Skip to content

Owner question: agent guards as a safety net, GitHub permissions as the hard limit #916

Open

Description

@scuttletoby

Recommendation: accept that the agent guard hooks (PR #847, fork-kit/agent-guards) are a safety net against mistakes, not a wall against a hostile agent. Rely on GitHub permissions for the hard limit.

Plan:

  1. fork-kit/agent-guards: the machine's agent guard hooks as a kit (#834 G4) #847 keeps the guards and widens them to catch the common ways an agent could start a second session with no hooks (by full program path, via npx, through python/node one-liners, and so on). The README and PR say plainly what they do not stop.
  2. The real limit is what the agent's GitHub login is allowed to do. Later step (own issue, after fork-kit/agent-guards: the machine's agent guard hooks as a kit (#834 G4) #847): agents use a login that cannot disable workflows or change rulesets at all, so no command, however it is written, can do it. Only the owner's own login can.
  3. The owner's approval to disable a workflow becomes tied to one exact command and is used up once, so a code that leaks into logs is worthless.

Why: the second review of #847 found many one-line ways around the "no hookless session" check. Checking command text can never catch every form; each fix only closes the forms someone thought of. Permissions on the login cannot be talked around.

What changes for you: nothing now. Step 2 later may mean one setup step for the agents' GitHub login (I will prepare it).

Accept or reject by replying "Owner decision: ..."

🤖 Generated with Claude Code

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

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions