[RFC] gogpu/ui — Enterprise-Grade GUI Toolkit for Go

[kolkov]

Overview

We are excited to announce gogpu/ui — an enterprise-grade GUI toolkit for Go, designed to enable building professional applications such as:

• IDEs (GoLand, VS Code class)
• Design Tools (Figma, Photoshop class)
• CAD Applications
• Professional Dashboards
• Chrome/Electron Replacement Apps

This RFC seeks community feedback on our architecture decisions and development roadmap before we begin Phase 1 implementation.

---

Key Differentiators

Feature	gogpu/ui	Fyne	Gio
CGO-free	Yes	No	Yes
WebGPU rendering	Yes	OpenGL	Direct GPU
Reactive state	Signals	Binding	Events
Layout engine	Flexbox + Grid	Custom	Flex
Virtualization	Yes (100K+ items)	Limited	Manual
IDE docking	Yes	No	No

---

Architecture

┌─────────────────────────────────────────────────────────────┐
│                    User Application                         │
├─────────────────────────────────────────────────────────────┤
│  theme/material3   │  theme/fluent   │  theme/cupertino     │
├─────────────────────────────────────────────────────────────┤
│  widgets/          │  docking/       │  animation/          │
│  Button, TextField │  DockingHost    │  Animation, Spring   │
├─────────────────────────────────────────────────────────────┤
│  layout/                            │  state/               │
│  VStack, HStack, Grid, Flexbox      │  Signals              │
├─────────────────────────────────────────────────────────────┤
│  core/                              │  event/               │
│  Widget, WidgetBase, Context        │  Mouse, Keyboard      │
├─────────────────────────────────────────────────────────────┤
│  gogpu/gg          │  gogpu/gogpu    │  coregx/signals      │
│  2D Graphics       │  Windowing      │  State Management    │
└─────────────────────────────────────────────────────────────┘

Full architecture details: docs/ARCHITECTURE.md

---

Planned API

package main

import (
    "github.com/gogpu/gogpu"
    "github.com/gogpu/ui/widgets"
    "github.com/gogpu/ui/layout"
    "github.com/coregx/signals"
)

func main() {
    app := gogpu.NewApp(gogpu.Config{
        Title:  "My Application",
        Width:  1280,
        Height: 720,
    })

    // Reactive state
    count := signals.New(0)

    // Declarative UI
    root := layout.VStack(
        widgets.Text("Counter Demo").FontSize(24),
        layout.HStack(
            widgets.Button("-").OnClick(func() {
                count.Set(count.Get() - 1)
            }),
            widgets.Text(signals.Computed(func() string {
                return fmt.Sprintf("Count: %d", count.Get())
            })),
            widgets.Button("+").OnClick(func() {
                count.Set(count.Get() + 1)
            }),
        ).Spacing(8),
    ).Spacing(16).Padding(24)

    app.SetRoot(root)
    app.Run()
}

---

Development Roadmap

Phase	Version	Description	Estimated LOC
Phase 1	v0.1.0	MVP: Core, layout, events	~12K
Phase 2	v0.2.0	Beta: Widgets, Material 3	~10K
Phase 3	v0.3.0	RC: Virtualization, animation	~10K
Phase 4	v1.0.0	Production: Docking, a11y, themes	~23K
Total			~55K LOC

Full roadmap: ROADMAP.md

---

Versioning Strategy

We follow a conservative v0.x.x strategy:

• Stay on v0.x.x as long as possible (API can evolve)
• v1.0.0 only when API is stable for 1+ year
• Avoid v2.0.0 at all costs (requires /v2 import path)

Backward Compatibility Patterns:

• Functional options for configuration
• Interface extension for capabilities
• Config structs with zero-value defaults
• internal/ for implementation details
• experimental/ for unstable features

Full policy: docs/VERSIONING.md

---

Architecture Update (December 25, 2025)

After conducting deep research on 7 leading Rust UI frameworks (GPUI/Zed, Xilem, Floem/Lapce, iced, egui, Slint, Dioxus) and analyzing modern UI patterns, we have finalized key architectural decisions:

Confirmed Design Decisions

Decision	Choice	Rationale
Reactivity Model	Fine-grained signals	O(affected) updates — only widgets using changed signal re-render. Proven by Floem (Lapce editor).
Styling API	Tailwind-style builders	Type-safe, IDE autocomplete, optimized for AI coding agents.
Layout Engine	Flexbox with incremental cache	Industry standard (Taffy-like), dirty-subtree-only recomputation.
Accessibility	AccessKit schema	Cross-platform standard used by Xilem, egui.
Effect Handling	Batched updates	Multiple signal changes = single render (like React 18, SolidJS).

