Troubleshooting and recovery

Preserve before repair

Keep the request/run/version ID, exact error and relevant source identity. Do not delete ledgers, history or owner preferences to make a test pass. Work in an isolated copy/settings directory for reproduction. Never send credentials or unredacted private projects to support.

AI failure

If workspace preparation stalls, check local runtime availability and project binding; elapsed preparation is not inference progress. For protocol rejection inspect runtime diagnostics and installed schema/account compatibility before Retry. Keep the failed request and unsent draft. Do not silently downgrade models, switch account class or launch automatic retries.

History and results

Loading is not empty. Retry the failed exact page; do not synthesize missing parents. Stale callback responses from a previous design/version must not replace current selection. If a value is unavailable inspect exact-version artifacts; do not show zero. On conflicts, resolve all scalar/profile choices explicitly or cancel.

Saving, jobs and restart

A failed checkpoint leaves pending work visible; use Retain pending change after addressing the error. Preserve unsaved editor work before a design switch. A cancelled job needs terminal status confirmation. After restart inspect retained scheduler state and receipts; a new conversation is not recovery evidence. Historical/incompatible results remain labeled.

Platform problems

Confirm supported OS/architecture, runtime version, free storage and normal launcher. Keep owner/customer authentication separate. Use an existing rollback only with its compatibility instructions; do not overwrite unbacked-up data. Known startup, performance, platform, mail/legal/billing and signing gates remain separate from UI success.