncGenerateContentResponse":
+        resp: AsyncGenerateContentResponse = await self.llm.generate_content_async(**self._const_kwargs(messages))
+        usage = await self.aget_usage(messages, resp.text)
+        self._update_costs(usage)
+        return resp
+
+    async def acompletion(self, messages: list[dict]) -> dict:
+        return await self._achat_completion(messages)
+
+    async def _achat_completion_stream(self, messages: list[dict]) -> str:
+        resp: AsyncGenerateContentResponse = await self.llm.generate_content_async(
+            **self._const_kwargs(messages, stream=True)
+        )
+        collected_content = []
+        async for chunk in resp:
+            content = chunk.text
+            print(content, end="")
+            collected_content.append(content)
+
+        full_content = "".join(collected_content)
+        usage = await self.aget_usage(messages, full_content)
+        self._update_costs(usage)
+        return full_content
+
+    @retry(
+        stop=stop_after_attempt(3),
+        wait=wait_random_exponential(min=1, max=60),
+        after=after_log(logger, logger.level("WARNING").name),
+        retry=retry_if_exception_type(ConnectionError),
+        retry_error_callback=log_and_reraise,
+    )
+    async def acompletion_text(self, messages: list[dict], stream=False) -> str:
+        """response in async with stream or non-stream mode"""
+        if stream:
+            return await self._achat_completion_stream(messages)
+        resp = await self._achat_completion(messages)
+        return self.get_choice_text(resp)

metagpt/provider/llm_provider_registry.py

@@ -0,0 +1,34 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+"""
+@Time    : 2023/12/19 17:26
+@Author  : alexanderwu
+@File    : llm_provider_registry.py
+"""
+from metagpt.config import LLMProviderEnum
+
+
+class LLMProviderRegistry:
+    def __init__(self):
+        self.providers = {}
+
+    def register(self, key, provider_cls):
+        self.providers[key] = provider_cls
+
+    def get_provider(self, enum: LLMProviderEnum):
+        """get provider instance according to the enum"""
+        return self.providers[enum]()
+
+
+# Registry instance
+LLM_REGISTRY = LLMProviderRegistry()
+
+
+def register_provider(key):
+    """register provider to registry"""
+
+    def decorator(cls):
+        LLM_REGISTRY.register(key, cls)
+        return cls
+
+    return decorator

metagpt/provider/metagpt_api.py

@@ -0,0 +1,16 @@
+# -*- coding: utf-8 -*-
+"""
+@Time    : 2023/5/5 23:08
+@Author  : alexanderwu
+@File    : metagpt_api.py
+@Desc    : MetaGPT LLM provider.
+"""
+from metagpt.config import LLMProviderEnum
+from metagpt.provider import OpenAIGPTAPI
+from metagpt.provider.llm_provider_registry import register_provider
+
+
+@register_provider(LLMProviderEnum.METAGPT)
+class METAGPTAPI(OpenAIGPTAPI):
+    def __init__(self):
+        super().__init__()

metagpt/provider/metagpt_llm_api.py

@@ -1,279 +0,0 @@
-# -*- coding: utf-8 -*-
-"""
-@Time    : 2023/8/30
-@Author  : mashenquan
-@File    : metagpt_llm_api.py
-@Desc    : MetaGPT LLM related APIs
-"""
-
-from metagpt.provider.openai_api import OpenAIGPTAPI
-
-# from metagpt.provider.base_gpt_api import BaseGPTAPI
-# from metagpt.provider.openai_api import RateLimiter
-
-
-class MetaGPTLLMAPI(OpenAIGPTAPI):
-    """MetaGPT LLM api"""
-
-    def __init__(self):
-        super(MetaGPTLLMAPI, self).__init__()
-
-    # def __init__(self):
-    #     self.__init_openai(CONFIG)
-    #     self.llm = openai
-    #     self.model = CONFIG.openai_api_model
-    #     self.auto_max_tokens = False
-    #     self._cost_manager = CostManager()
-    #     RateLimiter.__init__(self, rpm=self.rpm)
-    #
-    # def __init_openai(self, config):
-    #     openai.api_key = config.openai_api_key
-    #     if config.openai_api_base:
-    #         openai.api_base = config.openai_api_base
-    #     if config.openai_api_type:
-    #         openai.api_type = config.openai_api_type
-    #         openai.api_version = config.openai_api_version
-    #     self.rpm = int(config.get("RPM", 10))
-    #
-    # async def _achat_completion_stream(self, messages: list[dict]) -> str:
-    #     response = await openai.ChatCompletion.acreate(**self._cons_kwargs(messages), stream=True)
-    #
-    #     # create variables to collect the stream of chunks
-    #     collected_chunks = []
-    #     collected_messages = []
-    #     # iterate through the stream of events
-    #     async for chunk in response:
-    #         collected_chunks.append(chunk)  # save the event response
-    #         choices = chunk["choices"]
-    #         if len(choices) > 0:
-    #             chunk_message = chunk["choices"][0].get("delta", {})  # extract the message
-    #             collected_messages.append(chunk_message)  # save the message
-    #             if "content" in chunk_message:
-    #                 print(chunk_message["content"], end="")
-    #     print()
-    #
-    #     full_reply_content = "".join([m.get("content", "") for m in collected_messages])
-    #     usage = self._calc_usage(messages, full_reply_content)
-    #     self._update_costs(usage)
-    #     return full_reply_content
-    #
-    # def _cons_kwargs(self, messages: list[dict], **configs) -> dict:
-    #     kwargs = {
-    #         "messages": messages,
-    #         "max_tokens": self.get_max_tokens(messages),
-    #         "n": 1,
-    #         "stop": None,
-    #         "temperature": 0.3,
-    #         "timeout": 3,
-    #     }
-    #     if configs:
-    #         kwargs.update(configs)
-    #
-    #     if CONFIG.openai_api_type == "azure":
-    #         if CONFIG.deployment_name and CONFIG.deployment_id:
-    #             raise ValueError("You can only use one of the `deployment_id` or `deployment_name` model")
-    #         elif not CONFIG.deployment_name and not CONFIG.deployment_id:
-    #             raise ValueError("You must specify `DEPLOYMENT_NAME` or `DEPLOYMENT_ID` parameter")
-    #         kwargs_mode = (
-    #             {"engine": CONFIG.deployment_name}
-    #             if CONFIG.deployment_name
-    #             else {"deployment_id": CONFIG.deployment_id}
-    #         )
-    #     else:
-    #         kwargs_mode = {"model": self.model}
-    #     kwargs.update(kwargs_mode)
-    #     return kwargs
-    #
-    # async def _achat_completion(self, messages: list[dict]) -> dict:
-    #     rsp = await self.llm.ChatCompletion.acreate(**self._cons_kwargs(messages))
-    #     self._update_costs(rsp.get("usage"))
-    #     return rsp
-    #
-    # def _chat_completion(self, messages: list[dict]) -> dict:
-    #     rsp = self.llm.ChatCompletion.create(**self._cons_kwargs(messages))
-    #     self._update_costs(rsp)
-    #     return rsp
-    #
-    # def completion(self, messages: list[dict]) -> dict:
-    #     # if isinstance(messages[0], Message):
-    #     #     messages = self.messages_to_dict(messages)
-    #     return self._chat_completion(messages)
-    #
-    # async def acompletion(self, messages: list[dict]) -> dict:
-    #     # if isinstance(messages[0], Message):
-    #     #     messages = self.messages_to_dict(messages)
-    #     return await self._achat_completion(messages)
-    #
-    # @retry(
-    #     wait=wait_random_exponential(min=1, max=60),
-    #     stop=stop_after_attempt(6),
-    #     after=after_log(logger, logger.level("WARNING").name),
-    #     retry=retry_if_exception_type(APIConnectionError),
-    #     retry_error_callback=log_and_reraise,
-    # )
-    # async def acompletion_text(self, messages: list[dict], stream=False) -> str:
-    #     """when streaming, print each token in place."""
-    #     if stream:
-    #         return await self._achat_completion_stream(messages)
-    #     rsp = await self._achat_completion(messages)
-    #     return self.get_choice_text(rsp)
-    #
-    # def _func_configs(self, messages: list[dict], **kwargs) -> dict:
-    #     """
-    #     Note: Keep kwargs consistent with the parameters in the https://platform.openai.com/docs/api-reference/chat/create
-    #     """
-    #     if "tools" not in kwargs:
-    #         configs = {
-    #             "tools": [{"type": "function", "function": GENERAL_FUNCTION_SCHEMA}],
-    #             "tool_choice": GENERAL_TOOL_CHOICE,
-    #         }
-    #         kwargs.update(configs)
-    #
-    #     return self._cons_kwargs(messages, **kwargs)
-    #
-    # def _chat_completion_function(self, messages: list[dict], **kwargs) -> dict:
-    #     rsp = self.llm.ChatCompletion.create(**self._func_configs(messages, **kwargs))
-    #     self._update_costs(rsp.get("usage"))
-    #     return rsp
-    #
-    # async def _achat_completion_function(self, messages: list[dict], **chat_configs) -> dict:
-    #     rsp = await self.llm.ChatCompletion.acreate(**self._func_configs(messages, **chat_configs))
-    #     self._update_costs(rsp.get("usage"))
-    #     return rsp
-    #
-    # def _process_message(self, messages: Union[str, Message, list[dict], list[Message], list[str]]) -> list[dict]:
-    #     """convert messages to list[dict]."""
-    #     if isinstance(messages, list):
-    #         messages = [Message(msg) if isinstance(msg, str) else msg for msg in messages]
-    #         return [msg if isinstance(msg, dict) else msg.to_dict() for msg in messages]
-    #
-    #     if isinstance(messages, Message):
-    #         messages = [messages.to_dict()]
-    #     elif isinstance(messages, str):
-    #         messages = [{"role": "user", "content": messages}]
-    #     else:
-    #         raise ValueError(
-    #             f"Only support messages type are: str, Message, list[dict], but got {type(messages).__name__}!"
-    #         )
-    #     return messages
-    #
-    # def ask_code(self, messages: Union[str, Message, list[dict]], **kwargs) -> dict:
-    #     """Use function of tools to ask a code.
-    #
-    #     Note: Keep kwargs consistent with the parameters in the https://platform.openai.com/docs/api-reference/chat/create
-    #
-    #     Examples:
-    #
-    #     >>> llm = OpenAIGPTAPI()
-    #     >>> llm.ask_code("Write a python hello world code.")
-    #     {'language': 'python', 'code': "print('Hello, World!')"}
-    #     >>> msg = [{'role': 'user', 'content': "Write a python hello world code."}]
-    #     >>> llm.ask_code(msg)
-    #     {'language': 'python', 'code': "print('Hello, World!')"}
-    #     """
-    #     messages = self._process_message(messages)
-    #     rsp = self._chat_completion_function(messages, **kwargs)
-    #     return self.get_choice_function_arguments(rsp)
-    #
-    # async def aask_code(self, messages: Union[str, Message, list[dict]], **kwargs) -> dict:
-    #     """Use function of tools to ask a code.
-    #
-    #     Note: Keep kwargs consistent with the parameters in the https://platform.openai.com/docs/api-reference/chat/create
-    #
-    #     Examples:
-    #
-    #     >>> llm = OpenAIGPTAPI()
-    #     >>> rsp = await llm.ask_code("Write a python hello world code.")
-    #     >>> rsp
-    #     {'language': 'python', 'code': "print('Hello, World!')"}
-    #     >>> msg = [{'role': 'user', 'content': "Write a python hello world code."}]
-    #     >>> rsp = await llm.aask_code(msg)   # -> {'language': 'python', 'code': "print('Hello, World!')"}
-    #     """
-    #     messages = self._process_message(messages)
-    #     rsp = await self._achat_completion_function(messages, **kwargs)
-    #     return self.get_choice_function_arguments(rsp)
-    #
-    # def _calc_usage(self, messages: list[dict], rsp: str) -> dict:
-    #     usage = {}
-    #     if CONFIG.calc_usage:
-    #         try:
-    #             prompt_tokens = count_message_tokens(messages, self.model)
-    #             completion_tokens = count_string_tokens(rsp, self.model)
-    #             usage["prompt_tokens"] = prompt_tokens
-    #             usage["completion_tokens"] = completion_tokens
-    #             return usage
-    #         except Exception as e:
-    #             logger.error("usage calculation failed!", e)
-    #     else:
-    #         return usage
-    #
-    # async def acompletion_batch(self, batch: list[list[dict]]) -> list[dict]:
-    #     """Return full JSON"""
-    #     split_batches = self.split_batches(batch)
-    #     all_results = []
-    #
-    #     for small_batch in split_batches:
-    #         logger.info(small_batch)
-    #         await self.wait_if_needed(len(small_batch))
-    #
-    #         future = [self.acompletion(prompt) for prompt in small_batch]
-    #         results = await asyncio.gather(*future)
-    #         logger.info(results)
-    #         all_results.extend(results)
-    #
-    #     return all_results
-    #
-    # async def acompletion_batch_text(self, batch: list[list[dict]]) -> list[str]:
-    #     """Only return plain text"""
-    #     raw_results = await self.acompletion_batch(batch)
-    #     results = []
-    #     for idx, raw_result in enumerate(raw_results, start=1):
-    #         result = self.get_choice_text(raw_result)
-    #         results.append(result)
-    #         logger.info(f"Result of task {idx}: {result}")
-    #     return results
-    #
-    # def _update_costs(self, usage: dict):
-    #     if CONFIG.calc_usage:
-    #         try:
-    #             prompt_tokens = int(usage["prompt_tokens"])
-    #             completion_tokens = int(usage["completion_tokens"])
-    #             self._cost_manager.update_cost(prompt_tokens, completion_tokens, self.model)
-    #         except Exception as e:
-    #             logger.error("updating costs failed!", e)
-    #
-    # def get_costs(self) -> Costs:
-    #     return self._cost_manager.get_costs()
-    #
-    # def get_max_tokens(self, messages: list[dict]):
-    #     if not self.auto_max_tokens:
-    #         return CONFIG.max_tokens_rsp
-    #     return get_max_completion_tokens(messages, self.model, CONFIG.max_tokens_rsp)
-    #
-    # def moderation(self, content: Union[str, list[str]]):
-    #     try:
-    #         if not content:
-    #             logger.error("content cannot be empty!")
-    #         else:
-    #             rsp = self._moderation(content=content)
-    #             return rsp
-    #     except Exception as e:
-    #         logger.error(f"moderating failed:{e}")
-    #
-    # def _moderation(self, content: Union[str, list[str]]):
-    #     rsp = self.llm.Moderation.create(input=content)
-    #     return rsp
-    #
-    # async def amoderation(self, content: Union[str, list[str]]):
-    #     try:
-    #         if not content:
-    #             logger.error("content cannot be empty!")
-    #         else:
-    #             rsp = await self._amoderation(content=content)
-    #             return rsp
-    #     except Exception as e:
-    #         logger.error(f"moderating failed:{e}")
-    #
-    # async def _amoderation(self, content: Union[str, list[str]]):
-    #     rsp = await self.llm.Moderation.acreate(input=content)
-    #     return rsp

