AstrBot 3.5.6__py3-none-any.whl → 4.7.0__py3-none-any.whl

astrbot/core/agent/runners/tool_loop_agent_runner.py

@@ -0,0 +1,400 @@
+import sys
+import traceback
+import typing as T
+
+from mcp.types import (
+    BlobResourceContents,
+    CallToolResult,
+    EmbeddedResource,
+    ImageContent,
+    TextContent,
+    TextResourceContents,
+)
+
+from astrbot import logger
+from astrbot.core.message.message_event_result import (
+    MessageChain,
+)
+from astrbot.core.provider.entities import (
+    LLMResponse,
+    ProviderRequest,
+    ToolCallsResult,
+)
+from astrbot.core.provider.provider import Provider
+
+from ..hooks import BaseAgentRunHooks
+from ..message import AssistantMessageSegment, Message, ToolCallMessageSegment
+from ..response import AgentResponseData
+from ..run_context import ContextWrapper, TContext
+from ..tool_executor import BaseFunctionToolExecutor
+from .base import AgentResponse, AgentState, BaseAgentRunner
+
+if sys.version_info >= (3, 12):
+    from typing import override
+else:
+    from typing_extensions import override
+
+
+class ToolLoopAgentRunner(BaseAgentRunner[TContext]):
+    @override
+    async def reset(
+        self,
+        provider: Provider,
+        request: ProviderRequest,
+        run_context: ContextWrapper[TContext],
+        tool_executor: BaseFunctionToolExecutor[TContext],
+        agent_hooks: BaseAgentRunHooks[TContext],
+        **kwargs: T.Any,
+    ) -> None:
+        self.req = request
+        self.streaming = kwargs.get("streaming", False)
+        self.provider = provider
+        self.final_llm_resp = None
+        self._state = AgentState.IDLE
+        self.tool_executor = tool_executor
+        self.agent_hooks = agent_hooks
+        self.run_context = run_context
+
+        messages = []
+        # append existing messages in the run context
+        for msg in request.contexts:
+            messages.append(Message.model_validate(msg))
+        if request.prompt is not None:
+            m = await request.assemble_context()
+            messages.append(Message.model_validate(m))
+        if request.system_prompt:
+            messages.insert(
+                0,
+                Message(role="system", content=request.system_prompt),
+            )
+        self.run_context.messages = messages
+
+    async def _iter_llm_responses(self) -> T.AsyncGenerator[LLMResponse, None]:
+        """Yields chunks *and* a final LLMResponse."""
+        if self.streaming:
+            stream = self.provider.text_chat_stream(**self.req.__dict__)
+            async for resp in stream:  # type: ignore
+                yield resp
+        else:
+            yield await self.provider.text_chat(**self.req.__dict__)
+
+    @override
+    async def step(self):
+        """Process a single step of the agent.
+        This method should return the result of the step.
+        """
+        if not self.req:
+            raise ValueError("Request is not set. Please call reset() first.")
+
+        if self._state == AgentState.IDLE:
+            try:
+                await self.agent_hooks.on_agent_begin(self.run_context)
+            except Exception as e:
+                logger.error(f"Error in on_agent_begin hook: {e}", exc_info=True)
+
+        # 开始处理，转换到运行状态
+        self._transition_state(AgentState.RUNNING)
+        llm_resp_result = None
+
+        async for llm_response in self._iter_llm_responses():
+            assert isinstance(llm_response, LLMResponse)
+            if llm_response.is_chunk:
+                if llm_response.result_chain:
+                    yield AgentResponse(
+                        type="streaming_delta",
+                        data=AgentResponseData(chain=llm_response.result_chain),
+                    )
+                elif llm_response.completion_text:
+                    yield AgentResponse(
+                        type="streaming_delta",
+                        data=AgentResponseData(
+                            chain=MessageChain().message(llm_response.completion_text),
+                        ),
+                    )
+                elif llm_response.reasoning_content:
+                    yield AgentResponse(
+                        type="streaming_delta",
+                        data=AgentResponseData(
+                            chain=MessageChain(type="reasoning").message(
+                                llm_response.reasoning_content,
+                            ),
+                        ),
+                    )
+                continue
+            llm_resp_result = llm_response
+            break  # got final response
+
+        if not llm_resp_result:
+            return
+
+        # 处理 LLM 响应
+        llm_resp = llm_resp_result
+
+        if llm_resp.role == "err":
+            # 如果 LLM 响应错误，转换到错误状态
+            self.final_llm_resp = llm_resp
+            self._transition_state(AgentState.ERROR)
+            yield AgentResponse(
+                type="err",
+                data=AgentResponseData(
+                    chain=MessageChain().message(
+                        f"LLM 响应错误: {llm_resp.completion_text or '未知错误'}",
+                    ),
+                ),
+            )
+
+        if not llm_resp.tools_call_name:
+            # 如果没有工具调用，转换到完成状态
+            self.final_llm_resp = llm_resp
+            self._transition_state(AgentState.DONE)
+            # record the final assistant message
+            self.run_context.messages.append(
+                Message(
+                    role="assistant",
+                    content=llm_resp.completion_text or "",
+                ),
+            )
