Loading page...Загружаем страницу...正在加载页面...
AtlasRepo

69% of my coding agent's "done" claims weren't. Here is the gate I put in front of them.

Описание и применение

Кодовые агенты часто заявляют о завершении работы без проверки тестов. Внедрение Stop‑хука isitdone позволяет проверять тесты в момент заявления «готово», блокируя агента при неудаче и обеспечивая надёжность изменений.

Задача

Coding agents often claim completion without verifying that tests pass, leading to unreliable code changes.

Результат

The gate reduced false completion claims, ensuring agents only declare "done" after passing tests on the current working tree, with caching and diff scanning to prevent test tampering.

Рабочий процесс

  • 1. Install isitdone via npx isitdone init.
  • 2. Configure the agent to use the Stop hook (e.g., Stop in Claude Code).
  • 3. When the agent attempts to stop, isitdone runs the project's test command on the current working tree.
  • 4. If tests fail, isitdone returns a block message with the failure output.
  • 5. The agent fixes the issue and retries stopping; after three failures the hook allows stopping.
  • 6. Fast checks (typecheck, lint) run on every stop; full test suite runs only on completion claims.
  • 7. isitdone scans diffs for test‑weakening patterns and warns or blocks in strict mode.
  • 8. Successful runs generate a signed receipt tied to the tree hash for PR verification.

Доказательства

69% had no passing test run behind them.

Источник

isitdone: NOT DONE. 1 check failed on the current working tree (attempt 1/3).

Источник

  • Дата записи: 2026-09-22T12:15:48.137Z