fukuyori/wtmux 4.0.0

on GitHub

[4.0.0] - 2026-09-14

This release reworks how wtmux decides how many cells a character occupies,
and lets the pane ConPTY come from a modern OpenConsole instead of the inbox
conhost.exe. Together they remove the one-column drift that combining
marks, VS16 emoji (❤️ ✔️ ⚠️), ZWJ sequences (👨‍👩‍👧), flags (🇯🇵) and halfwidth
dakuten (ｶﾞ) caused in positioned TUI output. Design notes and the
measurement procedure are in docs/design-width-model.md.

Background (measured 2026-09-14, Windows 11 26200)

wtmux sits between two ConPTY hops: the pane's ConPTY feeds it, and the host
terminal's ConPTY consumes its output. Each layer has its own width model.
Cell counts measured with DSR/CPR (tools/width_probe.py):

Two consequences drove the changes below:

• The inbox conhost counts width per code point and re-lays out the
  application's absolute cursor moves with its own count, so a CSI 10 G
  after à, ❤️, 🇯 or a Hangul jamo arrived at column 9 in wtmux's
  grid. Neither wtmux nor the host terminal can correct that after the fact.
• wtmux's renderer re-anchors the cursor around every multi-byte cell. A
  grapheme cluster split over two cells (🇯 + 🇵, ｶ + ﾞ) was therefore
  emitted as two positioned writes, and WezTerm dropped the character that
  followed such a pair (z in 🇯🇵z, w in ｶﾞw;
  tools/host_cluster_probe.py reproduces this in a bare WezTerm pane).

Ambiguous-width characters agree across every layer and were not changed.

Added

• Bundled modern ConPTY (Windows). The pane ConPTY can now be created
  through conpty.dll + OpenConsole.exe from microsoft/terminal (MIT),
  the same pair Windows Terminal and WezTerm ship. src/core/pty/conpty_api.rs
  resolves CreatePseudoConsole / ResizePseudoConsole /
  ClosePseudoConsole once, looking for conpty.dll in WTMUX_CONPTY_DIR,
  then next to wtmux.exe, and falling back to kernel32 (WTMUX_CONPTY=system
  forces the fallback). wtmux --version prints
  ConPTY: bundled (<path>) or ConPTY: system (kernel32 / conhost).
  OpenConsole passes the application's escape sequences through unchanged
  and forwards DSR/DA queries to wtmux, which answers from its own grid
  (Response::CursorPosition), so an application inside the pane sees
  exactly wtmux's widths. Verified end to end: inside a wtmux hosted by
  Windows Terminal, a child measured ❤️=2, 👨‍👩‍👧=2, 🇯🇵=2, ｶﾞ=2 — identical to
  Windows Terminal's own layout — and the CSI 10 G cases above land on
  column 10. All existing ConPTY harnesses (CJK repaint, closed-loop render,
  host hop, kitty and win32-input round trips) pass with OpenConsole
  1.24.2607.10001 and with the inbox conhost.
• Packages bundle the pair. The portable ZIP, Inno Setup installer, WiX
  MSI and MSIX now install conpty.dll, OpenConsole.exe and
  LICENSE-ConPTY.txt next to wtmux.exe when they are present in
  vendor\conpty\; without them the packages are built as before.
  vendor/conpty/README.md documents the source (NuGet
  Microsoft.Windows.Console.ConPTY, also attached to every Windows Terminal
  GitHub release; x64 files at runtimes/win-x64/native/conpty.dll and
  build/native/runtimes/x64/OpenConsole.exe). The binaries are
  Microsoft-signed and ignored by git. The Windows App Certification Kit
  reports no finding attributable to the two files (the pre-existing
  optional "blocked executables" note about cmd.exe / powershell.exe
  references and the high-DPI warning are unchanged).
• Host VS16 policy measured at startup. Windows Terminal 1.24 draws a
  text-default emoji followed by U+FE0F (❤️) in two cells, WezTerm 1.22 in
  one. Right after entering the alternate screen, ui::host_probe writes
  ❤\u{FE0F}, reads the cursor back with DSR (500 ms deadline) and clears
  the line; the lead cell is widened to two cells only when the host
  answered 2. Without an answer (no tty, headless) the one-cell layout of
  unicode-width is kept. The set of affected characters is generated from
  UCD emoji-data.txt (Emoji=Yes, Emoji_Presentation=No; 119 ranges,
  Unicode 17.0) by tools/gen_emoji_table.py into
  src/core/term/emoji_text_default.rs. The renderer already strips VS16
  from one-cell cells and keeps it on two-cell cells, so both policies stay
  consistent with what the host draws.
• Diagnostic harnesses (cargo test <name> -- --ignored --nocapture):
  conpty_ambiguous_width_probe runs tools/width_probe.py inside a real
  ConPTY and diffs its cell counts against char_width (modes: amb,
  all, seq; it answers forwarded DSR like a host);
  conpty_width_divergence_dump shows the raw bytes and the resulting grid
  for text whose width the layers disagree on; render_emoji_line_dump
  prints what the multi-pane renderer emits for such a row.
  tools/compare_probe.py diffs two probe results.

Fixed

• ZWJ sequences (👨‍👩‍👧) occupied one cell pair per emoji (six cells)
  while every terminal draws them as one two-cell cluster; the character
  after a ZWJ now joins the previous grapheme, also at the right margin,
  where a terminal never wraps inside a cluster.
• Regional indicator pairs (🇯🇵) were two one-cell glyphs emitted with a
  cursor move in between; WezTerm then dropped the character that followed.
  The pair is now one two-cell cluster written in one go
  (TerminalState::extend_cluster_before_cursor). A lone regional
  indicator stays one cell, as measured on both hosts.
• Halfwidth dakuten / handakuten (U+FF9E/U+FF9F) and the halfwidth
  Hangul filler (U+FFA0) were treated as zero-width; every measured terminal
  draws them in their own cell. Halfwidth katakana + (han)dakuten (ｶﾞ ﾊﾟ)
  form one two-cell cluster so the renderer writes them together.
• build-msix.ps1 ran cargo build --release; like the other packaging
  scripts it now only checks that the release binary exists.

Not changed / known limitations

• East Asian Ambiguous characters remain one cell; applying a locale-eaw
  style table on the host side would only create a mismatch with the
  ConPTY.
• Hangul jamo (U+1160–11FF, U+D7B0–D7FB) differ between hosts (WezTerm 0,
  Windows Terminal 1) and keep unicode-width's zero width.
• U+17A4 / U+17D8 are reported as 2 / 3 cells by unicode-width 0.1 while
  terminals draw 1; left for a crate update.
• The startup probe writes and clears one line on the alternate screen;
  it needs a tty that answers DSR.
• Without conpty.dll next to wtmux.exe the pane still runs in the inbox
  conhost, and its per-code-point layout of positioned output is not
  corrected; only the wtmux-side fixes above apply.