I am seeing a hit-testing issue with an Apps SDK widget in ChatGPT fullscreen mode.
Steps to reproduce:
- Open the widget in fullscreen mode on ChatGPT desktop web.
- Place an interactive button near the bottom-left of the widget.
- Make the browser window narrower while keeping the desktop layout.
- The ChatGPT footer appears with this message:
ChatGPT can make mistakes. Workspace data isn’t used to train models.
Actual behavior:
- The footer or its transparent container overlaps the widget.
- The widget button is still visible, but it cannot be clicked.
- It appears that the ChatGPT host layer receives the pointer event before it reaches the widget iframe.
- The problem depends on the window width. The button works again on wider windows.
I also tested window.openai.safeArea. The reported bottom inset does not appear to include the complete footer panel or its transparent hit area, so subscribing to changing safe-area values does not prevent the overlap.
Because the widget runs in a sandboxed iframe, it cannot inspect or measure the ChatGPT parent DOM. The only available workaround is to guess a fixed responsive offset or move all bottom controls elsewhere, which is fragile because the host footer size can change.
Expected behavior:
ChatGPT should either report the full blocked area through safeArea, resize the widget’s usable viewport, or avoid intercepting clicks outside the visible footer panel.
Questions:
- Should
window.openai.safeArea.insets.bottominclude this footer? - Is there another supported API for getting the footer or host-overlay bounds?
- Is the transparent footer container blocking pointer events outside its visible content a known issue?