From 99636dff2b661a965689c998587c481d137c18c8 Mon Sep 17 00:00:00 2001
From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Mon, 28 Sep 2026 23:50:24 +0000
Subject: [PATCH 1/6] fix(proxy): expand partial wildcard prefixes against
 provider-stripped model ids

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 .../health_check_helpers.py                   |  23 ++-
 litellm/proxy/auth/model_checks.py            |  51 +++---
 .../test_wildcard_partial_prefix_expansion.py | 149 ++++++++++++++++++
 .../proxy/auth/test_model_checks.py           |  64 +++++---
 .../test_health_check_helpers.py              |  18 +++
 5 files changed, 260 insertions(+), 45 deletions(-)
 create mode 100644 tests/e2e/test_wildcard_partial_prefix_expansion.py

diff --git a/litellm/litellm_core_utils/health_check_helpers.py b/litellm/litellm_core_utils/health_check_helpers.py
index a0f027cd58f..16342728774 100644
--- a/litellm/litellm_core_utils/health_check_helpers.py
+++ b/litellm/litellm_core_utils/health_check_helpers.py
@@ -3,7 +3,7 @@ Helper functions for health check calls.
 """
 
 import base64
-from collections.abc import Awaitable, Callable
+from collections.abc import Awaitable, Callable, Sequence
 from typing import TYPE_CHECKING, Final, Literal
 
 from litellm.llms.base_llm.ocr.transformation import DocumentType
@@ -39,6 +39,19 @@ def _ocr_health_check_document(model: str, custom_llm_provider: str) -> Document
     return native(model, custom_llm_provider)
 
 
+def _wildcard_health_check_models(wildcard_model: str, cheapest_models: Sequence[str]) -> tuple[str, ...]:
+    """Substitute each candidate id into the wildcard's dynamic segment so the health
+    check probes a name the wildcard route actually expands, instead of the bare
+    cost-map id (e.g. `databricks/system.ai.*` + `databricks/databricks-x` ->
+    `databricks/system.ai.databricks-x`)."""
+    provider_prefix, sep, wildcard_suffix = wildcard_model.partition("/")
+    if not sep or wildcard_suffix == "*":
+        return tuple(cheapest_models)
+    return tuple(
+        f"{provider_prefix}/{wildcard_suffix.replace('*', model.partition('/')[-1], 1)}" for model in cheapest_models
+    )
+
+
 class HealthCheckHelpers:
     @staticmethod
     async def ahealth_check_wildcard_models(
@@ -58,11 +71,9 @@ class HealthCheckHelpers:
             raise Exception(
                 f"Unable to health check wildcard model for provider {custom_llm_provider}. Add a model on your config.yaml or contribute here - https://github.com/BerriAI/litellm/blob/main/model_prices_and_context_window.json"
             )
-        if len(cheapest_models) > 1:
-            fallback_models = cheapest_models[1:]  # Pick the last 2 models from the shuffled list
-        else:
-            fallback_models = None
-        model_params["model"] = cheapest_models[0]
+        candidates: Final = _wildcard_health_check_models(wildcard_model=model, cheapest_models=cheapest_models)
+        fallback_models: Final = list(candidates[1:]) or None
+        model_params["model"] = candidates[0]
         model_params["litellm_logging_obj"] = litellm_logging_obj
         model_params["fallbacks"] = fallback_models
         model_params["max_tokens"] = model_params.get("max_tokens", 16)  # GPT-5 models require max_output_tokens >= 16
diff --git a/litellm/proxy/auth/model_checks.py b/litellm/proxy/auth/model_checks.py
index de2ca4762f1..944ed80280a 100644
--- a/litellm/proxy/auth/model_checks.py
+++ b/litellm/proxy/auth/model_checks.py
@@ -262,6 +262,24 @@ def _hydrate_litellm_credential_name(
     return litellm_params
 
 
+def _strip_known_provider_prefix(model: str, known_providers: frozenset[str]) -> str:
+    leading, sep, model_suffix = model.partition("/")
+    return model_suffix if sep and leading in known_providers else model
+
+
+def _apply_partial_wildcard_prefix(
+    provider_models: list[str], model_prefix: str, known_providers: frozenset[str]
+) -> list[str]:
+    prefix_filtered: Final = [
+        model
+        for model in provider_models
+        if _strip_known_provider_prefix(model, known_providers).startswith(model_prefix)
+    ]
+    if prefix_filtered:
+        return prefix_filtered
+    return [f"{model_prefix}{_strip_known_provider_prefix(model, known_providers)}" for model in provider_models]
+
+
 def get_known_models_from_wildcard(wildcard_model: str, litellm_params: LiteLLM_Params | None = None) -> list[str]:
     wildcard_model_to_expand: Final = (
         litellm_params.model
@@ -288,23 +306,24 @@ def get_known_models_from_wildcard(wildcard_model: str, litellm_params: LiteLLM_
 
     litellm_params = _hydrate_litellm_credential_name(litellm_params)
 
-    wildcard_models = get_provider_models(provider=provider, litellm_params=litellm_params)
+    provider_models: Final = get_provider_models(provider=provider, litellm_params=litellm_params)
 
-    if wildcard_models is None:
+    if provider_models is None:
         return []
-    if wildcard_suffix != "*":
-        ## CHECK IF PARTIAL FILTER e.g. `gemini-*`
-        model_prefix: Final = wildcard_suffix.replace("*", "")
 
-        is_partial_filter: Final = any(wc_model.startswith(model_prefix) for wc_model in wildcard_models)
-        if is_partial_filter:
-            filtered_wildcard_models = [wc_model for wc_model in wildcard_models if wc_model.startswith(model_prefix)]
-            wildcard_models = filtered_wildcard_models
-        else:
-            # add model prefix to wildcard models
-            wildcard_models = [f"{model_prefix}{model}" for model in wildcard_models]
+    known_providers: Final = frozenset(provider.value for provider in LlmProviders)
+
+    ## CHECK IF PARTIAL FILTER e.g. `gemini-*`
+    wildcard_models: Final = (
+        provider_models
+        if wildcard_suffix == "*"
+        else _apply_partial_wildcard_prefix(
+            provider_models=provider_models,
+            model_prefix=wildcard_suffix.replace("*", ""),
+            known_providers=known_providers,
+        )
+    )
 
