Upgrading to model classes

This upgrade discards AI configuration. There is no compatibility bridge. Until an operator recreates models and assigns a class, inference fails with ai_no_model. The AI enabled switch still applies.

Cloud removes profiles, their keys and headers, model access grants, the five old default/tool model settings, model-specific budget rules, project default models and each person's last model selection. All-model budget rules and conversation and inference history remain.

1. Before upgrading, record the old configuration using a JSON export without secrets. Have the provider API keys ready separately.
2. After upgrading, open Administration → AI (/admin/ai). The old AI tabs under Administration → Settings lead there. Create and test providers, then add models, prices, access grants and variants. Follow Models and providers.
3. Assign at least balanced under model classes. Configure the vision and audio slots where those tools are needed.
4. Recreate model-specific budget rules and project default model requests.
5. Change workflows with fixed models to a class or a new variant ID. Tell OpenAI-compatible API clients about the new IDs and the cloud/fast, cloud/balanced, cloud/thorough and cloud/maximum class identifiers.

Verify a chat, the tools in use and a representative workflow. Check that allowed people see the expected choices and that budgets cover the intended models. Old fixed IDs intentionally fail until replaced.