Claude Waves
iamphduc/claude-waves
Fluxo de código em ondas paralelas: planejamento → sprint → execução supervisionada
Ver no GitHubAgent que orquestra fluxos de trabalho multi-agente para desenvolvimento de código em ondas paralelas. Organiza sprints em fatias de código com propriedade não-sobreposta, executa engenheiros Claude em paralelo em worktrees isoladas, e oferece gates manuais ou modo autônomo com auto-merge e verificação de trunk.
Instalação
npx convez-ai iamphduc/claude-waves1 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
- Coordenar múltiplos agentes Claude em projetos grandes
- Implementar features em paralelo com verificação de qualidade entre ondas
- Automação de deployments com gates manuais ou autônomos
Veredito

Ferramenta robusta para equipes que precisam coordenar agentes IA em fluxos complexos de desenvolvimento. Exige setup inicial cuidadoso e compreensão de git/worktrees, mas compensa com paralelismo controlado e gates de qualidade explícitos entre ondas.
README
A tradução para o português deste README ainda está na fila.
claude-waves
Ship a plan in waves of parallel Claude engineers: strategy → sprint → parallel waves → review, with you as the merge gate between every wave.
A wave is a batch of slices with non-overlapping file ownership, built concurrently in isolated worktrees. Each wave lands behind you, trunk is verified, then the next wave dispatches — manually with /code, or unattended with /autopilot.
Install
Requires Claude Code, git, and an authenticated gh CLI. From your project root (new or existing repo), run:
curl -fsSL https://raw.githubusercontent.com/iamphduc/claude-waves/main/install.sh | bash
This copies into your project:
agents/→.claude/agents/skills/→.claude/skills/docs/→docs/
It merges into existing folders and overwrites same-named files, so review your working tree afterward (or copy the folders by hand if you prefer).
Everything the workflow needs ships in those folders, scaffolding included. Then fill two stubs for your project:
docs/codebase-structure.md— your codebase brief; the## Smoke recipesection is required (engineers use it to bring the app up and browser-verify each slice before shipping).docs/decisions.md— architectural decisions, as you make them.
Update
Refresh the agents/skills/templates and policy docs to the latest. It only overwrites files that ship in this repo — it never deletes anything and never touches your codebase-structure.md, decisions.md, plans, or sprints:
curl -fsSL https://raw.githubusercontent.com/iamphduc/claude-waves/main/update.sh | bash
Manual flow — you ride each wave
| Step | Skill | What happens |
|---|---|---|
| 1 | /plan | Planner interviews you, writes docs/plans/<slug>.md |
| 2 | /sprint [slug] | Drafts docs/sprints/<slug>.md — slices grouped into waves by file ownership |
| — | read the sprint doc | Your quality gate — catch bad wave grouping or overlapping file ownership before any engineer runs |
| 3 | /code [slug] | Runs the wave loop: one worktree per slice, all engineers in the wave dispatched at once, then halts for you to merge |
| — | merge the wave's PRs, reply continue | Next wave dispatches — repeat until the sprint's waves are done |
| 4 | reviewer (auto) | Code audit; opens a follow-up PR or returns PR: clean |
| — | merge review PR, reply continue | Sprint archives; continue chains into the next sprint |
Autonomous flow — the waves ride themselves
/autopilot [plan-slug] [--max-sprints=N] [--max-waves=N] [--max-runtime=Nh] runs the whole plan unattended: dispatches each wave, auto-merges clean PRs (escalating risky ones), verifies trunk between waves, chains sprints, and halts + notifies at each gate. Invoking it is your consent to the auto-merges. Criteria, defaults, and resume behavior live in docs/autonomous-policy.md.
┌────────────────────────────────── SPRINT LOOP (outer) ──────────────────────────────────┐
v │
┌───────────┐ ┌────────────┐ ┌────────────┐ ╔═══════════ WAVE LOOP (inner) ═══════════╗ ┌───────────┐ ┌───────────┐ │
│ /autopilot│──>│ Read policy│──>│ Read sprint│──>║ ┌──────────┐ ┌──────────┐ ┌───────┐ ║──>│ Reviewer │──>│ Archive │ │
│ plan-slug │ │ + bounds │ │ doc │ ║ │ Dispatch │──>│Auto-merge│──>│ Verify│ ║ │ +auto-mrg │ │ +mark │ │
└───────────┘ └────────────┘ └────────────┘ ║ │ engineers│ │clean PRs │ │(waves)│ ║ └───────────┘ │ plan row │ │
║ └──────────┘ └──────────┘ └───┬───┘ ║ └─────┬─────┘ │
║ ^ │ ║ │ │
║ └──── more waves <───────────┘ ║ │ │
╚═════════════════════════════════════════╝ │ │
┌─────────────────────────────┘ │
│ │
┌──────────────┐ planned rows left? │
│ Plan complete│<──── no ───────────────────────────────────────┴─────── yes ───────┐ │
└──────────────┘ ┌────────────v─────────────┐ │
│sprint-planner drafts next│──────┘
│sprint ──> (re-read doc) │
└──────────────────────────┘
Any policy gate at any step → halt + notify, then end the turn.
State on disk
docs/
|-- known-issues/*.md # durable constraints
|-- plans/<slug>.md # strategic plans
|-- sprints/
| |-- archive/ # completed sprints
| `-- <slug>.md # active sprint — status board + per-slice detail
|-- templates/ # doc templates the agents fill in
|-- autonomous-policy.md # /autopilot criteria + gates
|-- codebase-structure.md # codebase brief (you maintain)
|-- decisions.md # architectural decisions, authoritative (you maintain)
|-- engineer-protocol.md # engineer/reviewer contract
`-- handoff-queue.md # inter-agent comms — BLOCKED halts, PENDING defers, SOLVED informational