Hello Vibe

Summary

A Kiro IDE extension that helps coding beginners start from one technology they want to learn and finish a TypeScript service. While agents write the code, the user picks the real decisions that come up during development, and "the agent wrote the code" is recorded separately from "the user understood it". Built by a two-person team for the Korea University × AWS AI Innovators Challenge, where it advanced to the finals.

What I Did

• Built the initial structure of the Agent Panel in the Kiro extension, with Builder and Helper tabs and streamed responses
• Designed and built the screen flow from learning goal input through project candidates to Learning Spec confirmation
• Integrated the Core frontend SDK and blocked features when the Core connection was unavailable
• Connected the Builder, Helper, Decision and Evidence screens to the live local Core
• Wrote Korean guidance for each error code found in live testing, and reported backend issues
• Built the finals version's 'Open in Kiro chat' action and the window reload notice
• Fixed backend toolchain errors on Windows, and built and checked the Windows installer on a real PC

Key Features

Project discovery

Entering a technology to learn returns ten project candidates, which the user can narrow down with feedback.

Learning Spec

The user reviews, refines and confirms the Learning Spec, and can go back to project discovery.

Builder decision cards

Real decisions the Builder agent meets during development appear as cards, and work continues with the user's choice.

Read-only Helper

A read-only conversation for explanations that does not modify files.

Evidence and concept state

Evidence of learning is recorded, and concept state only moves up through four fixed stages; downgrades are refused.

Local storage with redaction

State is stored in SQLite on the user's computer, and known forms of tokens, keys and home paths are redacted before saving.

Kiro chat handoff

In the finals version, the confirmed project opens in Kiro chat, where building continues.

Tech Stack

Libraries and Tools Also Used

Other projects using the same technologySHOWROOMZDayce

Architecture

The Kiro extension starts a local Core automatically and talks to it over token-authenticated HTTP and SSE on 127.0.0.1. Core stores state in SQLite, runs four agents through the Kiro built-in agent or kiro-cli, and gives each agent role its own tools through MCP servers.

Problems & Solutions

1. Problem

   The per-user kiro-cli install was not found on a real Windows 11 PC

   Solution

   Added the per-user install path to the lookup and fixed 13 tests that failed only on Windows

2. Problem

   Generated projects stopped with a toolchain-changed error after every extension version bump

   Solution

   Changed the check to accept upgrades to a newer version of the same product

3. Problem

   In the finals version, Core refused panel Builder runs, so the start button ended in an error

   Solution

   Changed the start action to open the project in Kiro chat

4. Problem

   portable/node_modules was ignored by git, and line-ending conversion broke the release kit hashes

   Solution

   Tracked the files byte for byte with a .gitignore exception and .gitattributes settings

Outcomes

• Advanced to the finals of the Korea University × AWS AI Innovators Challenge
• Passed 807 automated tests on frontend version 0.0.18
• Confirmed app generation and launch, decision handling, and restore after restart with a real model on the Mac build (0.1.3)

Learnings

• This was my first extension. I learned to keep screen state in controllers, to fit a setup where the extension host and the webview only exchange messages.
• It was also my first time putting AI into a program's features. I learned that telling users what happened and what to do next when a model response fails or stops matters as much as building the feature.