What problem does it solve?
Wave-based task execution forces workers to wait for an entire wave to finish before new ready tasks can start, leaving agents idle and slowing spec completion. This experimental variant of the flow-next work pipeline admits a new ready task at every worker-return event instead of at wave boundaries, keeping up to three tasks in flight while preserving the canonical review and ship gates.
Core Features & Use Cases
- Rolling admission scheduler: Recomputes the ready frontier at every worker-return event, admitting tasks only when they pass five fail-closed conditions including transitive dependency checks and disjoint Touches declarations.
- Isolated per-task workspaces: Each worker runs in its own worktree with task-unique evidence paths, so concurrent tasks never write into a shared checkout and wrong admissions surface as merge conflicts rather than correctness loss.
- Conductor-owned review and completion: Review dispatch, integration, flowctl done, and plan-sync stage lines are owned by the conductor for every backend, with bounded retries and typed escalations on failure.
- Use Case: Run a multi-task spec where tasks touch disjoint files; instead of waiting for wave boundaries, the conductor keeps three workers busy continuously, integrates each return, and ships with a full-suite verification at quiesce.
Quick Start
Invoke /flow-next:work-rolling with a spec id to run the rolling-frontier scheduler over that spec's ready tasks.
