Problem:
`vim.lsp.util.rename()` deletes the buffers that are affected by
renaming. This has undesireable side effects. For example, when renaming
a directory, all buffers under that directory are deleted and windows
displaying those buffers are closed. Also, buffer options may change
after renaming.
Solution:
Rename the buffers with :saveas.
An alternative approach is to record all the relevant states and restore
it after renaming, but that seems to be more complex. In fact, the older
version was attempting to restore the states but only partially and
incorrectly.
Context:
Nvim catches errors from the user's `on_exit` and rpc handler callbacks
and prints the error message.
Problem:
Printing the error message uses Nvim api functions. But callbacks
mentioned above run in `:h lua-loop-callbacks` where most of `vim.api`
is not allowed, so Nvim itself raises error.
Solution:
`vim.schedule()` the error reporting when necessary.
- Added `@inlinedoc` so single use Lua types can be inlined into the
functions docs. E.g.
```lua
--- @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
```
- Marked many classes with with `@nodoc` or `(private)`.
We can eventually introduce these when we want to.
Problem:
vim._watch.watchdirs has terrible performance.
Solution:
- On linux use fswatch as a watcher backend if available.
- Add File watcher section to health:vim.lsp. Warn if watchfunc is
libuv-poll.
Previously rename would unconditionally read the to-be-renamed file from the
disk and write it to the disk. This is redundant in some cases
If the file is not already loaded, it's not attached to lsp client, so nvim
doesn't need to care about this file.
If the file is loaded but has no change, it doesn't need to be written.
Problem:
The documentation flow (`gen_vimdoc.py`) has several issues:
- it's not very versatile
- depends on doxygen
- doesn't work well with Lua code as it requires an awkward filter script to convert it into pseudo-C.
- The intermediate XML files and filters makes it too much like a rube goldberg machine.
Solution:
Re-implement the flow using Lua, LPEG and treesitter.
- `gen_vimdoc.py` is now replaced with `gen_vimdoc.lua` and replicates a portion of the logic.
- `lua2dox.lua` is gone!
- No more XML files.
- Doxygen is now longer used and instead we now use:
- LPEG for comment parsing (see `scripts/luacats_grammar.lua` and `scripts/cdoc_grammar.lua`).
- LPEG for C parsing (see `scripts/cdoc_parser.lua`)
- Lua patterns for Lua parsing (see `scripts/luacats_parser.lua`).
- Treesitter for Markdown parsing (see `scripts/text_utils.lua`).
- The generated `runtime/doc/*.mpack` files have been removed.
- `scripts/gen_eval_files.lua` now instead uses `scripts/cdoc_parser.lua` directly.
- Text wrapping is implemented in `scripts/text_utils.lua` and appears to produce more consistent results (the main contributer to the diff of this change).
Assert that the buffer number passed to apply_text_edits is fully
resolved (not 0 or null). Pass the known buffer number to
apply_text_edits from lsp.formatexpr().
Previously the LSP-Client object contained some fields that are also
in the client config, but for a lot of other fields, the config was used
directly making the two objects vaguely entangled with either not having
a clear role.
Now the config object is treated purely as config (read-only) from the
client, and any fields the client needs from the config are now copied
in as additional fields.
This means:
- the config object is no longet normalised and is left as the user
provided it.
- the client only reads the config on creation of the client and all
other implementations now read the clients version of the fields.
In addition, internal support for multiple callbacks has been added to
the client so the client tracking logic (done in lua.lsp) can be done
more robustly instead of wrapping the user callbacks which may error.
The dispatchers used by the RPC client should be defined in the client,
so they have been moved there. Due to this, it also made sense to move
all code related to client configuration and the creation of the RPC
client there too.
Now vim.lsp.start_client is significantly simplified and now mostly
contains logic for tracking open clients.
- Renamed client.new -> client.start
Problem:
If a rename results in a path that has missing parent directory(s), it
will fail.
Solution:
Do a recursive mkdir before attempting the rename.
Problem:
`vim.lsp.diagnostic.on_diagnostic` accepts an undocumented severity_limit
option which is widely used.
Solution:
Deprecate it in favour of `{min = severity}` used in `vim.diagnostic`.
Since this is undocumented, the schedule for removal is accelerated to
0.11.
Problem:
The LSP client code is implemented as a complicated closure-class
(class defined in a single function).
