Fix Gemini ACP model IDs with Google provider prefixes

src/benchflow/acp/runtime.py

@@ -269,6 +269,11 @@ def _format_acp_model(model: str, agent: str) -> str:
     models.dev provider prefix when the agent requires it.
     """
     bare = strip_provider_prefix(model)
+    # Gemini CLI accepts bare Google model IDs. ``google/`` is a models.dev
+    # provider prefix rather than a registered BenchFlow provider, so the
+    # generic normalizer intentionally leaves it alone.
+    if agent == "gemini" and model.startswith(("google/gemini-", "google/gemma-")):
+        bare = model.removeprefix("google/")
     agent_cfg = AGENTS.get(agent)
     if agent_cfg and agent_cfg.acp_model_format == "registered-provider/model":
         # Proxy mode: BenchFlow's LiteLLM proxy serves the model under the alias

src/benchflow/providers/litellm_config.py

@@ -375,6 +375,9 @@ def resolve_litellm_route(model: str, env: dict[str, str]) -> LiteLLMRoute:
     elif lower.startswith("gemini/"):
         upstream = model
         required = ("GEMINI_API_KEY",)
+    elif lower.startswith(("google/gemini-", "google/gemma-")):
+        upstream = f"gemini/{model.split('/', 1)[1]}"
+        required = ("GEMINI_API_KEY",)
     elif "gemini" in lower:
         upstream = f"gemini/{bare}"
         required = ("GEMINI_API_KEY",)

tests/test_acp.py

@@ -1176,22 +1176,39 @@ def _make_mocks():
 
     @pytest.mark.asyncio
     @pytest.mark.parametrize(
-        "model_in, expected_model",
+        "agent, model_in, expected_model",
         [
             # Registered vllm/ prefix stripped; HF org/model intact — this is
             # what pi-acp and other ACP agents need for downstream routing.
-            ("vllm/Qwen/Qwen3.5-35B-A3B", "Qwen/Qwen3.5-35B-A3B"),
-            ("zai/glm-5", "glm-5"),
+            ("test-agent", "vllm/Qwen/Qwen3.5-35B-A3B", "Qwen/Qwen3.5-35B-A3B"),
+            ("test-agent", "zai/glm-5", "glm-5"),
             # Bare HF ID (no registered prefix) passes through unchanged.
-            ("Qwen/Qwen3-Coder", "Qwen/Qwen3-Coder"),
+            ("test-agent", "Qwen/Qwen3-Coder", "Qwen/Qwen3-Coder"),
             # Vertex ADC provider — prefix stripped like any other registered one.
-            ("anthropic-vertex/claude-sonnet-4-6", "claude-sonnet-4-6"),
+            ("test-agent", "anthropic-vertex/claude-sonnet-4-6", "claude-sonnet-4-6"),
             # No prefix at all — unchanged.
-            ("claude-sonnet-4-6", "claude-sonnet-4-6"),
+            ("test-agent", "claude-sonnet-4-6", "claude-sonnet-4-6"),
+            # Gemini CLI expects bare Google Gemini and Gemma model IDs.
+            (
+                "gemini",
+                "google/gemini-3.1-flash-lite-preview",
+                "gemini-3.1-flash-lite-preview",
+            ),
+            ("gemini", "google/gemma-3-27b-it", "gemma-3-27b-it"),
+            ("gemini", "google/text-bison", "google/text-bison"),
+        ],
+        ids=[
+            "vllm-hf",
+            "zai",
+            "bare-hf",
+            "vertex",
+            "no-prefix",
+            "gemini-google",
+            "gemma-google",
+            "unrelated-google",
         ],
-        ids=["vllm-hf", "zai", "bare-hf", "vertex", "no-prefix"],
     )
-    async def test_model_id_selection(self, model_in, expected_model, tmp_path):
+    async def test_model_id_selection(self, agent, model_in, expected_model, tmp_path):
         from benchflow.acp.runtime import connect_acp
 
         mock_acp = self._make_mocks()
@@ -1203,7 +1220,7 @@ async def test_model_id_selection(self, model_in, expected_model, tmp_path):
         ):
             await connect_acp(
                 env=mock_env,
-                agent="test-agent",
+                agent=agent,
                 agent_launch="test-agent",
                 agent_env={},
                 sandbox_user=None,

tests/test_litellm_config.py

@@ -274,3 +274,24 @@ def test_proxy_config_no_responses_bridge_for_non_openai_upstream():
     config = litellm_proxy_config(route, master_key="sk-local")
     names = [entry["model_name"] for entry in config["model_list"]]
     assert not any(n.endswith("-responses-bridge") for n in names)
+
+
+@pytest.mark.parametrize(
+    "model,expected",
+    [
+        (
+            "google/gemini-3.1-flash-lite-preview",
+            "gemini/gemini-3.1-flash-lite-preview",
+        ),
+        ("google/gemma-3-27b-it", "gemini/gemma-3-27b-it"),
+        ("google/geminix-1", "gemini/google/geminix-1"),
+    ],
+)
+def test_google_model_normalizes_only_gemini_families(model, expected):
+    """Google Gemini/Gemma IDs get one LiteLLM provider prefix."""
+    route = resolve_litellm_route(
+        model,
+        {"GEMINI_API_KEY": "key"},
+    )
+    assert route.upstream_model == expected
+    assert route.litellm_params["model"] == expected

tests/test_litellm_hardening.py

@@ -83,6 +83,16 @@ def test_format_acp_model_passes_through_existing_provider_prefix():
     )
 
 
+def test_gemini_acp_model_strips_google_prefix():
+    """Guards Gemini model-resource 404s for current Google model IDs."""
+    from benchflow.acp.runtime import _format_acp_model
+
+    assert (
+        _format_acp_model("google/gemini-3.5-flash-lite", "gemini")
+        == "gemini-3.5-flash-lite"
+    )
+
+
 def test_mimo_registered_xiaomi_model_keeps_provider_prefix():
     from benchflow.acp.runtime import _format_acp_model