refactor(provider): use LiteLLM as unified LLM requester backend

- Replace 23+ individual requester implementations with unified litellmchat.py
  - Add litellm_provider field to 27 YAML manifests for provider routing
  - Delete redundant requester subclasses
  - Add unit tests for LiteLLMRequester (29 tests)
  - Fix num_retries parameter name (was max_retries)
  - Fix exception handling order for subclass exceptions

  LiteLLM provides unified API for 100+ providers, eliminating need for
  provider-specific requesters.

pyproject.toml

@@ -77,6 +77,7 @@ dependencies = [
     "pymilvus>=2.6.4",
     "pgvector>=0.4.1",
     "botocore>=1.42.39",
+    "litellm>=1.0.0",
 ]
 keywords = [
     "bot",

src/langbot/pkg/api/http/service/provider.py

@@ -139,7 +139,9 @@ class ModelProviderService:
 
         return {'llm_count': llm_count, 'embedding_count': embedding_count, 'rerank_count': rerank_count}
 
-    async def find_or_create_provider(self, requester: str, base_url: str, api_keys: list) -> str:
+    async def find_or_create_provider(
+        self, requester: str, base_url: str, api_keys: list
+    ) -> str:
         """Find existing provider or create new one"""
         # Try to find existing provider with same config
         result = await self.ap.persistence_mgr.execute_async(

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

@@ -4,6 +4,7 @@ import sqlalchemy
 import traceback
 
 from . import requester
+from .requesters import litellmchat
 from ...core import app
 from ...discover import engine
 from . import token
@@ -42,6 +43,13 @@ class ModelManager:
 
         requester_dict: dict[str, type[requester.ProviderAPIRequester]] = {}
         for component in self.requester_components:
+            # Skip components that use litellm_provider (they will use litellmchat.py instead)
+            if component.spec.get('litellm_provider'):
+                self.ap.logger.debug(
+                    f'Skipping Python class loading for {component.metadata.name} '
+                    f'(uses litellm_provider={component.spec.get("litellm_provider")})'
+                )
+                continue
             requester_dict[component.metadata.name] = component.get_python_component_class()
 
         self.requester_dict = requester_dict
@@ -260,13 +268,34 @@ class ModelManager:
         else:
             provider_entity = provider_info
 
+        # Get requester manifest to check for litellm_provider
+        requester_manifest = self.get_available_requester_manifest_by_name(provider_entity.requester)
+
+        # Build config from base_url
+        config = {'base_url': provider_entity.base_url}
+
+        # Check if requester manifest specifies litellm_provider
+        if requester_manifest and requester_manifest.spec.get('litellm_provider'):
+            # Use unified LiteLLMRequester with provider prefix
+            # Map litellm_provider (YAML spec) to custom_llm_provider (config)
+            config['custom_llm_provider'] = requester_manifest.spec['litellm_provider']
+            requester_inst = litellmchat.LiteLLMRequester(
+                ap=self.ap,
+                config=config,
+            )
+            self.ap.logger.debug(
+                f'Using LiteLLMRequester for {provider_entity.requester} '
+                f'with custom_llm_provider={config["custom_llm_provider"]}'
+            )
+        else:
+            # Use original requester class (for backward compatibility)
             if provider_entity.requester not in self.requester_dict:
                 raise provider_errors.RequesterNotFoundError(provider_entity.requester)
-
             requester_inst = self.requester_dict[provider_entity.requester](
                 ap=self.ap,
-            config={'base_url': provider_entity.base_url},
+                config=config,
             )
+
         await requester_inst.initialize()
 
         token_mgr = token.TokenManager(name=provider_entity.uuid, tokens=provider_entity.api_keys or [])

src/langbot/pkg/provider/modelmgr/requesters/302aichatcmpl.py

@@ -1,17 +0,0 @@
-from __future__ import annotations
-
-import typing
-import openai
-
-from . import chatcmpl
-
-
-class AI302ChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """302.AI ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://api.302.ai/v1',
-        'timeout': 120,
-    }

src/langbot/pkg/provider/modelmgr/requesters/302aichatcmpl.yaml

@@ -22,12 +22,9 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: openai
   support_type:
   - llm
   - text-embedding
   - rerank
   provider_category: maas
-execution:
-  python:
-    path: ./302aichatcmpl.py
-    attr: AI302ChatCompletions

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

@@ -1,370 +0,0 @@
-from __future__ import annotations
-
-import typing
-import json
-import platform
-import socket
-import anthropic
-import httpx
-
-from .. import errors, requester
-
-from ....utils import image
-import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
-import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
-import langbot_plugin.api.entities.builtin.provider.message as provider_message
-
-
-class AnthropicMessages(requester.ProviderAPIRequester):
-    """Anthropic Messages API 请求器"""
-
-    client: anthropic.AsyncAnthropic
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://api.anthropic.com',
-        'timeout': 120,
-    }
-
-    async def initialize(self):
-        # 兼容 Windows 缺失 TCP_KEEPINTVL 和 TCP_KEEPCNT 的问题
-        if platform.system() == 'Windows':
-            if not hasattr(socket, 'TCP_KEEPINTVL'):
-                socket.TCP_KEEPINTVL = 0
-            if not hasattr(socket, 'TCP_KEEPCNT'):
-                socket.TCP_KEEPCNT = 0
-        httpx_client = anthropic._base_client.AsyncHttpxClientWrapper(
-            base_url=self.requester_cfg['base_url'],
-            # cast to a valid type because mypy doesn't understand our type narrowing
-            timeout=typing.cast(httpx.Timeout, self.requester_cfg['timeout']),
-            limits=anthropic._constants.DEFAULT_CONNECTION_LIMITS,
-            follow_redirects=True,
-            trust_env=True,
-        )
-
-        self.client = anthropic.AsyncAnthropic(
-            api_key='',
-            http_client=httpx_client,
-            base_url=self.requester_cfg['base_url'],
-        )
-
-    async def invoke_llm(
-        self,
-        query: pipeline_query.Query,
-        model: requester.RuntimeLLMModel,
-        messages: typing.List[provider_message.Message],
-        funcs: typing.List[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> provider_message.Message:
-        self.client.api_key = model.provider.token_mgr.get_token()
-
-        args = extra_args.copy()
-        args['model'] = model.model_entity.name
-
-        # 处理消息
-
-        # system
-        system_role_message = None
-
-        for i, m in enumerate(messages):
-            if m.role == 'system':
-                system_role_message = m
-
-                break
-
-        if system_role_message:
-            messages.pop(i)
-
-        if isinstance(system_role_message, provider_message.Message) and isinstance(system_role_message.content, str):
-            args['system'] = system_role_message.content
-
-        req_messages = []
-
-        for m in messages:
-            if m.role == 'tool':
-                tool_call_id = m.tool_call_id
-
-                req_messages.append(
-                    {
-                        'role': 'user',
-                        'content': [
-                            {
-                                'type': 'tool_result',
-                                'tool_use_id': tool_call_id,
-                                'is_error': False,
-                                'content': [{'type': 'text', 'text': m.content}],
-                            }
-                        ],
-                    }
-                )
-
-                continue
-
-            msg_dict = m.dict(exclude_none=True)
-
-            if isinstance(m.content, str) and m.content.strip() != '':
-                msg_dict['content'] = [{'type': 'text', 'text': m.content}]
-            elif isinstance(m.content, list):
-                for i, ce in enumerate(m.content):
-                    if ce.type == 'image_base64':
-                        image_b64, image_format = await image.extract_b64_and_format(ce.image_base64)
-
-                        alter_image_ele = {
-                            'type': 'image',
-                            'source': {
-                                'type': 'base64',
-                                'media_type': f'image/{image_format}',
-                                'data': image_b64,
-                            },
-                        }
-                        msg_dict['content'][i] = alter_image_ele
-
-            if m.tool_calls:
-                for tool_call in m.tool_calls:
-                    msg_dict['content'].append(
-                        {
-                            'type': 'tool_use',
-                            'id': tool_call.id,
-                            'name': tool_call.function.name,
-                            'input': json.loads(tool_call.function.arguments),
-                        }
-                    )
-
-                del msg_dict['tool_calls']
-
-            req_messages.append(msg_dict)
-
-        args['messages'] = req_messages
-
-        if 'thinking' in args:
-            args['thinking'] = {'type': 'enabled', 'budget_tokens': 10000}
-
-        if funcs:
-            tools = await self.ap.tool_mgr.generate_tools_for_anthropic(funcs)
-
-            if tools:
-                args['tools'] = tools
-
-        try:
-            resp = await self.client.messages.create(**args)
-
-            args = {
-                'content': '',
-                'role': resp.role,
-            }
-            assert type(resp) is anthropic.types.message.Message
-
-            for block in resp.content:
-                if not remove_think and block.type == 'thinking':
-                    args['content'] = '<think>\n' + block.thinking + '\n</think>\n' + args['content']
-                elif block.type == 'text':
-                    args['content'] += block.text
-                elif block.type == 'tool_use':
-                    assert type(block) is anthropic.types.tool_use_block.ToolUseBlock
-                    tool_call = provider_message.ToolCall(
-                        id=block.id,
-                        type='function',
-                        function=provider_message.FunctionCall(name=block.name, arguments=json.dumps(block.input)),
-                    )
-                    if 'tool_calls' not in args:
-                        args['tool_calls'] = []
-                    args['tool_calls'].append(tool_call)
-
-            return provider_message.Message(**args)
-        except anthropic.AuthenticationError as e:
-            raise errors.RequesterError(f'api-key 无效: {e.message}')
-        except anthropic.BadRequestError as e:
-            raise errors.RequesterError(str(e.message))
-        except anthropic.NotFoundError as e:
-            if 'model: ' in str(e):
-                raise errors.RequesterError(f'模型无效: {e.message}')
-            else:
-                raise errors.RequesterError(f'请求地址无效: {e.message}')
-
-    async def invoke_llm_stream(
-        self,
-        query: pipeline_query.Query,
-        model: requester.RuntimeLLMModel,
-        messages: typing.List[provider_message.Message],
-        funcs: typing.List[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> provider_message.Message:
-        self.client.api_key = model.provider.token_mgr.get_token()
-
-        args = extra_args.copy()
-        args['model'] = model.model_entity.name
-        args['stream'] = True
-
-        # 处理消息
-
-        # system
-        system_role_message = None
-
-        for i, m in enumerate(messages):
-            if m.role == 'system':
-                system_role_message = m
-
-                break
-
-        if system_role_message:
-            messages.pop(i)
-
-        if isinstance(system_role_message, provider_message.Message) and isinstance(system_role_message.content, str):
-            args['system'] = system_role_message.content
-
-        req_messages = []
-
-        for m in messages:
-            if m.role == 'tool':
-                tool_call_id = m.tool_call_id
-
-                req_messages.append(
-                    {
-                        'role': 'user',
-                        'content': [
-                            {
-                                'type': 'tool_result',
-                                'tool_use_id': tool_call_id,
-                                'is_error': False,  # 暂时直接写false
-                                'content': [
-                                    {'type': 'text', 'text': m.content}
-                                ],  # 这里要是list包裹，应该是多个返回的情况？type类型好像也可以填其他的，暂时只写text
-                            }
-                        ],
-                    }
-                )
-
-                continue
-
-            msg_dict = m.dict(exclude_none=True)
-
-            if isinstance(m.content, str) and m.content.strip() != '':
-                msg_dict['content'] = [{'type': 'text', 'text': m.content}]
-            elif isinstance(m.content, list):
-                for i, ce in enumerate(m.content):
-                    if ce.type == 'image_base64':
-                        image_b64, image_format = await image.extract_b64_and_format(ce.image_base64)
-
-                        alter_image_ele = {
-                            'type': 'image',
-                            'source': {
-                                'type': 'base64',
-                                'media_type': f'image/{image_format}',
-                                'data': image_b64,
-                            },
-                        }
-                        msg_dict['content'][i] = alter_image_ele
-            if isinstance(msg_dict['content'], str) and msg_dict['content'] == '':
-                msg_dict['content'] = []  # 这里不知道为什么会莫名有个空导致content为字符
-            if m.tool_calls:
-                for tool_call in m.tool_calls:
-                    msg_dict['content'].append(
-                        {
-                            'type': 'tool_use',
-                            'id': tool_call.id,
-                            'name': tool_call.function.name,
-                            'input': json.loads(tool_call.function.arguments),
-                        }
-                    )
-
-                del msg_dict['tool_calls']
-
-            req_messages.append(msg_dict)
-        if 'thinking' in args:
-            args['thinking'] = {'type': 'enabled', 'budget_tokens': 10000}
-
-        args['messages'] = req_messages
-
-        if funcs:
-            tools = await self.ap.tool_mgr.generate_tools_for_anthropic(funcs)
-
-            if tools:
-                args['tools'] = tools
-
-        try:
-            role = 'assistant'  # 默认角色
-            # chunk_idx = 0
-            think_started = False
-            think_ended = False
-            finish_reason = False
-            tool_name = ''
-            tool_id = ''
-            async for chunk in await self.client.messages.create(**args):
-                content = ''
-                tool_call = {'id': None, 'function': {'name': None, 'arguments': None}, 'type': 'function'}
-                if isinstance(
-                    chunk, anthropic.types.raw_content_block_start_event.RawContentBlockStartEvent
-                ):  # 记录开始
-                    if chunk.content_block.type == 'tool_use':
-                        if chunk.content_block.name is not None:
-                            tool_name = chunk.content_block.name
-                        if chunk.content_block.id is not None:
-                            tool_id = chunk.content_block.id
-
-                        tool_call['function']['name'] = tool_name
-                        tool_call['function']['arguments'] = ''
-                        tool_call['id'] = tool_id
-
-                    if not remove_think:
-                        if chunk.content_block.type == 'thinking' and not remove_think:
-                            think_started = True
-                        elif chunk.content_block.type == 'text' and chunk.index != 0 and not remove_think:
-                            think_ended = True
-                        continue
-                elif isinstance(chunk, anthropic.types.raw_content_block_delta_event.RawContentBlockDeltaEvent):
-                    if chunk.delta.type == 'thinking_delta':
-                        if think_started:
-                            think_started = False
-                            content = '<think>\n' + chunk.delta.thinking
-                        elif remove_think:
-                            continue
-                        else:
-                            content = chunk.delta.thinking
-                    elif chunk.delta.type == 'text_delta':
-                        if think_ended:
-                            think_ended = False
-                            content = '\n</think>\n' + chunk.delta.text
-                        else:
-                            content = chunk.delta.text
-                    elif chunk.delta.type == 'input_json_delta':
-                        tool_call['function']['arguments'] = chunk.delta.partial_json
-                        tool_call['function']['name'] = tool_name
-                        tool_call['id'] = tool_id
-                elif isinstance(chunk, anthropic.types.raw_content_block_stop_event.RawContentBlockStopEvent):
-                    continue  # 记录raw_content_block结束的
-
-                elif isinstance(chunk, anthropic.types.raw_message_delta_event.RawMessageDeltaEvent):
-                    if chunk.delta.stop_reason == 'end_turn':
-                        finish_reason = True
-                elif isinstance(chunk, anthropic.types.raw_message_stop_event.RawMessageStopEvent):
-                    continue  # 这个好像是完全结束
-                else:
-                    # print(chunk)
-                    self.ap.logger.debug(f'anthropic chunk: {chunk}')
-                    continue
-
-                args = {
-                    'content': content,
-                    'role': role,
-                    'is_final': finish_reason,
-                    'tool_calls': None if tool_call['id'] is None else [tool_call],
-                }
-                # if chunk_idx == 0:
-                #     chunk_idx += 1
-                #     continue
-
-                # assert type(chunk) is anthropic.types.message.Chunk
-
-                yield provider_message.MessageChunk(**args)
-
-            # return llm_entities.Message(**args)
-        except anthropic.AuthenticationError as e:
-            raise errors.RequesterError(f'api-key 无效: {e.message}')
-        except anthropic.BadRequestError as e:
-            raise errors.RequesterError(str(e.message))
-        except anthropic.NotFoundError as e:
-            if 'model: ' in str(e):
-                raise errors.RequesterError(f'模型无效: {e.message}')
-            else:
-                raise errors.RequesterError(f'请求地址无效: {e.message}')

src/langbot/pkg/provider/modelmgr/requesters/anthropicmsgs.yaml

@@ -22,10 +22,7 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: anthropic
   support_type:
   - llm
   provider_category: manufacturer
-execution:
-  python:
-    path: ./anthropicmsgs.py
-    attr: AnthropicMessages

