lite-agent 0.3.0__py3-none-any.whl → 0.4.0__py3-none-any.whl

lite_agent/client.py

@@ -1,8 +1,12 @@
 import abc
-from typing import Any
+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
 
 
 class BaseLLMClient(abc.ABC):
@@ -18,6 +22,15 @@ class BaseLLMClient(abc.ABC):
     async def completion(self, messages: list[Any], tools: list[ChatCompletionToolParam] | None = None, tool_choice: str = "auto") -> Any:  # noqa: ANN401
         """Perform a completion request to the LLM."""
 
+    @abc.abstractmethod
+    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]:
+        """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
@@ -32,3 +45,25 @@ class LiteLLMClient(BaseLLMClient):
             api_base=self.api_base,
             stream=True,
         )
+
+    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]:
+        """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,
+        )

lite_agent/message_transfers.py

@@ -64,7 +64,15 @@ def _process_message_to_xml(message: dict | object) -> list[str]:
         # Handle Pydantic model format messages
         role = getattr(message, "role", "unknown")
         content = getattr(message, "content", "")
-        if isinstance(content, str):
+
+        # Handle new message format where content is a list
+        if isinstance(content, list):
+            # Extract text from content items
+            text_parts = [item.text for item in content if (hasattr(item, "type") and item.type == "text") or hasattr(item, "text")]
+            content_text = " ".join(text_parts)
+            if content_text:
+                xml_lines.append(f"  <message role='{role}'>{content_text}</message>")
+        elif isinstance(content, str):
             xml_lines.append(f"  <message role='{role}'>{content}</message>")
     elif hasattr(message, "type"):
         # Handle function call messages

lite_agent/processors/__init__.py

@@ -1,3 +1,4 @@
-from lite_agent.processors.stream_chunk_processor import StreamChunkProcessor
+from lite_agent.processors.completion_event_processor import CompletionEventProcessor
+from lite_agent.processors.response_event_processor import ResponseEventProcessor
 
-__all__ = ["StreamChunkProcessor"]
+__all__ = ["CompletionEventProcessor", "ResponseEventProcessor"]

