[kaia-cod]

Problem

Closes #1245.

When an inner validator of deep_iterable() or deep_mapping() fails, the error message names the whole attribute instead of the offending member, so on a nested structure you learn that something is wrong but not where:

@attrs.define
class C:
    x = attrs.field(validator=attrs.validators.deep_iterable(
        member_validator=attrs.validators.min_len(1),
        iterable_validator=None,
    ))

C(x=[[], [1]])
# ValueError: Length of 'x' must be >= 1: 0
#                        ^^^ no position

The same happens for deep_mapping() with both the key and the value validator:

C(m={"a": 1, "b": "nope"})
# TypeError: ('m' must be <class 'int'> (got 'nope') ...)
#             ^^^ no key

Root cause

_DeepIterable.__call__ and _DeepMapping.__call__ invoke the member validator directly with the outer Attribute object. An Attribute carries its name in .name, and every built-in validator formats its message from attr.name — so the inner failure can only ever say x / m.

Fix

src/attr/validators.py (+33/−4):

• add a tiny _MemberAttribute stand-in that forwards everything (__getattr__) to the parent attribute and overrides only name (→ "x[1]", "m[2]", "m['a']") and __repr__;
• _DeepIterable passes it as attr to member_validator;
• _DeepMapping builds it once per key from repr(key) and passes it to both key_validator and value_validator.

No public API change: the shim is private (_MemberAttribute), __slots__ is kept, and the object quacks like the attribute it wraps, so third-party validators that read attr.name, attr.type, attr.metadata, … keep working — they just get a more precise name.

After the change:

ValueError: Length of 'x[1]' must be >= 1: 0
TypeError: 'm[2]' must be <class 'str'> (got 2 that is a <class 'int'>) ...

Tests

tests/test_validators.py (+60/−1) — new TestInnerValidatorErrorMessages, 4 cases, one per call site:

• test_deep_iterable_names_the_failing_member
• test_deep_iterable_names_the_wrongly_typed_member
• test_deep_mapping_names_the_failing_value
• test_deep_mapping_names_the_failing_key

Each test asserts the exact message produced by the real validators (no mocks) and is expected to fail on main.

Red before green

Branch base: 9d9f94a043de7fc4012459853a4ed597a5577855 (main).

On the base commit with only the new tests applied:

$ pytest tests/test_validators.py -k TestInnerValidatorErrorMessages -q
FAILED ...::test_deep_iterable_names_the_failing_member
FAILED ...::test_deep_iterable_names_the_wrongly_typed_member
FAILED ...::test_deep_mapping_names_the_failing_value
FAILED ...::test_deep_mapping_names_the_failing_key
4 failed, 168 deselected          # exit 1

On the branch (single commit 2529f830da93e42fd7f6a5d09b4204e95a4ffcd0, tree 0634e0d744e7c0f51288023675e210417a8e0c72):

$ pytest tests/test_validators.py -k TestInnerValidatorErrorMessages -q
4 passed, 168 deselected          # exit 0

Quality gates (on the branch tree)

gate	command	result
full suite	pytest -q	1404 passed, 11 skipped, 1 xfailed — exit 0
lint	ruff check --fix --exit-non-zero-on-fix .	All checks passed! — exit 0
format	ruff format --exit-non-zero-on-format .	left unchanged — exit 0

Diff: exactly 2 files, +93/−5, 1 commit on top of main. The base and the branch head share the same parent, so the diff is clean with no unrelated churn.

One behavioural change worth reviewing

not_() validates exc_types through deep_iterable, so TestNot_::test_bad_exception_args now reports 'exc_types[0]' instead of 'exc_types'. That assertion is the single line changed in an existing test (+1/−1 in tests/), and it is arguably the improvement this issue asks for — but if you'd rather keep the old wording there, I can gate the shim on an option and update this PR.

Notes

• No changelog fragment yet — tell me the format you want (changelog.d/1245.change.md?) and I'll push it.
• I'm happy to squash/rebase, drop the _MemberAttribute shim in favour of attrs.evolve-style naming, or split the deep_mapping part into its own PR if you prefer a smaller diff.

Reviewer

@hynek — requesting your review on this PR: it's a 2-file, single-commit fix for #1245 with a red-before-green test for every call site (deep_iterable member, deep_mapping key and value). I'd value your call on the two points above (the exc_types[0] wording and whether you want a changelog fragment), and I'll adjust within the day.

[kaia-cod]

Reviewer request

@hynek — requesting your review on this PR. It is a 2-file, +93/−5 change on top of 9d9f94a (main), single commit 2529f83.

What it does — deep_iterable() / deep_mapping() now hand the inner validator a private _MemberAttribute shim, so an inner failure names the offending member (x[1], m[2], m['a']) instead of the whole attribute (x, m). No public API change: the shim forwards every attribute access (type, metadata, …) to the real Attribute and overrides only name / __repr__.

Red before green — base commit 9d9f94a with only tests/test_validators.py from this branch applied:

$ pytest tests/test_validators.py -k TestInnerValidatorErrorMessages -q
4 failed, 168 deselected        # exit 1

On the branch: 4 passed, 168 deselected (exit 0).

Gates on the branch tree (0634e0d)

gate	command	result
full suite	pytest -q	1404 passed, 11 skipped, 1 xfailed — exit 0
lint	ruff check --fix --exit-non-zero-on-fix .	All checks passed! — exit 0
format	ruff format --exit-non-zero-on-format .	97 files left unchanged — exit 0

One behavioural change worth your decision — not_() routes exc_types through deep_iterable, so TestNot_::test_bad_exception_args now reports 'exc_types[0]' instead of 'exc_types'. That single assertion is the only pre-existing line this PR touches; it is arguably the improvement #1245 asks for, but say the word and I will gate the shim behind an opt-in so the old wording stays there.

No changelog fragment yet — tell me the changelog.d/ naming you want and I will add it. Happy to squash/rebase or split the deep_mapping half into its own PR if a smaller diff is easier to review.