-    known_providers: Final = {provider.value for provider in LlmProviders}
     suffix_appended_wildcard_models: Final = []
     for model in wildcard_models:
         if not model.startswith(wildcard_provider_prefix):
@@ -314,11 +333,7 @@ def get_known_models_from_wildcard(wildcard_model: str, litellm_params: LiteLLM_
             # both, which would otherwise yield an uncallable "ollama_server1/ollama/gemma3:1b".
             # Only strip the leading segment when it is a known provider, so ids whose first
             # segment is an org rather than a provider (e.g. "meta-llama/Llama-3-8B") keep it.
-            leading, sep, model_suffix = model.partition("/")
-            if sep and leading in known_providers:
-                model = f"{wildcard_provider_prefix}/{model_suffix}"
-            else:
-                model = f"{wildcard_provider_prefix}/{model}"
+            model = f"{wildcard_provider_prefix}/{_strip_known_provider_prefix(model, known_providers)}"
         suffix_appended_wildcard_models.append(model)
     return suffix_appended_wildcard_models or []
 
diff --git a/tests/e2e/test_wildcard_partial_prefix_expansion.py b/tests/e2e/test_wildcard_partial_prefix_expansion.py
new file mode 100644
index 00000000000..48b265c2c62
--- /dev/null
+++ b/tests/e2e/test_wildcard_partial_prefix_expansion.py
@@ -0,0 +1,149 @@
+"""Regression: wildcard expansion of a partial provider prefix like
+`databricks/system.ai.*` must produce `databricks/system.ai.<endpoint>` ids,
+not `databricks/system.ai.databricks/databricks-<name>`.
+
+On main, get_known_models_from_wildcard prepends the literal prefix before the
+`*` ("system.ai.") to cost-map keys that already carry the `databricks/`
+provider prefix, so /v1/models and /model/info list names that Databricks
+rejects with "Invalid Unity Catalog name". See LIT-8910.
+
+The test boots its own proxy (the shared stack's model_list is fixed), with two
+wildcard deployments: `databricks/system.ai.*` under test and `databricks/*` as
+a control whose expansion must stay `databricks/databricks-*`.
+"""
+
+import socket
+import subprocess
+import sys
+import time
+from collections.abc import Generator
+from pathlib import Path
+from typing import Final, cast
+
+import pytest
+from e2e_http import URL, AuthHeaders, NoBody, get, probe, unwrap
+from models import ModelsListParams, ModelsListResponse
+from pydantic import BaseModel, ConfigDict
+
+REPO_ROOT: Final = Path(__file__).resolve().parents[2]
+
+MASTER_KEY: Final = "sk-1234"
+AUTH: Final = AuthHeaders(authorization=f"Bearer {MASTER_KEY}")
+
+CONFIG_YAML: Final = """\
+model_list:
+  - model_name: databricks/system.ai.*
+    litellm_params:
+      model: databricks/system.ai.*
+      api_key: fake-databricks-token
+      api_base: https://example.invalid
+  - model_name: databricks/*
+    litellm_params:
+      model: databricks/*
+      api_key: fake-databricks-token
+      api_base: https://example.invalid
+general_settings:
+  master_key: sk-1234
+"""
+
+
+class _ModelInfoParams(BaseModel):
+    model_config = ConfigDict(protected_namespaces=(), extra="ignore")
+    model: str = ""
+
+
+class _ModelInfoRow(BaseModel):
+    model_name: str
+    litellm_params: _ModelInfoParams = _ModelInfoParams()
+
+
+class _ModelInfoResponse(BaseModel):
+    data: list[_ModelInfoRow] = []
+
+
+def _free_port() -> int:
+    with socket.socket() as sock:
+        sock.bind(("127.0.0.1", 0))
+        return cast("tuple[str, int]", sock.getsockname())[1]
+
+
+@pytest.fixture(scope="module")
+def wildcard_proxy(tmp_path_factory: pytest.TempPathFactory) -> Generator[str]:
+    config_path = tmp_path_factory.mktemp("wildcard") / "config.yaml"
+    config_path.write_text(CONFIG_YAML)
+    port = _free_port()
+    proc = subprocess.Popen(
+        [
+            sys.executable,
+            str(REPO_ROOT / "litellm/proxy/proxy_cli.py"),
+            "--config",
+            str(config_path),
+            "--port",
+            str(port),
+        ],
+        cwd=REPO_ROOT,
+        env={
+            "PATH": "/usr/bin:/bin:/usr/local/bin",
+            "HOME": str(Path.home()),
+            "PYTHONPATH": str(REPO_ROOT),
+            "LITELLM_LOCAL_MODEL_COST_MAP": "true",
+            "LITELLM_DANGEROUSLY_PERMIT_WEAK_OR_UNSET_MASTER_KEY": "true",
+            "LITELLM_MASTER_KEY": MASTER_KEY,
+        },
+    )
+    base = f"http://127.0.0.1:{port}"
+    try:
+        deadline = time.time() + 180
+        while True:
+            alive = probe(URL(f"{base}/health/liveliness"), headers=NoBody(), params=NoBody(), timeout=5.0)
+            if alive.healthy:
+                break
+            if proc.poll() is not None:
+                pytest.fail(f"proxy exited during boot with {proc.returncode}")
+            assert time.time() < deadline, "proxy did not answer /health/liveliness within 180s"
+            time.sleep(2)
+        yield base
+    finally:
+        proc.terminate()
+        proc.wait(timeout=30)
+
+
+def _list_model_ids(base: str) -> list[str]:
+    listed = get(
+        URL(f"{base}/v1/models"),
+        headers=AUTH,
+        params=ModelsListParams(return_wildcard_routes=False),
+        response_type=ModelsListResponse,
+    )
+    return [entry.id for entry in unwrap(listed).data]
+
+
+def test_partial_prefix_wildcard_expands_to_unity_catalog_names(wildcard_proxy: str) -> None:
+    ids = _list_model_ids(wildcard_proxy)
+
+    system_ai_ids = [model_id for model_id in ids if model_id.startswith("databricks/system.ai.")]
+    assert system_ai_ids, f"no databricks/system.ai.* expansion in {ids}"
+
+    malformed = [model_id for model_id in system_ai_ids if "/" in model_id.removeprefix("databricks/system.ai.")]
+    assert not malformed, f"expanded ids embed the provider-prefixed cost-map key after 'system.ai.': {malformed}"
+
+
+def test_provider_wildcard_still_expands_to_cost_map_names(wildcard_proxy: str) -> None:
+    ids = _list_model_ids(wildcard_proxy)
+
+    databricks_ids = [model_id for model_id in ids if model_id.startswith("databricks/databricks-")]
+    assert databricks_ids, f"databricks/* did not expand to cost-map names: {ids}"
+
+
+def test_model_info_keeps_provider_model_for_expanded_deployments(wildcard_proxy: str) -> None:
+    info = unwrap(
+        get(
+            URL(f"{wildcard_proxy}/model/info"),
+            headers=AUTH,
+            params=NoBody(),
+            response_type=_ModelInfoResponse,
+        )
+    )
+
+    bad = [row.model_name for row in info.data if "system.ai.databricks/" in row.litellm_params.model]
+    assert not bad, f"litellm_params.model carries the corrupted expanded name: {bad}"
diff --git a/tests/test_litellm/proxy/auth/test_model_checks.py b/tests/test_litellm/proxy/auth/test_model_checks.py
index 13171a42cda..9204a69a89e 100644
--- a/tests/test_litellm/proxy/auth/test_model_checks.py
+++ b/tests/test_litellm/proxy/auth/test_model_checks.py
@@ -1,3 +1,4 @@
+from typing import Final
 from unittest.mock import patch
 
 import pytest
@@ -11,9 +12,7 @@ def test_get_team_models_for_all_models_and_team_only_models():
     model_access_groups = {}
     include_model_access_groups = False
 
