feat(provider): add pipeline reasoning controls

tests/unit_tests/api/service/test_model_service.py

@@ -17,12 +17,14 @@ import pytest
 from unittest.mock import AsyncMock, Mock
 from types import SimpleNamespace
 
+from langbot.pkg.api.http.context import ExecutionContext
 from langbot.pkg.api.http.service.model import (
     LLMModelsService,
     EmbeddingModelsService,
     RerankModelsService,
     _parse_provider_api_keys,
     _runtime_model_data,
+    _serialize_llm_model,
     _validate_provider_supports,
 )
 from langbot.pkg.api.http.service import model as model_service_module
@@ -64,15 +66,19 @@ def _create_mock_llm_model(
     abilities: list = None,
     context_length: int | None = None,
     extra_args: dict = None,
+    reasoning_config: dict = None,
 ) -> Mock:
     """Helper to create mock LLMModel entity."""
     model = Mock(spec=LLMModel)
+    model.workspace_uuid = WORKSPACE_UUID
     model.uuid = model_uuid
     model.name = name
     model.provider_uuid = provider_uuid
     model.abilities = abilities or []
     model.context_length = context_length
     model.extra_args = extra_args or {}
+    model.reasoning_config = reasoning_config or {'level': 'provider_default'}
+    model.prefered_ranking = 0
     return model
 
 
@@ -156,6 +162,26 @@ def _create_runtime_model_mgr() -> SimpleNamespace:
     return manager
 
 
+def _create_reasoning_runtime_provider(capabilities: dict) -> SimpleNamespace:
+    execution_context = ExecutionContext(
+        instance_uuid='instance-test',
+        workspace_uuid=WORKSPACE_UUID,
+        placement_generation=1,
+    )
+    return SimpleNamespace(
+        execution_context=execution_context,
+        provider_entity=ModelProvider(
+            workspace_uuid=WORKSPACE_UUID,
+            uuid='provider-uuid',
+            name='Reasoning Provider',
+            requester='openai',
+            base_url='https://api.openai.com',
+            api_keys=[],
+        ),
+        requester=SimpleNamespace(get_reasoning_capabilities=Mock(return_value=capabilities)),
+    )
+
+
 class TestParseProviderApiKeys:
     """Tests for _parse_provider_api_keys helper function."""
 
@@ -209,6 +235,42 @@ class TestRuntimeModelData:
         assert result['extra_args'] == {'temp': 0.7}
 
 
+class TestSerializeLLMModel:
+    def test_includes_runtime_reasoning_capabilities(self):
+        model = _create_mock_llm_model(
+            abilities=['reasoning'],
+            reasoning_config={'level': 'high'},
+        )
+        capabilities = {
+            'supported': True,
+            'levels': ['provider_default', 'low', 'high'],
+            'source': 'litellm',
+        }
+        runtime_model = SimpleNamespace(
+            model_entity=model,
+            provider=SimpleNamespace(
+                requester=SimpleNamespace(get_reasoning_capabilities=Mock(return_value=capabilities))
+            ),
+        )
+        ap = SimpleNamespace(
+            persistence_mgr=SimpleNamespace(
+                serialize_model=Mock(
+                    return_value={
+                        'uuid': model.uuid,
+                        'name': model.name,
+                        'reasoning_config': {'level': 'high'},
+                    }
+                )
+            ),
+            model_mgr=SimpleNamespace(llm_model_dict={('workspace', model.uuid): runtime_model}),
+        )
+
+        serialized = _serialize_llm_model(ap, model)
+
+        assert serialized['reasoning_config'] == {'level': 'high'}
+        assert serialized['reasoning_capabilities'] == capabilities
+
+
 class TestLLMModelsServiceGetLLMModels:
     """Tests for LLMModelsService.get_llm_models method."""
 
