Three files fix Astra agents that report false completion
SPEC.md sets acceptance criteria, CHECKPOINTS.md forces step-by-step verification, and ROLLBACK.md prevents stacking fixes on broken code, so agents stop failing silently while reporting success.
damn most gpt-6 astra agents are confidently shipping the wrong thing... the code runs. the files exist. the agent says “done.” but nobody defined what “done” actually means. three files fix this: • SPEC.md GOAL: working checkout MUST PASS: payment + refund DONE WHEN: 10/10 tests pass gives astra a finish line it can’t reinterpret halfway through the task • CHECKPOINTS.md [✓] database connected [✓] payment processed [ ] refund verified forces the agent to prove each critical step before moving forward • ROLLBACK.md IF tests fail twice: restore last working version log the failure change the approach stops astra from stacking new fixes on top of broken work the dangerous agent isn’t the one that fails. it’s the one that fails silently and still reports success. save this, then stop accepting “done” without proof ↓
This post quotes another post: x.com/mr_kozh/status/2097221341107925350