Use chat controls and live progress

Open Complexity. Beside model and reasoning selection, the context circle shows estimated request tokens and effective tenant/user limits. Cumulative billing input/output appears only with Advanced LLM mode, which defaults to on. It is not the current context size.

The runtime estimates with the model tokenizer. For tool overflow, it shows the largest contributing call IDs, names, and token estimates. Large email sets show a count and bounded ID list, without full bodies.

Seeded model context caps use 500000 tokens, bounded by the provider window. Explicit administrator overrides remain intact. Catalog upgrades apply this cap to inherited unlimited seeded models while preserving custom models and overrides. Sorter context has a separate scope.

Tool markers, primary shimmer, and thinking indicators reflect real streaming state. Tool preparation appears before execution. Chat follows global Island mode. History search matches word fragments in both titles and messages.

Open Settings → Appearance → AI chat to change newline mode, Advanced LLM mode, the AI pet, or its shape. The pet defaults to off for every user. Tenant defaults and personal overrides use the settings policy system and administration permissions.

The optional pet loads only when enabled and shows a working state. Run indicators show planning, thinking, shaping, and per-tool work. Theme and reduced-motion settings apply. These indicators do not prove mailbox provider confirmation. See streaming and keyboard controls.

The optional avatar uses bot-avatars@0.1.1. Progress indicators use thinking-orbs from the ^0.3.1 dependency range.