A tab opened before a deploy asks for the old hashed Landing CSS or JS
file, gets a 404, and the navigation dies with "Unable to preload CSS".
router.onError now loads the destination again, and vite:preloadError
reloads for imports outside the router. A sessionStorage stamp limits
this to one reload per 60 seconds so a chunk missing for another reason
still surfaces as an error.
https://bugs.extant2000.com/caelum/issues/160
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Register, Login, Order and the shop config fetch caught errors and only
showed them on screen. A new report-failure helper sends network, 5xx and
unexpected failures at level error with flow, status and error code tags.
Expected 4xx user errors are not sent. The caught message is never sent.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The cherry-picked copy still claimed more than ships:
- "Frames land as the shutter closes": POST exposures in
server/src/routes/sessions.js:256 waits on a timer, then writes an
archive_frames row with empty headers and no file.
- "weather state": server/src/index.js:185 starts weather-sim, a random
walk, for every site in an active session.
- "can never be double-booked": bookings.js:165 is a check before the
insert, with no database constraint. The copy now says an overlap with
a confirmed booking is refused.
- Frames "arrive with the observation attached": observation_id is
optional in archive.js:235, so only daemon uploads link one.
- Footer and sign-in line said "shared observatory network" and "your
observatory network".
Tests: landing-claims.test.js refuses these phrases, requires the
simulated line on both copies, and fails if sessions.js or index.js stop
simulating. Two of the new checks failed before this copy edit.
Ledger: fmvslaunt7.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The <title>, og and twitter tags still read "Caelum - Federated
Observatory Network", and the JSON-LD called it "a federated observatory
network". All now read:
Caelum - MIT-licensed observatory network software
MIT-licensed software for booking telescope time, running live sessions
and archiving frames, with the Nox typed DSL for telescope control. One
observatory runs it today. No driver has yet run against real equipment.
The landing route's meta in client/src/routes.js changes too, because
router.js writes it over the title and description on every navigation.
The og:image:alt drops its em dash.
Sources: LICENSE line 1; the owner's statement of 2026-10-29 for one
observatory; docs/verifying-accuracy.md, "Not verified yet", for hardware.
Two new tests in client/test/landing-claims.test.js read every head tag and
the route meta, refuse federat, open, the retracted phrases and dashes, and
require one title and one description across all of them. Both fail on
the old head.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The landing claimed "A shared network of real telescopes", "Open network",
"Real hardware" and sites "from backyard rooftops to mountain summits".
Caelum runs at one observatory, the developer's, and no driver in it has
run against real equipment. Both the Vue view and the crawler skeleton in
index.html now say so.
Sources per line:
- MIT-licensed: LICENSE line 1.
- One observatory runs it today: owner's statement, 2026-10-29.
- No driver has run against real equipment: docs/verifying-accuracy.md,
"Not verified yet", Hardware.
- Sites list telescopes, set visibility, approve bookings:
server/src/routes/network.js (discover, booking-requests).
- Alpaca and Stellarium drivers: server/src/services/hardware-driver.js.
The "Live on the network" heading and its pulsing dot go too: nothing on
the page is live. client/test/landing-claims.test.js fails on the old copy
(2 of 4 tests) and checks the hardware line against verifying-accuracy.md.
Ledger: fmvslaunt7 (The Studio).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Finding fmw149sl04. Real mode checked only sign-in and confirmed: true, then resolved any telescope or dome by name, so a signed-in account could drive another site's hardware.
The driver now takes the calling user and checks every op: canControlTelescope for a telescope, canEditSite for a dome. The route refuses with 403 before the first op when a string literal names a resource the caller has no rights on. Names repeat across sites, so a name resolves among the caller's rows, and two matches is an error.
canControl moved from routes/control.js to perms.js as canControlTelescope. It now passes an admin with no booking live, as the control.js header said it did.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Security blocker on PR 45, finding fmw10xsce27. GET /api/jobs handed
nox_program_source to a daemon with no Sun check, while the sequencer
dispatcher already filtered its bridge handout. The route now filters on the
row's target and the telescope's live solar_filter, using
services/sun-avoidance.js.
target_ra on observations is in degrees. The new checkSunSeparationRaDeg keeps
a null RA untestable, because null / 15 is 0 in JavaScript and would test RA 0h.