lite_agent/processors/completion_event_processor.py

@@ -0,0 +1,306 @@
+from collections.abc import AsyncGenerator
+from datetime import datetime, timezone
+from typing import Literal
+
+import litellm
+from aiofiles.threadpool.text import AsyncTextIOWrapper
+from litellm.types.utils import ChatCompletionDeltaToolCall, ModelResponseStream, StreamingChoices
+
+from lite_agent.loggers import logger
+from lite_agent.types import (
+    AgentChunk,
+    AssistantMessage,
+    AssistantMessageEvent,
+    AssistantMessageMeta,
+    AssistantTextContent,
+    CompletionRawEvent,
+    ContentDeltaEvent,
+    EventUsage,
+    FunctionCallDeltaEvent,
+    FunctionCallEvent,
+    MessageUsage,
+    NewAssistantMessage,
+    Timing,
+    TimingEvent,
+    ToolCall,
+    ToolCallFunction,
+    UsageEvent,
+)
+
+
+class CompletionEventProcessor:
+    """Processor for handling completion event"""
+
+    def __init__(self) -> None:
+        self._current_message: AssistantMessage | None = None
+        self.processing_chunk: Literal["content", "tool_calls"] | None = None
+        self.processing_function: str | None = None
+        self.last_processed_chunk: ModelResponseStream | None = None
+        self.yielded_content = False
+        self.yielded_function = set()
+        self._start_time: datetime | None = None
+        self._first_output_time: datetime | None = None
+        self._output_complete_time: datetime | None = None
+        self._usage_time: datetime | None = None
+        self._usage_data: dict[str, int] = {}
+
+    async def process_chunk(
+        self,
+        chunk: ModelResponseStream,
+        record_file: AsyncTextIOWrapper | None = None,
+    ) -> AsyncGenerator[AgentChunk, None]:
+        # Mark start time on first chunk
+        if self._start_time is None:
+            self._start_time = datetime.now(timezone.utc)
+
+        if record_file:
+            await record_file.write(chunk.model_dump_json() + "\n")
+            await record_file.flush()
+        yield CompletionRawEvent(raw=chunk)
+        usage_chunks = self.handle_usage_chunk(chunk)
+        if usage_chunks:
+            for usage_chunk in usage_chunks:
+                yield usage_chunk
+            return
+        if not chunk.choices:
+            return
+
+        choice = chunk.choices[0]
+        delta = choice.delta
+        if delta.tool_calls:
+            if not self.yielded_content:
+                self.yielded_content = True
+                end_time = datetime.now(timezone.utc)
+                latency_ms = None
+                output_time_ms = None
+                # latency_ms: 从开始准备输出到 LLM 输出第一个字符的时间差
+                if self._start_time and self._first_output_time:
+                    latency_ms = int((self._first_output_time - self._start_time).total_seconds() * 1000)
+                # output_time_ms: 从输出第一个字符到输出完成的时间差
+                if self._first_output_time and self._output_complete_time:
+                    output_time_ms = int((self._output_complete_time - self._first_output_time).total_seconds() * 1000)
+
+                usage = MessageUsage(
+                    input_tokens=self._usage_data.get("input_tokens"),
+                    output_tokens=self._usage_data.get("output_tokens"),
+                )
+                meta = AssistantMessageMeta(
+                    sent_at=end_time,
+                    latency_ms=latency_ms,
+                    total_time_ms=output_time_ms,
+                    usage=usage,
+                )
+                # Include accumulated text content in the message
+                content = []
+                if self._current_message and self._current_message.content:
+                    content.append(AssistantTextContent(text=self._current_message.content))
+
+                yield AssistantMessageEvent(
+                    message=NewAssistantMessage(
+                        content=content,
+                        meta=meta,
+                    ),
+                )
+            first_tool_call = delta.tool_calls[0]
+            tool_name = first_tool_call.function.name if first_tool_call.function else ""
+            if tool_name:
+                self.processing_function = tool_name
+        delta = choice.delta
+        if (
+            self._current_message
+            and self._current_message.tool_calls
+            and self.processing_function != self._current_message.tool_calls[-1].function.name
+            and self._current_message.tool_calls[-1].function.name not in self.yielded_function
+        ):
