Trackline：检测 Cursor 智能体是否偏离任务并可选择拦截
Original title: Trackline: see when the Cursor agent goes off task, and optionally stop it
The title and summary in the selected language are awaiting translation.
作者开发了 Trackline，接入 Cursor 的 hooks，把智能体的每个动作与用户请求和项目规则比对，标记写入 .env 或密钥、未提及的依赖包、超出指定范围的写入、远超请求规模的改动以及重复动作（智能体卡住）。
2026 年9 月 26 日 19:33 1
I built trackline to answer one question while an agent works: is it still doing what I asked?
It plugs into Cursor’s hooks and checks each action against your request and your project rules. It flags writes to secrets (.env, keys), packages you never named, writes outside the area you mentioned, changes far bigger than the request, and the same action repeated (stuck agent).
Setup:
npm install -g trackline
cd your-project
trackline init --host cursor
Then make one edit with the agent and run trackline status to confirm the hook ran.
Two things to know:
- Start the agent inside the project folder. Cursor loads project hooks from the project, so an agent started elsewhere isn’t watched.
- Blocking in Cursor is newer than in Claude Code and Codex. In my one live run it blocked correctly, didn’t retry, and told me to make the change by hand.
Repo: GitHub - yerinsabraham/trackline: Watches whether an AI agent is still doing what you asked. Runs beside Claude Code, Codex and Cursor, checks production agent traces, and gates CI on eval regressions. One vendor-neutral engine. · GitHub
Site: https://trackline.dev
Feedback welcome, especially on false alarms.
Source: Cursor Forum · Showcase · forum.cursor.com