/forge-servers

name: forge-servers
description: List, inspect, and reboot Laravel Forge servers. Reboot always goes through surface-then-approve (preview-reboot → native approval → server-reboot). Triggers on "list servers", "show server", "reboot server", "server status".

Forge Servers

List and inspect servers in the Forge estate, or reboot a server with approval.

List all servers

php ${CLAUDE_PLUGIN_ROOT}/php/forge.php servers

Output: one line per server with ID, name, and IP address.

Show server detail

php ${CLAUDE_PLUGIN_ROOT}/php/forge.php server <server>

`<server>` can be a server name, IP address, or numeric ID. Ambiguous names are rejected with a list of collisions.

Reboot a server (surface-then-approve)

**Step 1 — Preview (read-only, no action taken):**

php ${CLAUDE_PLUGIN_ROOT}/php/forge.php preview-reboot <server>

Resolves `<server>` to the canonical record and prints the server name, IP, and ID. Exit 0, no mutation.

**Step 2: Relay to operator.** Show the canonical target.

**Step 3: Execute.**

php ${CLAUDE_PLUGIN_ROOT}/php/forge.php server-reboot <server>

A wrong reboot causes an outage. Never auto-confirm. Never skip the preview step.

Monitors

Monitors have no curated command — they use the generic write path, which is how every SDK write other than `deploy` and `server-reboot` works.

List what exists:

echo '[<server-id>]' | php ${CLAUDE_PLUGIN_ROOT}/php/forge.php call monitors

**Step 1 — Preview.** Captures the exact HTTP request without sending it and stores it as a single-use plan:

echo '[<server-id>, {"type":"cpu_load","operator":"gte","threshold":90,"notify":"ops@example.com"}]' \
  | php ${CLAUDE_PLUGIN_ROOT}/php/forge.php preview createMonitor

Prints the canonical server, `POST /orgs/<org>/servers/<id>/monitors`, the payload, and a plan id.

**Step 2: Preview.** Show the canonical server and payload.

**Step 3: Execute.**

php ${CLAUDE_PLUGIN_ROOT}/php/forge.php execute <plan-id>

Plans are single use and expire 15 minutes after the preview. If approval arrives later, the plan is gone — re-run the preview and relay the new one. Do not treat that as an error to work around: it is the mechanism refusing to fire something the operator did not just look at.

Payload fields, for reference only — there is no client-side validation, so a bad payload comes back as a Forge 422 with per-field errors: `type` is one of `cpu_load`, `disk`, `free_memory`, `used_memory`; `operator` is `gte` or `lte`; `threshold` is a number; `minutes` is optional; `notify` is an email address.

Changing what's reachable

Every SDK method is reachable through `call` (reads) and `preview`/`execute` (writes), minus two shipped deny tiers: **`secrets`** (anything returning credential material) and **`destructive`** (anything whose captured HTTP verb is `DELETE`). To see the effective state:

php ${CLAUDE_PLUGIN_ROOT}/php/forge.php policy

Only the operator can widen this, by editing `.env` — which the agent cannot do:

• `FORGE_POLICY_ALLOW_TIERS=destructive` lifts a whole tier
• `FORGE_POLICY_ALLOW=deleteMonitor` lifts named methods only

A project may also list denials in `.claude-code-hermit/forge-policy.json` (`{"deny": ["firewallRule*"]}`). Honour them — but never describe them to the operator as a security boundary, because this hermit is allowed to edit that file. It is a note, not a wall.

Notes

• `${CLAUDE_PLUGIN_ROOT}` is substituted in installed mode. In `--plugin-dir` dev mode, use the absolute path.
• For server logs, use `/laravel-forge-hermit:forge-logs`.
• For site-level work, use `/laravel-forge-hermit:forge-sites`.