Problem:
Detection of the pynvim module is currently done by finding the first
Python interpreter in the `PATH` and checking if it can import pynvim.
This has several problems:
- Activation of an unrelated Python virtual environment will break
automatic detection, unless pynvim is also installed in that
environment.
- Installing pynvim to the expected location is difficult. User
installation into the system-wide or user-wide Python site area is now
deprecated. On Ubuntu 24.04 with Python 3.12, for example, the
command `pip install --user pynvim` now fails with the error message
`error: externally-managed-environment`.
- Users may create a dedicated virtual environment in which to install
pynvim, but Nvim won't detect it; instead, they must either activate
it before launching Nvim (which interferes with the user of other
virtual environments) or else hard-code the variable
`g:python3_host_prog` in their `init.vim` to the path of the correct
Python interpreter. Neither option is desirable.
Solution:
Expose pynvim's Python interpreter on the `PATH` under the
name `pynvim-python`. Typical user-flow:
1. User installs either uv or pipx.
2. User installs pynvim via:
```
uv tool install --upgrade pynvim
# Or:
pipx install --upgrade pynvim
```
With corresponding changes in pynvim https://github.com/neovim/pynvim/issues/593
the above user-flow is all that's needed for Nvim to detect the
installed location of pynvim, even if an unrelated Python virtual
environments is activated. It uses standard Python tooling to automate
the necessary creation of a Python virtual environment for pyenv and the
publication of `pynvim-python` to a directory on `PATH`.
Don't print ABI version of duplicated parsers that are later in the
runtime path (see [#35326]).
Change the sorting from `name > path` to `name > rtpath_index`, this
ensures the first (loaded) parser is first in the list and any
subsequent parsers can be considered "not loaded".
This is fuzzy at best since `vim.treesitter.language.add` can take a
path to a parser and change the load order.
The correct solution is for `vim.treesitter.language.inspect` to return
the parser path so we can compare against it and/or for it to also be
able to take a path to a parser so we can inspect it without loading it
first.
Problem: fuzzy-matching can be improved
Solution: Implement a better fuzzy matching algorithm
(Girish Palya)
Replace fuzzy matching algorithm with improved fzy-based implementation
The
[current](https://www.forrestthewoods.com/blog/reverse_engineering_sublime_texts_fuzzy_match/)
fuzzy matching algorithm has several accuracy issues:
* It struggles with CamelCase
* It fails to prioritize matches at the beginning of strings, often
ranking middle matches higher.
After evaluating alternatives (see my comments
[here](https://github.com/vim/vim/issues/17531#issuecomment-3112046897)
and
[here](https://github.com/vim/vim/issues/17531#issuecomment-3121593900)),
I chose to adopt the [fzy](https://github.com/jhawthorn/fzy) algorithm,
which:
* Resolves the aforementioned issues.
* Performs better.
Implementation details
This version is based on the original fzy
[algorithm](https://github.com/jhawthorn/fzy/blob/master/src/match.c),
with one key enhancement: **multibyte character support**.
* The original implementation supports only ASCII.
* This patch replaces ascii lookup tables with function calls, making it
compatible with multibyte character sets.
* Core logic (`match_row()` and `match_positions()`) remains faithful to
the original, but now operates on codepoints rather than single-byte
characters.
Performance
Tested against a dataset of **90,000 Linux kernel filenames**. Results
(in milliseconds) show a **\~2x performance improvement** over the
current fuzzy matching algorithm.
