Repository navigation
Conversation
page-fetch.e2e.mjs saves the videos only the page can load through the Save button and the name prompt: a service worker's video and a current-cookie video are the film byte for byte (SHA-256), their far end asked for through the page. A save cancelled about a quarter in says "Save cancelled!" (no "failed, save what there is?" question), leaves no file, and the next save is whole. Each save is of a video no earlier test played (a player takes over what the previous player of the same video loaded), and the film is sent no-store (Firefox's cache answered ranges an earlier test had loaded). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… screen readers
- Space presses a control the keyboard went to, as it does a <button>
(WebUtils.setupTabIndex). After a mouse click the control has the
focus without :focus-visible, and Space stays play/pause.
- The six menu buttons (audio, subtitles, quality, language, speed,
loop) say aria-haspopup and whether their menu is open
(aria-expanded, WebUtils.setExpanded in each openUI/closeUI).
- The seek bar is a named slider in the tab order ("Playback position",
a new string in all 16 locales), its aria-valuemax/now/valuetext
written once a second with the time's text; focused, the arrows seek
as everywhere in the player. It gets the keyboard focus ring.
No transition or animation added. Tests: playerA11y.test.mjs (Space,
aria-expanded, the slider's wiring) and specs/player-a11y.e2e.mjs in the
real player (Tab reaches the slider, its values follow a seek, ArrowRight
seeks from it; Space on a keyboard-focused button presses it, after a
click it plays; the subtitles button's aria-expanded).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The options page sets up its fields with WebUtils.setupTabIndex too, and a text field always has :focus-visible: the Space rule of the commit before swallowed every space typed in one (the mpv path, "1 2" in the speed limit became "12", measured on that build in the e2e). Inputs, text areas, selects and buttons keep Space to themselves. Tests: playerA11y.test.mjs (each kind of form control), and player-a11y.e2e.mjs on the options page (a box ticked and unticked once each, a space typed). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
docs/firefox-optimization-ui-review.md: what the two branches do, the research (the watcher's high findings for 153-160 checked against the code, documentId, the replaced player's measurement, the page transport, media APIs, the UI), the backlog with its five labels, the numbers, what was verified and what was not, and the steps of the owner's hand test. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…b.mjs) The player's speed was compared by turns several times (docs/notes/player.md), each time with a harness thrown away afterwards. `node tools/bench-ab.mjs --a main` (pnpm run bench:ab) builds a git ref in a worktree of its own (node_modules linked, same lockfile required) and this checkout, and runs tests/e2e/wdio.bench.conf.mjs: one server per build, the same fixtures, both builds in one e2e Firefox in ABBA order, a warm-up round thrown away, the media optionally on one shared slow line (--rate, --ttfb). A probe added before the page's scripts (browser.addInitScript) records player modules ready, first frame, playing, a seek to 80 s, a seek preview, and counts timer/frame callbacks, DOM changes, frames more than 50 ms apart and dropped frames while playing, paused, after a seek preview, once ended and in an idle player; --where names the places in the player's code that asked for the callbacks. logs/bench/ab-<time>.md/.json: medians, MAD, a verdict only beyond twice the noise and the measure's floor (tests/unit/benchStats.test.mjs). It never fails on a number. --clean removes its worktrees, the node_modules link first. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…oop while paused or ended Both found with tools/bench-ab.mjs, on main as well: - play() on an ended MP4 seeks the element to 0 by itself, past MP4Player's currentTime setter, the only caller of resetHLS. mp4box went on from the end, the start had left the buffer, and the video stood seeking at 0 for good: the play button after the end of any MP4 longer than about 40 s. Firefox's picture-in-picture seek controls seek the element the same way. onElementSeeking (the element's seeking event) checks such a seek as the setter does (reloadUnlessBuffered); one made through the setter is recognised (ownSeek) and not checked twice. - The main loop ran every 50 ms while paused, in each MP4 player (the seek preview is a second one, always paused), and every 1 ms once the video had ended (readyState below 3 counted as urgent): 41 and 250 timer callbacks a second for as long as the tab stayed open. Paused or ended it runs every 250 ms now, and a seek or a play wakes it at once (wakeLoop). Measured: 9.3 a second paused and ended; a seek while paused lands in 62-109 ms. tests/e2e/specs/mp4-loop-pace.e2e.mjs fails on main (41 a second paused; replay stuck at 0, readyState 1) and passes here. The whole web e2e suite: 48 of 48 spec files. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
