fix: gate reasoning_effort by LiteLLM model registry (closes #517) (#523)

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>

strix/config/models.py

@@ -99,3 +99,17 @@ def uses_chat_completions_tool_schema(model_name: str, settings: Settings) -> bo
     if model.startswith(("litellm/", "any-llm/")):
         return True
     return bool(settings.llm.api_base)
+
+
+def model_supports_reasoning(model_name: str) -> bool:
+    import litellm
+
+    name = model_name.strip().lower()
+    for prefix in ("litellm/", "any-llm/"):
+        if name.startswith(prefix):
+            name = name[len(prefix) :]
+            break
+    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"))

strix/core/inputs.py

@@ -8,7 +8,7 @@ from typing import TYPE_CHECKING, Any
 from agents.model_settings import ModelSettings
 from openai.types.shared import Reasoning
 
-from strix.config.models import DEFAULT_MODEL_RETRY
+from strix.config.models import DEFAULT_MODEL_RETRY, model_supports_reasoning
 
 
 if TYPE_CHECKING:
@@ -108,13 +108,19 @@ def build_scope_context(scan_config: dict[str, Any]) -> dict[str, Any]:
 
 def make_model_settings(
     reasoning_effort: ReasoningEffort | None,
+    *,
+    model_name: str,
 ) -> ModelSettings:
     # Anthropic + DeepSeek thinking reject ``tool_choice="required"`` outright
     # when reasoning is enabled; OpenAI o-series accepts both but doesn't need
     # the safety net. When reasoning is on we let the model self-select tools
     # and rely on the system prompt + the ``_finish_tool_use_behavior`` callback
     # to keep the loop converging on a lifecycle tool.
-    use_reasoning = reasoning_effort is not None and reasoning_effort != "none"
+    use_reasoning = (
+        reasoning_effort is not None
+        and reasoning_effort != "none"
+        and model_supports_reasoning(model_name)
+    )
     model_settings = ModelSettings(
         parallel_tool_calls=False,
         tool_choice=None if use_reasoning else "required",

strix/core/runner.py

@@ -153,7 +153,10 @@ async def run_strix_scan(
         is_whitebox = any(t.get("type") == "local_code" for t in targets)
         skills = list(scan_config.get("skills") or [])
         root_task = build_root_task(scan_config)
-        model_settings = make_model_settings(settings.llm.reasoning_effort)
+        model_settings = make_model_settings(
+            settings.llm.reasoning_effort,
+            model_name=resolved_model,
+        )
         run_config = RunConfig(
             model=resolved_model,
             model_settings=model_settings,