Runtime controls

The composer shows controls for the model, thinking level, optional ultrafast, tools, and context usage.

Runtime controls stay with the composer for the selected session.

Select a model

1. Select the model control.
2. Enter text in Filter models.
3. Select a model from the results.

The control and result rows use provider/model-id labels. Filtering also matches the model display name, provider, and model ID.

You cannot change the model during a run, compaction, submission, or runtime reload. Leyline records the selected model in the session and saves it as the pi default for new sessions.

Repeated model selections before the next conversation message update one Model changed from … to … divider. Returning to the original model removes it. Initial model setup does not add a divider.

Use Models & providers for credentials, custom endpoints, and catalog changes. Custom compatible models can remain selectable without an API key. Test connection is optional and does not change the selected model. Reload an existing session to apply saved changes to its model choices.

Select a thinking level

1. Select the thinking control.
2. Select an available level.

A model without reasoning support offers off only. Reasoning models can offer minimal, low, medium, high, xhigh, or max.

The xhigh and max levels appear only when the selected model supports them. Changing the model updates this list. Leyline records the thinking level in the session and saves the selection as the pi default.

Use ultrafast

The lowercase ultrafast chip appears beside the model and thinking controls for supported models and authentication:

• gpt-6-astra or gpt-6.1-sol on openai-codex Responses.
• The same models on direct openai Responses with API-key authentication.

Direct openai subscription authentication does not support this control. Local models do not support it.

ultrafast requests the premium service tier. It is off by default and uses premium pricing. The thinking level stays the same.

Click the chip while the runtime is idle to enable or disable it. The change is direct, with no confirmation dialog.

The choice applies only to the current runtime/session. Model changes, explicit runtime reloads, and reset navigation turn it off. A new fork starts with it off. Retry and prompt edits preserve the choice when the runtime and model stay unchanged.

Normal agent turns use the selected mode. Compaction and branch summaries remain Standard.

On Home, the choice stays staged until Leyline creates the session. Leyline applies it after the staged model and thinking level, before the first prompt.

Inspect enabled tools

Select the tool-count control, such as 12 tools. The Enabled tools list shows each active tool name.

The control can show 0 tools. The list is read-only.

Read context usage

The context indicator shows used tokens and the model context limit. Its bar changes to a warning state at 80 percent and a danger state at 95 percent.

Point to the indicator to see the percentage. Before a response reports usage, the indicator states that usage is unknown.

The context control is an indicator. It does not open a popover.

Inspect recorded usage

Select Session traces in the header to inspect timing, tokens, and costs for a captured run or operation. These records are separate from the composer's current context usage. Available values depend on provider usage reports. See Tracing.

Stage controls for a new session

The start screen loads runtime choices for the selected project. Model, thinking, and ultrafast choices remain staged until you create the session.

Leyline applies the staged model first, then the thinking level and ultrafast choice, before the first prompt.