+            tool_call = self._current_message.tool_calls[-1]
+            yield FunctionCallEvent(
+                call_id=tool_call.id,
+                name=tool_call.function.name,
+                arguments=tool_call.function.arguments or "",
+            )
+            self.yielded_function.add(tool_call.function.name)
+        if not self.is_initialized:
+            self.initialize_message(chunk, choice)
+        if delta.content and self._current_message:
+            # Mark first output time if not already set
+            if self._first_output_time is None:
+                self._first_output_time = datetime.now(timezone.utc)
+            self._current_message.content += delta.content
+            yield ContentDeltaEvent(delta=delta.content)
+        if delta.tool_calls is not None:
+            self.update_tool_calls(delta.tool_calls)
+            if delta.tool_calls and self.current_message.tool_calls:
+                tool_call = delta.tool_calls[0]
+                message_tool_call = self.current_message.tool_calls[-1]
+                yield FunctionCallDeltaEvent(
+                    tool_call_id=message_tool_call.id,
+                    name=message_tool_call.function.name,
+                    arguments_delta=tool_call.function.arguments or "",
+                )
+        if choice.finish_reason:
+            # Mark output complete time when finish_reason appears
+            if self._output_complete_time is None:
+                self._output_complete_time = datetime.now(timezone.utc)
+
+            if self.current_message.tool_calls:
+                tool_call = self.current_message.tool_calls[-1]
+                yield FunctionCallEvent(
+                    call_id=tool_call.id,
+                    name=tool_call.function.name,
+                    arguments=tool_call.function.arguments or "",
+                )
+            if not self.yielded_content:
+                self.yielded_content = True
+                end_time = datetime.now(timezone.utc)
+                latency_ms = None
+                output_time_ms = None
+                # latency_ms: 从开始准备输出到 LLM 输出第一个字符的时间差
+                if self._start_time and self._first_output_time:
+                    latency_ms = int((self._first_output_time - self._start_time).total_seconds() * 1000)
+                # output_time_ms: 从输出第一个字符到输出完成的时间差
+                if self._first_output_time and self._output_complete_time:
+                    output_time_ms = int((self._output_complete_time - self._first_output_time).total_seconds() * 1000)
+
+                usage = MessageUsage(
+                    input_tokens=self._usage_data.get("input_tokens"),
+                    output_tokens=self._usage_data.get("output_tokens"),
+                )
+                meta = AssistantMessageMeta(
+                    sent_at=end_time,
+                    latency_ms=latency_ms,
+                    total_time_ms=output_time_ms,
+                    usage=usage,
+                )
+                # Include accumulated text content in the message
+                content = []
+                if self._current_message and self._current_message.content:
+                    content.append(AssistantTextContent(text=self._current_message.content))
+
+                yield AssistantMessageEvent(
+                    message=NewAssistantMessage(
+                        content=content,
+                        meta=meta,
+                    ),
+                )
+        self.last_processed_chunk = chunk
+
+    def handle_usage_chunk(self, chunk: ModelResponseStream) -> list[AgentChunk]:
+        usage = getattr(chunk, "usage", None)
+        if usage:
+            # Mark usage time
+            self._usage_time = datetime.now(timezone.utc)
+            # Store usage data for meta information
+            self._usage_data["input_tokens"] = usage["prompt_tokens"]
+            self._usage_data["output_tokens"] = usage["completion_tokens"]
+
+            results = []
+
+            # First yield usage event
+            results.append(UsageEvent(usage=EventUsage(input_tokens=usage["prompt_tokens"], output_tokens=usage["completion_tokens"])))
+
+            # Then yield timing event if we have timing data
+            if self._start_time and self._first_output_time and self._output_complete_time:
+                latency_ms = int((self._first_output_time - self._start_time).total_seconds() * 1000)
+                output_time_ms = int((self._output_complete_time - self._first_output_time).total_seconds() * 1000)
+
+                results.append(
+                    TimingEvent(
+                        timing=Timing(
+                            latency_ms=latency_ms,
+                            output_time_ms=output_time_ms,
+                        ),
+                    ),
+                )
+
+            return results
+        return []
+
+    def initialize_message(self, chunk: ModelResponseStream, choice: StreamingChoices) -> None:
+        """Initialize the message object"""
+        delta = choice.delta
+        if delta.role != "assistant":
+            logger.warning("Skipping chunk with role: %s", delta.role)
+            return
+        self._current_message = AssistantMessage(
+            id=chunk.id,
+            index=choice.index,
+            role=delta.role,
+            content="",
+        )
+        logger.debug('Initialized new message: "%s"', self._current_message.id)
+
+    def update_content(self, content: str) -> None:
+        """Update message content"""
+        if self._current_message and content:
+            self._current_message.content += content
+
+    def _initialize_tool_calls(self, tool_calls: list[litellm.ChatCompletionMessageToolCall]) -> None:
+        """Initialize tool calls"""
+        if not self._current_message:
+            return
+
+        self._current_message.tool_calls = []
+        for call in tool_calls:
+            logger.debug("Create new tool call: %s", call.id)
+
+    def _update_tool_calls(self, tool_calls: list[litellm.ChatCompletionMessageToolCall]) -> None:
+        """Update existing tool calls"""
+        if not self._current_message:
+            return
+        if not hasattr(self._current_message, "tool_calls"):
+            self._current_message.tool_calls = []
+        if not self._current_message.tool_calls:
+            return
+        if not tool_calls:
+            return
+        for current_call, new_call in zip(self._current_message.tool_calls, tool_calls, strict=False):
+            if new_call.function.arguments and current_call.function.arguments:
+                current_call.function.arguments += new_call.function.arguments
+            if new_call.type and new_call.type == "function":
+                current_call.type = new_call.type
+            elif new_call.type:
+                logger.warning("Unexpected tool call type: %s", new_call.type)
+
+    def update_tool_calls(self, tool_calls: list[ChatCompletionDeltaToolCall]) -> None:
+        """Handle tool call updates"""
+        if not tool_calls:
+            return
+        for call in tool_calls:
+            if call.id:
+                if call.type == "function":
+                    new_tool_call = ToolCall(
+                        id=call.id,
+                        type=call.type,
+                        function=ToolCallFunction(
+                            name=call.function.name or "",
+                            arguments=call.function.arguments,
+                        ),
+                        index=call.index,
+                    )
+                    if self._current_message is not None:
+                        if self._current_message.tool_calls is None:
+                            self._current_message.tool_calls = []
+                        self._current_message.tool_calls.append(new_tool_call)
+                else:
+                    logger.warning("Unexpected tool call type: %s", call.type)
+            elif self._current_message is not None and self._current_message.tool_calls is not None and call.index is not None and 0 <= call.index < len(self._current_message.tool_calls):
+                existing_call = self._current_message.tool_calls[call.index]
+                if call.function.arguments:
+                    if existing_call.function.arguments is None:
+                        existing_call.function.arguments = ""
+                    existing_call.function.arguments += call.function.arguments
+            else:
+                logger.warning("Cannot update tool call: current_message or tool_calls is None, or invalid index.")
+
+    @property
+    def is_initialized(self) -> bool:
+        """Check if the current message is initialized"""
+        return self._current_message is not None
+
+    @property
+    def current_message(self) -> AssistantMessage:
+        """Get the current message being processed"""
+        if not self._current_message:
+            msg = "No current message initialized. Call initialize_message first."
+            raise ValueError(msg)
+        return self._current_message

