babyAGI(7)-babyCoder源码阅读3（任务执行相关agent）

本文链接：https://blog.csdn.net/QZXSY1/article/details/137359871

1. 任务分配agent

该agent会根据，objective和task的内容自动分配要执行的agent

code_writer_agent 编写新代码或增加新功能使用该agent
code_reactor_agent 任务涉及到修改或优化执行该agent
command_executor_agent 任务涉及到文件操作执行该agent

整个任务分配的agent分成了两部分，一部分是任务推荐建议的agent，给出推荐的agent的建议

def task_assigner_recommendation_agent(objective: str, task: str):
    prompt = f"""You are an AGI agent responsible for providing recommendations on which agent should be used to handle a specific task. Analyze the provided major objective of the project and a single task from the JSON checklist generated by the previous agent, and suggest the most appropriate agent to work on the task.

    The overall objective is: {objective}
    The current task is: {task}
    
    The available agents are:
    1. code_writer_agent: Responsible for writing code based on the task description.
    2. code_refactor_agent: Responsible for editing existing code.
    3. command_executor_agent: Responsible for executing commands and handling file operations, such as creating, moving, or deleting files.

    When analyzing the task, consider the following tips:
    - Pay attention to keywords in the task description that indicate the type of action required, such as "write", "edit", "run", "create", "move", or "delete".
    - Keep the overall objective in mind, as it can help you understand the context of the task and guide your choice of agent.
    - If the task involves writing new code or adding new functionality, consider using the code_writer_agent.
    - If the task involves modifying or optimizing existing code, consider using the code_refactor_agent.
    - If the task involves file operations, command execution, or running a script, consider using the command_executor_agent.

    Based on the task and overall objective, suggest the most appropriate agent to work on the task."""
    return openai_call(prompt, temperature=0.5, max_tokens=2000)

第二部分，是结合推荐的建议，给出一个确定的agent推荐。

def task_assigner_agent(objective: str, task: str, recommendation: str):
    prompt = f"""You are an AGI agent responsible for choosing the best agent to work on a given task. Your goal is to analyze the provided major objective of the project and a single task from the JSON checklist generated by the previous agent, and choose the best agent to work on the task.

    The overall objective is: {objective}
    The current task is: {task}

    Use this recommendation to guide you: {recommendation}
        
    The available agents are:
    1. code_writer_agent: Responsible for writing code based on the task description.
    2. code_refactor_agent: Responsible for editing existing code.
    2. command_executor_agent: Responsible for executing commands and handling file operations, such as creating, moving, or deleting files.

    Please consider the task description and the overall objective when choosing the most appropriate agent. Keep in mind that creating a file and writing code are different tasks. If the task involves creating a file, like "calculator.py" but does not mention writing any code inside it, the command_executor_agent should be used for this purpose. The code_writer_agent should only be used when the task requires writing or adding code to a file. The code_refactor_agent should only be used when the task requires modifying existing code.
    
    TLDR: To create files, use command_executor_agent, to write text/code to files, use code_writer_agent, to modify existing code, use code_refactor_agent.

    Choose the most appropriate agent to work on the task and return a JSON output with the following format: {{"agent": "agent_name"}}. ONLY return JSON output:"""
    return openai_call(prompt, temperature=0, max_tokens=2000)

2. 执行具体任务agent实现

2.1 执行命令agent

负责执行命令和处理文件操作，例如创建、移动或删除文件
这里还需要输入os的信息

def command_executor_agent(task: str, file_path: str):
    prompt = f"""You are an AGI agent responsible for executing a given command on the {os_version} OS. Your goal is to analyze the provided major objective of the project and a single task from the JSON checklist generated by the previous agent, and execute the command on the {os_version} OS. 

    The current task is: {task}
    File or folder name referenced in the task (relative file path): {file_path} 
    
    Based on the task, write the appropriate command to execute on the {os_version} OS. Make sure the command is relevant to the task and objective. For example, if the task is to create a new folder, the command should be 'mkdir new_folder_name'. Return the command as a JSON output with the following format: {{"command": "command_to_execute"}}. ONLY return JSON output:"""
    return openai_call(prompt, temperature=0, max_tokens=2000)

2.2 代码编写agent

这个函数需要输入三个参数

task 当前任务
isolated_context 该字段作为参考上下文，了解与任务相关的其他代码库部分
context_code_chunks 与该段代码存在相关性的代码
这段代码同样规定了函数中的一些参数

def code_writer_agent(task: str, isolated_context: str, context_code_chunks):
    prompt = f"""You are an AGI agent responsible for writing code to accomplish a given task. Your goal is to analyze the provided major objective of the project and a single task from the JSON checklist generated by the previous agent, and write the necessary code to complete the task.

    The current task is: {task}

    To help you make the code useful in this codebase, use this context as reference of the other pieces of the codebase that are relevant to your task. PAY ATTENTION TO THIS: {isolated_context}
    
    The following code chunks were found to be relevant to the task. You can use them as reference to write the code if they are useful. PAY CLOSE ATTENTION TO THIS: 
    {context_code_chunks}

    Note: Always use 'encoding='utf-8'' when opening files with open().
    
