Develop
Nvim :help pages, generated
from source
using the tree-sitter-vimdoc parser.
This reference describes design constraints and guidelines, for developing
Nvim applications or Nvim itself. See
dev-arch for discussion of Nvim's
architecture and internal concepts.
Most important things come first (roughly). Some items conflict; this is
intentional. A balance must be found.
The Neo bits of Nvim should make it a better Vim, without becoming a
completely different editor.
In matters of taste, prefer Vim/Unix tradition. If there is no relevant
Vim/Unix tradition, consider the "common case".
There is no limit to the features that can be added. Select new features
based on (1) what users ask for, (2) how much effort it takes to implement
and (3) someone actually implementing it.
Backwards compatibility is a feature. The RPC API in particular should
never break.
A feature that isn't documented is a useless feature. A patch for a new
feature must include the documentation.
Documentation should be comprehensive and understandable. Use examples.
Don't make the text unnecessarily long. Less documentation means that an
item is easier to find.
Keep Nvim small and fast. This directly affects versatility and usability.
Computers are becoming faster and bigger each year. Vim can grow too, but
no faster than computers are growing. Keep Vim usable on older systems.
Many users start Vim from a shell very often. Startup time must be short.
Commands must work efficiently. The time they consume must be as small as
possible. Useful commands may take longer.
Don't forget that some people use Vim over a slow connection. Minimize the
communication overhead.
Vim is a component among other components. Don't turn it into a massive
application, but have it work well together with other programs
("composability").
The source code should not become a mess. It should be reliable code.
Use comments in a useful way! Quoting the function name and argument names
is NOT useful. Do explain what they are for.
Porting to another platform should be made easy, without having to change
too much platform-independent code.
Use the object-oriented spirit: Put data and code together. Minimize the
knowledge spread to other parts of the code.
Nvim is not an operating system; instead it should be composed with other
tools or hosted as a component. Marvim once said: "Unlike Emacs, Nvim does not
include the kitchen sink... but it's good for plumbing."
A primary goal of Nvim is to allow extension of the editor without special
knowledge in the core. Some core functions are delegated to "providers"
implemented as external scripts.
Examples:
1. In the Vim source code, clipboard logic accounts for more than 1k lines of
C source code (ui.c), to perform two tasks that are now accomplished with
shell commands such as xclip or pbcopy/pbpaste.
2. Python scripting support: Vim has three files dedicated to embedding the
Python interpreter: if_python.c, if_python3.c and if_py_both.h. Together
these files sum about 9.5k lines of C source code. In contrast, Nvim Python
scripting is performed by an external host process implemented in ~2k lines
of Python.
The provider framework invokes Vimscript from C. It is composed of two
functions in eval.c:
eval_call_provider({name}, {method}, {arguments}, {discard}): Calls
provider#{name}#Call with {method} and {arguments}. If {discard} is true, any
value returned by the provider will be discarded and empty value will be
returned.
eval_has_provider(
{name}): Checks the
g:loaded_{name}_provider variable
which must be set to 2 by the provider script to indicate that it is
"enabled and working". Called by
has() to check if features are available.
For example, the Python provider is implemented by the
"autoload/provider/python.vim" script, which sets g:loaded_python_provider
to 2 only if a valid external Python host is found. Then has("python")
reflects whether Python support is working.
provider-reload
Sometimes a GUI or other application may want to force a provider to
"reload". To reload a provider, undefine its "loaded" flag, then use
:runtime to reload it:
:unlet g:loaded_clipboard_provider
:runtime autoload/provider/clipboard.vim
"Just say it". Avoid mushy, colloquial phrasing in all documentation
(docstrings, user manual, website materials, newsletters, …). Don't mince
words. Personality and flavor, used sparingly, are welcome--but in general,
optimize for the reader's time and energy: be "precise yet concise".
Prefer the active voice: "Foo does X", not "X is done by Foo".
Write docstrings (as opposed to inline comments) with present tense ("Gets"),
not imperative ("Get"). This tends to reduce ambiguity and improve clarity
by describing "What" instead of "How".
✅ OK:
/// Gets a highlight definition.
❌ NO:
/// Get a highlight definition.
Avoid starting docstrings with "The" or "A" unless needed to avoid
ambiguity. This is a visual aid and reduces noise.
