Terminal emulator comparison

Unicode, VT sequence and GUI feature support across terminal emulators — measured, not asserted. Every number below is reproduced by the harness in this repository.

Generated 2026-07-19 21:03:50 +0200 · ucs-detect pinned at ea4510a4bc6e + 0001-vs15-must-not-narrow.patch.

How this was produced

Not measured here

These rows are filled by running the same harness on macOS or Windows and committing the results.

Unicode support (measured)

Share of measurements whose cursor advance matched the expected column width. Higher is better.

• Wide — Characters that must occupy two columns.
• Narrow — Characters that must occupy exactly one column.
• VS16 — Emoji presentation selector. VS16 may promote a cluster to two columns, so the expectation is a width of 2.
• VS15 — Text presentation selector. VS15 must NOT change the cluster's width, so a wide base is still expected to measure 2. See the VS15 section of the report.
• ZWJ — Zero-width-joiner emoji sequences, which must measure as a single cluster.
• Flags (RI) — Regional-indicator pairs forming flag emoji.
• Lone RI — A single regional indicator, with no pair to join.
• Skin tone — Fitzpatrick skin-tone modifiers applied to a base emoji.

Capabilities (measured by probe)

Answered by the terminal during the run. A dash can mean not supported, not enabled by default, or the probe asked the wrong question; caveats follow the table.

Caveats

• Sixel graphics — xterm gates Sixel on its *graphics* ID rather than its terminal ID, and ships with it off. The harness now starts xterm with `decGraphicsID: vt340`, which enables Sixel without dropping the terminal below VT420 and losing the rectangular operations, so xterm's row reflects what it can do rather than what it does out of the box.
• OSC 52 clipboard — Detected by what the terminal advertises -- DA1 extension 52 and the XTGETTCAP `Ms` capability -- not by writing to the clipboard, which would raise a permission prompt. A terminal that implements OSC 52 without advertising it, or that ships it disabled for security reasons, therefore reads as "no".
• Styled underlines (CSI 4:x) — Asked through XTGETTCAP, so this measures whether the terminal *advertises* the capability, not whether it draws a curly underline. Several terminals that render them -- VTE, Konsole and Alacritty among the ones measured here -- do not answer the capability query, and the documented matrix records them as supporting the feature. Read a "no" as "not advertised".
• Underline colour (SGR 58) — Asked through XTGETTCAP, with the same caveat as styled underlines: this is what the terminal advertises, not what it renders.
• DECRQCRA (checksum of rectangular area) — xterm gates this on two things at once (charproc.c:5612): a VT420 terminal level, which is already its default, and `AllowWindowOps(ewGetChecksum)`, which is not. The harness now starts xterm with `allowWindowOps: true` and xterm answers. Note also that the reply's final byte is contested -- xterm answers `* y`, not `$ y` -- so a probe written against the other reading can still record a false negative.

DEC private modes (DECRQM)

n/a marks a terminal that answers no DECRQM at all, which is a fact about it rather than a gap in the measurement: Konsole can set and reset modes but never report one, so no mode row can be filled in for it by query. The documented matrix below covers those features instead.

Mouse reporting modes (DECRQM)

What DECRQM cannot answer (measured by probe)

Page memory, the DEC locator, right-to-left cursor movement and Glyph Protocol are not DEC private modes, so DECRQM cannot answer for them. These rows come from harness/vt_probe.py, which uses the sequence inside the terminal and reads the reply.

