Refactor Skill
A production-ready Agent Skill that builds recursive dependency-aware modular architecture from raw context using an atom-of-thought pipeline, semantic graph analysis, adaptive decomposition, and cycle/bridge extraction.
This repository ships a standalone implementation for:
- extracting semantic atoms from files,
- building semantic + dependency graphs,
- recursively splitting low-cohesion modules,
- creating explicit bridge artifacts for weak/long-range dependencies,
- generating a final directed-acyclic module index,
- and emitting auditable graph/atlas artifacts with confidence tracking.
Repository layout
SKILL.md: skill contract (metadata + instructions) loaded by agent skill loaders.scripts/: orchestration engine and implementation.tests/: validation and regression tests.references/: schema and process references used by the skill..claude-plugin/marketplace.json: Claude plugin marketplace descriptor..gitignore: repo hygiene.
Quickstart
python3 scripts/rlm_orchestrator.py --input <INPUT_DIR> --output <OUT_DIR>
Install for Codex/Claude/Cursor/other agents
Via ai-agent-skills (npx skills)
# Install this skill to all supported local agents
npx skills install Zpankz/refactor/refactor
# Install to only Claude's skills directory
npx skills install Zpankz/refactor/refactor --agent claude
# Dry-run preview
npx skills install Zpankz/refactor/refactor --agent claude --dry-run
Claude plugin install
This repo includes a Claude marketplace manifest. After publishing, a Claude user can add the marketplace and install the plugin:
/plugin marketplace add https://github.com/Zpankz/refactor
/plugin install refactor@refactor-marketplace
Supported runtime options
--mode analysis for audit-only graph extraction.--mode orchestration for recursive refinement and module extraction.- configurable thresholds:
--min-similarity, --target-confidence, --min-bridge-weight, --max-bridges, and max depth settings.
Security and scope
This project does not include any secrets, credentials, or deployment keys. It is deterministic and keeps model context output artifacts on disk for traceability.
Zpankz
