Load balancing

OMP resolves models from agent definitions, role aliases, and task.agentModelOverrides. An alias provides an alternative model name. Supership records logical seats, assignments of agents and models. Overrides within one run do not change global configuration or sibling sessions.

A modelRoles list defines an ordered model selection sequence. It does not distribute requests in rotation. It does not guarantee that a model request succeeds. If a provider fails, an explicitly configured fallback policy determines the alternative. A provider supplies model responses.

The old task-pool rotation, usage-database scraping, and per-call model argument on agent() are not part of this package. Supership does not read private OMP account tables to guess provider health.

Independent seats

Concurrent reviewers can use distinct seats and model assignments. Ultra requires both planning seats and both judge seats. Missing agents/models block the phase. The engine does not silently choose a default model or drop a reviewer.

The engine retains source and model origins for recovery. A declared fallback differs from an undeclared model switch. Invalid structured output receives one same-seat correction and one configured fallback attempt. Unresolved output blocks the action.

Limits

Live OMP concurrency, the number of simultaneous tasks, is the ceiling. A run can specify a lower ceiling. Supership does not raise the global OMP limit. WorkPool schedules repeated independent items. It does not rotate models.

No new finite token, cost, wall-time, or review-round cap applies by default. Tokens are units of model input and output. Wall-time is elapsed clock time.

Interactive planning asks about unspecified limits. Autonomous runs use configured defaults. Work pauses when it reaches a limit. Unknown pricing stays unknown. Active provider requests can exceed a measured budget before the next observable boundary.