Bug: Locked-page approval succeeds, but writes still request approval in ChatGPT connector

Superhuman Developers Central
1

Hi! I’ve recently been working with the Superhuman Docs connector in ChatGPT and have been running into issues with the “Allow locked page edits” tool. The tool has been working in Claude for me. Wondering if anyone else has run into this issue.

Environment: Superhuman Docs MCP in ChatGPT Work Mode, observed October 31, 2026.

Steps observed

  1. User explicitly authorizes locked-page/table edits, keeping them locked.

  2. Call allow_locked_page_edits({}).

  3. Receive: “This connection may now edit locked pages and tables”.

  4. Attempt table_columns_manage or table_view_manage.

  5. Receive LOCKED_PAGE_EDIT_NEEDS_CONFIRMATION again.

Expected: Successful consent allows authorized writes within its stated scope while preserving locks.

Evidence

  • Consent returned success twice, including one immediately followed by a write within the same client orchestration call.

  • Readback confirmed the changes were not applied.

  • Other reads and some row writes worked.

  • The structured error reports INVALID_ARGUMENT, obscuring the specific approval failure.