agentwright.app

Marketing site for Agentwright. Static HTML, deployed on Cloudflare Pages.

• index.html — landing page. It is a template: npm run build:blog fills its {{TOKENS}} (version, release date, line and test counts, model list, changelog, code excerpt, screenshots, product JSON-LD) from content/kit/ and writes the result to dist/index.html. An unknown token fails the build. Open the built page, not the template.

Kit preview (/inside/, /docs/, /changelog/)

content/kit/ holds the public preview of Agentwright Starter. It is generated, so don't edit it by hand. After every Starter release, run it against a git checkout of the Starter, then commit and deploy:

node scripts/sync-kit-preview.mjs /path/to/starter

This repository is public. Only the source files named in PREVIEW_FILES (lib/agent/loop.ts, lib/billing/metering.ts, evals/cases.ts) are copied in full, plus the buyer docs in DOCS, CHANGELOG.md, ci.yml and .env.example. Every other file appears by name and line count only (meta.json, with one-line notes from the Starter README's Layout block in layout.txt).

scripts/lib/kit-pages.mjs builds from it:

• /inside/: the file map with notes, all files with line counts, the three preview files in full, tests and CI
• /docs/ and /docs/setup/, /docs/deploy/, /docs/price-snapshots/, /docs/license-delivery/
• /changelog/: the Starter's changelog
• the landing page tokens (renderLanding, see index.html above)

public/shots/shots.json is optional: { version, items[] } with product screenshots or a video in public/shots/. When it exists, the landing page gets a screenshots section; version is the kit version that was photographed.

content/kit-runs/ is optional: runs.json ({ version, date }) plus test.txt and eval.txt, the real terminal output of npm test and npm run eval -- --model claude-haiku-5-5. /inside/ shows it only while version equals the synced kit version, so run both commands again after every Starter release.

content/compare/*.json builds one page per file at /<slug>/ (scripts/lib/compare-page.mjs). Every fact about another product carries its source URL, and checked is the date shown on the page and used as the sitemap lastmod. Re-check the sources every month and update checked.

Visit and click counts

The build injects a small beacon into every page. It posts the path, the referrer and any ?utm_source= to /api/hit (functions/api/hit.ts), which keeps only daily counters in the STATS KV namespace: page views, unique visitors, referrers, utm sources, and two click events per page, buy_click (a link to buy.polar.sh) and preview_click (a link whose href starts with /inside/). Use exactly those hrefs for buy and preview links, or the clicks are not counted. ?notrack=1 turns counting off in your own browser (?notrack=0 turns it back on). node scripts/stats.mjs [since YYYY-MM-DD] prints the totals (needs CLOUDFLARE_API_TOKEN and CLOUDFLARE_ACCOUNT_ID).

Launch-list signup

functions/api/subscribe.ts is a Cloudflare Pages Function behind the form on the landing page. It stores signups in the KV namespace bound as WAITLIST (Pages project → Settings → Bindings), with a honeypot, a 2-second minimum fill time and a per-IP limit of 5/hour. GET /api/subscribe reports whether the binding is present.

npm install
npm run typecheck      # type-checks functions/
npm run dev            # local preview with an in-memory WAITLIST KV
npm run deploy         # build:blog + wrangler pages deploy dist (needs CLOUDFLARE_API_TOKEN + CLOUDFLARE_ACCOUNT_ID)

Blog

Posts are Markdown files in content/posts/*.md (frontmatter: title, description, date, slug, tags). npm run build:blog writes the deployable site to dist/ (gitignored): index.html, blog/<slug>/index.html, blog/index.html, feed.xml, sitemap.xml, robots.txt. npm run deploy runs the build and uploads dist/ (wrangler picks up functions/ from the repo root). Drafts/outlines live in content/drafts/ and are not built. BLOG_OUT=/tmp/some-dir npm run build:blog builds into another folder instead of dist/ (for test builds).

Call to action per post

content/post-cta.json maps a post slug to { text, files[] }: the kit box under that post. File paths in files are shown as code. Every sentence must be true for the current kit. A slug that matches no post fails the build; a post without an entry gets a neutral default text.

Scheduled posts

A post goes live on its frontmatter date (Korea time) without a new deploy. The build renders every non-draft post, plus dist/_schedule.json and one snapshot of /blog/, feed.xml and sitemap.xml per future publish date under dist/_asof/<date>/. functions/_middleware.ts then returns 404 for posts dated after today and serves the snapshot for today. So queue posts by committing them with a future date and running npm run deploy once. draft: true posts are never built. BLOG_PREVIEW=1 npm run build:blog treats every post as live (local preview only, never deploy with it).

dev.to cross-posts

content/devto-schedule.json lists {slug, date, tags}. node scripts/devto-crosspost.mjs (run by a daily scheduled job, needs DEVTO_API_KEY) publishes at most one due post per run with canonical_url set to our post, skipping anything already on dev.to. Text between <!-- site-only --> and <!-- /site-only --> is removed for dev.to (no self-promotion there), and the AI disclosure + "questions go to the original post's comments" closing is appended. --dry-run writes the converted bodies to dist-devto/ for review.