mikehasa/golive-skill

Drafted from the split we settled on here, and opened as a **draft PR** for review before implementation is considered complete: https://github.com/hashgraph-online/hol-guard/pull/3199

What it contains:

- Three rules, one permission each, all `default_mode: review`: `golive apply` with `--confirm-destroy`, `--confirm-live` and `--confirm-dns`. The confirmation flags are the match condition (they are deliberately not declared as generic interspersed flags, so the required-flag check still sees them).
- Safe variants per rule: `apply --help`, and a read-only set naming `help`, `version`, `update-check`, `detect`, `menu`, `doctor`, `plan`, `teardown`, `verify`, `status`, `handoff`. `teardown` is the read-only inverse plan and `plan` prints the plan id, so they stay outside the destructive boundary; `golive apply` without a confirmation flag is outside these rules too.
- The description states what GoLive already enforces (approved plan id recomputed at apply time, first production deplo...

Thanks for clarifying. The rule should cover apply with the destructive confirmation flags, not treat teardown as deletion. The independent approval boundary is exactly the use case we are interested in. Would you be willing to submit that extension using the builder? We can review the destructive/read-only split before merge. Guide: https://github.com/hashgraph-online/hol-guard/blob/main/docs/guard/extension-builder/README.md

Thanks — this fits, and yes, in principle.

One correction to build the rule around: golive teardown is the read-only half. It builds the inverse plan (only resources golive can prove it created — its own DNS records, recorded webhook endpoints, issued sending keys, and host projects whose creation marker matches), prints a plan id, and turns everything else — adopted projects, unreadable zones, resources without a delete capability — into manual handoffs. The deletion happens later, in golive apply --plan --yes --confirm-destroy [--confirm-dns] [--confirm-live]. So a rule that only reviews teardown reviews intent; to pause the deletion it needs to cover apply with the risk flags, with teardown as the intent signal and help / version / detect / menu / doctor / plan / status / handoff as safe counterparts. (credentials never takes a secret value in argv by design, so a generic secret-exposure rule should not fire on it.)

Where a Guard extension genuinely adds something is the ...