unlazy

Why we included this project

Teams that keep getting 'done' reports which turn out to be half-finished work will recognize the exact failure this targets. As an agent skill it installs into Claude Code, Codex, or Cursor and makes the model write an acceptance ledger, GATES.md, before any real work begins, then run actual shell checks whose results decide whether a task counts as closed. Large builds get the Depth Tree treatment: the task is split N layers deep into leaves executed by fresh-context subagents, which avoids the stall that usually sets in near the 80 percent mark once the long context runs out. The approach is measured rather than just argued, with a controlled test showing 1.6 to 3.9 times more effort and 4 to 10 self-found defects caught before delivery. Everything runs on plain markdown, so it works wherever a skill can land, and an optional hook refuses to let the model declare victory while gates are still unmet.