fix(proxy): expand all-proxy-models sentinel in direct access lookup

[mubashir1osmani]

Relevant issues

Fixes #22791

Linear ticket

Pre-Submission checklist

• I have added meaningful tests
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Screenshots / Proof of Fix

Repro: as an admin, create an internal user provisioned with "All Proxy Models" and no team, then list models as that user (the exact call the Models + Endpoints page makes).

# 1) provision the user with the all-proxy-models sentinel
curl -s -X POST http://localhost:4000/user/new \
  -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"user_role":"internal_user","models":["all-proxy-models"]}'
# -> user_id ..., key sk-..., models ["all-proxy-models"]

# 2) list models as that user (Models + Endpoints page request)
curl -s "http://localhost:4000/v2/model/info?include_team_models=true" \
  -H "Authorization: Bearer sk-USERKEY" | jq '.total_count'

Before this change the second call returns total_count = 0 (empty Models + Endpoints page) even though the same proxy has 37 models and the admin sees all 37. After the change the user sees the full set of non-team proxy models, matching what the Playground / model dropdown already showed (that path goes through get_key_models, which already expands the sentinel, which is why the two views disagreed).

verified locally:

Type

🐛 Bug Fix

Changes

A user provisioned with "All Proxy Models" stores the literal all-proxy-models sentinel in user.models. On the read path behind /v1/model/info and /v2/model/info, get_direct_access_models looped over user.models and looked each entry up as a real model_name via llm_router.get_model_list(model_name=...). The sentinel matches no deployment, so the lookup returned nothing, every model got direct_access=false, and with no team membership _filter_models_to_user_accessible dropped the entire list. The result was an empty Models + Endpoints page for exactly the users who were granted access to everything.

The fix expands the sentinel to all non-team deployment ids via get_model_ids(exclude_team_models=True), the same call the PROXY_ADMIN branch in the caller already uses. Because both /v1/model/info and /v2/model/info share _populate_team_access_on_models, this fixes both. An empty user.models still resolves to no direct access, matching get_key_models semantics.

Tests in tests/test_litellm/proxy/proxy_server/test_team_model_name_translation.py cover the sentinel expansion in get_direct_access_models, the unchanged explicit-model path, and the full _populate_team_access_on_models -> _filter_models_to_user_accessible chain for an internal user with no teams; the chain test fails on the pre-fix code and passes after.

[greptile-apps]

Greptile Summary

Fixes a bug where internal users provisioned with the all-proxy-models sentinel in user.models saw an empty Models + Endpoints page. The sentinel was passed to llm_router.get_model_list(model_name=...), which matched no real deployment, so every model got direct_access=False and was then dropped by the filter.

• get_direct_access_models now short-circuits on the sentinel and delegates to llm_router.get_model_ids(exclude_team_models=True), the same call already used by the PROXY_ADMIN branch in _populate_team_access_on_models.
• The non-sentinel path is refactored to an equivalent walrus-operator list comprehension, and three focused mock tests cover the sentinel, explicit-model, and full access-chain scenarios.

Confidence Score: 5/5

Safe to merge — the change is tightly scoped to one helper function and both affected endpoints, the fix reuses the exact same router call already trusted in the admin branch, and a model_validator in LiteLLM_UserTable guarantees models is never None.

The logic change is minimal and correct: the sentinel check is guarded by the existing Pydantic validator that coerces a DB NULL to [], so the in test is always safe. The non-sentinel list comprehension is semantically identical to the original loop. Three new mock-only unit tests directly target the regression path and would fail against the pre-fix code, and the PR description includes a concrete reproduction script. No existing tests were weakened.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/proxy/proxy_server.py	Expands the all-proxy-models sentinel in get_direct_access_models by returning llm_router.get_model_ids(exclude_team_models=True) when the sentinel is detected, matching the PROXY_ADMIN branch behavior; also refactors the non-sentinel path to an equivalent list comprehension.
tests/test_litellm/proxy/proxy_server/test_team_model_name_translation.py	Adds three new mock-only unit tests: sentinel expansion, explicit model name path, and a full integration-style chain through _populate_team_access_on_models → _filter_models_to_user_accessible for an internal user with no teams; no real network calls.

_{Reviews (1): Last reviewed commit: "fix(proxy): expand all-proxy-models sent..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[mateo-berri]

LGTM; thanks!

[mateo-berri]

wow this is way better. nice