Solution:
Move LSP client code to a more conventional Lua class and move to a
separate file.
Problem: vim.lsp.tagfunc() causes an infinite loop.
This is a bug happened while introducing deferred loading.
Solution: Rename the private module to `vim.lsp._tagfunc`.
The benefit of this is that users only pay for what they use. If e.g.
only `vim.lsp.buf_get_clients()` is called then they don't need to load
all modules under `vim.lsp` which could lead to significant startuptime
saving.
Also `vim.lsp.module` is a bit nicer to user compared to
`require("vim.lsp.module")`.
This isn't used for some nested modules such as `filetype` as it breaks
tests with error messages such as "attempt to index field 'detect'".
It's not entirely certain the reason for this, but it is likely it is
due to filetype being precompiled which would imply deferred loading
isn't needed for performance reasons.
This function is used only in the `workspace/configuration` handler,
and does not warrant a public API because of its confusing return types.
The only caller `vim.lsp.handlers["workspace.configuration"]` is also
refactored to use `vim.tbl_get()` instead.
Typings introduced in #26032 and #26552 have a few conflicts, so we
merge and clean them up. We also fix some incorrect type annotation in
the `vim.lsp.rpc` package. See the associated PR for more details.
Summary:
- vim.rpc.Dispatchers -> vim.lsp.rpc.Dispatchers
- vim.lsp.rpc.Error -> lsp.ResponseError
- Revise docs
Problem:
Currently `deepcopy` hashes every single tables it copies so it can be
reused. For tables of mostly unique items that are non recursive, this
hashing is unnecessarily expensive
Solution:
Port the `noref` argument from Vimscripts `deepcopy()`.
The below benchmark demonstrates the results for two extreme cases of
tables of different sizes. One table that uses the same table lots of
times and one with all unique tables.
| test | `noref=false` (ms) | `noref=true` (ms) |
| -------------------- | ------------------ | ----------------- |
| unique tables (50) | 6.59 | 2.62 |
| shared tables (50) | 3.24 | 6.40 |
| unique tables (2000) | 23381.48 | 2884.53 |
| shared tables (2000) | 3505.54 | 14038.80 |
The results are basically the inverse of each other where `noref` is
much more performance on tables with unique fields, and `not noref` is
more performant on tables that reuse fields.
This reverts commit 5cb906e91c.
They were intentionally fast-tracked.
- `parse_snippet()` because of limited scope, and given that it's kinda
semi-broken (arbitrary formatting rules, not that useful for what it
was used for)
- `extract_completion_items()` doesn't work if we want to add the LSP
completionlist capability
- `text_document_completion_list_to_complete_items()` also doesn't work
for completionlist
Style improvements:
1. Anonymous classes derived from `StructureLiteralType` should have a
better name. The class name can be also nested. Examples:
```diff
----@field serverInfo? anonym1
+---@field serverInfo? lsp._anonym1.serverInfo
```
```diff
----@field insertTextModeSupport? anonym26
+---@field insertTextModeSupport? lsp._anonym26.completionItem.insertTextModeSupport
```
2. Add one separate empty line before each `@field` definition. Without
these, empty lines the doc can look confusing because descriptions
also may contain empty lines. See `lsp.CompletionItem` for example:
```lua
---The kind of this completion item. Based of the kind
---an icon is chosen by the editor.
---@field kind? lsp.CompletionItemKind
---Tags for this completion item.
---
---@since 3.15.0
---@field tags? lsp.CompletionItemTag[]
```
It might feel like "Tags for this completion item" belongs to `kind`,
not `tags` due to the lack of separator blank lines. The following
(after this commit) should look much better:
```diff
---The kind of this completion item. Based of the kind
---an icon is chosen by the editor.
---@field kind? lsp.CompletionItemKind
+---
---Tags for this completion item.
---
---@since 3.15.0
---@field tags? lsp.CompletionItemTag[]
```
3. Escape some LSP-specific annotations that can't be recognized by
lua-ls. It'd be better to make them visible in LSP hover doc windows.
Example: `@sample ...`.
Fixes:
1. A type may extend from more than one base types (as well as mixin
types). Previously only the first base class was being considered,
resulting incomplete base classes for `@class` definitions.