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

@@ -1,242 +0,0 @@
-from __future__ import annotations
-
-import typing
-import dashscope
-import openai
-
-from . import modelscopechatcmpl
-from .. import requester
-import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
-import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
-import langbot_plugin.api.entities.builtin.provider.message as provider_message
-
-
-class BailianChatCompletions(modelscopechatcmpl.ModelScopeChatCompletions):
-    """阿里云百炼大模型平台 ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://dashscope.aliyuncs.com/compatible-mode/v1',
-        'timeout': 120,
-    }
-
-    async def _closure_stream(
-        self,
-        query: pipeline_query.Query,
-        req_messages: list[dict],
-        use_model: requester.RuntimeLLMModel,
-        use_funcs: list[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> provider_message.Message | typing.AsyncGenerator[provider_message.MessageChunk, None]:
-        self.client.api_key = use_model.provider.token_mgr.get_token()
-
-        args = {}
-        args['model'] = use_model.model_entity.name
-
-        if use_funcs:
-            tools = await self.ap.tool_mgr.generate_tools_for_openai(use_funcs)
-
-            if tools:
-                args['tools'] = tools
-
-        # 设置此次请求中的messages
-        messages = req_messages.copy()
-
-        is_use_dashscope_call = False  # 是否使用阿里原生库调用
-        is_enable_multi_model = True  # 是否支持多轮对话
-        use_time_num = 0  # 模型已调用次数，防止存在多文件时重复调用
-        use_time_ids = []  # 已调用的ID列表
-        message_id = 0  # 记录消息序号
-
-        for msg in messages:
-            # print(msg)
-            if 'content' in msg and isinstance(msg['content'], list):
-                for me in msg['content']:
-                    if me['type'] == 'image_base64':
-                        me['image_url'] = {'url': me['image_base64']}
-                        me['type'] = 'image_url'
-                        del me['image_base64']
-                    elif me['type'] == 'file_url' and '.' in me.get('file_name', ''):
-                        # 1. 视频文件推理
-                        # https://bailian.console.aliyun.com/?tab=doc#/doc/?type=model&url=2845871
-                        file_type = me.get('file_name').lower().split('.')[-1]
-                        if file_type in ['mp4', 'avi', 'mkv', 'mov', 'flv', 'wmv']:
-                            me['type'] = 'video_url'
-                            me['video_url'] = {'url': me['file_url']}
-                            del me['file_url']
-                            del me['file_name']
-                            use_time_num += 1
-                            use_time_ids.append(message_id)
-                            is_enable_multi_model = False
-                        # 2. 语音文件识别, 无法通过openai的audio字段传递，暂时不支持
-                        # https://bailian.console.aliyun.com/?tab=doc#/doc/?type=model&url=2979031
-                        elif file_type in [
-                            'aac',
-                            'amr',
-                            'aiff',
-                            'flac',
-                            'm4a',
-                            'mp3',
-                            'mpeg',
-                            'ogg',
-                            'opus',
-                            'wav',
-                            'webm',
-                            'wma',
-                        ]:
-                            me['audio'] = me['file_url']
-                            me['type'] = 'audio'
-                            del me['file_url']
-                            del me['type']
-                            del me['file_name']
-                            is_use_dashscope_call = True
-                            use_time_num += 1
-                            use_time_ids.append(message_id)
-                            is_enable_multi_model = False
-            message_id += 1
-
-        # 使用列表推导式，保留不在 use_time_ids[:-1] 中的元素，仅保留最后一个多媒体消息
-        if not is_enable_multi_model and use_time_num > 1:
-            messages = [msg for idx, msg in enumerate(messages) if idx not in use_time_ids[:-1]]
-
-        if not is_enable_multi_model:
-            messages = [msg for msg in messages if 'resp_message_id' not in msg]
-
-        args['messages'] = messages
-        args['stream'] = True
-
-        # 流式处理状态
-        # tool_calls_map: dict[str, provider_message.ToolCall] = {}
-        chunk_idx = 0
-        thinking_started = False
-        thinking_ended = False
-        role = 'assistant'  # 默认角色
-
-        if is_use_dashscope_call:
-            response = dashscope.MultiModalConversation.call(
-                # 若没有配置环境变量，请用百炼API Key将下行替换为：api_key = "sk-xxx"
-                api_key=use_model.provider.token_mgr.get_token(),
-                model=use_model.model_entity.name,
-                messages=messages,
-                result_format='message',
-                asr_options={
-                    # "language": "zh", # 可选，若已知音频的语种，可通过该参数指定待识别语种，以提升识别准确率
-                    'enable_lid': True,
-                    'enable_itn': False,
-                },
-                stream=True,
-            )
-            content_length_list = []
-            previous_length = 0  # 记录上一次的内容长度
-            for res in response:
-                chunk = res['output']
-                # 解析 chunk 数据
-                if hasattr(chunk, 'choices') and chunk.choices:
-                    choice = chunk.choices[0]
-                    delta_content = choice['message'].content[0]['text']
-                    finish_reason = choice['finish_reason']
-                    content_length_list.append(len(delta_content))
-                else:
-                    delta_content = ''
-                    finish_reason = None
-
-                # 跳过空的第一个 chunk（只有 role 没有内容）
-                if chunk_idx == 0 and not delta_content:
-                    chunk_idx += 1
-                    continue
-
-                # 检查 content_length_list 是否有足够的数据
-                if len(content_length_list) >= 2:
-                    now_content = delta_content[previous_length : content_length_list[-1]]
-                    previous_length = content_length_list[-1]  # 更新上一次的长度
-                else:
-                    now_content = delta_content  # 第一次循环时直接使用 delta_content
-                    previous_length = len(delta_content)  # 更新上一次的长度
-
-                # 构建 MessageChunk - 只包含增量内容
-                chunk_data = {
-                    'role': role,
-                    'content': now_content if now_content else None,
-                    'is_final': bool(finish_reason) and finish_reason != 'null',
-                }
-
-                # 移除 None 值
-                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
-        else:
-            async for chunk in self._req_stream(args, extra_body=extra_args):
-                # 解析 chunk 数据
-                if hasattr(chunk, 'choices') and chunk.choices:
-                    choice = chunk.choices[0]
-                    delta = choice.delta.model_dump() if hasattr(choice, 'delta') else {}
-                    finish_reason = getattr(choice, 'finish_reason', None)
-                else:
-                    delta = {}
-                    finish_reason = None
-
-                # 从第一个 chunk 获取 role，后续使用这个 role
-                if 'role' in delta and delta['role']:
-                    role = delta['role']
-
-                # 获取增量内容
-                delta_content = delta.get('content', '')
-                reasoning_content = delta.get('reasoning_content', '')
-
-                # 处理 reasoning_content
-                if reasoning_content:
-                    # accumulated_reasoning += reasoning_content
-                    # 如果设置了 remove_think，跳过 reasoning_content
-                    if remove_think:
-                        chunk_idx += 1
-                        continue
-
-                    # 第一次出现 reasoning_content，添加 <think> 开始标签
-                    if not thinking_started:
-                        thinking_started = True
-                        delta_content = '<think>\n' + reasoning_content
-                    else:
-                        # 继续输出 reasoning_content
-                        delta_content = reasoning_content
-                elif thinking_started and not thinking_ended and delta_content:
-                    # reasoning_content 结束，normal content 开始，添加 </think> 结束标签
-                    thinking_ended = True
-                    delta_content = '\n</think>\n' + delta_content
-
-                # 处理工具调用增量
-                if delta.get('tool_calls'):
-                    for tool_call in delta['tool_calls']:
-                        if tool_call['id'] != '':
-                            tool_id = tool_call['id']
-                        if tool_call['function']['name'] is not None:
-                            tool_name = tool_call['function']['name']
-
-                        if tool_call['type'] is None:
-                            tool_call['type'] = 'function'
-                        tool_call['id'] = tool_id
-                        tool_call['function']['name'] = tool_name
-                        tool_call['function']['arguments'] = (
-                            '' if tool_call['function']['arguments'] is None else tool_call['function']['arguments']
-                        )
-
-                # 跳过空的第一个 chunk（只有 role 没有内容）
-                if chunk_idx == 0 and not delta_content and not reasoning_content and not delta.get('tool_calls'):
-                    chunk_idx += 1
-                    continue
-
-                # 构建 MessageChunk - 只包含增量内容
-                chunk_data = {
-                    'role': role,
-                    'content': delta_content if delta_content else None,
-                    'tool_calls': delta.get('tool_calls'),
-                    'is_final': bool(finish_reason),
-                }
-
-                # 移除 None 值
-                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
-                # return

src/langbot/pkg/provider/modelmgr/requesters/bailianchatcmpl.yaml

@@ -22,11 +22,8 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: openai
   support_type:
   - llm
   - rerank
   provider_category: maas
-execution:
-  python:
-    path: ./bailianchatcmpl.py
-    attr: BailianChatCompletions

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

@@ -1,702 +0,0 @@
-from __future__ import annotations
-
-import asyncio
-import typing
-
-import openai
-import openai.types.chat.chat_completion as chat_completion_module
-import httpx
-
-from .. import errors, requester
-import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
-import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
-import langbot_plugin.api.entities.builtin.provider.message as provider_message
-
-
-class OpenAIChatCompletions(requester.ProviderAPIRequester):
-    """OpenAI ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://api.openai.com/v1',
-        'timeout': 120,
-    }
-
-    async def initialize(self):
-        self.client = openai.AsyncClient(
-            api_key='',
-            base_url=self.requester_cfg['base_url'].replace(' ', ''),
-            timeout=self.requester_cfg['timeout'],
-            http_client=httpx.AsyncClient(trust_env=True, timeout=self.requester_cfg['timeout']),
-        )
-
-    def _mask_api_key(self, api_key: str | None) -> str:
-        if not api_key:
-            return ''
-        if len(api_key) <= 8:
-            return '****'
-        return f'{api_key[:4]}...{api_key[-4:]}'
-
-    def _infer_model_type(self, model_id: str) -> str:
-        normalized_model_id = (model_id or '').lower()
-        embedding_keywords = (
-            'embedding',
-            'embed',
-            'bge-',
-            'e5-',
-            'm3e',
-            'gte-',
-            'multilingual-e5',
-            'text-embedding',
-        )
-        return 'embedding' if any(keyword in normalized_model_id for keyword in embedding_keywords) else 'llm'
-
-    def _infer_model_abilities(self, item: dict[str, typing.Any], model_id: str) -> list[str]:
-        normalized_model_id = (model_id or '').lower()
-        abilities: set[str] = set()
-
-        def _flatten(value: typing.Any) -> list[str]:
-            if value is None:
-                return []
-            if isinstance(value, str):
-                return [value.lower()]
-            if isinstance(value, dict):
-                flattened: list[str] = []
-                for nested_value in value.values():
-                    flattened.extend(_flatten(nested_value))
-                return flattened
-            if isinstance(value, (list, tuple, set)):
-                flattened: list[str] = []
-                for nested_value in value:
-                    flattened.extend(_flatten(nested_value))
-                return flattened
-            return [str(value).lower()]
-
-        capability_tokens = _flatten(item.get('capabilities'))
-        capability_tokens.extend(_flatten(item.get('modalities')))
-        capability_tokens.extend(_flatten(item.get('input_modalities')))
-        capability_tokens.extend(_flatten(item.get('output_modalities')))
-        capability_tokens.extend(_flatten(item.get('supported_generation_methods')))
-        capability_tokens.extend(_flatten(item.get('supported_parameters')))
-        capability_tokens.extend(_flatten(item.get('architecture')))
-
-        combined_tokens = capability_tokens + [normalized_model_id]
-
-        vision_keywords = (
-            'vision',
-            'image',
-            'file',
-            'video',
-            'multimodal',
-            'vl',
-            'ocr',
-            'omni',
-        )
-        function_call_keywords = (
-            'function',
-            'tool',
-            'tools',
-            'tool_choice',
-            'tool_call',
-            'tool-use',
-            'tool_use',
-        )
-
-        if any(any(keyword in token for keyword in vision_keywords) for token in combined_tokens):
-            abilities.add('vision')
-
-        if any(any(keyword in token for keyword in function_call_keywords) for token in combined_tokens):
-            abilities.add('func_call')
-
-        return sorted(abilities)
-
-    def _normalize_modalities(self, value: typing.Any) -> list[str]:
-        normalized: list[str] = []
-
-        def _collect(item: typing.Any):
-            if item is None:
-                return
-            if isinstance(item, str):
-                for part in item.replace('->', ',').replace('+', ',').split(','):
-                    token = part.strip().lower()
-                    if token and token not in normalized:
-                        normalized.append(token)
-                return
-            if isinstance(item, dict):
-                for nested in item.values():
-                    _collect(nested)
-                return
-            if isinstance(item, (list, tuple, set)):
-                for nested in item:
-                    _collect(nested)
-                return
-
-        _collect(value)
-        return normalized
-
-    def _extract_scan_metadata(self, item: dict[str, typing.Any], model_id: str) -> dict[str, typing.Any]:
-        display_name = item.get('name')
-        if not isinstance(display_name, str) or not display_name.strip() or display_name == model_id:
-            display_name = ''
-
-        description = item.get('description')
-        if not isinstance(description, str) or not description.strip():
-            description = ''
-
-        context_length = item.get('context_length')
-        if context_length is None and isinstance(item.get('top_provider'), dict):
-            context_length = item['top_provider'].get('context_length')
-
-        if not isinstance(context_length, int):
-            try:
-                context_length = int(context_length) if context_length is not None else None
-            except (TypeError, ValueError):
-                context_length = None
-
-        input_modalities = self._normalize_modalities(item.get('input_modalities'))
-        output_modalities = self._normalize_modalities(item.get('output_modalities'))
-
-        if isinstance(item.get('architecture'), dict):
-            if not input_modalities:
-                input_modalities = self._normalize_modalities(item['architecture'].get('input_modalities'))
-            if not output_modalities:
-                output_modalities = self._normalize_modalities(item['architecture'].get('output_modalities'))
-
-        owned_by = item.get('owned_by')
-        if not isinstance(owned_by, str) or not owned_by.strip():
-            owned_by = ''
-
-        return {
-            'display_name': display_name or None,
-            'description': description or None,
-            'context_length': context_length,
-            'owned_by': owned_by or None,
-            'input_modalities': input_modalities,
-            'output_modalities': output_modalities,
-        }
-
-    async def scan_models(self, api_key: str | None = None) -> dict[str, typing.Any]:
-        headers = {}
-        if api_key:
-            headers['Authorization'] = f'Bearer {api_key}'
-
-        models_url = f'{self.requester_cfg["base_url"].rstrip("/")}/models'
-        async with httpx.AsyncClient(trust_env=True, timeout=self.requester_cfg['timeout']) as client:
-            response = await client.get(models_url, headers=headers)
-            response.raise_for_status()
-            payload = response.json()
-
-        models = []
-        for item in payload.get('data', []):
-            model_id = item.get('id')
-            if not model_id:
-                continue
-            models.append(
-                {
-                    'id': model_id,
-                    'name': model_id,
-                    'type': self._infer_model_type(model_id),
-                    'abilities': self._infer_model_abilities(item, model_id),
-                    **self._extract_scan_metadata(item, model_id),
-                }
-            )
-
-        models.sort(key=lambda item: (item['type'] != 'llm', item['name'].lower()))
-        return {
-            'models': models,
-            'debug': {
-                'request': {
-                    'method': 'GET',
-                    'url': models_url,
-                    'headers': {
-                        'Authorization': f'Bearer {self._mask_api_key(api_key)}' if api_key else '',
-                    },
-                },
-                'response': payload,
-            },
-        }
-
-    async def _req(
-        self,
-        args: dict,
-        extra_body: dict = {},
-    ) -> chat_completion_module.ChatCompletion:
-        return await self.client.chat.completions.create(**args, extra_body=extra_body)
-
-    async def _req_stream(
-        self,
-        args: dict,
-        extra_body: dict = {},
-    ):
-        async for chunk in await self.client.chat.completions.create(**args, extra_body=extra_body):
-            yield chunk
-
-    async def _make_msg(
-        self,
-        chat_completion: chat_completion_module.ChatCompletion,
-        remove_think: bool = False,
-    ) -> provider_message.Message:
-        if not isinstance(chat_completion, chat_completion_module.ChatCompletion):
-            raise TypeError(f'Expected ChatCompletion, got {type(chat_completion).__name__}: {chat_completion[:16]}')
-
-        chatcmpl_message = chat_completion.choices[0].message.model_dump()
-
-        # 确保 role 字段存在且不为 None
-        if 'role' not in chatcmpl_message or chatcmpl_message['role'] is None:
-            chatcmpl_message['role'] = 'assistant'
-
-        # 处理思维链
-        content = chatcmpl_message.get('content', '')
-        reasoning_content = chatcmpl_message.get('reasoning_content', None)
-
-        processed_content, _ = await self._process_thinking_content(
-            content=content, reasoning_content=reasoning_content, remove_think=remove_think
-        )
-
-        chatcmpl_message['content'] = processed_content
-
-        # 移除 reasoning_content 字段，避免传递给 Message
-        if 'reasoning_content' in chatcmpl_message:
-            del chatcmpl_message['reasoning_content']
-
-        message = provider_message.Message(**chatcmpl_message)
-
-        return message
-
-    async def _process_thinking_content(
-        self,
-        content: str,
-        reasoning_content: str = None,
-        remove_think: bool = False,
-    ) -> tuple[str, str]:
-        """处理思维链内容
-
-        Args:
-            content: 原始内容
-            reasoning_content: reasoning_content 字段内容
-            remove_think: 是否移除思维链
-
-        Returns:
-            (处理后的内容, 提取的思维链内容)
-        """
-        thinking_content = ''
-
-        # 1. 从 reasoning_content 提取思维链
-        if reasoning_content:
-            thinking_content = reasoning_content
-
-        # 2. 从 content 中提取 <think> 标签内容
-        if content and '<think>' in content and '</think>' in content:
-            import re
-
-            think_pattern = r'<think>(.*?)</think>'
-            think_matches = re.findall(think_pattern, content, re.DOTALL)
-            if think_matches:
-                # 如果已有 reasoning_content，则追加
-                if thinking_content:
-                    thinking_content += '\n' + '\n'.join(think_matches)
-                else:
-                    thinking_content = '\n'.join(think_matches)
-                # 移除 content 中的 <think> 标签
-                content = re.sub(think_pattern, '', content, flags=re.DOTALL).strip()
-
-        # 3. 根据 remove_think 参数决定是否保留思维链
-        if remove_think:
-            return content, ''
-        else:
-            # 如果有思维链内容，将其以 <think> 格式添加到 content 开头
-            if thinking_content:
-                content = f'<think>\n{thinking_content}\n</think>\n{content}'.strip()
-            return content, thinking_content
-
-    async def _closure_stream(
-        self,
-        query: pipeline_query.Query,
-        req_messages: list[dict],
-        use_model: requester.RuntimeLLMModel,
-        use_funcs: list[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> provider_message.MessageChunk:
-        self.client.api_key = use_model.provider.token_mgr.get_token()
-
-        args = {}
-        args['model'] = use_model.model_entity.name
-
-        if use_funcs:
-            tools = await self.ap.tool_mgr.generate_tools_for_openai(use_funcs)
-            if tools:
-                args['tools'] = tools
-
-        # 设置此次请求中的messages
-        messages = req_messages.copy()
-
-        # 检查vision
-        for msg in messages:
-            if 'content' in msg and isinstance(msg['content'], list):
-                for me in msg['content']:
-                    if me['type'] == 'image_base64':
-                        me['image_url'] = {'url': me['image_base64']}
-                        me['type'] = 'image_url'
-                        del me['image_base64']
-
-        args['messages'] = messages
-        args['stream'] = True
-
-        # 流式处理状态
-        # tool_calls_map: dict[str, provider_message.ToolCall] = {}
-        chunk_idx = 0
-        thinking_started = False
-        thinking_ended = False
-        role = 'assistant'  # 默认角色
-        tool_id = ''
-        tool_name = ''
-        # accumulated_reasoning = ''  # 仅用于判断何时结束思维链
-
-        async for chunk in self._req_stream(args, extra_body=extra_args):
-            # 解析 chunk 数据
-
-            if hasattr(chunk, 'choices') and chunk.choices:
-                choice = chunk.choices[0]
-                delta = choice.delta.model_dump() if hasattr(choice, 'delta') else {}
-
-                finish_reason = getattr(choice, 'finish_reason', None)
-            else:
-                delta = {}
-                finish_reason = None
-            # 从第一个 chunk 获取 role，后续使用这个 role
-            if 'role' in delta and delta['role']:
-                role = delta['role']
-
-            # 获取增量内容
-            delta_content = delta.get('content', '')
-            reasoning_content = delta.get('reasoning_content', '')
-
-            # 处理 reasoning_content
-            if reasoning_content:
-                # accumulated_reasoning += reasoning_content
-                # 如果设置了 remove_think，跳过 reasoning_content
-                if remove_think:
-                    chunk_idx += 1
-                    continue
-
-                # 第一次出现 reasoning_content，添加 <think> 开始标签
-                if not thinking_started:
-                    thinking_started = True
-                    delta_content = '<think>\n' + reasoning_content
-                else:
-                    # 继续输出 reasoning_content
-                    delta_content = reasoning_content
-            elif thinking_started and not thinking_ended and delta_content:
-                # reasoning_content 结束，normal content 开始，添加 </think> 结束标签
-                thinking_ended = True
-                delta_content = '\n</think>\n' + delta_content
-
-            # 处理 content 中已有的 <think> 标签（如果需要移除）
-            # if delta_content and remove_think and '<think>' in delta_content:
-            #     import re
-            #
-            #     # 移除 <think> 标签及其内容
-            #     delta_content = re.sub(r'<think>.*?</think>', '', delta_content, flags=re.DOTALL)
-
-            # 处理工具调用增量
-            # delta_tool_calls = None
-            if delta.get('tool_calls'):
-                for tool_call in delta['tool_calls']:
-                    if tool_call['id'] and tool_call['function']['name']:
-                        tool_id = tool_call['id']
-                        tool_name = tool_call['function']['name']
-                    else:
-                        tool_call['id'] = tool_id
-                        tool_call['function']['name'] = tool_name
-                    if tool_call['type'] is None:
-                        tool_call['type'] = 'function'
-
-            # 跳过空的第一个 chunk（只有 role 没有内容）
-            if chunk_idx == 0 and not delta_content and not reasoning_content and not delta.get('tool_calls'):
-                chunk_idx += 1
-                continue
-            # 构建 MessageChunk - 只包含增量内容
-            chunk_data = {
-                'role': role,
-                'content': delta_content if delta_content else None,
-                'tool_calls': delta.get('tool_calls'),
-                'is_final': bool(finish_reason),
-            }
-
-            # 移除 None 值
-            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
-
-    async def _closure(
-        self,
-        query: pipeline_query.Query,
-        req_messages: list[dict],
-        use_model: requester.RuntimeLLMModel,
-        use_funcs: list[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> tuple[provider_message.Message, dict]:
-        self.client.api_key = use_model.provider.token_mgr.get_token()
-
-        args = {}
-        args['model'] = use_model.model_entity.name
-
-        if use_funcs:
-            tools = await self.ap.tool_mgr.generate_tools_for_openai(use_funcs)
-
-            if tools:
-                args['tools'] = tools
-
-        # 设置此次请求中的messages
-        messages = req_messages.copy()
-
-        # 检查vision
-        for msg in messages:
-            if 'content' in msg and isinstance(msg['content'], list):
-                for me in msg['content']:
-                    if me['type'] == 'image_base64':
-                        me['image_url'] = {'url': me['image_base64']}
-                        me['type'] = 'image_url'
-                        del me['image_base64']
-
-        args['messages'] = messages
-
-        # 发送请求
-
-        resp = await self._req(args, extra_body=extra_args)
-        # 处理请求结果
-        message = await self._make_msg(resp, remove_think)
-
-        # Extract token usage from response
-        usage_info = {}
-        if hasattr(resp, 'usage') and resp.usage:
-            usage_info['input_tokens'] = resp.usage.prompt_tokens or 0
-            usage_info['output_tokens'] = resp.usage.completion_tokens or 0
-            usage_info['total_tokens'] = resp.usage.total_tokens or 0
-
-        return message, usage_info
-
-    async def invoke_llm(
-        self,
-        query: pipeline_query.Query,
-        model: requester.RuntimeLLMModel,
-        messages: typing.List[provider_message.Message],
-        funcs: typing.List[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> tuple[provider_message.Message, dict]:
-        """Invoke LLM and return message with usage info"""
-        req_messages = []  # req_messages 仅用于类内，外部同步由 query.messages 进行
-        for m in messages:
-            msg_dict = m.dict(exclude_none=True)
-            content = msg_dict.get('content')
-            if isinstance(content, list):
-                # 检查 content 列表中是否每个部分都是文本
-                if all(isinstance(part, dict) and part.get('type') == 'text' for part in content):
-                    # 将所有文本部分合并为一个字符串
-                    msg_dict['content'] = '\n'.join(part['text'] for part in content)
-            req_messages.append(msg_dict)
-
-        try:
-            msg, usage_info = await self._closure(
-                query=query,
-                req_messages=req_messages,
-                use_model=model,
-                use_funcs=funcs,
-                extra_args=extra_args,
-                remove_think=remove_think,
-            )
-            return msg, usage_info
-        except asyncio.TimeoutError:
-            raise errors.RequesterError('请求超时')
-        except openai.BadRequestError as e:
-            error_message = str(e.message) if hasattr(e, 'message') else str(e)
-            if 'context_length_exceeded' in str(e):
-                raise errors.RequesterError(f'上文过长，请重置会话: {error_message}')
-            else:
-                raise errors.RequesterError(f'请求参数错误: {error_message}')
-        except openai.AuthenticationError as e:
-            error_message = str(e.message) if hasattr(e, 'message') else str(e)
-            raise errors.RequesterError(f'无效的 api-key: {error_message}')
-        except openai.NotFoundError as e:
-            error_message = str(e.message) if hasattr(e, 'message') else str(e)
-            raise errors.RequesterError(f'请求路径错误: {error_message}')
-        except openai.RateLimitError as e:
-            error_message = str(e.message) if hasattr(e, 'message') else str(e)
-            raise errors.RequesterError(f'请求过于频繁或余额不足: {error_message}')
-        except openai.APIConnectionError as e:
-            error_message = f'连接错误: {str(e)}'
-            raise errors.RequesterError(error_message)
-        except openai.APIError as e:
-            error_message = str(e.message) if hasattr(e, 'message') else str(e)
-            raise errors.RequesterError(f'请求错误: {error_message}')
-
-    async def invoke_embedding(
-        self,
-        model: requester.RuntimeEmbeddingModel,
-        input_text: list[str],
-        extra_args: dict[str, typing.Any] = {},
-    ) -> tuple[list[list[float]], dict]:
-        """调用 Embedding API, returns (embeddings, usage_info)"""
-        self.client.api_key = model.provider.token_mgr.get_token()
-
-        args = {
-            'model': model.model_entity.name,
-            'input': input_text,
-        }
-
-        if model.model_entity.extra_args:
-            args.update(model.model_entity.extra_args)
-
-        args.update(extra_args)
-
-        try:
-            resp = await self.client.embeddings.create(**args)
-
-            # Extract usage info
-            usage_info = {}
-            if hasattr(resp, 'usage') and resp.usage:
-                usage_info['prompt_tokens'] = resp.usage.prompt_tokens or 0
-                usage_info['total_tokens'] = resp.usage.total_tokens or 0
-
-            return [d.embedding for d in resp.data], usage_info
-        except asyncio.TimeoutError:
-            raise errors.RequesterError('请求超时')
-        except openai.BadRequestError as e:
-            raise errors.RequesterError(f'请求参数错误: {e.message}')
-
-    async def invoke_llm_stream(
-        self,
-        query: pipeline_query.Query,
-        model: requester.RuntimeLLMModel,
-        messages: typing.List[provider_message.Message],
-        funcs: typing.List[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> provider_message.MessageChunk:
-        req_messages = []  # req_messages 仅用于类内，外部同步由 query.messages 进行
-        for m in messages:
-            msg_dict = m.dict(exclude_none=True)
-            content = msg_dict.get('content')
-            if isinstance(content, list):
-                # 检查 content 列表中是否每个部分都是文本
-                if all(isinstance(part, dict) and part.get('type') == 'text' for part in content):
-                    # 将所有文本部分合并为一个字符串
-                    msg_dict['content'] = '\n'.join(part['text'] for part in content)
-            req_messages.append(msg_dict)
-
-        try:
-            async for item in self._closure_stream(
-                query=query,
-                req_messages=req_messages,
-                use_model=model,
-                use_funcs=funcs,
-                extra_args=extra_args,
-                remove_think=remove_think,
-            ):
-                yield item
-
-        except asyncio.TimeoutError:
-            raise errors.RequesterError('请求超时')
-        except openai.BadRequestError as e:
-            if 'context_length_exceeded' in e.message:
-                raise errors.RequesterError(f'上文过长，请重置会话: {e.message}')
-            else:
-                raise errors.RequesterError(f'请求参数错误: {e.message}')
-        except openai.AuthenticationError as e:
-            raise errors.RequesterError(f'无效的 api-key: {e.message}')
-        except openai.NotFoundError as e:
-            raise errors.RequesterError(f'请求路径错误: {e.message}')
-        except openai.RateLimitError as e:
-            raise errors.RequesterError(f'请求过于频繁或余额不足: {e.message}')
-        except openai.APIError as e:
-            raise errors.RequesterError(f'请求错误: {e.message}')
-
-    async def invoke_rerank(
-        self,
-        model: requester.RuntimeRerankModel,
-        query: str,
-        documents: typing.List[str],
-        extra_args: dict[str, typing.Any] = {},
-    ) -> typing.List[dict]:
-        """Standard /rerank endpoint (Jina/Cohere/SiliconFlow/Voyage/DashScope compatible)
-
-        Supports extra_args from model.extra_args:
-        - rerank_url: full URL override (e.g. "https://dashscope.aliyuncs.com/compatible-api/v1/reranks")
-        - rerank_path: path override appended to base_url (e.g. "reranks" instead of default "rerank")
-        - Any other fields are merged into the request payload.
-        """
-        api_key = model.provider.token_mgr.get_token()
-        base_url = self.requester_cfg.get('base_url', '').rstrip('/')
-        timeout = self.requester_cfg.get('timeout', 120)
-
-        merged_args = {}
-        if model.model_entity.extra_args:
-            merged_args.update(model.model_entity.extra_args)
-        if extra_args:
-            merged_args.update(extra_args)
-
-        rerank_url = merged_args.pop('rerank_url', None)
-        rerank_path = merged_args.pop('rerank_path', 'rerank')
-        if not rerank_url:
-            rerank_url = f'{base_url}/{rerank_path}'
-
-        headers = {
-            'Content-Type': 'application/json',
-            'Authorization': f'Bearer {api_key}',
-        }
-
-        payload = {
-            'model': model.model_entity.name,
-            'query': query,
-            'documents': documents[:64],
-            'top_n': min(len(documents), 64),
-        }
-
-        if merged_args:
-            payload.update(merged_args)
-
-        try:
-            async with httpx.AsyncClient(trust_env=True, timeout=timeout) as client:
-                resp = await client.post(rerank_url, headers=headers, json=payload)
-                resp.raise_for_status()
-                data = resp.json()
-
-            results = self._parse_rerank_response(data)
-
-            if results:
-                scores = [r.get('relevance_score', 0.0) for r in results]
-                min_score = min(scores)
-                max_score = max(scores)
-                if max_score - min_score > 1e-6:
-                    for r in results:
-                        r['relevance_score'] = (r['relevance_score'] - min_score) / (max_score - min_score)
-
-            return results
-        except httpx.HTTPStatusError as e:
-            raise errors.RequesterError(f'Rerank request failed: {e.response.status_code} - {e.response.text}')
-        except httpx.TimeoutException:
-            raise errors.RequesterError('Rerank request timed out')
-        except Exception as e:
-            raise errors.RequesterError(f'Rerank request error: {str(e)}')
-
-    @staticmethod
-    def _parse_rerank_response(data: dict) -> typing.List[dict]:
-        """Parse rerank response from various providers.
-
-        Handles:
-        - Jina/Cohere/SiliconFlow: {"results": [{"index", "relevance_score"}]}
-        - Voyage AI: {"data": [{"index", "relevance_score"}]}
-        - DashScope: {"output": {"results": [{"index", "relevance_score"}]}}
-        """
-        if 'results' in data:
-            return data['results']
-        if 'data' in data:
-            return data['data']
-        if 'output' in data and isinstance(data['output'], dict):
-            return data['output'].get('results', [])
-        return []