The Sun is tested at start_time when that is in the future, since a daemon holds
a job until then. A held job stays PENDING.
The check reads the target columns only. The program text is not parsed, so a
program that slews elsewhere is not tested.
observatory-daemon.test.js declares a filter on its telescope. M31 falls inside
the limit each spring, so that suite would otherwise fail on those dates.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Finding fmw10xsce27. No slew path tested the Sun. services/sun-avoidance.js
is now the one check, called by slewTo, the Nox Alpaca alpacaSlew, the
hardware and Astra routes, and the sequencer dispatcher, whose bridge slews
out of the server's sight. telescopes.solar_filter (default 0) lifts the
limit; a site editor sets it on PUT /api/hardware/telescopes/:id/driver.
The 46 degrees is SunSeparationConstraint(46 * u.deg) in the UVEX definition
in m4opt 4.0.0, src/m4opt/missions/_uvex/__init__.py:69. That is a space
telescope keep-out value, not a ground eye-safety standard. Foukal, Solar
Astrophysics (2004), ch. 1, p. 13, gives the warning and no angle.
The Nox driver passed RA in degrees to drivers that take hours, so a real
mode slew pointed elsewhere and the check would have tested the wrong place.
It now divides by 15.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 2026-10-21 Node.js security release (critical RCE, fixed in 22.23.4)
reached production only after a rebuild with a fresh base image: the
mac1 and mac2 runners still held node:22 from 2+ weeks before. docker
build now uses --pull, and workflow_dispatch allows a rebuild without a
code change. Merging this rebuilds and redeploys on Node 22.23.4.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On 10-19 03:25 to 04:25 UTC the UPDATE api_tokens SET last_used_at in
workerAuth hit ER_LOCK_WAIT_TIMEOUT 9 times while the DB was stalled. The
rejection escaped the async middleware, so a valid sidecar request got no
response. The stamp is bookkeeping, so it is now logged and skipped.
GlitchTip CAELUM-4, CAELUM-5.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`/api/healthz` and `scripts/build-sitemap.mjs` were both open items in
backlog/READY.md with a branch each, cut by the daily sweep in May and June and
never merged. Seven of those branches carried three features between them, each
cut again on a later day; the duplicates are deleted and the newest of each pair
is here.
healthz goes in as written: `GET /api/healthz` answers `{ ok, db, version }`,
200 with the pool up, 503 with it down, 200 and `db: "disabled"` with no pool.
Its three tests pass. The route is beside `/api/health` rather than replacing
it: that one reports history and runs an external probe, and a liveness check
has to answer when those cannot.
The sitemap generator was broken, and silently. It read `client/src/router.js`
for `meta: { guest: true }`, and the route table has since moved to
`client/src/routes.js`, so it matched nothing and wrote an empty sitemap over
the committed seven-URL one. Running it was the diagnosis.
Three changes to it:
- it reads routes.js for the table and router.js for the ALWAYS_PUBLIC name
set, and treats public as the union. Nox, the shop, the status page and the
legal pages are reachable without a sign-in through that set alone and carry
no `guest: true`, so `guest` on its own found three routes out of thirteen.
- `checkout` is excluded by name. It is public so the payment flow works
without a sign-in, which is not a reason to send a crawler to it.
- it refuses to write fewer than six URLs and exits 1. The route table moved
once; the next move should fail loudly.
The suite has a test for that last case, and one asserting the committed
sitemap is not empty. 13 URLs now, against 7 committed by hand.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Closes#39. The sqlite boot smoke test passed a 21-character JWT_SECRET
and the server refuses anything under 32 at boot (main run 214). Each run
now generates one with openssl rand -hex 32.
The image job also runs on pull requests (build, hadolint, smoke test)
so a change like this is proven before merge; only main pushes and
deploys. Registry login uses the org REGISTRY_TOKEN: the Gitea job token
is refused by the registry (seen on hittkitchen run 226).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The mac1/mac2 build machines cannot reach port 80 on the Debian mirror
CDN, so apt-get update failed and the reduction image never built.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Translate .gitlab-ci.yml into .gitea/workflows/ci.yml with the same
coverage (client, lint, nox, server mysql + sqlite advisory, sidecar,
image build, hadolint, sqlite boot smoke test). Images go to the Gitea
registry as caelum and caelum-reduce. Sourcemaps upload straight to
https://bugs.extant2000.com. Deploy copies the compose and runs
deploy/gitea-deploy.sh on the box over SSH with a throwaway registry
login, waits for healthy, and checks gitSha plus the .map exposure on
the public edge. Repository links now point at git.extant2000.com.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Dockerfile: Vite's maps move from client/dist/assets to /app/sourcemaps,
outside the express.static root. caelum.network/assets/*.js.map answered
200 (889 KB for the index bundle) before this.
