28 releases
new 0.6.0 | Oct 21, 2024 |
---|---|
0.5.7 | Aug 13, 2024 |
0.5.6 | Jun 29, 2024 |
0.5.2 | Jan 15, 2024 |
0.1.3 | Nov 6, 2022 |
#29 in Command-line interface
10,338 downloads per month
Used in 5 crates
(4 directly)
6.5MB
28K
SLoC
r3bl_tui
Why R3BL?
R3BL TUI library & suite of apps focused on developer productivity
We are working on building command line apps in Rust which have rich text user interfaces (TUI). We want to lean into the terminal as a place of productivity, and build all kinds of awesome apps for it.
-
๐ฎ Instead of just building one app, we are building a library to enable any kind of rich TUI development w/ a twist: taking concepts that work really well for the frontend mobile and web development world and re-imagining them for TUI & Rust.
- Taking inspiration from things like React, SolidJS, Elm, iced-rs, Jetpack Compose, JSX, CSS, but making everything async (so they can be run in parallel & concurrent via Tokio).
- Even the thread running the main event loop doesn't block since it is async.
- Using proc macros to create DSLs to implement something inspired by CSS & JSX.
-
๐ We are building apps to enhance developer productivity & workflows.
- The idea here is not to rebuild
tmux
in Rust (separate processes mux'd onto a single terminal window). Rather it is to build a set of integrated "apps" (or "tasks") that run in the same process that renders to one terminal window. - Inside of this terminal window, we can implement things like "app" switching, routing, tiling layout, stacking layout, etc. so that we can manage a lot of TUI apps (which are tightly integrated) that are running in the same process, in the same window. So you can imagine that all these "app"s have shared application state. Each "app" may also have its own local application state.
- Here are some examples of the types of "app"s we plan to build (for which this
infrastructure acts as the open source engine):
- Multi user text editors w/ syntax highlighting.
- Integrations w/ github issues.
- Integrations w/ calendar, email, contacts APIs.
- The idea here is not to rebuild
All the crates in the r3bl-open-core
repo provide lots of useful
functionality to help you build TUI (text user interface) apps, along w/ general
niceties & ergonomics that all Rustaceans ๐ฆ can enjoy ๐.
Table of contents
- Introduction
- Changelog
- Learn how these crates are built, provide feedback
- Run the demo locally
- Nushell scripts to build, run, test etc.
- Input and output devices
- Life of an input event
- The window
- Layout and styling
- Grapheme support
- Lolcat support
- Issues and PRs
Introduction
You can build fully async TUI (text user interface) apps with a modern API that brings the best of the web frontend development ideas to TUI apps written in Rust:
- Reactive & unidirectional data flow architecture from frontend development (React, SolidJS, Elm, iced-rs, Jetpack Compose).
- Responsive design w/ CSS, flexbox like concepts.
- Declarative style of expressing styling and layouts.
And since this is using Rust and Tokio you get the advantages of concurrency and parallelism built-in. No more blocking the main thread for user input, for async middleware, or even rendering ๐.
This framework is loosely coupled and strongly coherent meaning that you can pick and choose whatever pieces you would like to use w/out having the cognitive load of having to grok all the things in the codebase. Its more like a collection of mostly independent modules that work well w/ each other, but know very little about each other.
Changelog
Please check out the changelog to see how the library has evolved over time.
Learn how these crates are built, provide feedback
To learn how we built this crate, please take a look at the following resources.
- If you like consuming video content, here's our YT channel. Please consider subscribing.
- If you like consuming written content, here's our developer site. Please consider subscribing to our newsletter.
- If you have questions, please join our discord server.
Run the demo locally
Once you've cloned the repo to a folder on your computer, you can run the examples you see in the video with the following commands:
cd tui/examples
cargo run --release --example demo
These examples cover the entire surface area of the TUI API. You can also take a look
at the tests in the source (tui/src/
) as well. A single nu
shell script run
in the tui
sub folder in the repo
allows you to easily build, run, test, and do so much more with the repo.
The
run
script works on Linux, macOS, and Windows. On Linux and macOS, you can simply run./run
instead ofnu run
.
Nushell scripts to build, run, test etc.
