feat(lsp): add textDocument/inlineCompletion support #14876

devmanuelli · 2025-11-26T11:32:28Z

Implements LSP 3.18 textDocument/inlineCompletion (ghost text).

Demo

Screen.Recording.-.Made.with.RecordCast.webm

Test

python3 -m venv .venv
source .venv/bin/activate
pip install pygls lsprotocol

test_lsp.py:

#!/usr/bin/env python3
from lsprotocol import types as lsp
from pygls.lsp.server import LanguageServer

server = LanguageServer("test", "v0.1")

@server.feature(lsp.TEXT_DOCUMENT_INLINE_COMPLETION, lsp.InlineCompletionOptions())
def inline_completion(params: lsp.InlineCompletionParams):
    return lsp.InlineCompletionList(items=[
        lsp.InlineCompletionItem(insert_text="# Hello!")
    ])

if __name__ == "__main__":
    server.start_io()

~/.config/helix/languages.toml:

[language-server.test-lsp]
command = "/path/to/.venv/bin/python3"
args = ["/path/to/test_lsp.py"]

[[language]]
name = "python"
language-servers = ["test-lsp"]

~/.config/helix/config.toml:

[editor]
inline-completion-timeout = 150

[keys.insert]
tab = "inline_completion_accept"
C-e = "inline_completion_dismiss"

P.S.: spread the word on reddit!!

devmanuelli · 2025-11-26T16:32:00Z

I'm doing a refactoring simplifying the code
EDIT: done. I'm gonna make a demo video

Implement LSP 3.18 inline completion (ghost text) feature: - Add InlineCompletion to LanguageServerFeature enum - Remove 'proposed' feature gate from lsp-types (official in LSP 3.18) - Add inline_completion client capability and request method - Store inline completion as InlineAnnotation on Document - Render ghost text via text_annotations system - Add inline_completion_accept and inline_completion_dismiss commands - Auto-trigger on document change with configurable debounce (inline_completion_timeout) - Clear completion on entering normal mode

taylorplewe · 2025-11-26T17:48:26Z

Nice!!! Can't wait to clone and try it out!

cunha · 2025-11-26T19:17:14Z

Is this sufficient to get copilot-language-server working? If so, if someone configures it successfully, it would be useful to get a "bootstrap" config posted here.

(It seems copilot-language-server also requires other endpoints like textDocument/didOpen, textDocument/didChange, and textDocument/didClose, but a quick search through the code indicates we may have these already.)

devmanuelli · 2025-11-26T19:50:56Z

@cunha of course we already have them. Tomorrow I will sign up github copilot and try, But it should work

devmanuelli · 2025-11-26T20:32:15Z

@cunha we have work to do but we are confident 💪
https://github.com/user-attachments/assets/0eb9e5d4-6284-4e46-9e85-c5a8c3c6a168
I've launched first this script copilot_auth.py to login via web interface.
Minimal config is

[[language]]
name = "python"
language-servers = ["copilot"]

[language-server.copilot]
command = "copilot-language-server"
args = ["--stdio"]

[language-server.copilot.config]
editorInfo = { name = "Helix", version = "25.01" }
editorPluginInfo = { name = "helix-copilot", version = "0.1.0" }

devmanuelli · 2025-11-26T21:47:20Z

Screen.Recording.-.Made.with.RecordCast.webm

@cunha enjoy ;)

synecdokey · 2025-11-26T22:16:52Z

I would expect the cursor to stay where it is and the ghost text to appear after, no?

devmanuelli · 2025-11-26T22:31:31Z

I would expect the cursor to stay where it is and the ghost text to appear after, no?

@synecdokey cursor at end shows where you'd land after accepting, it is not so bad.
Poll: 👍 if u want cursor to stay where it is; 👎 o.w.

