Plan

Plan gives the AI room to understand the Goal and decide how to approach it before implementation. The Goal may come from a user or another AI. Refine does not require a user-specified outcome, an acceptance checklist, or a prescribed planning method.

Shared execution preparation pins the current Goal, Round, step, node, and coherent Git lineage before Refine runs the selected Plan Skills. This applies after Todo admission, a direct Plan selection, and recovery of unusable generated work. Their instructions guide the AI’s judgment. The agents report success, failure, or an execution error; useful planning detail is optional. Multiple reports remain separate in Skill history and are supplied to implementation as context. There is no additional planning phase machine or checklist coverage rule.

Plan may edit the repository or request a supported workflow change when its Skill instructions and current user instructions call for it. Refine checks response identity and current workflow authority, retains results and process records, and follows the reported decision. Failed verdicts are not automatically retried; another attempt requires an explicit workflow decision. Unfinished work remains scheduled and can continue from retained artifacts once its former execution is proven stopped.

Historical structured plans, criticism, revisions, and implementation reports remain readable. They provide context without independently blocking current work. Selecting Plan retains the authored Round; recovery may choose a fresh execution branch to reproduce unusable generated work under that request. A separately submitted Round adds new authored work. The shared consistency contract defines preparation, invalidation, and supersession.