agent-kit-testing

The agent kit proving itself on a real codebase.

agent-kit/ holds a process an AI agent runs: the contract for changing code, the queue that decides what gets built, the gate that decides what ships, and the voice it writes in. The codebase it governs is a SvelteKit site about the kit. The site renders the kit's own files, so the docs cannot drift from the process they describe.

The loop under test

1. Feedback arrives. A person opens an issue, or an email becomes one.
2. The roadmap decides. agent-kit/ROADMAP.md holds the queue. Each item carries one testable promise and the evidence that proves it.
3. The queue becomes tickets. roadmap-sync.yml mirrors each item into a labelled GitHub issue. The roadmap stays the source; issues are its surface, and the single pool an agent works from.
4. The agent builds. It takes the top ready ticket and runs the weekly cycle: spec, build, prove, document, ship.
5. The gate decides. npm run verify answers whether the repo is healthy. No green, no push.
6. The work is told. The devlog records what happened. The blog-post skill turns it into a post. This site publishes it.

Maintenance and marketing run on the same cycle.

Stack

SvelteKit 2 with Svelte 5 in runes mode, TypeScript, Vite 8, and @sveltejs/adapter-static. Requires Node 22 or later.

Every page prerenders, so the build is a folder of files with no server.

Commands

npm install
npm run dev       # dev server
npm run build     # static build into build/
npm run verify    # the gate: lint, check, build, test

CI runs the same npm run verify.

Hosting

.github/workflows/pages.yml builds the site and publishes it to GitHub Pages on every push to main.

The repository Settings must have Pages → Source set to GitHub Actions. Without that the workflow builds and then fails to deploy.

The build uses relative URLs, so the site works at the root or under a subdirectory. The workflow still passes BASE_PATH from actions/configure-pages, which is what a project site needs for its absolute references. To build the subdirectory case locally:

BASE_PATH=/agent-kit-testing npm run build

The roadmap sync

node scripts/sync-issues.ts --dry-run   # print the plan, write nothing
node scripts/sync-issues.ts             # apply it

Applying needs GITHUB_TOKEN and GITHUB_REPOSITORY. In CI this runs from .github/workflows/roadmap-sync.yml on a push to main that touches the roadmap. On a pull request it prints the plan and writes nothing.

Issues it generates are overwritten on the next run. To change one, change agent-kit/ROADMAP.md.

Testing

Vitest runs two projects. server runs unit tests in Node. client runs component tests in a real Chromium through Playwright.

The component tests need a Chromium. Install one with npx playwright install chromium, or point at one already on the machine:

CHROMIUM_PATH=/path/to/chromium npm run verify

Without a Chromium the gate fails rather than skipping. That contradicts agent-kit/VERIFICATION.md and is item 1 on the queue.

Process

Start at agent-kit/ROUTING.md. It says which file governs which situation.

The kit is partly tailored. The roadmap and use cases describe this repository, but {{PLACEHOLDER}} values remain in other kit files. Finishing that is item 4 on the queue.