lite-agent 0.4.1__py3-none-any.whl → 0.5.0__py3-none-any.whl

lite_agent/agent.py

@@ -7,7 +7,7 @@ from funcall import Funcall
 from jinja2 import Environment, FileSystemLoader
 from litellm import CustomStreamWrapper
 
-from lite_agent.client import BaseLLMClient, LiteLLMClient
+from lite_agent.client import BaseLLMClient, LiteLLMClient, ReasoningConfig
 from lite_agent.loggers import logger
 from lite_agent.stream_handlers import litellm_completion_stream_handler, litellm_response_stream_handler
 from lite_agent.types import AgentChunk, FunctionCallEvent, FunctionCallOutputEvent, RunnerMessages, ToolCall, message_to_llm_dict, system_message_to_llm_dict
@@ -32,15 +32,21 @@ class Agent:
         handoffs: list["Agent"] | None = None,
         message_transfer: Callable[[RunnerMessages], RunnerMessages] | None = None,
         completion_condition: str = "stop",
+        reasoning: ReasoningConfig = None,
     ) -> None:
         self.name = name
         self.instructions = instructions
+        self.reasoning = reasoning
+
         if isinstance(model, BaseLLMClient):
             # If model is a BaseLLMClient instance, use it directly
             self.client = model
         else:
             # Otherwise, create a LitellmClient instance
-            self.client = LiteLLMClient(model=model)
+            self.client = LiteLLMClient(
+                model=model,
+                reasoning=reasoning,
+            )
         self.completion_condition = completion_condition
         self.handoffs = handoffs if handoffs else []
         self._parent: Agent | None = None
@@ -269,7 +275,12 @@ class Agent:
                 res.append(message)
         return res
 
-    async def completion(self, messages: RunnerMessages, record_to_file: Path | None = None) -> AsyncGenerator[AgentChunk, None]:
+    async def completion(
+        self,
+        messages: RunnerMessages,
+        record_to_file: Path | None = None,
+        reasoning: ReasoningConfig = None,
+    ) -> AsyncGenerator[AgentChunk, None]:
         # Apply message transfer callback if provided - always use legacy format for LLM compatibility
         processed_messages = messages
         if self.message_transfer:
@@ -284,6 +295,7 @@ class Agent:
             messages=self.message_histories,
             tools=tools,
             tool_choice="auto",  # TODO: make this configurable
+            reasoning=reasoning,
         )
 
         # Ensure resp is a CustomStreamWrapper
@@ -292,7 +304,12 @@ class Agent:
         msg = "Response is not a CustomStreamWrapper, cannot stream chunks."
         raise TypeError(msg)
 
-    async def responses(self, messages: RunnerMessages, record_to_file: Path | None = None) -> AsyncGenerator[AgentChunk, None]:
+    async def responses(
+        self,
+        messages: RunnerMessages,
+        record_to_file: Path | None = None,
+        reasoning: ReasoningConfig = None,
+    ) -> AsyncGenerator[AgentChunk, None]:
         # Apply message transfer callback if provided - always use legacy format for LLM compatibility
         processed_messages = messages
         if self.message_transfer:
@@ -306,6 +323,7 @@ class Agent:
             messages=self.message_histories,
             tools=tools,
             tool_choice="auto",  # TODO: make this configurable
+            reasoning=reasoning,
         )
         return litellm_response_stream_handler(resp, record_to=record_to_file)
 

lite_agent/client.py

@@ -1,25 +1,81 @@
 import abc
 import os
-from collections.abc import AsyncGenerator
 from typing import Any, Literal
 
 import litellm
-from litellm.types.llms.openai import ResponsesAPIStreamingResponse
 from openai.types.chat import ChatCompletionToolParam
 from openai.types.responses import FunctionToolParam
 
+ReasoningEffort = Literal["minimal", "low", "medium", "high"]
+ThinkingConfig = dict[str, Any] | None
+
+# 统一的推理配置类型
+ReasoningConfig = (
+    str
+    | dict[str, Any]  # {"type": "enabled", "budget_tokens": 2048} 或其他配置
+    | bool  # True/False 简单开关
+    | None  # 不启用推理
+)
+
+
+def parse_reasoning_config(reasoning: ReasoningConfig) -> tuple[ReasoningEffort | None, ThinkingConfig]:
+    """
+    解析统一的推理配置，返回 reasoning_effort 和 thinking_config。
+
+    Args:
+        reasoning: 统一的推理配置
+            - str: "minimal", "low", "medium", "high" -> reasoning_effort
+            - dict: {"type": "enabled", "budget_tokens": N} -> thinking_config
+            - bool: True -> "medium", False -> None
+            - None: 不启用推理
+
+    Returns:
+        tuple: (reasoning_effort, thinking_config)
+    """
+    if reasoning is None:
+        return None, None
+    if isinstance(reasoning, str):
+        # 字符串类型，使用 reasoning_effort
+        return reasoning, None
+    if isinstance(reasoning, dict):
+        # 字典类型，使用 thinking_config
+        return None, reasoning
+    if isinstance(reasoning, bool):
+        # 布尔类型，True 使用默认的 medium，False 不启用
+        return "medium" if reasoning else None, None
+    # 其他类型，默认不启用
+    return None, None
+
 