src/langbot/pkg/provider/modelmgr/requesters/chatcmpl.yaml

@@ -22,12 +22,10 @@ spec:
     type: integer
     required: true
     default: 120
+  # LiteLLM provider prefix - when set, uses unified LiteLLMRequester
+  litellm_provider: openai
   support_type:
   - llm
   - text-embedding
   - rerank
   provider_category: manufacturer
-execution:
-  python:
-    path: ./chatcmpl.py
-    attr: OpenAIChatCompletions

src/langbot/pkg/provider/modelmgr/requesters/coherererank.yaml

@@ -22,10 +22,7 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: cohere
   support_type:
   - rerank
   provider_category: manufacturer
-execution:
-  python:
-    path: ./chatcmpl.py
-    attr: OpenAIChatCompletions

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

@@ -1,17 +0,0 @@
-from __future__ import annotations
-
-import typing
-import openai
-
-from . import chatcmpl
-
-
-class CompShareChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """CompShare ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://api.modelverse.cn/v1',
-        'timeout': 120,
-    }

src/langbot/pkg/provider/modelmgr/requesters/compsharechatcmpl.yaml

@@ -22,10 +22,7 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: openai
   support_type:
   - llm
   provider_category: maas
-execution:
-  python:
-    path: ./compsharechatcmpl.py
-    attr: CompShareChatCompletions

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

@@ -1,67 +0,0 @@
-from __future__ import annotations
-
-import typing
-
-from . import chatcmpl
-from .. import errors, requester
-import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
-import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
-import langbot_plugin.api.entities.builtin.provider.message as provider_message
-
-
-class DeepseekChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """Deepseek ChatCompletion API 请求器"""
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://api.deepseek.com',
-        'timeout': 120,
-    }
-
-    async def _closure(
-        self,
-        query: pipeline_query.Query,
-        req_messages: list[dict],
-        use_model: requester.RuntimeLLMModel,
-        use_funcs: list[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> tuple[provider_message.Message, dict]:
-        self.client.api_key = use_model.provider.token_mgr.get_token()
-
-        args = {}
-        args['model'] = use_model.model_entity.name
-
-        if use_funcs:
-            tools = await self.ap.tool_mgr.generate_tools_for_openai(use_funcs)
-
-            if tools:
-                args['tools'] = tools
-
-        # 设置此次请求中的messages
-        messages = req_messages
-
-        # deepseek 不支持多模态，把content都转换成纯文字
-        for m in messages:
-            if 'content' in m and isinstance(m['content'], list):
-                m['content'] = ' '.join([c['text'] for c in m['content'] if 'text' in c])
-
-        args['messages'] = messages
-
-        # 发送请求
-        resp = await self._req(args, extra_body=extra_args)
-
-        # print(resp)
-
-        if resp is None:
-            raise errors.RequesterError('接口返回为空，请确定模型提供商服务是否正常')
-        # 处理请求结果
-        message = await self._make_msg(resp, remove_think)
-
-        # Extract token usage from response
-        usage_info = {}
-        if hasattr(resp, 'usage') and resp.usage:
-            usage_info['input_tokens'] = resp.usage.prompt_tokens or 0
-            usage_info['output_tokens'] = resp.usage.completion_tokens or 0
-            usage_info['total_tokens'] = resp.usage.total_tokens or 0
-
-        return message, usage_info

src/langbot/pkg/provider/modelmgr/requesters/deepseekchatcmpl.yaml

@@ -22,10 +22,7 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: deepseek
   support_type:
   - llm
   provider_category: manufacturer
-execution:
-  python:
-    path: ./deepseekchatcmpl.py
-    attr: DeepseekChatCompletions

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

@@ -1,205 +0,0 @@
-from __future__ import annotations
-
-import typing
-import httpx
-
-from . import chatcmpl
-
-import uuid
-
-from .. import requester
-import langbot_plugin.api.entities.builtin.provider.message as provider_message
-import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
-import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
-
-
-class GeminiChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """Google Gemini API 请求器"""
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://generativelanguage.googleapis.com/v1beta/openai',
-        'timeout': 120,
-    }
-
-    async def scan_models(self, api_key: str | None = None) -> dict[str, typing.Any]:
-        models_url = 'https://generativelanguage.googleapis.com/v1beta/models'
-        params = {'key': api_key} if api_key else {}
-
-        all_models: list[dict[str, typing.Any]] = []
-        next_page_token = ''
-        last_payload: dict[str, typing.Any] = {}
-
-        async with httpx.AsyncClient(trust_env=True, timeout=self.requester_cfg['timeout']) as client:
-            while True:
-                request_params = dict(params)
-                if next_page_token:
-                    request_params['pageToken'] = next_page_token
-
-                response = await client.get(models_url, params=request_params)
-                response.raise_for_status()
-                payload = response.json()
-                last_payload = payload
-
-                for item in payload.get('models', []):
-                    model_name = item.get('name', '')
-                    model_id = model_name.replace('models/', '', 1)
-                    if not model_id:
-                        continue
-
-                    supported_methods = item.get('supportedGenerationMethods', []) or []
-                    if 'embedContent' in supported_methods and 'generateContent' not in supported_methods:
-                        model_type = 'embedding'
-                    else:
-                        model_type = 'llm'
-
-                    all_models.append(
-                        {
-                            'id': model_id,
-                            'name': model_id,
-                            'type': model_type,
-                            'abilities': self._infer_model_abilities(item, model_id),
-                            'display_name': item.get('displayName') or None,
-                            'description': item.get('description') or None,
-                            'context_length': item.get('inputTokenLimit'),
-                            'input_modalities': self._normalize_modalities(item.get('inputModalities')),
-                            'output_modalities': self._normalize_modalities(item.get('outputModalities')),
-                        }
-                    )
-
-                next_page_token = payload.get('nextPageToken', '')
-                if not next_page_token:
-                    break
-
-        all_models.sort(key=lambda item: (item['type'] != 'llm', item['name'].lower()))
-        return {
-            'models': all_models,
-            'debug': {
-                'request': {
-                    'method': 'GET',
-                    'url': models_url,
-                    'query': {'key': self._mask_api_key(api_key)} if api_key else {},
-                },
-                'response': last_payload,
-            },
-        }
-
-    async def _closure_stream(
-        self,
-        query: pipeline_query.Query,
-        req_messages: list[dict],
-        use_model: requester.RuntimeLLMModel,
-        use_funcs: list[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> provider_message.MessageChunk:
-        self.client.api_key = use_model.provider.token_mgr.get_token()
-
-        args = {}
-        args['model'] = use_model.model_entity.name
-
-        if use_funcs:
-            tools = await self.ap.tool_mgr.generate_tools_for_openai(use_funcs)
-            if tools:
-                args['tools'] = tools
-
-        # 设置此次请求中的messages
-        messages = req_messages.copy()
-
-        # 检查vision
-        for msg in messages:
-            if 'content' in msg and isinstance(msg['content'], list):
-                for me in msg['content']:
-                    if me['type'] == 'image_base64':
-                        me['image_url'] = {'url': me['image_base64']}
-                        me['type'] = 'image_url'
-                        del me['image_base64']
-
-        args['messages'] = messages
-        args['stream'] = True
-
-        # 流式处理状态
-        # tool_calls_map: dict[str, provider_message.ToolCall] = {}
-        chunk_idx = 0
-        thinking_started = False
-        thinking_ended = False
-        role = 'assistant'  # 默认角色
-        tool_id = ''
-        tool_name = ''
-        # accumulated_reasoning = ''  # 仅用于判断何时结束思维链
-
-        async for chunk in self._req_stream(args, extra_body=extra_args):
-            # 解析 chunk 数据
-
-            if hasattr(chunk, 'choices') and chunk.choices:
-                choice = chunk.choices[0]
-                delta = choice.delta.model_dump() if hasattr(choice, 'delta') else {}
-
-                finish_reason = getattr(choice, 'finish_reason', None)
-            else:
-                delta = {}
-                finish_reason = None
-            # 从第一个 chunk 获取 role，后续使用这个 role
-            if 'role' in delta and delta['role']:
-                role = delta['role']
-
-            # 获取增量内容
-            delta_content = delta.get('content', '')
-            reasoning_content = delta.get('reasoning_content', '')
-
-            # 处理 reasoning_content
-            if reasoning_content:
-                # accumulated_reasoning += reasoning_content
-                # 如果设置了 remove_think，跳过 reasoning_content
-                if remove_think:
-                    chunk_idx += 1
-                    continue
-
-                # 第一次出现 reasoning_content，添加 <think> 开始标签
-                if not thinking_started:
-                    thinking_started = True
-                    delta_content = '<think>\n' + reasoning_content
-                else:
-                    # 继续输出 reasoning_content
-                    delta_content = reasoning_content
-            elif thinking_started and not thinking_ended and delta_content:
-                # reasoning_content 结束，normal content 开始，添加 </think> 结束标签
-                thinking_ended = True
-                delta_content = '\n</think>\n' + delta_content
-
-            # 处理 content 中已有的 <think> 标签（如果需要移除）
-            # if delta_content and remove_think and '<think>' in delta_content:
-            #     import re
-            #
-            #     # 移除 <think> 标签及其内容
-            #     delta_content = re.sub(r'<think>.*?</think>', '', delta_content, flags=re.DOTALL)
-
-            # 处理工具调用增量
-            # delta_tool_calls = None
-            if delta.get('tool_calls'):
-                for tool_call in delta['tool_calls']:
-                    if tool_call['id'] == '' and tool_id == '':
-                        tool_id = str(uuid.uuid4())
-                    if tool_call['function']['name']:
-                        tool_name = tool_call['function']['name']
-                    tool_call['id'] = tool_id
-                    tool_call['function']['name'] = tool_name
-                    if tool_call['type'] is None:
-                        tool_call['type'] = 'function'
-
-            # 跳过空的第一个 chunk（只有 role 没有内容）
-            if chunk_idx == 0 and not delta_content and not reasoning_content and not delta.get('tool_calls'):
-                chunk_idx += 1
-                continue
-            # 构建 MessageChunk - 只包含增量内容
-            chunk_data = {
-                'role': role,
-                'content': delta_content if delta_content else None,
-                'tool_calls': delta.get('tool_calls'),
-                'is_final': bool(finish_reason),
-            }
-
-            # 移除 None 值
-            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

