RForge - Orquestrador do Ecossistema R
Data-Wise/rforge
Orquestrador completo de pacotes R com validação CRAN e CI/CD integrados
Ver no GitHubMCP server para desenvolvimento de pacotes R com 33 comandos especializados em análise de dependências, validação, testes e preparação para submissão no CRAN. Integra-se nativamente ao Claude Code com módulos Python puros, dispensando dependências externas além de Python 3.10+.
Instalação
npx convez-ai Data-Wise/rforge1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Desenvolvimento e manutenção de pacotes R
- Preparação de submissões CRAN com análise pré-submissão completa
- Análise de dependências e saúde do ecossistema R
- Automação de testes, linting, documentação e cobertura de código
- Integração com workflows de CI/CD para pacotes R
Veredito

Ferramenta sólida e madura (v2.2.0) para quem trabalha com pacotes R. A absorção completa do MCP e o suporte a 33 comandos especializados a torna atraente para fluxos profissionais; a curva de aprendizado é acentuada, mas o retorno é significativo em automatização de ciclos complexos de desenvolvimento e submissão.
README
A tradução para o português deste README ainda está na fila.
RForge Plugin
R package ecosystem orchestrator for Claude Code — 33 commands, R-aware hooks, validation skills.
Self-contained R package analysis for Claude Code. As of v1.3.0 the plugin is fully self-sufficient — pure-Python lib/ modules handle discovery, dependencies, status, and init. No MCP server required.
What's new in v2.2.0
- 5 new
r:CRAN-submission commands:r:revdep,r:goodpractice,r:winbuilder,r:rhub,r:cran-prep— full pre-submission gate that runs document→lint→spell→urlcheck→test→coverage→check(--as-cran)→revdep, generatescran-comments.md, and returns aready/warn/blockedverdict. r:checkNOTE classifier: notes are now classified asspurious(expected on CRAN submission) orreal(needs attention) usingnotes_classifiedin the envelope.- Total: 28 → 33 commands.
What's new in v2.1.0
- 🔬 12 new
r:commands — full R package dev cycle + quality layer:r:load,r:document,r:test,r:coverage,r:build,r:install,r:site,r:cycle,r:lint,r:spell,r:urlcheck,r:style. - 🐍
lib/rcmd.py— new pure-Python module driving each command via lower-level R engines (rcmdcheck,pkgbuild,roxygen2,testthat,pkgload,covr,pkgdown,lintr,spelling,urlchecker,styler). JSON output, never regex-scraped. Optional engines degrade gracefully to 🟡 + install hint. - Total: 16 → 28 commands.
r:checkretrofit: now drives its report fromlib.rcmdrather than raw subprocess.
What's new in v2.0.0 (BREAKING)
- 🔀 3 commands renamed for cleaner namespacing —
/rforge:doc-check→/rforge:docs:check,/rforge:ecosystem-health→/rforge:health,/rforge:rpkg-check→/rforge:r:check. The other 13 commands are unchanged. Typing an old name produces a helpful rename-error pointing at the new name — no silent failures. Seedocs/migration/v2.0.0-rename.mdfor the full mapping table and asedrecipe to mass-update local scripts.
What's new in v1.3.0
- 🎯 MCP absorption complete —
rforge-mcphas been absorbed into the plugin. All 7 implemented tools now ship as pure-Pythonlib/modules. The MCP server is no longer required (and is being archived). Seedocs/migration/rforge-mcp-deprecation.md. - 🐍
lib/status.py— ecosystem health snapshot (DESCRIPTION+.STATUSparsing).python3 -m lib.status [--path .] [--format text|json]. - 🌱
lib/init.py— initialize~/.rforge/context.jsonfor cross-package state.python3 -m lib.init [--quick]. New/rforge:initcommand. - 📦 No runtime dependencies beyond Python 3.10+ — the plugin works on any system with a modern Python.
