Repository navigation
fix: the Flatpak lists cameras again - #173
Conversation
Copied from #170's branch. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Maintainer decision (2026-11-15, /quest-complete, taking the recommendation):
(Written by Claude Opus 5.5) |
|
Maintainer decision (2026-11-15, /quest-complete): reporting the regression to z-galaxy/zbus: ✅ don't post (rejected: post the drafted issue; post it offering a PR). moq keeps its ashpd 0.13 hold (moq-dev/moq#8861) until Ubuntu 24.04 ages out or zbus restores the staged handshake on its own. Remove this item from (Written by Claude Opus 5.5) |
Moves the pin to moq-dev/moq@6c726f455, which holds ashpd at 0.13 (moq-dev/moq#8861). moq-dev/moq#8851 and #8852 dropped the compile-trim features, so moq-audio's bus and moq-tokio's network stack always compile: the studio drops its bus, net, quic, and websocket features. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A replay copy declared every start between its groups skipped, so a group lost on the way read as media running on and no gap reached the player. The publish copy declared none, so a forwarded track lost the skip before the encoder's first group, and its decoder treated that group as a midstream join. Both now repeat the source's skips, the initial run included, and leave a hole a hole. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
No actionable comments were generated in the recent review. WalkthroughThe changes update media writing, forwarding, replay, and playback to use timestamp-based group starts. Forwarding and replay preserve source-declared skips and group ends, while gaps caused by missing groups remain distinct from skips. Timestamp conversions now propagate errors in replay, decoding, and media playback. Dependency settings and timestamped test fixtures also change. Merge Risk:
|
There was a problem hiding this comment.
ℹ️ No critical issues. One minor suggestion inline.
Reviewed changes
I reviewed the whole PR at fb6f3c7: the moq pin bump to a commit on moq's main (checked with the compare API, which reports behind), and the studio's catch-up to moq main's API.
- Pin and features: moves
moqto6c726f455, which includes the ashpd 0.13 / zbus 5 hold. It drops the trimmed-awaybus,net,quicandwebsocketfeature requests. Nothing injustfile,packaging/orREADME.mdstill names them. - Group starts and fallible time:
group::Info { start },append_group(start),close(),Duration::try_fromandTimestamp::unscaledare applied everywhere they're used. The untimed token and probe tracks get keys that only ever increase. - Skips in copies: the forwarded copy in
publish.rsand the replay take inreplay.rsnow repeat only the skips their source declared. A lost group stays a gap instead of a run where the media seems to continue. The source is local andappend_grouprecords its skip under the same lock as the group, so theis_skippedcheck can't miss a skip that hasn't arrived yet. - Replay timescale: a take keys its range in
hang::container::TIMESCALE, and it now refuses a track on any other timescale instead of misreading its range.
quest context could not run here, because the review shell has no quest or nix binary.
claude-opus-5-5 | 𝕏
The publish copy closed every group, so a forwarded source's last group never aged out after it paused and its readers lost the group's end. It now finishes a group at the end its source stated. A replay does the same for a group it holds whole; one cut at the span's end still only closes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
✅ No new issues found.
Reviewed changes
I reviewed 168ca2e, the only commit since the prior Pullfrog review at fb6f3c7.
- Group ends in the publish copy:
copyinpublish.rsnow ends a group withfinish(end)when the source gavegroup.end(), and only closes it otherwise. A paused source's last forwarded group now ages out ofFORWARDEDand keeps its presented end. - Group ends in the replay copy:
copy_heldrecordsendonly for a group it kept whole. A group cut at the span's end, or one still arriving (end()isNone), is still only closed. - Tests:
a_forwarded_track_keeps_its_skipsnow checks one finished end and two closed groups. The newa_take_keeps_where_its_whole_groups_endchecks that two whole groups keep their ends and that the group cut at 2.5 s has none.
claude-opus-5-5 | 𝕏
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The REVIEW page landed against the old pin: timestamps convert to a Duration fallibly now, a timestamp's scale is optional, and groups are addressed by their start. The strip keys groups by start and fails loud on an unscaled timestamp; a test drives it over a real take. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
✅ No new issues found.
Reviewed changes
I reviewed what changed since 168ca2e: the merge of main at 7fa9668 (#177, the REVIEW page) and 87b778f, which moves REVIEW's strip and replay onto moq main's API.
- Strip keys groups by
start:groupsinpast.rsnow keys each group bygroup.startinstead of its sequence, andDecoding.peakskeeps one entry per start. The cache and the "redecode the newest group" rule work as before. - Fallible time in the strip: the take's end goes through
Duration::try_from, so an unscaled end resets the strip with a held error instead of panicking. Theatclosure,as_micros()for the picture cache key, and the sound peaks now returnTimeOverflowas an error, which shows up as the strip'sheldorsounderror. - Optional scale in
replay::before: it readsend.scale()as anOption, with a context error for a timeline that has no timescale, and usesDuration::try_fromfor the end. - New test:
the_strip_shows_each_groups_keyframe_onceencodes 3 s of 30 fps with a keyframe every second, cuts a realTake, and checks that the strip shows keyframes at 0, 1 and 2 s and doesn't decode them again on a second sample. It needsopenh264, whichjust testenables through thefeatureslist, so CI runs it (theCheckjob was still pending when I reviewed).
