Reeper

Reeper safely adapts external GitHub repositories to your existing project instead of blindly copying their architecture.

It analyzes the source and target independently, builds an evidence-backed conflict matrix, interviews you one material decision at a time, writes an Integration Contract, waits for explicit approval, implements in isolation, and verifies the result against the contract.

Why Reeper

Coding agents are good at copying code and bad at noticing when the copied assumptions quietly replace your authentication, database, billing, deployment, design system, or operational conventions. Reeper turns repository adaptation into a durable reconciliation process.

Core principles:

• Target-preserving by default — existing project systems remain authoritative unless you approve replacement.
• Evidence before questions — Reeper inspects both repos first and asks only what code cannot answer.
• One conflict at a time — every material decision is recommended, answered, and recorded.
• Contract before code — implementation is blocked until you approve the Integration Contract.
• Untrusted source model — source scripts, installs, hooks, submodules, and binaries are not executed during analysis.
• Resumable artifacts — sessions survive context loss and can be resumed later.
• Provenance-aware — copied, adapted, replaced, rejected, and reimplemented components are tracked.

Commands

/reeper

A single entry point that routes to the workflow below and works against either install form.

/reeper https://github.com/example/source-repo  Bring its analytics dashboard into this app
/reeper resume
/reeper skillify .  Create a user-invoked command that audits our lead import pipeline

Plugin commands are namespaced (/reeper:import), so the bare /reeper is a separate one-file command. Claude Code resolves a top-level command only from ~/.claude/commands/<name>.md:

curl -sL https://raw.githubusercontent.com/leadgenjay/Reeper/main/commands/reeper.md \
  -o ~/.claude/commands/reeper.md

It prefers the plugin install when both are present, because the plugin also registers the read-only subagents. With neither installed it prints both install commands and stops.

/reeper:import

Import or adapt a repository into the current project.

/reeper:import https://github.com/example/source-repo
Bring its analytics dashboard into this app, but preserve our Supabase auth,
Stripe billing, shadcn components, and Vercel deployment.

Supported integration modes:

• transplant selected features
• reimplement behavior against target-native systems
• fork the source as the new base
• vendor an isolated library/package
• reference source patterns without copying code

/reeper:resume

Resume the most recent incomplete Reeper session without repeating completed analysis or interview decisions.

/reeper:resume

/reeper:skillify

Turn a repository or stable workflow into a callable Agent Skill or Claude Code plugin.

/reeper:skillify .
Create a user-invoked command that audits and repairs our lead import pipeline.

Install from GitHub

After this repository is published at leadgenjay/Reeper:

/plugin marketplace add leadgenjay/Reeper
/plugin install reeper@reeper

Then invoke:

/reeper:import <repo-url> [goal]

Install as a single Agent Skill

Reeper also ships as one flattened Agent Skill for people who install skills rather than plugins:

curl -sL 'https://leadgenjay.com/api/skills/install.sh?items=reeper' | bash

That form is a build artifact, not a second source of truth. Regenerate it with:

make export     # -> dist/claude-skills/skills/reeper/

The plugin lays its skills out as plugins/reeper/skills/<name>/SKILL.md and reaches the shared guides with ${CLAUDE_SKILL_DIR}/../../references/.... That climb is correct inside a plugin and wrong for a skill install, where every file lands in one directory, so scripts/export_marketplace_skill.py rewrites the paths, strips the per-skill frontmatter into workflows/*.md, and renders the router SKILL.md plus manifest.yaml from marketplace/. tests/test_marketplace_export.py fails the build if any emitted file escapes the skill root or points at a path that does not exist.

The plugin form remains the richer install: namespaced commands and automatically registered subagents.

Test locally

From Claude Code:

/plugin marketplace add /absolute/path/to/Reeper
/plugin install reeper@reeper
/plugin validate /absolute/path/to/Reeper

Or with the CLI:

claude plugin validate .

What Reeper creates in the target

Each adaptation gets a durable session:

.reeper/sessions/<session>/
├── manifest.json
├── intake.md
├── source-profile.md
├── target-profile.md
├── source-fingerprint.json
├── target-fingerprint.json
├── conflict-matrix.md
├── decisions.md
├── integration-contract.md
├── plan.md
├── tasks.md
├── provenance.json
└── verification.md

Local source checkouts and worktrees are ignored under .reeper/cache/ and .reeper/worktrees/.

Workflow

Source repo ─┐
             ├─ independent evidence-first profiles
Target repo ─┘
                    ↓
             conflict matrix
                    ↓
          one-question interview
                    ↓
          Integration Contract
                    ↓ explicit approval
          isolated implementation
                    ↓
        contract-based verification
                    ↓
       optional skill/plugin packaging

Safety model

During analysis Reeper may read and statically inspect files, manifests, git metadata, and packed repository output. It must not run external repository install, build, setup, migration, seed, hook, application, container, or arbitrary scripts before trust approval.

The bundled safe_clone.sh disables hooks, skips submodules, and skips Git LFS smudge. It still treats the cloned files as untrusted.

Requirements

• Claude Code with plugin support
• Git
• Python 3.10+
• Optional: Node/npm for Repomix (npx repomix@latest)

The core helper scripts use only the Python standard library.

Inspiration

Reeper combines ideas from:

• Repomix: remote repository packing, incremental codebase exploration, security-aware ingestion, and generated reference skills
• GitHub Spec Kit: durable phase artifacts, clarification, planning, task generation, and cross-artifact validation
• Superpowers: one-question-at-a-time design discovery and a hard implementation gate
• Anthropic Agent Skills and Claude Code plugins: progressive disclosure, invocable skills, supporting scripts/references, subagents, and marketplace distribution

Reeper is an independent project and does not include code copied from those projects.

Status

v0.1.0 is an initial functional plugin scaffold with the complete workflow, deterministic session/fingerprint/scaffolding helpers, validation, and tests. It should be exercised against real repositories before a stable release.

License

MIT