vendor/ was seeded by copying af_readability's tree, so both plugins shipped
ComposerAutoloaderInitb44cc79a… and ComposerStaticInitb44cc79a…. tt-rss's
Preferences page loads every plugin on disk regardless of whether it is enabled,
so the second require redeclared those classes and fataled the request.
tt-rss blames whichever plugin lost the race -- the reported error named
af_readability, which was the innocent party.
Verified both autoloaders now load in the same process and the vendored
Readability and Masterminds classes still resolve.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011sSJdftQx5bW5HZHgUKF3i
The community rule matches only div.usercontent; journal styles vary and
diziet's (which arrives via Planet Debian) uses div.entry-content, so those
posts fell through to Readability. The upstream file's own comment notes it
needs converting to a wildcard.
Found by the stale-rule diagnostic after a full re-import.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011sSJdftQx5bW5HZHgUKF3i
The community rule pins the class to 'aligncenter size-full wp-image-'. Posts
vary between aligncenter and alignnone, so it matched some articles and fell
through to Readability on the rest -- which on a comic means the strip vanishes
and the article still looks plausible.
Found by the stale-rule diagnostic against live data: 8 stored articles carried
"STALE" while a hand-tested article of the other alignment extracted fine.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011sSJdftQx5bW5HZHgUKF3i
Replaces subscribing feeds through a self-hosted Full-Text RSS proxy. Feed URLs
go back to being real feed URLs and extraction happens inside tt-rss, using the
same ftr-site-config rules the proxy used.
The engine in lib/ has no tt-rss dependencies, so rules can be developed and
audited from the command line; init.php is a thin adapter over it.
Two findings from measuring the real subscription first, both of which shaped
the design:
- Firecrawl's own onlyMainContent is far too coarse to extract with (73KB of
chrome on a Cloudflare post), but it is an excellent renderer. So it is used
for rawHtml only and the rule engine does the extraction.
- A body rule that stops matching after a redesign falls through to Readability
and still produces a plausible article, so the breakage is invisible. Every
extraction now records which rule matched and whether it fell back; auditing
the 34 live feeds surfaced five community rules that match nothing.
Custom rules included for the sites that needed them, including three comics
where the article is an image and text-scoring extractors return the wrong thing
or nothing at all.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011sSJdftQx5bW5HZHgUKF3i