src/langbot/pkg/provider/modelmgr/requesters/geminichatcmpl.yaml

@@ -22,10 +22,7 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: gemini
   support_type:
   - llm
   provider_category: manufacturer
-execution:
-  python:
-    path: ./geminichatcmpl.py
-    attr: GeminiChatCompletions

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

@@ -1,15 +0,0 @@
-from __future__ import annotations
-
-
-import typing
-
-from . import ppiochatcmpl
-
-
-class GiteeAIChatCompletions(ppiochatcmpl.PPIOChatCompletions):
-    """Gitee AI ChatCompletions API 请求器"""
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://ai.gitee.com/v1',
-        'timeout': 120,
-    }

src/langbot/pkg/provider/modelmgr/requesters/giteeaichatcmpl.yaml

@@ -22,12 +22,9 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: openai
   support_type:
   - llm
   - text-embedding
   - rerank
   provider_category: maas
-execution:
-  python:
-    path: ./giteeaichatcmpl.py
-    attr: GiteeAIChatCompletions

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

@@ -1,208 +0,0 @@
-from __future__ import annotations
-
-import openai
-import typing
-
-from . import chatcmpl
-from .. import requester
-import openai.types.chat.chat_completion as chat_completion
-import re
-import langbot_plugin.api.entities.builtin.provider.message as provider_message
-import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
-import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
-
-
-class JieKouAIChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """接口 AI ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://api.jiekou.ai/openai',
-        'timeout': 120,
-    }
-
-    is_think: bool = False
-
-    async def _make_msg(
-        self,
-        chat_completion: chat_completion.ChatCompletion,
-        remove_think: bool,
-    ) -> provider_message.Message:
-        chatcmpl_message = chat_completion.choices[0].message.model_dump()
-        # print(chatcmpl_message.keys(), chatcmpl_message.values())
-
-        # 确保 role 字段存在且不为 None
-        if 'role' not in chatcmpl_message or chatcmpl_message['role'] is None:
-            chatcmpl_message['role'] = 'assistant'
-
-        reasoning_content = chatcmpl_message['reasoning_content'] if 'reasoning_content' in chatcmpl_message else None
-
-        # deepseek的reasoner模型
-        chatcmpl_message['content'] = await self._process_thinking_content(
-            chatcmpl_message['content'], reasoning_content, remove_think
-        )
-
-        # 移除 reasoning_content 字段，避免传递给 Message
-        if 'reasoning_content' in chatcmpl_message:
-            del chatcmpl_message['reasoning_content']
-
-        message = provider_message.Message(**chatcmpl_message)
-
-        return message
-
-    async def _process_thinking_content(
-        self,
-        content: str,
-        reasoning_content: str = None,
-        remove_think: bool = False,
-    ) -> tuple[str, str]:
-        """处理思维链内容
-
-        Args:
-            content: 原始内容
-            reasoning_content: reasoning_content 字段内容
-            remove_think: 是否移除思维链
-
-        Returns:
-            处理后的内容
-        """
-        if remove_think:
-            content = re.sub(r'<think>.*?</think>', '', content, flags=re.DOTALL)
-        else:
-            if reasoning_content is not None:
-                content = '<think>\n' + reasoning_content + '\n</think>\n' + content
-        return content
-
-    async def _make_msg_chunk(
-        self,
-        delta: dict[str, typing.Any],
-        idx: int,
-    ) -> provider_message.MessageChunk:
-        # 处理流式chunk和完整响应的差异
-        # print(chat_completion.choices[0])
-
-        # 确保 role 字段存在且不为 None
-        if 'role' not in delta or delta['role'] is None:
-            delta['role'] = 'assistant'
-
-        reasoning_content = delta['reasoning_content'] if 'reasoning_content' in delta else None
-
-        delta['content'] = '' if delta['content'] is None else delta['content']
-        # print(reasoning_content)
-
-        # deepseek的reasoner模型
-
-        if reasoning_content is not None:
-            delta['content'] += reasoning_content
-
-        message = provider_message.MessageChunk(**delta)
-
-        return message
-
-    async def _closure_stream(
-        self,
-        query: pipeline_query.Query,
-        req_messages: list[dict],
-        use_model: requester.RuntimeLLMModel,
-        use_funcs: list[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> provider_message.Message | typing.AsyncGenerator[provider_message.MessageChunk, None]:
-        self.client.api_key = use_model.provider.token_mgr.get_token()
-
-        args = {}
-        args['model'] = use_model.model_entity.name
-
-        if use_funcs:
-            tools = await self.ap.tool_mgr.generate_tools_for_openai(use_funcs)
-
-            if tools:
-                args['tools'] = tools
-
-        # 设置此次请求中的messages
-        messages = req_messages.copy()
-
-        # 检查vision
-        for msg in messages:
-            if 'content' in msg and isinstance(msg['content'], list):
-                for me in msg['content']:
-                    if me['type'] == 'image_base64':
-                        me['image_url'] = {'url': me['image_base64']}
-                        me['type'] = 'image_url'
-                        del me['image_base64']
-
-        args['messages'] = messages
-        args['stream'] = True
-
-        # tool_calls_map: dict[str, provider_message.ToolCall] = {}
-        chunk_idx = 0
-        thinking_started = False
-        thinking_ended = False
-        role = 'assistant'  # 默认角色
-        async for chunk in self._req_stream(args, extra_body=extra_args):
-            # 解析 chunk 数据
-            if hasattr(chunk, 'choices') and chunk.choices:
-                choice = chunk.choices[0]
-                delta = choice.delta.model_dump() if hasattr(choice, 'delta') else {}
-                finish_reason = getattr(choice, 'finish_reason', None)
-            else:
-                delta = {}
-                finish_reason = None
-
-            # 从第一个 chunk 获取 role，后续使用这个 role
-            if 'role' in delta and delta['role']:
-                role = delta['role']
-
-            # 获取增量内容
-            delta_content = delta.get('content', '')
-            # reasoning_content = delta.get('reasoning_content', '')
-
-            if remove_think:
-                if delta['content'] is not None:
-                    if '<think>' in delta['content'] and not thinking_started and not thinking_ended:
-                        thinking_started = True
-                        continue
-                    elif delta['content'] == r'</think>' and not thinking_ended:
-                        thinking_ended = True
-                        continue
-                    elif thinking_ended and delta['content'] == '\n\n' and thinking_started:
-                        thinking_started = False
-                        continue
-                    elif thinking_started and not thinking_ended:
-                        continue
-
-            # delta_tool_calls = None
-            if delta.get('tool_calls'):
-                for tool_call in delta['tool_calls']:
-                    if tool_call['id'] and tool_call['function']['name']:
-                        tool_id = tool_call['id']
-                        tool_name = tool_call['function']['name']
-
-                    if tool_call['id'] is None:
-                        tool_call['id'] = tool_id
-                    if tool_call['function']['name'] is None:
-                        tool_call['function']['name'] = tool_name
-                    if tool_call['function']['arguments'] is None:
-                        tool_call['function']['arguments'] = ''
-                    if tool_call['type'] is None:
-                        tool_call['type'] = 'function'
-
-            # 跳过空的第一个 chunk（只有 role 没有内容）
-            if chunk_idx == 0 and not delta_content and not delta.get('tool_calls'):
-                chunk_idx += 1
-                continue
-
-            # 构建 MessageChunk - 只包含增量内容
-            chunk_data = {
-                'role': role,
-                'content': delta_content if delta_content else None,
-                'tool_calls': delta.get('tool_calls'),
-                'is_final': bool(finish_reason),
-            }
-
-            # 移除 None 值
-            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

src/langbot/pkg/provider/modelmgr/requesters/jiekouaichatcmpl.yaml

@@ -29,11 +29,8 @@ spec:
     type: int
     required: true
     default: 120
+  litellm_provider: openai
   support_type:
   - llm
   - text-embedding
   provider_category: maas
-execution:
-  python:
-    path: ./jiekouaichatcmpl.py
-    attr: JieKouAIChatCompletions

src/langbot/pkg/provider/modelmgr/requesters/jinarerank.yaml

@@ -22,10 +22,7 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: jina
   support_type:
   - rerank
   provider_category: manufacturer
-execution:
-  python:
-    path: ./chatcmpl.py
-    attr: OpenAIChatCompletions

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

@@ -0,0 +1,394 @@
+"""LiteLLM unified requester for chat, embedding, and rerank."""
+
+from __future__ import annotations
+
+import typing
+
+import litellm
+from litellm import acompletion, aembedding, arerank
+
+from .. import errors, requester
+import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
+import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
+import langbot_plugin.api.entities.builtin.provider.message as provider_message
+
+
+class LiteLLMRequester(requester.ProviderAPIRequester):
+    """LiteLLM unified API requester supporting chat, embedding, and rerank."""
+
+    default_config: dict[str, typing.Any] = {
+        'base_url': '',
+        'timeout': 120,
+        'custom_llm_provider': '',
+        'drop_params': False,
+        'num_retries': 0,
+        'api_version': '',
+    }
+
+    async def initialize(self):
+        """Initialize LiteLLM client settings."""
+        # LiteLLM doesn't require explicit client initialization
+        # Configuration is passed per-request via litellm params
+        pass
+
+    def _build_litellm_model_name(self, model_name: str, custom_llm_provider: str | None = None) -> str:
+        """Build LiteLLM model name with provider prefix if needed."""
+        provider = custom_llm_provider or self.requester_cfg.get('custom_llm_provider', '')
+        if provider:
+            # LiteLLM format: provider/model_name
+            return f'{provider}/{model_name}'
+        # If no custom provider, assume model_name already includes prefix or is OpenAI-compatible
+        return model_name
+
+    def _convert_messages(self, messages: typing.List[provider_message.Message]) -> 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 isinstance(content, list):
+                for part in content:
+                    if isinstance(part, dict) and part.get('type') == 'image_base64':
+                        part['image_url'] = {'url': part['image_base64']}
+                        part['type'] = 'image_url'
+                        del part['image_base64']
+
+            req_messages.append(msg_dict)
+
+        return req_messages
+
+    def _process_thinking_content(self, content: str, reasoning_content: str | None, remove_think: bool) -> str:
+        """Process thinking/reasoning content.
+
+        Args:
+            content: The main content from response
+            reasoning_content: Separate reasoning content from model
+            remove_think: If True, remove thinking markers; if False, preserve them
+
+        Returns:
+            Processed content string
+        """
+        # Extract and handle thinking tags
+        if content and 'CRETIRE_REASONING_BEGINk' in content and 'CRETIRE_REASONING_ENDk' in content:
+            import re
+
+            think_pattern = r'CRETIRE_REASONING_BEGINk(.*?)CRETIRE_REASONING_ENDk'
+
+            if remove_think:
+                # Remove thinking tags and their content from output
+                content = re.sub(think_pattern, '', content, flags=re.DOTALL).strip()
+            # else: preserve thinking content as-is
+
+        # Handle separate reasoning_content field
+        # Currently we don't include reasoning_content in user-facing output regardless of remove_think
+        # because it's typically internal model reasoning, not user-visible thinking
+        return content or ''
+
+    def _extract_usage(self, response) -> dict:
+        """Extract usage info from LiteLLM response."""
+        usage = response.usage
+        usage_info = {
+            'prompt_tokens': usage.prompt_tokens or 0,
+            'completion_tokens': usage.completion_tokens or 0,
+            'total_tokens': usage.total_tokens or 0,
+        }
+        # TODO: LangBot internal inconsistency - LLM monitoring uses input_tokens/output_tokens,
+        # while embedding monitoring uses prompt_tokens. Should unify in requester.py to use
+        # prompt_tokens (OpenAI native) consistently. After that, remove these compatibility aliases.
+        usage_info['input_tokens'] = usage_info['prompt_tokens']
+        usage_info['output_tokens'] = usage_info['completion_tokens']
+        return usage_info
+
+    def _build_common_args(self, args: dict, include_retry_params: bool = True) -> dict:
+        """Apply common requester config to args dict."""
+        if self.requester_cfg.get('base_url'):
+            args['api_base'] = self.requester_cfg['base_url']
+        if self.requester_cfg.get('timeout'):
+            args['timeout'] = self.requester_cfg['timeout']
+        if include_retry_params:
+            if self.requester_cfg.get('drop_params'):
+                args['drop_params'] = self.requester_cfg['drop_params']
+            if self.requester_cfg.get('num_retries'):
+                args['num_retries'] = self.requester_cfg['num_retries']
+            if self.requester_cfg.get('api_version'):
+                args['api_version'] = self.requester_cfg['api_version']
+        return args
+
+    def _handle_litellm_error(self, e: Exception) -> None:
+        """Convert LiteLLM exceptions to RequesterError. Never returns, always raises."""
+        # Check more specific exceptions first (they inherit from base exceptions)
+        if isinstance(e, litellm.ContextWindowExceededError):
+            raise errors.RequesterError(f'上下文长度超限: {str(e)}')
+        if isinstance(e, litellm.BadRequestError):
+            raise errors.RequesterError(f'请求参数错误: {str(e)}')
+        if isinstance(e, litellm.AuthenticationError):
+            raise errors.RequesterError(f'API key 无效: {str(e)}')
+        if isinstance(e, litellm.NotFoundError):
+            raise errors.RequesterError(f'模型或路径无效: {str(e)}')
+        if isinstance(e, litellm.RateLimitError):
+            raise errors.RequesterError(f'请求过于频繁或余额不足: {str(e)}')
+        if isinstance(e, litellm.Timeout):
+            raise errors.RequesterError(f'请求超时: {str(e)}')
+        if isinstance(e, litellm.APIConnectionError):
+            raise errors.RequesterError(f'连接错误: {str(e)}')
+        if isinstance(e, litellm.APIError):
+            raise errors.RequesterError(f'API 错误: {str(e)}')
+        raise errors.RequesterError(f'未知错误: {str(e)}')
+
+    async def invoke_llm(
+        self,
+        query: pipeline_query.Query,
+        model: requester.RuntimeLLMModel,
+        messages: typing.List[provider_message.Message],
+        funcs: typing.List[resource_tool.LLMTool] = None,
+        extra_args: dict[str, typing.Any] = {},
+        remove_think: bool = False,
+    ) -> tuple[provider_message.Message, dict]:
+        """Invoke LLM and return message with usage info."""
+        # DO NOT modify input messages - copy them
+        req_messages = self._convert_messages(messages)
+
+        model_name = self._build_litellm_model_name(model.model_entity.name)
+
+        api_key = model.provider.token_mgr.get_token()
+
+        args = {
+            'model': model_name,
+            'messages': req_messages,
+            'api_key': api_key,
+        }
+        self._build_common_args(args)
+        args.update(extra_args)
+
+        if funcs:
+            tools = await self.ap.tool_mgr.generate_tools_for_openai(funcs)
+            if tools:
+                args['tools'] = tools
+
+        try:
+            response = await acompletion(**args)
+
+            message_data = response.choices[0].message.model_dump()
+            if 'role' not in message_data or message_data['role'] is None:
+                message_data['role'] = 'assistant'
+
+            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)
+
+            if 'reasoning_content' in message_data:
+                del message_data['reasoning_content']
+
+            message = provider_message.Message(**message_data)
+            usage_info = self._extract_usage(response)
+
+            return message, usage_info
+
+        except Exception as e:
+            self._handle_litellm_error(e)
+
+    async def invoke_llm_stream(
+        self,
+        query: pipeline_query.Query,
+        model: requester.RuntimeLLMModel,
+        messages: typing.List[provider_message.Message],
+        funcs: typing.List[resource_tool.LLMTool] = None,
+        extra_args: dict[str, typing.Any] = {},
+        remove_think: bool = False,
+    ) -> provider_message.MessageChunk:
+        """Invoke LLM streaming and yield chunks."""
+        req_messages = self._convert_messages(messages)
+
+        model_name = self._build_litellm_model_name(model.model_entity.name)
+        api_key = model.provider.token_mgr.get_token()
+
+        args = {
+            'model': model_name,
+            'messages': req_messages,
+            'api_key': api_key,
+            'stream': True,
+        }
+        self._build_common_args(args)
+        args.update(extra_args)
+
+        if funcs:
+            tools = await self.ap.tool_mgr.generate_tools_for_openai(funcs)
+            if tools:
+                args['tools'] = tools
+
+        chunk_idx = 0
+        role = 'assistant'
+
+        try:
+            response = await acompletion(**args)
+            async for chunk in response:
+                if not hasattr(chunk, 'choices') or not chunk.choices:
+                    continue
+
+                choice = chunk.choices[0]
+                delta = choice.delta.model_dump() if hasattr(choice, 'delta') else {}
+                finish_reason = getattr(choice, 'finish_reason', None)
+
+                if 'role' in delta and delta['role']:
+                    role = delta['role']
+
+                delta_content = delta.get('content', '')
+                reasoning_content = delta.get('reasoning_content', '')
+
+                if reasoning_content:
+                    chunk_idx += 1
+                    continue
+
+                if chunk_idx == 0 and not delta_content and not delta.get('tool_calls'):
+                    chunk_idx += 1
+                    continue
+
+                chunk_data = {
+                    'role': role,
+                    'content': delta_content if delta_content else None,
+                    'tool_calls': delta.get('tool_calls'),
+                    'is_final': bool(finish_reason),
+                }
+
+                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
+
+        except Exception as e:
+            self._handle_litellm_error(e)
+
+    async def invoke_embedding(
+        self,
+        model: requester.RuntimeEmbeddingModel,
+        input_text: list[str],
+        extra_args: dict[str, typing.Any] = {},
+    ) -> tuple[list[list[float]], dict]:
+        """Invoke embedding and return vectors with usage info."""
+        model_name = self._build_litellm_model_name(model.model_entity.name)
+        api_key = model.provider.token_mgr.get_token()
+
+        args = {
+            'model': model_name,
+            'input': input_text,
+            'api_key': api_key,
+        }
+        self._build_common_args(args, include_retry_params=False)
+
+        if model.model_entity.extra_args:
+            args.update(model.model_entity.extra_args)
+
+        args.update(extra_args)
+
+        try:
+            response = await aembedding(**args)
+
+            embeddings = [d.embedding for d in response.data]
+            usage_info = self._extract_usage(response)
+
+            return embeddings, usage_info
+
+        except Exception as e:
+            self._handle_litellm_error(e)
+
+    async def invoke_rerank(
+        self,
+        model: requester.RuntimeRerankModel,
+        query: str,
+        documents: typing.List[str],
+        extra_args: dict[str, typing.Any] = {},
+    ) -> typing.List[dict]:
+        """Invoke rerank and return relevance scores."""
+        model_name = self._build_litellm_model_name(model.model_entity.name)
+        api_key = model.provider.token_mgr.get_token()
+
+        args = {
+            'model': model_name,
+            'query': query,
+            'documents': documents,
+            'api_key': api_key,
+            'top_n': min(len(documents), 64),
+        }
+        self._build_common_args(args, include_retry_params=False)
+
+        if model.model_entity.extra_args:
+            args.update(model.model_entity.extra_args)
+
+        args.update(extra_args)
+
+        try:
+            response = await arerank(**args)
+
+            results = []
+            for r in response.results:
+                results.append(
+                    {
+                        'index': r.get('index', 0),
+                        'relevance_score': r.get('relevance_score', 0.0),
+                    }
+                )
+
+            if results:
+                scores = [r['relevance_score'] for r in results]
+                min_score = min(scores)
+                max_score = max(scores)
+                if max_score - min_score > 1e-6:
+                    for r in results:
+                        r['relevance_score'] = (r['relevance_score'] - min_score) / (max_score - min_score)
+
+            return results
+
+        except Exception as e:
+            self._handle_litellm_error(e)
+
+    async def scan_models(self, api_key: str | None = None) -> dict[str, typing.Any]:
+        """Scan models supported by the provider."""
+        import httpx
+
+        base_url = self.requester_cfg.get('base_url', '').rstrip('/')
+        timeout = self.requester_cfg.get('timeout', 120)
+
+        if not base_url:
+            raise errors.RequesterError('Base URL required for model scanning')
+
+        headers = {}
+        if api_key:
+            headers['Authorization'] = f'Bearer {api_key}'
+
+        models_url = f'{base_url}/models'
+
+        try:
+            async with httpx.AsyncClient(trust_env=True, timeout=timeout) as client:
+                response = await client.get(models_url, headers=headers)
+                response.raise_for_status()
+                payload = response.json()
+
+            models = []
+            for item in payload.get('data', []):
+                model_id = item.get('id')
+                if not model_id:
+                    continue
+
+                # Infer model type
+                normalized_id = (model_id or '').lower()
+                embedding_keywords = ('embedding', 'embed', 'bge-', 'e5-', 'm3e', 'gte-', 'text-embedding')
+                model_type = 'embedding' if any(kw in normalized_id for kw in embedding_keywords) else 'llm'
+
+                models.append(
+                    {
+                        'id': model_id,
+                        'name': model_id,
+                        'type': model_type,
+                    }
+                )
+
+            models.sort(key=lambda x: (x['type'] != 'llm', x['name'].lower()))
+
+            return {'models': models}
+
+        except httpx.HTTPStatusError as e:
+            raise errors.RequesterError(f'Model scan failed: {e.response.status_code}')
+        except httpx.TimeoutException:
+            raise errors.RequesterError('Model scan timeout')
+        except Exception as e:
+            raise errors.RequesterError(f'Model scan error: {str(e)}')

