fix(provider): route reasoning controls by model family

src/langbot/pkg/provider/modelmgr/modelmgr.py

@@ -718,7 +718,10 @@ class ModelManager:
         provider_entity = self._coerce_provider(provider_info, context)
         requester_manifest = self.get_available_requester_manifest_by_name(provider_entity.requester)
         litellm_provider = self._get_litellm_provider_from_manifest(requester_manifest)
-        config = {'base_url': provider_entity.base_url}
+        config = {
+            'base_url': provider_entity.base_url,
+            'requester_name': provider_entity.requester,
+        }
 
         if litellm_provider:
             from .requesters import litellmchat

src/langbot/pkg/provider/modelmgr/reasoning.py

@@ -28,8 +28,20 @@ REASONING_LEVELS: tuple[str, ...] = (
 )
 DEFAULT_REASONING_CONFIG: dict[str, str] = {'level': 'provider_default'}
 
-_CONFLICTING_TOP_LEVEL_ARGS = {'reasoning_effort', 'thinking', 'reasoning'}
-_CONFLICTING_EXTRA_BODY_ARGS = {'thinking', 'enable_thinking', 'thinking_budget', 'reasoning'}
+_CONFLICTING_TOP_LEVEL_ARGS = {
+    'reasoning_effort',
+    'thinking',
+    'enable_thinking',
+    'thinking_budget',
+    'reasoning',
+}
+_CONFLICTING_EXTRA_BODY_ARGS = {
+    'reasoning_effort',
+    'thinking',
+    'enable_thinking',
+    'thinking_budget',
+    'reasoning',
+}
 
 
 def normalize_reasoning_config(value: typing.Any) -> dict[str, str]:

src/langbot/pkg/provider/modelmgr/requesters/litellmchat.py

@@ -176,7 +176,16 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
             'xai',
         }
     )
-    _INFERRED_TOGGLE_PROVIDERS = frozenset({'deepseek', 'ollama', 'volcengine'})
+    _REQUESTER_REASONING_FAMILIES = {
+        'openai-chat-completions': 'openai',
+        'anthropic-messages': 'anthropic',
+        'deepseek-chat-completions': 'deepseek',
+        'moonshot-chat-completions': 'kimi',
+        'moonshot-cn-chat-completions': 'kimi',
+        'bailian-chat-completions': 'qwen',
+        'doubao-chat-completions': 'doubao',
+        'mimo-chat-completions': 'mimo',
+    }
 
     default_config: dict[str, typing.Any] = {
         'base_url': '',
@@ -185,6 +194,7 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
         'drop_params': False,
         'num_retries': 0,
         'api_version': '',
+        'requester_name': '',
     }
 
     async def initialize(self):
@@ -342,31 +352,144 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
     def _supports_reasoning(self, model_name: str) -> bool:
         return self._safe_litellm_bool_helper('supports_reasoning', model_name)
 
-    def _reasoning_provider(self, model_name: str) -> str:
+    def _requester_name(self, model: requester.RuntimeLLMModel | None = None) -> str:
+        if model is not None:
+            provider_entity = getattr(getattr(model, 'provider', None), 'provider_entity', None)
+            name = getattr(provider_entity, 'requester', None)
+            if isinstance(name, str) and name:
+                return name.lower()
+        return str(self.requester_cfg.get('requester_name') or '').lower()
+
+    @staticmethod
+    def _infer_reasoning_family_from_model_name(model_name: str) -> str:
+        normalized_name = (model_name or '').lower()
+        basename = normalized_name.rsplit('/', 1)[-1]
+        if basename.startswith(('gpt-', 'chatgpt-', 'o1', 'o3', 'o4')):
+            return 'openai'
+        if basename.startswith('claude-'):
+            return 'anthropic'
+        if basename.startswith('deepseek-'):
+            return 'deepseek'
+        if basename.startswith(('kimi-', 'moonshot-')):
+            return 'kimi'
+        if basename.startswith(('qwen-', 'qwen3', 'qwq')):
+            return 'qwen'
+        if basename.startswith(('doubao-', 'seed-')):
+            return 'doubao'
+        if basename.startswith('mimo-'):
+            return 'mimo'
+        return ''
+
+    def _reasoning_family(
+        self,
+        model_name: str,
+        model: requester.RuntimeLLMModel | None = None,
+    ) -> str:
+        requester_name = self._requester_name(model)
+        if requester_name in {'new-api-chat-completions', 'volcark-chat-completions'}:
+            inferred_family = self._infer_reasoning_family_from_model_name(model_name)
+            if inferred_family:
+                return inferred_family
+            return 'volcengine' if requester_name == 'volcark-chat-completions' else ''
+
+        requester_family = self._REQUESTER_REASONING_FAMILIES.get(requester_name)
+        if requester_family:
+            return requester_family
+
+        inferred_family = self._infer_reasoning_family_from_model_name(model_name)
         provider = (self._get_custom_llm_provider() or '').lower()
+        if provider == 'openai':
+            return inferred_family or ('openai' if requester_name in {'', 'openai'} else '')
         if provider:
             return provider
+        return inferred_family
 
