add: main.py

main.py

@@ -0,0 +1,203 @@
+"""Code Review Agent with LangGraph and DeepAgents
+
+This script implements the assignment described in the prompt.  It uses
+* LangGraph to model the review cycle (draft → reflect → rewrite → reflect …)
+* DeepAgents to expose the whole workflow as a single LLM‑driven agent.
+* OpenRouter via langchain‑openai for all LLM calls.
+
+Run the demo with:
+
+```bash
+python main.py
+```
+
+The demo reviews a simple `sort_numbers` function and prints the draft review,
+the critic’s scores, and any rewritten sections.
+"""
+
+import os
+import asyncio
+from typing import TypedDict, Annotated, Dict
+
+from langchain_openai import ChatOpenAI
+from langchain_core.messages import HumanMessage
+from langchain.tools import tool
+from langchain.output_parsers import PydanticOutputParser
+from pydantic import BaseModel, Field
+
+from langgraph.graph import StateGraph, START, END
+from langgraph.graph.message import add_messages
+
+from deepagents import create_deep_agent
+from deepagents.backends import FilesystemBackend, LocalShellBackend, CompositeBackend
+
+# ---------------------------------------------------------------------------
+# 1. LLM configuration (OpenRouter)
+# ---------------------------------------------------------------------------
+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,
+)
+
+# ---------------------------------------------------------------------------
+# 2. State definition
+# ---------------------------------------------------------------------------
+class CodeReviewState(TypedDict):
+    code: str
+    draft_review: str
+    criteria_scores: Dict[str, int]  # {"pep8": 0-10, "type_hints": ..., "edge_cases": ..., "naming": ...}
+    weakest_criterion: str
+    verdict: str  # "ok" | "needs_revision"
+    round: int
+    max_rounds: int
+
+# ---------------------------------------------------------------------------
+# 3. Structured output for the critic (reflect node)
+# ---------------------------------------------------------------------------
+class CriticOutput(BaseModel):
+    pep8: int = Field(..., ge=0, le=10, description="Score for PEP8 compliance")
+    type_hints: int = Field(..., ge=0, le=10, description="Score for type hints usage")
+    edge_cases: int = Field(..., ge=0, le=10, description="Score for handling edge cases")
+    naming: int = Field(..., ge=0, le=10, description="Score for naming conventions")
+    weakest_criterion: str = Field(..., description="Criterion with the lowest score")
+    verdict: str = Field(..., description="'ok' or 'needs_revision'")
+
+critic_parser = PydanticOutputParser(pydantic_object=CriticOutput)
+
+# ---------------------------------------------------------------------------
+# 4. Graph nodes
+# ---------------------------------------------------------------------------
+async def draft_review(state: CodeReviewState) -> CodeReviewState:
+    prompt = (
+        "You are a senior Python reviewer.\n"
+        "Given the following function, write a concise code review (3–6 points).\n"
+        "Focus on style, correctness, and best practices.\n"
+        "Return the review as plain text.\n"
+        f"Function:\n{state['code']}"
+    )
+    review = await llm.ainvoke([HumanMessage(content=prompt)])
+    state['draft_review'] = review.content.strip()
+    return state
+
+async def reflect(state: CodeReviewState) -> CodeReviewState:
+    prompt = (
+        "You are a code quality critic.\n"
+        "Evaluate the following review against four criteria: PEP8, type hints, edge cases, naming.\n"
+        "Assign a score 0–10 for each criterion.\n"
+        "Identify the weakest criterion and decide if the review is "ok" or "needs_revision".\n"
+        "Return the results in JSON matching the following schema:\n"
+        f"{critic_parser.get_format_instructions()}\n"
+        f"Review:\n{state['draft_review']}"
+    )
+    result = await llm.ainvoke([HumanMessage(content=prompt)])
+    parsed = critic_parser.parse(result.content)
+    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:
+    # Rewrite only the section of the review that addresses the weakest criterion
+    prompt = (
+        "You are a code reviewer.\n"
+        "The current review is: \n"
+        f"{state['draft_review']}\n"
+        "The weakest criterion is: " + state['weakest_criterion'] + ".\n"
+        "Rewrite the review to strengthen this part, keeping the overall tone.\n"
+        "Return only the updated review text."
+    )
+    updated = await llm.ainvoke([HumanMessage(content=prompt)])
+    state['draft_review'] = updated.content.strip()
+    state['round'] += 1
+    return state
+
+# ---------------------------------------------------------------------------
+# 5. Build the LangGraph
+# ---------------------------------------------------------------------------
+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)
+
+    graph.set_entry_point("draft_review")
+    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",
+    )
+    graph.add_edge("rewrite", "reflect")
+    return graph
+
+# ---------------------------------------------------------------------------
+# 6. Tool that runs the graph
+# ---------------------------------------------------------------------------
+@tool
+def run_review(code: str) -> str:
+    """Run the full review cycle on the provided Python code."""
+    # Initial state
+    state: CodeReviewState = {
+        "code": code,
+        "draft_review": "",
+        "criteria_scores": {},
+        "weakest_criterion": "",
+        "verdict": "",
+        "round": 1,
+        "max_rounds": 2,
+    }
+    graph = build_graph()
+    # Execute graph synchronously
+    final_state = graph.invoke(state)
+    # Prepare a readable output
+    output = [
+        "--- Draft Review ---",
+        final_state["draft_review"],
+        "\n--- Critic Scores ---",
+        f"PEP8: {final_state['criteria_scores'].get('pep8', 'N/A')}\n"
+        f"Type Hints: {final_state['criteria_scores'].get('type_hints', 'N/A')}\n"
+        f"Edge Cases: {final_state['criteria_scores'].get('edge_cases', 'N/A')}\n"
+        f"Naming: {final_state['criteria_scores'].get('naming', 'N/A')}\n",
+        f"Verdict: {final_state['verdict']} (round {final_state['round']})",
+    ]
+    return "\n".join(output)
+
+# ---------------------------------------------------------------------------
+# 7. DeepAgents setup
+# ---------------------------------------------------------------------------
+backend = CompositeBackend([
+    LocalShellBackend(workspace_dir="./workspace"),
+    FilesystemBackend(),
+])
+
+agent = create_deep_agent(
+    model=llm,
+    tools=[run_review],
+    backend=backend,
+    system_prompt="You are a helpful code review assistant.",
+)
+
+# ---------------------------------------------------------------------------
+# 8. Demo CLI
+# ---------------------------------------------------------------------------
+async def main():
+    example_code = """
+    def sort_numbers(arr):
+        return sorted(arr)
+    """
+    result = await agent.ainvoke(
+        {"messages": [HumanMessage(content="Please review the following function:
+
+"" + example_code + "")]},
+        {"configurable": {"thread_id": "demo-session"}},
+    )
+    print(result["messages"][-1].content)
+
+if __name__ == "__main__":
+    asyncio.run(main())