Cursor's New Cloud-Agent Hooks Ship Broken

Cursor's beforeSubmitPrompt hook returns a rewritten prompt, then silently discards it before the model ever sees it, per Cursor's own acknowledgment. Anyone who wired a policy check into that hook since it shipped July 10 has been sending unfiltered prompts while the hook reports success.

The five hooks, beforeSubmitPrompt, afterAgentResponse, afterAgentThought, subagentStart, and stop, landed in .cursor/hooks.json on July 10, built for self-correcting cloud agent loops, per the changelog.

Inside cloud agent VMs, the hooks don't reliably fire, per user reports since the July 10 release. Cursor confirmed the gap; a fix is pending.

The quiet one is worse

beforeSubmitPrompt exists to let you inspect and rewrite a prompt before an agent acts on it, the natural place to bolt on a policy check. The hook runs, returns an edited prompt, and reports success. The model never sees the edit.

A missing guardrail is obvious. Nothing filters, and you notice. This one logs a pass and convinces you it's working when it isn't.

Whether afterAgentResponse, afterAgentThought, subagentStart, and stop share the silent-discard bug or just the VM reliability problem isn't confirmed. There's no fix timeline yet, so treat any beforeSubmitPrompt-based gate as decorative until the patch lands.