feat(provider): add pipeline reasoning controls (#2373)

* feat(provider): add pipeline reasoning controls

* fix(provider): preserve local agent model compatibility

* refactor(web): use shadcn reasoning slider

* fix(runtime): stabilize reasoning chat delivery

* fix(provider): route reasoning controls by model family

* fix(provider): handle hosted Kimi reasoning protocols

* fix(provider): map qwen reasoning levels to budgets

* fix(provider): preserve think tags in streamed reasoning

* fix(provider): preserve reasoning tool metadata

* style(provider): satisfy ruff checks after merge

* fix(persistence): preserve reasoning migration compatibility

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

@@ -7,7 +7,7 @@ import typing
 import litellm
 from litellm import acompletion, aembedding, arerank
 
-from .. import errors, requester
+from .. import errors, reasoning, requester
 from ....utils import httpclient
 import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
 import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
@@ -164,6 +164,39 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
 
     _EMBEDDING_MODEL_HINTS = ('embedding', 'embed', 'bge-', 'e5-', 'm3e', 'gte-', 'text-embedding')
     _RERANK_MODEL_HINTS = ('rerank', 're-rank', 're_rank')
+    _QWEN_DEDICATED_THINKING_MODELS = frozenset(
+        {
+            'qwen3.7-max-preview',
+            'qwen3.7-max-2026-05-17',
+        }
+    )
+    _QWEN_REASONING_BUDGETS = {
+        'low': 1024,
+        'medium': 4096,
+        'high': 8192,
+    }
+    _INFERRED_EFFORT_PROVIDERS = frozenset(
+        {
+            'anthropic',
+            'gemini',
+            'groq',
+            'mistral',
+            'openai',
+            'openrouter',
+            'together_ai',
+            'xai',
+        }
+    )
+    _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': '',
@@ -172,6 +205,7 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
         'drop_params': False,
         'num_retries': 0,
         'api_version': '',
+        'requester_name': '',
     }
 
     async def initialize(self):
@@ -201,7 +235,10 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
             return False
 
         provider = self._get_custom_llm_provider()
-        candidates: list[tuple[str, str | None]] = [(model_name, provider)]
+        candidates: list[tuple[str, str | None]] = [
+            (candidate, None) for candidate in self._metadata_model_candidates(model_name)
+        ]
+        candidates.append((model_name, provider))
         litellm_model_name = self._build_litellm_model_name(model_name)
         if litellm_model_name != model_name:
             candidates.append((litellm_model_name, None))
@@ -268,6 +305,14 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
                 deduped_candidates.append(candidate)
         return deduped_candidates
 
+    @staticmethod
+    def _metadata_model_candidates(model_name: str) -> list[str]:
+        """Return known equivalent model IDs used only for LiteLLM metadata lookup."""
+        normalized_model_name = (model_name or '').lower()
+        if normalized_model_name.startswith('mimo-v2.5'):
+            return [f'openrouter/xiaomi/{normalized_model_name}']
+        return []
+
     def _known_context_length_fallback(self, model_name: str) -> int | None:
         normalized_model_name = (model_name or '').lower()
         if normalized_model_name.startswith('deepseek-v4-'):
@@ -287,7 +332,8 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
         if not callable(helper):
             return self._known_context_length_fallback(model_name)
 
-        candidates = [model_name]
+        candidates = self._metadata_model_candidates(model_name)
+        candidates.append(model_name)
         litellm_model_name = self._build_litellm_model_name(model_name)
         if litellm_model_name != model_name:
             candidates.append(litellm_model_name)
@@ -314,6 +360,297 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
     def _supports_vision(self, model_name: str) -> bool:
         return self._safe_litellm_bool_helper('supports_vision', model_name)
 
