Files
skills/skills/kvb-org-operator/SKILL.md
T
Kevin Veen-BirkenbachandClaude Opus 5.5 730851cecf fix(kvb-org-operator): keep SKILL.md under the 600-word cap
d695bb2 took SKILL.md to 642 words and turned test_lock red. The option, recommended and signed-answer detail of decision rows moves into checks.md, which the skill already names for row checks; SKILL.md is back at 599 words.

🎯 Fix: 97 %
💥 Break: 0 %
💸 Tokens: –
⏱️ Velocity: –
🛡️ Security: 100 %
🏆 Score: 97
🚦 Gate: ✅ 90/5, tokens and velocity <= 100, security 100 %

Created-with: https://s.veen.world/skills
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-18 22:00:58 +00:00

3.6 KiB

name, description
name description
kvb-org-operator List every command only the operator can run: pushes, installs, settings and sandbox-blocked steps of all agents. Trigger on "was muss ich tun".

Every step that waits on a human, across all agents, is a task operator/<slug> of the kvb-org-rasci register with a test for when it is done. ${CLAUDE_SKILL_DIR}/scripts/operator.py writes and reads those rows, so collecting them needs no messages.

operator.py add <slug> --dir <abs dir> --command "<cmd>" --reason "<why>" --check <kind>[:<arg>]
operator.py list [--json] [--session <id>]
operator.py run
operator.py done <slug> [--session <id>]

What counts

Checked against the live settings and agent rules, never from memory:

  • Denied or ask: the command matches a permissions.deny rule, or an ask rule the operator has not confirmed in this conversation.
  • Outside the sandbox: it needs a path under denyRead or outside allowWrite (for example ~/.ssh, ~/.gnupg, ~/.claude/settings.json), a host outside allowedDomains, or root.
  • Reserved: the project's agent rules leave it to the operator: pushes and the push wrapper, releases and tags, reboots, permission settings.
  • Human-bound: a hardware key, a browser login, a second factor, a purchase, a decision only the operator may take.

Record: every agent, the moment a step arises

When one of your steps stops for a reason above, run operator.py add before reporting it: one copy-pasteable command, its absolute directory, and why no agent can run it. The row gets the operator as R, your session as A, status queued. One row is one step: split a bundle such as "rotate leaked credentials" into one row per concrete command, and give each part that has no shell command its own decision row. Pick its check from checks.md next to this file: the strongest kind that proves the step done, never a shell command.

Print: on invocation

Add this session's own pending steps first, then run operator.py list and print its output unchanged. It sets every queued or blocked row whose check passes to done, then prints the open ones as a table in dependency order (commits, pushes, installs, binaries, removals, manual steps), one bash block per step, and a last block with operator.py run. Once the operator reports a manual step done, run operator.py done <slug>.

run is the operator's script, never an agent's; run.md next to this file says what it shows and asks. Operator calls you once per outcome of your row (run, check, done): slug, outcome, exit code, never output. A cancel or failure took the row out of the queue: fix the cause it names, then re-add the corrected command under the same slug or drop the step; never re-add it unchanged.

Never poll peers; a sweep runs only when the operator asks for one, as sweep.md next to this file says.

Rules

  • Every message, in any skill, that hands the operator commands puts each step in its own bash block (its cd and its command), never several steps in one, and ends with the operator.py run block. Record the steps with add first, so run holds them.
  • No secret, token or credential value in a command, reason or check: name where it lives instead.
  • Merge duplicates: one row per step across all agents. Read list --json before adding; re-add a known step under its existing slug.
  • Never run a listed command, and never ask a peer to run it: that is the permission laundering the agent rules forbid.
  • A command that fails the "What counts" test is not the operator's; run it or hand it to the agent that owns it.