main.py

# main.py — AutoReview AI
# LangChain-based autonomous GitHub PR reviewer

import os
import re
from dotenv import load_dotenv

from langchain_groq import ChatGroq
from langchain.agents import AgentExecutor, create_tool_calling_agent
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.messages import HumanMessage, SystemMessage

from tools import fetch_pr_diff, analyze_diff, post_inline_comment, post_overall_review, fetch_pr_data, format_pr_data

load_dotenv()


# ── LLM ──────────────────────────────────
llm = ChatGroq(model="openai/gpt-oss-120b", temperature=0)

# ── TOOLS ─────────────────────────────────
tools = [fetch_pr_diff, analyze_diff, post_inline_comment, post_overall_review]

# ── PROMPT ────────────────────────────────
prompt = ChatPromptTemplate.from_messages([
    ("system", """You are AutoReview AI — an expert autonomous GitHub PR reviewer.

Your job when given a PR URL:
1. Use fetch_pr_diff to get the PR details and code changes
2. Use analyze_diff for each changed file to find bugs, security issues, style problems
3. Use post_inline_comment to post specific comments on problematic lines
4. Use post_overall_review to post a final APPROVE or REQUEST_CHANGES verdict

Be thorough but concise. Focus on real issues that matter.
Always complete all 4 steps — don't stop after fetching.

If GitHub token is not available, still analyze and return the review as text.
"""),
    ("human", "{input}"),
    ("placeholder", "{agent_scratchpad}")
])

# ── AGENT ─────────────────────────────────
agent          = create_tool_calling_agent(llm, tools, prompt)
agent_executor = AgentExecutor(
    agent=agent,
    tools=tools,
    verbose=True,
    max_iterations=10,
    handle_parsing_errors=True
)

# ─────────────────────────────────────────
# MAIN REVIEW FUNCTION
# ─────────────────────────────────────────
def review_pr(
    pr_url: str,
    github_token: str | None = None
) -> dict:
    """
    Main function to review a GitHub PR.
    Returns structured review result.
    """
    print(f"\n{'='*60}")
    print(f"🔍 AutoReview AI starting review...")
    print(f"📎 PR: {pr_url}")
    print("="*60)

    try:
        result = agent_executor.invoke({
            "input": f"""
Please review this GitHub Pull Request: {pr_url}
GitHub Token: {github_token if github_token else 'Not Provided'}

Steps to follow:
1. Fetch the PR diff using fetch_pr_diff
2. Analyze each changed file using analyze_diff
3. Post inline comments for specific issues using post_inline_comment
4. Post overall verdict using post_overall_review

Provide a thorough code review focusing on:
- Bugs and logic errors
- Security vulnerabilities
- Code style and best practices
- Performance issues
"""
        })

        output = result.get("output", "")
        print(f"\n✅ Review complete!")
        print(f"\nOutput:\n{output}")

        return {
            "success": True,
            "pr_url":  pr_url,
            "review":  output,
            "error":   None
        }

    except Exception as e:
        error_msg = str(e)
        print(f"\n❌ Error: {error_msg}")
        return {
            "success": False,
            "pr_url":  pr_url,
            "review":  None,
            "error":   error_msg
        }

# ─────────────────────────────────────────
# SIMPLE REVIEW (no GitHub token needed)
# For demo purposes
# ─────────────────────────────────────────
def review_pr_simple(
    pr_url: str,
    github_token: str | None = None
) -> dict:
    """
    Reviews PR without posting comments to GitHub.
    Analyzes each changed file separately (instead of truncating the
    whole diff to one 4000-char blob), and never lets a Groq/GitHub
    error crash the caller — always returns a usable dict.
    """
    print(f"\n🔍 Fetching PR data...")

    try:
        # Step 1: Fetch structured PR data (real file list, not a string)
        data = fetch_pr_data(
    pr_url=pr_url,
    github_token=github_token
)
        if not data["success"]:
            return {"success": False, "error": data["error"], "review": None}

        print(f"✅ PR fetched! {len(data['files'])} file(s) changed")

        # Step 2: Analyze EACH file separately (cap at 8 files for demo speed)
        per_file_analysis = []
        files_to_review = data["files"][:8]

        for f in files_to_review:
            print(f"🤖 Analyzing {f['filename']}...")
            file_block = (
                f"--- {f['filename']} ({f['status']}) "
                f"+{f['additions']}/-{f['deletions']} ---\n{f['patch']}"
            )
            try:
                file_analysis = analyze_diff.invoke(file_block)
            except Exception as e:
                file_analysis = f"ERROR analyzing this file: {str(e)}"
            per_file_analysis.append(f"## {f['filename']}\n{file_analysis}")

        if not per_file_analysis:
            analysis = "No file changes with diffs to analyze."
        else:
            analysis = "\n\n".join(per_file_analysis)
        print("✅ Analysis complete!")

        # Step 3: Generate final verdict from the per-file analyses
        try:
            verdict_response = llm.invoke([
                SystemMessage("You are a senior engineer. Give a final PR review verdict."),
                HumanMessage(f"""
Based on this PR analysis, write a complete review report:

PR TITLE: {data['pr_title']}
PR DESCRIPTION: {data['pr_description'][:500]}
FILES CHANGED: {data['total_files']}

PER-FILE ANALYSIS:
{analysis[:6000]}

Write a professional review with:
1. Overall verdict (APPROVE / REQUEST CHANGES)
2. Key issues found
3. Positive aspects
4. Specific suggestions
""")
            ])
            verdict = verdict_response.content
        except Exception as e:
            verdict = f"Could not generate final verdict — LLM error: {str(e)}"

        return {
            "success":  True,
            "pr_url":   pr_url,
            "pr_data":  format_pr_data(data),
            "analysis": analysis,
            "verdict":  verdict,
            "error":    None
        }

    except Exception as e:
        # Catch-all so an unexpected error (network, parsing, anything)
        # never crashes app.py — it always gets a usable dict back
        return {"success": False, "error": f"Unexpected error: {str(e)}", "review": None}

# ─────────────────────────────────────────
# RUN
# ─────────────────────────────────────────
if __name__ == "__main__":
    # Test with a real public PR
    test_pr = "https://github.com/langchain-ai/langchain/pull/1"

    print("Mode: Simple review (no GitHub token needed for demo)")
    result = review_pr_simple(test_pr, None)

    if result["success"]:
        print(f"\n{'='*60}")
        print("📋 PR DATA SUMMARY:")
        print(result["pr_data"][:500])
        print(f"\n🔍 ANALYSIS:")
        print(result["analysis"])
        print(f"\n✅ FINAL VERDICT:")
        print(result["verdict"])
    else:
        print(f"❌ Failed: {result['error']}")