Merging a branch drops the value of a field added in the branch on an object that existed before the branch #1043

Plugin Version

1.0.0 (also reproduced on main at b34ee80)

NetBox Version

4.7.4 (netbox-branching 1.2.3)

Python Version

3.14

Steps to Reproduce

1. In main, create a Custom Object Type with a text field label, and one object with label = existing.
2. Create a branch (iterative or squash merge strategy) and activate it.
3. In the branch, add a text field notes to the type.
4. In the branch, edit the existing object and set notes = merged.
5. Merge the branch.

Expected Behavior

In main, the existing object has notes = merged.

Observed Behavior

The merge succeeds with no error or warning. In main, the notes column exists, but the existing object has notes = NULL; the branch schema still holds merged. An object created in the branch with a value for the new field keeps its value.

Reproduced in a lab with the REST API and both merge strategies. Two test cases that fail on main (b34ee80), added next to _assert_field_added_in_branch_is_merged in tests/branching/test_merge.py:

FAIL: test_iterative_merge_updates_existing_object_with_field_added_in_branch
AssertionError: None != 'merged'
FAIL: test_squash_merge_updates_existing_object_with_field_added_in_branch
AssertionError: None != 'merged'

Suspected Cause

Two causes, one per strategy.

Iterative. The field CREATE replays through CustomObjectTypeField.save(), which calls get_model() (registering a class without the new field), adds the column, then calls clear_model_cache(). That evicts the type's entry from _model_cache, but the old class stays registered in apps.all_models. The object UPDATE then replays through netbox-branching's ObjectChange.apply(), which resolves changed_object_type.model_class(), so CustomObjectsPluginConfig.get_model(). That method returns the registered class unless it was rolled back, so the update runs on a class without the new field: update_object() sets a plain attribute and save() never writes the column. Traced in a NetBox shell during a merge of the lab reproduction, where the new field is named ticket:

APPLY create customobjecttypefield pk=27 class_id=404865456 has_ticket=False
APPLY update table7model pk=1 class_id=410558864 has_ticket=False
UPDATE_OBJECT class_id=410558864 has_ticket=False data={'ticket': 'CHG-1'}
  after: instance.ticket=CHG-1
APPLY create table7model pk=2 class_id=410558864 has_ticket=False

New objects are not affected because CustomObject.deserialize_object() builds a fresh model and stashes the data for _apply_deferred_co_field().

Squash. The squash orders DELETEs, UPDATEs, then CREATEs, so the object UPDATE runs before the field CREATE unless an edge says otherwise. add_custom_object_dependencies() finds an object's field refs through _collect_co_refs(), which reads model_class._field_objects; the main class does not know a field created in the branch, so no UPDATE → field CREATE edge is added.

Proposed Fix

1. In CustomObjectsPluginConfig.get_model(), rebuild a registered dynamic model (table<id>model) through CustomObjectType.get_model() when _model_cache has no entry for its own context (_context_branch_id, when it is the active context) and the type now has a field that the class lacks. The cached path stays a dictionary lookup with no query; the evicted path runs one exists() query. Rebuilding on eviction alone is not enough: model generation itself resolves other types through ContentType.model_class(), so an eviction-only check recurses during generation and replaces classes that other models still point to.
2. In add_custom_object_dependencies(), map the columns of each field created in the squash (_occupied_columns()) to its key, and add an UPDATE → field CREATE edge when an object's postchange_data writes one of them.

Both are needed for squash; the iterative case needs the first only. This depends on #1042: with change 1, test_choice_set_choices_mutated_in_branch_merge fails until that issue is fixed.

I have a fix and tests ready on a branch. I'd be happy to open the PR if you assign this issue to me.