metagpt/provider/open_llm_api.py

@@ -2,40 +2,39 @@
 # -*- coding: utf-8 -*-
 # @Desc   : self-host open llm model with openai-compatible interface
 
-from metagpt.config import CONFIG
-from metagpt.logs import logger
+
+from metagpt.config import CONFIG, LLMProviderEnum
+from metagpt.provider.llm_provider_registry import register_provider
 from metagpt.provider.openai_api import OpenAIGPTAPI, RateLimiter
-from metagpt.utils.cost_manager import CostManager
-
-
-class OpenLLMCostManager(CostManager):
-    """open llm model is self-host, it's free and without cost"""
-
-    def update_cost(self, prompt_tokens, completion_tokens, model):
-        """
-        Update the total cost, prompt tokens, and completion tokens.
-
-        Args:
-        prompt_tokens (int): The number of tokens used in the prompt.
-        completion_tokens (int): The number of tokens used in the completion.
-        model (str): The model used for the API call.
-        """
-        self.total_prompt_tokens += prompt_tokens
-        self.total_completion_tokens += completion_tokens
-
-        logger.info(
-            f"Max budget: ${CONFIG.max_budget:.3f} | "
-            f"prompt_tokens: {prompt_tokens}, completion_tokens: {completion_tokens}"
-        )
-        CONFIG.total_cost = self.total_cost
+
+# class OpenLLMCostManager(CostManager):
+#     """open llm model is self-host, it's free and without cost"""
+#
+#     def update_cost(self, prompt_tokens, completion_tokens, model):
+#         """
+#         Update the total cost, prompt tokens, and completion tokens.
+#
+#         Args:
+#         prompt_tokens (int): The number of tokens used in the prompt.
+#         completion_tokens (int): The number of tokens used in the completion.
+#         model (str): The model used for the API call.
+#         """
+#         self.total_prompt_tokens += prompt_tokens
+#         self.total_completion_tokens += completion_tokens
+#
+#         logger.info(
+#             f"Max budget: ${CONFIG.max_budget:.3f} | "
+#             f"prompt_tokens: {prompt_tokens}, completion_tokens: {completion_tokens}"
+#         )
+#         CONFIG.total_cost = self.total_cost
 
 
+@register_provider(LLMProviderEnum.OPEN_LLM)
 class OpenLLMGPTAPI(OpenAIGPTAPI):
     def __init__(self):
         self.__init_openllm(CONFIG)
         self.model = CONFIG.open_llm_api_model
         self.auto_max_tokens = False
-        self._cost_manager = OpenLLMCostManager()
         RateLimiter.__init__(self, rpm=self.rpm)
 
     def __init_openllm(self, config: "Config"):

metagpt/provider/openai_api.py

@@ -10,12 +10,15 @@
 """
 
 import asyncio
+import json
 import time
-from typing import Union
+from typing import List, Union
 
 import openai
-from openai import APIConnectionError, AsyncAzureOpenAI, AsyncOpenAI, RateLimitError
+from openai import APIConnectionError, AsyncOpenAI, AsyncStream, OpenAI, RateLimitError
+from openai._base_client import AsyncHttpxClientWrapper, SyncHttpxClientWrapper
 from openai.types import CompletionUsage
+from openai.types.chat import ChatCompletion, ChatCompletionChunk
 from tenacity import (
     after_log,
     retry,
@@ -24,13 +27,15 @@ from tenacity import (
     wait_random_exponential,
 )
 
-from metagpt.config import CONFIG
+from metagpt.config import CONFIG, Config, LLMProviderEnum
+from metagpt.const import DEFAULT_MAX_TOKENS
 from metagpt.logs import logger
-from metagpt.provider import LLMType
 from metagpt.provider.base_gpt_api import BaseGPTAPI
 from metagpt.provider.constant import GENERAL_FUNCTION_SCHEMA, GENERAL_TOOL_CHOICE
+from metagpt.provider.llm_provider_registry import register_provider
 from metagpt.schema import Message
 from metagpt.utils.cost_manager import Costs
+from metagpt.utils.exceptions import handle_exception
 from metagpt.utils.token_counter import (
     count_message_tokens,
     count_string_tokens,
@@ -74,25 +79,18 @@ See FAQ 5.8
     raise retry_state.outcome.exception()
 
 
+@register_provider(LLMProviderEnum.OPENAI)
 class OpenAIGPTAPI(BaseGPTAPI, RateLimiter):
     """
     Check https://platform.openai.com/examples for examples
     """
 
     def __init__(self):
-        self.model = CONFIG.openai_api_model
+        self.config: Config = CONFIG
+        self.__init_openai()
         self.auto_max_tokens = False
-        self.rpm = int(CONFIG.get("RPM", 10))
-        if CONFIG.openai_api_type == "azure":
-            # https://learn.microsoft.com/zh-cn/azure/ai-services/openai/how-to/migration?tabs=python-new%2Cdalle-fix
-            self._client = AsyncAzureOpenAI(
-                api_key=CONFIG.openai_api_key,
-                api_version=CONFIG.openai_api_version,
-                azure_endpoint=CONFIG.openai_api_base,
-            )
-        else:
-            # https://github.com/openai/openai-python#async-usage
-            self._client = AsyncOpenAI(api_key=CONFIG.openai_api_key, base_url=CONFIG.openai_api_base)
+        # https://github.com/openai/openai-python#async-usage
+        self._client = AsyncOpenAI(api_key=CONFIG.openai_api_key, base_url=CONFIG.openai_api_base)
         RateLimiter.__init__(self, rpm=self.rpm)
 
     async def _achat_completion_stream(self, messages: list[dict], timeout=3) -> str:
@@ -103,6 +101,59 @@ class OpenAIGPTAPI(BaseGPTAPI, RateLimiter):
             chunk_message = chunk.choices[0].delta.content or ""  # extract the message
             yield chunk_message
 
+    def __init_openai(self):
+        self.rpm = int(self.config.get("RPM", 10))
+        self._make_client()
+
+    def _make_client(self):
+        kwargs, async_kwargs = self._make_client_kwargs()
+        self.client = OpenAI(**kwargs)
+        self.async_client = AsyncOpenAI(**async_kwargs)
+
+    def _make_client_kwargs(self) -> (dict, dict):
+        kwargs = dict(api_key=self.config.openai_api_key, base_url=self.config.openai_base_url)
+        async_kwargs = kwargs.copy()
+
+        # to use proxy, openai v1 needs http_client
+        proxy_params = self._get_proxy_params()
+        if proxy_params:
+            kwargs["http_client"] = SyncHttpxClientWrapper(**proxy_params)
+            async_kwargs["http_client"] = AsyncHttpxClientWrapper(**proxy_params)
+
+        return kwargs, async_kwargs
+
+    def _get_proxy_params(self) -> dict:
+        params = {}
+        if self.config.openai_proxy:
+            params = {"proxies": self.config.openai_proxy}
+            if self.config.openai_base_url:
+                params["base_url"] = self.config.openai_base_url
+
+        return params
+
+    async def _achat_completion_stream(self, messages: list[dict]) -> str:
+        response: AsyncStream[ChatCompletionChunk] = await self.async_client.chat.completions.create(
+            **self._cons_kwargs(messages), stream=True
+        )
+
+        # create variables to collect the stream of chunks
+        collected_chunks = []
+        collected_messages = []
+        # iterate through the stream of events
+        async for chunk in response:
+            collected_chunks.append(chunk)  # save the event response
+            if chunk.choices:
+                chunk_message = chunk.choices[0].delta  # extract the message
+                collected_messages.append(chunk_message)  # save the message
+                if chunk_message.content:
+                    print(chunk_message.content, end="")
+        print()
+
+        full_reply_content = "".join([m.content for m in collected_messages if m.content])
+        usage = self._calc_usage(messages, full_reply_content)
+        self._update_costs(usage)
+        return full_reply_content
+
     def _cons_kwargs(self, messages: list[dict], timeout=3, **configs) -> dict:
         kwargs = {
             "messages": messages,
@@ -110,14 +161,10 @@ class OpenAIGPTAPI(BaseGPTAPI, RateLimiter):
             "n": 1,
             "stop": None,
             "temperature": 0.3,
+            "model": self.config.openai_api_model,
         }
         if configs:
             kwargs.update(configs)
-
-        if CONFIG.openai_api_type == "azure":
-            kwargs["model"] = CONFIG.deployment_id
-        else:
-            kwargs["model"] = self.model
         try:
             default_timeout = int(CONFIG.TIMEOUT) if CONFIG.TIMEOUT else 0
         except ValueError:
@@ -126,19 +173,17 @@ class OpenAIGPTAPI(BaseGPTAPI, RateLimiter):
 
         return kwargs
 
-    async def _achat_completion(self, messages: list[dict], timeout=3) -> dict:
+    async def _achat_completion(self, messages: list[dict], timeout=3) -> ChatCompletion:
         kwargs = self._cons_kwargs(messages, timeout=timeout)
-        rsp = await self._client.chat.completions.create(**kwargs)
+        rsp: ChatCompletion = await self._client.chat.completions.create(**kwargs)
         self._update_costs(rsp.usage)
-        return rsp.dict()
+        return rsp
 
-    def completion(self, messages: list[dict], timeout=3) -> dict:
+    def completion(self, messages: list[dict], timeout=3) -> ChatCompletion:
         loop = self.get_event_loop()
         return loop.run_until_complete(self.acompletion(messages, timeout=timeout))
 
-    async def acompletion(self, messages: list[dict], timeout=3) -> dict:
-        # if isinstance(messages[0], Message):
-        #     messages = self.messages_to_dict(messages)
+    async def acompletion(self, messages: list[dict], timeout=3) -> ChatCompletion:
         return await self._achat_completion(messages, timeout=timeout)
 
     @retry(
@@ -188,20 +233,20 @@ class OpenAIGPTAPI(BaseGPTAPI, RateLimiter):
 
         return self._cons_kwargs(messages=messages, timeout=timeout, **kwargs)
 
-    def _chat_completion_function(self, messages: list[dict], timeout=3, **kwargs) -> dict:
+    def _chat_completion_function(self, messages: list[dict], timeout=3, **kwargs) -> ChatCompletion:
         loop = self.get_event_loop()
         return loop.run_until_complete(self._achat_completion_function(messages=messages, timeout=timeout, **kwargs))
 