EDIT: done @synecdokey

- Add InlineCompletion struct to store display text, full insert text, and replace range - Calculate offset from LSP range to show only new text as ghost text - Use helix_core::Range directly instead of std::ops::Range - Handle replace range on accept for proper text replacement - Use safe slicing with get().unwrap_or_default()

helix-term/src/handlers/inline_completion.rs

Render inline completion ghost text directly at cursor position instead of using InlineAnnotation, which was causing the cursor to shift to the end of the ghost text. Also handles multiline ghost text by rendering each line separately.

devmanuelli · 2025-11-27T00:55:33Z

Screen.Recording.-.Made.with.RecordCast.webm

@synecdokey

cunha · 2025-11-27T23:07:29Z

Rocking this locally. Works smoothly. Also went through the code and don't have any suggestions.

Feature request: I would prefer to disable the automatic triggers for inline completions and instead have a keybind to trigger it on demand.

I looked into this. It seems that the more involved part is passing the type of completion (autotrigger vs manual) to the event handler (the current code passes in ()) to know when to ask for an inline completion (e.g., manual trigger with autotriggers disabled). However, while working on a draft, I found that there already exists a CompletionEvent for basically the same purpose, and the CompletionHandler::handle_event function has logic that we might want when handling inline completions too (e.g., cancelling ongoing completion requests when the user deletes text).

Given I have very limited understanding of the overall architecture, I'm not sure what is the best approach. I'm happy to work on it and submit a PR to https://github.com/devmanuelli/helix/tree/textDocument/inlineCompletion with some guidance. I'm not even sure these are the right things to think about, but we would need to decide on whether to reuse the existing CompletionEvent structure (would require differentiating normal completions from inline completions, possibly touching the existing code for normal completions) or, in a more extreme case, to make inline completions a special case for the existing CompletionHandler to minimize code duplication.

devmanuelli · 2025-11-27T23:16:05Z

I would prefer to disable the automatic triggers for inline completions and instead have a keybind to trigger it on demand.

@cunha
for each request or once for all the requests until u decide to disable it?

of course u can make PRs ;)

now I'm gonna push last commit where u can cycle among different completions from different LSPs

- Query all LSP servers with inline completion capability, not just first - Add InlineCompletions container with push/current/next/take_and_clear - Add inline_completion_next command to cycle through completions - Simplify accept logic using Transaction::change for both range cases - Fix cursor position race by computing fresh cursor in response handler - Add stale completion detection (discard if cursor moved past range)

devmanuelli · 2025-11-27T23:23:10Z

@ljahier I got a simpler way to solve it than yours. Probably InlineAnnotation field is not really needed

devmanuelli · 2025-11-27T23:25:44Z

Screen.Recording.-.Made.with.RecordCast.webm

@ljahier

cunha · 2025-11-28T13:13:50Z

I would prefer to disable the automatic triggers for inline completions and instead have a keybind to trigger it on demand.

@cunha for each request or once for all the requests until u decide to disable it?

I had a toggle in mind. This would allow disabing the automated triggers, but still allow for manual triggers. (The reasoning is that I find the frequent suggestions distracting, but would still like the autocomplete when there's some formulaic code to be written, or after typing out a comment describing what the next lines should do.)

ljahier · 2025-11-28T19:45:52Z

helix-term/src/handlers/inline_completion.rs

+            };
+
+            let offset_encoding = ls.offset_encoding();
+            tokio::spawn(async move {


I might be misunderstanding, but it looks like this handler spawns async tasks without any cancellation or coordination. We could end up with multiple on the fly requests completing out of order, and each one will push its own completion. That means stale requests might still update the UI or accumulate completions even though newer input has already been sent

No it cannot happen that "stale requests might still update the UI" (look at lines 76-79 and 91-94)

You're right ! I see the version check prevents stale completions after document changes. take_and_clear() is called synchronously before triggering new request

thomasaarholt · 2025-11-28T20:06:28Z

This is extremely cool, and I will build this and try it out!
It took me quite a few runs to actually catch what was going on in the screen recording. Could I ask you to zoom in a bit when recording in the future?

Process all items returned by the LSP server instead of only the first one, allowing users to cycle through multiple suggestions with next/prev commands.

devmanuelli · 2025-11-28T21:18:55Z

@ljahier fix little bug (previously, we only took the first suggestion, even though the server could have sent us several ones)

cosmincartas · 2025-11-30T14:26:49Z

I've tested it on my side and it looks good if the only LSP used is copilot. I'm getting some weird behavior when combining it with other LSPs such: "typescript-language-server" or "vue-language-server".

devmanuelli · 2025-11-30T14:38:59Z

I've tested it on my side and it looks good if the only LSP used is copilot. I'm getting some weird behavior when combining it with other LSPs such: "typescript-language-server" or "vue-language-server".