    Based on the task and objective, write the appropriate code to achieve the task. Make sure the code is relevant to the task and objective, and follows best practices. Return the code as a plain text output and NOTHING ELSE. Use identation and line breaks in the in the code. Make sure to only write the code and nothing else as your output will be saved directly to the file by other agent. IMPORTANT" If the task is asking you to write code to write files, this is a mistake! Interpret it and either do nothing or return  the plain code, not a code to write file, not a code to write code, etc."""
    return openai_call(prompt, temperature=0, max_tokens=2000)

2.3 代码重构agent

这个函数有四个参数

task_description 任务描述
existing_code_snippet 存在的任务片段，你需要重构的
isolated_context 该字段作为参考上下文，了解与任务相关的其他代码库部分
context_code_chunks 与该段代码存在相关性的代码

def code_refactor_agent(task_description: str, existing_code_snippet: str, context_chunks, isolated_context: str):

    prompt = f"""You are an AGI agent responsible for refactoring code to accomplish a given task. Your goal is to analyze the provided major objective of the project, the task descriptionm and refactor the code accordingly.

    The current task description is: {task_description}
    To help you make the code useful in this codebase, use this context as reference of the other pieces of the codebase that are relevant to your task: {isolated_context}

    Here are some context chunks that might be relevant to the task:
    {context_chunks}
    
    Existing code you should refactor: 
    {existing_code_snippet}
    
    Based on the task description, objective, refactor the existing code to achieve the task. Make sure the refactored code is relevant to the task and objective, follows best practices, etc.

    Return a plain text code snippet with your refactored code. IMPORTANT: JUST RETURN CODE, YOUR OUTPUT WILL BE ADDED DIRECTLY TO THE FILE BY OTHER AGENT. BE MINDFUL OF THIS:"""

    return openai_call(prompt, temperature=0, max_tokens=2000)

2.4 文件管理agent

这个函数主要用来根据任务和目标管理项目中的文件

objective 目标
task 任务
current_directory_files 当前文件夹下的文件
file_path 路径

def file_management_agent(objective: str, task: str, current_directory_files: str, file_path: str):
    prompt = f"""You are an AGI agent responsible for managing files in a software project. Your goal is to analyze the provided major objective of the project and a single task from the JSON checklist generated by the previous agent, and determine the appropriate file path and name for the generated code.

    The overall objective is: {objective}
    The current task is: {task}
    Specified file path (relative path from the current dir): {file_path}

    Make the file path adapted for the current directory files. The current directory files are: {current_directory_files}. Assume this file_path will be interpreted from the root path of the directory.

    Do not use '.' or './' in the file path.

    BE VERY SPECIFIC WITH THE FILES, AVOID FILE DUPLICATION, AVOID SPECIFYING THE SAME FILE NAME UNDER DIFFERENT FOLDERS, ETC.

    Based on the task, determine the file path and name for the generated code. Return the file path and name as a JSON output with the following format: {{"file_path": "file_path_and_name"}}. ONLY return JSON output:"""
    return openai_call(prompt, temperature=0, max_tokens=2000)

2.5 代码相关性agent

这个函数用来判断当前的代码片段是否和目标、任务描述相关

objective 目标
task_description 任务描述
code_chunk 代码段
返回0-10的值判断相关性

def code_relevance_agent(objective: str, task_description: str, code_chunk: str):
    prompt = f"""You are an AGI agent responsible for evaluating the relevance of a code chunk in relation to a given task. Your goal is to analyze the provided major objective of the project, the task description, and the code chunk, and assign a relevance score from 0 to 10, where 0 is completely irrelevant and 10 is highly relevant.

    The overall objective is: {objective}
    The current task description is: {task_description}
    The code chunk is as follows (line numbers included):
    {code_chunk}

    Based on the task description, objective, and code chunk, assign a relevance score between 0 and 10 (inclusive) for the code chunk. DO NOT OUTPUT ANYTHING OTHER THAN THE RELEVANCE SCORE AS A NUMBER."""

    relevance_score = openai_call(prompt, temperature=0.5, max_tokens=50)

2.6 根据人类的输入校正任务agent

该函数根据人类的输入来提高软件项目任务的质量，分析提供的任务，并根据人类的建议进行调整

task 任务
humman_feedback 人类反馈

def task_human_input_agent(task: str, human_feedback: str):
    prompt = f"""You are an AGI agent responsible for getting human input to improve the quality of tasks in a software project. Your goal is to analyze the provided task and adapt it based on the human's suggestions. The tasks should  start with either 'Run a command to...', 'Write code to...', or 'Edit existing code to...' depending on the agent that will execute the task.

    For context, this task will be executed by other AGI agents with the following characteristics:
    - code_writer_agent: Writes code snippets or functions and saves them to the appropriate files. This agent can also append code to existing files if required.
    - code_refactor_agent: Responsible for modifying and refactoring existing code to meet the requirements of the task.
    - command_executor_agent: Executes terminal commands for tasks such as creating directories, installing dependencies, etc.

    The current task is:
    {task}

    The human feedback is:
    {human_feedback}

    If the human feedback is empty, return the task as is. If the human feedback is saying to ignore the task, return the following string: <IGNORE_TASK>

    Note that your output will replace the existing task, so make sure that your output is a valid task that starts with one of the required phrases ('Run a command to...', 'Write code to...', 'Edit existing code to...').
    
    Please adjust the task based on the human feedback while ensuring it starts with one of the required phrases ('Run a command to...', 'Write code to...', 'Edit existing code to...'). Return the improved task as a plain text output and nothing else. Write only the new task."""

    return openai_call(prompt, temperature=0.3, max_tokens=200)