fix(proxy): stop team BYOK model name corruption on model edit

[yuneng-berri]

Relevant issues

Fixes #28382 (team BYOK model names mangled when editing the model in the UI). Consolidates the backend fix (#29001) and the Agent Builder UI fix (#27376, merged to this branch via #29729) so both land in staging together.

Linear ticket

LIT-2682

What's in this PR

Two commits, both already reviewed and merged into litellm_yj_june4:

1. Backend (Fix/team public model name overwrite #29001) -> team-scoped BYOK models store an internal routing key model_name_{team_id}_{uuid} and keep the public name in model_info.team_public_model_name. /v1 and /v2/model/info were leaking the internal key as model_name; the dashboard edit form bound to it and PATCHed it back on any non-rename save, and _get_public_model_name then treated the internal key as a new public name, overwriting team_public_model_name and rewriting the team models[] ACL with the mangled string. The fix adds a response-layer translation that swaps the internal key back to the public name on both /model/info endpoints (with the v1 loop rewritten so the non-mutating swap actually lands), plus a guard in _get_public_model_name that ignores an incoming internal-shaped or no-op model_name while still honoring genuine renames.

2. UI (fix(ui): key Agent Builder agent selection on model_info.id #29729) -> once two team-scoped agents can share a public name, the Agent Builder can no longer key selection on model_name. Selection, create, update and delete now resolve the agent by the stable model_info.id (falling back to model_name for config-defined agents), with a name-match fallback in the post-create path.

Pre-Submission checklist

• I have added meaningful tests
• My PR's scope is as isolated as possible; it only solves 1 specific problem

The backend half ships with Python tests (test_team_model_name_translation.py, additions to test_model_management_endpoints.py). The UI selection-by-id change is not yet covered by a vitest; a component test asserting two agents that share a public name stay individually selectable would be a good follow-up.

Screenshots / Proof of Fix

The backend fix was verified end-to-end against a live proxy with a real provider (xai/grok-4-1-fast-non-reasoning), toggling the two backend source files between the pre-fix and fixed state so the same harness shows before and after on the same fixtures.

Pre-fix (bug reproduced): a non-rename PATCH whose body echoes the internal model_name_{team_id}_{uuid} (exactly what the dashboard sends on a TPM-only save) overwrote team_public_model_name with the internal key and rewrote the team models[] ACL with it; /v1 and /v2/model/info both returned the internal key.

Post-fix (this branch): the same PATCH preserved team_public_model_name, left the team ACL intact, and still applied the edit (tpm=123); both /model/info endpoints returned the public name. Adjacent flows held in both states: genuine renames still work, routing by the public name via a team-scoped key returns 200, and non-team models are untouched.

UI smoke (suggested once deployed, for screenshots): create two team-scoped BYOK agents in different teams that share a public name, open http://localhost:4000/ui/?page=llm-playground -> Agent Builder, confirm both render as separate rows, then select / rename / delete one and confirm the action lands on the right agent.

Type

🐛 Bug Fix

[codecov]

Codecov Report

❌ Patch coverage is 96.00000% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...management_endpoints/model_management_endpoints.py	87.50%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

This PR fixes team BYOK model name corruption (issue #28382) through two coordinated changes: a backend write-path guard in _get_public_model_name that ignores internally-shaped routing keys echoed back by the dashboard, and a read-path translation in _translate_model_name_for_response that swaps the internal key back to the public name in both /v1/model/info and /v2/model/info responses. The UI half resolves agent selection by the stable model_info.id instead of model_name so two team-scoped agents with the same public name remain individually addressable.

• Backend write-path (model_management_endpoints.py): _get_public_model_name now rejects an incoming model_name that matches the internal model_name_{team_id}_{uuid} shape, and _update_existing_team_model_assignment always clears patch_data.model_name = None before returning to prevent any public name from overwriting the DB routing key.
• Backend read-path (proxy_server.py): _translate_model_name_for_response is applied as a presentation-only swap at both v1 and v2 endpoints; it returns a shallow copy and never mutates the router's in-memory routing table.
• UI (AgentBuilderView.tsx): Introduces getAgentSelectionKey (id-first, name fallback) across all create/update/delete/select code paths, with model_id-based resolution after a create to handle name collisions across teams.

