Motivation

Tracer turns SWE tasks into agent trajectories and, optionally, ready-to-train SFT data.

It is the trajectory-generation stage of the LegoFlow pipeline, sitting between task curation (curator) and model training (trainer). It rolls a coding agent out in reproducible containers, captures each rollout, and can convert successful trajectories into SFT data. One config.yaml pins the Harbor and swe_data_process runtimes and declares every external input and downstream output.

curator ─▶ tracer ─▶ trainer

tracer provides:

• Curator and Hugging Face task sources — local curator output is filtered by verifiable_tasks.txt when present, while compatible Hugging Face task packages can be staged directly
• Processed-task protection — a ledger records terminal tasks and start.sh converts that state into Harbor exclusions on the next launch
• A per-job LiteLLM proxy fronting your model API with OpenAI- and Anthropic-compatible endpoints
• Containerized rollouts at scale via Harbor, with configurable concurrency, retries, and timeouts
• Configurable rollout agents — a pinned custom Claude Code preset by default, with custom OpenCode and OpenHands alternatives when all preset fields are aligned
• Multi-scaffold SFT conversion — Claude Code, OpenCode, OpenHands SDK, and Terminus-2 converters with TQS V2 rule scoring and optional LLM scoring
• A live progress dashboard you can preview locally or publish to Cloudflare Pages

Where to go next

• Getting Started — operate through the plugin or run the underlying commands
• Core Concepts — tasks, trajectories, jobs, scaffolds, and SFT data
• Run Jobs — prepare tasks, start the proxy, and roll the agent out
• SFT Data — convert trajectories into training datasets
• Dashboard — monitor progress locally or online