Ghostty now has a release channel build configuration. Current valid
values are "tip" and "stable" but I imagine more will be added in the
future.
The release channel is inferred whether the version we specify with the
`-Dversion-string` build flag has a prerelease tag or not. If it does,
the release channel is "tip". If it doesn't, the release channel is
"stable".
This also adds a configuration to specify the release channel for
auto-updates for the macOS application.
This adds a new workflow for building and releasing _tagged versions_
of Ghostty. The workflow is triggered automatically by new tags in the
format of `vX.Y.Z` but can also be manually triggered by running the
workflow from the GitHub Actions UI.
Release artifacts are uploaded to a completely separate R2 bucket
with its own access policy, retention, API keys, and so on.
There is currently no way to switch between "channels" in the macOS
app. I will follow up with a separate commit to add this feature.
`std.fs.accessAbsolute` asserts if the user proposed path is absolute,
which we are seemingly passing as-is with no validating that it is.
When running with safety checks on, passing non-absolute path to
--working-directory will make ghostty crash.
I changed it to use `Dir.access`, which is just `accessAbsolute` without
the check.
This has the side effect of also allowing relative working directory.
The [fixterms](http://www.leonerd.org.uk/hacks/fixterms/) "Really
Special Keypresses" section suggests using CSI 1 ; Ps R for F3, but this
is also a valid cursor position report. The intention was to make
back-compatible changes, so this is fairly considered a specification
bug.
This changes F3 in legacy mode to send CSI 13 ; Ps ~ instead, this is a
variant listed in fixterms, is what kitty protocol uses, and lacks the
problematic overlap with cursor positions.
The KeyEncoder.zig unit test has been changed accordingly, and all tests
pass on my machine.
`std.fs.accessAbsolute` asserts if the user proposed path is absolute,
which we are seemingly passing as-is with no validating that it is.
When running with safety checks on, passing non-absolute path to
--working-directory will make ghostty crash.
I changed it to use `Dir.access`, which is just `accessAbsolute` without
the check.
This has the side effect of also allowing relative working directory.
The [fixterms](http://www.leonerd.org.uk/hacks/fixterms/) "Really
Special Keypresses" section suggests using CSI 1 ; Ps R for F3, but this
is also a valid cursor position report. The intention was to make back-
compatible changes, so this is fairly considered a specification bug.
This changes F3 in legacy mode to send CSI 13 ; Ps ~ instead, this is a
variant listed in fixterms, is what kitty protocol uses, and lacks the
problematic overlap with cursor positions.
The KeyEncoder.zig unit test has been changed accordingly, and all tests
pass on my machine.
The comment in `function_keys.zig` was missing the `>` character for the
sequence. I've confirmed that this was just the comment, Ghostty treats
the original as an SGR sequence, which it is. Conversely, it does treat
`\x1b[>4;2m` as activating modifyOtherKeys.
The comment in `function_keys.zig` was missing the `>` character for the
sequence. I've confirmed that this was just the comment, Ghostty treats
the original as an SGR sequence, which it is. Conversely, it does treat
`\x1b[>4:2m` as activating modifyOtherKeys.
- Simplifies and clarifies the math for how the bounding box for
rendered glyphs is computed
- Reduces margin from 2px between glyphs to 1px by only padding the
bottom and right side of each glyph
- Avoids excessive padding to glyph box when font thicken is enabled or
when using a synthetic bold (it was previously 4x as much padding as
necessary in some cases)
I tried adding a key binding as the docs suggested, but it seems like
the referenced action has since been renamed from
`toggle_window_decoration` to `toggle_window_decorations`.
Since I saw questions in other small PRs: I am fine with this being
relicensed under the MIT license.
There is no reason to and I do not know where this assumption came from.
It's very possible for a colored glyph to (intentionally!) exceed the
cell bounds, and we shouldn't be stopping this...
nixos vim module expects vim plugins to have their code organized as top
level directories. eg: `/ftdetect` rather than
`/share/vim/vimfiles/ftdetect`. Adding a separate output seems like a
convenient way to reference the plugin files in `programs.vim.plugins
...`. Interestingly neovim will find the plugin via directory traversal
making `ghostty.vim` and `ghostty` the same for
`programs.neovim.plugins`
I agree to re-license my commits to MIT
nix vim module expects vim plugins to have their code organized as top level
directories. eg: /ftdetect rather than /share/vim/vimfiles/ftdetect. Adding
a separate output seems like a convenient way to reference the plugin
files in `programs.vim.plugins ...`. Interestingly nvim will find the
plugin via nested directories making ghostty.vim and ghostty the same
for `programs.neovim.plugins`
This adds a default-false `-Demit-webdata` build option. This emits some
files we copy over to our website (in the future I'll automate this more
to make a PR or something).
This also removes the old `website/` folder, all of the content was
moved to the website repo which is more updated.
Bash doesn't redraw the leading lines in a multiline prompt so we mark
the last line as a secondary prompt (k=s) to prevent the preceding lines
from being erased by Ghostty after a resize.