- New main-only job sourcemaps-upload (stage deploy, needs image-build,
allow_failure). It copies bundles and maps out of the just-built image
and runs getsentry/sentry-cli:2.58.6 on the `web` network against
http://glitchtip:8000: releases new, then sourcemaps upload for release
caelum@<client version> with url prefix ~/assets. Token comes from
GLITCHTIP_AUTH_TOKEN and exists only in this job, so the vite plugin
still never uploads.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Telescope position loop: failures report per telescope through
services/report-throttle.js, one warning an hour with the Error and a
telescope tag, a log-stream line a minute in between, a recovery line.
- notifications: unreadable mute preferences log a warning with the Error
and still deliver. jobs: last_connected_at failures warn. control: a
slew note that fails to save warns.
- initDb retries are info; attempt 5 and every 10th after warn with the
Error. Column migration failures are error. public-feed, health-history,
spacedb index, db-upgrade, migrate, seed-shop, limits, telescope settle
and research fetches pass the Error.
- Research.vue: Aladin script load failures are info (third party), init
errors warn through utils/log.js; a failed load no longer sticks.
Client log.warn with an Error carries the stack as context.
- Tests: webhook ledger fault answers 500, Stripe's retry is processed
once, a duplicate still gets 200 (fails against the pre-fix handler).
Throttle test: two hours offline = 2 warnings, 118 info lines.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Server (@sentry/node 10.53.1):
- log.js sends every level to the structured log stream (enableLogs).
debug/info stay out of issues; warn opens a warning issue, error/fatal an
error/fatal issue. An Error at warn keeps the log line as the title and
carries type, code and stack as context. log.with({ tags, extra }).
- sentry.js: http integration reads no request bodies; beforeSend keeps an
allowlist of headers and drops cookies, query strings, bodies and IPs.
- Express error handler: 5xx becomes an issue tagged with method, route
pattern and status; 4xx goes to the log stream only.
- safeInterval passes the Error and a task tag; the repeat-count summary
is info.
- failJob logs retries at info and a permanent failure at warning, so
sidecar failures reach GlitchTip through the API.
- Stripe webhook: a ledger insert failure other than a duplicate key now
answers 500 so Stripe retries. It was acked as a duplicate, which left a
paid order pending with nothing logged.
- Stripe, CRM sync, upload queue and auth DB faults pass the Error.
Client (@sentry/vue 8.55.2):
- release is caelum@<package version> (was caelum@0.0.0).
- sendDefaultPii off, no username on the user, query strings and cookies
scrubbed, extension noise ignored.
- log.debug/info become breadcrumbs; they were issues.
- Sourcemap plugin defaults point at GlitchTip with the same release name.
Sidecar: optional sentry-sdk; ERROR and up become issues, INFO and up go
to the log stream, frame locals off so the worker token stays out.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Both DEFAULT_DSN constants (server and client) now point at the caelum
project on the self-hosted GlitchTip instance, which accepts the Sentry
protocol. SENTRY_DSN and VITE_SENTRY_DSN still override. The CSP already
allows https: in connect-src, and GlitchTip answers the browser preflight
with Access-Control-Allow-Origin: *.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
corsOrigin() threw a plain Error, so the error handler logged it to Sentry
and answered 500. CAELUMNETWORK-2V (10-13) was two POSTs to /api/nox/check
with `Origin: null` and `Sec-Fetch-Site: same-origin` from bots whose client
hints contradict their user agent (Surface Pro model on macOS). A real Chrome
on caelum.network sends the true origin and gets 200. The error now carries
status 403 and a quiet flag, and the handler returns it without logging.
Allowed origins and requests with no Origin are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The home page listed Calibrate and Photometry as "runs today", said
frames are calibrated and reduced on arrival, and the meta description
said Caelum "calibrates" its frames. On caelum.network (checked
2026-10-07) the reduce worker is deployed and has run once, on 14
September: header, extract and timing completed on some frames and five
have measured seeing, but calibrate was skipped on every frame (no
calibration masters exist there), photometry has produced no
measurement (no camera has gain or read noise set), and no frame has
finished reduction.
