agents_runtime::middleware

Trait AgentMiddleware

Source 
pub trait AgentMiddleware: Send + Sync {
    // Required methods
    fn id(&self) -> &'static str;
    fn modify_model_request<'life0, 'life1, 'life2, 'async_trait>(
        &'life0 self,
        ctx: &'life1 mut MiddlewareContext<'life2>,
    ) -> Pin<Box<dyn Future<Output = Result<()>> + Send + 'async_trait>>
       where Self: 'async_trait,
             'life0: 'async_trait,
             'life1: 'async_trait,
             'life2: 'async_trait;

    // Provided methods
    fn tools(&self) -> Vec<ToolBox> ⓘ { ... }
    fn before_tool_execution<'life0, 'life1, 'life2, 'life3, 'async_trait>(
        &'life0 self,
        _tool_name: &'life1 str,
        _tool_args: &'life2 Value,
        _call_id: &'life3 str,
    ) -> Pin<Box<dyn Future<Output = Result<Option<AgentInterrupt>>> + Send + 'async_trait>>
       where Self: 'async_trait,
             'life0: 'async_trait,
             'life1: 'async_trait,
             'life2: 'async_trait,
             'life3: 'async_trait { ... }
}
Expand description

Middleware hook that can register additional tools and mutate the model request prior to execution. Mirrors the Python AgentMiddleware contracts but keeps the interface async-first for future network calls.

Required Methods§

Source

fn id(&self) -> &'static str

Unique identifier for logging and diagnostics.

Source

fn modify_model_request<'life0, 'life1, 'life2, 'async_trait>( &'life0 self, ctx: &'life1 mut MiddlewareContext<'life2>, ) -> Pin<Box<dyn Future<Output = Result<()>> + Send + 'async_trait>>
where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait,

Apply middleware-specific mutations to the pending model request.

Provided Methods§

Source

fn tools(&self) -> Vec<ToolBox> ⓘ

Tools to expose when this middleware is active.

Source

fn before_tool_execution<'life0, 'life1, 'life2, 'life3, 'async_trait>( &'life0 self, _tool_name: &'life1 str, _tool_args: &'life2 Value, _call_id: &'life3 str, ) -> Pin<Box<dyn Future<Output = Result<Option<AgentInterrupt>>> + Send + 'async_trait>>
where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait, 'life3: 'async_trait,

Hook called before tool execution - can return an interrupt to pause execution.

This hook is invoked for each tool call before it executes, allowing middleware to intercept and pause execution for human review. If an interrupt is returned, the agent will save its state and wait for human approval before continuing.

§Arguments
  • tool_name - Name of the tool about to be executed
  • tool_args - Arguments that will be passed to the tool
  • call_id - Unique identifier for this tool call
§Returns
  • Ok(Some(interrupt)) - Pause execution and wait for human response
  • Ok(None) - Continue with tool execution normally
  • Err(e) - Error occurred during interrupt check

Implementors§