Confidence Score: 4/5

Safe to merge; the write-path guard and always-clear of patch_data.model_name together prevent DB corruption in all observed client behaviors.

The fix is well-designed with multiple defence layers: _is_internal_shape blocks the stale client path, _update_existing_team_model_assignment always clears patch_data.model_name at the end so the DB routing key is never overwritten, and _translate_model_name_for_response is presentation-only and returns a shallow copy. The main gap is that no test explicitly exercises the primary post-fix client path (UI sends public name on a non-rename edit), so the correctness of the combined path requires tracing across two functions.

The interaction between _get_public_model_name and _update_existing_team_model_assignment for the post-fix UI client behavior (line 525-531 of model_management_endpoints.py) is not covered by any test and merits a second look.

Important Files Changed

Filename	Overview
litellm/proxy/management_endpoints/model_management_endpoints.py	Rewrites _get_public_model_name to filter out internal-shaped routing keys and no-op names, preventing team ACL corruption on non-rename edits. Logic is correct; a minor test gap exists for the post-fix UI behavior path.
litellm/proxy/proxy_server.py	Adds _translate_model_name_for_response (presentation-layer swap of internal routing key to public name) and applies it in both /v1/model/info (via _get_proxy_model_info) and /v2/model/info. Fixes the in-place loop bug in the v1 path; does not mutate the router's in-memory list.
tests/test_litellm/proxy/management_endpoints/test_model_management_endpoints.py	Adds 8 unit tests and one async integration test for _get_public_model_name and _update_team_model_in_db. Coverage is thorough for pre-fix client payloads; one test docstring (test_get_public_model_name_ignores_internal_shape_patch_public) is misleading since the function never reads patch_data.model_info.team_public_model_name.
tests/test_litellm/proxy/proxy_server/test_team_model_name_translation.py	New test file with unit and async integration tests for _translate_model_name_for_response and the v1/v2 read-path endpoints. Tests are well-isolated with no real network calls, covering mutation-safety, passthrough for non-team rows, and edge cases like missing model_info.
ui/litellm-dashboard/src/components/playground/chat_ui/AgentBuilderView.tsx	Introduces getAgentSelectionKey (prefers stable model_info.id, falls back to model_name) and updates all selection, create, update, and delete paths to use it. Post-create selection now resolves by model_id from the API response to handle shared public names across teams.

Comments Outside Diff (1)

1. tests/test_litellm/proxy/management_endpoints/test_model_management_endpoints.py, line 331-358 (link)

   Misleading test docstring — patch_data.model_info.team_public_model_name is never read

   _get_public_model_name explicitly never reads patch_data.model_info.team_public_model_name (the docstring says "We deliberately do NOT read patch_data.model_info.team_public_model_name"). The test therefore passes vacuously — the team_public_model_name="model_name_test-team_realuuid" in patch_data has no effect on the outcome. The test actually verifies fall-through when patch_data.model_name is None, not that an internal-shaped value in patch_data.model_info is filtered. A future reader could be misled into thinking the function inspects and sanitises that field.

   Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

_{Reviews (1): Last reviewed commit: "fix(ui): key Agent Builder agent selecti..." | Re-trigger Greptile}

[greptile-apps]

No-op guard misses the primary post-fix scenario

The condition incoming != db_model.model_name is designed to catch the case where the dashboard echoes back the internal routing key (before the read-path fix). After the read-path fix is deployed, the dashboard will instead send the public name (e.g. "team-claude-sonnet"), which is always != the internal key "model_name_{team_id}_{uuid}". The guard therefore never fires on the new client and falls through to return the incoming public name. This is still safe because _update_existing_team_model_assignment later detects old_public_name == public_model_name and skips the ACL update, then always sets patch_data.model_name = None at line 685.

However, there is no test that covers this scenario end-to-end (UI sends public name on a non-rename edit), so the correctness of the combined path relies solely on a reader tracing through two functions.