Select models and thinking levels
Defaults and conversation selection
Section titled “Defaults and conversation selection”Settings stores a global default model/thinking level and remembers the last selection. Each conversation can change its model from the composer. A change during a run applies to the next provider request.
A model scope narrows the picker. Empty scope means every authenticated model; a nonempty list filters it. Stale scope entries remain in Settings as unavailable and are ignored until the model returns or the entry is removed.
Sub-agents use the lead’s model and thinking level unless their profile names its own, which can suit work such as parallel codebase research. The choice is made when a sub-agent is created; afterwards you can change its model and thinking level in its own view like any conversation. If an existing child reaches usage limits, edit that child directly: changing its profile or the lead’s model will not update it, and follow-ups keep its current configuration. Scoped profile configuration adds conversation-local model/thinking overrides for future children without changing reusable profiles.
Catalog limits
Section titled “Catalog limits”Nerve combines built-in and custom definitions. Current runtime resolution exposes complete catalogs for selected provider families; other providers can be limited to the first eight models. A missing compatible model can be added manually.
The public model record includes reasoning/thinking support, context window, and maximum output. Custom definitions can hold richer cost and input-modality metadata, but not every field reaches the model picker.
Thinking levels
Section titled “Thinking levels”The full vocabulary is:
off · minimal · low · medium · high · xhigh · max
Concrete support comes from model metadata and provider mappings. xhigh and max are only meaningful for selected model families. General runtime resolution clamps unsupported requests. Sub-agent profile configuration instead blocks known model/thinking incompatibility at creation, with a diagnostic prompting a supported choice; it must not silently downgrade an explicit profile choice.
Choosing well
Section titled “Choosing well”Use a fast model for small, well-scoped changes and a deeper supported thinking level for architecture or difficult debugging. Context size is not a quality score. Review provider price/terms rather than relying only on Nerve’s cumulative cost display.