Repository navigation
Releases
v1.9.0 — Groups on the People page, and moving a document
worker/schema.sql is unchanged. No setting changes, and the compiler's
output is unchanged (git diff v1.8.1 v1.9.0 -- packages/compile is empty), so
a host that pins only the compiler needs nothing from this release.
Groups. /admin/people/ called a person's audiences "Answers as" and took
them as free text. A typo (owners) was stored without complaint and matched no
question, so that person's Home read "0 waiting for you" and nothing said why.
The column and the form now say Groups and offer one checkbox per group
site.config.yml defines (its audiences, with each one's describes as the
hover text). The API refuses a name the site does not define and names the real
ones; a group the person already has still passes, so a renamed one can be
removed. If the manifest cannot be read the page falls back to the old text box
rather than locking the roster. Storage, the manifest and the MCP keep the word
audiences; new log entries read "groups changed".
The deployed-with account can be edited. It is listed without a row in the
table, so every change to it — its groups included — answered 404 "no such
person", and the owner could not be put in a group at all. Its first edit now
gives it a row (logged as "added"); it stays an admin whatever the row says.
POST /api/content/batch (rows sites). Several edits as one version, and
moves: { note, changes: [{path, base_revision, body?, status?}], moves: [{from, to, base_revision, body?}] }. A move could not be done before: the copy at the
new path and the original carry the same question slugs, so neither save
compiled. Every path is checked against the revision the editor opened and one
stale path refuses the batch; one compile, one D1 batch, all or nothing. A move
keeps the status and the question numbers (the lock keys them by slug, and
question ids are project:slug), re-points the document's comments, and removes
the old row with its revisions kept and a last revision noting where it went.
Tests: 426 → 436 (worker/test/admin.test.js, worker/test/rows.test.js);
each new guard fails when its bug is put back. Checked under wrangler dev --local on a band+rows copy: groups listed, a typo refused, the root put in a
group, a document moved (new page 200, old 404) and a stale batch writing nothing.
v1.8.0 — A private site's /llms.txt names the way in
worker/schema.sql is unchanged. No setting changes. Every private
instance's /llms.txt is rewritten by the next build; a public instance's is
untouched.
- The private notice said "the repository is the entry point and this URL is
not". On a private instance the repository is usually private too, so an
assistant was sent from a door it could not open to a door it was told not to
try — while/.well-known/oauth-authorization-serverand
/.well-known/oauth-protected-resourcehave been served anonymously since
v1.2.0, which is exactly what claude.ai reads to offer the site as a custom
connector. The notice now says this URL IS the way in and names/connect
and/mcp. - It also carries
project.taglineand says what kind of thing the site is —
documents with their open questions in the margin, numbers that never change,
figures computed rather than typed — so an assistant with no access can still
say what it is looking at instead of only "I need permission". And it says
plainly that nothing in the file describes the project's state, so none is
inferred from it. agentIndexRotgains the matching check: on a private instance the notice
must name/connectand/mcp. Three self-tests, and the existing one that
asserted a bare private notice was fine is now stricter, because it was not.
private instance it is the only thing a stranger is served, and the gate
asserts both. The project's name and tagline are named public fields.
Tests: 422 → 424, in both the default and the band+rows variant.
v1.7.0 — The page keeps room for the floating Ask button
worker/schema.sql is unchanged. No setting changes. Every page that
carries the assistant gets about 4.8rem of room at its foot; a site with
chat.enabled: false is untouched.
- The Ask button is
position: fixed, so the page could end underneath it. On
a phone it did: at the foot of a document the pill sat on the last contents
link and, because a fixed element takes the taps as well as the space, that
link could not be followed at all — with nothing about the page looking
wrong.body:has(.chat-open)now reserves the dock's height plus its gap,
so the room is on the page and the button stays where a thumb expects it. - The band layout had its own version of this fix — 5rem of bottom padding on
.version-foot— which would have stacked with the new one on a phone. It is
removed: one reservation, in the widget the dock belongs to, in both layouts
and at every width. chatDockRoom(test/detectors.js) checks the arithmetic rather than the
spelling: it readsmin-heightandbottomoff.chat-openand
padding-bottomoff the reservation, and fails when the room no longer
clears the button. The two halves live in different rules and neither reads
the other, so restyling the pill taller or trimming the padding would
otherwise put the overlap back with every test green. This is also why the
button now states its ownmin-heightinstead of leaving its height to the
padding and the font.
Tests: 417 → 422, in both the default and the band+rows variant.
v1.6.0 — A passthrough page cannot claim an address the site serves
worker/schema.sql is unchanged. No setting changes; nothing an instance
does today behaves differently.
- The gate refuses a
site.passthroughentry whose served address is already a
document's or one of this template's own pages (passthroughCollisionsin
test/detectors.js). Such an entry was two pages at one URL with no error
anywhere: in the classic layout the copied file lands on top of the compiled
page, in the band layout the Worker renders the document and never reaches
the file. The check readssite.config.ymlagainst the manifest, not the
built tree, because whichever page loses the collision is also the evidence
that there was one. It catches a directoryto(which keeps the source
file's name) and two entries at one address as well. - The list of pages this template builds is now one const in
test/gate.test.js
(TEMPLATE_PAGES), read by the orphan check and the new one. - TEMPLATE.md gains "What no test catches": binding probes (
env.ASSETS?.xis a
truthy RPC stub in workerd and 1101 on every page),project.*settings being
compiled into a version rather than read at render time, and this passthrough
trap. All three have broken a live site while the Node suite stayed green.
