Approvals and Command Safety for Rocky

Command approval is not red tape; it is how a team decides which actions Rocky can take autonomously and which actions need a human checkpoint.

Official Hermes Agent docs: https://hermes-agent.nousresearch.com/docs/user-guide/configuration

Approval model

• Manual approvals are best for destructive commands, secret handling, billing, permissions, and unfamiliar infrastructure.
• Smart approvals can reduce friction for routine reads and safe build checks when the environment is trusted.
• Approval-off modes should be reserved for well-bounded automation where rollback and logs are understood.
• A scheduled job should be scoped tightly enough that it does not need ad-hoc risky decisions.

Safe command boundaries

• Prefer read-only inspection before file mutation.
• Avoid shell shortcuts that delete, reset, or overwrite broad paths unless the user explicitly asked for that scope.
• Use established deploy commands rather than improvising infrastructure changes.
• Never type or print secrets in chat, screenshots, or public artifacts.
• For public sites, deploy only after local checks pass.

Pitfalls

• Letting a cron job create or modify other cron jobs without explicit instruction.
• Using secrets as examples in docs or screenshots.
• Treating a failed command as permission to fabricate output.
• Running broad cleanup commands in a project root without an exact target.

Verification steps

• State which checks were actually run.
• Confirm changed files are within the intended boundary.
• For deployment work, verify the live URL after the deploy.
• If approval or credentials block the safe path, report that blocker directly.