Autocomplete: dynamically switch to multliline completions

[valerybugakov]

Context

• TLDR: we now use the completion's first line to determine if it starts the new multiline syntax node. If it does, we continue streaming until the completion is truncated or we reach the token sample limit.
• Closes Tree-sitter: dynamically switch singleline completions to multiline if the first line indicates the block start #1619
• Adds a feature flag cody-autocomplete-dynamic-multline-completions to run an experiment on the updated logic.
• Adds an experimental setting, cody.autocomplete.experimental.dynamicMultilineCompletions to test the feature locally.
• Adds completionPostProcessLogger used for local debugging. It's disabled by default, and I will replace it with OpenTelemetry spans in a follow-up PR. I put a little effort into improving this logger because it's a throw-away code, but I need it here because debugging streaming-truncation bugs is much faster with it.
• Revamps the multiline truncation logic:
  • We no longer look for the next new sibling after the completion. This approach is unreliable because the parse tree often breaks because of incomplete code in the completion insert text.
  • Instead, we take the descendant at the current cursor position (or multiline trigger position), look for the farthest ancestor that starts at the same line, and then wait for a moment when completion generates more line than this syntax node has. This means we do not care about broken parse-tree after the node under the cursor:

// Start completion
function test() {
  █
}

// ---------------------------

// Continue streaming
function test() {
  █if (true) {
    console.log('hello world'
}

// ---------------------------

// Stop, because characters were added outside of `function test` syntax node boundaries
function test() {
  █if (true) {
    console.log('hello world'
}

function 

• Need to merge Autocomplete: fix parse completions edit positions #1891 first to reduce the diff in this PR.

Screenshots

Follow-ups

• Autocomplete: add analytics for dynamic switch to multiline completions #1923
• Autocomplete: add heuristics to limit the number of huge unhelpful completions #1924
• Autocomplete: add unit tests for dynamic switch to multliline completions #1920
• Autocomplete: replace completion post processing logger with OpenTelemetry spans #1921

Merged preparation PRs

• Autocomplete: truncate multiline completions by the next new sibling #1709
• Autocomplete: split doc context calc logic #1784
• DX: log errors with source maps #1888
• Autocomplete: fix parse completions edit positions #1891
• Autocomplete: use a smaller percentage to detect the same string #1893

Test plan

• Tested manually:
  • Ensure that code completions are generated without changes on this branch.
  • Ensure that code completions are dynamically switched to multiline based on the first line when cody.autocomplete.experimental.dynamicMultilineCompletions is set to true.
• Existing tests pass.
• The functionality is disabled by default. I'm going to add more unit tests for the new logic in a separate PR (see planned follow-ups section).

[valerybugakov]

Hey @philipp-spiess, could you take an early look at this? I still have outstanding TODOs, but the core implementation is finished.

[philipp-spiess]

I’m SO excited for this. Pre-emptive approval but will stamp once it's ready for review :)

[umpox]

@valerybugakov Should this trigger a multiline completion:

I ran it on the HumanEval multiline dataset we have - it's just Python but should show a significant difference here, but it didn't seem to be generating multiline completions for these cases.

I have cody.autocomplete.experimental.dynamicMultilineCompletions set to true. Do I need to do anything else?

[philipp-spiess]

@umpox I think this example won't trigger a multi-line completion since it already has content that has a deeper indentation.

Maybe that's a check we can bypass when manually triggering completions though? Can you try to comment this out: https://sourcegraph.com/github.com/sourcegraph/cody/-/blob/vscode/src/completions/detect-multiline.ts?L52%3A21-54%3A22

[umpox]

@philipp-spiess Yep that fixes it:

There does seem to be something that seems off though, surely this should still generate a multi line completion?:

🤔

[valerybugakov]

@umpox, thanks for sharing the screenshots. The issue should be fixed here: ed20acb

[philipp-spiess]

Let's ship it!

[olafurpg]

This is huge! Looking forward to trying this out 😄