Example: `InlayHintOptions` (should have both of `resolveProvider`
and `workDoneProgress`, the latter is from `WorkDoneProgressOptions`)
```diff
----@class lsp.InlayHintOptions
+---@class lsp.InlayHintOptions: lsp.WorkDoneProgressOptions
```
2. Remove `<200b>` (zero-width space) unicode characters.
3. Add the missing newline at EOF.
The purpose of this commit is to make diff clean and easy to read; to
see the diff resulted from actual changes in gen_lsp.lua, not from the
updated LSP protocol JSON data.
Ran: `nvim -l scripts/gen_lsp.lua gen --methods`
Based on 3.18.0 (2023-12-23)
The following functions should be removed in 0.12 according to the
deprecation strategy in MAINTAIN.md:
- vim.lsp.util.extract_completion_items()
- vim.lsp.util.parse_snippet()
- vim.lsp.util.text_document_completion_list_to_complete_items()
Use the get_language_id client option to resolve the filetype when
matching the document selector in a dynamic capability.
Co-authored-by: Mathias Fussenegger <f.mathias@zignar.net>
refactor(lsp): move glob parsing to vim.glob
Moving the logic for using vim.lpeg to create a match pattern from a
glob into `vim.glob`. There are several places in the LSP spec that
use globs, and it's very useful to have glob matching as a
generally-available utility.
- Improve CLI argument parsing, rejects invalid argument and commands as
early as possible. Also prints USAGE in the command line.
- No longer allows `--<outfile>`, use `--out <outfile>` instead.
- Print a little bit of verbose messages to better know what's going on
rather than remaining silent at all times.
- Add type annotation `gen_lsp._opt` to avoid type warnings.
refactor!: `vim.lsp.inlay_hint()` -> `vim.lsp.inlay_hint.enable()`
Problem:
The LSP specification allows inlay hints to include tooltips, clickable
label parts, and code actions; but Neovim provides no API to query for
these.
Solution:
Add minimal viable extension point from which plugins can query for
inlay hints in a range, in order to build functionality on top of.
Possible Next Steps
---
- Add `virt_text_idx` field to `vim.fn.getmousepos()` return value, for
usage in mappings of `<LeftMouse>`, `<C-LeftMouse>`, etc
`code_action` used the same parameters for all clients, which led to the
following warning and incorrect start/end column locations if using
clients with mixed encodings:
warning: multiple different client offset_encodings detected for
buffer, this is not supported yet
Problem: Currently there is no way of customizing behavior of
`declaration`, `definition`, `typeDefinition`, and `implementation`
methods in `vim.lsp.buf` when LSP server returns `Location`. Instead,
cursor jumps to that location directly.
Solution: Normalize LSP response to be `Location[]` for those four cases.
Fixes a regression from 5e5f5174e3
Until that commit we had a logic like this:
`local prefix = startbyte and line:sub(startbyte + 1) or line_to_cursor:sub(word_boundary)`
The commit changed the logic and no longer cut off the line at the cursor, resulting in a prefix that included trailing characters
Fixes https://github.com/neovim/neovim/issues/25177
I initially wanted to split this into a refactor commit to make it more
testable, but it appears that already accidentally fixed the issue by
normalizing lnum/col to 0-indexing
To be more in line with the specification:
> To support the evolution of enumerations the using side of an enumeration shouldn’t fail on an enumeration value it doesn’t know. It should simply ignore it as a value it can use and try to do its best to preserve the value on round trips
Follw up to 63b3408551
`is_pull` should be optional, otherwise it is an API change that
introduces warnings in consumers.
Also fixes the type annotation of `_client_pull_namespaces` where the
key is a string.
The haskell-language-server supports resolve only for a subset of code
actions. For many code actions trying to resolve the `edit` property
results in an error, but the unresolved action already contains a
command that can be executed without issue.
The protocol specification is unfortunately a bit vague about this,
and what the haskell-language-server does seems to be valid.
Example:
newtype Dummy = Dummy Int
instance Num Dummy where
Triggering code actions on "Num Dummy" and choosing "Add placeholders
for all missing methods" resulted in:
-32601: No plugin enabled for SMethod_CodeActionResolve, potentially available: explicit-fields, importLens, hlint, overloaded-record-dot
With this change it will insert the missing methods:
instance Num Dummy where
(+) = _
(-) = _
(*) = _
negate = _
abs = _
signum = _
fromInteger = _
Problem:
Users using `vim.lsp.start` directly (instead of nvim-lspconfig) need
more visibility for troubleshooting. For example, troubleshooting
unnecesary servers or servers that aren't attaching to expected buffers.
