Struct AgentRuntime 

pub struct AgentRuntime { /* private fields */ }

Streaming-first agent runtime kernel.

Ties together provider registry, context pipeline, tool runtime, compaction service, persistent stores, snapshot recovery, session gating, and input admission into a complete agent execution loop.

Implementations

impl AgentRuntime

pub fn new(extensions: Arc<Extensions>, policy: RuntimePolicy) -> AgentRuntime

Creates a new agent runtime from an Extensions facade and a RuntimePolicy.

Internally constructs a ProviderRegistry, RuntimeStore, ContextPipeline, and ToolRuntime from the extensions or their defaults. Use with_tool_registry to populate the tool runtime, and the with_* setters for optional components (event publisher, snapshot store).

pub fn with_tool_registry(self, registry: ToolRegistry) -> AgentRuntime

Replaces the tool runtime with one backed by the given registry.

pub fn with_event_publisher( self, publisher: Arc<dyn EventPublisher>, ) -> AgentRuntime

Sets an external event publisher for the agent runtime.

When set, every AgentEvent emitted during a run will also be published to the configured EventPublisher via fire-and-forget. Requires the queue feature.

pub fn with_snapshot_store( self, snapshot_store: Arc<dyn SnapshotStore>, ) -> AgentRuntime

Sets an optional snapshot store for FSM run recovery.

When configured, intermediate run state is persisted after each turn, allowing crashed runs to be resumed via resume.

pub fn extensions(&self) -> &Arc<Extensions>

Returns the composable Extensions facade.

Use this to register, replace, or look up providers, tools, context adapters, stores, and other pluggable elements by name. The facade is shared with the runtime’s internal references, so registering here is equivalent to using the with_* setters.

pub fn session_gate(&self) -> &SessionGate

Returns the session gate used to serialize concurrent runs per session.

pub fn subscribe(&self) -> Receiver<AgentEvent>

Subscribes to runtime events via a tokio broadcast receiver.

The receiver starts with the oldest event still in the buffer (capacity 256). Slow consumers that fall behind will miss events.

pub fn policy(&self) -> &RuntimePolicy

Returns the runtime policy.

pub fn tools(&self) -> &ToolRuntime

Returns the tool runtime.

pub fn providers(&self) -> &ProviderRegistry

Returns the provider registry.

pub fn context(&self) -> &ContextPipeline

Returns the context pipeline.

pub fn store(&self) -> &Arc<RuntimeStore>

Returns the runtime store.

pub fn compaction(&self) -> &CompactionService

Returns the compaction service.

pub fn snapshot_store(&self) -> Option<&Arc<dyn SnapshotStore>>

Returns the snapshot store, if configured.

pub fn sessions(&self) -> &dyn SessionStore

Returns the session store.

pub fn executions(&self) -> &dyn ExecutionStore

Returns the execution store.

pub fn runs(&self) -> &dyn RunStore

Returns the run store.

pub fn embeddings(&self) -> Option<&dyn EmbeddingStore>

Returns the embedding store, if configured.

pub fn artifacts(&self) -> Option<&dyn ArtifactStore>

Returns the artifact store, if configured.

pub async fn run(&self, request: RunRequest) -> Result<RunOutput, RuntimeError>

Executes an agent run to completion.

The run loop:

1. Creates or loads a session
2. Persists the user message
3. Iterates: build context → call model → persist response → execute tools → repeat
4. Returns the final run ID and finish reason

Errors

Returns RuntimeError on provider, store, context, or policy violations.

pub async fn resume(&self, run_id: RunId) -> Result<RunOutput, RuntimeError>

Resumes a crashed or halted agent run from its last saved snapshot.

Loads the snapshot by run_id, re-acquires the session lock, pushes a Run-level tool scope, and restarts the turn loop from the saved state.

Errors

Returns RuntimeError if the snapshot is not found, the session is busy, or resuming fails.