devteam

Persona-driven AI development team: orchestrator, team agents, review agents, skills, slash commands, and advisory hooks for Claude Code

Orchestrate multi-agent teams for parallel code review, hypothesis-driven debugging, and coordinated feature development using Claude Code's Agent Teams

Long-running agent harness with 5-layer memory architecture, GitHub integration, autonomous batch processing, Agent Teams with ATDD, 9 hooks (safety, quality gates, team coordination), and 6 Agent Skills

Self-orchestrating multi-agent development system — 8 specialized AI agents, parallel quality gates, and automated workflows. You say WHAT, the AI decides HOW.

Virtual development team: TDD, debugging, code review, backlog management, and proven workflow patterns

49 specialist skills + 39 slash commands for coding agents — orchestrator, backend, frontend, QA, security, deploy, detective-spec, static-analysis, skill-author, program-router, parallel-dispatcher, blog-publisher, blog-screenshot, canary-deployment, zoom-out, handoff-context, post-deploy-canary-monitor, pattern-conformity, research-prep, context-budget + spec-driven development, anti-AI writing, memory consolidation, executable YAML pipelines + insights dashboard (v2.18.0, 6 tabs). v2.34.0: vault de memoria UNIFICADO ao kit — instalar o kit agora CRIA o vault automaticamente (scripts/init-vault.mjs roda no install.sh: cria estrutura logs/architecture/secrets, CLAUDE.md com as regras de escrita, .gitignore protegendo secrets/, git init). Path PORTAVEL via scripts/vault-resolver.mjs: $CLAUDE_MEMORY_VAULT → ~/.claude-memory (novo padrao, vale Windows/Mac/Linux) → D:/claude-memory (legado). Antes o vault era montado a mao e o path hardcoded; agora kit+memoria sao uma coisa so. Idempotente (nao sobrescreve vault existente). v2.33.0: absorcao obsidian-second-brain (memoria AI-first) — policy memory-write-rules.md aplica anti-fabricacao ao VAULT (false-absence: busque exaustivamente antes de afirmar que nao existe nota/decisao — o failure mode mais comum; no-fabrication: TBD para desconhecido; recency markers '(as of YYYY-MM, fonte)' em claims externos; niveis de confianca) + convencao 'For future Claude' (preambulo de 2-3 linhas que o futuro-Claude le em 10s pra decidir relevancia, no skill 31 session-summary) + comando /reconcile-memory (detecta contradicoes no vault — decisoes revertidas/superadas nunca atualizadas — e resolve: mais novo+autoritativo vence com secao ## History, ambiguo vira flag, evolucao marca superseded). v2.32.0: pre-build-gate (UserPromptSubmit hook) leva o 'pare e decida antes de codar' de cada disciplina (que o /auto tem nas fases) para o MODO PASSIVO — detecta intencao de criacao no prompt e injeta o checklist da disciplina certa: acceptance-criteria (defina done antes de implementar), api-contract (formato de erro + status codes antes da 1a rota), schema-integrity (constraints/FK/indices antes do 1o INSERT), ui-design (ancora estetica antes de estilizar), deploy-readiness (healthcheck/env/graceful-shutdown). 4 novas rules path-scoped por disciplina (rules/backend/, rules/database/, rules/frontend/, rules/common/acceptance-criteria.md). v2.31.0: design-aware /auto — fase UI-DESIGN como gate (PLAN→[UI-DESIGN]→BUILD) que invoca skill 02-ui-ux-design e bloqueia build de arquivo visual ate a ancora estetica estar escolhida; rules/frontend/ui-design.md (path-scoped em .css/.tsx/public) proibe o default generico (#4f46e5 indigo + system-ui), forca escolher 1 ancora; coverage config virou gate HARD no /auto; scope inference (app→fullstack) movido pra rules/common para o modo passivo herdar. Validado por bench A/B real (bench/ab/, 3 rounds) que expos UI generica nos 3 bracos. v2.29.0: claim-verifier (PostToolUse — detecta afirmacoes sem evidencia: 'email enviado', 'deploy OK', 'teste passou'; passa livre se ha exit code 0/HTTP 200/query result) + context-turn-counter (UserPromptSubmit — compact a cada 25 turnos, handoff inteligente a cada 50 usando vault de memoria). v2.28.0: /spec-kit (SDD pipeline unificado specify→plan→tasks→implement + Adversarial Verifier inline), /insights (recomendacoes baseadas em telemetria dos hooks), /swarm com Phase 3 Adversarial Verify (Implementor vs Verifier com goals opostos, spec atualizada em real-time). v2.27.0: investigate-first guard (hook PreToolUse que impede a IA de perguntar o auto-descobrivel — gh user, branch, package manager, porta, versao de runtime — manda investigar primeiro) + policy investigate-first. v2.26.0: silent-failure-hunter (16o subagent, review-only: caca catch{} vazio, swallowed errors, fallbacks perigosos, stack traces perdidos, rollback faltando) + skill 49 context-budget (audita peso de contexto carregado por componente) + /context-budget. v2.25.0: rules system path-scoped (`.claude/rules/` com `paths:` glob, inspirado no ECC), bug-fix da allowlist de subagents, 5 skills stub reescritos com profundidade. v2.24.0: curador AUTONOMO de memoria (inspirado no Hermes Agent) — roda async no SessionStart, faz decay/archive/dedup em JS puro sem gastar LLM e delega so a parte semantica ao agente presente. v2.23.0: absorcao addozhang (skill 48 research-prep, Spring Boot playbook, mem9 patterns). v2.22.0: memory curator nudge. v2.21.0: context-cost guards. v2.20.0: skill 47 pattern-conformity. 16 dispatchable subagents.