The current setup is quite a mess, allocates a ton of unnecessary heap
data when rendering text or querying font info and not really caching
things well, this PR changes that.
Release Notes:
- N/A *or* Added/Fixed/Improved ...
Fixes#38709
## Summary
This implements the `extract_pull_request` method for GitLab's
`GitHostingProvider` trait, enabling merge request links to be shown in
the git blame hover popover.
## Implementation
The implementation parses GitLab MR references from commit messages
using a regex that matches two common patterns:
1. **Squash merge pattern**: `message (!123)` - When GitLab
squash-merges, it appends the MR number in parentheses
2. **Standard merge commit**: `See merge request group/project!123` -
The default merge commit message
## Tests Added
- `test_extract_merge_request_from_squash_commit` - Validates the
`(!123)` pattern
- `test_extract_merge_request_from_merge_commit` - Validates the `See
merge request` pattern
- `test_extract_merge_request_self_hosted` - Ensures it works with
self-hosted GitLab instances
- `test_extract_merge_request_no_match` - Confirms None is returned for
non-matching messages
## Related
This follows the same pattern as the GitHub provider implementation.
## Release Notes
- Added support for GitLab merge request links in git blame hover
popover
Git commands can fail when staging or unstaging a very large number of
files because each file path is passed as a separate CLI argument. Once
the argument list grows beyond the OS limit, git errors with Argument
list too long (os error 7).
Since git 2.26, pathspecs can be read from stdin, which lets us avoid
passing thousands of paths as arguments and bypass this limitation.
I looked for existing issues related to this but couldn’t find any.
Repro:
- Clone rust-lang repo
- Delete root tests folder
- Try to stage or unstage (40,000+ files)
- Git fails with Argument list too long (os error 7)
Release Notes:
- Fixed an issue where stage/unstage could fail when operating on a very
large number of files
Fixed git panel not update after commit or switch branch on Linux.
On macOS and Windows, the watcher implementation recursively watches
subdirectories.
On Linux, the watcher is non-recursive.
In Git worktree scenarios, only `<project>/.git` is watched, but the
actual worktree Git directory
`<project>/.git/worktrees/<worktree>` is not.
Therefore, Git operations such as commits or branch switches inside a
worktree do not emit watcher events on Linux, causing the Git panel to
stay out of sync.
Release Notes:
- Fixed git panel not update after commit or switch branch on Linux with
git worktrees.
Closes#15231
changes the priority in the way files are color-labelled..making sure
forcely added files are highlighted appropriately.
Release Notes:
- N/A *or* Added/Fixed/Improved ...
GitHub's commit API endpoint is rate limited to 60 requests/hour for
unauthenticated users. This causes avatar loading to fail after toggling
blame a few times.
This PR uses GitHub's CDN avatar endpoint
`https://avatars.githubusercontent.com/u/e?email={email}&s=128` instead,
which doesn't count against API rate limits. The author email is already
available from local git data (blame output), so
no API calls are needed.
- When author email is available, constructs the CDN URL directly (zero
API calls)
- Falls back to existing API-based behavior when email is unavailable
- Adds unit tests for URL construction
Closes#47590
## Test plan
- [x] `./script/clippy` passes
- [x] `cargo test -p git_hosting_providers` passes (89 tests including 3
new ones i added)
- [ ] Manual test: Open a file, toggle git blame, verify avatars load
without hitting rate limits
Release Notes:
- Fixed GitHub avatar rate limiting in git blame by using CDN endpoint
instead of API calls (#47590)
- Remove old attempt to sync scrolling
- Share a `ScrollAnchor` between the two sides, and be sure to resolve
it against the correct snapshot
- Allow either side to initiate an autoscroll request, and make sure
that request is processed in the same frame by the other side
Release Notes:
- N/A
---------
Co-authored-by: cameron <cameron.studdstreet@gmail.com>
Co-authored-by: Zed Zippy <234243425+zed-zippy[bot]@users.noreply.github.com>
Co-authored-by: Jakub <jakub@zed.dev>
* [x] capture and store teacher model's predicted cursor position
* [x] provide cursor position to student during distillation
* [x] eval cursor positions
* [x] parse and apply cursor position predictions at runtime
Release Notes:
- N/A
The bug occurred because `cx.observe_new::<Workspace>` would pass a
mutable `Window` that was also the new workspace's window. Later,
`fetch_files` would fail to read the newly created workspace from
`WorkspaceStore` because its window already had a mutable reference to
it.
The fix always passes the Settings UI's window handle to the
`fetch_files` function, and uses a `settings_window_handle.defer`
callback to call `fetch_files` after the newly created workspace's
window no longer has any mutable references to it.
Release Notes:
- settings_ui: Fixed editable project settings not being updated when a
new workspace is created
Fixes#8048
## Summary
In vim mode, pressing Escape to dismiss the buffer search now correctly
restores the cursor to its original position, rather than leaving it at
the first match.
## Problem
When using vim's `/` command to search:
1. User positions cursor at line X
2. User presses `/` to open search, types a query
3. Matches are highlighted, cursor may visually jump to first match
4. User presses Escape to dismiss without navigating
5. **Bug:** Cursor ends up at first match instead of line X
This breaks vim parity where Escape should cancel the search and restore
cursor position.
## Solution
The fix leverages the `focused()` callback in `vim.rs`, which is called
when the editor regains focus after the search bar is dismissed.