Why Tailwind-style API?

We evaluated API ergonomics from both developer and AI coding agent perspective (Claude, Copilot, Cursor). Key findings:

Metric	Tailwind-style	Struct-based	CSS-like
Tokens to generate	~150	~280	~200
Compile-time errors	100%	95%	60%
Context switching	0 files	0 files	2+ files
IDE autocomplete	Excellent	Good	Poor

Updated API Example

func UserCard(user User) Widget {
    return HStack(
        Avatar(user.ImageURL).Size(48).Rounded(24),
        VStack(
            Text(user.Name).Font(typography.Title).Color(theme.OnSurface),
            Text(user.Email).Font(typography.Body).Color(theme.OnSurfaceVariant),
        ).Gap(4).AlignStart(),
        Spacer(),
        Button("Follow").Variant(Outlined).OnClick(func() {
            followUser(user.ID)
        }),
    ).Padding(16).Gap(12).Background(theme.Surface).Rounded(12).Shadow(1)
}

What We Decided NOT to Implement

Pattern	Reason
View/Widget separation (Xilem-style)	Unnecessary with fine-grained signals
Elm/MVU architecture	Too verbose, message types for every action
CSS-like stylesheets	Runtime errors, split context
Multiple styling APIs	Consistency > flexibility

---

Questions for the Community

We would appreciate feedback on the following:

1. API Design

• Does the declarative API feel Go-idiomatic?
• Are there patterns from other Go libraries we should adopt?

2. Widget Priority

Which widgets are most critical for your use cases?

• DataGrid / Table with virtualization
• TreeView with lazy loading
• Rich text editor
• Code editor component
• Chart/Graph widgets
• Other (please comment)

3. Theme System

• Should we prioritize Material 3, Fluent, or platform-native themes?
• Interest in a CSS-like styling system?

4. Accessibility

• What accessibility features are most important?
• Experience with WCAG compliance in Go applications?

5. Migration

• Are you migrating from Fyne, Gio, or another toolkit?
• What pain points should we address?

---

Resources

Resource	Link
Repository	github.com/gogpu/ui
Roadmap	ROADMAP.md
Architecture	docs/ARCHITECTURE.md
Versioning Policy	docs/VERSIONING.md
Contributing	CONTRIBUTING.md

---

Get Involved

• Star the repository to show interest
• Comment below with your feedback
• Watch for updates on development progress
• Contribute design ideas and API feedback

We believe Go deserves an enterprise-grade GUI toolkit, and we are building it with the community input.

---

The gogpu team

[kolkov]

Angular 20/21 Patterns Proposals (December 25, 2025)

After researching Angular 20/21 architecture, we propose additional patterns that could enhance gogpu/ui. These are proposals for community discussion, not final decisions.

Proposed Patterns

Pattern	Description	Status
linkedSignal	Writable computed with reset capability	Proposed
Signal Forms	Form abstraction with reactive validation	Proposed
Context DI	Type-safe DI for infrastructure only	Proposed
Control Flow DSL	If/For/Switch builders	Proposed
Feature Stores	State management best practices	Proposed

1. linkedSignal

Derived value that can be overridden, then reset:

// Form field linked to source, but editable
name := signals.NewLinked(func() string { return user.Get().Name })
name.Set("Override")  // User edits
name.Reset()          // Back to derived value
name.IsDirty()        // Check if modified

2. Signal Forms

Form-level abstraction with validation:

form := forms.NewGroup(
    forms.Field("email", forms.Required(), forms.Email()),
    forms.Field("password", forms.Required(), forms.MinLength(8)),
)
TextField("Email").Bind(form.Field("email"))
Button("Submit").Disabled(signals.Not(form.Valid()))

3. Context DI (Infrastructure Only)

Type-safe injection for Theme, Logger, Config, i18n. Business logic via explicit parameters.

