Pattern Detection

Pattern detection is the core of Auto-Skill. It operates in multiple layers, each adding more context to raw tool sequences.

Layer 1: Sequence Matching

The foundation. SequenceMatcher uses a sliding-window approach to find repeated tool subsequences.

Algorithm:

1. Extract all subsequences of length 2–10 from each session
2. Count occurrences across sessions (deduplicated per session)
3. Rank by length (longer = more specific) then by frequency

Example:

Session 1: [Read, Edit, Bash, Grep]
Session 2: [Read, Edit, Bash, Grep, Read]
Session 3: [Read, Edit, Bash]

Detected: (Read, Edit, Bash) — 3 occurrences across 3 sessions

Only tool names are matched at this layer — inputs and outputs are used for confidence scoring.

Layer 2: Session Analysis

Goes beyond tool sequences to understand what you were doing.

The SessionAnalyzer classifies each session by:

• Intent — What was the goal? Detected from conversation keywords:

  • debug: "bug", "error", "fix", "not working"
  • implement: "create", "add", "build", "new feature"
  • refactor: "clean up", "reorganize", "optimize"
  • test: "test", "TDD", "coverage"
  • explore: "understand", "explain", "how does"

• Workflow Type — The overall approach: TDD, Refactor-Safe, Debug-Systematic, etc.

• Success Indicators — Tool success rates and whether the problem was resolved.

This context enriches generated skills with "When to Use" guidance.

Layer 3: Code Structure Analysis

The LSPAnalyzer parses your codebase to understand what the tools were operating on.

Capabilities:

• TypeScript/JavaScript — AST analysis (classes, functions, decorators, signatures)

Extracts:

• Symbol definitions (classes, functions, variables)
• Import/dependency graph
• Entry points (main functions, CLI commands)
• File-to-module relationships

This allows skills to reference specific code structures rather than generic file paths.

Layer 4: Design Pattern Recognition

The DesignPatternDetector identifies 18 patterns across three categories:

Architectural Patterns

MVC, Repository, Factory, Singleton, Strategy, Observer, Adapter, Dependency Injection

Coding Patterns

Error-First Handling, REST API Design, Async Pattern, Decorator Pattern, Context Manager, Builder Pattern

Workflow Patterns

TDD, Refactor-Safe, Debug-Systematic, Explore-Then-Implement

Each detection includes confidence, indicators (what was found), affected files, and code examples. These are embedded into the generated skill for architectural context.

Confidence Scoring

Every pattern gets a composite confidence score:

confidence = (
    occurrences  × 0.40    # How often it repeats
  + length       × 0.20    # Ideal: 3–5 tools per sequence
  + success_rate × 0.25    # How often it succeeds
  + recency      × 0.15    # Exponential decay over 7 days
)

Score breakdown:

• Occurrences: min(count / 5, 1.0) — saturates at 5 repetitions
• Length: Peaks at 3–5 tools, decreases for shorter or longer sequences
• Success rate: Direct 0.0–1.0 mapping
• Recency: Exponential decay with 7-day half-life

Hybrid boosts:

• +0.10 if Mental Model domain matches
• +0.05 if an external skill from Skills.sh is proven through adoption

Thresholds

Parameter	Default	What It Controls
min_occurrences	3	Minimum repetitions to consider
min_confidence	0.7	Minimum score to suggest
min_sequence_length	2	Shortest pattern
max_sequence_length	10	Longest pattern
lookback_days	7	Analysis time window

All configurable in auto-skill.local.md. See Configuration.