-    result = get_team_models(
-        team_models, proxy_model_list, model_access_groups, include_model_access_groups
-    )
+    result = get_team_models(team_models, proxy_model_list, model_access_groups, include_model_access_groups)
     combined_models = team_models + proxy_model_list
     assert set(result) == set(combined_models)
 
@@ -246,9 +245,7 @@ def test_get_key_models_does_not_mutate_input():
         ),
     ],
 )
-def test_get_complete_model_list_order(
-    key_models, team_models, proxy_model_list, model_list, expected
-):
+def test_get_complete_model_list_order(key_models, team_models, proxy_model_list, model_list, expected):
     """
     Test that get_complete_model_list preserves order
     """
@@ -401,9 +398,7 @@ def test_wildcard_credential_hydration_preserves_deployment_params(
         captured_params["api_key"] = litellm_params.api_key
         captured_params["api_version"] = litellm_params.api_version
         captured_params["credential_name"] = litellm_params.litellm_credential_name
-        captured_params["has_unexpected_field"] = hasattr(
-            litellm_params, "unexpected_field"
-        )
+        captured_params["has_unexpected_field"] = hasattr(litellm_params, "unexpected_field")
         return ["gpt-4o"]
 
     monkeypatch.setattr(model_checks, "get_provider_models", fake_get_provider_models)
@@ -448,9 +443,7 @@ def test_wildcard_custom_prefix_does_not_stack_provider_prefix(monkeypatch):
 
     result = get_known_models_from_wildcard(
         wildcard_model="ollama_server1/*",
-        litellm_params=LiteLLM_Params(
-            model="ollama_chat/*", custom_llm_provider="ollama_chat"
-        ),
+        litellm_params=LiteLLM_Params(model="ollama_chat/*", custom_llm_provider="ollama_chat"),
     )
 
     assert result == ["ollama_server1/gemma3:1b", "ollama_server1/llama3:8b"]
@@ -477,14 +470,47 @@ def test_wildcard_custom_prefix_keeps_org_segment_for_non_provider_first_segment
 
     result = get_known_models_from_wildcard(
         wildcard_model="my_hf/*",
-        litellm_params=LiteLLM_Params(
-            model="huggingface/*", custom_llm_provider="huggingface"
-        ),
+        litellm_params=LiteLLM_Params(model="huggingface/*", custom_llm_provider="huggingface"),
     )
 
     assert result == ["my_hf/meta-llama/Llama-3-8B"]
 
 
+def test_partial_prefix_wildcard_expands_against_stripped_provider_ids(monkeypatch):
+    """A partial wildcard like ``databricks/system.ai.*`` must splice the prefix before
+    ``*`` onto the provider-stripped model id, not onto the whole cost-map key. Prepending
+    to the provider-prefixed key yielded ``databricks/system.ai.databricks/databricks-x``,
+    an id no upstream accepts.
+    """
+    import litellm
+    from litellm.proxy.auth import model_checks
+    from litellm.proxy.auth.model_checks import get_known_models_from_wildcard
+    from litellm.types.router import LiteLLM_Params
+
+    provider_ids: Final = list(litellm.models_by_provider["databricks"])
+    monkeypatch.setattr(
+        model_checks,
+        "get_provider_models",
+        lambda provider, litellm_params=None: provider_ids,
+    )
+
+    result: Final = get_known_models_from_wildcard(
+        wildcard_model="databricks/system.ai.*",
+        litellm_params=LiteLLM_Params(
+            model="databricks/system.ai.*",
+            api_key="x",
+            api_base="https://example.invalid",
+        ),
+    )
+
+    expected: Final = [f"databricks/system.ai.{model_id.partition('/')[-1]}" for model_id in provider_ids]
+    assert result == expected
+    assert all(
+        model_id.startswith("databricks/system.ai.") and "/" not in model_id.removeprefix("databricks/system.ai.")
+        for model_id in result
+    )
+
+
 def test_wildcard_credential_hydration_preserves_missing_credential_name(
     monkeypatch,
 ):
@@ -927,9 +953,7 @@ def test_add_known_models_refreshes_models_by_provider_for_wildcard_expansion():
     assert fake_model not in litellm.models_by_provider["vertex_ai"]
     try:
         litellm.add_known_models(
-            model_cost_map={
-                fake_model: {"litellm_provider": "vertex_ai-language-models", "mode": "chat"}
-            }
+            model_cost_map={fake_model: {"litellm_provider": "vertex_ai-language-models", "mode": "chat"}}
         )
         assert fake_model in litellm.models_by_provider["vertex_ai"]
         assert litellm.models_by_provider is captured_reference
@@ -977,6 +1001,4 @@ def test_transcribe_is_a_known_provider_for_wildcard_expansion():
     assert "transcribe" in litellm.models_by_provider
     assert "transcribe/StartTranscriptionJob" in litellm.models_by_provider["transcribe"]
     assert get_provider_models("transcribe") == ["transcribe/StartTranscriptionJob"]
-    assert get_known_models_from_wildcard("transcribe/*") == [
-        "transcribe/StartTranscriptionJob"
-    ]
+    assert get_known_models_from_wildcard("transcribe/*") == ["transcribe/StartTranscriptionJob"]
diff --git a/tests/unit/litellm_core_utils/test_health_check_helpers.py b/tests/unit/litellm_core_utils/test_health_check_helpers.py
index 47c4576f91f..4925bcb0b5d 100644
--- a/tests/unit/litellm_core_utils/test_health_check_helpers.py
+++ b/tests/unit/litellm_core_utils/test_health_check_helpers.py
@@ -14,6 +14,7 @@ from litellm.constants import LITTELM_INTERNAL_HEALTH_SERVICE_ACCOUNT_NAME
 from litellm.litellm_core_utils.health_check_helpers import (
     IMAGE_EDIT_HEALTH_CHECK_PROMPT,
     HealthCheckHelpers,
+    _wildcard_health_check_models,
 )
 from litellm.main import ahealth_check
 from litellm.proxy._types import UserAPIKeyAuth
@@ -548,3 +549,20 @@ def test_ocr_health_check_document_raises_without_the_extension():
             _ocr_health_check_document(model="mistral/mistral-ocr-latest", custom_llm_provider="mistral")
     finally:
         NATIVE_OCR_HEALTH_CHECK_DOCUMENT.reset()
+
+
+def test_wildcard_health_check_models_partial_prefix_substitutes_stripped_id():
+    candidates: Final = _wildcard_health_check_models(
+        "databricks/system.ai.*",
+        ["databricks/databricks-gemini-3-1-flash-image", "databricks/databricks-gpt-5"],
+    )
+    assert candidates == (
+        "databricks/system.ai.databricks-gemini-3-1-flash-image",
+        "databricks/system.ai.databricks-gpt-5",
+    )
+
+
+def test_wildcard_health_check_models_bare_provider_and_star_wildcard_unchanged():
+    candidates: Final = ("databricks/databricks-gpt-5",)
+    assert _wildcard_health_check_models("databricks/*", candidates) == candidates
+    assert _wildcard_health_check_models("*", candidates) == candidates