@cosmincartas do they also have inlineCompletion capability?

EDIT: u are right thank u very much for ur time testing my pr. Changed this pr into a draft. I will fix it. If someone finds the bug before me, pls make a pr to my repo

EDIT: @cosmincartas try using this config - maybe the problem is tab

[keys.insert]
C-a = "inline_completion_accept"
C-e = "inline_completion_dismiss"
C-n = "inline_completion_next"
C-p = "inline_completion_prev"
C-space = "inline_completion_trigger"

If u find sth strange pls report it recording a video. Now I'm testing gopls and copilot together

EDIT: video demo with copilot and with last hx release

Screen.Recording.-.Made.with.RecordCast.webm

// W/OUT COPILOT STD RELEASE

Screen.Recording.-.Made.with.RecordCast.1.webm

@cosmincartas as soon as u tell me it tab was the problem I change pr status. Ty very much!

EDIT: fixed bad indentation in ghost text (tabs vs spaces...)

Screen.Recording.-.Made.with.RecordCast.2.webm

Switch inline completion ghost text from direct surface drawing to using the text annotation and decoration systems. This allows proper coexistence with diagnostics and other virtual text. Changes: - Mid-line: Use Overlay (first char) + InlineAnnotation (rest) to keep cursor in place while shifting diagnostics - EOL: Use Decoration to render ghost text without shifting cursor, return column offset so diagnostics still shift - Multi-line: Use LineAnnotation to reserve virtual lines + Decoration to render additional lines - Add OnModeSwitch hook to clear completions when leaving insert mode - Add documentation in docs/inline-completion-implementation.md

devmanuelli · 2025-11-30T23:42:14Z

I'll keep this PR open but it will undergo major refactoring.
I don't give up I will bring Github copilot into Helix!!
The goal is to fully integrate it with other Helix's features like diagnostics (ATM we overwrite them ...)

Screen.Recording.-.Made.with.RecordCast.3.webm

- Fix ghost text rendering one column to the right of cursor at EOL - Root cause: virt_off.col includes newline width (1), but cursor is ON the newline cell, not after it - Solution: subtract 1 from virt_off.col for EOL ghost text positioning - Also apply cursor style to first ghost character so it appears "on" the block cursor (matching mid-line behavior) - Update documentation with fix details

NSPC911 · 2025-12-01T13:08:43Z

minor issue... splits don't seem to be handled nicely

NSPC911 · 2025-12-01T13:22:40Z

@cunha we have work to do but we are confident 💪 github.com/user-attachments/assets/0eb9e5d4-6284-4e46-9e85-c5a8c3c6a168 I've launched first this script copilot_auth.py to login via web interface. Minimal config is
[[language]]
name = "python"
language-servers = ["copilot"]

[language-server.copilot]
command = "copilot-language-server"
args = ["--stdio"]

[language-server.copilot.config]
editorInfo = { name = "Helix", version = "25.01" }
editorPluginInfo = { name = "helix-copilot", version = "0.1.0" }

I couldn't get this working, but I used the environment variable GITHUB_COPILOT_TOKEN as suggested in github/copilot-language-server-release#3, and it works. There is a minor issue with this method though.

The Fine Grained Token doesn't work, and Classic Token doesn't use the proper initials (it needs ghu_<token>, not ghp_). Using gh auth token outputs a token with gho_, again, not ghu_. The only way to get the Copilot-supported token is by using helix-gpt.

If possible, I would also like a way to make inline_completion_accept a higher priority than interacting with the completion menu. As a previous VSCode user, the tab button handled the inline completion and is a higher priority than the popup menu

cunha · 2025-12-04T17:37:55Z

Just confirming I also get some spill-over onto nearly splits:

Haven't done much testing recently, but will likely be able to over the next few days.

Some info for people coming in and wanting to disable the auto triggers, the option is: (@devmanuelli unclear if we're updating this PR, but it seems documentation for this is missing.)