…found Three bugs on main, each checked against the code and shown by a test that failed before: - The next episode's continuation (autoplay, the carried fullscreen, no remembered place) stayed on the tab until some player took it: a next page with no player, a page the user went to himself, and the video opened there minutes later went fullscreen, played by itself and skipped its place. It holds only for the first page after the click now (continuationArrived; MPV's mpvNext is unchanged), and a next episode on another hostname keeps it (the new site's reset dropped it). - Escape in the page could not end a carried fullscreen whose player's frame went or reloaded: the background passed it to that frame only. The player answers whether it ended it (exitCarriedFullscreen returns true); otherwise the background does. - The page around an embed kept its own menu hidden after the player left the embed while the iframe stayed: only the player's leaving page released the overlay guard. A player frame's FRAME_REMOVED now tells each frame above to let go of the iframe on the way (RELEASE_OVERLAY_GUARD), unless another player lives under it. Tests: backgroundFullscreen (continuation to the next page only, on another site too; the exit when the player is gone or new), backgroundOverlayGuard, contentScript (the release), content-overlays e2e (the embed's page goes elsewhere: fails on main, passes here). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…for an audio-only preview What a review found in the MP4 fixes before: - ownSeek was set by the client's start, made at readyState 0, where no seek runs and no seeked ever clears it: a video played to its end without a seek still stood seeking at 0 when played again. It is set only when the element started a seek now. - An audio-only MP4's seek preview made no SourceBuffer (a preview plays no audio), its element stayed at HAVE_NOTHING and its loop waited for a first frame every 4 ms for the tab's life (236 callbacks a second). Such a preview stops once the metadata is in (nothingToPreview): 6.7 a second. mp4-loop-pace.e2e.mjs: both new tests fail on the build before and pass here. docs/notes: these, the background fixes, and what the review found and was not changed yet. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s own files On the slow line, each of the player's modules waited 150 ms too, and every start took 4.5 s in both builds: an extension's files come from its package at once. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Three alerts on #675, all in the A/B harness (development only, nothing ships): - Remote property injection: a scenario name from BENCH_SCENARIOS named a key of the results. The names now come from SCENARIOS itself; an unknown one is a failing test of its own. - Resource exhaustion: BENCH_TTFB_MS set a timer's delay unchecked. Every number the harness reads from the environment (rounds, warm-up, rate, first byte, phase lengths) is refused outside its range now (boundedNumber, unit-tested). - File system race: the build stamp was checked with existsSync and written later. It is read and must name the commit now, which also rebuilds a tree whose stamp is another commit's. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… calls for a closed tab Two more of the review's findings, each with a test that fails before: - A player inside a frame that went (an embed's page left, with the player in it) is forgotten with that frame. Its own FRAME_REMOVED can come after: its entry in playerPageFrames was still there, so the background took it for a player lost while Firefox had unloaded it (restorePlayerFrame), made it and its page frame again and counted it gone a second time. With two players in a tab, the next one to open was then told it was the tab's main player. Every frame forgotten with the one removed now leaves playerPageFrames and playingPlayers. backgroundFrameRemoved.test.mjs. - The toolbar button's calls (badge, title, icon) for a tab closed in between were rejected unhandled: the "Off" badge's second, a start that waited for its page. They go through setAction now, which takes the rejection. backgroundRobustness.test.mjs. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The MPV allowlist's or auto-enable list's start waits half a second for the new page (startWithTrackedLater) and then checked the tab's address was still the same string. A page that sets #t=... or an anchor right away is still the same page (isSamePageUrlChange), which no new start takes up, as the tab is on already: the start was lost, and the stream found before the address change went nowhere. It now goes on for the same page; a #/ or #! route is a new page and still cancels it. backgroundMpv.test.mjs fails before. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…oolbar calls) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n from their numbers The 12 Mbit/s run called HLS's first frame 475 ms better on the branch: the medians were that far apart, over twice the MAD. Its rounds were in two modes on both builds - about 1.47 s or 2.0-2.1 s - three slow of five on main, two on the branch; the order says nothing. A verdict now also needs the Wilcoxon-Mann-Whitney rank-sum test (exact over every split of the runs, ties at their mean rank; seeded sampling past 200,000 splits) below 0.05: five runs a side reach 0.008 at best, three only 0.1. The table shows p. The measures and the report moved to report.mjs, so `node tools/bench-ab.mjs --render <name>` prints an earlier run's report from its .json in logs/bench with the verdicts as they are now. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Twelve places forward and back (one before the last in the same part of the file), a sweep of quick moves and a move within a part, paused and playing: each must show (not seeking, readyState 2 or more, at the place) within 10 s. The A/B benchmark had timed out once in 21 waiting for the preview at 120 s, on main's player and on this branch's alike; its check looked only at 'seeked' and 'loadeddata'. Shown in 29-137 ms on this branch, 57-169 ms on main. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ts alone About one run in 20 timed out waiting for the preview at 120 s, on main's player and the branch's alike, while the preview did show (mp4-seek-preview.e2e.mjs: 340 places of 340 on both builds, 10 runs each): the frame can come after 'seeked' with no other event of the two checked. canplay and timeupdate count too, and the wait checks the state itself. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
mp4-loop-pace's audio-only test played mp4-shapes/long-audio-only.mp4, which only mp4-loading.e2e.mjs wrote: on a fresh checkout (the test branch's worktree) it failed with "the audio did not load", and a CI shard running it before mp4-loading would too. The ffmpeg helper that two specs each had a copy of is one module now (ensureMp4Shape, ensureLongAudioOnly), and mp4-loop-pace asks for its file itself. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…after a DRM refusal What the user saw: in MPV mode on a copy-protected video, the toolbar's On -> MPV (or a start by address) was refused with "!", and then the page's next stream (an ad or a piece of the film) went to mpv, the "!" was cleared and the page paused. Cause: openMpvWithSources had its own copy of the allowlist's hand-off and re-armed the once-per-page latch after every failure, a DRM refusal too; since #606 autoOpenInMpv keeps it after one. Fix: openMpvWithSources sends through autoOpenInMpv, so there is one latch handling. sendPlayedToMpv (the MPV key's play) keeps re-arming after a DRM refusal on purpose, now said in a comment and pinned by a test: kept, a repeat play only paused the page. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
What the user saw: nothing; a private window's playlists and the first 64 KB of its files, read for their length by the background, stayed in the normal disk cache after the private window closed. Cause: the background is no private window, and its length reads (StreamLengths.get) passed credentials 'omit' but no cache option. Fix: cache 'no-store', as MpvSubtitles' reads already do for the same reason. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With the silence skipper on, D, S and the speed preset keys changed the speed for a moment only: the OSD showed the new speed, and within 100 ms the skipper's loop put its regular speed back, which was also what got saved. Only a pick in the speed menu became the skipper's regular (or skip) speed. A speed key now goes through the same step as a menu pick (chooseSpeed). Speeds the player holds for a moment (holding the video, a GIF's 4x) and the skipper's own switches do not, so they never become the regular speed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Every seek (arrow keys, frame step, a click on the timeline) and every yield to a watched player cancelled the running downloads ahead, nearly finished ones too, which then came again from their first byte. The half-done check read the downloader's stats, which were null until FetchLoader's progress, and FetchLoader gives that once, after the whole body. The downloader now takes FetchLoader's own stats as its request goes out (loaded and total as the body comes), and both callers share one check. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…'s page What the user saw: nothing; a refused stream in a private window or a container tab had its page fetched again by the background with the normal window's cookies. What the page set landed in the normal cookie jar, a container's own login was not the one sent, and the request missed a private window's proxy. Cause: REQUEST_FRESH_LINK fetched the page (credentials 'include') with no check of the sender tab's window or cookie store; the background is neither private nor in a container. Fix: answer no link for a private window's tab or a container tab. The player there handles the refusal as before FreshLink. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A key bound to "+" (the German + key, the numpad +, Shift+= on a US keyboard) never fired, and two actions on it got no warning. getKeyString names that key by its character, so it is saved as '+' or 'Control++'; actionsForKey split the string at every '+', found an empty key and gave up before the exact match. The modifiers and the key are now read the way keyDisplayName reads them: the leading Meta/Control/Alt/Shift, each with its '+', and the rest is the key. 'Shift+' still names no key. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…is in A broken server or proxy that answers each range with 206, no Content-Range and the rest of the file had the whole rest read for every 1 MB range: held in RAM as one buffer, stored in the RAM budget, and only then cut to the range by MP4Player. The cut-to-range reading covered only a 200 (a server that ignores Range). A 206 without Content-Range is now read from the range's start up to the range's length, and the rest is cancelled. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Nothing called EventEmitter.logAllEvents (the only setter of its debug flag, read on every emit) or EmitterRelay.prependOutput, and nothing sends the channel counter worklet a 'configure' message: its handler called a configure() the class does not have. Removed, with the debug flag. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…s as the tab's player What the user saw: after Firefox unloaded the idle event page while a player played, a second player in the same tab was told it is the main player and played without a click, as the first did. Cause: restorePlayerFrame made the frame a player's again (isPlayer) but did not count it in tab.playerCount, which acceptPlayer's isMainPlayer goes by; its leaving still counted it gone. Fix: count the frame when it becomes a player's, in restorePlayerFrame and in acceptPlayer alike, so a frame restored and then loading its player anew is counted once, not twice. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… the player The player's settings page is an iframe in every player, and it loaded with the player: its scripts ran beside the player's own as it started. Measured with tools/bench-ab.mjs (2026-11-13, 7 rounds, MP4): the first frame came 137 ms later with it (945 ms against 808 ms without, p 0.004, 7 of 7 rounds); seeks, wake-ups and frame drops the same. It loads now when the page is idle after the first frame, at once when the settings window opens, and after 5 s in a player that shows nothing, so the GitHub build's update check still runs and shows its mark on the settings button. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…yer external links) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
What the user saw: with a video delay set on an MP4, the two extra audio players (and the audio analyzer's player) each buffered and appended the video as well, frame by frame: the MediaSource memory (about 30 s ahead and 10 s behind of video in each) and the per-frame work three times over. Cause: MP4Player never looked at isAudioOnly beyond making an <audio>: it made a video SourceBuffer, had mp4box cut every video frame into a segment, appended it, and chose where to read from by the video's keyframes. Fix: playedVideoTrack() gives no video track to a player for sound only of a file with sound; makeSourceBuffers, startSegmentation and the keyframe index (currentFragment) use it, as the seek preview leaves out the sound. The ranges' times still come from the video track, as the main player's do (the fragment store is shared). A file without sound keeps its picture in such a player, as before; the main player, the preview and an audio-only file are unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… too OPEN_EXTERNAL_URL answered only the player page, but the settings page's report and update buttons now open their pages through EnvUtils.openExternalURL as well, and in a player framed in a page that settings page, an iframe in the player, has no chrome.tabs either: the buttons did nothing there. The two pages are now matched by their exact address, not by its start. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… list What the user saw: nothing directly; each HLS segment hls.js aborted while the client also downloaded it stayed in HLSPlayer.activeRequests for the session (its loader and callbacks kept in memory), and the seek preview, the only reader of the list, took it for a segment still loading. Cause: a loader leaves the list in its download's onSuccess, onFail or onAbort. A download with other watchers tells only its last watcher that it was aborted (DownloadEntry.abortWatcher), so an aborted loader that was not the last heard nothing. Fix: HLSLoader.abort() takes its loader out of the player's list itself. DownloadEntry is unchanged: telling every watcher would reach the other players' and the saves' handlers. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…r fixes (MO-1..9) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The e2e test of the seek past one track's end (PE-2, MP4Player.seekMp4box) played a 20 s cut of the long fixture: the player loads 30 s ahead, so both tracks were appended to their ends before the seek, and the test passed on the old code too. It now plays the first 100 s with the sound cut to 98 s, and checks that neither track is appended to its end before the seek (the old code's stall reproduced with mp4box on a file of this shape, written by the same ffmpeg arguments). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…to a place Part of the fix before (PE-1, previewSeek): a preview reading on to a place past the duration it knew held every range it read until it got there. A player's hold is one the memory budget cannot let go of: a far hover into a long fragmented file without a mehd box could hold most of it in RAM, and stop the main player's downloads ahead. The ranges behind the place held are let go of as the preview reads on, as those behind the playhead are while a video plays; the request loop starts at the range of the last keyframe known before the place, so none is asked for again. The SourceBuffer's own trimming is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> # Conflicts: # tests/unit/FastStreamClient.test.mjs
No caller passes it (FastStreamClient, SyncedAudioPlayer), and DashPlayer, the one player that stored it, no longer does. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…layer reads DashPlayer, the one player that stored it, did not read it either (fix/vendor removed it). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
What the fix branches changed, recorded where a maintainer looks for it: keybinds.md (speed keys with the silence skipper, Escape, Space and Enter, a malformed settings file, the "+" key, save echoes), storage.md (the hand-off's files, a save's disk copy after its download, SaveFile in place of mp4merger's slices, FreshLink and the length reads in private and container tabs), player.md (206 without Content-Range, the 256 GiB cap, MP4Player's seek, preview and sound-only fixes, half-done downloads, the A-B and volume loops, partial-save lengths, OPEN_EXTERNAL_URL, the settings page after the first frame, the review's second round), mpv.md (one latch for every automatic hand-off), build-and-release.md (the update check, instagram_inject.js) and libraries.md (dash.js's segment cap). Statements the fixes made wrong are corrected in place, with the date. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…675) CodeQL flagged three of the review's new tests. loopMenuFrames ran LoopMenu's source through new Function to step around the generated gif.mjs: it imports the class now, gif.mjs and the player's elements stand-ins (vi.mock), as `pnpm test` syncs the generated modules first. visibilityAction ran handleVisibilityChange's source: it checks the switch's cases against the actions VisChangeActions declares instead, which fails for the PIP case the fix removed. playerFieldKeys matched comment lines with a nested repeat that can backtrack: it compares the listener's first statements with its comments left out. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ree has (1.7.4, 5.3.0) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ocus
Space presses a control only when the keyboard went to it (WebUtils.setupKeyPress, since the
review's fixes): the test's plain focus() after a mouse click is no keyboard focus in Firefox,
and Space played the video instead (CI, Windows playback 2/3). focus({focusVisible: true}) is
the focus Tab gives.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each entry of DOMElements.mjs is cast once where it is found: HTMLElement for the divs, HTMLIFrameElement for the options frame, NodeListOf<HTMLElement> for the status messages. Every element is in index.html from the start. The 437 casts at the use sites that only repeated this (to HTMLElement, Element, ?HTMLElement, and HTMLButtonElement on moreButton, a div) are gone, and one on a status message read from the typed list. The code's own null checks stay. Comments only: the minified code is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
