Skip to content

BACKENDS.md

Latest commit

 

History

History
89 lines (79 loc) · 6.01 KB

BACKENDS.md

File metadata and controls

89 lines (79 loc) · 6.01 KB
Raw

Emulator backends: the comparison behind decision 2

STABILITY.md §2 says vt100 stays and the attribute shadow stays on purpose. This is the comparison it rests on (#150), measured in September 2026 against the seven methods the Emulator trait needs — process, snapshot, mid_sequence, in_sync_update, input_modes, mode_state, set_size — and against the three places the backend, not the design, was the limit: blink/conceal/strikethrough (the shadow), mouse modes collapsing into one value, and exotic grapheme clusters rendering however the backend renders them.

The candidates

vt100 0.16.2 alacritty_terminal 0.26.0 wezterm-term
On crates.io yes yes no — git-only in the wezterm workspace; the tattoy-wezterm-term fork is a third party's
Licence MIT Apache-2.0 only MIT
MSRV 1.70 1.85.0, moves with Alacritty releases wezterm's, moves with wezterm
Source size 3,950 lines 11,739 lines larger; a workspace of crates (wezterm-cell, -escape-parser, -surface, termwiz)
Dependency tree (default-features = false) 9 crates: vte, unicode-width, itoa, and theirs 27 crates: vte, polling, rustix, rustix-openpty, signal-hook, parking_lot, regex-automata, base64, home… — the tty/event-loop half is not feature-gated image, serde, terminfo, url, lru, miniz_oxide, unicode-normalization, wezterm-bidi…
Headless driving Parser::process(bytes); Callbacks trait for what it does not handle Term::new(config, dims, EventListener) + vte::ansi::Processor::advance; VoidListener exists Terminal::new with a TerminalConfiguration and a writer; designed around a renderer
SGR blink (5/6) dropped dropped — no Flags::BLINK; Attr::Blink is not handled tracked
SGR conceal (8) dropped tracked (HIDDEN) tracked
SGR strikethrough (9) dropped tracked (STRIKEOUT) tracked
Extra attributes — double/curly/dotted/dashed underline, underline colour the same and more
Mouse modes one value (the last set) distinct flags: MOUSE_REPORT_CLICK, MOUSE_DRAG, MOUSE_MOTION, SGR_MOUSE; no 1005 distinct
Kitty keyboard, focus, bracketed paste, alt screen, insert, origin, LNM partial (see below) all as TermMode flags all
Grapheme clusters one char + width; zero-width chars appended to the previous cell one char + zerowidth: Vec<char>, same shape finl_unicode segmentation; the strongest of the three
Scrollback reflow on resize no yes (grid/resize.rs) yes
Unhandled-sequence hook Callbacks::unhandled_* — the basis of Screen::unsupported none; unknown sequences are silently ignored none
Release cadence slow: 0.16.2 (July 2025) is the newest with Alacritty, several a year, breaking freely with wezterm

What termlens already does in front of the backend

The Emulator trait was meant as a swap point, but the more useful thing it turned out to be is a place to stand in front of the backend with a staged stream both parsers receive. Everything below is termlens's own and would survive any swap unchanged:

  • Character sets (G0–G3, SO/SI, SS2/SS3, DEC Special Graphics, UK): glyph substitution before the parser.
  • Tab stops (HTS/TBC/CHT/CBT, and plain HT): rewritten to CHA.
  • Insert mode (IRM): rewritten to ICH sized in columns.
  • The unsupported record (Screen::unsupported): every sequence the backend did not implement, from Callbacks — the honesty layer that says "this grid may be wrong" instead of looking plausible.
  • Queries and their answers, the title, the clipboard, links, the counters, graphics payloads: the sequence tracker, not the emulator.
  • Blink, conceal, strikethrough: the shadow parser.

Why a swap does not pay today

  1. wezterm-term cannot be depended on. It is not published; deny.toml refuses git sources and crates.io refuses git dependencies. It would have to be vendored, which is the option the shadow was chosen over at 3,950 lines, now at many times that.
  2. alacritty_terminal does not close the gap it would be adopted for. It tracks two of the three attributes and drops blink. A swap keeps a shadow (or a fork) for one attribute, so the mechanism stays and only the reason shrinks. It also pulls an event loop and a tty layer termlens has its own of, triples the dependency count, moves the MSRV onto Alacritty's schedule, and — because it ignores what it does not implement — would cost Screen::unsupported its source.
  3. The shadow is not a performance problem. 40,000 lines through an 80x24 screen: 260 ms with it, 263 ms without; 284 ms against 282 ms when every line carries four SGR sequences. A terminal's throughput is the PTY round trip.
  4. The three limits have been answered in front of the backend, or accepted. Mouse modes: Screen::mouse_modes reports the set the application enabled from the tracker (#151). Graphemes: the unicode-torture fixture pins vt100's rendering rather than promising one, and no candidate promises Unicode-correct segmentation and the rest of this table.

What would reopen the question

  • vt100 gains blink, conceal and strikethrough (three spare bits in mode: u8, an 80-line patch): emu/shadow.rs deletes and convert_cell reads the flags. This is the outcome to watch for, and the module header says so.
  • A candidate appears that is published, tracks all three attributes, can be driven headless without an event loop, and reports the sequences it does not handle. Then the comparison is worth re-running against the trait, behind a feature flag, with the fixture suite as the judge — the ratatui fidelity test and the unicode-torture snapshot are the two that would move.
  • A consumer needs reflowed history. Neither the design nor the backend offers it today (Terminal::resize says why); a backend that does would be an argument, not a decision.