ai-drift 0.1.1

AI agent conversation drift analyzer - identifies gaps between what AI agents did and what users wanted

Drift

Quality assurance for AI-augmented codebases - validates that your project follows best practices for effective AI agent collaboration.

What It Does

Drift ensures your AI-augmented development environment is optimized for productivity. It validates both the quality of AI interactions and the health of your project configuration.

Think of it as: A comprehensive testing and linting tool for AI-first development - catching issues in conversation patterns, documentation quality, dependency management, and project structure.

Two-Level Validation

1. Conversation Quality Analysis Analyzes AI agent conversation logs to detect patterns where work diverged from user intent:

• Incomplete work and premature task abandonment
• Missed delegation opportunities to specialized agents
• Ignored skills, commands, or workflow automation
• Deviation from documented project guidelines

2. Project Structure Validation Programmatically validates your AI collaboration setup:

• Dependency health: Detects redundant transitive dependencies in commands, skills, and agents
• Link integrity: Validates all file references and resource links in documentation
• Completeness checks: Ensures skills, commands, and agents have required structure
• Consistency validation: Detects contradictions between commands and project guidelines
• Required files: Verifies essential configuration files exist (e.g., CLAUDE.md)

Key Features

• Multi-layered analysis: Combines LLM-based conversation analysis with fast programmatic validation
• Project-aware: Automatically discovers and validates commands, skills, agents, and MCP servers
• Flexible execution: Run all checks, or use --no-llm for fast programmatic-only validation
• Multi-provider: Anthropic API and AWS Bedrock with Claude models (Sonnet, Haiku)
• Multi-agent support: Currently supports Claude Code
• Rich output: Markdown with colors (for terminals) or structured JSON
• Configurable rules: Extensible YAML-based rule system for custom validations

Installation

View on PyPI: ai-drift

# Install from PyPI
pip install ai-drift

# Or with uv
uv pip install ai-drift

For development:

# Clone repository
git clone https://github.com/jarosser06/drift.git
cd drift
uv pip install -e ".[dev]"

Quick Start

Provider Setup

Option 1: Anthropic API (Recommended)

export ANTHROPIC_API_KEY=your-api-key

Option 2: AWS Bedrock

aws configure

Usage

# Run full analysis on latest conversation
drift

# Fast programmatic-only validation (no LLM calls)
drift --no-llm

# Analyze last 7 days with JSON output
drift --days 7 --format json

# Check specific rules only
drift --rules command_broken_links,skill_duplicate_dependencies

# Use different model for analysis
drift --model sonnet

# Disable caching for fresh analysis
drift --no-cache

# Use custom cache directory
drift --cache-dir /tmp/my-cache

Response Caching

Drift automatically caches LLM responses to reduce API costs and speed up re-analysis:

• Smart invalidation: Cache automatically invalidates when file content changes
• Content-based: Uses SHA-256 hashing to detect changes
• TTL support: Default 24-hour cache expiration (configurable)
• Per-file caching: Each file + rule combination cached separately

Configure in .drift.yaml:

cache_enabled: true          # Enable/disable caching (default: true)
cache_dir: .drift/cache      # Cache directory (default: .drift/cache)
cache_ttl: 86400             # TTL in seconds (default: 86400 = 24 hours)

CLI overrides:

• --no-cache: Disable caching for this run
• --cache-dir <path>: Use custom cache directory
• --no-parallel: Disable parallel execution of validation rules (use sequential execution)

Example Output

# Drift Analysis Results

## Summary
- Total conversations: 3
- Total rule violations: 5
- Rules checked: 12
- Rules passed: 7
- Rules warned: 2
- Rules failed: 3

## Rules Passed ✓

- **documentation_gap**: No issues found
- **command_broken_links**: All file references valid
- **skill_duplicate_dependencies**: No redundant dependencies
- **claude_md_missing**: CLAUDE.md exists

## Failures

### command_duplicate_dependencies

*Commands should only declare direct dependencies. Transitive dependencies are automatically included.*

**Bundle:** create-pr command
**Files:** .claude/commands/create-pr.md
**Issue:** Command declares both `pr-writing` skill and `github-operations` skill, but `pr-writing` already depends on `github-operations`
**Expected:** Remove `github-operations` from command dependencies

### skill_completeness

*Incomplete skills create confusion and slow development. Skills must be self-contained with clear examples.*

**Bundle:** testing skill
**Files:** .claude/skills/testing/SKILL.md
**Issue:** Skill references `./examples/test_example.py` which doesn't exist
**Expected:** Include referenced examples or remove broken references

## Warnings

### incomplete_work

*AI stopping before completing full scope wastes user time and breaks workflow momentum.*

**Session:** abc-123
**Agent Tool:** claude-code
**Turn:** 3
**Observed:** Implemented login form without validation
**Expected:** Complete login system with validation and error handling
**Context:** User had to explicitly request validation in next turn

Configuration

Default config auto-generates at ~/.config/drift/config.yaml. Override per-project with .drift.yaml.

