- Add homelab/glitchtip/files_api.py with GzipChunk fallback for sentry-cli 3.x
raw zip uploads; bind-mount into both glitchtip-web and glitchtip-worker
- Add release.yaml prune step to delete all but the 10 newest GlitchTip releases
- Reader page: remove dead code and simplify layout
Adds a full-screen listening mode accessible via the headphones button in the
mini-player bar. Moves voice selector, speed, auto-next, and sleep timer out of
the reader settings panel into the new overlay. Voices are stored in AudioStore
so ListeningMode can read them without prop drilling.
- Mount glitchtip_uploads named volume to /code/uploads on both
glitchtip-web and glitchtip-worker so chunk blobs written by the
web container are readable by the worker during assembly
- Set MEDIA_ROOT=/code/uploads explicitly on both services so the
Django storage path matches the volume mount and DB blob records
- New `ai_jobs` PocketBase collection tracks all long-running AI tasks
(batch-covers, chapter-names) with status, progress, and cancellation
- `handlers_aijobs.go`: GET/cancel endpoints for ai_jobs; centralised
cancel registry (moved from handlers_catalogue)
- Batch-covers and chapter-names SSE handlers now create/resume ai_job
records, support from_item/to_item ranges, and resume from items_done
on restart via job_id
- New `POST /api/admin/image-gen/auto-prompt`: generates an image prompt
from book description (cover) or chapter title (chapter) via LLM
- image-gen page: "Auto-prompt" button calls auto-prompt API when a slug
is selected; falls back gracefully if TextGen not configured
- text-gen chapter-names: from/to chapter range inputs + job ID display
- catalogue-tools batch-covers: from/to item range + resume job ID input
- pb-init-v3.sh: adds ai_jobs collection (idempotent)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add 'releases new' and 'releases finalize' steps around sourcemaps
upload in release.yaml — without an explicit 'releases new' call,
GlitchTip creates the release entry but associates 0 files.
Add root AGENTS.md (picked up by Claude, Cursor, Copilot, etc.) with
full project context: stack, repo layout, Gitea CI conventions,
GlitchTip DSN/upload flow, infra, and iOS notes.
On md+ screens the drawer is now right-aligned and 320px wide (w-80)
instead of full-viewport-width. The sticky header is pulled out of the
scroll container so it never scrolls away, and overflow-y-auto is
applied only to the chapter list itself so both mobile and desktop can
scroll through long chapter lists.
- Add @grafana/faro-web-sdk to UI; wire initializeFaro in hooks.client.ts
gated on PUBLIC_FARO_COLLECTOR_URL (no-op in dev)
- Add Grafana Alloy service (faro.receiver) to homelab compose;
Faro endpoint → alloy:12347 (faro.libnovel.cc via cloudflared)
- Add PUBLIC_FARO_COLLECTOR_URL env var to docker-compose.yml UI service
- Add Web Vitals dashboard (web-vitals.json): LCP/INP/CLS/TTFB/FCP p75
stats + LCP/TTFB time-series + Faro exception logs from Loki
- Fix runner.json: strip libnovel_ prefix from all metric names
- Fix backend.json: replace 5 dead http_client_* panels with
spanmetrics-based equivalents (Request Rate by Span Name + Latency
by Span Name p95)
- Fix OTel collector: add service.telemetry.metrics.address: 0.0.0.0:8888
so Prometheus can scrape collector self-metrics
- Add Grafana link to admin nav external tools; add admin_nav_grafana
message key to all 5 locale files; recompile paraglide
CF Workers AI has a ~4MB JSON body limit. Large cover images base64-encoded
can easily exceed this, causing Cloudflare to return a 502 Bad Gateway.
Added resizeRefImage() which scales the reference down so its longest side
≤ 768px (nearest-neighbour, re-encoded as JPEG) before passing it to the
GenerateImageFromReference call. Images already within the limit pass through
unchanged.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Llama 4 Scout returns `result.response` as an array of objects
[{"generated_text":"..."}] instead of a plain string. Decode into
json.RawMessage and try both shapes; fall back to generated_text[0].