-        normalized_name = (model_name or '').lower()
-        if '/' in normalized_name:
-            prefix = normalized_name.split('/', 1)[0]
-            if prefix in {
-                'anthropic',
-                'deepseek',
-                'gemini',
-                'groq',
-                'mistral',
-                'ollama',
-                'openai',
-                'openrouter',
-                'together_ai',
-                'volcengine',
-                'xai',
-            }:
-                return prefix
+    @staticmethod
+    def _is_anthropic_adaptive_model(model_name: str) -> bool:
+        basename = model_name.lower().rsplit('/', 1)[-1]
+        if 'mythos-preview' in basename:
+            return True
 
-        candidates = self._metadata_provider_candidates(model_name)
-        return candidates[0] if candidates else ''
+        parts = basename.split('-')
+        if len(parts) < 3 or parts[0] != 'claude':
+            return False
+        model_families = {'opus', 'sonnet', 'fable', 'mythos'}
+        if parts[1] in model_families:
+            if parts[2] == '5':
+                return True
+            return len(parts) >= 4 and parts[2] == '4' and parts[3] in {'6', '7', '8'}
+        return parts[1] == '5' and parts[2] in model_families
+
+    @staticmethod
+    def _is_anthropic_always_thinking_model(model_name: str) -> bool:
+        normalized_name = model_name.lower()
+        return any(marker in normalized_name for marker in ('fable-5', 'mythos-5', 'mythos-preview'))
+
+    @staticmethod
+    def _is_dedicated_qwen_thinking_model(model_name: str) -> bool:
+        normalized_name = model_name.lower().rsplit('/', 1)[-1]
+        return normalized_name.startswith('qwq') or '-thinking' in normalized_name
+
+    def _known_reasoning_levels(self, model_name: str, family: str) -> list[str] | None:
+        normalized_name = model_name.lower().rsplit('/', 1)[-1]
+
+        if family == 'deepseek' and normalized_name.startswith('deepseek-'):
+            if normalized_name.startswith('deepseek-v4-'):
+                return ['provider_default', 'disabled', 'low', 'high', 'xhigh', 'max']
+            if 'reasoner' in normalized_name or '-r1' in normalized_name:
+                return ['provider_default']
+            return ['provider_default', 'disabled', 'enabled']
+
+        if family == 'kimi':
+            if normalized_name.startswith('kimi-k3'):
+                return ['provider_default', 'low', 'high', 'max']
+            if normalized_name.startswith('kimi-k2.7-code'):
+                return ['provider_default']
+            if normalized_name.startswith(('kimi-k2.5', 'kimi-k2.6')):
+                return ['provider_default', 'disabled', 'enabled']
+            if 'thinking' in normalized_name:
+                return ['provider_default']
+
+        if family == 'qwen' and normalized_name.startswith(('qwen-', 'qwen3', 'qwq')):
+            if self._is_dedicated_qwen_thinking_model(normalized_name):
+                return ['provider_default']
+            return ['provider_default', 'disabled', 'enabled']
+
+        if family == 'doubao' and normalized_name.startswith(('doubao-', 'seed-')):
+            return ['provider_default', 'disabled', 'low', 'medium', 'high']
+
+        if family == 'mimo' and normalized_name.startswith(('mimo-v2.5',)):
+            return ['provider_default', 'disabled', 'enabled']
+
+        if family == 'anthropic' and normalized_name.startswith('claude-'):
+            levels = ['provider_default']
+            adaptive = self._is_anthropic_adaptive_model(normalized_name)
+            if adaptive and not self._is_anthropic_always_thinking_model(normalized_name):
+                levels.append('disabled')
+            levels.extend(['low', 'medium', 'high'])
+            if adaptive:
+                levels.extend(['xhigh', 'max'])
+            return levels
+
+        if family == 'openai' and normalized_name.startswith(('gpt-5', 'o1', 'o3', 'o4')):
+            return ['provider_default', 'low', 'medium', 'high']
+
+        return None
+
+    def _openai_reasoning_levels(self, model_name: str) -> list[str]:
+        model_info = self._safe_model_info(model_name)
+        levels = ['provider_default']
+        if model_info.get('supports_none_reasoning_effort') is True:
+            levels.append('disabled')
+        if model_info.get('supports_minimal_reasoning_effort') is True:
+            levels.append('minimal')
+        for level in ('low', 'medium', 'high'):
+            if model_info.get(f'supports_{level}_reasoning_effort') is not False:
+                levels.append(level)
+        for level in ('xhigh', 'max'):
+            if model_info.get(f'supports_{level}_reasoning_effort') is True:
+                levels.append(level)
+        return levels
 
