strands.vended_interventions.hitl.classifier

Built-in LLM risk classifier for the HumanInTheLoop handler.

Uses an inner agent with structured output to evaluate whether a tool call requires human approval based on risk criteria.

ClassifierResult

@dataclass
class ClassifierResult()

Result from a classifier evaluation.

@runtime_checkable
class HumanInTheLoopClassifier(Protocol)

Callable (sync or async) that decides whether a tool call requires human approval.

def __call__(event: BeforeToolCallEvent,
             **kwargs: Any) -> ClassifierResult | Awaitable[ClassifierResult]

Evaluate whether a tool call requires human approval.

Arguments:

Returns:

ClassifierResult indicating whether approval is required.

@dataclass
class LLMClassifierConfig()

Configuration for the built-in LLM risk classifier.

Arguments:

system_prompt - Risk criteria prompt. Defaults to a general-purpose risk prompt.
model - Model for risk evaluation. Defaults to the parent agent’s model.