4. Control Flow DSL

Declarative If/For/Switch builders alongside native Go control flow.

Questions for Discussion

1. Is linkedSignal useful for your use cases?
2. Should Signal Forms be core or separate package?
3. Is limiting Context DI to infrastructure the right approach?
4. Preference: Native Go vs Control Flow DSL?

We want your feedback before finalizing these decisions!

[kolkov]

Thank you for the first real-world feedback! This is exactly the kind of practical insight we need.

You've identified a key ergonomics problem. The verbose signals.Computed(func() string {...}) pattern is indeed painful for common use cases.

Proposed Solutions

We're considering several approaches that don't require go generate:

1. BindFormat helper

// Instead of signals.Computed(func() string { return fmt.Sprintf(...) })
Text().BindFormat("Count: %d", count)

2. Direct signal binding with auto-conversion

// Widget accepts Signal[T] and converts automatically
Text().Bind(count) // int → string via fmt.Sprint

3. Method chaining for conditional properties

Button("Click").
    Bind(count, "Count: %d").
    DisabledWhen(signals.LessThan(count, 0)).
    ClassWhen("highlight", signals.GreaterThan(count, 10))

4. Computed shorthand

// Helper for simple format cases
label := signals.Format("Count: %d", count)
Text().Bind(label)

Our Direction

We want to avoid go generate magic — it complicates debugging and IDE support. Instead, we'll provide ergonomic helpers for common patterns while keeping the explicit Computed() for complex cases.

Question for you: Which approach feels most natural? Would BindFormat + ClassWhen/DisabledWhen cover most of your use cases?

[sedyh]

Does the declarative API feel Go-idiomatic?

Hello. Is it possible to use functional options instead of fluent-like api? As far as I know gofmt doesn't really like multiline method chains and I find them pretty inconcinient to use in a single line with bigger fonts due to vision problems. Also struct options and similar to func options with-wrappers are most go-idiomatic way to build things in stdlib.

[crsolver]

Also if AI can do this, why is it building a cross platform library? why not use AI to build a native app for each platform directly? that would be easier.

[typenoob]

Also if AI can do this, why is it building a cross platform library? why not use AI to build a native app for each platform directly? that would be easier.

I don't think this repo do real valuable things. It neither provide a high-level api to create a GUI application, nor low-level api to create the GUI framework. All it can do is to run some demos which can be done even by invoking functions from vulkan or opengl's dlls.

[kolkov]

I hope that was sarcasm :) In case it wasn't — feel free to explore the codebase: 460K+ lines of Pure Go across the ecosystem, zero CGO, 5 rasterization algorithms in gg, full WebGPU HAL in wgpu, shader compiler in naga, and a GUI toolkit with 1500+ tests at 97% coverage. All open source.

[kivutar]

This ecosystem is working very well for me. It has rough edges yes, but the value is already great.

[kolkov]

This ecosystem is working very well for me. It has rough edges yes, but the value is already great.

@kivutar Thank you! We hope to reach v1.0.0 by December 1, 2026. But to achieve this, the entire community must help with testing, debugging, and bringing the ecosystem to enterprise level.

[kolkov]

Hi @sedyh, thank you for the thoughtful feedback!

If I understand correctly, you're suggesting we prioritize functional options over fluent/method-chaining APIs because:

1. gofmt doesn't handle multiline method chains well
2. Functional options are more Go-idiomatic (matching http.Server, grpc.Dial, etc.)
3. Single-line method chains can be hard to read with larger fonts (accessibility)

These are excellent points that deserve careful consideration.

