Context: ROADMAP.md → Admission queue with deferred pickup
Related: #331 (closed bug—this is the feature successor)

Component

API or orchestration

Describe the feature

When admission is at capacity, persist tasks in a durable queue instead of failing them (see #331 live epic failure). Automatically re-attempt admission and continue processing in FIFO order with optional priority lanes as concurrency becomes available.

Must preserve: cancel semantics, idempotency semantics.

Must expose: queue position and ETA in task status / bgagent status.

Use case

#331 demonstrated mass failure when fan-out releases more children than the per-user concurrency cap. Hard-failing subtasks wrecks entire epics (#247). Users expect queueing, not 429-style task death.

Proposed solution

1. New state or sub-status: QUEUED / SUBMITTED with admission_pending: true.
2. Durable queue (DynamoDB GSI or SQS) polled by admission worker.
3. createTaskCore returns 201 with queue position when capped—not terminal failure.
4. Cancel removes from queue; idempotency keys dedupe.
5. API + CLI expose queue_position, estimated wait.
6. Regression test reproducing Fan-out orchestration mass-fails against the per-user concurrency cap (no queue/retry) #331 scenario.

Other information

• Successor to closed Fan-out orchestration mass-fails against the per-user concurrency cap (no queue/retry) #331—link in issue body when filing.

• Design context: docs/design/ORCHESTRATOR.md.

• This might be a breaking change