Fixes #21763: Replace M2M selection field with separate add/remove fields

[jeremystretch]

Fixes: #21763

• Introduce M2MAddRemoveFields for conditional rendering M2M form fields as separate add/remove fields when an object has many (100+) assignments
• Employ M2MAddRemoveFields for ASN relations in the site and provider edit forms
• Update NetBoxModelForm to support the add/remove logic
• Add tests

[jeremystretch]

@claude /review Check that this PR provides a solution for bug #21763 and satisfies all of the "todo" criteria listed in the P description

[claude]

Claude finished @jeremystretch's task in 1m 41s —— View job

---

PR Review: Fixes #21763 - Replace M2M selection field with separate add/remove fields

Tasks:

• Gather context and understand the request
• Read changed files and analyze implementation
• Check TODO criteria satisfaction
• Post review

---

Overview

This PR addresses issue #21763 by introducing a new M2MAddRemoveFields rendering primitive that switches M2M form fields between two modes based on the number of existing assignments. When an object has ≥100 assigned M2M items (the THRESHOLD), separate "Add" and "Remove" fields are shown instead of a single potentially massive multi-select — avoiding browser crashes/slowdowns.

---

TODO Criteria Verification

All three criteria from the PR description are satisfied:

✅ Limit "add" field to objects not already assigned
In both ProviderForm.__init__() and SiteForm.__init__(), the add_asns field filters out already-assigned items using NetBox's negation filter syntax:

• circuits/forms/model_forms.py:61 — add_query_param('provider_id__n', self.instance.pk)
• dcim/forms/model_forms.py:197 — add_query_param('site_id__n', self.instance.pk)

Symmetrically, remove_asns is restricted to only currently-assigned items via the positive filter (provider_id=pk / site_id=pk).

✅ Render a single selection field if <100 objects assigned
M2MAddRemoveFields.THRESHOLD = 100 controls the cutoff. New objects (no pk) and existing objects with fewer than 100 assignments always use the simple mode. This matches the spec.

✅ Display total number of objects currently assigned
The count is surfaced via help_text on the remove field:

• circuits/forms/model_forms.py:64 — _("{count} ASNs currently assigned").format(count=count)
• dcim/forms/model_forms.py:200 — same pattern

---

Code Quality Notes

Architecture is clean and extensible
The M2MAddRemoveFields class in rendering.py is well-documented with a clear contract. The template tag handler in form_helpers.py:83-96 correctly dispatches between simple and add/remove rendering modes.

_post_clean scope expansion is sound
The original code only iterated _meta.local_many_to_many. The new version also handles ManyToManyRel (reverse M2M relationships), which is needed because asns on Site/Provider appears to be a reverse relation. The condition guards (if name in self.cleaned_data / elif f'add_{name}' in ...) ensure fields not present in the form are skipped. No regression risk.

_save_m2m override is correct
Since asns was removed from Meta.fields in both forms, Django's default _save_m2m() would skip it. The override correctly saves M2M fields not listed in Meta.fields without double-saving those that are listed (if not meta_fields or field_name not in meta_fields).

Minor observations (non-blocking):

1. Duplicated __init__ logic: ProviderForm and SiteForm have near-identical __init__ implementations. A reusable mixin (e.g., M2MAddRemoveMixin) could eliminate duplication for future forms adopting this pattern — but given there are only two instances currently, this is fine to defer.

2. Count display placement: The count is in help_text on the remove field, which places it below the remove widget. This is functional but somewhat indirect — users looking for the total might expect it near the top. Low-priority UX consideration.

3. No tests included: Per CLAUDE.md, PRs must include tests for new functionality. There are no test changes in this PR. Coverage for _post_clean add/remove logic and the threshold switching behavior would be valuable, though this may be addressed in follow-up work.

---

Verdict

The PR correctly implements all three TODO criteria and the approach is architecturally sound. The one notable gap is the absence of tests for the new logic.