Struct Scheduler 

pub struct Scheduler { /* private fields */ }

Concurrency-controlled agent scheduler. Held as Arc<Scheduler> and shared across orchestration coroutines.

Implementations

impl Scheduler

pub fn new( config: SchedulerConfig, registry: BackendRegistry, journal_callback: Option<Arc<dyn JournalCallback>>, ) -> Arc<Self>

pub fn config(&self) -> &SchedulerConfig

pub fn init_run( &self, run_id: RunId, event_capacity: usize, ) -> Receiver<AgentEvent>

Initialise per-run state. Must be called before any run_agent. Returns the broadcast receiver; further consumers use resubscribe().

pub fn init_run_with(&self, run_id: RunId, events: EventSender)

Initialise per-run state using an externally-owned event sender.

This lets the orchestration layer share a single event bus between the scheduler (AgentStarted/AgentDone, plus the RunContext handed to backends) and the runtime SDK (phase/log/pipeline/RunDone).

pub async fn run_agent( &self, run_id: RunId, task: AgentTask, backend_id: Option<&str>, ) -> Result<AgentResult, SchedulerError>

Schedule and run a single agent task: quota check → permit → retry loop → events. Cancellation flows via RunContext::cancel.

The agent span carries run_id/agent_id/phase_id/model so every log emitted on this task’s async path inherits them (see docs/design/program-logging.md).

pub async fn run_parallel( &self, run_id: RunId, tasks: Vec<(AgentTask, Option<String>)>, ) -> Vec<Result<AgentResult, SchedulerError>>

Run a batch of tasks concurrently (the parallel() primitive). Bounded by the same global semaphore; does not short-circuit on failure — results preserve input order.

pub fn cancel_agent(&self, run_id: RunId, agent_id: AgentId)

Cancel one agent (fires its token; the backend observes ctx.cancel).

pub fn cancel_run(&self, run_id: RunId)

Cancel the whole run (all child agent tokens fire).

pub fn active_concurrency(&self) -> usize

Current global active concurrency.

pub fn quota_used(&self, run_id: RunId) -> Option<u32>

Quota consumed by a run, if initialised.