✅ OK:
/// @param dirname Path fragment before `pend`
❌ NO:
/// @param dirname The path fragment before `pend`
Vim differences:
Do not prefix help tags with "nvim-". Use
vim_diff.txt to catalog
differences from Vim; no other distinction is necessary.
If a Vim feature is removed, delete its help section and move its tag to
vim_diff.txt.
Mention deprecated features in
deprecated.txt and delete their old doc.
Use consistent language.
"terminal" in a help tag always means "the embedded terminal emulator",
not "the user host terminal".
Use "tui-" to prefix help tags related to the host terminal, and "TUI"
in prose if possible.
Rough guidelines on where Lua documentation should end up:
Nvim API functions vim.api.nvim_* should be in api.txt.
If the module is big and not relevant to generic and lower-level Lua
functionality, then it's a strong candidate for separation. Example:
treesitter.txt
Otherwise, add them to lua.txt
For Nvim-owned docs, use the following strict subset of "vimdoc" to ensure
the help doc renders nicely in other formats (such as HTML:
https://neovim.io/doc/user ).
Strict "vimdoc" subset:
Use lists (like this!) prefixed with "-" or "•", for adjacent lines that you
don't want to auto-wrap. Lists are always rendered with "flow" layout
(soft-wrapped) instead of preformatted (hard-wrapped) layout common in
legacy :help docs.
Separate blocks (paragraphs) of content by a blank line.
Do not use indentation in random places—that prevents the page from using
"flow" layout. If you need a preformatted section, put it in
a
help-codeblock starting with ">".
Parameters and fields are documented as {foo}.
Optional parameters and fields are documented as {foo}?.
Nvim API documentation lives in the source code, as docstrings (doc
comments) on the function definitions. The
api :help is generated
from the docstrings defined in src/nvim/api/*.c.
Docstring format:
Lines start with ///
Special tokens start with @ followed by the token name:
@note, @param, @return
Markdown is supported.
Tags are written as [tag]().
References are written as [tag]
Use "```" for code samples.
Code samples can be annotated as vim or lua
Example: the help for
nvim_open_win() is generated from a docstring defined
in src/nvim/api/win_config.c like this:
/// Opens a new window.
/// ...
///
/// Example (Lua): window-relative float
///
/// ```lua
/// vim.api.nvim_open_win(0, false, {
/// relative='win',
/// row=3,
/// col=3,
/// width=12,
/// height=3,
/// })
/// ```
///
/// @param buffer Buffer to display
/// @param enter Enter the window
/// @param config Map defining the window configuration. Keys:
/// - relative: Sets the window layout, relative to:
/// - "editor" The global editor grid.
/// - "win" Window given by the `win` field.
/// - "cursor" Cursor position in current window.
/// ...
/// @param[out] err Error details, if any
///
/// @return Window handle, or 0 on error
Lua docstrings
dev-lua-doc
Lua documentation lives in the source code, as docstrings on the function
definitions. The
lua-vim :help is generated from the docstrings.
Docstring format:
Markdown is supported.
Tags are written as [tag]().
References are written as [tag]
Use "```" for code samples.
Code samples can be annotated as vim or lua
Use
@since <api-level> to note the
api-level when the function became
"stable". If
<api-level> is greater than the current stable release (or
0), it is marked as "experimental".
See scripts/util.lua for the mapping of api-level to Nvim version.
Use @nodoc to prevent documentation generation.
Use
@inlinedoc to inline
@class blocks into
@param blocks.
E.g.
--- Object with fields:
--- @class myOpts
--- @inlinedoc
---
--- Documentation for some field
--- @field somefield? integer
--- @param opts? myOpts
function foo(opts)
end
Will be rendered as:
foo({opts})
Parameters:
- {opts}? (table) Object with the fields:
- {somefield}? (integer) Documentation
for some field
Files declared as @meta are only used for typing and documentation (similar to "*.d.ts" typescript files).
Example: the help for
vim.paste() is generated from a docstring decorating
vim.paste in runtime/lua/vim/_editor.lua like this:
--- Paste handler, invoked by |nvim_paste()| when a conforming UI
--- (such as the |TUI|) pastes text into the editor.
---
--- Example: To remove ANSI color codes when pasting:
---
--- ```lua
--- vim.paste = (function()
--- local overridden = vim.paste
--- ...
--- end)()
--- ```
---
--- @since 12
--- @see |paste|
---
--- @param lines ...
--- @param phase ...
--- @returns false if client should cancel the paste.
