Cloudflare replaced the fast lane's Durable Object during the 00:00 UTC run.
The alarm restarted on a new instance while the old one carried on posting,
and the old run's results were lost because it could no longer write to its
own storage. Like the old cron, a run now takes a lock in the private R2
bucket (which a replaced instance can still write), refreshes it every 10
seconds and deletes it at the end; a run that finds a fresh lock waits and
checks again in 15 seconds. A scheduled run that never saved its results is
recorded as cut off.
Each updater now reports whether its published file changed, in its log
line and in the run's results, to see when fresh data really arrives at the
top of the hour.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Production writes data snapshots to the archive Space the Node server used
(splatoon3ink-archive) and binds the production compressor for the admin
panel.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
splat-mitmproxy now PUTs the captured /api/bullet_tokens response to
/bullet-token on each updater Worker, authenticated with its own
TOKEN_UPLOAD_TOKEN. The Worker caches it for the TOKEN_UPLOAD_REGION
account (US) and records an "iPad upload" event on the Nintendo auth row.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A failed run is no longer retried: the lane's next slot runs again (1-15
minutes later for the fast lane). A run that stops at its time budget with
work left queues a continuation in the same queue as manual runs, whether it
was scheduled or manual, so a lane's state is just its switch, next slot and
last run. Each alarm runs one thing, a due slot before a queued run, so a run
always has the alarm's whole time limit.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Pause and turning scheduling off did nearly the same thing, so pause is gone:
a lane with scheduling off runs only when started by hand. Every on/off control
in the admin panel (lanes, renewals, the daily archive run, social accounts) is
now a toggle switch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A credential renewal now counts as failed only when it ends without a new
bullet token. A step that failed and was made up for by a later one (znca
briefly unavailable, say) is kept as a warning and no longer goes to
Sentry.
Everything that runs on its own now has a switch in the admin panel:
scheduled renewals (one switch for every region) and the daily archive
compression join the lanes' scheduling switches. The compressor's starts
from AUTOMATIC_SCHEDULING_ENABLED: on in dev, off in production, so the
production compressor can be deployed before spectre stops.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Both Workers and their Durable Objects are wrapped with @sentry/cloudflare, so the shared
code's @sentry/core calls (failed updaters, social posts, archive snapshots) reach Sentry.
A lane run that fails outright, a failed credential renewal and a failed archive compression
are reported too. Durable Object RPC methods use instrumentPrototypeMethods; without it they
run with no Sentry client and their events are dropped.
The Node runner and the compressor's container no longer use Sentry; @sentry/node is gone.
POST /sentry-test sends one test event from the updater Worker and one from a Durable Object.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Compressor runs get an Archive row and filter on the timeline, a details view with the
run's log lines, and a Controls card with Compress now and Dry run. The updater reaches the
compressor through a COMPRESSOR binding to its Durable Object (dev only for now); without the
binding the panel leaves the archive out. Worker tests use a stand-in compressor Worker.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A separate compressor Worker runs app/data/ArchiveCompressor.mjs in a container (standard-1)
once a day at 00:30 UTC, against the archive Space. Its Durable Object starts the container
and keeps the last 30 runs with their log lines. An alarm every few seconds collects the
container's output from its log server and ends the run when the container reports its exit
code, so a run survives the object being evicted from memory.
Only dev is deployed; production waits until spectre stops compressing at the same time.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The modes mostly chose which lane's updaters ran, which the lanes now decide. The fast
lane keeps one option, translations (quick runs at :00-:04 skip them); the slow lane has
none. Runs, slots and manual requests carry the option instead of a mode, and the admin
panel shows runs as quick or full, with an "Include translations" box for manual fast
runs. Slots and runs saved with a mode run with the defaults.
The Node runner follows the same cadence: npm run splatnet:quick is the fast lane without
translations, splatnet the fast lane, and splatnet:all both lanes. The unused disableXRank
and disableFestivalRankings settings are removed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With ARCHIVE_DATA set, the Worker writes each published data file to the
archive Space, at the same YYYY/MM/DD/<date>.<time>.<name>.json keys, ACL,
and content type as the Node uploader (DataArchiver):
- S3ArchiveStorage signs PUTs with aws4fetch, tries three times with
backoff, and stops early on errors that will not improve.
- A snapshot that cannot be written is a warning and a Sentry event, never
a failed update.
- The fast lane holds its snapshots until after social posts (even failed
ones) and then uploads the newest copy of each key, so a slow archive
bucket never delays a post. The slow lane uploads as it goes, to avoid
holding large X-Rank files in memory.
- Dev writes to its own Space (splatoon3ink-archive-test); production has
no archive settings until cutover. Daily compression is still to come.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Timeline page showed four kinds of things on the time axis but only
lane runs in the list, and credential renewals could not be opened. Now
both draw from one set of items (lib/activity.js): lane runs, credential
renewals, social posts, and what is coming up next.
- The timeline scrolls back through the last 48 hours by dragging (with a
short glide after a flick), and loads earlier history six hours at a
time. Marks that would cover each other are stacked so each can be
clicked.
- The timeline and the list follow one cursor, so scrolling either moves
the other. A marker shows the moment the list is lined up with, and
"Now" returns both to the present. The list is windowed, with upcoming
items above a "now" line.
