Iam 12 . I Stopped Coding Features and Started Designing Experiences. Here Is KODA cursor v0.4.0.
I am 12 years old. My development machine is a POCO C55 ($150 USD). It is December 2nd. Yesterday, I shipped v0.3.1. It worked. But it felt like a prototype. Today, I shipped KODA v0.4.0. And for the first time, it
I am 12 years old.
My development machine is a POCO C55 ($150 USD).
It is December 2nd.
Yesterday, I shipped v0.3.1. It worked. But it felt like a prototype.
Today, I shipped KODA v0.4.0. And for the first time, it feels like a product.
Most solo founders obsess over adding AI capabilities. They add more models, more prompts, more complexity.
I did the opposite. In v0.4.0, I added zero new AI logic. Instead, I focused on three things that actually make software usable:
- Control (Settings Drawer)
- Choice (Code Master Model)
- Delight (UI Overhaul)
Because users donβt care if your LLM has 70 billion parameters if they canβt change the temperature without editing JSON files.
β‘ THE THREE PILLARS OF V0.4.0
1. Control: The Native Settings Drawer
Before v0.4.0, configuring KODA meant opening settings.json and typing keys like koda.autoApply. That is hostile UX.
Now, there is a slide-in drawer triggered by a gear icon βοΈ.
- 8 Configurable Options: Model selection, Temperature slider, Auto-apply toggle, Web Search default, Source visibility, Delete confirmation, Context file limit, Worker URL override.
- Live Reload: Change a setting? It applies instantly. No restart. No reload window.
- Two-Way Binding: Edit the drawer? It updates
settings.json. Editsettings.jsonmanually? The drawer updates instantly.
This isnβt just convenience. Itβs respect for the userβs time.
2. Choice: Introducing "Code Master"
Not every task needs the fastest model. Debugging complex race conditions or refactoring legacy monoliths requires reasoning, not just prediction.
So I added a fourth option: DeepSeek R1 Distill Qwen 32B.
| Mode | Model | Best For | Latency |
|---|---|---|---|
| Auto | GPT-OSS-120B | General purpose | 1.5β3s |
| Smart | GPT-OSS-120B | Quality focus | 1.5β3s |
| Fast | GPT-OSS-20B | Speed/Budget | <1.5s |
| Code Master | DeepSeek R1 | Reasoning/Debugging | 5β15s |
The Magic Trick:
DeepSeek R1 emits intermediate "thinking" tokens. Most tools dump these into the chat, confusing users. My Worker parses ONLY the final delta.content, discarding the reasoning trace.
Result? You get the wisdom of a reasoning model with the cleanliness of a standard chat.
3. Delight: The UI Overhaul
A tool should feel good to use. So I stripped away the noise.
- Icon-Only Header: Less text, more space. Hover reveals tooltips.
- Pulsing Streaming Indicator: Replaced the blinking block cursor with a smooth, breathing dot. It signals "alive" without distracting.
- Date-Grouped History: Conversations are no longer a flat list. They are grouped by "Today," "Yesterday," "Last Week."
- Search Box: Find any past chat instantly.
- Starter Chips: Empty state now offers clickable prompts (
/explain,/fix) so you never stare at a blank input box. - Cleaner Messages: User bubbles are compact right-aligned. Assistant messages are full-width left-aligned. Code blocks have subtle left-accent bars instead of heavy borders.
It looks less like a plugin and more like a native part of VS Code.
π οΈ THE ARCHITECTURE: STILL VANILLA JS
Despite all this polish, the core constraint remains: No Build Step.
- Total Size: ~22 KB.
- Dependencies: Zero.
- Language: Vanilla JavaScript.
How? By leveraging VS Codeβs native APIs smarter, not harder.
- Used
vscode.workspace.onDidChangeConfigurationfor live-reload settings. - Used CSS variables for theme-aware styling.
- Kept the Worker simple: One map for model routing, one stream parser for SSE.
Complexity is easy. Simplicity under pressure is hard.
π WHY THIS MATTERS
In v0.1.0, I proved I could build it.
In v0.2.0, I proved I could make it fast.
In v0.3.0, I proved I could make it stable.
In v0.4.0, I proved I can make it loved.
Software is not just code. It is the sum of interactions. Every pixel, every millisecond of latency, every configuration choice shapes the userβs trust.
By giving users control (Settings), choice (Models), and delight (UI), I transformed KODA from a "cool experiment" into a "daily driver."
π₯ GET THE POLISHED BUILD
Download v0.4.0 here: github.com/harun-sket/Koda-cursor/releases/tag/v0.4.0
- Grab
koda-cursor-0.4.0.vsix(22 KB). - Install via
Ctrl+Shift+Pβ "Extensions: Install from VSIX". - Restart VS Code.
- Click the Gear Icon βοΈ. Switch to Code Master.
- Paste a messy, buggy function. Type
/debug. - Watch it reason silently, then deliver the perfect fix.
Age doesnβt matter. Device doesnβt matter.
Craftsmanship matters.
See you in v0.5.0. Where we go multi-step agent loops. π―β¨
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.