Command | Description |
---|---|
nu run help |
See all the commands you can pass to the run script |
nu run examples |
Run all the examples |
nu run release-examples |
Run all the examples with the release binary |
nu run examples-with-flamegraph-profiling |
This will run the examples and generate a flamegraph at the end so you can see profile the performance of the app. This video has a walkthrough of how to use this |
nu run log |
View the log output. This video has a walkthrough of how to use this. |
nu run build |
Build |
nu run clean |
Clean |
nu run test |
Run tests |
nu run clippy |
Run clippy |
nu run docs |
Build docs |
nu run serve-docs |
Serve docs over VSCode Remote SSH session |
nu run rustfmt |
Run rustfmt |
The following commands will watch for changes in the source folder and re-run:
Command | Description |
---|---|
nu run watch-all-tests |
Watch all test |
nu run watch-one-test <test_name> |
Watch one test |
nu run watch-clippy |
Watch clippy |
nu run watch-macro-expansion-one-test <test_name> |
Watch macro expansion for one test |
There's also a run
script at the top level folder of the repo. It is intended to
be used in a CI/CD environment w/ all the required arguments supplied or in
interactive mode, where the user will be prompted for input.
Command | Description |
---|---|
nu run all |
Run all the tests, linting, formatting, etc. in one go. Used in CI/CD |
nu run build-full |
This will build all the crates in the Rust workspace. And it will install all the required pre-requisite tools needed to work with this crate (what install-cargo-tools does) and clear the cargo cache, cleaning, and then do a really clean build. |
nu run install-cargo-tools |
This will install all the required pre-requisite tools needed to work with this crate (things like cargo-deny , flamegraph will all be installed in one go) |
nu run check-licenses |
Use cargo-deny to audit all licenses used in the Rust workspace |
Here are some framework highlights:
- An easy to use and approachable API that is inspired by React, JSX, CSS, Elm. Lots
of components and things are provided for you so you don't have to build them from
scratch. This is a full featured component library including:
- Elm like architecture w/ unidirectional data flow. The state is mutable. Async
middleware functions are supported, and they communicate w/ the main thread and
the [App] using an async
tokio::mpsc
channel and signals. - CSS like declarative styling engine.
- CSS like flexbox like declarative layout engine which is fully responsive. You can resize your terminal window and everything will be laid out correctly.
- A terminal independent underlying rendering and painting engine (can use crossterm or termion or whatever you want).
- Markdown text editor w/ syntax highlighting support, metadata (tags, title, author, date), smart lists. This uses a custom Markdown parser and custom syntax highlighter. Syntax highlighting for code blocks is provided by the syntect crate.
- Modal dialog boxes. And autocompletion dialog boxes.
- Lolcat (color gradients) implementation w/ a rainbow color-wheel palette. All the color output is sensitive to the capabilities of the terminal. Colors are gracefully downgraded from truecolor, to ANSI256, to grayscale.
- Support for Unicode grapheme clusters in strings. You can safely use emojis, and other Unicode characters in your TUI apps.
- Support for mouse events.
- Elm like architecture w/ unidirectional data flow. The state is mutable. Async
middleware functions are supported, and they communicate w/ the main thread and
the [App] using an async
- The entire TUI framework itself supports concurrency & parallelism (user input, rendering, etc. are generally non blocking).
- It is fast! There are no needless re-renders, or flickering. Animations and color changes are smooth (check this out for yourself by running the examples). You can even build your TUI in layers (like z-order in a browser's DOM).
Examples to get you started
Video of the demo in action
Here's a video of a prototype of R3BL CMDR app built using this TUI engine.
How does layout, rendering, and event handling work in general?
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ โ
โ main.rs โ
โ โโโโโโโโโโโโโโโโโโโโ โ
โ GlobalData โโโโโโโโโโโโโโโโบโ window size โ โ
โ HasFocus โ offscreen buffer โ โ
โ ComponentRegistryMap โ state โ โ
โ App & Component(s) โ channel sender โ โ
โ โโโโโโโโโโโโโโโโโโโโ โ
โ โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
- The main struct for building a TUI app is your struct which implements the [App] trait.
- The main event loop takes an [App] trait object and starts listening for input events. It enters raw mode, and paints to an alternate screen buffer, leaving your original scroll back buffer and history intact. When you exit this TUI app, it will return your terminal to where you'd left off.
- The [main_event_loop] is where many global structs live which are shared across the
lifetime of your app. These include the following:
- [HasFocus]
- [ComponentRegistryMap]
- [GlobalData] which contains the following
- Global application state. This is mutable. Whenever an input event or signal is processed the entire [App] gets re-rendered. This is the unidirectional data flow architecture inspired by React and Elm.
- Your [App] trait impl is the main entry point for laying out the entire application.
Before the first render, the [App] is initialized (via a call to App::app_init),
and is responsible for creating all the [Component]s that it uses, and saving them
to the [ComponentRegistryMap].
- State is stored in many places. Globally at the [GlobalData] level, and also in [App], and also in [Component].
- This sets everything up so that App::app_render, App::app_handle_input_event, and App::app_handle_signal can be called at a later time.