-    async def _achat_completion_function(self, messages: list[dict], timeout=3, **chat_configs) -> dict:
+    async def _achat_completion_function(self, messages: list[dict], timeout=3, **chat_configs) -> ChatCompletion:
         kwargs = self._func_configs(messages=messages, timeout=timeout, **chat_configs)
-        rsp = await self._client.chat.completions.create(**kwargs)
+        rsp: ChatCompletion = await self._client.chat.completions.create(**kwargs)
         self._update_costs(rsp.usage)
-        return rsp.dict()
+        return rsp
 
     def _process_message(self, messages: Union[str, Message, list[dict], list[Message], list[str]]) -> list[dict]:
         """convert messages to list[dict]."""
         if isinstance(messages, list):
-            messages = [Message(msg) if isinstance(msg, str) else msg for msg in messages]
+            messages = [Message(content=msg) if isinstance(msg, str) else msg for msg in messages]
             return [msg if isinstance(msg, dict) else msg.to_dict() for msg in messages]
 
         if isinstance(messages, Message):
@@ -269,7 +314,35 @@ class OpenAIGPTAPI(BaseGPTAPI, RateLimiter):
                 logger.error(f"{self.model} usage calculation failed!", e)
         return CompletionUsage(prompt_tokens=0, completion_tokens=0, total_tokens=0)
 
-    async def acompletion_batch(self, batch: list[list[dict]], timeout=3) -> list[dict]:
+    def get_choice_function_arguments(self, rsp: ChatCompletion) -> dict:
+        """Required to provide the first function arguments of choice.
+
+        :return dict: return the first function arguments of choice, for example,
+            {'language': 'python', 'code': "print('Hello, World!')"}
+        """
+        try:
+            return json.loads(rsp.choices[0].message.tool_calls[0].function.arguments)
+        except json.JSONDecodeError:
+            return {}
+
+    def get_choice_text(self, rsp: ChatCompletion) -> str:
+        """Required to provide the first text of choice"""
+        return rsp.choices[0].message.content if rsp.choices else ""
+
+    def _calc_usage(self, messages: list[dict], rsp: str) -> CompletionUsage:
+        usage = CompletionUsage(prompt_tokens=0, completion_tokens=0, total_tokens=0)
+        if not CONFIG.calc_usage:
+            return usage
+
+        try:
+            usage.prompt_tokens = count_message_tokens(messages, self.model)
+            usage.completion_tokens = count_string_tokens(rsp, self.model)
+        except Exception as e:
+            logger.error(f"usage calculation failed!: {e}")
+
+        return usage
+
+    async def acompletion_batch(self, batch: list[list[dict]], timeout=3) -> list[ChatCompletion]:
         """Return full JSON"""
         split_batches = self.split_batches(batch)
         all_results = []
@@ -296,11 +369,9 @@ class OpenAIGPTAPI(BaseGPTAPI, RateLimiter):
         return results
 
     def _update_costs(self, usage: CompletionUsage):
-        if CONFIG.calc_usage:
+        if CONFIG.calc_usage and usage:
             try:
-                prompt_tokens = usage.prompt_tokens
-                completion_tokens = usage.completion_tokens
-                CONFIG.cost_manager.update_cost(prompt_tokens, completion_tokens, self.model)
+                CONFIG.cost_manager.update_cost(usage.prompt_tokens, usage.completion_tokens, self.model)
             except Exception as e:
                 logger.error("updating costs failed!", e)
 
@@ -316,19 +387,9 @@ class OpenAIGPTAPI(BaseGPTAPI, RateLimiter):
         loop = self.get_event_loop()
         loop.run_until_complete(self.amoderation(content=content))
 
+    @handle_exception
     async def amoderation(self, content: Union[str, list[str]]):
-        try:
-            if not content:
-                logger.error("content cannot be empty!")
-            else:
-                rsp = await self._amoderation(content=content)
-                return rsp
-        except Exception as e:
-            logger.error(f"moderating failed:{e}")
-
-    async def _amoderation(self, content: Union[str, list[str]]):
-        rsp = await self._client.moderations.create(input=content)
-        return rsp
+        return await self._client.moderations.create(input=content)
 
     async def close(self):
         """Close connection"""
@@ -349,8 +410,73 @@ class OpenAIGPTAPI(BaseGPTAPI, RateLimiter):
             else:
                 raise e
 
-    async def get_summary(self, text: str, max_words=200, keep_language: bool = False, **kwargs) -> str:
-        from metagpt.memory.brain_memory import BrainMemory
+    async def summarize(self, text: str, max_words=200, keep_language: bool = False, limit: int = -1, **kwargs) -> str:
+        max_token_count = DEFAULT_MAX_TOKENS
+        max_count = 100
+        text_length = len(text)
+        if limit > 0 and text_length < limit:
+            return text
+        summary = ""
+        while max_count > 0:
+            if text_length < max_token_count:
+                summary = await self._get_summary(text=text, max_words=max_words, keep_language=keep_language)
+                break
 
-        memory = BrainMemory(llm_type=LLMType.OPENAI.value, historical_summary=text, cacheable=False)
-        return await memory.summarize(llm=self, max_words=max_words, keep_language=keep_language)
+            padding_size = 20 if max_token_count > 20 else 0
+            text_windows = self.split_texts(text, window_size=max_token_count - padding_size)
+            part_max_words = min(int(max_words / len(text_windows)) + 1, 100)
+            summaries = []
+            for ws in text_windows:
+                response = await self._get_summary(text=ws, max_words=part_max_words, keep_language=keep_language)
+                summaries.append(response)
+            if len(summaries) == 1:
+                summary = summaries[0]
+                break
+
+            # Merged and retry
+            text = "\n".join(summaries)
+            text_length = len(text)
+
+            max_count -= 1  # safeguard
+        return summary
+
+    async def _get_summary(self, text: str, max_words=20, keep_language: bool = False):
+        """Generate text summary"""
+        if len(text) < max_words:
+            return text
+        if keep_language:
+            command = f".Translate the above content into a summary of less than {max_words} words in language of the content strictly."
+        else:
+            command = f"Translate the above content into a summary of less than {max_words} words."
+        msg = text + "\n\n" + command
+        logger.debug(f"summary ask:{msg}")
+        response = await self.aask(msg=msg, system_msgs=[])
+        logger.debug(f"summary rsp: {response}")
+        return response
+
+    @staticmethod
+    def split_texts(text: str, window_size) -> List[str]:
+        """Splitting long text into sliding windows text"""
+        if window_size <= 0:
+            window_size = DEFAULT_TOKEN_SIZE
+        total_len = len(text)
+        if total_len <= window_size:
+            return [text]
+
+        padding_size = 20 if window_size > 20 else 0
+        windows = []
+        idx = 0
+        data_len = window_size - padding_size
+        while idx < total_len:
+            if window_size + idx > total_len:  # 不足一个滑窗
+                windows.append(text[idx:])
+                break
+            # 每个窗口少算padding_size自然就可实现滑窗功能, 比如: [1, 2, 3, 4, 5, 6, 7, ....]
+            # window_size=3, padding_size=1：
+            # [1, 2, 3], [3, 4, 5], [5, 6, 7], ....
+            #   idx=2,  |  idx=5   |  idx=8  | ...
+            w = text[idx : idx + window_size]
+            windows.append(w)
+            idx += data_len
+
+        return windows

metagpt/provider/postprecess/base_postprecess_plugin.py

@@ -4,7 +4,6 @@
 
 from typing import Union
 
-from metagpt.logs import logger
 from metagpt.utils.repair_llm_raw_output import (
     RepairType,
     extract_content_from_output,
@@ -44,7 +43,7 @@ class BasePostPrecessPlugin(object):
 
     def run_retry_parse_json_text(self, content: str) -> Union[dict, list]:
         """inherited class can re-implement the function"""
-        logger.info(f"extracted json CONTENT from output:\n{content}")
+        # logger.info(f"extracted json CONTENT from output:\n{content}")
         parsed_data = retry_parse_json_text(output=content)  # should use output=content
         return parsed_data
 

metagpt/provider/spark_api.py

@@ -19,11 +19,13 @@ from wsgiref.handlers import format_date_time
 
 import websocket  # 使用websocket_client
 
-from metagpt.config import CONFIG
+from metagpt.config import CONFIG, LLMProviderEnum
 from metagpt.logs import logger
 from metagpt.provider.base_gpt_api import BaseGPTAPI
+from metagpt.provider.llm_provider_registry import register_provider
 
 
+@register_provider(LLMProviderEnum.SPARK)
 class SparkAPI(BaseGPTAPI):
     def __init__(self):
         logger.warning("当前方法无法支持异步运行。当你使用acompletion时，并不能并行访问。")

metagpt/provider/zhipuai_api.py

@@ -15,12 +15,12 @@ from tenacity import (
     wait_random_exponential,
 )
 
-from metagpt.config import CONFIG
+from metagpt.config import CONFIG, LLMProviderEnum
 from metagpt.logs import logger
 from metagpt.provider.base_gpt_api import BaseGPTAPI
+from metagpt.provider.llm_provider_registry import register_provider
 from metagpt.provider.openai_api import log_and_reraise
 from metagpt.provider.zhipuai.zhipu_model_api import ZhiPuModelAPI
-from metagpt.utils.cost_manager import CostManager
 
 
 class ZhiPuEvent(Enum):
@@ -30,6 +30,7 @@ class ZhiPuEvent(Enum):
     FINISH = "finish"
 
 
+@register_provider(LLMProviderEnum.ZHIPUAI)
 class ZhiPuAIGPTAPI(BaseGPTAPI):
     """
     Refs to `https://open.bigmodel.cn/dev/api#chatglm_turbo`
@@ -42,7 +43,6 @@ class ZhiPuAIGPTAPI(BaseGPTAPI):
         self.__init_zhipuai(CONFIG)
         self.llm = ZhiPuModelAPI
         self.model = "chatglm_turbo"  # so far only one model, just use it
-        self._cost_manager = CostManager()
 
     def __init_zhipuai(self, config: CONFIG):
         assert config.zhipuai_api_key
@@ -60,9 +60,9 @@ class ZhiPuAIGPTAPI(BaseGPTAPI):
             try:
                 prompt_tokens = int(usage.get("prompt_tokens", 0))
                 completion_tokens = int(usage.get("completion_tokens", 0))
-                self._cost_manager.update_cost(prompt_tokens, completion_tokens, self.model)
+                CONFIG.cost_manager.update_cost(prompt_tokens, completion_tokens, self.model)
             except Exception as e:
-                logger.error("zhipuai updats costs failed!", e)
+                logger.error(f"zhipuai updats costs failed! exp: {e}")
 
     def get_choice_text(self, resp: dict) -> str:
         """get the first text of choice from llm response"""

metagpt/repo_parser.py

@@ -336,4 +336,4 @@ def error():
 
 
 if __name__ == "__main__":
-    error()
+    main()

metagpt/roles/architect.py

@@ -8,7 +8,7 @@
 
 from metagpt.actions import WritePRD
 from metagpt.actions.design_api import WriteDesign
-from metagpt.roles import Role
+from metagpt.roles.role import Role
 
 
 class Architect(Role):
@@ -22,17 +22,16 @@ class Architect(Role):
         constraints (str): Constraints or guidelines for the architect.
     """
 
-    def __init__(
-        self,
-        name: str = "Bob",
-        profile: str = "Architect",
-        goal: str = "design a concise, usable, complete software system",
-        constraints: str = "make sure the architecture is simple enough and use appropriate open source libraries."
-        "Use same language as user requirement",
-    ) -> None:
-        """Initializes the Architect with given attributes."""
-        super().__init__(name, profile, goal, constraints)
+    name: str = "Bob"
+    profile: str = "Architect"
+    goal: str = "design a concise, usable, complete software system"
+    constraints: str = (
+        "make sure the architecture is simple enough and use  appropriate open source "
+        "libraries. Use same language as user requirement"
+    )
 
+    def __init__(self, **kwargs) -> None:
+        super().__init__(**kwargs)
         # Initialize actions specific to the Architect role
         self._init_actions([WriteDesign])
 

metagpt/roles/customer_service.py

@@ -5,6 +5,8 @@
 @Author  : alexanderwu
 @File    : sales.py
 """
+from typing import Optional
+
 from metagpt.roles import Sales
 
