Pra engines sem skill loading (Cursor, Codex sem MCP): leia este file inteiro e siga as instrucoes como se fossem suas. Voce nao precisa "invocar" — apenas obedeca. Cursor e Codex que tem xp-stack instalado via npm leem este SKILL.md em .cursor/rules/ ou .codex/skills/.

Local Waves Orchestrator — Setup

Legacy notice: Esta skill foi a recomendação canônica antes do lançamento do Agent View nativo do Claude Code (https://claude.com/blog/agent-view-in-claude-code, 11-mai-2026). Pra novos projetos use Agent View via Agent tool nativo (model: "sonnet" + isolation: "worktree" + prompt invoca caveman:caveman). local-waves segue funcional como fallback ou pra projetos que rodam fora do Claude Code (não-interativo, CI, etc.).

Set up a local parallel-wave orchestrator in your project. Optional, opt-in. Invoke explicitly when you decide you want this — the regular bootstrap does NOT install it.

Mental model

• Orchestrator = the Claude Code session you're running (Opus-class, auto mode). It's the brain.
• Script = mechanical hand. The orchestrator invokes it via the Bash tool.
• Workers = N instances of claude -p running Sonnet-class headless in separate git worktrees. They execute 1 task each in parallel.

Worker permissions: --permission-mode acceptEdits + --allowedTools with specific allowlist (no --dangerously-skip-permissions). If a worker hits something outside the allowlist OR needs a credential / business decision not provided, it stops, writes BLOCKERS.md in the worktree, commits WIP, and exits without opening a PR. The orchestrator collects this and presents to the Pilot.

Trade-off vs xp-stack:paperclip-orchestrator and Agent View