Who picked a lead up and when, so nothing sits waiting unnoticed. No
outcomes, no end states, nothing on join_leads changes: the lead stays the
record of what the family said and the claim is only who has it. One row
per claim in lead_claims; the current claim is the latest unreleased.
Taking over is release then claim, never a silent overwrite: 409 names
the holder. The holder or an owner may release. Every claim and release
lands in the action log; the summary carries an unclaimed count.
tests/smoke_admin.py 112 -> 122.
Until now nobody could be invited without a script and a forgotten
password was a locked account. The identity layer already had invites,
sessions and the capability map; this wires the levers to them.
Admin API (people:invite_leader and up; the break-glass token cannot act
here, it has no person): list people with roles, memberships, live
sessions and keys, plus open invites and what YOU may grant; invite with
the URL returned once (no outbound mail yet, hand it over yourself);
revoke an invite; replace roles and memberships in full (owner only);
disable and enable (owner only, never yourself, never the last owner;
disabling ends sessions now and keys die through can()); mint a one-time
24-hour reset link (admin may reset anyone but an owner). Every one logs
who, whom and what.
Site: /account grows details and change-password forms (current password
required; the other devices are signed out, this one stays). /reset/{token}
sets a password, burns the link, ends every session and signs the person
in. Expired, used, revoked and never-existed read the same.
tests/smoke_identity.py 92 -> 123. Driven end to end on a throwaway with
a DB copy: invite, accept, account forms, reset link used then reused
(410), disable (session dies, login 401), enable, roles.
calendar_write.owns() now accepts both site namespaces: the 32 events the
2026-08-29 seed stamped site73-<sha1>@greenlanescouts73.org and the site's
own @site73.greenlanescouts73.org. The rule exists to keep the site off
@band.us and off anything a person adds in a calendar client, not off its
own data. Seeded objects are written back at the seeder's object name so
an edit replaces in place; proven on a throwaway against the real store
(edit, still 32 objects, restored byte-for-byte).
auth_events becomes the one action log. Every P2 write - nearby create,
update, deactivate; unit meeting edit; setting change; announcement post
and take-down - now records who, when, and a one-line before -> after for
the fields that changed, alongside the login, key and calendar entries
already there. GET /api/admin/history (history:read, admin and above, not
scopable on a key) reads it newest first with a kind prefix filter and
before= paging; rowid breaks second-resolution ties.
tests/smoke_admin.py 102 -> 112.
The site becomes a second writer to scouts/site73, as the scoped scoutsite
principal (rw on that one collection, denied everywhere else by the rights
file). Two rules enforced in calendar_write.py, not left to callers: the
site owns only UIDs ending @site73.greenlanescouts73.org and refuses any
other before a network call, the same shape as band-cal-sync and @band.us;
and with no RADICALE_* configuration every write is a 503, never a silent
no-op.
The VEVENT layout matches seed.py exactly so the feed reverses it into the
row shape the public pages already render: all-day DTEND exclusive, TZID +
VTIMEZONE on timed events, 90-minute default for a timed one-day event,
noon on the end date for a timed multi-day one, CATEGORIES for the unit,
X-SCOUT73-BADGE, 75-octet folding. Reads come from the scout-calendar feed
(now carrying uid and recurring); rows the site created and that are not
part of a series are marked mine. Writes are logged to auth_events and
bust the page cache so a leader sees their event within the feed's minute.
DELETE really deletes - the calendar's history is Radicale's git log.
Endpoints under calendar:write. tests/smoke_admin.py 78 -> 102. Proven
against the real store on a 2036 probe (outside the feed window): create,
read back byte-for-byte, replace, delete, second delete 404, foreign UID
403 with no store call, unconfigured 503.