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Cover generation: editable prompt (pre-filled from title+summary),
img2img toggle to use existing cover as reference, Full editor link
- Chapter cover: editable prompt textarea
- Description: instructions input field, Full editor link
- Chapter names: editable pattern field, SSE streaming with live batch
progress, inline-editable title proposals, batch warning display
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Split chapter-name LLM requests into 100-chapter batches and stream
results back as SSE so large books (e.g. Shadow Slave: 2916 chapters)
never time out or truncate. Frontend shows live batch progress inline
and accumulates proposals as they arrive.
chapters_idx was missing created/updated columns (never defined in the
PocketBase schema), causing PocketBase to return 400 for any query
sorted by -created. recentlyUpdatedBooks() uses this sort.
- Add created date field to chapters_idx schema in pb-init-v3.sh
(also added via add_field for existing installations)
- Add idx_chapters_idx_created index for sort performance
- Set created timestamp on first insert in upsertChapterIdx so new
chapters are immediately sortable; existing records retain empty created
and will sort to the back (acceptable — only affects home page recency)
Without com.centurylinklabs.watchtower.enable=true the homelab watchtower
(running with --label-enable) silently skipped the runner container,
leaving it stuck on v2.5.60 while fixes accumulated on newer tags.
- Replace ?? with || when reading font_size so 0 falls back to 1.0
(affects GET /api/settings, layout.server.ts, +layout.svelte)
- Remove the explicit 'body.fontSize !== 0' exception in PUT /api/settings
validation so 0 is now correctly rejected as an invalid font size
- Add add_index helper + idx_chapters_idx_slug_number declaration to
scripts/pb-init-v3.sh (idempotent UNIQUE INDEX on chapters_idx)
marked() returns string | Promise<string>; the previous cast 'as string'
silently passed a Promise object, which Svelte rendered as nothing.
Free users saw blank content even though SSR HTML was correct.
- Fix upsertChapterIdx race: use conflict-retry pattern (mirrors WriteMetadata)
so concurrent goroutines don't double-POST the same chapter number
- Add DeduplicateChapters to BookWriter interface and Store implementation;
keeps the latest record per (slug, number) and deletes extras
- Wire POST /api/admin/dedup-chapters/{slug} handler in server.go
The aura-2-en model enforces a hard 2 000-character limit per request.
Chapters routinely exceed this, producing 413 errors.
GenerateAudio now splits the stripped text into ≤1 800-char chunks at
paragraph → sentence → space → hard-cut boundaries, calls the API once
per chunk, and concatenates the MP3 frames. Callers (runner, streaming
handler) are unchanged. StreamAudioMP3/WAV inherit the fix automatically
since they delegate to GenerateAudio.
The old vertical dropdown was too tall on mobile — 14 stacked groups
required heavy scrolling and had no visual hierarchy.
New design:
- Full-width bottom sheet slides from screen edge (natural mobile gesture)
- Drag handle + dimmed backdrop for clarity
- Two tabs split the settings: Reading (typography + layout) and Listening
(player style, speed, auto-next, sleep timer)
- Each row uses label-on-left + pill-group-on-right layout — saves one line
per setting and makes the list scannable at a glance
- Settings are grouped into titled cards (Typography, Layout, Player)
with dividers between rows instead of floating individual blocks
- Gear button moved to bottom-[4.5rem] to clear the mini-player bar
html and fetchingContent were captured with untrack() at mount time.
When SvelteKit re-ran the page load (triggered by the layout's settings PUT),
data.html updated but html stayed stale. The {#key} block in the layout then
destroyed and recreated the component, and on remount data.html was momentarily
empty so html became '' and the live-scrape fallback ran unnecessarily.
Fix:
- html is now $derived(scrapedHtml || data.html || '') — always tracks load
- scrapedHtml is a separate $state only set by the live-scrape fallback
- fetchingContent starts false; the fallback sets it true only when actually fetching
- translationStatus/translatingLang: dropped untrack() so they also react to re-runs
- Removed unused untrack import
Both vars were present in Doppler but never injected into the ui service
environment, causing all checkout requests to fail with 500 and webhooks
to be silently rejected.
