Struct LoopRunner 

pub struct LoopRunner<'a> { /* private fields */ }

The ReAct loop execution context.

Holds references to the agent configuration and tool registry and drives the Thought -> Action -> Observation -> FinalAnswer cycle. It never calls a model itself: you pass each model response to step (text ReAct) or step_response (native tool calling). Use Agent if you would rather not manage the borrows.

Implementations

impl<'a> LoopRunner<'a>

pub fn new(config: &'a AgentConfig, registry: &'a ToolRegistry) -> Self

Creates a new runner with its own conversation history.

pub fn with_system_prompt(self) -> Self

Adds the standard ReAct system prompt (format rules plus the tool list) as the first message. System messages are never evicted from history.

pub fn ask(&mut self, question: &str)

Adds the user’s question to the history.

pub fn step(&mut self, llm_response: &str) -> Result<ReActStep, AgentError>

Processes one text LLM response, dispatching the tool when it is an Action.

With config.recover_from_errors, an unknown tool or an unparseable reply becomes an Error: ... observation the model can correct, instead of an error that ends the loop.

Errors

• AgentError::MaxIterationsExceeded: more than config.max_iterations steps.
• AgentError::ParseError: the response does not match any ReAct format.
• AgentError::ToolNotFound: the requested tool is not in the registry.

pub fn step_at( &mut self, llm_response: &str, now_ms: f64, ) -> Result<ReActStep, AgentError>

Like step, and also enforces config.timeout_ms. Pass the current time in milliseconds (for example Date.now()); the first call starts the clock.

Errors

AgentError::Timeout once more than timeout_ms has passed since the first call, plus everything step returns.

pub fn step_response( &mut self, response_json: &str, ) -> Result<NativeStep, AgentError>

Processes one native tool-calling response: the raw JSON body from OpenAI Chat Completions or Responses, Anthropic Messages, or Ollama’s chat API. Every tool call in it is dispatched (arguments arrive as JSON; see ToolRegistry::dispatch_call) and the results are returned and kept in the history with their call ids.

Errors

AgentError::ParseError if the body is not a recognised response, AgentError::MaxIterationsExceeded, AgentError::ToolNotFound (unless recover_from_errors).

pub fn run_scripted(&mut self, responses: &[&str]) -> Result<String, AgentError>

Runs a pre-scripted sequence of LLM responses, returning the final answer.

Useful for offline simulation and deterministic testing.

Errors

• AgentError::MaxIterationsExceeded: sequence exhausted before a FinalAnswer.
• Any error from step.

pub fn transcript(&self) -> String

The conversation so far as one prompt string, ready to send to a completion-style model: system and user messages as-is, model turns as-is, tool results as Observation: ... lines.

pub fn history(&self) -> &ConversationHistory

Returns a reference to the accumulated conversation history.

pub fn iteration_count(&self) -> usize

Number of model responses processed so far.