Large batch work fails two ways: it stalls mid-flight on a question nobody is around to answer, or it guesses an owner-level call (naming, authorship, scope) and has to be redone. Each planning iteration sorts the next open unknown into one of two buckets — resolvable from the code, docs, or evidence (resolve it yourself), or genuinely the owner's call (queue it in DECISIONS.md) — then asks everything queued in ONE interview, locks the answers, and executes the entire batch with zero mid-execution questions. Distinct from decisive-pick-loop (collapse one decision into a committed recommendation yourself): this SEPARATES the decisions that are yours from the ones that are not, and gates the batch on locking the latter up front.
Loop
- Goal
- every owner-owned decision is locked before execution starts, and the batch then completes with zero mid-execution questions back to the owner
- Exit when
- DECISIONS.md has zero open decisions and the batch has completed without a single mid-execution question to the owner
Check between iterations