fix(toolchoice): default required tool choice on OpenAI-compatible custom endpoints

Reasoning models on OpenAI-compatible custom endpoints (e.g. GLM / Kimi)
often reason and reply with prose instead of emitting native tool calls
under the default auto tool choice, stalling the tool-driven scan loop
while burning tokens. Default force_required_tool_choice to auto (None):
enable tool_choice=required when an api_base is set (still gated to
models that accept it), off otherwise. Explicit 0/1 overrides.

strix/config/settings.py

@@ -36,8 +36,11 @@ class LlmSettings(BaseSettings):
         ),
     )
     reasoning_effort: ReasoningEffort = Field(default="high", alias="STRIX_REASONING_EFFORT")
-    force_required_tool_choice: bool = Field(
-        default=False,
+    # None = auto: force required tool choice on OpenAI-compatible custom
+    # endpoints (where reasoning models otherwise burn tokens thinking without
+    # ever calling a tool), off elsewhere. True/False overrides explicitly.
+    force_required_tool_choice: bool | None = Field(
+        default=None,
         alias="STRIX_FORCE_REQUIRED_TOOL_CHOICE",
     )
     prompt_cache: bool = Field(

strix/core/inputs.py

@@ -129,7 +129,8 @@ def make_model_settings(
     reasoning_effort: ReasoningEffort | None,
     *,
     model_name: str,
-    force_required_tool_choice: bool = False,
+    force_required_tool_choice: bool | None = None,
+    custom_api_base: bool = False,
     request_timeout: float | None = None,
     prompt_cache: bool = True,
 ) -> ModelSettings:
@@ -147,7 +148,14 @@ def make_model_settings(
         model_settings = model_settings.resolve(
             ModelSettings(reasoning=Reasoning(effort=reasoning_effort)),
         )
-    if force_required_tool_choice and _accepts_required_tool_choice(model_name):
+    # Strix is fully tool-driven, so a turn that returns prose instead of a tool
+    # call stalls the scan. Reasoning models on OpenAI-compatible custom
+    # endpoints (e.g. GLM / Kimi via cortecs) do exactly that under the default
+    # ``auto`` tool choice, so default to ``required`` there; ``None`` means auto.
+    use_required = (
+        custom_api_base if force_required_tool_choice is None else force_required_tool_choice
+    )
+    if use_required and _accepts_required_tool_choice(model_name):
         model_settings = model_settings.resolve(ModelSettings(tool_choice="required"))
 
     cache_extra_args = _prompt_cache_extra_args(model_name) if prompt_cache else None

strix/core/runner.py

@@ -248,6 +248,7 @@ async def run_strix_scan(
             settings.llm.reasoning_effort,
             model_name=resolved_model,
             force_required_tool_choice=settings.llm.force_required_tool_choice,
+            custom_api_base=bool(settings.llm.api_base),
             request_timeout=settings.llm.timeout,
             prompt_cache=settings.llm.prompt_cache,
         )

tests/test_inputs.py

@@ -255,6 +255,44 @@ def test_make_model_settings_forces_required_for_anyllm_routed_openai_model() ->
     assert settings.tool_choice == "required"
 
 
+def test_make_model_settings_auto_forces_required_on_custom_openai_endpoint() -> None:
+    # GLM / Kimi via cortecs: openai/-routed reasoning model on a custom base.
+    settings = make_model_settings(
+        None,
+        model_name="openai/glm-5.2",
+        custom_api_base=True,
+    )
+
+    assert settings.tool_choice == "required"
+
+
+def test_make_model_settings_auto_skips_required_without_custom_endpoint() -> None:
+    settings = make_model_settings(None, model_name="openai/glm-5.2")
+
+    assert settings.tool_choice is None
+
+
+def test_make_model_settings_explicit_false_overrides_custom_endpoint() -> None:
+    settings = make_model_settings(
+        None,
+        model_name="openai/glm-5.2",
+        force_required_tool_choice=False,
+        custom_api_base=True,
+    )
+
+    assert settings.tool_choice is None
+
+
+def test_make_model_settings_auto_skips_required_for_non_openai_custom_endpoint() -> None:
+    settings = make_model_settings(
+        None,
+        model_name="anthropic/claude-3-7-sonnet-latest",
+        custom_api_base=True,
+    )
+
+    assert settings.tool_choice is None
+
+
 def test_make_model_settings_sets_request_timeout() -> None:
     settings = make_model_settings(
         "none",

tests/test_runner_rate_limit.py

@@ -38,6 +38,7 @@ async def test_persistent_rate_limit_stops_gracefully(
             model="openai/gpt-4o",
             reasoning_effort="high",
             force_required_tool_choice=False,
+            api_base=None,
             timeout=300,
             prompt_cache=True,
         ),

tests/test_runner_root_prompt.py

@@ -46,6 +46,7 @@ def _patch_engine_scaffold(
             model="openai/gpt-4o",
             reasoning_effort="high",
             force_required_tool_choice=False,
+            api_base=None,
             timeout=300,
             prompt_cache=True,
         ),