[editor]
inline-completion-auto-trigger = false

devmanuelli · 2025-12-04T20:22:01Z

u are all right. Im confident the code I will push in the following days will be okay ;) just stay tuned and patient ;)

MohdShahulMalik · 2025-12-05T11:33:02Z

Hey, do this with work lsp-ai language server as well? I tried to set up the inline completions with the lsp-ai language server but it doesn't seem to work. May someone help me with this please.

rohitbishnoi · 2025-12-05T11:55:35Z

Hey, do this with work lsp-ai language server as well? I tried to set up the inline completions with the lsp-ai language server but it doesn't seem to work. May someone help me with this please.

lsp-ai doesn't seems to be maintained anymore

MohdShahulMalik · 2025-12-05T12:31:28Z

Then can you please suggest me some lsp that would work with this product apart from copilot as it's free tier has very limited tab completes. Or it's the only choice?

archseer · 2025-12-07T03:48:17Z

helix-term/src/ui/editor.rs

+        // Render ghost text at cursor position
+        if let Some(completion) = doc.inline_completions.current() {
+            if let Some(cursor_pos) = editor.cursor_cache.get(view, doc) {


Thanks for starting on this! To echo the comment on the linked issue, rather than modifying the rendering code we should be leveraging a mix of LineAnnotation (to modify the current line) and Decoration to display the virtual lines:

#13039 (comment)

Replace InlineAnnotation with multiple Overlays for mid-line inline completion display. This prevents the cursor from shifting when ghost text is shown, providing a better editing experience. - Use Overlay per character position (cursor to EOL) for mid-line case - Add overflow text rendering via Decoration for chars beyond line end - Apply suffix trimming only for display, keep ghost_text intact for accept - Simplify Document cache to single overlay vector

Clear ghost text when selection changes (e.g., arrow keys) to match expected editor behavior where completions dismiss on cursor movement.

For multi-line inline completions when cursor is mid-line: - First line shown via overlays at cursor (no cursor shift) - Additional lines shown as virtual lines below - Original rest-of-line content pushed to bottom of ghost text Single-line mid-line completions unchanged (overlay + overflow).

Don't append rest_of_line to the last ghost text line if it already appears anywhere in the ghost text (indicating it's part of the replacement content rather than content to preserve).

When displaying multi-line inline completions mid-line, the first line preview was only creating overlays for as many characters as the ghost text contained. If the ghost text first line was shorter than the rest-of-line content, trailing characters would show through. Fix by padding the preview with spaces to cover the full rest-of-line length, ensuring all original content is blanked out (since it's being pushed down to the last ghost text line). Also removes the no-longer-needed implementation docs file.

devmanuelli · 2025-12-09T21:33:03Z

Screen-Recording.mp4

gonna slim it but we are on the good way

devmanuelli mentioned this pull request Nov 26, 2025

LSP: textDocument/inlineCompletion #13039

Open

devmanuelli force-pushed the textDocument/inlineCompletion branch 2 times, most recently from 3417d52 to dcde385 Compare November 26, 2025 11:52

devmanuelli force-pushed the textDocument/inlineCompletion branch from dcde385 to 166dc56 Compare November 26, 2025 17:04

devmanuelli mentioned this pull request Nov 26, 2025

Add support for copilot #6865

Closed

devmanuelli force-pushed the textDocument/inlineCompletion branch from 991cc39 to 7be883f Compare November 26, 2025 21:57

devmanuelli force-pushed the textDocument/inlineCompletion branch from 7be883f to de0038c Compare November 26, 2025 22:20

devmanuelli force-pushed the textDocument/inlineCompletion branch from de0038c to 6ebd915 Compare November 26, 2025 23:24

ljahier reviewed Nov 27, 2025

View reviewed changes

helix-term/src/handlers/inline_completion.rs Outdated Show resolved Hide resolved

devmanuelli force-pushed the textDocument/inlineCompletion branch from 78c2a9e to d45fae1 Compare November 27, 2025 00:48

ljahier mentioned this pull request Nov 27, 2025

feat: lsp add cycling through inline completions devmanuelli/helix#1

Closed

devmanuelli force-pushed the textDocument/inlineCompletion branch from 65237d6 to 5b7d6c1 Compare November 27, 2025 23:21

ljahier reviewed Nov 28, 2025

View reviewed changes

ljahier approved these changes Nov 28, 2025

View reviewed changes

feat(lsp): store all inline completion items from LSP response

ff17249

Process all items returned by the LSP server instead of only the first one, allowing users to cycle through multiple suggestions with next/prev commands.