- Everything opens in the details panel: renewals with their errors and
log lines, posts with each account's result, upcoming items with when
and why. A selected upcoming or running item stays selected as it
starts and finishes. The selection is ?item=<key> in the address.
- Lane and credential controls move to a new Controls page, which links
back to the timeline. The Timeline page notes a paused or unscheduled
lane.
- The list filters by row and by problems; the timeline dims the rest.
Worker side:
- Scheduler keeps runs for 48 hours (HISTORY_MS) instead of the last 50,
with log lines stored apart from the run summaries so listing stays
cheap. New history({ from, to }) RPC; run() finds a run directly when
given its start time; posts() takes a time span.
- RegionAuth keeps 48 hours of renewals, each with its trigger and log
lines, and serves them by time span and one at a time.
- New GET /admin/api/activity (by time span) and /admin/api/auth/event.
Status carries only the newest activity, and log lines from renewals
name the renewal that printed them.
Checked in the local preview against the simulated API; not yet run
against a deployed Worker.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
A small Vue + Tailwind app in workers/updater/admin, served by the updater
Worker on its own hostname (admin.splatoon3.ink, admin.dev.splatoon3.ink)
behind Cloudflare Access:
- Timeline: both lanes, credential renewals, and social posts on one time
axis, a filterable run list, run details with per-updater timings and
the run's own log lines, lane controls (run now, pause, scheduling),
credential renewal, and a combined text log.
- Social: the post log with errors, per-account posting switches, when
each status last went to each account, and a preview that can post one
status by hand after a confirmation.
- Data: freshness of the core data files, the published file list, a
check that the site serves the new data, and a cache purge.
Worker side:
- src/admin: the Access JWT check (jose; fails closed and only answers on
ADMIN_HOSTNAME) and the API. Actions must be same-origin JSON POSTs, and
the app is served with a strict CSP from the ADMIN_ASSETS binding.
- Runs capture their console output (AsyncLocalStorage), so log lines
belong to the run that printed them. The Scheduler keeps a post log and
serves run details and log lines; RegionAuth keeps its last 24 renewals.
- A manual run from the panel only posts to social when asked to.
Shared code:
- StatusGeneratorManager records what was posted or failed, and skips
accounts turned off in private settings/social.json (the Node runner
honours this too). app/social/admin.mjs previews and posts one status.
- Updater results include their duration.
Also: npm run admin:dev runs the panel against a simulated API, deploy
scripts build the panel first, the dev environment's assets move to
dev-assets.splatoon3.ink, and the lockfile drops entries nothing depends
on any more.
Not yet run against a deployed Worker or a real Access application.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Durable Object alarms are stopped after 15 minutes, and a full X-Rank or
Splatfest rankings update can take longer than that:
- RunLimits takes an optional deadline. X-Rank season details, Splatfest
rankings, and the slow lane's updater loop stop starting new work once
it passes and mark the run incomplete; what is left updates in a later
run. The current X-Rank season goes first.
- The slow lane runs with a 10-minute budget. An incomplete run is
followed a minute later by a continuation (up to 10, never past the
lane's next regular slot), which does not count as a retry. Manual runs
continue the same way through a queued follow-up.
- Ignore Wrangler's local secrets files (.dev.vars*).
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Run the shared app/ updaters and social posters from workers/updater:
- Scheduler Durable Object lanes: "fast" keeps the Node cron cadence
(quick at :00-:04, default at :05/:10/:15/:20/:30/:45) and posts to
social after checking SITE_URL serves the new data; "slow" runs X-Rank,
Splatfest rankings, and avatar updates hourly. Bounded retries, queued
manual runs, pause, and persisted automatic-scheduling state.
- RegionAuth Durable Object per region: the single owner of each
Nintendo login, serving bullet tokens to both lanes and renewing them
hourly at :35-:38.
- A five-minute cron watchdog that only re-arms alarms, bearer-token
control endpoints, and dev/production environments. DO migrations keep
Scheduler at v1 to match the earlier dev deployment.
- Per-run limits (RunLimits) and per-account NsoClient locks instead of
module-level limiters, since Durable Objects sharing an isolate cannot
share in-flight requests. Updaters without localization rules no longer
rewrite locale files.
- 20 workerd tests (npm run updater:test, included in npm test) with all
network access blocked.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Groundwork for running the existing updaters and social posters in a Worker:
- Derive IDs with @noble/hashes shake256; the Workers runtime's node:crypto
lacks shake256. A test pins an ID produced by node:crypto.
- Split updaters into fast (schedules, gear, co-op, festivals, stages; run
concurrently) and slow (X-Rank, Splatfest rankings; run one at a time).
The cron posts after the fast updaters, before the slow ones.
- Fetch X-Rank seasons and detail types sequentially, cutting peak heap per
division from ~85 MB to ~38 MB.
- Move Splatfest rankings into FestivalRankingsUpdater, driven by the
published festivals data, so results never wait on ranking pagination.
- Coalesce concurrent downloads of the same SplatNet image.
- Bring over the Browser-binding screenshot work from develop (f7293b6):
Cloudflare captures use the BROWSER binding with retries, Browserless is
removed, and workers/screenshots offers capture-only local testing. The
local driver keeps puppeteer-core with an installed Chrome, always serves
dist/, and fails on page-reported screenshot errors.
- Remove the Docker image, compose files, and GHCR deploy workflow.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>