claude-opus-5-5 | 𝕏
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Brings in BUILD's layers and inspector (#174), Studio mode (#176), REVIEW and Settings' Recording tab (#177), the moq pin and Flatpak camera fix (#173), and the quest cleanup (#185). The copied source tracks take main's group starts and skips while still stopping at END STREAM; BUILD's picture frames the program by whether it is on air and a scene open off program as main does; Settings has both Stream and Recording; and the hotkeys add Go live and End stream to TAKE, AUTO, and Studio mode. Records the maintainer's choice for program-encoder-priority.md: release the sources' replay encoders when the program's cannot open. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Fixes the Flatpak's camera listing by moving the moq pin to a moq
mainthat holds the portals on zbus 5 (moq-dev/moq#8861), catches the studio up to moq main's API, and completes the questquest/m0/packaging/flatpak-cameras.md.Root cause
ashpd0.14 /zbus6 in build(deps): bump the cargo group with 18 updates moq#8695, a dependabot bump. feat!: catch the studio up to moq's main #163's catch-up brought that in.AUTH,NEGOTIATE_UNIX_FD,BEGIN,Hello) in one write.BEGIN. The bus'sOK/AGREE_UNIX_FDstill on their way back are then parsed as message headers:Invalid message header read, the proxy hangs up, and the client seesD-Bus handshake failed: Unexpected EOF during handshake.FLATPAK_ID, a workaround zbus 6 removed.The surface isn't supported by this adapterpanic is not a second bug.smoke.shfails on the camera listing and kills sway while the studio is still opening its window (ERROR_SURFACE_LOST_KHR6 ms later). With the fix, the window opens and closes cleanly.Changes
just upgrade-moqto moq-dev/moq@6c726f455, on moq'smainafter #8861 (ashpd0.13, a dependabot ignore, and a test that plays a 0.1.5 proxy). The studio's ownashpdis already 0.13, so the tree keeps onezbus(5.19).busandnetrequests and its ownquicandwebsocketfeatures, which only forwarded to moq-tokio's; nothing gated on them.validate/flatpak-camerasscratch branch): feat(moq-net)!: address groups by their start moq#8638's groups addressed by their start and the fallibleTimestamp, and feat(moq-video)!: capture listings and the display selection can grow moq#8742'sDisplaySelection. The replay now refuses a track off hang's timescale, since its range is keyed before the track describes itself.is_skipped(0..start). The replay copy now repeats only the skips its source declared, so a group lost on the way stays a gap instead of freezing the last picture. The publish copy, which declared none, now repeats them too, so a forwarded source keeps its encoder-first marker and its continuity. Tests:a_forwarded_track_keeps_its_skips,a_take_keeps_the_skips_it_was_given, anda_take_keeps_a_gap, whose lost group the old fixture had declared skipped.a_take_keeps_where_its_whole_groups_end, anda_forwarded_track_keeps_its_skipschecks ends too.past.rs) and replay (replay.rs) landed against the old pin. The strip keys its cached groups bystart, and its timestamp-to-moment mapping fails loud on an unscaled timestamp instead of panicking or guessing;replay::beforereads the end's now-optional scale with an error. Test:the_strip_shows_each_groups_keyframe_once, which drives the strip over a real take, where nothing covered it before.quest/m0/packaging/flatpak-cameras.md.Decisions
quicandwebsocketfeatures are removed rather than kept as empty no-ops, since moq-tokio no longer lets a build leave either transport out and nothing in the studio gated on them.Validation
I built xdg-dbus-proxy 0.1.5 from its release tarball and ran it in front of a private
dbus-daemon, filtered asflatpak rundoes. WithFLATPAK_IDset, I called the camera portal as moq-video does:Ok(present=false)Invalid message header read, then main's CI error verbatimOk(present=false)Ok(present=false)End to end on Ubuntu 24.04: the Flatpak workflow dispatched on head
87b778f, after merging the REVIEW page, passed on x86_64 and aarch64, behind the runner's xdg-dbus-proxy0.1.5-1ubuntu0.2, as it had on168ca2e(run). Both smoke logs showlisted devices cameras=Ok(0) displays=Ok(1) windows=Ok(1) audio=Ok(7). The headf5ecf28adds only main's quest-file changes from moq-dev/studio#185. It runs only on dispatch here, since this PR touches nopackaging/flatpakpath.just cipasses on head87b778f(351 tests), andquest checkonf5ecf28.Caused by
Bundlejobs)Follow-ups
Notes
Models
Size: L
Implementer: Claude Code, claude-opus-5-5, unknown effort (iterate, L)
Implementer: Claude Code, claude-opus-5-5, unknown effort (iterate, S)
(Written by Opus 5.5)
🤖 Generated with Claude Code
Review
Rewritten by
quest reviewafter each round.Head: 87b778f
Base: origin/main at 7fa9668
Risk: medium,
the base has no RISK.md to name what is always high or customer-facing, so every change is mediumClassified: medium by default
Reviewer: Codex, gpt-6.1-sol, xhigh effort
Session:
01a1e359-0b8f-7910-b7c7-baec7b7603cd, round 4Score: 10/10,
The incremental REVIEW adaptations consistently use the new timestamp and group-start APIs, with no new findings from code review; builds and tests were excluded by the review rules.Rounds: 4; first round 2 (medium 2); findings 2 (medium 2): fixed 2, refuted 0, open 0
Findings: none
Reviewer: Codex, gpt-6.1-sol, high effort, shadow
Session:
01a1e359-0b90-7ef1-8674-f95922368f2aRounds: 2; first round 1 (medium 1); findings 1 (medium 1): fixed 1, refuted 0, open 0