- Fix audio presign 404: MinIO upload is now synchronous before response is sent,
eliminating the race where presign was called before the file landed in MinIO
- Replace all Browserless usage with direct HTTP client across catalogue, metadata,
ranking, and browse — novelfire.net pages are server-rendered and don't need a
headless browser; direct is faster and more reliable
- Harden handleBrowse with 3-attempt retry loop, proper backoff, and full
browser-like headers to reduce 502s from novelfire.net bot detection
- Remove Browserless env vars (BROWSERLESS_URL/TOKEN/STRATEGY) from main.go;
add SCRAPER_TIMEOUT as a single timeout knob
- Clean up now-dead rejectResourceTypes var and Browserless-specific WaitFor/
RejectResourceTypes/GotoOptions fields from scraper calls
- ScrapeCatalogue: use li.novel-item (was div), extract href from outer <a> and title from h4.novel-title (was h3), detect next page via rel=next (was class=next)
- handleBrowse/triggerBrowseSnapshot: treat zero-byte MinIO cache entries as misses, skip storing empty SingleFile output
- runAsync: after a successful full-catalogue run, call ScrapeRanking and upsert each result into PocketBase ranking collection
novelfire.net chapter-list pages (/chapters?page=N) are server-rendered —
verified via curl. Switch urlClient to NewDirectHTTPClient alongside the
existing chapterClient. Remove BROWSERLESS_URL_STRATEGY env var and clean
up the now-irrelevant WaitFor/GotoOptions fields from both ScrapeChapterList
and ScrapeChapterListPage ContentRequests.
novelfire.net chapter content is server-rendered, so Browserless is not
needed. Add a dedicated chapterClient (always StrategyDirect) to Scraper
and use it in ScrapeChapterText, removing the now-irrelevant WaitFor /
RejectResourceTypes / GotoOptions fields from the ContentRequest.
- e2e fixture: replace single contentClient with directClient (plain HTTP)
for chapter/metadata/ranking + contentClient (Browserless) for urlClient
only — matches production wiring and is significantly faster
- server: add max_chars field to audio request body; truncates stripped text
to N runes before sending to Kokoro (used by e2e for quick TTS tests)
- fix: move RankingItem to scraper package to break novelfire→storage import
cycle; storage.RankingItem is now a type alias for backward compat
- fix: update stale New() call in novelfire integration test (missing args)
- fix: replace removed blob-ranking methods in storage integration test with
current per-item API (UpsertRankingItem/ListRankingItems/RankingLastUpdated)
- justfile: add test-e2e and e2e tasks
- Replace SetRanking/GetRanking/SetRankingPageHTML/GetRankingPageHTML blob methods
with WriteRankingItem/ReadRankingItems/RankingFreshEnough per-item operations
- Add 24h staleness gate in ScrapeRanking to skip re-scraping fresh data
- Add GET /api/ranking endpoint returning []RankingItem sorted by rank
- Remove RankingPageCacher interface and rankingCacheAdapter adapter
- Update integration tests to use new per-item upsert semantics
- Include e2e test suite (scraper/internal/e2e/)
The previous selectors were based on a hypothetical structure that does not
match the actual site. The real novelfire.net popular listing uses:
<li class="novel-item"> (not <div>)
<a href="/book/slug" title="Title">
<figure class="novel-cover"><img data-src="/path.jpg"></figure>
<h4 class="novel-title text2row">Title</h4>
</a>
</li>
And pagination uses <a rel="next"> (not <a class="next">).