NPM sets the header with $proxy_add_x_forwarded_for, which appends the
connecting address to whatever the client sent. The first hop is therefore
client-controlled and the last is the one NPM vouches for. Reading the
first hop would have let an outsider claim a LAN address with one header,
and admin_api.client_is_lan decides on it. Found while verifying the nginx
block removal; fixed before the block came out was relied on. Also
tightens login throttling, which used the same helper.
api_keys_from is a typed setting, lan (default, the historical rule) or
anywhere, checked in admin_api._auth against the first X-Forwarded-For hop
that NPM sets. It governs API keys only: a session is never restricted, the
console is a session from anywhere; and the break-glass token is ALWAYS
LAN-only, which is not a choice and so is not a setting. LAN ranges are
facts about the network and live in code; the docker range is included
because NPM, the routines and sibling containers reach the app from
arrstack_arr_net. An unparseable address is not LAN - the rule fails closed.
With this in place the location /api/admin/ block on NPM host 51 can come
out; the app enforces what it enforced, and the toggle never touches NPM.
tests/smoke_admin.py 58 -> 76. Proven on a throwaway site: key LAN 200,
key WAN 403 naming the setting, session WAN 200, token LAN 200, token WAN
403; set anywhere, key WAN 200 and token WAN still 403; bogus value 422.
A key is the person who minted it, narrowed to the scopes they chose. Only
the sha256 is stored; the full key is returned once. Scopes must be a subset
of the owner's capabilities at mint time and are enforced again at use time
inside identity.can(), the one place that decides, so a key never outlives
its owner's demotion and disabling a person disables their keys with no
separate flag. A key cannot carry apikeys:own or the owner powers, so it
cannot mint keys. Revoked rows stay; a foreign key id is 404, never 403.
Bearer keys are honoured ONLY on /api/admin. The rest of the site reads
sessions alone, so a scoped key never widens into a browser identity.
X-Admin-Token remains break glass and, having no person, cannot own a key.
/api/docs is generated from the router on every request: path, methods and
docstring from the route objects, and the capability read out of each
handler's own _auth() call so it cannot drift from the check. Gated on a new
api:docs capability (leader and above). GET /api/admin/whoami answers who the
API thinks you are and what you can do.
Tests: smoke_identity 66 -> 92, smoke_admin 53 -> 58 (registry has a
capability for every route, docs page renders every route). Driven end to
end on a throwaway site with a DB copy: mint, whoami via key, scoped 200s
and a 403 that names the narrowing, key-mints-key 403, garbage key 401,
admin token on /keys 403, key on /account is not a session, revoke then
401, second revoke 409.
MEETING_DAY, MEETING_DAYS, MEETING_DAY_ABBR, PACK_TIME, TROOP_TIME,
PACK_CLOCK and TROOP_CLOCK were the last thing a leader could not change
without a commit. identity.meeting_words() derives all seven from the units
rows; app.py refreshes the module globals from it on a 30-second TTL in a
middleware, and the page templates - f-strings that read those globals when a
route runs - are untouched. A row that cannot supply a word keeps the
default, so a half-filled unit degrades to today's copy rather than a blank.
The day words come from the pack row: the site's copy assumes both units
meet the same night, and if that changes the copy needs rewriting, not a
bigger constant. Proven on a throwaway container with a DB copy: patching
pack73 to Thursday 18:15 changed the homepage, cubs, troop and meta
description; patching back restored them. tests/smoke_identity.py 59 -> 66.
The console's 401 redirect carries next=/leaders/ and /login dropped it, so a
leader arriving cold landed on /account. next now rides the form as a hidden
field and is applied after a successful POST and when an already-signed-in
person hits /login. identity.safe_next() admits only a relative path with a
single leading slash - no scheme, no //, no backslash, no control characters -
so the login page cannot become an open redirect. Ten checks added to
tests/smoke_identity.py; the suite is 59 + 24 + 53, all green.
Nearby rows bump verified_at on every save and deactivate rather than delete.
Unit edits are meeting fields only, gated by unit:write_own scoped to the unit
in the URL. Settings are a typed key registry falling back to code defaults;
find-a-unit now reads its source name and URL from it. link_url and contact
reject attribute-breakout characters and the nearby renderer escapes quotes.
