Affected Repositories

PyAutoLens (library); PyAutoMind task bookkeeping only.

Branch Survey

PyAutoLens: clean main at 7197380; no conflicting claim; only local branch main.
Suggested branch: feature/mass-field-bare-fields
Worktree root: ~/Code/PyAutoLabs-wt/mass-field-bare-fields/

Implementation Steps

1. Capture baseline from repository root with PYAUTO_SKIP_WORKSPACE_VERSION_CHECK=1. Expect collection identifier c5cf98ea7733689dc8c0ede6939c1d83 and flat 36a0be37c9667958bafca2f01e487f80; reproduce rather than assume. Capture prior_count 14 and paths.
2. Add witness in test_autolens/analysis/analysis/test_analysis_lens.py; run red before source edits and retain traceback.
3. Add internal tracer_util.fields_list_from: None -> [], isinstance(ag.MassField) -> [fields], otherwise list(fields). Never test the shared MassProfileAggregate base.
4. In tracer._validate_fields wrap bare MassField after ModelInstance escape hatch, before string trap. Preserve every message. Bare Galaxy continues to get generic container rejection; [Galaxy] retains the targeted message. Direct generators remain rejected without consumption.
5. Tracer.init: validate first, normalize before _warn_if_no_light_is_behind_any_mass, store normalized list. Replace folds in sliced_tracer_from, analysis/analysis/lens.py, aggregator/tracer.py, analysis/result.py. Update constructor and slicing annotations.
6. Tests: bare/list storage and list ownership, deflections and nonzero MassSheet convergence, members/redshifts, geometry warning, JSON parity, slicing redshift snapping, bare Galaxy rejection, flat/collection paths/count/collection identifier; analysis witness and likelihood agreement to 1e-8; Result fields for flat/collection/absent/None/empty; aggregator regression for flat/collection/absent. Existing validation and identifier tests stay untouched. NumPy only.
7. Docs: helper module bullet; Tracer and validation docstrings; docs/api/galaxy.rst and modeling.rst; model_util.mass_field_from. Keep collection primary; internal helper stays out of public exports/autosummary.
8. Full suite: NUMBA_CACHE_DIR=/tmp/numba_cache MPLCONFIGDIR=/tmp/matplotlib python -m pytest test_autolens/. PR states additive API, witness traceback, unchanged collection evidence, and different identifiers by design.

Review conclusions

Five normalization sites exhaust autolens source. Serialization reads self.fields; JAX reconstructs dict without constructor re-entry. No PyAutoGalaxy changes. No standalone slam package in checkout. Human approved reviewed plan with 'proceed'.

Accept a bare MassField in fields= so the flat model form works

Type: feature
Target: PyAutoLens
Repos:

• PyAutoLens
  Themes:
• cluster
  Difficulty: small
  Autonomy: human-required
  Priority: normal
  Consequence: judge
  Epic: mass-field
  Filed: 2026-09-18
  Witness: with the branch installed, af.Collection(galaxies=..., fields=af.Model(al.MassField, redshift=0.5, shear=af.Model(al.mp.ExternalShear))) yields prior paths fields.shear.gamma_1/gamma_2 (not fields.field.shear.…), AnalysisImaging's tracer construction succeeds on its instance, and tracer.fields == [MassField(redshift=0.5, shear)]. On main the same model raises TypeError: 'MassField' object is not iterable. The collection form fields=af.Collection(field=field) is unchanged in paths, prior count and result identifier.

Original request

The following API: field = af.Model(al.MassField, redshift=0.5, shear=af.Model(al.mp.ExternalShear)) and model = af.Collection(
galaxies=af.Collection(lens=lens, source=source),
fields=af.Collection(field=field),
) produces model.fields.field.shear, is there a way to avoid the double API of field.shear?

and, after the options were reported back:

do this If you want the flat form to work, it's a small library change in PyAutoLens, in two places:

1. PyAutoLens/autolens/analysis/analysis/lens.py:138 — wrap a bare MassField in a list rather than calling list() on it.
2. PyAutoLens/autolens/lens/tracer.py:240 (_validate_fields) plus the self.fields = list(fields) line in Tracer.init — accept a single MassField.

Adding iter to MassField in PyAutoGalaxy would also satisfy list(), but it would yield the field's profiles, not the field — wrong for Tracer(fields=...). So the lens-side change is the right one.

Why

The shipped composition is fields=af.Collection(field=field), giving parameter
paths fields.field.shear.gamma_1. The middle level is a name the user must
invent for a collection that, in every workspace example today, holds exactly
one entry — so it reads as a stutter (fields.field) rather than as
information. The galaxy-attached form it replaced was one level shallower
(galaxies.lens.shear.gamma_1).

Measured against the installed stack 2026-09-18:

prior_count is 14 in all four. Only the flat form removes the level, and it is
the one that breaks: MassField and its base MassProfileAggregate define
neither __iter__ nor __getitem__, so list(instance.fields) raises, and
_validate_fields independently rejects a non-list/tuple non-ModelInstance.

Scope

Three edits, all in PyAutoLens:

1. autolens/analysis/analysis/lens.py (~line 138, _tracer_via_instance_from)
   — the fields = list(instance.fields) if getattr(instance, "fields", None) is not None else None block wraps a bare MassField in a list instead of
   calling list() on it.
2. autolens/lens/tracer.py _validate_fields (line 240) — accept a single
   MassField alongside the list/tuple/ModelInstance forms, keeping the
   string trap and the Galaxy-in-fields message intact.
3. autolens/lens/tracer.py Tracer.__init__ (~line 352) — self.fields = list(fields) normalises a single MassField to [field], preserving the
   "plain list, never None" invariant the members-walking properties rely on.

Explicitly not in scope: adding __iter__/__getitem__ to MassField in
PyAutoGalaxy. It would satisfy list() but yield the field's profiles, not
the field, which is wrong for Tracer(fields=...) and would silently produce a
tracer holding ExternalShear objects with no redshift.

Open questions for the plan

• Epic decision conflict. The mass-field epic's Decisions section reads
  "Model slot fields=, a collection; tracer argument fields=, a list. The
  analysis folds list(instance.fields) into Tracer(fields=...) exactly as it
  folds extra_galaxies / scaling_galaxies into galaxies." Accepting a bare
  field relaxes that and breaks symmetry with extra_galaxies /
  scaling_galaxies, which stay collection-only. Decide whether the relaxation
  is accepted as a second supported form (both work) and record it in the epic
  ledger.
• Which form the workspaces teach. Phase 3 (autolens_workspace +
  autolens_workspace_test, 217 + 87 files) has shipped as draft PRs using the
  collection form, and phase 5 (HowToLens, autolens_assistant) is unstarted. If
  the flat form becomes the taught idiom, that is a follow-up sweep, not this
  task. This task adds the capability only.
• Multi-field systems. The collection form stays mandatory for more than one
  field (per-plane LOS sheets, foreground/background tidal planes); the flat
  form is a single-field convenience. Error messages should say so.
• Whether Tracer(fields=<bare MassField>) should also be accepted directly, or
  only the model slot. Scope above assumes both, since _validate_fields is the
  shared door.

Notes

• Backwards compatibility is the epic's hard invariant: the collection form's
  paths, prior count and PyAutoFit result identifier must be untouched, and the
  galaxy-attached form stays unwarned.
• Diagnostic script behind the table:
  /tmp/claude-1000/-home-jammy-Code-PyAutoLabs/65d55a6f-8729-4857-be92-b098eeb4031d/scratchpad/field_variants.py
  (scratch, not committed).