Struct BashRiskHooks 

pub struct BashRiskHooks;

Always-on catastrophic-shell guard. Stateless — a unit struct is the entire type.

Implementations

impl BashRiskHooks

pub fn new() -> Self

Kept for symmetry with PlanModeHooks::new, so the wiring in the CLI reads Arc::new(BashRiskHooks::new()) like every other built-in hook.

Trait Implementations

impl Default for BashRiskHooks

fn default() -> Self

Returns the “default value” for a type.

impl Hooks for BashRiskHooks

fn before_tool_call<'life0, 'life1, 'async_trait>( &'life0 self, call: &'life1 ToolCall, ) -> Pin<Box<dyn Future<Output = Option<String>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait,

Some(reason) iff this is a bash call whose command is catastrophic; None (allow) otherwise — the hot path.

fn transform_tool_input<'life0, 'life1, 'async_trait>( &'life0 self, _call: &'life1 mut ToolCall, ) -> Pin<Box<dyn Future<Output = ()> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait,

May mutate the tool call’s arguments before it is executed (e.g. to rewrite a bash command through rtk). Mutations are what actually executes, so before_tool_call/after_tool_call observe the rewrite.

fn after_tool_call<'life0, 'life1, 'life2, 'async_trait>( &'life0 self, _call: &'life1 ToolCall, _out: &'life2 mut ToolOutput, ) -> Pin<Box<dyn Future<Output = ()> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait,

Runs after every tool execution; may mutate the output before it is recorded and sent to the LLM.

fn before_inbound<'life0, 'life1, 'life2, 'async_trait>( &'life0 self, _from: Option<&'life1 SessionId>, _request: &'life2 mut Request, ) -> Pin<Box<dyn Future<Output = Option<String>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait,

Policy for an inbound peer message (an A2A Notify/Interrupt/Wake): the mirror of Hooks::before_tool_call. May rewrite the request in place (the rewrite is what is delivered); Some(reason) drops it — the reason is returned to the sender when it is an ask. A message dropped here never reaches the context or starts a turn.

fn transform_context<'life0, 'life1, 'async_trait>( &'life0 self, _msgs: &'life1 mut Vec<AgentMessage>, ) -> Pin<Box<dyn Future<Output = ()> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait,

Runs at every turn start, after steering is drained, before streaming.

fn should_stop_after_turn<'life0, 'life1, 'async_trait>( &'life0 self, _ctx: &'life1 [AgentMessage], ) -> Pin<Box<dyn Future<Output = bool> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait,

Checked after tool execution completes; true ends the run with StopReason::Stop.

fn after_run<'life0, 'life1, 'async_trait>( &'life0 self, _ctx: &'life1 [AgentMessage], _stop: StopReason, ) -> Pin<Box<dyn Future<Output = ()> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait,

Runs once when a run finishes, with the final context and why it stopped. The A2A auto-report rides this (a worker forwards its result to its orchestrator); a no-op by default.