Our Initial Research

During architecture planning, we evaluated API approaches from both human and AI perspectives:

Metric	Fluent (Tailwind-style)	Struct/Options	CSS-like
Compile-time error detection	100%	95%	60%
IDE autocomplete	Excellent	Good	Poor
Human readability	Linear flow	Explicit	Split context
AI code generation accuracy	~97%	~93%	~85%

Fluent APIs scored well for both: humans get IDE autocomplete and linear reading, while AI agents generate fewer errors due to immediate compile-time feedback.

Context Matters: When Fluent Works Well

Fluent API works well for builder patterns — it's simply convenient:

// SQL query building (ozzo-dbx, relica)
db.Select("id", "name").
    From("users").
    Where(dbx.HashExp{"id": 100}).
    One(&user)

Methods can be called in any order (query generates at the end), and it's readable for both humans and AI.

Your Points Add Important Dimensions

For widget construction, Go stdlib takes a different approach:

• Functional options pattern (http.Server, grpc.Dial, slog.New)
• gofmt formats them cleanly across multiple lines
• Accessibility — critical for developers with vision needs
• Familiar to all Go developers

The 2026+ Consideration

We believe that in 2026 and beyond, many developers will transition to AI-augmented development workflows. This means we need to carefully balance:

1. Human ergonomics — gofmt, readability, Go idioms
2. AI agent ergonomics — compile-time errors, linear flow, minimal context
3. IDE experience — autocomplete, refactoring support

Options We're Considering

Option A: Functional Options Only (most Go-idiomatic)

btn := ui.NewButton(
    ui.WithText("Submit"),
    ui.WithPadding(16),
    ui.WithBackground(theme.Primary),
    ui.WithOnClick(handleClick),
)

• gofmt-friendly, stdlib pattern
• Best for accessibility
• Slightly more tokens for AI

Option B: Hybrid Approach (construction + styling)

// Construction: Functional Options
btn := ui.NewButton(
    ui.WithText("Submit"),
    ui.WithOnClick(handleClick),
)

// Styling: Fluent (like query builders)
btn.Padding(16).Background(theme.Primary).Rounded(8)

• Separates concerns (construction vs styling)
• Fluent styling is convenient (like query builders)
• Both patterns proven in production

Option C: Config Struct (for many options)

btn := ui.NewButton(ui.ButtonConfig{
    Text:      "Submit",
    Padding:   16,
    Background: theme.Primary,
    OnClick:   handleClick,
})

• Familiar, explicit
• Good for complex configurations
• Less composable than options

Our Approach

We will carefully evaluate all approaches before deciding, considering:

• Developer experience (humans first!)
• AI agent compatibility (2026+ reality)
• gofmt and tooling behavior
• Accessibility requirements
• Proven patterns (ozzo-dbx, relica show fluent works for DSLs)

We may end up with a thoughtful combination rather than a single approach. For example:

• Functional options for widget construction
• Fluent for styling DSL (convenient like query builders)
• Config structs for complex initialization

Thank you for raising the vision accessibility angle — this is exactly the kind of feedback that helps us build a better library for everyone.

What would work best for your use case? Would you prefer pure functional options, or could the hybrid approach (Option B) work?

Relica SQL Builder

[soypat]

While I'm here I'd like to raise awareness of the pure-go rewrite version of clay UI. Heapless, dead-simple, robust-as-you-can-get, easy-to-use UI.

https://github.com/soypat/glay

To understand how something so simple can work so well: https://www.youtube.com/watch?v=by9lQvpvMIc

Still missing text rendering facilities though. This approach would keep the codebase small, lean and efficient. Also no third party dependencies.

[kolkov]

Hi @crsolver,

I noticed you've been leaving similar comments across multiple platforms (GitHub, Dev.to, Reddit) under different accounts. That's unusual behavior for GitHub — first time I've seen this kind of cross-platform pattern here.

What's interesting is that your first comment on Dev.to was actually a thoughtful question about retained-mode GUI and raylib. We had a constructive exchange there. Then somewhere along the way, "AI slop" and "clanker" became your contribution style.

If you have genuine technical concerns — bring them as issues or PRs. That's how open source works. Criticize? Sure, but propose alternatives. Found a bug? Report it. Think the architecture is wrong? Show us a better approach.

