The point of ghost transactions is to avoid notifying language servers
about changes since the change is meant to be temporary. This is used
for completion while selecting items in the menu: updating the language
server would mess up incomplete completions.
When a document is changed by a ghost transaction the language server
will not be notified so its understanding of the document will not be
synchronized and any positions it sends may be out-of-date. So we should
avoid triggering a request for new document color information when a
document is changed by a ghost transaction.
Previously the `call` helper (and its related functions) returned a
`serde_json::Value` which was then decoded either later in the client
(see signature help and hover) or by the client's caller. This led to
some unnecessary boilerplate in the client:
let resp = self.call::<MyRequest>(params);
Some(async move { Ok(serde_json::from_value(resp.await?)?) })
and in the caller. It also allowed for mistakes with the types. The
workspace symbol request's calling code for example mistakenly decoded a
`lsp::WorkspaceSymbolResponse` as `Vec<lsp::SymbolInformation>` - one of
the untagged enum members (so it parsed successfully) but not the
correct type.
With this change, the `call` helper eagerly decodes the response to a
request as the `lsp::request::Request::Result` trait item. This is
similar to the old helper `request` (which has become redundant and has
been eliminated) but all work is done within the same async block which
avoids some awkward lifetimes. The return types of functions like
`Client::text_document_range_inlay_hints` are now more verbose but it is
no longer possible to accidentally decode as an incorrect type.
Additionally `Client::resolve_code_action` now uses the `call_with_ref`
helper to avoid an unnecessary clone.
This adds events for:
* a document being opened
* a document being closed
* a language server sending the initialized notification
* a language server exiting
and also moves some handling done for these scenarios into hooks,
generally moving more into helix-view. A hook is also added on
`DocumentDidChange` which sends the `text_document_did_change`
notification - this resolves a TODO in `document`.
This resolves a TODO in the core diagnostic module to refactor this
type. It was originally an alias of `LanguageServerId` for simplicity.
Refactoring as an enum is a necessary step towards introducing
"internal" diagnostics - diagnostics emitted by core features such as
a spell checker. Fully supporting this use-case will require further
larger changes to the diagnostic type, but the change to the provider
can be made first.
Note that `Copy` is not derived for `DiagnosticProvider` (as it was
previously because `LanguageServerId` is `Copy`). In the child commits
we will add the `identifier` used in LSP pull diagnostics which is a
string - not `Copy`.
This is the same change as 1c9a5bd366 but for:
* document symbols
* workspace symbols
* goto definition/declaration/.../references
* hover
Instead of bailing when one server fails, we log an error and continue
gathering items from the other responses.
When requesting code actions from multiple LSP servers,
rather than bailing as soon as an error is encountered,
instead log the error and then keep going so that successful
requests can be presented to the user.
This is a minor move that will make future refactors of code actions
simpler. We should be able to move nearly all code action functionality
into `helix-view`, save UI stuff like the `menu::Item` implementation
and dealings with the compositor.
if `config.toml` either does not have `editor.true-color` or sets
it to false, many (most?) themes stop being usable. when loading such a
theme, Helix falls back to the default theme, but didn't mention this
anywhere - even in `~/.cache/helix/helix.log` when run with `-v`.
if this occurs when reloading a theme at runtime with `:theme`, there's
a fairly helpful error about
> `theme requires true color support`
seems worth logging about this if it happens during startup too.
This matches the behavior described by the EditorConfig spec for its
`insert_final_newline` option:
> Editors must not insert newlines in empty files when saving those
> files, even if `insert_final_newline = true`.
Co-authored-by: Axlefublr <101342105+Axlefublr@users.noreply.github.com>
These match the equivalent options in VSCode. `trim_trailing_whitespace`
is also the name used by EditorConfig.
* `trim-final-newlines` trims any extra line endings after the final one
* `trim-trailing-whitespace` trims any trailing whitespace (but not
empty lines)
Render rulers before the cursor to ensure that the cursor, when over
a ruler, is not hidden from view. Without this, you typically end up
with 1) foreground text that is the same as the background if the
ruler doesn't already have a foreground and 2) no visible cursor,
because the ruler's background color took precedence. By moving the
rulers before the cursor, this ensures that the theme is still rendered
more or less the way one would visually expect things to turn out.
