Repository navigation
Measure Jim's audio and keep his volume - #442
Merged
Merged
Conversation
![cubic-dev-ai[bot]](https://avatars.githubusercontent.com/in/1082092?s=60&v=4){kind=small}
This comment was marked as resolved.
This comment was marked as resolved.
Sorry, something went wrong.
jserv
force-pushed
the
feat/preflight-audio-controls
branch
from
October 23, 2026 07:26
998a714 to
aab6677
Compare
jserv
changed the title
Transcript order and a Speaking pill could pass the browser smoke while the candidate received no audio packets or decoded only silence. Probe the public LiveKit receiver stats for the native track attached as Jim, and separately measure its decoded RMS during Speaking. Keep unavailable observations distinct from observed silence and retain the failed measurements in the capture before refusing the run. Bound receiver waits, sampling windows and subscription counts. Keep counter baselines within one receiver and Speaking interval, discard late results from replaced tracks, and require healthy evidence after resubscription. A generation whose track ends within its first second is superseded rather than judged on a cut-short window, while a longer one keeps its failure. A slow AudioContext resume does not hold up or end sampling. Do not infer physical audibility or word delivery from these counters. Install through CDP so the probe respects the real page CSP without requiring inline scripts or disabling the policy. Validate with controlled counter, silence, reset, timeout, cancellation and subscription tests, plus a Chromium WebRTC tone loopback under the restrictive CSP using the vendored SDK. Removing the late-track and Speaking-gap guards makes their regressions fail. The isolated full gate and final focused checks pass; live LiveKit/Gemini acceptance still needs an operator-selected isolated configuration and budget.
Candidates could route Jim to a speaker but could not attenuate or mute him locally. Remember a validated positive volume and keep mute within the current page, so reload restores the last positive level. Expose a labelled slider, a fixed-name mute toggle, a reset and a persistent muted indicator. A single step up from the muted slider unmutes at the saved level instead of saving 1%. Keep mute available when volume setters are unsupported, moving focus before hiding the unusable slider. Apply preferences whenever Jim's audio attaches or returns after a subscription change. Restore unsolicited SDK volume changes, including an unmute without a playback-status event. Reapply synchronously during Resume audio and after its promises settle so a stalled or stale startAudio cannot leave playback unmuted while the UI claims otherwise. Remove the element listener when its attachment goes away. Expanded audio controls and playback warnings could cover report-wait actions on a short phone screen. Bound and scroll the compact stage, and reserve its measured height in the wait instead of fixed padding. Keep recovery reachable by keyboard inside the clipped stage. Preference, readback, attachment, recovery and DOM checks pass, as do Chromium keyboard, reload and phone-layout checks. Removing either SDK unmute restoration or synchronous recovery reapply makes its regression fail. The full gate covers this and the audio probe together. Physical audible-level and remote Meet-capture observations remain operator acceptance work; the interface identifies that possible effect, and the Meet manual check now resets the volume before recording.
jserv
force-pushed
the
feat/preflight-audio-controls
branch
from
October 23, 2026 07:41
aab6677 to
5301df5
Compare
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.
Candidates could not turn Jim down or mute him, and the browser smoke could pass while no interviewer audio arrived. This adds two changes, one commit each. Preflight microphone and output selection is left to #411, which this PR no longer duplicates.
Interviewer audio diagnostics in the browser probe: LiveKit receiver counters and decoded RMS for the track attached as Jim are measured separately, so missing observations, transport silence and decoded silence are reported as different outcomes. A generation whose track ends before Speaking or within its first second is superseded rather than judged on a cut-short window, and the last generation must pass.
Jim volume and mute: a positive volume persists across reloads while mute stays session-only, a single step up from the muted slider restores the saved level, and preferences are reapplied whenever Jim's audio attaches, reconnects or recovers through Resume audio. The Meet manual check now resets the volume before recording, since tab audio follows it.
After #411 merges, this branch needs a rebase with one deliberate reconciliation: #411 keeps a fresh Jim element muted until output routing settles, while this PR reapplies the candidate's mute on attach and restores it on unsolicited volume changes. Combined naively, each undoes the other, so the rebase must hold the element muted during routing and then apply the volume preferences. The only textual conflict is the Meet manual-check step list.
Verified with scripts/test.sh: 1540 tests pass, including the Chromium volume, layout and WebRTC loopback probe tests. The only failing gate is skill-refs, caused by an untracked skill directory in the local checkout that is not part of this branch. Each commit passes its own suites. Not covered: audible level and a live Meet capture, which need hardware and an operator run.