+    def _supports_reasoning(self, model_name: str) -> bool:
+        return self._safe_litellm_bool_helper('supports_reasoning', model_name)
+
+    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 ''
+
+        # Bailian's compatible endpoint also hosts Kimi models. Keep those
+        # models on Kimi's ``thinking`` protocol instead of Qwen's
+        # ``enable_thinking`` protocol.
+        if requester_name == 'bailian-chat-completions':
+            inferred_family = self._infer_reasoning_family_from_model_name(model_name)
+            if inferred_family == 'kimi':
+                return inferred_family
+
+        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
+
+    @staticmethod
+    def _is_anthropic_adaptive_model(model_name: str) -> bool:
+        basename = model_name.lower().rsplit('/', 1)[-1]
+        if 'mythos-preview' in basename:
+            return True
+
+        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 in LiteLLMRequester._QWEN_DEDICATED_THINKING_MODELS
+            or normalized_name.startswith('qwq')
+            or '-thinking' in normalized_name
+        )
+
+    @staticmethod
+    def _supports_qwen_thinking_budget(model_name: str) -> bool:
+        """Return whether the documented Qwen3 family supports thinking_budget."""
+        normalized_name = model_name.lower().rsplit('/', 1)[-1]
+        return normalized_name.startswith('qwen3')
+
+    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):
+                if self._supports_qwen_thinking_budget(normalized_name):
+                    return ['provider_default', 'low', 'medium', 'high']
+                return ['provider_default']
+            if self._supports_qwen_thinking_budget(normalized_name):
+                return ['provider_default', 'disabled', 'low', 'medium', 'high']
+            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)
+        if not callable(helper):
+            return {}
+
+        candidates = [
+            *self._metadata_model_candidates(model_name),
+            model_name,
+            self._build_litellm_model_name(model_name),
+        ]
+        for candidate in candidates:
+            try:
+                info = helper(candidate)
+            except Exception:
+                continue
+            if isinstance(info, dict):
+                return info
+            model_dump = getattr(info, 'model_dump', None)
+            if callable(model_dump):
+                try:
+                    dumped = model_dump()
+                    if isinstance(dumped, dict):
+                        return dumped
+                except Exception:
+                    continue
+        return {}
+
+    def get_reasoning_capabilities(self, model: requester.RuntimeLLMModel) -> dict[str, typing.Any]:
+        model_name = model.model_entity.name
+        abilities = model.model_entity.abilities or []
+        detected = self._supports_reasoning(model_name)
+        declared = 'reasoning' in abilities
+        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()
+        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 family in self._INFERRED_EFFORT_PROVIDERS:
+            levels = ['provider_default', 'low', 'medium', 'high']
+        else:
+            levels = ['provider_default']
+
+        capabilities = {
+            'supported': True,
+            'levels': list(dict.fromkeys(levels)),
+            'source': 'litellm' if detected else ('provider' if known_levels is not None else 'manual'),
+        }
+        if family == 'qwen' and 'disabled' in capabilities['levels'] and 'enabled' not in capabilities['levels']:
+            capabilities['legacy_levels'] = ['enabled']
+        return capabilities
+
+    def _build_reasoning_args(self, model: requester.RuntimeLLMModel) -> dict[str, typing.Any]:
+        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
+
+        family = self._reasoning_family(model.model_entity.name, model)
+        if level == 'disabled':
+            if family in {'deepseek', 'kimi', 'mimo', 'doubao'}:
+                return {'extra_body': {'thinking': {'type': 'disabled'}}}
+            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 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 == 'qwen' and level in self._QWEN_REASONING_BUDGETS:
+            return {
+                'extra_body': {
+                    'enable_thinking': True,
+                    'thinking_budget': self._QWEN_REASONING_BUDGETS[level],
+                }
+            }
+        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):
@@ -344,6 +681,13 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
             )
             if supports_provider_reported_vision or self._supports_vision(model_id):
                 abilities.append('vision')
+            supports_provider_reported_reasoning = bool(
+                model_payload and model_payload.get('supports_reasoning') is True
+            )
+            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
 
             context_length = self._context_length_from_scan_payload(model_payload)
@@ -354,13 +698,51 @@ 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)
+
+                    # ``content`` is also used for the user-facing rendering.
+                    # Do not replay that rendered <think> wrapper alongside the
+                    # structured provider reasoning on the next request.
+                    if reasoning_content or thinking_blocks:
+                        content = msg_dict.get('content')
+                        if isinstance(content, str):
+                            msg_dict['content'] = self._strip_think(content)
+
+                    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:
@@ -421,6 +803,52 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
 
         return content or ''
 