- The App::app_render method is responsible for creating the layout by using [Surface] and [FlexBox] to arrange whatever [Component]'s are in the [ComponentRegistryMap].
- The App::app_handle_input_event method is responsible for handling events that are sent to the [App] trait when user input is detected from the keyboard or mouse. Similarly the App::app_handle_signal deals with signals that are sent from background threads (Tokio tasks) to the main thread, which then get routed to the [App] trait object. Typically this will then get routed to the [Component] that currently has focus.
Switching from shared memory to message passing architecture after v0.3.10
Versions of this crate <= 0.3.10
used shared memory to communicate between the
background threads and the main thread. This was done using the async Arc<RwLock<T>>
from tokio. The state storage, mutation, subscription (on change handlers) were all
managed by the
r3bl_redux
crate. The use of the Redux pattern, inspired by React, brought with it a lot of
overhead both mentally and in terms of performance (since state changes needed to be
cloned every time a change was made, and memcpy
or clone
is expensive).
Versions > 0.3.10
use message passing to communicate between the background threads
using the tokio::mpsc
channel (also async). This is a much easier and more
performant model given the nature of the engine and the use cases it has to handle. It
also has the benefit of providing an easy way to attach protocol servers in the future
over various transport layers (eg: TCP, IPC, etc.); these protocol servers can be used
to manage a connection between a process running the engine, and other processes
running on the same host or on other hosts, in order to handle use cases like
synchronizing rendered output, or state.
Here are some papers outlining the differences between message passing and shared memory for communication between threads.
Input and output devices
Dependency injection is used to inject the
required resources into the main_event_loop
function. This allows for easy testing
and for modularity and extensibility in the codebase. The r3bl_terminal_async
crate
shares the same infrastructure for input and output devices. In fact the
r3bl_core::InputDevice and r3bl_core::OutputDevice structs are in the r3bl_core
crate.
- The advantage of this approach is that for testing, test fixtures can be used to perform end to end testing of the TUI.
- This also facilitates some other interesting capabilities, such as preserving all the state for an application and make it span multiple applets (smaller apps, and their components). This makes the entire UI composable, and removes the monolithic approaches to building complex UI and large apps that may be comprised of many reusable components and applets.
- It is easy to swap out implementations of input and output devices away from
stdin
andstdout
while preserving all the existing code and functionality. This can produce some interesting headless apps in the future, where the UI might be delegated to a window using eGUI or iced-rs or wgpu.
Life of an input event
There is a clear separation of concerns in this module. To illustrate what goes where, and how things work let's look at an example that puts the main event loop front and center & deals w/ how the system handles an input event (key press or mouse).
- The diagram below shows an app that has 3 [Component]s for (flexbox like) layout & (CSS like) styling.
- Let's say that you run this app (by hypothetically executing
cargo run
). - And then you click or type something in the terminal window that you're running this app in.
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โIn band input event โ
โ โ
โ Input โโโบ [TerminalWindow] โ
โ Event โฒ โ โ
โ โ โผ [ComponentRegistryMap] stores โ
โ โ [App]โโโโโโโโโโโโโโโโโบ[Component]s at 1st render โ
โ โ โ โ
โ โ โ โ
โ โ โ โโโโโโโโบ id=1 has focus โ
โ โ โ โ โ
โ โ โโโโบ [Component] id=1 โโโโโโ โ
โ โ โ โ โ
โ โ โโโโบ [Component] id=2 โ โ
โ โ โ โ
โ default handler โ โ
โ โฒ โ โ
โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ
โ โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โOut of band app signal โ
โ โ
โ App โ
โ Signal โโโบ [App] โ
โ โ โ
โ โ โ
โ โโโโโโโโบUpdate state โ
โ main thread rerender โ
โ โ โ
โ โ โ
โ โโโโโโโบ[App] โ
โ โ โ
โ โโโโโโบ[Component]s โ
โ โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
Let's trace the journey through the diagram when an input even is generated by the
user (eg: a key press, or mouse event). When the app is started via cargo run
it
sets up a main loop, and lays out all the 3 components, sizes, positions, and then
paints them. Then it asynchronously listens for input events (no threads are blocked).
When the user types something, this input is processed by the main loop of
[TerminalWindow].
- The [Component] that is in [FlexBox] w/
id=1
currently has focus. - When an input event comes in from the user (key press or mouse input) it is routed to the [App] first, before [TerminalWindow] looks at the event.
- The specificity of the event handler in [App] is higher than the default input handler in [TerminalWindow]. Further, the specificity of the [Component] that currently has focus is the highest. In other words, the input event gets routed by the [App] to the [Component] that currently has focus ([Component] id=1 in our example).
