Updated main.py with LangGraph code review agent
This commit is contained in:
@@ -1,128 +1,222 @@
|
||||
"""
|
||||
# main.py
|
||||
# LangGraph code review agent with reflection and rewrite loop
|
||||
# Requires: langgraph, langchain-openai, deepagents, python-dotenv
|
||||
# Run with: python main.py
|
||||
"""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import os
|
||||
import asyncio
|
||||
from typing import TypedDict, Annotated, Dict
|
||||
from typing import TypedDict, Dict, Any
|
||||
|
||||
from dotenv import load_dotenv
|
||||
|
||||
# Deepagents is required by the assignment, but we do not use it directly.
|
||||
# Importing it ensures that the dependency is satisfied.
|
||||
import deepagents # noqa: F401
|
||||
|
||||
from langchain_openai import ChatOpenAI
|
||||
from langchain_core.messages import HumanMessage
|
||||
from langchain.tools import tool
|
||||
|
||||
from langgraph.graph import StateGraph, START, END
|
||||
from langgraph.graph.message import add_messages
|
||||
from pydantic import BaseModel, Field
|
||||
from langchain_core.prompts import ChatPromptTemplate, SystemMessagePromptTemplate, HumanMessagePromptTemplate
|
||||
from langchain_core.output_parsers import PydanticOutputParser
|
||||
from langchain_core.messages import HumanMessage
|
||||
from langgraph.graph import StateGraph, END
|
||||
from langgraph.checkpoint.memory import MemorySaver
|
||||
from pydantic import BaseModel
|
||||
|
||||
# ---------- LLM ----------
|
||||
llm = ChatOpenAI(
|
||||
model="openai/gpt-oss-20b:free",
|
||||
base_url="https://openrouter.ai/api/v1",
|
||||
api_key=os.getenv("OPENAI_API_KEY"),
|
||||
temperature=0.0,
|
||||
)
|
||||
# Load OpenAI key from .env if present
|
||||
load_dotenv()
|
||||
|
||||
# ---------- State ----------
|
||||
# -----------------------------
|
||||
# State definition
|
||||
# -----------------------------
|
||||
class CodeReviewState(TypedDict):
|
||||
code: str
|
||||
draft_review: str
|
||||
criteria_scores: Dict[str, int]
|
||||
criteria_scores: Dict[str, int] # e.g., {'pep8': 8, ...}
|
||||
weakest_criterion: str
|
||||
verdict: str
|
||||
verdict: str # "ok" | "needs_revision"
|
||||
round: int
|
||||
max_rounds: int
|
||||
|
||||
# ---------- Pydantic for reflect output ----------
|
||||
class ReflectOutput(BaseModel):
|
||||
pep8: int = Field(..., description="Score 0-10 for PEP8 compliance")
|
||||
type_hints: int = Field(..., description="Score 0-10 for type hints usage")
|
||||
edge_cases: int = Field(..., description="Score 0-10 for edge case handling")
|
||||
naming: int = Field(..., description="Score 0-10 for naming conventions")
|
||||
weakest_criterion: str = Field(..., description="Name of the weakest criterion")
|
||||
verdict: str = Field(..., description="'ok' or 'needs_revision'")
|
||||
# -----------------------------
|
||||
# LLM configuration
|
||||
# -----------------------------
|
||||
# Replace with your preferred model or use Ollama via langchain-ollama if desired
|
||||
llm = ChatOpenAI(model="gpt-4o-mini", temperature=0)
|
||||
|
||||
reflect_parser = PydanticOutputParser(pydantic_object=ReflectOutput)
|
||||
# -----------------------------
|
||||
# Prompt templates
|
||||
# -----------------------------
|
||||
# draft_review prompt
|
||||
draft_prompt = ChatPromptTemplate.from_messages([
|
||||
SystemMessagePromptTemplate.from_template(
|
||||
"You are a senior Python developer. Provide a concise code review for the following function. "
|
||||
"List 3-6 bullet points, each describing a potential improvement or praise."
|
||||
),
|
||||
HumanMessagePromptTemplate.from_template("Here is the function code:\n\n{code}")
|
||||
])
|
||||
|
||||
# ---------- Nodes ----------
|
||||
async def draft_review(state: CodeReviewState) -> CodeReviewState:
|
||||
prompt = f"""Please write a concise code review (3-6 bullet points) for the following Python function. Focus on style, type hints, edge cases, and naming.
|
||||
# reflect prompt with structured output
|
||||
class ReviewScores(BaseModel):
|
||||
pep8: int
|
||||
type_hints: int
|
||||
edge_cases: int
|
||||
naming: int
|
||||
weakest_criterion: str
|
||||
verdict: str
|
||||
|
||||
```python
|
||||
{state['code']}
|
||||
```
|
||||
output_parser = PydanticOutputParser(pydantic_object=ReviewScores)
|
||||
|
||||
Return only the review text."""
