completion fix

[cossonleo]

Completion will not work correct When We type chars between sended to completion request and recived the result. These chars will not be replaced by completion items.

[pickfire]

How to reproduce the issue? Oh, I just realized I have this issue.

[archseer]

I don't quite understand the issue, we track the initial trigger_offset when completion is triggered:

helix/helix-term/src/commands.rs

Line 3490 in 7eb013c

let trigger_offset = cursor;

Then we use that trigger offset when needed:

helix/helix-term/src/ui/completion.rs

Lines 111 to 117 in 7eb013c

	// Some LSPs just give you an insertText with no offset ¯\_(ツ)_/¯
	// in these cases we need to check for a common prefix and remove it
	let prefix = Cow::from(doc.text().slice(start_offset..trigger_offset));
	let text = text.trim_start_matches::<&str>(&prefix);
	Transaction::change(
	doc.text(),
	vec![(trigger_offset, trigger_offset, Some(text.into()))].into_iter(),

We also remove any extra text:

helix/helix-term/src/ui/completion.rs

Lines 133 to 134 in 7eb013c

	// if more text was entered, remove it
	doc.restore(view.id);

[archseer]

We might need to call .savepoint sooner, inside fn completion rather than here

helix/helix-term/src/ui/completion.rs

Lines 152 to 153 in 0062af6

	// initialize a savepoint
	doc.savepoint();

[cossonleo]

We might need to call .savepoint sooner, inside fn completion rather than here

helix/helix-term/src/ui/completion.rs

Lines 152 to 153 in 0062af6

	// initialize a savepoint
	doc.savepoint();

Yes
In the current implementation, there are still redundant characters.
It occurs when the completion request is sent and wait response. At this time, we type char again.

[cossonleo]

How to reproduce the issue?

Although it is a small regular event that affects normal editing

gopls 0.7.0
helix 0.6
.config/helix/config.toml

  idle-timeout = 40

[the-mikedavis]

I've been seeing the same completion problems with ElixirLS and ErlangLS. I'm not sure if it's the slowness of the LS or my low idle-timeout config but when I type and get completion as I type, C-n/C-ping to one of the candidates does sometimes end up with a suffix like the screenshot. I hadn't really locked down what triggered it but it seems like if I deliberately pause before pressing C-n/C-p it doesn't happen (on master).

I just took this branch for a spin though and it seems to solve my completion problems.

[the-mikedavis]

I ran with this change yesterday and I noticed that the completion menu seems to pop up less. I imagine that what's happening is that I trigger the idle-timeout between keystrokes and the keystrokes following the auto-complete request cancel that request. (I haven't really looked at the change yet though so take that description with a grain of salt)

edit: this is my own doing - going into verbose mode and tailing the logs shows that it's the server that's being slow.

[archseer]

Hey @cossonleo I want to finally get this merged but I'm trying to understand the fix. Does it immediately set a save point, then when completion data arrives it filters that based on new input?

[cossonleo]

Hey @cossonleo I want to finally get this merged but I'm trying to understand the fix. Does it immediately set a save point, then when completion data arrives it filters that based on new input?

Yes, Method Completion::new() will call recompute_filter(editor) to filter the data.

[archseer]

Looks like my comments didn't get posted last time: I think moving doc.savepoint() before the callback is the only change needed to fix the completion. Why is savepoint now storing version?

[cossonleo]

Looks like my comments didn't get posted last time: I think moving doc.savepoint() before the callback is the only change needed to fix the completion. Why is savepoint now storing version?

We should discard the completion's result responsed after trigger the next completion. So we need to store the version to determine whether to discard.

[pickfire]

I think we need to add one more check to show completion beside this check, it's to check whether it's running from a macro (or repeat_last_motion), I notice this erratic behavior of completion popping up (and sometimes not closing and sometimes quickly switching between multiple different completions) when I type ..

This is already an issue in master but since you have this patch I think might as well just put a comment here.

[sbromberger]

On my system, this patch causes completion to fail completely. Logs show

2022-10-06T17:19:39.010 helix_lsp::transport [ERROR] err: <- StreamClosed
2022-10-06T17:19:39.012 helix_lsp::transport [ERROR] err: <- StreamClosed

This is gopls 0.9.5 with go 1.19.2.

EDIT: with hx -vvv, completion works, but I still get the problems described in #4131.

[kirawi]

I think this is the code that handles it in VSCode: https://github.com/microsoft/vscode/blob/ba9399b83d048e780646896d5f81bcc08d759c6b/src/vs/editor/contrib/inlineCompletions/browser/inlineCompletionsModel.ts#L539-L548

[rcorre]

I don't have enough context to be confident in my conflict resolutions, but I merged this into master at rcorre@51f8b67 and it seems to fix #4851 for me.

[pascalkuthe]

Superseded by #6173