From 9398790ba23ff03a816d40f7bb959720ecc811ca Mon Sep 17 00:00:00 2001
From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Mon, 28 Sep 2026 23:54:32 +0000
Subject: [PATCH 2/6] fix(health): splice wildcard prefix onto
 provider-stripped health check candidates

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 .../health_check_helpers.py                   | 19 +++++++----
 .../proxy/auth/test_model_checks.py           | 33 ++++++++++++-------
 .../test_health_check_helpers.py              |  8 +++++
 3 files changed, 42 insertions(+), 18 deletions(-)

diff --git a/litellm/litellm_core_utils/health_check_helpers.py b/litellm/litellm_core_utils/health_check_helpers.py
index 16342728774..7ea9f9496d9 100644
--- a/litellm/litellm_core_utils/health_check_helpers.py
+++ b/litellm/litellm_core_utils/health_check_helpers.py
@@ -9,7 +9,7 @@ from typing import TYPE_CHECKING, Final, Literal
 from litellm.llms.base_llm.ocr.transformation import DocumentType
 from litellm.rust_bridge import runtime
 from litellm.rust_bridge.ocr.entrypoints import NATIVE_OCR_HEALTH_CHECK_DOCUMENT
-from litellm.types.utils import LIST_BATCHES_SUPPORTED_PROVIDERS
+from litellm.types.utils import LIST_BATCHES_SUPPORTED_PROVIDERS, LlmProviders
 
 if TYPE_CHECKING:
     from litellm.litellm_core_utils.litellm_logging import Logging
@@ -39,16 +39,23 @@ def _ocr_health_check_document(model: str, custom_llm_provider: str) -> Document
     return native(model, custom_llm_provider)
 
 
+def _strip_known_provider_prefix(model: str, known_providers: frozenset[str]) -> str:
+    leading, sep, model_suffix = model.partition("/")
+    return model_suffix if sep and leading in known_providers else model
+
+
 def _wildcard_health_check_models(wildcard_model: str, cheapest_models: Sequence[str]) -> tuple[str, ...]:
-    """Substitute each candidate id into the wildcard's dynamic segment so the health
-    check probes a name the wildcard route actually expands, instead of the bare
-    cost-map id (e.g. `databricks/system.ai.*` + `databricks/databricks-x` ->
-    `databricks/system.ai.databricks-x`)."""
     provider_prefix, sep, wildcard_suffix = wildcard_model.partition("/")
     if not sep or wildcard_suffix == "*":
         return tuple(cheapest_models)
+    known_providers: Final = frozenset(provider.value for provider in LlmProviders)
+    literal_prefix: Final = wildcard_suffix.replace("*", "")
+    stripped_ids: Final = tuple(_strip_known_provider_prefix(model, known_providers) for model in cheapest_models)
     return tuple(
-        f"{provider_prefix}/{wildcard_suffix.replace('*', model.partition('/')[-1], 1)}" for model in cheapest_models
+        f"{provider_prefix}/{stripped}"
+        if stripped.startswith(literal_prefix)
+        else f"{provider_prefix}/{wildcard_suffix.replace('*', stripped, 1)}"
+        for stripped in stripped_ids
     )
 
 
diff --git a/tests/test_litellm/proxy/auth/test_model_checks.py b/tests/test_litellm/proxy/auth/test_model_checks.py
index 9204a69a89e..866a8f42bdf 100644
--- a/tests/test_litellm/proxy/auth/test_model_checks.py
+++ b/tests/test_litellm/proxy/auth/test_model_checks.py
@@ -12,7 +12,9 @@ def test_get_team_models_for_all_models_and_team_only_models():
     model_access_groups = {}
     include_model_access_groups = False
 
-    result = get_team_models(team_models, proxy_model_list, model_access_groups, include_model_access_groups)
+    result = get_team_models(
+        team_models, proxy_model_list, model_access_groups, include_model_access_groups
+    )
     combined_models = team_models + proxy_model_list
     assert set(result) == set(combined_models)
 
@@ -245,7 +247,9 @@ def test_get_key_models_does_not_mutate_input():
         ),
     ],
 )
-def test_get_complete_model_list_order(key_models, team_models, proxy_model_list, model_list, expected):
+def test_get_complete_model_list_order(
+    key_models, team_models, proxy_model_list, model_list, expected
+):
     """
     Test that get_complete_model_list preserves order
     """
@@ -398,7 +402,9 @@ def test_wildcard_credential_hydration_preserves_deployment_params(
         captured_params["api_key"] = litellm_params.api_key
         captured_params["api_version"] = litellm_params.api_version
         captured_params["credential_name"] = litellm_params.litellm_credential_name
-        captured_params["has_unexpected_field"] = hasattr(litellm_params, "unexpected_field")
+        captured_params["has_unexpected_field"] = hasattr(
+            litellm_params, "unexpected_field"
+        )
         return ["gpt-4o"]
 
     monkeypatch.setattr(model_checks, "get_provider_models", fake_get_provider_models)
@@ -443,7 +449,9 @@ def test_wildcard_custom_prefix_does_not_stack_provider_prefix(monkeypatch):
 
     result = get_known_models_from_wildcard(
         wildcard_model="ollama_server1/*",
-        litellm_params=LiteLLM_Params(model="ollama_chat/*", custom_llm_provider="ollama_chat"),
+        litellm_params=LiteLLM_Params(
+            model="ollama_chat/*", custom_llm_provider="ollama_chat"
+        ),
     )
 
     assert result == ["ollama_server1/gemma3:1b", "ollama_server1/llama3:8b"]
@@ -470,18 +478,15 @@ def test_wildcard_custom_prefix_keeps_org_segment_for_non_provider_first_segment
 
     result = get_known_models_from_wildcard(
         wildcard_model="my_hf/*",
-        litellm_params=LiteLLM_Params(model="huggingface/*", custom_llm_provider="huggingface"),
+        litellm_params=LiteLLM_Params(
+            model="huggingface/*", custom_llm_provider="huggingface"
+        ),
     )
 
