Fixes#1037
Renderers must convert the internal Kitty graphics state to a GPU
texture for rendering. For performance reasons, renderers cache the GPU
state by image ID. Unfortunately, this meant that if an image was
replaced with the same ID and was already cached, it would never be
updated on the GPU.
This PR adds the transmission time to the cache. If the transmission
time differs, we assume the image changed and replace the image.
Fixes#1014
There were a couple overlapping issues here:
1. To determine the "unshifted" key, we were using `keyval_to_lower`.
This only works if the key is uppercase and not all layouts are
uppercase for the unshifted value.
2. If we couldn't case the unicode value of a key or unshifted key to
ASCII, we'd say the key was the same as the physical key. This is
incorrect because the physical key is always the layout of the
physical keyboard so for example with the Turkish layout on a US
keyboard, you'd get US letters when a key may only correspond to
non-US letters.
To fix the first issue, we are using map_keycode to map the hardware
keycode to all possible keyvals from it. We then use the currently
active layout to find the unshifted value.
To fix the scond issue, we no longer fallback to the physical key if
there was IM text or non-ASCII unicode values for the key.
I tested Turkish with #1014 and I tested Dvorak to make sure those
basics still work.
With the nixpkgs LLVM 17 PR (NixOS/nixpkgs#258614) now merged, we can
update this flake input to base off of master so that we can take
advantage of the built LLVM derivation when it's ready.
This only leaves Zig 0.12 which should mean a much reduced build time.
Note that of this writing, the derivations are still queued, so until
that happens, this update means a rebuild of LLVM 17 for anyone using
the ghostty package.
This fixes an issue where selections from the bottom-right to the
top-left (or top-left to bottom-right), in addition to some single-line
rectangle selections, were not working.
This works by handling situations where only one of the x or y
axes in the start or end points may need to be flipped to get the
correct top-left or bottom-right of a selection. We call these kinds of
orientations "mirrored", like you were looking in a mirror.
This also adds a small bit of logic that keeps these kinds of motions in
rectangle selection from selecting the character before or after it.
This has the current side-effect of anchoring a rectangle selection to
the original characters if you change directions during the selection,
something I will look at in a later commit.
Finally, this also removes rectangle select on double-click. I thought
this might be a good idea, but word select in rectangle mode really
does not work (the effect seems pretty erratic), and it's not
implemented in Kitty either.
Fixes#1008.
Fixes#1018Fixes#1020
This disables the "visibleAtLaunch" configuration in the xib and
manually shows the window when it loads. This lets us carefully control
what happens particularly when a window is full screen (native) and part
of Mission Control.
Previously, the behavior depended on the Settings.app "Prefer tabs
when opening documents" setting, but we didn't handle every behavior
correctly (see #1018 and #1020). I couldn't find a way to robustly
handle all cases because there are no published macOS APIs for
interacting with Mission Control...
Plus, terminals aren't really "documents" so it did confuse at least one
user that Ghostty would follow this configuration. We just incidently
did because we use native tabbing.
This PR takes full control into our own hands. Our behavior is now:
- If a new window is created from a native fullscreen window, the
new window is created into native fullscreen.
- If a new tab is created from a native fullscreen window, the
tab is added to the existing window and does not create a new space.
- If a window or tab is created from a non-fullscreen window, the
existing behaviors remain.
This adds in-terminal rendering for the powerline glyphs E0B4 and E0B6,
similar to how we are rendering the triangle shapes currently.
The circle glyphs use a much more complex rendering due to the nuances
of drawing them: we use a midpoint algorithm for drawing on a 4x
supersampled matrix, fill, and then downsample. We use the same
downsampling approach as is done in the arc box drawing code.
The midpoint variant we're using here is described by Dennis Yurichev:
https://yurichev.com/news/20220322_circle/, although there are similar
variants elsewhere (some cited at the bottom of his article).
Fixes#965
When processing keybindings that closed the surface (`close_surface`,
`close_window`), the surface and associated runtime structures would be
freed so we could segfault.
This PR introduces a new enum result for input events (only key for now)
that returns whether an event resulted in a close. In this case, callers
can properly return immediately and avoid writing to deallocated memory.