Tests: 411 → 417, in both the default and the band+rows variant.
v1.5.0 — The header lists the documents, in dropdown menus
worker/schema.sql is unchanged. Band layout only; off unless set.
project.header_menus: true: the header is the one place documents are
listed. Home, theproject.pillsdocuments as pills, then a dropdown per
document group; any listed document in no group goes in "More", so none
becomes unreachable. The rail drops its documents list and keeps "where things
stand", contents and comments. The menu holding the page you are on is marked.
Dropdowns are<details>: they work with no script; the widget only keeps one
open and closes them on Escape or a click elsewhere. On a phone a menu opens in
place, full width.project.doc_groupsentries may now name their documents exactly,
{label, docs: [slug, ...]}, in the order to show them, as well as by
{label, prefix}.
The variant build turns header menus on, so the whole gate checks pages that
carry them. Tests: 406 → 411.
v1.4.0 — Choose the header's pills; group the documents
worker/schema.sql is unchanged. Band layout only; both settings are
optional and default to what v1.3.1 did.
project.pills: the documents the header shows as pills, a list of slugs in
order. Unset, every listed document is a pill (as before), which stops
working once a site has more than a handful. A slug the viewer may not see,
or that is not a listed document, is skipped.project.doc_groups: a list of{prefix, label}, in order. The rail and
Home group the documents under those labels by slug prefix (runbooks/,
briefs/, ...); the rest come last, under "Other documents".
Also a gate fix: in the band layout the gate named a self-contained page served
as-is (site.passthrough, e.g. design/x.html) design/x.htmlindex.html, so the
gate's exemption for such pages never matched and their own classes failed "every
class used is styled". The variant build now carries one (red before the fix).
Tests: 402 → 406.
v1.3.1 — Band layout fixes from the first real site
worker/schema.sql is unchanged. Band layout only; classic is untouched.
From a design review of ctfg-collab at desktop, tablet and phone widths:
- Narrow screens: the Tools box is one wrapped row above the content
whateverproject.toolssays. Opened as a column it pushed every page's h1
down by a screen. The "Hide contents" switch is gone below desktop width,
where the rail is under the content and the switch covered the Tools row. - Admin pages have icons, so a closed (icons-only) Tools box still reaches
them; they used to vanish. - "Where things stand" counts the documents the site lists, as Home's card
does. It counted every compiled file (About, partials, posts, decisions), so
one page said both "1 document" and "6 documents". - Activity titles each post and decision once, as a linked h2; the item's
own h1 was a second, giant h1 inside a card. - Home has no breadcrumb: it was one crumb repeating the name beneath it.
- The account corner is one row; blockquotes are styled; "Why asked:" has its
space.
Tests: 400 → 402 (the crumb rule, the Activity heading rule, the admin icons).
v1.3.0 — The content export can be pulled by its target
worker/schema.sql is unchanged.
content.export.mode (push, the default and v1.1.0's behaviour, or pull).
In pull mode the target repository runs the nightly export itself through the
template's reusable workflow (.github/workflows/content-export-pull.yml),
with its own GITHUB_TOKEN: no deploy key, and no credential in any other
repository that can write to the target. The instance's own export workflow
does nothing. The site refuses to be exported into any repository but the one
its settings name. TEMPLATE.md has the target's workflow file.
Use it when the target's base branch cannot be protected (a private repository
on GitHub's free plan) or when a push to it deploys something.
Also: an unreachable site is reported in a sentence, not a stack trace.
Includes v1.2.1.
v1.2.1 — Fix: signed-in band pages threw in the Workers runtime
worker/schema.sql is unchanged. Take this if you run site.layout: band.
In v1.1.0 to v1.2.0, every page a SIGNED-IN person opened on a band-layout site
threw (Cloudflare error 1101), and MCP and the assistant read the unfiltered
binding. The code marked its own wrapper of env.ASSETS with a property and
probed the binding for it (env.ASSETS?.real). In the Workers runtime a binding
is an RPC stub, so asking it for any property it does not have returns a truthy
stub that throws when used. The tests passed because their ASSETS was a plain
object; a stranger never reached the code because a private site redirects them
first.
Fixed: the wrapper is recorded in a WeakMap and the binding is asked for
nothing but fetch. The band and rows tests now use a runtime-shaped binding
(worker/test/fixture.js rpcBinding): with the old code 39 of them go red.
Checked under wrangler dev (workerd) on the path that failed in production.
v1.2.0 — Sign in with the CMS
worker/schema.sql changed: one table added, auth_tickets_used
(CREATE TABLE IF NOT EXISTS). Apply it before turning the provider on; until
then a CMS sign-in is refused, safely, because the ticket cannot be burned.
A new sign-in provider for a CMS on another domain that hands identity across as
a signed, two-minute ticket (worker/src/cms.js; the pattern of
Civic-Tech-Field-Guide/ctfg-website's tenant pass). Like every provider it
proves an email and nothing else; the roster in D1 decides. It is on only when
both are set:
CMS_PASS_SECRET(secret, 32+ characters, the same value the CMS holds)CMS_URL(var,https://only), and optionallyCMS_LABELfor the button
A ticket is refused if it is for another host, from another sign-in attempt
(state, so nobody can sign you in as them), expired, claiming more than five
minutes, of another kind, or already used. The CMS must offer
GET /api/collab-pass/start?host=&state= and only mint for hosts it lists.
worker/test/cms.test.js holds a test vector the CMS's own test asserts too.
Nothing changes for an instance that does not set the two values.