     assert result == ["my_hf/meta-llama/Llama-3-8B"]
 
 
 def test_partial_prefix_wildcard_expands_against_stripped_provider_ids(monkeypatch):
-    """A partial wildcard like ``databricks/system.ai.*`` must splice the prefix before
-    ``*`` onto the provider-stripped model id, not onto the whole cost-map key. Prepending
-    to the provider-prefixed key yielded ``databricks/system.ai.databricks/databricks-x``,
-    an id no upstream accepts.
-    """
     import litellm
     from litellm.proxy.auth import model_checks
     from litellm.proxy.auth.model_checks import get_known_models_from_wildcard
@@ -953,7 +958,9 @@ def test_add_known_models_refreshes_models_by_provider_for_wildcard_expansion():
     assert fake_model not in litellm.models_by_provider["vertex_ai"]
     try:
         litellm.add_known_models(
-            model_cost_map={fake_model: {"litellm_provider": "vertex_ai-language-models", "mode": "chat"}}
+            model_cost_map={
+                fake_model: {"litellm_provider": "vertex_ai-language-models", "mode": "chat"}
+            }
         )
         assert fake_model in litellm.models_by_provider["vertex_ai"]
         assert litellm.models_by_provider is captured_reference
@@ -1001,4 +1008,6 @@ def test_transcribe_is_a_known_provider_for_wildcard_expansion():
     assert "transcribe" in litellm.models_by_provider
     assert "transcribe/StartTranscriptionJob" in litellm.models_by_provider["transcribe"]
     assert get_provider_models("transcribe") == ["transcribe/StartTranscriptionJob"]
-    assert get_known_models_from_wildcard("transcribe/*") == ["transcribe/StartTranscriptionJob"]
+    assert get_known_models_from_wildcard("transcribe/*") == [
+        "transcribe/StartTranscriptionJob"
+    ]
diff --git a/tests/unit/litellm_core_utils/test_health_check_helpers.py b/tests/unit/litellm_core_utils/test_health_check_helpers.py
index 4925bcb0b5d..2e010bdf293 100644
--- a/tests/unit/litellm_core_utils/test_health_check_helpers.py
+++ b/tests/unit/litellm_core_utils/test_health_check_helpers.py
@@ -566,3 +566,11 @@ def test_wildcard_health_check_models_bare_provider_and_star_wildcard_unchanged(
     candidates: Final = ("databricks/databricks-gpt-5",)
     assert _wildcard_health_check_models("databricks/*", candidates) == candidates
     assert _wildcard_health_check_models("*", candidates) == candidates
+
+
+def test_wildcard_health_check_models_partial_prefix_matching_literal_keeps_stripped_id():
+    assert _wildcard_health_check_models("openai/gpt-4*", ["gpt-4o-mini"]) == ("openai/gpt-4o-mini",)
+
+
+def test_wildcard_health_check_models_partial_prefix_splices_suffix_around_star():
+    assert _wildcard_health_check_models("openai/ft:*", ["gpt-4o-mini"]) == ("openai/ft:gpt-4o-mini",)

From b24ef8bde68cb1b4e56c7a9f8a58f1f208173b27 Mon Sep 17 00:00:00 2001
From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Tue, 29 Sep 2026 00:02:22 +0000
Subject: [PATCH 3/6] fix(health): use provider-stripped wildcard suffix at the
 ahealth_check call site

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 .../health_check_helpers.py                   | 15 ++++----
 .../test_health_check_helpers.py              | 34 +++++++++++++++----
 2 files changed, 37 insertions(+), 12 deletions(-)

diff --git a/litellm/litellm_core_utils/health_check_helpers.py b/litellm/litellm_core_utils/health_check_helpers.py
index 7ea9f9496d9..30e9b020245 100644
--- a/litellm/litellm_core_utils/health_check_helpers.py
+++ b/litellm/litellm_core_utils/health_check_helpers.py
@@ -44,17 +44,18 @@ def _strip_known_provider_prefix(model: str, known_providers: frozenset[str]) ->
     return model_suffix if sep and leading in known_providers else model
 
 
-def _wildcard_health_check_models(wildcard_model: str, cheapest_models: Sequence[str]) -> tuple[str, ...]:
-    provider_prefix, sep, wildcard_suffix = wildcard_model.partition("/")
-    if not sep or wildcard_suffix == "*":
+def _wildcard_health_check_models(
+    wildcard_suffix: str, custom_llm_provider: str, cheapest_models: Sequence[str]
+) -> tuple[str, ...]:
+    if wildcard_suffix == "*":
         return tuple(cheapest_models)
     known_providers: Final = frozenset(provider.value for provider in LlmProviders)
     literal_prefix: Final = wildcard_suffix.replace("*", "")
     stripped_ids: Final = tuple(_strip_known_provider_prefix(model, known_providers) for model in cheapest_models)
     return tuple(
-        f"{provider_prefix}/{stripped}"
+        f"{custom_llm_provider}/{stripped}"
         if stripped.startswith(literal_prefix)
-        else f"{provider_prefix}/{wildcard_suffix.replace('*', stripped, 1)}"
+        else f"{custom_llm_provider}/{wildcard_suffix.replace('*', stripped, 1)}"
         for stripped in stripped_ids
     )
 
@@ -78,7 +79,9 @@ class HealthCheckHelpers:
             raise Exception(
                 f"Unable to health check wildcard model for provider {custom_llm_provider}. Add a model on your config.yaml or contribute here - https://github.com/BerriAI/litellm/blob/main/model_prices_and_context_window.json"
             )
-        candidates: Final = _wildcard_health_check_models(wildcard_model=model, cheapest_models=cheapest_models)
+        candidates: Final = _wildcard_health_check_models(
+            wildcard_suffix=model, custom_llm_provider=custom_llm_provider, cheapest_models=cheapest_models
+        )
         fallback_models: Final = list(candidates[1:]) or None
         model_params["model"] = candidates[0]
         model_params["litellm_logging_obj"] = litellm_logging_obj
diff --git a/tests/unit/litellm_core_utils/test_health_check_helpers.py b/tests/unit/litellm_core_utils/test_health_check_helpers.py
index 2e010bdf293..1087d8be3ff 100644
--- a/tests/unit/litellm_core_utils/test_health_check_helpers.py
+++ b/tests/unit/litellm_core_utils/test_health_check_helpers.py
@@ -553,7 +553,8 @@ def test_ocr_health_check_document_raises_without_the_extension():
 
 def test_wildcard_health_check_models_partial_prefix_substitutes_stripped_id():
     candidates: Final = _wildcard_health_check_models(
-        "databricks/system.ai.*",
+        "system.ai.*",
+        "databricks",
         ["databricks/databricks-gemini-3-1-flash-image", "databricks/databricks-gpt-5"],
     )
     assert candidates == (
@@ -562,15 +563,36 @@ def test_wildcard_health_check_models_partial_prefix_substitutes_stripped_id():
     )
 
