AI Security & RBAC
Access Model
- Viewer/Developer/Admin can use AI chat features.
- Only Developer/Admin can confirm executable proposals.
Workspace Isolation
All AI entities are workspace-scoped:
- settings
- secrets
- MCP servers
- conversations/messages
- tool traces
- action proposals
Secret Handling
- Secrets are encrypted at rest using
AI_SECRETS_MASTER_KEY. - Raw secret values are never returned by APIs.
- Environment variable fallback is supported.
Proposal Confirmation
Executable SQL/dbt proposals are single-action records with status lifecycle:
pendingexecutedrejectedexpired
No proposal executes without explicit confirmation.
Auditing
AI interactions persist:
- prompts/responses
- tool traces
- proposal creation and resolution events