Add main.py implementing LangGraph code review agent
This commit is contained in:
@@ -1,105 +1,135 @@
|
||||
import asyncio
|
||||
import os
|
||||
import asyncio
|
||||
from typing import TypedDict, Dict
|
||||
|
||||
from langchain_openai import ChatOpenAI
|
||||
from langchain_core.messages import HumanMessage
|
||||
from langchain.tools import tool
|
||||
from langgraph.graph import StateGraph, START, END
|
||||
from typing import TypedDict, Annotated
|
||||
from langgraph.graph.message import add_messages
|
||||
from pydantic import BaseModel, Field
|
||||
from langchain_core.output_parsers import PydanticOutputParser
|
||||
|
||||
# LLM setup
|
||||
llm = ChatOpenAI(
|
||||
model="openai/gpt-4o-mini",
|
||||
model="openai/gpt-oss-20b:free",
|
||||
base_url="https://openrouter.ai/api/v1",
|
||||
api_key=os.getenv("OPENAI_API_KEY"),
|
||||
temperature=0.0,
|
||||
)
|
||||
|
||||
# Pydantic models for reflection
|
||||
class CriteriaScores(BaseModel):
|
||||
pep8: int
|
||||
type_hints: int
|
||||
edge_cases: int
|
||||
naming: int
|
||||
|
||||
class ReviewResult(BaseModel):
|
||||
scores: CriteriaScores
|
||||
weakest: str
|
||||
verdict: str
|
||||
|
||||
# Tool to run shell commands
|
||||
@tool
|
||||
def run_command(command: str) -> str:
|
||||
"""Execute a shell command and return its output."""
|
||||
try:
|
||||
result = os.popen(command).read()
|
||||
return result.strip() or "(no output)"
|
||||
except Exception as e:
|
||||
return f"Error: {e}"
|
||||
|
||||
# Draft review node
|
||||
async def draft_review(state: dict):
|
||||
code = state["code"]
|
||||
prompt = f"Write a concise code review for the following Python function. Provide 3-6 bullet points highlighting strengths and areas for improvement.\n\n{code}"
|
||||
response = llm.invoke([HumanMessage(content=prompt)])
|
||||
state["draft_review"] = response.content
|
||||
return state
|
||||
|
||||
# Reflect node
|
||||
async def reflect(state: dict):
|
||||
review = state["draft_review"]
|
||||
prompt = f"Score the following code review on 4 criteria: PEP8, type hints, edge cases, naming. Return JSON with keys pep8, type_hints, edge_cases, naming (0-10). Also provide the weakest criterion and verdict ('ok' if all >=7 else 'needs_revision').\n\n{review}"
|
||||
response = llm.invoke([HumanMessage(content=prompt)])
|
||||
try:
|
||||
data = ReviewResult.parse_raw(response.content)
|
||||
except Exception:
|
||||
# fallback simple parse
|
||||
data = ReviewResult(scores=CriteriaScores(pep8=5,type_hints=5,edge_cases=5,naming=5),weakest="pep8",verdict="needs_revision")
|
||||
state["criteria_scores"] = data.scores.dict()
|
||||
state["weakest_criterion"] = data.weakest
|
||||
state["verdict"] = data.verdict
|
||||
return state
|
||||
|
||||
# Rewrite node
|
||||
async def rewrite(state: dict):
|
||||
weakest = state["weakest_criterion"]
|
||||
review = state["draft_review"]
|
||||
prompt = f"Improve the code review focusing on the {weakest} aspect. Keep the rest unchanged.\n\n{review}"
|
||||
response = llm.invoke([HumanMessage(content=prompt)])
|
||||
state["draft_review"] = response.content
|
||||
state["round"] += 1
|
||||
return state
|
||||
|
||||
# Graph definition
|
||||
# State definition
|
||||
class CodeReviewState(TypedDict):
|
||||
code: str
|
||||
draft_review: str
|
||||
criteria_scores: dict
|
||||
criteria_scores: Dict[str, int]
|
||||
weakest_criterion: str
|
||||
verdict: str
|
||||
round: int
|
||||
max_rounds: int
|
||||
|
||||
workflow = StateGraph(CodeReviewState)
|
||||
workflow.add_node("draft", draft_review)
|
||||
workflow.add_node("reflect", reflect)
|
||||
workflow.add_node("rewrite", rewrite)
|
||||
workflow.add_conditional_edges(START, lambda _: "draft")
|
||||
workflow.add_conditional_edges("draft", lambda _: "reflect")
|
||||
workflow.add_conditional_edges("reflect", lambda s: "rewrite" if s["verdict"]=="needs_revision" and s["round"]<s["max_rounds"] else "END")
|
||||
workflow.add_conditional_edges("rewrite", lambda _: "reflect")
|
||||
graph = workflow.compile()
|
||||
# Reflect output model
|
||||
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 coverage")
|
||||
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'")
|
||||
|
||||
reflect_parser = PydanticOutputParser(pydantic_object=ReflectOutput)
|
||||
|
||||
# Dummy tool for agent
|
||||
@tool
|
||||
def echo_tool(query: str) -> str:
|
||||
return query
|
||||
|
||||
# Agent creation
|
||||
from deepagents import create_deep_agent
|
||||
from deepagents.backends import FilesystemBackend, LocalShellBackend, CompositeBackend
|
||||
backend = CompositeBackend(
|
||||
default=LocalShellBackend(root_dir="./workspace", virtual_mode=True, inherit_env=True),
|
||||
routes={},
|
||||
)
|
||||
agent = create_deep_agent(
|
||||
model=llm,
|
||||
tools=[echo_tool],
|
||||
backend=backend,
|
||||
system_prompt="You are a code review assistant.",
|
||||
)
|
||||
|
||||
# Node functions
|
||||
async def draft_review(state: CodeReviewState) -> CodeReviewState:
|
||||
prompt = f"""Write a concise code review (3-6 points) for the following Python function. Focus on style, correctness, and potential improvements.
