fzf consumed only the part of a sequence it recognized, and the rest was
typed into the query. CTRL-A sent as \e[97;5u became "97;5u", and CTRL-F
sent as \e[70;5u fired Home before typing ";5u". A BEL-terminated OSC
reply also aborted fzf, because the BEL that followed the typed payload
was read as CTRL-G.
Frame CSI and SS3 by their parameter and final byte ranges, OSC, DCS and
APC by their terminator. Drop the whole frame when the parser does not
recognize it, or recognizes only a prefix of it.
Take the second-chance read only when the sequence at the start of the
buffer is the only one and is unfinished. Otherwise it blocked until the
next key, holding back what followed, CTRL-G included.
Telling a terminal's sequence from an ALT key is the hard part, so these
keep what the parser made of them:
- An unterminated sequence, since that is how ALT-[, ALT-O, ALT-], ALT-P
and ALT-_ arrive
- A CSI or SS3 of four bytes or fewer. It could be an ALT key and typed
text, and rxvt sends keys of that size fzf does not know, such as \e[3^
- A final byte after an intermediate $. rxvt ends keys with $, as in
\e[7$ and \e[23$, so the byte after it is the next key
- A string sequence whose payload does not start like a reply: digits and
';' for OSC, '>|', '!|', '[01]$r' or '[01]+r' for DCS, 'G' and a key
for APC
- SOS and PM, which nothing sends
Within a string, an ESC ends it and introduces a sequence of its own, and
only OSC ends with BEL. The read loop does not wait for a string
terminator: that held ALT-], ALT-P and ALT-_ for ESCDELAY and let typed
bytes pile up until they looked like a sequence.
- When the cursor is not at column 0, Init() makes space so fzf draws on
the next line and the prompt line is left intact
- Close() undid that: it moved up onto the prompt line before ESC[J, so
the erase wiped a line fzf never wrote on
- Shells repaint afterwards, so it showed as a flicker of the last prompt
line. Reported on fish, also seen on bash and nushell. zsh erases the
same line but repaints fast enough to hide it
- Erase our own area first, then step back and restore the column, so
nothing on that line is touched, RPROMPT included
Fix#4913
- getch() charged EINTR against a 100 iteration limit meant for read
attempts, and the count only reset once a byte was read
- So signals arriving during an idle prompt drained it, and running out
returned getCharError, which became Event{Fatal} and exit status 2
- Only the cancellable path is affected, hence only --listen
- select() is not restarted by SA_RESTART, unlike the blocking read on
the non-cancellable path, so it is the only one that sees EINTR
- waitReadable() already retries EINTR without a limit
Fix#4917
- progress was set to 100 by every UpdateList, partial results included,
and never reset when a search started
- So a GET could report a new query, the previous result set, and
progress 100, and a client had no way to tell it was stale
- Now cleared where the search is requested and set to 100 only on the
final result. Query and the reset happen in the same locked section
that dumpStatus locks, so query == X with progress 100 means the
matches belong to X
Fix#4903
- extraLines() used borderLines(), which reports zero for BorderInline
- Divider drawn for an inline section still takes a line, and
resizeWindows() reserves it via addInline()
- So adaptive height came up one line short per inline section and the
list scrolled
- Affects --header-lines-border, --header-border and --footer-border
alike. All three set came up three lines short
- Without a list border the shape normalizes to BorderTop or BorderBottom,
which report a line, hence --no-list-border looked like a fix
seq 10 | fzf --height=~100% --list-border \
--header-lines=1 --header-lines-border=inline
Fix#4904
ESCDELAY=0 makes the retry count zero, so the wait added in d0377ed had
no effect and the reply still leaked. Past the introducer these bytes
cannot be the ESC key, so waiting for them adds no Escape latency.
- Wait at least defaultEscDelay, honour a larger ESCDELAY
- Lone ESC still honours ESCDELAY=0, measured 3ms to exit either way
Fix#4899
Read loop dropped its escDelay retry budget after every successful byte,
so a sequence split across reads reached the parser as a fragment, parsed
as ALT-[ with the remainder left behind as query text.
- fzf queries DECRQM at startup since dab626b, so a terminal answering
late leaked "?2004;2$y" into the query
- Same split leaked modified keys and mouse sequences: CTRL-UP left "5A",
SGR mouse left "0;1;1M"
- Bound unchanged, a stall longer than escDelay still falls back to ALT
Fix#4899
escapeArg only doubles a backslash that precedes a quote or ends the
argument, the same rule as syscall.EscapeArg, and @ is not a cmd
metacharacter so it is not caret-escaped. The Windows expectations here
still asked for both. Checked by round-tripping each entry through
cmd.exe into a program's argv: all ten come back byte for byte.