**Key insight:** When search starts via `/`, the cursor position is
saved in `SearchState.prior_selections`. When search is *submitted* with
Enter, `search_submit()` drains these selections. But when search is
*dismissed* with Escape, they remain.
So in `focused()`, if:
- `prior_selections` is non-empty, AND
- The search bar's `is_dismissed()` returns true
...then we know the user dismissed the search (Escape) rather than
submitted it (Enter), and we restore the cursor.
## Why not handle `buffer_search::Dismiss` directly?
The initial approach tried to register a vim handler for the `Dismiss`
action. This didn't work because when Escape is pressed, the search bar
(which has focus) handles the `Cancel` action internally and calls its
`dismiss()` method directly—it doesn't dispatch `Dismiss` through the
action system. The vim handler registered on the editor was never
invoked.
## Test Plan
- Added `test_search_dismiss_restores_cursor` — verifies cursor
restoration when search is dismissed
- Added `test_search_dismiss_restores_cursor_no_matches` — verifies
behavior when query has no matches
- All 455 vim tests pass
- Manual testing confirms fix works with both `/` and `cmd-f`
## Release Notes
- Fixed vim mode: cursor now returns to original position when
dismissing buffer search with Escape (#8048)
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Introduces a separate JSON schema for project settings that excludes
user-only settings like `auto_update`, `telemetry`, `vim_mode`, etc.
This provides more accurate autocomplete and validation when editing
`.zed/settings.json`.
- Add `SettingsStore::project_json_schema`
- Map `.zed/settings.json` to `zed://schemas/project_settings` schema URL
Release Notes:
- Improved autocomplete for the project settings file
(`.zed/settings.json`) to only include settings that are valid at the
project level, excluding user-only settings.
This is still only a debugging/evaluation version (which is why it has
e.g. hardcoded stuff). It successfully ran a few times triggered
manually, so we're replacing the broken version on `main` with this to
see how it fares on incoming issues.
Release Notes:
- N/A
We used to just use the anchor resolution function that allows resolving
multiple anchors in one loop before. That has a lot of overhead though
when we only have a single anchor to resolve, so instead we just
specialize that case now as resolving a single anchor to an offset is a
super common operation.
Release Notes:
- N/A *or* Added/Fixed/Improved ...
Closes#45198
Release Notes:
- Fixed an issue where "Reopen Closed Item" would fail to reopen tabs
that had been converted from preview to non-preview before being closed.
---------
Co-authored-by: Smit Barmase <heysmitbarmase@gmail.com>
This branch:
1. teaches `--diff `command line option to to recurse into folders if
provided.
2. Adds a `MultiDiffView` that shows _all_ changed files in a single,
scrollable view.
This is necessary to provide a smooth user experience for `jj` and Zed
users who wish to use Zed as their jj difftool.
I'm not fully sure how this change interacts with
https://github.com/zed-industries/zed/pull/44936, or what plans y'all
have in mind.
Here's a screenshot of the resulting behavior:
<img width="1090" height="950" alt="Screenshot 2025-12-17 at 9 10 52 AM"
src="https://github.com/user-attachments/assets/8efd09b4-974f-4059-9f94-539c484c6d4a"
/>
I setup zed to handle jj diffs by adding the following to my jj config:
```toml
[aliases]
zdiff = ["diff", "--tool", "zed"]
[merge-tools.zed]
program = "/Users/dbarsky/Developer/zed/target/debug/cli"
# omit diff-invocation-mode to keep the default (JJ passes two dirs)
diff-args = [
"--zed", "/Users/dbarsky/Developer/zed/target/debug/zed",
"--wait",
"--new",
"--diff",
"$left",
"$right",
]
```
Release Notes:
- `--diff`, if provided with folders instead of files, will recurse into
those directories.
- Added a `MultiDiffView`, which will show all changed files within a
single, scrollable view.
**AI Disclosure**: Pretty much all of this code was written using Codex
with GPT-5.1-Codex. I edited by hand and tested this.
---------
Co-authored-by: Lukas Wirth <lukas@zed.dev>
## Summary
Add documentation for the default value of `include_warnings` in
diagnostics settings. This makes it consistent with other settings that
document their default values.
## Test plan
- Documentation-only change, no functional changes
Release Notes:
- N/A
Previously, the 'Add Virtual Environment' path browser started at the
filesystem root (`/`). Since virtual environments are almost always
located within the project directory, starting at the worktree root
provides feels a bit nicer to me.
Release Notes:
- Python: Improved the `Add Virtual Environment` path browser to start
at the workspace root.
The `Select Toolchain Path` picker would skip directories that exist on
my system and display `open this directory` multiple times. `Personal`
is missing:
<img width="569" height="225" alt="SCR-20260129-tvcc"
src="https://github.com/user-attachments/assets/a78db373-fdfb-439d-a688-a4c81763c745"
/>
```
~/Projects> ls
╭───┬──────────┬──────┬────────┬─────────────╮
│ # │ name │ type │ size │ modified │
├───┼──────────┼──────┼────────┼─────────────┤
│ 0 │ Personal │ dir │ 1.5 kB │ 2 weeks ago │
│ 1 │ Work │ dir │ 1.0 kB │ 2 hours ago │
╰───┴──────────┴──────┴────────┴─────────────╯
~/Projects>
```
Release Notes:
- Fixed a bug where directories could be missing from the `Select
Toolchain Path` path picker
Oleksiy Syvokon