Or build your own ecosystem the way you think it should be done. That's the beauty of open source — you can fork, improve, or start fresh. 250K+ lines of working code is there for anyone to learn from, critique, or build upon.

But commenting "AI slop" without any technical substance across multiple platforms isn't criticism — it's just noise.

The door is open for real contributions. 🚪

[crsolver]

well, they are different websites, so off course I use different accounts for each, how's that unusual? You wouldn't find it unusual if you were a real software engineer.

[soypat]

I'd like to back @crsolver on the matter of the quality of responses and interaction in my short time in the gogpu ecosystem. My interaction was limited to gogpu/naga#6 but it was enough to for me to leave gogpu with an unfavourable perception of the project. @kolkov real human interaction goes a long way to creating trust between humans. Really hope you invest time into actually using your own words and building a personality the internet community can get a feel for rather than just using an AI to reply to comments. This is important so that we as interested parties of the project get an idea for who provides the technical leadership (is it you, or the computer?). Wishing you the best for gogpu, really looks awesome!

[crsolver]

This guy doesn't understand anything about his own project, because he hasn't written any of it. cogentcore/webgpu#14 (comment)

[beikege]

Does it support Android?

[crsolver]

You might want to look for other options, or wait until they achieve AGI

[kolkov]

Hi @beikege,

Short answer: not yet.

Longer answer: Android support is technically feasible, but there are some fundamental challenges I want to be transparent about.

Current Status

Right now gogpu supports Windows, macOS, and Linux (X11/Wayland). The GPU backends (Vulkan, GLES) could theoretically work on Android since the APIs exist there, but the platform layer (windowing, input, lifecycle) is the blocker.

The CGO Problem

After researching this thoroughly, I found that all production Go graphics libraries — Gio, Fyne, Ebitengine — require CGO for Android. The reason is architectural:

• On desktop, Go is the main process. We call system APIs directly.
• On Android, the Java VM is the host. Go code runs as a guest library loaded via JNI.

This means:

• Entry point is JNI_OnLoad(), not main() — requires C-callable exports
• Activity lifecycle (onCreate, onPause, etc.) comes through JNI callbacks
• ANativeWindow handle is passed from Java via JNI

There's no pure Go workaround for this. Even purego uses CGO internally on Android.

gogpu's philosophy has been zero-CGO for easy cross-compilation. Adding Android would mean accepting CGO as an Android-only requirement.

Potential Paths

Approach	Trade-offs
CGO + JNI (standard)	Full support, but breaks zero-CGO philosophy
WebView + WASM	No CGO, but WebGPU in Android WebView is still inconsistent
Termux	Works today as Linux app, but niche use case
Hybrid	Go for logic, native Kotlin/Java for UI — no GPU rendering from Go

Estimated Effort

If we go the CGO route:

• Platform layer (JNI, Activity lifecycle, input): ~2,000 LOC
• Vulkan/GLES Android surface support: ~500 LOC
• Build integration (gomobile): ~1 week

Total: roughly 4-5 weeks of focused work.

Current Priority

We're stabilizing the desktop platforms first. Android is on the radar but not in the immediate roadmap. The ecosystem (wgpu, naga, gg) needs to mature before expanding to mobile.

That said, I'm curious about your use case. Are you:

• Evaluating gogpu for a specific project?
• Interested in contributing Android support?
• Just exploring what's possible?

This would help me understand the demand better.

[kolkov]

Update: After more research, I found a more elegant approach.

Minimal Bootstrap Architecture

Instead of ~2,000 lines of CGO (like Gio/Fyne), we could use a "minimal bootstrap" pattern:

┌─────────────────────────────────────┐
│  bootstrap.c (~80 lines)            │
│  - JNI entry points                 │
│  - Get ANativeWindow* from Java     │
│  - Pass to Go via callback          │
└─────────────────┬───────────────────┘
                  │ uintptr
                  ▼
┌─────────────────────────────────────┐
│  100% Pure Go (goffi/purego)        │
│  - vkCreateAndroidSurfaceKHR()      │
│  - All Vulkan rendering             │
│  - All application logic            │
└─────────────────────────────────────┘

This preserves gogpu's "Pure Go" philosophy — only ~80 lines of C as a bridge, everything else remains Go.

