app/agents/code_review_agent.py

import asyncio

from app.core.llm import call_llm_for_json
from app.core.logger import get_logger
from app.core.prompts import AUTO_FIX_PROMPT, CODE_REVIEW_PROMPT
from app.models.repository import RepositoryMetadata
from app.models.review import AutoFix, ReviewSuggestions
from app.tools.file_scanner import read_source_samples, get_python_files
from app.tools.github_tool import get_changed_files
import os

logger = get_logger(__name__)


async def run(local_path: str, metadata: RepositoryMetadata) -> ReviewSuggestions:
    """
    Perform AI-powered code review on repository.
    """
    logger.info("Starting Code Review Agent", extra={"path": local_path})

    # Step 1 - Get Python files
    base_sha = metadata.base_sha
    head_sha = metadata.head_sha

    changed_files = []
    if base_sha and head_sha:
        changed_files = get_changed_files(local_path, base_sha, head_sha)

    #Step 2 - Read source code samples

    all_files = changed_files or [
    f for f in get_python_files(local_path)
    if not os.path.basename(f).startswith("test_")
    and not os.path.basename(f).endswith("_test.py")
]

    source_samples = read_source_samples(
        local_path,
        max_files=4,
        max_chars=2500,
        file_list=all_files,
    )

    if not source_samples:
        return ReviewSuggestions(
            summary="No source files found to review.", overall_score=0.0
        )

    # Step 3 - Call LLM and parse JSON in one shot
    source_code = "\n\n".join(source_samples)
    prompt = CODE_REVIEW_PROMPT.format(source_code=source_code)
    llm_data = await call_llm_for_json(prompt, task="code_review")

    review = ReviewSuggestions(
        solid_violations=llm_data.get("solid_violations", []),
        duplicate_code=llm_data.get("duplicate_code", []),
        refactor_suggestions=llm_data.get("refactor_suggestions", []),
        overall_score=float(llm_data.get("overall_score", 5.0)),
        summary=llm_data.get("summary", ""),
    )

    # Generate auto-fixes for top 3 violations (cap LLM calls)
    auto_fixes: list[AutoFix] = []
    top_findings = (review.solid_violations + review.refactor_suggestions)[:3]
    if top_findings and source_samples:
        first_file_content = source_samples[0]  # already read
        fix_tasks = [
            call_llm_for_json(
                AUTO_FIX_PROMPT.format(
                    source_code=first_file_content[:1500], finding=finding
                )
            )
            for finding in top_findings
        ]
        fix_results = await asyncio.gather(*fix_tasks, return_exceptions=True)
        for res in fix_results:
            if isinstance(res, dict) and res.get("fixed_snippet"):
                auto_fixes.append(AutoFix(**res))

    review = review.model_copy(update={"auto_fixes": auto_fixes})

    # Second pass — if score is poor, ask LLM for a prioritised fix list
    if review.overall_score < 5.0:
        logger.info(
            "Low score detected, running focused follow-up pass",
            extra={"score": review.overall_score},
        )
        followup_prompt = (
            f"The following code scored {review.overall_score}/10 in a review.\n\n"
            f"Top violations: {review.solid_violations[:3]}\n\n"
            f'Return ONLY a JSON object with one key: "priority_fixes": a list of up to 3 strings, '
            f"each describing the single most impactful change to make first."
        )
        followup_data = await call_llm_for_json(followup_prompt)
        priority_fixes = followup_data.get("priority_fixes", [])
        if isinstance(priority_fixes, list) and priority_fixes:
            review = review.model_copy(
                update={
                    "refactor_suggestions": priority_fixes + review.refactor_suggestions
                }
            )
    logger.info("Code review complete", extra={"score": review.overall_score})
    return review