LegoFlow

Blockscurator

Getting Started

Curator wraps the pinned repos/swegen checkout in the LegoFlow block layout (recorded as a gitlink under blocks/curator/repos/swegen). You drive it through the curator-plugin Claude Code plugin: each slash command runs its own preflight, confirmation, and bookkeeping, so you operate the block through commands rather than raw scripts.

Prerequisites

  • Docker — task validation and execution run in containerized Harbor tasks.
  • Python — the swegen package installs from repos/swegen.
  • GitHub tokens — for PR collection and PR metadata lookup.
  • OpenAI-compatible LLM endpoint — PR evaluation and instruction generation.
  • Claude-compatible task model — the task-completion and verification stage.

Fastest start: one prompt

Launch Claude from inside blocks/curator/ (so the block-local plugin loads), then describe the run in plain language. The agent fills the environment, runs preflight, and chains the plugins for you — you don't even need to edit config.yaml. Grab the copy-paste prompt from Run Generation → The simplest way.

The stages

Whether you prompt for them or run them yourself, the pipeline is five commands:

  • /curator:setup — one-time: init the submodule, build the venv, install swegen.
  • /curator:check — read-only preflight (config, tokens, a live LLM call, Docker).
  • /curator:collect-prs — collect GitHub PR IDs as a background job.
  • /curator:create-tasks — generate tasks from collected PRs and verify each.
  • /curator:dashboard — dataset analytics (difficulty + tags), not run progress.

A sixth command, /curator:run, is a compatibility adapter that forwards to /curator:create-tasks so every block exposes the same /<block>:run verb.

Next

  • Run Generation — the full flow, per-command reference, config.yaml schema, and manual commands.
  • Core Concepts — PR pools, task skeletons, validation, and manifests.
  • Outputs — task directories, manifests, logs, and resuming after moving nodes.

On this page