The template writes to os.CreateTemp("", "fzf-temp-*"); the fzf-preview-
prefix went away with the --tmux work. The digit count was also pinned at
nine, but CreateTemp's suffix has no fixed width.
The file is opened under t.TempDir() and never closed, so the cleanup
that TempDir registers cannot remove it on Windows, where an open file
cannot be unlinked. The test itself passes; it is the cleanup that fails
the run. history_test.go and options_test.go already close theirs.
Ask the terminal whether the mode is already on (DECRQM) and put it back
that way on exit. Forcing it off broke pasting in shells that run fzf
from a line editor widget, which enable the mode only when the editor
starts. Terminals that do not answer fall back to disabling it.
- Startup queries go out in one write, cursor position last. Every
terminal answers DSR, so its reply bounds the wait: a paste reply
still missing by then means the terminal does not know the query.
- Bound the first read with select(2). Terminals that never answer
escape sequences, such as FreeBSD virtual terminals, blocked startup
until a key was pressed, and that keystroke was then discarded.
Fix#4887Fix#2860Fix#976
ToRunes aliases the rune array, and the editing actions append into
t.input in place when the cursor is not at the end, so the keystrokes
edit the item. Non-ASCII items only, ASCII gets a fresh slice.
printf '한글abcde\n' | fzf --bind 'ctrl-y:replace-query'
ctrl-y, Left, BSpace, ctrl-u -> 한글abcee
Runes and ToRunes are now documented read-only. A stale fold bit was the
other symptom, letting the prefilter reject an item the general path
matches.
The scan only ran for ASCII patterns, so searching CJK text with a CJK
query still built the full score matrix. Scan for one byte of the pattern
rune and verify all four.
Which byte matters. Every ASCII rune contributes three zero bytes, so
U+AE00 scanned by its zero low byte hits on nearly every character of an
ASCII-heavy line. Pick a byte that cannot occur in an ASCII rune, else
any non-zero one.
A non-ASCII pattern rune is safe only when no other rune lowercases onto
it. Uncased is not sufficient: U+00DF has no simple uppercase, yet U+1E9E
lowercases to it, so the foldable set is excluded too.
Measured on 1.4M-line corpora, with a non-ASCII query:
- Every line CJK: 5.1x to 10.2x
- Mostly-ASCII paths behind a Hangul prefix: 5.6x to 6.0x, and 1.2x
where every line matches so nothing can be rejected
- ASCII queries unchanged, kept off the non-inlinable guard
normalizeRune guarded with 0x00C0..0xFF61, which does not exclude Hangul,
CJK or Cyrillic, so every rune of those scripts hashed into the map only
to miss. Every key of the map folds to ASCII, so the bitmap added for the
rune prefilter rejects them without a lookup.
- Non-ASCII queries 1.21x where every line is CJK, 1.05x on mostly-ASCII
paths behind a Hangul prefix
- ASCII queries unchanged, the prefilter already skips Phase 2 for them
- Normalization share of query time for a non-ASCII query: 17.2% -> 0%
asciiFuzzyIndex gave up on non-ASCII lines, so every item ran the full
score matrix. A []rune is a fixed 4-byte stride, so the SIMD byte
scanners can run over it directly: find the low byte, then confirm
4-byte alignment and three zero bytes.
Case folding and normalization can turn a non-ASCII rune into the ASCII
char being searched, which the scan cannot see. ToChars now flags lines
holding such a rune and those keep the old path. Normalization is
Latin-only, so Hangul, CJK, Cyrillic, Greek, Hebrew, Arabic, Thai, kana
and emoji never set the flag.
Chars had no spare padding, so inBytes moves into a flags byte.
Measured on 1.4M-line corpora:
- Mostly-ASCII paths behind a Hangul prefix: 'conf' 1.8x, 'binutils'
2.9x, 'ltversion' 4.0x, no-match 8.4x
- Every line CJK: 17x on both matching and non-matching queries
- ASCII input unchanged, non-ASCII patterns not covered yet
utf8.DecodeRune already fast-paths ASCII, but it is too complex to inline
(cost 201 against a budget of 80), so a mostly-ASCII line pays one call
per byte just to be told the byte is ASCII.
Only the run after the first non-ASCII byte reaches the decode loop, so
the gain depends on where that byte falls.