@@ -580,6 +642,66 @@ class TestLLMModelsServiceCreateLLMModel:
         ap.provider_service.find_or_create_provider.assert_called_once()
         assert result_uuid is not None
 
+    async def test_create_llm_model_validates_explicit_reasoning_level(self):
+        ap = SimpleNamespace()
+        ap.persistence_mgr = SimpleNamespace(execute_async=AsyncMock(return_value=_create_mock_result([])))
+        runtime_provider = _create_reasoning_runtime_provider(
+            {
+                'supported': True,
+                'levels': ['provider_default', 'low', 'high'],
+                'source': 'litellm',
+            }
+        )
+        ap.model_mgr = _create_runtime_model_mgr()
+        ap.model_mgr.provider_dict = {'provider-uuid': runtime_provider}
+
+        service = LLMModelsService(ap)
+        await service.create_llm_model(
+            WORKSPACE_UUID,
+            {
+                'uuid': 'reasoning-model',
+                'name': 'Reasoning Model',
+                'provider_uuid': 'provider-uuid',
+                'abilities': ['reasoning'],
+                'reasoning_config': {'level': 'high'},
+                'extra_args': {},
+            },
+            preserve_uuid=True,
+            auto_set_to_default_pipeline=False,
+        )
+
+        runtime_entity = ap.model_mgr.load_llm_model_with_provider.await_args.args[1]
+        assert runtime_entity.reasoning_config == {'level': 'high'}
+
+    async def test_create_llm_model_rejects_unsupported_reasoning_before_insert(self):
+        ap = SimpleNamespace()
+        ap.persistence_mgr = SimpleNamespace(execute_async=AsyncMock())
+        runtime_provider = _create_reasoning_runtime_provider(
+            {
+                'supported': True,
+                'levels': ['provider_default'],
+                'source': 'manual',
+            }
+        )
+        ap.model_mgr = _create_runtime_model_mgr()
+        ap.model_mgr.provider_dict = {'provider-uuid': runtime_provider}
+
+        service = LLMModelsService(ap)
+        with pytest.raises(ValueError, match='Available levels: provider_default'):
+            await service.create_llm_model(
+                WORKSPACE_UUID,
+                {
+                    'name': 'Unknown Reasoning Model',
+                    'provider_uuid': 'provider-uuid',
+                    'abilities': ['reasoning'],
+                    'reasoning_config': {'level': 'high'},
+                    'extra_args': {},
+                },
+                auto_set_to_default_pipeline=False,
+            )
+
+        ap.persistence_mgr.execute_async.assert_not_awaited()
+
 
 class TestLLMModelsServiceUpdateLLMModel:
     """Tests for LLMModelsService.update_llm_model method."""
@@ -595,7 +717,10 @@ class TestLLMModelsServiceUpdateLLMModel:
         ap.model_mgr.remove_llm_model = AsyncMock()
         ap.model_mgr.load_llm_model_with_provider = AsyncMock(return_value=Mock())
 
-        ap.persistence_mgr.execute_async = AsyncMock()
+        existing_model = _create_mock_llm_model()
+        ap.persistence_mgr.execute_async = AsyncMock(
+            side_effect=[_create_mock_result(first_item=existing_model), _create_mock_result()]
+        )
 
         service = LLMModelsService(ap)
         service.get_llm_model = AsyncMock(return_value=_existing_llm_data())
@@ -623,7 +748,8 @@ class TestLLMModelsServiceUpdateLLMModel:
         ap.model_mgr.provider_dict = {}  # Empty
         ap.model_mgr.remove_llm_model = AsyncMock()
 
-        ap.persistence_mgr.execute_async = AsyncMock()
+        existing_model = _create_mock_llm_model()
+        ap.persistence_mgr.execute_async = AsyncMock(return_value=_create_mock_result(first_item=existing_model))
 
         service = LLMModelsService(ap)
         service.get_llm_model = AsyncMock(return_value=_existing_llm_data('nonexistent-provider'))

tests/unit_tests/provider/test_litellmchat.py

@@ -1304,6 +1304,7 @@ class TestScanModels:
         )
         requester._supports_function_calling = Mock(side_effect=lambda model_id: model_id == 'gpt-4o')
         requester._supports_vision = Mock(side_effect=lambda model_id: model_id == 'gpt-4o')
+        requester._supports_reasoning = Mock(side_effect=lambda model_id: model_id == 'o3')
         requester._safe_context_length = Mock(side_effect=lambda model_id: 128000 if model_id == 'gpt-4o' else None)
 
         mock_response = Mock()
@@ -1311,6 +1312,7 @@ class TestScanModels:
             return_value={
                 'data': [
                     {'id': 'gpt-4o'},
+                    {'id': 'o3'},
                     {'id': 'text-embedding-3-small'},
                     {'id': 'bge-reranker-v2'},
                 ]
@@ -1327,6 +1329,7 @@ class TestScanModels:
         by_id = {model['id']: model for model in result['models']}
         assert by_id['gpt-4o']['abilities'] == ['func_call', 'vision']
         assert by_id['gpt-4o']['context_length'] == 128000
+        assert by_id['o3']['abilities'] == ['reasoning']
         assert by_id['text-embedding-3-small']['type'] == 'embedding'
         assert by_id['bge-reranker-v2']['type'] == 'rerank'
 
