Namespace-based Access Control & RBAC

[DarlingtonDeveloper]

Overview

Multi-agent isolation via namespaces. Deferred — all agents currently see the full graph.

Scope

• Namespace model: nodes get a namespace field, queries filter by agent's allowed namespaces
• Three modes: open (default, current), namespace, rbac
• Namespace inheritance (agent in "kai" inheriting "shared" can read both)
• Write vs read permissions per namespace
• gRPC x-cortex-agent-id header for namespace resolution

Config

[access]
mode = "namespace"

[[access.namespaces]]
name = "kai"
agents = ["kai", "dutybound"]
inherit = ["shared"]

[[access.namespaces]]
name = "shared"
agents = ["*"]
write = ["kai"]

Why Deferred

Currently one graph with cooperative agents — isolation is premature. Needed when external users run multi-tenant setups.

Spec

Was part of specs/07e-access-control.md (stripped out). Original design in specs/07-standalone-product.md.

Priority

Post-launch. Required for multi-tenant and Cortex Cloud.