- Since it is not guaranteed that some [Component] will have focus, this input event can then be handled by [App], and if not, then by [TerminalWindow]'s default handler. If the default handler doesn't process it, then it is simply ignored.
- In this journey, as the input event is moved between all these different entities, each entity decides whether it wants to handle the input event or not. If it does, then it returns an enum indicating that the event has been consumed, else, it returns an enum that indicates the event should be propagated.
An input event is processed by the main thread in the main event loop. This is a synchronous operation and thus it is safe to mutate state directly in this code path. This is why there is no sophisticated locking in place. You can mutate the state directly in
Life of a signal (aka "out of band event")
This is great for input events which are generated by the user using their keyboard or
mouse. These are all considered "in-band" events or signals, which have no delay or
asynchronous behavior. But what about "out of band" signals or events, which do have
unknown delays and asynchronous behaviors? These are important to handle as well. For
example, if you want to make an HTTP request, you don't want to block the main thread.
In these cases you can use a tokio::mpsc
channel to send a signal from a background
thread to the main thread. This is how you can handle "out of band" events or signals.
To provide support for these "out of band" events or signals, the [App] trait has a
method called App::app_handle_signal. This is where you can handle signals that are
sent from background threads. One of the arguments to this associated function is a
signal
. This signal needs to contain all the data that is needed for a state
mutation to occur on the main thread. So the background thread has the responsibility
of doing some work (eg: making an HTTP request), getting some information as a result,
and then packaging that information into a signal
and sending it to the main thread.
The main thread then handles this signal by calling the App::app_handle_signal
method. This method can then mutate the state of the [App] and return an
[EventPropagation] enum indicating whether the main thread should repaint the UI or
not.
So far we have covered what happens when the [App] receives a signal. Who sends this
signal? Who actually creates the tokio::spawn
task that sends this signal? This can
happen anywhere in the [App] and [Component]. Any code that has access to [GlobalData]
can use the r3bl_core::send_signal! macro to send a signal in a background task.
However, only the [App] can receive the signal and do something with it, which is
usually apply the signal to update the state and then tell the main thread to repaint
the UI.
Now that we have seen this whirlwind overview of the life of an input event, let's look at the details in each of the sections below.
The window
The main building blocks of a TUI app are:
- [TerminalWindow] - You can think of this as the main "window" of the app. All the content of your app is painted inside of this "window". And the "window" conceptually maps to the screen that is contained inside your terminal emulator program (eg: tilix, Terminal.app, etc). Your TUI app will end up taking up 100% of the screen space of this terminal emulator. It will also enter raw mode, and paint to an alternate screen buffer, leaving your original scroll back buffer and history intact. When you exit this TUI app, it will return your terminal to where you'd left off. You don't write this code, this is something that you use.
- [App] - This is where you write your code. You pass in a [App] to the [TerminalWindow] to bootstrap your TUI app. You can just use [App] to build your app, if it is a simple one & you don't really need any sophisticated layout or styling. But if you want layout and styling, now we have to deal with [FlexBox], [Component], and r3bl_core::TuiStyle.
Layout and styling
Inside of your [App] if you want to use flexbox like layout and CSS like styling you can think of composing your code in the following way:
- [App] is like a box or container. You can attach styles and an id here. The id has to be unique, and you can reference as many styles as you want from your stylesheet. Yes, cascading styles are supported! ๐ You can put boxes inside of boxes. You can make a container box and inside of that you can add other boxes (you can give them a direction and even relative sizing out of 100%).
- As you approach the "leaf" nodes of your layout, you will find [Component] trait objects. These are black boxes which are sized, positioned, and painted relative to their parent box. They get to handle input events and render [RenderOp]s into a [RenderPipeline]. This is kind of like virtual DOM in React. This queue of commands is collected from all the components and ultimately painted to the screen, for each render! Your app's state is mutable and is stored in the [GlobalData] struct. You can handle out of band events as well using the signal mechanism.
Component, ComponentRegistry, focus management, and event routing
Typically your [App] will look like this:
#[derive(Default)]
pub struct AppMain {
// Might have some app data here as well.
// Or `_phantom: std::marker::PhantomData<(State, AppSignal)>,`
}
As we look at [Component] & [App] more closely we will find a curious thing [ComponentRegistry] (that is managed by the [App]). The reason this exists is for input event routing. The input events are routed to the [Component] that currently has focus.
The [HasFocus] struct takes care of this. This provides 2 things:
- It holds an
id
of a [FlexBox] / [Component] that has focus. - It also holds a map that holds a r3bl_core::Position for each
id
. This is used to represent a cursor (whatever that means to your app & component). This cursor is maintained for eachid
. This allows a separate cursor for each [Component] that has focus. This is needed to build apps like editors and viewers that maintains a cursor position between focus switches.
