Where does the bug appear (feature/product)?
Somewhere else…
Describe the Bug
Cloud agents repeatedly report “working” without real end-to-end tests: three days of fix-on-fix on an MCP/OAuth integration. Problem whit Grok & Cursor
Body:
Over the past three days I used Cursor cloud agents to build and fix an MCP server with OAuth (Supabase OAuth Server behind a Vercel proxy, clients ChatGPT, Claude Code and Codex). The quality of the fixes and the code was not acceptable, and I want this to improve in upcoming releases.
What went wrong:
The agents declared things “working” or “aligned” based on script tests and CI that talked directly to the Supabase functions. The real browser flow through the public domain kept failing (CSRF_INVALID on the consent screen). The cause turned out to be the Vercel proxy overwriting the CSP header with form-action ‘self’, which blocked the redirect to the client callback. No test ever went through the public URL, so CI was green while production was broken.
My own binding playbook for MCP/OAuth projects, stored in a GitHub repo, was not read or followed. It explicitly requires manual end-to-end tests in the real clients (ChatGPT, Claude) before release.
Every round fixed one layer and left the next one untouched. The real root cause was only found after several production deploys, by reading the logs, which should have been step one.
A parallel agent session deployed a wrong version of a file to production. Nothing stopped two agents from working on the same problem at the same time.
Repeated promises of “from now on we’ll do better”, without visible improvement.
What I expect from Cursor agents:
Test end to end the way a real user does, through the public URL and the real client, before calling anything “working”. Check the headers the browser actually receives, not only the upstream response.
Read and follow project documentation and playbooks that the user points to before starting.
Find the root cause from logs before shipping a fix.
Guard against parallel agents deploying conflicting code for the same project.
Deliver thorough, correct code on the first attempt instead of fix-on-fix.
Steps to Reproduce
I hope future updates will deliver a better experience.
Operating System
Windows 10/11
iOS
Version Information
For Grokbot & Cursor
Does this stop you from using Cursor
No - Cursor works, but with this issue