import argparse import sys from rich.panel import Panel from rich.console import Console from langchain_core.messages import HumanMessage from langgraph.checkpoint.memory import MemorySaver from langgraph.prebuilt import create_react_agent from ra_aid.tools import ( ask_expert, run_shell_command, run_programming_task, emit_research_notes, emit_plan, emit_related_files, emit_task, emit_expert_context, get_memory_value, emit_key_facts, delete_key_facts, emit_key_snippets, delete_key_snippets, emit_research_subtask, request_implementation, read_file_tool, fuzzy_find_project_files, ripgrep_search, list_directory_tree, swap_task_order ) from ra_aid.env import validate_environment from ra_aid.tools.memory import _global_memory, get_related_files, one_shot_completed from ra_aid import print_agent_output, print_stage_header, print_task_header, print_error from ra_aid.prompts import ( RESEARCH_PROMPT, PLANNING_PROMPT, IMPLEMENTATION_PROMPT, ) from ra_aid.exceptions import TaskCompletedException import time from anthropic import APIError, APITimeoutError, RateLimitError, InternalServerError from ra_aid.llm import initialize_llm # Read-only tools that don't modify system state READ_ONLY_TOOLS = [ emit_related_files, emit_key_facts, delete_key_facts, emit_key_snippets, delete_key_snippets, read_file_tool, fuzzy_find_project_files, ripgrep_search ] # Tools that can modify files or system state MODIFICATION_TOOLS = [ run_shell_command, run_programming_task ] # Common tools used across multiple agents COMMON_TOOLS = READ_ONLY_TOOLS + [] # Expert-specific tools EXPERT_TOOLS = [ emit_expert_context, ask_expert ] # Research-specific tools RESEARCH_TOOLS = [ list_directory_tree, emit_research_subtask, run_shell_command, emit_research_notes, one_shot_completed ] def parse_arguments(): parser = argparse.ArgumentParser( description='RA.Aid - AI Agent for executing programming and research tasks', formatter_class=argparse.RawDescriptionHelpFormatter, epilog=''' Examples: ra-aid -m "Add error handling to the database module" ra-aid -m "Explain the authentication flow" --research-only ''' ) parser.add_argument( '-m', '--message', type=str, help='The task or query to be executed by the agent' ) parser.add_argument( '--research-only', action='store_true', help='Only perform research without implementation' ) parser.add_argument( '--provider', type=str, default='anthropic', choices=['anthropic', 'openai', 'openrouter', 'openai-compatible'], help='The LLM provider to use' ) parser.add_argument( '--model', type=str, help='The model name to use (required for non-Anthropic providers)' ) parser.add_argument( '--cowboy-mode', action='store_true', help='Skip interactive approval for shell commands' ) parser.add_argument( '--expert-provider', type=str, default='openai', choices=['anthropic', 'openai', 'openrouter', 'openai-compatible'], help='The LLM provider to use for expert knowledge queries (default: openai)' ) parser.add_argument( '--expert-model', type=str, help='The model name to use for expert knowledge queries (required for non-OpenAI providers)' ) args = parser.parse_args() # Set default model for Anthropic, require model for other providers if args.provider == 'anthropic': if not args.model: args.model = 'claude-3-5-sonnet-20241022' elif not args.model: parser.error(f"--model is required when using provider '{args.provider}'") # Validate expert model requirement if args.expert_provider != 'openai' and not args.expert_model: parser.error(f"--expert-model is required when using expert provider '{args.expert_provider}'") return args # Create console instance console = Console() # Create individual memory objects for each agent research_memory = MemorySaver() planning_memory = MemorySaver() implementation_memory = MemorySaver() def get_research_tools(research_only: bool = False, expert_enabled: bool = True) -> list: """Get the list of research tools based on mode and whether expert is enabled.""" # Start with read-only tools tools = READ_ONLY_TOOLS.copy() # Add research tools except run_shell_command research_tools = [t for t in RESEARCH_TOOLS if t != run_shell_command] tools.extend(research_tools) # Add modification tools if not research_only if not research_only: tools.extend(MODIFICATION_TOOLS) tools.append(request_implementation) # Add expert tools if enabled if expert_enabled: tools.extend(EXPERT_TOOLS) return tools def get_planning_tools(expert_enabled: bool = True) -> list: """Get the list of planning tools based on whether expert is enabled.""" # Start with common tools tools = COMMON_TOOLS.copy() # Add planning-specific tools planning_tools = [ emit_plan, emit_task, swap_task_order ] tools.extend(planning_tools) # Add expert tools if enabled if expert_enabled: tools.extend(EXPERT_TOOLS) return tools def get_implementation_tools(expert_enabled: bool = True) -> list: """Get the list of implementation tools based on whether expert is enabled.""" # Start with common tools tools = COMMON_TOOLS.copy() # Add modification tools since it's not research-only tools.extend(MODIFICATION_TOOLS) # Add expert tools if enabled if expert_enabled: tools.extend(EXPERT_TOOLS) return tools def is_informational_query() -> bool: """Determine if the current query is informational based on implementation_requested state.""" return _global_memory.get('config', {}).get('research_only', False) or not is_stage_requested('implementation') def is_stage_requested(stage: str) -> bool: """Check if a stage has been requested to proceed.""" if stage == 'implementation': return len(_global_memory.get('implementation_requested', [])) > 0 return False def run_agent_with_retry(agent, prompt: str, config: dict): """Run an agent with retry logic for internal server errors.""" max_retries = 20 base_delay = 1 # Initial delay in seconds for attempt in range(max_retries): try: for chunk in agent.stream( {"messages": [HumanMessage(content=prompt)]}, config ): print_agent_output(chunk) break except (InternalServerError, APITimeoutError, RateLimitError, APIError) as e: if attempt == max_retries - 1: raise RuntimeError(f"Max retries ({max_retries}) exceeded. Last error: {str(e)}") delay = base_delay * (2 ** attempt) # Exponential backoff error_type = e.