     def _safe_model_info(self, model_name: str) -> dict[str, typing.Any]:
         helper = getattr(litellm, 'get_model_info', None)
@@ -400,83 +523,91 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
         abilities = model.model_entity.abilities or []
         detected = self._supports_reasoning(model_name)
         declared = 'reasoning' in abilities
-        provider = self._reasoning_provider(model_name)
-        inferred = provider in self._INFERRED_EFFORT_PROVIDERS | self._INFERRED_TOGGLE_PROVIDERS
-        supported = detected or declared or inferred
+        family = self._reasoning_family(model_name, model)
+        known_levels = self._known_reasoning_levels(model_name, family)
+        supported = detected or declared or known_levels is not None
         if not supported:
             return reasoning.default_reasoning_capabilities()
 
         normalized_name = model_name.lower()
-        levels = ['provider_default']
-
-        if provider == 'deepseek':
-            if 'reasoner' not in normalized_name and '-r1' not in normalized_name:
-                levels.append('disabled')
-            levels.append('enabled')
-        elif provider == 'volcengine':
-            levels.extend(['disabled', 'enabled'])
-        elif provider == 'ollama':
+        if family == 'openai':
+            levels = self._openai_reasoning_levels(model_name)
+        elif known_levels is not None:
+            levels = known_levels
+        elif family == 'anthropic':
+            levels = ['provider_default', 'low', 'medium', 'high']
+        elif family in {'deepseek', 'qwen', 'mimo', 'volcengine'}:
+            levels = ['provider_default', 'disabled', 'enabled']
+        elif family == 'doubao':
+            levels = ['provider_default', 'disabled', 'low', 'medium', 'high']
+        elif family == 'ollama':
+            levels = ['provider_default']
             levels.append('disabled')
             if normalized_name.startswith('gpt-oss') or '/gpt-oss' in normalized_name:
                 levels.extend(['low', 'medium', 'high'])
             else:
                 levels.append('enabled')
-        elif not detected:
-            levels.extend(['low', 'medium', 'high'])
+        elif family in self._INFERRED_EFFORT_PROVIDERS:
+            levels = ['provider_default', 'low', 'medium', 'high']
         else:
-            model_info = self._safe_model_info(model_name)
-            supports_none = model_info.get('supports_none_reasoning_effort') is True
-            if provider == 'anthropic':
-                supports_none = True
-            if provider == 'gemini' and 'gemini-3' in normalized_name:
-                supports_none = False
-            if supports_none:
-                levels.append('disabled')
-
-            for level in ('minimal', 'low', 'medium', 'high'):
-                flag = model_info.get(f'supports_{level}_reasoning_effort')
-                if flag is not False:
-                    levels.append(level)
-            for level in ('xhigh', 'max'):
-                if model_info.get(f'supports_{level}_reasoning_effort') is True:
-                    levels.append(level)
+            levels = ['provider_default']
 
         return {
             'supported': True,
             'levels': list(dict.fromkeys(levels)),
-            'source': 'litellm' if detected else ('provider' if inferred else 'manual'),
+            'source': 'litellm' if detected else ('provider' if known_levels is not None else 'manual'),
         }
 
     def _build_reasoning_args(self, model: requester.RuntimeLLMModel) -> dict[str, typing.Any]:
-        raw_config = getattr(model, 'reasoning_config_override', None)
-        if raw_config is None:
-            raw_config = getattr(model.model_entity, 'reasoning_config', None)
-        if not isinstance(raw_config, dict):
-            raw_config = None
-        config = reasoning.normalize_reasoning_config(raw_config)
-        level = config['level']
+        level = self._reasoning_level(model)
         if level == 'provider_default':
             return {}
 
+        config = {'level': level}
         capabilities = self.get_reasoning_capabilities(model)
         try:
             reasoning.validate_reasoning_capabilities(config, capabilities, model.model_entity.name)
         except ValueError as exc:
             raise errors.RequesterError(str(exc)) from exc
 
-        provider = self._reasoning_provider(model.model_entity.name)
+        family = self._reasoning_family(model.model_entity.name, model)
         if level == 'disabled':
-            if provider == 'deepseek':
+            if family in {'deepseek', 'kimi', 'mimo', 'doubao'}:
                 return {'extra_body': {'thinking': {'type': 'disabled'}}}
-            if provider == 'volcengine':
+            if family == 'qwen':
+                return {'extra_body': {'enable_thinking': False}}
+            if family == 'volcengine':
+                return {'extra_body': {'thinking': {'type': 'disabled'}}}
+            if family == 'anthropic':
                 return {'thinking': {'type': 'disabled'}}
             return {'reasoning_effort': 'none'}
         if level == 'enabled':
-            if provider in {'deepseek', 'volcengine'}:
-                return {'thinking': {'type': 'enabled'}}
+            if family in {'deepseek', 'kimi', 'mimo', 'volcengine'}:
+                return {'extra_body': {'thinking': {'type': 'enabled'}}}
+            if family == 'qwen':
+                return {'extra_body': {'enable_thinking': True}}
             return {'reasoning_effort': 'low'}
+        if family == 'deepseek':
+            return {
+                'extra_body': {
+                    'thinking': {'type': 'enabled'},
+                    'reasoning_effort': level,
+                }
+            }
         return {'reasoning_effort': level}
 