Another thing to keep in mind is that the [App] and [TerminalWindow] is persistent between re-renders.
Input event specificity
[TerminalWindow] gives [App] first dibs when it comes to handling input events. ComponentRegistry::route_event_to_focused_component can be used to route events directly to components that have focus. If it punts handling this event, it will be handled by the default input event handler. And if nothing there matches this event, then it is simply dropped.
Rendering and painting
The R3BL TUI engine uses a high performance compositor to render the UI to the
terminal. This ensures that only "pixels" that have changed are painted to the
terminal. This is done by creating a concept of PixelChar
which represents a single
"pixel" in the terminal screen at a given col and row index position. There are only
as many PixelChar
s as there are rows and cols in a terminal screen. And the index
maps directly to the position of the pixel in the terminal screen.
Offscreen buffer
Here is an example of what a single row of rendered output might look like in a row of
the OffscreenBuffer
. This diagram shows each PixelChar
in row_index: 1
of the
OffscreenBuffer
. In this example, there are 80 columns in the terminal screen. This
actual log output generated by the TUI engine when logging is enabled.
row_index: 1
000 S โโโโโโโโณโโโโโโโโ001 P 'j'โfgโbg 002 P 'a'โfgโbg 003 P 'l'โfgโbg 004 P 'd'โfgโbg 005 P 'k'โfgโbg
006 P 'f'โfgโbg 007 P 'j'โfgโbg 008 P 'a'โfgโbg 009 P 'l'โfgโbg 010 P 'd'โfgโbg 011 P 'k'โfgโbg
012 P 'f'โfgโbg 013 P 'j'โfgโbg 014 P 'a'โfgโbg 015 P 'โ'โrev 016 S โโโโโโโโณโโโโโโโโ017 S โโโโโโโโณโโโโโโโโ
018 S โโโโโโโโณโโโโโโโโ019 S โโโโโโโโณโโโโโโโโ020 S โโโโโโโโณโโโโโโโโ021 S โโโโโโโโณโโโโโโโโ022 S โโโโโโโโณโโโโโโโโ023 S โโโโโโโโณโโโโโโโโ
024 S โโโโโโโโณโโโโโโโโ025 S โโโโโโโโณโโโโโโโโ026 S โโโโโโโโณโโโโโโโโ027 S โโโโโโโโณโโโโโโโโ028 S โโโโโโโโณโโโโโโโโ029 S โโโโโโโโณโโโโโโโโ
030 S โโโโโโโโณโโโโโโโโ031 S โโโโโโโโณโโโโโโโโ032 S โโโโโโโโณโโโโโโโโ033 S โโโโโโโโณโโโโโโโโ034 S โโโโโโโโณโโโโโโโโ035 S โโโโโโโโณโโโโโโโโ
036 S โโโโโโโโณโโโโโโโโ037 S โโโโโโโโณโโโโโโโโ038 S โโโโโโโโณโโโโโโโโ039 S โโโโโโโโณโโโโโโโโ040 S โโโโโโโโณโโโโโโโโ041 S โโโโโโโโณโโโโโโโโ
042 S โโโโโโโโณโโโโโโโโ043 S โโโโโโโโณโโโโโโโโ044 S โโโโโโโโณโโโโโโโโ045 S โโโโโโโโณโโโโโโโโ046 S โโโโโโโโณโโโโโโโโ047 S โโโโโโโโณโโโโโโโโ
048 S โโโโโโโโณโโโโโโโโ049 S โโโโโโโโณโโโโโโโโ050 S โโโโโโโโณโโโโโโโโ051 S โโโโโโโโณโโโโโโโโ052 S โโโโโโโโณโโโโโโโโ053 S โโโโโโโโณโโโโโโโโ
054 S โโโโโโโโณโโโโโโโโ055 S โโโโโโโโณโโโโโโโโ056 S โโโโโโโโณโโโโโโโโ057 S โโโโโโโโณโโโโโโโโ058 S โโโโโโโโณโโโโโโโโ059 S โโโโโโโโณโโโโโโโโ
060 S โโโโโโโโณโโโโโโโโ061 S โโโโโโโโณโโโโโโโโ062 S โโโโโโโโณโโโโโโโโ063 S โโโโโโโโณโโโโโโโโ064 S โโโโโโโโณโโโโโโโโ065 S โโโโโโโโณโโโโโโโโ
066 S โโโโโโโโณโโโโโโโโ067 S โโโโโโโโณโโโโโโโโ068 S โโโโโโโโณโโโโโโโโ069 S โโโโโโโโณโโโโโโโโ070 S โโโโโโโโณโโโโโโโโ071 S โโโโโโโโณโโโโโโโโ
072 S โโโโโโโโณโโโโโโโโ073 S โโโโโโโโณโโโโโโโโ074 S โโโโโโโโณโโโโโโโโ075 S โโโโโโโโณโโโโโโโโ076 S โโโโโโโโณโโโโโโโโ077 S โโโโโโโโณโโโโโโโโ
078 S โโโโโโโโณโโโโโโโโ079 S โโโโโโโโณโโโโโโโโ080 S โโโโโโโโณโโโโโโโโspacer [ 0, 16-80 ]
When RenderOps
are executed and used to create an OffscreenBuffer
that maps to the
size of the terminal window, clipping is performed automatically. This means that it
isn't possible to move the caret outside of the bounds of the viewport (terminal
window size). And it isn't possible to paint text that is larger than the size of the
offscreen buffer. The buffer really represents the current state of the viewport.