Why This Works

CGO is only mandatory for JNI callbacks (entry points from Java). Once we have the native window pointer, all GPU operations can use purego / goffi to call Vulkan directly.

Additional Context

• Go 1.26 reduces CGO overhead by ~30% (release notes)
• Projects like xlab/android-go and JNIGI show this pattern works
• There's an open issue from 2014 about Android without CGO — still unresolved

Timeline

This is still not immediate priority, but the minimal bootstrap approach makes it more feasible than I initially described. A PoC could be done in 1-2 weeks when we get to it.

If you're interested in Android specifically, let me know your use case — it helps prioritize.

[kolkov]

I've created a dedicated discussion for Android support: Android Support — Research & Community Interest

Feel free to share your use case and follow updates there.

[kolkov]

Ecosystem Update: Ready for gogpu/ui Development

January 30, 2026 Release Summary

We've completed a major ecosystem integration that provides the foundation for gogpu/ui:

Package	Version	Key Feature
gpucontext	v0.4.0	TextureDrawer, TextureCreator, TouchEventSource interfaces
wgpu	v0.12.0	BufferRowLength fix (correct aspect ratio)
naga	v0.9.0	Shader compiler improvements
gg	v0.22.0	gpucontext.TextureDrawer integration in ggcanvas
gogpu	v0.14.0	Context.AsTextureDrawer() implementation

---

What This Enables

Unified Rendering Interface

The gpucontext.TextureDrawer interface allows any 2D graphics library to render into any GPU window without direct dependencies:

// gpucontext/texture.go
type TextureDrawer interface {
    DrawTexture(tex Texture, x, y float32) error
    TextureCreator() TextureCreator
}

type TextureCreator interface {
    NewTextureFromRGBA(width, height int, data []byte) (Texture, error)
}

This is the exact pattern needed for gogpu/ui:

• UI widgets render to gg.Context (CPU)
• ggcanvas uploads to GPU texture
• gogpu.Context displays in window

---

Working Example

See examples/gg_integration/main.go:

app := gogpu.NewApp(gogpu.DefaultConfig().
    WithTitle("GoGPU + gg Integration").
    WithSize(800, 600))

var canvas *ggcanvas.Canvas

app.OnDraw(func(dc *gogpu.Context) {
    if canvas == nil {
        canvas, _ = ggcanvas.New(app.GPUContextProvider(), w, h)
    }

    ctx := canvas.Context()
    ctx.SetRGB(1, 0, 0)
    ctx.DrawCircle(400, 300, 100)
    ctx.Fill()

    // Render to gogpu window (one line!)
    canvas.RenderTo(dc.AsTextureDrawer())
})

---

Architecture for gogpu/ui

User Application
      │
layout/VStack, HStack  │  widgets/Button, Text
      │
ui.Renderer            │  state/Signals
      │
ggcanvas.Canvas (CPU→GPU bridge)
      │
gg.Context (2D rasterization)
      │
gpucontext.TextureDrawer ←── gogpu.Context.AsTextureDrawer()
      │
gogpu/wgpu (Pure Go WebGPU)
      │
Vulkan │ Metal │ DX12 │ GLES │ Software

---

What's Next

With these interfaces in place, we can now start gogpu/ui Phase 1:

1. Widget base — Widget interface, WidgetBase struct
2. Layout engine — VStack, HStack, Flexbox
3. Basic widgets — Text, Button, TextField
4. Event dispatch — Mouse, Keyboard, Touch (via gpucontext.TouchEventSource)
5. Theme system — Material 3 baseline

The ecosystem is ready. Let's build the UI toolkit!

https://dev.to/kolkov/gogpu-unified-2d3d-graphics-integration-in-pure-go-djf

---

Releases:

• https://github.com/gogpu/gpucontext/releases/tag/v0.4.0
• https://github.com/gogpu/wgpu/releases/tag/v0.12.0
• https://github.com/gogpu/naga/releases/tag/v0.9.0
• https://github.com/gogpu/gg/releases/tag/v0.22.0
• https://github.com/gogpu/gogpu/releases/tag/v0.14.0

[celer]

This project is fascinating to me, as someone who has previously written a golang vulkan library, and someone who currently owns a 2D/3D CAD project written in C++, I have lots of background doing golang development and deeply miss it and it's tooling, but also I recognize that AI assisted programming is the future.