+    @staticmethod
+    def _reasoning_config_value(model: requester.RuntimeLLMModel) -> typing.Any:
+        raw_config = getattr(model, 'reasoning_config_override', None)
+        if raw_config is None:
+            raw_config = getattr(model.model_entity, 'reasoning_config', None)
+        if not isinstance(raw_config, dict):
+            return None
+        return raw_config
+
+    def _reasoning_level(self, model: requester.RuntimeLLMModel) -> str:
+        return reasoning.normalize_reasoning_config(self._reasoning_config_value(model))['level']
+
     def _infer_model_type(self, model_id: str) -> str:
         normalized_id = (model_id or '').lower()
         if any(kw in normalized_id for kw in self._RERANK_MODEL_HINTS):
@@ -510,7 +641,9 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
             supports_provider_reported_reasoning = bool(
                 model_payload and model_payload.get('supports_reasoning') is True
             )
-            if supports_provider_reported_reasoning or self._supports_reasoning(model_id):
+            family = self._reasoning_family(model_id)
+            supports_known_reasoning = self._known_reasoning_levels(model_id, family) is not None
+            if supports_provider_reported_reasoning or supports_known_reasoning or self._supports_reasoning(model_id):
                 abilities.append('reasoning')
             scanned_model['abilities'] = abilities
 
@@ -522,13 +655,43 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
 
         return scanned_model
 
-    def _convert_messages(self, messages: typing.List[provider_message.Message]) -> list[dict]:
+    def _convert_messages(
+        self,
+        messages: typing.List[provider_message.Message],
+        reasoning_family: str = '',
+        include_reasoning_context: bool = True,
+    ) -> list[dict]:
         """Convert LangBot messages to LiteLLM/OpenAI format."""
         req_messages = []
         for m in messages:
             msg_dict = m.dict(exclude_none=True)
             content = msg_dict.get('content')
 
+            if msg_dict.get('role') == 'assistant' and reasoning_family:
+                provider_fields = msg_dict.get('provider_specific_fields')
+                if isinstance(provider_fields, dict):
+                    cleaned_provider_fields = dict(provider_fields)
+                    reasoning_content = cleaned_provider_fields.pop('reasoning_content', None)
+                    thinking_blocks = cleaned_provider_fields.pop('thinking_blocks', None)
+
+                    if include_reasoning_context:
+                        if reasoning_family == 'anthropic' and thinking_blocks:
+                            msg_dict['thinking_blocks'] = thinking_blocks
+                        elif reasoning_family in {
+                            'deepseek',
+                            'kimi',
+                            'qwen',
+                            'doubao',
+                            'mimo',
+                            'volcengine',
+                        } and isinstance(reasoning_content, str):
+                            msg_dict['reasoning_content'] = reasoning_content
+
+                    if cleaned_provider_fields:
+                        msg_dict['provider_specific_fields'] = cleaned_provider_fields
+                    else:
+                        msg_dict.pop('provider_specific_fields', None)
+
             if isinstance(content, list):
                 converted_parts = []
                 for part in content:
@@ -819,7 +982,13 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
         stream: bool = False,
     ) -> dict:
         """Build common completion arguments for invoke_llm and invoke_llm_stream."""
-        req_messages = self._convert_messages(messages)
+        reasoning_family = self._reasoning_family(model.model_entity.name, model)
+        reasoning_level = self._reasoning_level(model)
+        req_messages = self._convert_messages(
+            messages,
+            reasoning_family=reasoning_family,
+            include_reasoning_context=reasoning_level != 'disabled',
+        )
         model_name = self._build_litellm_model_name(model.model_entity.name)
         api_key = model.provider.token_mgr.get_token()
 