@@ -1374,8 +1377,8 @@ class TestScanModels:
         )
 
         with patch.object(litellmchat.litellm, 'get_model_info') as mock_get_model_info:
-            mock_get_model_info.side_effect = (
-                lambda model: {'max_input_tokens': 131072} if model == 'moonshot/moonshot-v1-128k' else {}
+            mock_get_model_info.side_effect = lambda model: (
+                {'max_input_tokens': 131072} if model == 'moonshot/moonshot-v1-128k' else {}
             )
 
             assert requester._safe_context_length('moonshot-v1-128k') == 131072
@@ -1404,8 +1407,8 @@ class TestScanModels:
         )
 
         with patch.object(litellmchat.litellm, 'supports_function_calling') as mock_supports_function_calling:
-            mock_supports_function_calling.side_effect = (
-                lambda model, custom_llm_provider=None: model == 'moonshot/kimi-k2.6' and custom_llm_provider is None
+            mock_supports_function_calling.side_effect = lambda model, custom_llm_provider=None: (
+                model == 'moonshot/kimi-k2.6' and custom_llm_provider is None
             )
 
             assert requester._supports_function_calling('kimi-k2.6') is True

tests/unit_tests/provider/test_model_service.py

@@ -249,7 +249,11 @@ async def test_updated_llm_model_is_immediately_usable_by_local_agent_pipeline()
         'ai': {
             'runner': {'runner': 'local-agent'},
             'local-agent': {
-                'model': {'primary': model_uuid, 'fallbacks': []},
+                'model': {
+                    'primary': model_uuid,
+                    'fallbacks': [],
+                    'reasoning': {model_uuid: 'high'},
+                },
                 'prompt': [],
                 'knowledge-bases': [],
             },
@@ -293,3 +297,130 @@ async def test_updated_llm_model_is_immediately_usable_by_local_agent_pipeline()
     candidates = await LocalAgentRunner._get_model_candidates(runner, processed_query)
 
