Struct AgentRunner 

pub struct AgentRunner<S: CheckpointStore> { /* private fields */ }

The robust agent runner — drives the tool call loop with retries, compaction, and multi-format parsing.

Implementations

impl<S: CheckpointStore> AgentRunner<S>

pub fn new(runner: Runner<S>, config: RunnerConfig) -> Self

Create a new AgentRunner wrapping an existing Runner.

pub fn with_hook_registry(self, registry: Arc<HookRegistry>) -> Self

Set the hook registry for lifecycle hooks (SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, SessionEnd).

pub fn with_approval_checker(self, checker: Arc<dyn ApprovalChecker>) -> Self

Set an approval checker (PreToolUse). When set, each tool call is checked; if false, the tool is blocked.

pub fn usage(&self) -> &UsageAccumulator

Get the cumulative usage statistics

pub fn add_tool(self, tool: impl Tool + 'static) -> Self

Register a tool for use in the loop.

pub fn add_tools(self, tools: Vec<Arc<dyn Tool>>) -> Self

Register multiple tools at once.

pub fn inner(&self) -> &Runner<S>

Access the underlying runner.

pub fn inner_mut(&mut self) -> &mut Runner<S>

Access the underlying runner mutably.

pub async fn run( &mut self, callable: &dyn Callable, input: &str, project_dir: Option<&Path>, ) -> Result<LoopOutcome>

Run the robust agent loop.

This is the core method — it replaces both AgenticLoop::run (which was a skeleton) and LlmCallable::run (which had a basic loop but no retries, compaction, or multi-format parsing).

Flow

1. Call the callable with current input
2. Parse the response for tool calls (JSON → XML → Markdown)
3. If tool calls found: execute them, append results to history, loop
4. If no tool calls: return the final response
5. Throughout: check limits, retry on errors, compact history, checkpoint

impl AgentRunner<InMemoryCheckpointStore>

pub fn default_new() -> Self

Create a new AgentRunner with default in-memory store and default config.

pub fn with_config(config: RunnerConfig) -> Self

Create with a specific config.