toolchains
This is prototype V4~ish of Hannah building an extensible event-driven dataflow environment. It's a JS-only implementation of my previous version, Planager. I originally began this project with the goal of using it for digital fabrication, but the feature set general enough that I could see it be used for many applications.
Running Locally
Running Toolchains locally is easy!
- Make sure you have Node installed.
- Clone the repository locally and
cd
into it npm install
installs the dependenciesnpm run dev
starts the local development server
Todo List
- core
- add/remove tools
- pan/zoom workspace
- drag tools
- add/remove pipes
- draw pipes between ports
- send data from outports to inports
- pipes update when tools moved
- remove pipes between ports
- save/load toolchain state
- save/load workspace view (pan zoom)
- example dropdown
- basic port type checking
- compound tools
- mobile pan zoom
- QoL things
- tool layering - any selected tool should be above others
- add tool to coordinates in current view, rather than 0,0
- clear current toolchain
- name toolchain
- pipe coloring for boolean/null values
- cache toolchains to local storage?
- adaptive pipe routing? current beziers are fine for most cases but not great for e.g. cycles.
- tool alignment/layout commands, should be available under selection
context menu
- align left, right, center
- vertical/horizontal even spread
- fade/adjust dotted background on scroll
- tool ui
- set initial width and height from tool config
- render template tool ui
- ports
- header
- state pane
- edit displayname
- collapse ui
- resize handle
- toolbar context menu
- remove tool
- collapse ui elements (hide ports when not active)
- Edit tool display name
- show state inspector
- workspace ui
- tool library
- background dots
- box select tools
- move box selection
- background context menu - maybe like grasshopper's commands
- selection context menu
- background annotations
Debug Panel / Toolchain Inspector
- current view coords and scale (in workspace coords)
- current mouse coords (in workspace coords)
- highlight tool/pipe when hovered
- inspect current port values in debug pane
Future Directions
Ideas for features that I think would be cool to implement.
- In-workspace tool definition editor. This is easily one of the future features I am most excited about working on.
- Tool suggestions. Right clicking a port gives a dropdown of tools it could
be connected to. For example, an inport with type
boolean
could suggest a boolean toggle switch tool. Clicking on the suggestion would add the tool and connect it automatically. - Auto tool configuration. Ports with built-in value checking could suggest
tools that are auto-configured for them. For example, the saturation inport
for HSL should accept a float between 0 and 100. The tool suggestion could be
a number slider where
min = 0
,max = 100
,step = 0.001
, anddisplayName = "saturation"
.
Open Questions
Things I haven't decided how to implement yet, or for which I may not have an ideal solution.
How to handle tool style scoping?- I made a custom lit directive which renders each tool's UI to a shadow DOM that is created when the tool is added to the toolchain. Unsure if this will cause performance issues related to render optimizations down the road, but it seems to work fine right now.
- How to handle in-tool DOM querying?
- Querying the DOM from within the tool function doesn't work because the view
is rendered to the shadow DOM. It is possible to use
lit-html
s'createRef
directive to manage a reference to elements defined in the tool's view template, but I don't like this approach because it relies on an external library. I've thought of a few ways of implementing this, one of which is to pass in the root shadow DOM element to the tool definition function. This would let the tool query its own DOM directly. However, this means that the shadow DOM node would need to exist before initializing the tool. Open question indeed.
- Querying the DOM from within the tool function doesn't work because the view
is rendered to the shadow DOM. It is possible to use
- How to handle tool dependencies? (stylesheets, libraries, assets, etc)
- I definitely don't want tool dependencies to be installed/bundled with the
toolchain environment. Currently, tools with external dependencies can
dynamically import them from a CDN. This works locally, but causes some
issues when using vite for bundling that I haven't looked into yet. For
example, the
editor
tool uses CodeMirror and the dynamic imports in the built/deployed version fail.
- I definitely don't want tool dependencies to be installed/bundled with the
toolchain environment. Currently, tools with external dependencies can
dynamically import them from a CDN. This works locally, but causes some
issues when using vite for bundling that I haven't looked into yet. For
example, the
- Where to place the state inspection pane?
- In Planager, I had the tool's state pane be a dropdown from the tool view. This only looks nice when the tool view is the right size, otherwise it overflows, is too narrow/wide, etc. I'm not sure where to put it in the workspace interface, especially as I think it should be possible to have multiple state panes open at once. It might be a panel that has some visual indication it is connected to a tool, like a bounding box around both of them?
- How to handle tool view resizing?
- Tools should be able to request resizing of their view. In Planager they arbitrarily expanded the view pane, which is not ideal because they have a fixed position and would then overlap other tools if they got too big. One approach would be to move other tools in the toolchain to accomodate changes in view size. I think this would be nice in some cases, but annoying in others.
- Should there be different tool types?
- My gut says no. stateless vs stateful tools? Mini (no-view) tools?
Tool Definition
Current approach: when a tool is imported, we import the default export from the tool's file. This should be a object containing the tool's configuration and a function which returns its lifecycle methods.
Tool Configuration Object
This is subject to change. See current tool implementations in tools/
for
current examples.
displayName
- The tool's display name, shown in its' view toolbarwidth
- UI initial width in pixelsheight
- UI initial height in pixelsstate
- JSON-serializable object containing initial state valuesinports
- JSON-serializable object defining the ports, their types, and their display names.outports
- thinking of removing the concept of outports alltogether. outports could just be state.