     assert [model.model_entity.uuid for model in candidates] == [model_uuid]
+    assert candidates[0].reasoning_config_override == {'level': 'high'}
+
+
+@pytest.mark.asyncio
+async def test_local_agent_applies_reasoning_per_fallback_model():
+    execution_context = ExecutionContext(
+        instance_uuid='instance-test',
+        workspace_uuid='workspace-test',
+        placement_generation=1,
+    )
+    provider = Mock(
+        execution_context=execution_context,
+        provider_entity=persistence_model.ModelProvider(
+            workspace_uuid='workspace-test',
+            uuid='provider',
+            name='provider',
+            requester='openai',
+            base_url='https://example.com',
+            api_keys=[],
+        ),
+    )
+    primary = requester.RuntimeLLMModel(
+        execution_context,
+        persistence_model.LLMModel(
+            workspace_uuid='workspace-test',
+            uuid='primary-model',
+            name='primary',
+            provider_uuid='provider',
+            abilities=['reasoning'],
+            extra_args={},
+        ),
+        provider,
+    )
+    fallback = requester.RuntimeLLMModel(
+        execution_context,
+        persistence_model.LLMModel(
+            workspace_uuid='workspace-test',
+            uuid='fallback-model',
+            name='fallback',
+            provider_uuid='provider',
+            abilities=['reasoning'],
+            extra_args={},
+        ),
+        provider,
+    )
+    models = {'primary-model': primary, 'fallback-model': fallback}
+    runner = SimpleNamespace(
+        ap=SimpleNamespace(
+            model_mgr=SimpleNamespace(
+                get_model_by_uuid=AsyncMock(side_effect=lambda _context, model_uuid: models[model_uuid]),
+            ),
+            logger=Mock(),
+        )
+    )
+    query = SimpleNamespace(
+        use_llm_model_uuid='primary-model',
+        variables={'_fallback_model_uuids': ['fallback-model']},
+        pipeline_config={
+            'ai': {
+                'local-agent': {
+                    'model': {
+                        'primary': 'primary-model',
+                        'fallbacks': ['fallback-model'],
+                        'reasoning': {
+                            'primary-model': 'low',
+                            'fallback-model': 'high',
+                        },
+                    }
+                }
+            }
+        },
+        _execution_context=execution_context,
+    )
+
+    candidates = await LocalAgentRunner._get_model_candidates(runner, query)
+
+    assert [candidate.reasoning_config_override for candidate in candidates] == [
+        {'level': 'low'},
+        {'level': 'high'},
+    ]
+
+
+def test_local_agent_rejects_invalid_pipeline_reasoning_level():
+    execution_context = ExecutionContext(
+        instance_uuid='instance-test',
+        workspace_uuid='workspace-test',
+        placement_generation=1,
+    )
+    provider = Mock(
+        execution_context=execution_context,
+        provider_entity=persistence_model.ModelProvider(
+            workspace_uuid='workspace-test',
+            uuid='provider',
+            name='provider',
+            requester='openai',
+            base_url='https://example.com',
+            api_keys=[],
+        ),
+    )
+    model = requester.RuntimeLLMModel(
+        execution_context,
+        persistence_model.LLMModel(
+            workspace_uuid='workspace-test',
+            uuid='primary-model',
+            name='primary',
+            provider_uuid='provider',
+            abilities=['reasoning'],
+            extra_args={},
+        ),
+        provider,
+    )
+    query = SimpleNamespace(
+        pipeline_config={
+            'ai': {
+                'local-agent': {
+                    'model': {
+                        'primary': 'primary-model',
+                        'fallbacks': [],
+                        'reasoning': {'primary-model': 'turbo'},
+                    }
+                }
+            }
+        }
+    )
+
+    with pytest.raises(ValueError, match='Unsupported reasoning level'):
+        LocalAgentRunner._apply_pipeline_reasoning_config(query, model)

