gMAS / src /tools /base.py
Артём Боярских
chore: initial commit
3193174
"""
Base classes for agent tools.
Tools are used via Native Function Calling (OpenAI/Anthropic API).
If an agent has tools, it ALWAYS uses them on every call.
"""
import json
import re
from abc import ABC, abstractmethod
from collections.abc import Callable
from typing import Any
from pydantic import BaseModel, Field
class ToolCall(BaseModel):
"""Tool invocation request."""
name: str
arguments: dict[str, Any] = Field(default_factory=dict)
@classmethod
def parse_from_response(cls, response: str) -> list["ToolCall"]:
r"""
Parse tool calls from an LLM response.
Supports two formats:
1. XML-like tags: <tool_call>{"name": "...", "arguments": {...}}</tool_call>
2. Markdown code blocks: ```tool_call\n{"name": "...", "arguments": {...}}\n```
Args:
response: Text response from the LLM
Returns:
List of ToolCall objects
"""
calls: list[ToolCall] = []
# Pattern for XML-like tags
xml_pattern = r"<tool_call>\s*(\{.*?\})\s*</tool_call>"
xml_matches = re.findall(xml_pattern, response, re.DOTALL)
for match in xml_matches:
try:
data = json.loads(match)
if isinstance(data, dict) and "name" in data:
calls.append(cls(name=data["name"], arguments=data.get("arguments", {})))
except (json.JSONDecodeError, ValueError):
# Skip invalid JSON
pass
# Pattern for markdown code blocks
code_block_pattern = r"```tool_call\s*\n(\{.*?\})\s*\n```"
code_matches = re.findall(code_block_pattern, response, re.DOTALL)
for match in code_matches:
try:
data = json.loads(match)
if isinstance(data, dict) and "name" in data:
calls.append(cls(name=data["name"], arguments=data.get("arguments", {})))
except (json.JSONDecodeError, ValueError):
# Skip invalid JSON
pass
return calls
class ToolResult(BaseModel):
"""Tool execution result."""
tool_name: str
success: bool = True
output: str = ""
error: str | None = None
def to_message(self) -> str:
"""Format the result for insertion into a prompt."""
if self.success:
return f'<tool_result name="{self.tool_name}">\n{self.output}\n</tool_result>'
return f'<tool_error name="{self.tool_name}">\n{self.error}\n</tool_error>'
class BaseTool(ABC):
"""
Abstract base class for tools.
All tools must inherit from this class and implement
the name, description, and execute methods.
"""
@property
@abstractmethod
def name(self) -> str:
"""Unique tool name."""
...
@property
@abstractmethod
def description(self) -> str:
"""Tool description for the LLM."""
...
@property
def parameters_schema(self) -> dict[str, Any]:
"""JSON Schema of the tool parameters."""
return {"type": "object", "properties": {}}
@abstractmethod
def execute(self, **kwargs: Any) -> ToolResult:
"""Execute the tool with the given arguments."""
...
def to_openai_schema(self) -> dict[str, Any]:
"""Serialize the tool to the OpenAI function calling format."""
return {
"type": "function",
"function": {
"name": self.name,
"description": self.description,
"parameters": self.parameters_schema,
},
}
class ToolRegistry:
"""
Tool registry for agents.
Example:
from tools import get_registry, CodeInterpreterTool
# Get the global registry
registry = get_registry()
registry.register(CodeInterpreterTool())
# Or create your own
my_registry = ToolRegistry()
my_registry.register(ShellTool())
"""
def __init__(self):
"""Initialize an empty registry."""
self._tools: dict[str, BaseTool] = {}
def register(self, tool: BaseTool) -> "ToolRegistry":
"""
Register a tool.
Args:
tool: Tool instance.
Returns:
self for method chaining.
"""
self._tools[tool.name] = tool
return self
def function(
self,
func: Callable | None = None,
*,
name: str | None = None,
description: str | None = None,
) -> Callable:
"""
Decorator for registering a function as a tool.
Example:
@registry.function
def my_tool(arg: str) -> str:
\"\"\"Tool description.\"\"\"
return arg.upper()
"""
def decorator(f: Callable) -> Callable:
tool_name = name or getattr(f, "__name__", "unnamed")
tool_desc = description or getattr(f, "__doc__", None) or f"Function {tool_name}"
from .function_calling import FunctionWrapper
wrapper = FunctionWrapper(
func=f,
tool_name=tool_name,
tool_description=tool_desc,
)
self._tools[tool_name] = wrapper
return f
if func is not None:
return decorator(func)
return decorator
def get(self, name: str) -> BaseTool | None:
"""Get a tool by name."""
return self._tools.get(name)
def has(self, name: str) -> bool:
"""Check whether a tool is registered."""
return name in self._tools
def execute(self, call: ToolCall) -> ToolResult:
"""Execute a tool call."""
tool = self._tools.get(call.name)
if tool is None:
return ToolResult(
tool_name=call.name,
success=False,
error=f"Tool '{call.name}' not found",
)
try:
return tool.execute(**call.arguments)
except (ValueError, KeyError, TypeError, AttributeError) as e:
return ToolResult(
tool_name=call.name,
success=False,
error=str(e),
)
def execute_all(self, calls: list[ToolCall]) -> list[ToolResult]:
"""Execute multiple tool calls."""
return [self.execute(call) for call in calls]
def list_tools(self) -> list[str]:
"""Get the list of registered tool names."""
return list(self._tools.keys())
def get_tools(self, tool_names: list[str] | None = None) -> list[BaseTool]:
"""
Get tools by name.
