It turns out this wasn’t an AI model problem at all, but a repo-discovery failure. Because Antigravity is designed as an agentic platform that needs to understand your entire workspace to function, the whole pipeline stalls if its internal Git parser can’t resolve your repository metadata. In my case, the repo had the worktreeConfig extension enabled, which triggered a compatibility error because the platform’s parser didn’t support that specific Git extension.
The fix was to remove the extensions.worktreeConfig entry and run a prune to clean up stale metadata, reverting the repo to a standard format. The main trade-off is that I lost the ability to have unique Git settings for individual worktrees, though the worktrees themselves still function normally for standard development. If you’re seeing silent failures, it’s likely your repo metadata is confusing the workspace resolution logic.
For more context, the Google AI 2025 announcements explain Antigravity’s architecture as an agent-first platform that requires successful repository mapping to initiate tasks. You can also find technical details in the official Git documentation, which explains why certain parsers refuse the worktreeConfig extension to maintain compatibility with older repository formats.