Project: Code Review Agent

Code Input
    ↓
[Parallel Specialist Reviews]
    ├── Security Reviewer
    ├── Bug Hunter
    ├── Performance Analyzer
    └── Code Quality Reviewer
    ↓
[Synthesis Node] — Merge, deduplicate, prioritize
    ↓
Structured Review Report

# code_review_agent.py
import asyncio
import json
from typing import TypedDict, Annotated
import operator
from langchain_openai import ChatOpenAI
from langchain_core.messages import SystemMessage, HumanMessage
from langgraph.graph import StateGraph, END
from langgraph.types import Send
from dotenv import load_dotenv

load_dotenv()

llm = ChatOpenAI(model="gpt-4o", temperature=0)

# === STATE ===

class ReviewIssue(TypedDict):
    severity: str      # critical, high, medium, low, info
    category: str      # security, bug, performance, quality
    title: str
    description: str
    location: str      # Line number or code reference
    suggestion: str    # What to do instead
    code_fix: str      # Optional corrected code

class ReviewState(TypedDict):
    code: str
    language: str
    context: str  # e.g., "this is a public API endpoint"
    issues: Annotated[list[ReviewIssue], operator.add]  # Accumulated from all reviewers
    summary: str
    overall_score: int  # 1-10

# === SPECIALIST REVIEWERS ===

REVIEWER_PROMPTS = {
    "security": """You are a security engineer reviewing code for vulnerabilities.

Focus ONLY on security issues:
- SQL injection, XSS, CSRF, command injection
- Authentication and authorization gaps
- Sensitive data exposure (credentials in code, logs)
- Input validation and sanitization
- Cryptographic issues
- Dependency vulnerabilities (obvious unsafe patterns)

For each issue found, return JSON:
{
  "issues": [
    {
      "severity": "critical|high|medium|low",
      "category": "security",
      "title": "Brief issue name",
      "description": "What the vulnerability is and why it's dangerous",
      "location": "Line X or function name",
      "suggestion": "How to fix it",
      "code_fix": "Corrected code snippet (if simple enough to show)"
    }
  ]
}

If no security issues found, return: {"issues": []}
Return ONLY valid JSON.""",

    "bugs": """You are a senior engineer reviewing code for bugs and logic errors.

Focus ONLY on:
- Logic errors and incorrect behavior
- Off-by-one errors
- Null pointer / undefined reference risks
- Race conditions
- Incorrect error handling (swallowed exceptions, wrong error messages)
- Edge cases not handled (empty arrays, zero values, negative numbers)
- Type mismatches or conversion issues
- Incorrect boolean logic

For each issue found, return JSON with this structure:
{
  "issues": [
    {
      "severity": "critical|high|medium|low",
      "category": "bug",
      "title": "Brief issue name",
      "description": "What's wrong and what would happen at runtime",
      "location": "Line X or function name",
      "suggestion": "How to fix it",
      "code_fix": "Corrected code snippet"
    }
  ]
}

Return ONLY valid JSON.""",

    "performance": """You are a performance engineer reviewing code for inefficiencies.

Focus ONLY on:
- N+1 query problems (loops with database calls)
- Unnecessary repeated computation (move outside loops)
- Missing caching for expensive operations
- Inefficient data structures (list search when dict lookup needed)
- Blocking operations that should be async
- Missing pagination or limits on data fetching
- Memory leaks or excessive memory usage

For each issue found, return JSON:
{
  "issues": [
    {
      "severity": "high|medium|low|info",
      "category": "performance",
      "title": "Brief issue name",
      "description": "The performance problem and its impact at scale",
      "location": "Line X or function name",
      "suggestion": "How to fix it",
      "code_fix": "Optimized code snippet"
    }
  ]
}

Return ONLY valid JSON.""",

    "quality": """You are a senior engineer reviewing code for maintainability and quality.

Focus ONLY on:
- Functions that are too long (> 40 lines doing multiple things)
- Poor naming (variables, functions, classes)
- Missing or incorrect documentation (complex code without explanation)
- Unnecessary complexity (simpler approach exists)
- Repeated code that should be extracted
- Incorrect use of language features or libraries
- Missing type annotations where expected
- Dead code or unused variables

For each issue found, return JSON:
{
  "issues": [
    {
      "severity": "medium|low|info",
      "category": "quality",
      "title": "Brief issue name",
      "description": "What's wrong and why it matters for maintainability",
      "location": "Line X or function name",
      "suggestion": "How to improve it",
      "code_fix": "Improved code snippet (if helpful)"
    }
  ]
}

Return ONLY valid JSON."""
}

def run_specialist_review(state: dict) -> ReviewState:
    """Run a single specialist review. Called in parallel for each specialist."""
    review_type = state["review_type"]
    code = state["code"]
    language = state["language"]
    context = state["context"]
    
    print(f"   🔍 Running {review_type} review...")
    
    response = llm.invoke([
        SystemMessage(content=REVIEWER_PROMPTS[review_type]),
        HumanMessage(content=f"""Language: {language}
Context: {context}

Code to review:
```{language}
{code}
```""")
    ])
    
    try:
        parsed = json.loads(response.content)
        issues = parsed.get("issues", [])
        print(f"   ✓ {review_type}: Found {len(issues)} issue(s)")
        return {"issues": issues}
    except json.JSONDecodeError:
        print(f"   ⚠️ {review_type}: Failed to parse response")
        return {"issues": []}

# === SYNTHESIS NODE ===

def synthesize_review(state: ReviewState) -> ReviewState:
    """Combine all issues into a structured review."""
    print("\n📊 Synthesizing review...")
    
    issues = state["issues"]
    
    # Deduplicate similar issues
    seen_titles = set()
    unique_issues = []
    for issue in issues:
        if issue["title"] not in seen_titles:
            unique_issues.append(issue)
            seen_titles.add(issue["title"])
    
    # Sort by severity
    severity_order = {"critical": 0, "high": 1, "medium": 2, "low": 3, "info": 4}
    sorted_issues = sorted(unique_issues, key=lambda x: severity_order.get(x.get("severity", "info"), 5))
    
    # Count by severity
    counts = {"critical": 0, "high": 0, "medium": 0, "low": 0, "info": 0}
    for issue in sorted_issues:
        sev = issue.get("severity", "info")
        counts[sev] = counts.get(sev, 0) + 1
    
    # Calculate score (10 = perfect, deduct based on severity)
    score = 10
    score -= counts["critical"] * 3
    score -= counts["high"] * 2
    score -= counts["medium"] * 1
    score -= counts["low"] * 0.25
    score = max(1, min(10, round(score)))
    
    # Generate summary
    issues_text = json.dumps(sorted_issues, indent=2)
    
    summary_response = llm.invoke([
        SystemMessage(content="You summarize code reviews concisely for developers."),
        HumanMessage(content=f"""Code: {state['language']} — {state['context']}

Issues found:
- Critical: {counts['critical']}
- High: {counts['high']}  
- Medium: {counts['medium']}
- Low: {counts['low']}

Top issues:
{issues_text[:2000]}

Write a 2-3 sentence executive summary of this code review. 
Lead with the most important finding, then overall assessment.""")
    ])
    
    return {
        **state,
        "issues": sorted_issues,
        "summary": summary_response.content,
        "overall_score": score
    }

# === ROUTING: FAN-OUT TO PARALLEL SPECIALISTS ===

def route_to_specialists(state: ReviewState):
    """Send the same code to all specialists simultaneously."""
    return [
        Send("specialist_review", {
            "review_type": review_type,
            "code": state["code"],
            "language": state["language"],
            "context": state["context"]
        })
        for review_type in ["security", "bugs", "performance", "quality"]
    ]

# === BUILD GRAPH ===

def build_code_review_agent():
    graph = StateGraph(ReviewState)
    
    graph.add_node("specialist_review", run_specialist_review)
    graph.add_node("synthesize", synthesize_review)
    
    graph.add_conditional_edges("__start__", route_to_specialists, ["specialist_review"])
    graph.add_edge("specialist_review", "synthesize")
    graph.add_edge("synthesize", END)
    
    return graph.compile()

# === REPORT FORMATTING ===

def format_review_report(state: ReviewState) -> str:
    """Format the review into a readable markdown report."""
    issues = state["issues"]
    score = state["overall_score"]
    
    score_bar = "🟢" * min(score, 10) if score >= 7 else "🟡" * min(score, 10) if score >= 4 else "🔴" * min(score, 10)
    
    lines = [
        f"# Code Review Report",
        f"",
        f"## Summary",
        f"{state['summary']}",
        f"",
        f"**Overall Score:** {score}/10 {score_bar}",
        f"",
    ]
    
    # Group by severity
    for severity in ["critical", "high", "medium", "low", "info"]:
        severity_issues = [i for i in issues if i.get("severity") == severity]
        if not severity_issues:
            continue
        
        emoji = {"critical": "🚨", "high": "⚠️", "medium": "⚡", "low": "💡", "info": "ℹ️"}.get(severity, "")
        lines.append(f"## {emoji} {severity.title()} Issues ({len(severity_issues)})")
        lines.append("")
        
        for issue in severity_issues:
            lines.append(f"### {issue['title']}")
            lines.append(f"**Category:** {issue.get('category', 'unknown')} | **Location:** {issue.get('location', 'N/A')}")
            lines.append(f"")
            lines.append(issue['description'])
            lines.append(f"")
            lines.append(f"**Suggestion:** {issue['suggestion']}")
            
            if issue.get("code_fix"):
                lines.append(f"")
                lines.append(f"```")
                lines.append(issue["code_fix"])
                lines.append(f"```")
            lines.append("")
    
    return "\n".join(lines)

# === MAIN ===

def review_code(code: str, language: str = "python", context: str = "general application code") -> str:
    agent = build_code_review_agent()
    
    print(f"\n{'='*60}")
    print(f"🔍 Starting Code Review")
    print(f"   Language: {language}")
    print(f"   Context: {context}")
    print(f"   Code length: {len(code)} characters")
    print(f"{'='*60}")
    
    result = agent.invoke({
        "code": code,
        "language": language,
        "context": context,
        "issues": [],
        "summary": "",
        "overall_score": 10
    })
    
    report = format_review_report(result)
    print(f"\n✅ Review complete: {len(result['issues'])} issues found, score: {result['overall_score']}/10")
    return report


if __name__ == "__main__":
    # Example: review a Python function
    sample_code = '''
def get_user(user_id):
    query = f"SELECT * FROM users WHERE id = {user_id}"
    result = db.execute(query)
    password = result[0]["password"]
    print(f"Found user: {result[0]['email']}, password: {password}")
    return result[0]

def process_orders(user_id):
    user = get_user(user_id)
    orders = []
    for order_id in range(1000):
        order = db.query(f"SELECT * FROM orders WHERE id = {order_id}")
        if order and order["user_id"] == user_id:
            orders.append(order)
    return orders
'''
    
    report = review_code(
        code=sample_code,
        language="python",
        context="public API endpoint handling user authentication"
    )
    print(report)
    
    with open("code_review.md", "w") as f:
        f.write(report)
    print("\n📄 Review saved to code_review.md")