mjschock's picture
Refactor agent.py and graph.py to enhance agent functionality and logging. Introduce Configuration class for managing parameters, improve state handling in AgentRunner, and update agent graph to support step logging and user interaction. Add new tests for agent capabilities and update requirements for code formatting tools.
401799d unverified
raw
history blame
5.89 kB
"""Define the agent graph and its components."""
import logging
import os
import uuid
from typing import Dict, List, Optional, TypedDict, Union, cast
import yaml
from langchain_core.language_models import BaseChatModel
from langchain_core.messages import AIMessage, HumanMessage, SystemMessage
from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.runnables import RunnableConfig
from langgraph.graph import END, StateGraph
from langgraph.prebuilt import ToolExecutor, ToolNode
from langgraph.types import interrupt
from smolagents import CodeAgent, LiteLLMModel, ToolCallingAgent
from configuration import Configuration
from tools import tools
# Configure logging
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
# Enable LiteLLM debug logging only if environment variable is set
import litellm
if os.getenv("LITELLM_DEBUG", "false").lower() == "true":
litellm.set_verbose = True
logger.setLevel(logging.DEBUG)
else:
litellm.set_verbose = False
logger.setLevel(logging.INFO)
# Configure LiteLLM to drop unsupported parameters
litellm.drop_params = True
# Load default prompt templates from local file
current_dir = os.path.dirname(os.path.abspath(__file__))
prompts_dir = os.path.join(current_dir, "prompts")
yaml_path = os.path.join(prompts_dir, "code_agent.yaml")
with open(yaml_path, "r") as f:
prompt_templates = yaml.safe_load(f)
# Initialize the model and agent using configuration
config = Configuration()
model = LiteLLMModel(
api_base=config.api_base,
api_key=config.api_key,
model_id=config.model_id,
)
agent = CodeAgent(
add_base_tools=True,
max_steps=1, # Execute one step at a time
model=model,
prompt_templates=prompt_templates,
tools=tools,
verbosity_level=logging.DEBUG,
)
class AgentState(TypedDict):
"""State for the agent graph."""
messages: List[Union[HumanMessage, AIMessage, SystemMessage]]
question: str
answer: Optional[str]
step_logs: List[Dict]
is_complete: bool
step_count: int
class AgentNode:
"""Node that runs the agent."""
def __init__(self, agent: CodeAgent):
"""Initialize the agent node with an agent."""
self.agent = agent
def __call__(
self, state: AgentState, config: Optional[RunnableConfig] = None
) -> AgentState:
"""Run the agent on the current state."""
# Log current state
logger.info("Current state before processing:")
logger.info(f"Messages: {state['messages']}")
logger.info(f"Question: {state['question']}")
logger.info(f"Answer: {state['answer']}")
# Get configuration
cfg = Configuration.from_runnable_config(config)
logger.info(f"Using configuration: {cfg}")
# Log execution start
logger.info("Starting agent execution")
# Run the agent
result = self.agent.run(state["question"])
# Log result
logger.info(f"Agent execution result type: {type(result)}")
logger.info(f"Agent execution result value: {result}")
# Update state
new_state = state.copy()
new_state["messages"].append(AIMessage(content=result))
new_state["answer"] = result
new_state["step_count"] += 1
# Log updated state
logger.info("Updated state after processing:")
logger.info(f"Messages: {new_state['messages']}")
logger.info(f"Question: {new_state['question']}")
logger.info(f"Answer: {new_state['answer']}")
return new_state
class StepCallbackNode:
"""Node that handles step callbacks and user interaction."""
def __call__(
self, state: AgentState, config: Optional[RunnableConfig] = None
) -> AgentState:
"""Handle step callback and user interaction."""
# Get configuration
cfg = Configuration.from_runnable_config(config)
# Log the step
step_log = {
"step": state["step_count"],
"messages": [msg.content for msg in state["messages"]],
"question": state["question"],
"answer": state["answer"],
}
state["step_logs"].append(step_log)
try:
# Use interrupt for user input
user_input = interrupt(
"Press 'c' to continue, 'q' to quit, or 'i' for more info: "
)
if user_input.lower() == "q":
state["is_complete"] = True
return state
elif user_input.lower() == "i":
logger.info(f"Current step: {state['step_count']}")
logger.info(f"Question: {state['question']}")
logger.info(f"Current answer: {state['answer']}")
return self(state, config) # Recursively call for new input
elif user_input.lower() == "c":
return state
else:
logger.warning("Invalid input. Please use 'c', 'q', or 'i'.")
return self(state, config) # Recursively call for new input
except Exception as e:
logger.warning(f"Error during interrupt: {str(e)}")
return state
def build_agent_graph(agent: AgentNode) -> StateGraph:
"""Build the agent graph."""
# Initialize the graph
workflow = StateGraph(AgentState)
# Add nodes
workflow.add_node("agent", agent)
workflow.add_node("callback", StepCallbackNode())
# Add edges
workflow.add_edge("agent", "callback")
workflow.add_conditional_edges(
"callback",
lambda x: END if x["is_complete"] else "agent",
{True: END, False: "agent"},
)
# Set entry point
workflow.set_entry_point("agent")
return workflow.compile()
# Initialize the agent graph
agent_graph = build_agent_graph(AgentNode(agent))