 
-def test_wildcard_health_check_models_bare_provider_and_star_wildcard_unchanged():
+def test_wildcard_health_check_models_star_wildcard_unchanged():
     candidates: Final = ("databricks/databricks-gpt-5",)
-    assert _wildcard_health_check_models("databricks/*", candidates) == candidates
-    assert _wildcard_health_check_models("*", candidates) == candidates
+    assert _wildcard_health_check_models("*", "databricks", candidates) == candidates
 
 
 def test_wildcard_health_check_models_partial_prefix_matching_literal_keeps_stripped_id():
-    assert _wildcard_health_check_models("openai/gpt-4*", ["gpt-4o-mini"]) == ("openai/gpt-4o-mini",)
+    assert _wildcard_health_check_models("gpt-4*", "openai", ["gpt-4o-mini"]) == ("openai/gpt-4o-mini",)
 
 
 def test_wildcard_health_check_models_partial_prefix_splices_suffix_around_star():
-    assert _wildcard_health_check_models("openai/ft:*", ["gpt-4o-mini"]) == ("openai/ft:gpt-4o-mini",)
+    assert _wildcard_health_check_models("ft:*", "openai", ["gpt-4o-mini"]) == ("openai/ft:gpt-4o-mini",)
+
+
+@pytest.mark.asyncio
+async def test_ahealth_check_partial_prefix_wildcard_probes_spliced_model_id():
+    from litellm.litellm_core_utils.llm_request_utils import pick_cheapest_chat_models_from_llm_provider
+
+    cheapest: Final = pick_cheapest_chat_models_from_llm_provider(custom_llm_provider="databricks", n=3)
+    stripped: Final = tuple(model_id.partition("/")[-1] for model_id in cheapest)
+
+    captured: dict[str, object] = {}
+
+    async def _capture(**kwargs):
+        captured.update(kwargs)
+
+    with patch("litellm.acompletion", side_effect=_capture):
+        await ahealth_check(
+            model_params={"model": "databricks/system.ai.*", "api_key": "x", "api_base": "https://example.invalid"},
+            mode="chat",
+        )
+
+    assert captured["model"] == f"databricks/system.ai.{stripped[0]}"
+    assert captured["fallbacks"] == [f"databricks/system.ai.{model_id}" for model_id in stripped[1:]]

From 4248af21c961c7be54ebec6b3fd5ad9db2526efb Mon Sep 17 00:00:00 2001
From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Tue, 29 Sep 2026 00:09:26 +0000
Subject: [PATCH 4/6] test(e2e): spawn wildcard repro proxy with isolated
 interpreter flag

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 tests/e2e/test_wildcard_partial_prefix_expansion.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/tests/e2e/test_wildcard_partial_prefix_expansion.py b/tests/e2e/test_wildcard_partial_prefix_expansion.py
index 48b265c2c62..dbe85028e10 100644
--- a/tests/e2e/test_wildcard_partial_prefix_expansion.py
+++ b/tests/e2e/test_wildcard_partial_prefix_expansion.py
@@ -75,6 +75,7 @@ def wildcard_proxy(tmp_path_factory: pytest.TempPathFactory) -> Generator[str]:
     proc = subprocess.Popen(
         [
             sys.executable,
+            "-P",
             str(REPO_ROOT / "litellm/proxy/proxy_cli.py"),
             "--config",
             str(config_path),

From a5b18431f7779a5b934c0283177d5f9743de3d65 Mon Sep 17 00:00:00 2001
From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Tue, 29 Sep 2026 03:11:00 +0000
Subject: [PATCH 5/6] fix(health): filter wildcard health probes to ids
 matching the literal prefix

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 .../health_check_helpers.py                   | 20 +++++++++----------
 .../test_wildcard_partial_prefix_expansion.py |  4 +++-
 .../test_health_check_helpers.py              | 14 +++++++++++--
 3 files changed, 24 insertions(+), 14 deletions(-)

diff --git a/litellm/litellm_core_utils/health_check_helpers.py b/litellm/litellm_core_utils/health_check_helpers.py
index 30e9b020245..692435b3910 100644
--- a/litellm/litellm_core_utils/health_check_helpers.py
+++ b/litellm/litellm_core_utils/health_check_helpers.py
@@ -45,19 +45,17 @@ def _strip_known_provider_prefix(model: str, known_providers: frozenset[str]) ->
 
 
 def _wildcard_health_check_models(
-    wildcard_suffix: str, custom_llm_provider: str, cheapest_models: Sequence[str]
+    wildcard_suffix: str, custom_llm_provider: str, candidate_models: Sequence[str]
 ) -> tuple[str, ...]:
     if wildcard_suffix == "*":
-        return tuple(cheapest_models)
+        return tuple(candidate_models[:3])
     known_providers: Final = frozenset(provider.value for provider in LlmProviders)
     literal_prefix: Final = wildcard_suffix.replace("*", "")
-    stripped_ids: Final = tuple(_strip_known_provider_prefix(model, known_providers) for model in cheapest_models)
-    return tuple(
-        f"{custom_llm_provider}/{stripped}"
-        if stripped.startswith(literal_prefix)
-        else f"{custom_llm_provider}/{wildcard_suffix.replace('*', stripped, 1)}"
-        for stripped in stripped_ids
-    )
+    stripped_ids: Final = tuple(_strip_known_provider_prefix(model, known_providers) for model in candidate_models)
+    matching: Final = [stripped for stripped in stripped_ids if stripped.startswith(literal_prefix)]
+    if matching:
+        return tuple(f"{custom_llm_provider}/{stripped}" for stripped in matching[:3])
+    return tuple(f"{custom_llm_provider}/{wildcard_suffix.replace('*', stripped, 1)}" for stripped in stripped_ids[:3])
 
 
 class HealthCheckHelpers:
@@ -74,13 +72,13 @@ class HealthCheckHelpers:
         )
 