|
||||
|
||||
```python
|
||||
{state['code']}
|
||||
```
|
||||
|
||||
Return only the review text."""
|
||||
response = await agent.ainvoke({"messages": [HumanMessage(content=prompt)]}, {"configurable": {"thread_id": "draft-review"}})
|
||||
review_text = response["messages"][-1].content
|
||||
state["draft_review"] = review_text
|
||||
return state
|
||||
|
||||
async def reflect(state: CodeReviewState) -> CodeReviewState:
|
||||
prompt = f"""You are a senior reviewer. Evaluate the following review text against four criteria: PEP8, type hints, edge cases, naming. Assign each a score 0-10. Identify the weakest criterion and give a verdict: 'ok' if all scores >=7, else 'needs_revision'. Return a JSON with keys: pep8, type_hints, edge_cases, naming, weakest_criterion, verdict.
|
||||
|
||||
Review:
|
||||
{state['draft_review']}"""
|
||||
response = await agent.ainvoke({"messages": [HumanMessage(content=prompt)]}, {"configurable": {"thread_id": "reflect"}})
|
||||
json_text = response["messages"][-1].content
|
||||
try:
|
||||
parsed = reflect_parser.parse(json_text)
|
||||
except Exception:
|
||||
parsed = ReflectOutput(pep8=5, type_hints=5, edge_cases=5, naming=5, 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(state: CodeReviewState) -> CodeReviewState:
|
||||
crit = state["weakest_criterion"]
|
||||
prompt = f"""Improve the review section that addresses the weakest criterion '{crit}'. Provide a more detailed point for that criterion. Keep the rest of the review unchanged.
|
||||
|
||||
Current review:
|
||||
{state['draft_review']}"""
|
||||
response = await agent.ainvoke({"messages": [HumanMessage(content=prompt)]}, {"configurable": {"thread_id": "rewrite"}})
|
||||
new_review = response["messages"][-1].content
|
||||
state["draft_review"] = new_review
|
||||
state["round"] += 1
|
||||
return state
|
||||
|
||||
# Graph definition
|
||||
from langgraph.graph import StateGraph
|
||||
|
||||
graph = StateGraph(CodeReviewState)
|
||||
|
||||
graph.add_node("draft_review", draft_review)
|
||||
graph.add_node("reflect", reflect)
|
||||
graph.add_node("rewrite", rewrite)
|
||||
|
||||
graph.set_entry_point("draft_review")
|
||||
graph.add_edge("draft_review", "reflect")
|
||||
|
||||
graph.add_conditional_edges(
|
||||
"reflect",
|
||||
lambda x: "END" if x["verdict"] == "ok" else "rewrite" if x["round"] < x["max_rounds"] else "END",
|
||||
)
|
||||
|
||||
graph.add_edge("rewrite", "reflect")
|
||||
app = graph.compile()
|
||||
|
||||
# Demo function
|
||||
async def run_demo():
|
||||
code = """
|
||||
def sort_numbers(arr):
|
||||
return sorted(arr)
|
||||
"""
|
||||
state: CodeReviewState = {
|
||||
"code": code,
|
||||
async def demo():
|
||||
sample_code = """def sort_numbers(arr):
|
||||
return sorted(arr)"""
|
||||
init_state: CodeReviewState = {
|
||||
"code": sample_code,
|
||||
"draft_review": "",
|
||||
"criteria_scores": {},
|
||||
"weakest_criterion": "",
|
||||
@@ -107,9 +137,18 @@ def sort_numbers(arr):
|
||||
"round": 0,
|
||||
"max_rounds": 2,
|
||||
}
|
||||
final = await graph.ainvoke(state)
|
||||
print("Final review:\n", final["draft_review"])
|
||||
print("Scores:", final["criteria_scores"])
|
||||
result = await app.ainvoke(init_state)
|
||||
print("--- Draft 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--- Final Review After Rewrite ---")
|
||||
print(result["draft_review"])
|
||||
print("\n--- Final Scores ---")
|
||||
print(result["criteria_scores"])
|
||||
|
||||
if __name__ == "__main__":
|
||||
asyncio.run(run_demo())
|
||||
asyncio.run(demo())
|
||||
|
||||
Reference in New Issue
Block a user