Problem

Current review system is inefficient:

• Re-reviews entire PR on each push (wastes 60-80% of tokens)
• Reviews trivial changes (formatting, imports, lockfiles)
• Provides redundant feedback on unchanged code
• Poor developer experience with repeated suggestions

Solution

Implement smart incremental reviews with two modes:

Stateless Mode (No DB required)

• File filtering: Skip auto-generated files, lockfiles, trivial formatting changes
• Change classification: Detect meaningful vs cosmetic changes using patterns
• Commit analysis: Parse commit messages for change context
• Expected savings: 30-50% token reduction

Stateful Mode (Database enabled)

• Incremental diffs: Only review changes since last review (git diff last_sha..head_sha)
• Suggestion tracking: Store and deduplicate previous suggestions
• Review state: Track last reviewed SHA per PR
• Expected savings: 60-80% token reduction

Implementation

Phase 1 (Low risk): Stateless improvements

Skip these file patterns

SKIP_FILES = ['.lock', 'package-lock.json', 'yarn.lock']
TRIVIAL_PATTERNS = [r'^\s$', r'^(import|from)\s+', r'^\s*(#|//)']

Phase 2: Stateful improvements

• Add ReviewState model to track last reviewed SHA
• Add ReviewSuggestionHistory for deduplication
• Modify dispatcher to use incremental diffs

Benefits

• Cost reduction: 30-80% fewer tokens used
• Faster reviews: Skip processing trivial files
• Better UX: No redundant suggestions for unchanged code
• Backwards compatible: Opt-in via config flags

Configuration

ENABLE_SMART_FILTERING=true
ENABLE_INCREMENTAL_REVIEWS=true # stateful only
TRIVIAL_CHANGE_THRESHOLD=3