         # this is a wildcard model, we need to pick a random model from the provider
-        cheapest_models = pick_cheapest_chat_models_from_llm_provider(custom_llm_provider=custom_llm_provider, n=3)
+        cheapest_models = pick_cheapest_chat_models_from_llm_provider(custom_llm_provider=custom_llm_provider, n=10_000)
         if len(cheapest_models) == 0:
             raise Exception(
                 f"Unable to health check wildcard model for provider {custom_llm_provider}. Add a model on your config.yaml or contribute here - https://github.com/BerriAI/litellm/blob/main/model_prices_and_context_window.json"
             )
         candidates: Final = _wildcard_health_check_models(
-            wildcard_suffix=model, custom_llm_provider=custom_llm_provider, cheapest_models=cheapest_models
+            wildcard_suffix=model, custom_llm_provider=custom_llm_provider, candidate_models=cheapest_models
         )
         fallback_models: Final = list(candidates[1:]) or None
         model_params["model"] = candidates[0]
diff --git a/tests/e2e/test_wildcard_partial_prefix_expansion.py b/tests/e2e/test_wildcard_partial_prefix_expansion.py
index dbe85028e10..1281e9d921b 100644
--- a/tests/e2e/test_wildcard_partial_prefix_expansion.py
+++ b/tests/e2e/test_wildcard_partial_prefix_expansion.py
@@ -146,5 +146,7 @@ def test_model_info_keeps_provider_model_for_expanded_deployments(wildcard_proxy
         )
     )
 
-    bad = [row.model_name for row in info.data if "system.ai.databricks/" in row.litellm_params.model]
+    expanded = [row for row in info.data if row.model_name.startswith("databricks/system.ai.")]
+    assert expanded, "model/info returned no expanded rows for databricks/system.ai.*"
+    bad = [row.model_name for row in expanded if "system.ai.databricks/" in row.litellm_params.model]
     assert not bad, f"litellm_params.model carries the corrupted expanded name: {bad}"
diff --git a/tests/unit/litellm_core_utils/test_health_check_helpers.py b/tests/unit/litellm_core_utils/test_health_check_helpers.py
index 1087d8be3ff..e5bd43369ec 100644
--- a/tests/unit/litellm_core_utils/test_health_check_helpers.py
+++ b/tests/unit/litellm_core_utils/test_health_check_helpers.py
@@ -564,14 +564,24 @@ def test_wildcard_health_check_models_partial_prefix_substitutes_stripped_id():
 
 
 def test_wildcard_health_check_models_star_wildcard_unchanged():
-    candidates: Final = ("databricks/databricks-gpt-5",)
-    assert _wildcard_health_check_models("*", "databricks", candidates) == candidates
+    candidates: Final = ("databricks/a", "databricks/b", "databricks/c", "databricks/d")
+    assert _wildcard_health_check_models("*", "databricks", candidates) == (
+        "databricks/a",
+        "databricks/b",
+        "databricks/c",
+    )
 
 
 def test_wildcard_health_check_models_partial_prefix_matching_literal_keeps_stripped_id():
     assert _wildcard_health_check_models("gpt-4*", "openai", ["gpt-4o-mini"]) == ("openai/gpt-4o-mini",)
 
 
+def test_wildcard_health_check_models_partial_prefix_filters_candidates_to_literal_prefix():
+    assert _wildcard_health_check_models(
+        "gpt-4*", "openai", ["gpt-5-nano", "gpt-4o-mini", "gpt-4.1-mini", "gpt-4o"]
+    ) == ("openai/gpt-4o-mini", "openai/gpt-4.1-mini", "openai/gpt-4o")
+
+
 def test_wildcard_health_check_models_partial_prefix_splices_suffix_around_star():
     assert _wildcard_health_check_models("ft:*", "openai", ["gpt-4o-mini"]) == ("openai/ft:gpt-4o-mini",)
 

From f5735591d7ba0b9df41e56bab8eeb1010a9d6188 Mon Sep 17 00:00:00 2001
From: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Date: Tue, 29 Sep 2026 03:18:48 +0000
Subject: [PATCH 6/6] test(integration): move partial wildcard expansion repro
 onto the shared proxy

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---
 .../test_wildcard_partial_prefix_expansion.py | 152 ------------------
 .../test_wildcard_partial_prefix_expansion.py |  71 ++++++++
 2 files changed, 71 insertions(+), 152 deletions(-)
 delete mode 100644 tests/e2e/test_wildcard_partial_prefix_expansion.py
 create mode 100644 tests/integration/routing/test_wildcard_partial_prefix_expansion.py

