[Josiah Vinson (jovinson-ms)]

Problem

In the generated model runtime (packages/http-client-python/generator/pygen/codegen/templates/model_base.py.jinja2), _RestField.__get__ deserializes list/dict/set values into a new container, writes that new container back into obj._data[rest_name], and sets _deserialized_<rest_name>. Any container a caller already got through obj["rest_name"] is now detached from _data, so mutations made through it are silently dropped when the model is serialized.

Repro

src = SourceStorageLocation(location="https://x.blob.core.windows.net/c", prefix="p", extensions=[".txt"])
filters = src["extensions"]
_ = src.extensions          # replaces _data["extensions"] with a new list
filters.append(".csv")
print(src.as_dict()["extensions"])

• Expected: ['.txt', '.csv']. Dict access and property access share one backing container.
• Actual: ['.txt']. The append is lost.

Observed with @azure-tools/typespec-python 0.63.10 / @typespec/http-client-python 0.39.0.

What this PR adds (tests only, no fix)

In tests/unit/test_model_base_serialization.py:

• test_list_getitem_reference_survives_attribute_access: fails (list[str], the case above)
• test_dict_getitem_reference_survives_attribute_access: fails (dict[str, str] variant)
• test_list_attribute_reference_mutation_serializes: passes. Guard: property access first, then mutate, still serializes correctly and must keep working after a fix.

The failing tests are intentionally left plainly failing (not xfail). This package's tests have no xfail convention, and the goal of this draft is to show a red test that the fix turns green. If maintainers prefer, they can be switched to @pytest.mark.xfail(strict=True, reason=...).

A .chronus changeset with changeKind: internal is included because packages/http-client-python/tests/** is not excluded by changedFiles.

Local run

model_base.py.jinja2 was rendered (unbranded, corehttp) into a temporary specialwords._utils.model_base package. unit/test_model_base_serialization.py was then run with --noconftest, because its conftest auto-starts the Spector mock server:

FAILED test_list_getitem_reference_survives_attribute_access - AssertionError: assert ['.txt'] == ['.txt', '.csv']
FAILED test_dict_getitem_reference_survives_attribute_access - AssertionError: assert {'a': '1'} == {'a': '1', 'b': '2'}
2 failed, 123 passed

The guard test and all existing tests in that file pass.

Likely location and suggested fix direction

_RestField.__get__ in model_base.py.jinja2:

if isinstance(deserialized, (dict, list, set)):
    obj._data[self._rest_name] = deserialized   # <-- replaces the container
    object.__setattr__(obj, cache_attr, True)
    return deserialized

Suggested direction:

• If deserialization didn't convert anything (deserialized == item and same type), keep the existing _data container (return item) instead of replacing it.
• Otherwise, update the existing container in place so outstanding references stay attached, e.g. item[:] = deserialized for lists, item.clear(); item.update(deserialized) for dicts/sets.

Found during review of Azure/azure-sdk-for-python#49627.

Created by Copilot on behalf of Josiah Vinson (Josiah Vinson (@jovinson-ms))

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@typespec/http-client-python@12359

commit: 08f5b6c

[github-actions]

All changed packages have been documented.

• ✅ @typespec/http-client-python

Show changes

@typespec/http-client-python - internal ✏️

Add failing unit tests reproducing a container returned by __getitem__ becoming detached after property access in generated model_base.

[github-actions]

Python emitter diff

Baseline gh:721154b2ab55bfc78ab4850e86c40487f66241f2 vs this PR.

No changes to generated output.

Rendered diff: inline on the run summary, or the emitter-diff-html artifact.

Informational check (eng/emitter-diff); does not block the PR.

[azure-sdk-automation]

You can try these changes here

🛝 Playground	🌐 Website	🛝 VSCode Extension