Solution:
Mention attached buffers in the `:checkhealth lsp` report.
Example:
vim.lsp: Active Clients ~
- clangd (id=1, root_dir=~/dev/neovim, attached_to=[7])
- lua_ls (id=2, root_dir=~/dev/neovim, attached_to=[10])
Fixes#24339
rust-analyzer sends "Invalid offset" error in such cases. Some other
servers handle it specially.
LSP spec mentions that "A range is comparable to a selection in an
editor". Most editors don't handle trailing newlines the same way
Neovim/Vim does, it's clearly visible if it's present or not. With that
in mind it's understandable why sending end position as simply the start
of the line after the last one is considered invalid in such cases.
This fixes the issue where the LspNotify handlers for inlay_hint /
diagnostics would end up refreshing all attached clients.
The handler would call util._refresh, which called
vim.lsp.buf_request, which calls the method on all attached clients.
Now util._refresh takes an optional client_id parameter, which is used
to specify a specific client to update.
This commit also fixes util._refresh's handling of the `only_visible`
flag. Previously if `only_visible` was false, two requests would be made
to the server: one for the visible region, and one for the entire file.
Co-authored-by: Stanislav Asunkin <1353637+stasjok@users.noreply.github.com>
Co-authored-by: Mathias Fußenegger <mfussenegger@users.noreply.github.com>
Problem: luals returns stricter diagnostics with bundled luarc.json
Solution: Improve some function and type annotations:
* use recognized uv.* types
* disable diagnostic for global `vim` in shared.lua
* docs: don't start comment lines with taglink (otherwise LuaLS will interpret it as a type)
* add type alias for lpeg pattern
* fix return annotation for `vim.secure.trust`
* rename local Range object in vim.version (shadows `Range` in vim.treesitter)
* fix some "missing fields" warnings
* add missing required fields for test functions in eval.lua
* rename lsp meta files for consistency
PR #23689 assumes `client.config.capabilities.workspace.didChangeWatchedFiles`
exists when checking `dynamicRegistration`, but thats's true only if it was
passed to `vim.lsp.start{_client}`.
This caused #23806 (still an issue in v0.9.1; needs manual backport), but #23681
fixed it by defaulting `config.capabilities` to `make_client_capabilities` if
not passed to `vim.lsp.start{_client}`.
However, the bug resurfaces on HEAD if you provide a non-nil `capabilities` to
`vim.lsp.start{_client}` with missing fields (e.g: not made via
`make_client_capabilities`).
From what I see, the spec says such missing fields should be interpreted as an
absence of the capability (including those indicated by missing sub-fields):
https://microsoft.github.io/language-server-protocol/specifications/lsp/3.17/specification/#clientCapabilities
Also, suggest `vim.empty_dict()` for an empty dict in
`:h vim.lsp.start_client()` (`{[vim.type_idx]=vim.types.dictionary}`
no longer works anyway, probably since the cjson switch).
Problem:
- Notifications are missing from `lsp.Methods`.
- Need a way to represent `$/` prefixed methods.
Solution:
- Generate notifications.
- Use "dollar_" prefix for `$/` methods.
Problem:
"Failed to delete autocmd" error when deleting LspNotify autocmd. #24456
Solution:
Change a few things in the inlay_hint and diagnostic LSP code:
1. Re-introduce the `enabled` flag for the buffer state tables. Previously I was
relying on the presence of an autocmd id in the state table to track whether
inlay_hint / diagnostic was enabled for a buffer. There are two reasons why
this doesn't work well:
- Each time inlay_hint / diagnostic is enabled, we call `nvim_buf_attach` on
the buffer, resulting in multiple `on_reload` or `on_detach` callbacks being
registered.
- Commands like `bwipeout` delete buffer local autocmds, sometimes before our
`on_detach` callbacks have a chance to delete them first. This causes the
- Use module local enabled state for diagnostic as well. bwipeout can race
with on_detach callbacks for deleting autocmds. Error referenced in #24456.