+            try:
+                await self.agent_hooks.on_agent_done(self.run_context, llm_resp)
+            except Exception as e:
+                logger.error(f"Error in on_agent_done hook: {e}", exc_info=True)
+
+        # 返回 LLM 结果
+        if llm_resp.result_chain:
+            yield AgentResponse(
+                type="llm_result",
+                data=AgentResponseData(chain=llm_resp.result_chain),
+            )
+        elif llm_resp.completion_text:
+            yield AgentResponse(
+                type="llm_result",
+                data=AgentResponseData(
+                    chain=MessageChain().message(llm_resp.completion_text),
+                ),
+            )
+
+        # 如果有工具调用，还需处理工具调用
+        if llm_resp.tools_call_name:
+            tool_call_result_blocks = []
+            for tool_call_name in llm_resp.tools_call_name:
+                yield AgentResponse(
+                    type="tool_call",
+                    data=AgentResponseData(
+                        chain=MessageChain(type="tool_call").message(
+                            f"🔨 调用工具: {tool_call_name}"
+                        ),
+                    ),
+                )
+            async for result in self._handle_function_tools(self.req, llm_resp):
+                if isinstance(result, list):
+                    tool_call_result_blocks = result
+                elif isinstance(result, MessageChain):
+                    result.type = "tool_call_result"
+                    yield AgentResponse(
+                        type="tool_call_result",
+                        data=AgentResponseData(chain=result),
+                    )
+            # 将结果添加到上下文中
+            tool_calls_result = ToolCallsResult(
+                tool_calls_info=AssistantMessageSegment(
+                    tool_calls=llm_resp.to_openai_to_calls_model(),
+                    content=llm_resp.completion_text,
+                ),
+                tool_calls_result=tool_call_result_blocks,
+            )
+            # record the assistant message with tool calls
+            self.run_context.messages.extend(
+                tool_calls_result.to_openai_messages_model()
+            )
+
+            self.req.append_tool_calls_result(tool_calls_result)
+
+    async def step_until_done(
+        self, max_step: int
+    ) -> T.AsyncGenerator[AgentResponse, None]:
+        """Process steps until the agent is done."""
+        step_count = 0
+        while not self.done() and step_count < max_step:
+            step_count += 1
+            async for resp in self.step():
+                yield resp
+
+    async def _handle_function_tools(
+        self,
+        req: ProviderRequest,
+        llm_response: LLMResponse,
+    ) -> T.AsyncGenerator[MessageChain | list[ToolCallMessageSegment], None]:
+        """处理函数工具调用。"""
+        tool_call_result_blocks: list[ToolCallMessageSegment] = []
+        logger.info(f"Agent 使用工具: {llm_response.tools_call_name}")
+
+        # 执行函数调用
+        for func_tool_name, func_tool_args, func_tool_id in zip(
+            llm_response.tools_call_name,
+            llm_response.tools_call_args,
+            llm_response.tools_call_ids,
+        ):
+            try:
+                if not req.func_tool:
+                    return
+                func_tool = req.func_tool.get_func(func_tool_name)
+                logger.info(f"使用工具：{func_tool_name}，参数：{func_tool_args}")
+
+                if not func_tool:
+                    logger.warning(f"未找到指定的工具: {func_tool_name}，将跳过。")
+                    tool_call_result_blocks.append(
+                        ToolCallMessageSegment(
+                            role="tool",
+                            tool_call_id=func_tool_id,
+                            content=f"error: 未找到工具 {func_tool_name}",
+                        ),
+                    )
+                    continue
+
+                valid_params = {}  # 参数过滤：只传递函数实际需要的参数
+
+                # 获取实际的 handler 函数
+                if func_tool.handler:
+                    logger.debug(
+                        f"工具 {func_tool_name} 期望的参数: {func_tool.parameters}",
+                    )
+                    if func_tool.parameters and func_tool.parameters.get("properties"):
+                        expected_params = set(func_tool.parameters["properties"].keys())
+
+                        valid_params = {
+                            k: v
+                            for k, v in func_tool_args.items()
+                            if k in expected_params
+                        }
+
+                    # 记录被忽略的参数
+                    ignored_params = set(func_tool_args.keys()) - set(
+                        valid_params.keys(),
+                    )
+                    if ignored_params:
+                        logger.warning(
+                            f"工具 {func_tool_name} 忽略非期望参数: {ignored_params}",
+                        )
+                else:
+                    # 如果没有 handler（如 MCP 工具），使用所有参数
+                    valid_params = func_tool_args
+
+                try:
+                    await self.agent_hooks.on_tool_start(
+                        self.run_context,
+                        func_tool,
+                        valid_params,
+                    )
+                except Exception as e:
+                    logger.error(f"Error in on_tool_start hook: {e}", exc_info=True)
+
+                executor = self.tool_executor.execute(
+                    tool=func_tool,
+                    run_context=self.run_context,
+                    **valid_params,  # 只传递有效的参数
+                )
+
