[BUG] Twinny formatter misidentifies split lines as duplicates in tightly wrapped code

[ton-An]

Describe the bug
When using a formatter with a short line length, single logical lines of code are often split across multiple physical lines. For example:

when(
  () => mockAuthenticationRepository.getCurrentUserId(),
).thenAnswer((_) async => const Right(tUserId));

when(
  () => mockLocationDataRepository.getLocationsByDate(
    start: any(named: 'start'),
    end: any(named: 'end'),
  ),
).thenAnswer((_) async => Right(tLocations));

In this case, if the FIM model generates a completion that begins with when(, the Twinny formatter at

twinny/src/extension/completion-formatter.ts

Line 249 in b50702f

protected preventDuplicateLine(): this {

incorrectly identifies it as a duplicate line and removes it. This results in an empty formatted completion:

[INFO] *** Twinny completion triggered for file: file:///Users/anton/Desktop/Projects/Location-History/location_history/app/test/features/location_tracking/domain/usecases/get_locations_by_date_test.dart ***
      Original completion: when(
      Formatted completion: 
      Max Lines: 20
      Use file context: true
      Completed lines count 1
      Using custom FIM template fim.bhs?: false

To Reproduce
1. Open a test file where logical lines are split due to short formatter line limits.
2. Trigger a completion that starts with the beginning of a multi-line expression (e.g., when().
3. Observe that the completion is removed as a “duplicate” despite being valid.

Here is a file with just that

Expected behavior
The Twinny formatter should treat logically connected lines of code—those split across multiple physical lines due to formatting constraints—as a single unit. This would prevent it from mistakenly identifying individual fragments (like when() as duplicates and ensure valid completions are preserved.

API Provider
LMStudio

Chat or Auto Complete?
FIM

Model Name
deepseek-coder-v2-lite-base-mlx

Desktop:

• OS: MacOS 15.3.2
• VSCode Version: 1.100.0-insider

[ton-An]

I will look into this over the next few days to see how easy/difficult it is to fix this.

[ton-An]

Unfortunately, this is quite tricky to implement because VS Code doesn’t provide an API to format a document with a custom line length (like 500). The formatting behavior—including line wrapping—is entirely controlled by the user’s configured formatter, and the extension API doesn’t allow us to override those settings dynamically.

As a result, a reliable fix would require handling each formatter individually (e.g. Prettier, Dart, ESLint).

A better fix is probably to make multi line completion a bit more reliable. @rjmacarthy Do you know the reason the multi line completion is not that reliable? I could look at fixing that up.

[yuzcat01]

I agree with that direction — if we can make FIM's multi-line completions more robust, or improve its ability to recognize logical lines (even when they're split across physical lines), we could likely avoid this issue.

I’ve been thinking about a few ideas :

1.Would it be possible to merge logical lines before preventDuplicateLine() runs? For example, if a line starts with when( and doesn't close its parentheses, maybe we could temporarily merge the indented lines that follow before checking for duplicates?

2.Could we use heuristic rules to detect common multi-line expressions? In Dart, constructs like when().thenAnswer() or calls like mock...Repository.get...() are pretty common and might be safely exempted from the “duplicate first line” check.

[ton-An]

I agree with that direction — if we can make FIM's multi-line completions more robust, or improve its ability to recognize logical lines (even when they're split across physical lines), we could likely avoid this issue.

I’ve been thinking about a few ideas :

1.Would it be possible to merge logical lines before preventDuplicateLine() runs? For example, if a line starts with when( and doesn't close its parentheses, maybe we could temporarily merge the indented lines that follow before checking for duplicates?

2.Could we use heuristic rules to detect common multi-line expressions? In Dart, constructs like when().thenAnswer() or calls like mock...Repository.get...() are pretty common and might be safely exempted from the “duplicate first line” check.

Yeah, I’ve actually been thinking along the same lines as your second idea. But I feel like that’s a bit of a separate thing from the overall reliability of multi-line completions. Unless I’m missing something?