Maestro实践

GitHub - Doriandarko/maestro: A framework for Claude Opus to intelligently orchestrate subagents.

初步使用了一下，可能是我使用不当，感觉并没有很厉害

1)克隆代码库

git clone GitHub - Doriandarko/maestro: A framework for Claude Opus to intelligently orchestrate subagents. 2)我用的是ollama

maestro-ollama.py

import os
import re
from datetime import datetime
import json
from rich.console import Console
from rich.panel import Panel
import ollama
from ollama import Client  # Import the Ollama client
import argparse

# Only for the first time run based on the model you want to use
# ollama.pull('llama3:70b')
# ollama.pull('llama3:8b')
# ollama.pull('llama3:8b')
# ollama.pull('llama3:70b-instruct')
# ollama.pull('llama3:instruct')

# Define model identifiers as variables at the top of the script
ORCHESTRATOR_MODEL = 'qwen:14b'
SUBAGENT_MODEL = 'deepseek-coder:latest'
REFINER_MODEL = 'llama3:latest'

# check and pull models if they don't exist yet
for model in [ORCHESTRATOR_MODEL, SUBAGENT_MODEL, REFINER_MODEL]:
    try:
        print(f"Checking for model: {model}")
        m = ollama.show(model)
    except ollama._types.ResponseError as e:
        print(f"Pulling model from ollama: {model}")
        ollama.pull(model)

# Initialize the Ollama client
client = Client(host='http://localhost:11434')

console = Console()

def opus_orchestrator(objective, file_content=None, previous_results=None):
    console.print(f"\n[bold]Calling Ollama Orchestrator for your objective[/bold]")
    previous_results_text = "\n".join(previous_results) if previous_results else "None"
    if file_content:
        console.print(Panel(f"File content:\n{file_content}", title="[bold blue]File Content[/bold blue]", title_align="left", border_style="blue"))
    
    response = client.chat(
        model=ORCHESTRATOR_MODEL,
        messages=[
            {
                "role": "user",
                "content": f"Based on the following objective{' and file content' if file_content else ''}, and the previous sub-task results (if any), please break down the objective into the next sub-task, and create a concise and detailed prompt for a subagent so it can execute that task. Focus solely on the objective and avoid engaging in casual conversation with the subagent.\n\nWhen dealing with code tasks, make sure to check the code for errors and provide fixes and support as part of the next sub-task. If you find any bugs or have suggestions for better code, please include them in the next sub-task prompt.\n\nPlease assess if the objective has been fully achieved. If the previous sub-task results comprehensively address all aspects of the objective, include the phrase 'The task is complete:' at the beginning of your response. If the objective is not yet fully achieved, break it down into the next sub-task and create a concise and detailed prompt for a subagent to execute that task.\n\nObjective: {objective}" + (f'\nFile content:\n{file_content}' if file_content else '') + f"\n\nPrevious sub-task results:\n{previous_results_text}"
            }
        ]
    )
    
    response_text = response['message']['content']
    console.print(Panel(response_text, title="[bold green]Ollama Orchestrator[/bold green]", title_align="left", border_style="green", subtitle="Sending task to Ollama sub-agent 👇"))
    return response_text, file_content

def haiku_sub_agent(prompt, previous_haiku_tasks=None, continuation=False):
    if previous_haiku_tasks is None:
        previous_haiku_tasks = []

    continuation_prompt = "Continuing from the previous answer, please complete the response."
    if continuation:
        prompt = continuation_prompt

    # Compile previous tasks into a readable format
    previous_tasks_summary = "Previous Sub-agent tasks:\n" + "\n".join(f"Task: {task['task']}\nResult: {task['result']}" for task in previous_haiku_tasks)
    
    # Append previous tasks summary to the prompt
    full_prompt = f"{previous_tasks_summary}\n\n{prompt}"

    # Ensure prompt is not empty
    if not full_prompt.strip():
        raise ValueError("Prompt cannot be empty")

    response = client.chat(
        model=SUBAGENT_MODEL, 
        messages=[{"role": "user", "content": full_prompt}]
    )
    
    response_text = response['message']['content']
    
    if len(response_text) >= 4000:  # Threshold set to 4000 as a precaution
        console.print("[bold yellow]Warning:[/bold yellow] Output may be truncated. Attempting to continue the response.")
        continuation_response_text = haiku_sub_agent(continuation_prompt, previous_haiku_tasks, continuation=True)
        response_text += continuation_response_text

    console.print(Panel(response_text, title="[bold blue]Ollama Sub-agent Result[/bold blue]", title_align="left", border_style="blue", subtitle="Task completed, sending result to Ollama Orchestrator 👇"))
    return response_text

def opus_refine(objective, sub_task_results, filename, projectname, continuation=False):
    console.print("\nCalling Ollama to provide the refined final output for your objective:")
    
