AI Cockpit

The AI Cockpit is the operator's surface for tuning the AI that runs across the app — guest reply suggestions, Pulse Copilot, task-template assistance, message analysis, accounting agents.

Open it from the V28 sidebar → AI Cockpit.

Sections

• Agents editor — per-specialist (accounting / messaging / operations / knowledge / property / general) prompt + model + thinking-budget + tool-rounds + plan-then-execute.
• Prompts editor — the existing llm_prompts editor for the named slugs (message_analysis, wa_chat_assistant, etc.).
• Cost dashboard — per-turn drill-down of accounting agent and Pulse Copilot runs.

What lives here vs Admin → LLM Prompts

The legacy Admin → LLM Prompts page (covered separately in admin/llm-prompts.md) edits the same llm_prompts table. The AI Cockpit Prompts tab is the V28-native version of that surface; the legacy page is still available and still works.

Going forward, all per-tenant AI configuration is being moved into the AI Cockpit.

Per-tenant from day one

Everything in the AI Cockpit is scoped to your tenant. Editing the accounting agent's system prompt in flatsbratislava does not change the accounting agent for another tenant. Each tenant tunes their own agents.

All edits go through an audit log — see who changed an agent's model or prompt and when, via Admin → Audit Log (see admin/activity-log.md).

Implements: gh#415 (Agents editor), gh#414 (cost dashboard turn-grouped drill-down), gh#412 (cost reduction underlying work).