strands.vended_interventions.hitl.classifier

Built-in risk classifiers for the HumanInTheLoop handler.

The LLM classifier uses an inner agent with structured output; the decision model classifier asks any DecisionModel a single boolean question. Both evaluate whether a tool call requires human approval against the same risk criteria.

ClassifierResult

@dataclass
class ClassifierResult()

Defined in: src/strands/vended_interventions/hitl/classifier.py:27

Result from a classifier evaluation.

HumanInTheLoopClassifier

@runtime_checkable
class HumanInTheLoopClassifier(Protocol)

Defined in: src/strands/vended_interventions/hitl/classifier.py:35

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

__call__

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

Defined in: src/strands/vended_interventions/hitl/classifier.py:38

Evaluate whether a tool call requires human approval.

Arguments:

• event - The tool call event under evaluation.
• **kwargs - Additional keyword arguments for future extensibility.

Returns:

ClassifierResult indicating whether approval is required.

LLMClassifierConfig

@dataclass
class LLMClassifierConfig()

Defined in: src/strands/vended_interventions/hitl/classifier.py:52

Configuration for the built-in LLM risk classifier.

Kept for backwards compatibility. Prefer passing a DecisionModel as the classifier for new code.

Arguments:

• system_prompt - Risk criteria prompt. Defaults to a general-purpose risk prompt.
• model - Model for risk evaluation. Resolution order: this model > agent.aux_model

  agent.model.