<!--
Thank you for contributing to Ruff! To help us out with reviewing,
please consider the following:
- Does this pull request include a summary of the change? (See below.)
- Does this pull request include a descriptive title?
- Does this pull request include references to any relevant issues?
-->
## Summary
This PR introduces the `ruff_server` crate and a new `ruff server`
command. `ruff_server` is a re-implementation of
[`ruff-lsp`](https://github.com/astral-sh/ruff-lsp), written entirely in
Rust. It brings significant performance improvements, much tighter
integration with Ruff, a foundation for supporting entirely new language
server features, and more!
This PR is an early version of `ruff_lsp` that we're calling the
**pre-release** version. Anyone is more than welcome to use it and
submit bug reports for any issues they encounter - we'll have some
documentation on how to set it up with a few common editors, and we'll
also provide a pre-release VSCode extension for those interested.
This pre-release version supports:
- **Diagnostics for `.py` files**
- **Quick fixes**
- **Full-file formatting**
- **Range formatting**
- **Multiple workspace folders**
- **Automatic linter/formatter configuration** - taken from any
`pyproject.toml` files in the workspace.
Many thanks to @MichaReiser for his [proof-of-concept
work](https://github.com/astral-sh/ruff/pull/7262), which was important
groundwork for making this PR possible.
## Architectural Decisions
I've made an executive choice to go with `lsp-server` as a base
framework for the LSP, in favor of `tower-lsp`. There were several
reasons for this:
1. I would like to avoid `async` in our implementation. LSPs are mostly
computationally bound rather than I/O bound, and `async` adds a lot of
complexity to the API, while also making harder to reason about
execution order. This leads into the second reason, which is...
2. Any handlers that mutate state should be blocking and run in the
event loop, and the state should be lock-free. This is the approach that
`rust-analyzer` uses (also with the `lsp-server`/`lsp-types` crates as a
framework), and it gives us assurances about data mutation and execution
order. `tower-lsp` doesn't support this, which has caused some
[issues](https://github.com/ebkalderon/tower-lsp/issues/284) around data
races and out-of-order handler execution.
3. In general, I think it makes sense to have tight control over
scheduling and the specifics of our implementation, in exchange for a
slightly higher up-front cost of writing it ourselves. We'll be able to
fine-tune it to our needs and support future LSP features without
depending on an upstream maintainer.
## Test Plan
The pre-release of `ruff_server` will have snapshot tests for common
document editing scenarios. An expanded test suite is on the roadmap for
future version of `ruff_server`.
## Summary
This PR adds a new `cell` field to the JSON output format which
indicates the Notebook cell this diagnostic (and fix) belongs to. It
also updates the location for the diagnostic and fixes as per the
`NotebookIndex`. It will be used in the VSCode extension to display the
diagnostic in the correct cell.
The diagnostic and edit start and end source locations are translated
for the notebook as per the `NotebookIndex`. The end source location for
an edit needs some special handling.
### Edit end location
To understand this, the following context is required:
1. Visible lines in Jupyter Notebook vs JSON array strings: The newline
is part of the string in the JSON format. This means that if there are 3
visible lines in a cell where the last line is empty then the JSON would
contain 2 strings in the source array, both ending with a newline:
**JSON format:**
```json
[
"# first line\n",
"# second line\n",
]
```
**Notebook view:**
```python
1 # first line
2 # second line
3
```
2. If an edit needs to remove an entire line including the newline, then
the end location would be the start of the next row.
To remove a statement in the following code:
```python
import os
```
The edit would be:
```
start: row 1, col 1
end: row 2, col 1
```
Now, here's where the problem lies. The notebook index doesn't have any
information for row 2 because it doesn't exists in the actual notebook.
The newline was added by Ruff to concatenate the source code and it's
removed before writing back. But, the edit is computed looking at that
newline.
This means that while translating the end location for an edit belong to
a Notebook, we need to check if both the start and end location belongs
to the same cell. If not, then the end location should be the first
character of the next row and if so, translate that back to the last
character of the previous row. Taking the above example, the translated
location for Notebook would be:
```
start: row 1, col 1
end: row 1, col 10
```
## Test Plan
Add test cases for notebook output in the JSON format and update
existing snapshots.
