Struct SubAgent 

pub struct SubAgent { /* private fields */ }

The sub-agent tool.

Constructed with:

• workspace: path for sandboxed tools
• provider: the LLM provider (shared Arc from parent)
• all_tools: the full tool registry from which the sub-agent can draw
• max_depth: absolute depth limit (env-configured)
• current_depth: how deep we already are (passed from parent)
• permission_hook: optional permission hook inherited from the parent agent

Implementations

impl SubAgent

pub fn new( workspace: impl Into<PathBuf>, provider: Arc<dyn LlmProvider>, all_tools: ToolRegistry, max_depth: usize, current_depth: usize, permission_hook: Option<Arc<dyn PermissionHook>>, ) -> Self

Trait Implementations

impl Tool for SubAgent

fn is_readonly_for_args(&self, arguments: &Value) -> bool

Returns true when subagent_type is "explore", making the dispatch layer treat this call as read-only and eligible for parallel execution.

fn spec(&self) -> ToolSpec

fn execute<'life0, 'async_trait>( &'life0 self, arguments: Value, ) -> Pin<Box<dyn Future<Output = Result<String>> + Send + 'async_trait>>

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

fn side_effect_class(&self) -> ToolSideEffect

Classify this tool’s observable side-effects. Default is the most conservative value (External) so any unannotated tool is treated as risky on resume. Override to ReadOnly or Mutating for built-in tools; MCP tools derive this from their annotations.

fn is_readonly(&self) -> bool

Convenience: a tool is read-only iff it classifies as ReadOnly. Used by the parallel-dispatch path in agent.rs. Override only if you have an unusual reason (you almost never should — override side_effect_class instead and let this default through).