feat: Schedule through named resource pools #319

New issue

Closed

opened 2026-07-13 19:46:12 +02:00 by dries · 0 comments

dries commented

2026-07-13 19:46:12 +02:00

Owner

Parent

#311

What to build

Schedule command and agent attempts under a required run concurrency cap and optional named resource capacities, so expensive compilers, agents, and commit operations can be bounded independently.

Acceptance criteria

Every published runnable workflow requires a positive concurrency cap.
Definitions can declare named resource pools with positive capacities and nodes can request units from multiple pools.
The scheduler atomically acquires all required resources before creating an active attempt.
Competing ready nodes cannot oversubscribe run concurrency or any named pool.
Resources release only after an attempt reaches a settled state.
Fair scheduling prevents a ready node from starving indefinitely under ordinary contention.
Resource waits and held capacities are durable enough for restart reconciliation and visible in the run UI.
Optional cost, token, and duration limits aggregate descendant work; omission means unlimited.
Tests cover atomic acquisition, contention, cancellation, failure, restart, and configured budget stops.

Blocked by

#314
#315

## Parent #311 ## What to build Schedule command and agent attempts under a required run concurrency cap and optional named resource capacities, so expensive compilers, agents, and commit operations can be bounded independently. ## Acceptance criteria - [ ] Every published runnable workflow requires a positive concurrency cap. - [ ] Definitions can declare named resource pools with positive capacities and nodes can request units from multiple pools. - [ ] The scheduler atomically acquires all required resources before creating an active attempt. - [ ] Competing ready nodes cannot oversubscribe run concurrency or any named pool. - [ ] Resources release only after an attempt reaches a settled state. - [ ] Fair scheduling prevents a ready node from starving indefinitely under ordinary contention. - [ ] Resource waits and held capacities are durable enough for restart reconciliation and visible in the run UI. - [ ] Optional cost, token, and duration limits aggregate descendant work; omission means unlimited. - [ ] Tests cover atomic acquisition, contention, cancellation, failure, restart, and configured budget stops. ## Blocked by - #314 - #315