- Add /subscribe route with hero, benefits list, and pricing cards
(annual featured with Save 33% badge, monthly secondary)
- Add 'Pro' link in nav for non-Pro users
- Add 'See plans' link in profile subscription section
- i18n keys across en/fr/id/pt/ru for all subscribe strings
Note: checkout still requires POLAR_API_TOKEN with checkouts:write scope.
Regenerate the token at polar.sh to fix the 502 error on subscribe buttons.
PocketTTS emits 16-bit PCM WAV (16 kHz mono). WAV is natively supported
on all browsers including iOS/macOS Safari, so the ffmpeg MP3 transcode
is unnecessary for the streaming path.
Using format=wav for PocketTTS voices eliminates the ffmpeg subprocess
startup delay (~200–400 ms) and a pipeline stage, giving lower latency
to first audio frame. Kokoro and CF AI continue using MP3 (they output
MP3 natively or via the OpenAI-compatible endpoint).
The runner (MinIO storage) is unaffected — it still stores MP3 for
space efficiency.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously all voices waited for full TTS generation before first byte
reached the browser (POST → poll → presign flow). For long chapters this
meant 2–5 minutes of silence.
New flow for Kokoro and PocketTTS:
- tryPresign fast path unchanged (audio in MinIO → seekable presigned URL)
- On cache miss: set audioEl.src to /api/audio-stream/{slug}/{n} and mark
status=ready immediately — audio starts playing within seconds
- Backend streams bytes to browser while concurrently uploading to MinIO;
subsequent plays use the fast path
New SvelteKit route: GET /api/audio-stream/[slug]/[n]
- Proxies backend handleAudioStream (already implemented in Go)
- Same 3 chapters/day free paywall as POST route
- HEAD handler for paywall pre-check (no side effects, no counter increment)
so AudioPlayer can surface upgrade CTA before pointing <audio> at URL
CF AI voices keep the old POST+poll flow (batch-only API, no real streaming
benefit, preserves the generating progress bar UX).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- AudioPlayer: add 'compact' playerStyle prop — slim seekable player with
progress bar, skip ±15/30s, play/pause circle button, and speed cycle
- Chapter reader settings gear panel: new Listening section with player
style picker (Standard/Compact), speed control (0.75–2×), auto-next
toggle, and sleep timer — all persisted in reader_layout_v1 localStorage
- audioStore.speed/autoNext/sleep now accessible directly from settings
panel without opening the audio player
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Without explicit setActionHandler('play'/'pause'/...) the browser uses
default handling which on iOS Safari stops working after ~1 min of
pause in background/locked state (AudioSession gets suspended and the
lock screen button can't resume it without a direct user gesture).
Registering handlers in the layout (where audioEl lives) ensures:
- play/pause call audioEl directly — iOS treats this as a trusted
gesture and allows .play() even from the lock screen
- seekbackward/seekforward map to ±15s / ±30s
- playbackState stays in sync so the lock screen shows the right icon
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Default max_tokens to 4096 for chapter-names so large chapter lists
are not cut off mid-JSON by the model's token limit
- Rewrite system prompt to clarify placeholder semantics ({n} = number,
{scene} = scene hint) and explicitly forbid echoing the number inside
the title field — prevents "Chapter 1 - 1: ..." style duplications
- UI: surface raw_response in the error area when chapters:[] is returned
so the admin can see what the model actually produced
Adds backend handlers and SvelteKit UI for an admin text generation tool.
The tool lets admins propose and apply AI-generated chapter titles and book
descriptions using Cloudflare Workers AI (12 LLM models, model selector shared
across both tabs).
Both vars were in Doppler but never forwarded via docker-compose.yml, causing
the image generation handler to return 503 "CFAI_ACCOUNT_ID/CFAI_API_TOKEN missing".
Committed hooks directory (.githooks/pre-commit) auto-runs `npm run paraglide`
and force-stages the generated JS output whenever ui/messages/*.json files are
staged, preventing svelte-check CI failures from stale paraglide output.
Added `just setup` recipe to configure core.hooksPath for new contributors.