Changes:
- ScrapeRanking: use li.novel-item, h4.novel-title, figure.novel-cover,
img[data-src]; strip base64 placeholder covers
- hasNextPageLink(): new helper walking all <a> nodes for rel="next"
- Import golang.org/x/net/html for Node.Attr access
- Test fixtures rewritten to match real structure (li/h4/rel=next)
- Status and genres removed from ranking items (not present on listing page)
Verified end-to-end: ranking.json written with 24 items on first fetch
- Switch ScrapeRanking to novelfire.net/genre-all/sort-popular URL and updated DOM selectors (div.novel-item, h3.novel-title, div.genres)
- Replace 5 hardcoded refresh buttons with dynamic 100-page paginator (smart ellipsis via rankingPageNums)
- Add RankingPageCacher interface and writer methods to cache raw HTML per page under static/books/_ranking_cache/page-N.html
- ScrapeRanking serves from disk cache on hit and writes to cache on miss, skipping Browserless round-trip
- Thread writer as PageCacher through novelfire.New and main.go
- Add TestScrapeRanking_CacheHit and TestScrapeRanking_CacheMiss tests
- Fix channel drain goroutine deadlock in handleRankingRefresh: replace
'for { ... continue ... if nil break }' with 'for A != nil || B != nil'
so the select never blocks on two nil channels
- Switch ScrapeRanking from urlClient (Browserless) to client (direct HTTP)
since novelfire.net/ranking is fully server-rendered — no JS needed
- Add ranking unit tests (single page, multi-page, empty page, write round-trip)
- Add .gitea/workflows/ci.yaml: lint, test, build jobs with commented-out
Docker image push step for when runner has Docker available
- ScrapeRanking now accepts a maxPages int parameter (0 = all pages).
Each page is fetched strictly sequentially; the next page is only
requested after every entry from the current page has been sent,
so there is no pre-fetching or look-ahead.
Pagination stops automatically when no next-page link is present
or when the rank-novels container is absent/empty.
- The ranking URL pattern follows the existing catalogue convention:
/ranking?page=N (next-page link detection as the stop condition).
- Server: handleRankingRefresh reads an optional 'pages' form field
and passes it to ScrapeRanking. Timeout scales at 90 s/page.
- UI: Refresh Rankings button is now a small form with a numeric
'Pages' input (default 1), letting the user choose how many pages
to pull in one refresh without touching the server config.
Two bugs caused the 'Refresh Rankings' button to silently fail in production:
1. ScrapeRanking was using the plain HTTP client (s.client) instead of the
browserless content client (s.urlClient). The /ranking page requires
JavaScript rendering, so a plain fetch returned HTML without any novel
entries. Now uses s.urlClient so the page is fully rendered before scraping.
2. handleRankingRefresh was synchronous, holding the HTTP connection open for
up to 60 s while scraping. Reverse proxies and HTMX timeouts closed the
connection before the scrape finished. Rewritten to the same async pattern
used for book scraping: POST /ranking/refresh returns immediately with a
polling badge; the browser polls GET /ui/ranking/status every 3 s; when
the goroutine finishes the status endpoint sends HX-Redirect to /ranking.
- Add Kokoro-FastAPI TTS integration to the chapter reader UI:
- Browser-side MSE streaming with paragraph-level click-to-start
- Voice selector, speed slider, auto-next with prefetch of the next chapter
- New GET /ui/chapter-text endpoint that strips Markdown and serves plain text
- Add ranking page (novelfire /ranking scraper, WriteRanking/ReadRankingItems
in writer, GET /ranking + POST /ranking/refresh + GET /ranking/view routes)
with local-library annotation and one-click scrape buttons
- Add StrategyDirect (plain HTTP client) as a new browser strategy; the
default strategy is now 'direct' for chapter fetching and 'content'
for chapter-list URL retrieval (split via BROWSERLESS_URL_STRATEGY)
- Fix chapter numbering bug: numbers are now derived from the URL path
(/chapter-N) rather than list position, correcting newest-first ordering
- Add 'refresh <slug>' CLI sub-command to re-scrape a book from its saved
source_url without knowing the original URL
- Extend NovelScraper interface with RankingProvider (ScrapeRanking)
- Tune scraper timeouts: wait-for-selector reduced to 5 s, GotoOptions
timeout set to 60 s, content/scrape client defaults raised to 90 s
- Add cover extraction fix (figure.cover > img rather than bare img.cover)
- Add AGENTS.md and .aiignore for AI tooling context
- Add integration tests for browser client and novelfire scraper (build
tag: integration) and unit tests for chapterNumberFromURL and pagination
