diff --git a/skills/agent-ops/scoped-privileged-command-key/SKILL.md b/skills/agent-ops/scoped-privileged-command-key/SKILL.md
new file mode 100644
index 0000000..4d7f248
--- /dev/null
+++ b/skills/agent-ops/scoped-privileged-command-key/SKILL.md
@@ -0,0 +1,89 @@
+---
+name: scoped-privileged-command-key
+description: "Give an agent a narrow, audited set of privileged host commands (e.g. Proxmox qm) via an SSH forced-command wrapper with an argument-by-argument allowlist — without handing over root. Two-gate model: CR approval AND wrapper both must pass."
+version: 1.0.0
+author: Claude Code (Agent Alpha, agent-ops session)
+license: MIT
+platforms: [linux]
+prerequisites:
+  env_vars: []
+  commands: [ssh, ssh-keygen, qm]
+metadata:
+  hermes:
+    tags: [agent-ops, SAIOP, security, least-privilege, proxmox, ssh, forced-command, wrapper]
+---
+
+# Scoped privileged-command key (SSH forced-command allowlist)
+
+## When to use
+
+An agent needs to run a **small, fixed set** of privileged commands on a host
+(e.g. clone/reconfigure/start one specific VM) but must NOT get general root.
+The approval gate decides *whether*; this pattern bounds *what the transport can
+physically do*, independent of the approval gate. Both gates must pass.
+(Built for `agentops-cr@` on pve1/pve2 — CR-012 T10 / CR-015.)
+
+## The pattern
+
+A dedicated SSH key whose `authorized_keys` entry pins a **forced command** and
+strips everything else:
+
+    command="/usr/local/sbin/cr-exec-wrapper",no-port-forwarding,no-agent-forwarding,no-X11-forwarding,no-pty <pubkey>
+
+The wrapper is the security boundary. It reads the request from
+`$SSH_ORIGINAL_COMMAND`, **never `eval`s it**, and permits ONLY an explicit
+allowlist of command *shapes*, validated argument-by-argument. Fail-closed;
+every call (allowed and refused) is logged.
+
+## Procedure
+
+1. Generate a dedicated keypair; private key 0600 root on the caller
+   (`/etc/agent-ops/keys/<name>`), never on a laptop, never in git.
+2. Write the wrapper (root:root 0755). It MUST:
+   - reject any char outside a strict whitelist before parsing —
+     `case "$cmd" in *[!A-Za-z0-9\ ._:=,/+-]*) deny ;; esac` — this kills
+     injection (`; | & $ backtick > < \ newline`);
+   - tokenise with `read -r -a`, then match **exactly one** permitted shape,
+     validating every arg against fixed allowlists (target IDs, storage, size
+     ceilings, fixed values);
+   - be fail-closed (unmatched → refuse, non-zero exit);
+   - log every call with a UTC timestamp to an append-only log;
+   - on match, `exec <realcmd> "${a[@]:1}"` (array — no re-splitting, no eval).
+3. Pin the forced command in `authorized_keys` on the target host(s) only
+   (exclude unstable/quorum-only nodes, e.g. pve3).
+4. Keep an explicit ID/target allowlist (e.g. one VMID). Adding an entry later
+   is itself a reviewed change.
+
+## Gotchas / guardrails
+
+1. **Never include destructive shapes** (`destroy`, `delete`, `--force`) in the
+   allowlist. Prefer a constrained `clone` from a fixed template over a
+   free-form create.
+2. **Parse arg-by-arg, never substring-match** the command — substring checks
+   are trivially bypassable.
+3. **Off-by-one arity is the common bug.** The target ID at `a[2]` shifts every
+   later index; count tokens carefully (`qm set 144 --scsi1 local-lvm:20` is 5
+   tokens, not 4). Bugs here fail *open* if you're not fail-closed.
+4. **The key is transport, not authority.** The CR-approval gate
+   (`cr_authorize.py`) still governs *whether* a shape may run. Both gates must
+   pass — the wrapper adds no authority the approval gate hasn't granted.
+
+## Verification
+
+Run the NEGATIVE set FIRST and confirm each is refused + logged, BEFORE trusting
+the key:
+- a destroy/delete shape → refused
+- a target ID outside the allowlist → refused
+- a wrong storage / oversize disk → refused
+- a shell metacharacter (`... ; rm -rf /tmp/x`) → refused (metachar)
+
+Only then run the POSITIVE allowed shapes and confirm they execute and log
+`ALLOWED`.
+
+## Provenance
+
+Learned building the ATU app-delivery platform (CR-015, unblocking CR-014
+webapp-01). `agentops-cr@` permits only: clone template 9000 → VMID 144, disk
+resize/add on local-lvm (≤ ceiling), ipconfig/ciuser/boot set, start, status —
+and refuses everything else, fail-closed, logged to `/var/log/cr-exec.log`.
+Author: Agent Alpha, 2026-08-14.