fix(models): reasoning support follows the model, not the route

model_supports_reasoning only consulted LiteLLM's map under the exact route
name, so openai/glm-5.3 (mapped only as zai/... and openrouter/...) silently
dropped STRIX_REASONING_EFFORT. Fall back to any provider's entry for the
same bare model name.

strix/config/models.py

@@ -837,7 +837,17 @@ def model_supports_reasoning(model_name: str) -> bool:
     entry = litellm.model_cost.get(name)
     if entry is None and "/" in name:
         entry = litellm.model_cost.get(name.rsplit("/", 1)[1])
-    return bool(entry and entry.get("supports_reasoning"))
+    if entry is not None:
+        return bool(entry.get("supports_reasoning"))
+    # Reasoning is a property of the model, not of the route: a model only
+    # mapped under other providers (``zai/glm-5.3`` for ``openai/glm-5.3``)
+    # still reasons through an OpenAI-compatible gateway.
+    bare_name = name.rsplit("/", 1)[-1]
+    return any(
+        bool(candidate.get("supports_reasoning"))
+        for key, candidate in litellm.model_cost.items()
+        if key.endswith(f"/{bare_name}") and isinstance(candidate, dict)
+    )
 
 
 def is_recommended_or_frontier_model(model_name: str) -> bool:

tests/test_models.py

@@ -12,6 +12,7 @@ from strix.config.models import (
     _NonStreamingModel,
     _TurnGuardModel,
     is_recommended_or_frontier_model,
+    model_supports_reasoning,
     request_timeout_extra_args,
     routes_through_litellm,
     supports_strict_tool_schemas,
@@ -23,6 +24,19 @@ def test_recommended_models_are_accepted(model_name: str) -> None:
     assert is_recommended_or_frontier_model(model_name)
 
 
+@pytest.mark.parametrize(
+    "model_name",
+    ["openai/glm-5.3", "glm-5.3", "hosted_vllm/glm-5.3", "openai/kimi-k3", "openai/qwen3.8-max"],
+)
+def test_reasoning_support_follows_the_model_not_the_route(model_name: str) -> None:
+    assert model_supports_reasoning(model_name)
+
+
+def test_reasoning_support_rejects_unknown_models() -> None:
+    assert not model_supports_reasoning("openai/no-such-model-xyz")
+    assert not model_supports_reasoning("openai/gpt-4.1")
+
+
 def test_request_timeout_extra_args_positive() -> None:
     assert request_timeout_extra_args(300) == {"timeout": 300}
     assert request_timeout_extra_args(10) == {"timeout": 10}