 class BaseLLMClient(abc.ABC):
     """Base class for LLM clients."""
 
-    def __init__(self, *, model: str, api_key: str | None = None, api_base: str | None = None, api_version: str | None = None):
+    def __init__(
+        self,
+        *,
+        model: str,
+        api_key: str | None = None,
+        api_base: str | None = None,
+        api_version: str | None = None,
+        reasoning: ReasoningConfig = None,
+    ):
         self.model = model
         self.api_key = api_key
         self.api_base = api_base
         self.api_version = api_version
 
+        # 处理推理配置
+        self.reasoning_effort, self.thinking_config = parse_reasoning_config(reasoning)
+
     @abc.abstractmethod
-    async def completion(self, messages: list[Any], tools: list[ChatCompletionToolParam] | None = None, tool_choice: str = "auto") -> Any:  # noqa: ANN401
+    async def completion(
+        self,
+        messages: list[Any],
+        tools: list[ChatCompletionToolParam] | None = None,
+        tool_choice: str = "auto",
+        reasoning: ReasoningConfig = None,
+        **kwargs: Any,  # noqa: ANN401
+    ) -> Any:  # noqa: ANN401
         """Perform a completion request to the LLM."""
 
     @abc.abstractmethod
@@ -28,42 +84,95 @@ class BaseLLMClient(abc.ABC):
         messages: list[dict[str, Any]],  # Changed from ResponseInputParam
         tools: list[FunctionToolParam] | None = None,
         tool_choice: Literal["none", "auto", "required"] = "auto",
-    ) -> AsyncGenerator[ResponsesAPIStreamingResponse, None]:
+        reasoning: ReasoningConfig = None,
+        **kwargs: Any,  # noqa: ANN401
+    ) -> Any:  # noqa: ANN401
         """Perform a response request to the LLM."""
 
 
 class LiteLLMClient(BaseLLMClient):
-    async def completion(self, messages: list[Any], tools: list[ChatCompletionToolParam] | None = None, tool_choice: str = "auto") -> Any:  # noqa: ANN401
+    def _resolve_reasoning_params(
+        self,
+        reasoning: ReasoningConfig,
+    ) -> tuple[ReasoningEffort | None, ThinkingConfig]:
+        """解析推理配置参数。"""
+        if reasoning is not None:
+            return parse_reasoning_config(reasoning)
+
+        # 使用实例默认值
+        return self.reasoning_effort, self.thinking_config
+
+    async def completion(
+        self,
+        messages: list[Any],
+        tools: list[ChatCompletionToolParam] | None = None,
+        tool_choice: str = "auto",
+        reasoning: ReasoningConfig = None,
+        **kwargs: Any,  # noqa: ANN401
+    ) -> Any:  # noqa: ANN401
         """Perform a completion request to the Litellm API."""
-        return await litellm.acompletion(
-            model=self.model,
-            messages=messages,
-            tools=tools,
-            tool_choice=tool_choice,
-            api_version=self.api_version,
-            api_key=self.api_key,
-            api_base=self.api_base,
-            stream=True,
+
+        # 处理推理配置参数
+        final_reasoning_effort, final_thinking_config = self._resolve_reasoning_params(
+            reasoning,
         )
 
+        # Prepare completion parameters
+        completion_params = {
+            "model": self.model,
+            "messages": messages,
+            "tools": tools,
+            "tool_choice": tool_choice,
+            "api_version": self.api_version,
+            "api_key": self.api_key,
+            "api_base": self.api_base,
+            "stream": True,
+            **kwargs,
+        }
+
+        # Add reasoning parameters if specified
+        if final_reasoning_effort is not None:
+            completion_params["reasoning_effort"] = final_reasoning_effort
+        if final_thinking_config is not None:
+            completion_params["thinking"] = final_thinking_config
+
+        return await litellm.acompletion(**completion_params)
+
     async def responses(
         self,
         messages: list[dict[str, Any]],  # Changed from ResponseInputParam
         tools: list[FunctionToolParam] | None = None,
         tool_choice: Literal["none", "auto", "required"] = "auto",
-    ) -> AsyncGenerator[ResponsesAPIStreamingResponse, None]:
+        reasoning: ReasoningConfig = None,
+        **kwargs: Any,  # noqa: ANN401
+    ) -> Any:  # type: ignore[return]  # noqa: ANN401
         """Perform a response request to the Litellm API."""
 
