More reliable preview recovery

This week focuses on recovering unhealthy previews and keeping their connections working.

Recover before finishing

An agent run now attempts to repair an unhealthy preview before marking its work complete. This helps avoid a finished conversation leaving you with an app that still needs recovery.

Keep connections moving

We fixed recovery for stalled orchestrator heartbeats and corrected resource limits for database proxies. These changes address cases where a preview or its attached database could become unavailable even though the application itself had not changed.