What problem does it solve? Teams running multiple AI subscriptions waste frontier-model budget on mechanical work and stall when one account hits its rate cap while others sit idle. This Skill provides a routing layer that picks the cheapest capable model with headroom and keeps a shared ledger of usage across accounts. ## Core Features & Use Cases - Capability-ranked routing: coord-engine route --needs code,long-context ranks models cheapest-capable-first, filtered to accounts with current headroom, using a frozen capability taxonomy (code, architecture, writing, long-context, vision, fast, tool-use). - Cross-account cap ledger: declared rolling windows per account track spend; a real throttle event zeroes that account's headroom for the window and flags it for cap calibration. - Outcome-driven demotion: logging --model --task-class --outcome demotes a (model, task_class) pair after repeated rework/escalated outcomes, so the router stops picking models that keep failing a task class. - Use Case: A coordinator agent on a multi-harness team needs code review done; it runs route --needs code, dispatches to the top-ranked model on the account with headroom, and logs the outcome so the fleet ledger stays honest. ## Quick Start Install the fulcra and coord-engine CLIs, run coord-engine atc init to declare your subscription accounts, then ask the agent to route a task with coord-engine route solo --needs code before dispatching it.