2. Change the `LspDetach` autocmd to run each time (i.e., remove the `once`
flag). Since we're only registering autocmds once per buffer now, we
need to make sure that we set the enabled flag properly each time the LSP
client detaches from the buffer.
- Remove `once` from the LspDetach autocmds for inlay_hint and diagnostic.
We only set up the autocmd once now. Gets removed when buffer is deleted.
3. Have the `LspNotify` handler also refresh the inlay_hint / diagnostics when
receiving the `textDocument/didOpen` event. Before this point, the LSP
backend doesn't have the contents of the buffer, so can't provide inlay hints
or diagnostics.
Downsides of this approach:
* When inlay_hint / diagnostics are disabled on a buffer, it will continue to
receive `LspNotify` events for that buffer. The callback exits early since the
`enabled` flag is false.
Alternatives:
* Can we wrap the call to `nvim_del_autocmd` in `pcall` to swallow any errors
resulting from trying to delete the autocmd?
Fixes#24456
Helped-by: Maria José Solano <majosolano99@gmail.com>
Problem:
Content that has codeblocks with different languages, results in
multiple calls to:
syntax include vim syntax/vim.vim
which raises error:
E403: syntax sync: line continuations pattern specified twice
Before ba8f19ebb6, this was avoided by
using pcall() to ignore the error.
Solution:
Restore the use of pcall() to ignore the error.
We plan to replace this logic with a treesitter approach, so this is
good enough for now.
Fix#24431
* fix(lsp): replace @private with @nodoc for public client functions
To prevent lua-ls warnings in plugins which use the functions.
* fix(lsp): remove duplicate type annotations/class definitions
These annotations became duplicate with https://github.com/neovim/neovim/pull/23750
When adding `workspace/didChangeWorkspaceFolders` support to my [language server](https://github.com/elixir-tools/next-ls), I noticed that when neovim removes a workspace, it sends an empty table (which is serialized to an empty JSON array) for the value in the `added` field.
This does not follow the spec; the `added` table should just be empty.
The following error led me to this discovery. Note the payload includes `"added" => [[]]`:
```
22:46:48.476 [error] LSP Exited.
Last message received: handle_notification %{"jsonrpc" => "2.0", "method" => "workspace/didChangeWorkspaceFolders", "params" => %{"event" => %{"added" => [[]], "removed" => [%{"name" => "/Users/mitchell/src/gen_lsp", "uri" => "file:///Users/mitchell/src/gen_lsp"}]}}}
** (MatchError) no match of right hand side value: {:error, %{"params" => %{"event" => %{"added" => [error: "expected a map"]}}}}
(gen_lsp 0.4.0) lib/gen_lsp.ex:265: anonymous fn/4 in GenLSP.loop/3
(gen_lsp 0.4.0) lib/gen_lsp.ex:292: GenLSP.attempt/3
(stdlib 5.0.2) proc_lib.erl:241: :proc_lib.init_p_do_apply/3
```
In the case you hit this warning in a buffer (like with C++ and clangd),
this message potentially fires over and over again making it difficult
to use the editor at all.
* docs(lua): teach lua2dox how to table
* docs(lua): teach gen_vimdoc.py about local functions
No more need to mark local functions with @private
* docs(lua): mention @nodoc and @meta in dev-lua-doc
* fixup!
Co-authored-by: Justin M. Keyes <justinkz@gmail.com>
---------
Co-authored-by: Justin M. Keyes <justinkz@gmail.com>
Problem:
Bash language server returns "hover" markdown content that starts with
a code fence and info string of `man` preceded by whitespace, which Nvim
does not render properly.
See 0ee73c53ce/server/src/server.ts (L821C15-L821C15)
```typescript
function getMarkdownContent(documentation: string, language?: string): LSP.MarkupContent {
return {
value: language
? // eslint-disable-next-line prefer-template
['``` ' + language, documentation, '```'].join('\n')
: documentation,
kind: LSP.MarkupKind.Markdown,
}
}
```
For example,
```
``` man
NAME
git - the stupid content tracker
```
```
If I remove the white space, then it is properly formatted.
```
```man instead of ``` man
```
Per CommonMark Spec https://spec.commonmark.org/0.30/#info-string
whitespace is allowed before and after the `info string` which
identifies the language in a codeblock.