 # from metagpt.actions import SearchAndSummarize
@@ -24,5 +26,11 @@ DESC = """
 
 
 class CustomerService(Sales):
-    def __init__(self, name="Xiaomei", profile="Human customer service", desc=DESC, store=None):
-        super().__init__(name, profile, desc=desc, store=store)
+    name: str = "Xiaomei"
+    profile: str = "Human customer service"
+    desc: str = DESC
+
+    store: Optional[str] = None
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)

metagpt/roles/engineer.py

@@ -16,6 +16,7 @@
 @Modified By: mashenquan, 2023-12-5. Enhance the workflow to navigate to WriteCode or QaEngineer based on the results
     of SummarizeCode.
 """
+
 from __future__ import annotations
 
 import json
@@ -67,24 +68,25 @@ class Engineer(Role):
         use_code_review (bool): Whether to use code review.
     """
 
-    def __init__(
-        self,
-        name: str = "Alex",
-        profile: str = "Engineer",
-        goal: str = "write elegant, readable, extensible, efficient code",
-        constraints: str = "the code should conform to standards like google-style and be modular and maintainable. "
-        "Use same language as user requirement",
-        n_borg: int = 1,
-        use_code_review: bool = False,
-    ) -> None:
-        """Initializes the Engineer role with given attributes."""
-        super().__init__(name, profile, goal, constraints)
-        self.use_code_review = use_code_review
+    name: str = "Alex"
+    profile: str = "Engineer"
+    goal: str = "write elegant, readable, extensible, efficient code"
+    constraints: str = (
+        "the code should conform to standards like google-style and be modular and maintainable. "
+        "Use same language as user requirement"
+    )
+    n_borg: int = 1
+    use_code_review: bool = False
+    code_todos: list = []
+    summarize_todos = []
+
+    def __init__(self, **kwargs) -> None:
+        super().__init__(**kwargs)
+
         self._init_actions([WriteCode])
         self._watch([WriteTasks, SummarizeCode, WriteCode, WriteCodeReview, FixBug])
         self.code_todos = []
         self.summarize_todos = []
-        self.n_borg = n_borg
         self._next_todo = any_to_name(WriteCode)
 
     @staticmethod
@@ -307,4 +309,5 @@ class Engineer(Role):
 
     @property
     def todo(self) -> str:
+        """AgentStore uses this attribute to display to the user what actions the current role should take."""
         return self._next_todo

metagpt/roles/product_manager.py

@@ -7,10 +7,11 @@
 @Modified By: mashenquan, 2023/11/27. Add `PrepareDocuments` action according to Section 2.2.3.5.1 of RFC 135.
 """
 
+
 from metagpt.actions import UserRequirement, WritePRD
 from metagpt.actions.prepare_documents import PrepareDocuments
 from metagpt.config import CONFIG
-from metagpt.roles import Role
+from metagpt.roles.role import Role
 from metagpt.utils.common import any_to_name
 
 
@@ -25,23 +26,13 @@ class ProductManager(Role):
         constraints (str): Constraints or limitations for the product manager.
     """
 
-    def __init__(
-        self,
-        name: str = "Alice",
-        profile: str = "Product Manager",
-        goal: str = "efficiently create a successful product",
-        constraints: str = "use same language as user requirement",
-    ) -> None:
-        """
-        Initializes the ProductManager role with given attributes.
+    name: str = "Alice"
+    profile: str = "Product Manager"
+    goal: str = "efficiently create a successful product that meets market demands and user expectations"
+    constraints: str = "utilize the same language as the user requirements for seamless communication"
 
-        Args:
-            name (str): Name of the product manager.
-            profile (str): Role profile.
-            goal (str): Goal of the product manager.
-            constraints (str): Constraints or limitations for the product manager.
-        """
-        super().__init__(name, profile, goal, constraints)
+    def __init__(self, **kwargs) -> None:
+        super().__init__(**kwargs)
 
         self._init_actions([PrepareDocuments, WritePRD])
         self._watch([UserRequirement, PrepareDocuments])
@@ -61,4 +52,5 @@ class ProductManager(Role):
 
     @property
     def todo(self) -> str:
+        """AgentStore uses this attribute to display to the user what actions the current role should take."""
         return self._todo

metagpt/roles/project_manager.py

@@ -5,9 +5,10 @@
 @Author  : alexanderwu
 @File    : project_manager.py
 """
+
 from metagpt.actions import WriteTasks
 from metagpt.actions.design_api import WriteDesign
-from metagpt.roles import Role
+from metagpt.roles.role import Role
 
 
 class ProjectManager(Role):
@@ -21,23 +22,16 @@ class ProjectManager(Role):
         constraints (str): Constraints or limitations for the project manager.
     """
 
-    def __init__(
-        self,
-        name: str = "Eve",
-        profile: str = "Project Manager",
-        goal: str = "break down tasks according to PRD/technical design, generate a task list, and analyze task "
-        "dependencies to start with the prerequisite modules",
-        constraints: str = "use same language as user requirement",
-    ) -> None:
-        """
-        Initializes the ProjectManager role with given attributes.
+    name: str = "Eve"
+    profile: str = "Project Manager"
+    goal: str = (
+        "break down tasks according to PRD/technical design, generate a task list, and analyze task "
+        "dependencies to start with the prerequisite modules"
+    )
+    constraints: str = "use same language as user requirement"
+
+    def __init__(self, **kwargs) -> None:
+        super().__init__(**kwargs)
 
-        Args:
-            name (str): Name of the project manager.
-            profile (str): Role profile.
-            goal (str): Goal of the project manager.
-            constraints (str): Constraints or limitations for the project manager.
-        """
-        super().__init__(name, profile, goal, constraints)
         self._init_actions([WriteTasks])
         self._watch([WriteDesign])

metagpt/roles/qa_engineer.py

@@ -14,6 +14,8 @@
 @Modified By: mashenquan, 2023-12-5. Enhance the workflow to navigate to WriteCode or QaEngineer based on the results
     of SummarizeCode.
 """
+
+
 from metagpt.actions import DebugError, RunCode, WriteTest
 from metagpt.actions.summarize_code import SummarizeCode
 from metagpt.config import CONFIG
@@ -30,21 +32,23 @@ from metagpt.utils.file_repository import FileRepository
 
 
 class QaEngineer(Role):
-    def __init__(
-        self,
-        name="Edward",
-        profile="QaEngineer",
-        goal="Write comprehensive and robust tests to ensure codes will work as expected without bugs",
-        constraints="The test code you write should conform to code standard like PEP8, be modular, easy to read and maintain",
-        test_round_allowed=5,
-    ):
-        super().__init__(name, profile, goal, constraints)
-        self._init_actions(
-            [WriteTest]
-        )  # FIXME: a bit hack here, only init one action to circumvent _think() logic, will overwrite _think() in future updates
+    name: str = "Edward"
+    profile: str = "QaEngineer"
+    goal: str = "Write comprehensive and robust tests to ensure codes will work as expected without bugs"
+    constraints: str = (
+        "The test code you write should conform to code standard like PEP8, be modular, " "easy to read and maintain"
+    )
+    test_round_allowed: int = 5
+    test_round: int = 0
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+        # FIXME: a bit hack here, only init one action to circumvent _think() logic,
+        #  will overwrite _think() in future updates
+        self._init_actions([WriteTest])
         self._watch([SummarizeCode, WriteTest, RunCode, DebugError])
         self.test_round = 0
-        self.test_round_allowed = test_round_allowed
 
     async def _write_test(self, message: Message) -> None:
         src_file_repo = CONFIG.git_repo.new_file_repository(CONFIG.src_workspace)
@@ -111,7 +115,8 @@ class QaEngineer(Role):
         )
         run_code_context.code = None
         run_code_context.test_code = None
-        recipient = parse_recipient(result.summary)  # the recipient might be Engineer or myself
+        # the recipient might be Engineer or myself
+        recipient = parse_recipient(result.summary)
         mappings = {"Engineer": "Alex", "QaEngineer": "Edward"}
         self.publish_message(
             Message(
@@ -178,4 +183,4 @@ class QaEngineer(Role):
     async def _observe(self, ignore_memory=False) -> int:
         # This role has events that trigger and execute themselves based on conditions, and cannot rely on the
         # content of memory to activate.
-        return await super(QaEngineer, self)._observe(ignore_memory=True)
+        return await super()._observe(ignore_memory=True)

metagpt/roles/researcher.py

@@ -9,7 +9,7 @@ import asyncio
 
 from pydantic import BaseModel
 
-from metagpt.actions import CollectLinks, ConductResearch, WebBrowseAndSummarize
+from metagpt.actions import Action, CollectLinks, ConductResearch, WebBrowseAndSummarize
 from metagpt.actions.research import get_research_system_text
 from metagpt.const import RESEARCH_PATH
 from metagpt.logs import logger
@@ -62,24 +62,45 @@ class Researcher(Role):
         else:
             topic = msg.content
 
-        research_system_text = get_research_system_text(topic, self.language)
+        research_system_text = self.research_system_text(topic, todo)
         if isinstance(todo, CollectLinks):
             links = await todo.run(topic, 4, 4)
-            ret = Message("", Report(topic=topic, links=links), role=self.profile, cause_by=todo)
+            ret = Message(
+                content="", instruct_content=Report(topic=topic, links=links), role=self.profile, cause_by=todo
+            )
         elif isinstance(todo, WebBrowseAndSummarize):
             links = instruct_content.links
             todos = (todo.run(*url, query=query, system_text=research_system_text) for (query, url) in links.items())
             summaries = await asyncio.gather(*todos)
             summaries = list((url, summary) for i in summaries for (url, summary) in i.items() if summary)
-            ret = Message("", Report(topic=topic, summaries=summaries), role=self.profile, cause_by=todo)
+            ret = Message(
+                content="", instruct_content=Report(topic=topic, summaries=summaries), role=self.profile, cause_by=todo
+            )
         else:
             summaries = instruct_content.summaries
             summary_text = "\n---\n".join(f"url: {url}\nsummary: {summary}" for (url, summary) in summaries)
             content = await self._rc.todo.run(topic, summary_text, system_text=research_system_text)
-            ret = Message("", Report(topic=topic, content=content), role=self.profile, cause_by=self._rc.todo)
+            ret = Message(
+                content="",
+                instruct_content=Report(topic=topic, content=content),
+                role=self.profile,
+                cause_by=self._rc.todo,
+            )
         self._rc.memory.add(ret)
         return ret
 
+    def research_system_text(self, topic, current_task: Action) -> str:
+        """BACKWARD compatible
+        This allows sub-class able to define its own system prompt based on topic.
+        return the previous implementation to have backward compatible
+        Args:
+            topic:
+            language:
+
+        Returns: str
+        """
+        return get_research_system_text(topic, self.language)
+
     async def react(self) -> Message:
         msg = await super().react()
         report = msg.instruct_content

metagpt/roles/role.py

@@ -19,20 +19,33 @@
 @Modified By: mashenquan, 2023-11-4. According to the routing feature plan in Chapter 2.2.3.2 of RFC 113, the routing
     functionality is to be consolidated into the `Environment` class.
 """
+
 from __future__ import annotations
 
 from enum import Enum
-from typing import Iterable, Set, Type
+from pathlib import Path
+from typing import Any, Iterable, Set, Type
 
 from pydantic import BaseModel, Field
 
-from metagpt.actions import Action, ActionOutput, UserRequirement
+from metagpt.actions import Action, ActionOutput
+from metagpt.actions.action import action_subclass_registry
 from metagpt.actions.action_node import ActionNode
+from metagpt.actions.add_requirement import UserRequirement
+from metagpt.const import SERDESER_PATH
 from metagpt.llm import LLM, HumanProvider
 from metagpt.logs import logger
 from metagpt.memory import Memory
+from metagpt.provider.base_gpt_api import BaseGPTAPI
 from metagpt.schema import Message, MessageQueue
-from metagpt.utils.common import any_to_name, any_to_str
+from metagpt.utils.common import (
+    any_to_name,
+    any_to_str,
+    import_class,
+    read_json_file,
+    role_raise_decorator,
+    write_json_file,
+)
 from metagpt.utils.repair_llm_raw_output import extract_state_value_from_output
 
 PREFIX_TEMPLATE = """You are a {profile}, named {name}, your goal is {goal}, and the constraint is {constraints}. """
@@ -75,34 +88,21 @@ class RoleReactMode(str, Enum):
         return [item.value for item in cls]
 
 
-class RoleSetting(BaseModel):
-    """Role properties"""
-
-    name: str
-    profile: str
-    goal: str
-    constraints: str
-    desc: str
-    is_human: bool
-
-    def __str__(self):
-        return f"{self.name}({self.profile})"
-
-    def __repr__(self):
-        return self.__str__()
-
-
 class RoleContext(BaseModel):
     """Role Runtime Context"""
 
