Adding an instruction to request feedback before completing a task is what makes this work at all, and it is the difference between an agent that marks a broken build finished and one that shows you the running app first. For multi-step start commands, put them in a task runner and point the saved command at that rather than trying to chain them in one string.
A small MCP server that opens a feedback window: the agent describes what it did, you look, optionally run the project's own command, and type back what is wrong. Built for desktop-app work, where checking the result means actually using the thing.
- `user_feedback` — takes a project directory and the agent's summary, and returns your reply
- A per-project `.user-feedback.json` holding the command to run, saved from the UI
- An execute-automatically option, so the command runs the moment the window opens instead of waiting for a click
- A human-in-the-loop step that works in Cline and Cursor without either of them supporting one natively
Uv installed globally, and the repository cloned somewhere your client can point at. Add the server to your client config with a generous timeout — the call is open for as long as you take to answer — and auto-approve the feedback tool so it does not need a confirmation click every time.