@@ -855,7 +1024,7 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
                 args['extra_body'] = {**existing_extra_body, **reasoning_extra_body}
             else:
                 args.update(reasoning_args)
-            if 'reasoning_effort' in reasoning_args and self._reasoning_provider(model.model_entity.name) == 'openai':
+            if 'reasoning_effort' in reasoning_args and self._get_custom_llm_provider() == 'openai':
                 allowed_openai_params = args.get('allowed_openai_params') or []
                 if not isinstance(allowed_openai_params, (list, tuple, set)):
                     raise errors.RequesterError('allowed_openai_params must be an array')

tests/unit_tests/provider/test_reasoning_control.py

@@ -19,6 +19,7 @@ def _runtime_model(
     level: str = 'provider_default',
     name: str = 'reasoning-model',
     abilities: list[str] | None = None,
+    requester_name: str | None = None,
 ) -> requester.RuntimeLLMModel:
     execution_context = ExecutionContext(
         instance_uuid='instance-test',
@@ -40,7 +41,7 @@ def _runtime_model(
             workspace_uuid='workspace-test',
             uuid='provider-test',
             name='provider',
-            requester='openai',
+            requester=requester_name or request.requester_cfg.get('requester_name') or 'custom-requester',
             base_url='https://example.com',
             api_keys=[],
         ),
@@ -50,8 +51,14 @@ def _runtime_model(
     return requester.RuntimeLLMModel(execution_context, entity, provider)
 
 
-def _requester(provider: str = '') -> LiteLLMRequester:
-    return LiteLLMRequester(SimpleNamespace(), {'custom_llm_provider': provider})
+def _requester(provider: str = '', requester_name: str = '') -> LiteLLMRequester:
+    return LiteLLMRequester(
+        SimpleNamespace(),
+        {
+            'custom_llm_provider': provider,
+            'requester_name': requester_name,
+        },
+    )
 
 
 def test_reasoning_config_normalization_and_conflicts():
@@ -73,46 +80,35 @@ def test_reasoning_config_normalization_and_conflicts():
             ['reasoning'],
             {'extra_body': {'thinking_budget': 1024}},
         )
+    assert reasoning.find_reasoning_arg_conflicts(
+        {
+            'enable_thinking': True,
+            'extra_body': {'reasoning_effort': 'high'},
+        }
+    ) == ['enable_thinking', 'extra_body.reasoning_effort']
 
 
-def test_manual_reasoning_model_exposes_conservative_effort_levels(monkeypatch):
+def test_manual_reasoning_model_without_known_protocol_stays_conservative(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'],
+        'levels': ['provider_default'],
         'source': 'manual',
     }
 
 
-def test_provider_protocol_exposes_reasoning_for_unknown_model(monkeypatch):
-    request = _requester('openai')
+def test_openai_protocol_does_not_mark_unknown_models_as_reasoning(monkeypatch):
+    request = _requester('openai', 'openai-chat-completions')
     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'],
@@ -120,33 +116,34 @@ def test_unknown_unmarked_model_without_provider_stays_safe(monkeypatch):
     }
 
 
-def test_mimo_native_model_uses_known_equivalent_litellm_metadata(monkeypatch):
-    request = _requester('openai')
+def test_unknown_unmarked_model_without_provider_stays_safe(monkeypatch):
+    request = _requester()
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
 
-    def supports_reasoning(model: str, custom_llm_provider: str | None = None) -> bool:
-        return model == 'openrouter/xiaomi/mimo-v2.5'
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name='unknown-model', abilities=[]))
 
-    def get_model_info(model: str) -> dict:
-        if model == 'openrouter/xiaomi/mimo-v2.5':
-            return {'supports_reasoning': True}
-        raise ValueError('unknown model')
+    assert capabilities == {
+        'supported': False,
+        'levels': ['provider_default'],
+        'source': 'unknown',
+    }
 
-    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=[])
-    )
+def test_mimo_exposes_off_on_without_fake_effort_levels(monkeypatch):
+    request = _requester('openai', 'mimo-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    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',
+        'levels': ['provider_default', 'disabled', 'enabled'],
+        'source': 'provider',
     }
 
 
 def test_openai_reasoning_levels_follow_litellm_metadata(monkeypatch):
-    request = _requester('openai')
+    request = _requester('openai', 'openai-chat-completions')
     monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
     monkeypatch.setattr(
         request,
@@ -172,34 +169,199 @@ def test_openai_reasoning_levels_follow_litellm_metadata(monkeypatch):
     ]
 