__class__.__name__ print_error(f"Encountered {error_type}: {str(e)}. Retrying in {delay} seconds... (Attempt {attempt + 1}/{max_retries})") time.sleep(delay) continue def run_implementation_stage(base_task, tasks, plan, related_files, model, expert_enabled: bool): """Run implementation stage with a distinct agent for each task.""" if not is_stage_requested('implementation'): print_stage_header("Implementation Stage Skipped") return print_stage_header("Implementation Stage") # Get tasks directly from memory, maintaining order by ID task_list = [task for _, task in sorted(_global_memory['tasks'].items())] print_task_header(f"Found {len(task_list)} tasks to implement") for i, task in enumerate(task_list, 1): print_task_header(task) # Create a unique memory instance for this task task_memory = MemorySaver() # Create a fresh agent for each task task_agent = create_react_agent(model, get_implementation_tools(expert_enabled=expert_enabled), checkpointer=task_memory) # Construct task-specific prompt task_prompt = IMPLEMENTATION_PROMPT.format( plan=plan, key_facts=get_memory_value('key_facts'), key_snippets=get_memory_value('key_snippets'), task=task, related_files="\n".join(related_files), base_task=base_task ) # Run agent for this task run_agent_with_retry(task_agent, task_prompt, {"configurable": {"thread_id": "abc123"}, "recursion_limit": 100}) def run_research_subtasks(base_task: str, config: dict, model, expert_enabled: bool): """Run research subtasks with separate agents.""" subtasks = _global_memory.get('research_subtasks', []) if not subtasks: return print_stage_header("Research Subtasks") # Get tools for subtask agents (excluding emit_research_subtask and implementation) research_only = _global_memory.get('config', {}).get('research_only', False) subtask_tools = [ t for t in get_research_tools(research_only=research_only, expert_enabled=expert_enabled) if t.name not in ['emit_research_subtask'] ] for i, subtask in enumerate(subtasks, 1): print_task_header(f"Research Subtask {i}/{len(subtasks)}") # Create fresh memory and agent for each subtask subtask_memory = MemorySaver() subtask_agent = create_react_agent( model, subtask_tools, checkpointer=subtask_memory ) # Run the subtask agent subtask_prompt = f"Base Task: {base_task}\nResearch Subtask: {subtask}\n\n{RESEARCH_PROMPT}" run_agent_with_retry(subtask_agent, subtask_prompt, config) def main(): """Main entry point for the ra-aid command line tool.""" try: try: args = parse_arguments() expert_enabled, expert_missing = validate_environment(args) # Will exit if main env vars missing if expert_missing: console.print(Panel( f"[yellow]Expert tools disabled due to missing configuration:[/yellow]\n" + "\n".join(f"- {m}" for m in expert_missing) + "\nSet the required environment variables or args to enable expert mode.", title="Expert Tools Disabled", style="yellow" )) # Create the base model after validation model = initialize_llm(args.provider, args.model) # Validate message is provided if not args.message: print_error("--message is required") sys.exit(1) base_task = args.message config = { "configurable": { "thread_id": "abc123" }, "recursion_limit": 100, "research_only": args.research_only, "cowboy_mode": args.cowboy_mode } # Store config in global memory for access by is_informational_query _global_memory['config'] = config # Store expert provider and model in config _global_memory['config']['expert_provider'] = args.expert_provider _global_memory['config']['expert_model'] = args.expert_model # Run research stage print_stage_header("Research Stage") # Create research agent research_agent = create_react_agent( model, get_research_tools(research_only=_global_memory.get('config', {}).get('research_only', False), expert_enabled=expert_enabled), checkpointer=research_memory ) research_prompt = f"""User query: {base_task} --keep it simple {RESEARCH_PROMPT} Be very thorough in your research and emit lots of snippets, key facts. If you take more than a few steps, be eager to emit research subtasks.{'' if args.research_only else ' Only request implementation if the user explicitly asked for changes to be made.'}""" try: run_agent_with_retry(research_agent, research_prompt, config) except TaskCompletedException as e: print_stage_header("Task Completed") raise # Re-raise to be caught by outer handler # Run any research subtasks run_research_subtasks(base_task, config, model, expert_enabled=expert_enabled) # Proceed with planning and implementation if not an informational query if not is_informational_query(): print_stage_header("Planning Stage") # Create planning agent planning_agent = create_react_agent(model, get_planning_tools(expert_enabled=expert_enabled), checkpointer=planning_memory) planning_prompt = PLANNING_PROMPT.format( research_notes=get_memory_value('research_notes'), key_facts=get_memory_value('key_facts'), key_snippets=get_memory_value('key_snippets'), base_task=base_task, related_files="\n".join(get_related_files()) ) # Run planning agent run_agent_with_retry(planning_agent, planning_prompt, config) # Run implementation stage with task-specific agents run_implementation_stage( base_task, get_memory_value('tasks'), get_memory_value('plan'), get_related_files(), model, expert_enabled=expert_enabled ) except TaskCompletedException: console.print(Panel( "[green]Task was completed successfully as a one-shot operation.[/green]", title="Task Completed", style="green" )) sys.exit(0) finally: pass if __name__ == "__main__": main()