Caveats

• Reports a page number — A terminal can answer DECXCPR without the third (page) parameter. That is recorded here as "no": it answers the report but has no page memory to describe.
• DEC page memory (more than one page) — Confirmed by moving with PPA and asking DECXCPR which page the cursor is on, so a terminal that accepts PPA and ignores it is not counted.
• Pages reached — The walk stops at 32, well above any implementation found, and the probe records whether it hit that ceiling. Contour reaches 15, which matches its source exactly: MaxPageCount is 16 and the last index is reserved for the alternate screen, leaving 15 DEC-addressable pages. An earlier ceiling of 8 understated it, and the number here was the probe's limit rather than the terminal's.
• DEC locator (DECELR / DECRQLP) — A terminal with no locator answers nothing at all. xterm's case is a build option, not a runtime one: OPT_DEC_LOCATOR is `#undef` by default in xtermcfg.h and needs `--enable-dec-locator` at configure time, so a stock xterm has no locator code compiled in and its "no" is correct for the binary measured here, while the documented row records that the implementation exists in the source.
• Locator reports a position — DECLRP with Pe=0 means "locator unavailable" -- the sequence is implemented but no pointing device answered. Under a headless display that is the expected reply, so this row separates "implements the locator" from "a device was there".
• Glyph Protocol — Detection is the protocol's own `s` verb — `ESC _ 25a1 ; s ESC \` — whose reply is defined to be the way an application finds out. Glyph Protocol rides on APC, which terminals are required to ignore when they do not recognise the message, so a terminal without it answers nothing and that silence is the "no" here. The protocol is new (v1.0 April 2026, v1.9 by May) and Rio is its first implementation, so a dash in this row is the expected reading for everyone else rather than a gap.
• Glyph payload formats accepted — The `fmt=` list from the same reply, printed rather than reduced to yes/no because which formats a terminal takes is the answer to the question. `glyf` is the monochrome outline every v1 terminal must accept; `colrv0` and `colrv1` are the two OpenType colour models, added in v1.2. An empty list means the terminal speaks the protocol but will reject every registration.
• Glyph registration is stored — Functional, not declarative — the same separation the DECRQM behaviour check makes below. The probe registers a 29-byte `glyf` triangle at U+100000 and then asks the `q` verb who covers that codepoint, counting it only when the answer names `glossary`. U+100000 is chosen because no known system font covers it, so the coverage reply cannot be a font answering in the registration's place. A terminal that acknowledged the registration with `status=0` and kept nothing would pass the row above and fail this one. The probe clears the slot again.
• DECRLM actually moves the cursor leftwards — Functional, not declarative. The probe writes one Hebrew letter from column 40 with DECRLM (mode 34) reset and again with it set, and compares the cursor advance. Only a reversal counts: +1 then -1 is honoured, +1 then +1 means the mode is recognised and does nothing. Measuring both directions also proves the terminal was answering DSR 6 at all, so a "no" here cannot be a silent timeout.

Every DEC private mode any terminal supports

105 modes, collected from the measurements rather than a hand-kept list. Modes that no terminal recognised are omitted.

Modes a terminal declares but does not honour

DECRQM reports what a terminal recognises; a functional probe reports what it does. Where this report has both for the same feature, the two are compared. A terminal listed here answers a mode query affirmatively and then does not act on the mode — the one failure a mode query cannot detect by construction, and worse for an application than an honest “not recognised”, because there is nothing left to test.

• Declaring DECRLM tells an application it may rely on right-to-left layout. A terminal that answers DECRQM affirmatively and then advances the cursor rightwards is worse than one that answers "not recognised", because the application has no way to find out.

Where measurement and documentation disagree

9 of the features covered by both a runtime probe and the documented matrix disagree. Every case runs the same way — the source says yes, the running terminal says no — and each has a reason worth knowing. Some are shipped disabled: xterm and Alacritty gate OSC 52 behind a setting. The rest are implemented but not advertised: the underline probes ask XTGETTCAP, and VTE, Konsole, Alacritty and Rio draw styled or coloured underlines perfectly well without answering the capability query. Two former entries are gone because the harness now launches xterm correctly — Sixel is gated on its graphics ID and DECRQCRA on a window-op permission, neither of which is a conformance level. Where they still differ, the documented column is the better guide to what the terminal can do, and the measured column to what it will admit to.

VT sequences and extensions (documented)

Compiled from each terminal's source tree and documentation rather than measured, so it also covers terminals this machine cannot run. Where a row overlaps the measured tables above -- the DEC mode rows especially -- the two were derived independently and can be read against each other.

colors

graphics

hyperlinks

input

other

shell

unicode

window

GUI and user-facing features (documented)

Compiled from official documentation and source.

accessibility

config

interaction

layout

other

platform

rendering

session