STDLIB DESIGN GUIDELINES dev-lua
Keep the core Lua modules
lua-stdlib simple. Avoid elaborate OOP or
pseudo-OOP designs. Plugin authors just want functions to call, not a big,
fancy inheritance hierarchy.
Avoid requiring or returning special objects in the Nvim stdlib. Plain
tables or values are easier to serialize, easier to construct from literals,
easier to inspect and print, and inherently compatible with all Lua plugins.
(This guideline doesn't apply to opaque, non-data objects like vim.cmd.)
stdlib functions should follow these common patterns:
Accept iterable instead of only table.
Note: in some cases iterable doesn't make sense, e.g. spair() sorts the
input by definition, so there is no reason for it to accept an iterable,
because the input needs to be "reified"; it can't operate on a "stream".
Return an iterable (generator) instead of table, if possible.
Mimic the pairs() or ipairs() interface if the function is intended for
use in a
for-in loop.
dev-error-patterns
To communicate failure to a consumer, choose from these patterns (in order of
preference):
1.
retval, errmsg
When failure is normal, or when it is practical for the consumer to
continue (fallback) in some other way. See
lua-result-or-message.
2. optional result, no errormsg
Special case of 1. When there is only a single case of "doesn't exist"
(e.g. cache lookup, dict lookup).
3. error("no luck")
For invalid state ("must not happen"), when failure is exceptional, or at
a low level where the consumers are unlikely to handle it in a meaningful
way. Advantage is that propagation happens for free and it's harder to
accidentally swallow errors. (E.g. using
uv_handle/pipe:write() without checking return values is common.)
4. on_error callback
For async and "visitors" traversing a graph, where many errors may be
collected while work continues.
5. vim.notify (sometimes with optional opts.silent (async, visitors ^))
High-level / application-level messages. End-user invokes these directly.
Where possible, these patterns apply to _both_ Lua and the API:
When accepting a buffer id, etc., 0 means "current buffer", nil means "all
buffers". Likewise for window id, tabpage id, etc.
Any function signature that accepts a callback (example:
table.foreach())
should place it as the LAST parameter (after opts), if possible (or ALWAYS
for "continuation callbacks"—functions called exactly once).
Improves readability by placing the less "noisy" arguments near the start.
Consistent with luv.
Useful for future async lib which transforms functions of the form
function(<args>, cb(<ret)>)) => function(<args>) -> <ret>.
Example:
-- ✅ OK:
filter(…, opts, function() … end)
-- ❌ NO:
filter(function() … end, …, opts)
-- ❌ NO:
filter(…, function() … end, opts)
"Enable" ("toggle") interface and behavior:
enable(…, nil) and enable(…, {buf=nil}) are synonyms and control the
the "global" enablement of a feature.
is_enabled(nil) and is_enabled({buf=nil}), likewise, query the
global state of the feature.
enable(…, {buf: number}) sets a buffer-local "enable" flag.
is_enabled({buf: number}), likewise, queries the buffer-local state of
the feature.
Transformation functions should also have "filter" functionality (when
appropriate): when the function returns a nil value it excludes (filters
out) its input, else the transformed value is used.
When adding an API, check the following:
What precedents did you draw from? How does your solution compare to them?
Does your new API allow future expansion? How? Or why not?
Is the new API similar to existing APIs? Do we need to deprecate the old ones?
Did you cross-reference related concepts in the docs?
Avoid "mutually exclusive" parameters--via constraints or limitations, if
necessary. For example nvim_create_autocmd() has mutually exclusive
"callback" and "command" args; but the "command" arg could be eliminated by
simply not supporting Vimscript function names, and treating a string
"callback" arg as an Ex command (which can call Vimscript functions). The
"buffer" arg could also be eliminated by treating a number "pattern" as
a buffer number.
Avoid functions that depend on cursor position, current buffer, etc. Instead
the function should take a position parameter, buffer parameter, etc.
API (libnvim/RPC): exposes low-level internals, or fundamental things (such
as nvim_exec_lua()) needed by clients or C consumers.
Lua stdlib = high-level functionality that builds on top of the API.
Naming is exceedingly important: the name of a thing is the primary interface
for uses it, discusses it, searches for it, shares it... Consistent
naming in the stdlib, API, and UI helps both users and developers discover and
intuitively understand related concepts ("families"), and reduces cognitive
burden. Discoverability encourages code re-use and likewise avoids redundant,
overlapping mechanisms, which reduces code surface-area, and thereby minimizes
bugs...