tests/unit_tests/provider/test_reasoning_control.py

@@ -0,0 +1,408 @@
+from __future__ import annotations
+
+from types import SimpleNamespace
+from unittest.mock import AsyncMock
+
+import pytest
+
+import langbot_plugin.api.entities.builtin.provider.message as provider_message
+from langbot.pkg.api.http.context import ExecutionContext
+from langbot.pkg.entity.persistence import model as persistence_model
+from langbot.pkg.provider.modelmgr import errors, reasoning, requester
+from langbot.pkg.provider.modelmgr.requesters import litellmchat
+from langbot.pkg.provider.modelmgr.requesters.litellmchat import LiteLLMRequester
+from langbot.pkg.provider.runners.localagent import _StreamAccumulator
+
+
+def _runtime_model(
+    request: LiteLLMRequester,
+    level: str = 'provider_default',
+    name: str = 'reasoning-model',
+    abilities: list[str] | None = None,
+) -> requester.RuntimeLLMModel:
+    execution_context = ExecutionContext(
+        instance_uuid='instance-test',
+        workspace_uuid='workspace-test',
+        placement_generation=1,
+    )
+    entity = persistence_model.LLMModel(
+        workspace_uuid='workspace-test',
+        uuid='reasoning-model',
+        name=name,
+        provider_uuid='provider-test',
+        abilities=abilities if abilities is not None else ['reasoning'],
+        reasoning_config={'level': level},
+        extra_args={},
+    )
+    provider = SimpleNamespace(
+        execution_context=execution_context,
+        provider_entity=persistence_model.ModelProvider(
+            workspace_uuid='workspace-test',
+            uuid='provider-test',
+            name='provider',
+            requester='openai',
+            base_url='https://example.com',
+            api_keys=[],
+        ),
+        requester=request,
+        token_mgr=SimpleNamespace(),
+    )
+    return requester.RuntimeLLMModel(execution_context, entity, provider)
+
+
+def _requester(provider: str = '') -> LiteLLMRequester:
+    return LiteLLMRequester(SimpleNamespace(), {'custom_llm_provider': provider})
+
+
+def test_reasoning_config_normalization_and_conflicts():
+    assert reasoning.normalize_reasoning_config(None) == {'level': 'provider_default'}
+    assert reasoning.normalize_reasoning_config({}) == {'level': 'provider_default'}
+    assert reasoning.validate_reasoning_config(
+        {'level': 'high'},
+        ['reasoning'],
+        {},
+    ) == {'level': 'high'}
+
+    with pytest.raises(ValueError, match='Unsupported reasoning level'):
+        reasoning.normalize_reasoning_config({'level': 'turbo'})
+    with pytest.raises(ValueError, match='reasoning ability'):
+        reasoning.validate_reasoning_config({'level': 'low'}, [], {})
+    with pytest.raises(ValueError, match='extra_body.thinking_budget'):
+        reasoning.validate_reasoning_config(
+            {'level': 'low'},
+            ['reasoning'],
+            {'extra_body': {'thinking_budget': 1024}},
+        )
+
+
+def test_manual_reasoning_model_exposes_conservative_effort_levels(monkeypatch):
+    request = _requester()
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request))
+
+    assert capabilities == {
+        'supported': True,
+        'levels': ['provider_default', 'low', 'medium', 'high'],
+        'source': 'manual',
+    }
+
+
+def test_provider_protocol_exposes_reasoning_for_unknown_model(monkeypatch):
+    request = _requester('openai')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: pytest.fail('metadata should not be queried'))
+
+    capabilities = request.get_reasoning_capabilities(
+        _runtime_model(request, name='future-reasoning-model', abilities=[])
+    )
+
+    assert capabilities == {
+        'supported': True,
+        'levels': ['provider_default', 'low', 'medium', 'high'],
+        'source': 'provider',
+    }
+
+
+def test_unknown_unmarked_model_without_provider_stays_safe(monkeypatch):
+    request = _requester()
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(
+        _runtime_model(request, name='unknown-model', abilities=[])
+    )
+
+    assert capabilities == {
+        'supported': False,
+        'levels': ['provider_default'],
+        'source': 'unknown',
+    }
+
+
+def test_mimo_native_model_uses_known_equivalent_litellm_metadata(monkeypatch):
+    request = _requester('openai')
+
+    def supports_reasoning(model: str, custom_llm_provider: str | None = None) -> bool:
+        return model == 'openrouter/xiaomi/mimo-v2.5'
+
+    def get_model_info(model: str) -> dict:
+        if model == 'openrouter/xiaomi/mimo-v2.5':
+            return {'supports_reasoning': True}
+        raise ValueError('unknown model')
+
+    monkeypatch.setattr(litellmchat.litellm, 'supports_reasoning', supports_reasoning)
+    monkeypatch.setattr(litellmchat.litellm, 'get_model_info', get_model_info)
+
+    capabilities = request.get_reasoning_capabilities(
+        _runtime_model(request, name='mimo-v2.5', abilities=[])
+    )
+
+    assert capabilities == {
+        'supported': True,
+        'levels': ['provider_default', 'minimal', 'low', 'medium', 'high'],