-    env: "Environment" = Field(default=None)
-    msg_buffer: MessageQueue = Field(default_factory=MessageQueue)  # Message Buffer with Asynchronous Updates
+    # # env exclude=True to avoid `RecursionError: maximum recursion depth exceeded in comparison`
+    env: "Environment" = Field(default=None, exclude=True)
+    # TODO judge if ser&deser
+    msg_buffer: MessageQueue = Field(
+        default_factory=MessageQueue, exclude=True
+    )  # Message Buffer with Asynchronous Updates
     memory: Memory = Field(default_factory=Memory)
     # long_term_memory: LongTermMemory = Field(default_factory=LongTermMemory)
     state: int = Field(default=-1)  # -1 indicates initial or termination state where todo is None
-    todo: Action = Field(default=None)
+    todo: Action = Field(default=None, exclude=True)
     watch: set[str] = Field(default_factory=set)
-    news: list[Type[Message]] = Field(default=[])
+    news: list[Type[Message]] = Field(default=[], exclude=True)  # TODO not used
     react_mode: RoleReactMode = (
         RoleReactMode.REACT
     )  # see `Role._set_react_mode` for definitions of the following two attributes
@@ -127,35 +127,154 @@ class RoleContext(BaseModel):
         return self.memory.get()
 
 
-class Role:
+role_subclass_registry = {}
+
+
+class Role(BaseModel):
     """Role/Agent"""
 
-    def __init__(self, name="", profile="", goal="", constraints="", desc="", is_human=False):
-        self._llm = LLM() if not is_human else HumanProvider()
-        self._setting = RoleSetting(
-            name=name, profile=profile, goal=goal, constraints=constraints, desc=desc, is_human=is_human
-        )
+    name: str = ""
+    profile: str = ""
+    goal: str = ""
+    constraints: str = ""
+    desc: str = ""
+    is_human: bool = False
+
+    _llm: BaseGPTAPI = Field(default_factory=LLM)
+    _role_id: str = ""
+    _states: list[str] = []
+    _actions: list[Action] = []
+    _rc: RoleContext = Field(default_factory=RoleContext)
+    _subscription: tuple[str] = set()
+
+    # builtin variables
+    recovered: bool = False  # to tag if a recovered role
+    latest_observed_msg: Message = None  # record the latest observed message when interrupted
+    builtin_class_name: str = ""
+
+    _private_attributes = {
+        "_llm": LLM() if not is_human else HumanProvider(),
+        "_role_id": _role_id,
+        "_states": [],
+        "_actions": [],
+        "_rc": RoleContext(),
+        "_subscription": set(),
+    }
+
+    __hash__ = object.__hash__  # support Role as hashable type in `Environment.members`
+
+    class Config:
+        arbitrary_types_allowed = True
+        exclude = ["_llm"]
+
+    def __init__(self, **kwargs: Any):
+        for index in range(len(kwargs.get("_actions", []))):
+            current_action = kwargs["_actions"][index]
+            if isinstance(current_action, dict):
+                item_class_name = current_action.get("builtin_class_name", None)
+                for name, subclass in action_subclass_registry.items():
+                    registery_class_name = subclass.__fields__["builtin_class_name"].default
+                    if item_class_name == registery_class_name:
+                        current_action = subclass(**current_action)
+                        break
+                kwargs["_actions"][index] = current_action
+
+        super().__init__(**kwargs)
+
+        # 关于私有变量的初始化  https://github.com/pydantic/pydantic/issues/655
+        self._private_attributes["_llm"] = LLM() if not self.is_human else HumanProvider()
+        self._private_attributes["_role_id"] = str(self._setting)
+        self._private_attributes["_subscription"] = {any_to_str(self), self.name} if self.name else {any_to_str(self)}
+
+        for key in self._private_attributes.keys():
+            if key in kwargs:
+                object.__setattr__(self, key, kwargs[key])
+                if key == "_rc":
+                    _rc = RoleContext(**kwargs["_rc"])
+                    object.__setattr__(self, "_rc", _rc)
+            else:
+                if key == "_rc":
+                    # # Warning, if use self._private_attributes["_rc"],
+                    # # self._rc will be a shared object between roles, so init one or reset it inside `_reset`
+                    object.__setattr__(self, key, RoleContext())
+                else:
+                    object.__setattr__(self, key, self._private_attributes[key])
+
         self._llm.system_prompt = self._get_prefix()
-        self._states = []
-        self._actions = []
-        self._role_id = str(self._setting)
-        self._rc = RoleContext(watch={any_to_str(UserRequirement)})
-        self._subscription = {any_to_str(self), name} if name else {any_to_str(self)}
+
+        # deserialize child classes dynamically for inherited `role`
+        object.__setattr__(self, "builtin_class_name", self.__class__.__name__)
+        self.__fields__["builtin_class_name"].default = self.__class__.__name__
+
+        if "actions" in kwargs:
+            self._init_actions(kwargs["actions"])
+
+        self._watch(kwargs.get("watch") or [UserRequirement])
+
+    def __init_subclass__(cls, **kwargs: Any) -> None:
+        super().__init_subclass__(**kwargs)
+        role_subclass_registry[cls.__name__] = cls
 
     def _reset(self):
-        self._states = []
-        self._actions = []
+        object.__setattr__(self, "_states", [])
+        object.__setattr__(self, "_actions", [])
+
+    @property
+    def _setting(self):
+        return f"{self.name}({self.profile})"
+
+    def serialize(self, stg_path: Path = None):
+        stg_path = (
+            SERDESER_PATH.joinpath(f"team/environment/roles/{self.__class__.__name__}_{self.name}")
+            if stg_path is None
+            else stg_path
+        )
+
+        role_info = self.dict(exclude={"_rc": {"memory": True, "msg_buffer": True}, "_llm": True})
+        role_info.update({"role_class": self.__class__.__name__, "module_name": self.__module__})
+        role_info_path = stg_path.joinpath("role_info.json")
+        write_json_file(role_info_path, role_info)
+
+        self._rc.memory.serialize(stg_path)  # serialize role's memory alone
+
+    @classmethod
+    def deserialize(cls, stg_path: Path) -> "Role":
+        """stg_path = ./storage/team/environment/roles/{role_class}_{role_name}"""
+        role_info_path = stg_path.joinpath("role_info.json")
+        role_info = read_json_file(role_info_path)
+
+        role_class_str = role_info.pop("role_class")
+        module_name = role_info.pop("module_name")
+        role_class = import_class(class_name=role_class_str, module_name=module_name)
+
+        role = role_class(**role_info)  # initiate particular Role
+        role.set_recovered(True)  # set True to make a tag
+
+        role_memory = Memory.deserialize(stg_path)
+        role.set_memory(role_memory)
+
+        return role
 
     def _init_action_system_message(self, action: Action):
-        action.set_prefix(self._get_prefix(), self.profile)
+        action.set_prefix(self._get_prefix())
+
+    def set_recovered(self, recovered: bool = False):
+        self.recovered = recovered
+
+    def set_memory(self, memory: Memory):
+        self._rc.memory = memory
+
+    def init_actions(self, actions):
+        self._init_actions(actions)
 
     def _init_actions(self, actions):
         self._reset()
         for idx, action in enumerate(actions):
             if not isinstance(action, Action):
-                i = action("", llm=self._llm)
+                ## 默认初始化
+                i = action(name="", llm=self._llm)
             else:
-                if self._setting.is_human and not isinstance(action.llm, HumanProvider):
+                if self.is_human and not isinstance(action.llm, HumanProvider):
                     logger.warning(
                         f"is_human attribute does not take effect, "
                         f"as Role's {str(action)} was initialized using LLM, "
@@ -211,7 +330,7 @@ class Role:
     def _set_state(self, state: int):
         """Update the current state."""
         self._rc.state = state
-        logger.debug(self._actions)
+        logger.debug(f"actions={self._actions}, state={state}")
         self._rc.todo = self._actions[self._rc.state] if state >= 0 else None
 
     def set_env(self, env: "Environment"):
@@ -221,35 +340,10 @@ class Role:
         if env:
             env.set_subscription(self, self._subscription)
 
-    @property
-    def profile(self):
-        """Get the role description (position)"""
-        return self._setting.profile
-
-    @property
-    def name(self):
-        """Get virtual user name"""
-        return self._setting.name
-
     @property
     def subscription(self) -> Set:
         """The labels for messages to be consumed by the Role object."""
-        return self._subscription
-
-    @property
-    def desc(self):
-        """Return role `desc`, read only"""
-        return self._setting.desc
-
-    @property
-    def goal(self):
-        """Return role `goal`, read only"""
-        return self._setting.goal
-
-    @property
-    def constraints(self):
-        """Return role `constraints`, read only"""
-        return self._setting.constraints
+        return set(self._subscription)
 
     @property
     def action_count(self):
@@ -258,16 +352,25 @@ class Role:
 
     def _get_prefix(self):
         """Get the role prefix"""
-        if self._setting.desc:
-            return self._setting.desc
-        return PREFIX_TEMPLATE.format(**self._setting.dict())
+        if self.desc:
+            return self.desc
+        return PREFIX_TEMPLATE.format(
+            **{"profile": self.profile, "name": self.name, "goal": self.goal, "constraints": self.constraints}
+        )
 
     async def _think(self) -> bool:
         """Consider what to do and decide on the next course of action. Return false if nothing can be done."""
         if len(self._actions) == 1:
             # If there is only one action, then only this one can be performed
             self._set_state(0)
+
             return True
+
+        if self.recovered and self._rc.state >= 0:
+            self._set_state(self._rc.state)  # action to run from recovered state
+            self.recovered = False  # avoid max_react_loop out of work
+            return True
+
         prompt = self._get_prefix()
         prompt += STATE_TEMPLATE.format(
             history=self._rc.history,
@@ -275,10 +378,11 @@ class Role:
             n_states=len(self._states) - 1,
             previous_state=self._rc.state,
         )
-        # print(prompt)
+
         next_state = await self._llm.aask(prompt)
         next_state = extract_state_value_from_output(next_state)
         logger.debug(f"{prompt=}")
+
         if (not next_state.isdigit() and next_state != "-1") or int(next_state) not in range(-1, len(self._states)):
             logger.warning(f"Invalid answer of state, {next_state=}, will be set to -1")
             next_state = -1
@@ -292,7 +396,7 @@ class Role:
     async def _act(self) -> Message:
         logger.info(f"{self._setting}: ready to {self._rc.todo}")
         response = await self._rc.todo.run(self._rc.important_memory)
-        if isinstance(response, ActionOutput) or isinstance(response, ActionNode):
+        if isinstance(response, (ActionOutput, ActionNode)):
             msg = Message(
                 content=response.content,
                 instruct_content=response.instruct_content,
@@ -308,15 +412,30 @@ class Role:
 
         return msg
 
+    def _find_news(self, observed: list[Message], existed: list[Message]) -> list[Message]:
+        news = []
+        # Warning, remove `id` here to make it work for recover
+        observed_pure = [msg.dict(exclude={"id": True}) for msg in observed]
+        existed_pure = [msg.dict(exclude={"id": True}) for msg in existed]
+        for idx, new in enumerate(observed_pure):
+            if new["cause_by"] in self._rc.watch and new not in existed_pure:
+                news.append(observed[idx])
+        return news
+
     async def _observe(self, ignore_memory=False) -> int:
         """Prepare new messages for processing from the message buffer and other sources."""
         # Read unprocessed messages from the msg buffer.
         news = self._rc.msg_buffer.pop_all()
+        if self.recovered:
+            news = [self.latest_observed_msg] if self.latest_observed_msg else []
+        else:
+            self.latest_observed_msg = news[-1] if len(news) > 0 else None  # record the latest observed msg
+
         # Store the read messages in your own memory to prevent duplicate processing.
         old_messages = [] if ignore_memory else self._rc.memory.get()
         self._rc.memory.add_batch(news)
         # Filter out messages of interest.
-        self._rc.news = [n for n in news if n.cause_by in self._rc.watch and n not in old_messages]
+        self._rc.news = self._find_news(news, old_messages)
 
         # Design Rules:
         # If you need to further categorize Message objects, you can do so using the Message.set_meta function.
@@ -347,7 +466,7 @@ class Role:
         Use llm to select actions in _think dynamically
         """
         actions_taken = 0
-        rsp = Message("No actions taken yet")  # will be overwritten after Role _act
+        rsp = Message(content="No actions taken yet")  # will be overwritten after Role _act
         while actions_taken < self._rc.max_react_loop:
             # think
             await self._think()
@@ -361,7 +480,8 @@ class Role:
 
     async def _act_by_order(self) -> Message:
         """switch action each time by order defined in _init_actions, i.e. _act (Action1) -> _act (Action2) -> ..."""
-        for i in range(len(self._states)):
+        start_idx = self._rc.state if self._rc.state >= 0 else 0  # action to run from recovered state
+        for i in range(start_idx, len(self._states)):
             self._set_state(i)
             rsp = await self._act()
         return rsp  # return output from the last action
@@ -369,7 +489,7 @@ class Role:
     async def _plan_and_act(self) -> Message:
         """first plan, then execute an action sequence, i.e. _think (of a plan) -> _act -> _act -> ... Use llm to come up with the plan dynamically."""
         # TODO: to be implemented
-        return Message("")
+        return Message(content="")
 
     async def react(self) -> Message:
         """Entry to one of three strategies by which Role reacts to the observed Message"""
@@ -386,16 +506,17 @@ class Role:
         """A wrapper to return the most recent k memories of this role, return all when k=0"""
         return self._rc.memory.get(k=k)
 