- 70-rune ASCII line: 145ns -> 42ns in the decode loop
- Ingestion of 1.4M mostly-ASCII paths behind a Hangul prefix, where the
loop covers the whole line: 377ms -> 233ms
- The same paths with the Hangul at the end, where it covers six bytes:
207ms -> 196ms
- Break-even sits at ~100% non-ASCII runes: still 1.01x at 95%. Only a
line holding no ASCII byte at all loses, by ~0.14ns per rune, ~5% of
the loop
- Capacity was byte length, over-allocating by bytes-per-rune (2-4x)
- Count non-continuation bytes with SWAR before allocating
- Invalid bytes undercount, never overcount, so append covers the gap
- Query performance unchanged, this is a memory fix
- The gain tracks bytes-per-rune, the cost tracks how much of the line
follows the first non-ASCII byte, so the two move independently
Measured on 1.4M-line corpora:
- Every line CJK: RSS 362MB -> 255MB, ingestion -5%
- Mostly-ASCII paths behind a Hangul prefix: RSS 556MB -> 533MB,
ingestion +3.5%, the counting pass covering the whole line
- The same paths with the Hangul at the end: RSS 563MB -> 535MB,
ingestion +0.9%, the counting pass covering six bytes
chafa ends each row of a Kitty Unicode placeholder image with CUB + IND
instead of a newline, so that an image drawn at a column offset survives
ONLCR, which would rewrite the newline as CR NL and pull the cursor back
to column 0. Both sequences were dropped, so the image arrived as a single
line and was re-wrapped, shifting every row after the first by the width
of the wrap sign. Without wrapping it was truncated to its first row.
The column is tracked across the breaks and restored with padding, so an
image indented by the preview command keeps its indent.
Fix#4885
--preview-window ~N draws the header in its own pass, resuming the body
in the middle of an image: an image arrives as one line but takes up many
rows. ~N is typically global and meant for text, while the same preview
command also renders images.
tmux ends a command at an argument whose last character is ';', so a
--border-label ending in one was stored truncated. escapeTmuxSeparator
only covered a value that was exactly ';'.
fzf --popup --border-label 'foo;' -> @fzf-border-label was 'foo'
--border-label of a tmux floating pane was stored in the pane title and
read back by pane-border-format as '#{pane_title}', so any program in the
pane could replace the border text by setting the title. On Zellij the
name was only set when the pane was created. Either way the built-in
label actions could not touch it, as fzf draws no border of its own
there.
- tmux: hold the label in pane-scoped @fzf-border-label. Set the option
even without a label, so a later change-border-label has somewhere to
write. Drop select-pane -T; the pane title is left to the user
- Zellij: run 'zellij action rename-pane' on label change. 'zellij run'
has no option to set the environment of the new pane, so the command
exports the target itself
- change-border-label and transform-border-label now update the native
border, through the setter nativeLabelSetter picks by multiplexer
- The pane is named in a __FZF_INTERNAL_ variable that runProxy withholds
from the environment replay, so the same name in the outer environment
cannot redirect the update to another pane
A Kitty graphics command is an APC sequence, which tmux takes as a
request to set the pane title. A bare one therefore never reaches the
terminal and overwrites the title on the way, which is visible on the
border of a floating pane. Programs are expected to wrap it in the tmux
passthrough sequence themselves, but 'kitten icat --clear' does not,
unlike the image data it emits right after.
Sixel and iTerm2 sequences are left alone; tmux renders the former
itself and neither affects the title.
Related: #4870
- Vary normalize in the fuzz targets; document why the exhaustive sweep
keeps it fixed (no-op for ASCII input)
- Seed the fuzz targets with patterns of the right length; single-char
seeds were all two-char and got rejected
- Note that the disable* hook forbids parallel test execution
For two ASCII characters, rows 0 and 1 of the score matrix collapse to
scalar running state, so Phase 2 and Phase 3 fuse into one pass with no
score arrays. withPos stores the two rows for the backtrace. Up to 1.4x
on two-char queries, the most common multi-char length.
Verify both fast paths against the general algorithm with exhaustive
(every short string over a class-complete alphabet) and fuzz tests,
runnable via the new make fuzz target.
Precompute OR of all sort keys; a byte position that is zero across every
key contributes a no-op pass. Default two-criteria setup leaves the low 32
bits zero, so 4 of 8 passes are skipped without a histogram scan. 2x on the
sort, 3-8% on high-match queries. No regression when all bytes are used.
For single ASCII character, jump between occurrences with vectorized
search and score each directly from bonus matrix instead of running
full pipeline with prefilter and score arrays. 2x faster on 100K-line
benchmark corpus (5.4ms -> 2.5ms). Matters most on first keystroke
where candidate set is largest and result cache cannot help.
