AI_REVIEW_SETUP.md

AI Review Setup Guide

Critical Fix Applied

The AI Review system has been fixed to use the Anthropic API directly instead of trying to invoke a non-existent CLI command.

What Was Wrong

Previously, the code tried to call:

claude code --prompt-file /path/to/file

This command does not exist and was causing all AI reviews to fail with:

error: unknown option '--prompt-file'

The system would catch this error and return an empty fallback review, which is why you saw:

• filesReviewed: 9 (files were correctly read)
• totalIssues: 0 (AI never actually reviewed them)
• Response came immediately (because it was just returning fallback)

What Was Fixed

Modified Files:

• AIReviewService.ts:1-5 - Added Anthropic SDK import
• AIReviewService.ts:65-78 - Added Anthropic client initialization
• AIReviewService.ts:449-502 - Replaced broken CLI call with proper API call
• index.ts:2 - Added dotenv to load environment variables
• .env.example - Added ANTHROPIC_API_KEY requirement

Key Changes:

1. Now uses @anthropic-ai/sdk package (already installed)
2. Calls anthropic.messages.create() with proper API parameters
3. Uses Claude Sonnet 4 model (claude-sonnet-4-20250514)
4. Properly extracts text response and handles errors
5. Logs detailed information about API calls

Setup Instructions

Step 1: Get Your Anthropic API Key

1. Visit: https://console.anthropic.com/settings/keys
2. Log in or create an account
3. Click "Create Key"
4. Copy your API key (starts with sk-ant-)

Step 2: Create .env File

Create a .env file in the project root:

cd /Users/highgarden/Developments/AI/HighReview
cp .env.example .env

Edit .env and add your API key:

# HighReview Environment Configuration

# Anthropic API Key (Required for AI Code Review)
ANTHROPIC_API_KEY=sk-ant-your-actual-api-key-here

# Optional: Set port if you want to use a different port
# PORT=8765

Step 3: Restart Server

# If server is running, stop it (Ctrl+C)
cd apps/cli
npm run dev

You should now see:

🚀 Starting HighReview server...
✓ Server running at http://localhost:8765

No longer see the warning: [AI Review] ANTHROPIC_API_KEY not set. AI reviews will use fallback mode.

Step 4: Test AI Review

1. Open HighReview in browser: http://localhost:5273
2. Navigate to a PR
3. Click "Start Review" with AI options enabled
4. Wait for real AI review (will take 30-60 seconds for proper analysis)
5. Check browser console - you should see:

   [AI Review] Calling Claude AI API...
   [AI Review] Prompt size: XXXXX characters
   [AI Review] AI response received: {
     responseLength: XXXX,
     stopReason: 'end_turn',
     inputTokens: XXXX,
     outputTokens: XXXX
   }

Verification

Server Logs (Should See)

[AI Review] Starting review for worktree: /path/to/worktree
[AI Review] Got diff from origin/branch-name, 8881 bytes
[AI Review] Found 9 changed files from origin/branch-name
[AI Review] Read 9 file contents
[AI Review] Calling Claude AI API...
[AI Review] Prompt size: 125000 characters
[AI Review] AI response received: {
  responseLength: 5234,
  stopReason: 'end_turn',
  inputTokens: 32145,
  outputTokens: 1234
}
[AI Review] Review completed: { filesReviewed: 9, totalIssues: 15 }

Browser Console (Should See)

[EnhancedAIReviewPanel] Rendering with data: {
  hasReview: true,
  filesReviewed: 9,
  totalIssues: 15,        // <-- Should NOT be 0!
  criticalIssues: 3,
  warnings: 7,
  suggestions: 5,
  hasChangeIntents: true,
  hasCallStacks: true,
  hasImpactAnalysis: true
}

What AI Review Now Sends to Claude

The AI now receives:

1. All changed files list (file paths)
2. Complete git diff (showing exact changes)
3. Full file contents (up to 20 files, 50KB each) for context
4. Selected review options:
   • Change intent analysis (file/block level)
   • Call stack visualization (flowchart/sequence diagrams)
   • Impact analysis (module/project/dependency scope)
   • Semantic diff (moved code, refactoring detection)
   • Custom prompts
5. Current conversation context (if available)
6. Related files (for impact analysis)

All of this is packaged into a comprehensive prompt and sent to Claude Sonnet 4 for analysis.