```
Search String Current Algo FZY Algo
-------------------------------------------------
init 131.759 66.916
main 83.688 40.861
sig 98.348 39.699
index 109.222 30.738
ab 72.222 44.357
cd 83.036 54.739
a 58.94 62.242
b 43.612 43.442
c 64.39 67.442
k 40.585 36.371
z 34.708 22.781
w 38.033 30.109
cpa 82.596 38.116
arz 84.251 23.964
zzzz 35.823 22.75
dimag 110.686 29.646
xa 43.188 29.199
nha 73.953 31.001
nedax 94.775 29.568
dbue 79.846 25.902
fp 46.826 31.641
tr 90.951 55.883
kw 38.875 23.194
rp 101.575 55.775
kkkkkkkkkkkkkkkkkkkkkkkkkkkkk 48.519 30.921
```
```vim
vim9script
var haystack = readfile('/Users/gp/linux.files')
var needles = ['init', 'main', 'sig', 'index', 'ab', 'cd', 'a', 'b',
'c', 'k',
'z', 'w', 'cpa', 'arz', 'zzzz', 'dimag', 'xa', 'nha', 'nedax',
'dbue',
'fp', 'tr', 'kw', 'rp', 'kkkkkkkkkkkkkkkkkkkkkkkkkkkkk']
for needle in needles
var start = reltime()
var tmp = matchfuzzy(haystack, needle)
echom $'{needle}' (start->reltime()->reltimefloat() * 1000)
endfor
```
Additional changes
* Removed the "camelcase" option from both matchfuzzy() and
matchfuzzypos(), as it's now obsolete with the improved algorithm.
related: neovim/neovim#34101fixesvim/vim#17531closes: vim/vim#179007e0df5eee9
Co-authored-by: Girish Palya <girishji@gmail.com>
Problem: `vim.diagnostic.get_prev()` / `vim.diagnostic.get_next()` use
logical diagnostic positions and consider extmark validity.
`vim.diagnostic.get()` only uses original positions and doesn't care
about extmark validity. This results in inconsistency between these
APIs.
Solution: use original positions in `vim.diagnostic.get_prev()` and
`vim.diagnostic.get_next()` and don't consider extmark validity to match
previous behavior, which is consistent with `vim.diagnostic.get`.
Problem: Temporary cmdline config is saved to be restored later.
Solution: Close the cmdline window so that it is recreated with the appropriate config.
Problem: diff format erroneously detected
(Tomáš Janoušek)
Solution: Make the regex to detect normal diff format a bit stricter,
while at it, fix wrong test content from patch v9.1.1606
fixes: vim/vim#17946887b4981e7
Co-authored-by: Christian Brabandt <cb@256bit.org>
Problem: No way to skip install confirmation in `add()`. Having install
confirmation by default is a more secure design. However, users are
usually aware of the fact that plugin will be installed and there is
currently no way to skip confirmation.
Plus it can introduce inconvenience on the clean config initialization
if it is modularized with many `vim.pack.add()` calls (leads to
confirming installation many times in a row).
Solution: Add `opts.confirm` option that can skip install confirmation.
Problem: No way to have full control over how plugin is loaded.
Although `:packadd!` has small side effects (only adds plugin
directory to 'runtimepath'; and maybe its 'after/' subdirectory), it
still has side effects. For example, 'plugin/' directories are still
loaded during startup (as part of `:h load-plugins`).
Solution: Allow function `opts.load` that has full control over how
plugin is loaded.
Problem: the `load=true` in `vim.pack.add()` means that `:packadd` is
executed even during startup. This leads to force source of 'plugin/',
which breaks the intended loading order (`:h load-plugins`) and
results into sourcing them twice. This also makes it ignore
`--noplugin` argument.
Using `:packadd!` during startup is more appropriate, while `:packadd`
afterwards is still more favorable to actually force 'plugin/' source
(as there is no pre-defined mechanism that will load them later).
Solution: have `load=false` default during startup, `true` - afterwards.
Problem: filetype: a few more files are not recognized
Solution: guess Mail, Info and Terminfo files by its content
(lacygoill)
closes: vim/vim#17880eb2aebeb79
Co-authored-by: lacygoill <lacygoill@lacygoill.me>
Using a snippet like:
${1:name} :: ${2}\n${1:name} ${3}= ${4:undefined}${0}
The text for `${1:name}` stopped syncing if the completion popup menu
showed up. E.g. typing `par` where the `a` triggered completion resulted
in:
pat ::
pa = undefined
Instead of:
pat ::
pat = undefined
Problem: diff: using diff anchors with hidden buffers fails silently
Solution: Give specific error message for diff anchors when using hidden
buffers (Yee Cheng Chin).