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

@@ -0,0 +1,64 @@
+apiVersion: v1
+kind: LLMAPIRequester
+metadata:
+  name: litellm-chat
+  label:
+    en_US: LiteLLM (Unified)
+    zh_Hans: LiteLLM (统一请求器)
+  icon: litellm.svg
+spec:
+  config:
+  - name: base_url
+    label:
+      en_US: Base URL
+      zh_Hans: 基础 URL
+    type: string
+    required: false
+    default: ''
+  - name: timeout
+    label:
+      en_US: Timeout
+      zh_Hans: 超时时间
+    type: integer
+    required: true
+    default: 120
+  - name: custom_llm_provider
+    label:
+      en_US: Custom Provider
+      zh_Hans: 自定义 Provider
+    type: string
+    required: false
+    default: ''
+    description:
+      en_US: Force provider type (e.g., anthropic, openai, gemini)
+      zh_Hans: 强制指定 provider 类型（如 anthropic, openai, gemini）
+  - name: drop_params
+    label:
+      en_US: Drop Unsupported Params
+      zh_Hans: 丢弃不支持参数
+    type: boolean
+    required: false
+    default: false
+  - name: num_retries
+    label:
+      en_US: Number of Retries
+      zh_Hans: 重试次数
+    type: integer
+    required: false
+    default: 0
+  - name: api_version
+    label:
+      en_US: API Version
+      zh_Hans: API 版本
+    type: string
+    required: false
+    default: ''
+  support_type:
+  - llm
+  - text-embedding
+  - rerank
+  provider_category: unified
+execution:
+  python:
+    path: ./litellmchat.py
+    attr: LiteLLMRequester

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

@@ -1,17 +0,0 @@
-from __future__ import annotations
-
-import typing
-import openai
-
-from . import chatcmpl
-
-
-class LmStudioChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """LMStudio ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'http://127.0.0.1:1234/v1',
-        'timeout': 120,
-    }

src/langbot/pkg/provider/modelmgr/requesters/lmstudiochatcmpl.yaml

@@ -22,11 +22,8 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: openai
   support_type:
   - llm
   - text-embedding
   provider_category: self-hosted
-execution:
-  python:
-    path: ./lmstudiochatcmpl.py
-    attr: LmStudioChatCompletions

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

@@ -1,561 +0,0 @@
-from __future__ import annotations
-
-import asyncio
-import typing
-
-import openai
-import openai.types.chat.chat_completion as chat_completion
-import httpx
-
-from .. import entities, errors, requester
-import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
-import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
-import langbot_plugin.api.entities.builtin.provider.message as provider_message
-
-
-class ModelScopeChatCompletions(requester.ProviderAPIRequester):
-    """ModelScope ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://api-inference.modelscope.cn/v1',
-        'timeout': 120,
-    }
-
-    async def initialize(self):
-        self.client = openai.AsyncClient(
-            api_key='',
-            base_url=self.requester_cfg['base_url'],
-            timeout=self.requester_cfg['timeout'],
-            http_client=httpx.AsyncClient(trust_env=True, timeout=self.requester_cfg['timeout']),
-        )
-
-    def _mask_api_key(self, api_key: str | None) -> str:
-        if not api_key:
-            return ''
-        if len(api_key) <= 8:
-            return '****'
-        return f'{api_key[:4]}...{api_key[-4:]}'
-
-    def _infer_model_type(self, model_id: str) -> str:
-        normalized_model_id = (model_id or '').lower()
-        embedding_keywords = (
-            'embedding',
-            'embed',
-            'bge-',
-            'e5-',
-            'm3e',
-            'gte-',
-            'multilingual-e5',
-            'text-embedding',
-        )
-        return 'embedding' if any(keyword in normalized_model_id for keyword in embedding_keywords) else 'llm'
-
-    def _infer_model_abilities(self, item: dict[str, typing.Any], model_id: str) -> list[str]:
-        normalized_model_id = (model_id or '').lower()
-        abilities: set[str] = set()
-
-        def _flatten(value: typing.Any) -> list[str]:
-            if value is None:
-                return []
-            if isinstance(value, str):
-                return [value.lower()]
-            if isinstance(value, dict):
-                flattened: list[str] = []
-                for nested_value in value.values():
-                    flattened.extend(_flatten(nested_value))
-                return flattened
-            if isinstance(value, (list, tuple, set)):
-                flattened: list[str] = []
-                for nested_value in value:
-                    flattened.extend(_flatten(nested_value))
-                return flattened
-            return [str(value).lower()]
-
-        capability_tokens = _flatten(item.get('capabilities'))
-        capability_tokens.extend(_flatten(item.get('modalities')))
-        capability_tokens.extend(_flatten(item.get('input_modalities')))
-        capability_tokens.extend(_flatten(item.get('output_modalities')))
-        capability_tokens.extend(_flatten(item.get('supported_generation_methods')))
-        capability_tokens.extend(_flatten(item.get('supported_parameters')))
-        capability_tokens.extend(_flatten(item.get('architecture')))
-
-        combined_tokens = capability_tokens + [normalized_model_id]
-
-        vision_keywords = ('vision', 'image', 'file', 'video', 'multimodal', 'vl', 'ocr', 'omni')
-        function_call_keywords = ('function', 'tool', 'tools', 'tool_choice', 'tool_call', 'tool-use', 'tool_use')
-
-        if any(any(keyword in token for keyword in vision_keywords) for token in combined_tokens):
-            abilities.add('vision')
-
-        if any(any(keyword in token for keyword in function_call_keywords) for token in combined_tokens):
-            abilities.add('func_call')
-
-        return sorted(abilities)
-
-    def _normalize_modalities(self, value: typing.Any) -> list[str]:
-        normalized: list[str] = []
-
-        def _collect(item: typing.Any):
-            if item is None:
-                return
-            if isinstance(item, str):
-                for part in item.replace('->', ',').replace('+', ',').split(','):
-                    token = part.strip().lower()
-                    if token and token not in normalized:
-                        normalized.append(token)
-                return
-            if isinstance(item, dict):
-                for nested in item.values():
-                    _collect(nested)
-                return
-            if isinstance(item, (list, tuple, set)):
-                for nested in item:
-                    _collect(nested)
-                return
-
-        _collect(value)
-        return normalized
-
-    def _extract_scan_metadata(self, item: dict[str, typing.Any], model_id: str) -> dict[str, typing.Any]:
-        display_name = item.get('name')
-        if not isinstance(display_name, str) or not display_name.strip() or display_name == model_id:
-            display_name = ''
-
-        description = item.get('description')
-        if not isinstance(description, str) or not description.strip():
-            description = ''
-
-        context_length = item.get('context_length')
-        if context_length is None and isinstance(item.get('top_provider'), dict):
-            context_length = item['top_provider'].get('context_length')
-
-        if not isinstance(context_length, int):
-            try:
-                context_length = int(context_length) if context_length is not None else None
-            except (TypeError, ValueError):
-                context_length = None
-
-        input_modalities = self._normalize_modalities(item.get('input_modalities'))
-        output_modalities = self._normalize_modalities(item.get('output_modalities'))
-
-        if isinstance(item.get('architecture'), dict):
-            if not input_modalities:
-                input_modalities = self._normalize_modalities(item['architecture'].get('input_modalities'))
-            if not output_modalities:
-                output_modalities = self._normalize_modalities(item['architecture'].get('output_modalities'))
-
-        owned_by = item.get('owned_by')
-        if not isinstance(owned_by, str) or not owned_by.strip():
-            owned_by = ''
-
-        return {
-            'display_name': display_name or None,
-            'description': description or None,
-            'context_length': context_length,
-            'owned_by': owned_by or None,
-            'input_modalities': input_modalities,
-            'output_modalities': output_modalities,
-        }
-
-    async def scan_models(self, api_key: str | None = None) -> dict[str, typing.Any]:
-        headers = {}
-        if api_key:
-            headers['Authorization'] = f'Bearer {api_key}'
-
-        models_url = f'{self.requester_cfg["base_url"].rstrip("/")}/models'
-        async with httpx.AsyncClient(trust_env=True, timeout=self.requester_cfg['timeout']) as client:
-            response = await client.get(models_url, headers=headers)
-            response.raise_for_status()
-            payload = response.json()
-
-        models = []
-        for item in payload.get('data', []):
-            model_id = item.get('id')
-            if not model_id:
-                continue
-            models.append(
-                {
-                    'id': model_id,
-                    'name': model_id,
-                    'type': self._infer_model_type(model_id),
-                    'abilities': self._infer_model_abilities(item, model_id),
-                    **self._extract_scan_metadata(item, model_id),
-                }
-            )
-
-        models.sort(key=lambda item: (item['type'] != 'llm', item['name'].lower()))
-        return {
-            'models': models,
-            'debug': {
-                'request': {
-                    'method': 'GET',
-                    'url': models_url,
-                    'headers': {
-                        'Authorization': f'Bearer {self._mask_api_key(api_key)}' if api_key else '',
-                    },
-                },
-                'response': payload,
-            },
-        }
-
-    async def _req(
-        self,
-        query: pipeline_query.Query,
-        args: dict,
-        extra_body: dict = {},
-        remove_think: bool = False,
-    ) -> list[dict[str, typing.Any]]:
-        args['stream'] = True
-
-        chunk = None
-
-        pending_content = ''
-
-        tool_calls = []
-
-        resp_gen: openai.AsyncStream = await self.client.chat.completions.create(**args, extra_body=extra_body)
-
-        chunk_idx = 0
-        thinking_started = False
-        thinking_ended = False
-        tool_id = ''
-        tool_name = ''
-        message_delta = {}
-        async for chunk in resp_gen:
-            if not chunk or not chunk.id or not chunk.choices or not chunk.choices[0] or not chunk.choices[0].delta:
-                continue
-
-            delta = chunk.choices[0].delta.model_dump() if hasattr(chunk.choices[0], 'delta') else {}
-            reasoning_content = delta.get('reasoning_content')
-            # 处理 reasoning_content
-            if reasoning_content:
-                # accumulated_reasoning += reasoning_content
-                # 如果设置了 remove_think，跳过 reasoning_content
-                if remove_think:
-                    chunk_idx += 1
-                    continue
-
-                # 第一次出现 reasoning_content，添加 <think> 开始标签
-                if not thinking_started:
-                    thinking_started = True
-                    pending_content += '<think>\n' + reasoning_content
-                else:
-                    # 继续输出 reasoning_content
-                    pending_content += reasoning_content
-            elif thinking_started and not thinking_ended and delta.get('content'):
-                # reasoning_content 结束，normal content 开始，添加 </think> 结束标签
-                thinking_ended = True
-                pending_content += '\n</think>\n' + delta.get('content')
-
-            if delta.get('content') is not None:
-                pending_content += delta.get('content')
-
-            if delta.get('tool_calls') is not None:
-                for tool_call in delta.get('tool_calls'):
-                    if tool_call['id'] != '':
-                        tool_id = tool_call['id']
-                    if tool_call['function']['name'] is not None:
-                        tool_name = tool_call['function']['name']
-                    if tool_call['function']['arguments'] is None:
-                        continue
-                    tool_call['id'] = tool_id
-                    tool_call['name'] = tool_name
-                    for tc in tool_calls:
-                        if tc['index'] == tool_call['index']:
-                            tc['function']['arguments'] += tool_call['function']['arguments']
-                            break
-                    else:
-                        tool_calls.append(tool_call)
-
-            if chunk.choices[0].finish_reason is not None:
-                break
-        message_delta['content'] = pending_content
-        message_delta['role'] = 'assistant'
-
-        message_delta['tool_calls'] = tool_calls if tool_calls else None
-        return [message_delta]
-
-    async def _make_msg(
-        self,
-        chat_completion: list[dict[str, typing.Any]],
-    ) -> provider_message.Message:
-        chatcmpl_message = chat_completion[0]
-
-        # 确保 role 字段存在且不为 None
-        if 'role' not in chatcmpl_message or chatcmpl_message['role'] is None:
-            chatcmpl_message['role'] = 'assistant'
-
-        message = provider_message.Message(**chatcmpl_message)
-
-        return message
-
-    async def _closure(
-        self,
-        query: pipeline_query.Query,
-        req_messages: list[dict],
-        use_model: requester.RuntimeLLMModel,
-        use_funcs: list[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> tuple[provider_message.Message, dict]:
-        self.client.api_key = use_model.provider.token_mgr.get_token()
-
-        args = {}
-        args['model'] = use_model.model_entity.name
-
-        if use_funcs:
-            tools = await self.ap.tool_mgr.generate_tools_for_openai(use_funcs)
-
-            if tools:
-                args['tools'] = tools
-
-        # 设置此次请求中的messages
-        messages = req_messages.copy()
-
-        # 检查vision
-        for msg in messages:
-            if 'content' in msg and isinstance(msg['content'], list):
-                for me in msg['content']:
-                    if me['type'] == 'image_base64':
-                        me['image_url'] = {'url': me['image_base64']}
-                        me['type'] = 'image_url'
-                        del me['image_base64']
-
-        args['messages'] = messages
-
-        # 发送请求
-        resp = await self._req(query, args, extra_body=extra_args, remove_think=remove_think)
-
-        # 处理请求结果
-        message = await self._make_msg(resp)
-
-        # ModelScope uses streaming, usage info not available
-        usage_info = {}
-
-        return message, usage_info
-
-    async def _req_stream(
-        self,
-        args: dict,
-        extra_body: dict = {},
-    ) -> chat_completion.ChatCompletion:
-        async for chunk in await self.client.chat.completions.create(**args, extra_body=extra_body):
-            yield chunk
-
-    async def _closure_stream(
-        self,
-        query: pipeline_query.Query,
-        req_messages: list[dict],
-        use_model: requester.RuntimeLLMModel,
-        use_funcs: list[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> provider_message.Message | typing.AsyncGenerator[provider_message.MessageChunk, None]:
-        self.client.api_key = use_model.provider.token_mgr.get_token()
-
-        args = {}
-        args['model'] = use_model.model_entity.name
-
-        if use_funcs:
-            tools = await self.ap.tool_mgr.generate_tools_for_openai(use_funcs)
-
-            if tools:
-                args['tools'] = tools
-
-        # 设置此次请求中的messages
-        messages = req_messages.copy()
-
-        # 检查vision
-        for msg in messages:
-            if 'content' in msg and isinstance(msg['content'], list):
-                for me in msg['content']:
-                    if me['type'] == 'image_base64':
-                        me['image_url'] = {'url': me['image_base64']}
-                        me['type'] = 'image_url'
-                        del me['image_base64']
-
-        args['messages'] = messages
-        args['stream'] = True
-
-        # 流式处理状态
-        # tool_calls_map: dict[str, provider_message.ToolCall] = {}
-        chunk_idx = 0
-        thinking_started = False
-        thinking_ended = False
-        role = 'assistant'  # 默认角色
-        # accumulated_reasoning = ''  # 仅用于判断何时结束思维链
-
-        async for chunk in self._req_stream(args, extra_body=extra_args):
-            # 解析 chunk 数据
-            if hasattr(chunk, 'choices') and chunk.choices:
-                choice = chunk.choices[0]
-                delta = choice.delta.model_dump() if hasattr(choice, 'delta') else {}
-                finish_reason = getattr(choice, 'finish_reason', None)
-            else:
-                delta = {}
-                finish_reason = None
-
-            # 从第一个 chunk 获取 role，后续使用这个 role
-            if 'role' in delta and delta['role']:
-                role = delta['role']
-
-            # 获取增量内容
-            delta_content = delta.get('content', '')
-            reasoning_content = delta.get('reasoning_content', '')
-
-            # 处理 reasoning_content
-            if reasoning_content:
-                # accumulated_reasoning += reasoning_content
-                # 如果设置了 remove_think，跳过 reasoning_content
-                if remove_think:
-                    chunk_idx += 1
-                    continue
-
-                # 第一次出现 reasoning_content，添加 <think> 开始标签
-                if not thinking_started:
-                    thinking_started = True
-                    delta_content = '<think>\n' + reasoning_content
-                else:
-                    # 继续输出 reasoning_content
-                    delta_content = reasoning_content
-            elif thinking_started and not thinking_ended and delta_content:
-                # reasoning_content 结束，normal content 开始，添加 </think> 结束标签
-                thinking_ended = True
-                delta_content = '\n</think>\n' + delta_content
-
-            # 处理 content 中已有的 <think> 标签（如果需要移除）
-            # if delta_content and remove_think and '<think>' in delta_content:
-            #     import re
-            #
-            #     # 移除 <think> 标签及其内容
-            #     delta_content = re.sub(r'<think>.*?</think>', '', delta_content, flags=re.DOTALL)
-
-            # 处理工具调用增量
-            if delta.get('tool_calls'):
-                for tool_call in delta['tool_calls']:
-                    if tool_call['id'] != '':
-                        tool_id = tool_call['id']
-                    if tool_call['function']['name'] is not None:
-                        tool_name = tool_call['function']['name']
-
-                    if tool_call['type'] is None:
-                        tool_call['type'] = 'function'
-                    tool_call['id'] = tool_id
-                    tool_call['function']['name'] = tool_name
-                    tool_call['function']['arguments'] = (
-                        '' if tool_call['function']['arguments'] is None else tool_call['function']['arguments']
-                    )
-
-            # 跳过空的第一个 chunk（只有 role 没有内容）
-            if chunk_idx == 0 and not delta_content and not reasoning_content and not delta.get('tool_calls'):
-                chunk_idx += 1
-                continue
-
-            # 构建 MessageChunk - 只包含增量内容
-            chunk_data = {
-                'role': role,
-                'content': delta_content if delta_content else None,
-                'tool_calls': delta.get('tool_calls'),
-                'is_final': bool(finish_reason),
-            }
-
-            # 移除 None 值
-            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
-            # return
-
-    async def invoke_llm(
-        self,
-        query: pipeline_query.Query,
-        model: entities.LLMModelInfo,
-        messages: typing.List[provider_message.Message],
-        funcs: typing.List[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> provider_message.Message:
-        req_messages = []  # req_messages 仅用于类内，外部同步由 query.messages 进行
-        for m in messages:
-            msg_dict = m.dict(exclude_none=True)
-            content = msg_dict.get('content')
-            if isinstance(content, list):
-                # 检查 content 列表中是否每个部分都是文本
-                if all(isinstance(part, dict) and part.get('type') == 'text' for part in content):
-                    # 将所有文本部分合并为一个字符串
-                    msg_dict['content'] = '\n'.join(part['text'] for part in content)
-            req_messages.append(msg_dict)
-
-        try:
-            return await self._closure(
-                query=query,
-                req_messages=req_messages,
-                use_model=model,
-                use_funcs=funcs,
-                extra_args=extra_args,
-                remove_think=remove_think,
-            )
-        except asyncio.TimeoutError:
-            raise errors.RequesterError('请求超时')
-        except openai.BadRequestError as e:
-            if 'context_length_exceeded' in e.message:
-                raise errors.RequesterError(f'上文过长，请重置会话: {e.message}')
-            else:
-                raise errors.RequesterError(f'请求参数错误: {e.message}')
-        except openai.AuthenticationError as e:
-            raise errors.RequesterError(f'无效的 api-key: {e.message}')
-        except openai.NotFoundError as e:
-            raise errors.RequesterError(f'请求路径错误: {e.message}')
-        except openai.RateLimitError as e:
-            raise errors.RequesterError(f'请求过于频繁或余额不足: {e.message}')
-        except openai.APIError as e:
-            raise errors.RequesterError(f'请求错误: {e.message}')
-
-    async def invoke_llm_stream(
-        self,
-        query: pipeline_query.Query,
-        model: requester.RuntimeLLMModel,
-        messages: typing.List[provider_message.Message],
-        funcs: typing.List[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> provider_message.MessageChunk:
-        req_messages = []  # req_messages 仅用于类内，外部同步由 query.messages 进行
-        for m in messages:
-            msg_dict = m.dict(exclude_none=True)
-            content = msg_dict.get('content')
-            if isinstance(content, list):
-                # 检查 content 列表中是否每个部分都是文本
-                if all(isinstance(part, dict) and part.get('type') == 'text' for part in content):
-                    # 将所有文本部分合并为一个字符串
-                    msg_dict['content'] = '\n'.join(part['text'] for part in content)
-            req_messages.append(msg_dict)
-
-        try:
-            async for item in self._closure_stream(
-                query=query,
-                req_messages=req_messages,
-                use_model=model,
-                use_funcs=funcs,
-                extra_args=extra_args,
-                remove_think=remove_think,
-            ):
-                yield item
-
-        except asyncio.TimeoutError:
-            raise errors.RequesterError('请求超时')
-        except openai.BadRequestError as e:
-            if 'context_length_exceeded' in e.message:
-                raise errors.RequesterError(f'上文过长，请重置会话: {e.message}')
-            else:
-                raise errors.RequesterError(f'请求参数错误: {e.message}')
-        except openai.AuthenticationError as e:
-            raise errors.RequesterError(f'无效的 api-key: {e.message}')
-        except openai.NotFoundError as e:
-            raise errors.RequesterError(f'请求路径错误: {e.message}')
-        except openai.RateLimitError as e:
-            raise errors.RequesterError(f'请求过于频繁或余额不足: {e.message}')
-        except openai.APIError as e:
-            raise errors.RequesterError(f'请求错误: {e.message}')

