The nightly client-deluge profile has failed since 61c7ad3 with "Failed
to
add to deluge: Failed to fetch torrent file: torrent fetch refused:
address
is not public". Shelfmark now fetches a .torrent only from the
configured
Prowlarr/Newznab/Jackett origin or from a public address, and the mock
Prowlarr's release linked straight to http://mock-aa/payload.torrent, a
private address on another origin. qBittorrent and Transmission hid this
by
handing the URL to the client; Deluge needs the file and failed.
Real Prowlarr rewrites indexer links to its own /{id}/download endpoint.
The
mock now does the same: the torznab link uses the host Shelfmark reached
it
on, and /{id}/download fetches the .torrent from mock-aa. Every client
now
gets the file through Shelfmark's own fetch.
tests/download/test_prowlarr_mock_torrent_fetch.py runs the real
extract_torrent_info against the mock in normal CI, so the link cannot
drift
off the Prowlarr origin again unnoticed; it fails on the old mock with
the
same error.
The destination guard from the oceanofpdf/SSRF fix (f837506) trusts only
hosts in
the mirror lists and refuses any other host that resolves to a private
address.
The full e2e profile has failed on every PR since then: its AA mirror
was mock-aa,
but the slow-download links pointed at cf.mock.test (172.30.0.11). The
guard refused
the slow-download page before the internal Chrome bypasser ever started,
so the
download ended with "All download sources failed". Adding cf.mock.test
as a second
AA mirror would not help: AAMirrorSelector.rewrite would then move the
gated link
onto mock-aa, and the test would pass without Chrome solving anything.
The profile
now uses the gate as its only AA mirror. The cloudflare mock takes
CF_GATED_PATHS,
and the full profile challenges only /slow_download/, so search, detail
and the
final file pass through while the download still needs Chrome. The
unused
SLOW_DOWNLOAD_BASE indirection is gone.
The same allowlist missed a supported setup: Primary Mirror
(AA_BASE_URL) may pin a
URL that is not in the mirror list, and network uses it as the active AA
base. With
it on a LAN address, every AA request was refused. The guard now trusts
the pinned
URL too (mirrors.get_pinned_aa_base_url), and
has_aa_mirror_configuration reuses
the same helper.
download_url also ran the guard a second time outside its try block,
which repeated
the DNS lookup and could raise instead of returning None. The duplicate
call is gone.
CaliBrain