AI engagement

The mental model: an AI is a room member that you can @mention, give a power level, kick, redact, and instruct. What makes it different from a human member is when it speaks and how much agency you give it. Those are the engagement settings.

Engagement modes#

A room's engagement mode controls when the AI speaks spontaneously:

What happens when you invite an AI to a new room. Group rooms seed always by default; 1-on-1 rooms (DM with AI, sidekick, AI session) always seed always regardless of preference. Change the group-room default for your account under Settings → AI → When you invite an AI to a new group room.

Change a single room's mode any time with PUT /v1/rooms/{room_id}/ai/engagement or via the Engagement menu in the chat header.

Response modes#

How the AI's reply lands in the timeline:

• streaming — tokens arrive live, building one message in place. Best for short replies and code blocks. The whole reply lives as a single message in the timeline.
• conversational — the AI's reply is buffered and flushed as separate messages on paragraph boundaries. Best for long-form replies; feels like a person typing follow-ups.

Response mode is per-(room × AI) so a single room with two AIs can have them on different modes. Change via the toggle in the chat header or with PUT /v1/rooms/{room_id}/ai/response-mode.

System prompt overrides#

Every AI has a default system prompt configured by an admin. In a specific room you can override it with PUT /v1/rooms/{room_id}/ai/system-prompt or by editing the System field under Engagement → Settings. The override is additive: room-level instructions are appended to the AI's base prompt for that room only.

A common pattern is a base "you are a friendly assistant" prompt plus a room-level "you are debugging a payment integration; assume Stripe context" override.

Plugins#

The AI can call plugins to do work it can't do with words alone. The standard set is:

• web_search — query the live web (SearXNG-backed).
• web — fetch a specific URL.
• scaiwave.notes — find, read, create, update notes.
• scaiwave.todos — manage cross-workspace todos.
• scaiwave.drive — search ScaiDrive (file storage).
• scaiwave.knowledge — RAG against tenant-private knowledge.
• scaiwave.calculator — exact arithmetic.
• scaiwave.calendar — calendar lookups.
• scaiwave.memory — long-term memory across conversations.
• scaiwave.scailink — delegate to specialist tools registered by other ScaiLabs services.
• scaiwave.workspace_bridge — write into a workspace-bridged external service.
• scaiwave.skills — skill library invocation.

Each plugin is enabled or disabled per-tenant by an admin. Per-room overrides are possible (POST /v1/rooms/{room_id}/plugins/{plugin_id}/toggle).

The AI sees plugin tools in its function-calling schema; it decides when to use them based on the question. The harness rules around how the AI handles tool failures (e.g. trying a fallback before giving up) live in the planner and tool-etiquette system prompt.

Sampling parameters#

Per AI: temperature, top_p, max_tokens, frequency_penalty, presence_penalty, and (for reasoning models) reasoning_effort. Tune from Engagement → Sampling or via the API at PUT /v1/rooms/{room_id}/ai/temperature etc. Per-room overrides stack on top of the AI's defaults; the prompt studio shows you exactly what's being sent.

Stopping an AI mid-response#

• Click Stop in the chat header.
• Send another message — engagement cancels the in-flight generation.
• POST /v1/rooms/{room_id}/ai/stop.

Where to go next#

• Sidekicks — delegate long tasks.
• Planner — multi-step plans.
• Plugins and tools — the AI's toolbox.
• Tutorial: Tune AI response mode.
• Tutorial: Use the prompt studio.
• API: AI.