In general, look for precedent when choosing a name, that is, look at existing
(non-deprecated) functions. In particular, see below...
dev-name-common
Use existing common
{verb} names (actions) if possible:
add: Appends or inserts into a collection
attach: Listens to something to get events from it (TODO: rename to "on"?)
call: Calls a function
cancel: Cancels or dismisses an event or interaction, typically
user-initiated and without error. (Compare "abort", which
cancels and signals error/failure.)
clear: Clears state but does not destroy the container
create: Creates a new (non-trivial) thing (TODO: rename to "def"?)
del: Deletes a thing (or group of things)
detach: Dispose attached listener (TODO: rename to "un"?)
enable: Enables/disables functionality. Signature should be
enable(enable?:boolean, filter?:table).
eval: Evaluates an expression
exec: Executes code, may return a result
fmt: Formats
get: Gets things. Two variants (overloads):
1. get<T>(id: int): T returns one item.
2. get<T>(filter: dict): T[] returns a list.
has: Checks for the presence of an item, feature, etc.
inspect: Presents a high-level, often interactive, view
is_enabled: Checks if functionality is enabled.
open: Opens something (a buffer, window, …)
parse: Parses something into a structured form
set: Sets a thing (or group of things)
start: Spin up a long-lived process. Prefer "enable" except when
"start" is obviously more appropriate.
stop: Inverse of "start". Teardown a long-lived process.
try_{verb}: Best-effort operation, failure returns null or error obj
Do NOT use these deprecated verbs:
contains: Prefer "has".
disable: Prefer enable(enable: boolean).
exit: Prefer "cancel" (or "stop" if appropriate).
is_disabled: Prefer is_enabled().
list: Redundant with "get"
notify: Redundant with "print", "echo"
show: Redundant with "print", "echo"
toggle: Prefer enable(not is_enabled()).
Use consistent names for
{topic} in API functions: buffer is called "buf"
everywhere, not "buffer" in some places and "buf" in others.
buf: Buffer
cmd: Command
cmdline: Command-line UI or input
fn: Function
hl: Highlight
pos: Position
proc: System process
tabpage: Tabpage
win: Window
Do NOT use these deprecated nouns:
buffer Use "buf" instead
callback Use on_foo instead
command Use "cmd" instead
window Use "win" instead
dev-name-events
Use the "on_" prefix to name event-handling callbacks and also the interface for
"registering" such handlers (on_key). The dual nature is acceptable to avoid
a confused collection of naming conventions for these related concepts.
Editor
events (autocommands) are historically named like:
{Noun}{Event}
Use this format to name API (RPC) events:
nvim_{noun}_{event-name}_event
Example:
nvim_buf_changedtick_event
dev-api-name
Use this format to name new RPC
API functions:
nvim_{topic}_{verb}_{arbitrary-qualifiers}
Do not add new nvim_buf/nvim_win/nvim_tabpage APIs, unless you are certain the
concept will NEVER be applied to more than one "scope". That is,
{topic}
should be the TOPIC ("ns", "extmark", "option", …) that acts on the scope(s)
(buf/win/tabpage/global), it should NOT be the scope. Instead the scope should
be a parameter (typically manifest as mutually-exclusive buf/win/… flags like
nvim_get_option_value(), or less commonly as a
scope: string field like
nvim_get_option_info2()).
Example: nvim_get_keymap('v') operates in a global context (first
parameter is not a Buffer). The "get" verb indicates that it gets anything
matching the given filter parameter. A "list" verb is unnecessary because
nvim_get_keymap('') (empty filter) returns all items.
Example: nvim_buf_del_mark acts on a Buffer object (the first parameter)
and uses the "del" {verb}.
dev-namespace-name
Use namespace names like
nvim.foo.bar:
vim.api.nvim_create_namespace('nvim.lsp.codelens')
dev-augroup-name
Use autocommand group names like
nvim.foo.bar:
vim.api.nvim_create_augroup('nvim.treesitter.dev')
Prefer adding a single nvim_{topic}_{verb}_… interface for a given topic.