Cost Considerations

Anthropic API pricing (as of January 2025):

• Input: $3 per million tokens
• Output: $15 per million tokens

Typical PR review:

• Input: 30,000-50,000 tokens (~$0.10-0.15)
• Output: 2,000-5,000 tokens (~$0.03-0.08)
• Total per review: ~$0.13-0.23

For 100 reviews per month: ~$13-23

Troubleshooting

Error: "ANTHROPIC_API_KEY not set"

Solution: Create .env file with your API key (see Step 2 above)

Error: "401 Unauthorized"

Possible causes:

1. Invalid API key - check for typos
2. API key expired - generate new key
3. API key not from Anthropic - must start with sk-ant-

Solution:

# Verify your .env file
cat /Users/highgarden/Developments/AI/HighReview/.env

# Should show:
ANTHROPIC_API_KEY=sk-ant-...

AI Review Still Returns Empty Results

Check server logs for:

[AI Review] Claude AI API call failed: [error details]

Common issues:

1. Network connectivity - check internet connection
2. API rate limit - wait and retry
3. Invalid API key - verify in console.anthropic.com
4. Prompt too large - try with smaller PR

Solution:

# Check server logs in terminal running `npm run dev`
# Look for detailed error information

AI Review Takes Too Long (> 2 minutes)

This is now expected behavior for proper AI analysis!

• Small PRs (< 500 lines): 20-40 seconds
• Medium PRs (500-2000 lines): 40-90 seconds
• Large PRs (2000+ lines): 90-180 seconds

If it takes > 3 minutes, check:

1. API timeout settings (currently 120 seconds)
2. Network latency
3. PR size (very large PRs may need chunking)

Model Configuration

Current settings in AIReviewService.ts:466-470:

const message = await this.anthropic.messages.create({
  model: "claude-sonnet-4-20250514", // Latest Sonnet - balance of speed/quality
  max_tokens: 16000, // Increased for detailed reviews
  temperature: 0.3, // Lower = more focused and consistent
  messages: [
    /* prompt */
  ],
});

Why Claude Sonnet 4:

• Better code understanding than Sonnet 3.5
• Faster than Opus
• Lower cost than Opus
• Excellent at structured output (matches our format)

To use a different model: Edit AIReviewService.ts:466:

model: 'claude-opus-4-20250514',  // For highest quality (slower, more expensive)
// or
model: 'claude-sonnet-3-5-20241022',  // For faster reviews (slightly lower quality)

Feature Verification Checklist

After setup, verify these features work:

• AI Review returns actual issues (not totalIssues: 0)
• Critical Issues section populated
• Warnings section populated
• Suggestions section populated
• Change Intent analysis visible (if enabled)
• Call Stack visualization with Mermaid diagrams (if enabled)
• Impact Analysis showing affected areas (if enabled)
• Moved Code detection (if enabled)
• Refactoring detection (if enabled)
• Review summary in Korean (if language set to 'ko')

Additional Resources

• Anthropic API Documentation
• Claude Sonnet 4 Model Card
• HighReview Code Navigation Guide
• Final Implementation Summary

Support

If AI Review still doesn't work after following this guide:

1. Check server terminal logs for detailed error messages
2. Check browser console for frontend errors
3. Verify .env file exists and contains valid API key
4. Test API key directly with curl:

   curl https://api.anthropic.com/v1/messages \
     --header "x-api-key: $ANTHROPIC_API_KEY" \
     --header "anthropic-version: 2023-06-01" \
     --header "content-type: application/json" \
     --data '{
       "model": "claude-sonnet-4-20250514",
       "max_tokens": 1024,
       "messages": [{"role": "user", "content": "Hello"}]
     }'

---

Fixed Date: 2026-01-28 Fixed By: Claude Sonnet 4.5 Status: ✅ Production Ready