AI Agents · Architecture Notes

Architecture diagram: three entrances go through a single CLI to operate per-worktree environments, with hooks managing the lifecycle

Disposable isolated dev environments with Claude Code worktree hooks

git worktree only isolates code. Spawn a dedicated DB, dynamic ports, and a dev server per worktree, and let WorktreeCreate / SessionStart / SessionEnd hooks manage the lifecycle automatically. Design principles that settled after running this on three projects.

2026-08-10

Architecture diagram: a session declares scope and permissions, handling multiple repositories with distinct read / write-local / write-push permissions

Enforce with config, not attention — scope and permission guardrails for coding agents

Enforcing multi-repository AI agent work through declared scopes and permissions: per-session guardrail design, a self-restraint that keeps the agent away from its own permission definitions, and a double lock on push.

2026-06-30