Skip to content

user-guide.md

Latest commit

 

History

History
107 lines (69 loc) · 2.62 KB

user-guide.md

File metadata and controls

107 lines (69 loc) · 2.62 KB
Raw

User Guide

Feedback Learning

The action also learns from your positive feedback (👍 reactions) to improve confidence scoring over time.

How It Works

  • Thumbs Up (👍): Indicates the finding is valuable

The system tracks:

  • Which findings get positive reactions
  • Provider agreement on findings
  • Confidence levels over time

Configuration

Feedback learning is controlled by these environment variables:

- LEARNING_ENABLED: true # Enable feedback learning
- LEARNING_MIN_FEEDBACK_COUNT: 5 # Min feedback before adjusting confidence
- LEARNING_LOOKBACK_DAYS: 30 # How far back to look for feedback

See the main README for full configuration options.

Tips for Effective Review

Managing Noise

If you're seeing too many low-confidence findings, consider:

  1. Quiet Mode: Filters findings below a confidence threshold

    - QUIET_MODE_ENABLED: true
- QUIET_MIN_CONFIDENCE: 0.6 # Only show findings ≥60% confidence

  2. Severity Filtering: Only show critical/major findings

    - INLINE_MIN_SEVERITY: major # Skip minor findings in inline comments

  3. Comment Limits: Cap the number of inline comments

    - INLINE_MAX_COMMENTS: 20 # Maximum inline comments per review

Working with the Summary Comment

The action posts a summary comment on each PR with:

  • Overall statistics (critical/major/minor counts)
  • All findings grouped by severity
  • Performance metrics (duration, cost, providers used)

On incremental reviews, the summary comment is updated in place rather than creating a new comment each time.

Advanced Usage

Dry Run Mode

Test the action without posting comments:

- DRY_RUN: true

The action will:

  • Run the full review
  • Generate all findings and comments
  • Log what it would have posted (check Action logs)
  • NOT actually post any comments to GitHub

Custom Severity Thresholds

Control which findings appear in inline comments vs. summary only:

# Only post critical findings as inline comments
- INLINE_MIN_SEVERITY: critical

# Require high provider agreement for inline comments
- INLINE_MIN_AGREEMENT: 0.7 # 70% of providers must agree

Performance Optimization

For large PRs, optimize review speed:

# Enable incremental mode (reviews only changed files since last run)
- INCREMENTAL_ENABLED: true

# Use faster models for AST analysis
- ENABLE_AST_ANALYSIS: false # Disable AST if not needed

# Limit concurrent provider calls
- PROVIDER_MAX_PARALLEL: 3 # Reduce parallel requests

See analytics.md for tracking costs and performance.