Tools

工具系统提供了 LLM 工具调用请求与实际执行之间的接口。

源码位置：llm_harness.core.tools

核心类

BaseTool (ABC)

class BaseTool(ABC):
    name: ClassVar[str]
    description: ClassVar[str]
    input_model: ClassVar[type[BaseModel]]

    async def execute(self, arguments: BaseModel, context: ToolExecutionContext) -> ToolResult: ...
    def is_read_only(self, arguments: BaseModel) -> bool: ...
    def to_api_schema(self, api_format: str = "anthropic") -> dict[str, Any]: ...
    def to_openai_schema(self) -> dict[str, Any]: ...

ToolRegistry

class ToolRegistry:
    def register(self, tool: BaseTool) -> None: ...
    def unregister(self, name: str) -> None: ...
    def has(self, name: str) -> bool: ...
    def get(self, name: str) -> BaseTool | None: ...
    def list_tools(self) -> list[BaseTool]: ...
    def to_api_schema(self, api_format: str = "anthropic") -> list[dict[str, Any]]: ...

ToolExecutionContext

@dataclass
class ToolExecutionContext:
    cwd: Path                              # 工作目录
    metadata: dict[str, Any]               # session_key、account 等

ToolResult

@dataclass(frozen=True)
class ToolResult:
    output: str                            # 工具输出文本
    is_error: bool = False                 # 执行是否失败
    metadata: dict[str, Any]               # 任意元数据

ToolFactory

class ToolFactory:
    def __init__(self, *, sandbox=None, memory=None, swarm=None, bus=None, skills=None, harness_tool_names=None): ...
    def register(self, name: str, builder: Callable[[], BaseTool | None]) -> None: ...
    def build(self, name: str) -> BaseTool | None: ...

内置工具

工具	名称	依赖	只读
ReadFileTool	read_file	sandbox	是
WriteFileTool	write_file	sandbox	否
EditFileTool	edit_file	sandbox	否
ExecTool	exec	sandbox	否
GlobTool	glob	sandbox	是
GrepTool	grep	sandbox	是
WebSearchTool	web_search	无	是
WebFetchTool	web_fetch	无	是
MemoryReadTool	memory_read	memory	是
MemoryWriteTool	memory_write	memory	否
AgentTool	agent	swarm, bus	否
SendMessageTool	send_message	swarm	否
TaskStopTool	task_stop	swarm	否
SkillTool	skill	skills	是
AskUserQuestionTool	ask_user_question	无	否

实现自定义工具

from pydantic import BaseModel, Field
from llm_harness.core.tools.base import BaseTool, ToolExecutionContext, ToolResult

class GreetInput(BaseModel):
    name: str = Field(description="要问候的名称")

class GreetTool(BaseTool):
    name: ClassVar[str] = "greet"
    description: ClassVar[str] = "按名称问候某人。"
    input_model: ClassVar[type[BaseModel]] = GreetInput

    async def execute(self, args: GreetInput, ctx: ToolExecutionContext) -> ToolResult:
        return ToolResult(output=f"Hello, {args.name}!")

    @staticmethod
    def is_read_only(args: GreetInput) -> bool:
        return True