src/langbot/pkg/provider/modelmgr/requesters/modelscopechatcmpl.yaml

@@ -29,10 +29,7 @@ spec:
     type: int
     required: true
     default: 120
+  litellm_provider: openai
   support_type:
   - llm
   provider_category: maas
-execution:
-  python:
-    path: ./modelscopechatcmpl.py
-    attr: ModelScopeChatCompletions

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

@@ -1,67 +0,0 @@
-from __future__ import annotations
-
-import typing
-
-
-from . import chatcmpl
-from .. import requester
-import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
-import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
-import langbot_plugin.api.entities.builtin.provider.message as provider_message
-
-
-class MoonshotChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """Moonshot ChatCompletion API 请求器"""
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://api.moonshot.cn/v1',
-        'timeout': 120,
-    }
-
-    async def _closure(
-        self,
-        query: pipeline_query.Query,
-        req_messages: list[dict],
-        use_model: requester.RuntimeLLMModel,
-        use_funcs: list[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> tuple[provider_message.Message, dict]:
-        self.client.api_key = use_model.provider.token_mgr.get_token()
-
-        args = {}
-        args['model'] = use_model.model_entity.name
-
-        if use_funcs:
-            tools = await self.ap.tool_mgr.generate_tools_for_openai(use_funcs)
-
-            if tools:
-                args['tools'] = tools
-
-        # 设置此次请求中的messages
-        messages = req_messages
-
-        # deepseek 不支持多模态，把content都转换成纯文字
-        for m in messages:
-            if 'content' in m and isinstance(m['content'], list):
-                m['content'] = ' '.join([c['text'] for c in m['content']])
-
-        # 删除空的，不知道干嘛的，直接删了。
-        # messages = [m for m in messages if m["content"].strip() != "" and ('tool_calls' not in m or not m['tool_calls'])]
-
-        args['messages'] = messages
-
-        # 发送请求
-        resp = await self._req(args, extra_body=extra_args)
-
-        # 处理请求结果
-        message = await self._make_msg(resp, remove_think)
-
-        # Extract token usage from response
-        usage_info = {}
-        if hasattr(resp, 'usage') and resp.usage:
-            usage_info['input_tokens'] = resp.usage.prompt_tokens or 0
-            usage_info['output_tokens'] = resp.usage.completion_tokens or 0
-            usage_info['total_tokens'] = resp.usage.total_tokens or 0
-
-        return message, usage_info

src/langbot/pkg/provider/modelmgr/requesters/moonshotchatcmpl.yaml

@@ -22,10 +22,7 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: moonshot
   support_type:
   - llm
   provider_category: manufacturer
-execution:
-  python:
-    path: ./moonshotchatcmpl.py
-    attr: MoonshotChatCompletions

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

@@ -1,17 +0,0 @@
-from __future__ import annotations
-
-import typing
-import openai
-
-from . import chatcmpl
-
-
-class NewAPIChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """New API ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'http://localhost:3000/v1',
-        'timeout': 120,
-    }

src/langbot/pkg/provider/modelmgr/requesters/newapichatcmpl.yaml

@@ -22,11 +22,8 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: openai
   support_type:
   - llm
   - text-embedding
   provider_category: maas
-execution:
-  python:
-    path: ./newapichatcmpl.py
-    attr: NewAPIChatCompletions

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

@@ -1,314 +0,0 @@
-from __future__ import annotations
-
-import asyncio
-import os
-import typing
-from typing import Union, Mapping, Any, AsyncIterator
-import uuid
-import json
-
-import ollama
-import httpx
-
-from .. import errors, requester
-import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
-import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
-import langbot_plugin.api.entities.builtin.provider.message as provider_message
-
-REQUESTER_NAME: str = 'ollama-chat'
-
-
-class OllamaChatCompletions(requester.ProviderAPIRequester):
-    """Ollama平台 ChatCompletion API请求器"""
-
-    client: ollama.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'http://127.0.0.1:11434',
-        'timeout': 120,
-    }
-
-    async def initialize(self):
-        os.environ['OLLAMA_HOST'] = self.requester_cfg['base_url']
-        self.client = ollama.AsyncClient(timeout=self.requester_cfg['timeout'])
-
-    def _infer_model_type(self, model_id: str) -> str:
-        normalized_model_id = (model_id or '').lower()
-        embedding_keywords = ('embedding', 'embed', 'bge-', 'e5-', 'm3e', 'gte-', 'text-embedding')
-        return 'embedding' if any(keyword in normalized_model_id for keyword in embedding_keywords) else 'llm'
-
-    def _infer_model_abilities(self, item: dict[str, typing.Any], model_id: str) -> list[str]:
-        normalized_model_id = (model_id or '').lower()
-        abilities: set[str] = set()
-        details = item.get('details', {}) or {}
-        families = details.get('families', []) or []
-        tokens = [normalized_model_id, str(details.get('family', '')).lower()]
-        tokens.extend(str(family).lower() for family in families)
-
-        if any(keyword in token for token in tokens for keyword in ('vision', 'vl', 'omni', 'llava', 'ocr')):
-            abilities.add('vision')
-        if any(keyword in token for token in tokens for keyword in ('tool', 'function')):
-            abilities.add('func_call')
-        return sorted(abilities)
-
-    async def scan_models(self, api_key: str | None = None) -> dict[str, typing.Any]:
-        del api_key
-        models_url = f'{self.requester_cfg["base_url"].rstrip("/")}/api/tags'
-
-        async with httpx.AsyncClient(trust_env=True, timeout=self.requester_cfg['timeout']) as client:
-            response = await client.get(models_url)
-            response.raise_for_status()
-            payload = response.json()
-
-        models: list[dict[str, typing.Any]] = []
-        for item in payload.get('models', []):
-            model_id = item.get('model') or item.get('name')
-            if not model_id:
-                continue
-            models.append(
-                {
-                    'id': model_id,
-                    'name': item.get('name', model_id),
-                    'type': self._infer_model_type(model_id),
-                    'abilities': self._infer_model_abilities(item, model_id),
-                }
-            )
-
-        models.sort(key=lambda item: (item['type'] != 'llm', item['name'].lower()))
-        return {
-            'models': models,
-            'debug': {
-                'request': {
-                    'method': 'GET',
-                    'url': models_url,
-                },
-                'response': payload,
-            },
-        }
-
-    async def _req(
-        self,
-        args: dict,
-    ) -> Union[Mapping[str, Any], AsyncIterator[Mapping[str, Any]]]:
-        return await self.client.chat(**args)
-
-    async def _closure(
-        self,
-        query: pipeline_query.Query,
-        req_messages: list[dict],
-        use_model: requester.RuntimeLLMModel,
-        use_funcs: list[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> provider_message.Message:
-        args = extra_args.copy()
-        args['model'] = use_model.model_entity.name
-
-        messages: list[dict] = req_messages.copy()
-        for msg in messages:
-            if 'content' in msg and isinstance(msg['content'], list):
-                text_content: list = []
-                image_urls: list = []
-                for me in msg['content']:
-                    if me['type'] == 'text':
-                        text_content.append(me['text'])
-                    elif me['type'] == 'image_base64':
-                        image_urls.append(me['image_base64'])
-
-                msg['content'] = '\n'.join(text_content)
-                msg['images'] = [url.split(',')[1] for url in image_urls]
-            if 'tool_calls' in msg:  # LangBot 内部以 str 存储 tool_calls 的参数，这里需要转换为 dict
-                for tool_call in msg['tool_calls']:
-                    tool_call['function']['arguments'] = json.loads(tool_call['function']['arguments'])
-        args['messages'] = messages
-
-        args['tools'] = []
-        if use_funcs:
-            tools = await self.ap.tool_mgr.generate_tools_for_openai(use_funcs)
-            if tools:
-                args['tools'] = tools
-
-        resp = await self._req(args)
-        message: provider_message.Message = await self._make_msg(resp)
-        return message
-
-    async def _make_msg(self, chat_completions: ollama.ChatResponse) -> provider_message.Message:
-        message: ollama.Message = chat_completions.message
-        if message is None:
-            raise ValueError("chat_completions must contain a 'message' field")
-
-        ret_msg: provider_message.Message = None
-
-        if message.content is not None:
-            ret_msg = provider_message.Message(role='assistant', content=message.content)
-        if message.tool_calls is not None and len(message.tool_calls) > 0:
-            tool_calls: list[provider_message.ToolCall] = []
-
-            for tool_call in message.tool_calls:
-                tool_calls.append(
-                    provider_message.ToolCall(
-                        id=uuid.uuid4().hex,
-                        type='function',
-                        function=provider_message.FunctionCall(
-                            name=tool_call.function.name,
-                            arguments=json.dumps(tool_call.function.arguments),
-                        ),
-                    )
-                )
-            ret_msg.tool_calls = tool_calls
-
-        return ret_msg
-
-    async def _prepare_messages(
-        self,
-        messages: typing.List[provider_message.Message],
-    ) -> list[dict]:
-        """Prepare messages for Ollama API request."""
-        req_messages: list = []
-        for m in messages:
-            msg_dict: dict = m.dict(exclude_none=True)
-            content: Any = msg_dict.get('content')
-            if isinstance(content, list):
-                if all(isinstance(part, dict) and part.get('type') == 'text' for part in content):
-                    msg_dict['content'] = '\n'.join(part['text'] for part in content)
-            req_messages.append(msg_dict)
-        return req_messages
-
-    async def invoke_llm(
-        self,
-        query: pipeline_query.Query,
-        model: requester.RuntimeLLMModel,
-        messages: typing.List[provider_message.Message],
-        funcs: typing.List[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> provider_message.Message:
-        req_messages = await self._prepare_messages(messages)
-        try:
-            return await self._closure(
-                query=query,
-                req_messages=req_messages,
-                use_model=model,
-                use_funcs=funcs,
-                extra_args=extra_args,
-                remove_think=remove_think,
-            )
-        except asyncio.TimeoutError:
-            raise errors.RequesterError('请求超时')
-
-    async def invoke_llm_stream(
-        self,
-        query: pipeline_query.Query,
-        model: requester.RuntimeLLMModel,
-        messages: typing.List[provider_message.Message],
-        funcs: typing.List[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> provider_message.MessageChunk:
-        req_messages = await self._prepare_messages(messages)
-
-        try:
-            args = extra_args.copy()
-            args['model'] = model.model_entity.name
-
-            # Process messages for Ollama format
-            msgs: list[dict] = req_messages.copy()
-            for msg in msgs:
-                if 'content' in msg and isinstance(msg['content'], list):
-                    text_content: list = []
-                    image_urls: list = []
-                    for me in msg['content']:
-                        if me['type'] == 'text':
-                            text_content.append(me['text'])
-                        elif me['type'] == 'image_base64':
-                            image_urls.append(me['image_base64'])
-                    msg['content'] = '\n'.join(text_content)
-                    msg['images'] = [url.split(',')[1] for url in image_urls]
-                if 'tool_calls' in msg:
-                    for tool_call in msg['tool_calls']:
-                        tool_call['function']['arguments'] = json.loads(tool_call['function']['arguments'])
-            args['messages'] = msgs
-
-            args['tools'] = []
-            if funcs:
-                tools = await self.ap.tool_mgr.generate_tools_for_openai(funcs)
-                if tools:
-                    args['tools'] = tools
-
-            args['stream'] = True
-
-            chunk_idx = 0
-            thinking_started = False
-            thinking_ended = False
-            role = 'assistant'
-
-            async for chunk in await self.client.chat(**args):
-                message: ollama.Message = chunk.message
-                done = chunk.done
-
-                delta_content = message.content or ''
-                reasoning_content = getattr(message, 'thinking', '') or ''
-
-                # Handle reasoning/thinking content
-                if reasoning_content:
-                    if remove_think:
-                        chunk_idx += 1
-                        continue
-
-                    if not thinking_started:
-                        thinking_started = True
-                        delta_content = '<think>\n' + reasoning_content
-                    else:
-                        delta_content = reasoning_content
-                elif thinking_started and not thinking_ended and delta_content:
-                    thinking_ended = True
-                    delta_content = '\n</think>\n' + delta_content
-
-                # Handle tool calls
-                tool_calls_data = None
-                if message.tool_calls:
-                    tool_calls_data = []
-                    for tc in message.tool_calls:
-                        tool_calls_data.append(
-                            {
-                                'id': uuid.uuid4().hex,
-                                'type': 'function',
-                                'function': {
-                                    'name': tc.function.name,
-                                    'arguments': json.dumps(tc.function.arguments),
-                                },
-                            }
-                        )
-
-                # Skip empty first chunk
-                if chunk_idx == 0 and not delta_content and not reasoning_content and not tool_calls_data:
-                    chunk_idx += 1
-                    continue
-
-                chunk_data = {
-                    'role': role,
-                    'content': delta_content if delta_content else None,
-                    'tool_calls': tool_calls_data,
-                    'is_final': bool(done),
-                }
-                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
-
-        except asyncio.TimeoutError:
-            raise errors.RequesterError('请求超时')
-
-    async def invoke_embedding(
-        self,
-        model: requester.RuntimeEmbeddingModel,
-        input_text: list[str],
-        extra_args: dict[str, typing.Any] = {},
-    ) -> list[list[float]]:
-        return (
-            await self.client.embed(
-                model=model.model_entity.name,
-                input=input_text,
-                **extra_args,
-            )
-        ).embeddings

src/langbot/pkg/provider/modelmgr/requesters/ollamachat.yaml

@@ -22,11 +22,8 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: ollama
   support_type:
   - llm
   - text-embedding
   provider_category: self-hosted
-execution:
-  python:
-    path: ./ollamachat.py
-    attr: OllamaChatCompletions