This follows a pattern used in the signature help request for example.
Moving the json deserialization into the return future of
`text_document_hover` makes the types easier for callers to work with.
The info log within `process_request_response` duplicated the body of
the JSON message printed earlier by the transport which was confusing.
The error log in the completion handler was easy to hit during normal
use and is not actually an error - dropping is the graceful way to
handle changes occurring while completion requests are in flight.
This is not consequential now but when we switch to the new highlighter
we will want the theme to be set (and the loader's `scopes` to be set
based on the theme) before parsing a document. Previously `set_theme`
came after the loading of documents, so documents would be missing
locals highlights after being loaded and before the first edit.
`:lsp-stop` should consider only the set of active language servers for
a document. `:lsp-restart` though may be used to start up a language
server that crashed or was manually stopped, so it needs to consider the
language servers in config instead.
This change inlines the `valid_lang_servers` function into `:lsp-stop`
and `:lsp-restart` and changes `:lsp-restart` to check the doc's config
rather than active language servers. `:lsp-restart` now also does not
need to clone the language server names as strings since it borrows from
the config and arguments rather than `Document`. The completer has also
been split into two - one matching active language servers, used by
`:lsp-stop`, and the other matching configured language servers, used by
`:lsp-restart`.
This also removes the part of `:lsp-restart` which bailed if a language
server failed to be restarted (for example because it is not installed).
There might be multiple language servers configured for a language and
only one installed. In that case the `:lsp-restart` should be considered
successful even if not all servers could be started. Bailing prevented
any language servers which could start from being attached to the
document. Instead errors are collected and emitted at the end.
Previously all register selection info boxes had "Registers" as the
title. That was particularly confusing for `copy_between_registers`
which presents two info boxes back-to-back.
The application held onto these since their introduction in ce97a2f0 but
the Arcs are duplicated between Application and Editor - we can store it
only on Editor without issue.
There is no functional change to the move - it's just moving the code
into helix-view under a new method `Editor::handle_lsp_diagnostics` -
thought there is a typo fix, the removal of an unnecessary clone (for
the document's language config) and the removal of some nesting.
Co-authored-by: Sofus Addington <sofus@addington.dk>
This causes the infobox to disappear even when you type a non-character
key like escape. For example `"<esc>` now clears the infobox where
before it was left hanging.
Path completion items always have documentation but future core (i.e.
non-LSP) completions may not always have documentation - for example
word completion from the current buffer.
Completions are not specific to LSP anymore. In the child commits we
will expand on the types in this module so this refactor is done
eagerly to minimize changes later.
Previously LSP notifications were sent within a future and most callers
used a `tokio::spawn` to send the notification, silently discarding any
failures like problems serializing parameters or sending on the channel.
It's possible that tokio could schedule futures out of intended order
though which could cause notifications where order is important, like
document synchronization, to become partially shuffled. This change
removes the future wrapper and logs all internal failures.
Also included in this commit is the same change for `Client::reply`
which was also unnecessarily wrapped in a future.
Co-authored-by: Pascal Kuthe <pascalkuthe@pm.me>
This improves the display of the keymap popup for example, so that if
you bind a key like `C-x = ":buffer-close"` under the `<space>` menu,
the infobox shows "Close the current buffer." rather than `:buffer-close
[]`.
This simplifies the hover component by eagerly converting all
`lsp::Hover` responses into `Markdown`s. Previously we cached the
current `Markdown` and created a new `Markdown` when switching the
active response. Instead we can consume the `lsp::Hover` and avoid some
clones of its inner types.
The Hover component is used as the inner contents of a Popup. The Popup
should be doing calculations based on whether popup_borders is
configured and not Hover. This fixes an issue with hover rendering when
the popup border option is enabled for popups.
Fixes#12742
The language server may return `None` for a definition/reference
request. The parent commits introduced a regression for these commands
when a server did not provide locations. With this change a server may
respond with `null` and its locations will instead not be considered.
Fixes#12732
This covers all goto-definition-like commands: declaration, definition,
type definition and implementation.
Closes#11689
Co-authored-by: j <junglerobba@jngl.one>
<https://github.com/helix-editor/helix/pull/11486> introduced a Location
type in the LSP commands module which unified helpers like
`jump_to_location`. This change moves `OffsetEncoding` onto that type.