lite_agent/processors/response_event_processor.py

@@ -0,0 +1,205 @@
+from collections.abc import AsyncGenerator
+from datetime import datetime, timezone
+from typing import Any
+
+from aiofiles.threadpool.text import AsyncTextIOWrapper
+from litellm.types.llms.openai import (
+    ContentPartAddedEvent,
+    FunctionCallArgumentsDeltaEvent,
+    FunctionCallArgumentsDoneEvent,
+    OutputItemAddedEvent,
+    OutputItemDoneEvent,
+    OutputTextDeltaEvent,
+    ResponseCompletedEvent,
+    ResponsesAPIStreamEvents,
+    ResponsesAPIStreamingResponse,
+)
+
+from lite_agent.types import (
+    AgentChunk,
+    AssistantMessageEvent,
+    AssistantMessageMeta,
+    ContentDeltaEvent,
+    EventUsage,
+    FunctionCallEvent,
+    NewAssistantMessage,
+    ResponseRawEvent,
+    Timing,
+    TimingEvent,
+    UsageEvent,
+)
+
+
+class ResponseEventProcessor:
+    """Processor for handling response events"""
+
+    def __init__(self) -> None:
+        self._messages: list[dict[str, Any]] = []
+        self._start_time: datetime | None = None
+        self._first_output_time: datetime | None = None
+        self._output_complete_time: datetime | None = None
+        self._usage_time: datetime | None = None
+        self._usage_data: dict[str, Any] = {}
+
+    async def process_chunk(
+        self,
+        chunk: ResponsesAPIStreamingResponse,
+        record_file: AsyncTextIOWrapper | None = None,
+    ) -> AsyncGenerator[AgentChunk, None]:
+        # Mark start time on first chunk
+        if self._start_time is None:
+            self._start_time = datetime.now(timezone.utc)
+
+        if record_file:
+            await record_file.write(chunk.model_dump_json() + "\n")
+            await record_file.flush()
+
+        yield ResponseRawEvent(raw=chunk)
+
+        events = self.handle_event(chunk)
+        for event in events:
+            yield event
+
+    def handle_event(self, event: ResponsesAPIStreamingResponse) -> list[AgentChunk]:  # noqa: PLR0911
+        """Handle individual response events"""
+        if event.type in (
+            ResponsesAPIStreamEvents.RESPONSE_CREATED,
+            ResponsesAPIStreamEvents.RESPONSE_IN_PROGRESS,
+            ResponsesAPIStreamEvents.OUTPUT_TEXT_DONE,
+            ResponsesAPIStreamEvents.CONTENT_PART_DONE,
+        ):
+            return []
+
+        if isinstance(event, OutputItemAddedEvent):
+            self._messages.append(event.item)  # type: ignore
+            return []
+
+        if isinstance(event, ContentPartAddedEvent):
+            latest_message = self._messages[-1] if self._messages else None
+            if latest_message and isinstance(latest_message.get("content"), list):
+                latest_message["content"].append(event.part)
+            return []
+
+        if isinstance(event, OutputTextDeltaEvent):
+            # Mark first output time if not already set
+            if self._first_output_time is None:
+                self._first_output_time = datetime.now(timezone.utc)
+
+            latest_message = self._messages[-1] if self._messages else None
+            if latest_message and isinstance(latest_message.get("content"), list):
+                latest_content = latest_message["content"][-1]
+                if "text" in latest_content:
+                    latest_content["text"] += event.delta
+                    return [ContentDeltaEvent(delta=event.delta)]
+            return []
+
+        if isinstance(event, OutputItemDoneEvent):
+            item = event.item
+            if item.get("type") == "function_call":
+                return [
+                    FunctionCallEvent(
+                        call_id=item["call_id"],
+                        name=item["name"],
+                        arguments=item["arguments"],
+                    ),
+                ]
+            if item.get("type") == "message":
+                # Mark output complete time when message is done
+                if self._output_complete_time is None:
+                    self._output_complete_time = datetime.now(timezone.utc)
+
+                content = item.get("content", [])
+                if content and isinstance(content, list) and len(content) > 0:
+                    end_time = datetime.now(timezone.utc)
+                    latency_ms = None
+                    output_time_ms = None
+                    # latency_ms: 从开始准备输出到 LLM 输出第一个字符的时间差
+                    if self._start_time and self._first_output_time:
+                        latency_ms = int((self._first_output_time - self._start_time).total_seconds() * 1000)
+                    # output_time_ms: 从输出第一个字符到输出完成的时间差
+                    if self._first_output_time and self._output_complete_time:
+                        output_time_ms = int((self._output_complete_time - self._first_output_time).total_seconds() * 1000)
+
+                    meta = AssistantMessageMeta(
+                        sent_at=end_time,
+                        latency_ms=latency_ms,
+                        output_time_ms=output_time_ms,
+                        input_tokens=self._usage_data.get("input_tokens"),
+                        output_tokens=self._usage_data.get("output_tokens"),
+                    )
+                    return [
+                        AssistantMessageEvent(
+                            message=NewAssistantMessage(content=[], meta=meta),
+                        ),
+                    ]
+
+        elif isinstance(event, FunctionCallArgumentsDeltaEvent):
+            if self._messages:
+                latest_message = self._messages[-1]
+                if latest_message.get("type") == "function_call":
+                    if "arguments" not in latest_message:
+                        latest_message["arguments"] = ""
+                    latest_message["arguments"] += event.delta
+            return []
+
+        elif isinstance(event, FunctionCallArgumentsDoneEvent):
+            if self._messages:
+                latest_message = self._messages[-1]
+                if latest_message.get("type") == "function_call":
+                    latest_message["arguments"] = event.arguments
+            return []
+
+        elif isinstance(event, ResponseCompletedEvent):
+            usage = event.response.usage
+            if usage:
+                # Mark usage time
+                self._usage_time = datetime.now(timezone.utc)
+                # Store usage data for meta information
+                self._usage_data["input_tokens"] = usage.input_tokens
+                self._usage_data["output_tokens"] = usage.output_tokens
+                # Also store usage time for later calculation
+                self._usage_data["usage_time"] = self._usage_time
+
+                results = []
+
+                # First yield usage event
+                results.append(
+                    UsageEvent(
+                        usage=EventUsage(
+                            input_tokens=usage.input_tokens,
+                            output_tokens=usage.output_tokens,
+                        ),
+                    ),
+                )
+
+                # Then yield timing event if we have timing data
+                if self._start_time and self._first_output_time and self._output_complete_time:
+                    latency_ms = int((self._first_output_time - self._start_time).total_seconds() * 1000)
+                    output_time_ms = int((self._output_complete_time - self._first_output_time).total_seconds() * 1000)
+
+                    results.append(
+                        TimingEvent(
+                            timing=Timing(
+                                latency_ms=latency_ms,
+                                output_time_ms=output_time_ms,
+                            ),
+                        ),
+                    )
+
+                return results
+
+        return []
+
+    @property
+    def messages(self) -> list[dict[str, Any]]:
+        """Get the accumulated messages"""
+        return self._messages
+
+    def reset(self) -> None:
+        """Reset the processor state"""
+        self._messages = []
+        self._start_time = None
+        self._first_output_time = None
+        self._output_complete_time = None
+        self._usage_time = None
+        self._usage_data = {}