Skip to main content

setup-sandcastle

Configure Sandcastle for agent delegation -- headless run(), HITL interactive(), parallel sandboxes, branches. Use when delegating to autonomous agents, parallelizing implementation, or running interactive reviews.

Datos de origen

Repositorio
redpanda-data/ui-harness
Última actividad en el origen
27 de mayo de 2026 a las 14:03
Idioma detectado de SKILL.md
inglés
Estrellas
34
Forks
4

Opciones de instalación

De forma predeterminada está seleccionado el prompt que primero revisa el origen. Puedes cambiar a un comando directo o descargar una copia local.

Revisa los archivos de origen

Lee SKILL.md y los archivos complementarios que muestra SkillsMP antes de decidir si quieres instalarlo.

Explorador de archivos
2 archivos

Mostrando SKILL.md

SKILL.md
Instrucciones de origen · Vista previa de solo lectura
name
setup-sandcastle
description
Configure Sandcastle for agent delegation -- headless run(), HITL interactive(), parallel sandboxes, branches. Use when delegating to autonomous agents, parallelizing implementation, or running interactive reviews.
# Setup Sandcastle [Sandcastle](https://github.com/mattpocock/sandcastle) orchestrate agents in sandboxes with branch strategies. Two modes: - **`run()`** -- headless (`--print`), stream-JSON parsed. CI, batch, overnight. - **`interactive()`** -- full TUI passthrough (stdin/stdout/stderr). Human watch + intervene. HITL review, pair-review, local dev. Both modes: hooks fire inside each session. Dev lifecycle enforced regardless of launch method. **Capabilities:** task picking (GitHub issues -> one agent per issue), parallel N agents in isolated sandboxes, HITL review with full TUI, `noSandbox()` for git worktrees only, hooks in each session, branch strategies (head, merge-to-head, branch). ```mermaid sequenceDiagram participant GH as GitHub Issues participant Main as main.ts participant A1 as Agent 1 (run) participant An as Agent N (run) participant Rev as Reviewer (interactive) GH->>Main: Fetch labeled issues par Headless parallel execution Main->>A1: Issue #1 (Docker sandbox) A1->>A1: development-lifecycle (6 phases) and Main->>An: Issue #N (Docker sandbox) An->>An: development-lifecycle (6 phases) end A1->>Rev: Branch for HITL review Note over Rev: Human watches TUI, can intervene Rev->>Main: Approved -- merge to head ``` ## Steps ### 1. Install ```bash bun add -D @ai-hero/sandcastle && bunx sandcastle init ``` ### 2. Configure `.sandcastle/.env` ``` ANTHROPIC_API_KEY=sk-ant-... ``` ### 3. Choose launch mode | Scenario | Mode | Sandbox | |---|---|---| | CI/batch/overnight | `run()` | `docker()` | | Parallel 5+ issues | `run()` | `docker()` | | Local dev quick review | `interactive()` | `noSandbox()` | | Pair-review with human | `interactive()` | `docker()` or `noSandbox()` | | Single interactive session | `interactive()` | `noSandbox()` | ### 4. Orchestration Script See [REFERENCE.md](REFERENCE.md) for templates: headless batch, HITL review, mixed pipelines. ### 5. Run ```bash bunx tsx .sandcastle/main.ts ``` Each agent: read issue -> development-lifecycle -> hooks enforce patterns -> commit -> review -> merge. See [REFERENCE.md](REFERENCE.md) for templates + prompt patterns.
Ver en GitHub