AI Code Review Bot — PR Analysis

## Task
GitHub App that automatically reviews pull requests using LLMs.

## Requirements
- Platform: GitHub App (webhook-based)
- Runtime: Node.js serverless function or container
- LLM: Claude or GPT-4 for analysis

## Workflow
```
1. PR opened/updated → webhook fires
2. Fetch diff via GitHub API
3. For each changed file:
   a. Get full file content (not just diff)
   b. Analyze with LLM: security, bugs, performance, style
   c. Generate inline comments at specific lines
4. Post review via GitHub API (REQUEST_CHANGES or APPROVE)
5. Post summary comment with overview
```

## LLM Prompt Strategy
```
System: You are a senior code reviewer. Analyze this diff for:
- Security vulnerabilities (injection, auth bypass, secret exposure)
- Logic errors and potential bugs
- Performance issues (N+1 queries, memory leaks, unnecessary computation)
- Best practice violations

For each issue, respond with JSON:
{
  "path": "src/api/users.ts",
  "line": 42,
  "severity": "critical|warning|suggestion",
  "message": "SQL injection risk: user input passed directly to query",
  "suggestion": "Use parameterized query: db.query($1, [userId])"
}
```

## Implementation Notes
1. Chunk large diffs — don't exceed context window
2. Filter: skip lock files, generated code, images
3. Rate limit: max 50 comments per review
4. Deduplicate: don't comment on the same pattern twice
5. Include a /reviewbot ignore comment to skip files
6. Store review history to avoid re-reviewing unchanged files