swegen

swegen is the data-curation stage. It collects PRs from configured GitHub repositories, scores them with an LLM, generates Harbor-compatible SWE tasks, and verifies each one with NOP and Oracle checks before adding its ID to the per-language verifiable_tasks.txt manifest.

Full docs: swe-swegen-docs.pages.dev/docs

Long-form docs live there

The swegen block ships its own complete documentation site — getting started, PR collection, task generation, NOP/Oracle verification, and the full reference. This page is just an orientation; follow the link above for everything else.

At a glance

Inputs: github_tokens, llm_api (api_key, api_base_url, pr_model, task_model).
Output: swe_tasks_dir — subblock/swegen/artifacts/swe_tasks/{lang}-cc/. The authoritative manifest is verifiable_tasks.txt; only task IDs in that file have passed NOP/Oracle.
Runs: Local (CPU + Docker). Long-running — hours per language.

How to run

/swegen:setup     # clone repos, build envs, prepare default config
/swegen:check     # preflight: tokens, LLM endpoint, Docker, schema
/swegen:run       # full pipeline: collect PRs → generate tasks → verify
/swegen:dashboard # progress dashboard

Reference

Block contract: subblock/swegen/CLAUDE.md
Full docs site: swe-swegen-docs.pages.dev/docs

swegen

At a glance

How to run

Reference

On this page