What's new in V2

V1 is one job, one attempt, gone when the process exits. V2 adds durable, multi-step workflows on top of it — additively. Nothing about Queue, Job, Priority, or how V1 behaves has changed; a V1 consumer who never touches workflows sees zero difference after upgrading.

New: durable workflows

• Workflow / StepDef. An ordered list of steps under a stable id, each with its own RetryPolicy and optional timeout.
• WorkflowEngine. Drives a workflow’s steps through the same Queue V1 jobs use — it’s the only thing that calls Queue::push for a workflow step. Retries a failed step per its RetryPolicy, entirely on its own; Queue itself stays exactly as retry-agnostic as it’s always been.
• Waiting on external input. A step can return WaitForInput to block on something outside the process — an approval, a webhook, a human. run pauses the whole workflow there rather than failing it; WorkflowEngine::resume supplies the missing input and continues through the remaining steps.
• Non-retryable failures. A step can return NonRetryable to fail the whole workflow immediately, skipping the rest of its RetryPolicy’s attempts — for errors that retrying can’t fix (bad input, a 4xx from a downstream API).
• Per-step timeouts. StepDef::timeout is now enforced — an attempt that runs past its timeout is treated as a failure and retried like any other. The original job keeps running in the background after a timeout; see Durable store and the CHANGELOG for the exact tradeoff this makes.
• Cancellation. WorkflowEngine::cancel idempotently stops a workflow. A step attempt already pushed to the queue isn’t aborted, but no further step runs afterward.
• DurableStore. A trait for where step status and event history live. SqliteStore ships as the one concrete implementation, behind the optional sqlite Cargo feature — off by default, so a consumer who only wants the raw Queue never pulls in rusqlite. It also records each event’s timestamp (SqliteStore::load_events_with_timestamps), with existing databases migrated automatically.
• Recovery. agentq::recover finds steps whose worker died mid-lease (via DurableStore::expired_leases) and re-admits them, once you’ve re-registered the workflow’s bodies with WorkflowEngine:: register_workflow for the current process.

See the Durable workflows guide for how the pieces fit together, and Durable store for how SqliteStore is actually structured.

Install

[dependencies]
agentq = { version = "0.2", features = ["sqlite"] }

Without the sqlite feature, nothing changes from V1 — rusqlite never enters your dependency tree, and none of the new public types cost you anything you don’t use.

What didn’t change

• Queue, Job, Priority, lanes, backpressure, idempotency — identical to V1. See Guarantees for what still does and doesn’t hold at that layer.
• The raw Queue still does not persist anything or retry automatically. Those are workflow-layer features now, not Queue features — see Upcoming for what’s still actually planned for Queue itself.