Covered by tests/smoke_admin.py, 53 checks in-process.
documents.visible() now takes the viewer's unit set. A members document is
served and listed only to a signed-in member of the matching unit; 'both'
reaches any member; owner and admin reach everything including unit types
added later. Everyone else gets 404, never 403.
The gate moved INSIDE find(), so there is one path from a slug to a file and
no route can forget to check. listed() and visible() stay separate functions.
admin_api takes a session first and falls back to X-Admin-Token as break
glass. Still fails closed: no session and no ADMIN_TOKEN is 503. Capability,
not role, decides per route. announcements.created_by now comes from the
session and ignores any value in the request body.
tests/smoke_documents.py, 24 checks, including the invariant that the index
can never list something serving would refuse.
Adds identity.py (schema, capability map, scrypt passwords, invites,
sessions, login throttle, boot seed) and auth.py (login, invite
acceptance, account page). app.py gains two imports and one wiring
block at EOF; no existing behaviour changes.
Units are a table seeded from the site constants. Roles split: leader
and member per unit in memberships, owner and admin site-wide in
people.global_role, so a unit added later cannot under-grant an admin.
tests/smoke_identity.py covers the rules that are invisible when wrong:
single-use invites, reissue revoking the prior link, expiry, idle and
absolute session bounds, throttling, and the capability split. 49 checks.
The hero button said 'Join us this fall' while the nav button two
inches above it said 'Join us'. That is the whole problem in one
screenshot: strings that go stale on a schedule, scattered.
Seven values at the top of app.py now hold the program year, the
meeting day, and the two meeting times. Twelve inline meeting-time
strings and four program-year labels read from them, so a time change
is one edit rather than a grep. The block carries a note that this is
not a CMS: anything with a date in it belongs in the calendar feed,
and evergreen copy stays inline where git log can answer what the
site said during recruitment.
Two strings were not stale, they were false out of season:
'There is nothing to do between now and Tuesday' on the thanks page
and 'See you Tuesday' on the join CTA both confidently invited
families to a meeting that did not exist through July. Now 'we will
be in touch with the next meeting date' and 'See you on a Tuesday'.
The properly correct fix reads the next meeting from the calendar
feed and this is not that, but it stops the site making a promise
it cannot keep in August.
The chocolate booth does not run all winter. Three places now say
in season.
'fall grade' becomes 'current grade' on the join label and in
SHEET_HEADERS. The header is only written to an empty sheet, so the
live sheet would have kept saying 'fall grade' forever; sheet_append
now repairs cell E1 in place on the next lead, the same way it
already backfills the Comments header.
thanks_body became an f-string. It had no bare braces, so the
conversion is safe.
Two faults on a phone, both found on the live site.
The single-line clamp truncated a venue change at 'the Deep Creek
pa...', cutting the only part of the notice that mattered. Collapsed
now clamps to two lines, three under 560px. A SINGLE notice does not
clamp at all - it has no expand affordance, so clamping it loses text
with no way to get it back.
The banner scrolled away while the nav stayed pinned. That was a
deliberate call and it was wrong: a cancellation is worth more
viewport than the nav is. Banner and nav now share one sticky wrapper
so they pin together.
Expanded is exempt via :has() - an open three-notice banner pinned to
a phone viewport is the thing the original call was trying to avoid,
and it is still right about that. Open scrolls, collapsed sticks. The
40vh cap stays as the floor where :has() is unsupported.
The one string on this site where a deploy is the wrong latency is
"tonight's meeting is cancelled, the lot is flooded" at 4pm on a
Tuesday. That is a record with a lifecycle, not site copy, so it gets
a table and a write path rather than a commit.
store.py
announcements table, created by the existing IF NOT EXISTS path so
there is no migration. ends_at is REQUIRED: an announcement that
never expires is site copy, and site copy belongs in the repo where
it has a diff. Nothing is hard-deleted; taking one down early sets
revoked_at, so what the site said and when survives.
