This Week 3 build is SoloAgent, a desktop AI harness experiment for AI-assisted coding workflows.

There is no live site yet, only the repository.
What I built
- Open source repo: nearbycoder/soloagent
- New article: Building SoloAgent to understand AI harnesses
- New project page: SoloAgent
Repo pull-down tool audit
I pulled the repo and inspected the current manifests/docs to verify the tooling.
- App/runtime: Electron 39, React 19, TypeScript 5, Node.js 22+, SQLite
- AI runtime:
@tanstack/ai,@tanstack/ai-react,@tanstack/ai-openai,@tanstack/ai-client - Terminal/diff:
node-pty,ghostty-web,@pierre/diffs - UI/state: Tailwind CSS 4, Radix UI,
lucide-react,motion,zustand - Quality/build:
electron-vite, Vite,electron-builder, ESLint, Prettier, Vitest - External CLIs: Git, GitHub CLI (
gh), Codex CLI
Why this mattered
This was a practical way to map where AI harness complexity actually sits: state boundaries, tool-call UX, terminal process management, and git operations that hold up under real project workflows.