+    @staticmethod
+    def _thinking_blocks_text(thinking_blocks: typing.Any) -> str:
+        if not isinstance(thinking_blocks, list):
+            return ''
+        parts = []
+        for block in thinking_blocks:
+            if isinstance(block, dict):
+                text = block.get('thinking')
+            else:
+                text = getattr(block, 'thinking', None)
+            if isinstance(text, str) and text:
+                parts.append(text)
+        return ''.join(parts)
+
+    @classmethod
+    def _merge_thinking_blocks(
+        cls,
+        current: list[dict[str, typing.Any]],
+        incoming: typing.Any,
+    ) -> list[dict[str, typing.Any]]:
+        """Merge Anthropic thinking block fragments emitted by a stream."""
+        if not isinstance(incoming, list):
+            return current
+        merged = [dict(block) for block in current]
+        for raw_block in incoming:
+            block = cls._as_dict(raw_block)
+            if not block:
+                continue
+            block_type = block.get('type')
+            if block_type == 'redacted_thinking':
+                merged.append(block)
+                continue
+
+            text = block.get('thinking') if isinstance(block.get('thinking'), str) else ''
+            signature = block.get('signature')
+            if merged and merged[-1].get('type') == 'thinking' and not merged[-1].get('signature'):
+                merged[-1]['thinking'] = f'{merged[-1].get("thinking", "")}{text}'
+                if signature:
+                    merged[-1]['signature'] = signature
+            elif merged and signature and merged[-1].get('signature') == signature:
+                if text and text != merged[-1].get('thinking', ''):
+                    merged[-1]['thinking'] = f'{merged[-1].get("thinking", "")}{text}'
+            else:
+                merged.append(block)
+        return merged
+
     @staticmethod
     def _normalize_usage(usage: typing.Any) -> dict:
         """Normalize a LiteLLM/OpenAI usage object into a plain token dict.
@@ -651,7 +1079,13 @@ 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()
 
@@ -670,6 +1104,29 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
             args.update(model.model_entity.extra_args)
         args.update(extra_args)
 
+        reasoning_args = self._build_reasoning_args(model)
+        if reasoning_args:
+            conflicts = reasoning.find_reasoning_arg_conflicts(model.model_entity.extra_args)
+            conflicts.extend(reasoning.find_reasoning_arg_conflicts(extra_args))
+            if conflicts:
+                raise errors.RequesterError(
+                    'reasoning_config conflicts with advanced parameters: ' + ', '.join(dict.fromkeys(conflicts))
+                )
+            reasoning_extra_body = reasoning_args.get('extra_body')
+            if isinstance(reasoning_extra_body, dict):
+                existing_extra_body = args.get('extra_body') or {}
+                if not isinstance(existing_extra_body, dict):
+                    raise errors.RequesterError('extra_body must be an object')
+                args.update({key: value for key, value in reasoning_args.items() if key != 'extra_body'})
+                args['extra_body'] = {**existing_extra_body, **reasoning_extra_body}
+            else:
+                args.update(reasoning_args)
+            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')
+                args['allowed_openai_params'] = list(dict.fromkeys([*allowed_openai_params, 'reasoning_effort']))
+
         if funcs:
             tools = await self.ap.tool_mgr.generate_tools_for_openai(funcs)
             if tools:
@@ -699,10 +1156,21 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
 
             content = message_data.get('content', '')
             reasoning_content = message_data.get('reasoning_content', None)
-            message_data['content'] = self._process_thinking_content(content, reasoning_content, remove_think)
+            thinking_blocks = message_data.get('thinking_blocks')
+            if reasoning_content or thinking_blocks:
+                provider_fields = dict(message_data.get('provider_specific_fields') or {})
+                if reasoning_content:
+                    provider_fields['reasoning_content'] = reasoning_content
+                if thinking_blocks:
+                    provider_fields['thinking_blocks'] = thinking_blocks
+                message_data['provider_specific_fields'] = provider_fields
+            display_reasoning = reasoning_content or self._thinking_blocks_text(thinking_blocks) or None
+            message_data['content'] = self._process_thinking_content(content, display_reasoning, remove_think)
 
             if 'reasoning_content' in message_data:
                 del message_data['reasoning_content']
+            if 'thinking_blocks' in message_data:
+                del message_data['thinking_blocks']
 
             message = provider_message.Message(**message_data)
             usage_info = self._extract_usage(response)
@@ -728,6 +1196,9 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
         role = 'assistant'
         tool_call_state: dict[int, dict[str, typing.Any]] = {}
         think_state = _ThinkStripState() if remove_think else None
+        reasoning_started = False
+        reasoning_closed = False
+        thinking_blocks_state: list[dict[str, typing.Any]] = []
 
         try:
             response = await acompletion(**args)
@@ -758,28 +1229,63 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
                 if 'role' in delta and delta['role']:
                     role = delta['role']
 