Scrolling has to be handled by the component itself (an example of this is the editor
component).
Each PixelChar
can be one of 4 things:
- Space. This is just an empty space. There is no flickering in the TUI engine. When a new offscreen buffer is created, it is fulled w/ spaces. Then components paint over the spaces. Then the diffing algorithm only paints over the pixels that have changed. You don't have to worry about clearing the screen and painting, which typically will cause flickering in terminals. You also don't have to worry about printing empty spaces over areas that you would like to clear between renders. All of this handled by the TUI engine.
- Void. This is a special pixel that is used to indicate that the pixel should be ignored. It is used to indicate a wide emoji is to the left somewhere. Most terminals don't support emojis, so there's a discrepancy between the display width of the character and its index in the string.
- Plain text. This is a normal pixel which wraps a single character that maybe a
grapheme cluster segment. Styling information is encoded in each
PixelChar::PlainText
and is used to paint the screen via the diffing algorithm which is smart enough to "stack" styles that appear beside each other for quicker rendering in terminals.
Render pipeline
The following diagram provides a high level overview of how apps (that contain components, which may contain components, and so on) are rendered to the terminal screen.
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ Container โ
โ โ
โ โโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโ โ
โ โ Col 1 โ โ Col 2 โ โ
โ โ โ โ โ โ
โ โ โ โ โโโโโโโโโผโโผโโโโโโโโโโโบ RenderPipelineโโโโโโ
โ โ โ โ โ โ โ
โ โ โ โ โ โ โ
โ โ โโโโโโโโผโโโผโโโโโโโโโโโโโโผโโผโโโโโโโโโโโบ RenderPipelineโโ โ
โ โ โ โ โ โ โ โ
โ โ โ โ โ โ โผ + โผ
โ โ โ โ โ โ โโโโโโโโโโโโโโโโโโโโโโโ
โ โโโโโโโโโโโโโโโ โโโโโโโโโโโโโโโ โ โ โ
โ โ โ OffscreenBuffer โ
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ โ
โโโโโโโโโโโโโโโโโโโโโโโ
Each component produces a RenderPipeline
, which is a map of ZOrder
and
Vec<RenderOps>
. RenderOps
are the instructions that are grouped together, such as
move the caret to a position, set a color, and paint some text.
Inside of each RenderOps
the caret is stateful, meaning that the caret position is
remembered after each RenderOp
is executed. However, once a new RenderOps
is
executed, the caret position reset just for that RenderOps
. Caret position is not
stored globally. You should read more about "atomic paint operations" in the
RenderOp
documentation.
Once a set of these RenderPipeline
s have been generated, typically after the user
enters some input event, and that produces a new state which then has to be rendered,
they are combined and painted into an OffscreenBuffer
.
First render
The paint.rs
file contains the paint
function, which is the entry point for all
rendering. Once the first render occurs, the OffscreenBuffer
that is generated is
saved to GlobalSharedState
. The following table shows the various tasks that have to
be performed in order to render to an OffscreenBuffer
. There is a different code
path that is taken for ANSI text and plain text (which includes StyledText
which is
just plain text with a color). Syntax highlighted text is also just StyledText
.
UTF-8 | Task |
---|---|
Y | convert RenderPipeline to List<List<PixelChar>> (OffscreenBuffer ) |
Y | paint each PixelChar in List<List<PixelChar>> to stdout using OffscreenBufferPainterImplCrossterm |
Y | save the List<List<PixelChar>> to GlobalSharedState |
Currently only crossterm
is supported for actually painting to the terminal. But
this process is really simple making it very easy to swap out other terminal libraries
such as termion
, or even a GUI backend, or some other custom output driver.