What's new in v1.2.0
- 🛒 Marketplace install — one-shot setup via
/plugin marketplace add Data-Wise/rforge(no clone, no symlinks). Seedocs/configuration.mdfor tunable options (CRAN mirror, vignette engine, R version pin, CLAUDE.md budget). - 🪝 R-aware
PreToolUsehook — four rules that fire on everyWrite/Edit: blocks hand-edits to roxygen-generatedman/*.Rd, warns whenR/*.Redits may need NAMESPACE/DESCRIPTION sync, warns on non-SemVerDESCRIPTIONVersion bumps, warns on writes outside the active worktree. Diagnostic, not adversarial — only theman/*.Rdrule blocks. Seedocs/hooks-and-skills.md. - 🔍
description-syncvalidation skill — pure-shell sanity check thatDESCRIPTIONVersion matches the top entry inNEWS.md/CHANGELOG.md. Catches the most common pre-CRAN release-prep failure. No R required. - 📐 Plugin Surface architecture diagram — new Mermaid diagram in
docs/architecture.mdshowing how marketplace, config, commands, agents, hooks, and skills relate.
Full changelog: CHANGELOG.md.
Quick Start
# Quick analysis (< 30 seconds)
/rforge:analyze "Update RMediation bootstrap algorithm"
# Ultra-fast (< 10 seconds)
/rforge:quick
# CRAN submission gate (per-package)
/rforge:r:cran-prep
# Ecosystem rollup (2-5 minutes)
/rforge:thorough "Prepare for CRAN release"
Features
✨ Auto-delegation - Recognizes task patterns, selects appropriate tools
⚡ Parallel execution - Invokes multiple lib/ modules simultaneously
📊 Live progress - Real-time updates as tools complete
🎯 Smart synthesis - Combines results into actionable summary
🧠 ADHD-friendly - Fast feedback, clear structure, visual progress
🐍 Pure-Python lib/ - lib/discovery.py, lib/deps.py, lib/status.py, lib/init.py (pure Python, no R); lib/rcmd.py runs R engines for the r:* dev-cycle + CRAN commands. No MCP server, no Node.js. See docs/lib-modules.md.
How It Works
User Request
↓
Pattern Recognition (CODE_CHANGE, BUG_FIX, etc.)
↓
Tool Selection (impact, tests, docs, health)
↓
Parallel `lib/` invocations (python3 -m lib.* in subprocess)
↓
Results Synthesis (impact + quality + maintenance + next steps)
↓
Actionable Summary
Skills
/rforge:analyze
Balanced analysis with recommendations (< 30 seconds)
/rforge:analyze "Update code"
/rforge:quick
Ultra-fast status check (< 10 seconds)
/rforge:quick
/rforge:thorough
Comprehensive analysis with R CMD check (2-5 minutes)
/rforge:thorough "Prepare for CRAN"
Requirements
- Python 3.10+ (the
lib/modules run viapython3 -m lib.<tool>) - R Environment
- R >= 4.0.0
- devtools package (optional, for
/rforge:thorough) - testthat package (optional)
- covr package (optional, for coverage analysis)
- Claude Code CLI or Claude Desktop
- Plugin works in both environments
Migrating from v1.2.x? If you have an
mcpServers.rforgeentry in~/.claude/settings.json, you can remove it — v1.3.0 no longer needs the MCP server. Seedocs/migration/rforge-mcp-deprecation.md.
Installation
Option 1: Claude Code Marketplace (Recommended)
From inside Claude Code:
/plugin marketplace add Data-Wise/rforge
/plugin install rforge
The marketplace install reads .claude-plugin/marketplace.json from the
repository, fetches the plugin into ~/.claude/plugins/rforge, and wires
it up automatically. Works on macOS, Linux, and Windows. Update later with
/plugin update rforge.