+        'source': 'litellm',
+    }
+
+
+def test_openai_reasoning_levels_follow_litellm_metadata(monkeypatch):
+    request = _requester('openai')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(
+        request,
+        '_safe_model_info',
+        lambda _: {
+            'supports_none_reasoning_effort': True,
+            'supports_minimal_reasoning_effort': False,
+            'supports_low_reasoning_effort': True,
+            'supports_xhigh_reasoning_effort': True,
+        },
+    )
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name='gpt-5'))
+
+    assert capabilities['source'] == 'litellm'
+    assert capabilities['levels'] == [
+        'provider_default',
+        'disabled',
+        'low',
+        'medium',
+        'high',
+        'xhigh',
+    ]
+
+
+def test_reasoning_argument_translation(monkeypatch):
+    openai_request = _requester('openai')
+    monkeypatch.setattr(openai_request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(
+        openai_request,
+        '_safe_model_info',
+        lambda _: {'supports_none_reasoning_effort': True},
+    )
+
+    assert openai_request._build_reasoning_args(_runtime_model(openai_request, 'provider_default')) == {}
+    assert openai_request._build_reasoning_args(_runtime_model(openai_request, 'disabled')) == {
+        'reasoning_effort': 'none'
+    }
+    assert openai_request._build_reasoning_args(_runtime_model(openai_request, 'high')) == {'reasoning_effort': 'high'}
+
+    deepseek_request = _requester('deepseek')
+    monkeypatch.setattr(deepseek_request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(deepseek_request, '_safe_model_info', lambda _: {})
+    assert deepseek_request._build_reasoning_args(
+        _runtime_model(deepseek_request, 'enabled', name='deepseek-chat')
+    ) == {'thinking': {'type': 'enabled'}}
+
+
+def test_pipeline_reasoning_override_takes_precedence(monkeypatch):
+    request = _requester('openai')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+    model = _runtime_model(request, 'high', name='gpt-5')
+
+    model.reasoning_config_override = {'level': 'provider_default'}
+    assert request._build_reasoning_args(model) == {}
+
+    model.reasoning_config_override = {'level': 'low'}
+    assert request._build_reasoning_args(model) == {'reasoning_effort': 'low'}
+
+
+def test_deepseek_provider_is_inferred_from_model_name(monkeypatch):
+    request = _requester()
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name='deepseek-chat'))
+
+    assert capabilities['levels'] == ['provider_default', 'disabled', 'enabled']
+
+
+def test_always_on_reasoning_models_do_not_offer_disabled(monkeypatch):
+    deepseek_request = _requester('deepseek')
+    monkeypatch.setattr(deepseek_request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(deepseek_request, '_safe_model_info', lambda _: {})
+    deepseek_capabilities = deepseek_request.get_reasoning_capabilities(
+        _runtime_model(deepseek_request, name='deepseek-r1')
+    )
+    assert deepseek_capabilities['levels'] == ['provider_default', 'enabled']
+
+    gemini_request = _requester('gemini')
+    monkeypatch.setattr(gemini_request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(
+        gemini_request,
+        '_safe_model_info',
+        lambda _: {'supports_none_reasoning_effort': True},
+    )
+    gemini_capabilities = gemini_request.get_reasoning_capabilities(_runtime_model(gemini_request, name='gemini-3-pro'))
+    assert 'disabled' not in gemini_capabilities['levels']
+    with pytest.raises(errors.RequesterError, match='not supported'):
+        gemini_request._build_reasoning_args(_runtime_model(gemini_request, 'disabled', name='gemini-3-pro'))
+
+
+def test_toggle_and_effort_provider_capabilities(monkeypatch):
+    ollama_request = _requester('ollama')
+    monkeypatch.setattr(ollama_request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(ollama_request, '_safe_model_info', lambda _: {})
+
+    toggle_capabilities = ollama_request.get_reasoning_capabilities(_runtime_model(ollama_request, name='qwen3'))
+    assert toggle_capabilities['levels'] == [
+        'provider_default',
+        'disabled',
+        'enabled',
+    ]
+    assert ollama_request._build_reasoning_args(_runtime_model(ollama_request, 'enabled', name='qwen3')) == {
+        'reasoning_effort': 'low'
+    }
+
+    effort_capabilities = ollama_request.get_reasoning_capabilities(_runtime_model(ollama_request, name='gpt-oss:20b'))
+    assert effort_capabilities['levels'] == [
+        'provider_default',
+        'disabled',
+        'low',
+        'medium',
+        'high',
+    ]
+    assert ollama_request._build_reasoning_args(_runtime_model(ollama_request, 'high', name='gpt-oss:20b')) == {
+        'reasoning_effort': 'high'
+    }
+
+    volcengine_request = _requester('volcengine')
+    monkeypatch.setattr(volcengine_request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(volcengine_request, '_safe_model_info', lambda _: {})
+    assert volcengine_request._build_reasoning_args(
+        _runtime_model(volcengine_request, 'disabled', name='doubao-seed')