 
-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},
+def test_anthropic_adaptive_and_always_on_profiles(monkeypatch):
+    request = _requester('anthropic', 'anthropic-messages')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    adaptive = request.get_reasoning_capabilities(_runtime_model(request, name='claude-sonnet-4-6', abilities=[]))
+    assert adaptive['levels'] == [
+        'provider_default',
+        'disabled',
+        'low',
+        'medium',
+        'high',
+        'xhigh',
+        'max',
+    ]
+
+    always_on = request.get_reasoning_capabilities(_runtime_model(request, name='claude-fable-5', abilities=[]))
+    assert 'disabled' not in always_on['levels']
+
+    legacy = request.get_reasoning_capabilities(_runtime_model(request, name='claude-3-5-sonnet', abilities=[]))
+    assert legacy['levels'] == ['provider_default', 'low', 'medium', 'high']
+
+
+def test_deepseek_profiles_match_model_generation(monkeypatch):
+    request = _requester('deepseek', 'deepseek-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    assert request.get_reasoning_capabilities(_runtime_model(request, name='deepseek-v4-flash', abilities=[]))[
+        'levels'
+    ] == ['provider_default', 'disabled', 'low', 'high', 'xhigh', 'max']
+    assert request.get_reasoning_capabilities(_runtime_model(request, name='deepseek-chat', abilities=[]))[
+        'levels'
+    ] == ['provider_default', 'disabled', 'enabled']
+    assert request.get_reasoning_capabilities(_runtime_model(request, name='deepseek-r1', abilities=[]))['levels'] == [
+        'provider_default'
+    ]
+
+
+@pytest.mark.parametrize(
+    ('model_name', 'expected_levels'),
+    [
+        ('kimi-k3', ['provider_default', 'low', 'high', 'max']),
+        ('kimi-k2.7-code', ['provider_default']),
+        ('kimi-k2.6', ['provider_default', 'disabled', 'enabled']),
+        ('kimi-k2.5', ['provider_default', 'disabled', 'enabled']),
+    ],
+)
+def test_kimi_profiles(model_name, expected_levels, monkeypatch):
+    request = _requester('openai', 'moonshot-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name=model_name, abilities=[]))
+
+    assert capabilities['levels'] == expected_levels
+
+
+def test_qwen_mixed_and_dedicated_thinking_profiles(monkeypatch):
+    request = _requester('openai', 'bailian-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    mixed = request.get_reasoning_capabilities(_runtime_model(request, name='qwen-plus', abilities=[]))
+    dedicated = request.get_reasoning_capabilities(
+        _runtime_model(request, name='qwen3-235b-a22b-thinking-2507', abilities=[])
     )
 