    response = client.chat(
        model=REFINER_MODEL,
        messages=[
            {
                "role": "user",
                "content": "Objective: " + objective + "\n\nSub-task results:\n" + "\n".join(sub_task_results) + "\n\nPlease review and refine the sub-task results into a cohesive final output. Add any missing information or details as needed.\n\nWhen working on code projects, ONLY AND ONLY IF THE PROJECT IS CLEARLY A CODING ONE, please provide the following:\n\n1. Project Name: Create a concise and appropriate project name that fits the project based on what it's creating. The project name should be no more than 20 characters long.\n\n2. Folder Structure: Provide the folder structure as a valid JSON object, where each key represents a folder or file, and nested keys represent subfolders. Use null values for files. Ensure the JSON is properly formatted without any syntax errors. Please make sure all keys are enclosed in double quotes, and ensure objects are correctly encapsulated with braces, separating items with commas as necessary. Wrap the JSON object in <folder_structure> tags.\n\n3. Code Files: For each code file, include ONLY the file name, NEVER EVER USE THE FILE PATH OR ANY OTHER FORMATTING. YOU ONLY USE THE FOLLOWING format 'Filename: <filename>' followed by the code block enclosed in triple backticks, with the language identifier after the opening backticks, like this:\n\npython\n<code>\n\n\nFocus solely on the objective and avoid engaging in casual conversation. Ensure the final output is clear, concise, and addresses all aspects of the objective.​"
            }
        ]
    )
    
    response_text = response['message']['content']
    
    if len(response_text) >= 4000:  # Threshold set to 4000 as a precaution
        console.print("[bold yellow]Warning:[/bold yellow] Output may be truncated. Attempting to continue the response.")
        continuation_response_text = opus_refine(objective, sub_task_results, filename, projectname, continuation=True)
        response_text += continuation_response_text

    console.print(Panel(response_text, title="[bold green]Final Output[/bold green]", title_align="left", border_style="green"))
    return response_text

def create_folder_structure(project_name, folder_structure, code_blocks):
    try:
        os.makedirs(project_name, exist_ok=True)
        console.print(Panel(f"Created project folder: [bold]{project_name}[/bold]", title="[bold blue]Project Folder Creation[/bold blue]", title_align="left", border_style="blue"))
    except OSError as e:
        console.print(Panel(f"Error creating project folder: [bold]{project_name}[/bold]\nError: {e}", title="[bold red]Project Folder Creation Error[/bold red]", title_align="left", border_style="red"))

    create_folders_and_files(project_name, folder_structure, code_blocks)

def create_folders_and_files(current_path, structure, code_blocks):
    for key, value in structure.items():
        path = os.path.join(current_path, key)
        if isinstance(value, dict):
            try:
                os.makedirs(path, exist_ok=True)
                console.print(Panel(f"Created folder: [bold]{path}[/bold]", title="[bold blue]Folder Creation[/bold blue]", title_align="left", border_style="blue"))
                create_folders_and_files(path, value, code_blocks)
            except OSError as e:
                console.print(Panel(f"Error creating folder: [bold]{path}[/bold]\nError: {e}", title="[bold red]Folder Creation Error[/bold red]", title_align="left", border_style="red"))
        else:
            code_content = next((code for file, code in code_blocks if file == key), None)
            if code_content:
                try:
                    with open(path, 'w') as file:
                        file.write(code_content)
                    console.print(Panel(f"Created file: [bold]{path}[/bold]", title="[bold green]File Creation[/bold green]", title_align="left", border_style="green"))
                except IOError as e:
                    console.print(Panel(f"Error creating file: [bold]{path}[/bold]\nError: {e}", title="[bold red]File Creation Error[/bold red]", title_align="left", border_style="red"))
            else:
                console.print(Panel(f"Code content not found for file: [bold]{key}[/bold]", title="[bold yellow]Missing Code Content[/bold yellow]", title_align="left", border_style="yellow"))

def read_file(file_path):
    with open(file_path, 'r') as file:
        content = file.read()
    return content
   
def has_task_data():
    return os.path.exists('task_data.json')

def read_task_data():
    with open('task_data.json', 'r') as file:
        task_data = json.load(file)
    return task_data


def write_task_data(task_data):
    with open('task_data.json', 'w') as file:
        json.dump(task_data, file)


continue_from_last_task = False
tmp_task_data = {}

# parse args
parser = argparse.ArgumentParser()
parser.add_argument('-p', '--prompt', type=str, help='Please enter your objective with or without a text file path')
args = parser.parse_args()

if args.prompt is not None:
    objective = args.prompt
else:
    # Check if there is a task data file
    if has_task_data():
        continue_from_last_task = input("Do you want to continue from the last task? (y/n): ").lower() == 'y'

    if continue_from_last_task:
        tmp_task_data = read_task_data()
        objective = tmp_task_data['objective']
        task_exchanges = tmp_task_data['task_exchanges']
        console.print(Panel(f"Resuming from last task: {objective}", title="[bold blue]Resuming from last task[/bold blue]", title_align="left", border_style="blue"))
    else:
        # Get the objective from user input
        objective = input("Please enter your objective with or without a text file path: ")
        tmp_task_data['objective'] = objective
        tmp_task_data['task_exchanges'] = []