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

@@ -1,25 +0,0 @@
-from __future__ import annotations
-
-import typing
-import openai
-
-from . import modelscopechatcmpl
-
-
-class OpenRouterChatCompletions(modelscopechatcmpl.ModelScopeChatCompletions):
-    """OpenRouter ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://openrouter.ai/api/v1',
-        'timeout': 120,
-    }
-
-    async def scan_models(self, api_key: str | None = None) -> dict[str, typing.Any]:
-        original_base_url = self.requester_cfg.get('base_url', '')
-        self.requester_cfg['base_url'] = 'https://openrouter.ai/api/v1'
-        try:
-            return await super().scan_models(api_key)
-        finally:
-            self.requester_cfg['base_url'] = original_base_url

src/langbot/pkg/provider/modelmgr/requesters/openrouterchatcmpl.yaml

@@ -22,12 +22,9 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: openrouter
   support_type:
   - llm
   - text-embedding
   - rerank
   provider_category: maas
-execution:
-  python:
-    path: ./openrouterchatcmpl.py
-    attr: OpenRouterChatCompletions

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

@@ -1,208 +0,0 @@
-from __future__ import annotations
-
-import openai
-import typing
-
-from . import chatcmpl
-from .. import requester
-import openai.types.chat.chat_completion as chat_completion
-import re
-import langbot_plugin.api.entities.builtin.provider.message as provider_message
-import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
-import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
-
-
-class PPIOChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """欧派云 ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://api.ppinfra.com/v3/openai',
-        'timeout': 120,
-    }
-
-    is_think: bool = False
-
-    async def _make_msg(
-        self,
-        chat_completion: chat_completion.ChatCompletion,
-        remove_think: bool,
-    ) -> provider_message.Message:
-        chatcmpl_message = chat_completion.choices[0].message.model_dump()
-        # print(chatcmpl_message.keys(), chatcmpl_message.values())
-
-        # 确保 role 字段存在且不为 None
-        if 'role' not in chatcmpl_message or chatcmpl_message['role'] is None:
-            chatcmpl_message['role'] = 'assistant'
-
-        reasoning_content = chatcmpl_message['reasoning_content'] if 'reasoning_content' in chatcmpl_message else None
-
-        # deepseek的reasoner模型
-        chatcmpl_message['content'] = await self._process_thinking_content(
-            chatcmpl_message['content'], reasoning_content, remove_think
-        )
-
-        # 移除 reasoning_content 字段，避免传递给 Message
-        if 'reasoning_content' in chatcmpl_message:
-            del chatcmpl_message['reasoning_content']
-
-        message = provider_message.Message(**chatcmpl_message)
-
-        return message
-
-    async def _process_thinking_content(
-        self,
-        content: str,
-        reasoning_content: str = None,
-        remove_think: bool = False,
-    ) -> tuple[str, str]:
-        """处理思维链内容
-
-        Args:
-            content: 原始内容
-            reasoning_content: reasoning_content 字段内容
-            remove_think: 是否移除思维链
-
-        Returns:
-            处理后的内容
-        """
-        if remove_think:
-            content = re.sub(r'<think>.*?</think>', '', content, flags=re.DOTALL)
-        else:
-            if reasoning_content is not None:
-                content = '<think>\n' + reasoning_content + '\n</think>\n' + content
-        return content
-
-    async def _make_msg_chunk(
-        self,
-        delta: dict[str, typing.Any],
-        idx: int,
-    ) -> provider_message.MessageChunk:
-        # 处理流式chunk和完整响应的差异
-        # print(chat_completion.choices[0])
-
-        # 确保 role 字段存在且不为 None
-        if 'role' not in delta or delta['role'] is None:
-            delta['role'] = 'assistant'
-
-        reasoning_content = delta['reasoning_content'] if 'reasoning_content' in delta else None
-
-        delta['content'] = '' if delta['content'] is None else delta['content']
-        # print(reasoning_content)
-
-        # deepseek的reasoner模型
-
-        if reasoning_content is not None:
-            delta['content'] += reasoning_content
-
-        message = provider_message.MessageChunk(**delta)
-
-        return message
-
-    async def _closure_stream(
-        self,
-        query: pipeline_query.Query,
-        req_messages: list[dict],
-        use_model: requester.RuntimeLLMModel,
-        use_funcs: list[resource_tool.LLMTool] = None,
-        extra_args: dict[str, typing.Any] = {},
-        remove_think: bool = False,
-    ) -> provider_message.Message | typing.AsyncGenerator[provider_message.MessageChunk, None]:
-        self.client.api_key = use_model.provider.token_mgr.get_token()
-
-        args = {}
-        args['model'] = use_model.model_entity.name
-
-        if use_funcs:
-            tools = await self.ap.tool_mgr.generate_tools_for_openai(use_funcs)
-
-            if tools:
-                args['tools'] = tools
-
-        # 设置此次请求中的messages
-        messages = req_messages.copy()
-
-        # 检查vision
-        for msg in messages:
-            if 'content' in msg and isinstance(msg['content'], list):
-                for me in msg['content']:
-                    if me['type'] == 'image_base64':
-                        me['image_url'] = {'url': me['image_base64']}
-                        me['type'] = 'image_url'
-                        del me['image_base64']
-
-        args['messages'] = messages
-        args['stream'] = True
-
-        # tool_calls_map: dict[str, provider_message.ToolCall] = {}
-        chunk_idx = 0
-        thinking_started = False
-        thinking_ended = False
-        role = 'assistant'  # 默认角色
-        async for chunk in self._req_stream(args, extra_body=extra_args):
-            # 解析 chunk 数据
-            if hasattr(chunk, 'choices') and chunk.choices:
-                choice = chunk.choices[0]
-                delta = choice.delta.model_dump() if hasattr(choice, 'delta') else {}
-                finish_reason = getattr(choice, 'finish_reason', None)
-            else:
-                delta = {}
-                finish_reason = None
-
-            # 从第一个 chunk 获取 role，后续使用这个 role
-            if 'role' in delta and delta['role']:
-                role = delta['role']
-
-            # 获取增量内容
-            delta_content = delta.get('content', '')
-            # reasoning_content = delta.get('reasoning_content', '')
-
-            if remove_think:
-                if delta['content'] is not None:
-                    if '<think>' in delta['content'] and not thinking_started and not thinking_ended:
-                        thinking_started = True
-                        continue
-                    elif delta['content'] == r'</think>' and not thinking_ended:
-                        thinking_ended = True
-                        continue
-                    elif thinking_ended and delta['content'] == '\n\n' and thinking_started:
-                        thinking_started = False
-                        continue
-                    elif thinking_started and not thinking_ended:
-                        continue
-
-            # delta_tool_calls = None
-            if delta.get('tool_calls'):
-                for tool_call in delta['tool_calls']:
-                    if tool_call['id'] and tool_call['function']['name']:
-                        tool_id = tool_call['id']
-                        tool_name = tool_call['function']['name']
-
-                    if tool_call['id'] is None:
-                        tool_call['id'] = tool_id
-                    if tool_call['function']['name'] is None:
-                        tool_call['function']['name'] = tool_name
-                    if tool_call['function']['arguments'] is None:
-                        tool_call['function']['arguments'] = ''
-                    if tool_call['type'] is None:
-                        tool_call['type'] = 'function'
-
-            # 跳过空的第一个 chunk（只有 role 没有内容）
-            if chunk_idx == 0 and not delta_content and not delta.get('tool_calls'):
-                chunk_idx += 1
-                continue
-
-            # 构建 MessageChunk - 只包含增量内容
-            chunk_data = {
-                'role': role,
-                'content': delta_content if delta_content else None,
-                'tool_calls': delta.get('tool_calls'),
-                'is_final': bool(finish_reason),
-            }
-
-            # 移除 None 值
-            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

src/langbot/pkg/provider/modelmgr/requesters/ppiochatcmpl.yaml

@@ -29,11 +29,8 @@ spec:
     type: int
     required: true
     default: 120
+  litellm_provider: openai
   support_type:
   - llm
   - text-embedding
   provider_category: maas
-execution:
-  python:
-    path: ./ppiochatcmpl.py
-    attr: PPIOChatCompletions

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

@@ -1,17 +0,0 @@
-from __future__ import annotations
-
-import openai
-import typing
-
-from . import chatcmpl
-
-
-class QHAIGCChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """启航 AI ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://api.qhaigc.com/v1',
-        'timeout': 120,
-    }

src/langbot/pkg/provider/modelmgr/requesters/qhaigcchatcmpl.yaml

@@ -29,11 +29,8 @@ spec:
     type: int
     required: true
     default: 120
+  litellm_provider: openai
   support_type:
   - llm
   - text-embedding
   provider_category: maas
-execution:
-  python:
-    path: ./qhaigcchatcmpl.py
-    attr: QHAIGCChatCompletions

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

@@ -1,32 +0,0 @@
-from __future__ import annotations
-
-import openai
-import typing
-
-from . import chatcmpl
-import openai.types.chat.chat_completion as chat_completion
-
-
-class ShengSuanYunChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """胜算云(ModelSpot.AI) ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://router.shengsuanyun.com/api/v1',
-        'timeout': 120,
-    }
-
-    async def _req(
-        self,
-        args: dict,
-        extra_body: dict = {},
-    ) -> chat_completion.ChatCompletion:
-        return await self.client.chat.completions.create(
-            **args,
-            extra_body=extra_body,
-            extra_headers={
-                'HTTP-Referer': 'https://langbot.app',
-                'X-Title': 'LangBot',
-            },
-        )

src/langbot/pkg/provider/modelmgr/requesters/shengsuanyun.yaml

@@ -29,11 +29,8 @@ spec:
     type: int
     required: true
     default: 120
+  litellm_provider: openai
   support_type:
   - llm
   - text-embedding
   provider_category: maas
-execution:
-  python:
-    path: ./shengsuanyun.py
-    attr: ShengSuanYunChatCompletions

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

@@ -1,17 +0,0 @@
-from __future__ import annotations
-
-import typing
-import openai
-
-from . import chatcmpl
-
-
-class SiliconFlowChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """SiliconFlow ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://api.siliconflow.cn/v1',
-        'timeout': 120,
-    }

src/langbot/pkg/provider/modelmgr/requesters/siliconflowchatcmpl.yaml

@@ -22,12 +22,9 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: siliconflow
   support_type:
   - llm
   - text-embedding
   - rerank
   provider_category: maas
-execution:
-  python:
-    path: ./siliconflowchatcmpl.py
-    attr: SiliconFlowChatCompletions

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

@@ -1,17 +0,0 @@
-from __future__ import annotations
-
-import typing
-import openai
-
-from . import chatcmpl
-
-
-class LangBotSpaceChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """LangBot Space ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://api.langbot.cloud/v1',
-        'timeout': 120,
-    }

src/langbot/pkg/provider/modelmgr/requesters/spacechatcmpl.yaml

@@ -22,11 +22,8 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: openai
   support_type:
   - llm
   - text-embedding
   provider_category: maas
-execution:
-  python:
-    path: ./spacechatcmpl.py
-    attr: LangBotSpaceChatCompletions

src/langbot/pkg/provider/modelmgr/requesters/tokenpony.yaml

@@ -22,11 +22,8 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: openai
   support_type:
   - llm
   - text-embedding
   provider_category: maas
-execution:
-  python:
-    path: ./tokenponychatcmpl.py
-    attr: TokenPonyChatCompletions

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

@@ -1,17 +0,0 @@
-from __future__ import annotations
-
-import typing
-import openai
-
-from . import chatcmpl
-
-
-class TokenPonyChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """TokenPony ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://api.tokenpony.cn/v1',
-        'timeout': 120,
-    }

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

@@ -1,17 +0,0 @@
-from __future__ import annotations
-
-import typing
-import openai
-
-from . import chatcmpl
-
-
-class VolcArkChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """火山方舟大模型平台 ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://ark.cn-beijing.volces.com/api/v3',
-        'timeout': 120,
-    }

src/langbot/pkg/provider/modelmgr/requesters/volcarkchatcmpl.yaml

@@ -22,10 +22,7 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: openai
   support_type:
   - llm
   provider_category: maas
-execution:
-  python:
-    path: ./volcarkchatcmpl.py
-    attr: VolcArkChatCompletions

src/langbot/pkg/provider/modelmgr/requesters/voyageairerank.yaml

@@ -22,10 +22,7 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: voyage
   support_type:
   - rerank
   provider_category: manufacturer
-execution:
-  python:
-    path: ./chatcmpl.py
-    attr: OpenAIChatCompletions

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

@@ -1,17 +0,0 @@
-from __future__ import annotations
-
-import typing
-import openai
-
-from . import chatcmpl
-
-
-class XaiChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """xAI ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://api.x.ai/v1',
-        'timeout': 120,
-    }

src/langbot/pkg/provider/modelmgr/requesters/xaichatcmpl.yaml

@@ -22,10 +22,7 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: xai
   support_type:
   - llm
   provider_category: manufacturer
-execution:
-  python:
-    path: ./xaichatcmpl.py
-    attr: XaiChatCompletions

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

@@ -1,17 +0,0 @@
-from __future__ import annotations
-
-import typing
-import openai
-
-from . import chatcmpl
-
-
-class ZhipuAIChatCompletions(chatcmpl.OpenAIChatCompletions):
-    """智谱AI ChatCompletion API 请求器"""
-
-    client: openai.AsyncClient
-
-    default_config: dict[str, typing.Any] = {
-        'base_url': 'https://open.bigmodel.cn/api/paas/v4',
-        'timeout': 120,
-    }

src/langbot/pkg/provider/modelmgr/requesters/zhipuaichatcmpl.yaml

@@ -22,10 +22,7 @@ spec:
     type: integer
     required: true
     default: 120
+  litellm_provider: zhipu
   support_type:
   - llm
   provider_category: manufacturer
-execution:
-  python:
-    path: ./zhipuaichatcmpl.py
-    attr: ZhipuAIChatCompletions

tests/unit_tests/provider/__init__.py

@@ -0,0 +1 @@
+"""Provider requester tests"""

