v0.8.24 — statusline wiring, savings milestones, honest mode costs

[0.8.24] - 2026-10-20

Wiring the badge into a host, and the defects an adversarial pass over that code
found. Each fix was reproduced first and is pinned by a test.

Added

• mcptoon statusline --install / --uninstall / --check wires the badge into
  Claude Code. The one-line badge existed since 0.8.22 but nothing put it where a
  user would see it, so "I installed mcptoon and cannot tell it is doing anything"
  stayed true. --install points statusLine.command in ~/.claude/settings.json
  at mcptoon statusline; --uninstall removes exactly that entry; --check
  reports the state and exits non-zero when not wired, so doctor and scripts can
  gate on it. Idempotent, refuses to overwrite a different statusLine, and never
  touches any other key.
• Every host tier now carries a "mcptoon is doing this" signal, not just the
  savings number. A user who sees the gateway forward a call but never sees the
  saving has a tool they cannot justify keeping — the worry is not "I cannot see
  it", it is "I see it forwarding and cannot see it saving, so I will delete it".
  Three surfaces close that gap, one per host tier: (1) the status-line badge
  (Claude Code's status bar) already shows calls and tokens together; (2) the
  pointer block written into AGENTS.md-style files (Codex, Gemini CLI, Qwen
  Code, Zed, Crush, opencode, Claude Code) now also states that the saving is
  definition-side and tells the model to run mcptoon statusline for the live
  receipt — it used to talk about skills only; (3) the serve-leg instructions
  now asks the model for the per-turn mcptoon statusline presence receipt
  (calls, tokens saved, mode) whenever the turn routed an upstream call, alongside
  the unchanged savings report. All three follow config.resolve_lang(), so the
  text is Chinese on a Chinese machine and English otherwise — the same chain
  (explicit lang setting > MCPTOON_LANG > OS UI language > LANG > English)
  every other surface already uses. The pointer block's heading stays English on
  purpose: it is the anchor _strip_skill_pointer and the idempotency check key
  on, and the strip path recognises both language bodies, so switching language
  never leaves an un-strippable block behind (pinned byte-for-byte).
• uninstall now says what you LOSE, not only what files vanish. It deleted the
  whole cache directory — which holds usage.json and activity.jsonl, the running
  tally a user watched climb for weeks — while printing only a list of paths. The
  plan now ends with "You will lose: your savings history — N routed call(s) over D
  day(s) since ", and --keep-history clears the rebuildable schema cache but
  keeps the two history files, so the number survives the uninstall.
• mcptoon savings — milestones a person can feel. report answers "how much";
  this answers "was it worth keeping". It shows the token milestones crossed (1M …
  1B …), the same figure in human terms (≈ N novels / books / copies of the Bible),
  and the next milestone with the distance to go. Every number is derived from
  footer._cumulative_estimate — the same estimate the footer and report use, never
  a second drifting one — and labelled as the lower bound it is.
• mcptoon statusline --guide wires the badge into every host, not just Claude
  Code. Only Claude Code has a documented status-bar schema, so --install wires
  one host; --guide prints the exact one-liners for tmux (status-right), bash/zsh
  (PS1), fish, and the "run it in a side terminal" pattern for hosts with no status
  bar at all. Bilingual, and honest that the command is identical everywhere.
• doctor now reports the status line. It tested every server and config path
  but never asked whether the badge was actually visible — so a healthy-but-invisible
  install read as "all good". The new line says wired / not-wired / no-Claude-Code,
  teaches statusline --install when the host is present, and is informational only
  (never bumps the issue count or the exit code).
• welcome and quickstart point at the badge. Both explained the saving and
  pointed at mcptoon status; neither said how to keep it visible. The welcome card
  gained a "See it / 看得见" row (only when servers exist — nothing to watch
  otherwise), and the quickstart celebration gained a statusline --install line.
• sync --self --dry names the roster it scanned. It used to hide every
  only_when_present host that was absent, so a user could not tell "scanned 12,
  found 3" from "only knows about 3". The preview now ends with "Scanned N known
  host(s) — M found on this machine" and names the absent ones, so "not found" for a
  host the user DOES use is a signal to check a path.
• report dates the lifetime total. A count of 2,920 says nothing about how long
  the user has trusted the tool. track_call now stamps a first_ts on the first
  call (additive; legacy files fall back to the oldest windowed call), and the
  cumulative section prints "since · N day(s) · ~M calls/day".
• scripts/cleanroom_check.sh — prove the SHIPPED artifact works. The existing
  cleanroom runs the lifecycle suite against a source install on 3.12 only. This
  script builds the real wheel, installs THAT into a clean container (no repo on
  sys.path), and runs a small smoke suite across every Python in $PY_VERSIONS
  (default 3.10 + 3.12) — catching wheel-packing drift and a 3.10 regression that an
  editable install on the dev's 3.12 cannot see.

