BuildProof CLI 核實 AI 助手聲稱測試通過
My AI assistant kept saying "all tests pass". So I built a gate that checks.
billbin1969-debug 建立 BuildProof CLI，在 AI 助手與程式碼庫之間設置測試閘門，只有真實測試套件確實執行並通過，變更才會獲接受。它會檢查被削弱的斷言、跳過的測試及為強行通過而修改的測試檔，遭拒變更會回復至最後一個正常狀態；每次獲接受的執行均有可供任何人驗證的 ed25519 簽署證明。
If you code with AI assistants, you've probably seen this: the agent says "done, all tests pass", and then you find out the tests never ran, were skipped, or were quietly edited so they would pass.
I hit this often enough that I stopped trusting the claim and built a tool to check it: BuildProof.
What it does
BuildProof is a CLI that sits between the AI and your repo:
- Fail-closed test gate - a change is accepted only if the real test suite actually ran and passed. If anything is unclear, it counts as a failure, not a pass.
- Guard - detects tampering: weakened assertions, skipped tests, or test files edited to force a pass.
- Rollback - a rejected change is rolled back to the last good state.
- Signed certificates - every accepted run gets an ed25519-signed proof that anyone can verify.
Why "fail-closed"
Most CI setups fail open: if a step is missing or misconfigured, the pipeline can still go green. With AI agents writing more of the code, that gap gets exploited, usually by accident. BuildProof flips the default. No real proof, no pass.
Try it
npm i -g buildproof
There's a 14-day free trial, and plans start at $15/month.
Website: https://buildproof.mgrprosnetwork.com
Your turn
I'd really like feedback, especially if you can think of ways an AI agent could still sneak a fake "pass" past it. Tell me in the comments.
來源：DEV Community · dev.to