-                delta_content = delta.get('content', '')
-                reasoning_content = delta.get('reasoning_content', '')
+                delta_content = delta.get('content') or ''
+                reasoning_content = delta.get('reasoning_content') or ''
+                provider_fields = dict(delta.get('provider_specific_fields') or {})
+                raw_thinking_blocks = delta.get('thinking_blocks')
+                if raw_thinking_blocks:
+                    thinking_blocks_state = self._merge_thinking_blocks(thinking_blocks_state, raw_thinking_blocks)
+                    provider_fields['thinking_blocks'] = thinking_blocks_state
+                thinking_blocks_text = self._thinking_blocks_text(raw_thinking_blocks)
+                display_reasoning_content = reasoning_content or thinking_blocks_text
 
                 # Handle reasoning_content based on remove_think flag
                 if reasoning_content:
+                    provider_fields['reasoning_content'] = reasoning_content
                     if remove_think:
-                        # Skip reasoning content when remove_think is True
-                        chunk_idx += 1
-                        continue
+                        delta_content = delta_content or None
                     else:
-                        # Use reasoning_content as the displayed content
-                        delta_content = reasoning_content
+                        # Stream explicit markers so downstream adapters and
+                        # the debug page see the same format as non-streaming
+                        # responses.
+                        if not reasoning_started:
+                            delta_content = '<think>\n'
+                            reasoning_started = True
+                        else:
+                            delta_content = ''
+                        delta_content += display_reasoning_content
+                        if delta.get('content'):
+                            delta_content += f'\n</think>\n{delta.get("content")}'
+                            reasoning_closed = True
+
+                elif display_reasoning_content:
+                    if remove_think:
+                        delta_content = delta_content or None
+                    else:
+                        if not reasoning_started:
+                            delta_content = '<think>\n'
+                            reasoning_started = True
+                        else:
+                            delta_content = ''
+                        delta_content += display_reasoning_content
+                        if delta.get('content'):
+                            delta_content += f'\n</think>\n{delta.get("content")}'
+                            reasoning_closed = True
+
+                elif delta_content and not remove_think and reasoning_started and not reasoning_closed:
+                    delta_content = f'\n</think>\n{delta_content}'
+                    reasoning_closed = True
+
+                if finish_reason and not remove_think and reasoning_started and not reasoning_closed:
+                    delta_content = f'{delta_content}\n</think>\n'
+                    reasoning_closed = True
 
                 if think_state is not None and delta_content:
                     delta_content = think_state.feed(delta_content)
-                    if not delta_content:
-                        chunk_idx += 1
-                        continue
 
                 tool_calls = self._normalize_stream_tool_calls(delta.get('tool_calls'), tool_call_state)
 
-                if chunk_idx == 0 and not delta_content and not tool_calls:
+                if not delta_content and not tool_calls and not provider_fields and not finish_reason:
                     chunk_idx += 1
                     continue
 
@@ -791,13 +1297,20 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
                 }
 
                 # Preserve provider_specific_fields from delta (e.g., Gemini thought_signatures)
-                if delta.get('provider_specific_fields'):
-                    chunk_data['provider_specific_fields'] = delta['provider_specific_fields']
+                if provider_fields:
+                    chunk_data['provider_specific_fields'] = provider_fields
 
                 chunk_data = {k: v for k, v in chunk_data.items() if v is not None}
                 yield provider_message.MessageChunk(**chunk_data)
                 chunk_idx += 1
 
+            if reasoning_started and not reasoning_closed:
+                yield provider_message.MessageChunk(
+                    role=role,
+                    content='\n</think>\n',
+                    is_final=True,
+                )
+
             if think_state is not None:
                 pending_content = think_state.flush()
                 if pending_content: