agents/tool.py

from __future__ import annotations

import inspect
import json
from collections.abc import Awaitable
from dataclasses import dataclass
from typing import TYPE_CHECKING, Any, Callable, Literal, Union, overload

from openai.types.responses.file_search_tool_param import Filters, RankingOptions
from openai.types.responses.response_computer_tool_call import (
    PendingSafetyCheck,
    ResponseComputerToolCall,
)
from openai.types.responses.response_output_item import LocalShellCall, McpApprovalRequest
from openai.types.responses.tool_param import CodeInterpreter, ImageGeneration, Mcp
from openai.types.responses.web_search_tool import Filters as WebSearchToolFilters
from openai.types.responses.web_search_tool_param import UserLocation
from pydantic import ValidationError
from typing_extensions import Concatenate, NotRequired, ParamSpec, TypedDict

from . import _debug
from .computer import AsyncComputer, Computer
from .exceptions import ModelBehaviorError
from .function_schema import DocstringStyle, function_schema
from .items import RunItem
from .logger import logger
from .run_context import RunContextWrapper
from .strict_schema import ensure_strict_json_schema
from .tool_context import ToolContext
from .tool_guardrails import ToolInputGuardrail, ToolOutputGuardrail
from .tracing import SpanError
from .util import _error_tracing
from .util._types import MaybeAwaitable

if TYPE_CHECKING:
    from .agent import Agent, AgentBase

ToolParams = ParamSpec("ToolParams")

ToolFunctionWithoutContext = Callable[ToolParams, Any]
ToolFunctionWithContext = Callable[Concatenate[RunContextWrapper[Any], ToolParams], Any]
ToolFunctionWithToolContext = Callable[Concatenate[ToolContext, ToolParams], Any]

ToolFunction = Union[
    ToolFunctionWithoutContext[ToolParams],
    ToolFunctionWithContext[ToolParams],
    ToolFunctionWithToolContext[ToolParams],
]


@dataclass
class FunctionToolResult:
    tool: FunctionTool
    """The tool that was run."""

    output: Any
    """The output of the tool."""

    run_item: RunItem
    """The run item that was produced as a result of the tool call."""


@dataclass
class FunctionTool:
    """A tool that wraps a function. In most cases, you should use  the `function_tool` helpers to
    create a FunctionTool, as they let you easily wrap a Python function.
    """

    name: str
    """The name of the tool, as shown to the LLM. Generally the name of the function."""

    description: str
    """A description of the tool, as shown to the LLM."""

    params_json_schema: dict[str, Any]
    """The JSON schema for the tool's parameters."""

    on_invoke_tool: Callable[[ToolContext[Any], str], Awaitable[Any]]
    """A function that invokes the tool with the given context and parameters. The params passed
    are:
    1. The tool run context.
    2. The arguments from the LLM, as a JSON string.

    You must return a string representation of the tool output, or something we can call `str()` on.
    In case of errors, you can either raise an Exception (which will cause the run to fail) or
    return a string error message (which will be sent back to the LLM).
    """

    strict_json_schema: bool = True
    """Whether the JSON schema is in strict mode. We **strongly** recommend setting this to True,
    as it increases the likelihood of correct JSON input."""

    is_enabled: bool | Callable[[RunContextWrapper[Any], AgentBase], MaybeAwaitable[bool]] = True
    """Whether the tool is enabled. Either a bool or a Callable that takes the run context and agent
    and returns whether the tool is enabled. You can use this to dynamically enable/disable a tool
    based on your context/state."""