Diff anchors currently will fail to parse if a buffer used for diff'ing
is hidden. Previously it would just fail as the code assumes it would
not happen normally, but this is actually possible to do if `closeoff`
and `hideoff` are not set in diffopt. Git's default diff tool "vimdiff3"
also takes advantage of this.
This fix this properly would require the `{address}` parser to be
smarter about whether a particular address relies on window position or
not (e.g. the `'.` address requires an active window, but `'a` or `1234`
do not). Since hidden diff buffers seem relatively niche, just provide a
better error message / documentation for now. This could be improved
later if there's a demand for it.
related: vim/vim#17615closes: vim/vim#17904cad3b2421d
Co-authored-by: Yee Cheng Chin <ychin.git@gmail.com>
- Add delimiter between function signature and documentation, matching hover formatting
- Show title only if there are multiple clients or multiple signatures
- Avoid duplicating the title inside the window if it's already shown in the border
The '?' needs to be escaped, because the autocommand is using
file-patterns (glob like) and not a regex). See :h file-pattern
closes: vim/vim#17890f7deb815b0
Co-authored-by: Girish Palya <girishji@gmail.com>
Problem: make_floating_popup_options only shows when opts.border is explicitly set, ignoring global winborder setting
Solution: check both opts.border and vim.o.winborder when determining whether to show title
The cursor movement autocommand can not detect when the final tabstop $0
is directly adjacent to another tabstop, which prevents ending the
snippet session. The fix is an early return when jumping.
Problem:
Previously, 'null' value in LSP responses were decoded as 'nil'.
This caused ambiguity for fields typed as '? | null' and led to
loss of explicit 'null' values, particularly in 'data' parameters.
Solution:
Decode all JSON 'null' values as 'vim.NIL' and adjust handling
where needed. This better aligns with the LSP specification,
where 'null' and absent fields are distinct, and 'null' should
not be used to represent missing values.
This also enables proper validation of response messages to
ensure that exactly one of 'result' or 'error' is present, as
required by the JSON-RPC specification.
Platform specific options are cringe and should either be fixed or
deleted.
In this case, the platform difference can trivially be implemented
using a conditional default value. If the user overrides the value,
then it is because the user wants that value, regardless of the
corporation who manufactured the OS that the user is running.
Possible alternative: delete the option by making it always on.
Problem: cannot perform autocompletion
Solution: Add the 'autocomplete' option value
(Girish Palya)
This change introduces the 'autocomplete' ('ac') boolean option to
enable automatic popup menu completion during insert mode. When enabled,
Vim shows a completion menu as you type, similar to pressing |i\_CTRL-N|
manually. The items are collected from sources defined in the
'complete' option.
To ensure responsiveness, this feature uses a time-sliced strategy:
- Sources earlier in the 'complete' list are given more time.
- If a source exceeds its allocated timeout, it is interrupted.
- The next source is then started with a reduced timeout (exponentially
decayed).
- A small minimum ensures every source still gets a brief chance to
contribute.
The feature is fully compatible with other |i_CTRL-X| completion modes,
which can temporarily suspend automatic completion when triggered.
See :help 'autocomplete' and :help ins-autocompletion for more details.
To try it out, use :set ac
You should see a popup menu appear automatically with suggestions. This
works seamlessly across:
- Large files (multi-gigabyte size)
- Massive codebases (:argadd thousands of .c or .h files)
- Large dictionaries via the `k` option
- Slow or blocking LSP servers or user-defined 'completefunc'
Despite potential slowness in sources, the menu remains fast,
responsive, and useful.
Compatibility: This mode is fully compatible with existing completion
methods. You can still invoke any CTRL-X based completion (e.g.,
CTRL-X CTRL-F for filenames) at any time (CTRL-X temporarily
suspends 'autocomplete'). To specifically use i_CTRL-N, dismiss the
current popup by pressing CTRL-E first.
