1.2 KiB
R4.2 P2 Agent Fleet / Multi-Agent Scheduling / Fault Isolation
P2 adds a durable fleet decision layer inside TaskPlane without replacing TaskPlane claim/lease ownership. Worker registration/heartbeat remains authoritative for liveness. TaskPlane claim CAS remains authoritative for ownership. Probe-before-reassignment remains mandatory after an expired lease. P1 checkpoint/result replay remains authoritative after worker restart.
FleetScheduler adds durable worker capacity slots, weighted scope fairness, starvation aging,
affinity/anti-affinity labels, worker circuit breakers, per-task/worker quarantine and idempotent
dispatch reservations. Fleet state is colocated in the existing TaskPlane SQLite database and is
created idempotently; no new daemon or database is introduced.
A fleet reservation is only a scheduling decision. It does not claim a Task. Callers must perform the normal TaskPlane claim and may then bind the returned claim ID to the reservation. This preserves the existing ownership and recovery invariants.
Failure recovery never bypasses probe evidence. When TaskPlane recovery places work back in a claimable state, P2 may quarantine the failed worker and prefer another worker/failure domain.