Provider Configuration

Configure LLM providers in .drift.yaml:

Anthropic API:

providers:
  anthropic:
    provider: anthropic
    params:
      api_key_env: ANTHROPIC_API_KEY  # optional, defaults to ANTHROPIC_API_KEY

models:
  claude-sonnet:
    provider: anthropic
    model_id: claude-sonnet-4-5-20250929
    params:
      max_tokens: 4096
      temperature: 0.0
  claude-haiku:
    provider: anthropic
    model_id: claude-haiku-4-5-20251001
    params:
      max_tokens: 4096
      temperature: 0.0

AWS Bedrock:

providers:
  bedrock:
    provider: bedrock
    params:
      region: us-east-1

models:
  sonnet:
    provider: bedrock
    model_id: us.anthropic.claude-3-5-sonnet-20240620-v1:0
    params:
      max_tokens: 4096
      temperature: 0.0

Validation Categories

Conversation Analysis Rules (LLM-based)

Analyze AI agent conversation patterns:

• incomplete_work - AI stopped before completing full scope
• agent_delegation_miss - AI did work manually instead of using agents
• skill_ignored - AI didn't use available skills
• workflow_bypass - User manually executed steps that commands automate
• prescriptive_deviation - AI ignored explicit workflow documentation
• no_agents_configured - Project lacks agent definitions

Project Validation Rules (Programmatic)

Fast validation without LLM calls:

• command_duplicate_dependencies - Redundant transitive skill dependencies in commands
• skill_duplicate_dependencies - Redundant transitive dependencies in skills
• agent_duplicate_dependencies - Redundant transitive dependencies in agents
• command_broken_links - Broken file references in command documentation
• skill_broken_links - Broken file references in skill documentation
• agent_broken_links - Broken file references in agent documentation
• claude_md_missing - Missing CLAUDE.md configuration file
• skill_completeness - Skills missing essential structure or examples
• agent_completeness - Agents missing scope definition or dependencies
• command_completeness - Commands missing execution steps or prerequisites
• command_consistency - Commands contradicting project guidelines

Custom Rule Example

Add custom rules to .drift.yaml:

rule_definitions:
  command_broken_links:
    description: "Commands contain broken file references"
    scope: project_level
    document_bundle:
      bundle_type: command
      file_patterns:
        - .claude/commands/*.md
      bundle_strategy: individual
    phases:
      - name: check_links
        type: markdown_link
        description: "Validate all markdown links and file paths"
        failure_message: "Found broken links"
        expected_behavior: "All file references should be valid"
        params:
          check_local_files: true
          check_external_urls: false

CLI Options

• --format (-f): Output format (markdown or json)
• --scope (-s): Analysis scope (conversation, project, or all)
• --agent-tool (-a): Specific agent tool to analyze (e.g., claude-code)
• --rules (-r): Comma-separated list of specific rules to check
• --latest: Analyze only the latest conversation
• --days (-d): Analyze conversations from last N days
• --all: Analyze all conversations
• --model (-m): Override model for analysis (sonnet, haiku)
• --no-llm: Skip LLM-based rules, run only programmatic validation (fast)
• --no-parallel: Disable parallel execution of validation rules
• --project (-p): Project path (defaults to current directory)
• --verbose (-v): Increase verbosity (-v, -vv, -vvv)

Development

# Run tests (requires 90%+ coverage)
./test.sh

# Run linters
./lint.sh

# Auto-fix formatting
./lint.sh --fix

Use Cases

During Development:

• Run drift --no-llm before commits to catch broken links and dependency issues
• Validate skill/command documentation is complete and consistent

In CI/CD:

• Enforce documentation quality standards
• Prevent broken resource references from merging

Periodic Reviews:

• Analyze conversation patterns to identify workflow improvements
• Find opportunities to better leverage agents, skills, and commands
• Ensure project customizations are being utilized

Project Setup:

• Validate new AI agent configurations
• Ensure documentation follows best practices
• Catch structural issues before they impact productivity

License

MIT