fix(llm): wire RetryingLLMClient with MAX_RETRIES in LLM factory

agentkit/config.py

@@ -71,6 +71,11 @@ class Settings(BaseSettings):
         gt=0,
         description="Maximum execution steps allowed per agent run.",
     )
+    MAX_RETRIES: int = Field(
+        default=3,
+        ge=0,
+        description="Maximum retries on transient LLM errors.",
+    )
     RUN_TIMEOUT_S: int = Field(
         default=60,
         gt=0,

agentkit/llm/factory.py

@@ -8,6 +8,7 @@
 from agentkit.llm.fake import FakeLLMClient
 from agentkit.llm.gemini import GeminiClient
 from agentkit.llm.openai import OpenAIClient
+from agentkit.llm.retry import RetryingLLMClient
 
 logger = logging.getLogger(__name__)
 
@@ -59,7 +60,7 @@ def create_llm_client_from_settings(settings: Settings) -> LLMClient:
         settings: Application Settings instance.
 
     Returns:
-        Configured LLMClient instance.
+        Configured LLMClient instance (wrapped with RetryingLLMClient for live providers).
     """
     provider = settings.LLM_PROVIDER.strip().lower()
     api_key: str | None = None
@@ -69,8 +70,16 @@ def create_llm_client_from_settings(settings: Settings) -> LLMClient:
     elif provider == "openai":
         api_key = settings.OPENAI_API_KEY
 
-    return create_llm_client(
+    client = create_llm_client(
         provider=provider,
         api_key=api_key,
         model=settings.LLM_MODEL,
     )
+
+    if provider in ("gemini", "openai") and settings.MAX_RETRIES > 0:
+        return RetryingLLMClient(
+            client=client,
+            max_retries=settings.MAX_RETRIES,
+        )
+
+    return client

tests/unit/test_llm_factory.py

@@ -5,6 +5,7 @@
 from agentkit.llm.fake import FakeLLMClient
 from agentkit.llm.gemini import GeminiClient
 from agentkit.llm.openai import OpenAIClient
+from agentkit.llm.retry import RetryingLLMClient
 
 
 def test_create_fake_client() -> None:
@@ -47,24 +48,43 @@ def test_create_from_settings_gemini() -> None:
         LLM_PROVIDER="gemini",
         LLM_MODEL="gemini-2.0-flash",
         GEMINI_API_KEY="test-gemini-key",
+        MAX_RETRIES=3,
     )
     client = create_llm_client_from_settings(settings)
-    assert isinstance(client, GeminiClient)
+    assert isinstance(client, RetryingLLMClient)
+    assert isinstance(client._client, GeminiClient)
     assert client.model == "gemini-2.0-flash"
+    assert client._max_retries == 3
 
 
 def test_create_from_settings_openai() -> None:
     settings = Settings(
         LLM_PROVIDER="openai",
         LLM_MODEL="gpt-4o",
         OPENAI_API_KEY="test-openai-key",
+        MAX_RETRIES=2,
     )
     client = create_llm_client_from_settings(settings)
-    assert isinstance(client, OpenAIClient)
+    assert isinstance(client, RetryingLLMClient)
+    assert isinstance(client._client, OpenAIClient)
     assert client.model == "gpt-4o"
+    assert client._max_retries == 2
+
+
+def test_create_from_settings_zero_retries() -> None:
+    settings = Settings(
+        LLM_PROVIDER="gemini",
+        LLM_MODEL="gemini-2.0-flash",
+        GEMINI_API_KEY="test-gemini-key",
+        MAX_RETRIES=0,
+    )
+    client = create_llm_client_from_settings(settings)
+    assert isinstance(client, GeminiClient)
+    assert not isinstance(client, RetryingLLMClient)
 
 
 def test_create_from_settings_fake() -> None:
     settings = Settings(LLM_PROVIDER="fake")
     client = create_llm_client_from_settings(settings)
     assert isinstance(client, FakeLLMClient)
+    assert not isinstance(client, RetryingLLMClient)