devmanuelli requested a review from ljahier November 28, 2025 21:19

ljahier approved these changes Nov 28, 2025

View reviewed changes

devmanuelli marked this pull request as draft November 30, 2025 14:46

fix(ui): expand tabs in ghost text using document tab width

f817a78

the-mikedavis linked an issue Nov 30, 2025 that may be closed by this pull request

LSP: textDocument/inlineCompletion #13039

Open

bruwozniak mentioned this pull request Dec 3, 2025

LSP textDocument/inlineCompletion support gj1118/helix#28

Open

archseer requested changes Dec 7, 2025

View reviewed changes

devmanuelli added 5 commits December 9, 2025 22:01

fix(ui): dismiss inline completion on cursor movement

0b04aaa

Clear ghost text when selection changes (e.g., arrow keys) to match expected editor behavior where completions dismiss on cursor movement.

fix(ui): avoid duplicate rest-of-line in multi-line ghost text

395df7d

Don't append rest_of_line to the last ghost text line if it already appears anywhere in the ghost text (indicating it's part of the replacement content rather than content to preserve).

Uh oh!

feat(lsp): add textDocument/inlineCompletion support #14876

Are you sure you want to change the base?

feat(lsp): add textDocument/inlineCompletion support #14876

Conversation

devmanuelli commented Nov 26, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Demo

Test

Uh oh!

devmanuelli commented Nov 26, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

taylorplewe commented Nov 26, 2025

Uh oh!

cunha commented Nov 26, 2025

Uh oh!

devmanuelli commented Nov 26, 2025

Uh oh!

devmanuelli commented Nov 26, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

devmanuelli commented Nov 26, 2025

Uh oh!

synecdokey commented Nov 26, 2025

Uh oh!

devmanuelli commented Nov 26, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

devmanuelli commented Nov 27, 2025

Uh oh!

cunha commented Nov 27, 2025

Uh oh!

devmanuelli commented Nov 27, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

devmanuelli commented Nov 27, 2025

Uh oh!

devmanuelli commented Nov 27, 2025

Uh oh!

cunha commented Nov 28, 2025

Uh oh!

ljahier Nov 28, 2025

Choose a reason for hiding this comment

Uh oh!

devmanuelli Nov 28, 2025

Choose a reason for hiding this comment

Uh oh!

ljahier Nov 28, 2025

Choose a reason for hiding this comment

Uh oh!

thomasaarholt commented Nov 28, 2025

Uh oh!

devmanuelli commented Nov 28, 2025

Uh oh!

cosmincartas commented Nov 30, 2025

Uh oh!

devmanuelli commented Nov 30, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

devmanuelli commented Nov 30, 2025

Uh oh!

NSPC911 commented Dec 1, 2025

Uh oh!

NSPC911 commented Dec 1, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cunha commented Dec 4, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

devmanuelli commented Dec 4, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

MohdShahulMalik commented Dec 5, 2025

Uh oh!

rohitbishnoi commented Dec 5, 2025

Uh oh!

MohdShahulMalik commented Dec 5, 2025

Uh oh!

devmanuelli commented Nov 26, 2025 •

edited

Loading

devmanuelli commented Nov 26, 2025 •

edited

Loading

devmanuelli commented Nov 26, 2025 •

edited

Loading

devmanuelli commented Nov 26, 2025 •

edited

Loading

devmanuelli commented Nov 27, 2025 •

edited

Loading

devmanuelli commented Nov 30, 2025 •

edited

Loading

NSPC911 commented Dec 1, 2025 •

edited

Loading

cunha commented Dec 4, 2025 •

edited

Loading

devmanuelli commented Dec 4, 2025 •

edited

Loading

devmanuelli commented Dec 9, 2025 •

edited

Loading