diff --git a/tests/e2e/test_wildcard_partial_prefix_expansion.py b/tests/e2e/test_wildcard_partial_prefix_expansion.py
deleted file mode 100644
index 1281e9d921b..00000000000
--- a/tests/e2e/test_wildcard_partial_prefix_expansion.py
+++ /dev/null
@@ -1,152 +0,0 @@
-"""Regression: wildcard expansion of a partial provider prefix like
-`databricks/system.ai.*` must produce `databricks/system.ai.<endpoint>` ids,
-not `databricks/system.ai.databricks/databricks-<name>`.
-
-On main, get_known_models_from_wildcard prepends the literal prefix before the
-`*` ("system.ai.") to cost-map keys that already carry the `databricks/`
-provider prefix, so /v1/models and /model/info list names that Databricks
-rejects with "Invalid Unity Catalog name". See LIT-8910.
-
-The test boots its own proxy (the shared stack's model_list is fixed), with two
-wildcard deployments: `databricks/system.ai.*` under test and `databricks/*` as
-a control whose expansion must stay `databricks/databricks-*`.
-"""
-
-import socket
-import subprocess
-import sys
-import time
-from collections.abc import Generator
-from pathlib import Path
-from typing import Final, cast
-
-import pytest
-from e2e_http import URL, AuthHeaders, NoBody, get, probe, unwrap
-from models import ModelsListParams, ModelsListResponse
-from pydantic import BaseModel, ConfigDict
-
-REPO_ROOT: Final = Path(__file__).resolve().parents[2]
-
-MASTER_KEY: Final = "sk-1234"
-AUTH: Final = AuthHeaders(authorization=f"Bearer {MASTER_KEY}")
-
-CONFIG_YAML: Final = """\
-model_list:
-  - model_name: databricks/system.ai.*
-    litellm_params:
-      model: databricks/system.ai.*
-      api_key: fake-databricks-token
-      api_base: https://example.invalid
-  - model_name: databricks/*
-    litellm_params:
-      model: databricks/*
-      api_key: fake-databricks-token
-      api_base: https://example.invalid
-general_settings:
-  master_key: sk-1234
-"""
-
-
-class _ModelInfoParams(BaseModel):
-    model_config = ConfigDict(protected_namespaces=(), extra="ignore")
-    model: str = ""
-
-
-class _ModelInfoRow(BaseModel):
-    model_name: str
-    litellm_params: _ModelInfoParams = _ModelInfoParams()
-
-
-class _ModelInfoResponse(BaseModel):
-    data: list[_ModelInfoRow] = []
-
-
-def _free_port() -> int:
-    with socket.socket() as sock:
-        sock.bind(("127.0.0.1", 0))
-        return cast("tuple[str, int]", sock.getsockname())[1]
-
-
-@pytest.fixture(scope="module")
-def wildcard_proxy(tmp_path_factory: pytest.TempPathFactory) -> Generator[str]:
-    config_path = tmp_path_factory.mktemp("wildcard") / "config.yaml"
-    config_path.write_text(CONFIG_YAML)
-    port = _free_port()
-    proc = subprocess.Popen(
-        [
-            sys.executable,
-            "-P",
-            str(REPO_ROOT / "litellm/proxy/proxy_cli.py"),
-            "--config",
-            str(config_path),
-            "--port",
-            str(port),
-        ],
-        cwd=REPO_ROOT,
-        env={
-            "PATH": "/usr/bin:/bin:/usr/local/bin",
-            "HOME": str(Path.home()),
-            "PYTHONPATH": str(REPO_ROOT),
-            "LITELLM_LOCAL_MODEL_COST_MAP": "true",
-            "LITELLM_DANGEROUSLY_PERMIT_WEAK_OR_UNSET_MASTER_KEY": "true",
-            "LITELLM_MASTER_KEY": MASTER_KEY,
-        },
-    )
-    base = f"http://127.0.0.1:{port}"
-    try:
-        deadline = time.time() + 180
-        while True:
-            alive = probe(URL(f"{base}/health/liveliness"), headers=NoBody(), params=NoBody(), timeout=5.0)
-            if alive.healthy:
-                break
-            if proc.poll() is not None:
-                pytest.fail(f"proxy exited during boot with {proc.returncode}")
-            assert time.time() < deadline, "proxy did not answer /health/liveliness within 180s"
-            time.sleep(2)
-        yield base
-    finally:
-        proc.terminate()
-        proc.wait(timeout=30)
-
-
-def _list_model_ids(base: str) -> list[str]:
-    listed = get(
-        URL(f"{base}/v1/models"),
-        headers=AUTH,
-        params=ModelsListParams(return_wildcard_routes=False),
-        response_type=ModelsListResponse,
-    )
-    return [entry.id for entry in unwrap(listed).data]
-
-
-def test_partial_prefix_wildcard_expands_to_unity_catalog_names(wildcard_proxy: str) -> None:
-    ids = _list_model_ids(wildcard_proxy)
-
-    system_ai_ids = [model_id for model_id in ids if model_id.startswith("databricks/system.ai.")]
-    assert system_ai_ids, f"no databricks/system.ai.* expansion in {ids}"
-
-    malformed = [model_id for model_id in system_ai_ids if "/" in model_id.removeprefix("databricks/system.ai.")]
-    assert not malformed, f"expanded ids embed the provider-prefixed cost-map key after 'system.ai.': {malformed}"
-
-
-def test_provider_wildcard_still_expands_to_cost_map_names(wildcard_proxy: str) -> None:
-    ids = _list_model_ids(wildcard_proxy)
-
-    databricks_ids = [model_id for model_id in ids if model_id.startswith("databricks/databricks-")]
-    assert databricks_ids, f"databricks/* did not expand to cost-map names: {ids}"
-
-
-def test_model_info_keeps_provider_model_for_expanded_deployments(wildcard_proxy: str) -> None:
-    info = unwrap(
-        get(
-            URL(f"{wildcard_proxy}/model/info"),
-            headers=AUTH,
-            params=NoBody(),
-            response_type=_ModelInfoResponse,
-        )
-    )
-
-    expanded = [row for row in info.data if row.model_name.startswith("databricks/system.ai.")]
-    assert expanded, "model/info returned no expanded rows for databricks/system.ai.*"
-    bad = [row.model_name for row in expanded if "system.ai.databricks/" in row.litellm_params.model]
-    assert not bad, f"litellm_params.model carries the corrupted expanded name: {bad}"
diff --git a/tests/integration/routing/test_wildcard_partial_prefix_expansion.py b/tests/integration/routing/test_wildcard_partial_prefix_expansion.py
new file mode 100644
index 00000000000..01206a2163a
--- /dev/null
+++ b/tests/integration/routing/test_wildcard_partial_prefix_expansion.py
@@ -0,0 +1,71 @@
+"""Partial-prefix wildcards like ``databricks/system.ai.*`` must splice the literal
+prefix onto provider-stripped cost-map ids, never onto the provider-prefixed key."""
+
+from collections.abc import Iterator
+from typing import Final
+
+import pytest
+
+from integration._support.client import Gateway, object_value, string_value
+
+PATTERNS: Final = ("databricks/system.ai.*", "databricks/*")
+
+
+def _registered_model_id(gateway: Gateway, pattern: str) -> str:
+    created: Final = object_value(
+        gateway.post(
+            "/model/new",
+            {
+                "model_name": pattern,
+                "litellm_params": {
+                    "model": pattern,
+                    "api_key": "integration-wildcard-key",
+                    "api_base": "https://example.invalid",
+                },
+            },
+        )
+    )
+    return string_value(object_value(created["model_info"])["id"])
+
+
+@pytest.fixture
+def wildcard_deployments(gateway: Gateway) -> Iterator[None]:
+    with gateway.scenario() as scenario:
+        for pattern in PATTERNS:
+            scenario.cleanups.callback(scenario.delete_model, _registered_model_id(gateway, pattern))
+        yield
+
+
+def _listed_model_ids(gateway: Gateway) -> tuple[str, ...]:
+    data: Final = gateway.get("/v1/models")["data"]
+    assert isinstance(data, list)
+    return tuple(string_value(object_value(entry)["id"]) for entry in data)
+
+
+def test_partial_prefix_wildcard_expands_to_unity_catalog_names(gateway: Gateway, wildcard_deployments: None) -> None:
+    prefixed: Final = tuple(
+        model_id for model_id in _listed_model_ids(gateway) if model_id.startswith("databricks/system.ai.")
+    )
+    assert prefixed
+    assert all("/" not in model_id.removeprefix("databricks/system.ai.") for model_id in prefixed)
+
+
+def test_provider_wildcard_still_expands_to_cost_map_names(gateway: Gateway, wildcard_deployments: None) -> None:
+    assert any(model_id.startswith("databricks/databricks-") for model_id in _listed_model_ids(gateway))
+
+
+def test_model_info_keeps_provider_model_for_expanded_deployments(gateway: Gateway, wildcard_deployments: None) -> None:
+    data: Final = gateway.get("/model/info")["data"]
+    assert isinstance(data, list)
+    expanded: Final = [
+        row
+        for row in (object_value(entry) for entry in data)
+        if string_value(row["model_name"]).startswith("databricks/system.ai.")
+    ]
+    assert expanded, "model/info returned no expanded rows for databricks/system.ai.*"
+    bad: Final = [
+        row["model_name"]
+        for row in expanded
+        if "system.ai.databricks/" in string_value(object_value(row["litellm_params"])["model"])
+    ]
+    assert not bad, f"litellm_params.model carries the corrupted expanded name: {bad}"