◈OpenTree DashboardSign in

$ date -I # 2026-07-25

What agent-written pages look like, and what that implies for a host

A host for agent output should be designed around what agents actually emit, not around what a website is. The shape is consistent enough to design for: one file, no build, no assets directory, and a readership of one.

A note on numbers we do not have

The page this one is modelled on quotes aggregate statistics from a live customer fleet. This is a self-hosted deployment; it has no fleet to aggregate, and inventing percentages to match would be the exact failure this repository tests against. What follows is the shape the implementation is built for and why — check it against your own sites once you have some.

Single file, self-contained

Agent output is nearly always one document with its CSS and JavaScript inline, because that is the only form that renders anywhere without a build step. So the storage model is one object per site rather than a tree, and file_count is 1 unless a zip was uploaded. Nothing in the publish path resolves relative asset paths, because in the common case there are none.

Small enough that the transfer is not the problem

Inline charts and a few hundred rows of data still land well under a megabyte. That is why content is stored as an object and served directly rather than chunked, and why the preview card is drawn from the stored bytes rather than from a build artifact. Size limits exist as a deployment variable in wrangler.toml, not as a plan tier.

Four input formats, one path

InputWhat happens
HTMLStored and served byte for byte.
MarkdownRendered to a self-contained HTML document at publish time, so /raw, the title extraction and the preview card all read the same rendered copy.
SVGStored unchanged and served as image/svg+xml; the strict CSP is what makes that safe rather than rewriting the bytes.
ZipUnpacked into a multi-file site; this is the only case with more than one file.

The distribution is a link sent to a person, so the defaults are the opposite of a marketing site's: an unguessable token instead of a slug, noindex and noai on every viewer response, no-referrer so the token does not leak into a third party's logs, and an expiry that runs by default.

It contains real data more often than you would like

A page built from a live query embeds the query's results, and sometimes the credential that ran it. The publish path scans content before it is stored and blocks on a hit unless the caller overrides, which is the one place this deployment deliberately makes publishing fail rather than succeed quietly.

What the author gets back

Because the artifact is a deliverable rather than a page, the interesting measurement is whether it was read, not how many hits it got: opens, unique viewers, dwell time, scroll depth, read-to-end rate and a ten-segment attention map, at GET /sites/:idOrSlug/engagement. No session replay, no keystrokes, no IP addresses in storage — only a hash.

Can I publish a multi-file site?

Yes — upload a zip and it becomes a multi-file site with file_count greater than one. Anything that needs a server-side runtime is out of scope; this hosts documents.

What happens to markdown exactly?

It is converted to HTML when it is published, not when it is read. Raw HTML inside the markdown is escaped and link protocols are filtered, so a <script> written in markdown is visible rather than executed.

Is there a size limit?

There is a configured maximum in the deployment's variables. It is a resource guard, not a plan boundary — raise it in wrangler.toml and redeploy.

Do you keep the previous revision?

No. Replacing content overwrites the same stored object. If you need history, keep it where the artifact was generated.

Keep reading