I wish I could have written my CAD software in golang and gogpu offers that potential. I think the "Smart Coding" approach is also potentially a major accelerator and a major risk. My challenge with my current CAD software is that AI face plants on helping me with it, due to the tech stack I'm using, C++ and the overall complexity - resulting in my current velocity being pretty poor.

I've been playing with having AI write GUI desktop apps for me and I can see that the specific GUI library can have a big impact on the velocity and context usage of the AI agents. I can pretty much ask for complex QT6 apps and get what I want, out of Qwen3.6-27B, but I can't ask for a complex fyne.io app because the AI uses up all it's context just trying to learn how to use the framework.

I'm OK with @kolkov's approach with gogpu/ui - although now that I've sat down and played with the example apps it's clear to me for this to be a usable project a bunch of humans are going to have to sit down and scrutinize the actual UI experience as it still has lots of problems - this is an area I believe requires deep coordination with humans and AI - it's hard for AI to understand why a UI interaction feels bad.

Anyways this project holds a lot of potential - but also a lot of risk but it's going to require humans to be in the loop for testing to tell the AI what needs to be improved with the UI experience.

What I would like to do is try to adopt @kolkov "Smart Coding" approach and try a reboot of my CAD software utilizing gogpu - but I don't think I could do this without other humans involved in growing gogpu/ui, et al to be more mature, and without also more time learning about how to apply "Smart Coding" to my approach. I did vibe code the start of a scene tree for my CAD software using gogpu/wgpu - which went OK and I see the potential.

[kolkov]

@celer Thank you for the thoughtful analysis -- and welcome! Your background (Vulkan library + 2D/3D CAD in C++) is exactly the kind of perspective we need.

You're right that humans in the loop are essential for UI polish. Good news: we've recently assembled a small team of contributors and testers:

Active contributors:

• @darkliquid -- ported Quake 1 engine from C to Go, renders on gogpu/wgpu Vulkan (Ironwail-go demo #163)
• @jdbann -- 3 merged Metal PRs (cull mode, depth textures, vertex buffer indices)
• @lkmavi -- macOS platform owner (4 merged PRs: window delegate, tabbing, close lifecycle, multi-window)
• @k-chimi -- Metal compute pipeline fix + input fix
• @snakeru -- particle simulation example
• @timzifer -- Metal bind group fix
• @Carmen-Shannon -- WriteTexture API

Active testers/reviewers:

• @unxed -- rewriting Far Manager in Go on our stack. Real desktop app exercising text rendering, keyboard, HiDPI daily. Found and helped fix keyboard layout bugs (Support multiple keyboard layouts #227, Русские кавычки-«ёлочки» не вводятся через AltGr #233).
• @paulie-g -- C/asm background, X11 expertise, IBM Model M. Wants to build a Petri net/state machine IDE.
• @joeblew999 -- building offline AI & 3D CAD systems. Requested and validated WASM support (WASM in browser #70).

The lower-level libraries (wgpu, naga, gg, gpucontext) are now stable enough -- today we shipped v0.36.1 with unified xkbcommon for proper international keyboard support. So the foundation is solid and we're actively shifting focus to gogpu/ui polish and 3D.

Your CAD use case is very relevant -- @joeblew999 is in a similar space. The 3D scene tree via gogpu/wgpu is the right starting point, and once gogpu/ui matures, you'd have native Go GUI panels alongside your 3D viewport. If you hit specific UI issues while experimenting, please file them -- that "this interaction feels bad" feedback is gold.

Re: "Smart Coding" -- happy to share more about the workflow if useful. The key insight: AI does bulk coding, humans validate behavior and UX. Exactly what you described.