ListeningKitVoice notes
Every voice note matt was sent, or sent himself, on X, Discord and Slack is transcribed with AssemblyAI and written into the conversation store. log search, voice and the SubQ matrices then see what was said instead of a bare file name.
What it looks like in the store
Before, an X voice note was only its marker, and a Discord one was an empty message:
(audio: 9035F3EB-8899-42CE-8BF7-E6AD2C4F152E.m4a)After log transcribe:
(voice note: "<what was said, word for word>")How audio is detected
Once, for every platform, in the shared message layer: python/platformKit/media.py for the engines and src/platformKit/media for the commands (rule 126). An attachment is audio when its content type is audio/* or video/ogg, its extension is .ogg .oga .opus .m4a .mp3 .wav .aac .webm, or the platform says so itself: Discord's voice-message flag (8192) with a duration, or a waveform; X Chat's audio media; Slack's slack_audio file subtype. A video or image type is never audio. Every platform's messages keep their attachments, and log new platform writes that shape into a new engine; check:platforms refuses an engine whose messages drop them.
Where each voice note comes from
| Platform | How log finds it | How the file is fetched |
|---|---|---|
| X | The (audio: <name>) marker in the saved X Chat conversation, matched to its media in the account's X Chat store (inboxSync.json) | Already saved by the X Chat read, or fetched through the account's one X session: the conversation is read again with its audio events reopened (rule 105) |
| Slack | A file the read or watch keeps on the message | Through gabriel's own authenticated transport (rule 125), when the chat is read |
| Discord | Discord's own sound search (/channels/<id>/messages/search?has=sound), one chat at a time through log's Discord engine | From its fresh attachment url, through the same engine; Discord's urls expire, so the search is what gets a fresh one |
Any refusal from a platform (a 429, a 401, anything that isn't an answer) stops that platform's part of the run. Nothing is retried, and what was found before it is kept (rule 94).
AssemblyAI
Each file goes up with POST /v2/upload, its transcript is asked for with POST /v2/transcript (language detected), and it is polled with GET /v2/transcript/<id> every 3 s until it is completed or error, at most 15 minutes. The key is ASSEMBLYAI_API_KEY or ASSEMBLY_AI_KEY: the environment, .env.local in the repo, ~/.buzz/.env.local or ~/filmcraft/.env.local. It is never printed. Without it log transcribe refuses with assemblyAiKeyMissing before it reads anything.
Each transcript is kept once by the sha256 of its file in ~/.logCli/transcripts/<hash>.json, and ~/.logCli/transcripts/messages.json ties each message (platform, account, message id, file name) to its file's hash. A file already transcribed is never sent again. Saving a conversation (an export, twitterDms, searchIndex) writes the kept transcripts back in, so a fresh read never loses them.
New voice notes
| Where | What happens |
|---|---|
slackRead | Each audio file is saved through gabriel's transport and transcribed |
discordRead | Each audio attachment is saved through the engine and transcribed, and the message shows (voice note: "…") |
twitterDms with or conversation | Each X Chat audio file the read saved is transcribed in the thread |
| The Discord watch (inbox hook) | A voice message pushed by the gateway is saved and transcribed before it reaches the agent's >>> NEW DISCORD MESSAGE alert |
| The X inbox alert | Shows the kept transcript; one not transcribed yet says so, and reading the chat transcribes it |
Commands
log transcribe --listlog transcribelog transcribe --platform discord --account sourcedandsoldtothehighestbidderlog transcribe file ~/Downloads/note.m4aThe run ends by updating the search index with the conversations that changed (--noIndex skips it). Code: src/media/voiceNotes, src/cli/transcribe. Spec: docs/specs/voiceNotes.md.