Subsequent render
Since the OffscreenBuffer
is cached in GlobalSharedState
a diff to be performed
for subsequent renders. And only those diff chunks are painted to the screen. This
ensures that there is no flicker when the content of the screen changes. It also
minimizes the amount of work that the terminal or terminal emulator has to do put the
PixelChar
s on the screen.
How does the editor component work?
The EditorComponent
struct can hold data in its own memory, in addition to relying
on the state.
- It has an
EditorEngine
which holds syntax highlighting information, and configuration options for the editor (such as multiline mode enabled or not, syntax highlighting enabled or not, etc.). Note that this information lives outside of the state. - It also implements the
Component<S, AS>
trait. - However, for the reusable editor component we need the data representing the
document being edited to be stored in the state (
EditorBuffer
) and not inside of theEditorComponent
itself.- This is why the state must implement the trait
HasEditorBuffers
which is where the document data is stored (the key is the id of the flex box in which the editor component is placed). - The
EditorBuffer
contains the text content in aVec
ofUnicodeString
. Where each line is represented by aUnicodeString
. It also contains the scroll offset, caret position, and file extension for syntax highlighting.
- This is why the state must implement the trait
In other words,
EditorEngine
-> This goes inEditorComponent
- Contains the logic to process keypresses and modify an editor buffer.
EditorBuffer
-> This goes in theState
- Contains the data that represents the document being edited. This contains the caret (insertion point) position and scroll position. And in the future can contain lots of other information such as undo / redo history, etc.
Here are the connection points w/ the impl of Component<S, AS>
in EditorComponent
:
handle_event(global_data: &mut GlobalData<S, AS>, input_event: InputEvent, has_focus: &mut HasFocus)
- Can simply relay the arguments to
EditorEngine::apply(state.editor_buffer, input_event)
which will return anotherEditorBuffer
. - Return value can be dispatched to the store via an action
UpdateEditorBuffer(EditorBuffer)
.
- Can simply relay the arguments to
render(global_data: &mut GlobalData<S, AS>, current_box: FlexBox, surface_bounds: SurfaceBounds, has_focus: &mut HasFocus,)
- Can simply relay the arguments to
EditorEngine::render(state.editor_buffer)
- Which will return a
RenderPipeline
.
- Can simply relay the arguments to
Painting the caret
Definitions:
-
Caret
- the block that is visually displayed in a terminal which represents the insertion point for whatever is in focus. While only one insertion point is editable for the local user, there may be multiple of them, in which case there has to be a way to distinguish a local caret from a remote one (this can be done w/ bg color). -
Cursor
- the global "thing" provided in terminals that shows by blinking usually where the cursor is. This cursor is moved around and then paint operations are performed on various different areas in a terminal window to paint the output of render operations.
There are two ways of showing cursors which are quite different (each w/ very different constraints).
-
Using a global terminal cursor (we don't use this).
- Both termion::cursor and crossterm::cursor support this. The cursor has lots of effects like blink, etc.
- The downside is that there is one global cursor for any given terminal window.
And this cursor is constantly moved around in order to paint anything (eg:
MoveTo(col, row), SetColor, PaintText(...)
sequence).
-
Paint the character at the cursor w/ the colors inverted (or some other bg color) giving the visual effect of a cursor.
- This has the benefit that we can display multiple cursors in the app, since this is not global, rather it is component specific. For the use case requiring google docs style multi user editing where multiple cursors need to be shown, this approach can be used in order to implement that. Each user for eg can get a different caret background color to differentiate their caret from others.
- The downside is that it isn't possible to blink the cursor or have all the other "standard" cursor features that are provided by the actual global cursor (discussed above).
How do modal dialog boxes work?
A modal dialog box is different than a normal reusable component. This is because:
- It paints on top of the entire screen (in front of all other components, in
ZOrder::Glass, and outside of any layouts using
FlexBox
es). - Is "activated" by a keyboard shortcut (hidden otherwise). Once activated, the user can accept or cancel the dialog box. And this results in a callback being called w/ the result.
So this activation trigger must be done at the App
trait impl level (in the
app_handle_event()
method). Also, when this trigger is detected it has to:
- When a trigger is detected, send a signal via the channel sender (out of band) so that it will show when that signal is processed.
- When the signal is handled, set the focus to the dialog box, and return a
EventPropagation::ConsumedRerender
which will re-render the UI w/ the dialog box on top.
