Motivation

Training a strong coding agent is a multi-stage pipeline — curate tasks, roll out trajectories, do SFT, then RL. Each stage needs its own data, its own environment, and its own GPUs, and the handoffs between them quietly break every time someone runs them by hand.

We built SWE-Lego-Live so the whole pipeline behaves like one project. Every stage is a self-contained block with the same shape — a config.yaml, a scripts/start.sh, an artifacts/ tree — wired into a parent by name. Run one block, run the whole tree, or swap a block out; the contracts hold.

GitHub PRs ─▶ swegen ─▶ trajgen ─▶ sft ─▶ rl

SWE-Lego-Live provides:

• A block abstraction that makes every pipeline stage swappable, restartable, and replayable
• swegen — curated SWE tasks generated and NOP/Oracle-verified directly from GitHub PRs
• trajgen — containerized agent rollouts at scale on Harbor, with pluggable agent scaffolds
• sft — supervised fine-tuning on the resulting trajectories (LLaMA-Factory + DeepSpeed ZeRO-3)
• rl — online RL on SWE-bench through Harbor + vLLM + verl
• Per-run archives with config, scripts, and repo SHAs frozen for every execution
• A Claude Code plugin (/root:check, /root:run, /root:create) that drives the tree end-to-end

Where to go next

• Getting Started — the three skills (:setup, :check, :run) used to drive every block
• Block Design — the abstraction the whole pipeline is built on
• Blocks: swegen · trajgen · sft · rl — each links out to that block's full docs site
• Reference — config.yaml schema and the archive format