If Antigravity App leaves you waiting for a thinking model and then reports a failed request, check the update button in the window title bar. Google's November 9 release notes list a fix in v2.31.0 for requests that failed when a thinking model took more than a minute to begin responding. That is a failure fix, not a promise to finish an answer within a minute.
The update control is in the title bar; the fix is for the App
The changelog identifies an update button in the window title bar. When an update is offered, use that control and follow the application's prompts. The correction is listed under Antigravity App, not as a fix for the CLI, IDE, or every other interface. Google also says releases arrive gradually and can take a few days to reach everyone.
A separate correction in the same release addresses an agent left waiting because an artifact requesting feedback hid its Proceed button when Always Proceed was enabled. That is different from a model taking time to start its response. When investigating a stall, distinguish a failed request from a review awaiting action or a task still being processed.
Inspect the turn before repeating it
For coding tasks, the final reply should not be your only basis for deciding whether to start over. Google's review documentation says the Review pane displays changes in the project's Git repository, including agent edits. Its scope is not limited to the output of a successful answer.
To narrow the inspection, click Review on the file-change card beneath the relevant response. The pane filters to that turn and displays For Turn alongside its prompt. Check the affected files before choosing to continue, retry, or commit. A changed file is evidence of an edit, not proof that the requested work passed its checks.
Be particularly careful with Discard unstaged changes. Google documents that it takes effect without a confirmation prompt and also deletes untracked files from disk. Staged changes remain. It is not a harmless way to dismiss the review.
Our recommendation is to repeat the same task after updating. Keep the model, application version, task, and error details, then inspect the resulting diff. If work stops again, record whether the application reports a failed request or is waiting for review, so the next troubleshooting step addresses the right part of the workflow.