fix(needs_fixes): 1 исправлений, 0 отстояно — main.py
This commit is contained in:
@@ -1,37 +1,15 @@
|
||||
"""
|
||||
# main.py
|
||||
# LangGraph code review agent with reflection and rewrite loop.
|
||||
# Implements the task specification without any deepagents dependency.
|
||||
# Uses OpenRouter via langchain-openai.
|
||||
|
||||
import os
|
||||
import asyncio
|
||||
from typing import TypedDict, Annotated, Dict
|
||||
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
|
||||
from dotenv import load_dotenv
|
||||
|
||||
load_dotenv()
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# 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
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# LLM setup
|
||||
# ---------------------------------------------------------------------------
|
||||
# ---------- LLM ----------
|
||||
llm = ChatOpenAI(
|
||||
model="openai/gpt-oss-20b:free",
|
||||
base_url="https://openrouter.ai/api/v1",
|
||||
@@ -39,134 +17,128 @@ llm = ChatOpenAI(
|
||||
temperature=0.0,
|
||||
)
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Structured output models for reflect node
|
||||
# ---------------------------------------------------------------------------
|
||||
class ReflectOutput(BaseModel):
|
||||
pep8: int = Field(..., ge=0, le=10)
|
||||
type_hints: int = Field(..., ge=0, le=10)
|
||||
edge_cases: int = Field(..., ge=0, le=10)
|
||||
naming: int = Field(..., ge=0, le=10)
|
||||
weakest_criterion: str = Field(...)
|
||||
verdict: str = Field(..., regex="^(ok|needs_revision)$")
|
||||
# ---------- 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
|
||||
|
||||
reflect_parser = PydanticOutputParser(pydantic_object=ReflectOutput)
|
||||
# ---------- 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'")
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Node implementations
|
||||
# ---------------------------------------------------------------------------
|
||||
async def draft_review_node(state: CodeReviewState) -> CodeReviewState:
|
||||
"""Generate an initial code review with 3–6 bullet points."""
|
||||
prompt = f"""
|
||||
You are a senior Python developer. You will write a concise code review for the following function. Provide 3 to 6 bullet points, each starting with a dash.
|
||||
parser = PydanticOutputParser(pydantic_object=ReflectionOutput)
|
||||
|
||||
Function code:
|
||||
{state['code']}
|
||||
# ---------- Nodes ----------
|
||||
|
||||
Review:"""
|
||||
response = await llm.ainvoke([HumanMessage(content=prompt)])
|
||||
state['draft_review'] = response.content.strip()
|
||||
return state
|
||||
|
||||
async def reflect_node(state: CodeReviewState) -> CodeReviewState:
|
||||
"""Critic evaluates the draft review on 4 criteria and returns structured scores."""
|
||||
prompt = f"""
|
||||
You are a code review critic. Evaluate the following draft review on the four criteria below, assigning a score from 0 (worst) to 10 (excellent). Return the scores and the weakest criterion in a JSON format matching the schema:
|
||||
|
||||
{reflect_parser.get_format_instructions()}
|
||||
|
||||
Draft review:
|
||||
{state['draft_review']}
|
||||
|
||||
Scores:"""
|
||||
response = await llm.ainvoke([HumanMessage(content=prompt)])
|
||||
try:
|
||||
parsed = reflect_parser.parse(response.content)
|
||||
except Exception as e:
|
||||
# Fallback: treat as all zeros
|
||||
parsed = ReflectOutput(pep8=0, type_hints=0, edge_cases=0, naming=0, weakest_criterion="pep8", verdict="needs_revision")
|
||||
state['criteria_scores'] = {
|
||||
"pep8": parsed.pep8,
|
||||
"type_hints": parsed.type_hints,
|
||||
"edge_cases": parsed.edge_cases,
|
||||
"naming": parsed.naming,
|
||||
}
|
||||
state['weakest_criterion'] = parsed.weakest_criterion
|
||||
state['verdict'] = parsed.verdict
|
||||
return state
|
||||
|
||||
async def rewrite_node(state: CodeReviewState) -> CodeReviewState:
|
||||
"""Rewrite the part of the review that addresses the weakest criterion."""
|
||||
prompt = f"""
|
||||
You are a senior Python developer. The following code review has been identified as weak in the criterion: {state['weakest_criterion']}. Rewrite only the section of the review that addresses this criterion, improving clarity and depth. Keep the rest of the review unchanged.
|
||||
|
||||
Original review:
|
||||
{state['draft_review']}
|
||||
|
||||
Rewritten review:"""
|
||||
response = await llm.ainvoke([HumanMessage(content=prompt)])
|
||||
# Replace only the weak section. For simplicity, we replace the whole review.
|
||||
state['draft_review'] = response.content.strip()
|
||||
state['round'] += 1
|
||||
return state
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Graph construction
|
||||
# ---------------------------------------------------------------------------
|
||||
def create_graph() -> StateGraph:
|
||||
graph = StateGraph(CodeReviewState)
|
||||
graph.add_node("draft_review", draft_review_node)
|
||||
graph.add_node("reflect", reflect_node)
|
||||
graph.add_node("rewrite", rewrite_node)
|
||||
|
||||
# Entry point
|
||||
graph.set_entry_point("draft_review")
|
||||
|
||||
# Transitions
|
||||
graph.add_edge("draft_review", "reflect")
|
||||
graph.add_conditional_edges(
|
||||
"reflect",
|
||||
lambda state: "rewrite" if state["verdict"] == "needs_revision" and state["round"] < state["max_rounds"] else "END",
|
||||
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
|
||||
)
|
||||
graph.add_edge("rewrite", "reflect")
|
||||
response = llm.invoke([HumanMessage(content=prompt)])
|
||||
state["draft_review"] = response.content
|
||||
return state
|
||||
|
||||
return graph
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# CLI helper
|
||||
# ---------------------------------------------------------------------------
|
||||
async def run_review(code: str, max_rounds: int = 2) -> CodeReviewState:
|
||||
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": code,
|
||||
"draft_review": "",
|
||||
"code": demo_code.strip(),
|
||||
"draft_review": "", # will be filled
|
||||
"criteria_scores": {},
|
||||
"weakest_criterion": "",
|
||||
"verdict": "",
|
||||
"round": 0,
|
||||
"max_rounds": max_rounds,
|
||||
"max_rounds": 2,
|
||||
}
|
||||
graph = create_graph()
|
||||
final_state = await graph.astate(initial_state)
|
||||
return final_state
|
||||
|
||||
# ---------------------------------------------------------------------------
|
||||
# Demo main
|
||||
# ---------------------------------------------------------------------------
|
||||
if __name__ == "__main__":
|
||||
sample_code = """
|
||||
def sort_numbers(arr):
|
||||
return sorted(arr)
|
||||
"""
|
||||
result = asyncio.run(run_review(sample_code))
|
||||
print("\n=== Initial Draft Review ===")
|
||||
result = await graph.ainvoke(initial_state)
|
||||
print("\n--- Final Review ---")
|
||||
print(result["draft_review"])
|
||||
print("\n=== Scores ===")
|
||||
print(result["criteria_scores"])
|
||||
print("\n=== Verdict ===")
|
||||
print(result["verdict"])
|
||||
if result["verdict"] == "needs_revision":
|
||||
print("\n=== Rewritten Review ===")
|
||||
print(result["draft_review"]) # after last rewrite
|
||||
print("\n=== Updated Scores ===")
|
||||
print(result["criteria_scores"])
|
||||
""
|
||||
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())
|
||||
|
||||
Reference in New Issue
Block a user