henriquescastilho Capability Router — EOI-OS
Index resumido pra Claude rotear pro tool certo SEM varrer 791 skill descriptions.
PLUGINS (59 ativos) por capability
Comunicação / Productivity
- productivity@knowledge-work-plugins: Slack, Notion, Asana, Linear, Jira, ClickUp, Atlassian, Monday, Gmail, Google Calendar
- enterprise-search: Guru, Gmail, Calendar cross-tool search
- commit-commands: git commit, push, PR workflow
Data / Analytics
- data@knowledge-work-plugins: Snowflake, Databricks, BigQuery, Hex, Amplitude, Definite
- posthog@claude-plugins-official: analytics, feature flags, experiments
Code Intelligence
- github: PRs, issues, repos (HTTP MCP)
- understand-anything: knowledge graph +
/understand-*commands - plugin-dev: plugin development toolkit
- firecrawl: web scrape/crawl/search
- context7: up-to-date library docs lookup
Security
- security-guidance: auto-review code for vulns (writes hooks)
- semgrep: static analysis security
- sonatype-guide: supply chain dep security
- ruflo-security-audit: security review, dependency scanning, CVE
- ruflo-aidefence: PII detection, prompt injection defense
Testing / QA
- playwright: browser automation + E2E
- pr-review-toolkit: 6 specialized PR review agents
- code-review: PR review with confidence scoring
- ruflo-testgen: test gap detection + generation
- coderabbit: AI code review
Observability / Monitoring
- sentry: error tracking, debug prod
- claude-mem@thedotmack: cross-session memory (3 MCP tools)
- ruflo-observability: structured logging + tracing + metrics
- ruflo-cost-tracker: token usage + cost per agent
Multi-agent / Orchestration
- ruflo-swarm: agent teams, hierarchical/mesh/adaptive topologies
- ruflo-autopilot: autonomous /loop task completion
- ruflo-intelligence: self-learning SONA patterns
- pro-workflow: 18 hook events + 5 agents + reference guides
- superpowers: TDD, debugging, collaboration patterns
Memory / RAG
- pinecone: vector DB (HTTP MCP)
- ruflo-rag-memory: HNSW search, AgentDB, semantic retrieval
- ruflo-agentdb: AgentDB memory controllers
- ruflo-ruvector: self-learning vector DB (HNSW, FlashAttention)
- claude-mem: cross-session memory bridge
Documentation
- ruflo-docs: doc generation + drift detection
- claude-md-management: CLAUDE.md audit + improvement
- ruflo-adr: ADR lifecycle (Architecture Decision Records)
- skill-creator: create/improve/measure skills
Deployment / Infra
- vercel: build/deploy web apps + agents (33 hooks)
- ECC: harness performance (64 agents + 261 skills + hooks)
- ruflo-jujutsu: advanced git workflows + risk scoring
- ruflo-workflows: visual workflow automation
Domain-specific
- ruflo-neural-trader: trading models (LSTM/Transformer/N-BEATS) — D2 NOT-RECOMMENDED for DME
- ruflo-market-data: market feeds (financial)
- ruflo-iot-cognitum: IoT device management
- ruflo-ruvllm: local LLM inference
MCP SERVERS standalone (user scope)
- codegraph (
codegraph serve --mcp) → explore codebase. 47% menos tokens vs grep+read. Use SEMPRE quando explore/scout codebase grande. - cve-mcp → CVE/EPSS/CISA KEV/Shodan/VirusTotal lookup (27 tools). Use pra security audits + threat intel.
