Skip to content
Back to skills

Mac Dispatch

ASecurity

Route work that needs the operator's Mac (local scripts, signed-in apps, launchd jobs, Mac-only CLIs) through the existing dispatch: a GitHub issue whose body contains `<!-- agent:ready -->`, or a new row in the operator's existing herdr Google Sheet. Use whenever a cloud thread can't do a step itself because it needs the Mac. Never ask the operator to run it, and never create a new sheet.

  • 31 stars
  • 0 votes
  • 0 copies
  • 0 views
  • Added November 18, 2026
ai-agentsgobashgitapi

Works with

  • claude code
  • cli
  • api

Security analysis

A100/100

Scanned November 18, 2026

npx -y skills add IgorGanapolsky/ThumbGate --skill mac-dispatch --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Mac Dispatch?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Mac Dispatch
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/igorganapolsky-mac-dispatch/badge)](https://www.skillsdirectory.com/skills/igorganapolsky-mac-dispatch)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: mac-dispatch
description: >
  Route work that needs the operator's Mac (local scripts, signed-in apps,
  launchd jobs, Mac-only CLIs) through the existing dispatch: a
  GitHub issue whose body contains `<!-- agent:ready -->`, or a new row in the operator's existing herdr
  Google Sheet. Use whenever a cloud thread can't do a step itself because it
  needs the Mac. Never ask the operator to run it, and never create a new sheet.
---

# Mac Dispatch — Queue Mac Work, Don't Hand It To The Operator

The Mac is wired in. Cloud threads queue work for it; the Mac picks it up and
reports back. The operator should never be asked to "run this on your Mac".

## 0. Is it really Mac work?

Try the cloud routes first (`blocked-action-fallback` steps 1-4). It is Mac
work only when it needs local credentials, signed-in apps or browsers, local
files, launchd, or a CLI that exists only on the Mac.

## 1. Pick the queue (both are live; don't ask which is on)

| Queue | Use for | How |
|-------|---------|-----|
| **GitHub issue with `<!-- agent:ready -->`** (default) | anything a `claude -p` run in a repo folder can do | see step 2 |
| **herdr Google Sheet** (the operator's existing sheet; URL in project memory) | tasks the operator's herdr panes already pull from the sheet | append one row via the Google Sheets connector, same columns as existing rows |

Rules for the sheet:

- **NEVER create another spreadsheet**, tab, or copy. Append rows to the
  existing sheet only. If the sheet URL isn't in memory, use the GitHub issue
  queue instead of making a sheet.
- Read the header row and 2-3 existing rows first; match their columns,
  status values and wording exactly.
- Don't edit or delete other rows. Only update a row you added.

## 2. Queue a GitHub issue

Requirements (from `docs/guides/mac-dispatch-loop.md`):

- Repo `IgorGanapolsky/ThumbGate`. The body must contain `<!-- agent:ready -->`. Do not add a label.
- The issue must be authored by the operator's account (the allow list); the
  Mac loop ignores issues from anyone else. Create it with the session's
  GitHub tools, which act as the operator.
- The body is the whole prompt: goal, exact steps or commands, acceptance
  criteria, and what evidence to post back.
- Optional `### Working directory` section; default `~/workspace/ThumbGate`,
  must be under `~/workspace`.

```bash
gh api repos/IgorGanapolsky/ThumbGate/issues -f title="[Agent] <task>" \
  -f body=$'<!-- agent:ready -->\n\n<prompt>\n\n### Working directory\n\n~/workspace/ThumbGate\n\n### Acceptance criteria\n\n- <evidence to post>'
```

Search open issues for `<!-- agent:ready -->` first, so the same task isn't queued twice.

## 3. Write a prompt the Mac can finish unattended

- The run uses the operator's own Claude Code permissions and cannot answer
  prompts. Anything not already allowed is denied. Don't ask it to bypass
  permissions, and never add `--dangerously-skip-permissions` or
  `bypassPermissions` (the loop refuses them anyway; see `safety-boundaries`).
- One task per issue. Include the do-not-contact check for any outreach
  (`outreach-do-not-contact`).
- Ask for `STATUS: DONE` or `STATUS: BLOCKED` with receipts
  (`receipts-or-blocked`).

## 4. Follow through

- Record the issue URL (or sheet row) as the receipt for "queued".
- Results come back as an evidence comment. A finished run has `<!-- agent:done -->`
  in the body. A run that could not finish has `<!-- agent:blocked -->` in the body.
  Read the comment before claiming anything; the marker alone is not proof.
- `<!-- agent:blocked -->` on a permission denial: report it once as BLOCKED with the
  denied tool. Adding allow rules is the operator's call; don't add them
  yourself.
- Requeue a failed run by putting `<!-- agent:ready -->` back in the body and
  removing `<!-- agent:done -->`, `<!-- agent:blocked -->`, and `<!-- agent:claimed -->`,
  at most once, after fixing the prompt. Do not add a label.
- Queued is not done. Report "queued on the Mac: <link>" until the evidence
  comment arrives.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

View all in ai-agents →

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.