LOCALAI_WATCHDOG_BUSY had no matching timeout setting, so the role could only ever use LocalAI's built-in 5 minutes. On the throttled Tesla T4 a legitimate request can take more than 6 minutes, so default to 10 minutes, which also matches the ingress proxy-read-timeout. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>