Skip to content

Voice channel session logging

How to surface "what's happening in the voice channel right now" in the paired text channel without flooding it with new messages.

Status: Design

Not yet shipped.

Motivation

While the familiar is active in a voice channel, operators (and other participants) want to see what's being said: live transcripts, the bot's replies, session status. The naive approach — one Discord message per transcript line — drowns the text channel in noise within seconds. Discord rate-limits also kick in long before a busy conversation settles down. We need a layout that shows a rolling "what's happening right now" view and a scrollable transcript, without putting either on the main channel feed.

Sketch

Combined approach: live-edited embed + thread

  1. On voice join, the bot posts one embed message in the paired text channel. That's the only top-level message the session creates.
  2. A thread is opened from that message, named with the session timestamp (Session — 2026-04-11 21:07). All transcript entries and session events go into the thread — the main channel stays clean.
  3. The embed is live-edited to show current session state: who's speaking right now, the last exchange, session duration, and message count. Edits are debounced (batch for ~2 seconds) to stay well under Discord's ~5 edits/sec rate limit.
  4. The thread receives append-only entries: timestamped transcript lines, bot responses, notable events (join, leave, interruption, Twitch event acknowledgement).
  5. On voice leave, the embed is updated to "Session ended — {duration}, {message_count} exchanges", and the thread is archived.

This keeps the main channel to one message per session while preserving full history in the thread. The session logging writer pass reads the thread content at session end to generate its memory summary.

Alternatives considered

  • Single pinned message, live-edited, no thread. Simpler, but loses history the moment it's overwritten. Suitable only if persistence moves entirely to the memory directory and nobody needs to scroll back in Discord itself.
  • Thread-only, no top-level embed. Skips the live-state view. Operators then have to open the thread to see if the bot is still active, which is a worse UX.
  • One message per transcript line, no embed, no thread. Rejected — floods the channel and hits rate limits immediately.

Implementation notes

  • Associated text channel. py-cord's VoiceChannel.text_channel property gives the paired channel in stage / voice channels (a newer Discord feature). For regular voice channels without that property, the operator configures a text channel ID on the channel config sidecar.
  • Edit debouncing. Buffer embed updates for ~2 seconds before flushing to avoid Discord's edit rate limits.
  • Thread message format. Short, regular, timestamp + speaker + text. No distinction between voice and text input:
    • [12:34] User: "can you tell me a story"
    • [12:34] Bot: "Once upon a time..."
  • Inline image noting. Images sent during the session are noted inline in the thread: [12:34] User: "what do you think of this?" [image: screenshot.png]. Vision input handling is a Voice input future-work item.
  • Privacy opt-out. Sensitive content (private conversations) should respect a configurable opt-out so logging can be disabled per-familiar or per-channel. Likely lives on ChannelConfig as a session_logging: bool field.

Non-goals

  • Logging outside Discord. This page is only about surfacing the session inside the Discord text channel. Persistent on-disk memory of the session is a separate concern — see Session logging.
  • A "what happened while I was away" catch-up feature. Re-opening the archived thread is the catch-up feature.

Open questions

  • Thread auto-archive behaviour. Discord auto-archives threads after inactivity (1h / 1d / 3d / 1w depending on server boost level). Do we explicitly archive on voice leave, or let Discord do it? Explicit is cleaner for the "session ended" marker.
  • What to show in the embed. A minimal set: "currently speaking: X", "last utterance: ...", "last reply: ...", "session started: ...", "exchanges: N". More may clutter; less may feel stale. Measure in practice.
  • Multi-session overlap. If the bot joins a different voice channel during an existing session (or is restarted mid-session), what happens to the old embed? Probably edit it to "interrupted / ended".