> The line with the opening code fence may optionally contain some text
> following the code fence; this is trimmed of leading and trailing
> spaces or tabs and called the [info
> string](https://spec.commonmark.org/0.30/#info-string). If the [info
> string](https://spec.commonmark.org/0.30/#info-string) comes after
> a backtick fence, it may not contain any backtick characters. (The
> reason for this restriction is that otherwise some inline code would
> be incorrectly interpreted as the beginning of a fenced code block.)
Solution:
Adjust stylize_markdown() to allow whitespace before codeblock info.
Problem: `hl_mode` for inlay hints is `combine`, causing bugs like
inlay hints using highlights from the previous character
(#24152, #24068)
Solution: Don't use hl_mode=combine for inlay hints.
Problem:
Showing an error via vim.notify() makes it awkward for callers such as
lsp/handlers.lua to avoid showing redundant errors.
Solution:
Return the message instead of showing it. Let the caller decide whether
and when to show the message.
---
Rejected experiment: move vim.ui.open() to vim.env.open()
Problem:
`vim.ui` is where user-interface "providers" live, which can be
overridden. It would also be useful to have a "providers" namespace for
platform-specific features such as "open", clipboard, python, and the other
providers listed in `:help providers`. We could overload `vim.ui` to
serve that purpose as the single "providers" namespace, but
`vim.ui.nodejs()` for example seems awkward.
Solution:
`vim.env` currently has too narrow of a purpose. Overload it to also be
a namespace for `vim.env.open`.
diff --git a/runtime/lua/vim/_meta.lua b/runtime/lua/vim/_meta.lua
index 913f1fe20348..17d05ff37595 100644
--- a/runtime/lua/vim/_meta.lua
+++ b/runtime/lua/vim/_meta.lua
@@ -37,8 +37,28 @@ local options_info = setmetatable({}, {
end,
})
-vim.env = setmetatable({}, {
- __index = function(_, k)
+vim.env = setmetatable({
+ open = setmetatable({}, {
+ __call = function(_, uri)
+ print('xxxxx'..uri)
+ return true
+ end,
+ __tostring = function()
+ local v = vim.fn.getenv('open')
+ if v == vim.NIL then
+ return nil
+ end
+ return v
+ end,
+ })
+ },
+ {
+ __index = function(t, k, ...)
+ if k == 'open' then
+ error()
+ -- vim.print({...})
+ -- return rawget(t, k)
+ end
local v = vim.fn.getenv(k)
if v == vim.NIL then
return nil
Problem: in #24046 the signature of buf.clear_references() changed, which
indirectly breaks callers that were passing "ignored" args.
Solution: because util.buf_clear_references() already defaulted to "current buffer",
the change to buf.clear_references() isn't actually needed, so just revert it.
- fix lint / analysis warnings
- locations_to_items(): get default offset_encoding from active client
- character_offset(): get default offset_encoding from active client
Commit 37079fc moved inlay_hint to vim.lsp() but in the process did
missed converting a call to disable/enable which are now local.
Fixes the below error when trying to toggle inlay hints.
E5108: Error executing lua /usr/local/share/nvim/runtime/lua/vim/lsp/inlay_hint.lua:248: attempt to call field 'disable' (a nil value)
stack traceback:
/usr/local/share/nvim/runtime/lua/vim/lsp/inlay_hint.lua:248: in function 'toggle'
/usr/local/share/nvim/runtime/lua/vim/lsp/inlay_hint.lua:310: in function 'inlay_hint'
[string ":lua"]:1: in main chunk
Problem:
vim_lsp_inlayhint: Error executing lua: .../lsp/_inlay_hint.lua:249: attempt to index field 'applied' (a nil value)
Solution:
Assign {} to bufstates.applied in on_reload
fixes#24172
Problem:
On running `zig fmt` manually, the on_lines callback and the
server both detach (for some reason), and both of them call
`clear()`. This fixes it, otherwise the second one to detach
has an error in `reset_timer` since the bufstate doesn't exist
Solution:
* exit early in clear if `bufstates[bufnr]` is nil
* set bufstatte.enabled to true on reload instead of making bufstate nil
Problem:
The decoration provider clears the whole buffer then redraws all the hints every
time the window was redrawn. This may lead to an infinite loop.
Solution:
Store the last applied version for a line and only clear and redraw the line if
the buffer version has changed.
