H8: prefetched tiles and file blobs could be evicted under storage pressure
(worsened by opaque tile responses inflating the quota ~7MB each), blanking the
offline map right when a traveler needs it. Request persistent storage at app
init so the browser exempts our caches from eviction. We deliberately keep tile
requests no-cors (a cors switch would break self-hosted/custom tile providers
without CORS headers), so persistence is the safe mitigation rather than
de-opaquing responses.
H9: Mapbox GL users had no offline map at all — no runtimeCaching matched the
Mapbox hosts. Add a StaleWhileRevalidate rule for api.mapbox.com /
*.tiles.mapbox.com so visited areas are available offline (best-effort; full
pre-download still requires the Leaflet renderer, now documented).
- new sync/persistentStorage.ts requestPersistentStorage(), called from main.tsx
- vite.config: mapbox-tiles SW cache rule
- MapViewAuto / tilePrefetcher comments document the offline-maps policy
- tests for the persist helper (granted / already-persisted / absent / rejects)
setRefetchCallback was dead code, so on reconnect the queue flushed and Dexie
re-seeded but the open trip's Zustand store was never refreshed — a
collaborator's edits made while we were offline didn't appear until navigating
away and back.
- new tripStore.hydrateActiveTrip(): silent refresh of the active trip's
collaborative state (days/places/packing/todo/budget/reservations/files),
no resetTrip and no isLoading toggle so there's no splash on reconnect
- syncTriggers wires setRefetchCallback to it (WS layer awaits the flush hook
first) and re-hydrates open trips after the online-event syncAll; cleared on
unregister
- websocket exposes getActiveTrips() for the online-event path
- tests: refetch wiring + ordering, silent hydrate without reset/splash
Blob cache previously leaked forever: clearTripData omitted it, entries had
no trip discriminator, and there was no size/count bound, so file blobs
survived trip eviction and could starve the map-tile cache for quota.
- BlobCacheEntry gains tripId + bytes; Dexie v3 adds a tripId index with a
backfill upgrade (legacy rows -> tripId -1, bytes from blob.size)
- clearTripData purges the trip's blobs in-transaction
- enforceBlobBudget() evicts oldest-by-cachedAt past 200 entries / 100 MB
- tripSyncManager threads tripId/bytes into puts and enforces the budget
Closes BLOCKER B4 — three reinforcing paths could serve one account's
cached data to the next user on a shared device:
- The Workbox 'api-data' cache keyed trip/user-scoped GETs by URL only
(cookie-blind). Changed to NetworkOnly; offline reads come from the
per-user IndexedDB cache via the repo layer instead.
- IndexedDB had no per-user scoping. The Dexie connection is now scoped
per user (trek-offline-u<id>) behind a Proxy so the ~19 importers keep a
stable binding; login opens the user DB, logout deletes it and returns
to the anonymous DB.
- logout() was fire-and-forget and racy: background flush/syncAll could
re-seed the DB after the wipe. It is now async and ordered — close an
auth gate, unregister sync triggers, disconnect, clear caches, delete
the user DB — and flush()/syncAll() bail when the gate is closed.
Closes BLOCKER B5 — the offline map was blank for most real trips:
- The Workbox 'map-tiles' cache held only 1000 entries while the prefetcher
budgeted ~3413, so prefetched tiles were evicted on arrival. Both caps are
now a coherent 12288 (~180 MB), kept in sync with cross-referencing comments.
- prefetchTilesForTrip skipped a trip entirely when its all-zooms estimate
exceeded the cap, so region/road-trip bboxes got no tiles. Removed the
all-or-nothing guard; prefetchTiles already fills zooms low→high and stops at
the budget, so large trips now cache the zooms that fit instead of nothing.
Closes three offline BLOCKERs from the PWA audit:
- B1: offline edits/deletes of an offline-created entity were lost. The
negative temp id was baked into the PUT/DELETE url and never rewritten
after the CREATE returned a real id, so dependents 404'd and were dropped.
Dependents now carry a {id} placeholder + tempEntityId; flush builds a
tempId->realId map and durably rewrites still-queued dependents on CREATE
success (survives flush boundaries / reloads).
- B2: tempId = -(Date.now()) collided within a millisecond, overwriting an
optimistic row. Replaced with a monotonic nextTempId() minter.
- B3: any 4xx marked the mutation failed with no rollback and no signal, and
the badge ignored failed rows. Terminal failures now roll back the phantom
optimistic CREATE; 401/408/425/429 are treated as retryable; failedCount()
is surfaced in OfflineBanner (red pill) and OfflineTab.
