Meeting mode

Hold Ctrl+Win and press M, or use the tray item, or the record button on the Meetings tab. Deixis records the microphone and — when it can — the system output, so both halves of a call are transcribed. Press again to stop.

A session lands in %USERPROFILE%\.deixis\sessions\<timestamp>\:

Notes need a cleanup endpoint whose Worker serves /summarize (see Settings, and where cleanup runs); without one the meeting still records and transcript.md is still written — notes are deliberately the last, failure-tolerant step, and the Regenerate notes button retries them.

Each session is named from what was said in it, so a month of meetings is a list of subjects rather than thirty timestamps. Deixis asks a model for a short phrase, from the opening of the transcript only — a local model if you have configured one, otherwise the same /summarize endpoint the notes use. With neither, it falls back to the first substantial sentence somebody said. A name you type yourself is never overwritten, by any of them, ever. Naming is the most failure-tolerant step there is: when nothing can name a session it simply shows its date, which is what every meeting showed before the feature existed.

While a meeting runs, it says so — the tray icon carries a red dot, its menu counts the elapsed time, and the Meetings item in the app's side rail shows a recording dot and the clock from whichever page you are on. The floating bar shows it too, unless you have set it to Hidden.

Worth knowing:

• Wear headphones if you can. Loopback captures what the speakers play, but your microphone hears it too, so on speakers "Them" bleeds into your own transcript. Echo cancellation is out of scope.
• Loopback can be unavailable (exclusive-mode devices, some virtual cards). Deixis falls back to microphone-only, says so, and records it in meta.json rather than failing the meeting.
• Speakers are channels, not people. This version labels "Me" and "Them"; it does not yet split multiple remote voices apart.
• Dictation always wins. Transcription of a meeting yields to any dictation you start, so push-to-talk stays fast while an hour-long recording runs.
• A crash costs at most a minute. WAVs are finalised per segment; rebuild the transcript with deixis-core --transcribe <session-dir>.