Struct ConditionalGuardrail 

pub struct ConditionalGuardrail { /* private fields */ }

Predicate-gated guardrail — only runs the inner guardrail when the predicate returns true for the tool name.

Use for patterns like “apply this guardrail only to MCP tools” or “only check bash tool calls”.

The predicate receives the tool name for pre_tool/post_tool hooks. For pre_llm/post_llm hooks (no tool name), the inner guardrail always runs.

Implementations

impl ConditionalGuardrail

pub fn new( inner: Arc<dyn Guardrail>, predicate: Arc<dyn Fn(&str) -> bool + Send + Sync>, ) -> Self

Create a guardrail that only activates for tool calls matching predicate.

Trait Implementations

impl Guardrail for ConditionalGuardrail

fn name(&self) -> &str

Human-readable name for this guardrail, used in events and audit. Override to attribute which guardrail fired in logs.

fn pre_llm( &self, request: &mut CompletionRequest, ) -> Pin<Box<dyn Future<Output = Result<(), Error>> + Send + '_>>

Called before each LLM call. Can mutate the request (e.g., inject safety instructions, redact content). Err aborts the run.

fn post_llm( &self, response: &mut CompletionResponse, ) -> Pin<Box<dyn Future<Output = Result<GuardAction, Error>> + Send + '_>>

Called after each LLM response. Can inspect and mutate the response (e.g. redact PII in ContentBlock::Text blocks before it reaches the caller, audit log, or downstream tools).

fn pre_tool( &self, call: &ToolCall, ) -> Pin<Box<dyn Future<Output = Result<GuardAction, Error>> + Send + '_>>

Called before each tool execution. Can deny individual tool calls. Deny returns a ToolResult::error for that call. Err aborts the run.

fn post_tool( &self, call: &ToolCall, output: &mut ToolOutput, ) -> Pin<Box<dyn Future<Output = Result<(), Error>> + Send + '_>>

Called after each tool execution (after truncation). Can mutate the output (e.g., redact sensitive data). Err converts to a tool error (consistent with tool execution errors — the agent loop continues).

fn set_turn(&self, _turn: usize)

Called by the agent loop before each guardrail evaluation to provide the current turn number. Stateful guardrails (e.g., BehavioralMonitorGuardrail) can override this to track turn context.