+    @role_raise_decorator
     async def run(self, with_message=None):
         """Observe, and think and act based on the results of the observation"""
         if with_message:
             msg = None
             if isinstance(with_message, str):
-                msg = Message(with_message)
+                msg = Message(content=with_message)
             elif isinstance(with_message, Message):
                 msg = with_message
             elif isinstance(with_message, list):
-                msg = Message("\n".join(with_message))
+                msg = Message(content="\n".join(with_message))
             if not msg.cause_by:
                 msg.cause_by = UserRequirement
             self.put_message(msg)
@@ -430,6 +551,7 @@ class Role:
 
     @property
     def todo(self) -> str:
+        """AgentStore uses this attribute to display to the user what actions the current role should take."""
         if self._actions:
             return any_to_name(self._actions[0])
         return ""

metagpt/roles/sales.py

@@ -5,30 +5,33 @@
 @Author  : alexanderwu
 @File    : sales.py
 """
+
+from typing import Optional
+
 from metagpt.actions import SearchAndSummarize
 from metagpt.roles import Role
 from metagpt.tools import SearchEngineType
 
 
 class Sales(Role):
-    def __init__(
-        self,
-        name="Xiaomei",
-        profile="Retail sales guide",
-        desc="I am a sales guide in retail. My name is Xiaomei. I will answer some customer questions next, and I "
-        "will answer questions only based on the information in the knowledge base."
-        "If I feel that you can't get the answer from the reference material, then I will directly reply that"
-        " I don't know, and I won't tell you that this is from the knowledge base,"
-        "but pretend to be what I know. Note that each of my replies will be replied in the tone of a "
-        "professional guide",
-        store=None,
-    ):
-        super().__init__(name, profile, desc=desc)
-        self._set_store(store)
+    name: str = "Xiaomei"
+    profile: str = "Retail sales guide"
+    desc: str = "I am a sales guide in retail. My name is Xiaomei. I will answer some customer questions next, and I "
+    "will answer questions only based on the information in the knowledge base."
+    "If I feel that you can't get the answer from the reference material, then I will directly reply that"
+    " I don't know, and I won't tell you that this is from the knowledge base,"
+    "but pretend to be what I know. Note that each of my replies will be replied in the tone of a "
+    "professional guide"
+
+    store: Optional[str] = None
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self._set_store(self.store)
 
     def _set_store(self, store):
         if store:
-            action = SearchAndSummarize("", engine=SearchEngineType.CUSTOM_ENGINE, search_func=store.asearch)
+            action = SearchAndSummarize(name="", engine=SearchEngineType.CUSTOM_ENGINE, search_func=store.asearch)
         else:
             action = SearchAndSummarize()
         self._init_actions([action])

metagpt/roles/searcher.py

@@ -7,6 +7,9 @@
 @Modified By: mashenquan, 2023-11-1. According to Chapter 2.2.1 and 2.2.2 of RFC 116, change the data type of
         the `cause_by` value in the `Message` to a string to support the new message distribution feature.
 """
+
+from pydantic import Field
+
 from metagpt.actions import ActionOutput, SearchAndSummarize
 from metagpt.actions.action_node import ActionNode
 from metagpt.logs import logger
@@ -27,15 +30,13 @@ class Searcher(Role):
         engine (SearchEngineType): The type of search engine to use.
     """
 
-    def __init__(
-        self,
-        name: str = "Alice",
-        profile: str = "Smart Assistant",
-        goal: str = "Provide search services for users",
-        constraints: str = "Answer is rich and complete",
-        engine=SearchEngineType.SERPAPI_GOOGLE,
-        **kwargs,
-    ) -> None:
+    name: str = Field(default="Alice")
+    profile: str = Field(default="Smart Assistant")
+    goal: str = "Provide search services for users"
+    constraints: str = "Answer is rich and complete"
+    engine: SearchEngineType = SearchEngineType.SERPAPI_GOOGLE
+
+    def __init__(self, **kwargs) -> None:
         """
         Initializes the Searcher role with given attributes.
 
@@ -46,12 +47,12 @@ class Searcher(Role):
             constraints (str): Constraints or limitations for the searcher.
             engine (SearchEngineType): The type of search engine to use.
         """
-        super().__init__(name, profile, goal, constraints, **kwargs)
-        self._init_actions([SearchAndSummarize(engine=engine)])
+        super().__init__(**kwargs)
+        self._init_actions([SearchAndSummarize(engine=self.engine)])
 
     def set_search_func(self, search_func):
         """Sets a custom search function for the searcher."""
-        action = SearchAndSummarize("", engine=SearchEngineType.CUSTOM_ENGINE, search_func=search_func)
+        action = SearchAndSummarize(name="", engine=SearchEngineType.CUSTOM_ENGINE, search_func=search_func)
         self._init_actions([action])
 
     async def _act_sp(self) -> Message:
@@ -59,7 +60,7 @@ class Searcher(Role):
         logger.info(f"{self._setting}: ready to {self._rc.todo}")
         response = await self._rc.todo.run(self._rc.memory.get(k=0))
 
-        if isinstance(response, ActionOutput) or isinstance(response, ActionNode):
+        if isinstance(response, (ActionOutput, ActionNode)):
             msg = Message(
                 content=response.content,
                 instruct_content=response.instruct_content,

metagpt/roles/tutorial_assistant.py

@@ -42,17 +42,7 @@ class TutorialAssistant(Role):
         self.main_title = ""
         self.total_content = ""
         self.language = language
-
-    async def _think(self) -> None:
-        """Determine the next action to be taken by the role."""
-        if self._rc.todo is None:
-            self._set_state(0)
-            return
-
-        if self._rc.state + 1 < len(self._states):
-            self._set_state(self._rc.state + 1)
-        else:
-            self._rc.todo = None
+        self._set_react_mode(react_mode="by_order")
 
     async def _handle_directory(self, titles: Dict) -> Message:
         """Handle the directories for the tutorial document.
@@ -75,8 +65,6 @@ class TutorialAssistant(Role):
             for second_dir in first_dir[key]:
                 directory += f"  - {second_dir}\n"
         self._init_actions(actions)
-        self._rc.todo = None
-        return Message(content=directory)
 
     async def _act(self) -> Message:
         """Perform an action as determined by the role.
@@ -90,7 +78,8 @@ class TutorialAssistant(Role):
             self.topic = msg.content
             resp = await todo.run(topic=self.topic)
             logger.info(resp)
-            return await self._handle_directory(resp)
+            await self._handle_directory(resp)
+            return await super().react()
         resp = await todo.run(topic=self.topic)
         logger.info(resp)
         if self.total_content != "":
@@ -98,17 +87,8 @@ class TutorialAssistant(Role):
         self.total_content += resp
         return Message(content=resp, role=self.profile)
 
-    async def _react(self) -> Message:
-        """Execute the assistant's think and actions.
-
-        Returns:
-            A message containing the final result of the assistant's actions.
-        """
-        while True:
-            await self._think()
-            if self._rc.todo is None:
-                break
-            msg = await self._act()
+    async def react(self) -> Message:
+        msg = await super().react()
         root_path = TUTORIAL_PATH / datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
         await File.write(root_path, f"{self.main_title}.md", self.total_content.encode("utf-8"))
         return msg

metagpt/schema.py

@@ -12,16 +12,18 @@
         between actions.
         3. Add `id` to `Message` according to Section 2.2.3.1.1 of RFC 135.
 """
+
 from __future__ import annotations
 
 import asyncio
 import json
 import os.path
 import uuid
+from abc import ABC
 from asyncio import Queue, QueueEmpty, wait_for
 from json import JSONDecodeError
 from pathlib import Path
-from typing import Dict, List, Optional, Set, Type, TypedDict, TypeVar
+from typing import Any, Dict, List, Optional, Set, Type, TypedDict, TypeVar
 
 from pydantic import BaseModel, Field
 
@@ -35,8 +37,13 @@ from metagpt.const import (
     TASK_FILE_REPO,
 )
 from metagpt.logs import logger
-from metagpt.utils.common import any_to_str, any_to_str_set
+from metagpt.utils.common import any_to_str, any_to_str_set, import_class
 from metagpt.utils.exceptions import handle_exception
+from metagpt.utils.serialize import (
+    actionoutout_schema_to_mapping,
+    actionoutput_mapping_to_str,
+    actionoutput_str_to_mapping,
+)
 
 
 class RawMessage(TypedDict):
@@ -97,41 +104,31 @@ class Message(BaseModel):
 
     id: str  # According to Section 2.2.3.1.1 of RFC 135
     content: str
-    instruct_content: BaseModel = Field(default=None)
+    instruct_content: BaseModel = None
     role: str = "user"  # system / user / assistant
     cause_by: str = ""
     sent_from: str = ""
     send_to: Set = Field(default_factory={MESSAGE_ROUTE_TO_ALL})
 
-    def __init__(
-        self,
-        content,
-        instruct_content=None,
-        role="user",
-        cause_by="",
-        sent_from="",
-        send_to=MESSAGE_ROUTE_TO_ALL,
-        **kwargs,
-    ):
-        """
-        Parameters not listed below will be stored as meta info, including custom parameters.
-        :param content: Message content.
-        :param instruct_content: Message content struct.
-        :param cause_by: Message producer
-        :param sent_from: Message route info tells who sent this message.
-        :param send_to: Specifies the target recipient or consumer for message delivery in the environment.
-        :param role: Message meta info tells who sent this message.
-        """
-        super().__init__(
-            id=uuid.uuid4().hex,
-            content=content,
-            instruct_content=instruct_content,
-            role=role,
-            cause_by=any_to_str(cause_by),
-            sent_from=any_to_str(sent_from),
-            send_to=any_to_str_set(send_to),
-            **kwargs,
+    def __init__(self, content: str = "", **kwargs):
+        ic = kwargs.get("instruct_content", None)
+        if ic and not isinstance(ic, BaseModel) and "class" in ic:
+            # compatible with custom-defined ActionOutput
+            mapping = actionoutput_str_to_mapping(ic["mapping"])
+
+            actionnode_class = import_class("ActionNode", "metagpt.actions.action_node")  # avoid circular import
+            ic_obj = actionnode_class.create_model_class(class_name=ic["class"], mapping=mapping)
+            ic_new = ic_obj(**ic["value"])
+            kwargs["instruct_content"] = ic_new
+
+        kwargs["id"] = kwargs.get("id", uuid.uuid4().hex)
+        kwargs["content"] = kwargs.get("content", content)
+        kwargs["cause_by"] = any_to_str(
+            kwargs.get("cause_by", import_class("UserRequirement", "metagpt.actions.add_requirement"))
         )
+        kwargs["sent_from"] = any_to_str(kwargs.get("sent_from", ""))
+        kwargs["send_to"] = any_to_str_set(kwargs.get("send_to", {MESSAGE_ROUTE_TO_ALL}))
+        super(Message, self).__init__(**kwargs)
 
     def __setattr__(self, key, val):
         """Override `@property.setter`, convert non-string parameters into string parameters."""
@@ -145,8 +142,26 @@ class Message(BaseModel):
             new_val = val
         super().__setattr__(key, new_val)
 
+    def dict(self, *args, **kwargs) -> "DictStrAny":
+        """overwrite the `dict` to dump dynamic pydantic model"""
+        obj_dict = super(Message, self).dict(*args, **kwargs)
+        ic = self.instruct_content
+        if ic:
+            # compatible with custom-defined ActionOutput
+            schema = ic.schema()
+            # `Documents` contain definitions
+            if "definitions" not in schema:
+                # TODO refine with nested BaseModel
+                mapping = actionoutout_schema_to_mapping(schema)
+                mapping = actionoutput_mapping_to_str(mapping)
+
+                obj_dict["instruct_content"] = {"class": schema["title"], "mapping": mapping, "value": ic.dict()}
+        return obj_dict
+
     def __str__(self):
         # prefix = '-'.join([self.role, str(self.cause_by)])
+        if self.instruct_content:
+            return f"{self.role}: {self.instruct_content.dict()}"
         return f"{self.role}: {self.content}"
 
     def __repr__(self):
@@ -164,6 +179,7 @@ class Message(BaseModel):
     @handle_exception(exception_type=JSONDecodeError, default_return=None)
     def load(val):
         """Convert the json string to object."""
+
         try:
             m = json.loads(val)
             id = m.get("id")
@@ -205,11 +221,22 @@ class AIMessage(Message):
         super().__init__(content=content, role="assistant")
 
 
-class MessageQueue:
+class MessageQueue(BaseModel):
     """Message queue which supports asynchronous updates."""
 
