feat: allow using llama-stack-library-client from verifications (#2238)

Having to run (and re-run) a server while running verifications can be
annoying while you are iterating on code. This makes it so you can use
the library client -- and because it is OpenAI client compatible, it all
works.

## Test Plan

```
pytest -s -v tests/verifications/openai_api/test_responses.py \
   --provider=stack:together \
   --model meta-llama/Llama-4-Scout-17B-16E-Instruct
```

tests/verifications/conftest.py

@@ -25,6 +25,11 @@ def pytest_addoption(parser):
         action="store",
         help="Provider to use for testing",
     )
+    parser.addoption(
+        "--model",
+        action="store",
+        help="Model to use for testing",
+    )
 
 
 pytest_plugins = [

tests/verifications/openai_api/conftest.py

@@ -16,6 +16,11 @@ def pytest_generate_tests(metafunc):
             metafunc.parametrize("model", [])
             return
 
+        model = metafunc.config.getoption("model")
+        if model:
+            metafunc.parametrize("model", [model])
+            return
+
         try:
             config_data = _load_all_verification_configs()
         except (OSError, FileNotFoundError) as e:

tests/verifications/openai_api/fixtures/fixtures.py

@@ -12,6 +12,8 @@ import pytest
 import yaml
 from openai import OpenAI
 
+from llama_stack import LlamaStackAsLibraryClient
+
 # --- Helper Functions ---
 
 
@@ -81,7 +83,7 @@ def verification_config():
         pytest.fail(str(e))  # Fail test collection if config loading fails
 
 
-@pytest.fixture
+@pytest.fixture(scope="session")
 def provider(request, verification_config):
     provider = request.config.getoption("--provider")
     base_url = request.config.getoption("--base-url")
@@ -100,12 +102,14 @@ def provider(request, verification_config):
     return provider
 
 
-@pytest.fixture
+@pytest.fixture(scope="session")
 def base_url(request, provider, verification_config):
-    return request.config.getoption("--base-url") or verification_config["providers"][provider]["base_url"]
+    return request.config.getoption("--base-url") or verification_config.get("providers", {}).get(provider, {}).get(
+        "base_url"
+    )
 
 
-@pytest.fixture
+@pytest.fixture(scope="session")
 def api_key(request, provider, verification_config):
     provider_conf = verification_config.get("providers", {}).get(provider, {})
     api_key_env_var = provider_conf.get("api_key_var")
@@ -122,11 +126,21 @@ def model_mapping(provider, providers_model_mapping):
     return providers_model_mapping[provider]
 
 
-@pytest.fixture
+@pytest.fixture(scope="session")
-def openai_client(base_url, api_key):
+def openai_client(base_url, api_key, provider):
     # Simplify running against a local Llama Stack
-    if "localhost" in base_url and not api_key:
+    if base_url and "localhost" in base_url and not api_key:
         api_key = "empty"
+    if provider.startswith("stack:"):
+        parts = provider.split(":")
+        if len(parts) != 2:
+            raise ValueError(f"Invalid config for Llama Stack: {provider}, it must be of the form 'stack:<config>'")
+        config = parts[1]
+        client = LlamaStackAsLibraryClient(config, skip_logger_removal=True)
+        if not client.initialize():
+            raise RuntimeError("Initialization failed")
+        return client
+
     return OpenAI(
         base_url=base_url,
         api_key=api_key,