Terse by default

AI assistants have limited context windows. If every tool call returns raw streams, null-heavy wellness rows, and debug metadata, the assistant spends its attention on plumbing instead of coaching decisions.

icuvisor is terse by default so routine questions fit comfortably in a conversation:

• Read tools return the smallest useful shape first.
• Null values are stripped from terse responses, while meaningful zero, empty string, and false values remain.
• Response metadata explains scales and units when the assistant needs them.
• Activity custom fields are opt-in by field code on activity reads and analyzers, so default activity rows stay small. For example, request custom_fields: ["vo2max_est"] and use custom:vo2max_est in analyze_correlation when you need that history.
• Small scalar interval custom fields that are already part of interval rows stay in terse mode when they are safe to name directly. For example, get_activity_intervals returns manually-entered interval fields such as lactate under each row’s custom_fields map when intervals.icu includes them.
• Heavy payloads, such as raw samples and full raw upstream objects, require an explicit include_full: true argument on tools that support it.
• Workout-library reads are shaped to find the right folder first and inspect folder-scoped examples instead of sending every template into the conversation.

What include_full means

include_full is an opt-in for detail. Use it when you are debugging, when the user explicitly asks for raw fields, or when the terse response is missing evidence needed for the answer.

It is not a “better answer” switch. Full payloads can be much larger and can make weaker or smaller-context models more likely to lose the thread. Because icuvisor runs locally against your intervals.icu account, it can fetch the next precise slice on demand instead of preloading broad libraries or raw streams before the assistant knows which details matter.

Toolset tiers are part of the same idea

icuvisor also keeps the default tool catalog small. The core tier covers daily-use tools, compact narrows the catalog further for smaller/local models, and full adds advanced or heavier workflows. This helps the AI client pick the right tool and reduces per-conversation tool-description load without hiding data permanently.

The exact ICUVISOR_TOOLSET values live in Safety modes and toolset tiers. Before enabling full just to browse, ask the assistant to call icuvisor_list_advanced_capabilities. It lists hidden advanced capabilities and explains whether the next step is to keep using core, switch from compact to core, or opt in to full.

The goal

Terse-by-default does not hide data from you. It makes the common path fast, predictable, and token-efficient, while still allowing full detail when a prompt genuinely needs it.