Skip to content

Repository files navigation

intent-drift

A skill that uses the Intent Alignment Engine to analyze intent drift in AI-assisted development.

🎯 Purpose

Detects when AI coding agents begin solving a different problem than originally requested by monitoring alignment between:

  • Original goals
  • Current execution plans
  • File changes and behavior

🛠️ Features

  • Evidence-based drift detection using multiple providers
  • Explainable assessments with detailed evidence tracking
  • Real-time monitoring and timeline tracking
  • Pluggable architecture for custom evidence providers
  • Type-safe with comprehensive validation
  • Exportable reports in multiple formats

🚀 Quick Start

# Import into any Claude Code agent
cd ~/.claude/skills/intent-drift
./analyze-code

Usage Examples

# Basic usage
/intent-drift
--original-goal "Reduce application memory usage"
--current-plan "Optimize startup performance"
--context "Edited: main.py, startup.py"

# With auto-collection of git context
/intent-drift
--original-goal "Improve response time"
--current-plan "Add database indexing"
--auto-context

📊 Analysis Output

Intent Alignment Report

Overall Alignment: 68%
Status: Moderate Drift
Confidence: 89%

Evidence:
✓ Goal partially overlaps
✓ Constraints remain satisfied
⚠ Edited files primarily affect startup logic
⚠ Implementation no longer targets memory allocation

Risk: High - additional work unlikely to improve memory usage

Recommendation: Pause and confirm alignment before continuing

🏗️ Architecture

intent-drift/
├── __init__.py              # Skill entrypoint
├── analyzer.py              # Core analysis logic
├── providers/               # Evidence providers
├── exporters/               # Report exporters (text, markdown, json)
├── config/                  # Configuration defaults
├── docs/                    # Usage documentation
└── examples/                # Usage examples

🔧 Configuration

Required Configuration

# config/defaults.yaml
analysis:
  threshold: 75              # Minimum alignment score (%)
  confidence: 80            # Minimum confidence (%)
  providers:
    enabled:               # Which providers to use
      - goal_provider
      - constraint_provider
      - execution_provider
      - scope_provider
  
  evidence_providers:
    goal_provider:
      weight: 0.25
      thresholds:
        match_score: 80
        drift_score: 60
    
    constraint_provider:
      weight: 0.20
      thresholds:
        violation_score: 90
        partial_compliance: 70

Customization

# Edit config file
nano ~/.claude/skills/intent-drift/config/user.yaml

# Reset to defaults
./analyze-code --reset-config

📁 Integration

With Git Repos

Automatically analyzes:

  • Git diffs between commit points
  • File modification patterns
  • Commit message trends
  • Branch divergence

With Codebase Features

Analyzes:

  • Type checking evidence
  • Build system outputs
  • Test coverage changes
  • Performance metrics

🔌 Extending the Skill

Adding New Evidence Providers

# New providers go in providers/
class CustomEvidenceProvider:
    def __init__(self):
        self.name = "custom_provider"
        self.weight = 0.15
    
    def collect(self, context):
        # Implementation
        return [Evidence(...)]

Custom Export Formats

# New exporters go in exporters/
class CsvExporter:
    def export(self, report, output_path):
        # CSV implementation
        pass

📚 Documentation

See the docs/ directory for:

🤝 Contributing

See CONTRIBUTING.md for:

  • Code style guidelines
  • Testing requirements
  • Documentation standards

📄 License

MIT License - see LICENSE file for details.

🙏 Acknowledgments

Based on the Intent Alignment Engine by Shaurya Gangrade.

About

An AI skill that measures the drift of intent of an AI model.

Resources

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages