The worked example: the live lensing instance#

Everything on this site is running today, developing the PyAuto astronomy stack. This page is that instance, labelled as the example — when a concept page says “a library” or “a workspace”, this is what it concretely means upstream. Your organism replaces every name below with your own.

The body#

The live repos.yaml declares ~25 repos. The load-bearing categories:

  • Libraries: PyAutoConf (shared config), PyAutoFit (Bayesian inference), PyAutoArray (data structures), PyAutoGalaxy (galaxy modelling), PyAutoLens (strong lensing), PyAutoReduce (data reduction) — a dependency chain released to PyPI nightly when there is new activity.

  • Workspaces: autofit_workspace, autogalaxy_workspace, autolens_workspace — user-facing examples, notebook-generated, version-pinned against their libraries.

  • Test workspaces: the *_workspace_test trio — regression, smoke and parity scripts the release pipeline runs.

  • HowTos: HowToFit / HowToGalaxy / HowToLens — lecture-style courses.

  • Assistant: autolens_assistant — the curated knowledge pack that makes any AI assistant a strong-lensing expert.

  • Memory: sub-wikis on lensing, black holes, detector calibration, inference methods and galaxy evolution, with a ~600-paper bibliography layer.

A task, end to end#

A representative real task (2026): a prompt file feature/autoarray/psf_oversampling.md written as free prose — “PSF blurring lives in @PyAutoArray/.../convolver.py; modeling should convolve at higher resolution than the image; here’s roughly how” — with typos and half-decisions left in. start_dev classified it (feature, library), planned it at two levels, opened a tracked issue, and claimed a worktree. Development touched PyAutoArray and PyAutoGalaxy; the ship gate ran both test suites, the review faculty, and Heart’s verdict; one PR landed per repo; the workspaces then gained a demonstration script in a follow-up workspace task. The registry recorded every state transition, and the task retired to complete.md.

That is the whole pitch: the input was a paragraph of intent, and every step from there — planning, isolation, testing, gating, releasing, bookkeeping — was the organism’s machinery, with a human approving the plan and the merge.

Scale, honestly#

This system runs tens of active repos with a single human maintainer. In a typical recent quarter the framework organs themselves took a few hundred commits — which is exactly why the stability disclaimer exists and why adoption is a fork, not a dependency.