Args:
tool_names: List of tool names (None = all).
Returns:
List of BaseTool objects.
"""
if tool_names is None:
return list(self._tools.values())
return [self._tools[name] for name in tool_names if name in self._tools]
def to_openai_schemas(self, tool_names: list[str] | None = None) -> list[dict[str, Any]]:
"""
Get schemas in the OpenAI function calling API format.
Args:
tool_names: List of tool names (None = all).
Returns:
List of schemas in OpenAI tools API format.
"""
tools = self.get_tools(tool_names)
return [tool.to_openai_schema() for tool in tools]
def get_tools_for_agent(self, tool_names: list[str]) -> list[BaseTool]:
"""
Get tools for an agent by name.
Args:
tool_names: List of tool names.
Returns:
List of BaseTool objects (only those that exist).
"""
return [self._tools[name] for name in tool_names if name in self._tools]
def to_schemas(self, tool_names: list[str] | None = None) -> list[dict[str, Any]]:
"""
Get simplified tool schemas.
Args:
tool_names: List of tool names (None = all).
Returns:
List of schemas in simplified format.
"""
tools = self.get_tools(tool_names)
return [
{
"name": tool.name,
"description": tool.description,
"parameters": tool.parameters_schema,
}
for tool in tools
]
def format_tools_prompt(self, tool_names: list[str] | None = None) -> str:
"""
Build a text prompt describing available tools.
Args:
tool_names: List of tool names (None = all).
Returns:
String with tool descriptions for the prompt.
"""
tools = self.get_tools(tool_names)
if not tools:
return "No tools available."
lines = ["Available tools:"]
for tool in tools:
lines.append(f"\n- {tool.name}: {tool.description}")
params = tool.parameters_schema.get("properties", {})
if params:
for param_name, param_info in params.items():
param_type = param_info.get("type", "any")
param_desc = param_info.get("description", "")
lines.append(f" - {param_name} ({param_type}): {param_desc}")
lines.append("\nTo use a tool, respond with:")
lines.append("<tool_call>")
lines.append('{"name": "tool_name", "arguments": {...}}')
lines.append("</tool_call>")
return "\n".join(lines)
# Global tool registry
_global_registry: ToolRegistry | None = None
def get_registry() -> ToolRegistry:
"""
Get the global tool registry.
Creates the registry on the first call (singleton).
Example:
from tools import get_registry, ShellTool
registry = get_registry()
registry.register(ShellTool())
"""
global _global_registry # noqa: PLW0603
if _global_registry is None:
_global_registry = ToolRegistry()
return _global_registry
def register_tool(tool: BaseTool) -> BaseTool:
"""
Register a tool in the global registry.
Example:
from tools import register_tool, ShellTool
register_tool(ShellTool()) # Now available globally
"""
get_registry().register(tool)
return tool
def tool(
func: Callable | None = None,
*,
name: str | None = None,
description: str | None = None,
) -> Callable:
"""
Decorator for registering a function as a tool in the global registry.
Example:
from tools import tool
@tool
def fibonacci(n: int) -> str:
'''Calculate the n-th Fibonacci number.'''
a, b = 0, 1
for _ in range(n):
a, b = b, a + b
return str(a)
# Now 'fibonacci' is available globally via get_registry()
"""
return get_registry().function(func, name=name, description=description)
# ============================================================
# Tool factory — creating tools from a dict config
# ============================================================
# Factory registry: tool_name -> callable(config_dict) -> BaseTool
_tool_factories: dict[str, Callable[..., BaseTool]] = {}
def register_tool_factory(tool_name: str, factory: Callable[..., BaseTool]) -> None:
"""
Register a factory for creating a tool by name from a config.
Args:
tool_name: Tool name (e.g. "web_search").
factory: Function that accepts **kwargs and returns a BaseTool.
Example:
register_tool_factory("web_search", lambda **kw: WebSearchTool(**kw))
"""
_tool_factories[tool_name] = factory
def create_tool_from_config(config: dict[str, Any]) -> BaseTool | None:
"""
Create a tool from a dict config.
The config must contain a "name" key (tool name).
All other keys are passed as constructor parameters.
Args:
config: Settings dictionary, e.g.:
{"name": "web_search", "use_selenium": True}
Returns:
BaseTool or None if no factory is found.
Example:
tool = create_tool_from_config({
"name": "web_search",
"use_selenium": True,
"selenium_config": {"headless": True, "browser": "chrome"},
})
"""
name = config.get("name") or config.get("tool") or config.get("id")
if not name:
return None
factory = _tool_factories.get(name)
if factory is None:
return None
# Remove identification keys, keep only parameters
params = {k: v for k, v in config.items() if k not in ("name", "tool", "id")}
return factory(**params)