-    assert openai_request._build_reasoning_args(_runtime_model(openai_request, 'provider_default')) == {}
-    assert openai_request._build_reasoning_args(_runtime_model(openai_request, 'disabled')) == {
+    assert mixed['levels'] == ['provider_default', 'disabled', 'enabled']
+    assert dedicated['levels'] == ['provider_default']
+
+
+def test_doubao_exposes_documented_effort_range(monkeypatch):
+    request = _requester('openai', 'doubao-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(
+        _runtime_model(request, name='doubao-seed-2-1-pro-260628', abilities=[])
+    )
+
+    assert capabilities['levels'] == ['provider_default', 'disabled', 'low', 'medium', 'high']
+
+
+@pytest.mark.parametrize(
+    ('model_name', 'expected_levels'),
+    [
+        ('gpt-5', ['provider_default', 'low', 'medium', 'high']),
+        (
+            'claude-sonnet-4-6',
+            ['provider_default', 'disabled', 'low', 'medium', 'high', 'xhigh', 'max'],
+        ),
+        ('deepseek-v4-flash', ['provider_default', 'disabled', 'low', 'high', 'xhigh', 'max']),
+        ('kimi-k2.6', ['provider_default', 'disabled', 'enabled']),
+        ('qwen-plus', ['provider_default', 'disabled', 'enabled']),
+        ('doubao-seed-2-1-pro-260628', ['provider_default', 'disabled', 'low', 'medium', 'high']),
+        ('mimo-v2.5', ['provider_default', 'disabled', 'enabled']),
+    ],
+)
+def test_new_api_infers_upstream_protocol_from_model_name(model_name, expected_levels, monkeypatch):
+    request = _requester('openai', 'new-api-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name=model_name, abilities=[]))
+
+    assert capabilities['levels'] == expected_levels
+
+
+@pytest.mark.parametrize(
+    ('provider', 'requester_name', 'model_name'),
+    [
+        ('openai', 'openai-chat-completions', 'gpt-5'),
+        ('anthropic', 'anthropic-messages', 'claude-sonnet-4-6'),
+        ('deepseek', 'deepseek-chat-completions', 'deepseek-v4-flash'),
+        ('openai', 'mimo-chat-completions', 'mimo-v2.5'),
+        ('openai', 'moonshot-chat-completions', 'kimi-k2.6'),
+        ('openai', 'bailian-chat-completions', 'qwen-plus'),
+        ('openai', 'doubao-chat-completions', 'doubao-seed-2-1-pro-260628'),
+        ('openai', 'new-api-chat-completions', 'deepseek-v4-flash'),
+    ],
+)
+def test_scanned_known_models_gain_reasoning_ability(provider, requester_name, model_name, monkeypatch):
+    request = _requester(provider, requester_name)
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(request, '_supports_function_calling', lambda _: False)
+    monkeypatch.setattr(request, '_supports_vision', lambda _: False)
+    monkeypatch.setattr(request, '_safe_context_length', lambda _: None)
+
+    scanned = request._enrich_scanned_model(model_name)
+
+    assert scanned['abilities'] == ['reasoning']
+
+
+def test_new_api_unknown_alias_stays_conservative(monkeypatch):
+    request = _requester('openai', 'new-api-chat-completions')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(
+        _runtime_model(request, name='company-internal-alias', abilities=[])
+    )
+
+    assert capabilities == {
+        'supported': False,
+        'levels': ['provider_default'],
+        'source': 'unknown',
+    }
+
+
+def test_reasoning_argument_translation(monkeypatch):
+    openai_request = _requester('openai', 'openai-chat-completions')
+    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, 'disabled', name='gpt-5')) == {
         '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 _: {})
+    anthropic_request = _requester('anthropic', 'anthropic-messages')
+    assert anthropic_request._build_reasoning_args(
+        _runtime_model(anthropic_request, 'disabled', name='claude-sonnet-4-6')
+    ) == {'thinking': {'type': 'disabled'}}
+
+    deepseek_request = _requester('deepseek', 'deepseek-chat-completions')
     assert deepseek_request._build_reasoning_args(
-        _runtime_model(deepseek_request, 'enabled', name='deepseek-chat')
-    ) == {'thinking': {'type': 'enabled'}}
-    assert deepseek_request._build_reasoning_args(
-        _runtime_model(deepseek_request, 'disabled', name='deepseek-chat')
-    ) == {'extra_body': {'thinking': {'type': 'disabled'}}}
+        _runtime_model(deepseek_request, 'high', name='deepseek-v4-flash')
+    ) == {
+        'extra_body': {
+            'thinking': {'type': 'enabled'},
+            'reasoning_effort': 'high',
+        }
+    }
+
+    kimi_request = _requester('openai', 'moonshot-chat-completions')
+    assert kimi_request._build_reasoning_args(_runtime_model(kimi_request, 'enabled', name='kimi-k2.6')) == {
+        'extra_body': {'thinking': {'type': 'enabled'}}
+    }
+    assert kimi_request._build_reasoning_args(_runtime_model(kimi_request, 'high', name='kimi-k3')) == {
+        'reasoning_effort': 'high'
+    }
+
+    qwen_request = _requester('openai', 'bailian-chat-completions')
+    assert qwen_request._build_reasoning_args(_runtime_model(qwen_request, 'disabled', name='qwen-plus')) == {
+        'extra_body': {'enable_thinking': False}
+    }
+
+    doubao_request = _requester('openai', 'doubao-chat-completions')
+    assert doubao_request._build_reasoning_args(
+        _runtime_model(doubao_request, 'high', name='doubao-seed-2-1-pro-260628')
+    ) == {'reasoning_effort': 'high'}
+
+    mimo_request = _requester('openai', 'mimo-chat-completions')
+    assert mimo_request._build_reasoning_args(_runtime_model(mimo_request, 'disabled', name='mimo-v2.5')) == {
+        'extra_body': {'thinking': {'type': 'disabled'}}
+    }
 
 
 def test_pipeline_reasoning_override_takes_precedence(monkeypatch):
-    request = _requester('openai')
+    request = _requester('openai', 'openai-chat-completions')
     monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
     monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
     model = _runtime_model(request, 'high', name='gpt-5')
@@ -211,24 +373,14 @@ def test_pipeline_reasoning_override_takes_precedence(monkeypatch):
     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')
+    deepseek_request = _requester('deepseek', 'deepseek-chat-completions')
     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']
+    assert deepseek_capabilities['levels'] == ['provider_default']
 
     gemini_request = _requester('gemini')
     monkeypatch.setattr(gemini_request, '_supports_reasoning', lambda _: True)
@@ -243,8 +395,8 @@ def test_always_on_reasoning_models_do_not_offer_disabled(monkeypatch):
         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')
+def test_non_target_provider_capabilities_remain_supported(monkeypatch):
+    ollama_request = _requester('ollama', 'ollama')
     monkeypatch.setattr(ollama_request, '_supports_reasoning', lambda _: False)
     monkeypatch.setattr(ollama_request, '_safe_model_info', lambda _: {})
 
@@ -270,12 +422,12 @@ def test_toggle_and_effort_provider_capabilities(monkeypatch):
         'reasoning_effort': 'high'
     }
 
-    volcengine_request = _requester('volcengine')
+    volcengine_request = _requester('volcengine', 'volcark-chat-completions')
     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'}}
+    ) == {'extra_body': {'thinking': {'type': 'disabled'}}}
 
 
 def test_explicit_unsupported_level_raises(monkeypatch):
@@ -288,19 +440,20 @@ def test_explicit_unsupported_level_raises(monkeypatch):
 
 
 def test_provider_inference_rejects_levels_outside_conservative_profile(monkeypatch):
-    request = _requester('openai')
+    request = _requester('openai', 'openai-chat-completions')
     monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
 
     with pytest.raises(errors.RequesterError, match='Available levels: provider_default, low, medium, high'):
-        request._build_reasoning_args(_runtime_model(request, 'xhigh', abilities=[]))
+        request._build_reasoning_args(_runtime_model(request, 'xhigh', name='gpt-5', abilities=[]))
 
