convez.ai
AgenteDevOps & InfraestruturaAvançado
StChiotis

Claude-LFE

StChiotis/Claude-LFE

Engenharia agnóstica confiável: agents Claude que mantêm intenção em produção

Ver no GitHub
BomPluginScore · de 0 a 100

Framework de engenharia confiável para agents Claude que implementa princípios de confiabilidade de sistemas críticos (FMEA, RCM, poka-yoke) em software com IA. Mantém a intenção humana através de múltiplas sessões via documentação como fonte de verdade, inspeção mecanizada e checkpoint de estado.

Instalação

$npx convez-ai StChiotis/Claude-LFE

1 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

Estrelas
7
Forks
1
Acompanhando
0
Saúde
100%
Último commit
há 4 meses
Versão
v1.2.0
Issues abertas
1
Contribuidores
2

Casos de uso

  • Construir agents de longa duração que não perdem contexto ou objetivo entre sessões
  • Implementar workflows automáticos críticos que precisam auditoria e rastreabilidade
  • Estruturar código de IA com disciplina de engenharia confiável (FMEA, inspeção contínua)

Veredito

Solução arquitetural robusta para quem precisa levar agents Claude a produção sem degradação. Traz overhead deliberado em troca de confiabilidade mecanizada — não para protótipos rápidos, mas para sistemas que precisam manter intenção ao longo do tempo.

README

A tradução para o português deste README ainda está na fila.

Claude-LFE

CI tests license release use this template

Intro deck

Want to see the full introduction deck? Visit stchiotis.github.io/Claude-LFE.intro


Making AI reliable, session after session.

The bottleneck in agentic software isn't capability — it's trust. A model that nails the prototype will, across dozens of sessions, lose intent, re-litigate settled decisions, and sprawl into spaghetti. Claude-LFE is a ready-to-clone Claude Code scaffold that keeps that from happening: the human stays on the wheel, documentation is the source of truth, and reliability is mechanically enforced and measured — not asserted.

It's deliberately slower. That's the trade: overhead you choose to pay so the work holds up in production, not just in the demo. The discipline is borrowed from reliability engineering in marine and biotech — FMEA, RCM, poka-yoke — ported onto AI.

Built on the Library-First Engineering framework.

📖 The long-form argument — this thesis is laid out in full in "The Bottleneck in Agentic Software Isn't Capability. It's Trust" (Towards AI, Jun 2026). The full text also ships in-repo as MEDIUM ARTICLE.md so the rationale travels with the template — for human readers and for any agent auditing the repo.


What's included

ComponentLocationPurpose
LFE skills.agents/skills/ + .claude/skills/ mirrorFull LFE skill set covering bootstrap, planning, building, inspection, archival, hygiene, recovery, and navigation
Claude Code hooks.claude/hooks/Persona-discipline + plan-critique write gates, coordination-file frontmatter validators (with sub-validators per file type), state-aware SessionStart reminder, checkpoint-flip state mutator
Inspector specialist skills.agents/skills/lfe-*-check/ + lfe-dep-audit/ + lfe-mutation-verify/Security (OWASP Top-10), performance, complexity, dependency-audit, and mutation-reasoning passes the Inspector runs as in-chat skills
statusLine.claude/statusline.mjsLive pipeline state rendered below the Claude Code prompt
LFE protocol docs.docs/protocol/PERSONAS, ASSEMBLY_LINE, GOVERNANCE, COORDINATION_FILES, LOOP_ARCHITECTURE
Sync toolingscripts/ + .githooks/sync-claude-skills.mjs + pre-commit drift guard
Quality layer.docs/quality/PROTOCOL_DEBT, RETENTION_RUNBOOK, inspector-config
Skill-accuracy harness.agents/skills/_evals/ + .claude/lib/skill-eval*.mjs + /lfe-skill-evalMeasures whether the five defect-catching reasoning skills — security, performance, complexity, mutation-reasoning, and plan-critique (the checks whose job is to find a problem, not the ~18 workflow skills) — actually catch planted defects; renders .docs/quality/skill-eval-scorecard.md

Getting started

# 1. Clone into your new project folder
git clone <this-repo> my-new-project
cd my-new-project

# 2. Activate the pre-commit drift guard
npm install

# 3. Open in Claude Code and bootstrap
/lfe-boot

# 4. Run the Day 0 domain extraction interview
/lfe-extract-domain

After /lfe-extract-domain completes, CONTEXT.md and .docs/domain/domain-knowledge.md will be populated with your project's domain language and then you're ready for your first mission.

Fresh-clone note: CONTEXT.md is intentionally absent in a fresh clone — it is not missing or broken. /lfe-extract-domain creates it from your Day-0 domain interview. Likewise, .docs/domain/ ships with only a Shelf Index (.docs/domain/README.md) explaining what belongs there; the domain files are created on Day 0.


Key docs

DocumentPurpose
LLM_AGENT_GUIDE.mdCanonical agent rules: workflow, skills, coordination layer, project bindings
pipeline_status.mdLive session cursor — active persona, mission state, coordination file tracker
.docs/protocol/LFE protocol: personas, assembly line, governance, coordination file schema
CONTEXT.mdDomain vocabulary (populated by /lfe-extract-domain)
.docs/README.mdDocumentation floor map

Framework philosophy

"Thinking in the Human · Processing in the AI · Truth in the Documentation"

See README.md on the public LFE repo for the full philosophy.


Author

Built by Stylianos Chiotis — reliability engineering (marine · biotech) ported onto AI. Connect or reach out: LinkedIn · intro deck · Writing: The Trust Bottleneck · Library-First Engineering.


License

MIT

Ver o original no GitHub

Bom para

Relacionados