* fix(mcp): MCP RFC compliant for more strict clients
* fix(mcp): serve flat /.well-known/oauth-protected-resource for ChatGPT reconnect
Clients such as ChatGPT probe the flat well-known URL on every fresh discovery
cycle (i.e. after a full disconnect/reconnect where cached OAuth state is cleared).
The SDK's mcpAuthMetadataRouter only serves the path-based form
/.well-known/oauth-protected-resource/mcp, so the flat probe returned 404.
Without the resource metadata, ChatGPT fell back to the issuer URL as the
resource parameter (https://…/ instead of https://…/mcp). The authorize handler
then rejected it with invalid_target and redirected back to ChatGPT's callback
with an error — showing the user the TREK home page instead of the consent form.
Add an explicit GET handler for the flat URL that returns the same protected
resource metadata, so the resource URI is discovered correctly on the first probe.
* fix(mcp): fix OAuth popup blank page — SW denylist and COOP header
Service worker was intercepting /oauth/authorize navigate requests
(not in denylist), serving index.html, and React Router's catch-all
redirected to / instead of the SDK authorize handler.
Helmet's default COOP: same-origin isolated the /oauth/consent popup
from its cross-origin opener, making window.opener null and breaking
the popup-based OAuth completion signal for ChatGPT and similar clients.
* fix(ntfy): encode non-Latin-1 header values with RFC 2047 to prevent ByteString crash
Todo/trip names containing chars like → or € (and non-Latin-1 locale templates
for Czech, Chinese, Russian, etc.) caused the Fetch API to throw when setting
the ntfy Title header. Apply RFC 2047 base64 encoded-word encoding for any
header value containing chars above U+00FF; ntfy decodes this automatically.
* docs(mcp): document Cloudflare bot detection blocking ChatGPT MCP requests
Add Cloudflare WAF note to MCP-Setup and a full troubleshooting entry covering
root cause (IP reputation + UA heuristics), free-plan limitation (disable Bot
Fight Mode entirely, with explicit warning), and paid-plan WAF skip rule with
the full expression syntax and path table for all MCP/OAuth/.well-known routes.
* fix(pwa): detect upstream proxy auth challenges and recover gracefully
Behind Cloudflare Zero Trust or Pangolin, cross-origin auth redirects on
/api/* calls surface as CORS errors (error.response === undefined) that
the existing 401 interceptor never catches, leaving the PWA stuck with
network-error toasts instead of re-authenticating.
New connectivity module probes /api/health every 30s using fetch with
cache:no-store and inspects Content-Type to reliably detect whether the
server is reachable vs intercepted by an upstream proxy.
axios interceptor changes:
- On !error.response + navigator.onLine: run probeNow(); if the health
probe also fails (proxy is intercepting all requests), trigger a guarded
window.location.reload() so the edge proxy can intercept the top-level
navigation and run its auth flow (covers CF Access and Pangolin 302 mode)
- On error.response status 401 with text/html body: same reload path,
covering Pangolin header-auth extended compatibility mode which returns
401+HTML instead of a 302 redirect. TREK own 401s are always JSON so
there is no collision with the existing AUTH_REQUIRED branch.
- sessionStorage flag prevents reload loops; cleared on any successful
response so the guard resets after re-auth.
/api/health excluded from SW NetworkFirst cache (vite.config.js regex)
and Cache-Control: no-store added server-side so probes always hit the
network and cannot be served stale from the 24h api-data cache.
LoginPage caches last-known appConfig in localStorage so the SSO button
renders in OIDC+UN/PW dual mode even when the config fetch is intercepted
by the proxy. Auto-redirect to IdP skipped when config comes from cache
to avoid redirect loops while the proxy is challenging.
Fixes discussion #836.
* fix(files): add bottom-nav padding to files tab wrapper on mobile
* fix(budget): expose toolbar on mobile so users can add budget categories
* fix(pwa): unregister SW before proxy-reauth reload so Pangolin can challenge
WorkBox's NavigationRoute served the cached SPA shell on window.location.reload(),
meaning Pangolin/CF Access never saw the navigation and the app was left stuck
showing stale offline data. Unregistering the SW first lets the navigation reach
the network so the upstream proxy can run its auth flow.
Also rebuilds server/public with corrected sw.js (health excluded from
NetworkFirst, /oauth/ and /.well-known/ added to NavigationRoute denylist).
* chore: remove committed build artifacts from server/public
Dockerfile and Proxmox community script both rebuild client/dist and copy
it into server/public at build time — committed artifacts were never used.
Replace with .gitkeep and add server/public/* to .gitignore.
* chore: add build-from-sources script
jubnl