// Example configuration object - subject to change
const config = {
displayName: "Counter",
width: 200,
height: 200,
inports: {
text: {
type: "string",
},
},
outports: {
text: {
type: "bool",
},
},
state: {
colors: [],
currentColor: "#ffff00",
},
};
Tool Lifecycle Callbacks
-
init
- called when tool added to toolchain, before the tool view is added to the DOM -
render
- called during the workspace render loop. Returns somethinglit-html
can render, such as anhtml
tagged template literal. -
postInit
- called after the tool's view has been added to the DOM -
inportsUpdated
- called when one of the outports connected to a tool's inport changes. -
stateUpdated
- called when one of the tool's state values change. Could use this for built-in -
onResize
- called on changes to tool view bounds -
onZoom
/onScale
- called on changes to workspace zoom -
saveState
- called when saving a toolchain. Should return a JSON-serializable version of the state. -
loadState
- called when loading a tool, before the first render. Should accept a JSON-serialized representation of the state and initialize state accordingly
Global Toolchain Callbacks
These are callbacks that would be passed into the tool function. They allow the tool to communicate with the toolchain runtime environment.
-
log
- logs a message to a toolchain-wide console -
resize
- resizes the tool view -
notify
- sends a notification - perhaps using the browser notification API. -
isMobile
- checks if the toolchain is running in a mobile browser
Example Tool Definition
import { html } from "lit-html";
const config = {
inports: {
text: {
type: "string",
value: "asdf",
},
num: {
type: "number",
value: 57,
},
bool: {
type: "boolean",
value: false,
},
},
outports: {
text: {
type: "string",
value: "asdf",
},
num: {
type: "number",
value: 57,
},
bool: {
type: "boolean",
value: false,
},
},
state: {
colors: [],
currentColor: "#ffff00",
num: 500000,
floatieBoi: 25.8,
obj: { currentColor: "#ffff00", num: 500000, floatieBoi: 25.8 },
},
ui: {
displayName: "Test",
width: "200px",
height: "200px",
},
};
function test(inports, outports, state) {
let localVar = Math.floor(Math.random() * 40);
const init = () => {
shuffle();
};
const resize = () => {};
const shuffle = (e) => {
state.colors = new Array(30).fill(0).map(() => {
return `#${Math.floor(Math.random() * 16777215).toString(16)}`;
});
};
const render = () => {
return html`<style>
.container {
display: flex;
flex-direction: column;
height: 100%;
}
.grid-container {
display: grid;
grid-template-columns: repeat(auto-fit, minmax(30px, 1fr));
grid-auto-rows: auto;
flex: 1;
}
.grid-container > div {
width: 100%;
}
.grid-container > div:hover {
background-color: var(--pink);
}
button {
width: 100%;
}
</style>
<div class="container">
<div class="grid-container">
${state.colors.map(
(color) => html`<div style="background-color: ${color};"></div>`
)}
</div>
<button @click=${shuffle}>Shuffle!</button>
</div>`;
};
return { init, resize, render };
}
export default { config, tool: test };
Icebox Features
These are things that are probably not worth taking the time to do, but I'll list them anyway as I think of them.
- pipe layering - A foreground tool's pipes would be drawn over a background tool if they are overlapping. The actual benefits from doing this are small and it would require making another SVG layer and tool view layer, if not multiple.
Project History
V0: NoFlo. I initially tried to use NoFlo to handle dataflow. I quickly decided that relying on an existing dataflow framework was too limiting, particularly because I wanted to be able to easily write custom tool views to visualize and modify tool state - something that was not easily acheivable with the framework.
V1: Client-side JS + React. Despite knowing some JavaScript, this version was mostly about me learning JavaScript and React. Many mistakes were made and terrible code was written. Thinking back on it now, it's clear how little I knew at the start.
V2: React + Python Backend. I realized many of the example applications I wanted to build could easily be built from Python modules. For example, I wanted to work with the AxiDraw plotter, which has a great Python library for controlling it. I also had far more experience writing Python than I did JavaScript. So, I wrote a Python library for managing the data flow and the toolchain data structure. This version dynamically creates websocket connections with the frontend, which pass port and state data between the back- and front-ends. The frontend react components for each tool received proxy objects for this data as props, and could modify the port/state values (which automatically sent data to the backend). I again learned a lot building this version.
V3: Lit + Python Backend. React's reliance on the virtual DOM made it hard to build tools that used libraries that manipulate the DOM. Every workaround felt hacky, such as needing to use a React-specific version of some library (e.g. D3) that might not be well maintained or documented. Given that one of my core goals is extensibility, particularly by others who might not be familiar with the nuances of a virtual DOM, I thought it better to scrap React in favor of a more external-library friendly component system. So I switched to Lit, which extends HTMLElement rather than relying on JSX + a virtual DOM layer. (This also meant I no longer had to deal with JSX transpilation, a big win).
V4: Vanilla JS + lit-html
. Working within a component system began to be
more annoying than beneficial, particularly when it came to managing the
toolchain state. I also piloted V3 of the system in a small workshop study, and
one thing that became clear was that requiring the full stack was far clunkier
than desired. It also complicated building and deployment. So I started over in
vanilla JS with just the HTML templating and rendering functionality provided by
the standalone lit-html
library. I am far happier with the current
implementation, and see many clear paths toward implementing fun and interesting
features.