# Check if the input contains a file path
if "./" in objective or "/" in objective:
    # Extract the file path from the objective
    file_path = re.findall(r'[./\w]+\.[\w]+', objective)[0]
    # Read the file content
    with open(file_path, 'r') as file:
        file_content = file.read()
    # Update the objective string to remove the file path
    objective = objective.split(file_path)[0].strip()
else:
    file_content = None

task_exchanges = []
haiku_tasks = []

while True:
    # Call Orchestrator to break down the objective into the next sub-task or provide the final output
    previous_results = [result for _, result in task_exchanges]
    if not task_exchanges:
        # Pass the file content only in the first iteration if available
        opus_result, file_content_for_haiku = opus_orchestrator(objective, file_content, previous_results)
    else:
        opus_result, _ = opus_orchestrator(objective, previous_results=previous_results)

    if "The task is complete:" in opus_result:
        # If Opus indicates the task is complete, exit the loop
        final_output = opus_result.replace("The task is complete:", "").strip()
        break
    else:
        sub_task_prompt = opus_result
        # Append file content to the prompt for the initial call to haiku_sub_agent, if applicable
        if file_content_for_haiku and not haiku_tasks:
            sub_task_prompt = f"{sub_task_prompt}\n\nFile content:\n{file_content_for_haiku}"
        # Call haiku_sub_agent with the prepared prompt and record the result
        sub_task_result = haiku_sub_agent(sub_task_prompt, haiku_tasks)
        # Log the task and its result for future reference
        haiku_tasks.append({"task": sub_task_prompt, "result": sub_task_result})
        # Record the exchange for processing and output generation
        task_exchanges.append((sub_task_prompt, sub_task_result))
        # Update the task data with the new task exchanges
        tmp_task_data['task_exchanges'] = task_exchanges
        # Save the task data to a JSON file for resuming later
        write_task_data(tmp_task_data)
        # Prevent file content from being included in future haiku_sub_agent calls
        file_content_for_haiku = None

# Create the .md filename
sanitized_objective = re.sub(r'\W+', '_', objective)
timestamp = datetime.now().strftime("%H-%M-%S")

# Call Opus to review and refine the sub-task results
refined_output = opus_refine(objective, [result for _, result in task_exchanges], timestamp, sanitized_objective)

# Extract the project name from the refined output
project_name_match = re.search(r'Project Name: (.*)', refined_output)
project_name = project_name_match.group(1).strip() if project_name_match else sanitized_objective

# Extract the folder structure from the refined output
folder_structure_match = re.search(r'<folder_structure>(.*?)</folder_structure>', refined_output, re.DOTALL)
folder_structure = {}
if folder_structure_match:
    json_string = folder_structure_match.group(1).strip()
    try:
        folder_structure = json.loads(json_string)
    except json.JSONDecodeError as e:
        console.print(Panel(f"Error parsing JSON: {e}", title="[bold red]JSON Parsing Error[/bold red]", title_align="left", border_style="red"))
        console.print(Panel(f"Invalid JSON string: [bold]{json_string}[/bold]", title="[bold red]Invalid JSON String[/bold red]", title_align="left", border_style="red"))

# Extract code files from the refined output
code_blocks = re.findall(r'Filename: (\S+)\s*```[\w]*\n(.*?)\n```', refined_output, re.DOTALL)

# Create the folder structure and code files
create_folder_structure(project_name, folder_structure, code_blocks)

# Truncate the sanitized_objective to a maximum of 50 characters
max_length = 25
truncated_objective = sanitized_objective[:max_length] if len(sanitized_objective) > max_length else sanitized_objective

# Update the filename to include the project name
filename = f"{timestamp}_{truncated_objective}.md"

# Prepare the full exchange log
exchange_log = f"Objective: {objective}\n\n"
exchange_log += "=" * 40 + " Task Breakdown " + "=" * 40 + "\n\n"
for i, (prompt, result) in enumerate(task_exchanges, start=1):
    exchange_log += f"Task {i}:\n"
    exchange_log += f"Prompt: {prompt}\n"
    exchange_log += f"Result: {result}\n\n"

exchange_log += "=" * 40 + " Refined Final Output " + "=" * 40 + "\n\n"
exchange_log += refined_output

console.print(f"\n[bold]Refined Final output:[/bold]\n{refined_output}")

with open(filename, 'w') as file:
    file.write(exchange_log)
print(f"\nFull exchange log saved to {filename}")

我用的是

ORCHESTRATOR_MODEL = 'qwen:14b'

SUBAGENT_MODEL = 'deepseek-coder:latest'

REFINER_MODEL = 'llama3:latest'
这三个模型，其中

• 调用 Orchestrator 模型分解任务。
• 调用 Subagent 模型执行子任务。
• 调用 Refiner 模型整合子任务结果。

比如我允许之后输入一个生成贪吃蛇小游戏

他就会生成对应的md文档和对应的文件夹结构，文档里面有对应的文件的内容，需要你自己创建文件并复制粘贴，我运行了一下，效果非常lllllllllllllllllllow

可能是我模型选的不对，而且还是量化过的，就这吧