Skip to content

AGENTS.md

Latest commit

 

History

History
95 lines (71 loc) · 4.43 KB

AGENTS.md

File metadata and controls

95 lines (71 loc) · 4.43 KB
Raw

PyAutoGalaxy Workspace Test — Agent Instructions

This is the integration-test suite for PyAutoGalaxy, run on the build server to verify the core library works end-to-end. It is not a user-facing workspace — see ../autogalaxy_workspace for examples and tutorials. These are the canonical, agent-agnostic instructions for this repo.

Dependencies: autogalaxy, autofit, autoarray, numba.

Repository Structure

scripts/                     Integration-test scripts run on the build server
  imaging/ interferometer/   CCD imaging / interferometer tests; each dataset folder
  ellipse/ multi/            holds visualization/, jax_grad/ and jax_likelihood/ subfolders
  misc/                      Dataset-agnostic tests:
    aggregator/              Results database aggregator tests (pytest-based)
    model_composition/       Model-composition tests
    jax_assertions/          JAX assertion tests
    latent/                  Latent-variable tests
failed/                      One log per failing script
config/ output/              YAML config and runtime fit results

Running Scripts

Run a single script directly from the repo root (with no env applied, searches run for real). The aggregator tests are pytest-based and run from their directory:

python scripts/imaging/model_fit.py
NUMBA_CACHE_DIR=/tmp/numba_cache MPLCONFIGDIR=/tmp/matplotlib python -m pytest scripts/misc/aggregator -v

Testing

On CI, smoke_tests.yml gates every PR on Python 3.12 and 3.13. The gate runs the smoke runner (the definition of green):

python .github/scripts/run_smoke.py

It executes the curated entries in smoke_tests.txt, applying per-entry environment from config/build/profile_smoke.yaml. That file sets fast-mode defaults for every entry — PYAUTO_TEST_MODE=2 (skip the sampler), PYAUTO_SMALL_DATASETS=1 (cap grids/masks), PYAUTO_FAST_PLOTS=1 — with per-script unset/override blocks where a test genuinely needs a real sampler run or full-resolution data. So CI is not "searches run for real" by default; it is fast-mode with targeted exceptions. A failure under these flags signals a real problem.

Sandboxed / restricted runs

If numba or matplotlib cannot write to the default cache locations, point them at writable dirs:

NUMBA_CACHE_DIR=/tmp/numba_cache MPLCONFIGDIR=/tmp/matplotlib python scripts/imaging/model_fit.py

Aggregator & JAX Testing

  • Aggregator (scripts/misc/aggregator/) — MockSearch + MockSamples build mock model-fit results, scrape them into an SQLite database, then exercise each aggregator class (GalaxiesAgg, FitImagingAgg, ImagingAgg, InterferometerAgg, EllipsesAgg, FitEllipseAgg, MultipolesAgg, …) via its generator methods.
  • JAX (<dataset>/jax_likelihood/, <dataset>/jax_grad/, misc/jax_assertions/) — exercises the xp=jnp path that library unit tests (NumPy-only) never touch, via fitness._vmap. See the PyAutoArray deep dive ../PyAutoArray/docs/agents/jax_and_decorators.md for the boundary patterns.

Bulk-edit safety

When editing the same region across many scripts in one pass, only rewrite the targeted region. Never produce a whole-file write unless you have read the entire current file — a whole-file write from a header skim silently deletes every section below the header.

Related Repos

  • Source libs: ../PyAutoGalaxy, ../PyAutoArray, ../PyAutoFit, ../PyAutoNerves.
  • ../autogalaxy_workspace — the user-facing workspace; ../HowToGalaxy — the tutorial series.
  • ../PyAutoHands — CI / build tooling.
  • ../autolens_assistant — science-assistant workspace (literature wiki).

Task Workflows

When a library change lands, run the smoke suite, read any [FAIL] entries, and update the affected test scripts to the new API (preserving intent). Never edit a script to mask a real regression — if a library bug surfaces, flag it for the source repo rather than papering over it. Note in your PR any change that affects sibling repos (autogalaxy_workspace, the source libraries).

Never rewrite history

Never rewrite pushed history on any repo with a remote — no git init over a tracked repo, no force-push to main, no fresh-start "Initial commit", no filter-repo / filter-branch / rebase -i on pushed branches. To get a clean tree: git fetch origin && git reset --hard origin/main && git clean -fd.