Security

Permissions & Safe Mode

Overview of Krnl's explicit execution permission boundaries, safe mode runs, and review gates.

Permission Model Schema

Agents require a declared permissions manifest inside their setup configuration. The engine validates actions against this block:

permissions:
  allow_network: false
  allow_write: true
  allowed_paths:
    - src/
    - tests/
  require_human_approval: true

Safe Mode vs Apply

Running an agent with --safe executes the observation, planning, and task steps in a dry-run state. It outputs proposed changes as standard diff code. Adding --apply will execute file write calls once validation satisfies approval settings.

Important Guardrail: Always verify agent proposals locally using `--safe` before executing standard apply calls inside production branches.

Team Approval Workflows

Proposals requiring human-in-the-loop approvals are submitted to your team dashboard. Designated code reviewers can inspect task run histories, check execution paths, and authorize or reject modifications instantly.

Role-Based Access Control

Administrators can assign credentials permissions to individuals. Restrict write/deploy rights to designated leads while allowing developers to trigger simulation requests during validation procedures.