Data model

Summary of the schema notes in CLAUDE.md. The full design, with the reasons for each choice, is NEW_DATABASE_STRUCTURE.md (written in Italian).

Group	Tables
Tenant	orgs, orgs_members, orgs_invites, profiles, api_keys
Canvas	projects, canvases, nodes, nodes_connections, node_runs, assets, canvas_events
Agents	chat_threads, chat_messages, ai_calls

Tables of the earlier social product (brands, products, social_accounts, posts, post_sources, scheduled_posts, the ad_* tables, social_posts, influencers) still exist in the schema but are not part of the canvas product and are not documented here.

Constraints worth knowing

• org_id is on every org table; row-level security uses auth_org_ids(). The service-role key bypasses it, so every use is declared in src/lib/server/db/service-role-uses.ts.
• nodes.data is jsonb, shaped per node type.
• nodes.version: position is last-write-wins; data uses optimistic concurrency (where version = expected; zero rows is a conflict).
• nodes.deleted_at: soft delete; every read filters deleted_at is null.
• actor_kind, actor_id, agent_key wherever an action is recorded. For an agent, actor_id is the user it acts for.
• nodes_connections.target_handle: a video with start and end frames has two inputs.
• Realtime publishes nodes, nodes_connections, canvases, node_runs with replica identity full.
• Types are generated (npm run db:types), never hand-written.