activeing123/mcptoon v0.8.33

v0.8.33 — installed-server registry follows --dir/XDG; skill counts state their caliber

on GitHub

Fixed

• The installed-server registry now follows --dir, MCPTOON_HOME and XDG
  (issue #25, re-raised by @examosa). mcptoon install recorded the servers it
  had installed in ~/.mcp-cli/pro_installed.json, a path resolved from a
  module-level constant — so it ignored all three. Everything else in the state
  dir moved where the user asked and this one file did not, which meant a
  --dir X run wrote half its state inside X and half into $HOME. It is now
  resolved the way the rest of config.py resolves paths: MCPTOON_INSTALLED_FILE
  wins when set; an existing legacy ~/.mcp-cli/pro_installed.json is kept,
  never orphaned, so nobody's installed servers look uninstalled after upgrading;
  the default is STATE_DIR/pro_installed.json, beside handlers/ and
  toggles.json. installer no longer caches the path at import time — a
  constant cannot follow --dir, which is parsed after import — and uninstall
  reports the same path it cleans. 6 new tests run each case in a fresh
  interpreter under a fabricated $HOME, the only honest way to observe
  import-time path resolution.
• Every skill count now names what it counted (issue #40). Four surfaces
  answered "how many skills?" with four different numbers and none said which
  set it had measured, so each one looked wrong while every surface was
  internally consistent: 453 (top-level skills/<name>/ dirs), 467 (index
  entries minus alias cards), 483 (every SKILL.md under the source,
  recursively — 30 of them nested), 486 (distinct SKILL.md across the four
  agent view roots, the footer's number). The fix labels the caliber instead of
  collapsing the numbers: skills doctor prints a calibers: line reconciling
  all four, and its header says skill dirs rather than a bare skills;
  status says index entries (recursive, deduped, N aliases); skills index
  says indexed N entries (M alias cards, K skills); quickstart says
  indexed N entries routable. count_calibers() reads the footer's own
  sidecar for the view count rather than walking the roots again — a hand-rolled
  walk reported 958 against the footer's 486 here, because two of the four roots
  are whole-directory links to the source. A count it cannot take is omitted,
  never reported as 0. Reported by @activeing123.

Docs

• The wheel-size claim was six releases stale. Every surface that stated it
  said 534KB, which was accurate for 0.8.26 (547,084 B on PyPI) and had not
  been re-measured since; 0.8.32 ships 574,555 B, so the honest figure was
  ~561KB. Corrected on all nine surfaces plus the repository description, and
  534KB moved into RETIRED_KB so the old value stays pinned as a known-past
  number rather than a wrong current one.