nodecode-meeting
Meeting Overlay: the live meeting transcript and the records saved from it, as tools. A Nodecode rewrite of taxbax/meeting-overlay by taxbax.
Get it
Tell the agent "add the meeting cell", or run
nodecode add nc://meeting
Set it up
No key, no account, no network. The cell reads a transcript directory and writes records; both have defaults, so it works the moment it is on.
The section, as ~/.nodecode/config.jsonc carries it:
| Setting | Default | What it does |
|---|---|---|
dir |
~/.nodecode/meeting |
where rolling.txt and full.txt are read from and appended to |
records |
(empty) | where a saved record goes; empty means Records/ inside dir |
window_chars |
20000 |
how much of the newest transcript rolling.txt keeps |
stale_seconds |
30 |
how old the transcript may be before status calls the capture dead |
Point dir at whatever your transcriber already writes, or let the cell be the one that writes it. The line format is the original's: [HH:MM:SS] [mic] what was said.
Use it
Ask the agent, or call the verbs from eval:
(meeting:note-line "we should ship the patch on Friday" :source "audio") (meeting:transcript :lines 30) (meeting:status) (meeting:save "Weekly sync" :notes "ship Friday; budget approved")
The four verbs are also the four tools the model calls — meeting_append, meeting_transcript, meeting_status, meeting_save — and /meeting prints where capture stands (/meeting save TITLE writes the record).
- A transcript.
meeting_appendtakes one line at a time, taggedmic,audio(the other side of a call),phone, or a tag of your own.rolling.txtkeeps the newestwindow_chars;full.txtkeeps everything. Anything that can append a line can feed it — a transcriber, a script, or you, dictating. - Reading it.
meeting_transcriptanswers the newest lines with the transcript's age, and saysSTALEwhen it is older thanstale_seconds, so an answer is never invented from words spoken an hour ago. - Records.
meeting_savewrites one markdown record underRecords/YYYY/MM/week-YYYY-MM-DD/: the title, the notes, and the transcript. The tree is the original's, so an Obsidian vault of meeting records reads the same. - An echo. The same words arriving twice — your mic and the call's own audio — are one line, not two; the phone's mic is the preferred source and always wins.
What is in it
| File | What it is |
|---|---|
cell.lisp |
the section, the four verbs, the four tools, /meeting, the primer |
package.lisp |
the MEETING package, so the verbs are callable from eval |
test/ |
offline tests: a scratch transcript directory, the four tools through the real dispatch, the refusals |
config.example.jsonc |
the section, commented |
What the original had and this does not: the macOS whisper daemon and its BlackHole audio routing, the launchd agents, the phone PWA and the desktop pane. Transcript and records are what a cell can carry; on macOS the original's listen.py can still be the thing that appends the lines.