    # Tool-specific guardrails
    tool_input_guardrails: list[ToolInputGuardrail[Any]] | None = None
    """Optional list of input guardrails to run before invoking this tool."""

    tool_output_guardrails: list[ToolOutputGuardrail[Any]] | None = None
    """Optional list of output guardrails to run after invoking this tool."""

    def __post_init__(self):
        if self.strict_json_schema:
            self.params_json_schema = ensure_strict_json_schema(self.params_json_schema)


@dataclass
class FileSearchTool:
    """A hosted tool that lets the LLM search through a vector store. Currently only supported with
    OpenAI models, using the Responses API.
    """

    vector_store_ids: list[str]
    """The IDs of the vector stores to search."""

    max_num_results: int | None = None
    """The maximum number of results to return."""

    include_search_results: bool = False
    """Whether to include the search results in the output produced by the LLM."""

    ranking_options: RankingOptions | None = None
    """Ranking options for search."""

    filters: Filters | None = None
    """A filter to apply based on file attributes."""

    @property
    def name(self):
        return "file_search"


@dataclass
class WebSearchTool:
    """A hosted tool that lets the LLM search the web. Currently only supported with OpenAI models,
    using the Responses API.
    """

    user_location: UserLocation | None = None
    """Optional location for the search. Lets you customize results to be relevant to a location."""

    filters: WebSearchToolFilters | None = None
    """A filter to apply based on file attributes."""

    search_context_size: Literal["low", "medium", "high"] = "medium"
    """The amount of context to use for the search."""

    @property
    def name(self):
        return "web_search"


@dataclass
class ComputerTool:
    """A hosted tool that lets the LLM control a computer."""

    computer: Computer | AsyncComputer
    """The computer implementation, which describes the environment and dimensions of the computer,
    as well as implements the computer actions like click, screenshot, etc.
    """

    on_safety_check: Callable[[ComputerToolSafetyCheckData], MaybeAwaitable[bool]] | None = None
    """Optional callback to acknowledge computer tool safety checks."""

    @property
    def name(self):
        return "computer_use_preview"


@dataclass
class ComputerToolSafetyCheckData:
    """Information about a computer tool safety check."""

    ctx_wrapper: RunContextWrapper[Any]
    """The run context."""

    agent: Agent[Any]
    """The agent performing the computer action."""

    tool_call: ResponseComputerToolCall
    """The computer tool call."""

    safety_check: PendingSafetyCheck
    """The pending safety check to acknowledge."""


@dataclass
class MCPToolApprovalRequest:
    """A request to approve a tool call."""

    ctx_wrapper: RunContextWrapper[Any]
    """The run context."""

    data: McpApprovalRequest
    """The data from the MCP tool approval request."""


class MCPToolApprovalFunctionResult(TypedDict):
    """The result of an MCP tool approval function."""

    approve: bool
    """Whether to approve the tool call."""

    reason: NotRequired[str]
    """An optional reason, if rejected."""


MCPToolApprovalFunction = Callable[
    [MCPToolApprovalRequest], MaybeAwaitable[MCPToolApprovalFunctionResult]
]
"""A function that approves or rejects a tool call."""


@dataclass
class HostedMCPTool:
    """A tool that allows the LLM to use a remote MCP server. The LLM will automatically list and
    call tools, without requiring a round trip back to your code.
    If you want to run MCP servers locally via stdio, in a VPC or other non-publicly-accessible
    environment, or you just prefer to run tool calls locally, then you can instead use the servers
    in `agents.mcp` and pass `Agent(mcp_servers=[...])` to the agent."""

    tool_config: Mcp
    """The MCP tool config, which includes the server URL and other settings."""

    on_approval_request: MCPToolApprovalFunction | None = None
    """An optional function that will be called if approval is requested for an MCP tool. If not
    provided, you will need to manually add approvals/rejections to the input and call
    `Runner.run(...)` again."""

    @property
    def name(self):
        return "hosted_mcp"


@dataclass
class CodeInterpreterTool:
    """A tool that allows the LLM to execute code in a sandboxed environment."""

    tool_config: CodeInterpreter
    """The tool config, which includes the container and other settings."""

    @property
    def name(self):
        return "code_interpreter"


@dataclass
class ImageGenerationTool:
    """A tool that allows the LLM to generate images."""

    tool_config: ImageGeneration
    """The tool config, which image generation settings."""

    @property
    def name(self):
        return "image_generation"


@dataclass
class LocalShellCommandRequest:
    """A request to execute a command on a shell."""

    ctx_wrapper: RunContextWrapper[Any]
    """The run context."""

    data: LocalShellCall
    """The data from the local shell tool call."""


LocalShellExecutor = Callable[[LocalShellCommandRequest], MaybeAwaitable[str]]
"""A function that executes a command on a shell."""


@dataclass
class LocalShellTool:
    """A tool that allows the LLM to execute commands on a shell.

    For more details, see:
    https://platform.openai.com/docs/guides/tools-local-shell
    """

    executor: LocalShellExecutor
    """A function that executes a command on a shell."""

    @property
    def name(self):
        return "local_shell"


Tool = Union[
    FunctionTool,
    FileSearchTool,
    WebSearchTool,
    ComputerTool,
    HostedMCPTool,
    LocalShellTool,
    ImageGenerationTool,
    CodeInterpreterTool,
]
"""A tool that can be used in an agent."""


def default_tool_error_function(ctx: RunContextWrapper[Any], error: Exception) -> str:
    """The default tool error function, which just returns a generic error message."""
    return f"An error occurred while running the tool. Please try again. Error: {str(error)}"


ToolErrorFunction = Callable[[RunContextWrapper[Any], Exception], MaybeAwaitable[str]]


@overload
def function_tool(
    func: ToolFunction[...],
    *,
    name_override: str | None = None,
    description_override: str | None = None,
    docstring_style: DocstringStyle | None = None,
    use_docstring_info: bool = True,
    failure_error_function: ToolErrorFunction | None = None,
    strict_mode: bool = True,
    is_enabled: bool | Callable[[RunContextWrapper[Any], AgentBase], MaybeAwaitable[bool]] = True,
) -> FunctionTool:
    """Overload for usage as @function_tool (no parentheses)."""
    ...


@overload
def function_tool(
    *,
    name_override: str | None = None,
    description_override: str | None = None,
    docstring_style: DocstringStyle | None = None,
    use_docstring_info: bool = True,
    failure_error_function: ToolErrorFunction | None = None,
    strict_mode: bool = True,
    is_enabled: bool | Callable[[RunContextWrapper[Any], AgentBase], MaybeAwaitable[bool]] = True,
) -> Callable[[ToolFunction[...]], FunctionTool]:
    """Overload for usage as @function_tool(...)."""
    ...


def function_tool(
    func: ToolFunction[...] | None = None,
    *,
    name_override: str | None = None,
    description_override: str | None = None,
    docstring_style: DocstringStyle | None = None,
    use_docstring_info: bool = True,
    failure_error_function: ToolErrorFunction | None = default_tool_error_function,
    strict_mode: bool = True,
    is_enabled: bool | Callable[[RunContextWrapper[Any], AgentBase], MaybeAwaitable[bool]] = True,
) -> FunctionTool | Callable[[ToolFunction[...]], FunctionTool]:
    """
    Decorator to create a FunctionTool from a function. By default, we will:
    1. Parse the function signature to create a JSON schema for the tool's parameters.
    2. Use the function's docstring to populate the tool's description.
    3. Use the function's docstring to populate argument descriptions.
    The docstring style is detected automatically, but you can override it.

    If the function takes a `RunContextWrapper` as the first argument, it *must* match the
    context type of the agent that uses the tool.