         os.environ["DISABLE_AIOHTTP_TRANSPORT"] = "True"
 
-        return await litellm.aresponses(
-            model=self.model,
-            input=messages,  # type: ignore[arg-type]
-            tools=tools,
-            tool_choice=tool_choice,
-            api_version=self.api_version,
-            api_key=self.api_key,
-            api_base=self.api_base,
-            stream=True,
-            store=False,
+        # 处理推理配置参数
+        final_reasoning_effort, final_thinking_config = self._resolve_reasoning_params(
+            reasoning,
         )
+
+        # Prepare response parameters
+        response_params = {
+            "model": self.model,
+            "input": messages,  # type: ignore[arg-type]
+            "tools": tools,
+            "tool_choice": tool_choice,
+            "api_version": self.api_version,
+            "api_key": self.api_key,
+            "api_base": self.api_base,
+            "stream": True,
+            "store": False,
+            **kwargs,
+        }
+
+        # Add reasoning parameters if specified
+        if final_reasoning_effort is not None:
+            response_params["reasoning_effort"] = final_reasoning_effort
+        if final_thinking_config is not None:
+            response_params["thinking"] = final_thinking_config
+
+        return await litellm.aresponses(**response_params)  # type: ignore[return-value]

lite_agent/runner.py

@@ -168,13 +168,14 @@ class Runner:
         """Collect all chunks from an async generator into a list."""
         return [chunk async for chunk in stream]
 
-    def run(
+    def run(  # noqa: PLR0913
         self,
         user_input: UserInput,
         max_steps: int = 20,
         includes: Sequence[AgentChunkType] | None = None,
         context: "Any | None" = None,  # noqa: ANN401
         record_to: PathLike | str | None = None,
+        agent_kwargs: dict[str, Any] | None = None,
     ) -> AsyncGenerator[AgentChunk, None]:
         """Run the agent and return a RunResponse object that can be asynchronously iterated for each chunk."""
         includes = self._normalize_includes(includes)
@@ -188,9 +189,16 @@ class Runner:
             case _:
                 # Handle single message (BaseModel, TypedDict, or dict)
                 self.append_message(user_input)  # type: ignore[arg-type]
-        return self._run(max_steps, includes, self._normalize_record_path(record_to), context=context)
+        return self._run(max_steps, includes, self._normalize_record_path(record_to), context=context, agent_kwargs=agent_kwargs)
 
-    async def _run(self, max_steps: int, includes: Sequence[AgentChunkType], record_to: Path | None = None, context: Any | None = None) -> AsyncGenerator[AgentChunk, None]:  # noqa: ANN401
+    async def _run(
+        self,
+        max_steps: int,
+        includes: Sequence[AgentChunkType],
+        record_to: Path | None = None,
+        context: Any | None = None,  # noqa: ANN401
+        agent_kwargs: dict[str, Any] | None = None,
+    ) -> AsyncGenerator[AgentChunk, None]:
         """Run the agent and return a RunResponse object that can be asynchronously iterated for each chunk."""
         logger.debug(f"Running agent with messages: {self.messages}")
         steps = 0
@@ -213,11 +221,24 @@ class Runner:
             logger.debug(f"Step {steps}: finish_reason={finish_reason}, is_finish()={is_finish()}")
             # Convert to legacy format only when needed for LLM communication
             # This allows us to keep the new format internally but ensures compatibility
+            # Extract agent kwargs for reasoning configuration
+            reasoning = None
+            if agent_kwargs:
+                reasoning = agent_kwargs.get("reasoning")
+
             match self.api:
                 case "completion":
-                    resp = await self.agent.completion(self.messages, record_to_file=record_to)
+                    resp = await self.agent.completion(
+                        self.messages,
+                        record_to_file=record_to,
+                        reasoning=reasoning,
+                    )
                 case "responses":
-                    resp = await self.agent.responses(self.messages, record_to_file=record_to)
+                    resp = await self.agent.responses(
+                        self.messages,
+                        record_to_file=record_to,
+                        reasoning=reasoning,
+                    )
                 case _:
                     msg = f"Unknown API type: {self.api}"
                     raise ValueError(msg)
@@ -367,11 +388,6 @@ class Runner:
                 msg = "Cannot continue running without a valid last message from the assistant."
                 raise ValueError(msg)
 
