WEBSITE_WORKSPACE.md, plus the parts of the README that were still
describing the old layout or a manual step the code doesn't need
anymore (the key-ring paragraph, mainly - that one was actively
wrong now, not just stale).
continuous_integration.yaml: six blocking checks, then a separate
publish per environment so a Vite build never gets tagged for the
wrong one, then a call into last commit's deploy workflow. Along the
way: the lint list had drifted (two problems not in the requirement,
one already fixed), and the Umami-origin gate needed a variable pair
of its own since the backend's side of that comparison lives on the
host, not in CI. Pinned the two vulnerable packages while at it.
deploy-scp.yaml: download, back up the production database before
touching anything, upload into a timestamped release, symlink the
persistent website in, switch current and restart, verify /health
with retries, prune old releases only once that check passes. No
sudo, no container actions, one shared interface U5 will call next.
Infrastructure design for U6: single Pi split by directory not port,
systemd --user services instead of sudo, a releases/current/shared
layout that keeps the customer website outside every atomic switch,
and a health check through the existing reverse proxy before pruning
anything. Flags that user-level systemd needs enable-linger or the
service dies the moment the deploy SSH session closes.
Records the functional design for the two remaining application units,
before any of their code exists.
Security headers have to come from the application, because relying on
nginx or IIS configuration is exactly what this deployment model rules
out. Strict applies to /admin, /api/v1 and /health; a relaxed policy
applies to the public website, which this repository does not author.
The strict policy needs style-src 'unsafe-inline'. That is not a
shortcut: Radix positions dropdowns and dialogs with inline style
attributes recalculated per click and scroll position, and CSP nonces
apply only to style elements, never to style attributes. No nonce- or
hash-based variant leaves the admin UI working. The exception is bounded
to styles — script-src stays closed, which is where XSS actually lives.
The website's policy is enforcing rather than absent, so every
HTML-serving path carries a CSP and no exception has to be recorded. It
still blocks external script origins, so it remains a real boundary.
HSTS is skipped in development: browsers remember it per host and
localhost is shared with unrelated projects. Every other header applies
locally, so a CSP violation surfaces while developing.
For observability, browser error reports tunnel through the API rather
than going to Sentry directly. Ad blockers block Sentry domains, which
loses errors precisely for the users most likely to have browser
oddities. The tunnel forwards only to the host derived from the
configured DSN — a caller-supplied destination would turn an anonymous
endpoint into a request-forgery primitive.
Two consequences of the chosen options are recorded rather than left
implicit:
Enabling SendDefaultPii attaches request headers, and this application
carries two standing credentials in them. Besides the refreshToken
cookie, X-Master-Api-Key would have been sent to a third party on every
error raised during a master/slave call. The scrub list removes the
whole Cookie header, Authorization, X-Master-Api-Key and the request
body.
Console logging at Information plus structured logging to Sentry would,
taken literally, mean one Sentry event per request — exhausting the free
plan within hours and burying real errors in request noise. The
thresholds are split: console keeps Information, Sentry takes warnings
and above as events with Information as breadcrumbs, so every event
arrives carrying the trail that led to it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HHoJpxYXzHACSQguHrC5fw
Nothing here is visible in normal operation. Its whole purpose is that
swapping the release directory on deploy cannot silently destroy state.
Data Protection secures the API keys that authenticate master/slave
communication. Two separate defaults would each have destroyed them:
keys are held on the filesystem, which a release swap discards, and the
application discriminator is derived from the content root path, which
changes with every release directory — so even keys stored in a database
would have stopped being derivable. Keys now live in
ApplicationDbContext and the discriminator is a fixed constant.
Losing them produces no error. It produces stored keys that no longer
decrypt, which presents as an apparent network fault between a Master
and its slaves and is easily misdiagnosed. That is also why the tests
assert the resulting configuration rather than the registration: the
XmlRepository must be the EF one and the discriminator must be the
constant, plus a round-trip proving a value encrypted before a deploy is
readable after one. A test that only checked "Data Protection is
registered" would have passed in the broken case too.
Both modules previously called AddDataProtection() themselves. Module
registration runs after the host's, so those calls re-registered the
configuration chain and would have overridden the persistent store while
IDataProtector still resolved. They are removed, with a comment at each
site — the deletion otherwise looks like a regression. Each module's own
test project now guards against it being reintroduced.
ApplicationDbContext also migrates itself at startup. Deploy targets
offer no CLI, so migrations cannot be a manual step on the server.
Failures are classified rather than treated alike: a connection failure
means the database is not up yet, normal when the app and the database
start together after a reboot, and is retried with backoff; a migration
failure means something is broken and fails at once. Either way the
process does not start, which is what makes the liveness health check
trustworthy — an application that cannot reach its schema never answers
/health, so monitoring goes red instead of reporting a healthy instance
that cannot serve a request.
The cost of migrating without a human gate is that migrations must stay
forward-compatible and non-destructive, since rollback is "redeploy the
previous release". The new migration is purely additive.
Also wires this and the preceding hosting commit into both hosts, as
they touch the same lines of Program.cs.
Two constraints are enforced by documentation rather than code, and
belong in the deployment instructions: the key table must never be
pruned, and only one instance may migrate a given database at a time.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HHoJpxYXzHACSQguHrC5fw
Prepares the single-host layout for deployment. The customer's public
website moves from wwwroot/ to wwwroot/web/, so a CMS deploy can no
longer overwrite content it does not own: with the website in its own
directory, the release directory can be swapped without touching it.
Each front-end gets its own file provider, and both tolerate a missing
directory at startup — a fresh deployment has no website until a
separate workspace deploys one, and the CMS must still serve /admin and
the API. When the website's index.html is absent, an embedded
placeholder is served instead of a 404, which also doubles as proof the
CMS itself is running. The placeholder is embedded in the assembly
rather than shipped into wwwroot/web/, because that directory is owned
and overwritten by the website workspace.
Adds GET /health for uptime monitoring. It reports infrastructure
liveness only and is deliberately NOT the same thing as
/api/v1/Availability/status or /api/v1/System/capabilities: those are
CMS domain state that also serve the master/slave protocol. A healthy
instance can be switched off by design, and a switched-on instance can
be unhealthy, so conflating them would alert on business state and stay
silent on real outages. /health is on the availability gate's bypass
list for the same reason.
Fixes a real defect in the gate's admin bypass. It parsed the bearer
token with ReadJwtToken, which reads claims without verifying the
signature, so an unauthenticated caller could forge an unsigned token
carrying an Owner role claim and bypass the gate that suspends a
customer's site. Protected endpoints still rejected them, so nothing
leaked — but the gate itself was bypassable. The token is now fully
validated against the same parameters as the bearer scheme, resolved
from one shared source so the two cannot drift apart.
Host wiring for these changes lands with the data-durability commit,
since both units touch the same lines of Program.cs.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HHoJpxYXzHACSQguHrC5fw
Sluijsens