Phase 3 writes C rows only from column F[r], but tie-breaking in the
backtrace could read row i+1 at column j+1 left of F[i+1]. With reused
slab, that cell holds data from previously processed item, so highlight
positions in equal-score ties depended on processing order.
Windows has no SIGWINCH and notifyOnResize was an empty TODO, so fzf
running in --height mode never noticed console size changes. Poll the
console screen buffer dimensions every 100ms and push a signal through
the same channel SIGWINCH uses on Unix. Full-screen mode is unaffected
as it goes through tcell which emits its own resize events.
Decision must not depend on threshold chain resolution
(computePreviewSize consults noSeparatorLine); separator is hidden only
when every spec in the chain displays a 'next' preview window with a
border line facing the input section. Otherwise err on the side of
showing it.
- Without active previewer, preview position is resolved from first spec
alone, so 'next' in a threshold alternative does not create input
window and cannot hide the separator
- Empty preview command is undecidable; preview(...) action can display
ad-hoc window at 'next' position at any time
- change-preview-window relayouts or repaints when separator visibility
changes even if the active spec is unchanged, without dropping preview
window forced by preview(...) action
Visible input border suppressed separator even when it draws no line
toward list section (bottom in default layout, top in reverse, vertical).
Suppress only when input border has a line on the list side; otherwise
fall through to adjacent section check.
- Requires header window to have content before treating it as facing
section; hasHeaderWindow returns true with empty header when input
border is visible
- Extract inputBorderFacesList and resolveInputBorderShape shared by
options and terminal
- Add tests for separator suppression rules
- Applies when adjacent section (header, header lines, or list) draws
border line facing input section
# No separator below header border
fzf --style full --input-border none --header foo
# Separator shown; nothing separates input from list
fzf --style full --input-border none --header foo --no-header-border
- Evaluated dynamically so change-header/toggle-header update it
- Inline header is drawn inside list border and not adjacent to input;
decide from section actually next to it
- Separator is kept when a preview window can be at 'next' position,
right next to the input section
Blocking run-shell suspends client key processing until the command
exits, so a floating pane deadlocks; popup input bypasses the suspended
path. Fall back to display-popup when TMUX_PANE is not set.
The native border is the handle for moving and resizing the pane with
the mouse, so use it by default, consistent with tmux. fzf draws its
own border only when a border style is explicitly specified with
--border. Extract the shared native-border decision into a helper.
Prevents another user on a shared TMPDIR from planting a file or a
symbolic link at the predictable path while fzf is running, like the
exit status file of the floating pane. Reject an empty become command.
The pane sets the options itself before running fzf, so that they are
in place no matter how quickly the command exits, targeted at
$TMUX_PANE; the default target would resolve to the active pane of
the session's current window.
- pane-border-format is set to '#{pane_title}' so that the label is
displayed on the border when pane-border-status is enabled;
pane-border-status itself is a window option in released tmux
versions and is left alone
- When a border style is explicitly specified with --border, a popup
is used instead of a floating pane so that the fzf-drawn border is
the only border shown; give 'border-native' to force a floating pane
- 'none' and 'line' are treated as no border; fzf draws no box for
either, so the label is displayed on the native border
- Remove 'border-fzf' which is now redundant; it was never released
- The title is escaped for select-pane -T which expands format
expressions; a lone ';' is escaped as tmux would treat it as a
command separator
- The label is skipped when ANSI stripping leaves an empty string
- --border-label-pos is ignored
- Fix remain-on-exit set on the original pane instead of the floating
pane
Unlike a popup, a floating pane is not modal; you can switch to other
panes and windows while fzf is running, move and resize the pane with
the mouse, zoom it to fullscreen, and use copy-mode in it.
- Floating pane always has a native border, so 'border-native' is
implied; give new 'border-fzf' option to fall back to a popup where
fzf draws its own border
- Popup is also used on tmux versions below 3.7, or when the window is
too small to fit a floating pane
- new-pane does not block until the command finishes and does not
propagate the exit status; block on a wait-for channel signaled by
the pane and pass the exit status through a temporary file
- Watchdog process signals the channel when the pane is closed
abnormally (e.g. kill-pane)
- Kill the pane when the proxy process is interrupted, like a popup
dying with its client
- Unzoom the window before creating the floating pane; doing so over a
zoomed window crashes the tmux server on 3.7b, and newer versions of
tmux unzoom the window anyway
- Floating pane size excludes the border and the position is that of
the content area; treat the requested size as the total footprint
including the border for consistency with popups
- Close the pane on exit even when remain-on-exit is on
- Pre-create the exit status file with O_EXCL to prevent tampering on
a shared TMPDIR
Junegunn Choi