Nova Sprint coordinates AI work across models and harnesses

Summary

Nova Sprint is an open-source system for coordinating AI collaborators that run in different models, sessions, and harnesses. Its authors say that relying on chat alone made assignments, dependencies, handoffs, reviews, and interrupted sessions unreliable, leaving a human to act as scheduler. Nova Sprint stores the plan, assignments, attempts, reviews, state transitions, and landing status outside the conversations. A runner enforces the workflow, releases work when dependencies land, assigns ready cards to free capacity, sends completed attempts to review, recovers interrupted operations, and rejects stale results from older assignments. The system represents work with bounded cards, work streams, dependencies, and sentinel gates. These primitives support sequential work, parallel streams, prerequisite joins, phased rollouts, and manual decisions. Automatic sentinels can release when required work lands, while manual sentinels wait for a coordinator. Different AI collaborators can work at their own pace, with concurrency limits and configured model and harness routes; fleets can provide additional capacity. Ordinary cards move through implementation, review, checks, and landing, while sentinel gates do not consume a worker. The AI coordinator handles planning, findings, and exceptions, but escalates decisions outside its authority to a human. The repository is presented as open source under the MIT license and available for use now.