Affected Repositories

• PyAutoLens (primary, library — ships first)
• autolens_workspace (workspace)
• euclid_strong_lens_modeling_pipeline (workspace)

Branch Survey

Suggested branch: feature/coolest-observation-grid
Work Classification: Both (library first)
Worktree root: ~/Code/PyAutoLabs-wt/coolest-observation-grid/

Implementation Steps

1. PyAutoLens — autolens/interop/coolest.py

to_coolest(galaxies, file_path, cosmology=None, mode="MAP", pixel_size=0.1, shape_native=None, dataset=None, on_unsupported="raise", metadata=None) -> str

• Observation grid: if dataset (anything with .shape_native and .pixel_scales) is given, pixel_size = dataset.pixel_scales[0], shape_native = dataset.shape_native. With shape_native=(ny, nx) build lazy.PixelatedRegularGrid(field_of_view_x=(-nx*ps/2, nx*ps/2), field_of_view_y=(-ny*ps/2, ny*ps/2), num_pix_x=nx, num_pix_y=ny) and pass lazy.Observation(pixels=grid). num_pix_x is columns (shape_native[1]). No fits_path (check_external_files=False already). Neither given → unchanged zeros plus warnings.warn that COOLEST plotting needs a pixel grid.
• Unsupported components: wrap the coolest_dict_from_light / coolest_dict_from_mass calls; on ValueError with on_unsupported="skip" append {"galaxy": "galaxy_<i>", "profile": <name>} to metadata["skipped_profiles"] and continue; "raise" re-raises (existing behaviour, workspace_test round-trip untouched); other values → ValueError. A Basis (MGE) is itself a LightProfile in galaxy.cls_list_from(cls=LightProfile) — skip it as one entry Basis(Gaussian x N). A Pixelization is not a LightProfile; record it via galaxy.cls_list_from(cls=Pixelization) as Pixelization(<mesh cls>). A galaxy with everything skipped still exports as an entity with empty models.
• Docstring: new params, the metadata key.

Tests test_autolens/interop/test_coolest.py: observation block from shape_native=(10, 12), pixel_size=0.05 → fov_x (-0.3, 0.3), fov_y (-0.25, 0.25), num_pix 12/10, instrument pixel size == fov_x span / num_pix_x; dataset= path; no grid → warning, zeros; MGE + pixelized galaxy default raises, skip writes the mass model and records both; from_coolest on the skipped template returns the mass-only tracer.

2. autolens_workspace — scripts/guides/coolest_interop.py

Export cell passes shape_native + pixel_size (or dataset=), prints the observation block, one sentence on why COOLEST needs the grid. New short section "Unsupported profiles" after __Supported Profiles__. Regenerate the notebook per this workspace's convention; no unrelated notebook diffs. autolens_workspace_test/scripts/misc/interop/coolest_round_trip.py only if the new warning makes the smoke run noisy.

3. euclid_strong_lens_modeling_pipeline

• util.py AnalysisImaging.save_results: after wcs.json, a module-level helper (e.g. coolest_json_from(...)) wraps al.interop.coolest.to_coolest(galaxies=result.max_log_likelihood_tracer, file_path=paths._files_path / "coolest", dataset=self.dataset, on_unsupported="skip", metadata={"pipeline": ..., "search": paths.name, "dataset": <name>}). Guarded like the wcs.json block: ImportError and any other exception → logging.warning("coolest.json: ..."), never fails the fit.
• scripts/tools/build_inspect.py process_dataset: best-effort coolest.json (from initial_lens_model/vis_pix member files/coolest.json) and coolest_sersic.json (from sersic_lens_model/vis) via the existing zip/dir member streaming. catalogue/README.md table: two new rows; build_inspection_bundle.sh header comment updated.
• README.md: final section "COOLEST output" (after "Visualization"): every search writes files/coolest.json (MAP template: full mass model, Sersic light; MGE and pixelized components under meta.skipped_profiles); bundle collects coolest.json / coolest_sersic.json; pip install autolens[coolest] (add [coolest] to the Getting Started install line); missing package = logged warning. Citation Galan et al. 2023, JOSS 8(88) 5567, doi:10.21105/joss.05567 — BibTeX copied from the installed coolest METADATA; link the workspace guide. AGENTS.md: one line in the output-contract list if present.
• Tests: tests/test_coolest_output.py (not slow, JAX-free) — helper on a Basis+Pixelization+SIE+shear tracer writes JSON whose grid matches a 20x20 / 0.1" mock dataset and whose metadata lists the skipped profiles; monkeypatched ImportError → warning, no file. Extend the slow run-level test that asserts wcs.json to assert files/coolest.json. smoke_tests.txt unchanged.
• Pre-run checklist (human): pip install coolest in the RAL venv /mnt/ral/jnightin/PyAuto — HPCPullPyAuto syncs library mains, not new pip deps.

