Skip to content

ADR-0011: Agent permission profiles and no generic shell tool

  • Status: Accepted
  • Date: 2026-07-15

Context

Agent clients need useful project operations, but prompt content and model output cannot be trusted as an authorization boundary.

Decision

Enforce application-level observe, develop, maintain, and admin profiles scoped by provider and project. Default to observe. Annotate MCP tools with accurate risk/idempotency metadata, audit every agent mutation, and keep destructive tools disabled by default. Never expose a generic unrestricted shell MCP tool; agents may invoke only approved typed project actions.

Consequences

Permissions remain effective independent of prompts or client behavior. Providers receive less ambient power and users can review agent activity. Adding a tool requires a use case, risk classification, bounded contract, and permission tests.

Use this pageLast verified Jul 17, 2026
Copy-ready MarkdownOpen sourceReport a docs issue