+                _final_resp: CallToolResult | None = None
+                async for resp in executor:  # type: ignore
+                    if isinstance(resp, CallToolResult):
+                        res = resp
+                        _final_resp = resp
+                        if isinstance(res.content[0], TextContent):
+                            tool_call_result_blocks.append(
+                                ToolCallMessageSegment(
+                                    role="tool",
+                                    tool_call_id=func_tool_id,
+                                    content=res.content[0].text,
+                                ),
+                            )
+                            yield MessageChain().message(res.content[0].text)
+                        elif isinstance(res.content[0], ImageContent):
+                            tool_call_result_blocks.append(
+                                ToolCallMessageSegment(
+                                    role="tool",
+                                    tool_call_id=func_tool_id,
+                                    content="返回了图片(已直接发送给用户)",
+                                ),
+                            )
+                            yield MessageChain(type="tool_direct_result").base64_image(
+                                res.content[0].data,
+                            )
+                        elif isinstance(res.content[0], EmbeddedResource):
+                            resource = res.content[0].resource
+                            if isinstance(resource, TextResourceContents):
+                                tool_call_result_blocks.append(
+                                    ToolCallMessageSegment(
+                                        role="tool",
+                                        tool_call_id=func_tool_id,
+                                        content=resource.text,
+                                    ),
+                                )
+                                yield MessageChain().message(resource.text)
+                            elif (
+                                isinstance(resource, BlobResourceContents)
+                                and resource.mimeType
+                                and resource.mimeType.startswith("image/")
+                            ):
+                                tool_call_result_blocks.append(
+                                    ToolCallMessageSegment(
+                                        role="tool",
+                                        tool_call_id=func_tool_id,
+                                        content="返回了图片(已直接发送给用户)",
+                                    ),
+                                )
+                                yield MessageChain(
+                                    type="tool_direct_result",
+                                ).base64_image(resource.blob)
+                            else:
+                                tool_call_result_blocks.append(
+                                    ToolCallMessageSegment(
+                                        role="tool",
+                                        tool_call_id=func_tool_id,
+                                        content="返回的数据类型不受支持",
+                                    ),
+                                )
+                                yield MessageChain().message("返回的数据类型不受支持。")
+
+                    elif resp is None:
+                        # Tool 直接请求发送消息给用户
+                        # 这里我们将直接结束 Agent Loop。
+                        # 发送消息逻辑在 ToolExecutor 中处理了。
+                        logger.warning(
+                            f"{func_tool_name} 没有没有返回值或者将结果直接发送给用户，此工具调用不会被记录到历史中。"
+                        )
+                        self._transition_state(AgentState.DONE)
+                    else:
+                        # 不应该出现其他类型
+                        logger.warning(
+                            f"Tool 返回了不支持的类型: {type(resp)}，将忽略。",
+                        )
+
+                try:
+                    await self.agent_hooks.on_tool_end(
+                        self.run_context,
+                        func_tool,
+                        func_tool_args,
+                        _final_resp,
+                    )
+                except Exception as e:
+                    logger.error(f"Error in on_tool_end hook: {e}", exc_info=True)
+            except Exception as e:
+                logger.warning(traceback.format_exc())
+                tool_call_result_blocks.append(
+                    ToolCallMessageSegment(
+                        role="tool",
+                        tool_call_id=func_tool_id,
+                        content=f"error: {e!s}",
+                    ),
+                )
+
+        # 处理函数调用响应
+        if tool_call_result_blocks:
+            yield tool_call_result_blocks
+
+    def done(self) -> bool:
+        """检查 Agent 是否已完成工作"""
+        return self._state in (AgentState.DONE, AgentState.ERROR)
+
+    def get_final_llm_resp(self) -> LLMResponse | None:
+        return self.final_llm_resp

astrbot/core/agent/tool.py

@@ -0,0 +1,285 @@
+from collections.abc import Awaitable, Callable
+from typing import Any, Generic
+
+import jsonschema
+import mcp
+from deprecated import deprecated
+from pydantic import Field, model_validator
+from pydantic.dataclasses import dataclass
+
+from .run_context import ContextWrapper, TContext
+
+ParametersType = dict[str, Any]
+ToolExecResult = str | mcp.types.CallToolResult
+
+
+@dataclass
+class ToolSchema:
+    """A class representing the schema of a tool for function calling."""
+
+    name: str
+    """The name of the tool."""
+
+    description: str
+    """The description of the tool."""
+
+    parameters: ParametersType
+    """The parameters of the tool, in JSON Schema format."""
+
+    @model_validator(mode="after")
+    def validate_parameters(self) -> "ToolSchema":
+        jsonschema.validate(
+            self.parameters, jsonschema.Draft202012Validator.META_SCHEMA
+        )
+        return self
+
+
+@dataclass
+class FunctionTool(ToolSchema, Generic[TContext]):
+    """A callable tool, for function calling."""
+
+    handler: Callable[..., Awaitable[Any]] | None = None
+    """a callable that implements the tool's functionality. It should be an async function."""
+
+    handler_module_path: str | None = None
+    """
+    The module path of the handler function. This is empty when the origin is mcp.
+    This field must be retained, as the handler will be wrapped in functools.partial during initialization,
+    causing the handler's __module__ to be functools
+    """
+    active: bool = True
+    """
+    Whether the tool is active. This field is a special field for AstrBot.
+    You can ignore it when integrating with other frameworks.
+    """
+
+    def __repr__(self):
+        return f"FuncTool(name={self.name}, parameters={self.parameters}, description={self.description})"
+
+    async def call(self, context: ContextWrapper[TContext], **kwargs) -> ToolExecResult:
+        """Run the tool with the given arguments. The handler field has priority."""
+        raise NotImplementedError(
+            "FunctionTool.call() must be implemented by subclasses or set a handler."
+        )
+
+
+@dataclass
+class ToolSet:
+    """A set of function tools that can be used in function calling.
+
+    This class provides methods to add, remove, and retrieve tools, as well as
+    convert the tools to different API formats (OpenAI, Anthropic, Google GenAI).
+    """
+
+    tools: list[FunctionTool] = Field(default_factory=list)
+
+    def empty(self) -> bool:
+        """Check if the tool set is empty."""
+        return len(self.tools) == 0
+
+    def add_tool(self, tool: FunctionTool):
+        """Add a tool to the set."""
+        # 检查是否已存在同名工具
+        for i, existing_tool in enumerate(self.tools):
+            if existing_tool.name == tool.name:
+                self.tools[i] = tool
+                return
+        self.tools.append(tool)
+
+    def remove_tool(self, name: str):
+        """Remove a tool by its name."""
+        self.tools = [tool for tool in self.tools if tool.name != name]
+
+    def get_tool(self, name: str) -> FunctionTool | None:
+        """Get a tool by its name."""
+        for tool in self.tools:
+            if tool.name == name:
+                return tool
+        return None
+
+    @deprecated(reason="Use add_tool() instead", version="4.0.0")
+    def add_func(
+        self,
+        name: str,
+        func_args: list,
+        desc: str,
+        handler: Callable[..., Awaitable[Any]],
+    ):
+        """Add a function tool to the set."""
+        params = {
+            "type": "object",  # hard-coded here
+            "properties": {},
+        }
+        for param in func_args:
+            params["properties"][param["name"]] = {
+                "type": param["type"],
+                "description": param["description"],
+            }
+        _func = FunctionTool(
+            name=name,
+            parameters=params,
+            description=desc,
+            handler=handler,
+        )
+        self.add_tool(_func)
+
+    @deprecated(reason="Use remove_tool() instead", version="4.0.0")
+    def remove_func(self, name: str):
+        """Remove a function tool by its name."""
+        self.remove_tool(name)
+
+    @deprecated(reason="Use get_tool() instead", version="4.0.0")
+    def get_func(self, name: str) -> FunctionTool | None:
+        """Get all function tools."""
+        return self.get_tool(name)
+
+    @property
+    def func_list(self) -> list[FunctionTool]:
+        """Get the list of function tools."""
+        return self.tools
+
+    def openai_schema(self, omit_empty_parameter_field: bool = False) -> list[dict]:
+        """Convert tools to OpenAI API function calling schema format."""
+        result = []
+        for tool in self.tools:
+            func_def = {
+                "type": "function",
+                "function": {
+                    "name": tool.name,
+                    "description": tool.description,
+                },
+            }
+
+            if (
+                tool.parameters and tool.parameters.get("properties")
+            ) or not omit_empty_parameter_field:
+                func_def["function"]["parameters"] = tool.parameters
+
+            result.append(func_def)
+        return result
+
+    def anthropic_schema(self) -> list[dict]:
+        """Convert tools to Anthropic API format."""
+        result = []
+        for tool in self.tools:
+            input_schema = {"type": "object"}
+            if tool.parameters:
+                input_schema["properties"] = tool.parameters.get("properties", {})
+                input_schema["required"] = tool.parameters.get("required", [])
+            tool_def = {
+                "name": tool.name,
+                "description": tool.description,
+                "input_schema": input_schema,
+            }
+            result.append(tool_def)
+        return result
+
+    def google_schema(self) -> dict:
+        """Convert tools to Google GenAI API format."""
+
+        def convert_schema(schema: dict) -> dict:
+            """Convert schema to Gemini API format."""
+            supported_types = {
+                "string",
+                "number",
+                "integer",
+                "boolean",
+                "array",
+                "object",
+                "null",
+            }
+            supported_formats = {
+                "string": {"enum", "date-time"},
+                "integer": {"int32", "int64"},
+                "number": {"float", "double"},
+            }
+
+            if "anyOf" in schema:
+                return {"anyOf": [convert_schema(s) for s in schema["anyOf"]]}
+
+            result = {}
+
+            if "type" in schema and schema["type"] in supported_types:
+                result["type"] = schema["type"]
+                if "format" in schema and schema["format"] in supported_formats.get(
+                    result["type"],
+                    set(),
+                ):
+                    result["format"] = schema["format"]
+            else:
+                result["type"] = "null"
+
+            support_fields = {
+                "title",
+                "description",
+                "enum",
+                "minimum",
+                "maximum",
+                "maxItems",
+                "minItems",
+                "nullable",
+                "required",
+            }
+            result.update({k: schema[k] for k in support_fields if k in schema})
+
+            if "properties" in schema:
+                properties = {}
+                for key, value in schema["properties"].items():
+                    prop_value = convert_schema(value)
+                    if "default" in prop_value:
+                        del prop_value["default"]
+                    properties[key] = prop_value
+
+                if properties:
+                    result["properties"] = properties
+
+            if "items" in schema:
+                result["items"] = convert_schema(schema["items"])
+
+            return result
+
+        tools = []
+        for tool in self.tools:
+            d: dict[str, Any] = {
+                "name": tool.name,
+                "description": tool.description,
+            }
+            if tool.parameters:
+                d["parameters"] = convert_schema(tool.parameters)
+            tools.append(d)
+
+        declarations = {}
+        if tools:
+            declarations["function_declarations"] = tools
+        return declarations
+
+    @deprecated(reason="Use openai_schema() instead", version="4.0.0")
+    def get_func_desc_openai_style(self, omit_empty_parameter_field: bool = False):
+        return self.openai_schema(omit_empty_parameter_field)
+
+    @deprecated(reason="Use anthropic_schema() instead", version="4.0.0")
+    def get_func_desc_anthropic_style(self):
+        return self.anthropic_schema()
+
+    @deprecated(reason="Use google_schema() instead", version="4.0.0")
+    def get_func_desc_google_genai_style(self):
+        return self.google_schema()
+
+    def names(self) -> list[str]:
+        """获取所有工具的名称列表"""
+        return [tool.name for tool in self.tools]
+
+    def __len__(self):
+        return len(self.tools)
+
+    def __bool__(self):
+        return len(self.tools) > 0
+
+    def __iter__(self):
+        return iter(self.tools)
+
+    def __repr__(self):
+        return f"ToolSet(tools={self.tools})"
+
+    def __str__(self):
+        return f"ToolSet(tools={self.tools})"

astrbot/core/agent/tool_executor.py

@@ -0,0 +1,17 @@
+from collections.abc import AsyncGenerator
+from typing import Any, Generic
+
+import mcp
+
+from .run_context import ContextWrapper, TContext
+from .tool import FunctionTool
+
+
+class BaseFunctionToolExecutor(Generic[TContext]):
+    @classmethod
+    async def execute(
+        cls,
+        tool: FunctionTool,
+        run_context: ContextWrapper[TContext],
+        **tool_args,
+    ) -> AsyncGenerator[Any | mcp.types.CallToolResult, None]: ...