Setup check

The engine is setup-check.ps1 in the claude-dotfiles clone; this skill runs it and does the one
thing it cannot: change desktop routines, which only a Claude session's scheduled-tasks tools can
reach. Terms (local build, setup check, anchor PC) are in that repo's CONTEXT.md.

Build every path from $env:USERPROFILE and quote it: the plugin is built on one PC and runs on
all of them. Commands below are for the PowerShell tool.

1. Find the clone

$d = Join-Path $env:USERPROFILE 'Claude\Projects\Meta\claude-dotfiles'
if (-not (Test-Path -LiteralPath (Join-Path $d '.git'))) { gh repo clone surreptakos/claude-dotfiles "$d" }
git -C "$d" pull --ff-only

A failed clone ends the run: report gh's error and the bare-PC steps in the clone's README
(gh auth login first). The anchor name is read from this checkout, so a failed pull, or a clone
not on master, is reported as "anchor name may be stale" and the run continues on what is there.

Done when $d\setup-check.ps1 exists.

2. Run the engine with fixes

powershell.exe -NoProfile -ExecutionPolicy Bypass -File "$d\setup-check.ps1" -Fix

Windows PowerShell 5.1, always, even where pwsh exists. Keep the whole report. Exit 0 is clean, 1
means a STOP remains (normal on a new PC; carry on), 2 means the check could not run: quote its
last line and stop.

Done when you hold the report and its exit code.

3. Act on the routine findings

Read two lines under Projects: <PC> is the anchor PC or <PC> is not the anchor PC, and the
desktop-routines line. The routine ids are the comma-separated list after its colon. An ok
routine line, a not audited line or a -- line means nothing to act on; report it and skip to
step 4.

Run this from a session signed in to the routines' owner: the tools write to the signed-in
account, and the audit counts a routine only under its owner (~/.claude/accounts.json).

Routines change through create_scheduled_task and update_scheduled_task alone. A routine that
looks wrong becomes an owner to-do: never delete a routine or edit a registry file.

Anchor PC - N of M desktop routines not registered here: <ids>. Call list_scheduled_tasks
once, then for each id:

$r = Join-Path $env:USERPROFILE 'Claude\Projects\Meta\aac-routines'
py -3 "$r\scripts\desktop_routines.py" show <id>

• Not in the task list: call create_scheduled_task with the output's create_scheduled_task
  object as printed, except that a C:\Users\<name> home in its prompt becomes this PC's
  $env:USERPROFILE value. Then, if registry.enabled is false, update_scheduled_task with
  enabled: false.
• In the list but disabled while registry.enabled is true: update_scheduled_task with
  enabled: true.
• Otherwise leave it.
• The tool takes no cwd, model, permissionMode or useWorktree. When registry holds any,
  add one owner step: set them on that routine in the app's routine settings, with cwd under
  this PC's %USERPROFILE%.

Any other PC - N desktop routines live on a PC that is not the anchor: <ids>. For each id,
update_scheduled_task with taskId: <id> and enabled: false.

Then re-run the engine without -Fix (same command, flag dropped). Done when its routine line is
ok, or names only ids whose registry.enabled is false (the audit counts enabled routines
only, so a routine the store keeps disabled always reads as not registered). Any other id still
named is a finding to report with the tool's error.

4. Report

Report from the last engine run: the step 3 re-run when there was one, else step 2's. Drop its
Run /setup-check ... to-do item: this run is that. Print, in order:

1. The PC and whether it is the anchor, quoting the engine's line.
2. What the routine step did, one line per id (created, disabled, enabled, left), or "nothing".
3. The Owner to-do: the engine's numbered items, then the steps from step 3, numbered on
   through. One step per item, its exact command under it in a code block; an item the engine gave
   no command keeps its sentence alone.
4. The engine's last line (<n> STOP, <n> !!) and its exit code.