-            last_message = self.messages[-1]
-            if not (isinstance(last_message, NewAssistantMessage) or (hasattr(last_message, "role") and getattr(last_message, "role", None) == "assistant")):
-                msg = "Cannot continue running without a valid last message from the assistant."
-                raise ValueError(msg)
-
             resp = self._run(max_steps=max_steps, includes=includes, record_to=self._normalize_record_path(record_to), context=context)
             async for chunk in resp:
                 yield chunk

{lite_agent-0.4.1.dist-info → lite_agent-0.5.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lite-agent
-Version: 0.4.1
+Version: 0.5.0
 Summary: A lightweight, extensible framework for building AI agent.
 Author-email: Jianqi Pan <jannchie@gmail.com>
 License: MIT

{lite_agent-0.4.1.dist-info → lite_agent-0.5.0.dist-info}/RECORD

@@ -1,11 +1,11 @@
 lite_agent/__init__.py,sha256=Swuefee0etSiaDnn30K2hBNV9UI3hIValW3A-pRE7e0,338
-lite_agent/agent.py,sha256=t4AYlw3aF2DCPXf2W3s7aow0ql1ON5O2Q8VVuyoN6UI,22936
+lite_agent/agent.py,sha256=M0U59KpMy6OGFje6yZuQCYVGr4oBboRwbtImPF59o2w,23314
 lite_agent/chat_display.py,sha256=b0sUH3fkutc4e_KAKH7AtPu2msyLloNIAiWqCNavdds,30533
-lite_agent/client.py,sha256=m2jfBPIsleMZ1QCczjyHND-PIF17kQh4RTuf5FaipGM,2571
+lite_agent/client.py,sha256=HG-NbTIUSFAUAPjRow3TFYJxvTc6Y4bdT2oJWIJNEEk,5963
 lite_agent/loggers.py,sha256=XkNkdqwD_nQGfhQJ-bBWT7koci_mMkNw3aBpyMhOICw,57
 lite_agent/message_transfers.py,sha256=9qucjc-uSIXvVfhcmVRC_0lp0Q8sWp99dV4ReCh6ZlI,4428
 lite_agent/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-lite_agent/runner.py,sha256=ACZuFJ2dNpdg4Tzeg-bl4Th1X14uhHJdELcBWe5E_Us,40155
+lite_agent/runner.py,sha256=U7eVNAJ_VLwgbPPpn-vggSgvBmFl8wMMFWn3mWCsDow,40423
 lite_agent/processors/__init__.py,sha256=ybpAzpMBIE9v5I24wIBZRXeaOaPNTmoKH13aofgNI6Q,234
 lite_agent/processors/completion_event_processor.py,sha256=8fQYRofgBd8t0V3oUakTOmZdv5Q9tCuzADGCGvVgy0k,13442
 lite_agent/processors/response_event_processor.py,sha256=CElJMUzLs8mklVqJtoLiVu-NTq0Dz2NNd9YdAKpjgE0,8088
@@ -18,6 +18,6 @@ lite_agent/types/__init__.py,sha256=QKuhjFWRcpAlsBK9JYgoCABpoQExwhuyGudJoiiqQfs,
 lite_agent/types/events.py,sha256=mFMqV55WWJbPDyb_P61nd3qMLpEnwZgVY6NTKFkINkg,2389
 lite_agent/types/messages.py,sha256=c7nTIWqXNo562het_vaWcZvsoy-adkARwAYn4JNqm0c,9897
 lite_agent/types/tool_calls.py,sha256=Xnut8-2-Ld9vgA2GKJY6BbFlBaAv_n4W7vo7Jx21A-E,260
-lite_agent-0.4.1.dist-info/METADATA,sha256=iQIr1OAdiVK5Ad6Uho65OpqS1u4YC9sOaoxKZ1FssOs,3456
-lite_agent-0.4.1.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-lite_agent-0.4.1.dist-info/RECORD,,
+lite_agent-0.5.0.dist-info/METADATA,sha256=20K2Xirnyawl1uN_I8TLcuGlgRjNhs04hz2BtDDRnbM,3456
+lite_agent-0.5.0.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+lite_agent-0.5.0.dist-info/RECORD,,