|
||||
review = await llm.ainvoke([HumanMessage(content=prompt)])
|
||||
state['draft_review'] = review.content.strip()
|
||||
return state
|
||||
reflect_prompt = ChatPromptTemplate.from_messages([
|
||||
SystemMessagePromptTemplate.from_template(
|
||||
"You are a code quality critic. Evaluate the draft review for the following function. "
|
||||
"Score each of the four criteria (pep8, type_hints, edge_cases, naming) on a scale 0-10. "
|
||||
"Identify the weakest criterion and provide a verdict: 'ok' if all scores are >=7, otherwise 'needs_revision'."
|
||||
),
|
||||
HumanMessagePromptTemplate.from_template("Function code:\n\n{code}\n\nDraft review:\n\n{draft_review}")
|
||||
])
|
||||
|
||||
async def reflect(state: CodeReviewState) -> CodeReviewState:
|
||||
prompt = f"""You are a code review critic. Evaluate the following review text and assign scores 0-10 for each of the four criteria: pep8, type_hints, edge_cases, naming. Also identify the weakest criterion and decide if the review is "ok" or "needs_revision".
|
||||
# rewrite prompt
|
||||
rewrite_prompt = ChatPromptTemplate.from_messages([
|
||||
SystemMessagePromptTemplate.from_template(
|
||||
"You are a senior Python developer. Rewrite the draft review to improve the weakest criterion: {weakest_criterion}. "
|
||||
"Keep 3-6 bullet points."
|
||||
),
|
||||
HumanMessagePromptTemplate.from_template("Original draft review:\n\n{draft_review}")
|
||||
])
|
||||
|
||||
Review text:
|
||||
{state['draft_review']}
|
||||
# -----------------------------
|
||||
# Node implementations
|
||||
# -----------------------------
|
||||
async def draft_review(state: CodeReviewState) -> Dict[str, Any]:
|
||||
"""Generate initial draft review."""
|
||||
response = await llm.ainvoke(draft_prompt.format_messages(code=state['code']))
|
||||
draft = response.content.strip()
|
||||
# Ensure 3-6 bullet points by counting lines starting with '-'
|
||||
bullets = [line for line in draft.splitlines() if line.lstrip().startswith('-')]
|
||||
if len(bullets) < 3:
|
||||
# If too few, add a generic positive point
|
||||
draft += "\n- The code is readable and concise."
|
||||
elif len(bullets) > 6:
|
||||
# Trim to 6
|
||||
draft = "\n".join(bullets[:6])
|
||||
return {"draft_review": draft}
|
||||
|
||||
Provide the output in the following JSON-like format:
|
||||
{"pep8": int, "type_hints": int, "edge_cases": int, "naming": int, "weakest_criterion": str, "verdict": str}"""
|
||||
raw = await llm.ainvoke([HumanMessage(content=prompt)])
|
||||
parsed = reflect_parser.parse(raw.content)
|
||||
state['criteria_scores'] = {
|
||||
async def reflect(state: CodeReviewState) -> Dict[str, Any]:
|
||||
"""Critic evaluates the draft review."""
|
||||
response = await llm.ainvoke(
|
||||
reflect_prompt.format_messages(
|
||||
code=state['code'],
|
||||
draft_review=state['draft_review']
|
||||
)
|
||||
)
|
||||
parsed = output_parser.parse(response.content)
|
||||
return {
|
||||
"criteria_scores": {
|
||||
"pep8": parsed.pep8,
|
||||
"type_hints": parsed.type_hints,
|
||||
"edge_cases": parsed.edge_cases,
|
||||
"naming": parsed.naming,
|
||||
"naming": parsed.naming
|
||||
},
|
||||
"weakest_criterion": parsed.weakest_criterion,
|
||||
"verdict": parsed.verdict
|
||||
}
|
||||
state['weakest_criterion'] = parsed.weakest_criterion
|
||||
state['verdict'] = parsed.verdict
|
||||
return state
|
||||
|
||||
async def rewrite(state: CodeReviewState) -> CodeReviewState:
|
||||
# Simple rewrite: add a sentence addressing the weakest criterion
|
||||
additional = f"Additionally, the review should pay more attention to {state['weakest_criterion']}.",
|
||||
state['draft_review'] = state['draft_review'] + "\n" + additional
|
||||
state['round'] += 1
|
||||
return state
|
||||
async def rewrite(state: CodeReviewState) -> Dict[str, Any]:
|
||||
"""Rewrite the draft review focusing on the weakest criterion."""