There is a question about where does the response from the user (once a dialog is
shown) go? This seems as though it would be different in nature from an
EditorComponent
but it is the same. Here's why:
- The
EditorComponent
is always updating its buffer based on user input, and there's no "handler" for when the user performs some action on the editor. The editor needs to save all the changes to the buffer to the state. This requires the trait boundHasEditorBuffers
to be implemented by the state. - The dialog box seems different in that you would think that it doesn't always
updating its state and that the only time we really care about what state the dialog
box has is when the user has accepted something they've typed into the dialog box
and this needs to be sent to the callback function that was passed in when the
component was created. However, due to the reactive nature of the TUI engine, even
before the callback is called (due to the user accepting or cancelling), while the
user is typing things into the dialog box, it has to be updating the state,
otherwise, re-rendering the dialog box won't be triggered and the user won't see
what they're typing. This means that even intermediate information needs to be
recorded into the state via the
HasDialogBuffers
trait bound. This will hold stale data once the dialog is dismissed or accepted, but that's ok since the title and text should always be set before it is shown.- Note: it might be possible to save this type of intermediate data in
ComponentRegistry::user_data
. And it is possible forhandle_event()
to return aEventPropagation::ConsumedRerender
to make sure that changes are re-rendered. This approach may have other issues related to having both immutable and mutable borrows at the same time to some portion of the component registry if one is not careful.
- Note: it might be possible to save this type of intermediate data in
Two callback functions
When creating a new dialog box component, two callback functions are passed in:
on_dialog_press_handler()
- this will be called if the user choose no, or yes (w/ their typed text).on_dialog_editors_changed_handler()
- this will be called if the user types something into the editor.
How to use this dialog to make an HTTP request & pipe the results into a selection area?
So far we have covered the use case for a simple modal dialog box. In order to provide
auto-completion capabilities, via some kind of web service, there needs to be a
slightly more complex version of this. This is where the DialogEngineConfigOptions
struct comes in. It allows us to create a dialog component and engine to be configured
w/ the appropriate mode - simple or autocomplete.
In autocomplete mode, an extra "results panel" is displayed, and the layout of the dialog is different on the screen. Instead of being in the middle of the screen, it starts at the top of the screen. The callbacks are the same.
How to make HTTP requests
Crates like reqwest
and hyper
(which is part of Tokio) will work. Here's a link
that shows the pros and cons of using each:
Custom Markdown (MD) parsing and custom syntax highlighting
The code for parsing and syntax highlighting is in [try_parse_and_highlight].
A custom Markdown parser is provided to provide some extensions over the standard Markdown syntax. The parser code is in the [parse_markdown()] function. Here are some of the extensions:
- Metadata title (eg:
@title: <title_text>
). Similar to front matter. - Metadata tags (eg:
@tags: <tag1>, <tag2>
). - Metadata authors (eg:
@authors: <author1>, <author2>
). - Metadata date (eg:
@date: <date>
).
Some other changes are adding support for smart lists. These are lists that span multiple lines of text. And indentation levels are tracked. This information is used to render the list items in a way that is visually appealing.
- The code for parsing smart lists is in [parse_smart_list].
- The code for syntax highlighting is in StyleUSSpanLines::from_document.
Also, syntect
crate is still used by the editor component
EditorEngineApi::render_engine to syntax highlight the text inside code blocks of
Markdown documents.
An alternative approach to doing this was considered using the crate markdown-rs
,
but we decided to implement our own parser using
nom
since it was
streaming and used less CPU and memory.
Grapheme support
Unicode is supported (to an extent). There are some caveats. The r3bl_core::UnicodeString struct has lots of great information on this graphemes and what is supported and what is not.
Lolcat support
An implementation of lolcat color wheel is provided. Here's an example.
use r3bl_core::*;
use r3bl_tui::*;
let mut lolcat = LolcatBuilder::new()
.set_color_change_speed(ColorChangeSpeed::Rapid)
.set_seed(1.0)
.set_seed_delta(1.0)
.build();
let content = "Hello, world!";
let unicode_string = UnicodeString::from(content);
let lolcat_mut = &mut lolcat;
let st = lolcat_mut.colorize_to_styled_texts(&unicode_string);
lolcat.next_color();
This r3bl_core::Lolcat that is returned by build()
is safe to re-use.
- The colors it cycles through are "stable" meaning that once constructed via the
builder (which sets the speed, seed, and delta that
determine where the color wheel starts when it is used). For eg, when used in a
dialog box component that re-uses the instance, repeated calls to the
render()
function of this component will produce the same generated colors over and over again. - If you want to change where the color wheel "begins", you have to change the speed, seed, and delta of this r3bl_core::Lolcat instance.
Issues and PRs
Please report any issues to the issue tracker. And if you have any feature requests, feel free to add them there too ๐.
License: Apache-2.0
Dependencies
~19โ32MB
~492K SLoC