    Args:
        func: The function to wrap.
        name_override: If provided, use this name for the tool instead of the function's name.
        description_override: If provided, use this description for the tool instead of the
            function's docstring.
        docstring_style: If provided, use this style for the tool's docstring. If not provided,
            we will attempt to auto-detect the style.
        use_docstring_info: If True, use the function's docstring to populate the tool's
            description and argument descriptions.
        failure_error_function: If provided, use this function to generate an error message when
            the tool call fails. The error message is sent to the LLM. If you pass None, then no
            error message will be sent and instead an Exception will be raised.
        strict_mode: Whether to enable strict mode for the tool's JSON schema. We *strongly*
            recommend setting this to True, as it increases the likelihood of correct JSON input.
            If False, it allows non-strict JSON schemas. For example, if a parameter has a default
            value, it will be optional, additional properties are allowed, etc. See here for more:
            https://platform.openai.com/docs/guides/structured-outputs?api-mode=responses#supported-schemas
        is_enabled: Whether the tool is enabled. Can be a bool or a callable that takes the run
            context and agent and returns whether the tool is enabled. Disabled tools are hidden
            from the LLM at runtime.
    """

    def _create_function_tool(the_func: ToolFunction[...]) -> FunctionTool:
        schema = function_schema(
            func=the_func,
            name_override=name_override,
            description_override=description_override,
            docstring_style=docstring_style,
            use_docstring_info=use_docstring_info,
            strict_json_schema=strict_mode,
        )

        async def _on_invoke_tool_impl(ctx: ToolContext[Any], input: str) -> Any:
            try:
                json_data: dict[str, Any] = json.loads(input) if input else {}
            except Exception as e:
                if _debug.DONT_LOG_TOOL_DATA:
                    logger.debug(f"Invalid JSON input for tool {schema.name}")
                else:
                    logger.debug(f"Invalid JSON input for tool {schema.name}: {input}")
                raise ModelBehaviorError(
                    f"Invalid JSON input for tool {schema.name}: {input}"
                ) from e

            if _debug.DONT_LOG_TOOL_DATA:
                logger.debug(f"Invoking tool {schema.name}")
            else:
                logger.debug(f"Invoking tool {schema.name} with input {input}")

            try:
                parsed = (
                    schema.params_pydantic_model(**json_data)
                    if json_data
                    else schema.params_pydantic_model()
                )
            except ValidationError as e:
                raise ModelBehaviorError(f"Invalid JSON input for tool {schema.name}: {e}") from e

            args, kwargs_dict = schema.to_call_args(parsed)

            if not _debug.DONT_LOG_TOOL_DATA:
                logger.debug(f"Tool call args: {args}, kwargs: {kwargs_dict}")

            if inspect.iscoroutinefunction(the_func):
                if schema.takes_context:
                    result = await the_func(ctx, *args, **kwargs_dict)
                else:
                    result = await the_func(*args, **kwargs_dict)
            else:
                if schema.takes_context:
                    result = the_func(ctx, *args, **kwargs_dict)
                else:
                    result = the_func(*args, **kwargs_dict)

            if _debug.DONT_LOG_TOOL_DATA:
                logger.debug(f"Tool {schema.name} completed.")
            else:
                logger.debug(f"Tool {schema.name} returned {result}")

            return result

        async def _on_invoke_tool(ctx: ToolContext[Any], input: str) -> Any:
            try:
                return await _on_invoke_tool_impl(ctx, input)
            except Exception as e:
                if failure_error_function is None:
                    raise

                result = failure_error_function(ctx, e)
                if inspect.isawaitable(result):
                    return await result

                _error_tracing.attach_error_to_current_span(
                    SpanError(
                        message="Error running tool (non-fatal)",
                        data={
                            "tool_name": schema.name,
                            "error": str(e),
                        },
                    )
                )
                return result

        return FunctionTool(
            name=schema.name,
            description=schema.description or "",
            params_json_schema=schema.params_json_schema,
            on_invoke_tool=_on_invoke_tool,
            strict_json_schema=strict_mode,
            is_enabled=is_enabled,
        )

    # If func is actually a callable, we were used as @function_tool with no parentheses
    if callable(func):
        return _create_function_tool(func)

    # Otherwise, we were used as @function_tool(...), so return a decorator
    def decorator(real_func: ToolFunction[...]) -> FunctionTool:
        return _create_function_tool(real_func)

    return decorator