Diagnosed against a fixture rebuilt from the live yummyanime.tv page, with the
ad churn the real site produces. Under that churn the resolver reported
frame=0, hasVideo=false while the video demonstrably existed two frame levels
down: one ad slot tearing down mid-call makes Chromium reject the whole
allFrames sweep, and the resolver then silently fell back to the top frame and
never looked again.
v3.1.2 did not have this failure because webNavigation.getAllFrames() gave it
an explicit frame list. That list is now rebuilt without the permission: every
content script that messages the background carries sender.frameId, so the
background keeps a per-tab registry of frames it has seen and the resolver asks
any frame the sweep missed directly. One rejected probe now costs one frame
instead of the whole page.
Two supporting fixes fell out of the same investigation. Frames reported hidden
by an ancestor that could inspect them directly — the 0x0 same-origin wrapper an
anime host parks unwatched mirrors in — are now excluded without waiting for the
postMessage visibility handshake, which was the tie the resolver kept failing to
break. And leaf frames with no video and no nested frames are left out of that
handshake entirely, so a churning ad slot can no longer make every phase wait on
a frame that is already gone.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Reproduced against the live yummyanime.tv layout, which is:
top (no video)
├── visible same-origin wrapper 830x498 -> cross-origin player
├── hidden same-origin wrapper 0x0 -> cross-origin mirror
└── hidden cross-origin trailer 0x0
Three things kept that page from ever settling on a target.
Equally-ranked players were a hard failure. Several mirrors or dubs loaded at
once is an ordinary layout for these sites, and refusing to activate made them
unusable. The resolver now holds the top frame and waits for one of them to
start playing, which is the signal that breaks the tie.
Inconclusive probes moved the target. A page whose players are still loading
resolves differently from one call to the next, and every difference triggered
a full teardown and reinjection, so activation never finished — the popup sat
on "activating" with nothing in the log. A probe that finds no video now leaves
the target where it is.
The visibility handshake expired mid-probe. Its listener lived 1000ms while the
probe sequence is six separate executeScript round trips; on a heavy page it
was gone before the answer arrived, leaving every frame's visibility unknown —
the exact state that makes two players look equal. It now outlives the sequence.
A settled failure also no longer reports itself as "activating".
Covered by two fixtures built from the real page: one where the player exists
up front, and one where the host only creates it on play, asserting the target
is promoted into the deep cross-origin frame without touching the popup again.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Google Drive and YummyAnime host their player in a cross-origin iframe. The
3.1.2 targeting work reached those frames but misdiagnosed and destabilized
them in four separate ways. No manifest permission is added or restored;
webNavigation stays removed.
Access diagnosis was inferred, not measured. Every frame probe error was
swallowed, and any origin that failed to answer was reported as missing host
access. A slow or still-loading player frame therefore produced
"Host access required for youtube.googleapis.com" for an origin the extension
already held. The resolver now asks permissions.contains() before raising an
access error, and treats a granted-but-unresponsive origin as a retry, not a
user decision.
Probes were unbounded. Every executeScript in the resolver now runs under a
timeout, so one unreachable frame can no longer stall an activation, and the
retry budget drops from eight passes to three.
The chat overlay followed the player into its frame, which rendered it on top
of the video and scoped closing and minimizing to that frame. It is now always
installed in the tab's top document, with all chat traffic routed to frame 0,
while only the playback controller goes into the selected media frame.
Nested targets reactivated continuously. Every heartbeat and content event
revalidated the target with a full teardown and reinjection, and the media
monitor treated ordinary play, pause and buffering as frame layout changes.
Both paths now reactivate only when the selected frame or document actually
moves.
Also restores the audio-route retention that keeps a deselected tab audible:
createMediaElementSource() can only be called once per element, so a
reinjected content script must adopt the existing route rather than rebuild it.
Verified with 90 unit tests, 40 browser E2E tests including two new
Drive-shaped fixtures that assert the controller lands in the player frame
while the chat stays in the top document, and npm run verify.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The evidence for the frame-observation fixes was a reimplementation of the
logic measured in a browser console, not the shipped code. Two extension
specs now cover the observable behaviour end to end: a player frame that
swaps its document, and the same one level deeper.
The nested case fails against the pre-fix content.js and passes now. The
top-level case already passed before the fix, so that fix removed dead
observer registrations without changing what a user could see; recorded
here so the distinction is not lost.
Also adds bench-finder.mjs, which measures the extracted shipped finder
instead of a transcription of it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The weighted score summed incomparable units, so size could outvote traits
that disqualify an element outright. Measured on a real page: a display:none
preload reports its full 1080p intrinsic size and scored 2073600, beating a
visible unmuted player at 509920.
Selection now compares an ordered list of signals, highest priority first:
has a source, is rendered, is not a silent background loop, rendered size
bucket, is playing, has controls, duration. Rendered size replaces intrinsic
resolution, and mute state is gone from the ranking entirely: it is a viewer
preference, not evidence about which element is the player.
It stays a ranking rather than a filter, so a page of only bad candidates
still yields one and findVideo never returns null where a video exists.
The new tests/e2e suite runs the shipped finder against real fixture pages
and drives the packed extension for injection, reinjection and remote
play/pause/seek into a first-party frame. All five scoring scenarios fail
against the previous implementation.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Timo