Ranking is urgent first, then most recent. Recency alone would let a
routine Wednesday notice bury a Tuesday cancellation still live.
Two caps, enforced here rather than in the route so the future panel
inherits them: 200 characters, and 3 live at once. Both reject rather
than truncate. Clipping a cancellation mid-sentence is worse than
making someone shorten it, and a 4th live notice is a signal nobody
is expiring things rather than something to render.
app.py
announcement_bar() above the sticky nav. Native <details>, no
JavaScript, which matters on a read-only rootfs with no build step.
Collapsed clamps to one line with a count; expanded lists all of
them and caps at 40vh. One notice renders with no chevron and no
count: the common case must not look like a widget.
FAILS OPEN. The admin API fails closed because it serves family
phone numbers. This is the opposite case, and a broken announcement
must never take down the public homepage.
Colours are the existing note and rust token pairs from the brand
standard. No new colour enters the palette.
admin_api.py
GET/POST/DELETE on /api/admin/announcements. Leads stay read-only;
announcements are the deliberate exception, because being mutable
and expiring is the entire feature rather than a guess at a process.
list returns a computed state per row (live, scheduled, expired,
revoked, over_cap) so "why is my notice not showing" is answerable
from the API and not from the homepage.
The app is the only process on this box accepting unauthenticated input from
the internet and it was running as root in a writable container. Now uid 10001,
read_only with a tmpfs /tmp, cap_drop ALL and no-new-privileges. Host-side
/srv/scout-website/data and the service account key are chowned to 10001.
Verified on a throwaway container: every route, the admin API, spam rejection,
and a real submission writing to both the jsonl and SQLite.
Three bot submissions landed 2026-08-30 with both selects carrying the
placeholder label "Select one..." - required= is browser-only and a direct
POST skips it. Adds a hidden honeypot (answers 303 so the bot sees success)
and server-side checks on name, email, children, interested_in and source.
Rejections are logged with the client IP and never touch any store.
Internal papers need a durable link before member login exists. 'unlisted'
serves a document at its slug but keeps it off /documents and sends
X-Robots-Tag: noindex so it stays out of search results.
Serving and listing are now separate questions: visible() decides whether a
document can be served at all and stays the seam login attaches to, listed()
decides whether it shows on the index. 'members' remains hidden AND
unservable, so the weaker state cannot be mistaken for the gate.
This is obscurity, not access control, and both the README and the manifest
say so. An unlisted link is forwardable.
Files live outside the repo at /srv/scout-website-assets/docs (bind-mounted
read-only at /docs), the same arrangement as the photos, published by a
manifest.json beside them. Adding a document is a file drop plus a manifest
entry - the app re-reads the manifest on mtime change, so no rebuild and no
redeploy.
The manifest maps a stable slug to a filename, so next year's permission slip
can replace this year's without breaking a link already printed on a flyer.
Documents are served through /documents/{slug} rather than from a static
mount, and NOT placed under /app/static, which is public forever. That is the
point: when member login exists it plugs into documents.visible() and no
public URL moves. The visibility field already carries 'members', which today
is hidden from the index and 404s rather than 403s, since a 403 would
advertise a document we cannot yet gate.
Events now come from EVENTS_FEED_URL (the scout-calendar container,
which serves Radicale mike/site73 as JSON). The last good copy is
cached to /data/events-cache.json and served whenever the feed is down
or empty; 'Nothing on the books' remains the cold-start floor only.
The round-trip test proved the feed reproduces all 32 rows of
app/events.json byte-identically, so nothing visible changes today.
Removes app/events.json, the first-boot seed copy, and the local load
path; README calendar section rewritten; compose gains EVENTS_FEED_URL
from the Portainer stack env.
On a phone the /join grid collapses to one column, so the success
banner rendered below the entire pitch column and the 303 landed the
parent at the top of an apparently unchanged page. The empty form
underneath then read as a failed send.
