Carry speaker assignments into the live transcript engine by ComputelessComputer · Pull Request #7286 · fastrepl/anarlog

ComputelessComputer · 2026-09-03T09:29:02Z

Naming a speaker mid-meeting only relabeled the main transcript view through
a JavaScript overlay; the live engine kept emitting unlabeled segments, so
the meeting float and plugin snapshot still showed "Speaker N", and a
listener refresh started a fresh engine that knew nothing about the pick.

The desktop config sync now follows the captured session's newest transcript
and pushes its resolved identity assignments with the participant list. The
engine folds them into segment building the same way the settled render does
(hints first, participant channel defaults after), re-renders the retained
window on an identity update so on-screen segments relabel without waiting
for the next stream response, and seeds a refreshed listener with the same
assignments. Assignment changes never trigger a listener refresh.

Note

Medium Risk
Touches live transcription identity resolution and capture config sync across desktop and listener-core; mis-synced assignments could mislabel speakers in live UI, but behavior is heavily tested and failures degrade to unnamed capture.

Overview
Live capture now receives the same speaker identity assignments as the settled transcript render, so meeting float and plugin live segments show real names instead of generic speaker labels after mid-meeting picks or after a capture restart.

The desktop live capture config sync watches the active session’s newest transcript (words + speaker hints), resolves assignments via the same render path as the main UI, and debounces updateCaptureConfig with speaker_assignments alongside participants and languages. It waits for the first transcript snapshot before pushing (an empty list clears engine-held names), falls back to unnamed capture if the transcript query fails early, and re-pushes when captureGenerationBySession changes on the same session.

On the Rust side, speaker_assignments flows through session params, CaptureConfigUpdate, and listener startup via LiveTranscriptEngine::with_speaker_assignments. Identity updates call update_identities, merge persisted hints with participant channel defaults when building segments, re-render the retained window so on-screen segments relabel immediately, and emit segment deltas without waiting for the next stream chunk. Assignment-only config changes do not force a listener refresh (only language or expected speaker-count changes do).

^{Reviewed by Cursor Bugbot for commit 7f4b649. Bugbot is set up for automated code reviews on this repo. Configure here.}

netlify · 2026-09-03T09:29:24Z

✅ Deploy Preview for anarlog canceled.

Name	Link
🔨 Latest commit	`7f4b649`
🔍 Latest deploy log	https://app.netlify.com/projects/anarlog/deploys/6a9960984c912400083c1158

Naming a speaker mid-meeting only relabeled the main transcript view through a JavaScript overlay; the live engine kept emitting unlabeled segments, so the meeting float and plugin snapshot still showed "Speaker N", and a listener refresh started a fresh engine that knew nothing about the pick. The desktop config sync now follows the captured session's newest transcript and pushes its resolved identity assignments with the participant list. The engine folds them into segment building the same way the settled render does (hints first, participant channel defaults after), re-renders the retained window on an identity update so on-screen segments relabel without waiting for the next stream response, and seeds a refreshed listener with the same assignments. Assignment changes never trigger a listener refresh.

The config signature that dedupes live capture pushes outlived the capture: restarting on the same session recomputed the same config, the push was skipped, and the fresh engine, which always starts with no speaker assignments, never got the names. Scope the signature to the store's capture generation so a new capture always pushes once.

cursor

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 75df9ba. Configure here.}

The wait-for-snapshot gate held every push until the transcript query answered, and a query error or a rejected subscription never opened it, so a failed transcript read also froze participant and language updates for the rest of the session. A failure before the first rows now counts as an empty snapshot and the config runs without names; a failure after them keeps the last known assignments.