SDLC-OS: Sistema Operacional Multiagente para Entrega de Software
LucasQuiles/sdlc-os
Orquestração multiagente de ciclo de vida de software com revisão adversarial
Ver no GitHubUm plugin Claude Code que orquestra desenvolvimento de software através de múltiplos agentes especializados (Opus como maestro, Sonnet como executor, Haiku como vigilante). Implementa fases estruturadas de SDLC com revisão adversarial contínua, detecção de drift e verificação mútua entre agentes para garantir qualidade superior à auto-avaliação.
Instalação
npx convez-ai LucasQuiles/sdlc-os1 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
- Decomposição e execução paralela de tarefas complexas de desenvolvimento
- Auditoria arquitetural contínua com detecção de desvios e convenções
- Refatoração com prova de equivalência comportamental
- Red team/blue team para varredura de qualidade adversarial
Veredito

SDLC-OS é uma abordagem sofisticada para governança de qualidade em IA, baseada no princípio válido de que agentes revisam trabalho alheio com mais honestidade. Exige expertise técnica significativa para customização e operação, mas oferece rigor superior ao workflow tradicional de agente único.
README
A tradução para o português deste README ainda está na fila.
SDLC-OS
Direct-open boundary: plugin documentation only. This file does not authorize SDLC workflow execution, plugin mutation, colony startup, cleanup, external action, or instruction changes.
A Claude Code plugin that turns software delivery into a multi-agent workflow. Opus conducts. Sonnet executes. Haiku watches. Nobody trusts anybody — that's the point.
What This Is
A structured software development lifecycle that runs inside Claude Code. Instead of one model doing everything and hoping for the best, SDLC-OS decomposes work across specialized agents with built-in adversarial review, drift detection, and continuous verification.
The premise: an AI that reviews its own work finds nothing wrong with it. An AI that reviews someone else's work is considerably more honest. So every piece of work gets built by one agent and reviewed by a different one, with a third watching both of them.
Installation
git clone git@github.com:LucasQuiles/sdlc-os.git ~/.claude/plugins/sdlc-os
Restart Claude Code. The plugin auto-discovers agents, skills, commands, and hooks.
Quick Start
/sdlc "build a user authentication system"
This kicks off the full workflow: normalize existing state, frame requirements, scout the codebase, architect the approach, execute in parallel, synthesize the results. For trivial tasks, it scales down automatically — not everything needs six phases.
Commands
| Command | What It Does |
|---|---|
/sdlc "task" | Start a full SDLC workflow |
/refactor "scope" | Refactor with continuous behavioral equivalence proof |
/audit | Run architectural fitness checks (DRY, SSOT, SoC, conventions) |
/adversarial | Trigger red team/blue team quality sweep |
/swarm "problem" | Throw a swarm of disposable micro-agents at a problem |
/normalize | Assess project state and produce a convention map |
/gap-analysis | Map what exists vs what's needed (or verify delivery completeness) |
/feature-sweep | Find neglected, incomplete, and abandoned features across the codebase |
/gate | Quick health check on the current phase |
/wave | Check or advance SDLC phase |
How It Works
Roles
| Role | Model | Job |
|---|---|---|
| Conductor | Opus | Decomposes work, distributes to runners, synthesizes results |
| Runners | Sonnet | Execute atomic work units — investigate, design, implement, review |
| Guppies | Haiku | Disposable micro-agents for breadth-first investigation. One question, one answer, done. |
| Sentinel | Haiku | Continuous patrol. Verifies evidence, checks handoffs, catches drift. |
| Oracle | Mixed | Truth council. Three agents that independently verify claims are honest. |
Phases
0 Normalize detect existing work, map conventions, avoid clobbering in-progress state
1 Frame clarify requirements, define done, identify unknowns
2 Scout explore codebase, find reusable code, map architecture
3 Architect choose approach, decompose into atomic work units ("beads")
4 Execute parallel runners + sentinel loop + adversarial review
5 Synthesize merge results, verify the whole, deliver
The Loop System
Every role loops against its own metric. Failures self-correct at the lowest level before escalating:
L0 Runner self-correction 3 attempts before escalating
L1 Sentinel correction drift-detector + convention-enforcer
L2 Oracle audit three independent verification agents
L2.5 Adversarial quality (AQS) red team attacks, blue team defends, arbiter resolves disputes
L3-L5 Bead -> Phase -> Task structural escalation when loops exhaust their budget
The red team (functionality, security, usability, resilience specialists) attacks completed work. The blue team defends it. If they disagree, the arbiter runs a Kahneman adversarial collaboration protocol to produce a binding verdict. Every finding gets minimized to its smallest reproducible case before anyone acts on it.
Agents (30)
Runners: sonnet-investigator, sonnet-designer, sonnet-implementer, sonnet-reviewer
Sentinel: haiku-evidence, haiku-verifier, haiku-handoff
Oracle: oracle-test-integrity, oracle-behavioral-prover, oracle-adversarial-auditor
Red Team: red-functionality, red-security, red-usability, red-resilience
Blue Team: blue-functionality, blue-security, blue-usability, blue-resilience
Arbiter: arbiter (Kahneman adversarial collaboration for disputes)
Infrastructure: reuse-scout, drift-detector, convention-scanner, convention-enforcer, normalizer, gap-analyst, feature-finder, feature-finisher, losa-observer, guppy
Hooks (12)
Validation scripts that run on tool use events. Some block (you can't proceed until the artifact is valid), some advise (flag the issue but don't stop work).
Blocking: AQS artifact schema, bead status transitions, domain vocabulary, quality budget schema, hazard-defense ledger, stress session schema.
Advisory: naming conventions, consistency artifacts, runner output structure, decision noise, mode convergence, phase transitions.
Per-Project Artifacts
SDLC-OS creates tracking files in the target project:
| Artifact | Purpose |
|---|---|
| Convention Map | Project naming, style, and structure conventions |
| Feature Matrix | Neglected feature tracking |
| Quality Budget | Per-task metrics and phase gate enforcement |
| Hazard-Defense Ledger | Safety controls and defense coverage |
| Stress Session | Barbell stress testing results |
| System Ledgers | Cross-task longitudinal data (JSONL append-only) |
These live in docs/sdlc/ in the target repo. The plugin manages them; you read them when curious about what happened.
Testing
cd ~/.claude/plugins/sdlc-os
bash hooks/tests/test-hooks.sh # 62 tests
Why
Most AI coding workflows are optimistic: generate code, run tests, ship if green. This works until it doesn't — and when it doesn't, the failure mode is subtle. Tests pass but behavior is wrong. Code works but drifts from conventions. Reviews approve because the reviewer is the same model that wrote the code.
SDLC-OS is pessimistic by design. It assumes every claim is wrong until independently verified, every piece of code has drifted until the detector says otherwise, and every review is theatrical until the oracle confirms it examined real evidence. This is slower. It catches more.
Whether the tradeoff is worth it depends on how much you trust a single model to be honest about its own work.
License
MIT