Window census

The instrument is skills/app-control/scripts/window-census.sh. Its mechanics live in app-control § Window census: status vocabulary, file columns, and how each skin is isolated from the last. This skill is the user interaction around a run: what to confirm, how to run it, and what to report. Read that section once before the first run in a session.

1. Scope the request

Turn the request into the script's arguments:

Request	Arguments
named skins	aquamp corona "2222-cPro__Bento" (anything launch.sh takes)
one or more families	--all wal / --all classic,original,metal
the whole corpus	--all all
"with screenshots"	add --shots (a PNG per opened window, plus shots.zip)

Run window-census.sh --list <families> to get the count; it launches nothing. On 2026-09-27 the corpus was 307 skins: 26 classic, 14 Original, 7 Metal, 80 .wal, 180 .wmz. Budget about a minute per skin, so the full corpus takes 5–6 hours.

2. Confirm before starting (anything over a handful of skins)

State all of this in one message and wait for the user's go-ahead:

• Skin count and time estimate.
• It takes over the machine. The script clicks the menu bar and raises the app for the whole run. The user should not use the mouse or keyboard meanwhile, or clicks land in the wrong app.
• It quits any running debug build, including another session's. Name the pid if one is up (pgrep -lf 'debug/NullPlayer'). The installed app is never touched.
• Whether to clear the shared size keys first. Until ~/.claude/plans/skin-window-size-isolation.md lands, the debug domain holds hostedInteriorSize2.* sizes that every skin inherits. A true-defaults baseline needs them gone. Ask; don't delete on your own. If the user agrees: defaults read NullPlayer | grep -o '"hostedInteriorSize2\.[^"]*"' | tr -d '"' | while read -r k; do defaults delete NullPlayer "$k"; done, before the run, because the census saves the domain as its baseline at start.
• The output directory. Default /tmp/np-window-census/<timestamp>/. For a corpus run, pick a stable path such as /tmp/np-window-census/corpus-<date>, so an interrupted run resumes into it.

3. Run it

• Freeze the tree. Make no source edits during the run. The first launch builds, and an edit mid-build aborts it.
• More than ~10 skins: hand it to a worker (a Haiku subagent) with the exact command. It buys context, not time. The worker runs it in the foreground with the longest Bash timeout, never run_in_background, because a detached sweep is killed mid-build and still exits 0. If the timeout ends the call, re-run the same command with the same --out. Finished skins are skipped and the defaults baseline is reused, so repeat until every skin has a CENSUS line.
• Handle stdout lines as follows:
  • A CENSUS … launch failed line is recorded in skins.tsv; the run carries on.
  • app-exited means the app crashed during that skin. Note it and don't re-run it silently.

4. Report

The user wants findings, not a file dump. Read skins.tsv and windows.tsv (plain TSV with one header row, so awk -F'\t' or sqlite3 -tabs works) and answer with:

1. The largest default sizes. Per window item, rank the skins by w×h among role=item, status=opened rows, and say which family each extreme belongs to. Compare with the classic size (275×Skin.scaleFactor = 344 pt wide at 100%); the rule in the size-isolation plan is that every native window defaults to it.
2. Main window size per family: the main_w/main_h spread from skins.tsv.
3. Problems:
   • reachable=0 rows;
   • residue=1 rows, where a window didn't close back to the baseline;
   • launch failures and app-exited;
   • enabled items that read no-window: skin-drawn panels the census can't see, so name them rather than count them as broken.
4. The paths: report.md for reading, the TSVs for queries, and shots.zip if taken.

Keep the per-skin tables in report.md. The answer is the summary and the outliers.

Debugging a live defect

A census row is a measurement, not a diagnosis. To chase why a window opens where it does, read live-ui-testing and then app-control, and drive that one skin by hand. The reference workflow is winamp-modern-skin-guide/reference/harness.md § Debugging a live defect.