Struct AgentRunner 

pub struct AgentRunner { /* private fields */ }

Drives a session forward, multi-turn, until the model emits no tool call or the iteration cap is reached.

Clone is intentional: Arc<AgentRunner> is awkward because every field is already cheaply cloneable, and the TUI spawns runner work onto its own task which needs an owned value.

Implementations

impl AgentRunner

pub fn new( provider: Arc<dyn LlmProvider>, tools: ToolRegistry, model: impl Into<String>, ) -> Self

Construct a runner with the V1 default policy (ApprovalPolicy::DenyDestructive).

pub fn model_handle(&self) -> Arc<Mutex<String>>

Share the live model handle so the TUI /model command can swap the active model mid-session (Phase EE).

pub fn provider_handle(&self) -> Arc<Mutex<Arc<dyn LlmProvider>>>

Share the live provider handle so the TUI /provider command can swap the whole backend mid-session (Phase LL).

pub fn skill_toggle_handle(&self) -> Arc<Mutex<HashSet<String>>>

Share a live skill-toggle set with the caller. Skill ids present in the set are skipped during per-turn matching, letting the TUI enable / disable skills mid-session (Phase X).

pub fn last_request_handle(&self) -> Arc<Mutex<Option<String>>>

Share a handle to the most-recent-request snapshot. The TUI /inspect panel (Phase Y) reads the pretty-printed JSON written here on every turn.

pub fn with_approval(self, policy: ApprovalPolicy) -> Self

Override the approval policy.

pub fn with_max_iterations(self, n: usize) -> Self

Override the per-turn iteration cap.

pub fn with_skills(self, skills: Vec<Skill>) -> Self

Register a skill catalogue. Active skills are matched per turn against the latest user message + the (optional) project slug.

pub fn with_project(self, project: impl Into<String>) -> Self

Set the project slug used for project-pattern skill triggers.

pub fn with_auto_retrieve( self, enabled: bool, top_k: usize, min_len: usize, ) -> Self

Enable retrieve-then-generate. Before each turn, pre-fetch RAG context for the latest user message via the rag_search MCP tool and inject it as a system block — helps weaker models that don’t reliably call the tool themselves on interrogative messages. The agent stays free to call read_document / find_related to dig deeper; this only pre-loads. No-op when no …__rag_search tool is registered, on slash-commands, or on messages shorter than min_len chars. top_k is clamped to 1..=10.

pub async fn run(&self, messages: Vec<Message>) -> Result<TurnResult>

Run the loop and return the full transcript.

Equivalent to AgentRunner::run_streaming with a discarded event channel — convenient for tests and batch callers that don’t need progressive UI updates.

pub async fn run_streaming( &self, messages: Vec<Message>, events: Sender<TurnEvent>, ) -> Result<TurnResult>

Run the loop, emitting TurnEvents on events as they happen.

events is dropped when the function returns, which signals the caller that the run is complete (a receive on the matching receiver will then yield None).

pub async fn summarize(&self, history: &[Message]) -> Result<String>

Summarize a slice of conversation into a single compact paragraph (Phase BB). One-shot, tool-free, non-streaming — used by the TUI auto-compaction to fold old turns into a system note.

Trait Implementations

impl Clone for AgentRunner

fn clone(&self) -> AgentRunner

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.