4. euclid_dr1 science clone (/mnt/c/Users/Jammy/Science/euclid_dr1), after the pipeline PR merges

git merge origin/main (never push); pytest tests/test_coolest_output.py -q; hpc/batch_cpu/submit_reload_coolest copied from submit_reload_wcs (array 0,2-9, no --seed); output.force_pickle_overwrite: true toggle + revert commits; ledger page wiki/project/2026-09-<dd>-coolest-json-reload-pass.md. RAL leg (push, submit, pull, rebuild dr1_sep1 bundle, confirm coolest.json per lens) only on the human's go.

Key Files

• PyAutoLens/autolens/interop/coolest.py — to_coolest observation grid + skip policy
• PyAutoLens/test_autolens/interop/test_coolest.py — tests
• autolens_workspace/scripts/guides/coolest_interop.py — guide
• euclid_strong_lens_modeling_pipeline/util.py — save_results hook + helper
• euclid_strong_lens_modeling_pipeline/scripts/tools/build_inspect.py — bundle collection
• euclid_strong_lens_modeling_pipeline/README.md, catalogue/README.md — docs + citation
• euclid_strong_lens_modeling_pipeline/tests/test_coolest_output.py — new tests

Verification

• PyAutoLens: pytest test_autolens/interop -q; load a written template with coolest's JSON serializer and assert observation.pixels.num_pix_x == 100, field_of_view_x == (-5.0, 5.0); render ModelPlotter once if coolest.api.plotting imports.
• Workspace: run the guide end-to-end (round-trip deflection difference still ~1e-12); notebook regenerated.
• Pipeline: pytest -q -m "not slow", pytest -q -m slow, smoke initial_lens_model.py under PYAUTO_TEST_MODE, build the bundle on the test-mode output and confirm coolest.json is collected; open one coolest.json and check mass model, meta.skipped_profiles, and grid = shape × pixel scale.

COOLEST: fix the empty Observation pixel grid, emit coolest.json from the Euclid pipeline by default, carry it into euclid_dr1

Type: feature
Target: PyAutoLens
Repos:

• PyAutoLens
• autolens_workspace
• euclid_strong_lens_modeling_pipeline
  Themes:
• coolest
• interop
• euclid
  Difficulty: medium
  Autonomy: supervised
  Priority: high
  Status: formalised
  Consequence: judge
  Filed: 2026-09-16

Request (verbatim)

Hi Aaron,
The models and results for the 3,000 DR1 lenses described in ECLIPSE C are here: Next week I'm planning to model the full DR1 sample (~15,000 lenses). I've also added COOLEST outputs to PyAutoLens, which are described here:
https://github.com/PyAutoLabs/autolens_workspace/blob/main/scripts/guides/coolest_interop.py
These aren't included in the 3,000 models above. it's a new feature that I'll use for the full run next week.
The COOLEST guide and the conversion code above are fairly vibe-coded and have had less human oversight than I'd normally like. If you could take a look and confirm that everything makes sense before I blitz through the full DR1 sample next week, that'd be massively appreciated!

Hey James,

I just wanted to let you know that I checked out the COOLEST conversion code. I noticed one small problem with the json it created. In the "observation/pixels" attribute, "field_of_view_x(/y)" and "num_pix_x(/y)" are all automatically set to 0. It appears that the al.interop.coolest.to_coolest() imports an empty COOLEST Observation object with these values preset to 0.

When I try to plot with these values in the COOLEST api, however, COOLEST doesn't know how to handle them. After editing the values in the json so that they result in a pixel scale equal to the one found in the "instrument" attribute, COOLEST makes a nice-looking plot. This was the only issue I was able to find so far!

Can you 1) Check and fix the issue Aaron brought up; 2) ensure COolest is supported documenting and working on the euclid_strong_lens_modeling_pipeline, dont make it front and centre but ensure runs output coolest by default and that the bottom of the README.md mentions this with citation this includes ensuring its output as a .json file when we do catalogue generation; 3) Add Coolest support to the euclid_dr1 science project and ensure its in the catalogue, if it isnt already!

Scope

1. PyAutoLens — to_coolest builds a PixelatedRegularGrid for lazy.Observation from
   shape_native + pixel_size (or a dataset=), warns when no grid is given, and gains
   on_unsupported="skip" which exports the mass model plus any Sersic light and records the
   omitted components (MGE Basis, pixelized source) under meta.skipped_profiles. Tests.
2. autolens_workspace — the COOLEST guide passes a pixel grid and documents the skip policy;
   notebook regenerated.
3. euclid_strong_lens_modeling_pipeline — every search writes files/coolest.json from
   save_results (guarded like wcs.json, missing package is a logged warning); the
   inspection bundle collects coolest.json / coolest_sersic.json; README gains a final
   "COOLEST output" section with the Galan et al. 2023 JOSS citation and autolens[coolest]
   in the install line; unit + slow tests.
4. euclid_dr1 science clone (after the pipeline PR merges) — merge pipeline main, add
   hpc/batch_cpu/submit_reload_coolest, ledger page; RAL reload leg on the human's go.

Decision (user-confirmed at planning): unsupported light/source components are skipped and
recorded, not converted. Pixel-grid export (MGE → PixelatedRegularGrid, pixelized source →
IrregularGrid) is the follow-up prompt draft/feature/autolens/coolest_pixel_grid_export.md.