Kudzu: Framework SDLC Agêntico para Claude Code
pat-eason/kudzu
SDLC automático com human-in-the-loop em decisões que importam
Ver no GitHubFramework que orquestra o ciclo completo de desenvolvimento (pesquisa, planejamento, implementação) com subagentes especializados. Você toma decisões em 4 gates críticos; o resto (testes, reviews, PRs, docs, integração com Linear) roda automaticamente.
Instalação
npx convez-ai pat-eason/kudzu1 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
- Pesquisa técnica antes de implementar features complexas
- Triagem automática de bugs com recomendações
- Decomposição de PRDs em chunks de trabalho
- Pipeline de build com code review e QA automáticos
- Decisões arquiteturais (Temporal vs BullMQ, etc)
Veredito

Ferramenta robusta para times que precisam acelerar delivery sem perder controle. Requer setup inicial (3 configs) e disciplina no uso dos gates, mas devolve tempo real ao eliminar overhead de orquestração manual e artifacts (PRs, docs, testes).
README
A tradução para o português deste README ainda está na fila.
Kudzu
Kudzu grows everywhere. So does your workflow.
An agentic SDLC framework for Claude Code. Three commands cover the full cycle.
/kudzu:research → research a concept, triage a bug, evaluate a decision
/kudzu:plan → decompose an approved PRD into implementation chunks
/kudzu:implement → run the build pipeline, chunk by chunk
Everything else — subagent orchestration, code review, QA, security analysis, documentation, Linear updates, PR creation — runs automatically. You get pulled in at four defined gates where your judgment actually matters.
Named for the vine that covers everything in western NC.
Install
/plugin marketplace add denworks/kudzu
/plugin install kudzu@kudzu-marketplace
/kudzu:setup
That's it. /kudzu:setup walks you through creating the three config files.
Commands
| Command | Description |
|---|---|
/kudzu:research [concept] | Research anything before building it |
/kudzu:plan | Decompose approved PRD into implementation chunks |
/kudzu:implement [target] | Run build pipeline |
/kudzu:status [full] | Current project state |
/kudzu:checkpoint [notes] | Save session state |
/kudzu:gate [N] [decision] | Record a HITL gate decision |
/kudzu:setup | First-run config initialization |
/kudzu:implement targets
| Target | What it does |
|---|---|
| (empty) | Run next pending chunk |
chunk-N | Run specific chunk |
all | Run all remaining chunks automatically |
auto | Run all chunks, no Gate 4 pauses — auto-resolves defensively |
docs | Documentation-only pass |
bug: [description] | Bug fix mode |
tickets | Sync open issues to Linear |
/kudzu:research modes
Kudzu detects intent from your input:
| Input | Mode |
|---|---|
/kudzu:research "build a BullMQ observability dashboard" | Full DISCOVER — researchers + PRD |
/kudzu:research "triage: workers dropping jobs silently" | Bug triage |
/kudzu:research "decide: Temporal vs BullMQ for orchestration" | Decision research |
/kudzu:research "doc audit" | Documentation gap analysis |
/kudzu:research "feature: add webhook delivery" | Feature scoping |
Configuration
Three config layers, merged at every session start by /kudzu:config.
Project values override machine, machine overrides global.
~/.claude/kudzu/config.global.md — Your identity and working style
Committed to your personal dotfiles. Same on every machine.
ENGINEER_NAME=Pat
ENGINEER_GITHUB=yourhandle
ADVERSARY_STRICTNESS=high
BIAS_TOWARD_INTERFACES=true
SECURITY_AUTO_TRIGGER_PATTERNS=auth,token,credential,secret,permission,role,key,password
~/.claude/kudzu/config.machine.md — Machine-specific credentials
Never committed. Fill in fresh on each machine.
LINEAR_TEAM_ID= # LP Architecture team ID (on work machine)
LINEAR_ASSIGNEE_DEFAULT= # your Linear user ID
GITHUB_REVIEWERS= # ntsianos,teammate (on work machine)
NOTION_ENABLED=false
.kudzu/config.project.md — Per-project config
Committed with the repo. No secrets. Auto-picked up by teammates.
PROJECT_NAME=Kenai
GITHUB_REPO=luxury-presence/kenai
LINEAR_TEAM_ID= # overrides machine default for this project
PROJECT_KEY_LIBS=bullmq@5,ioredis@5
PROJECT_MESSAGE_BUS=bullmq
PROJECT_AUTH=spicedb
HITL gates
You decide at four points. Everywhere else, Kudzu runs automatically.
| Gate | After | You decide |
|---|---|---|
| 1 | Research complete | Worth writing a PRD? |
| 2 | PRD drafted | Does this spec capture what I want? |
| 3 | Architecture planned | Will this decomposition produce good code? |
| 4 | Chunk reviewed (conditional) | Only on failures, security issues, chunk 1, or load-bearing chunks |
Record decisions with /kudzu:gate:
/kudzu:gate 3 approved
/kudzu:gate 4 fix_and_retry "null check missing at auth.ts:47"
Multi-machine setup (e.g. work laptop)
# Install on work machine
/plugin marketplace add denworks/kudzu
/plugin install kudzu@kudzu-marketplace
/kudzu:setup
# config.global.md — copy from your dotfiles (same content, all machines)
# config.machine.md — fill in LP-specific values:
# LINEAR_TEAM_ID = LP Architecture team ID
# GITHUB_REVIEWERS = ntsianos,jp-handle
# NOTION_ENABLED = true
For each LP project, add project config once:
cd /path/to/lp-project
# Create project config (if not already committed):
cp ~/.claude/kudzu/ # kudzu:setup handles this
nano .kudzu/config.project.md
git add .kudzu/config.project.md
git commit -m "chore: add Kudzu project config"
Teammates who install Kudzu globally pick up the project config automatically when they clone the repo.
What gets committed where
| File | Where | Committed? |
|---|---|---|
~/.claude/kudzu/config.global.md | Personal dotfiles | ✓ Yes |
~/.claude/kudzu/config.machine.md | Local only | ✗ Never |
.kudzu/config.project.md | Project repo | ✓ Yes |
.kudzu/<type>/<topic>/PRD.md, ARCH_DECISIONS.md | Project repo | Recommended |
.kudzu/<type>/<topic>/IMPL_SPECS/, REVIEW_SPECS/ | Project repo | Recommended |
.kudzu/<type>/<topic>/CONTEXT.md | Project repo | Your call |
Agent team
Kudzu orchestrates 13 specialist agents across four tiers:
| Agent | Model | Role |
|---|---|---|
| Project Planner | Opus | Concept → synthesis → PRD |
| Researcher Pro | Opus | Pro-bias research subagent |
| Researcher Con | Opus | Adversarial research subagent |
| Concept Reviewer | Opus | Viability + decomposition review |
| Software Architect | Opus | PRD → chunks + specs (5 modes) |
| Software Engineer | Sonnet | Spec → code |
| Code Reviewer | Sonnet | Review + escalation |
| QA Analyst | Sonnet | Test coverage |
| Security Analyst | Opus | Security + scalability |
| Documentarian | Sonnet | Doc gap analysis |
| Technical Writer | Haiku | Actual writing |
| Project Manager | Sonnet | Linear ownership |
| Checkpoint | Sonnet | State management |
Requirements
- Claude Code ≥ 2.0.0
- Optional: Linear MCP (direct ticket creation)
- Optional: GitHub MCP (direct PR creation)
- Optional: Notion MCP (direct doc updates)
Contributing
See CONTRIBUTING.md.
License
MIT — see LICENSE.