fix(config): don't set litellm.api_key global when LLM_API_BASE is configured (#1095)

When LLM_API_BASE is set, the model routes through openai-agents'
LitellmModel which always forwards api_key=self.api_key (None by
default) as an explicit kwarg on every litellm.acompletion() call.
Setting litellm.api_key as a module-level global on top of that caused:

  TypeError: acompletion() got multiple values for keyword argument
             'api_key'           (anthropic/ prefix path)
  TypeError: AsyncCompletions.create() got an unexpected keyword
             argument 'api_key' (openai/ prefix path)

Fix: guard _configure_litellm_default('api_key', ...) behind
'if not llm.api_base'. The key still reaches the provider via
_mirror_api_key_to_provider_env (sets the provider-specific *_API_KEY
env var e.g. ANTHROPIC_API_KEY) and set_default_openai_key, so
skipping the module-level global is safe.

Fixes #1095

strix/config/models.py

@@ -562,7 +562,26 @@ def configure_sdk_model_defaults(settings: Settings) -> None:
     _configure_openrouter_attribution(llm.model)
     if llm.api_key:
         set_default_openai_key(llm.api_key, use_for_tracing=False)
-        _configure_litellm_default("api_key", llm.api_key)
+        # Do NOT set litellm.api_key as a module-level global when a custom
+        # api_base is configured. When api_base is active the model resolves
+        # through openai-agents' LitellmModel, which always forwards
+        # api_key=self.api_key (None by default, because Strix never populates
+        # it on the instance) as an *explicit* kwarg in every
+        # litellm.acompletion() call.  litellm's internal dispatch then merges
+        # the module-level global into the same kwargs dict that already
+        # carries the explicit kwarg, producing:
+        #
+        #   TypeError: acompletion() got multiple values for keyword argument
+        #              'api_key'          (anthropic/ prefix path)
+        #   TypeError: AsyncCompletions.create() got an unexpected keyword
+        #              argument 'api_key' (openai/ prefix path)
+        #
+        # The key is still delivered to the provider via
+        # _mirror_api_key_to_provider_env (sets the provider-specific
+        # *_API_KEY env var, e.g. ANTHROPIC_API_KEY) and set_default_openai_key
+        # above, so skipping the global default here is safe.
+        if not llm.api_base:
+            _configure_litellm_default("api_key", llm.api_key)
         _mirror_api_key_to_provider_env(llm.model, llm.api_key)
     if llm.api_base:
         os.environ["OPENAI_BASE_URL"] = llm.api_base

tests/test_llm_api_base_no_duplicate_key.py

@@ -0,0 +1,129 @@
+"""Regression tests for issue #1095.
+
+LLM CONNECTION FAILED — duplicate api_key keyword argument when LLM_API_BASE
+is set to a custom endpoint.
+
+When LLM_API_BASE is configured the model resolves through openai-agents'
+LitellmModel, which always forwards api_key=self.api_key (None by default) as
+an *explicit* kwarg on every litellm.acompletion() call.  If litellm.api_key
+is also set as a module-level global by configure_sdk_model_defaults, litellm's
+internal dispatch merges the global into the same kwargs dict that already
+carries the explicit kwarg and raises:
+
+    TypeError: acompletion() got multiple values for keyword argument 'api_key'
+                                   (anthropic/ / litellm/ prefix path)
+    TypeError: AsyncCompletions.create() got an unexpected keyword argument
+               'api_key'            (openai/ prefix path)
+
+The fix: skip _configure_litellm_default("api_key", ...) when api_base is set.
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING
+
+import litellm
+import pytest
+
+from strix.config import loader
+from strix.config.loader import load_settings
+from strix.config.models import configure_sdk_model_defaults
+
+
+if TYPE_CHECKING:
+    from collections.abc import Iterator
+
+
+_ENV_KEYS = [
+    # Primary Strix settings
+    "STRIX_LLM",
+    "LLM_API_KEY",
+    "LLM_API_BASE",
+    # All aliases that LlmSettings maps to api_base (settings.py AliasChoices),
+    # so configure_sdk_model_defaults()'s os.environ["OPENAI_BASE_URL"] write
+    # in one test does not bleed into the next test's load_settings() call.
+    "OPENAI_API_BASE",
+    "OPENAI_BASE_URL",
+    "LITELLM_BASE_URL",
+    "OLLAMA_API_BASE",
+]
+
+
+@pytest.fixture(autouse=True)
+def _reset(monkeypatch: pytest.MonkeyPatch) -> Iterator[None]:
+    """Isolate litellm.api_key and the settings cache between tests."""
+    for key in _ENV_KEYS:
+        monkeypatch.delenv(key, raising=False)
+    monkeypatch.setattr(loader, "_cached", None)
+    monkeypatch.setattr(loader, "_override", None)
+
+    saved_api_key = litellm.api_key
+    litellm.api_key = None
+    try:
+        yield
+    finally:
+        litellm.api_key = saved_api_key
+
+
+
+# ---------------------------------------------------------------------------
+# Core regression: litellm.api_key must NOT be set when api_base is active
+# ---------------------------------------------------------------------------
+
+
+def test_litellm_api_key_global_not_set_when_api_base_configured(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Issue #1095 (anthropic/ path): litellm.api_key must remain None when
+    LLM_API_BASE is set, so that LitellmModel's explicit api_key=None kwarg
+    does not collide with a module-level global."""
+    monkeypatch.setenv("STRIX_LLM", "anthropic/deepseek-main")
+    monkeypatch.setenv("LLM_API_KEY", "placeholder")
+    monkeypatch.setenv("LLM_API_BASE", "http://192.168.1.18:8081")
+
+    configure_sdk_model_defaults(load_settings())
+
+    # The module-level litellm.api_key must NOT be set; the key reaches the
+    # provider through provider-specific env vars (_mirror_api_key_to_provider_env).
+    assert litellm.api_key is None, (
+        "litellm.api_key was set as a module-level global even though "
+        "LLM_API_BASE is configured. This causes "
+        "'acompletion() got multiple values for keyword argument api_key' "
+        "(issue #1095)."
+    )
+
+
+def test_litellm_api_key_global_not_set_for_openai_prefix_with_custom_base(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Issue #1095 (openai/ path): same guard applies regardless of the
+    STRIX_LLM prefix used."""
+    monkeypatch.setenv("STRIX_LLM", "openai/deepseek-main")
+    monkeypatch.setenv("LLM_API_KEY", "placeholder")
+    monkeypatch.setenv("LLM_API_BASE", "http://192.168.1.18:8081/v1")
+
+    configure_sdk_model_defaults(load_settings())
+
+    assert litellm.api_key is None, (
+        "litellm.api_key was set as a module-level global even though "
+        "LLM_API_BASE is configured (openai/ prefix path). This causes "
+        "'AsyncCompletions.create() got an unexpected keyword argument api_key'"
+        " (issue #1095)."
+    )
+
+
+def test_litellm_api_key_global_set_when_no_api_base(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """Positive case: litellm.api_key IS set when no custom api_base is given
+    (standard direct-to-provider path)."""
+    monkeypatch.setenv("STRIX_LLM", "openai/gpt-5.4")
+    monkeypatch.setenv("LLM_API_KEY", "sk-real-key")
+    # LLM_API_BASE is intentionally absent
+
+    configure_sdk_model_defaults(load_settings())
+
+    assert litellm.api_key == "sk-real-key", (
+        "litellm.api_key should be set globally when no custom api_base is "
+        "configured (standard direct-to-provider path)."
+    )