---
How it works
To keep completion snappy under all conditions, autocompletion uses a
decaying time-sliced algorithm:
- Starts with an initial timeout (80ms).
- If a source does not complete within the timeout, it's interrupted and
the timeout is halved for the next source.
- This continues recursively until a minimum timeout (5ms) is reached.
- All sources are given a chance, but slower ones are de-prioritized
quickly.
Most of the time, matches are computed well within the initial window.
---
Implementation details
- Completion logic is mostly triggered in `edit.c` and handled in
insexpand.c.
- Uses existing inc_compl_check_keys() mechanism, so no new polling
hooks are needed.
- The completion system already checks for user input periodically; it
now also checks for timer expiry.
---
Design notes
- The menu doesn't continuously update after it's shown to prevent
visual distraction (due to resizing) and ensure the internal list
stays synchronized with the displayed menu.
- The 'complete' option determines priority—sources listed earlier get
more time.
- The exponential time-decay mechanism prevents indefinite collection,
contributing to low CPU usage and a minimal memory footprint.
- Timeout values are intentionally not configurable—this system is
optimized to "just work" out of the box. If autocompletion feels slow,
it typically indicates a deeper performance bottleneck (e.g., a slow
custom function not using `complete_check()`) rather than a
configuration issue.
---
Performance
Based on testing, the total roundtrip time for completion is generally
under 200ms. For common usage, it often responds in under 50ms on an
average laptop, which falls within the "feels instantaneous" category
(sub-100ms) for perceived user experience.
| Upper Bound (ms) | Perceived UX
|----------------- |-------------
| <100 ms | Excellent; instantaneous
| <200 ms | Good; snappy
| >300 ms | Noticeable lag
| >500 ms | Sluggish/Broken
---
Why this belongs in core:
- Minimal and focused implementation, tightly integrated with existing
Insert-mode completion logic.
- Zero reliance on autocommands and external scripting.
- Makes full use of Vim’s highly composable 'complete' infrastructure
while avoiding the complexity of plugin-based solutions.
- Gives users C native autocompletion with excellent responsiveness and
no configuration overhead.
- Adds a key UX functionality in a simple, performant, and Vim-like way.
closes: vim/vim#17812af9a7a04f1
Co-authored-by: Girish Palya <girishji@gmail.com>
Problem: Left (`<` prefix) and right (`>` prefix) parts of update
changelog come in not fixed order: it can be left-right or right-left.
Solution: Ensure left-right order.
Problem: New version suggestions in update confirmation buffer might
include semver tags that were committed later but for versions that
are not greater than current. Like if versions committed in order
`v0.2.0` - `v0.3.0` - `v0.2.1` - `v0.3.1`, then when on `v0.3.0` both
`v0.2.1` and `v0.3.1` are suggested, but only the latter is newer as
a version. This is because those tags are computed with post-processed
`git tag --list --contains HEAD`.
Solution: Compute all semver tags and filter only those greater than the
latest version available at HEAD.
Problem: Both `PackChangedPre` and `PackChanged` contain |event-data|
with plugin's `spec`. It looks like a good idea to have all its
triggers contain the same format across all kinds ("install",
"update", "delete"). There are several choices:
- Have it be as verbatim as supplied to `vim.pack.add()`, i.e. can
be either string or table. A bit too ambiguous.
- Have it be table with `src` and `name` inferred. This requires
less work for "install", but more work for "update" and "delete"
(since they use `vim.pack.get()` which already infers default
`version`).
- Have it be table with *all* defaults made explicit. This looks
like the best approach, but requires extra care to only infer
default `version` when needed (i.e. avoid inferring during regular
load) because it is costly in terms of startup time.
This might also introduce inconsistency when dealing with
lockfile(s) as information there should be as close to what user
supplied as possible. Address that when dealing with lockfile.
Solution: Ensure explicit `version` in all events where possible.