-    def __init__(self):
-        self._queue = Queue()
+    _queue: Queue = Field(default_factory=Queue)
+
+    _private_attributes = {"_queue": Queue()}
+
+    class Config:
+        arbitrary_types_allowed = True
+
+    def __init__(self, **kwargs: Any):
+        for key in self._private_attributes.keys():
+            if key in kwargs:
+                object.__setattr__(self, key, kwargs[key])
+            else:
+                object.__setattr__(self, key, Queue())
 
     def pop(self) -> Message | None:
         """Pop one message from the queue."""
@@ -257,16 +284,16 @@ class MessageQueue:
         return json.dumps(lst)
 
     @staticmethod
-    def load(i) -> "MessageQueue":
+    def load(data) -> "MessageQueue":
         """Convert the json string to the `MessageQueue` object."""
         queue = MessageQueue()
         try:
-            lst = json.loads(i)
+            lst = json.loads(data)
             for i in lst:
                 msg = Message(**i)
                 queue.push(msg)
         except JSONDecodeError as e:
-            logger.warning(f"JSON load failed: {i}, error:{e}")
+            logger.warning(f"JSON load failed: {data}, error:{e}")
 
         return queue
 
@@ -275,7 +302,7 @@ class MessageQueue:
 T = TypeVar("T", bound="BaseModel")
 
 
-class BaseContext(BaseModel):
+class BaseContext(BaseModel, ABC):
     @classmethod
     @handle_exception
     def loads(cls: Type[T], val: str) -> Optional[T]:

metagpt/startup.py

@@ -7,7 +7,7 @@ import typer
 
 from metagpt.config import CONFIG
 
-app = typer.Typer()
+app = typer.Typer(add_completion=False)
 
 
 @app.command()
@@ -22,13 +22,17 @@ def startup(
     inc: bool = typer.Option(default=False, help="Incremental mode. Use it to coop with existing repo."),
     project_path: str = typer.Option(
         default="",
-        help="Specify the directory path of the old version project to fulfill the " "incremental requirements.",
+        help="Specify the directory path of the old version project to fulfill the incremental requirements.",
+    ),
+    reqa_file: str = typer.Option(
+        default="", help="Specify the source file name for rewriting the quality assurance code."
     ),
-    reqa_file: str = typer.Option(default="", help="Specify the source file name for rewriting the quality test code."),
     max_auto_summarize_code: int = typer.Option(
         default=0,
-        help="The maximum number of times the 'SummarizeCode' action is automatically invoked, with -1 indicating unlimited. This parameter is used for debugging the workflow.",
+        help="The maximum number of times the 'SummarizeCode' action is automatically invoked, with -1 indicating "
+        "unlimited. This parameter is used for debugging the workflow.",
     ),
+    recover_path: str = typer.Option(default=None, help="recover the project from existing serialized storage"),
 ):
     """Run a startup. Be a boss."""
     from metagpt.roles import (
@@ -40,30 +44,31 @@ def startup(
     )
     from metagpt.team import Team
 
-    # Use in the PrepareDocuments action according to Section 2.2.3.5.1 of RFC 135.
-    CONFIG.project_path = project_path
-    if project_path:
-        inc = True
-        project_name = project_name or Path(project_path).name
-    CONFIG.project_name = project_name
-    CONFIG.inc = inc
-    CONFIG.reqa_file = reqa_file
-    CONFIG.max_auto_summarize_code = max_auto_summarize_code
+    CONFIG.update_via_cli(project_path, project_name, inc, reqa_file, max_auto_summarize_code)
 
-    company = Team()
-    company.hire(
-        [
-            ProductManager(),
-            Architect(),
-            ProjectManager(),
-        ]
-    )
+    if not recover_path:
+        company = Team()
+        company.hire(
+            [
+                ProductManager(),
+                Architect(),
+                ProjectManager(),
+            ]
+        )
 
-    if implement or code_review:
-        company.hire([Engineer(n_borg=5, use_code_review=code_review)])
+        if implement or code_review:
+            company.hire([Engineer(n_borg=5, use_code_review=code_review)])
 
-    if run_tests:
-        company.hire([QaEngineer()])
+        if run_tests:
+            company.hire([QaEngineer()])
+    else:
+        # # stg_path = SERDESER_PATH.joinpath("team")
+        stg_path = Path(recover_path)
+        if not stg_path.exists() or not str(stg_path).endswith("team"):
+            raise FileNotFoundError(f"{recover_path} not exists or not endswith `team`")
+
+        company = Team.deserialize(stg_path=stg_path)
+        idea = company.idea  # use original idea
 
     company.invest(investment)
     company.run_project(idea)

metagpt/subscription.py

@@ -19,7 +19,7 @@ class SubscriptionRunner(BaseModel):
 
         >>> async def trigger():
         ...     while True:
-        ...         yield Message("the latest news about OpenAI")
+        ...         yield Message(content="the latest news about OpenAI")
         ...         await asyncio.sleep(3600 * 24)
 
         >>> async def callback(msg: Message):

metagpt/team.py

@@ -7,24 +7,31 @@
 @Modified By: mashenquan, 2023/11/27. Add an archiving operation after completing the project, as specified in
         Section 2.2.3.3 of RFC 135.
 """
+
 import warnings
+from pathlib import Path
 
 from pydantic import BaseModel, Field
 
 from metagpt.actions import UserRequirement
 from metagpt.config import CONFIG
-from metagpt.const import MESSAGE_ROUTE_TO_ALL
+from metagpt.const import MESSAGE_ROUTE_TO_ALL, SERDESER_PATH
 from metagpt.environment import Environment
 from metagpt.logs import logger
 from metagpt.roles import Role
 from metagpt.schema import Message
-from metagpt.utils.common import NoMoneyException
+from metagpt.utils.common import (
+    NoMoneyException,
+    read_json_file,
+    serialize_decorator,
+    write_json_file,
+)
 
 
 class Team(BaseModel):
     """
-    Team: Possesses one or more roles (agents), SOP (Standard Operating Procedures), and a platform for instant messaging,
-    dedicated to perform any multi-agent activity, such as collaboratively writing executable code.
+    Team: Possesses one or more roles (agents), SOP (Standard Operating Procedures), and a env for instant messaging,
+    dedicated to env any multi-agent activity, such as collaboratively writing executable code.
     """
 
     env: Environment = Field(default_factory=Environment)
@@ -34,6 +41,38 @@ class Team(BaseModel):
     class Config:
         arbitrary_types_allowed = True
 
+    def serialize(self, stg_path: Path = None):
+        stg_path = SERDESER_PATH.joinpath("team") if stg_path is None else stg_path
+
+        team_info_path = stg_path.joinpath("team_info.json")
+        write_json_file(team_info_path, self.dict(exclude={"env": True}))
+
+        self.env.serialize(stg_path.joinpath("environment"))  # save environment alone
+
+    @classmethod
+    def recover(cls, stg_path: Path) -> "Team":
+        return cls.deserialize(stg_path)
+
+    @classmethod
+    def deserialize(cls, stg_path: Path) -> "Team":
+        """stg_path = ./storage/team"""
+        # recover team_info
+        team_info_path = stg_path.joinpath("team_info.json")
+        if not team_info_path.exists():
+            raise FileNotFoundError(
+                "recover storage meta file `team_info.json` not exist, "
+                "not to recover and please start a new project."
+            )
+
+        team_info: dict = read_json_file(team_info_path)
+
+        # recover environment
+        environment = Environment.deserialize(stg_path=stg_path.joinpath("environment"))
+        team_info.update({"env": environment})
+
+        team = Team(**team_info)
+        return team
+
     def hire(self, roles: list[Role]):
         """Hire roles to cooperate"""
         self.env.add_roles(roles)
@@ -77,6 +116,7 @@ class Team(BaseModel):
     def _save(self):
         logger.info(self.json(ensure_ascii=False))
 
+    @serialize_decorator
     async def run(self, n_round=3, auto_archive=True):
         """Run company until target round or no money"""
         while n_round > 0:
@@ -84,7 +124,7 @@ class Team(BaseModel):
             n_round -= 1
             logger.debug(f"max {n_round=} left.")
             self._check_balance()
+
             await self.env.run()
-        if auto_archive and CONFIG.git_repo:
-            CONFIG.git_repo.archive()
+        self.env.archive(auto_archive)
         return self.env.history

metagpt/tools/azure_tts.py

@@ -1,10 +1,10 @@
 #!/usr/bin/env python
 # -*- coding: utf-8 -*-
 """
-@Time    : 2023/8/17
-@Author  : mashenquan
+@Time    : 2023/6/9 22:22
+@Author  : Leo Xiao
 @File    : azure_tts.py
-@Desc    : azure TTS OAS3 api, which provides text-to-speech functionality
+@Modified by: mashenquan, 2023/8/17. Azure TTS OAS3 api, which provides text-to-speech functionality
 """
 import asyncio
 import base64

metagpt/tools/code_interpreter.py

@@ -46,7 +46,6 @@ class OpenCodeInterpreter(object):
         interpreter.auto_run = auto_run
         interpreter.model = CONFIG.openai_api_model or "gpt-3.5-turbo"
         interpreter.api_key = CONFIG.openai_api_key
-        # interpreter.api_base = CONFIG.openai_api_base
         self.interpreter = interpreter
 
     def chat(self, query: str, reset: bool = True):

metagpt/tools/moderation.py

@@ -5,6 +5,7 @@
 @Author  : zhanglei
 @File    : moderation.py
 """
+import asyncio
 from typing import Union
 
 from metagpt.llm import LLM
@@ -14,16 +15,6 @@ class Moderation:
     def __init__(self):
         self.llm = LLM()
 
-    def moderation(self, content: Union[str, list[str]]):
-        resp = []
-        if content:
-            moderation_results = self.llm.moderation(content=content)
-            results = moderation_results.results
-            for item in results:
-                resp.append(item.flagged)
-
-        return resp
-
     async def amoderation(self, content: Union[str, list[str]]):
         resp = []
         if content:
@@ -35,6 +26,13 @@ class Moderation:
         return resp
 
 
-if __name__ == "__main__":
+async def main():
     moderation = Moderation()
-    print(moderation.moderation(content=["I will kill you", "The weather is really nice today", "I want to hit you"]))
+    rsp = await moderation.amoderation(
+        content=["I will kill you", "The weather is really nice today", "I want to hit you"]
+    )
+    print(rsp)
+
+
+if __name__ == "__main__":
+    asyncio.run(main())

metagpt/tools/web_browser_engine_selenium.py

@@ -111,6 +111,8 @@ def _gen_get_driver_func(browser_type, *args, executable_path=None):
         options.add_argument("--headless")
         options.add_argument("--enable-javascript")
         if browser_type == "chrome":
+            options.add_argument("--disable-gpu")  # This flag can help avoid renderer issue
+            options.add_argument("--disable-dev-shm-usage")  # Overcome limited resource problems
             options.add_argument("--no-sandbox")
         for i in args:
             options.add_argument(i)

metagpt/utils/common.py

@@ -13,15 +13,21 @@ from __future__ import annotations
 
 import ast
 import contextlib
+import importlib
 import inspect
+import json
 import os
 import platform
 import re
+import sys
+import traceback
+import typing
 from pathlib import Path
-from typing import Callable, List, Tuple, Union
+from typing import Any, Callable, List, Tuple, Union, get_args, get_origin
 
 import aiofiles
 import loguru
+from pydantic.json import pydantic_encoder
 from tenacity import RetryCallState, _utils
 
 from metagpt.config import CONFIG
@@ -43,6 +49,12 @@ def check_cmd_exists(command) -> int:
     return result
 
 
+def require_python_version(req_version: tuple[int]) -> bool:
+    if not (2 <= len(req_version) <= 3):
+        raise ValueError("req_version should be (3, 9) or (3, 10, 13)")
+    return True if sys.version_info > req_version else False
+
+
 class OutputParser:
     @classmethod
     def parse_blocks(cls, text: str):
@@ -130,8 +142,32 @@ class OutputParser:
             parsed_data[block] = content
         return parsed_data
 
