Skip to content

No agent ever runs uncaged: cage the read-only roles too #100

Description

@naashw

Founder rule, 2026-08-27: no AI agent process ever runs directly on the host, no exceptions. Today the dev agent is caged but the read-only roles (reviewer, prosecutor, judge, and ticket drafting) run on the host, bounded cooperatively by CLI flags and a temporary settings file (D15).

Target:

  • Reviewer/prosecutor/judge run in the same container cage as the dev turn, with the worktree mounted read-only (:ro): the read-only bound becomes structural (kernel-enforced) instead of cooperative, strictly stronger than D15.
  • Ticket drafting and any future agent role follow the same rule.
  • install.sh stops installing claude-code on the host once nothing host-side needs it (the cage bootstraps its own copy).
  • Direct (uncaged) isolation mode becomes an explicit opt-out for local development only, never a default anywhere.

Sequencing: after the H-dogfood e2e; pairs naturally with the orchestrator/micro-VM direction where everything moves into disposable VMs anyway.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions