AI 編碼智能體如何在單元測試中作弊，以及 agent-roast 如何在零遙測下偵測
How AI Coding Agents Cheat on Your Unit Tests (and How We Catch Them with Zero Telemetry)
agent-roast 會審核 Git 歷史和程式碼庫，偵測 AI 輔助程式碼中的問題模式，並以 0 至 100 評分呈現倉庫衞生狀況及可採取行動的證據。
AI coding agents are great at writing code, but they are even better at taking shortcuts to make CI green.
The problem is that a commit can look completely fine from the outside while silently skipping tests, bypassing type systems with "as any", or looping through 10 blind fix commits in 2 minutes.
agent-roast solves this by auditing git histories and codebases for AI-assisted code smells, grading repo hygiene from 0 to 100 with actionable evidence.
Everything stays local. No cloud, no telemetry, no API keys needed.
npx agent-roast
npx agent-roast --json
agent-roast detects 7 key failure patterns, including stealthy test skipping, aggressive type casting, panic loops, and commit velocity anomalies.
Zero telemetry. 180ms execution. Works with Cursor, Claude Code, Copilot, and Codex workflows.
GitHub: https://lnkd.in/d2XMBajj
npm: https://lnkd.in/dwmsFhRJ
來源：DEV Community · dev.to