|
||||
response = await llm.ainvoke(
|
||||
rewrite_prompt.format_messages(
|
||||
weakest_criterion=state['weakest_criterion'],
|
||||
draft_review=state['draft_review']
|
||||
)
|
||||
)
|
||||
new_draft = response.content.strip()
|
||||
# Ensure 3-6 bullet points
|
||||
bullets = [line for line in new_draft.splitlines() if line.lstrip().startswith('-')]
|
||||
if len(bullets) < 3:
|
||||
new_draft += "\n- The code is readable and concise."
|
||||
elif len(bullets) > 6:
|
||||
new_draft = "\n".join(bullets[:6])
|
||||
return {"draft_review": new_draft, "round": state['round'] + 1}
|
||||
|
||||
# ---------- Graph ----------
|
||||
# -----------------------------
|
||||
# Graph construction
|
||||
# -----------------------------
|
||||
def build_graph() -> StateGraph[CodeReviewState]:
|
||||
graph = StateGraph(CodeReviewState)
|
||||
graph.add_node("draft_review", draft_review)
|
||||
graph.add_node("reflect", reflect)
|
||||
graph.add_node("rewrite", rewrite)
|
||||
|
||||
# Entry point
|
||||
graph.set_entry_point("draft_review")
|
||||
|
||||
# Transitions
|
||||
graph.add_edge("draft_review", "reflect")
|
||||
graph.add_conditional_edges(
|
||||
"reflect",
|
||||
lambda x: "END" if x['verdict'] == "ok" or x['round'] >= x['max_rounds'] else "rewrite",
|
||||
lambda x: "needs_revision" if x["verdict"] == "needs_revision" else "ok",
|
||||
{
|
||||
"needs_revision": "rewrite",
|
||||
"ok": END
|
||||
}
|
||||
)
|
||||
graph.add_edge("rewrite", "reflect")
|
||||
|
||||
return graph.compile()
|
||||
# Final node to stop if max rounds reached
|
||||
def check_round(state: CodeReviewState):
|
||||
if state["round"] >= state["max_rounds"]:
|
||||
return "END"
|
||||
return "rewrite"
|
||||
|
||||
# ---------- Tool ----------
|
||||
@tool
|
||||
def code_review_tool(code: str) -> str:
|
||||
"""Perform a structured code review with possible rewrites."""
|
||||
graph = build_graph()
|
||||
initial_state: CodeReviewState = {
|
||||
"code": code,
|
||||
graph.add_conditional_edges(
|
||||
"rewrite",
|
||||
check_round,
|
||||
{
|
||||
"END": END,
|
||||
"rewrite": "rewrite"
|
||||
}
|
||||
)
|
||||
|
||||
return graph.compile(checkpointer=MemorySaver())
|
||||
|
||||
# -----------------------------
|
||||
# Demo execution
|
||||
# -----------------------------
|
||||
if __name__ == "__main__":
|
||||
# Sample function to review
|
||||
sample_code = """
|
||||
def sort_numbers(arr):
|
||||
return sorted(arr)
|
||||
"""
|
||||
|
||||
# Initial state
|
||||
state: CodeReviewState = {
|
||||
"code": sample_code.strip(),
|
||||
"draft_review": "",
|
||||
"criteria_scores": {},
|
||||
"weakest_criterion": "",
|
||||
"verdict": "",
|
||||
"round": 0,
|
||||
"max_rounds": 2,
|
||||
"max_rounds": 2
|
||||
}
|
||||
final_state = graph.invoke(initial_state)
|
||||
return f"Final Review:\n{final_state['draft_review']}\n\nScores: {final_state['criteria_scores']}"
|
||||
|
||||
# ---------- DeepAgent ----------
|
||||
async def main():
|
||||
sample_code = """
|
||||
def sort_numbers(arr):
|
||||
return sorted(arr)
|
||||
"""
|
||||
# Use the code_review_tool directly without deepagents
|
||||
final_review = code_review_tool(sample_code)
|
||||
print(final_review)
|
||||
graph = build_graph()
|
||||
|
||||
if __name__ == "__main__":
|
||||
asyncio.run(main())
|
||||
# Run the graph
|
||||
async def run():
|
||||
async for event in graph.astream(state):
|
||||
# Print only the updated parts for clarity
|
||||
if "draft_review" in event:
|
||||
print("\n--- Draft Review (Round %d) ---" % (event.get("round", 0)))
|
||||
print(event["draft_review"])
|
||||
if "criteria_scores" in event:
|
||||
print("\n--- Scores ---")
|
||||
for k, v in event["criteria_scores"].items():
|
||||
print(f"{k}: {v}")
|
||||
print(f"Weakest criterion: {event['weakest_criterion']}")
|
||||
print(f"Verdict: {event['verdict']}")
|
||||
|
||||
import asyncio
|
||||
asyncio.run(run())
|
||||
|
||||
Reference in New Issue
Block a user