Calibrate and Photometry now read "built, not live yet", with a note
saying why. Header, Extract and Timing keep "runs today". The step
copy, the archive card, the research lede, the crawler skeleton and
the meta descriptions match.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Say what Pulse records, that it keeps rotating codes instead of the IP
address, that it sets no cookie of its own and skips GPC and Do Not Track
browsers, and that a click on one of our emails sets a cookie that links
visits to our record of the person.
The policy said Caelum runs no analytics trackers; it now says no
third-party advertising trackers and names Pulse.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This one looked like it already worked. The path returned HTTP 200, so an
estate-wide check that only reads status codes counts it as present. What it
returned was the SPA shell: the catch-all answers any non-/api path with
index.html, so the file a researcher or a scanner looks for was a page of
HTML. A 404 would have been more honest.
SECURITY.md was no better. It told a reporter to "email the project
maintainers at the address on the GitLab project page", which is not an
address. It now names security@caelum.network, which was created as an alias
and confirmed to answer 250 in an SMTP conversation from off the mail server.
The route is registered before the SPA block so the catch-all cannot swallow
it, and Expires is computed per request: RFC 9116 makes it mandatory and tells
consumers to disregard a file that has passed it, so a hardcoded date is a
promise to come back and edit it in a year. Policy points at SECURITY.md,
which is publicly readable.
Verified on a real request: 200, text/plain; charset=utf-8. The production
container sets no PUBLIC_BASE_URL, so the canonical falls back to
https://caelum.network, which is correct.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Uploaded five real SDSS frames of M51 and M63 so the pipeline had genuine sky
to reduce. Extract worked: 513-702 sources per frame, seeing 2.7-3.1 px, which
at SDSS's 0.396 arcsec/px is about 1.2 arcsec. Both bugs below were invisible
to synthetic fixtures and to every test we had.
## A numeric keyword written as a quoted string was discarded
SDSS frames carry `EXPTIME = '53.907456'` — a STRING card, which is legal FITS
and common in real archives. extractFrameFields took numbers only, so
exposure_time landed null on all five frames and the timing stage failed with
"the frame lacks DATE-OBS or EXPTIME" on frames whose headers plainly had both.
num() now accepts a string that is ENTIRELY a number. 'N/A' and '12 deg' stay
null rather than becoming NaN and 12.
## The source catalogue was silently dropped
express.json()'s default limit is 100 KB. A rich real field is hundreds of
detections, and the upload came back "request entity too large" — so the
extract stage reported done while its entire catalogue went nowhere. Zero
frame_sources rows from 3,021 detected sources.
The worker path now parses with a 25 MB limit, mounted BEFORE the global parser
because body-parser marks a request parsed and later parsers skip it, so the
first matching one decides the limit. Scoped to that path so no ordinary route
accepts a body this size.
Mutation-tested: reverting num() to numbers-only fails a test.
Server tests 463.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
server-test:mysql failed on MR !21 with "Cannot find package 'better-sqlite3'"
— on the leg that gates deploy, for a backend it does not use.
better-sqlite3 is a native addon whose binary prebuild-install fetches during
npm ci. It sat in optionalDependencies, so a failed fetch was SKIPPED SILENTLY
and only surfaced forty lines into the test output as ERR_MODULE_NOT_FOUND,
looking like a code fault rather than an install one.
It was never optional. Eight test files set DB_KIND=sqlite themselves —
control-simulator, research-object, sqlite-backend, reduction-worker, messages,
reduction-pipeline, observatory-daemon and tonight — so BOTH CI legs need it
whatever their own DB_KIND says. The production image needs it too: image-test
boots the image with DB_KIND=sqlite, and the running container has it.
Moved to dependencies and the lockfile regenerated, so npm ci now fails on the
install rather than continuing into a suite that cannot run. A preflight after
npm ci keeps the diagnosis one line long instead of forty, and says the thing
that is not obvious: the suite needs it whatever DB_KIND says.
Verified: the preflight exits 0 when it loads and 1 when it does not, the CI
yaml lints to 15 jobs, and the server suite is still 460.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>