Task queue ports (TaskQueue, TaskStore, TaskWorker) with a local virtual-thread implementation. Payloads are data-only, delivery is at-least-once, cancellation is cooperative — the contract is written for the cluster path (DB store with SKIP LOCKED + lease) from day one, even though the local implementation would not need it. See agents/doc/architecture/concepts/11-task-queue.md.

The agent runtime's core: agent and session domain, the ports around it (chat, tools, token counting, memory strategies, repositories) and the use cases — the turn loop, prompt assembly, tool dispatch. Free of storage, HTTP, ...

Postgres-backed stores for the task queue: JdbcTaskStore (SKIP LOCKED claims, leases for crash recovery) and JdbcChannelStore (store-assigned seq per channel). Plain JDBC on a DataSource — no Spring required.

The task queue as a cluster: ONE jar, two roles. The master serves the dashboard, submits tasks and starts/monitors worker JVMs as OS processes; workers run the unchanged LocalTaskQueue against the shared Postgres store.

Demo app for mc-task-queue: pick a task type, submit it with parameters, and watch tasks and channels live on a streaming board (semantic-ui SSE). Flagship task: crawl a website as a parent/child task tree, one page per child, ...

Cron for the task queue, as an addon rather than a queue feature: a TaskSchedule is a long-lived definition in its own store, and every firing becomes an ordinary TaskSubmission. Includes a dependency-free cron parser.