trading_agent.agent.loop_guard

Stops a model from calling the same tool with the same arguments over and over in one turn.

Small models sometimes get stuck repeating one call. For a trading agent that could mean sending the same order again and again, so the guard refuses a call once it has been made too many times in a turn. The risk engine (milestone 9) also limits orders per minute; the guard catches the loop earlier and for every tool.

Typical usage example:

guard = LoopGuard(maximum_identical_calls=3) guard.start_turn() if guard.allows(call): ...

LoopGuard

LoopGuard(maximum_identical_calls)

Counts identical tool calls within the current turn.

Creates the guard.

Parameters:

Name	Type	Description	Default
maximum_identical_calls	int	How many times one tool may be called with the same arguments in a turn.	required

Source code in src/trading_agent/agent/loop_guard.py

def __init__(self, maximum_identical_calls: int):
    """Creates the guard.

    Args:
        maximum_identical_calls (int): How many times one tool may be called with the same arguments in a turn.
    """
    self._maximum_identical_calls = maximum_identical_calls
    self._counts = {}

start_turn

start_turn()

Forgets the previous turn's calls.

Source code in src/trading_agent/agent/loop_guard.py

def start_turn(self) -> None:
    """Forgets the previous turn's calls."""
    self._counts = {}

allows

allows(call)

Records a call and says whether it may run.

Parameters:

Name	Type	Description	Default
call	ToolCall	The call the model made.	required

Returns:

Type	Description
bool	False once the same tool and arguments have been seen more than the maximum in this turn.

Source code in src/trading_agent/agent/loop_guard.py

def allows(self, call: messages.ToolCall) -> bool:
    """Records a call and says whether it may run.

    Args:
        call (messages.ToolCall): The call the model made.

    Returns:
        bool: False once the same tool and arguments have been seen more than the maximum in this turn.
    """
    key = call.name + ' ' + json.dumps(call.arguments, sort_keys=True)
    self._counts[key] = self._counts.get(key, 0) + 1
    return self._counts[key] <= self._maximum_identical_calls

refusal_text

refusal_text(call)

The message sent to the model instead of running a refused call.

Parameters:

Name	Type	Description	Default
call	ToolCall	The refused call.	required

Returns:

Type	Description
str	An explanation the model can act on.

Source code in src/trading_agent/agent/loop_guard.py

def refusal_text(self, call: messages.ToolCall) -> str:
    """The message sent to the model instead of running a refused call.

    Args:
        call (messages.ToolCall): The refused call.

    Returns:
        str: An explanation the model can act on.
    """
    return f'Refused: you have already called {call.name} with these exact arguments {self._maximum_identical_calls} times in this turn. Use the earlier results, try different arguments, or give your answer.'