+    @staticmethod
+    def extract_content(text, tag="CONTENT"):
+        # Use regular expression to extract content between [CONTENT] and [/CONTENT]
+        extracted_content = re.search(rf"\[{tag}\](.*?)\[/{tag}\]", text, re.DOTALL)
+
+        if extracted_content:
+            return extracted_content.group(1).strip()
+        else:
+            return "No content found between [CONTENT] and [/CONTENT] tags."
+
+    @staticmethod
+    def is_supported_list_type(i):
+        origin = get_origin(i)
+        if origin is not List:
+            return False
+
+        args = get_args(i)
+        if args == (str,) or args == (Tuple[str, str],) or args == (List[str],):
+            return True
+
+        return False
+
     @classmethod
     def parse_data_with_mapping(cls, data, mapping):
+        if "[CONTENT]" in data:
+            data = cls.extract_content(text=data)
         block_dict = cls.parse_blocks(data)
         parsed_data = {}
         for block, content in block_dict.items():
@@ -198,7 +234,7 @@ class OutputParser:
                 result = ast.literal_eval(structure_text)
 
                 # Ensure the result matches the specified data type
-                if isinstance(result, list) or isinstance(result, dict):
+                if isinstance(result, (list, dict)):
                     return result
 
                 raise ValueError(f"The extracted structure is not a {data_type}.")
@@ -437,6 +473,81 @@ def general_after_log(i: "loguru.Logger", sec_format: str = "%0.3f") -> typing.C
     return log_it
 
 
+def read_json_file(json_file: str, encoding=None) -> list[Any]:
+    if not Path(json_file).exists():
+        raise FileNotFoundError(f"json_file: {json_file} not exist, return []")
+
+    with open(json_file, "r", encoding=encoding) as fin:
+        try:
+            data = json.load(fin)
+        except Exception:
+            raise ValueError(f"read json file: {json_file} failed")
+    return data
+
+
+def write_json_file(json_file: str, data: list, encoding=None):
+    folder_path = Path(json_file).parent
+    if not folder_path.exists():
+        folder_path.mkdir(parents=True, exist_ok=True)
+
+    with open(json_file, "w", encoding=encoding) as fout:
+        json.dump(data, fout, ensure_ascii=False, indent=4, default=pydantic_encoder)
+
+
+def import_class(class_name: str, module_name: str) -> type:
+    module = importlib.import_module(module_name)
+    a_class = getattr(module, class_name)
+    return a_class
+
+
+def import_class_inst(class_name: str, module_name: str, *args, **kwargs) -> object:
+    a_class = import_class(class_name, module_name)
+    class_inst = a_class(*args, **kwargs)
+    return class_inst
+
+
+def format_trackback_info(limit: int = 2):
+    return traceback.format_exc(limit=limit)
+
+
+def serialize_decorator(func):
+    async def wrapper(self, *args, **kwargs):
+        try:
+            result = await func(self, *args, **kwargs)
+            return result
+        except KeyboardInterrupt:
+            logger.error(f"KeyboardInterrupt occurs, start to serialize the project, exp:\n{format_trackback_info()}")
+        except Exception:
+            logger.error(f"Exception occurs, start to serialize the project, exp:\n{format_trackback_info()}")
+        self.serialize()  # Team.serialize
+
+    return wrapper
+
+
+def role_raise_decorator(func):
+    async def wrapper(self, *args, **kwargs):
+        try:
+            return await func(self, *args, **kwargs)
+        except KeyboardInterrupt as kbi:
+            logger.error(f"KeyboardInterrupt: {kbi} occurs, start to serialize the project")
+            if self.latest_observed_msg:
+                self._rc.memory.delete(self.latest_observed_msg)
+            # raise again to make it captured outside
+            raise Exception(format_trackback_info(limit=None))
+        except Exception:
+            if self.latest_observed_msg:
+                logger.warning(
+                    "There is a exception in role's execution, in order to resume, "
+                    "we delete the newest role communication message in the role's memory."
+                )
+                # remove role newest observed msg to make it observed again
+                self._rc.memory.delete(self.latest_observed_msg)
+            # raise again to make it captured outside
+            raise Exception(format_trackback_info(limit=None))
+
+    return wrapper
+
+
 @handle_exception
 async def aread(file_path: str) -> str:
     """Read file asynchronously."""

metagpt/utils/exceptions.py

@@ -21,6 +21,7 @@ def handle_exception(
     _func: Callable[..., ReturnType] = None,
     *,
     exception_type: Union[Type[Exception], Tuple[Type[Exception], ...]] = Exception,
+    exception_msg: str = "",
     default_return: Any = None,
 ) -> Callable[..., ReturnType]:
     """handle exception, return default value"""
@@ -32,8 +33,9 @@ def handle_exception(
                 return await func(*args, **kwargs)
             except exception_type as e:
                 logger.opt(depth=1).error(
-                    f"Calling {func.__name__} with args: {args}, kwargs: {kwargs} failed: {e}, "
-                    f"stack: {traceback.format_exc()}"
+                    f"{e}: {exception_msg}, "
+                    f"\nCalling {func.__name__} with args: {args}, kwargs: {kwargs} "
+                    f"\nStack: {traceback.format_exc()}"
                 )
                 return default_return
 

metagpt/utils/get_template.py

@@ -8,10 +8,10 @@
 from metagpt.config import CONFIG
 
 
-def get_template(templates, format=CONFIG.prompt_format):
-    selected_templates = templates.get(format)
+def get_template(templates, schema=CONFIG.prompt_schema):
+    selected_templates = templates.get(schema)
     if selected_templates is None:
-        raise ValueError(f"Can't find {format} in passed in templates")
+        raise ValueError(f"Can't find {schema} in passed in templates")
 
     # Extract the selected templates
     prompt_template = selected_templates["PROMPT_TEMPLATE"]

metagpt/utils/make_sk_kernel.py

@@ -21,14 +21,12 @@ def make_sk_kernel():
     if CONFIG.openai_api_type == "azure":
         kernel.add_chat_service(
             "chat_completion",
-            AzureChatCompletion(CONFIG.deployment_name, CONFIG.openai_api_base, CONFIG.openai_api_key),
+            AzureChatCompletion(CONFIG.deployment_name, CONFIG.openai_base_url, CONFIG.openai_api_key),
         )
     else:
         kernel.add_chat_service(
             "chat_completion",
-            OpenAIChatCompletion(
-                CONFIG.openai_api_model, CONFIG.openai_api_key, org_id=None, endpoint=CONFIG.openai_api_base
-            ),
+            OpenAIChatCompletion(CONFIG.openai_api_model, CONFIG.openai_api_key),
         )
 
     return kernel

metagpt/utils/repair_llm_raw_output.py

@@ -253,7 +253,7 @@ def retry_parse_json_text(output: str) -> Union[list, dict]:
         if CONFIG.repair_llm_output is True, the _aask_v1 and the retry_parse_json_text will loop for {x=3*3} times.
             it's a two-layer retry cycle
     """
-    logger.debug(f"output to json decode:\n{output}")
+    # logger.debug(f"output to json decode:\n{output}")
 
     # if CONFIG.repair_llm_output is True, it will try to fix output until the retry break
     parsed_data = CustomDecoder(strict=False).decode(output)

metagpt/utils/serialize.py

@@ -4,13 +4,11 @@
 
 import copy
 import pickle
-from typing import Dict, List
 
-from metagpt.actions.action_output import ActionOutput
-from metagpt.schema import Message
+from metagpt.utils.common import import_class
 
 
-def actionoutout_schema_to_mapping(schema: Dict) -> Dict:
+def actionoutout_schema_to_mapping(schema: dict) -> dict:
     """
     directly traverse the `properties` in the first level.
     schema structure likes
@@ -35,14 +33,31 @@ def actionoutout_schema_to_mapping(schema: Dict) -> Dict:
         if property["type"] == "string":
             mapping[field] = (str, ...)
         elif property["type"] == "array" and property["items"]["type"] == "string":
-            mapping[field] = (List[str], ...)
+            mapping[field] = (list[str], ...)
         elif property["type"] == "array" and property["items"]["type"] == "array":
-            # here only consider the `List[List[str]]` situation
-            mapping[field] = (List[List[str]], ...)
+            # here only consider the `list[list[str]]` situation
+            mapping[field] = (list[list[str]], ...)
     return mapping
 
 
-def serialize_message(message: Message):
+def actionoutput_mapping_to_str(mapping: dict) -> dict:
+    new_mapping = {}
+    for key, value in mapping.items():
+        new_mapping[key] = str(value)
+    return new_mapping
+
+
+def actionoutput_str_to_mapping(mapping: dict) -> dict:
+    new_mapping = {}
+    for key, value in mapping.items():
+        if value == "(<class 'str'>, Ellipsis)":
+            new_mapping[key] = (str, ...)
+        else:
+            new_mapping[key] = eval(value)  # `"'(list[str], Ellipsis)"` to `(list[str], ...)`
+    return new_mapping
+
+
+def serialize_message(message: "Message"):
     message_cp = copy.deepcopy(message)  # avoid `instruct_content` value update by reference
     ic = message_cp.instruct_content
     if ic:
@@ -56,11 +71,12 @@ def serialize_message(message: Message):
     return msg_ser
 
 
-def deserialize_message(message_ser: str) -> Message:
+def deserialize_message(message_ser: str) -> "Message":
     message = pickle.loads(message_ser)
     if message.instruct_content:
         ic = message.instruct_content
-        ic_obj = ActionOutput.create_model_class(class_name=ic["class"], mapping=ic["mapping"])
+        actionnode_class = import_class("ActionNode", "metagpt.actions.action_node")  # avoid circular import
+        ic_obj = actionnode_class.create_model_class(class_name=ic["class"], mapping=ic["mapping"])
         ic_new = ic_obj(**ic["value"])
         message.instruct_content = ic_new
 

metagpt/utils/token_counter.py

@@ -7,6 +7,7 @@
 ref1: https://github.com/openai/openai-cookbook/blob/main/examples/How_to_count_tokens_with_tiktoken.ipynb
 ref2: https://github.com/Significant-Gravitas/Auto-GPT/blob/master/autogpt/llm/token_counter.py
 ref3: https://github.com/hwchase17/langchain/blob/master/langchain/chat_models/openai.py
+ref4: https://ai.google.dev/models/gemini
 """
 import tiktoken
 
@@ -16,6 +17,8 @@ TOKEN_COSTS = {
     "gpt-3.5-turbo-0613": {"prompt": 0.0015, "completion": 0.002},
     "gpt-3.5-turbo-16k": {"prompt": 0.003, "completion": 0.004},
     "gpt-3.5-turbo-16k-0613": {"prompt": 0.003, "completion": 0.004},
+    "gpt-35-turbo": {"prompt": 0.0015, "completion": 0.002},
+    "gpt-35-turbo-16k": {"prompt": 0.003, "completion": 0.004},
     "gpt-3.5-turbo-1106": {"prompt": 0.001, "completion": 0.002},
     "gpt-4-0314": {"prompt": 0.03, "completion": 0.06},
     "gpt-4": {"prompt": 0.03, "completion": 0.06},
@@ -25,6 +28,7 @@ TOKEN_COSTS = {
     "gpt-4-1106-preview": {"prompt": 0.01, "completion": 0.03},
     "text-embedding-ada-002": {"prompt": 0.0004, "completion": 0.0},
     "chatglm_turbo": {"prompt": 0.0, "completion": 0.00069},  # 32k version, prompt + completion tokens=0.005￥/k-tokens
+    "gemini-pro": {"prompt": 0.00025, "completion": 0.0005},
 }
 
 
@@ -34,6 +38,8 @@ TOKEN_MAX = {
     "gpt-3.5-turbo-0613": 4096,
     "gpt-3.5-turbo-16k": 16384,
     "gpt-3.5-turbo-16k-0613": 16384,
+    "gpt-35-turbo": 4096,
+    "gpt-35-turbo-16k": 16384,
     "gpt-3.5-turbo-1106": 16384,
     "gpt-4-0314": 8192,
     "gpt-4": 8192,
@@ -43,6 +49,7 @@ TOKEN_MAX = {
     "gpt-4-1106-preview": 128000,
     "text-embedding-ada-002": 8192,
     "chatglm_turbo": 32768,
+    "gemini-pro": 32768,
 }
 
 
@@ -56,6 +63,8 @@ def count_message_tokens(messages, model="gpt-3.5-turbo-0613"):
     if model in {
         "gpt-3.5-turbo-0613",
         "gpt-3.5-turbo-16k-0613",
+        "gpt-35-turbo",
+        "gpt-35-turbo-16k",
         "gpt-3.5-turbo-16k",
         "gpt-3.5-turbo-1106",
         "gpt-4-0314",