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.

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,
         ),