convez.ai
AgenteAutomaçãoAvançado
sighup

Claude Workflow

sighup/claude-workflow

Spec-driven development com paralelização e validação estruturada

Ver no GitHub
InicialPluginScore · de 0 a 100

Agent de desenvolvimento orientado por especificações que transforma ideias em implementações validadas usando grafos de tarefas com dependências e execução paralela de agentes. Integra pesquisa, planejamento, execução e validação em um fluxo único e automatizado.

Instalação

$npx convez-ai sighup/claude-workflow

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
15
Forks
1
Acompanhando
0
Saúde
28%
Último commit
há 2 meses
Versão
v3.7.2
Issues abertas
3
Contribuidores
2

Casos de uso

  • Desenvolvimento iterativo de features com validação contínua
  • Execução paralela de tarefas interdependentes com múltiplos agentes
  • Code review automatizado com análise particionada por preocupações
  • Geração e execução de testes E2E com auto-fix de falhas
  • Desenvolvimento multi-feature simultâneo via worktrees

Veredito

Ferramenta sofisticada para quem precisa automatizar ciclos completos de desenvolvimento com múltiplos agentes. Exige compreensão de fluxos complexos e CLI (jq, gh), mas oferece valor significativo em projetos que demandam paralelização e validação estruturada.

README

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

claude-workflow

A Claude Code plugin that unifies spec-driven development, autonomous task execution, and parallel agent dispatch into a single workflow. Takes a feature from idea to validated implementation using structured specifications, dependency-aware task graphs, and evidence-based verification.

Install

From Git (recommended)

# Add the marketplace
claude plugin marketplace add https://github.com/sighup/claude-workflow.git

# Install at project scope (shared with team via .claude/settings.json)
claude plugin install claude-workflow@claude-workflow --scope project

# Or install at user scope (personal, across all projects)
claude plugin install claude-workflow@claude-workflow --scope user

Interactive installation

/plugin
# Navigate to Marketplaces tab → Add → paste the git URL
# Then go to Discover tab → select claude-workflow → choose scope

Workflow

Interactive (inside Claude)

[/cw-research]  →  /cw-spec  →  [/cw-gherkin]  →  /cw-plan  →  /cw-dispatch  →  /cw-validate

Each step can also be run independently. /cw-execute handles single-task execution for manual or shell-scripted loops. /cw-review adds a code review gate and /cw-testing generates and runs E2E tests.

Worktrees (manual parallel development)

Use /cw-worktree to develop multiple features simultaneously. Each worktree gets its own feature branch and isolated task list (via .claude/settings.local.json). Tasks persist in ~/.claude/tasks/{worktree-name}/, enabling seamless resume across sessions. See examples/workflows.md for the full multi-terminal walkthrough.

Skills

SkillPurpose
/cw-researchExplore codebase across 5 dimensions and produce a structured research report with /cw-spec meta-prompt
/cw-specGenerate structured specification with demoable units and proof artifacts
/cw-planTransform spec into native task graph with dependencies and metadata
/cw-executeExecute one task using the 11-step protocol (orient → commit → clean exit)
/cw-dispatchSpawn parallel subagent workers for independent tasks (no setup required)
/cw-dispatch-teamPersistent agent team with lead coordination for parallel task execution
/cw-validateRun 7 validation gates and produce a coverage matrix report
/cw-reviewReview implementation for bugs, security issues, and quality; creates fix tasks
/cw-review-teamConcern-partitioned team review — each reviewer sees all files through a specialized lens (security, correctness, spec compliance)
/cw-testingE2E testing with auto-fix — generate tests from specs, execute, and fix failures
cw-gherkinGenerate Gherkin BDD scenarios from spec acceptance criteria; called automatically by cw-spec
/cw-worktreeManage git worktrees for multi-feature parallel development

Prerequisites

jq is required by the worktree hooks and cw-status. The gh CLI is needed for PR creation.

Most skills work out of the box. /cw-dispatch-team uses Claude Code agent teams which requires two env vars:

  1. Enable the experimental feature flag (user-level, applies to all projects):
// ~/.claude/settings.json
{
  "env": {
    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
  }
}
  1. Set the task list ID (project-level, unique per project):
// .claude/settings.json
{
  "env": {
    "CLAUDE_CODE_TASK_LIST_ID": "your-project-name"
  }
}

Note: /cw-worktree create sets CLAUDE_CODE_TASK_LIST_ID automatically in .claude/settings.local.json — no manual configuration needed for worktree-based workflows.

/cw-dispatch (subagent workers) needs no setup and is the recommended default. /cw-plan will offer both options after task graph creation.

/cw-testing supports multiple backends. To use the playwright-bdd backend (Gherkin → Playwright, CI-friendly):

npm install --save-dev playwright-bdd @playwright/test
npx playwright install

Optional integrations

herdr

herdr is a terminal-native agent multiplexer (Linux/macOS). When it is installed and running, /cw-worktree create will automatically open a Claude session in the new worktree's herdr pane instead of asking you to open a new terminal. /cw-worktree open <name> does the same retrospectively, focusing the existing pane if one already exists.

If herdr is not installed, not running, or you set CW_DISABLE_HERDR=1, every worktree command works exactly as before — the integration is silently skipped and the manual cd ... && claude instructions are printed instead. Nothing in the plugin requires herdr.

To give the in-pane Claude session full knowledge of herdr's CLI (workspaces, panes, wait output, etc.), install herdr's own skill into Claude Code after installing herdr:

herdr integration install claude

That skill is maintained by the herdr project (AGPL-3.0) and self-disables outside a herdr pane via the HERDR_ENV=1 guard, so it stays out of your way on hosts without herdr.

Task Metadata

Every task on the board carries self-contained metadata enabling autonomous execution:

{
  "task_id": "T01",
  "spec_path": "docs/specs/01-spec-auth/01-spec-auth.md",
  "scope": {
    "files_to_create": ["src/auth/login.ts"],
    "files_to_modify": ["src/routes/index.ts"],
    "patterns_to_follow": ["src/routes/health.ts"]
  },
  "requirements": [
    { "id": "R01.1", "text": "POST /auth/login accepts credentials", "testable": true }
  ],
  "proof_artifacts": [
    { "type": "test", "command": "npm test -- src/auth/login.test.ts", "expected": "All pass" }
  ],
  "commit": { "template": "feat(auth): add login endpoint" },
  "verification": {
    "pre": ["npm run lint"],
    "post": ["npm test"]
  },
  "role": "implementer",
  "complexity": "standard",
  "model": null
}
Ver o original no GitHub

Bom para

Relacionados