+    ) == {'thinking': {'type': 'disabled'}}
+
+
+def test_explicit_unsupported_level_raises(monkeypatch):
+    request = _requester()
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+
+    with pytest.raises(errors.RequesterError, match='Available levels: provider_default'):
+        request._build_reasoning_args(_runtime_model(request, 'high', abilities=[]))
+
+
+def test_provider_inference_rejects_levels_outside_conservative_profile(monkeypatch):
+    request = _requester('openai')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    with pytest.raises(errors.RequesterError, match='Available levels: provider_default, low, medium, high'):
+        request._build_reasoning_args(_runtime_model(request, 'xhigh', abilities=[]))
+
+
+@pytest.mark.asyncio
+async def test_completion_args_reject_reasoning_extra_arg_conflicts(monkeypatch):
+    request = _requester('openai')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+    model = _runtime_model(request, 'high')
+    model.model_entity.extra_args = {'reasoning_effort': 'low'}
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+
+    with pytest.raises(errors.RequesterError, match='conflicts with advanced parameters'):
+        await request._build_completion_args(model, [])
+
+
+@pytest.mark.asyncio
+async def test_openai_compatible_reasoning_effort_is_explicitly_allowed(monkeypatch):
+    request = _requester('openai')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+    model = _runtime_model(request, 'high', name='deepseek-v4-flash')
+    model.model_entity.extra_args = {'allowed_openai_params': ['custom_extension']}
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+
+    args = await request._build_completion_args(model, [])
+
+    assert args['reasoning_effort'] == 'high'
+    assert args['allowed_openai_params'] == ['custom_extension', 'reasoning_effort']
+
+
+@pytest.mark.asyncio
+async def test_provider_default_does_not_allow_or_send_reasoning_effort():
+    request = _requester('openai')
+    model = _runtime_model(request, 'provider_default', name='deepseek-v4-flash')
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+
+    args = await request._build_completion_args(model, [])
+
+    assert 'reasoning_effort' not in args
+    assert 'allowed_openai_params' not in args
+
+
+class _Dumpable:
+    def __init__(self, data: dict):
+        self.data = data
+
+    def model_dump(self) -> dict:
+        return dict(self.data)
+
+
+@pytest.mark.asyncio
+async def test_non_stream_reasoning_content_is_preserved(monkeypatch):
+    request = _requester('deepseek')
+    request._build_completion_args = AsyncMock(return_value={})
+    response = SimpleNamespace(
+        choices=[
+            SimpleNamespace(
+                message=_Dumpable(
+                    {
+                        'role': 'assistant',
+                        'content': 'answer',
+                        'reasoning_content': 'private reasoning',
+                    }
+                )
+            )
+        ],
+        usage=None,
+    )
+    monkeypatch.setattr(litellmchat, 'acompletion', AsyncMock(return_value=response))
+
+    message, _ = await request.invoke_llm(None, _runtime_model(request), [], remove_think=True)
+
+    assert message.content == 'answer'
+    assert message.provider_specific_fields == {'reasoning_content': 'private reasoning'}
+
+
+@pytest.mark.asyncio
+async def test_stream_reasoning_round_trip_with_hidden_display(monkeypatch):
+    request = _requester('deepseek')
+    request._build_completion_args = AsyncMock(return_value={})
+
+    async def chunks():
+        yield SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=_Dumpable({'role': 'assistant', 'reasoning_content': 'private '}),
+                    finish_reason=None,
+                )
+            ],
+            usage=None,
+        )
+        yield SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=_Dumpable({'content': 'answer'}),
+                    finish_reason='stop',
+                )
+            ],
+            usage=None,
+        )
+
+    monkeypatch.setattr(litellmchat, 'acompletion', AsyncMock(return_value=chunks()))
+    accumulator = _StreamAccumulator(remove_think=True)
+    emitted: provider_message.MessageChunk | None = None
+
+    async for chunk in request.invoke_llm_stream(
+        None,
+        _runtime_model(request),
+        [],
+        remove_think=True,
+    ):
+        emitted = accumulator.add(chunk) or emitted
+
+    assert emitted is not None
+    assert emitted.content == 'answer'
+    assert emitted.provider_specific_fields == {'reasoning_content': 'private '}

tests/unit_tests/provider/test_requester_base.py

@@ -400,6 +400,7 @@ def test_runtime_llm_model_initialization(runtime_llm_model, fake_persistence_da
     assert model.model_entity.abilities == model_entity.abilities
     assert model.model_entity.extra_args == model_entity.extra_args
     assert model.provider is not None
+    assert model.reasoning_config_override is None
 
 
 def test_runtime_llm_model_provider_ref(runtime_llm_model):