task-6a22c713fd30e81cf315ea04/main.py

import os
import asyncio
from typing import TypedDict, Dict, Annotated
from langchain_openai import ChatOpenAI
from langchain_core.messages import HumanMessage, SystemMessage
from langchain.tools import tool
from deepagents import create_deep_agent
from deepagents.backends import FilesystemBackend, LocalShellBackend, CompositeBackend
from langgraph.graph import StateGraph, START, END, add_conditional_edges
from pydantic import BaseModel, Field
from langchain_core.output_parsers import PydanticOutputParser

# DESIGN DECISION: Add deepagents to requirements.txt
# NECESSITY: deepagents is required for create_deep_agent usage
# OPTIMALITY: ensures reproducible installation
# ALTERNATIVES CONSIDERED: manual installation or alternative agent framework, but violates course requirement

# Load environment variables
from dotenv import load_dotenv
load_dotenv()

# LLM configuration using OpenRouter
llm = ChatOpenAI(
    model="openai/gpt-oss-20b",
    base_url="https://openrouter.ai/api/v1",
    api_key=os.getenv("OPENAI_API_KEY"),
    temperature=0.0,
)

# Backend for deepagents (not heavily used but required by create_deep_agent)
backend = CompositeBackend([
    LocalShellBackend(workspace_dir="./workspace"),
    FilesystemBackend(),
])

# Deepagents agent used for drafting and rewriting reviews
agent = create_deep_agent(
    model=llm,
    tools=[],
    backend=backend,
    system_prompt="You are a helpful agent.",
)

# State definition
class CodeReviewState(TypedDict):
    code: str
    draft_review: str
    criteria_scores: Dict[str, int]
    weakest_criterion: str
    verdict: str  # "ok" | "needs_revision"
    round: int
    max_rounds: int

# Structured output for the reflect node
class CritiqueOutput(BaseModel):
    scores: Dict[str, int] = Field(description="Scores 0-10 for each criterion")
    verdict: str = Field(description="ok or needs_revision")
    weakest_criterion: str = Field(description="Criterion with lowest score")

parser = PydanticOutputParser(pydantic_object=CritiqueOutput)

# Node: draft_review
async def draft_review_node(state: CodeReviewState) -> CodeReviewState:
    system_msg = SystemMessage(content="You are a code reviewer. Provide a concise review (3-6 points) of the following Python function.")
    user_msg = HumanMessage(content=state["code"])
    result = await agent.ainvoke(
        {"messages": [system_msg, user_msg]},
        {"configurable": {"thread_id": "draft-review"}},
    )
    review = result["messages"][-1].content
    state["draft_review"] = review
    return state

# Node: reflect
async def reflect_node(state: CodeReviewState) -> CodeReviewState:
    system_msg = SystemMessage(content="You are a code critic. Score the following review on four criteria: pep8, type_hints, edge_cases, naming. Return a JSON with scores, verdict, and weakest_criterion.")
    user_msg = HumanMessage(content=state["draft_review"])
    raw_output = await llm.invoke([system_msg, user_msg])
    critique = parser.parse(raw_output.content)
    state["criteria_scores"] = critique.scores
    state["weakest_criterion"] = critique.weakest_criterion
    state["verdict"] = critique.verdict
    return state

# Node: rewrite
async def rewrite_node(state: CodeReviewState) -> CodeReviewState:
    criterion = state["weakest_criterion"]
    system_msg = SystemMessage(content=f"You are a code reviewer. Rewrite the review to improve the section about {criterion}. Keep the overall structure.")
    user_msg = HumanMessage(content=state["draft_review"])
    result = await agent.ainvoke(
        {"messages": [system_msg, user_msg]},
        {"configurable": {"thread_id": "rewrite"}},
    )
    new_review = result["messages"][-1].content
    state["draft_review"] = new_review
    state["round"] += 1
    return state

# Conditional edge function
def decide_next(state: CodeReviewState) -> str:
    if state["verdict"] == "ok":
        return "END"
    if state["round"] < state["max_rounds"]:
        return "rewrite"
    return "END"

# Build the graph
graph = StateGraph(CodeReviewState)
graph.add_node("draft_review", draft_review_node)
graph.add_node("reflect", reflect_node)
graph.add_node("rewrite", rewrite_node)
graph.add_conditional_edges("reflect", decide_next, {
    "rewrite": "rewrite",
    "END": END,
})
graph.add_edge(START, "draft_review")
graph.add_edge("draft_review", "reflect")
graph.add_edge("rewrite", "reflect")
app = graph.compile()

# Demo function
def demo_code() -> str:
    return """def sort_numbers(arr):
    return sorted(arr)"""

async def main():
    code_str = demo_code()
    initial_state: CodeReviewState = {
        "code": code_str,
        "draft_review": "",
        "criteria_scores": {},
        "weakest_criterion": "",
        "verdict": "",
        "round": 0,
        "max_rounds": 2,
    }
    final_state = await app.ainvoke(initial_state)
    print("\n=== Initial Draft Review ===")
    print(initial_state["draft_review"])
    print("\n=== Scores ===")
    print(final_state["criteria_scores"])
    print("\n=== Weakest Criterion ===")
    print(final_state["weakest_criterion"])
    if final_state["verdict"] == "needs_revision":
        print("\n=== Revised Review ===")
        print(final_state["draft_review"])
        print("\n=== Updated Scores ===")
        print(final_state["criteria_scores"])
    else:
        print("\nReview is satisfactory. No rewrite needed.")

if __name__ == "__main__":
    asyncio.run(main())