unwind

Unwind

Skills library for reverse engineering codebases. Produces complete, machine-readable documentation and phased rebuild plans to reliably re-build the service or application in a new technology or modernised framework.

Purpose

Generate documentation that enables an AI agent to rebuild your system in a different language or framework while maintaining:

• External API contract compatibility
• Business logic accuracy
• Data model integrity

Unwind is hybrid: a deterministic scanner (@unwind/core, built on tree-sitter) produces the verifiable ground truth — file inventory, structural symbols, import graph, and a first-pass layer assignment — and LLM specialists add the semantic rebuild documentation. Completeness is then verified by set arithmetic (scan − docs), not asserted. Symbol extraction supports TypeScript/JavaScript, Python, Rust, Java, and C#; other languages get file-level coverage, and if Node/pnpm is unavailable Unwind falls back to a pure-LLM flow.

It can also execute the rebuild: uw-build interviews you about scope and order, dispatches technology-agnostic builder agents that reproduce each layer's [MUST] contracts in the target stack, then re-scans the rebuilt repo and diffs it against the source to measure completeness — a before/after picture, again by set arithmetic rather than assertion.

Quick Start

Install

/plugin marketplace add cliftonc/unwind
/plugin install unwind@cliftonc

Restart Claude Code after installation.

Use

New here? Just run the entry point:

Use unwind:uw-start   # orients you, checks prereqs, drives the whole pipeline

Or run the phases manually:

1. Use unwind:uw-scan              # deterministic scan → architecture.md (+ scan-manifest.json)
2. Use unwind:uw-dashboard   # visualize it — works straight after the scan
3. Review docs/unwind/architecture.md
4. Use unwind:uw-analyze # seed → analyze → verify coverage → complete
5. Use unwind:uw-plan   # → REBUILD-PLAN.md
6. Use unwind:uw-dashboard   # re-open to explore coverage, priorities & contracts
7. Use unwind:uw-build   # execute the rebuild in the target stack (+ verification graph)

You can run unwind:uw-dashboard any time after step 1 — right after the scan it shows the structural graph (everything scanned); after analysis it fills in coverage and the MUST/SHOULD/DON'T priorities.

The first run builds the scanner automatically (pnpm install && pnpm build via ensure_unwind_core). It needs Node + pnpm; without them, Unwind falls back to a pure-LLM flow.

Output:

• docs/unwind/REBUILD-PLAN.md - Strategic rebuild approach
• docs/unwind/layers/*/ - Detailed layer analysis (folder per layer)
• docs/unwind/rebuild-graph.json - Knowledge graph for the dashboard
• docs/unwind/rebuild-verification-graph.json - Source→target completeness (after uw-build)
• docs/unwind/.cache/ - Deterministic artifacts: scan-manifest.json (ground truth), seeds/ (per-layer checklists), coverage/ (per-layer coverage reports), rebuild-state.json (rebuild ledger)

---

Updating

/plugin marketplace update cliftonc
/plugin uninstall unwind
/plugin install unwind@cliftonc

---

How it works

Unwind interleaves deterministic scripts (which own the verifiable facts) with LLM sub-agents (which add semantic judgment). The deterministic layer is what makes "did we document everything?" a checkable question instead of a hopeful one.

Step by step