Model

Model settings control which model answers, how much reasoning it uses, and how Teaflask manages long conversations.

Set the default and fallbacks

Open the agent's Model & access settings. Model routing is one ordered list:

• The first model is the default answering model.
• Later models are fallbacks, tried in order when a provider is throttled, overloaded, or unavailable.

Use Add model to extend the list and reorder it to change the default or fallback priority. With only one model, there is no fallback.

Fallback routing also applies after a model has started streaming its response: when a reply breaks mid-stream and the provider stays down, the next eligible model takes over, the broken model's partial text is withdrawn, and the backup's answer is the only one the conversation keeps.

Set how much each model thinks

Every row in the routing list carries its own Thinking switch and level — the primary and each fallback alike. There is no single effort setting for the list: a fallback never inherits the model above it, so no row is asked for a level its own provider cannot speak.

The level menu offers only the levels that row's model supports (Low, Medium, High, Extra high, Max). A model whose floor keeps thinking between tool calls also offers Between tool calls. When you add a model, its row starts from the primary row's level, adjusted to the nearest level that model supports; after that you edit each row on its own.

A row's thinking can only be switched off on a model that has a true off. On a model that always thinks, the switch stays on and says so — "Always thinks; lowest level is Low," or "Always thinks between tool calls."

When a row's thinking is off, Teaflask does not fall back to the provider's API default: it sends that model's floor — the model's true off where it has one, otherwise its lightest declared setting.

Higher effort can use more tokens and take longer.

Let users choose

Enable Let users choose the model to let users select the answering model and thinking level for their conversation. This is off by default.

The picker offers models from your routing list, not the whole catalog. With a single-model list, users can choose only that model and its supported effort levels.

Without a selection, the configured default answers. With a selection, the chosen model is tried first; the other models remain fallbacks in your configured order. The selection applies to that conversation, not to the agent's default for other users.

Returning to the default makes the conversation follow your configured model again. If you disable user choice or remove a selected model from the routing list, the conversation returns to the default.

Manage conversation context

There is no context strategy to pick: every agent runs on the one conversation manager, and Teaflask keeps a long conversation inside the answering model's context window for you.

Before every model call it measures the conversation against that model's usable window (the window less the room the reply needs). Once the conversation passes about 85% of that usable window, the oldest messages are folded into notes — back to about half the window — and the recent turns are kept whole. The folded messages are not deleted: the agent can read them back word for word while the conversation continues.

These settings manage what the model sees during a conversation. Memory is configured separately.

Save and test your changes in the Playground, then publish a revision to apply them to users.