Fixed

• sync_claims.py now owns the line- and module-count claims, not just test
  totals. After a refactor added a module, the docs still said "31,410 lines … 45
  modules" while the tree was 31,979 / 46; the module count is compared exactly, and
  nothing auto-wrote either, so they only moved when someone hand-edited five files.
  The script now measures src/mcptoon/**/*.py with the guard's own ruler and writes
  both — numbers only, wording untouched (respects the README freeze).

• mcptoon statusline no longer prints the savings footer to stderr. It is
  machine-consumed (a host reads its stdout as the status line), so the footer was
  both a wasted catalog walk on every prompt and a corrupting extra stream. It
  joins footer-facts/report in the footer-silent set.

• The status-line badge was hard-coded English. On a machine configured for
  Chinese it still printed tools / saved / calls, the one surface a user stares
  at all day. It now follows resolve_lang(), and the 80-column guard measures
  display columns (a CJK glyph is two) rather than len(), so neither language
  overflows.

• mcptoon statusline and mcptoon footer-facts no longer leak the first-run
  banner into their output. On a fresh machine the 16-line welcome card (and, on
  the first command, the self-heal's "skill installed" lines) were printed to stdout
  ahead of the badge. A host renders stdout as the status line, so the bar came out
  as box-drawing; and footer-facts is the command serve tells the model to paste
  verbatim, so the banner went into the chat too. Both commands are now exempt from
  the welcome and self-heal, leaving statusline at one line and footer-facts at
  three.

• statusline --install / --uninstall is now genuinely byte-reversible. It
  re-serialised the whole file with json.dumps, so a 4-space file came back
  2-space, ["Bash"] was exploded one-per-line, and a minified file was re-indented
  — the "we only added one key" promise was silently false. Removal now cuts the
  exact text span that was spliced in back out of the original bytes, so a round
  trip restores the file character for character across 2-space, 4-space, tab,
  minified, no-trailing-newline, CRLF, nested-same-key and CJK layouts. A file that
  existed as {} before we filled it is restored, not deleted.

• A settings.json that is a directory is a clean error, not a traceback.

• The statusline messages follow the configured language. They were English-only
  on a project where every other user-facing command is bilingual.

• statusline and footer-facts no longer re-tokenize the whole catalog on every
  prompt. Each call recomputed the full/slim token totals for every cached schema —
  ~440 ms on a 209-tool catalog, paid on every prompt because statusline renders
  each turn and footer-facts is meant to be run each turn. The totals are now
  memoised on a stat-only fingerprint (in memory and in a sidecar), so an unchanged
  catalog costs a stat walk. Measured: a warm facts() dropped from ~430 ms to
  ~22 ms. The sidecar never recreates the cache directory that mcptoon uninstall
  deletes.

• skills list --usage no longer shows yesterday's date before 08:00. The
  "last used" stamp was written with datetime.now(timezone.utc) while every other
  surface (usage.py, the activity summary) reads the local calendar day. On a
  UTC+8 machine that made the two disagree for the first eight hours of every day,
  so the column read as yesterday's date in the morning. It now uses the same local
  datetime.now() as the rest of the codebase.