Groups and scheduling

Groups and scheduling

A Worker can belong to many Mission Groups without merging their contexts. It maintains a distinct Event inbox, Cursor, and Work queue for each Group, and those local projections remain rebuildable from accepted Events (MWP-WRK-008).

Group A Events → Group A queue ┐
Group B Events → Group B queue ├→ Worker Scheduler → isolated execution slots
Group C Events → Group C queue ┘

Who controls priority

The Coordinator supplies requested urgency, deadline, and business impact for its Mission. The Worker Scheduler has final control over cross-Group ordering; one Coordinator cannot force its Mission ahead of unrelated Groups (MWP-WRK-009).

The Scheduler applies weighted fairness and anti-starvation. Effective order considers Organization priority class, deadlines, queue aging, per-Group quotas, risk gates, and available resources. Pure unbounded highest-priority-first scheduling is excluded by MWP-WRK-010.

Isolation and schedule disclosure

Each active slot isolates Mission context, credentials, checkpoints, tool budgets, and side-effect keys from every other slot (MWP-WRK-011).

A Worker can publish estimated start and completion windows, confidence, capacity status, and calculation time without exposing another Group’s identity, content, assignments, or exact global queue position (MWP-WRK-012).

Ownership, leases, and safe preemption

Acceptance begins an Ownership Epoch and a start-by deadline. Active execution uses an Execution Lease bound to the Agent, Session Epoch, WorkItem, Ownership Epoch, stable Lease ID, and expiry (MWP-WRK-013).

Priority changes reorder queued work immediately. Active work reaches a safe, idempotent checkpoint before releasing its slot:

1. the Scheduler requests preemption;
2. the Worker reaches a safe checkpoint;
3. the Worker records the checkpoint and pauses;
4. the execution slot is released; and
5. the WorkItem returns to its Group queue.

This sequence is the normative cooperative-preemption model in MWP-WRK-015.

Recovery by replay

Event delivery is at least once. A restarted Worker restores per-Group Cursors and queues from snapshots and ordered replay, deduplicates by Event ID, and does not advance a durable Cursor over a sequence gap (MWP-WRK-029). Acknowledgement reports the highest contiguous durable position without deleting authoritative Group history (MWP-WRK-030).

MissionWeaveProtocol defines ordering within each Group, not one global order across Groups.