Example:
nvim_ns_add(
ns_id: int,
filter: {
handle: integer (buf/win/tabpage id)
scope: "global" | "win" | "buf" | "tabpage"
}
): { ok: boolean }
nvim_ns_get(
ns_id: int,
filter: {
handle: integer (buf/win/tabpage id)
scope: "global" | "win" | "buf" | "tabpage"
}
): { ids: int[] }
nvim_ns_del(
ns_id: int,
filter: {
handle: integer (buf/win/tabpage id)
scope: "global" | "win" | "buf" | "tabpage"
}
): { ok: boolean }
Anti-Example:
Creating separate
nvim_xx,
nvim_buf_xx,
nvim_win_xx, and
nvim_tabpage_xx, functions all for the same
xx topic, requires 4x the
amount of documentation, tests, boilerplate, and interfaces, which users must
comprehend, maintainers must maintain, etc. Thus the following is NOT
recommended (compare these 12(!) functions to the above 3 functions):
nvim_add_ns(…)
nvim_buf_add_ns(…)
nvim_win_add_ns(…)
nvim_tabpage_add_ns(…)
nvim_del_ns(…)
nvim_buf_del_ns(…)
nvim_win_del_ns(…)
nvim_tabpage_del_ns(…)
nvim_get_ns(…)
nvim_buf_get_ns(…)
nvim_win_get_ns(…)
nvim_tabpage_get_ns(…)
api-client
API clients wrap the Nvim
API to provide idiomatic "SDKs" for their
respective platforms (see
jargon). You can build a new API client for your
favorite platform or programming language.
node-client pynvim
These clients can be considered the "reference implementation" for API clients:
API clients exist to hide msgpack-rpc details. The wrappers can be
automatically generated by reading the
api-metadata from Nvim.
api-mapping
Clients should call
nvim_set_client_info() after connecting, so users and
plugins can detect the client by handling the
ChanInfo event. This avoids
the need for special variables or other client hints.
Clients should handle
nvim_error_event notifications, which will be sent
if an async request to nvim was rejected or caused an error.
API client packages should NOT be named something ambiguous like "neovim" or
"python-client". Use "nvim" as a prefix/suffix to some other identifier
following ecosystem conventions.
For example, Python packages tend to have "py" in the name, so "pynvim" is
a good name: it's idiomatic and unambiguous. If the package is named "neovim",
it confuses users, and complicates documentation and discussions.
Examples of API-client package names:
✅ OK: nvim-racket
✅ OK: pynvim
❌ NO: python-client
❌ NO: neovim_
API client implementation guidelines
Separate the transport layer from the rest of the library.
rpc-connecting
Use a MessagePack library that implements at least version 5 of the
MessagePack spec, which supports the BIN and EXT types used by Nvim.
Use a single-threaded event loop library/pattern.
Use a fiber/coroutine library for the language being used for implementing
a client. These greatly simplify concurrency and allow the library to
expose a blocking API on top of a non-blocking event loop without the
complexity that comes with preemptive multitasking.
Don't assume anything about the order of responses to RPC requests.
Clients should expect requests, which must be handled immediately because
Nvim is blocked while waiting for the client response.
Clients should expect notifications, but these can be handled "ASAP" (rather
than immediately) because they won't block Nvim.
External UIs should be aware of the
api-contract. In particular, future
versions of Nvim may add new items to existing events. The API is strongly
backwards-compatible, but clients must not break if new (optional) fields are
added to existing events.
External UIs are expected to implement these common features:
Call
nvim_set_client_info() after connecting, so users and plugins can
detect the UI by handling the
ChanInfo event. This avoids the need for
special variables and UI-specific config files (gvimrc, macvimrc, …).
Cursor style (shape, color) should conform to the
'guicursor' properties
delivered with the mode_info_set UI event.
Send the ALT/META ("Option" on macOS) key as a
<M- chord.
Send the "super" key (Windows key, Apple key) as a
<D- chord.
Avoid mappings that conflict with the Nvim keymap-space; GUIs have many new
chords (<C-,> <C-Enter> <C-S-x> <D-x>) and patterns ("shift shift") that do
not potentially conflict with Nvim defaults, plugins, etc.
Consider the "option_set"
ui-global event as a hint for other GUI
behaviors. Various UI-related options (
'guifont',
'ambiwidth', …) are
published in this event. See also "mouse_on", "mouse_off".
UIs generally should NOT set
$NVIM_APPNAME (unless explicitly requested by
the user).
Support the text decorations/attributes given by
ui-event-hl_attr_define.
The "url" attr should be presented as a clickable hyperlink.
Handle the "restart" UI event so that
:restart works.