This commit removes the CI step that installs LLVM on Windows because:
1. It was constantly failing when I worked on [1]
```text
Failed in attempting to update the source: winget
The `msstore` source requires that you view the following agreements before using.
Terms of Transaction: https://aka.ms/microsoft-store-terms-of-transaction
The source requires the current machine's 2-letter geographic region to be sent to
the backend service to function properly (ex. "US").
Failed when searching source: winget
An unexpected error occurred while executing the command:
0x8a15000f : Data required by the source is missing
No packages were found among the working sources.
```
2. Actually, we don't need to install it since the Windows Github
action image already includes it. See [2]
[1]: https://github.com/infinilabs/coco-app/pull/946
[2]: https://github.com/actions/runner-images/blob/main/images/windows/Windows2025-Readme.md#language-and-runtime
This commit refactors the code that evenly collects documents from query
sources to let it collect at least 2 documents in every source, which
could correct the case when `max_hits_per_source` is 0. This was possible
with the previous impl, no longer allowed after this commit
* refactor: custom_version_comparator() now compares semantic versions
Previously, when comparing 2 versions, custom_version_comparator() only
compared their build numbers, which was incorrect. See this case:
```text
0.8.0-2500 -> 0.9.0-SNAPSHOT-2501 -> 0.8.1-2502
```
Coco adopts SemVer[1], and according to the specification, "0.8.1-2502"
is older than "0.9.0-SNAPSHOT-2501" even though it has a larger build
number.
This commit refactors it to compare the semantic versions.
[1]: Even though Coco uses SemVer, our version string does not follow the
spec. In this implementation, we use `to_semver()` to do the conversion, see
the code comments for more details.
* correct comments
Extract a procedure that calls convert_pages() to process HTML files
into a function. It is used in both install/store.rs and
install/local_extension.rs, doing this avoids code duplication.
This commit:
1. Bump dep tauri_nspenel to v2.1
2. On macOS, our main window is not a window but panel. Previously, we
were using window.show() and window.hide() in show_coco() and
hide_coco(). In this commit, we switch from window.show/hide to
panel.show/hide
Co-authored-by: ayang <473033518@qq.com>
* chore: use a custom log directory
This commit changes our log dirctory from the one set by Tauri to
a custom one. It is mostly the same as Tauri’s one, except that
the "{bundleIdentifier}" will be "Coco AI" rather than the real
identifier.
We are doing this because our bundle ID ("rs.coco.app") ends with ".app", log
directory "/Users/xxx/Library/Logs/rs.coco.app" is mistakenly thought as an
application by Finder on macOS, making it inconvenient to open. We do not want
to change the bundle identifier. The data directory, which stores all the data, still
references it. So doing that will be a breaking change. Using a custom log
directory make more sense.
* fmt
Take the 'Spotify Control' extension as an example:
- Spotify Control
- Toggle Play/Pause
- Next Track
- Previous Track
Previously, these sub-extensions were only returned when the query string
matched them, and thus counterintuitively, searching for 'Spotify Control' would
not hit them.
This commit changes that behavior: when a main extension (of type Extension)
matches the query, all of its sub-extensions are now included in the results.
When the expression to evaluate contains only a number, the result is
guaranteed to be this number. In this case, we no longer evaluate the
expression as telling users that "x = x" is meaningless.
Found another case where the `NextDisplay` command does not work (I said
another because the bug that commit ca71f07f3a3cc[1] fixed was also found
by playing with the `NextDisplay` command). After debugging, the root cause
of the issue is that the macOS API `AXUIElementSetAttributeValue()` does
not work in the expected way.
> When I execute the `NextDisplay` command to move the focused window from
> a big display (2560x1440) to a small display (1440*900), the window size
> could be set to 1460.
I don't know why it does not work so the only thing we can do is to retry.
Luckily, retrying works, at least in my tests.
[1]: ca71f07f3a
This commit fixes the issue that when the current desktop contains more
than 1 windows, moving the focused window via `NextDesktop` and
`PreviousDesktop` won't work.
How? By adding 2 missing `sleep()` functions:
1. b91a18dbb8/Silica/Sources/SIWindow.m (L242)
2. b91a18dbb8/Silica/Sources/SIWindow.m (L249)
Also, this commit improves the implementation by resetting the mouse position.
`NextDesktop` and `PreviousDesktop` are implemented by emulating mouse and
keyboard events, draging the focused window and switching to the corresponding
desktop. To make a window draggable, we have to move the mouse to the window's
traffic light area. It is disturbing to not move the mouse back so this commit
implements it.
This commit fixes a bug that most Window Management extension commands
won't work if you:
1. operate the focused window from another display
2. and they are adjacent
To reproduce this:
say you have 2 displays
1. Put the focused window on a non-main display, maximize the window
2. Move the mourse to the main display, making it the active display
3. Launch Coco, then execute the `TopHalf` command
The focused window will be moved to the main display, while it should
stay in the non-main display.
The root cause of the issue is that the previous implementation of
`intersects()` didn't handle an edge case correctly, adjavent rectangles
should not be considered overlapping. This commit replaces the buggy
implementation with the `CGRectIntersectsRect()` function from macOS
core graphics library.