Option 2: Homebrew (macOS)
# Add the Data-Wise tap
brew tap data-wise/tap
# Install rforge plugin
brew install rforge
The Homebrew formula automatically:
- Installs the plugin to
~/.claude/plugins/rforge - Makes it available in Claude Code CLI and Claude Desktop
Option 3: npm (When published)
# Install from npm (after publishing)
npm install -g @data-wise/rforge-plugin
# Plugin will auto-install to ~/.claude/plugins/rforge
Option 4: Manual Installation (Local Development)
For Claude Code CLI and Claude Desktop:
# Clone the repository
git clone https://github.com/Data-Wise/rforge.git
cd rforge
# Install in development mode (symlink - changes reflected immediately)
ln -s $(pwd) ~/.claude/plugins/rforge
# Or install in production mode (copy - stable)
cp -r . ~/.claude/plugins/rforge
Installation locations:
- Plugin directory:
~/.claude/plugins/rforge - Commands:
~/.claude/plugins/rforge/commands/ - Agent:
~/.claude/plugins/rforge/agents/orchestrator.md - Lib:
~/.claude/plugins/rforge/lib/
Verify Installation
Step 1: Check plugin
# Check plugin directory exists
ls -la ~/.claude/plugins/rforge
# Verify plugin.json
cat ~/.claude/plugins/rforge/.claude-plugin/plugin.json
Step 2: Check Python
python3 --version # Expect 3.10+
Step 3: Test end-to-end
# Navigate to an R package
cd ~/projects/r-packages/active/RMediation
# Start Claude Code
claude
# Test a command
/rforge:status
Expected output:
📊 RMediation - Single Package
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Version: 2.4.0
Tests: 187 passing
Health: 87/100 (B+)
Using in Claude Code CLI
After installation, commands are immediately available:
# Navigate to any R package
cd ~/my-r-package
# Start Claude Code
claude
# Use rforge commands
/rforge:analyze "Update bootstrap algorithm"
/rforge:status
/rforge:quick
Using in Claude Desktop App
RForge automatically loads when you open Claude Desktop. Commands work the same way:
- Open Claude Desktop app
- Navigate conversation to an R package directory (or specify path)
- Use slash commands:
/rforge:analyze,/rforge:status, etc.
Migration
- From v1.2.x: the
mcpServers.rforgeentry in~/.claude/settings.jsonis no longer needed in v1.3.0. Remove it manually (we don't auto-edit user settings). Seedocs/migration/rforge-mcp-deprecation.mdfor the full migration table.
Migration from rforge-orchestrator (historical, pre-v1.3.0)
Only relevant if you skipped v1.2.x and are coming from a much older
rforge-orchestratorsetup. v1.3.0 doesn't use any MCP server, so the end state is just: remove these entries entirely.
If your ~/.claude/settings.json has either an rforge-orchestrator or rforge-mcp entry under mcpServers, delete it:
// Delete BOTH of these patterns if present:
{
"mcpServers": {
"rforge-orchestrator": { ... },
"rforge-mcp": { ... }
}
}
In v1.3.0+ the plugin runs entirely in-process via lib/ modules — there's no MCP server to configure. The plugin loads via /plugin install rforge; no settings.json entries are required.
For the full transition narrative, see docs/migration/rforge-mcp-deprecation.md.
Pattern Recognition
The orchestrator automatically detects what you're doing:
| Your Request | Pattern | Tools Used |
|---|---|---|
| "Update algorithm" | CODE_CHANGE | impact, tests, docs, health |
| "Add function" | NEW_FUNCTION | detect, tests, docs |
| "Fix bug" | BUG_FIX | tests, impact |
| "Update docs" | DOCUMENTATION | docs, detect |
| "Release 2.1.0" | RELEASE | health, impact, tests, docs |
Example Outputs
Quick Analysis
⚡ Quick analysis running...
