[jnovinger]

Fixes: #812

Field type markers (Devices.device_type = DeviceTypes, Permissions.users = [Users], Devices.config_context = JsonField) are plain class attributes. When a payload omits one of those fields, normal attribute lookup finds the class attribute, so device.device_type returns the DeviceTypes class and Record.__getattr__ never runs to lazy-load it. The post-hydration probe in __getattr__ that the issue points at is never reached for these keys.

Record.__init_subclass__ now wraps each marker in a non-data descriptor. Class access still returns the marked type, so _parse_values() and any external code reading Devices.device_type see no change. Instance access raises AttributeError, so an omitted field lazy-loads like any other field and raises if the server omits it too.

The probe also changes from getattr()/hasattr() to object.__getattribute__. It no longer re-enters __getattr__ twice for a key that is still missing after hydration, and it still returns falsy values and re-runs properties.

Behavior changes for the release notes:

• Reading an omitted marker field on a record with a URL now fetches the full object (or raises AttributeError) instead of returning the class.
• hasattr(record, "device_type") on an unhydrated record fetches like any other missing field. dir() driven introspection (inspect.getmembers, debugger panes) can trigger the same fetch.
• Code that checked record.field is SomeClass to detect absence needs hasattr or try/except AttributeError instead.
• Any class body attribute holding a Record subclass, a JsonField style class, or [RecordSubclass] is treated as a marker. Plugin code that keeps a non-field reference like that should read it through type(self). docs/extensions.md now says so.

Devices.has_details = True and Interfaces.interface_connection are dead: __init__ always overrides has_details, and no NetBox serializer emits interface_connection. I left both in place because removing them saves no requests and changes class-level surface. I can remove them in a follow-up, but I am leaning against it.

This touches pynetbox/core/response.py alongside #794, so whichever lands second will need a rebase.

[jnovinger]

@claude /review Only report on issues of potential concern; we don't need commentary on anything that doesn't require further consideration.

[github-actions]

Claude finished @jnovinger's task in 6m 18s —— View job

---

Review

• Read changed files
• Analyze implementation
• Post review findings

---

pynetbox/core/response.py

_FieldTypeMarker descriptor design

The non-data descriptor (no __set__) is the right tool here. setattr(self, k, v) in _parse_values writes to instance.__dict__, which takes precedence over non-data descriptors in Python's lookup order, so present fields are returned from __dict__ and absent fields fall through to the descriptor (which raises). Clean.

_is_field_type_marker uses hasattr(value, "_json_field") not a value check

hasattr(value, "_json_field") returns True for any class carrying that attribute regardless of its value. A class with _json_field = False would still be treated as a field-type marker and wrapped. Since _json_field is an internal convention and the only classes that carry it are intentional JsonField-style markers, this is unlikely to bite anyone, but getattr(value, "_json_field", False) would be more precise.

__getattr__ change fixes a second, pre-existing bug

The old getattr(self, k, None) + hasattr(self, k) probe had an undocumented failure mode for properties: if a property raised AttributeError before hydration and then succeeded after, getattr(self, k, None) saved None as ret, and hasattr (returning True) caused __getattr__ to return None rather than the property value. object.__getattribute__ re-invokes the property descriptor so it gets the correct result. test_property_reruns_after_hydration covers this, but it may be worth a brief note in the release notes since it's a quiet behavior fix on top of the marker fix.

Interfaces.interface_connection = InterfaceConnection is now wrapped

The PR acknowledges this is dead code (no NetBox serializer emits interface_connection). With the descriptor in place, any access to interface.interface_connection on a record that has a URL will now trigger one fetch attempt (capped by has_details). That is worse than before (was a silent no-op returning the class), but since the field is never in any payload, no fetch completes successfully and AttributeError is raised. The risk is only if someone was using record.interface_connection is InterfaceConnection as an existence check, which seems unlikely. Leaving it is fine for now per the PR rationale.

Devices.has_details = True not wrapped (correct) but still confusing

has_details = True on Devices is a bool, so _is_field_type_marker correctly skips it. Record.__init__ always writes self.has_details = False as an instance attribute, shadowing the class attribute immediately. The class attribute is harmless but misleading; __init_subclass__ does not interact with it. Deferring cleanup is fine.

---

tests/unit/test_response.py

Coverage is thorough — missing field, present field, list marker, JsonField marker, single-underscore marker, inheritance, dynamic subclass creation, pickle round-trip, lazy-load, server-omits-field-after-fetch, hasattr-fetches-once, nested brief record, falsy-value return, and property-reruns-after-hydration are all exercised.

One observation: test_hasattr_on_missing_marker_fetches_once patches Request.get with return_value=iter([...]). If the mock were called a second time it would return an exhausted iterator and full_details() would raise StopIteration. The test is correct in asserting call_count == 1, but if the caching breaks the test failure message would be a confusing StopIteration from inside the context manager rather than a clear assertion failure. A minor robustness issue in the test itself — not in the production code.

---

docs/extensions.md

The new sentence is correct but is doing a lot of work in one line. It reads well enough for existing contributors but may be hard for a first-time extension author to parse. A follow-up to split it into two or three shorter bullet points would help, but it is not a blocker.

---

Summary: The core mechanism is sound. The only items worth a follow-up are (1) getattr → getattr/hasattr precision in _is_field_type_marker, (2) the docs sentence density, and (3) the dead has_details/interface_connection class attributes — all acknowledged in the PR description and none blocking merge.

[jnovinger]

_is_field_type_marker uses hasattr(value, "_json_field") not a value check

This matches the three hasattr(lookup, "_json_field") checks in _parse_values(). Switching only the new one would let the two disagree about what counts as a JsonField.

__getattr__ change fixes a second, pre-existing bug

The old probe did not have this bug. After hydration, getattr(self, k, None) re-ran the property and returned its value. It only fell back to None when the property still raised, and then hasattr was also False, so it raised anyway. test_property_reruns_after_hydration guards against a regression to a self.__dict__ lookup, which would skip the property.

One observation: test_hasattr_on_missing_marker_fetches_once patches Request.get with return_value=iter([...]).

The #808 tests in this file use the same mock pattern, and a second fetch still fails the test. I am keeping it consistent with those.

The new sentence is correct but is doing a lot of work in one line.

Split into two bullets in ddfb71a.