types/player-window.d.ts adds the player's client to Window, optional (main.mjs deletes it at beforeunload, and other pages have none), so window.fastStream type-checks directly. The 26 casts of window that the code's own checks make needless are gone; the 13 in recieveSources() and loadOptions(), which read the client without a check, keep FastStreamWindow, now defined from the declaration. Comments and a declaration only: the minified code is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…sts fewer, same code) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Firefox leaves fullscreen as soon as another tab comes to the front, and sendToPageInFront switches to the extension's tab to read each answer: the first toggle's fullscreen was over before the test looked, and the second toggle could not be tested. Both toggles now go out on timers from the extension's tab while the site's tab stays in front, and the page records each fullscreen change. Checked against the old compare (the document's fullscreenElement): answers enter, enter; seen host - the test fails there. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…as on main The review's PC-1 fix made the "spare a download at least half in" check work (its stats were null until the whole body was in, so it never held). The A/B of the finished test branch then showed a seek to 80 s in an MP4 on a 12 Mbit/s line a second slower, 1557 -> 2547 ms in 7 of 7 rounds (local files: no change): the spared downloads, outside the new place's window, shared the line with the ones the seek waited for. The same build with only the check off seeks in 1543 ms (2370 -> 1543, 5 of 5 rounds, p 0.008). The check is gone from cancelIfCheap and setYield, and StandardDownloader is as on main. The tests say the measured choice: a download 900 of 1000 bytes in is cancelled by a seek and by a yield (both fail with the check back); the two older tests that set stats by hand to keep one nearly done now expect it cancelled. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Set on the element alone, the rate went back to the client's 1x at the next
canplay (FastStreamClient's CANPLAY handler), which comes again after every
stall: on a busy runner the 160 s clip could not end within the 60 s wait ("the
video did not reach its end at 8x", CI Windows and once locally, each passing on
its retry). The state with its rate now goes into the message if it ever times out.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…page-fetch says where a save stood
page-qualities ("switches to another where the video is", from 2026-10-24) stopped looking
as soon as the time moved on, before its watcher had noted any 'playing' or 'timeupdate'
of the new video (firstPlayed -1, CI Linux, passed on its retry). It now waits for a
noted time as well; the assertion is unchanged.
page-fetch's worker save once never began on the Linux runner (no banner, no download,
the player never logged its save; passed on its retry). The cause is not in the log, so
a timeout now reports the open dialog and the save manager's askingToSave and
makingDownload.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…d the owner's decisions Sections 4-8 rewritten: the final A/B of main against the test branch (local and a 12 Mbit/s line, 7 rounds), the whole-code review's fixes by area with the behaviour changes marked, the half-done check measured and taken back, the workflow hardening, the type cleanup, sizes, local and CI verification with the flakes looked at, the manual test steps for the test branch, the merge order with the rename, and the behaviour questions for the owner. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Nawid3333
changed the title
Owner
Author
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Draft, not for merging yet: this waits for the owner's test and decision. Merging it would release, because a push to main releases.
The full report is in
docs/firefox-optimization-ui-review.md: research, backlog, numbers, the review's fixes, verification, the manual test steps, the merge order and the owner's decisions (section 8). The watcher is #674, the rename #676. All open PRs except the mpv update #634 are merged together ontest/all-prs, the build to test.What it changes
ext-specs/page-fetch.e2e.mjs): a service-worker video and a current-cookie video are saved byte for byte; a cancelled save leaves no file.aria-haspopup/aria-expanded; the seek bar is a named slider ("Playback position", in all 16 locales) with arrow keys and a focus ring.tools/bench-ab.mjs): two builds by turns, slow-line option, rank-sum verdicts.gh;update-prs.yml's concurrency on its job.window.fastStreamtyped once; 1,384 → 1,001 casts, the minified code of all 22 files the same.Numbers (main → test branch, 7 rounds, Firefox 160)
SaveFile.mjs). Permissions unchanged;instagram_inject.jsis web-accessible to Instagram only.Checked
build:keep,lint:amo/lint:github(0/0),verify:vtt,verify:fsaunpack, workflow tests in WSL.🤖 Generated with Claude Code