?sent=1 now returns a dedicated confirmation body: banner first, what
happens next, and no form to resubmit.
Two real leads on 2026-08-26 landed in the DB with no Sheet row and no mirror
row at all: the empty-env window meant the Sheet call never ran. Because
/mirrors/failed looked for state='failed', a mirror that was never attempted
read exactly like one that was never owed, and the summary reported a clean
failed_mirrors: {} while two families were missing from the sheet.
/join now writes a `pending` mirror row for every target BEFORE attempting any
of them, so the gap between "owed" and "done" is a row rather than an absence.
set_mirror_pending never downgrades an attempted mirror. failed_mirror_records
covers failed and pending, since both need replaying, and summary reports
pending_mirrors alongside failed_mirrors.
sheet_append also stamped the Sheet with datetime.now(), so replaying a lead
filed it under the day of the replay rather than the day the family submitted.
It now takes the submission timestamp off the record and only falls back to now
when there isn't one.
records was a leads table wearing a generic name. It carried display_name /
email / phone, which only mean anything for a lead, plus status / assigned_to /
notes, which were a guess at an outreach process that has not been designed.
Contacting a family is one-to-many, so three columns on the lead row was always
the wrong shape for it.
join_leads is one row per /join submission with one column per form field, two
timestamps, and payload holding the submission verbatim. Outreach gets its own
table when the process is actually known.
Also splits heard_from from heard_from_detail. app.py folded source_place /
source_other into a composed "Other: ..." string and threw the raw answer away,
which is the half that tells you which daycare the lead came from. The composed
value is still built for the Sheet and the ntfy push.
admin API: /records -> /leads, and PATCH is gone since a lead now has nothing
mutable on it. mirrors and meta are unchanged.
app/__pycache__/app.cpython-313.pyc was committed and went stale the moment
app.py changed. It was inert because the image runs 3.12 and the bytecode was
3.13, but a stale .pyc that happened to match the interpreter would be a very
quiet way to run the wrong code.
Also ignores service_account.json, which is bind-mounted in at runtime and has
no business being committable.
Form submissions now land in /data/scout73.db before anything leaves the
box. The Google Sheet and the ntfy push become mirrors whose per-record
outcome is recorded, so a failed sheet write is replayable instead of
surviving only as a push telling you to retype it.
The table is a generic record store keyed by 'kind', with workflow columns
and an audit trail, so RSVPs and other forms can land in the same place
later without a migration. Admin panel talks to /api/admin over HTTP and
never opens the DB file.
- app/store.py schema, writes, reads, one-time leads.jsonl backfill
- app/admin_api.py token-gated API; fails closed when ADMIN_TOKEN is unset
- app/app.py three hunks: imports, boot init, join_post
- compose ADMIN_TOKEN passed through from the Portainer stack env
The five troop entries were placeholders written to get the site up. Troop
73's real calendar lives in BAND, so those are replaced by what BAND
actually holds:
- First Court of Honor, Klondike Winter Derby, Troop Spring Campout,
Founder's Day, River Canoe Trip
+ Canoe Day (9/1), Gettysburg Trip (11/21), Lock In (12/12-13),
Cabin Camping at French Creek (1/16-17)
Excluded from BAND: six events already past (the /calendar page has no
date filter), the weekly Tuesday meeting series (runs to May 2027, and
the schema has no recurrence field - it is covered in page copy), and
the 10/17 Cub Scout Campout, which is the same event as the existing
Fall Campout entry and that one carries a start time.
Pack and both entries are untouched; BAND has no pack events.
Pack runs 6:00-7:15, so the troop follows at 7:30 with a buffer. Updates
all eight troop time references across the footer, /troop and /calendar.
The two 7:00 AM references (Green Lane farmers market chocolate booth)
are deliberately unchanged.
Planned timing, not yet established practice.
thethreemagi
Mike Wichers