✅ Done! (8.2 seconds)
📊 QUICK SUMMARY:
✅ Impact: 2 packages (MEDIUM)
✅ Tests: 187/187 passing
⚠️ Docs: NEWS.md needs update
✅ Health: 87/100 (B+)
Full Analysis
🎯 IMPACT: MEDIUM
• 2 packages affected (mediate, sensitivity)
• Estimated cascade: 4 hours
✅ QUALITY: EXCELLENT
• Tests: 187/187 passing (94% coverage)
• CRAN: Clean
📝 MAINTENANCE: 2 items
• NEWS.md needs entry
• Vignette example outdated
📋 NEXT STEPS:
1. Implement changes (3 hours)
2. Auto-fix documentation
3. Run cascade for dependents
Architecture
┌─────────────────────────────────────┐
│ Claude Code Session │
│ │
│ ┌───────────────────────────────┐ │
│ │ RForge Orchestrator Plugin │ │
│ │ │ │
│ │ • Pattern Recognition │ │
│ │ • Tool Selection │ │
│ │ • Parallel Execution │ │
│ │ • Progress Display │ │
│ │ • Results Synthesis │ │
│ └───────┬───────────────────────┘ │
│ │ │
└──────────┼──────────────────────────┘
│ Parallel python3 -m lib.* calls
↓
┌──────┴──────┬──────┬──────┐
↓ ↓ ↓ ↓
[discovery] [deps] [status] [init]
↓ ↓ ↓ ↓
(8s) (5s) (3s) (7s)
│ │ │ │
└─────────────┴──────┴──────┘
│
↓
Results synthesized
by orchestrator
Performance
| Mode | Tools | Time | Use Case |
|---|---|---|---|
| Quick | 4 quick tools | ~10s | Status check |
| Analyze | 4 quick + synthesis | ~30s | Daily dev |
| Thorough | Background R | 2-5m | Pre-release |
Time savings: Parallel execution = 4x faster than sequential!
ADHD-Friendly Design
- Fast feedback - Results in seconds, not minutes
- Clear structure - Consistent output format
- Visual progress - See what's happening
- Actionable - Always provides next steps
- Interruptible - Can cancel/resume anytime
- Incremental - Results stream as they complete
Troubleshooting
"python3: command not found"
# Verify Python 3.10+ is on PATH
python3 --version
# macOS: install via Homebrew
brew install python@3.12
"Package not detected"
# Run from package directory
cd /path/to/package
# Or specify path explicitly
/rforge:analyze --package /path/to/package
"Analysis too slow"
- Use
/rforge:quickfor fast status - Use
/rforge:analyzefor balanced speed/depth - Only use
/rforge:thoroughwhen needed
Configuration
Plugin settings in plugin.json:
{
"settings": {
"default_mode": "quick", // quick, analyze, or thorough
"parallel_execution": true, // Run tools in parallel
"show_progress": true, // Show progress bars
"auto_synthesize": true // Auto-generate summary
}
}
Development
Plugin structure:
~/.claude/plugins/rforge/
├── .claude-plugin/
│ ├── plugin.json # Plugin manifest (v2.2.0)
│ ├── marketplace.json # Marketplace install metadata
│ ├── config.json # User-tunable options (CRAN mirror, etc.)
│ ├── hooks/
│ │ └── pretooluse.py # R-aware Write/Edit guard (4 rules)
│ └── skills/
│ └── validation/
│ └── description-sync.md # DESCRIPTION ↔ NEWS.md drift check
├── commands/ # 33 slash commands (/rforge:*)
├── agents/
│ └── orchestrator.md # Pattern recognition + delegation
├── lib/ # Pure-Python analysis modules
│ ├── discovery.py # Package detection + ecosystem layout
│ ├── deps.py # Dependency graph + impact
│ ├── status.py # DESCRIPTION + .STATUS health snapshot
│ ├── init.py # ~/.rforge/context.json initializer
│ ├── rcmd.py # R dev-cycle + quality + CRAN-submission engines (v2.2.0)
│ └── formatters.py # Output formatting helpers
└── docs/ # User-facing docs
Contributing
Ideas for improvement:
- Add caching for repeated analyses
- Track user preferences for tool selection
- Add more pattern types
- Improve time estimates
- Add result export (markdown, JSON)
License
MIT
Links
- v1.3.0 absorbed the prior
rforge-mcpprototype (local-only, never published to GitHub or npm). Seedocs/migration/rforge-mcp-deprecation.mdfor the migration narrative. - Claude Code: https://claude.com/code
- Documentation: See
docs/folder
Version: 2.2.0 Status: Active development Compatibility: Claude Code 0.1.0+