Five new reader settings (persisted in localStorage as reader_layout_v1):
- Read mode: Scroll (default) vs Pages — paginated mode splits content
into viewport-height pages, navigate with tap left/right, arrow keys,
or Prev/Next buttons. Recalculates on content change.
- Line spacing: Tight (1.55) / Normal (1.85) / Loose (2.2) via
--reading-line-height CSS var on :root.
- Reading width: Narrow (58ch) / Normal (72ch) / Wide (90ch) via
--reading-max-width CSS var on :root.
- Paragraph style: Spaced (default) vs Indented (text-indent: 2em,
tight margin — book-like feel).
- Focus mode: hides audio player, language switcher, bottom nav and
comments so only the text remains.
Scroll progress bar: thin 2px brand-colored bar fixed at top of
viewport in scroll mode, fills as you read through the chapter.
All options added to the floating settings gear panel.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Admin changelog page now fetches releases live from the public Gitea
API (https://gitea.kalekber.cc) instead of a baked releases.json file
- Remove /static/releases.json from ui/.gitignore (no longer generated)
- Add Gitea to admin sidebar external tools (admin_nav_gitea i18n key,
all 5 locales, compiled paraglide JS force-added)
- release.yaml: remove 'Fetch releases from Gitea API' step; extract
release title and body from the tagged commit message for Gitea releases
New feed_*.js and nav_feed.js outputs from npm run paraglide were not tracked
because src/lib/paraglide/.gitignore ignores everything by default. Force-add
them so CI can run svelte-check without running paraglide compile first.
Also include recentlyUpdatedBooks fallback improvements from working tree.
Fix "Recently Updated" showing stale books: replace meta_updated
sorting (only changes on metadata writes) with chapters_idx sorted
by -created, so the section now reflects actual chapter activity.
Add per-section show/hide toggles on the home page, persisted in
localStorage via Svelte 5 $state. Each section header gets a small
hide button; hidden sections appear as restore chips above the footer.
Toggleable: Recently Updated, Browse by Genre, From Following.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Mobile action area was cluttered with Continue, Start ch.1, bookmark,
stars, and shelf dropdown all competing in a single unstructured row.
- Split mobile CTAs into two clear rows:
Row 1: primary read button(s) only (Continue / Start from ch.1)
Row 2: bookmark icon + shelf dropdown + star rating inline
- 'Start from ch.1' no longer stretches to flex-1 when Continue is
present — it's a compact secondary button instead
- Stars and shelf dropdown moved out of the CTA row into their own line
Genre tags were plain <span> elements with no interaction. Tapping
'fantasy' or 'action' now navigates to /catalogue?genre=fantasy,
pre-selecting the genre filter on the catalogue page.
The save-settings $effect was firing on the initial data load because
settingsApplied was set to true synchronously in the apply effect, then
currentTheme/fontFamily/fontSize were written in the same tick — causing
the save effect to immediately fire with uninitialized default values
(theme: "", fontFamily: "", fontSize: 0), producing a 400 error.
- Add settingsDirty flag, set via setTimeout(0) after initial apply so
the save effect is blocked for the first load and only runs on real
user-driven changes
- Also accept empty string / 0 as 'not provided' in PUT /api/settings
validation as a defensive backstop
OAuth login was silently failing: upsertUserSession() queried the
device_fingerprint column which didn't exist in the user_sessions
collection, PocketBase returned 400, the fallback authSessionId was
never written to the DB, and isSessionRevoked() immediately revoked
the cookie on first load after the OAuth redirect.
- scripts/pb-init-v3.sh: add device_fingerprint text field to the
user_sessions create block (new installs) and add an idempotent
add_field migration line (existing installs)
Audio jobs were stuck pending because the homelab runner was
connecting to its own local Redis instead of the prod VPS Redis.
- homelab/docker-compose.yml: change hardcoded REDIS_ADDR=redis:6379
to ${REDIS_ADDR} so Doppler injects rediss://redis.libnovel.cc:6380
(the Caddy TLS proxy that bridges the homelab runner to prod Redis)