 
 @pytest.mark.asyncio
 async def test_completion_args_reject_reasoning_extra_arg_conflicts(monkeypatch):
-    request = _requester('openai')
+    request = _requester('openai', 'openai-chat-completions')
     monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
     monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
-    model = _runtime_model(request, 'high')
+    model = _runtime_model(request, 'high', name='gpt-5')
     model.model_entity.extra_args = {'reasoning_effort': 'low'}
     model.provider.token_mgr.get_token = lambda: 'test-token'
 
@@ -310,10 +463,8 @@ async def test_completion_args_reject_reasoning_extra_arg_conflicts(monkeypatch)
 
 @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')
+    request = _requester('openai', 'moonshot-chat-completions')
+    model = _runtime_model(request, 'high', name='kimi-k3')
     model.model_entity.extra_args = {'allowed_openai_params': ['custom_extension']}
     model.provider.token_mgr.get_token = lambda: 'test-token'
 
@@ -325,7 +476,7 @@ async def test_openai_compatible_reasoning_effort_is_explicitly_allowed(monkeypa
 
 @pytest.mark.asyncio
 async def test_provider_default_does_not_allow_or_send_reasoning_effort():
-    request = _requester('openai')
+    request = _requester('openai', 'new-api-chat-completions')
     model = _runtime_model(request, 'provider_default', name='deepseek-v4-flash')
     model.provider.token_mgr.get_token = lambda: 'test-token'
 
@@ -337,7 +488,7 @@ async def test_provider_default_does_not_allow_or_send_reasoning_effort():
 
 @pytest.mark.asyncio
 async def test_deepseek_disabled_thinking_is_merged_into_extra_body(monkeypatch):
-    request = _requester('deepseek')
+    request = _requester('deepseek', 'deepseek-chat-completions')
     monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
     model = _runtime_model(request, 'disabled', name='deepseek-chat')
     model.model_entity.extra_args = {'extra_body': {'custom_extension': True}}
@@ -351,6 +502,68 @@ async def test_deepseek_disabled_thinking_is_merged_into_extra_body(monkeypatch)
     }
 
 
+@pytest.mark.asyncio
+async def test_openai_compatible_reasoning_history_is_promoted_for_tool_continuity():
+    request = _requester('openai', 'mimo-chat-completions')
+    model = _runtime_model(request, 'enabled', name='mimo-v2.5')
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+    history = [
+        provider_message.Message(
+            role='assistant',
+            content='',
+            provider_specific_fields={'reasoning_content': 'prior reasoning'},
+        )
+    ]
+
+    args = await request._build_completion_args(model, history)
+
+    assert args['messages'][0]['reasoning_content'] == 'prior reasoning'
+    assert 'provider_specific_fields' not in args['messages'][0]
+
+
+@pytest.mark.asyncio
+async def test_disabling_reasoning_removes_previous_reasoning_context():
+    request = _requester('openai', 'mimo-chat-completions')
+    model = _runtime_model(request, 'disabled', name='mimo-v2.5')
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+    history = [
+        provider_message.Message(
+            role='assistant',
+            content='answer',
+            provider_specific_fields={'reasoning_content': 'prior reasoning'},
+        )
+    ]
+
+    args = await request._build_completion_args(model, history)
+
+    assert 'reasoning_content' not in args['messages'][0]
+    assert 'provider_specific_fields' not in args['messages'][0]
+
+
+@pytest.mark.asyncio
+async def test_anthropic_history_promotes_thinking_blocks_instead_of_reasoning_content():
+    request = _requester('anthropic', 'anthropic-messages')
+    model = _runtime_model(request, 'high', name='claude-sonnet-4-6')
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+    thinking_blocks = [{'type': 'thinking', 'thinking': 'prior reasoning', 'signature': 'sig'}]
+    history = [
+        provider_message.Message(
+            role='assistant',
+            content='',
+            provider_specific_fields={
+                'reasoning_content': 'prior reasoning',
+                'thinking_blocks': thinking_blocks,
+            },
+        )
+    ]
+
+    args = await request._build_completion_args(model, history)
+
+    assert args['messages'][0]['thinking_blocks'] == thinking_blocks
+    assert 'reasoning_content' not in args['messages'][0]
+    assert 'provider_specific_fields' not in args['messages'][0]
+
+
 class _Dumpable:
     def __init__(self, data: dict):
         self.data = data