tests/unit_tests/provider/test_litellmchat.py

@@ -0,0 +1,635 @@
+"""
+Tests for LiteLLMRequester - unified requester for chat, embedding, and rerank.
+
+These tests verify:
+- Parameter building and LiteLLM API calls
+- Response processing and usage extraction
+- Error handling and exception translation
+- Model name building with provider prefix
+"""
+
+import pytest
+from unittest.mock import Mock, AsyncMock, patch
+
+import litellm
+
+from langbot.pkg.provider.modelmgr.requesters import litellmchat
+from langbot.pkg.provider.modelmgr import errors
+
+
+class MockRuntimeModel:
+    """Mock RuntimeLLMModel for testing"""
+
+    def __init__(self, model_name: str = 'gpt-4o', api_key: str = 'test-key'):
+        self.model_entity = Mock()
+        self.model_entity.name = model_name
+        self.model_entity.extra_args = {}
+        self.provider = Mock()
+        self.provider.token_mgr = Mock()
+        self.provider.token_mgr.get_token = Mock(return_value=api_key)
+
+
+class MockRuntimeEmbeddingModel:
+    """Mock RuntimeEmbeddingModel for testing"""
+
+    def __init__(self, model_name: str = 'text-embedding-3-small', api_key: str = 'test-key'):
+        self.model_entity = Mock()
+        self.model_entity.name = model_name
+        self.model_entity.extra_args = {}
+        self.provider = Mock()
+        self.provider.token_mgr = Mock()
+        self.provider.token_mgr.get_token = Mock(return_value=api_key)
+
+
+class MockRuntimeRerankModel:
+    """Mock RuntimeRerankModel for testing"""
+
+    def __init__(self, model_name: str = 'cohere/rerank-english-v3.0', api_key: str = 'test-key'):
+        self.model_entity = Mock()
+        self.model_entity.name = model_name
+        self.model_entity.extra_args = {}
+        self.provider = Mock()
+        self.provider.token_mgr = Mock()
+        self.provider.token_mgr.get_token = Mock(return_value=api_key)
+
+
+class TestBuildLiteLLMModelName:
+    """Test _build_litellm_model_name method"""
+
+    def test_no_provider_prefix(self):
+        """Test model name without provider prefix"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={'custom_llm_provider': ''})
+        result = requester._build_litellm_model_name('gpt-4o')
+        assert result == 'gpt-4o'
+
+    def test_with_provider_prefix(self):
+        """Test model name with provider prefix"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={'custom_llm_provider': 'openai'})
+        result = requester._build_litellm_model_name('gpt-4o')
+        assert result == 'openai/gpt-4o'
+
+    def test_override_provider(self):
+        """Test override provider via parameter"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={'custom_llm_provider': 'openai'})
+        result = requester._build_litellm_model_name('claude-3', custom_llm_provider='anthropic')
+        assert result == 'anthropic/claude-3'
+
+
+class TestExtractUsage:
+    """Test _extract_usage method"""
+
+    def test_extract_usage_with_data(self):
+        """Test extraction with valid usage data"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        response = Mock()
+        response.usage = Mock()
+        response.usage.prompt_tokens = 100
+        response.usage.completion_tokens = 50
+        response.usage.total_tokens = 150
+
+        result = requester._extract_usage(response)
+
+        assert result['prompt_tokens'] == 100
+        assert result['completion_tokens'] == 50
+        assert result['total_tokens'] == 150
+        assert result['input_tokens'] == 100  # Compatibility alias
+        assert result['output_tokens'] == 50  # Compatibility alias
+
+    def test_extract_usage_with_zero_values(self):
+        """Test extraction when values are 0"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        response = Mock()
+        response.usage = Mock()
+        response.usage.prompt_tokens = 0
+        response.usage.completion_tokens = 0
+        response.usage.total_tokens = 0
+
+        result = requester._extract_usage(response)
+
+        assert result['prompt_tokens'] == 0
+        assert result['completion_tokens'] == 0
+
+
+class TestProcessThinkingContent:
+    """Test _process_thinking_content method"""
+
+    def test_no_thinking_markers(self):
+        """Test content without thinking markers"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        result = requester._process_thinking_content('Hello world', None, remove_think=True)
+        assert result == 'Hello world'
+
+    def test_remove_thinking_markers(self):
+        """Test removing thinking markers when remove_think=True"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        content = 'CRETIRE_REASONING_BEGINkLet me think...CRETIRE_REASONING_ENDk The answer is 42.'
+        result = requester._process_thinking_content(content, None, remove_think=True)
+        assert result == 'The answer is 42.'
+
+    def test_preserve_thinking_markers(self):
+        """Test preserving thinking markers when remove_think=False"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        content = 'CRETIRE_REASONING_BEGINkLet me think...CRETIRE_REASONING_ENDk The answer is 42.'
+        result = requester._process_thinking_content(content, None, remove_think=False)
+        assert 'CRETIRE_REASONING_BEGINk' in result
+        assert 'The answer is 42.' in result
+
+    def test_empty_content(self):
+        """Test empty content"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        result = requester._process_thinking_content('', None, remove_think=True)
+        assert result == ''
+
+
+class TestBuildCommonArgs:
+    """Test _build_common_args method"""
+
+    def test_build_args_with_all_params(self):
+        """Test building args with all config params"""
+        requester = litellmchat.LiteLLMRequester(
+            ap=Mock(),
+            config={
+                'base_url': 'https://api.openai.com/v1',
+                'timeout': 60,
+                'drop_params': True,
+                'num_retries': 3,
+                'api_version': '2024-01-01',
+            },
+        )
+
+        args = {}
+        requester._build_common_args(args)
+
+        assert args['api_base'] == 'https://api.openai.com/v1'
+        assert args['timeout'] == 60
+        assert args['drop_params'] == True
+        assert args['num_retries'] == 3
+        assert args['api_version'] == '2024-01-01'
+
+    def test_build_args_without_retry_params(self):
+        """Test building args without retry params for embedding/rerank"""
+        requester = litellmchat.LiteLLMRequester(
+            ap=Mock(),
+            config={
+                'base_url': 'https://api.openai.com/v1',
+                'timeout': 60,
+                'num_retries': 3,
+            },
+        )
+
+        args = {}
+        requester._build_common_args(args, include_retry_params=False)
+
+        assert args['api_base'] == 'https://api.openai.com/v1'
+        assert args['timeout'] == 60
+        assert 'num_retries' not in args
+
+
+class TestHandleLiteLLMError:
+    """Test _handle_litellm_error method"""
+
+    def test_bad_request_error(self):
+        """Test BadRequestError translation"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        # Create proper LiteLLM exception with required args
+        error = litellm.BadRequestError(message='test error', model='gpt-4o', llm_provider='openai')
+
+        with pytest.raises(errors.RequesterError) as exc_info:
+            requester._handle_litellm_error(error)
+
+        assert '请求参数错误' in str(exc_info.value)
+
+    def test_authentication_error(self):
+        """Test AuthenticationError translation"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        error = litellm.AuthenticationError(message='invalid key', model='gpt-4o', llm_provider='openai')
+
+        with pytest.raises(errors.RequesterError) as exc_info:
+            requester._handle_litellm_error(error)
+
+        assert 'API key 无效' in str(exc_info.value)
+
+    def test_rate_limit_error(self):
+        """Test RateLimitError translation"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        error = litellm.RateLimitError(message='rate limited', model='gpt-4o', llm_provider='openai')
+
+        with pytest.raises(errors.RequesterError) as exc_info:
+            requester._handle_litellm_error(error)
+
+        assert '请求过于频繁' in str(exc_info.value)
+
+    def test_timeout_error(self):
+        """Test Timeout translation"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        error = litellm.Timeout(message='timeout', model='gpt-4o', llm_provider='openai')
+
+        with pytest.raises(errors.RequesterError) as exc_info:
+            requester._handle_litellm_error(error)
+
+        assert '请求超时' in str(exc_info.value)
+
+    def test_context_window_error(self):
+        """Test ContextWindowExceededError translation"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        error = litellm.ContextWindowExceededError(message='context too long', model='gpt-4o', llm_provider='openai')
+
+        with pytest.raises(errors.RequesterError) as exc_info:
+            requester._handle_litellm_error(error)
+
+        assert '上下文长度超限' in str(exc_info.value)
+
+    def test_unknown_error(self):
+        """Test unknown error translation"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        with pytest.raises(errors.RequesterError) as exc_info:
+            requester._handle_litellm_error(Exception('unknown'))
+
+        assert '未知错误' in str(exc_info.value)
+
+
+class TestInvokeLLM:
+    """Test invoke_llm method"""
+
+    @pytest.mark.asyncio
+    async def test_invoke_llm_basic(self):
+        """Test basic LLM invocation"""
+        mock_ap = Mock()
+        mock_ap.tool_mgr = Mock()
+        mock_ap.tool_mgr.generate_tools_for_openai = AsyncMock(return_value=None)
+
+        requester = litellmchat.LiteLLMRequester(
+            ap=mock_ap,
+            config={
+                'base_url': 'https://api.openai.com/v1',
+                'timeout': 60,
+            },
+        )
+
+        model = MockRuntimeModel('gpt-4o', 'test-api-key')
+
+        # Mock LiteLLM response
+        mock_response = Mock()
+        mock_response.choices = [Mock()]
+        mock_response.choices[0].message = Mock()
+        mock_response.choices[0].message.model_dump = Mock(
+            return_value={
+                'role': 'assistant',
+                'content': 'Hello! How can I help you?',
+            }
+        )
+        mock_response.usage = Mock()
+        mock_response.usage.prompt_tokens = 10
+        mock_response.usage.completion_tokens = 20
+        mock_response.usage.total_tokens = 30
+
+        import langbot_plugin.api.entities.builtin.provider.message as provider_message
+
+        messages = [provider_message.Message(role='user', content='Hello')]
+
+        # Patch acompletion at the import location
+        with patch.object(litellmchat, 'acompletion', new_callable=AsyncMock, return_value=mock_response):
+            result_msg, usage = await requester.invoke_llm(
+                query=None,
+                model=model,
+                messages=messages,
+            )
+
+            assert result_msg.role == 'assistant'
+            assert result_msg.content == 'Hello! How can I help you?'
+            assert usage['prompt_tokens'] == 10
+            assert usage['completion_tokens'] == 20
+
+    @pytest.mark.asyncio
+    async def test_invoke_llm_with_tools(self):
+        """Test LLM invocation with function calling"""
+        mock_ap = Mock()
+        mock_ap.tool_mgr = Mock()
+        mock_ap.tool_mgr.generate_tools_for_openai = AsyncMock(
+            return_value=[{'type': 'function', 'function': {'name': 'get_weather'}}]
+        )
+
+        requester = litellmchat.LiteLLMRequester(ap=mock_ap, config={})
+
+        model = MockRuntimeModel('gpt-4o', 'test-api-key')
+
+        mock_response = Mock()
+        mock_response.choices = [Mock()]
+        mock_response.choices[0].message = Mock()
+        mock_response.choices[0].message.model_dump = Mock(
+            return_value={
+                'role': 'assistant',
+                'content': None,
+                'tool_calls': [
+                    {'id': 'call_123', 'type': 'function', 'function': {'name': 'get_weather', 'arguments': '{}'}}
+                ],
+            }
+        )
+        mock_response.usage = Mock()
+        mock_response.usage.prompt_tokens = 15
+        mock_response.usage.completion_tokens = 10
+        mock_response.usage.total_tokens = 25
+
+        import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
+        import langbot_plugin.api.entities.builtin.provider.message as provider_message
+
+        messages = [provider_message.Message(role='user', content='What is the weather?')]
+        # Create proper LLMTool with all required fields
+        funcs = [Mock(spec=resource_tool.LLMTool)]
+        funcs[0].name = 'get_weather'
+        funcs[0].description = 'Get weather'
+
+        with patch.object(litellmchat, 'acompletion', new_callable=AsyncMock, return_value=mock_response):
+            result_msg, usage = await requester.invoke_llm(
+                query=None,
+                model=model,
+                messages=messages,
+                funcs=funcs,
+            )
+
+            assert result_msg.tool_calls is not None
+
+    @pytest.mark.asyncio
+    async def test_invoke_llm_error_handling(self):
+        """Test LLM invocation error handling"""
+        mock_ap = Mock()
+        mock_ap.tool_mgr = Mock()
+        mock_ap.tool_mgr.generate_tools_for_openai = AsyncMock(return_value=None)
+
+        requester = litellmchat.LiteLLMRequester(ap=mock_ap, config={})
+
+        model = MockRuntimeModel('gpt-4o', 'test-api-key')
+
+        import langbot_plugin.api.entities.builtin.provider.message as provider_message
+
+        messages = [provider_message.Message(role='user', content='Hello')]
+
+        error = litellm.AuthenticationError(message='invalid key', model='gpt-4o', llm_provider='openai')
+
+        with patch.object(litellmchat, 'acompletion', new_callable=AsyncMock, side_effect=error):
+            with pytest.raises(errors.RequesterError) as exc_info:
+                await requester.invoke_llm(
+                    query=None,
+                    model=model,
+                    messages=messages,
+                )
+
+            assert 'API key 无效' in str(exc_info.value)
+
+
+class TestInvokeEmbedding:
+    """Test invoke_embedding method"""
+
+    @pytest.mark.asyncio
+    async def test_invoke_embedding_basic(self):
+        """Test basic embedding invocation"""
+        requester = litellmchat.LiteLLMRequester(
+            ap=Mock(),
+            config={
+                'base_url': 'https://api.openai.com/v1',
+            },
+        )
+
+        model = MockRuntimeEmbeddingModel('text-embedding-3-small', 'test-api-key')
+
+        # Mock LiteLLM embedding response
+        mock_response = Mock()
+        mock_response.data = [
+            Mock(embedding=[0.1, 0.2, 0.3]),
+            Mock(embedding=[0.4, 0.5, 0.6]),
+        ]
+        mock_response.usage = Mock()
+        mock_response.usage.prompt_tokens = 20
+        mock_response.usage.completion_tokens = 0
+        mock_response.usage.total_tokens = 20
+
+        with patch.object(litellmchat, 'aembedding', new_callable=AsyncMock, return_value=mock_response):
+            embeddings, usage = await requester.invoke_embedding(
+                model=model,
+                input_text=['Hello', 'World'],
+            )
+
+            assert len(embeddings) == 2
+            assert embeddings[0] == [0.1, 0.2, 0.3]
+            assert embeddings[1] == [0.4, 0.5, 0.6]
+            assert usage['prompt_tokens'] == 20
+
+
+class TestInvokeRerank:
+    """Test invoke_rerank method"""
+
+    @pytest.mark.asyncio
+    async def test_invoke_rerank_basic(self):
+        """Test basic rerank invocation"""
+        requester = litellmchat.LiteLLMRequester(
+            ap=Mock(),
+            config={
+                'base_url': 'https://api.cohere.ai',
+            },
+        )
+
+        model = MockRuntimeRerankModel('rerank-english-v3.0', 'test-api-key')
+
+        # Mock LiteLLM rerank response
+        mock_response = Mock()
+        mock_response.results = [
+            {'index': 0, 'relevance_score': 0.95},
+            {'index': 1, 'relevance_score': 0.3},
+            {'index': 2, 'relevance_score': 0.8},
+        ]
+
+        with patch.object(litellmchat, 'arerank', new_callable=AsyncMock, return_value=mock_response):
+            results = await requester.invoke_rerank(
+                model=model,
+                query='What is the capital of France?',
+                documents=['Paris is the capital.', 'London is a city.', 'France is in Europe.'],
+            )
+
+            assert len(results) == 3
+            # Scores should be normalized
+            assert results[0]['index'] == 0
+            assert results[0]['relevance_score'] >= 0 and results[0]['relevance_score'] <= 1
+
+    @pytest.mark.asyncio
+    async def test_invoke_rerank_normalization(self):
+        """Test rerank score normalization"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        model = MockRuntimeRerankModel('rerank-english-v3.0', 'test-api-key')
+
+        # Mock response with varying scores
+        mock_response = Mock()
+        mock_response.results = [
+            {'index': 0, 'relevance_score': 0.9},
+            {'index': 1, 'relevance_score': 0.1},
+        ]
+
+        with patch.object(litellmchat, 'arerank', new_callable=AsyncMock, return_value=mock_response):
+            results = await requester.invoke_rerank(
+                model=model,
+                query='test query',
+                documents=['doc1', 'doc2'],
+            )
+
+            # After normalization: 0.9 -> 1.0, 0.1 -> 0.0
+            assert results[0]['relevance_score'] == 1.0
+            assert results[1]['relevance_score'] == 0.0
+
+    @pytest.mark.asyncio
+    async def test_invoke_rerank_single_document(self):
+        """Test rerank with single document (no normalization needed)"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        model = MockRuntimeRerankModel('rerank-english-v3.0', 'test-api-key')
+
+        mock_response = Mock()
+        mock_response.results = [
+            {'index': 0, 'relevance_score': 0.5},
+        ]
+
+        with patch.object(litellmchat, 'arerank', new_callable=AsyncMock, return_value=mock_response):
+            results = await requester.invoke_rerank(
+                model=model,
+                query='test query',
+                documents=['doc1'],
+            )
+
+            assert len(results) == 1
+            # Single score stays as is (min==max, no normalization)
+            assert results[0]['relevance_score'] == 0.5
+
+
+class TestConvertMessages:
+    """Test _convert_messages method"""
+
+    def test_convert_simple_message(self):
+        """Test converting simple text message"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        import langbot_plugin.api.entities.builtin.provider.message as provider_message
+
+        messages = [provider_message.Message(role='user', content='Hello')]
+        result = requester._convert_messages(messages)
+
+        assert len(result) == 1
+        assert result[0]['role'] == 'user'
+        assert result[0]['content'] == 'Hello'
+
+    def test_convert_message_with_image_base64(self):
+        """Test converting message with image_base64 content"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        import langbot_plugin.api.entities.builtin.provider.message as provider_message
+
+        messages = [
+            provider_message.Message(
+                role='user',
+                content=[
+                    {'type': 'text', 'text': 'What is in this image?'},
+                    {'type': 'image_base64', 'image_base64': 'data:image/png;base64,abc123'},
+                ],
+            )
+        ]
+        result = requester._convert_messages(messages)
+
+        assert len(result) == 1
+        content = result[0]['content']
+        assert isinstance(content, list)
+        # Check image_base64 converted to image_url
+        image_part = [p for p in content if p.get('type') == 'image_url'][0]
+        assert 'image_url' in image_part
+        assert image_part['image_url']['url'] == 'data:image/png;base64,abc123'
+
+    def test_convert_message_with_multiple_text_parts(self):
+        """Test converting message with multiple text parts (LiteLLM handles this)"""
+        requester = litellmchat.LiteLLMRequester(ap=Mock(), config={})
+
+        import langbot_plugin.api.entities.builtin.provider.message as provider_message
+
+        messages = [
+            provider_message.Message(
+                role='user',
+                content=[
+                    {'type': 'text', 'text': 'Hello'},
+                    {'type': 'text', 'text': 'World'},
+                ],
+            )
+        ]
+        result = requester._convert_messages(messages)
+
+        assert len(result) == 1
+        # LiteLLM handles multiple text parts, we pass them through
+        assert isinstance(result[0]['content'], list)
+
+
+class TestScanModels:
+    """Test scan_models method"""
+
+    @pytest.mark.asyncio
+    async def test_scan_models_basic(self):
+        """Test basic model scanning"""
+        requester = litellmchat.LiteLLMRequester(
+            ap=Mock(),
+            config={
+                'base_url': 'https://api.openai.com/v1',
+                'timeout': 60,
+            },
+        )
+
+        # Mock httpx response
+        mock_response = Mock()
+        mock_response.json = Mock(
+            return_value={
+                'data': [
+                    {'id': 'gpt-4o'},
+                    {'id': 'text-embedding-3-small'},
+                    {'id': 'gpt-3.5-turbo'},
+                ]
+            }
+        )
+        mock_response.raise_for_status = Mock()
+
+        with patch('httpx.AsyncClient') as mock_client:
+            mock_client.return_value.__aenter__ = AsyncMock(return_value=Mock())
+            mock_client.return_value.__aenter__.return_value.get = AsyncMock(return_value=mock_response)
+
+            result = await requester.scan_models(api_key='test-key')
+
+            assert 'models' in result
+            assert len(result['models']) == 3
+            # Check LLM models are first
+            assert result['models'][0]['type'] == 'llm'
+            # Check embedding model is detected
+            embedding_models = [m for m in result['models'] if m['type'] == 'embedding']
+            assert len(embedding_models) == 1
+
+    @pytest.mark.asyncio
+    async def test_scan_models_no_base_url(self):
+        """Test scan_models without base_url raises error"""
+        requester = litellmchat.LiteLLMRequester(
+            ap=Mock(),
+            config={
+                'base_url': '',
+            },
+        )
+
+        with pytest.raises(errors.RequesterError) as exc_info:
+            await requester.scan_models()
+
+        assert 'Base URL required' in str(exc_info.value)
+
+
+if __name__ == '__main__':
+    pytest.main([__file__, '-v'])