Our previously attempt at this (#1973) was flawed. Instead, we now just
re-issue the OSC "133;A" command with a 'k=s' (secondary) kind at the
end of our prompt string.
This isn't a great solution because it stomps on the prompt's "133;B"
command (end of prompt and start of user input), but it's sufficient for
now and only applies in the multiline prompt case.
Going forward, we should revisit our semantic prompt implementation. Our
row-based approach is too limiting; lines can have multiple markers, and
markers should be recorded with their full coordinates so they can form
ranges.
See: https://per.bothner.com/blog/2019/shell-integration-proposal/Fixes: #1961
Bash doesn't redraw the leading lines in a multiline prompt so we mark
the last line as a secondary prompt (k=s) to prevent the preceding lines
from being erased by Ghostty after a resize.
Our previously attempt at this (#1973) was flawed. Instead, we now just
re-issue the OSC "133;A" command with a 'k=s' (secondary) kind at the
end of our prompt string.
This isn't a great solution because it stomps on the prompt's "133;B"
command (end of prompt and start of user input), but it's sufficient for
now and only applies in the multiline prompt case.
Going forward, we should revisit our semantic prompt implementation. Our
row-based approach is too limiting; lines can have multiple markers, and
markers should be recorded with their full coordinates so they can form
ranges.
See: https://per.bothner.com/blog/2019/shell-integration-proposal/
This prevents forcibly injecting any configured shell integration into
the command's environment. Automatic shell detection is enabled, unless
shell integration has been explicitly disabled.
This prevents forcibly injecting any configured shell integration into
the command's environment. Automatic shell detection is enabled, unless
shell integration has been explicitly disabled.
Our semantic prompts are row-based, so the last prompt marker set on a
row "wins". In the case of at least our bash shell integration, this
means that consecutive prompt lines will all be marked as .input (OSC
133;B -- end-of-prompt, start of input).
Previously, clearPrompt() identified the current prompt's "area" by
searching upward from the current row until it encounters a .prompt
marker or some command output. In the bash case, .prompt is never the
dominant ("last") marker, so clearPrompt() would aggressively clear all
immediately preceding consecutive prompts.
With this change, we'll stop searching upwards when we encounter some
command output, a .prompt marker, _or another .input marker_. That last
case prevents clearPrompt() from unintentionally clearing earlier prompt
lines.
There may be improvements we can make to the way that our bash shell
integration emits semantic prompt markers, but I think this logic is
generally sound for all cases, and it specifically improves the current
bash prompt-clearing experience.
---
Before and after, after resizing the terminal window to trigger a
reflow:
<img width="504"
src="https://github.com/user-attachments/assets/91aa652e-e262-445a-8eed-7268c0d66428"
/>
<img width="510"
src="https://github.com/user-attachments/assets/b43e0aa4-0e24-4f4e-9b2c-b6b07c8f8c77"
/>
Our semantic prompts are row-based, so the last prompt marker set on a
row "wins". In the case of at least our bash shell integration, this
means that consecutive prompt lines will all be marked as .input (OSC
133;B -- end-of-prompt, start of input).
Previously, clearPrompt() identified the current prompt's "area" by
searching upward from the current row until it encounters a .prompt
marker or some command output. In the bash case, .prompt is never the
dominant ("last") marker, so clearPrompt() would aggressively clear all
immediately preceding consecutive prompts.
With this change, we'll stop searching upwards when we encounter some
command output, a .prompt marker, _or another .input marker_. That last
case prevents clearPrompt() from unintentionally clearing earlier prompt
lines.
There may be improvements we can make to the way that our bash shell
integration emits semantic prompt markers, but I think this logic is
generally sound for all cases, and it specifically improves the current
bash prompt-clearing experience.
A typo in the fish completions (that was likely copied to the zsh and
bash completions) prevented CLI actions from using enums as option
values because the completions tried to access non-existent fields from
type introspection. This doesn't cause any problems _now_ because no CLI
action uses an enum as an option value. However as soon as you try and
add one the completions fail to compile.
This patch fixes the incorrect field reference. It also adds the ability
to have _optional_ enums as option values.
cc @jparise @anund @anmolw
A typo in the fish completions (that was likely copied to the zsh and
bash completions) prevented CLI actions from using enums as option
values because the completions tried to access non-existent fields from
type introspection. This doesn't cause any problems _now_ because no CLI
action uses an enum as an option value. However as soon as you try and
add one the completions fail to compile.
This patch fixes the incorrect field reference. It also adds the ability
to have _optional_ enums as option values.
It turns out this happens more than I thought. Most notably, apple's
fallback font for a bunch of symbols, `Apple Gothic`, has no OS/2 table
🙃. I've gone ahead and fixed it so that both CoreText and FreeType
handle this gracefully.
Speaking of graceful handling, this problem came to attention due to our
very *un*-graceful handling of when a font fails to load due to a
metrics calculation error. It results in a bunch of blanked out text,
very bad, we should fix that.
*At some point I'll try to collect a group of test fonts with all sorts
of weirdnesses to exercise the weird edge cases in our handling.*
