arch-epic

Wraps $arch-step to orchestrate goals too big for one canonical plan. The skill takes a prose goal, proposes a plain-English decomposition whose count follows proof gates rather than a preset range, gets user approval, and then drives each sub-plan through arch-step's new → auto-plan → implement-loop → audit-implementation arc. After each sub-plan completes, a fresh Claude, Codex, or Grok critic subprocess inspects the shipped work for scope drift against the approved North Star. If the critic finds a requirement needed to preserve approved scope or a non-trivial scope change, the skill halts and asks the user how to preserve the scope by extending the current sub-plan or inserting a new sub-plan. Otherwise it advances to the next sub-plan.

The normal lane is interactive and re-entrant: arch-epic invokes or observes arch-step commands in the visible session one transition at a time.