- claude-skills-mcp (uvx) → discovery de skills disponíveis
- mem-hybrid → memory backend
- magic (@21st-dev) → magic component generation
- n8n-mcp → n8n workflow integration
SKILLS por categoria (~791 total)
EOI-OS first-party (~25 — design DME doctrine)
- criar-tirinhas / criar-tirinhas-care-sa → planilhas médicas cirurgias
- pre-deploy-validator → 23-item checklist antes de prod
- deploy-check → quick pre-deploy checks
- deslop → remove código com cara de IA
- graphify → knowledge graph de codebase
- mcp-builder → criar MCP servers
- notebooklm → Google NotebookLM API
- rebuild-skills, auto-update, verification-before-completion
- repo-fit-analyzer → analisa repos do GitHub vs config EOI-OS
Design (taste-skill 13 — anti-slop UI)
- design-taste-frontend → landing pages, portfolios premium
- brandkit → brand-kit boards, logo systems, identity
- gpt-taste → editorial typography + GSAP motion
- minimalist-ui, industrial-brutalist-ui, high-end-visual-design
- imagegen-frontend-mobile/web → premium screen mockups
- redesign-existing-projects → upgrade existing UI
- stitch-design-taste → Google Stitch DESIGN.md
- image-to-code, full-output-enforcement
Cybersec (mukul975 754 — 26 domínios)
- Cloud Security (60), Threat Hunting (55), Threat Intel (50), Web App (42), Network (40), Malware (39), Forensics (37), SOC Ops (36), IAM (35), + 17 outros
- Trigger words: "audit security", "threat hunt", "CVE", "OWASP", "incident response", "forensics", "pentest", "red team", "compliance"
Plugin-derived (~309 entries)
- firecrawl 10, sentry 26, posthog 55, pro-workflow 11, superpowers 14
- ECC ~416 (testing, security, backend patterns, languages-specific)
- ruflo 4×4 (core/swarm/rag/security)
- plugin-dev 7, sonatype-guide, semgrep, supabase, etc
AGENTS customs (8 EOI-OS + 90+ plugin-derived)
EOI-OS (escolha por intent — per CLAUDE.md routing table)
- scout (Sonnet 4.6) → buscar/explorar código
- deployer (Sonnet 4.6) → pre-deploy validation
- implementer (Opus 4.8) → implementar feature/fix bug
- reviewer (Opus 4.8) → review código
- tester (Opus 4.8) → testes
- security-auditor (Opus 4.8) → audit segurança
- architect (Fable 5) → arquitetura, PRD, decisão complexa
- infra-architect (Fable 5) → infraestrutura, cloud, networking
Plugin-derived (top picks)
- feature-dev:feature-dev → guided feature development workflow
- pr-review-toolkit: (6 agents)* → code-reviewer, silent-failure-hunter, comment-analyzer, pr-test-analyzer, type-design-analyzer, code-simplifier
- pro-workflow:orchestrator → wire commands+agents+skills for complex features
- claude-md-management:claude-md-improver → audit/improve CLAUDE.md
- ECC agents (64) → planning, code-review, security analysis, backend, frontend, testing
HOW TO USE (routing patterns)
| User intent | Tool/skill recomendado |
|---|---|
| "audita esse repo: " | skill repo-fit-analyzer |
| "explora esse codebase" | MCP codegraph_explore (NÃO grep+read loops) |
| "checa CVEs do " | MCP cve-mcp tools |
| "deploy" / "manda pra prod" | skill pre-deploy-validator ANTES |
| "limpa código com cara de IA" | skill deslop |
| "faz UI/landing premium" | skill design-taste-frontend ou gpt-taste |
| "threat hunting" / "incident response" | cybersec skills (754) categoria threat |
| "swarm de agents" / "paralelizar" | ruflo-swarm:swarm-init ou dispatching-parallel-agents |
| "memory cross-session" | claude-mem MCP ou ruflo-rag-memory:memory-bridge |
| "commit" / "push" / "PR" | plugin commit-commands:commit-push-pr |
| "criar plugin/skill" | plugin-dev:create-plugin ou skill-creator:skill-creator |
| "criar tirinhas (médico)" | skill criar-tirinhas / criar-tirinhas-care-sa |
| "knowledge graph" | skill graphify OU understand-anything@understand-anything |
REGRA OURO
Quando user pede algo:
- Match contra HOW TO USE table primeiro
- Se não match: scan categorias relevantes desta skill (não as 791 descriptions)
- Se ainda não match: invoke skill específica que parece mais alinhada
- NUNCA invoke múltiplos plugins/skills "no caso de" — escolha 1-2 e use bem
Anti-pattern: invocar 5 skills "to be safe" — inflaciona context, custa tokens, confunde response.