opencode.ai site

Read the design-system rules in DESIGN.md (also shown at /dev/components#rules) before you change UI or copy.

Public pages

When you add, remove or rename a public page, keep /sitemap.xml, llms.txt, search and the page's .md version in sync; run the sitemap test (bun test test/sitemap.test.ts from this directory).
/sitemap.xml is the only sitemap in robots.txt. It is an index of sitemapFiles in src/lib/sitemap.ts: this app's /sitemap-pages.xml, the V2 docs sitemap /docs/sitemap.xml, and the Data and Console sitemaps.
/sitemap-pages.xml (src/routes/sitemap-pages.xml.ts) lists every page route automatically. To leave a page out, add it to sitemapExclusions in src/lib/sitemap.ts with a reason. A route with parameters needs an entry in dynamicRoutes there, or an exclusion.
Do not add a static public/sitemap.xml, because it would hide the generated one.
The view-source comment's FOR HUMANS rows start with the main nav in the same order; they're generated from the nav data, so update the nav, not the comment.

Do not use colons in site copy (headings, leads, body text, buttons, captions, alt text). Rewrite the sentence instead, for example "The easiest way to start is to download OpenCode", not "The easiest way to start: download OpenCode". Colons in code, commands, URLs, times and data labels (such as "HEX") are fine.
Write statements, not rhetorical questions. Use "For more, bring your own subscription", not "Want more? Bring your own subscription". Real questions are fine where the reader is asking them, such as FAQ entries.

A button that opens a menu, popover, dropdown, sheet or dialog sets aria-expanded="true" while it is open and is styled from that attribute, not hover: filled and split buttons fill with the link blue (like the product cards' split chevron), neutral buttons keep their pressed grey, and text or nav triggers take the current-item colour with the chevron flipped; Escape closes and returns focus to the trigger.

Every page and the docs report to the Console's funnel through src/lib/funnel.ts: entry-client.tsx and src/docs/client.ts start it, and it loads the Console's tracker (/console/funnel/v1.js) with src/lib/funnel-config.ts once the visitor first presses a key or the pointer, or the page is idle. Keep the stub free of imports, and imported only by entry-client.tsx on the site, so it stays in the entry chunk and the config and the tracker stay out of every page's first JavaScript.
The tracker reads the markup: data-section names a part of a page (section.viewed, and every click's section), data-funnel names a control, data-funnel-copy (with data-funnel-method) marks what a copy button or a selection inside copies, data-funnel-surface names a page's surface (the 404 page), and data-funnel="ignore" leaves a control out. Give a new page a surface in surfaceOf (test/funnel.test.ts fails until you do) and its main parts a data-section.
data-section names, data-funnel slugs and surfaces are analytics keys: rename one only together with the reports that read it.
Never send what a visitor typed. Code that reports its own events (search, the Markdown shortcut) pushes them onto window.__ocFunnel (Funnel.track in the docs), with counts and lengths, never text.

This service is the website and the V2 docs, deployed as its own Worker with cf (cloudflare.config.ts, script/build.ts). Never add Wrangler or SST here.
The V2 docs at /docs are a self-contained feature in src/docs/ (read src/docs/AGENTS.md). They prerender their pages with the site's components and build the site's stylesheets into their own files under /docs/_build, so a docs page never loads the site's built files and a site page never loads the docs'. A /docs URL without a V2 page redirects to the same path on the V1 docs (/v1/docs, then docs.opencode.ai).
Old URLs kept for outside links live in src/lib/legacy.ts, each commented with why it exists and who still uses it; nothing may link to them. Short links (/discord, /feishu, /desktop-feedback) are ordinary routes.
The site also serves fixed-URL machine files: /config.json and /openapi.json (written by bun run generate), /tui.json (public/), the V2 install script /v2/install and config schema /v2/cli.json (both stay at /v2), and /oauth/opencode/client.json and /.well-known/apple-app-site-association (src/lib/meta.ts). Their bytes and headers are a contract with clients.
Heavy media (video, audio, zips, 3D models) is never a Worker asset: link it with mediaUrl() (src/lib/media.ts), which points into the files bucket behind /files.
The Worker holds no GitHub token (see README.md).