`SymbolInformationItem` and `PickerDiagnostic` already had fields for
carrying the offset encoding. We would want a similar setup for goto
definition/references as well (for supporting multiple language servers
with that feature) but those use the `Location` type. By moving
`OffsetEncoding` onto `Location` we make future changes to allow
mulitple language servers possible for goto definition/references
features and also simplify some calls for symbols and diagnostics.
The positions passed to `Transaction::change_by_selection` should be
character indexes. `edit_diff` is meant to track the number of
characters that should be deleted to erase the file name that has been
typed so far (if any). Mistakenly this was using `str::len` which is
the byte count. This fixes a bug that could cause more text to be
deleted than intended or a panic when completing a directory with
multi-byte characters like 'éclair'.
This change also moves the `edit_diff` binding out of the loop since
it's now performing some non-trivial work (counting characters, where
before it was just accessing the pre-computed number of bytes).
Previously you could use `<A-a><A-b>` to jump to a label "ab". We should
not treat characters with modifiers the same as characters without.
With this change the `<A-a>` input exits out of the jumping on-next-key.
Fixes#12695
The new `RopeSliceExt::ceil_char_boundary` from the parent commits can
be used to implement `RopeSliceExt::byte_to_next_char` when used with
`RopeSlice::byte_to_char`. That function had only one caller and that
caller will eventually disappear when we switch to Ropey v2 and drop
character indexing, so we can drop `byte_to_next_char` now and replace
its caller with `byte_to_char` plus `ceil_char_boundary`.
This change keeps the unit tests for `byte_to_next_char` and checks them
against a polyfill of `byte_to_char` plus `ceil_char_boundary` to ensure
that `byte_to_next_char`'s intended behavior is not changed.
This is a good example use-case of the `floor_char_boundary` and
`ceil_char_boundary` functions added in the parent commit. In the
single-width, single-selection case in `goto_file` we cap the search
to either the current line or 1000 bytes before or after the cursor
(whichever case comes earlier). That byte index might not lie on a
character boundary so it needs to be fixed to either the prior or
later boundary.
* fix: panic when pressing `*` at the end of the file
chore: remove incorrect additions
* docs: add info comment
* test: add new syntax to add a selection at the final character
* test: `*` panics when after the last char
* test: move into a more appopriate module
* test: fix failing
* test: account for Windows test suite
* test: choose a different strategy for custom syntax
* test: do not modify the syntax
* style: remove newline
---------
Co-authored-by: Nikita Revenco <154856872+NikitaRevenco@users.noreply.github.com>
This should help debug formatting failures when using external
formatters in the future. Previously we didn't log anything when an
external formatter failed despite having a custom error type for it.
This is partially a style commit:
* Pull more bindings out the `change_by_selection` closure like the
line-ending string and the comment tokens used for continuation.
* Prefer `Editor::config` to `Document`'s config.
The rest is changes to places where `insert_newline` may allocate.
The first is to move `new_text` out of the `change_by_selection`
closure, reusing it between iterations. This is not necessarily always
an improvement as we need to clone the text for the return type of the
closure. `SmartString`'s `From<String>` implementation reuses the
allocation when the string is too long to inline and drops it if it is
short enough to inline though which can be wasteful. `From<&String>`
clones the string's allocation only when it is too long to be inlined,
so we save on allocations for any `new_text` short enough to be inlined.
The rest is changes to `new_text.reserve_exact`. Previously calls to
this function in this block mixed up character and byte indexing by
treating the length of the line-ending as 1. `reserve_exact` takes a
number of bytes to reserve and that may be 2 when `line_ending` is a
CRLF. A call to `reserve_exact` is also added to the branch used when
continuing line comments.
Joffrey Bluthé
Sumandora
dependabot[bot]
RoloEdits
Jason Fuchs
Dimitri Sabadie
Keir Lawson
Michael Davis
Nick
Nik Revenco
Ian Hobson
iximeow
Noel Cower
Alexander Brassel
may
David Vogt
Roberto Vidal
Axlefublr
Harishankar G
T. Li
Pascal Kuthe
renshyle
Zi How Poh
kyfanc
Kristoffer Plagborg Bak Sørensen
magackame
Darshan Kumawat
Denys Rybalka
Yomain
TornaxO7
Jose Alvarez
janos-r