Heiryn Docs

Auto Model

Auto is the default chat model mode when no personal or server model has been selected.

In Auto mode, Heiryn starts with the configured standard model: GLM 5.3 Flash. It switches to the configured pro model, Claude Haiku 5.5, only when one of these conditions applies:

  • the bot owner explicitly requests the Rohan host-level tools with --rohan;
  • at least one tool explicitly marked as dangerous is loaded for the request; or
  • the requester uses Retry after an interrupted query.

Long conversation history does not trigger the Pro model. History is compacted when it approaches its context budget instead.

This dynamic switch only happens while the resolved model is still Auto. If you select a model explicitly, Heiryn keeps that model for the request instead of upgrading it automatically, including for Rohan, dangerous tools, and retries after interrupted queries.

For example, selecting GLM 5.3 Flash explicitly is different from leaving the setting on Auto:

  • Auto: starts on GLM 5.3 Flash and may switch to Claude Haiku 5.5.
  • GLM 5.3 Flash: stays on GLM 5.3 Flash.

Resolution Order #

The chat model for a request is resolved in this order:

  1. A one-message model suffix.
  2. Your personal model preference.
  3. The server default model.
  4. Auto mode.

Server admins can set the server default model from the dashboard or with /community_chat_model. Users can set their personal preference from the dashboard or with a persistent %%model suffix.

When To Use Auto #

Auto is usually best when you want the bot to balance cost and quality without manual intervention. It keeps normal requests on the standard model and uses the Pro model only for the conditions listed above. Requester or staff status alone, long history, and ordinary safe tools do not trigger an upgrade.

Choose an explicit alias when you want a predictable cost category or the behavior of its current mapping. Explicit choices disable the automatic pro-model upgrade for that scope, but an alias may map to a newer qualifying model on a later request under the model alias policy.