Fix: propagate top-level llm.provider to per-model configs

The `provider` field (e.g. "claude_code") set at the top `llm:` level was
omitted from the shared config that LLMConfig propagates down to each model,
so it never reached the individual LLMModelConfig entries. As a result the
ensemble ignored `provider: "claude_code"` and fell back to the OpenAI
backend, crashing with "Missing credentials ... OPENAI_API_KEY" — which
made the examples/claude_code_quickstart example unusable out of the box.

Add `provider` to the shared_config dicts in both LLMConfig.__post_init__
and rebuild_models. Propagation uses update_model_params (overwrite=False),
so an explicit per-model `provider` still takes precedence.

Add regression tests asserting that a top-level provider reaches every
evolution and evaluator model, that per-model overrides win, that the
default remains None, and that the ensemble builds ClaudeCodeLLM end to end.

Author: froucoux

openevolve/config.py

@@ -174,6 +174,7 @@ def __post_init__(self):
 
         # Update models with shared configuration values
         shared_config = {
+            "provider": self.provider,
             "api_base": self.api_base,
             "api_key": self.api_key,
             "temperature": self.temperature,
@@ -228,6 +229,7 @@ def rebuild_models(self) -> None:
 
         # Update models with shared configuration values
         shared_config = {
+            "provider": self.provider,
             "api_base": self.api_base,
             "api_key": self.api_key,
             "temperature": self.temperature,

tests/test_claude_code_llm.py

@@ -133,6 +133,81 @@ def test_max_budget_usd_from_dict(self):
         self.assertEqual(config.llm.models[0].max_budget_usd, 3.0)
 
 
+class TestProviderPropagation(unittest.TestCase):
+    """A top-level ``llm.provider`` must reach each model in the ensemble.
+
+    Regression test for a bug where ``provider`` was omitted from the shared
+    config propagated to per-model configs, so ``provider: "claude_code"`` set
+    at the ``llm:`` level was silently dropped and every model fell back to the
+    OpenAI backend (crashing with "Missing credentials").
+    """
+
+    def test_top_level_provider_propagates_to_models(self):
+        from openevolve.config import Config
+
+        config = Config.from_dict(
+            {
+                "llm": {
+                    "provider": "claude_code",
+                    "models": [
+                        {"name": "sonnet", "weight": 0.8},
+                        {"name": "haiku", "weight": 0.2},
+                    ],
+                }
+            }
+        )
+        self.assertTrue(config.llm.models, "expected models to be configured")
+        for model in config.llm.models:
+            self.assertEqual(model.provider, "claude_code")
+        # Evaluator models default to the evolution models and must inherit too.
+        for model in config.llm.evaluator_models:
+            self.assertEqual(model.provider, "claude_code")
+
+    def test_per_model_provider_overrides_top_level(self):
+        from openevolve.config import Config
+
+        config = Config.from_dict(
+            {
+                "llm": {
+                    "provider": "claude_code",
+                    "models": [
+                        {"name": "sonnet", "weight": 0.5},
+                        {"name": "gpt-4o", "weight": 0.5, "provider": "openai"},
+                    ],
+                }
+            }
+        )
+        providers = {m.name: m.provider for m in config.llm.models}
+        self.assertEqual(providers["sonnet"], "claude_code")
+        self.assertEqual(providers["gpt-4o"], "openai")
+
+    def test_default_provider_is_none(self):
+        from openevolve.config import Config
+
+        config = Config.from_dict(
+            {"llm": {"models": [{"name": "gpt-4o", "weight": 1.0}]}}
+        )
+        self.assertIsNone(config.llm.models[0].provider)
+
+    def test_ensemble_builds_claude_code_from_config(self):
+        from openevolve.config import Config
+        from openevolve.llm.ensemble import LLMEnsemble
+
+        config = Config.from_dict(
+            {
+                "llm": {
+                    "provider": "claude_code",
+                    "models": [{"name": "sonnet", "weight": 1.0}],
+                }
+            }
+        )
+        ensemble = LLMEnsemble(config.llm.models)
+        self.assertTrue(
+            all(isinstance(m, ClaudeCodeLLM) for m in ensemble.models),
+            "ensemble should build ClaudeCodeLLM instances from top-level provider",
+        )
+
+
 class TestProviderRegistry(unittest.TestCase):
     def test_claude_code_in_registry(self):
         from openevolve.llm.ensemble import _PROVIDER_REGISTRY