145 lines
5.1 KiB
Python
145 lines
5.1 KiB
Python
import os
|
|
import asyncio
|
|
from typing import TypedDict, Annotated
|
|
|
|
from langchain_openai import ChatOpenAI
|
|
from langchain_core.messages import HumanMessage, AIMessage
|
|
from langchain_core.output_parsers import PydanticOutputParser
|
|
from langchain_core.pydantic_v1 import BaseModel, Field
|
|
from langgraph.graph import StateGraph, START, END
|
|
from langgraph.graph.message import add_messages
|
|
|
|
# ---------- 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,
|
|
)
|
|
|
|
# ---------- State ----------
|
|
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 reflect ----------
|
|
class ReflectionOutput(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 handling edge cases")
|
|
naming: int = Field(description="Score 0-10 for naming conventions")
|
|
weakest_criterion: str = Field(description="Criterion with lowest score")
|
|
verdict: str = Field(description="'ok' or 'needs_revision'")
|
|
|
|
parser = PydanticOutputParser(pydantic_object=ReflectionOutput)
|
|
|
|
# ---------- Nodes ----------
|
|
|
|
def draft_review_node(state: CodeReviewState) -> CodeReviewState:
|
|
code = state["code"]
|
|
prompt = (
|
|
"You are a senior Python developer.\n"
|
|
"Given the following function, write a concise code review (3-6 bullet points).\n"
|
|
"Focus on style, correctness, edge cases, and naming.\n"
|
|
f"Function:\n{code}\n\nReview:" # LLM will output review
|
|
)
|
|
response = llm.invoke([HumanMessage(content=prompt)])
|
|
state["draft_review"] = response.content
|
|
return state
|
|
|
|
|
|
def reflect_node(state: CodeReviewState) -> CodeReviewState:
|
|
review = state["draft_review"]
|
|
code = state["code"]
|
|
prompt = (
|
|
"You are an automated code review critic.\n"
|
|
"Given the code and its review, assign a score 0-10 for each of the following criteria:\n"
|
|
"- pep8: PEP8 compliance\n"
|
|
"- type_hints: use of type hints\n"
|
|
"- edge_cases: handling of edge cases\n"
|
|
"- naming: clarity of names\n"
|
|
"Return the scores, the weakest criterion, and a verdict ('ok' if all scores >=7, else 'needs_revision').\n"
|
|
f"Code:\n{code}\n\nReview:\n{review}\n\nOutput in JSON with fields: pep8, type_hints, edge_cases, naming, weakest_criterion, verdict."
|
|
)
|
|
response = llm.invoke([HumanMessage(content=prompt)])
|
|
try:
|
|
out = parser.parse(response.content)
|
|
except Exception as e:
|
|
# Fallback: simple parsing if JSON is not strict
|
|
import json
|
|
out = json.loads(response.content)
|
|
state["criteria_scores"] = {
|
|
"pep8": out.pep8,
|
|
"type_hints": out.type_hints,
|
|
"edge_cases": out.edge_cases,
|
|
"naming": out.naming,
|
|
}
|
|
state["weakest_criterion"] = out.weakest_criterion
|
|
state["verdict"] = out.verdict
|
|
return state
|
|
|
|
|
|
def rewrite_node(state: CodeReviewState) -> CodeReviewState:
|
|
weakest = state["weakest_criterion"]
|
|
review = state["draft_review"]
|
|
code = state["code"]
|
|
prompt = (
|
|
"You are a senior Python developer tasked with improving a code review.\n"
|
|
f"The current review is:\n{review}\n\nThe weakest criterion is '{weakest}'.\n"
|
|
"Rewrite only the part of the review that addresses this criterion, making it stronger and more specific.\n"
|
|
"Keep the rest of the review unchanged.\n"
|
|
"Output only the updated review."
|
|
)
|
|
response = llm.invoke([HumanMessage(content=prompt)])
|
|
state["draft_review"] = response.content
|
|
state["round"] = state.get("round", 0) + 1
|
|
return state
|
|
|
|
# ---------- Graph ----------
|
|
builder = StateGraph(CodeReviewState)
|
|
builder.add_node("draft_review", draft_review_node)
|
|
builder.add_node("reflect", reflect_node)
|
|
builder.add_node("rewrite", rewrite_node)
|
|
|
|
builder.set_entry_point("draft_review")
|
|
builder.add_edge("draft_review", "reflect")
|
|
builder.add_conditional_edges(
|
|
"reflect",
|
|
lambda x: "rewrite" if x["verdict"] == "needs_revision" and x["round"] < x["max_rounds"] else "END",
|
|
)
|
|
builder.add_edge("rewrite", "reflect")
|
|
builder.add_edge("END", END)
|
|
|
|
graph = builder.compile()
|
|
|
|
# ---------- Demo ----------
|
|
async def main():
|
|
demo_code = """
|
|
def sort_numbers(arr):
|
|
return sorted(arr)
|
|
"""
|
|
initial_state: CodeReviewState = {
|
|
"code": demo_code.strip(),
|
|
"draft_review": "", # will be filled
|
|
"criteria_scores": {},
|
|
"weakest_criterion": "",
|
|
"verdict": "",
|
|
"round": 0,
|
|
"max_rounds": 2,
|
|
}
|
|
result = await graph.ainvoke(initial_state)
|
|
print("\n--- Final Review ---")
|
|
print(result["draft_review"])
|
|
print("\n--- Scores ---")
|
|
for k, v in result["criteria_scores"].items():
|
|
print(f"{k}: {v}")
|
|
print(f"Verdict: {result['verdict']}")
|
|
|
|
if __name__ == "__main__":
|
|
asyncio.run(main())
|